Deploy safely without secrets

.gitattributes

@@ -0,0 +1,10 @@
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.gpkg filter=lfs diff=lfs merge=lfs -text
+*.ttl filter=lfs diff=lfs merge=lfs -text
+*.faiss filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.sst filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.csv filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,27 @@
+# Ignore Python environment files (Do NOT upload to Git!)
+*.dll
+*.lib
+*.exe
+*.so
+__pycache__/
+.venv/
+env/
+venv/
+anaconda/
+
+# Ignore massive Knowledge Graph database files (We will host the .ttl on a HF Dataset instead)
+agent_kg_db/
+*.sst
+
+# Ignore massive GIS files and Models (We will download these via Python at runtime)
+*.gpkg
+*.ttl
+*.pth
+
+# Ignore generated outputs
+agent_detection_outputs/
+*.tif
+
+# Ignore secrets
+.env
+*.env

Dockerfile

@@ -0,0 +1,45 @@
+# Use a lightweight Python 3.10 base image
+FROM python:3.10-slim
+
+# Hugging Face Spaces requires running as a non-root user (UID 1000)
+RUN useradd -m -u 1000 user
+
+# Set working directory and change ownership to the non-root user
+WORKDIR /app
+RUN chown user:user /app
+
+# Switch to the non-root user
+USER user
+ENV HOME=/home/user \
+    PATH=/home/user/.local/bin:$PATH \
+    PYTHONPATH=/app
+
+# Install system dependencies (must switch to root temporarily)
+USER root
+RUN apt-get update && apt-get install -y \
+    build-essential \
+    libgdal-dev \
+    git \
+    git-lfs \
+    && rm -rf /var/lib/apt/lists/*
+USER user
+
+# Copy requirements and install via standard pip
+# (Pip handles Linux wheels perfectly, unlike Windows where Conda is needed)
+COPY --chown=user:user requirements.txt .
+RUN pip install --no-cache-dir --user -r requirements.txt
+
+# Install PyTorch Geometric dependencies (Sparse/Scatter) for Linux CPU natively
+RUN pip install --no-cache-dir --user torch-scatter torch-sparse -f https://data.pyg.org/whl/torch-2.2.0+cpu.html
+
+# Copy the rest of the application
+COPY --chown=user:user . .
+
+# Ensure correct permissions for the Pyoxigraph Knowledge Graph storage
+RUN chmod -R 777 /app || true
+
+# Expose port for Hugging Face Spaces
+EXPOSE 7860
+
+# Command to run the FastAPI server
+CMD ["python", "chatbot/chat_server.py", "--host", "0.0.0.0", "--port", "7860"]

README.md

@@ -0,0 +1,52 @@
+# Glacial Lake Outburst Flood (GLOF) Agentic System
+
+An advanced, multimodal AI agent designed to monitor, forecast, and explain Glacial Lake Outburst Flood (GLOF) risks in real-time. This system fuses Deep Learning computer vision, Graph Neural Networks (GNN), XGBoost, and an Agentic Retrieval-Augmented Generation (RAG) backend to act as a complete early warning and analysis system.
+
+## Features
+*   **Agentic RAG Backend**: Processes natural language queries to search scientific literature, traverse a semantic Knowledge Graph, or trigger live predictive models.
+*   **Dual-Inference Forecasting**: 
+    *   **XGBoost**: Predicts GLOF susceptibility based on historical static geomorphology and live 90-day ERA5 weather data.
+    *   **Graph Neural Network (GConvGRU)**: A spatiotemporal physics-informed model that simulates surface runoff and snowmelt cascading through the downstream river network.
+*   **Live Geospatial UI**: A web-based frontend integrating Mapbox and Google Earth Engine (GEE) to visualize glacial lakes, river hierarchies, building infrastructure, and real-time inference results.
+*   **Hybrid Knowledge Graph**: Combines dense vector searching (BAAI/bge-m3) with deterministic SPARQL/Graph traversal (RDFLib) to ensure physically accurate answers without LLM hallucination.
+
+## Setup & Installation
+
+### Prerequisites
+*   [Miniconda](https://docs.anaconda.com/free/miniconda/) or Anaconda installed.
+*   Google Earth Engine account/credentials.
+*   Hugging Face API Token.
+
+### 1. Create the Environment
+Due to complex C++ dependencies for PyTorch Geometric on Windows, use the provided Conda environment file:
+```bash
+conda env create -f environment.yml
+conda activate glof_agent
+```
+
+### 2. Environment Variables
+You must set your Hugging Face API token in your environment (or create a `.env` file):
+```bash
+setx HF_TOKEN "your_hf_token_here"
+```
+
+### 3. Run the Application
+You can start both the backend FastAPI server and the frontend web interface simply by running:
+```bash
+run_application.bat
+```
+*(Alternatively, you can run `python chatbot/chat_server.py` and serve the `website/` folder manually).*
+
+## Repository Structure
+*   `chatbot/`: The core Agentic RAG logic, routing, Earth Engine fetching, and dual-inference scripts.
+*   `models/`: Pre-trained weights for the GNN (`.pth`), Lake Segmentation U-Net (`.pth`), and XGBoost (`.joblib`).
+*   `gis_data/`: Geopackages (`.gpkg`) containing the static river networks, glacial lake boundaries, and building footprints.
+*   `rag/`: Vector indices and the semantic Knowledge Graph (`.ttl`).
+*   `website/`: The HTML/JS/CSS code for the interactive Mapbox frontend.
+
+## Git Large File Storage (LFS)
+This repository uses Git LFS to host heavy model weights and `.gpkg` files. If you clone this repository, make sure you have Git LFS installed to pull the actual files rather than their pointer references:
+```bash
+git lfs install
+git lfs pull
+```

chatbot/.vscode/settings.json

@@ -0,0 +1,4 @@
+{
+    "python-envs.defaultEnvManager": "ms-python.python:conda",
+    "python-envs.defaultPackageManager": "ms-python.python:conda"
+}

chatbot/agentic_rag/agent.py

@@ -0,0 +1,202 @@
+import sys
+import torch # CRITICAL: Import torch before other libraries to fix WinError 1114
+import re
+from pathlib import Path
+import os
+from dotenv import load_dotenv
+
+# Setup paths to import existing modules
+PROJECT_ROOT = Path(r"D:\Rubel\M.Tech\MTP\Phase 2")
+if str(PROJECT_ROOT) not in sys.path:
+    sys.path.insert(0, str(PROJECT_ROOT))
+RAG_ROOT = PROJECT_ROOT / "chatbot" / "rag"
+if str(RAG_ROOT) not in sys.path:
+    sys.path.insert(0, str(RAG_ROOT))
+
+# Load environment variables
+env_path = PROJECT_ROOT / "chatbot" / "rag" / ".env"
+load_dotenv(dotenv_path=env_path)
+
+from chatbot.agentic_rag.tools import TOOLS, get_tool_descriptions
+
+def agent_chat(messages: list[dict[str, str]], *, model: str, max_tokens: int, temperature: float, stop: list[str] | None = None) -> str:
+    """Localized LLM call with support for stop sequences to prevent hallucinations."""
+    from huggingface_hub import InferenceClient
+    
+    token = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACEHUB_API_TOKEN") or os.getenv("HUGGINGFACE_API_KEY")
+    if not token:
+        raise ValueError("Missing HF token. Put HF_TOKEN=... in .env or your environment.")
+
+    client = InferenceClient(model=model, token=token)
+    
+    # We use stop sequences to force the LLM to stop after generating a tool call
+    response = client.chat_completion(
+        messages=messages,
+        max_tokens=max_tokens,
+        temperature=temperature,
+        stop=stop
+    )
+    
+    try:
+        content = response.choices[0].message.content
+    except AttributeError:
+        content = response["choices"][0]["message"]["content"]
+    
+    return content
+
+SYSTEM_PROMPT = """You are an autonomous Geospatial Reasoning Agent specialized in Glacial Lake Outburst Floods (GLOFs).
+Your goal is to answer complex questions by reasoning step-by-step and calling specific tools to gather evidence.
+
+You have access to the following tools:
+{tool_descriptions}
+
+Use the following format strictly:
+
+Question: the input question you must answer
+Thought: you should always think about what to do next. Decompose the problem if necessary.
+Action: the name of the tool to use. Must be one of: {tool_names}
+Action Input: the exact string argument to pass to the tool
+Observation: the result of the tool
+... (this Thought/Action/Action Input/Observation can repeat N times)
+Thought: I now know the final answer based on the evidence.
+Final Answer: the final answer to the original input question. 
+
+Rules:
+1. For specific lakes or basins, you must ONLY answer using the data returned by the tools. Do not hallucinate data or guess Lake IDs. 
+2. For general conceptual, scientific, or threshold questions (e.g., "What causes GLOFs?", "weather thresholds", "hazard index"), you MUST use BOTH the `query_knowledge_graph` tool AND the `search_scientific_papers` tool to gather comprehensive background context from all available sources before formulating your answer. Do NOT use specific Lake IDs from tool examples as "representative" cases unless specifically asked.
+3. If the user provides a numeric Basin ID or Sub-catchment ID (e.g., '170067682'), you MUST first use `query_lakes_by_subcatchment` or `execute_python_code` to find the actual lakes in that basin. NEVER guess which lake belongs to a basin.
+4. If a tool fails or returns no data, try a different approach or tool.
+5. If the user asks a multi-step question, you must first find the lake properties, then find its risk, etc.
+6. Your output MUST contain "Action:" and "Action Input:" if you want to use a tool.
+7. ALWAYS STOP writing after you output "Action Input:". Wait for the Observation.
+8. HYDROLOGICAL REASONING & CASCADING GLOFS: Water flows downhill. If asked which other lakes are affected by a GLOF from a source lake, first find the source lake's sub-catchment using `query_downstream_rivers_kg`. Then use `query_lakes_by_subcatchment` which will return all nearby lakes AND their elevations. A lake can only be flooded if its elevation is LOWER than the source lake. Just compare the elevations returned by the tool directly. Do NOT write Python code for this!
+9. FLOOD SCENARIOS: If the user asks about the impact of a lake breach, buildings at risk, or "what happens if X lake fails", ALWAYS use the `simulate_glof_flood_impact` tool.
+10. SATELLITE DETECTION: ONLY use the `detect_lakes_and_calculate_area` tool IF the user specifically requests "new satellite detection", "area trend analysis", or "calculate area from recent imagery". This tool is SLOW. You must pass 4 comma-separated arguments: 'lake_id, mode, start, end'. 'mode' must be 'yearly' (e.g. 2021) or 'monthly' (e.g. 2024-06). Example: "GL087099E27798N, monthly, 2024-06, 2025-01".
+11. CONCEPTUAL QUESTIONS: Do not pick a random Lake ID from tool descriptions (like GL088561E28014N) to illustrate a general point. Use general scientific principles or findings from `search_scientific_papers`.
+12. COORDINATE QUERIES: If the user provides lat/lon coordinates (e.g., "near 28.98 lat, 88.70 lon") and asks for averages, counts, or properties, you MUST use `execute_python_code` to buffer the point and query the GeoPackage. Do NOT say you cannot do it.
+13. NAMED RIVER BASINS (Ganga, Indus, Brahmaputra, etc.): The GeoPackage does NOT contain basin names, only numeric Sub-catchment IDs. For questions about distributions, statistics, or properties within a named basin (e.g., "Elevation distribution in the Indus basin"), you MUST first use `search_scientific_papers`. These distributions are often pre-calculated in the scientific literature. If you need to perform custom analysis, use the papers to find which numeric Sub-catchment IDs belong to that basin first.
+14. DATA INTEGRITY: NEVER manually type out or hardcode long lists of numbers, coordinates, or Lake IDs into a Python script (e.g., `elevations = [4239, 4201, ...]`). This is prone to error and truncation. Instead, use Python to query the GeoPackage file directly (`gpd.read_file`) and filter the data using IDs or ranges found in previous steps.
+15. NAMED BASIN PRIORITIZATION: For questions about distributions or statistics in named basins/sub-basins (e.g., 'Indus Middle', 'Upper Ganga'), do NOT waste steps looking for numeric Sub-catchment IDs. Instead, use `search_scientific_papers` to find the pre-calculated statistics or tables directly. Only use `execute_python_code` if the papers fail to provide the summary or if the user asks for very specific lake-level calculations.
+16. SUMMARY TABLE PREFERENCE: When using `search_scientific_papers`, look specifically for keywords like "Summary", "Distribution", "Table X", or "Statistical Analysis". Distributions (like elevation ranges) are ALWAYS provided as summary tables in the reports. NEVER attempt to manually count or extract individual lake elevations from long Annexure lists if a summary table exists.
+17. DATA MISIDENTIFICATION WARNING: Be extremely careful with column headers. Do NOT confuse Serial Numbers, Lake IDs, or numeric codes with physical properties like Elevation, Area, or Volume. Always verify which column contains the actual physical value.
+18. Once you have enough information, output "Final Answer:" followed by your final synthesized response.
+
+"""
+
+def run_reasoning_agent(question: str, max_steps: int = 8):
+    tool_names = list(TOOLS.keys())
+    tool_descriptions = get_tool_descriptions()
+    
+    prompt = SYSTEM_PROMPT.format(
+        tool_descriptions=tool_descriptions,
+        tool_names=", ".join(tool_names)
+    )
+    
+    messages = [{"role": "system", "content": prompt}]
+    messages.append({"role": "user", "content": f"Question: {question}"})
+    
+    print("="*60)
+    print(f"[Agent Started] Question: {question}")
+    print("="*60)
+    
+    for step in range(max_steps):
+        # We tell the model to STOP after it provides the Action Input.
+        # This prevents it from hallucinating the "Observation:" field.
+        try:
+            response = agent_chat(
+                messages, 
+                model="Qwen/Qwen2.5-72B-Instruct", 
+                max_tokens=1024, 
+                temperature=0.1,
+                stop=["Observation:", "Observation"]
+            )
+        except Exception as e:
+            print(f"LLM Error: {e}")
+            break
+        
+        # We append the LLM's response to the conversation history
+        messages.append({"role": "assistant", "content": response})
+        print(f"\n{response}\n")
+        
+        # Check if the agent reached the final answer
+        if "Final Answer:" in response:
+            print("="*60)
+            print("[Agent Finished Successfully]")
+            print("="*60)
+            return
+            
+        # Parse Tool Call using Regex. Use re.DOTALL to capture multi-line code blocks
+        action_match = re.search(r"Action:\s*(.*?)(?:\n|$)", response)
+        action_input_match = re.search(r"Action Input:\s*(.*?)(?:\n(?=Observation:)|$)", response, re.DOTALL)
+        
+        if action_match and action_input_match:
+            action_name = action_match.group(1).strip()
+            action_input = action_input_match.group(1).strip()
+            
+            # Clean up YAML pipe or markdown blocks often added by LLMs
+            if action_input.startswith("|"):
+                action_input = action_input[1:].strip()
+            
+            # Remove ```python, ```py, or just ``` at the start
+            action_input = re.sub(r"^```[a-zA-Z]*\n?", "", action_input)
+            
+            if action_input.endswith("```"):
+                action_input = action_input[:-3]
+            
+            action_input = action_input.strip()
+            
+            # Strip triple double-quotes if the LLM wraps the code in them
+            if action_input.startswith('"""') and action_input.endswith('"""'):
+                action_input = action_input[3:-3].strip()
+                
+            # Strip quotes if the LLM added them (only for single-line strings)
+            if "\n" not in action_input and action_input.startswith(("'", '"')) and action_input.endswith(("'", '"')):
+                action_input = action_input[1:-1]
+                
+            if action_name in TOOLS:
+                print(f"\n[EXECUTING TOOL]: {action_name}")
+                print(f"--- Input ---\n{action_input}\n-------------")
+                try:
+                    observation = TOOLS[action_name](action_input)
+                except Exception as e:
+                    observation = f"Tool Execution Error: {str(e)}"
+            else:
+                observation = f"Error: {action_name} is not a valid tool. Choose from: {', '.join(tool_names)}"
+                
+            # Truncate print output for observation so console isn't flooded
+            print_obs = observation if len(observation) < 300 else observation[:300] + "... [truncated]"
+            print(f"[OBSERVATION]: {print_obs}")
+            
+            # Append REAL Observation to prompt so LLM can read it
+            messages.append({"role": "user", "content": f"Observation: {observation}"})
+        else:
+            # If the LLM format is broken but no final answer
+            warning_msg = "Format Error: You must output 'Action:' and 'Action Input:' or 'Final Answer:'."
+            print(f"[WARNING]: {warning_msg}")
+            messages.append({"role": "user", "content": warning_msg})
+
+    print("="*60)
+    print("[Agent Stopped] Reached maximum steps without a final answer.")
+    print("="*60)
+
+if __name__ == "__main__":
+    print("\n" + "="*60)
+    print("Welcome to the Geospatial Reasoning Agent (GLOF Expert)")
+    print("Type 'exit' or 'quit' to end the session.")
+    print("="*60 + "\n")
+    
+    while True:
+        try:
+            user_question = input("\nYour Question: ").strip()
+            if user_question.lower() in ['exit', 'quit']:
+                print("Exiting agent. Goodbye!")
+                break
+            if not user_question:
+                continue
+                
+            run_reasoning_agent(user_question)
+        except KeyboardInterrupt:
+            print("\nExiting agent. Goodbye!")
+            break
+        except Exception as e:
+            print(f"\nAn error occurred: {e}")

chatbot/agentic_rag/tools.py

@@ -0,0 +1,681 @@
+import sys
+import torch # CRITICAL: Import torch first to prevent WinError 1114 on Windows
+import os
+import math
+from datetime import date, datetime
+from pathlib import Path
+import warnings
+import contextlib
+import io
+import re
+
+# Suppress noisy warnings
+warnings.filterwarnings("ignore")
+
+# Heavy GIS/ML Imports (Top-level to prevent DLL conflicts on Windows)
+try:
+    import numpy as np
+    import pandas as pd
+    import geopandas as gpd
+    import rasterio
+    import fiona
+    import pyproj
+    import ee
+    import geemap
+    import matplotlib.pyplot as plt
+    import segmentation_models_pytorch as smp
+    from rasterio.features import rasterize
+    from scipy.ndimage import label
+    from shapely.geometry import box, Point
+    from shapely.ops import transform as shapely_transform
+except Exception as e:
+    print(f"Warning: Failed to load some GIS/ML libraries: {e}")
+
+# Setup paths to import existing modules
+DEPLOYMENT_ROOT = Path(__file__).resolve().parent.parent.parent
+PROJECT_ROOT = DEPLOYMENT_ROOT
+if str(PROJECT_ROOT) not in sys.path:
+    sys.path.insert(0, str(PROJECT_ROOT))
+RAG_ROOT = PROJECT_ROOT / "chatbot" / "rag"
+if str(RAG_ROOT) not in sys.path:
+    sys.path.insert(0, str(RAG_ROOT))
+
+from chatbot.rag.kg_store import KgStore
+from chatbot.rag.hybrid_graphrag import retrieve_vector_context
+# The import block handles potential missing tools gracefully
+try:
+    from chatbot.dual_inference import explain_glof_prediction
+except ImportError:
+    explain_glof_prediction = None
+
+KG_STORE_DIR = PROJECT_ROOT / "chatbot" / "rag" / "kg_store"
+INDEX_DIR = PROJECT_ROOT / "chatbot" / "rag" / "index"
+
+# Configuration for GIS Data
+DATA_DIR = PROJECT_ROOT / "gis_data"
+LAKES_PATH = DATA_DIR / "stage2_dataset_15.gpkg"
+RIVERS_PATH = DATA_DIR / "river_network.gpkg"
+BUILDINGS_PATH = DATA_DIR / "final_merged_clipped_buildings.gpkg"
+
+# --- Detection Tool Constants ---
+CKPT_PATH = PROJECT_ROOT / "models" / "lake_segmentation" / "finetuned_model_best.pth"
+DETECTION_OUT_DIR = PROJECT_ROOT / "agent_detection_outputs"
+
+# Normalization stats for Detection model
+S2_MEAN = torch.tensor([857.6, 1044.0, 1356.7, 1574.4, 1786.2, 2076.5, 2215.2, 2277.8, 2348.5, 2243.7, 2665.0, 2217.2]).view(12, 1, 1)
+S2_STD  = torch.tensor([626.3, 709.2, 748.7, 825.5, 773.5, 786.1, 810.8, 834.9, 833.6, 729.6, 875.7, 851.3]).view(12, 1, 1)
+S1_MEAN = torch.tensor([-9.25, -18.00]).view(2, 1, 1)
+S1_STD  = torch.tensor([5.90, 5.92]).view(2, 1, 1)
+DEM_MEAN = torch.tensor([4299.8]).view(1, 1, 1)
+DEM_STD  = torch.tensor([901.0]).view(1, 1, 1)
+
+# We create a persistent global dictionary so variables persist between python executions
+GLOBAL_PYTHON_STATE = {}
+
+def execute_python_code(code: str) -> str:
+    """
+    Use this tool to execute arbitrary Python code for spatial reasoning and data analysis.
+    SCHEMA GUIDE:
+    - Primary dataset: r'D:\\Rubel\\M.Tech\\MTP\\Phase 2\\model_data\\Final data\\stage2_dataset_15.gpkg'
+    - Dataset CRS: ESRI:102025 (Asia North Albers, units=meters).
+    - Elevation Columns: Use 'Elevation' or 'surface_elevation_m' for lake elevation.
+    - LAKE TYPE MAPPING: Use 'lake_type' column with these codes:
+      * M(e): End-moraine Dammed
+      * M(l): Lateral Moraine Dammed
+      * M(lg): Lateral Moraine Dammed (with ice)
+      * M(o): Other Moraine Dammed
+      * O: Other Glacial Lake
+      * E(c): Cirque Erosion
+      * E(o): Other Glacial Erosion
+      * E(v): Glacier Trough Valley Erosion
+    - MULTI-LAKE QUERY (CRITICAL FOR CASCADING GLOFS): To compare elevations of multiple lakes, pass their GLAKE_IDs in a list.
+      Example:
+      ```python
+      import geopandas as gpd
+      lakes = gpd.read_file(r'D:\\Rubel\\M.Tech\\MTP\\Phase 2\\model_data\\Final data\\stage2_dataset_15.gpkg')
+      target_ids = ['GL088814E27991N', 'GL088743E27992N']
+      subset = lakes[lakes['GLAKE_ID'].isin(target_ids)]
+      print(subset[['GLAKE_ID', 'Elevation']].to_string(index=False))
+      ```
+    - COORDINATE QUERIES (CRITICAL): User provides Lat/Lon (WGS84). You MUST re-project the point to the dataset CRS (ESRI:102025) before calculating distances.
+      Example:
+      ```python
+      import geopandas as gpd
+      from shapely.geometry import Point
+      lakes = gpd.read_file(r'D:\\Rubel\\M.Tech\\MTP\\Phase 2\\model_data\\Final data\\stage2_dataset_15.gpkg')
+      # 1. Create point in WGS84 (EPSG:4326), then re-project to dataset's projected CRS (ESRI:102025)
+      user_pt = gpd.GeoDataFrame([{'geometry': Point(88.70, 28.98)}], crs="EPSG:4326").to_crs(lakes.crs).geometry.iloc[0]
+      # 2. Query within 10km (10000 meters)
+      nearby = lakes[lakes.geometry.distance(user_pt) < 10000]
+      if not nearby.empty:
+          avg_elev = nearby['Elevation'].mean()
+          print(f"Found {len(nearby)} lakes. Average Elevation: {avg_elev:.2f}m")
+      else:
+          print("No lakes found within 10km of coordinates.")
+      ```
+    Input: A valid, executable Python script. Use print() for results.
+    """
+    # Clean up markdown formatting if the LLM adds it
+    code = code.strip()
+    if code.startswith("```python"):
+        code = code[9:]
+    elif code.startswith("```"):
+        code = code[3:]
+    if code.endswith("```"):
+        code = code[:-3]
+    
+    # Capture standard output
+    output_buffer = io.StringIO()
+    try:
+        with contextlib.redirect_stdout(output_buffer):
+            # Execute the code in the persistent global state
+            exec(code, GLOBAL_PYTHON_STATE)
+        output = output_buffer.getvalue()
+        if not output.strip():
+            return "Code executed successfully, but nothing was printed. Use print() to output results."
+        return output
+    except Exception as e:
+        import traceback
+        return f"Python Error:\n{traceback.format_exc()}"
+
+def get_dataset_info(dummy_input: str = "") -> str:
+    """
+    Returns the schema and summary statistics of the glacial lake dataset.
+    Use this BEFORE writing Python code if you are unsure about column names or unique values.
+    Input: Any string (ignored).
+    """
+    try:
+        lakes = gpd.read_file(LAKES_PATH)
+        info = [f"Dataset Schema (Top 20 columns of {len(lakes.columns)}):", ", ".join(lakes.columns[:20])]
+        info.append("\nSample lake_type values: " + ", ".join(lakes['lake_type'].dropna().unique()[:10]))
+        info.append(f"\nTotal Records: {len(lakes)}")
+        return "\n".join(info)
+    except Exception as e:
+        return f"Error reading schema: {str(e)}"
+
+def query_lakes_by_subcatchment(subcatchment_id: str) -> str:
+    """
+    Finds all glacial lakes belonging to a specific sub-catchment or basin ID.
+    Use this when the user provides a numeric basin ID (e.g., '170067682').
+    Returns a list of Lake IDs and their basic properties (Elevation, CAGR, Area, Type).
+    Input: The numeric sub-catchment ID.
+    """
+    try:
+        # Normalize input
+        sid_str = subcatchment_id.strip().replace("'", "").replace('"', "")
+        sid = float(sid_str)
+        lakes = gpd.read_file(LAKES_PATH)
+        # Handle potential variations in column name
+        col = 'sub-catchment_id' if 'sub-catchment_id' in lakes.columns else 'sub-catchment'
+        in_basin = lakes[lakes[col] == sid]
+        if in_basin.empty:
+            return f"No lakes found in sub-catchment {subcatchment_id}."
+        
+        # Sort by CAGR to help the agent find the "highest"
+        if 'area_cagr_total_pct' in in_basin.columns:
+            in_basin = in_basin.sort_values(by='area_cagr_total_pct', ascending=False)
+            
+        res = [f"Found {len(in_basin)} lake(s) in sub-catchment {subcatchment_id}:"]
+        for _, row in in_basin.head(15).iterrows():
+            cagr = row.get('area_cagr_total_pct', 0.0)
+            elev = row.get('Elevation', row.get('surface_elevation_m', 'N/A'))
+            if isinstance(elev, float): elev = f"{elev:.1f}m"
+            res.append(f"- ID: {row['GLAKE_ID']} | Elev: {elev} | CAGR: {cagr:.4f}% | Type: {row.get('lake_type', 'N/A')}")
+        return "\n".join(res)
+    except Exception as e:
+        return f"Error querying sub-catchment: {str(e)}"
+
+def query_lake_properties_kg(lake_id: str) -> str:
+    """
+    Use this tool to get exact, structured facts about a specific glacial lake from the Knowledge Graph.
+    This includes area, volume, elevation, dam type, and growth trends (CAGR).
+    Input: The exact Lake ID (e.g., 'GL088561E28014N').
+    Returns: A structured string of lake properties.
+    """
+    kg = KgStore(KG_STORE_DIR)
+    props = kg.lake_properties_by_id(lake_id)
+    if not props:
+        return f"Lake ID {lake_id} not found in the Knowledge Graph."
+    return "\n".join(props)
+
+def query_downstream_rivers_kg(lake_id: str) -> str:
+    """
+    Use this tool to perform topological reasoning. It finds which rivers are connected 
+    to a lake and traces the downstream path.
+    Input: The exact Lake ID.
+    Returns: A list of connected rivers and downstream exposure elements.
+    """
+    kg = KgStore(KG_STORE_DIR)
+    res = kg.associated_river_for_lake_text(lake_id)
+    if not res:
+        return f"No downstream river connection found for Lake {lake_id}."
+    return res
+def is_tabular_query(query: str) -> bool:
+    lowered = query.lower()
+    has_table_signal = any(term in lowered for term in ["table", "distribution", "range", "stat", "count", "number of"])
+    has_category_signal = any(term in lowered for term in ["elevation", "area", "type", "moraine", "erosion"])
+    return has_table_signal and has_category_signal
+
+def search_scientific_papers(query: str) -> str:
+    """
+    Use this tool to search ingested PDFs, scientific papers, and literature reports.
+    Use this when you need background knowledge, causes, mechanisms, or historical facts.
+    Returns excerpts and structured Markdown tables from scientific literature.
+    Input: A search query string (e.g., 'Table for elevation distribution in Ganga basin').
+    """
+    # Adjust parameters for tabular/statistical queries to favor BM25 (keywords)
+    top_k = 5
+    alpha = 0.65
+    if is_tabular_query(query):
+        top_k = 10  # Increase depth to catch tables near headers
+        alpha = 0.45 # Favor BM25 (keywords) for structured table lookup
+        print(f"DEBUG: Tabular query detected. Using hybrid alpha={alpha}, top_k={top_k}")
+
+    from chatbot.rag.hybrid_graphrag import retrieve_vector_context
+    context, evidence = retrieve_vector_context(
+        query, 
+        index_dir=INDEX_DIR, 
+        top_k=top_k,
+        dense_top_k=30,
+        bm25_top_k=30,
+        alpha=alpha
+    )
+    if not context:
+        return "No relevant scientific papers found for this query."
+    return context
+
+
+def get_glof_risk_forecast(lake_id: str) -> str:
+    """
+    Use this tool to run live machine learning predictions (XGBoost & GNN) for a glacial lake.
+    It returns the current GLOF susceptibility probability, hazard index, and recent weather anomalies.
+    Input: The exact Lake ID.
+    Returns: A forecast report string.
+    """
+    if not explain_glof_prediction:
+        return "Forecasting model is offline or unavailable."
+    try:
+        explanation = explain_glof_prediction(lake_id)
+        if not explanation:
+             return f"No forecast data available for {lake_id}."
+        return explanation
+    except Exception as e:
+        return f"Error running forecast for {lake_id}: {str(e)}"
+
+def get_regional_glof_forecast(input_str: str) -> str:
+    """
+    Runs a batched GLOF risk forecast for all lakes within a specified radius (buffer) 
+    of a central point (lat/lon). This is much faster than querying lakes individually.
+    Use this when the user asks about risk in a "region", "area", "basin", or "near [coordinates]".
+    Input: A comma-separated string "lat, lon, buffer_km" (e.g., '27.895, 87.013, 10.0').
+    """
+    try:
+        parts = [p.strip() for p in input_str.split(',')]
+        if len(parts) < 2:
+            return "Error: Regional forecast requires at least 'lat, lon'. Example: '27.895, 87.013, 5.0'"
+        
+        lat = float(parts[0])
+        lon = float(parts[1])
+        buffer_km = float(parts[2]) if len(parts) > 2 else 50.0
+        
+        from chatbot.dual_inference import explain_regional_glof
+        return explain_regional_glof(lat, lon, buffer_km)
+    except Exception as e:
+        return f"Error parsing regional forecast input: {str(e)}"
+
+def gaussian_window_2d(size, sigma=None, device="cpu"):
+    import torch
+    if sigma is None: sigma = size / 4.0
+    coords = torch.arange(size, device=device, dtype=torch.float32) - (size - 1) / 2.0
+    g1 = torch.exp(-(coords ** 2) / (2.0 * sigma ** 2))
+    g2 = g1[:, None] * g1[None, :]
+    return g2 / g2.max()
+
+def predict_sliding_window(model, tensor, window=224, stride=64, batch_size=4, device="cuda"):
+    import torch
+    """Runs inference over a large image using an overlapping sliding window."""
+    _, H, W = tensor.shape
+    pad_h = max(0, window - H)
+    pad_w = max(0, window - W)
+    if pad_h or pad_w:
+        tensor = torch.nn.functional.pad(tensor.unsqueeze(0), (0, pad_w, 0, pad_h), mode='reflect').squeeze(0)
+        _, H, W = tensor.shape
+
+    starts_y = list(range(0, H - window + 1, stride))
+    starts_x = list(range(0, W - window + 1, stride))
+    if starts_y[-1] + window < H: starts_y.append(H - window)
+    if starts_x[-1] + window < W: starts_x.append(W - window)
+    coords = [(y, x) for y in starts_y for x in starts_x]
+
+    weights = gaussian_window_2d(window, sigma=window/6.0, device=device)
+    pred_logits = torch.zeros((H, W), device=device, dtype=torch.float32)
+    pred_weights = torch.zeros((H, W), device=device, dtype=torch.float32)
+
+    for offset in range(0, len(coords), batch_size):
+        chunk = coords[offset: offset + batch_size]
+        batch = torch.stack([tensor[:, y:y+window, x:x+window] for y, x in chunk]).to(device)
+        
+        with torch.amp.autocast(device_type="cuda", enabled=(device == "cuda" or str(device) == "cuda")):
+            logits = model(batch)
+            
+        for i, (y, x) in enumerate(chunk):
+            pred_logits[y:y+window, x:x+window] += logits[i, 0].float() * weights
+            pred_weights[y:y+window, x:x+window] += weights
+
+    prob = torch.sigmoid(pred_logits / pred_weights.clamp_min(1e-6)).cpu().numpy()
+    if pad_h or pad_w:
+        prob = prob[:H-pad_h, :W-pad_w]
+    return prob
+
+def detect_lakes_and_calculate_area(input_str: str) -> str:
+    """
+    Perform deep-learning based glacial lake detection and area trend analysis using Sentinel-2, Sentinel-1, and DEM data from GEE.
+    WARNING: This tool is slow. USE ONLY IF requested by the user.
+    Input format: "lake_id, mode, start, end"
+    Modes: 'yearly' or 'monthly'.
+    Example Yearly: 'GL088561E28014N, yearly, 2021, 2025'
+    Example Monthly: 'GL088561E28014N, monthly, 2024-06, 2025-01'
+    """
+    import numpy as np
+    import pandas as pd
+    import torch
+    import segmentation_models_pytorch as smp
+    import ee
+    import geemap
+    import rasterio
+    import matplotlib.pyplot as plt
+    from rasterio.features import rasterize
+    from scipy.ndimage import label
+    from shapely.geometry import box
+    from shapely.ops import transform as shapely_transform
+    import geopandas as gpd
+    import pyproj
+    from datetime import datetime
+    from dateutil.relativedelta import relativedelta
+    from pathlib import Path
+
+    parts = [p.strip() for p in input_str.split(',')]
+    if len(parts) != 4:
+        return "Invalid input format. Use: 'lake_id, mode, start, end'. Example: 'GL088561E28014N, monthly, 2024-06, 2025-01'"
+    
+    lake_id, mode, start_str, end_str = parts[0], parts[1].lower(), parts[2], parts[3]
+
+    try:
+        LAKES_PATH = Path(__file__).resolve().parent.parent.parent / "gis_data" / "stage2_dataset_15.gpkg"
+        DETECTION_OUT_DIR = Path(__file__).resolve().parent.parent.parent / "agent_detection_outputs"
+        
+        # 1. Load Lake Geometry
+        lakes_db = gpd.read_file(LAKES_PATH)
+        target = lakes_db[lakes_db['GLAKE_ID'].str.upper() == lake_id.upper()]
+        if target.empty: return f"Lake {lake_id} not found."
+        
+        target_wgs84 = target.to_crs("EPSG:4326")
+        bbox = list(target_wgs84.total_bounds)
+        bbox = [bbox[0]-0.01, bbox[1]-0.01, bbox[2]+0.01, bbox[3]+0.01]
+        
+        lake_geom = target_wgs84.geometry.iloc[0]
+
+        # 2. Time Periods
+        time_periods = []
+        if mode == 'yearly':
+            s_year, e_year = int(start_str), int(end_str)
+            for y in range(s_year, e_year + 1):
+                time_periods.append({"label": str(y), "start": f"{y}-07-01", "end": f"{y}-10-31"})
+        elif mode == 'monthly':
+            s_date = datetime.strptime(start_str, "%Y-%m")
+            e_date = datetime.strptime(end_str, "%Y-%m")
+            curr_date = s_date
+            while curr_date <= e_date:
+                next_month = curr_date + relativedelta(months=1)
+                end_of_month = next_month - relativedelta(days=1)
+                time_periods.append({
+                    "label": curr_date.strftime("%Y-%m"),
+                    "start": curr_date.strftime("%Y-%m-%d"),
+                    "end": end_of_month.strftime("%Y-%m-%d")
+                })
+                curr_date = next_month
+        else:
+            return "Invalid mode. Use 'yearly' or 'monthly'."
+
+        # 3. GEE Init & Model Load
+        import os
+        import json
+        ee_creds_json = os.environ.get("EE_SERVICE_ACCOUNT_JSON")
+        try:
+            if ee_creds_json:
+                creds_dict = json.loads(ee_creds_json)
+                credentials = ee.ServiceAccountCredentials(creds_dict['client_email'], key_data=ee_creds_json)
+                ee.Initialize(credentials, project='ee-rubelsarkar2') # Project fallback if needed, but best to let ServiceAccount auth work
+            else:
+                ee.Initialize()
+        except Exception:
+            ee.Authenticate()
+            ee.Initialize()
+
+        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        model = smp.Unet(encoder_name="se_resnext50_32x4d", encoder_weights=None, in_channels=15, classes=1, activation=None)
+        
+        ckpt_path = Path(__file__).resolve().parent.parent.parent / "models" / "lake_segmentation" / "finetuned_model_best.pth"
+        if not ckpt_path.exists(): return f"Model weights not found at {ckpt_path}"
+        
+        ckpt = torch.load(ckpt_path, map_location=device)
+        model.load_state_dict(ckpt["state_dict"])
+        model.to(device).eval()
+
+        results = []
+        out_dir = DETECTION_OUT_DIR / lake_id
+        out_dir.mkdir(parents=True, exist_ok=True)
+
+        for period in time_periods:
+            date_str = period['label']
+            aoi = ee.Geometry.BBox(*bbox)
+            
+            s2 = ee.ImageCollection("COPERNICUS/S2_SR_HARMONIZED").filterBounds(aoi).filterDate(period['start'], period['end']).median()
+            s1 = ee.ImageCollection("COPERNICUS/S1_GRD").filterBounds(aoi).filterDate(period['start'], period['end']).select(['VV', 'VH']).median()
+            dem = ee.ImageCollection("COPERNICUS/DEM/GLO30").filterBounds(aoi).select('DEM').mosaic()
+            composite = ee.Image.cat([s2.select(['B1', 'B2', 'B3', 'B4', 'B5', 'B6', 'B7', 'B8', 'B8A', 'B9', 'B11', 'B12']), s1, dem]).clip(aoi).toFloat()
+            
+            tif_path = out_dir / f"{date_str}_comp.tif"
+            import contextlib, io as io2
+            with contextlib.redirect_stdout(io2.StringIO()), contextlib.redirect_stderr(io2.StringIO()):
+                geemap.download_ee_image(composite, str(tif_path), region=aoi, crs="EPSG:32645", scale=10)
+
+            with rasterio.open(tif_path) as ds:
+                arr = ds.read()
+                transform = ds.transform
+                pixel_area = abs(transform.a * transform.e) / 1e6
+                
+                t = torch.from_numpy(np.nan_to_num(arr)).float()
+                
+                S2_MEAN = torch.tensor([857.6, 1044.0, 1356.7, 1574.4, 1786.2, 2076.5, 2215.2, 2277.8, 2348.5, 2243.7, 2665.0, 2217.2]).view(12, 1, 1)
+                S2_STD  = torch.tensor([626.3, 709.2, 748.7, 825.5, 773.5, 786.1, 810.8, 834.9, 833.6, 729.6, 875.7, 851.3]).view(12, 1, 1)
+                S1_MEAN = torch.tensor([-9.25, -18.00]).view(2, 1, 1)
+                S1_STD  = torch.tensor([5.90, 5.92]).view(2, 1, 1)
+                DEM_MEAN = torch.tensor([4299.8]).view(1, 1, 1)
+                DEM_STD  = torch.tensor([901.0]).view(1, 1, 1)
+                
+                t[0:12] = torch.clamp((t[0:12] - S2_MEAN) / S2_STD, -10, 10)
+                t[12:14] = torch.clamp((t[12:14] - S1_MEAN) / S1_STD, -10, 10)
+                t[14:15] = torch.clamp((t[14:15] - DEM_MEAN) / DEM_STD, -10, 10)
+                
+                with torch.no_grad():
+                    prob_mask = predict_sliding_window(model, t, device=device)
+                    
+                binary_mask = prob_mask >= 0.8
+                
+                # Visual Overlay
+                rgb = np.stack([arr[3], arr[2], arr[1]], axis=-1)
+                rgb_img = np.zeros_like(rgb, dtype=np.uint8)
+                for i in range(3):
+                    channel = rgb[:, :, i]
+                    valid = channel[np.isfinite(channel) & (channel > 0)]
+                    if valid.size > 0:
+                        lo, hi = np.percentile(valid, [2, 98])
+                        if hi <= lo: hi = lo + 1.0
+                        rgb_img[:, :, i] = (np.clip((channel - lo) / (hi - lo), 0, 1) * 255).astype(np.uint8)
+                
+                overlay_color = np.zeros((*binary_mask.shape, 3), dtype=np.uint8)
+                overlay_color[binary_mask] = [0, 220, 255]
+                overlay = (0.7 * rgb_img.astype(np.float32) + 0.3 * overlay_color.astype(np.float32)).astype(np.uint8)
+                plt.imsave(str(out_dir / f"{date_str}_prediction.png"), overlay)
+                
+                labeled, num = label(binary_mask)
+                counts = np.bincount(labeled.ravel())
+
+                project = pyproj.Transformer.from_crs("EPSG:4326", ds.crs, always_xy=True).transform
+                geom_proj = shapely_transform(project, lake_geom)
+                zone = rasterize([(geom_proj, 1)], out_shape=binary_mask.shape, transform=transform, fill=0)
+                
+                intersect_ids = np.unique(labeled[zone == 1])
+                intersect_ids = intersect_ids[intersect_ids > 0]
+                area = counts[intersect_ids].sum() * pixel_area
+                results.append({"date": date_str, "area_km2": float(area)})
+
+        summary = [f"--- Satellite Lake Detection Report: {lake_id} ({mode.upper()}) ---"]
+        area_str = " -> ".join([f"{r['date']}: {r['area_km2']:.3f}km2" for r in results])
+        summary.append(f"Trend: {area_str}")
+        summary.append(f"Files saved to: {out_dir.resolve()}")
+        return "\n".join(summary)
+
+    except Exception as e:
+        import traceback
+        return f"Detection Error: {str(e)}\n{traceback.format_exc()}"
+
+def simulate_glof_flood_impact(lake_id: str) -> str:
+
+    """
+    Runs a GIS simulation to estimate GLOF impact.
+    1. Calculates peak discharge using Huggel's empirical formula.
+    2. Traces the flood zone 50km downstream along the river network.
+    3. Identifies specific buildings at risk by fixing CRS projection issues.
+    Input: The exact Lake ID.
+    Returns: A text report of the estimated discharge and count of buildings at risk.
+    """
+    import geopandas as gpd
+    from shapely.geometry import box
+    
+    print(f"Simulating flood impact for {lake_id}...")
+    
+    try:
+        # 1. Load Lake Data
+        lakes = gpd.read_file(LAKES_PATH)
+        target_lake = lakes[lakes['GLAKE_ID'].str.upper() == lake_id.upper()]
+        if target_lake.empty:
+            return f"Lake {lake_id} not found in the spatial database."
+        
+        lake_geom = target_lake.geometry.iloc[0]
+        metric_crs = lakes.crs
+        
+        # 2. Get Volume and Peak Discharge
+        # Prioritize existing volume data
+        if 'volume_m3' in target_lake.columns and not target_lake['volume_m3'].isna().iloc[0]:
+            est_volume = target_lake['volume_m3'].iloc[0]
+            source_note = " (from database)"
+        else:
+            # Fallback to empirical formula using correct area column name
+            lake_area = target_lake['area_m2'].iloc[0] if 'area_m2' in target_lake.columns else 100000
+            est_volume = 0.04 * (lake_area ** 1.2)
+            source_note = " (estimated from area)"
+            
+        # Empirical Peak Discharge (Huggel et al., 2002 approximation)
+        q_max = 0.0048 * (est_volume ** 0.896)
+        
+        # 3. Find Downstream Path (Improved Topological Tracing)
+        rivers = gpd.read_file(RIVERS_PATH)
+        if rivers.crs != metric_crs:
+            rivers = rivers.to_crs(metric_crs)
+            
+        # Better Tracing Logic:
+        # 1. Use KG to find the associated river if possible
+        kg = KgStore(KG_STORE_DIR)
+        link = kg.associated_river_for_lake(lake_id)
+        start_river_id = None
+        if link and link.get('river_id'):
+            try:
+                # Handle cases like "30002024.0" or "30002024"
+                start_river_id = int(float(link['river_id']))
+            except (ValueError, TypeError):
+                pass
+        
+        # 2. If no KG link or ID mismatch, find the nearest river segment
+        if start_river_id is None or start_river_id not in rivers['river_id'].values:
+            # Spatial fallback: Find nearest segment
+            distances = rivers.geometry.distance(lake_geom)
+            nearest_idx = distances.idxmin()
+            start_river_id = rivers.loc[nearest_idx, 'river_id']
+            print(f"Using nearest river segment fallback: {start_river_id}")
+        else:
+            print(f"Using KG-linked river segment: {start_river_id}")
+
+        # 3. Trace downstream using 'downstream_line_ids' topology
+        def trace_downstream(df, start_id, max_dist=50000):
+            path_ids = []
+            curr_id = start_id
+            total_dist = 0
+            # Pre-filter for performance: create a lookup dictionary
+            river_lookup = df.set_index('river_id')[['downstream_line_ids', 'geometry']].to_dict('index')
+            
+            while curr_id in river_lookup and total_dist < max_dist:
+                if curr_id in path_ids: break # Cycle detection
+                path_ids.append(curr_id)
+                row = river_lookup[curr_id]
+                total_dist += row['geometry'].length
+                
+                next_ids_str = str(row['downstream_line_ids'])
+                if not next_ids_str or next_ids_str == 'None' or next_ids_str == '':
+                    break
+                try:
+                    # Take first downstream branch (primary flow path)
+                    curr_id = int(float(next_ids_str.split(',')[0]))
+                except (ValueError, TypeError):
+                    break
+            return path_ids
+
+        path_ids = trace_downstream(rivers, start_river_id)
+        potential_path = rivers[rivers['river_id'].isin(path_ids)]
+        
+        if potential_path.empty:
+            flood_zone = lake_geom.buffer(2000) 
+            path_desc = "Radial 2km hazard zone (no rivers found nearby)."
+        else:
+            # Create a 1km hazard zone along connected rivers
+            union_rivers = potential_path.union_all() if hasattr(potential_path, 'union_all') else potential_path.unary_union
+            flood_zone = union_rivers.buffer(1000)
+            path_desc = f"Downstream path along {len(potential_path)} topological river segments (1km hazard buffer)."
+            
+        # 4. Identify Buildings at Risk (CRITICAL CRS FIX)
+        # Bbox for read_file MUST match the file's CRS (EPSG:4326)
+        flood_zone_4326 = gpd.GeoSeries([flood_zone], crs=metric_crs).to_crs("EPSG:4326").iloc[0]
+        bbox = flood_zone_4326.bounds
+        
+        print(f"Loading buildings in bbox: {bbox}")
+        buildings = gpd.read_file(BUILDINGS_PATH, bbox=bbox)
+        
+        if buildings.empty:
+            return f"--- GLOF IMPACT SIMULATION REPORT: {lake_id} ---\nEstimated Qmax: {q_max:.2f} m3/s\nNote: No buildings found in the mapped {path_desc}"
+            
+        if buildings.crs != metric_crs:
+            buildings = buildings.to_crs(metric_crs)
+            
+        at_risk = buildings[buildings.geometry.intersects(flood_zone)]
+        
+        report = [
+            f"--- GLOF IMPACT SIMULATION REPORT: {lake_id} ---",
+            f"Estimated Peak Discharge (Qmax): {q_max:.2f} m3/s",
+            f"Estimated Lake Volume: {est_volume:,.0f} m3",
+            f"Hazard Path: {path_desc}",
+            f"Total Buildings at Risk: {len(at_risk)} structure(s).",
+        ]
+        
+        if not at_risk.empty:
+            if 'type' in at_risk.columns:
+                types = at_risk['type'].fillna('unknown').value_counts().to_dict()
+                type_str = ", ".join([f"{count} {t}" for t, count in types.items()])
+                report.append(f"Inundated Structures: {type_str}")
+            
+            # Distance to nearest building
+            min_dist = at_risk.geometry.distance(lake_geom).min()
+            report.append(f"Nearest structure is {min_dist/1000:.2f} km downstream.")
+        
+        return "\n".join(report)
+
+    except Exception as e:
+        import traceback
+        return f"Simulation Error: {str(e)}\n{traceback.format_exc()}"
+
+def query_knowledge_graph(query: str) -> str:
+    """
+    Use this tool to search the semantic Knowledge Graph for definitions, concepts, triggers, risks, and thresholds.
+    Use this when the user asks about "thresholds", "causes", "hazard index", "weather variables", or ontological classes.
+    Input: The search query string.
+    """
+    try:
+        kg = KgStore(KG_STORE_DIR)
+        context = kg.context_for_question(query)
+        if not context:
+            return "No relevant information found in the Knowledge Graph for this query."
+        return context
+    except Exception as e:
+        return f"KG Error: {str(e)}"
+
+# Dictionary mapping tool names to functions for the Agent loop
+TOOLS = {
+    "get_dataset_info": get_dataset_info,
+    "execute_python_code": execute_python_code,
+    "query_lakes_by_subcatchment": query_lakes_by_subcatchment,
+    "query_lake_properties_kg": query_lake_properties_kg,
+    "query_downstream_rivers_kg": query_downstream_rivers_kg,
+    "query_knowledge_graph": query_knowledge_graph,
+    "search_scientific_papers": search_scientific_papers,
+    "get_glof_risk_forecast": get_glof_risk_forecast,
+    "get_regional_glof_forecast": get_regional_glof_forecast,
+    "simulate_glof_flood_impact": simulate_glof_flood_impact,
+    "detect_lakes_and_calculate_area": detect_lakes_and_calculate_area
+}
+
+def get_tool_descriptions() -> str:
+    """Helper to inject tool descriptions into the prompt."""
+    desc = []
+    for name, func in TOOLS.items():
+        desc.append(f"Tool Name: {name}\nDescription: {func.__doc__.strip()}\n")
+    return "\n---\n".join(desc)

chatbot/chat_server.py

@@ -0,0 +1,185 @@
+from __future__ import annotations
+
+import os
+
+import argparse
+import contextlib
+import io
+import json
+import sys
+import traceback
+from functools import partial
+from http.server import SimpleHTTPRequestHandler, ThreadingHTTPServer
+from pathlib import Path
+from urllib.parse import urlparse
+
+
+PROJECT_ROOT = Path(__file__).resolve().parent.parent
+WEBSITE_DIR = PROJECT_ROOT / "website"
+RAG_DIR = PROJECT_ROOT / "chatbot" / "rag"
+
+def download_heavy_files():
+    """Downloads the massive 4GB files from the Hugging Face Dataset at startup."""
+    try:
+        from huggingface_hub import hf_hub_download
+        print("Checking for heavy model weights and GIS databases...")
+        
+        # Dictionary of destination paths to HF Dataset filenames
+        files_to_download = {
+            PROJECT_ROOT / "gis_data" / "final_merged_clipped_buildings.gpkg": "final_merged_clipped_buildings.gpkg",
+            PROJECT_ROOT / "chatbot" / "graphrag" / "glof_kg_populated_codex.ttl": "glof_kg_populated_codex.ttl",
+            PROJECT_ROOT / "models" / "lake_segmentation" / "finetuned_model_best.pth": "finetuned_model_best.pth",
+            PROJECT_ROOT / "models" / "gnn" / "gnn" / "glof_gnn_weights_final.pth": "glof_gnn_weights_final.pth"
+        }
+        
+        for local_path, hf_filename in files_to_download.items():
+            if not local_path.exists():
+                print(f"Downloading {hf_filename} from Hugging Face Dataset...")
+                local_path.parent.mkdir(parents=True, exist_ok=True)
+                downloaded_path = hf_hub_download(
+                    repo_id="freate16/glof-database", 
+                    repo_type="dataset", 
+                    filename=hf_filename,
+                    local_dir=str(local_path.parent),
+                    local_dir_use_symlinks=False
+                )
+                print(f"Successfully downloaded to {downloaded_path}")
+    except Exception as e:
+        print(f"Warning: Could not download heavy files (this is normal if you haven't uploaded them yet): {e}")
+
+# Run the download check before starting the server
+download_heavy_files()
+
+if str(RAG_DIR) not in sys.path:
+    sys.path.insert(0, str(RAG_DIR))
+if str(PROJECT_ROOT) not in sys.path:
+    sys.path.insert(0, str(PROJECT_ROOT))
+
+
+class ChatHandler(SimpleHTTPRequestHandler):
+    server_version = "GLOFChatServer/1.0"
+
+    def end_headers(self) -> None:
+        self.send_header("Access-Control-Allow-Origin", "*")
+        self.send_header("Access-Control-Allow-Headers", "Content-Type")
+        self.send_header("Access-Control-Allow-Methods", "GET, POST, OPTIONS")
+        super().end_headers()
+
+    def do_OPTIONS(self) -> None:
+        self.send_response(204)
+        self.end_headers()
+
+    def do_GET(self) -> None:
+        path = urlparse(self.path).path
+        if path == "/api/health":
+            self._send_json({"ok": True, "service": "glof-chatbot"})
+            return
+        super().do_GET()
+
+    def do_POST(self) -> None:
+        path = urlparse(self.path).path
+        if path != "/api/chat":
+            self.send_error(404, "Unknown endpoint")
+            return
+
+        try:
+            payload = self._read_json()
+            question = str(payload.get("question", "")).strip()
+            route = str(payload.get("route", "auto")).strip() or "auto"
+            if not question:
+                self._send_json({"error": "Question is required."}, status=400)
+                return
+
+            response = self._run_chatbot(question, route)
+            self._send_json(response)
+        except SystemExit as exc:
+            self._send_json({"error": str(exc)}, status=500)
+        except Exception as exc:
+            self._send_json(
+                {
+                    "error": str(exc),
+                    "traceback": traceback.format_exc(),
+                },
+                status=500,
+            )
+
+    def _read_json(self) -> dict:
+        length = int(self.headers.get("Content-Length", "0"))
+        if length > 20000:
+            raise ValueError("Request body is too large.")
+        raw = self.rfile.read(length).decode("utf-8")
+        return json.loads(raw or "{}")
+
+    def _run_chatbot(self, question: str, route: str) -> dict:
+        from chatbot.agentic_rag.agent import run_reasoning_agent
+
+        log_buffer = io.StringIO()
+        with contextlib.redirect_stdout(log_buffer):
+            try:
+                run_reasoning_agent(question, max_steps=8)
+            except Exception as e:
+                print(f"Agent crashed: {e}")
+
+        # Extract the final answer and logs
+        output_str = log_buffer.getvalue()
+
+        final_answer = "No final answer reached by the agent."
+        logs = []
+        for line in output_str.splitlines():
+            line = line.strip()
+            if not line: continue
+            if "Final Answer:" not in line:
+                logs.append(line)
+
+        # Robust multi-line extraction
+        if "Final Answer:" in output_str:
+            # Split by Final Answer:, take the second half
+            after_final = output_str.split("Final Answer:", 1)[1]
+            # The agent prints '============================================================'
+            # and '[Agent Finished Successfully]' after the final answer. We split by the first '=' line to trim it.
+            final_answer = after_final.split("=======")[0].strip()
+
+        return {
+            "answer": final_answer,
+            "route": "agentic",
+            "has_kg_context": True,
+            "logs": logs[-40:]
+        }
+
+    def _send_json(self, payload: dict, status: int = 200) -> None:
+        body = json.dumps(payload, ensure_ascii=False, indent=2).encode("utf-8")
+        self.send_response(status)
+        self.send_header("Content-Type", "application/json; charset=utf-8")
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        self.wfile.write(body)
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Serve the GLOF website and chatbot API locally.")
+    parser.add_argument("--host", default="127.0.0.1")
+    parser.add_argument("--port", type=int, default=8000)
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    if not WEBSITE_DIR.exists():
+        raise SystemExit(f"Website directory not found: {WEBSITE_DIR}")
+
+    handler = partial(ChatHandler, directory=str(WEBSITE_DIR))
+    server = ThreadingHTTPServer((args.host, args.port), handler)
+    url = f"http://{args.host}:{args.port}"
+    print(f"Serving website: {WEBSITE_DIR}")
+    print(f"Chat API: {url}/api/chat")
+    print(f"Open: {url}/")
+    try:
+        server.serve_forever()
+    except KeyboardInterrupt:
+        print("\nServer stopped.")
+    finally:
+        server.server_close()
+
+
+if __name__ == "__main__":
+    main()

chatbot/dual_inference.py

@@ -0,0 +1,717 @@
+import os
+import sys
+import torch # CRITICAL: Import torch first to prevent WinError 1114 on Windows
+import numpy as np
+import pandas as pd
+import datetime as dt
+import geopandas as gpd
+import joblib
+import duckdb
+import shap
+import logging
+
+# Ensure project root is in sys.path
+SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+PROJECT_ROOT = os.path.dirname(SCRIPT_DIR)
+if PROJECT_ROOT not in sys.path:
+    sys.path.insert(0, PROJECT_ROOT)
+
+# ==========================================
+# Paths
+# ==========================================
+STATIC_DATA_PATH = os.path.join(PROJECT_ROOT, "gis_data", "stage2_dataset_15.gpkg")
+XGB_MODEL_PATH = os.path.join(PROJECT_ROOT, "models", "xgb_best_results", "best_f2_xgb_model.joblib")
+
+# ==========================================
+# Unified Data Fetch & Imputation
+# ==========================================
+def impute_raw_window(raw_window: np.ndarray) -> np.ndarray:
+    """
+    Scans the (T, N, F) window fetched by Earth Engine. The last 5 days are usually
+    NaN due to the lag. We fill them using the rolling mean of the 7 days prior.
+    """
+    T, N, F = raw_window.shape
+    imputed_window = raw_window.copy()
+
+    for n in range(N):
+        for f in range(F):
+            series = imputed_window[:, n, f]
+            valid_idx = np.where(~np.isnan(series))[0]
+            if len(valid_idx) == 0:
+                imputed_window[:, n, f] = 0.0
+                continue
+
+            last_valid = valid_idx[-1]
+            if last_valid < T - 1:
+                start_hist = max(0, last_valid - 6)
+                hist_mean = np.nanmean(series[start_hist:last_valid+1])
+                imputed_window[last_valid+1:, n, f] = hist_mean
+
+    return imputed_window
+
+def fetch_unified_weather(glake_id: str):
+    """
+    1. Isolates the GNO node and its k-hop subgraph.
+    2. Fetches 90-day ERA5 data (required for XGBoost).
+    3. Imputes the last 5 days lag using rolling means.
+    4. Extracts the Central Node's data as a DataFrame for XGBoost.
+    5. Returns the full tensor array for GNN (sliced to 60 days later), DataFrame for XGBoost, and context.
+    """
+    import torch
+    from gnn.inference import (
+        initialize_earth_engine, load_model_and_graph, load_graph_nodes,
+        build_k_hop_subgraph, fetch_era5_land_daily_for_nodes,
+        NODE_HOPS, FEATURE_COLS
+    )
+
+    print("\n[1/3] 🌍 Initializing Google Earth Engine for Live ERA5 Retrieval...")
+    initialize_earth_engine()
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+    model, edge_index = load_model_and_graph(device)
+    nodes_gdf = load_graph_nodes()
+
+    lake_row = nodes_gdf[nodes_gdf["GLAKE_ID"] == glake_id]
+    if lake_row.empty:
+        raise ValueError(f"No GNN node mapping found for {glake_id}")
+
+    center_node = int(lake_row["node_index"].iloc[0])
+    node_ids, sub_edge_index, center_local = build_k_hop_subgraph(edge_index, center_node, hops=NODE_HOPS)      
+    sub_nodes = nodes_gdf.set_index("node_index").loc[node_ids].reset_index()
+
+    target_date = dt.date.today()
+    # We explicitly fetch 90 days for XGBoost, even though GNN only needs 60.
+    raw_window, all_dates = fetch_era5_land_daily_for_nodes(sub_nodes, target_date, history_days=90)
+
+    print("      >> Imputing recent 5-day lag with 7-day rolling mean...")
+    imputed_window = impute_raw_window(raw_window)
+
+    # Extract only the central lake's sequence to a pandas dataframe for XGBoost
+    center_ts = imputed_window[:, center_local, :]
+    df_records = []
+    # XGBoost needs the first 90 days (history)
+    for t_idx in range(90):
+        row_dict = {"GLAKE_ID": glake_id, "date": all_dates[t_idx]}
+        for f_idx, col_name in enumerate(FEATURE_COLS):
+            row_dict[col_name] = center_ts[t_idx, f_idx]
+        df_records.append(row_dict)
+
+    weather_df = pd.DataFrame(df_records)
+
+    return imputed_window, weather_df, model, edge_index, sub_edge_index, center_local, device, all_dates       
+
+# ==========================================
+# XGBoost Pipeline
+# ==========================================
+def run_xgboost_inference(glake_id: str, weather_seq_df: pd.DataFrame):
+    print(f"\n[2/3] 🌳 Running XGBoost GLOF Context Model for {glake_id}...")
+
+    temp_csv = os.path.join(SCRIPT_DIR, f"temp_{glake_id}_weather.csv")
+    weather_seq_df.to_csv(temp_csv, index=False)
+
+    query = f"""
+    WITH era5 AS (
+        SELECT
+            GLAKE_ID,
+            CAST("date" AS DATE) AS obs_date,
+            total_precipitation_sum,
+            snowmelt_sum,
+            surface_runoff_sum,
+            sub_surface_runoff_sum,
+            snowfall_sum,
+            temperature_2m,
+            dewpoint_temperature_2m,
+            snow_depth_water_equivalent,
+            surface_pressure
+        FROM read_csv_auto('{temp_csv}')
+    ),
+    lake_targets AS (
+        SELECT GLAKE_ID, MAX(obs_date) as event_date
+        FROM era5 GROUP BY GLAKE_ID
+    )
+    SELECT
+        t.GLAKE_ID,
+
+        -- 3d
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 2 DAYS AND t.event_date THEN e.total_precipitation_sum END) AS total_precipitation_sum_3d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 2 DAYS AND t.event_date THEN e.snowmelt_sum END) AS snowmelt_sum_3d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 2 DAYS AND t.event_date THEN e.surface_runoff_sum END) AS surface_runoff_sum_3d,
+
+        -- 7d
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.total_precipitation_sum END) AS total_precipitation_sum_7d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.snowmelt_sum END) AS snowmelt_sum_7d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.surface_runoff_sum END) AS surface_runoff_sum_7d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.sub_surface_runoff_sum END) AS sub_surface_runoff_sum_7d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.snowfall_sum END) AS snowfall_sum_7d,
+        AVG(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.temperature_2m END) AS temperature_2m_mean_7d,
+        AVG(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.dewpoint_temperature_2m END) AS dewpoint_temperature_2m_mean_7d,
+        AVG(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.snow_depth_water_equivalent END) AS snow_depth_water_equivalent_mean_7d,
+        AVG(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.surface_pressure END) AS surface_pressure_mean_7d,
+
+        -- 30d
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 29 DAYS AND t.event_date THEN e.total_precipitation_sum END) AS total_precipitation_sum_30d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 29 DAYS AND t.event_date THEN e.snowmelt_sum END) AS snowmelt_sum_30d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 29 DAYS AND t.event_date THEN e.surface_runoff_sum END) AS surface_runoff_sum_30d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 29 DAYS AND t.event_date THEN e.sub_surface_runoff_sum END) AS sub_surface_runoff_sum_30d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 29 DAYS AND t.event_date THEN e.snowfall_sum END) AS snowfall_sum_30d,
+        AVG(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 29 DAYS AND t.event_date THEN e.temperature_2m END) AS temperature_2m_mean_30d,
+        AVG(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 29 DAYS AND t.event_date THEN e.snow_depth_water_equivalent END) AS snow_depth_water_equivalent_mean_30d,
+
+        -- 90d
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 89 DAYS AND t.event_date THEN e.total_precipitation_sum END) AS total_precipitation_sum_90d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 89 DAYS AND t.event_date THEN e.snowmelt_sum END) AS snowmelt_sum_90d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 89 DAYS AND t.event_date THEN e.sub_surface_runoff_sum END) AS sub_surface_runoff_sum_90d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 89 DAYS AND t.event_date THEN e.snowfall_sum END) AS snowfall_sum_90d,
+        AVG(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 89 DAYS AND t.event_date THEN e.temperature_2m END) AS temperature_2m_mean_90d,
+        AVG(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 89 DAYS AND t.event_date THEN e.snow_depth_water_equivalent END) AS snow_depth_water_equivalent_mean_90d
+    FROM lake_targets t
+    JOIN era5 e ON t.GLAKE_ID = e.GLAKE_ID
+    GROUP BY t.GLAKE_ID
+    """
+    con = duckdb.connect()
+    agg_weather_df = con.execute(query).df()
+    con.close()
+    if os.path.exists(temp_csv): os.remove(temp_csv)
+
+    static_df = pd.DataFrame(gpd.read_file(STATIC_DATA_PATH).drop(columns="geometry"))
+    lake_static = static_df[static_df['GLAKE_ID'].str.upper() == glake_id.upper()].copy()
+    if lake_static.empty:
+        return {"error": f"Static data for {glake_id} not found."}
+
+    xgb_input = pd.merge(lake_static, agg_weather_df, on="GLAKE_ID", how="inner")
+
+    try:
+        xgb_input['precip_anomaly_7d_90d'] = xgb_input['total_precipitation_sum_7d'] / (xgb_input['total_precipitation_sum_90d'] / 90 * 7 + 1e-5)
+        xgb_input['snowmelt_anomaly_7d_30d'] = xgb_input['snowmelt_sum_7d'] / (xgb_input['snowmelt_sum_30d'] / 30 * 7 + 1e-5)
+        xgb_input['temp_spike_7d_over_30d'] = xgb_input['temperature_2m_mean_7d'] - xgb_input['temperature_2m_mean_30d']
+    except Exception as e:
+        pass
+
+    model = joblib.load(XGB_MODEL_PATH)
+    features_used = model.feature_names_in_
+
+    for col in features_used:
+        if col not in xgb_input.columns:
+            xgb_input[col] = 0.0
+
+    X = xgb_input[features_used]
+    prob = float(model.predict_proba(X)[0, 1])
+
+    explainer = shap.TreeExplainer(model)
+    shap_values = explainer.shap_values(X)
+
+    feature_impacts = list(zip(features_used, shap_values[0]))
+    positive_drivers = sorted([f for f in feature_impacts if f[1] > 0], key=lambda x: x[1], reverse=True)[:3]   
+
+    return {
+        "probability": prob,
+        "top_drivers": [f"{feat}: {val:.3f}" for feat, val in positive_drivers],
+        "driver_names": [f[0] for f in positive_drivers]
+    }
+
+
+def run_xgboost_batch(weather_seq_df: pd.DataFrame):
+    """Run XGBoost model for multiple GLAKE_IDs at once using DuckDB aggregations."""
+    print("\n[2/3] 🌳 Running batched XGBoost for multiple lakes...")
+    temp_csv = os.path.join(SCRIPT_DIR, f"temp_batch_weather.csv")
+    weather_seq_df.to_csv(temp_csv, index=False)
+
+    query = f"""
+    WITH era5 AS (
+        SELECT
+            GLAKE_ID,
+            CAST("date" AS DATE) AS obs_date,
+            total_precipitation_sum,
+            snowmelt_sum,
+            surface_runoff_sum,
+            sub_surface_runoff_sum,
+            snowfall_sum,
+            temperature_2m,
+            dewpoint_temperature_2m,
+            snow_depth_water_equivalent,
+            surface_pressure
+        FROM read_csv_auto('{temp_csv}')
+    ),
+    lake_targets AS (
+        SELECT GLAKE_ID, MAX(obs_date) as event_date
+        FROM era5 GROUP BY GLAKE_ID
+    )
+    SELECT
+        t.GLAKE_ID,
+        -- aggregations identical to single-run
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 2 DAYS AND t.event_date THEN e.total_precipitation_sum END) AS total_precipitation_sum_3d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 2 DAYS AND t.event_date THEN e.snowmelt_sum END) AS snowmelt_sum_3d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 2 DAYS AND t.event_date THEN e.surface_runoff_sum END) AS surface_runoff_sum_3d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.total_precipitation_sum END) AS total_precipitation_sum_7d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.snowmelt_sum END) AS snowmelt_sum_7d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.surface_runoff_sum END) AS surface_runoff_sum_7d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.sub_surface_runoff_sum END) AS sub_surface_runoff_sum_7d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.snowfall_sum END) AS snowfall_sum_7d,
+        AVG(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.temperature_2m END) AS temperature_2m_mean_7d,
+        AVG(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.dewpoint_temperature_2m END) AS dewpoint_temperature_2m_mean_7d,
+        AVG(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.snow_depth_water_equivalent END) AS snow_depth_water_equivalent_mean_7d,
+        AVG(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 6 DAYS AND t.event_date THEN e.surface_pressure END) AS surface_pressure_mean_7d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 29 DAYS AND t.event_date THEN e.total_precipitation_sum END) AS total_precipitation_sum_30d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 29 DAYS AND t.event_date THEN e.snowmelt_sum END) AS snowmelt_sum_30d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 29 DAYS AND t.event_date THEN e.surface_runoff_sum END) AS surface_runoff_sum_30d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 29 DAYS AND t.event_date THEN e.sub_surface_runoff_sum END) AS sub_surface_runoff_sum_30d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 29 DAYS AND t.event_date THEN e.snowfall_sum END) AS snowfall_sum_30d,
+        AVG(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 29 DAYS AND t.event_date THEN e.temperature_2m END) AS temperature_2m_mean_30d,
+        AVG(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 29 DAYS AND t.event_date THEN e.snow_depth_water_equivalent END) AS snow_depth_water_equivalent_mean_30d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 89 DAYS AND t.event_date THEN e.total_precipitation_sum END) AS total_precipitation_sum_90d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 89 DAYS AND t.event_date THEN e.snowmelt_sum END) AS snowmelt_sum_90d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 89 DAYS AND t.event_date THEN e.sub_surface_runoff_sum END) AS sub_surface_runoff_sum_90d,
+        SUM(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 89 DAYS AND t.event_date THEN e.snowfall_sum END) AS snowfall_sum_90d,
+        AVG(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 89 DAYS AND t.event_date THEN e.temperature_2m END) AS temperature_2m_mean_90d,
+        AVG(CASE WHEN e.obs_date BETWEEN t.event_date - INTERVAL 89 DAYS AND t.event_date THEN e.snow_depth_water_equivalent END) AS snow_depth_water_equivalent_mean_90d
+    FROM lake_targets t
+    JOIN era5 e ON t.GLAKE_ID = e.GLAKE_ID
+    GROUP BY t.GLAKE_ID
+    """
+    con = duckdb.connect()
+    try:
+        agg_weather_df = con.execute(query).df()
+    finally:
+        con.close()
+    if os.path.exists(temp_csv):
+        os.remove(temp_csv)
+
+    static_df = pd.DataFrame(gpd.read_file(STATIC_DATA_PATH).drop(columns="geometry"))
+    xgb_input = pd.merge(static_df, agg_weather_df, on="GLAKE_ID", how="inner")
+    if xgb_input.empty:
+        return {}
+
+    try:
+        xgb_input['precip_anomaly_7d_90d'] = xgb_input['total_precipitation_sum_7d'] / (xgb_input['total_precipitation_sum_90d'] / 90 * 7 + 1e-5)
+        xgb_input['snowmelt_anomaly_7d_30d'] = xgb_input['snowmelt_sum_7d'] / (xgb_input['snowmelt_sum_30d'] / 30 * 7 + 1e-5)
+        xgb_input['temp_spike_7d_over_30d'] = xgb_input['temperature_2m_mean_7d'] - xgb_input['temperature_2m_mean_30d']
+    except Exception:
+        pass
+
+    model = joblib.load(XGB_MODEL_PATH)
+    features_used = model.feature_names_in_
+
+    for col in features_used:
+        if col not in xgb_input.columns:
+            xgb_input[col] = 0.0
+
+    X = xgb_input[features_used]
+    probs = model.predict_proba(X)[:, 1]
+
+    explainer = shap.TreeExplainer(model)
+    shap_values = explainer.shap_values(X)
+
+    results = {}
+    for i, gid in enumerate(xgb_input['GLAKE_ID'].values):
+        prob = float(probs[i])
+        # shap_values may be list or ndarray depending on SHAP version
+        try:
+            sv = shap_values[i]
+        except Exception:
+            sv = shap_values[1][i]
+        feature_impacts = list(zip(features_used, sv))
+        positive_drivers = sorted([f for f in feature_impacts if f[1] > 0], key=lambda x: x[1], reverse=True)[:3]
+        results[gid] = {
+            "probability": prob,
+            "top_drivers": [f"{feat}: {val:.3f}" for feat, val in positive_drivers],
+            "driver_names": [f[0] for f in positive_drivers]
+        }
+
+    return results
+
+# ==========================================
+# GNN Pipeline (Dynamic Thresholds)
+# ==========================================
+def run_gnn_inference(imputed_window, model, sub_edge_index, center_local, device):
+    import torch
+    from gnn.inference import build_model_input, TARGET_COLS, HISTORY_DAYS
+
+    logging.getLogger(__name__).debug("Running GNN Physical Forecast for center_local=%s on device=%s", center_local, device)
+
+    # 1. Forward pass
+    # imputed_window is 93 days (90 hist + 3 forecast placeholder).
+    # GNN expects 60 days history. So we take the LAST 60 days of history.
+    # HISTORY_DAYS is 60 (imported from gnn.inference).
+    
+    # Slice to take the last 60 days of history + the forecast days
+    start_slice = 90 - HISTORY_DAYS # 90 - 60 = 30
+    gnn_input_window = imputed_window[start_slice:, :, :]
+    
+    x_tensor, _ = build_model_input(gnn_input_window)
+    x_tensor = x_tensor.to(device)
+    sub_edge_index = sub_edge_index.to(device)
+    with torch.no_grad():
+        pred = model(x_tensor, sub_edge_index).squeeze(0).cpu().numpy()
+
+    # FIX: Clip negative predictions to 0.0 for physical consistency
+    pred = np.maximum(pred, 0.0)
+
+    center_pred = pred[center_local]
+    # 2. Dynamic Threshold Logic (use surface runoff historical stats)
+    runoff_idx = TARGET_COLS.index("surface_runoff_sum")
+    from gnn.inference import FEATURE_COLS
+    feature_runoff_idx = FEATURE_COLS.index("surface_runoff_sum")
+
+    # Use the same sliced window for historical stats
+    historical_runoff = gnn_input_window[:HISTORY_DAYS, center_local, feature_runoff_idx]
+    hist_mean = np.mean(historical_runoff)
+    hist_std = np.std(historical_runoff) + 1e-8
+    dynamic_threshold = hist_mean + (3.0 * hist_std)
+
+    # 3. Assess the Forecasts: compute mean across forecast steps for each target
+    forecasts = {}
+    for i, target_name in enumerate(TARGET_COLS):
+        # center_pred shape: (out_steps, out_vars)
+        try:
+            target_vals = center_pred[:, i]
+        except Exception:
+            target_vals = center_pred[:, i] if center_pred.ndim >= 2 else np.array([center_pred[i]])
+        forecasts[f"{target_name}_mean"] = float(np.mean(target_vals))
+        forecasts[f"{target_name}_max"] = float(np.max(target_vals))
+
+    forecasted_runoff_mean = forecasts.get("surface_runoff_sum_mean", 0.0)
+    risk_level = "HIGH" if forecasted_runoff_mean > dynamic_threshold else "LOW"
+
+    out = {
+        "risk_level": risk_level,
+        "dynamic_threshold": dynamic_threshold,
+        "historical_mean": hist_mean,
+        "historical_std": hist_std,
+    }
+    out.update(forecasts)
+    return out
+
+# ==========================================
+# Unified Runner
+# ==========================================
+def explain_glof_prediction(glake_id: str):
+    """
+    Executes BOTH models sequentially and returns natural language evidence.
+    """
+    try:
+        # Step 1: Shared Live Fetch via GEE
+        (imputed_window, weather_seq_df,
+         gnn_model, edge_index, sub_edge_index,
+         center_local, device, all_dates) = fetch_unified_weather(glake_id)
+
+        # Step 2: XGBoost
+        xgb_res = run_xgboost_inference(glake_id, weather_seq_df)
+        if "error" in xgb_res:
+            return f"Error running models: {xgb_res.get('error')}"
+
+        # Step 3: GNN
+        print("\n[3/3] 🌀 Running GNN Physical Forecast Model...", flush=True)
+        gnn_res = run_gnn_inference(imputed_window, gnn_model, sub_edge_index, center_local, device)
+
+    except Exception as e:
+        import traceback
+        traceback.print_exc()
+        return f"Error running dual inference models for {glake_id}: {str(e)}"
+
+    prob_pct = xgb_res['probability'] * 100
+    top_drivers = ", ".join(xgb_res.get("driver_names", []))
+
+    risk = gnn_res.get('risk_level', "UNKNOWN")
+    dyn_threshold = gnn_res.get('dynamic_threshold', 0)
+    surf_mean = gnn_res.get('surface_runoff_sum_mean')
+    subsurf_mean = gnn_res.get('sub_surface_runoff_sum_mean')
+    snowmelt_mean = gnn_res.get('snowmelt_sum_mean')
+
+    explanation = (
+        f"XGBoost Prediction: {prob_pct:.1f}% probability of GLOF. This is the primary signal.\n"
+        f"Primary Driving Features (SHAP): {top_drivers}\n\n"
+        f"GNN Physical Summary: the model predicts surface runoff {surf_mean:.5f} m, sub-surface runoff {subsurf_mean:.5f} m, and snowmelt {snowmelt_mean:.5f} m over the next 3 days. "
+        f"Overall GNN trigger: {risk} (threshold={dyn_threshold:.5f}, based on 3x the 60-day runoff std).\n\n"  
+        f"This lake looks low-risk from the XGBoost score, but the GNN suggests stronger runoff activity, so it is worth watching rather than treating as an immediate alarm. If both models are high, that would be a stronger warning sign."
+    )
+
+    return explanation
+
+def explain_regional_glof(lat: float, lon: float, buffer_km: float):
+    """
+    Finds all lakes within a radius of (lat, lon) and runs the dual-inference pipeline.
+    """
+    from shapely.geometry import Point
+    # Import GNN helpers
+    from gnn.inference import (
+        initialize_earth_engine,
+        fetch_era5_land_daily_for_nodes,
+        load_model_and_graph,
+        load_graph_nodes,
+        build_k_hop_subgraph,
+        NODE_HOPS,
+        FEATURE_COLS,
+        HISTORY_DAYS,
+    )
+
+    # Load full static geodata and graph node mapping
+    if not os.path.exists(STATIC_DATA_PATH):
+        return "Error: Lake geometry database not found."
+
+    lakes_gdf = gpd.read_file(STATIC_DATA_PATH)
+    metric_crs = lakes_gdf.crs
+
+    pt = Point(lon, lat)
+    pt_gdf = gpd.GeoDataFrame(geometry=[pt], crs="EPSG:4326").to_crs(metric_crs)
+    pt_metric = pt_gdf.geometry.iloc[0]
+
+    buffered_area = pt_metric.buffer(buffer_km * 1000.0)
+    regional_lakes = lakes_gdf[lakes_gdf.geometry.intersects(buffered_area)]
+
+    if regional_lakes.empty:
+        return f"No lakes found within {buffer_km}km of {lat}, {lon}."
+
+    regional_lakes = regional_lakes.reset_index(drop=True)
+    max_lakes = 20
+    if len(regional_lakes) > max_lakes:
+        regional_lakes = regional_lakes.head(max_lakes)
+
+    nodes_all = load_graph_nodes()
+    import torch
+    model_tmp, edge_index = load_model_and_graph(torch.device('cpu'))
+
+    candidate_nodes = nodes_all[nodes_all['GLAKE_ID'].isin(regional_lakes['GLAKE_ID'])].reset_index(drop=True)  
+    if candidate_nodes.empty:
+        return f"No mapped graph nodes for lakes within {buffer_km}km of {lat}, {lon}."
+
+    union_node_ids = set()
+    for node_idx in candidate_nodes['node_index'].astype(int).values:
+        node_ids, _, _ = build_k_hop_subgraph(edge_index, int(node_idx), hops=NODE_HOPS)
+        union_node_ids.update(node_ids)
+
+    regional_nodes = nodes_all[nodes_all['node_index'].isin(sorted(union_node_ids))].reset_index(drop=True)     
+    if regional_nodes.empty:
+        return f"No mapped graph nodes for lakes within {buffer_km}km of {lat}, {lon}."
+
+    print("\n[1/3] 🌍 Initializing Google Earth Engine for Live ERA5 Retrieval...", flush=True)
+    initialize_earth_engine()
+    target_date = dt.date.today()
+    # Fetch 90 days for XGBoost compatibility
+    raw_window, all_dates = fetch_era5_land_daily_for_nodes(regional_nodes, target_date, history_days=90)
+    imputed_window = impute_raw_window(raw_window)
+
+    # Build weather dataframe for XGBoost batch (first 90 days history only)
+    records = []
+    for local_idx in range(imputed_window.shape[1]):
+        gid = regional_nodes.iloc[local_idx]['GLAKE_ID']
+        for t_idx in range(90):
+            row = {"GLAKE_ID": gid, "date": all_dates[t_idx]}
+            for f_idx, col in enumerate(FEATURE_COLS):
+                row[col] = float(imputed_window[t_idx, local_idx, f_idx]) if not np.isnan(imputed_window[t_idx, local_idx, f_idx]) else None
+            records.append(row)
+
+    weather_df = pd.DataFrame.from_records(records)
+
+    print("\n[2/3] 🌳 Running batched XGBoost for multiple lakes...", flush=True)
+    xgb_results = run_xgboost_batch(weather_df)
+
+    import torch
+    model, edge_index = load_model_and_graph(torch.device('cpu'))
+
+    per_lake_results = []
+    node_index_to_local = {int(r['node_index']): i for i, r in regional_nodes.iterrows()}
+    print("\n[3/3] 🌀 Running GNN Physical Forecast Model...", flush=True)
+    for local_idx, node_row in regional_nodes.iterrows():
+        glake_id = node_row['GLAKE_ID']
+        xres = xgb_results.get(glake_id)
+        if xres:
+            prob_pct = xres['probability'] * 100
+            xgb_text = f"XGBoost Prediction: {prob_pct:.1f}% probability of GLOF."
+        else:
+            xgb_text = "XGBoost Prediction: not available for this lake."
+
+        center_node = int(node_row['node_index'])
+        node_ids, sub_edge_index, center_local = build_k_hop_subgraph(edge_index, center_node, hops=NODE_HOPS)  
+
+        if all(n in node_index_to_local for n in node_ids):
+            positions = [node_index_to_local[n] for n in node_ids]
+            sub_window = imputed_window[:, positions, :]
+            gnn_res = run_gnn_inference(sub_window, model, sub_edge_index, center_local, torch.device('cpu'))   
+        else:
+            try:
+                (sub_imputed, _, fetched_model, _, sub_edge_index, center_local, device, _) = fetch_unified_weather(glake_id)
+                gnn_res = run_gnn_inference(sub_imputed, fetched_model, sub_edge_index, center_local, device)   
+            except Exception as e:
+                gnn_res = {"error": str(e)}
+
+        if not isinstance(gnn_res, dict) or 'error' in gnn_res:
+            err = gnn_res.get('error') if isinstance(gnn_res, dict) else str(gnn_res)
+            gnn_text = f"GNN Prediction: error - {err}"
+        else:
+            surf_mean = gnn_res.get('surface_runoff_sum_mean')
+            risk = gnn_res.get('risk_level', 'UNKNOWN')
+            if surf_mean is None:
+                gnn_text = f"GNN summary: overall trigger {risk}."
+            else:
+                subsurf_mean = gnn_res.get('sub_surface_runoff_sum_mean')
+                snowmelt_mean = gnn_res.get('snowmelt_sum_mean')
+                gnn_text = (
+                    f"GNN summary: surface runoff {surf_mean:.5f} m, sub-surface runoff {subsurf_mean:.5f} m, " 
+                    f"snowmelt {snowmelt_mean:.5f} m; overall trigger {risk}."
+                )
+
+        per_lake_results.append({"GLAKE_ID": glake_id, "xgb": xres, "gnn": gnn_res, "xgb_text": xgb_text, "gnn_text": gnn_text})
+
+    summary_lines = ["XGB SUMMARY"]
+    ranked_results = sorted(
+        per_lake_results,
+        key=lambda item: item['xgb']['probability'] if item.get('xgb') else -1.0,
+        reverse=True,
+    )
+    max_reported = min(8, len(ranked_results))
+    for item in ranked_results[:max_reported]:
+        gid = item['GLAKE_ID']
+        prob = item['xgb']['probability'] * 100 if item['xgb'] else None
+        if prob is not None:
+            line = f"- {gid}: {prob:.1f}%"
+        else:
+            line = f"- {gid}: N/A"
+        summary_lines.append(line)
+
+    if len(ranked_results) > max_reported:
+        summary_lines.append(f"- ... {len(ranked_results) - max_reported} more lakes omitted for brevity")      
+
+    gnn_high = sum(1 for item in per_lake_results if item.get('gnn', {}).get('risk_level') == 'HIGH')
+    gnn_low = sum(1 for item in per_lake_results if item.get('gnn', {}).get('risk_level') == 'LOW')
+    valid_runoffs = [item['gnn']['surface_runoff_sum_mean'] for item in per_lake_results if item.get('gnn') and 'surface_runoff_sum_mean' in item['gnn'] and item['gnn']['surface_runoff_sum_mean'] is not None]
+    gnn_summary_lines = ["GNN SUMMARY"]
+    if valid_runoffs:
+        subsurf_vals = [item['gnn'].get('sub_surface_runoff_sum_mean') for item in per_lake_results if item.get('gnn') and 'sub_surface_runoff_sum_mean' in item['gnn']]
+        snowmelt_vals = [item['gnn'].get('snowmelt_sum_mean') for item in per_lake_results if item.get('gnn') and 'snowmelt_sum_mean' in item['gnn']]
+        gnn_summary_lines.append(f"- HIGH: {gnn_high}, LOW: {gnn_low}")
+        gnn_summary_lines.append(f"- Surface runoff (avg): {float(np.mean(valid_runoffs)):.5f}, max: {float(np.max(valid_runoffs)):.5f}")
+        if subsurf_vals:
+            gnn_summary_lines.append(f"- Sub-surface runoff (avg): {float(np.mean(subsurf_vals)):.5f}")
+        if snowmelt_vals:
+            gnn_summary_lines.append(f"- Snowmelt (avg): {float(np.mean(snowmelt_vals)):.5f}")
+        gnn_summary_lines.append("")
+        gnn_summary_lines.append("Summary: the GNN predicts 3-day surface runoff, sub-surface runoff, and snowmelt. 'HIGH' means the surface runoff is above a 3x-standard-deviation threshold from the 60-day history; use this as a supporting signal after XGBoost.")
+    else:
+        gnn_summary_lines.append("- GNN results unavailable.")
+
+    return "\n".join(summary_lines) + "\n\n" + "\n".join(gnn_summary_lines)
+
+
+def explain_top_glof_lakes(limit: int = 5):
+    """Run the dual pipeline for all lakes and return the top-N by XGBoost probability."""
+    from gnn.inference import (
+        initialize_earth_engine,
+        fetch_era5_land_daily_for_nodes,
+        load_model_and_graph,
+        load_graph_nodes,
+        build_k_hop_subgraph,
+        NODE_HOPS,
+        FEATURE_COLS,
+        HISTORY_DAYS,
+    )
+
+    if not os.path.exists(STATIC_DATA_PATH):
+        return "Error: Lake geometry database not found."
+
+    nodes_all = load_graph_nodes()
+    if nodes_all.empty:
+        return "Error: No graph nodes available for lake ranking."
+
+    print("\n[1/3] 🌍 Initializing Google Earth Engine for Live ERA5 Retrieval...", flush=True)
+    initialize_earth_engine()
+
+    import torch
+    model, edge_index = load_model_and_graph(torch.device("cpu"))
+
+    target_date = dt.date.today()
+    node_batches = [nodes_all.iloc[i : i + 100].reset_index(drop=True) for i in range(0, len(nodes_all), 100)]  
+    per_lake_results = []
+    node_index_to_local = {int(r["node_index"]): i for i, r in nodes_all.iterrows()}
+
+    print("\n[2/3] 🌳 Running batched XGBoost for multiple lakes...", flush=True)
+    for batch_index, batch_nodes in enumerate(node_batches, start=1):
+        print(f"Batch {batch_index}/{len(node_batches)}: fetching ERA5 for {len(batch_nodes)} lakes...", flush=True)
+        raw_window, all_dates = fetch_era5_land_daily_for_nodes(batch_nodes, target_date, history_days=90)
+        imputed_window = impute_raw_window(raw_window)
+        batch_node_index_to_local = {int(r["node_index"]): i for i, r in batch_nodes.iterrows()}
+
+        records = []
+        for local_idx in range(imputed_window.shape[1]):
+            gid = batch_nodes.iloc[local_idx]["GLAKE_ID"]
+            for t_idx in range(90):
+                row = {"GLAKE_ID": gid, "date": all_dates[t_idx]}
+                for f_idx, col in enumerate(FEATURE_COLS):
+                    value = imputed_window[t_idx, local_idx, f_idx]
+                    row[col] = float(value) if not np.isnan(value) else None
+                records.append(row)
+
+        weather_df = pd.DataFrame.from_records(records)
+        xgb_results = run_xgboost_batch(weather_df)
+
+        print(f"Batch {batch_index}/{len(node_batches)}: running GNN...", flush=True)
+        for _, node_row in batch_nodes.iterrows():
+            glake_id = node_row["GLAKE_ID"]
+            xres = xgb_results.get(glake_id)
+            if xres:
+                xgb_text = f"XGBoost Prediction: {xres['probability'] * 100:.1f}% probability of GLOF."
+            else:
+                xgb_text = "XGBoost Prediction: not available for this lake."
+
+            center_node = int(node_row["node_index"])
+            node_ids, sub_edge_index, center_local = build_k_hop_subgraph(edge_index, center_node, hops=NODE_HOPS)
+            positions = [batch_node_index_to_local[n] for n in node_ids if n in batch_node_index_to_local]      
+            if len(positions) != len(node_ids):
+                gnn_res = {"error": "Missing regional nodes in all-lake fetch."}
+            else:
+                start_slice = 90 - HISTORY_DAYS
+                sub_window = imputed_window[start_slice:, positions, :]
+                gnn_res = run_gnn_inference(sub_window, model, sub_edge_index, center_local, torch.device("cpu"))
+
+            if not isinstance(gnn_res, dict) or "error" in gnn_res:
+                err = gnn_res.get("error") if isinstance(gnn_res, dict) else str(gnn_res)
+                gnn_text = f"GNN Prediction: error - {err}"
+            else:
+                surf_mean = gnn_res.get("surface_runoff_sum_mean")
+                risk = gnn_res.get("risk_level", "UNKNOWN")
+                if surf_mean is None:
+                    gnn_text = f"GNN summary: overall trigger {risk}."
+                else:
+                    subsurf_mean = gnn_res.get("sub_surface_runoff_sum_mean")
+                    snowmelt_mean = gnn_res.get("snowmelt_sum_mean")
+                    gnn_text = (
+                        f"GNN summary: surface runoff {surf_mean:.5f} m, sub-surface runoff {subsurf_mean:.5f} m, "
+                        f"snowmelt {snowmelt_mean:.5f} m; overall trigger {risk}."
+                    )
+
+            per_lake_results.append({"GLAKE_ID": glake_id, "xgb": xres, "gnn": gnn_res, "xgb_text": xgb_text, "gnn_text": gnn_text})
+
+    ranked_results = sorted(
+        per_lake_results,
+        key=lambda item: item["xgb"]["probability"] if item.get("xgb") else -1.0,
+        reverse=True,
+    )
+
+    top_results = ranked_results[:limit]
+    if not top_results:
+        return "No lakes were ranked."
+
+    lines = [f"Top {len(top_results)} glacial lakes by forecast GLOF probability:"]
+    for idx, item in enumerate(top_results, start=1):
+        prob = item["xgb"]["probability"] * 100 if item.get("xgb") else None
+        gnn = item.get("gnn", {})
+        risk = gnn.get("risk_level", "UNKNOWN") if isinstance(gnn, dict) else "UNKNOWN"
+        surf_mean = gnn.get("surface_runoff_sum_mean") if isinstance(gnn, dict) else None
+        if prob is None:
+            lines.append(f"{idx}. {item['GLAKE_ID']}: probability unavailable; GNN trigger {risk}.")
+        else:
+            if surf_mean is None:
+                lines.append(f"{idx}. {item['GLAKE_ID']}: XGBoost probability {prob:.1f}%; GNN trigger {risk}.")
+            else:
+                lines.append(
+                    f"{idx}. {item['GLAKE_ID']}: XGBoost probability {prob:.1f}%; "
+                    f"GNN surface runoff {surf_mean:.5f} m; trigger {risk}."
+                )
+
+    return "\n".join(lines)
+
+if __name__ == "__main__":
+    import argparse
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--lake", default="GL11")
+    args = parser.parse_args()
+    print(explain_glof_prediction(args.lake))

chatbot/graphrag/catalog-v001.xml

@@ -0,0 +1,4 @@
+<?xml version="1.0" encoding="UTF-8" standalone="no"?>
+<catalog prefer="public" xmlns="urn:oasis:names:tc:entity:xmlns:xml:catalog">
+    <group id="Folder Repository, directory=, recursive=true, Auto-Update=true, version=2" prefer="public"/>
+</catalog>

chatbot/rag/README_RAG.md

@@ -0,0 +1,204 @@
+# GLOF Markdown RAG Pipeline
+
+This pipeline is designed for the Markdown files in `rag/data`.
+
+It does four things:
+
+1. Splits Markdown by headings.
+2. Keeps tables and figures as special chunks.
+3. Embeds chunks with `BAAI/bge-m3`.
+4. Retrieves with hybrid dense FAISS + BM25 keyword search.
+
+## 1. Install Dependencies
+
+From the project root:
+
+```powershell
+python -m venv .venv
+.\.venv\Scripts\Activate.ps1
+pip install -r rag\requirements-rag.txt
+```
+
+If `python` opens the Microsoft Store or is not available, use `uv`:
+
+```powershell
+$env:UV_CACHE_DIR = ".uv-cache"
+uv venv .venv --python 3.12
+.\.venv\Scripts\Activate.ps1
+uv pip install -r rag\requirements-rag.txt
+```
+
+The first index build will download `BAAI/bge-m3`, so it needs internet access.
+
+## 2. Chunk the Markdown Files
+
+```powershell
+python rag\chunk_markdown.py `
+  --data-dir rag\data `
+  --out rag\artifacts\chunks.jsonl `
+  --max-text-tokens 850 `
+  --text-overlap-tokens 120 `
+  --table-rows-per-chunk 35
+```
+
+Recommended defaults for this corpus:
+
+```text
+Text chunks: 850 tokens
+Text overlap: 120 tokens
+Table chunks: 35 rows, table header repeated
+Figure chunks: one figure/caption block per chunk
+```
+
+Each JSONL row has:
+
+```json
+{
+  "id": "chunk id",
+  "text": "Document/Section metadata plus chunk text",
+  "metadata": {
+    "source_file": "data/...",
+    "document_title": "...",
+    "section_path": "...",
+    "chunk_type": "text/table/figure",
+    "line_start": 1,
+    "line_end": 20
+  }
+}
+```
+
+## 3. Build BGE-M3 Embedding Index
+
+```powershell
+python rag\build_bge_m3_index.py `
+  --chunks rag\artifacts\chunks.jsonl `
+  --index-dir rag\index `
+  --model-name BAAI/bge-m3 `
+  --batch-size 8
+```
+
+Outputs:
+
+```text
+rag/index/dense.faiss
+rag/index/bm25.pkl
+rag/index/chunks.jsonl
+rag/index/manifest.json
+rag/index/embeddings.npy
+```
+
+If your GPU has enough memory, increase `--batch-size`.
+
+## 4. Search the Index
+
+```powershell
+python rag\search_hybrid.py `
+  --index-dir rag\index `
+  --query "Which glacial lakes in India showed more than 40 percent increase in August 2025?" `
+  --top-k 8
+```
+
+For table-heavy or number-heavy queries, hybrid search is important because BM25 catches exact terms such as lake IDs, years, basin names, and `>50 Ha`.
+
+Use optional reranking:
+
+```powershell
+python rag\search_hybrid.py `
+  --index-dir rag\index `
+  --query "What are the methods used for GLOF hazard assessment?" `
+  --top-k 8 `
+  --rerank-model BAAI/bge-reranker-v2-m3
+```
+
+## 5. Using Retrieved Chunks in a Chatbot
+
+Take the top retrieved chunks and pass them to your LLM as context:
+
+```text
+You are a GLOF research assistant. Answer only from the context.
+If the answer is not present, say you do not know.
+Cite source file and line numbers.
+
+Context:
+[chunk 1]
+[chunk 2]
+...
+
+Question:
+...
+```
+
+Use `metadata.source_file`, `metadata.section_path`, `metadata.line_start`, and `metadata.line_end` for citations.
+
+## 6. Build The Knowledge Graph Store
+
+The TTL file is large, so do not parse it on every question. Build a persistent Oxigraph store once:
+
+```powershell
+python rag\build_kg_store.py `
+  --ttl graphrag\glof_kg_populated_codex.ttl `
+  --store-dir rag\kg_store
+```
+
+Use `--force` only if you want to delete and rebuild the store.
+
+## 7. Hybrid GraphRAG With Qwen Through Hugging Face
+
+Put your token in `.env`:
+
+```text
+HF_TOKEN=your_huggingface_token
+```
+
+Then ask a question:
+
+```powershell
+python rag\hybrid_graphrag.py `
+  --question "What are the main GLOF triggering mechanisms and how are they related in the knowledge graph?" `
+  --route auto `
+  --model Qwen/Qwen3-4B-Instruct-2507
+```
+
+Routes:
+
+```text
+auto    chooses vector, KG, or both
+vector  uses document chunks only
+kg      uses the TTL knowledge graph only
+both    combines document chunks and KG evidence
+```
+
+Examples:
+
+```powershell
+python rag\hybrid_graphrag.py `
+  --question "Which glacial lakes showed more than 40 percent increase in August 2025?" `
+  --route vector
+```
+
+```powershell
+python rag\hybrid_graphrag.py `
+  --question "Explain how GLOFTrigger, IceAvalancheTrigger, and RockAvalancheTrigger are connected." `
+  --route kg `
+  --show-context
+```
+
+```powershell
+python rag\hybrid_graphrag.py `
+  --question "Why do moraine-dammed lakes have high GLOF risk? Use both reports and the ontology." `
+  --route both
+```
+
+Geospatial KG lookup:
+
+```powershell
+python rag\hybrid_graphrag.py `
+  --question "Give information about 5 glacial lakes that are around 34.0926 degree lat and 75.4972 degree lon" `
+  --route kg `
+  --show-context
+```
+
+The graph uses `:latitude` and `:longitude` as WGS84 decimal degrees for representative points. Some `geo:asWKT`
+polygons are also stored in longitude-latitude degree order, while a few older/source geometries may appear in projected
+coordinates. For nearest-lake lookup, the code uses the decimal-degree `:latitude`/`:longitude` values and computes
+haversine distance.

chatbot/rag/build_bge_m3_index.py

@@ -0,0 +1,132 @@
+from __future__ import annotations
+
+import argparse
+import json
+import pickle
+import re
+import shutil
+from pathlib import Path
+
+
+TOKEN_RE = re.compile(r"[A-Za-z0-9_./+-]+")
+
+
+def require_dependencies() -> None:
+    missing: list[str] = []
+    try:
+        import faiss  # noqa: F401
+    except Exception:
+        missing.append("faiss-cpu")
+    try:
+        import sentence_transformers  # noqa: F401
+    except Exception:
+        missing.append("sentence-transformers")
+    try:
+        import rank_bm25  # noqa: F401
+    except Exception:
+        missing.append("rank-bm25")
+    if missing:
+        raise SystemExit(
+            "Missing dependencies: "
+            + ", ".join(missing)
+            + "\nInstall them with: pip install -r rag/requirements-rag.txt"
+        )
+
+
+def load_chunks(path: Path) -> list[dict]:
+    chunks: list[dict] = []
+    with path.open("r", encoding="utf-8") as handle:
+        for line in handle:
+            if line.strip():
+                chunks.append(json.loads(line))
+    if not chunks:
+        raise ValueError(f"No chunks found in {path}")
+    return chunks
+
+
+def tokenize_for_bm25(text: str) -> list[str]:
+    return [token.lower() for token in TOKEN_RE.findall(text)]
+
+
+def save_json(path: Path, payload: dict) -> None:
+    path.write_text(json.dumps(payload, indent=2, ensure_ascii=False), encoding="utf-8")
+
+
+def build_dense_index(chunks: list[dict], model_name: str, batch_size: int, index_dir: Path) -> tuple[int, int]:
+    import faiss
+    import numpy as np
+    from sentence_transformers import SentenceTransformer
+
+    try:
+        model = SentenceTransformer(model_name, trust_remote_code=True)
+    except TypeError:
+        model = SentenceTransformer(model_name)
+
+    texts = [chunk["text"] for chunk in chunks]
+    embeddings = model.encode(
+        texts,
+        batch_size=batch_size,
+        normalize_embeddings=True,
+        show_progress_bar=True,
+        convert_to_numpy=True,
+    ).astype("float32")
+
+    if embeddings.ndim != 2:
+        raise ValueError(f"Expected 2D embeddings, got shape {embeddings.shape}")
+
+    dimension = int(embeddings.shape[1])
+    index = faiss.IndexFlatIP(dimension)
+    index.add(embeddings)
+    faiss.write_index(index, str(index_dir / "dense.faiss"))
+    np.save(index_dir / "embeddings.npy", embeddings)
+    return len(chunks), dimension
+
+
+def build_bm25_index(chunks: list[dict], index_dir: Path) -> None:
+    from rank_bm25 import BM25Okapi
+
+    tokenized_corpus = [tokenize_for_bm25(chunk["text"]) for chunk in chunks]
+    bm25 = BM25Okapi(tokenized_corpus)
+    with (index_dir / "bm25.pkl").open("wb") as handle:
+        pickle.dump(
+            {
+                "bm25": bm25,
+                "tokenized_corpus": tokenized_corpus,
+            },
+            handle,
+        )
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Build dense BGE-M3 + BM25 indexes for Markdown RAG chunks.")
+    parser.add_argument("--chunks", type=Path, default=Path("rag/artifacts/chunks.jsonl"))
+    parser.add_argument("--index-dir", type=Path, default=Path("rag/index"))
+    parser.add_argument("--model-name", default="BAAI/bge-m3")
+    parser.add_argument("--batch-size", type=int, default=8)
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    require_dependencies()
+    args.index_dir.mkdir(parents=True, exist_ok=True)
+
+    chunks = load_chunks(args.chunks)
+    shutil.copyfile(args.chunks, args.index_dir / "chunks.jsonl")
+    chunk_count, dimension = build_dense_index(chunks, args.model_name, args.batch_size, args.index_dir)
+    build_bm25_index(chunks, args.index_dir)
+
+    manifest = {
+        "model_name": args.model_name,
+        "chunk_count": chunk_count,
+        "embedding_dimension": dimension,
+        "dense_index": "dense.faiss",
+        "bm25_index": "bm25.pkl",
+        "chunks": "chunks.jsonl",
+    }
+    save_json(args.index_dir / "manifest.json", manifest)
+    print(json.dumps({"index_dir": str(args.index_dir), **manifest}, indent=2))
+
+
+if __name__ == "__main__":
+    main()

chatbot/rag/build_kg_store.py

@@ -0,0 +1,25 @@
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+from kg_store import KgStore
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Build a persistent Oxigraph SPARQL store from the GLOF TTL file.")
+    parser.add_argument("--ttl", type=Path, default=Path("graphrag/glof_kg_populated_codex.ttl"))
+    parser.add_argument("--store-dir", type=Path, default=Path("rag/kg_store"))
+    parser.add_argument("--force", action="store_true", help="Delete and rebuild the existing store directory.")
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    manifest = KgStore.build(args.ttl, args.store_dir, force=args.force)
+    print(json.dumps(manifest, indent=2))
+
+
+if __name__ == "__main__":
+    main()

chatbot/rag/chunk_markdown.py

@@ -0,0 +1,604 @@
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import re
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Iterable
+
+
+HEADING_RE = re.compile(r"^(#{1,6})\s+(.+?)\s*$")
+TABLE_LINE_RE = re.compile(r"^\s*\|.*\|\s*$")
+TABLE_CAPTION_RE = re.compile(
+    r"^\s*(?:\*\*)?\s*(?:Table|TABLE)\s+[A-Za-z0-9. -]+(?::|\.|\s*$).*?(?:\*\*)?\s*$"
+)
+FIGURE_RE = re.compile(
+    r"^\s*(?:Image\s+/.+?\s+description:|(?:\*\*)?\s*Figure\s+\d+[A-Za-z0-9.:-]*\b)",
+    re.IGNORECASE,
+)
+WORD_RE = re.compile(r"\w+|[^\w\s]", re.UNICODE)
+LAKE_ID_RE = re.compile(r"\b\d{2}[_A-Z0-9]{3,12}[_A-Z0-9]*\b")
+GENERIC_TITLE_RE = re.compile(
+    r"^(abstract|arstract|introduction|summary|executive summary|foreword|acknowledgements?|keywords?)$",
+    re.IGNORECASE,
+)
+SKIP_SECTION_RE = re.compile(
+    r"^(contents|references|bibliography|further reading|acknowledgements?|acknowledgments?|"
+    r"declaration of competing interest|declaration of competing interests|competing interests?|"
+    r"conflict of interest|data availability)$",
+    re.IGNORECASE,
+)
+
+
+NOISE_PATTERNS = [
+    re.compile(r"^\s*\*{3,}\s*$"),
+    re.compile(r"^\s*-{3,}\s*$"),
+    re.compile(r"^\s*nrsc\s*$", re.IGNORECASE),
+    re.compile(r"^\s*\d+\s*$"),
+    re.compile(r"^\s*[ivxlcdm]{1,8}\s*$", re.IGNORECASE),
+    re.compile(r"^\s*National Remote Sensing Centre,\s*ISRO,\s*Hyderabad\s*\d*\s*$", re.IGNORECASE),
+    re.compile(r"^\s*\d+\s+National Remote Sensing Centre,\s*ISRO,\s*Hyderabad\s*$", re.IGNORECASE),
+
+]
+
+
+@dataclass
+class ChunkConfig:
+    max_text_tokens: int = 850
+    text_overlap_tokens: int = 120
+    min_text_tokens: int = 40
+    table_rows_per_chunk: int = 35
+    model_name: str | None = "BAAI/bge-m3"
+
+
+class TokenCounter:
+    def __init__(self, model_name: str | None = None) -> None:
+        self.tokenizer = None
+        if model_name:
+            try:
+                from transformers import AutoTokenizer
+
+                self.tokenizer = AutoTokenizer.from_pretrained(model_name, local_files_only=True)
+            except Exception:
+                self.tokenizer = None
+
+    def count(self, text: str) -> int:
+        if not text:
+            return 0
+        if self.tokenizer is not None:
+            return len(self.tokenizer.encode(text, add_special_tokens=False))
+        return len(WORD_RE.findall(text))
+
+
+def strip_markdown(text: str) -> str:
+    text = text.strip()
+    text = re.sub(r"^\*+|\*+$", "", text)
+    text = re.sub(r"\*{1,3}", " ", text)
+    text = re.sub(r"`+", "", text)
+    text = re.sub(r"<br\s*/?>", " ", text, flags=re.IGNORECASE)
+    text = re.sub(r"<[^>]+>", "", text)
+    return re.sub(r"\s+", " ", text).strip()
+
+
+def clean_line(line: str) -> str:
+    line = line.replace("\ufeff", "")
+    line = re.sub(r"<br\s*/?>", " ", line, flags=re.IGNORECASE)
+    line = re.sub(r"\s+$", "", line)
+    return line
+
+
+def is_noise_line(line: str) -> bool:
+    stripped = strip_markdown(line)
+    if not stripped:
+        return False
+    return any(pattern.match(stripped) for pattern in NOISE_PATTERNS)
+
+
+def is_table_line(line: str) -> bool:
+    return bool(TABLE_LINE_RE.match(line))
+
+
+def is_table_separator(line: str) -> bool:
+    cells = [cell.strip() for cell in line.strip().strip("|").split("|")]
+    return bool(cells) and all(re.fullmatch(r":?-{3,}:?", cell or "") for cell in cells)
+
+
+def is_heading(line: str) -> tuple[int, str] | None:
+    match = HEADING_RE.match(line)
+    if not match:
+        return None
+    return len(match.group(1)), strip_markdown(match.group(2))
+
+
+def looks_like_table_caption(line: str) -> bool:
+    return bool(TABLE_CAPTION_RE.match(line.strip())) or bool(
+        re.match(r"^Table\s+[A-Za-z0-9. -]+(?::|\.|\s|$)", strip_markdown(line), re.IGNORECASE)
+    )
+
+
+def looks_like_figure(line: str) -> bool:
+    return bool(FIGURE_RE.match(line))
+
+
+def title_from_filename(path: Path) -> str:
+    stem = path.stem
+    stem = stem.replace("_extraction", "").replace("_gemini", "")
+    stem = stem.replace("('", "").replace("', '.pdf')", "")
+    stem = stem.replace("_", " ")
+    return strip_markdown(stem)
+
+
+def normalize_document_title(path: Path, first_heading: str | None) -> str:
+    if first_heading and not GENERIC_TITLE_RE.match(first_heading):
+        return first_heading
+    return title_from_filename(path)
+
+
+def make_chunk_id(source_file: str, chunk_type: str, line_start: int, line_end: int, index: int) -> str:
+    raw = f"{source_file}:{chunk_type}:{line_start}:{line_end}:{index}"
+    return hashlib.sha1(raw.encode("utf-8")).hexdigest()[:16]
+
+
+def heading_path(headings: list[str | None]) -> str:
+    return " > ".join(heading for heading in headings if heading)
+
+
+def should_skip_section(headings: list[str | None]) -> bool:
+    return any(bool(heading and SKIP_SECTION_RE.match(heading.strip())) for heading in headings)
+
+
+def prefix_content(
+    body: str,
+    document_title: str,
+    section_path: str,
+    chunk_type: str,
+    label: str | None = None,
+) -> str:
+    lines = [f"Document: {document_title}"]
+    if section_path:
+        lines.append(f"Section: {section_path}")
+    lines.append(f"Type: {chunk_type}")
+    if label:
+        lines.append(label)
+    lines.append("")
+    lines.append(body.strip())
+    return "\n".join(lines).strip()
+
+
+def split_oversized_paragraph(paragraph: str, counter: TokenCounter, max_tokens: int) -> list[str]:
+    sentences = re.split(r"(?<=[.!?])\s+", paragraph)
+    pieces: list[str] = []
+    current: list[str] = []
+    for sentence in sentences:
+        if counter.count(sentence) > max_tokens:
+            words = WORD_RE.findall(sentence)
+            for start in range(0, len(words), max_tokens):
+                pieces.append(" ".join(words[start : start + max_tokens]))
+            continue
+        candidate = " ".join(current + [sentence]).strip()
+        if current and counter.count(candidate) > max_tokens:
+            pieces.append(" ".join(current).strip())
+            current = [sentence]
+        else:
+            current.append(sentence)
+    if current:
+        pieces.append(" ".join(current).strip())
+    return [piece for piece in pieces if piece]
+
+
+def overlap_tail(text: str, overlap_tokens: int) -> str:
+    if overlap_tokens <= 0:
+        return ""
+    words = WORD_RE.findall(text)
+    return " ".join(words[-overlap_tokens:])
+
+
+def split_text(text: str, counter: TokenCounter, max_tokens: int, overlap_tokens: int) -> list[str]:
+    text = re.sub(r"\n{3,}", "\n\n", text.strip())
+    if not text:
+        return []
+    if counter.count(text) <= max_tokens:
+        return [text]
+
+    paragraphs = [paragraph.strip() for paragraph in re.split(r"\n\s*\n", text) if paragraph.strip()]
+    expanded: list[str] = []
+    for paragraph in paragraphs:
+        if counter.count(paragraph) > max_tokens:
+            expanded.extend(split_oversized_paragraph(paragraph, counter, max_tokens))
+        else:
+            expanded.append(paragraph)
+
+    chunks: list[str] = []
+    current = ""
+    for paragraph in expanded:
+        candidate = f"{current}\n\n{paragraph}".strip() if current else paragraph
+        if current and counter.count(candidate) > max_tokens:
+            chunks.append(current.strip())
+            tail = overlap_tail(current, overlap_tokens)
+            current = f"{tail}\n\n{paragraph}".strip() if tail else paragraph
+        else:
+            current = candidate
+    if current:
+        chunks.append(current.strip())
+    return chunks
+
+
+def parse_table_header(table_lines: list[str]) -> tuple[list[str], str | None]:
+    if not table_lines:
+        return [], None
+    header = [strip_markdown(cell) for cell in table_lines[0].strip().strip("|").split("|")]
+    separator = table_lines[1] if len(table_lines) > 1 and is_table_separator(table_lines[1]) else None
+    return header, separator
+
+
+def extract_caption_from_buffer(buffer: list[tuple[int, str]]) -> str | None:
+    for _, line in reversed(buffer[-4:]):
+        if line.strip() and looks_like_table_caption(line):
+            return strip_markdown(line)
+    return None
+
+
+def trim_caption_from_buffer(buffer: list[tuple[int, str]], caption: str | None) -> None:
+    if not caption:
+        return
+    for idx in range(len(buffer) - 1, max(-1, len(buffer) - 5), -1):
+        if strip_markdown(buffer[idx][1]) == caption:
+            del buffer[idx]
+            return
+
+
+def infer_caption_from_previous(lines: list[str], table_start: int) -> str | None:
+    for idx in range(table_start - 1, max(-1, table_start - 5), -1):
+        line = lines[idx].strip()
+        if not line:
+            continue
+        if looks_like_table_caption(line):
+            return strip_markdown(line)
+        break
+    return None
+
+
+def extract_entities(text: str) -> dict[str, list[str]]:
+    BASINS = ["Indus", "Ganga", "Brahmaputra"]
+    SUBBASINS = [
+        "Teesta", "Satluj", "Chenab", "Jhelum", "Ravi", "Beas", "Shyok", "Gandak",
+        "Ghaghara", "Kosi", "Sarda", "Upper Ganga", "Yamuna", "Gilgit", "Indus upper",
+        "Indus middle", "Amo Chu", "Dibang", "Dihang", "Jia Bharali", "Lhasa Tsangpo",
+        "Lohit", "Lower Yarlung Tsangpo", "Manas", "Puna Tsang Chu", "Subansiri",
+        "Upper Yarlung Tsangpo"
+    ]
+    COUNTRIES = ["India", "China", "Nepal", "Bhutan", "Myanmar"]
+
+    def find_terms(terms: list[str]) -> list[str]:
+        if not terms:
+            return []
+        pattern = r"\b(?:" + "|".join(re.escape(t) for t in terms) + r")\b"
+        return sorted({m.group(0) for m in re.finditer(pattern, text, flags=re.IGNORECASE)})
+
+    basins = find_terms(BASINS)
+    subbasins = find_terms(SUBBASINS)
+    countries = find_terms(COUNTRIES)
+    lake_ids = sorted(set(LAKE_ID_RE.findall(text)))
+
+    return {
+        "basins": basins[:20],
+        "subbasins": subbasins[:20],
+        "countries": countries[:20],
+        "lake_ids": lake_ids[:50],
+    }
+
+
+def collect_table(lines: list[str], start_idx: int) -> tuple[list[str], int]:
+    table_lines: list[str] = []
+    idx = start_idx
+    while idx < len(lines) and is_table_line(lines[idx]):
+        table_lines.append(clean_line(lines[idx]))
+        idx += 1
+    return table_lines, idx
+
+
+def collect_figure(lines: list[str], start_idx: int) -> tuple[list[str], int]:
+    figure_lines = [clean_line(lines[start_idx])]
+    idx = start_idx + 1
+    while idx < len(lines):
+        line = clean_line(lines[idx])
+        if not line.strip():
+            break
+        if is_heading(line) or is_table_line(line):
+            break
+        # Keep short caption continuations with the figure, but avoid swallowing a whole paragraph.
+        if looks_like_figure(line) or len(figure_lines) < 3:
+            figure_lines.append(line)
+            idx += 1
+            continue
+        break
+    return figure_lines, idx
+
+
+def emit_text_chunks(
+    chunks: list[dict],
+    buffer: list[tuple[int, str]],
+    *,
+    source_file: str,
+    document_title: str,
+    section_path: str,
+    headings: list[str | None],
+    counter: TokenCounter,
+    config: ChunkConfig,
+) -> None:
+    if not buffer:
+        return
+
+    line_start = buffer[0][0]
+    line_end = buffer[-1][0]
+    body = "\n".join(line for _, line in buffer).strip()
+    body = re.sub(r"\n{3,}", "\n\n", body)
+    if not body:
+        return
+
+    text_chunks = split_text(body, counter, config.max_text_tokens, config.text_overlap_tokens)
+    for piece_index, piece in enumerate(text_chunks):
+        if counter.count(piece) < config.min_text_tokens and len(text_chunks) > 1:
+            continue
+        content = prefix_content(piece, document_title, section_path, "text")
+        metadata = {
+            "source_file": source_file,
+            "document_title": document_title,
+            "section_path": section_path,
+            "section_headings": [heading for heading in headings if heading],
+            "chunk_type": "text",
+            "line_start": line_start,
+            "line_end": line_end,
+            "token_count_estimate": counter.count(content),
+            **extract_entities(content),
+        }
+        chunks.append(
+            {
+                "id": make_chunk_id(source_file, "text", line_start, line_end, piece_index),
+                "text": content,
+                "metadata": metadata,
+            }
+        )
+
+
+def emit_table_chunks(
+    chunks: list[dict],
+    table_lines: list[str],
+    *,
+    caption: str | None,
+    source_file: str,
+    document_title: str,
+    section_path: str,
+    headings: list[str | None],
+    line_start: int,
+    counter: TokenCounter,
+    config: ChunkConfig,
+) -> None:
+    if not table_lines:
+        return
+
+    columns, separator = parse_table_header(table_lines)
+    header = table_lines[0]
+    body_start = 2 if separator else 1
+    body_rows = table_lines[body_start:]
+    rows_per_chunk = max(1, config.table_rows_per_chunk)
+    row_groups = [body_rows[i : i + rows_per_chunk] for i in range(0, len(body_rows), rows_per_chunk)] or [[]]
+
+    for group_index, rows in enumerate(row_groups):
+        table_body = [header]
+        if separator:
+            table_body.append(separator)
+        table_body.extend(rows)
+        label = f"Table: {caption}" if caption else "Table"
+        body = "\n".join(table_body)
+        content = prefix_content(body, document_title, section_path, "table", label)
+        row_start = group_index * rows_per_chunk + 1
+        row_end = row_start + len(rows) - 1 if rows else row_start
+        metadata = {
+            "source_file": source_file,
+            "document_title": document_title,
+            "section_path": section_path,
+            "section_headings": [heading for heading in headings if heading],
+            "chunk_type": "table",
+            "table_caption": caption,
+            "columns": columns,
+            "table_row_start": row_start,
+            "table_row_end": row_end,
+            "line_start": line_start,
+            "line_end": line_start + len(table_lines) - 1,
+            "token_count_estimate": counter.count(content),
+            **extract_entities(content),
+        }
+        chunks.append(
+            {
+                "id": make_chunk_id(source_file, "table", line_start, line_start + len(table_lines) - 1, group_index),
+                "text": content,
+                "metadata": metadata,
+            }
+        )
+
+
+def emit_figure_chunk(
+    chunks: list[dict],
+    figure_lines: list[str],
+    *,
+    source_file: str,
+    document_title: str,
+    section_path: str,
+    headings: list[str | None],
+    line_start: int,
+    counter: TokenCounter,
+) -> None:
+    body = "\n".join(figure_lines).strip()
+    caption = next((strip_markdown(line) for line in figure_lines if re.match(r"^\s*(?:\*\*)?\s*Figure\s+\d+", line, re.IGNORECASE)), None)
+    label = f"Figure: {caption}" if caption else "Figure"
+    content = prefix_content(body, document_title, section_path, "figure", label)
+    metadata = {
+        "source_file": source_file,
+        "document_title": document_title,
+        "section_path": section_path,
+        "section_headings": [heading for heading in headings if heading],
+        "chunk_type": "figure",
+        "figure_caption": caption,
+        "line_start": line_start,
+        "line_end": line_start + len(figure_lines) - 1,
+        "token_count_estimate": counter.count(content),
+        **extract_entities(content),
+    }
+    chunks.append(
+        {
+            "id": make_chunk_id(source_file, "figure", line_start, line_start + len(figure_lines) - 1, 0),
+            "text": content,
+            "metadata": metadata,
+        }
+    )
+
+
+def chunk_markdown_file(path: Path, data_dir: Path, counter: TokenCounter, config: ChunkConfig) -> list[dict]:
+    raw_text = path.read_text(encoding="utf-8", errors="replace")
+    lines = [clean_line(line) for line in raw_text.splitlines()]
+    chunks: list[dict] = []
+    headings: list[str | None] = [None] * 6
+    first_heading: str | None = None
+    document_title = normalize_document_title(path, first_heading)
+    source_file = str(path.relative_to(data_dir.parent)).replace("\\", "/")
+    text_buffer: list[tuple[int, str]] = []
+
+    def flush_text() -> None:
+        emit_text_chunks(
+            chunks,
+            text_buffer,
+            source_file=source_file,
+            document_title=document_title,
+            section_path=heading_path(headings),
+            headings=headings,
+            counter=counter,
+            config=config,
+        )
+        text_buffer.clear()
+
+    idx = 0
+    while idx < len(lines):
+        line = lines[idx]
+        line_no = idx + 1
+
+        heading = is_heading(line)
+        if heading:
+            flush_text()
+            level, title = heading
+            if first_heading is None:
+                first_heading = title
+                document_title = normalize_document_title(path, first_heading)
+            headings[level - 1] = title
+            for reset_idx in range(level, len(headings)):
+                headings[reset_idx] = None
+            idx += 1
+            continue
+
+        if is_noise_line(line):
+            idx += 1
+            continue
+
+        if should_skip_section(headings):
+            idx += 1
+            continue
+
+        if is_table_line(line):
+            caption = extract_caption_from_buffer(text_buffer) or infer_caption_from_previous(lines, idx)
+            trim_caption_from_buffer(text_buffer, caption)
+            flush_text()
+            table_lines, next_idx = collect_table(lines, idx)
+            emit_table_chunks(
+                chunks,
+                table_lines,
+                caption=caption,
+                source_file=source_file,
+                document_title=document_title,
+                section_path=heading_path(headings),
+                headings=headings,
+                line_start=line_no,
+                counter=counter,
+                config=config,
+            )
+            idx = next_idx
+            continue
+
+        if looks_like_figure(line):
+            flush_text()
+            figure_lines, next_idx = collect_figure(lines, idx)
+            emit_figure_chunk(
+                chunks,
+                figure_lines,
+                source_file=source_file,
+                document_title=document_title,
+                section_path=heading_path(headings),
+                headings=headings,
+                line_start=line_no,
+                counter=counter,
+            )
+            idx = next_idx
+            continue
+
+        text_buffer.append((line_no, line))
+        idx += 1
+
+    flush_text()
+    return chunks
+
+
+def write_jsonl(path: Path, rows: Iterable[dict]) -> int:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    count = 0
+    with path.open("w", encoding="utf-8") as handle:
+        for row in rows:
+            handle.write(json.dumps(row, ensure_ascii=False) + "\n")
+            count += 1
+    return count
+
+
+def build_chunks(data_dir: Path, config: ChunkConfig) -> list[dict]:
+    counter = TokenCounter(config.model_name)
+    all_chunks: list[dict] = []
+    for path in sorted(data_dir.glob("*.md")):
+        all_chunks.extend(chunk_markdown_file(path, data_dir, counter, config))
+    return all_chunks
+
+
+def summarize(chunks: list[dict]) -> dict[str, int]:
+    summary: dict[str, int] = {"total": len(chunks)}
+    for chunk in chunks:
+        chunk_type = chunk["metadata"]["chunk_type"]
+        summary[chunk_type] = summary.get(chunk_type, 0) + 1
+    return summary
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Chunk Markdown files for a GLOF RAG chatbot.")
+    parser.add_argument("--data-dir", type=Path, default=Path("rag/data"), help="Directory containing Markdown files.")
+    parser.add_argument("--out", type=Path, default=Path("rag/artifacts/chunks.jsonl"), help="Output JSONL path.")
+    parser.add_argument("--model-name", default="BAAI/bge-m3", help="Tokenizer model used for token-aware splitting.")
+    parser.add_argument("--max-text-tokens", type=int, default=850)
+    parser.add_argument("--text-overlap-tokens", type=int, default=120)
+    parser.add_argument("--min-text-tokens", type=int, default=40)
+    parser.add_argument("--table-rows-per-chunk", type=int, default=35)
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    config = ChunkConfig(
+        max_text_tokens=args.max_text_tokens,
+        text_overlap_tokens=args.text_overlap_tokens,
+        min_text_tokens=args.min_text_tokens,
+        table_rows_per_chunk=args.table_rows_per_chunk,
+        model_name=args.model_name,
+    )
+    chunks = build_chunks(args.data_dir, config)
+    count = write_jsonl(args.out, chunks)
+    print(json.dumps({"output": str(args.out), "written": count, "summary": summarize(chunks)}, indent=2))
+
+
+if __name__ == "__main__":
+    main()

chatbot/rag/data/('1-s2.0-S0022169423002536-main', '.pdf')_extraction.md

@@ -0,0 +1,366 @@
+
+## ABSTRACT
+
+This manuscript was handled by Emmanouil Anagnostou, Editor-in-Chief.
+
+Deglaciation due to atmospheric warming has led to the formation and expansion of numerous glacial lakes, especially in the eastern Himalaya. Many of these glacial lakes are susceptible to glacial lake outburst floods (GLOFs), which can cause far-reaching impacts on downstream infrastructure and livelihoods. This study is a comprehensive assessment of GLOF susceptibility, hazard, exposure, vulnerability, and risk for four potentially dangerous glacial lakes (Bechung Tsho, Raphstreng Tsho, Thorthomi Tsho, and Lugge Tsho) located in the Lunana glacier complex of the Phochu basin in Bhutan. Exposure and risk assessments were based on modelled GLOF hydrodynamics, infrastructure data, population and housing census data. Among the four glacial lakes, Thorthormi Tsho and Lugge Tsho are relatively more susceptible to outburst floods than Raphstreng Tsho and Bechung Tsho. Outflow flood volumes from these lakes range between  $6 \times 10^5$  and  $3 \times 10^8$  m<sup>3</sup> which can potentially impact over 16,000 people, two hydropower projects, numerous other infrastructures, and agricultural land up to 150 km downstream of the lakes. The GLOF exposed elements are largely in Punakha and Wangdue Phodrang districts, which are located 90 and 100 km downstream of the Lunana glacier complex respectively. Among 17 subdistrict blocks within the basin, one (Lunana) lies in a very high GLOF risk area, while 9 others are in the high GLOF risk zone. The study highlights the importance of multi-source data in improving the knowledge of downstream GLOF risk and serves as a base for improving GLOF risk reduction strategies in high mountain regions.
+
+## 1. Introduction
+
+Glacial lakes can be potential hazard sources as outburst floods originating from these lakes can cause far-reaching damage to infrastructure and people (Allen et al., 2015; Zheng et al., 2021b). Glacial lakes are often dammed by unstable moraine or glacier ice and are surrounded by destabilizing permafrost slopes or hanging glaciers (Otto, 2019; Richardson and Reynolds, 2000). Slope failure originating in the vicinity of a glacial lake can result in a mass movement that can enter a lake, capable of displacing the lake's water creating an impulse wave that overtops the frontal dam. The instability of slopes and glacial lakes
+
+are further aggravated in young and seismically active alpine mountains (Armbruster et al., 1978), such as Himalaya, where hazard chains/cascades are common (Kirschbaum et al., 2019; Sharma et al., 2022). The rapid expansion of glacial lakes resulting from glacier retreat increases the exposure of the lakes to glacial lake outburst flood (GLOF) triggers from the surrounding catchment, where slopes are destabilized due to permafrost degradation making it more susceptible to slope failures and rockfalls (Grämiger et al., 2017). In addition, as lakes grow depending on the basin geometry, the risk of overtopping failures may increase. Thus, lake growth increases the probability of GLOFs being triggered (Rounce et al., 2017, 2016) and increases the potential flood volume
+
+E-mail address: guoqing.zhang@itpcas.ac.cn (G. Zhang).
+
+<sup>\*</sup> Corresponding author.
+
+(Fujita et al., 2013), which are among the main factors contributing to increased GLOF hazard (Westoby et al., 2015). These intrinsic and extrinsic hazard drivers can cause largely unpredictable, rapid, and deadly draining of lake water, which may transform into debris or hyper-concentrated flow as the flood entrains sediment (Richardson and Reynolds, 2000; Sattar et al., 2022).
+
+Since GLOFs are high-magnitude and largely unpredictable catastrophic events, they can cause immense damage and disruption to the lives and infrastructures in the downstream settlements (Allen et al., 2015; Carey et al., 2011; Zheng et al., 2021b). It is estimated that 15 million people globally are exposed to the threat of GLOF, with populations in High Mountain Asia the most exposed, with around 1 million people living within 10 km of a glacial lake (Taylor et al. 2023). Numerous GLOF events have occurred in the past across the world, causing damage to infrastructures and the loss of thousands of lives (Carrivick and Tweed, 2016; Veh et al., 2022). For example, the Jinweng Co GLOF event in 2020 in the southeastern Tibetan Plateau destroyed ten residential houses, washed away eight bridges, and 43.9 km of road, also causing damage to several other structures and agricultural land (Zheng et al., 2021b). In 2013, the debris flow resulting from Chorabari lake combined with cloud-burst-induced landslides and flooding in the Indian state of Uttarakhand caused over 6,000 deaths and heavy damage to infrastructure (Allen et al., 2015). In the Peruvian Cordillera Blanca, GLOFs in 1941 and 1945 killed over 6,000 people, washed away  $\sim$ 35 % of Huaraz city and other essential infrastructures, including hydroelectric power stations (Carey et al., 2011). Bhutan alone has recorded at least 18 GLOF events since the 1950s, the latest being the breaching of subsidiary lake II of Thorthormi Tsho on 20 June 2019 (Gurung et al., 2017; Komori et al., 2012; National Center for Hydrology and Meteorology [NCHM], 2019b; Rinzin et al., 2021). At least seven of these past GLOF events in Bhutan have caused varying degrees of infrastructure damage and loss of lives in the downstream settlements. The most serious was the 1994 GLOF event of Lugge Tsho, amongst best-known contemporary GLOF events in the Himalaya (Veh et al., 2019b), which led to around 21 fatalities, seriously damaged Punakha Dzong, the capital administrative center of Bhutan, and around 90 houses (Komori et al., 2012; Rinzin et al., 2021; Watanbe and Daniel, 1996). All these tragic and disastrous past GLOF experiences underpin the need for systemic flood disaster risk reduction (DRR) activities for which detailed downstream GLOF hazard and risk assessment provides the necessary first step.
+
+GLOFs often involve a cascade of different processes, variably termed a process chain or hazard cascade (Mergili et al., 2020; Zheng et al., 2021b), which can increase the spatial reach and destructive power of the event (Sattar et al., 2022). In the Himalaya, GLOFs frequently start with rapid mass movement into the lake, often due to slope failure from the mountain flank above the lake, leading to displacement waves, overtopping the dam (Komori et al., 2012; Nie et al., 2018). The overtopping waves cause moraine dam incision and dam failure resulting in a sudden discharge of lake water. As the flood propagates farther downstream, it transforms into debris flow and hyper-concentrated or secondary debris flow, depending on the geologic and topographic characteristics of the river channel and available sediment (GAPHAZ, 2017; Schneider et al., 2014; Westoby et al., 2014; Worni et al., 2014). These complex GLOF process chains present manifold challenges to accurately capturing its physical characteristics by numerical modelling and to a growing extent, the whole hazard cascade. However, emerging models can either better represent phase changes or sediment-laden events and their application to past events has increasingly enabled a better understanding of GLOFs. Some examples of commonly employed mass movement models include RAMMS (Christen et al., 2010), r.avaflow (Mergili et al., 2017), IBER (Bladé et al., 2014), FLOW3D (Flow Science, 2012), D-Claw (Iverson and George, 2014) and, primarily for clear water floods, HEC-RAS (CEIWR-HEC, 2021). Two or more of these models can be combined appropriately to model each process chain separately and sequentially (Frey et al., 2018; Schneider et al., 2014;
+
+Worni et al., 2014), or the entire process chain can be integrated into a single model framework (Mergili et al., 2017). However, due to computation cost and uncertainty of input parameters for practical application of numerically sophisticated models (Westoby et al., 2014), computationally less complex models like HEC-RAS 2D are often used to capture first-order hydrodynamics and impacts of GLOF (Maskey et al., 2020; Sattar et al., 2021a; Sattar et al., 2021b).
+
+Population and infrastructures continue to grow in GLOF-prone areas despite the increasing threat from GLOFs (Allen et al., 2022, Carrivick and Tweed, 2016, Taylor et al. 2023). This is especially true for Bhutan, where most of the settlements (National Statistics Bureau [NSB], 2018) and major hydropower projects (the mainstay of the country's economy) are concentrated along the glacier-fed river valleys; for example, Punatsangchu hydroelectric dams (PHP-I and -II) are fed by meltwater from glaciers in the Lunana region (Ministry of Economic Affairs [MOEA], 2021). Although the positive trend in the frequency of GLOFs has decayed since the 1970s (Veh et al., 2022), some studies have forecasted a 3 to 7-fold increase in GLOF hazard and downstream risk in the Himalaya due to future deglaciation (Allen et al., 2016; Zheng et al., 2021a). The assessment at the global scale has pointed out that Bhutan and Nepal have the greatest national-level economic impact from GLOF (Carrivick and Tweed, 2016), and in absolute normalized GLOF risk scores, Bhutan is ranked 13th in the world (Taylor et al. 2023).
+
+The increasing GLOF threat coupled with growing downstream exposure merits an evaluation of systematic downstream risk combining GLOF hydraulics and population and census housing data. While ample existing studies offer comprehensive information about the GLOF process chains and its effects on downstream communities (Frey et al., 2018; Kougkoulos et al., 2018; Sattar et al., 2022; Sattar et al., 2021b; Schneider et al., 2014), these studies have not included the social vulnerability, due to which the GLOF downstream risk has remained far from being understood. Although Allen et al. (2016) have included the social vulnerability to evaluate GLOF downstream risk in the Indian state of Himachal Pradesh, the GLOF risk was evaluated based on the rough estimation of GLOF paths in a GIS environment, falling short of including GLOF hydraulics and intensity. Taylor et al. (2023) assessed GLOF risk, including exposure and vulnerability, but for consistency used open access simplified data and assumptions, for example, populations within a 1 km buffer of river centerlines, and GLOF reach as a fixed 50 km downstream of a lake. Lunana glacier complex, located in the headwater of Punatsangchu has five proglacial lakes, of which four were reported to have high to very high outburst susceptibility in terms of both potential trigger mechanisms and flood volume (Nagai et al., 2017; Rinzin et al., 2021). Therefore, for the first time, this study offers a detailed GLOF risk assessment, including underlying components of hazard, vulnerability, and exposure, using multi-source data including lake bathymetry, OpenStreetMap, and population and housing census data, to improve confidence in the knowledge of GLOF susceptibility and downstream risk from these lakes. This study is guided by the following three main objectives: i) to evaluate the potential GLOF hazard under different moraine-breaching scenarios from the glacial lakes in the Lunana glacier complex; ii) to identify the exposed population, infrastructures, and agricultural land in the valley; and iii) to conduct a detailed downstream risk assessment by considering socio-economic factors along the potentially inundated areas of Punatsangchu and Phochu river basins.
+
+## 2. Study area: significance of the Lunana glacier complex
+
+Bhutan is a small landlocked country in the Himalaya, where the landscapes are largely dominated by rugged terrains and high mountains. The Himalaya stretch along the entire northern part of the country and host large reserves of glaciers. A total of 1583 ( $\sim$ 1487 km²) glaciers, of which 219 ( $\sim$ 951 km²) debris-covered, and 1364 ( $\sim$ 536 km²) cleanice glaciers have been reported in the Bhutan Himalaya (Nagai et al., 2016). Bhutanese glaciers have undergone significant area and mass loss
+
+since the 1980s. For example, the glacier area shrunk by  $23.3 \pm 0.9$  % during 1980–2010 (Bajracharya et al., 2014). Further, between 2000 and 2020, glacier mass loss occurred at the rate of 0.47 m w.e.yr $^{-1}$ , higher than in other neighboring eastern Himalayan ( $\sim$ 0.33 m w.e.yr $^{-1}$ ) and Nyainqentanglha ( $\sim$ 0.46 m w.e.yr $^{-1}$ ) regions (Hugonnet et al., 2021). Under the future global 1.5 °C warming scenario, the eastern Himalaya, where Bhutan is located, is projected to warm by  $\sim$  1.9 °C, which is predicted to cause a  $\sim$  60 % of glacier mass loss by the end of 21st century (Kraaijenbrink et al., 2017).
+
+Region-wide studies such as the Himalayan scale (Nie et al., 2018), High Mountain Asia (Wang et al., 2020), and Third Pole (Zhang et al., 2015), all have indicated that the eastern Himalaya is one of the regions with the highest concentration of glacial lakes. Using high-resolution ALOS PRISM and AVNIR-2 satellite imagery from 2006 to 2011, Nagai et al. (2017) documented 733 glacial lakes (~83.6 km²) in the Bhutan Himalaya, identified 226 lakes with positive potential flood volume (PFV) (0.04–33.7  $\times$  106 m³), 33 as having larger PFV (>5  $\times$  106 m³), and two with very large-scale PFV (>30  $\times$  106 m³). A more recent inventory (Rinzin et al., 2021) using Sentinel-1, Sentinel-2, and declassified Corona KH-4 data reported 2574 (156.63  $\pm$  7.9 km²) glacial lakes in 2020, from which 1118 (~82.92 km²) were classified as glacier-fed lakes. Their inventory also revealed that Bhutanese glacial lakes have expanded by ~26.5 km² (+17 %) since the 1960s (Rinzin et al., 2021). Further, 65 lakes were found to have high outburst susceptibility based
+
+on multiple attributes such as lake area and the likelihood of mass movement into the lake (Rinzin et al., 2021). The two inventories reported different numbers and statistics of glacial lakes in the Bhutan Himalaya due to differences in the approach and criteria employed in mapping glacial lakes and defining potentially dangerous glacial lakes. However, there is a common consensus that the Phochu basin, where the Lunana glacier complex is located contains the largest proportion of glacial lakes and the majority of potentially dangerous glacial lakes (PDGLs). For example, Rinzin et al. (2021) reported the highest proportion (22.6 %) of glacier-fed lakes in the Phochu basin and identified it as the inland basin with the highest number of PDGLs. The two glacial lakes with larger-scale potential flood volume identified by Nagai et al. (2017) are both located in the Phochu basin. Likewise, nearly half (8) of 17 PDGLs in Bhutan, first identified by ICIMOD (Mool et al., 2001) and later verified by NCHM (2019) are located in the Phochu basin. Bhutan is a dense cluster of glacial lakes in the Himalaya, and the Phochu basin is a global hotspot of potential GLOF risk (Taylor et al., 2023).
+
+Four proglacial lakes, namely, Bechung Tsho, Raphstreng Tsho, Lugge Tsho, and Thorthormi Tsho in the Lunana glacier complex of Phochu basin in the Bhutan Himalaya were selected for detailed GLOF susceptibility evaluation and downstream hazard and risk assessment (Fig. 1). These four glacial lakes are the primary source of water for the Phochu river, which coalesces with Mochu and forms Punatsangchu, one of the biggest rivers in Bhutan. A first-order hazard assessment had
+
+Image /page/2/Figure/6 description: A detailed topographical map of the Lunana glacier complex and downstream settlements in Bhutan. The main map, labeled (a), shows a mountainous region with elevation ranging from 352 to 8590 meters. It displays rivers, lakes, and Gewog boundaries. Key settlements along a river are numbered: 1. Lunana, 2. Punakha, 3. Wangdue Phodrang, 4. PHP-I, and 5. PHP-II. Two inset maps in the top-left corner show the location of the area within High Mountain Asia and Bhutan. On the right, four panels, (b) through (e), show close-ups of specific lakes from the Lunana complex with their respective areas: (b) Area: 0.5 km², (c) Area: 1.3 km², (d) Area: 4.3 km², and (e) Area: 1.6 km². The map includes latitude and longitude lines and a scale bar for 20 km.
+
+Fig. 1. Lunana glacier complex and downstream settlements. Map (a) shows settlements downstream of the Lunana glacier complex along the Phochu and Punatsangchu basins. Downstream settlements are presented following the subdistrict block administrative boundary. Panels (b–e) show four glacial lakes, which we studied here: Bechung Tsho, Raphstreng Tsho, Lugge Tsho, and Thorthormi Tsho, respectively. Background image: ALOS-PALSAR DEM.
+
+already revealed that all these lakes are highly (susceptibility level higher than 0.8) or very highly (susceptibility level higher than 0.6) susceptible to outburst flood (Rinzin et al., 2021). The Lunana glacier complex is also the catchment where at least three past GLOF events have been recorded since 1990, for example, the most devastating 1994 GLOF from Lugge Tsho (Watanbe and Daniel, 1996), and the recent breaching of subsidiary lake II of Thorthormi Tsho in July 2019 (National Center for Hydrology and Meteorology [NCHM], 2019a). Although the artificial draining of Thorthormi Tsho initiated in 2008 reduced the lake water level by 5 m, the lake is still deemed as highly dangerous (NCHM, 2019b). The investigation with persistent scatterer interferometry (PSI) has revealed an unstable left moraine wall for Thorthormi Tsho and significant movement of the side moraine wall of Lugge Tsho (Wangchuk et al., 2022). Although assessment with PSI by Wangchuk et al. (2022) has found no significant movement of the thin side moraine between Thorthormi Tsho and Raphstreng Tsho, a detailed field investigation in 2019 has reported active sliding of this thin sidemoraine (NCHM, 2019a), suggesting a threat of a future cascading flood from Thorthormi Tsho and Raphstreng Tsho. Lunana, the only community-dwelling within the proximity of glaciers in Bhutan, is located about 5-20 km away from these glacial lakes. Punakha Dzong, the former capital, with significant cultural heritage, is situated at the confluence of Phochu and Mochu, about 90 km downstream. Punatsangchu basin also hosts the two largest hydroelectric projects, Punatsangchu hydroelectric project (PHP)-I (1200 MW) and -II (1020 MW), at 110 and 125 km downstream respectively which are expected to be commissioned soon (MOEA, 2021). Taylor et al. (2023) state that the Punatsangchu catchment is the 4<sup>th</sup> most at risk from GLOF catchments in the world. Located along the Phochu and Punatsangchu riverbanks there are also two district capital towns: Khuruthang town (90 km downstream) for Punakha and Bajo town (100 km downstream) for Wangdue Phodrang districts. This high GLOF potential combined with high downstream exposure means that investigating the hazard and risk status becomes critically important (Fig. 1).
+
+## 3. Data and method
+
+### 3.1. Remote sensing and field data
+
+Declassified satellite data, Corona KH-4 and KH-9 were used to manually digitize glacier and glacial lake boundaries in the 1960s and 1970s, respectively (Table 1). To correct the image distortion of Corona and Hexagon data, the co-registration using ground control points from Sentinel-2 imagery was performed (see Rinzin et al., 2021). A Landsat series (TM, ETM, and OLI) were used to map glacial lakes and glaciers from 1990 to 2020. Advanced Land Observing Satellite (ALOS)-Phased Array type L-band Synthetic Aperture Radar (PALSAR) DEM with 12.5 m spatial resolution was used as a source of terrain information for the GLOF routing. We modified terrain data along the river flow path by adding a line vector and cutting a channel through the artificial flood plain using the Terrain Modification Tool in the RAS Mapper (see supplementary material for detail). The ALOS-PALSAR DEM was also used to map the lake surrounding topographic potential for ice and rock avalanching and steepness of the moraine dam. The bathymetry data for three glacial lakes (Bechung Tsho, Rapshtreng Tsho, and Lugge Tsho) in the Lunana glacier complex were collected during the field expedition from 15 September to 20 October 2021. A Hummingbird 999SI combo (echo sounder) with an inbuilt Global Positioning System was used to collect the bathymetry data. The echo sounder transmits sound waves coupled with coordinate information into the lake bottom and records the returning echoes on a sonar chart with depth information. The lake depth data collected along the transverse transects were interpolated to the entire lake surface. The field-measured lake bathymetry was used to calculate the flood breach volume and determine the bed topography of the lakes.
+
+We used the Hindu Kush Himalaya-scale Landsat-based land cover
+
+**Table 1**Summary of all data used for lake outburst flood susceptibility assessment, downstream hazard, exposure, vulnerability, and risk assessment.
+
+| No. | Period         | Source                        | Resolution (m) | Purpose                                                                                                        |
+|-----|----------------|-------------------------------|----------------|----------------------------------------------------------------------------------------------------------------|
+| 1   | 1965–1968      | Corona KH-4                   | ~2.7           | Map glacial lakes and glaciers in the 1960 s                                                                   |
+| 2   | 1973–1974      | Hexagon KH-9                  | 6–9            | Map glacial lakes and glaciers in the 1970 s                                                                   |
+| 3   | 1990–2020      | Landsat (OLI, ETM, TM)        | 30             | Map glaciers and glacial lakes during 1990–2020                                                                |
+| 4   | 2006–2011      | ALOS PALSAR                   | 12.5           | Source of terrain data<br>for multi-scenario flood<br>modelling and lake<br>surrounding topography<br>analysis |
+| 5   | 2022/02/<br>18 | OpenStreetMap                 | NA             | Quantify structures<br>exposed to the<br>simulated GLOFs                                                       |
+| 6   | 2000 &<br>2018 | Land use and land<br>cover    | 30             | Determine Manning's<br>coefficient and area of<br>agricultural land (<br>ICIMOD, 2021)                         |
+| 7   | 2017           | Population and housing census | NA             | Determine downstream<br>settlement social<br>vulnerability level (<br>NSB, 2018)                               |
+| 8   | 2010–2020      | Glacier mass<br>balance       | 100            | Determine glacier mass<br>balance (Hugonnet<br>et al., 2021)                                                   |
+| 9   | 2000–2018      | Glacier velocity<br>data      | 240            | Determine glacier flow<br>velocity (Gardner et al.,<br>2019)                                                   |
+| 10  | 2021           | Lake bathymetry               | NA             | Calculate potential<br>flood volume                                                                            |
+
+data of 2018 (from ICIMOD) to quantify the area of agricultural land inundated by the modelled GLOF (ICIMOD, 2021). The same data was also used to derive pixel-based Manning's n value for flood routing. We used time-series of Land Ice Velocity and Elevation (ITS\_LIVE) (Gardner et al., 2019) to calculate the flow velocity of glaciers. The glacier mass loss was calculated using the global-scale glacier mass loss dataset available at a  $1^{\circ} \times 1^{\circ}$  grid scale (Hugonnet et al., 2021). The global-scale consensus ice thickness datasets and corresponding surface DEM (Farinotti et al., 2019) were also used to calculate glacier bed overdeepening and determine future glacial lake extent.
+
+OpenStreetMap data was used to count the number of buildings exposed to the multiple GLOF scenarios modelled in the study (data as of 15 June 2022). Bhutan population and housing 2017 census data were used to determine the social vulnerability level of the exposed population (NSB, 2018). The 2020 livestock statistics from the Ministry of Agriculture and Forestry, Bhutan (Renewable Natural Resources Statistics Division [RNR-SD], 2020) were used to quantify the number of livestock exposed to the hazard from GLOFs. The granularity of these data will be further discussed below.
+
+### 3.2. Glacial lake outburst susceptibility assessment
+
+The GLOF susceptibility assessment for the lakes in the Lunana glacier complex was performed based on a total of 18 conditioning and triggering factors, broadly categorized into atmospheric, cryospheric, geotechnical, and geomorphic factors (after GAPHAZ, 2017) (see Supplementary Table S1). Each factor was either observed or quantified using the presented data, including remote sensing, DEM, and field observation.
+
+The topographic potential of the surrounding slopes for landslide/rockfall and ice/snow avalanche (slope greater than or equal  $30^{\circ}$  and run-out trajectory slope greater than or equal to  $14^{\circ}$ ) were mapped using a GIS-based modelling approach (Allen et al., 2019; Romstad et al., 2008) (Fig. 2). Steep Lakefront Area (SLA) (depression angle  $>10^{\circ}$  from
+
+Image /page/4/Figure/2 description: A methodological flowchart is presented in six numbered sections, detailing the process for assessing Glacial Lake Outburst Flood (GLOF) susceptibility, hazard, exposure, vulnerability, and risk.
+
+1. Glacial lake outburst flood susceptibility: This flowchart shows that data from Landsat series, Hexagon KH-9, Corona KH-4, ALOS-PALSAR DEM, Google Earth Imagery, and Field photos are used to analyze factors like glacier & lake area change, mass balance, glacier velocity (leading to cryospheric factor), topographic potential area, moraine dam stability, and catchment area. These analyses inform intermediate factors such as calving front, avalanche, rockfall/landslide, and geotechnical & geomorphological factors, which all contribute to determining GLOF susceptibility.
+
+2. Dam breach and GLOF routing: This section shows that ALOS-PALSAR DEM and Bathymetry Survey data are used to calculate peak discharge, flood velocity, and flood depth, which together determine the flood intensity.
+
+3. Hazard: A 2x3 matrix shows the hazard level based on Intensity (High, Medium) and Probability (High, Medium, Low). The combinations result in High (H), Medium (M), or Low (L) hazard levels, represented by red, orange, and green colors respectively. For example, High Intensity and High Probability result in a High Hazard level.
+
+4. Exposure: A flowchart indicates that flood intensity affects people, buildings, agricultural land, motorways, and bridges. The overall exposure is rated as High (H), Medium (M), or Low (L), shown in a red, orange, and green bar.
+
+5. Vulnerability: This section shows that population and housing census data are used to create a social vulnerability index, which is then rated as High (H), Medium (M), or Low (L) in a red, orange, and green bar.
+
+6. Risk: A 3x3 matrix determines the risk level based on Exposure (High, Medium, Low) and Hazard (High, Medium, Low). The matrix also aligns Vulnerability levels with Exposure levels. The resulting risk is categorized as High (H), Medium (M), or Low (L), represented by dark red, orange, and green squares. For instance, High Exposure and High Hazard result in a High Risk level.
+
+Fig. 2. Methodological flow chart adopted for GLOF susceptibility, downstream hazard, exposure, vulnerability, and risk assessments for the potentially dangerous glacial lakes in the Lunana glacier complex of the Bhutan Himalaya: downstream GLOF hazard, vulnerability, exposure, and risk are categorized into three classes; high (H), medium (M) and low (L). For the hazard computation, the GLOF intensity used here is translated from the modelled flood depth: high (more than 1 m) and medium (less than 1 m) (after GAPHAZ, 2017).
+
+the flat lake surface), which provides information on the steepness of the moraine dam, was computed following the approach of Fujita et al. (2013). Hydrological characteristics such as total area of catchment and stream density were calculated from AlOS-PALSAR DEM. Other attributes, such as calving potential, which require finer resolution observation, were validated based on expert interpretation of Google Earth imagery and field photos (Fig. 3). The parent glacier extent was vectorized by modifying RGI v6.0 glacier outlines (RGI Consortium, 2017) and through our expert observation in a false-color composite of Landsat series imagery (OLI, TM, and ETM) for 1990–2020, Corona KH-4 for the 1960s, and Hexagon KH-9 for the 1970s. The same scenes were also used for digitizing the glacial lake boundaries manually. The glacier bed overdeepenings, as an indicator of future lake extent, were calculated by subtracting glacier ice thickness from the surface DEM (Farinotti et al., 2019) following Linsbauer et al. (2016).
+
+Based on the degree to which these factors can condition or trigger a GLOF from a given lake, inferred through either expert judgment or calculated value, the above factors were classified into high, medium, and low susceptibility levels (Allen et al., 2022; GAPHAZ, 2017) with the corresponding score of 3, 2, 1, respectively. The total GLOF susceptibility score of the lake was calculated as the sum of scores across all factors. The susceptibility score was finally used to compare the relative outburst susceptibility level of the four lakes considered here.
+
+### 3.3. Dam breach and GLOF routing
+
+Hydrologic Engineering Center's River Analysis System (HEC-RAS v
+
+6.0), a 2-dimensional (2D) dam breach model was used to simulate a series of potential GLOF events from the lakes in the Lunana glacier complex. Although limited to clear-water flow assumptions, HEC-RAS has been widely used for hydrodynamic modelling of GLOF in the Himalaya, such as in Nepal (Khanal et al., 2015; Sattar et al., 2021b), Tibetan Plateau (Wang et al., 2015), India (Sattar et al., 2021a) and Bhutan (Maurer et al., 2020; Osti et al., 2013). The model solves the 2D Saint Venant equation to calculate the unsteady flow hydraulics of GLOFs (CEIWR-HEC, 2021). The sine wave progression breach model was employed where initial breach forms slowly and with a further gradual incision the outflow velocities and shear stress change which leads to an increase in the outflow discharge (Sattar et al., 2021b). For the majority of moraine-dammed lakes, the outburst events are caused by displacement waves from mass movement followed by accompanying breach development and dam break (GAPHAZ, 2017). Therefore, we assumed that the dam breach was instigated by overtopping due to the mass movement into the lake, with subsequent dam incision and degradation leading to the outburst event. In HEC-RAS this is implemented with the dam failure mode considered as overtopping, where dam breach starts at the top of the dam and progresses toward the base of the moraine as erosion occurs (CEIWR-HEC, 2021). Breach parameters like breach width and formation time for each GLOF scenario were calculated using Froehlich (1995) (Eqs. (1) and (2)). This set of equations has been used previously in calculating breach parameters for GLOF modelling (Anacona et al., 2015; Sattar et al., 2021b; Wang et al.,
+
+Image /page/5/Figure/2 description: A figure displaying four photographs of different glacial lakes, each with highlighted geological hazards. A legend at the top defines the colored, dashed boxes used in the photos: black for 'Calving front', blue for 'Avalanche', orange for 'Landslide/Rockfall', and purple for 'Unstable moraine'.
+
+The top-left photo is of 'Bechung Tsho' on 11/10/2021, showing highlighted areas for avalanches and landslides/rockfalls, with an inset image pointing to the calving front.
+
+The top-right photo is of 'Lugge Tsho' on 8/10/2021, indicating a calving front and a landslide/rockfall area, with an inset showing another view with avalanche and landslide zones.
+
+The bottom-left photo is of 'Raphstreng Tsho' on 11/10/2021, with several areas marked for potential landslides/rockfalls and an avalanche.
+
+The bottom-right photo is of 'Thorthormi Tsho' on 11/10/2021, showing potential avalanche and landslide/rockfall zones, with an inset highlighting an unstable moraine.
+
+Fig. 3. Field photographs show the surrounding conditions that make lakes susceptible to GLOF, including calving front, avalanche, landslide/rockfall, and unstable moraine dam. Photo courtesv: NCHM. Bhutan.
+
+$$B_{w} = 0.1803k_{0}(V_{w})^{0.32}(h_{b})^{0.14}$$
+(1)
+
+$$T_f = 0.00254 (V_w)^{0.53} (h_b)^{-0.9}$$
+ (2)
+
+where  $B_w$  is breach width,  $T_f$  is breach formation time,  $V_w$  is breach volume, and  $h_b$  is breach height.
+
+For the forward modelling, a series of 12 GLOF scenarios were simulated, considering large, moderate, and small magnitude events originating from each lake: Bechung Tsho, Raphstreng Tsho, Thorthormi Tsho, and Lugge Tsho. For each simulation, a unique set of parameters were considered: breach width  $(B_w)$ , breach formation time  $(T_f)$  and breach volume  $(V_w)$ , and breach height  $(h_b)$  while keeping the downstream valley geomorphology and river hydrological conditions constant. For most past GLOF events in the Himalaya, it is unlikely that the entire moraine fails in case of an outburst event (Maurer et al., 2020; Veh et al., 2019a). Therefore, for the high magnitude GLOF scenario routing, we constructed lake breach models assuming  $h_b$  starting at the crest of the dam down to the mean depth of the lakes (calculated based on the field-measured bathymetry). The  $h_b$  for the moderate magnitude is assumed to be half of the large-magnitude  $h_b$ . Likewise, the smallmagnitude  $h_b$  was calculated as half of the moderate magnitude. Breach volumes were calculated from the field bathymetries as the volume of the water above the lowest point of breaching following the conservative trend in the previous studies, such as of Sattar et al. (2021b). However, owing to the lack of bathymetry data, the  $V_w$  for Thorthormi Tsho was calculated as the product of area and potential lowering height (Hp) following (equation (3)) Fujita et al. (2013). Hp is the anticipated water lowering height after SLA is removed due to moraine dam failure (Fujita et al., 2013). Here, it is worth noting that the GLOF was simulated as a clear water flood and did not account for the complex process chain of a GLOF, such as debris loading from erosion and sediment deposition along the downstream channel. We used pixelbased Manning's n value according to the latest Landsat-derived land use and land cover data (LULC) (ICIMOD, 2021). Depending on different classes of LULC, the pixel-based Manning's *n* ranges from 0.035 to 0.120.
+
+$$PFV = \min[Hp; Dm]A \tag{3}$$
+
+where, *PFV* is potential flood volume, *Hp* is potential height lowering, *Dm* is the mean depth of the lake, and A is a lake area.
+
+We reconstructed the 1994 GLOF event of Lugge Tsho to understand the GLOF hydrodynamics in the Punatsangchu basin before conducting the multi-scenario modelling from the lakes in the Lunana glacier complex. There are limited field and eyewitness information on the 1994 GLOF event, so we drew upon sparse data available in the previous literature to define essential input parameters and initial conditions. For example, the moraine dam breach height was set to 18 m based on the height estimated by Fujita et al. (2008). Likewise, the dam breach width was set to 105 m and the dam crest width to 28 m both of which were constrained by Koike and Takenaka (2012). Combining field measurements and remote sensing observations, Fujita et al. (2008) have estimated a GLOF volume of 17.2  $\pm$  5.3  $\times$  10<sup>6</sup> m<sup>3</sup>. Accordingly, we simulated six scenarios of GLOF by changing flood volume between 15  $\times$  10<sup>6</sup> and 20  $\times$  10<sup>6</sup> m<sup>3</sup> (each time increasing by 1 million m<sup>3</sup>) and keeping all other initial breach parameters constant. Other essential information available for the 1994 Lugge GLOF is the Peak flow at Wangdue Phodrang hydrological station (1800–2500 m<sup>3</sup>/s) (Yamada, 2000; Osti et al., 2013), and flood wave arrival time at Punakha through eyewitness (5.1 hr) and estimated using seismometer (4.75 hr) (Maurer et al., 2020; Watanbe and Daniel, 1996). So, we measured peak flow at Wangdue Phodrang hydrological station and flood reach time at Punakha.
+
+### 3.4. Downstream risk assessment
+
+The risk from any disaster event is represented as a function of hazard, exposure, and vulnerability (IPCC, 2014). Here, GLOF modelling
+
+results were first translated into a hazard map. In doing so, the flow height from each scenario modelled GLOF was converted into two classes of intensity: high intensity (more than 1 m) and medium intensity (less than 1 m) (Frey et al., 2018; GAPHAZ, 2017). Later, the hazard map was computed by combining these intensities with the probability of occurrence (low, medium, and high corresponding to potential large, moderate, and small magnitude events, respectively). We combined intensity and probability maps by using the simple matrix approach (GAPHAZ, 2017). Since all the lakes examined here are situated in the same basin and pose a threat to the downstream regions, we evaluate the combined hazard for all four glacial lakes in the Lunana glacier complex. This final hazard map was created by initially producing four GLOF hazard maps (one for each lake) with three hazard levels given as high, medium, and low. Later, a simple matrix technique was used to merge the individual lake's hazard map. The combined hazard map is classified into four hazard levels: very high (when all (four) high hazard level from each lake intersects), high (when at least three high hazard level intersect or two high hazard level and two medium levels intersect), medium (at least two medium hazard level intersect) and low (when all low hazard levels intersect).
+
+Exposure is the presence of people, livelihood, environmental services, and other infrastructure along the flood plain (Allen et al., 2019). Flood exposure calculation involves two components: delineating the flood hazard map and the population, structures, livestock, and agricultural land intersecting with that flood hazard map (Tate et al., 2021). The OpenStreetMap data were used to quantify any buildings, roads, and bridges that intersect with the extent of flood hazard under different scenarios. While there was good coverage of OpenStreetMap in our study area, we manually updated 75 buildings based on high-resolution Google Earth imagery. To estimate the GLOF exposed population we divided the total population by the number of buildings in the 17 identified subdistrict blocks to roughly appraise the number of people per exposed building, following Kougkoulos et al. (2018). The population data were extracted from Bhutan population and housing census 2017 data (NSB, 2018). The same approach was used to calculate the total number of exposed livestock using Bhutan's annual livestock statistics 2020 from the Ministry of Agriculture and Forestry (RNR-SD, 2020). The inundated agricultural land was delineated by overlaying GLOF extent on the 2018 Landsat 8-based LULC (ICIMOD, 2021). The exposure level for each subdistrict block was classified into three classes depending mainly on the number of exposed buildings as high (more than 50), medium (10–50), and low (up to 10 buildings with agricultural land, road, and bridge damage). We identified buildings as highly important since their destruction often directly relates to human casualties during disasters like GLOF (Allen et al., 2016).
+
+Based on Bhutanese population and housing census 2017 data, 15 social vulnerability indicators, and their values were selected, representing a wide range of social and economic conditions of the people (NSB, 2018) (Table 2). Here, we did not differentiate the indicator into positive or negative dependence, since we uniformly chose indicators with positive dependence, where increasing indicator value increases vulnerability. For example, we chose the percentage of households without a car for the indicator, household ownership of the car. Each indicator value of the subdistrict block (Table S2) was normalized to a value ranging from 0 (low vulnerability) to 1 (high vulnerability) with impartial weighting. The individual vulnerability index (VI) of each subdistrict block was calculated as the average across all normalized scores. The vulnerability index for each subdistrict block was also classified into very high (0.75-1), high (0.5-0.75), medium (0.25-0.5), and low (0-0.25). Finally, hazard, vulnerability, and exposure were combined using the simple matrix method to produce the risk level of each subdistrict block (Fig. 2). Here, we treated hazard, exposure, and vulnerability equally and no weighting was used.
+
+### Table 2
+
+List of indicators and corresponding represented components considered for calculating the social vulnerability index for the downstream settlements. The indicators are extracted from the Bhutan Population and Housing Census 2017 data (NSB, 2018), following the convention used by Allen et al. (2016) and Cutter et al. (2003). The absolute and normalized values for each indicator are also provided in supplementary Table S2.
+
+| No. | Social vulnerability indicator                                            | Component represented                                    |
+|-----|---------------------------------------------------------------------------|----------------------------------------------------------|
+| 1   | Old age dependence ratio                                                  | Sensitivity, capacity to prepare,<br>respond and recover |
+| 2   | Child dependence ratio                                                    | Sensitivity, capacity to prepare,<br>respond and recover |
+| 3   | Disability prevalence rate                                                | Sensitivity, capacity to prepare,<br>respond and recover |
+| 4   | Female population                                                         | Sensitivity, capacity to prepare,<br>respond and recover |
+| 5   | Household without high school education<br>attainment (adequate literacy) | Capacity to prepare, respond<br>and recover              |
+| 6   | Household with insufficiency in food                                      | Sensitivity, capacity to prepare,<br>respond and recover |
+| 7   | Rented household                                                          | Capacity to recover                                      |
+| 8   | Household without a reliable source of<br>water                           | Capacity to prepare, respond<br>and recover              |
+| 9   | Household without an improved<br>sanitation                               | Capacity to recover                                      |
+| 10  | Household with more than a 30-minute<br>walk to the road point            | Capacity to prepare, respond<br>and recover              |
+| 11  | Household without a smartphone                                            | Capacity to prepare and<br>respond                       |
+| 12  | Household living in the house without<br>housing concrete wall            | Capacity to recover                                      |
+| 13  | Household without a car                                                   | Capacity to prepare, respond<br>and recover              |
+| 14  | Household without a TV                                                    | Capacity to prepare and<br>respond                       |
+| 15  | Household without electricity as a main<br>source of energy               | Capacity to prepare and<br>respond                       |
+
+## 4. Results
+
+### 4.1. Outburst susceptibility of glacial lake
+
+The total GLOF susceptibility score (Sc) for the Thorthormi Tsho (Sc = 47) and Lugge Tsho (Sc = 45) was comparatively higher than the Bechung Tsho (Sc = 37) and Raphstreng Tsho (Sc = 34), suggesting that the former two lakes have higher chances of producing GLOF than the latter two. Since all lakes are located within the small catchment (99.3 km<sup>2</sup>), they had uniform susceptibility scores of 8 resulting from atmospheric factors. Seven of the 18 susceptibility factors considered here were cryospheric factors. With a total score of 17, Thorthormi Tsho had the highest susceptibility score from cryospheric factors, followed by Lugge Tsho with 16. This high score is because, firstly, the likelihood of ice avalanche, an important cryospheric factor, was highest for the Thorthormi Tsho as its total area susceptible to avalanche (4.3 km<sup>2</sup>) is higher than other lakes, almost by 10-fold. Likewise, for the Lugge Tsho, the meltwater ponds on the moraine indicated a possible degradation of the ice core or permafrost in its terminal moraine dam. Lugge Tsho is also in direct contact with the parent glacier and has a calving front measuring up to 360 m (Fig. 3) indicating the high possibility of impact from the calving. The Thorthormi glacier surface flow velocity (9.37 m a<sup>-1</sup>), especially in the lower terminus, is almost 2–3-fold higher than other glaciers. The other cryospheric factors, such as frontal glacier retreat, mass loss, and lake expansion, were also higher for Thorthormi and Lugge Tsho (Fig. 4 and Supplementary Table S1). Additionally, the modelling of glacier bed topography suggested that Lugge Tsho will expand up to 1.83 km<sup>2</sup> (+11.6 %) Supplementary Fig. S3).
+
+The total susceptibility scores across nine geotechnical and geomorphologic factors ranked Lugge Tsho with the highest GLOF susceptibility level, followed by Thorthormi Tsho with a score of 23 and 20, respectively. Again, this higher score is because most of the calculated
+
+Image /page/7/Figure/2 description: This is a figure from a scientific paper, labeled "Fig. 4", illustrating glacier retreat and glacial lake expansion between 1966 and 2020 for four different locations: Rapstreng, Thorthormi, Lugge, and Bechung. The figure is organized into four rows, one for each location. Each row contains a series of four satellite images from different years (1966, 1990, 2010, 2020, with specific dates in 1994 for Lugge) and a corresponding line graph. The images visually show the decrease in glacier size (outlined in blue) and the increase in lake size (outlined in yellow) over time. The line graphs plot Lake area (km²) in green on the left y-axis and Glacier length (km) in blue on the right y-axis, against the year on the x-axis. For Rapstreng, the lake area grows from near 0 to 1.25 km², while the glacier length shrinks from about 5.5 to 4.2 km. For Thorthormi, the lake area increases from 0 to about 3.5 km², and the glacier length decreases from about 9.2 to 6.2 km. For Lugge, the graph shows a sharp drop in lake area on 7/10/1994 due to an outburst flood, after which it begins to increase again. For Bechung, the lake area grows from 0 to 0.4 km², while the glacier length decreases from about 6.6 to 5.8 km. A legend at the bottom clarifies that the blue outline represents the glacier and the yellow outline represents the lake.
+
+Fig. 4. Glacier retreat and glacial lake expansion between 1966 and 2020 in the Lunana glacier complex. Lugge Tsho area recession after the 1994 outburst flood is also shown. The background image: Corona KH-4 (1968) and Landsat series (1990–2020).
+
+values and expert judgment for these factors were in favour of Lugge Tsho and Thorthormi Tsho, as discussed below. All four glacial lakes in the Luanna complex have the most imminent threat from probable mass movements from the surrounding slopes (clearly visible in field photos; Fig. 3). By comparison, the total area susceptible to the mass movement for Thorthormi Tsho is  $10.8 \text{ km}^2$  which is 2–4-fold higher than other lakes making it relatively more susceptible to the impact of the mass movement (Supplementary Table S1). The individual catchment area for the lake ranges from 11.8– $46.2 \text{ km}^2$ . However, owing to the largest catchment area ( $46.2 \text{ km}^2$ ) and high stream order, Lugge Tsho appears to be most exposed, allowing us to assign higher scores for the Lugge Tsho (Supplementary Table S1). Likewise, the field photos showed that Thorthormi Tsho's moraine dam width, especially toward Raphstreng Tsho (right side), is very thin (30–40 m) (Fig. 3). On the other hand,
+
+Raphstreng Tsho's moraine width is thick (500 m) as observed in Google Earth high-resolution imagery. Moraine damming Bechung Tsho seems very wide (700 m) with a good vegetation cover, seemingly the most stable among the lakes in the Lunana glacier complex.
+
+### 4.2. GLOF discharge and hydraulics: scenario modelling of the lakes in the Lunana glacier complex
+
+The glacial lakes in the Lunana glacier complex have potential flood volumes ranging from  $6\times 10^5$  to  $3\times 10^8$  m³, which could produce floods with a peak discharge of 200–76,000 m³/s. Thorthormi Tsho has a comparatively higher GLOF magnitude than the other three lakes (Fig. 5). The resulting peak flow mainly depends on the flood magnitude at the starting point and gradually decreases as the flood routes
+
+Image /page/8/Figure/2 description: A figure from a scientific paper illustrating breach parameters and outflow hydrographs for four different glacial lakes (Tsho) under three different scenarios: Large, Moderate, and Small. The figure is a 4x3 grid where rows represent the lakes (Thorthormi Tsho, Lugge Tsho, Raphstreng Tsho, Bechung Tsho) and columns represent the scenarios. Each cell contains a schematic of a dam breach showing breach width (Bw) and height (Hb), along with a corresponding hydrograph of discharge versus time. For Thorthormi Tsho, the parameters for Large, Moderate, and Small scenarios are respectively: V=303.5, 151.5, 75.8 (x10⁶ m³); T=1.7, 2.2, 2.9 hr; Bw=293.7, 206.2, 144.8 m; Hb=70.8, 35.4, 17.7 m; with peak discharges around 75,000, 22,000, and 9,000 m³/s. For Lugge Tsho, the parameters are: V=80.2, 40.1, 20.1 (x10⁶ m³); T=1.2, 1.7, 2.0 hr; Bw=178.1, 126.0, 87.9 m; Hb=48, 24, 12 m; with peak discharges around 30,000, 12,000, and 6,000 m³/s. For Raphstreng Tsho, the parameters are: V=61.1, 30.5, 15.3 (x10⁶ m³); T=1.1, 1.4, 1.8 hr; Bw=161.6, 113.5, 79.7 m; Hb=45.4, 22.7, 11.4 m; with peak discharges around 25,000, 10,000, and 5,000 m³/s. For Bechung Tsho, the parameters are: V=4.9, 2.4, 0.6 (x10⁶ m³); T=1.1, 1.4, 1.8 hr; Bw=54.5, 41.0, 26.9 m; Hb=10.4, 5.2, 2.6 m; with peak discharges around 1200, 450, and 200 m³/s. The hydrographs for Large and Moderate scenarios span 200 minutes, while those for the Small scenario span 400 minutes.
+
+**Fig. 5.** Illustration of breach parameters: breach width  $(B_w)$ , breach height  $(H_b)$ , breach formation time  $(T_f)$ , and outflow hydrograph at the dam for four typical proglacial lakes in the Lunana glacier complex under three different outbursts GLOF magnitude scenarios: large, moderate, and small.
+
+downstream. For example, peak flow from Thorthormi Tsho at Lunana, 10 km from the Lunana glacier complex, is 9000–55,000 m<sup>3</sup>/s while it decreases to 4200-20,000 m<sup>3</sup>/s when the flood arrives PHP-II, about 125 km downstream (Fig. 6). The flood from Bechung is the smallest of all, attenuating before reaching Toedwang in Punakha, the first major settlement after Lunana. So, the flood hydraulics hereafter is presented only for the other three lakes (Raphstreng Tsho, Lugge Tsho, and Thorthormi Tsho). Unlike peak flow, the disparities in inundation depth and velocity at different downstream sites were also influenced by the channel topography and flood plain (Fig. 7). All 12 simulated GLOFs arrive at Lunana (which is about 10 km downstream) between 0.5 hr (Thorthormi large scenario) to 4.5 hr (Bechung small scenario) (Table 3), and depth ranges between 1 m and 25 m, while flow velocity is 3–14 m s<sup>-1</sup> (Table 3). The nine GLOFs (3 each from Raphstreng Tsho, Lugge Tsho, and Thorthormi Tsho) take ~2.5 to 4.75 hr to reach Punakha Dzong and flow with a velocity of 4.7–8.0 m s<sup>-1</sup> and flow depth of 6.9-27.8 m (Table 3). The flood arrival time increases to 3.0-6.0 hr as the floods arrive at Wangdue Phodrang hydrological station. Here, due to the low gradient (slope =  $\sim 0.02^{\circ}$ ) flood path, the floods flow with the minimum velocity  $(1.7-3.2 \text{ m s}^{-1})$  while their depth becomes the maximum (8.0-29.2 m) (Table 3). The flood proceeds farther downstream and arrives at PHP-I within 4.0 to 8.25 hr with a flow velocity of 1.7– $7.0~m~s^{-1}$  and depth of 9.0–26.2~m~(Table~3). Taking around 4.5-10.5 hr, the flood flows at 3.7 m s<sup>-1</sup> at PHP-II while their depth ranges between 7.0 and 29.0 m (Table 3). The detailed flow velocity and arrival time across various stations along the river centreline is also indicated in Fig. 7.
+
+### 4.3. Downstream hazard, exposure, and risk
+
+GLOF downstream risk was computed by combining hazard, exposure, and vulnerability. Fig. 8 is the excerpt of the hazard map from Lunana and Wangdue Phodrang. The simulated flood (up to 150 km downstream) from the lakes in the Lunana glacier complex showed that the flood would inundate about 42.8 km<sup>2</sup> downstream area, out of which 15.6 % fall under the very high hazard zone while 42.4 %, 14.5 %, and 27.4 % fall in the high, medium, and low hazard levels, respectively. In terms of total land area, Lunana suffers the maximum (11.3 km<sup>2</sup>) inundation, followed by Toedwang (5.1 km<sup>2</sup>). Subdistrict blocks were classified into different GLOF hazard units based on which hazard level map intersected with them. Lunana, Kazhi, and Toedwang lie in very high hazard level zones, so three of them were classified as very high GLOF hazard level subdistrict blocks. The 13 others overlap with the high-hazard level map, so, these 13 were assigned high-hazard levels. Only one (Nahi) does not intersect with the hazard map, given a zerohazard level (Fig. 9a).
+
+The GLOF exposure assessment showed that the simulated clearwater floods from the glacial lakes in the Lunana glacier complex can cause substantial inundation to two hydropower projects (PHP-I and PHP-I) and numerous structures and agricultural land across different downstream sites. The floods expose up to 2,721 buildings which convert to 16,062 people and 6,157 livestock. Besides 11.8 km² of agricultural land, 101.4 km of roads (including highways, farms, and feeder roads) and 71 bridges are also exposed to the different hazard levels of the simulated floods. Of the total, only 4 buildings (all in
+
+Journal of Hydrology 619 (2023) 129311
+
+Image /page/9/Figure/2 description: A figure, labeled Fig. 6, displays a grid of flow hydrographs for three different Glacial Lake Outburst Flood (GLOF) scenarios. The grid has four columns, representing four proglacial lakes (Thorthormi Tsho, Lugge Tsho, Raphstreng Tsho, and Bechung Tsho), and five rows, representing downstream sites (Lunana, Punakha dzong, Wangdue hydro. sta, PHP-I, and PHP-II). Each graph plots Flow (m³/s) on the y-axis against Time (min) on the x-axis, from 0 to 1000 minutes. Three scenarios are shown in each plot: Large (red line), Moderate (black line), and Small (blue line). Generally, as the flood wave moves downstream (from top to bottom row), the peak flow decreases and the time to peak increases. For the Thorthormi Tsho lake, the large scenario peak flow decreases from about 55,000 m³/s at Lunana to 18,000 m³/s at PHP-II. For Lugge Tsho, it decreases from 20,000 m³/s to 5,000 m³/s. For Raphstreng Tsho, it decreases from 18,000 m³/s to about 3,500 m³/s. For the Bechung Tsho lake, a flood wave is only shown at the Lunana site, with a peak of about 850 m³/s for the large scenario; for all downstream sites, a note indicates 'Flood wave terminates before reaching here'.
+
+Fig. 6. Flow hydrograph of three different scenarios of the GLOF magnitudes: large, moderate and small at each site downstream of the four typical proglacial lakes.
+
+Lunana), 0.01 km<sup>2</sup> of agricultural land, and 4 bridges (all in Lunana) were located within the very high-hazard flood zone; none of the roads fall within the very high-hazard zone (detail in Table 4). These exposed structures and agricultural land were primarily located in the subdistrict blocks of Wangdue Phodrang and Punakha districts, about 80–100 km
+
+downstream of the Lunana glacier complex. For example, the highest number of buildings are exposed in Guma (753), followed by Thedtsho (639). Guma also suffers from the longest stretch (15.5 km) of road inundation. Likewise, Dzomi is modelled to likely suffer maximum (2.1 km²) damage to agricultural land, while the highest number of bridges
+
+Image /page/10/Figure/2 description: A figure with eight graphs, labeled (a) through (h), arranged in two columns and four rows. The figure illustrates flood reach time and velocity for three different Glacial Lake Outburst Flood (GLOF) scenarios (Large, Moderate, and Small) from four different lakes: Bechung tsho, Raphstreng tsho, Lugge tsho, and Thorthormi tsho. A legend indicates that the Large scenario is represented by a red line, Moderate by a blue line, and Small by a green line.
+
+The left column of graphs (a, b, c, d) shows Flood Reach Time. The x-axis represents 'Station: distance from the lakes (m)' from 0 to 150,000. The y-axis represents 'Time (hr)' from 0 to 20, with the values increasing downwards. 
+- Graph (a) for Bechung tsho shows the flood traveling less than 100,000 m in over 20 hours.
+- Graphs (b) for Raphstreng tsho, (c) for Lugge tsho, and (d) for Thorthormi tsho show the flood traveling the full 150,000 m. In these cases, the 'Large' scenario is the fastest, taking approximately 8-11 hours, while the 'Small' scenario is the slowest, taking 12-18 hours.
+
+The right column of graphs (e, f, g, h) shows Velocity. The x-axis is also 'Station: distance from the lakes (m)' from 0 to 150,000. The right y-axis is 'Velocity (m s⁻¹)' from 0 to 25. The left y-axis, shown in graph (e), is 'Elevation m a.s.l.' from 0 to 4000. Graph (e) also includes a light green line showing the elevation profile, which decreases from over 4000 m to near 0 m over the 150,000 m distance. These graphs have two vertical grey shaded bands.
+- Graph (e) for Bechung tsho shows velocities generally below 15 m/s.
+- Graphs (f) for Raphstreng tsho, (g) for Lugge tsho, and (h) for Thorthormi tsho show fluctuating velocities along the river path, with peaks exceeding 25 m/s, particularly within the first 25,000 m. Generally, the 'Large' scenario has the highest velocity and the 'Small' scenario has the lowest.
+
+Fig. 7. Flood reach time (a–d) and velocity (e–h) for every 300 m along the river centreline of three different GLOF scenarios (large, moderate, and small) from the lakes in the Lunana glacier complex. Plotting the elevation profile (see e) along the river centreline shows how flow velocity changes depending on the elevation gradient at different places (shaded with grey color in e–h).
+
+(11) exposure would occur in Rubisa (Table 4). Although no road is exposed in Lunana since it is only connected by non-vehicle access, 11.8 km of the footpath is being inundated. There are, however, current plans to extend the road network towards Lunana which will change exposure and risk. For display purposes, we classified exposure into high, medium, and low levels for each subdistrict block. Based on our classification approach (see method), 10 subdistrict blocks were classified as highly exposed, while 3 (Athang, Rubisa, and Patshaling) and 1 (Gase Tshowongm) were labeled as medium and low, respectively. The exposure level was zero for the other three, including Kazhi, Largyab,
+
+and Nahi, since no structure or agricultural land intersects with the flood hazard map (Fig. 9b).
+
+The calculated normalized score for the social vulnerability index for the 17 subdistrict blocks ranges from 0.3 to 0.7. A total of six, including Lunana, Largyab, Dzomi, Kabisa, Nahi, and Kazhi, had high vulnerability levels. Only Gase Tshowongm had a low vulnerability level, while the other 10 exhibited medium-level vulnerability (Fig. 9c). The risk assessment produced by combining hazard, exposure, and vulnerability revealed that only Lunana gewog is located in a very high-risk zone. Nine of the other subdistrict blocks lie in high-risk categories, and four
+
+Table 3
+Flood arrival time, flow velocity, and depth at five downstream sites under three scenarios: large (L), moderate (M), and small (S) GLOF magnitudes.
+
+| Flood hydraulics             | Lake       | Magnitude | Downstream | site    |                     |       |        |
+|------------------------------|------------|-----------|------------|---------|---------------------|-------|--------|
+|                              |            |           | Lunana     | Punakha | Wangdue<br>Phodrang | PHP-I | PHP-II |
+| Flood wave arrival time (hr) | Thorthormi | L         | 0.5        | 2.5     | 3.0                 | 4.0   | 4.5    |
+|                              |            | M         | 0.75       | 3.0     | 4.0                 | 5.0   | 6.0    |
+|                              |            | S         | 1.0        | 4.25    | 5.25                | 7.0   | 8.0    |
+|                              | Lugge      | L         | 1.0        | 3.75    | 4.5                 | 5.75  | 7.0    |
+|                              |            | M         | 1.25       | 4.75    | 5.5                 | 7.25  | 8.75   |
+|                              |            | S         | 1.25       | 5.5     | 6.5                 | 8.75  | 10.75  |
+|                              | Raphstreng | L         | 0.75       | 3.5     | 4.25                | 5.75  | 7.0    |
+|                              |            | M         | 0.75       | 4.25    | 5.25                | 7.0   | 8.75   |
+|                              |            | S         | 1.0        | 4.75    | 6.0                 | 8.25  | 10.5   |
+|                              | Bechung    | L         | 1.4        |         |                     |       |        |
+|                              | -          | M         | 2.1        |         |                     |       |        |
+|                              |            | S         | 3.25       |         |                     |       |        |
+| Flood velocity (m/s)         | Thorthormi | L         | 14.2       | 8.0     | 3.2                 | 7.0   | 8.9    |
+| •                            |            | M         | 10.2       | 6.1     | 2.2                 | 5.3   | 7.4    |
+|                              |            | S         | 7.4        | 5.0     | 1.7                 | 3.5   | 6.3    |
+|                              | Lugge      | L         | 12.3       | 7.1     | 3.2                 | 3.2   | 6.0    |
+|                              | 00         | M         | 9.0        | 6.0     | 2.4                 | 2.5   | 5.2    |
+|                              |            | S         | 7.4        | 5.0     | 2.0                 | 2.0   | 5.0    |
+|                              | Raphstreng | L         | 12.3       | 6.7     | 3.0                 | 3.1   | 5.2    |
+|                              | r 0        | M         | 9.7        | 5.6     | 2.2                 | 2.2   | 4.3    |
+|                              |            | S         | 8.0        | 4.7     | 1.7                 | 1.7   | 3.7    |
+|                              | Bechung    | L         | 6.0        |         |                     |       |        |
+|                              |            | M         | 4.0        |         |                     |       |        |
+|                              |            | S         | 3.0        |         |                     |       |        |
+| Flood depth (m)              | Thorthormi | L         | 25.0       | 28.0    | 29.2                | 26.2  | 29.1   |
+| 1                            |            | M         | 14.7       | 18.0    | 20.4                | 18.7  | 20.0   |
+|                              |            | S         | 10.0       | 11.4    | 14.0                | 13.3  | 12.8   |
+|                              | Lugge      | L         | 13.0       | 13.4    | 15.3                | 16.6  | 15.0   |
+|                              | - 00 -     | M         | 9.2        | 10.0    | 11.0                | 12.6  | 10.0   |
+|                              |            | S         | 7.2        | 7.3     | 8.6                 | 10.1  | 7.0    |
+|                              | Raphstreng | L         | 12.4       | 12.1    | 13.6                | 14.0  | 13.4   |
+|                              | 1 0        | M         | 9.2        | 9.0     | 1.0                 | 10.6  | 9.3    |
+|                              |            | S         | 6.0        | 7.0     | 8.0                 | 9.0   | 7.0    |
+|                              | Bechung    | L         | 2.5        | ,       |                     |       |        |
+|                              |            | M         | 1.7        |         |                     |       |        |
+|                              |            | S         | 1.0        |         |                     |       |        |
+
+are in the medium-risk zone. The remaining 3 (Kazhi, Largyab, and Nahi) were classified as zero-risk areas since they have zero exposure and hazard (Fig. 9d).
+
+## 5. Discussion
+
+### 5.1. Glacial lakes outburst susceptibility
+
+The comprehensive analysis of wide-ranging cryospheric, geotechnical, and geomorphological factors showed that Thorthormi Tsho and Lugge Tsho are relatively more susceptible to GLOF than Raphstreng Tsho and Bechung Tsho. This finding resonates well with the earlier GLOF susceptibility assessments undertaken in the Bhutan Himalaya. The first-order GIS-based susceptibility assessment by Rinzin et al. (2021) ranked Thorthormi Tsho as the most dangerous glacial lake in the Bhutan Himalaya. Also, the detailed field investigation by NCHM (2019a) has reported active sliding of a thin moraine between Thorthormi Tsho and Raphstreng Tsho, thus, increasing the susceptibility to GLOF events in the future. The comparatively higher velocity of the Thorthormi glacier, especially in the frontal area, indicates active sliding of ice, the same being identified as the likely cause of Thorthormi subsidiary lake-II draining in 2019 according to NCHM (2019a). Calving of ice into the lake would further increase its hydrostatic gradient and GLOF volume (GAPHAZ, 2017; Rounce et al., 2016) and calving activities can create displacement waves as the calved ice impacts the lake. The high GLOF susceptibility of Lugge Tsho also relates well with the
+
+findings from the earlier studies (NCHM, 2019a; Rinzin et al., 2021; Wangchuk et al., 2022). This is evident as Lugge Tsho has produced the most catastrophic past GLOF events in the Bhutan Himalaya (Komori et al., 2012; Maurer et al., 2020). There is evidence of multiple cases of GLOF events from the same lake in the Himalaya, e.g., Cirenma Co in the Central Himalaya produced three GLOFs in 1964, 1981, and 1983, respectively (Wang et al., 2015). Such experience from the Himalaya combined with the high GLOF susceptibility of Lugge Tsho suggests that GLOF occurrence in the future cannot be neglected and is a subject of further investigation. It is also an actively growing glacial lake in the region, with the potential to grow in the future (Linsbauer et al., 2016), implying that it is likely to become more exposed with time to surrounding slope failure (Rounce et al., 2017; Rounce et al., 2016). Also, Thorthormi Tsho and Lugge Tsho both being in contact with their parents' glaciers means they would exacerbate the melting of the corresponding glacier through increasing mechanical calving activities and subaqueous melting, which in turn would cause enhanced expansion of the lake through glacier-lake interaction. This interaction would make Thorthormi Tsho and Lugge Tsho more susceptible to GLOF while also increasing potential flood volume (King et al., 2018, 2019). Thus, we highly recommend continuous monitoring of Thorthormi Tsho and Lugge Tsho. Nevertheless, we cannot discount the GLOF susceptibility of Raphstreng Tsho and Bechung Tsho entirely since they also have one or more factors indicating high outburst susceptibility (e.g., high potential flood volume of Raphstreng Tsho). Bechung Tsho is also expected to grow in the future by expanding towards its upstream direction and can
+
+Image /page/12/Figure/2 description: The image displays two satellite maps, labeled (a) and (b) in the caption, showing hazard levels in different river valleys. The top map, (a), shows a barren, mountainous area around Thanza village. It is marked with longitude from 90°11'30"E to 90°13'0"E and latitude from 28°4'30"N to 28°5'0"N. A scale bar indicates 0.5 km. A color-coded overlay shows hazard levels, with a legend defining green as Low, yellow as Medium, orange as High, and dark red as Very high. The 'Very high' and 'High' zones are concentrated in the main river channel, with 'Medium' and 'Low' zones extending to the banks. The bottom map, (b), shows the area around Wangdue Phodrang, which has more vegetation and a visible settlement. It is marked with longitude from 89°53'0"E to 89°54'0"E and latitude from 27°29'30"N to 27°31'30"N, with a similar 0.5 km scale bar. The hazard overlay here shows 'High' (orange) in the river channel, with 'Medium' (yellow) and 'Low' (green) zones on the banks, partially covering the settlement. The 'Very high' level is not present in this map. The caption identifies map (a) as Thanza village (7 km from the lakes) and map (b) as Wangdue Phodrang (100 km from the lakes).
+
+Fig. 8. The excerpt of the hazard map from Thanza village of Lunana (7 km from the lakes) (a) and Wangdue Phodrang (100 km from the lakes) (b) is derived from multiple GLOF scenarios originating from the Lunana glacier complex. It should be noted that the GLOF from the Bechung Tsho ultimately terminates before arriving Wangude Phodrang area, and so no very high hazard level is identified in this area based on our classification approach. The background image: ESRI Basemap Imagery.
+
+Journal of Hydrology 619 (2023) 129311
+
+Image /page/13/Figure/2 description: A figure from a scientific paper by S. Rinzin et al., published in the Journal of Hydrology, showing four maps labeled (a), (b), (c), and (d). These maps illustrate the hazard, exposure, vulnerability, and risk levels in 2020 for 17 subdistrict blocks in a mountainous region. The maps are gridded with latitude lines from 27°0'N to 28°0'N and longitude lines from 90°0'E to 90°30'E, and include a 50 km scale bar.
+
+Map (a) shows the Hazard level. The legend indicates 'High' (orange) and 'Very high' (dark red). The subdistricts LN, TW, and KZ are colored dark red for 'Very high' hazard, while the rest are orange for 'High' hazard.
+
+Map (b) shows the Exposure level. The legend has four categories: 'No exposure' (white), 'Low' (green), 'Medium' (yellow), and 'High' (orange). Subdistricts KZ and LM have 'No exposure'. GW has 'Low' exposure. AT and PT have 'Medium' exposure. The remaining subdistricts have 'High' exposure.
+
+Map (c) shows the Vulnerability level. The legend has three categories: 'Low' (green), 'Medium' (yellow), and 'High' (orange). Subdistrict GW has 'Low' vulnerability. TW, CH, DZ, AT, and PT have 'Medium' vulnerability. The rest of the subdistricts have 'High' vulnerability.
+
+Map (d) shows the Risk level. The legend has four categories: 'No risk' (white), 'Medium' (yellow), 'High' (orange), and 'Very high' (dark red). Subdistricts KZ and LM have 'No risk'. GW, AT, and PT have 'Medium' risk. Subdistrict LN has 'Very high' risk. The remaining subdistricts have 'High' risk.
+
+Fig. 9. The hazard (a), exposure (b), vulnerability (c), and risk (d) levels in 2020 of 17 subdistrict blocks downstream of the Lunana glacier complex under three different scenarios of GLOF magnitude (large, medium, and small) from each lake. The risk is calculated as a function of hazard, exposure, and vulnerability. The 17 subdistrict blocks are: Lunana (LN), Kazhi (KZ), Toedwang (TD), Chhubu (CH), Guma (GM), Dzomi (DZ), Lingmukha (LM), Kabisa (KB), Thedtsho (TD), Nahi (NH), Gase Tshogongm (GG), Gase Tshowongm (GW), Rubisa (RB), Darkar (DK), Athang (AT), Patshaling (PT), Largyab (LG). Background image: ALOS-PALSAR DEM.
+
+become more exposed to GLOF triggers like an avalanche and landslide/rockfall in the future (Richardson and Reynolds, 2000). Therefore, attention might be needed to monitor Raphstreng Tsho and Bechung Tsho. The NCHM has reported an increasing temperature trend in Bhutan since the 1990s. There has been a decreasing trend in precipitation, but Bhutan experienced high precipitation variability over the
+
+same period (Dorji and Tamang, 2019), leading to numerous extreme weather events in Bhutan (NCHM, 2022). This increasing frequency of extreme precipitation (Dorji and Tamang, 2019) added to increasing temperature would further amplify the GLOF susceptibility level of glacial lakes in the Lunana complex (GAPHAZ, 2017). However, since all lakes are located in the same basin, if the most dangerous glacial lakes
+
+Table 4
+Buildings, people, livestock, road, bridges, and agricultural land within flood hazard level: very high (VH), high (H), medium (M) and low (L).
+
+| No. | Name        | Building (count) |     |     |      | People (count) |      |      |      | Agricultural land (km2) |      |      |      | Livestock (count) |       |       |       | Bridge (count) |    |    |    | Road (km) |     |     |     |
+|-----|-------------|------------------|-----|-----|------|----------------|------|------|------|-------------------------|------|------|------|-------------------|-------|-------|-------|----------------|----|----|----|-----------|-----|-----|-----|
+|     |             | VH               | H   | M   | L    | VH             | H    | M    | L    | VH                      | H    | M    | L    | VH                | H     | M     | L     | VH             | H  | M  | L  | VH        | H   | M   | L   |
+| 1   | Athang      | 0                | 8   | 7   | 11   | 0              | 48   | 42   | 66   | 0                       | 0.4  | 0.1  | 0.2  | 0                 | 18.4  | 16.1  | 25.3  | 0              | 4  | 0  | 1  | 0         | 1.3 | 0.8 | 2.0 |
+| 2   | Chhubu      | 0                | 5   | 6   | 97   | 0              | 30   | 36   | 582  | 0                       | 0.4  | 0.2  | 0.5  | 0                 | 11.5  | 13.8  | 223.1 | 0              | 3  | 1  | 2  | 0         | 1.3 | 1.4 | 4.2 |
+| 3   | Darkar      | 0                | 49  | 42  | 50   | 0              | 294  | 252  | 300  | 0                       | 0.5  | 0.2  | 0.3  | 0                 | 112.7 | 96.6  | 115   | 0              | 7  | 1  | 2  | 0         | 5.3 | 3.2 | 5.0 |
+| 4   | Dzomi       | 0                | 152 | 56  | 75   | 0              | 912  | 336  | 450  | 0                       | 0.8  | 0.3  | 1.0  | 0                 | 349.6 | 128.8 | 172.5 | 0              | 3  | 2  | 1  | 0         | 7.9 | 1.9 | 1.0 |
+| 5   | Gase Tgongm | 0                | 36  | 22  | 30   | 0              | 216  | 132  | 180  | 0                       | 0.3  | 0.1  | 0.1  | 0                 | 82.8  | 50.6  | 69    | 0              | 3  | 2  | 1  | 0         | 3.9 | 1.8 | 2.5 |
+| 6   | Gase Twongm | 0                | 0   | 2   | 3    | 0              | 0    | 12   | 18   | 0                       | 0.05 | 0.03 | 0.04 | 0                 | 0     | 4.6   | 6.9   | 0              | 1  | 0  | 1  | 0         | 0.3 | 0.4 | 1.5 |
+| 7   | Guma        | 0                | 172 | 127 | 454  | 0              | 1032 | 762  | 2724 | 0                       | 0.4  | 0.2  | 1.1  | 0                 | 395.6 | 292.1 | 1044  | 0              | 2  | 0  | 0  | 0         | 3.4 | 3.1 | 9.0 |
+| 8   | Kabisa      | 0                | 22  | 11  | 37   | 0              | 132  | 66   | 222  | 0                       | 0.3  | 0.1  | 0.3  | 0                 | 50.6  | 25.3  | 85.1  | 0              | 0  | 0  | 0  | 0         | 2.1 | 0.7 | 1.2 |
+| 9   | Kazhi       | 0                | 0   | 0   | 0    | 0              | 0    | 0    | 0    | 0                       | 0    | 0    | 0    | 0                 | 0     | 0     | 0     | 0              | 0  | 0  | 0  | 0         | 0   | 0   | 0   |
+| 10  | Largyab     | 0                | 0   | 0   | 0    | 0              | 0    | 0    | 0    | 0                       | 0    | 0    | 0    | 0                 | 0     | 0     | 0     | 0              | 0  | 0  | 0  | 0         | 0   | 0   | 0   |
+| 11  | Lingmukha   | 0                | 73  | 39  | 101  | 0              | 438  | 234  | 606  | 0                       | 0    | 0    | 0    | 0                 | 167.9 | 89.7  | 232.3 | 0              | 1  | 0  | 0  | 0         | 0   | 0   | 0   |
+| 12  | Lunana      | 1                | 17  | 20  | 21   | 6              | 102  | 120  | 126  | 0                       | 0.2  | 0.05 | 0.1  | 2.3               | 39.1  | 46    | 48.3  | 3              | 0  | 1  | 0  | 0         | 4.5 | 1.0 | 1.2 |
+| 13  | Nahi        | 0                | 0   | 0   | 0    | 0              | 0    | 0    | 0    | 0                       | 0    | 0    | 0    | 0                 | 0     | 0     | 0     | 0              | 0  | 0  | 0  | 0         | 0   | 0   | 0   |
+| 14  | Patshaling  | 0                | 8   | 3   | 23   | 0              | 48   | 18   | 138  | 0                       | 0.1  | 0.01 | 0.01 | 0                 | 18.4  | 6.9   | 52.9  | 0              | 2  | 1  | 0  | 0         | 0.5 | 0.1 | 0.5 |
+| 15  | Rubisa      | 0                | 5   | 9   | 18   | 0              | 30   | 54   | 108  | 0                       | 0.8  | 0.2  | 0.2  | 0                 | 11.5  | 20.7  | 41.4  | 0              | 10 | 3  | 2  | 0         | 4.0 | 1.2 | 2.0 |
+| 16  | Thedtsho    | 0                | 199 | 116 | 324  | 0              | 1194 | 696  | 1944 | 0                       | 0.7  | 0.2  | 0.4  | 0                 | 457.7 | 266.8 | 745.2 | 0              | 2  | 1  | 0  | 0         | 5.0 | 2.8 | 5.9 |
+| 17  | Toedwang    | 3                | 109 | 59  | 55   | 18             | 654  | 354  | 330  | 0.01                    | 0.4  | 0.1  | 0.3  | 6.9               | 250.7 | 135.7 | 126.5 | 1              | 4  | 2  | 1  | 0.03      | 3.6 | 1.1 | 3   |
+|     | Aggregate   | 4                | 855 | 519 | 1299 | 24             | 5130 | 3114 | 7794 | 0.01                    | 5.35 | 1.78 | 4.64 | 9.2               | 1967  | 1194  | 2988  | 4              | 42 | 14 | 11 | 0         | 43  | 19  | 39  |
+|     | Total       | 2677             |     |     |      | 16,062         |      |      |      | 11.8                    |      |      |      | 6157              |       |       |       | 71             |    |    |    | 101.4     |     |     |     |
+
+are monitored (e.g., Thorthormi Tsho), and proper disaster management plans are implemented in the downstream settlement, possible downstream damage from the floods from any lakes in the basin could be highly reduced.
+
+### 5.2. GLOF risk on downstream settlement
+
+Our simulation of the 1994 GLOF by setting breach volume at  $20.0 \times 10^6$  yielded a peak flow of 2,129 m³/s at Wangdue Phodrang, within the range of estimated peak flow (1800 to 2500 m³/s) of the 1994 GLOF at the same station (Osti et al., 2013). Likewise, our simulated GLOF takes 4.4 hr to reach Punakha, which also relates well with the arrival time noted by eyewitnesses and the reconstructed arrival time (4.75 hr) using a seismometer (Maurer et al., 2020). Simulations with varying breach times between 0.25 and 1 hr result in a significant difference in peak discharge at the dam, but both arrival time and peak flow remain fairly consistent by the time the flood wave reaches Punakha and Wangdue Phodrang (Supplementary Fig. S2). This model calibration with the GLOF events of 1994 in the same basin reasonably indicates that our model setup is suitable for modelling different scenarios of future GLOF for risk assessment purposes.
+
+The study combines hazard, exposure, and vulnerability to determine the risk posed by the potentially dangerous lakes in the Lunana glacier complex to the downstream region along the Phochu and Punatsangchu banks. The study reaffirms the earlier studies such as Osti et al. (2013) and Koike and Takenaka (2012) that Bhutanese people are inevitably highly exposed to the GLOF. Most of the modelled GLOFs (8/12) from the lakes produce a flood of magnitude higher than the 1994 GLOF of Lugge Tsho (Fujita et al., 2008; Maurer et al., 2020; Osti et al., 2013; Watanbe and Daniel, 1996). Although there is no history of floods of such high magnitudes in Bhutan, with climate change-induced deglaciation and cryospheric system change in the high mountains across the world in recent decades (GAPHAZ, 2017), these potential events in the future cannot be disregarded. Our results offer new insights into the impacts of large-magnitude GLOF events in a typical basin in the Himalaya.
+
+Our study showed that one among 17 subdistrict blocks (Lunana) lies in the very high GLOF risk zone and 9 others in a high-risk zone. Some of the villages in Lunana, like Thanza village of Lunana, are located very close to the lakes in the Lunana glacier complex (just about 2-7 km). The large and medium scenario GLOFs from Thorthormi Tsho take only 15-30 min to reach some part of Lunana. This also means that the Lunana region is likely impact by all GLOF events with little warning, suggesting high-magnitude damage to the exposed structures (most of which are not built with the concrete wall (NSB, 2018). The capacity to prepare for, respond to, and recover from such a disaster is challenging for Lunana because of its remote location and high social vulnerability level. The nine high-risk subdistrict block clusters (Fig. 9) are located in the Punakha and Wangdue Phodrang districts, about 80-100 km from the Lunana glacier complex. The long travel time for GLOF allows these subdistrict blocks to potentially have more warning time in a GLOF event. Even the large scenario GLOF from the Thorthormi Tsho takes about 2.5 and 3 hr to arrive in Punakha and Wangdue Phodrang, respectively (Table 2). But this is only possible if there is a flood early warning system installed along the riverbank and if authorities and communities are well trained in response procedures, or it becomes entirely reliant on ad-hoc community-to-community communication. Flood early warning instruments have been installed along the banks of Phochu and Punatsangchu (NCHM, 2021), and continuous monitoring and maintenance of these instruments are highly recommended (Huggel et al., 2020). However, due to the high density of settlement, the inundation of infrastructures including buildings and roads is very high in Punakha and Wangdue Phodrang areas; 85 % of exposed buildings and 59 % of the bridge are located within these nine subdistrict blocks (Table 4). Our results suggest that the Lunana subdistrict block needs the highest level of attention for GLOF disaster risk reduction. Also, the
+
+farther downstream settlements in the Punakha and Wangdue Phodrang districts, such as Guma and Thedtsho need to be prioritized (Fig. 9).
+
+Investigations elsewhere in the world have revealed that eventual damage from disasters not only depend on the magnitude of the hazard (GLOF intensity in this case) but also on the social vulnerability level (crucial for resilience, preparedness, and response capacity of the people) (Cutter and Finch, 2008; Flanagan et al., 2018). Similarly, the importance of data from multiple sources (GLOF hazard data, subdistrict level socio-economic condition, and settlement density) for the downstream GLOF risk assessment is clearly portrayed in this study. Gase Tshogongm falls in a high-hazard zone, but with low vulnerability and exposure level, so it has been classified as a medium GLOF-risk subdistrict block. Likewise, Kazhi, Nahi, and Largyab have high social vulnerability and fall in a high GLOF hazard zone; however, all of them have zero exposure, which allowed us to rule out the GLOF risk in these subdistrict blocks under current conditions. These findings help us improve the GLOF risk knowledge in the Punatsangchu and Phochu basins, which is an essential preliminary assessment for the implementation of an early warning system (Huggel et al., 2020) and risk reduction prioritization. The GLOF early warning system has been put in place by National Centre for Hydrology and Meteorology and is mainly based on experience from the 1994 GLOF event of Lugge Tsho (NCHM, 2021). By contrast, our downstream risk assessment is based on the multiple GLOF scenario modelling results, most of which have a higher magnitude than the 1994 GLOF event. Thus, our result provides new insights for consolidating and improving the existing early warning system. To this end, this study is valuable for the comprehensive and careful design of GLOF risk management strategy, both hard engineering schemes such as improving structural measures (e.g., construction of flood retention walls, identification of safe/evacuation zone, lowering lake levels) and enhancing the response and recovery capacity (e.g., GLOF risk advocacy programs, training populations for GLOF evacuation) in the affected downstream community.
+
+### 5.3. GLOF impact on the hydroelectric project
+
+Punatsangchu basin hosts the two largest hydroelectric power projects, PHP-I (1200 MW) and PHP-II (1020 MW), both of which are at the advanced stage of construction (MOEA, 2021). With a cost estimate of 1.74 (Nu. 133.20) billion USD for PHP-I and 0.95 (Nu. 72.9) billion USD for PHP-II; these two projects are record-high investment projects in Bhutan. In the latest inventory, Punatsangchu headwaters have a total of 180 glacial lakes of larger than 0.05 km<sup>2</sup> (Rinzin et al., 2021). The glacial lakes present an imminent threat of outburst flood to these two hydroelectric projects. Potential GLOFs can impact these hydroelectric power projects and lead to significant economic loss and potentially endanger human lives. While the potential impact of GLOFs on the hydroelectric project has been undertaken at a large scale (Schwanghart et al., 2016), this study presents a pilot-based detailed assessment of the dangerous glacial lakes and their impacts on hydropower projects in Bhutan. Except for GLOFs from the Bechung Tsho, all potential GLOFs intercept the hydroelectric projects and may present a significant risk in the future, e.g., at PHP-II, the peak ranges from 870 m<sup>3</sup>/s (Raphstreng small magnitude) to 19000 m<sup>3</sup>/s (Thorthormi large magnitude). These hydroelectric power projects are inevitably exposed to GLOFs from the lakes in the Lunana glacier complex. However, their risk will depend on the design standard of the dams. The feasibility assessment done before the commencement of these projects by the Japan International Cooperation Agency (2001) has considered peak flow up to 2176 m<sup>3</sup>/s. So, a high flood magnitude (>20,000 m<sup>3</sup>/s) from Thorthormi Tsho is an order of magnitude higher than the design flood of the hydropower projects. A flood of such magnitude may therefore cause severe damage to the reservoir and dam, causing overtopping, excessive sedimentation, outages, equipment damage, and cascading impacts on the low-lying settlements. The high concentration of sediment in the Punatsangchu river from the 1994 GLOF event of Lugge Tsho is said to have persisted up to a
+
+month after the event (Watanbe and Daniel, 1996). The massive landslide-induced flood cascade in 2003 from the Tsatichhu river, a north-easterly flowing tributary of the Kurichu river in eastern Bhutan, resulted in an estimated loss of 1.1 million USD due to the disruption of the 60 MW Kurichu hydroelectric power plant (Dunning et al., 2006). The disruption of PHP-I and PHP-II due to GLOFs from the Lunana glacier complex could result in greater financial implications since their installed capacity is about 2-order magnitudes greater than the Kurichu hydroelectric power plant. Several past GLOF events in the Himalaya have caused implacable damage to the hydroelectric power downstream, including GLOFs that occurred in the neighbouring regions. For example, the 1985 GLOF of Dig Tsho in eastern Nepal (Vuichard and Zimmermann, 1987) and the 1981 GLOF of Cirenma Co (Wang et al., 2015), and the 2016 Gongbatongsha GLOF (Sattar et al., 2022) in the Poiqu River basin of the central Himalaya. All these experiences from the neighbouring Himalaya and Bhutan suggest that GLOFs from the lakes in the Lunana glacier complex may also cause potential damage to PHP-I and PHP-II.
+
+### 5.4. GLOF modelling approach and uncertainty
+
+The input dam breach parameters are the essential initial condition determining the magnitude of flood hydraulics and the translated hazard and risk maps (Sattar et al., 2021a; Wahl, 2004). Previous studies have found outflow hydraulics like peak discharge are highly sensitive to breach formation time ( $T_f$ ) (Basheer et al., 2017; Sattar et al., 2021a). While there are no models to predict the uncertainty of GLOF breach parameters, previous studies (Sattar et al., 2021a; Somos-Valenzuela et al., 2015) have used embankment dam-based equations (Wahl, 2004) to predict uncertainty ranges of moraine dam breach parameters. When Wahl (2004) is used here, the value of upper bound uncertainty dam breach width for the Thorthormi Tsho is more than double (704.5 m) the actual calculated value (293.7 m), while the lower bound decreases to less than half (116.7 m). There is an extensive range of uncertainty in other breach parameters like  $T_f$  (see supplementary Table 2 for detail). Also, due to the lack of a clear historical record, scenario definition for GLOFs cannot typically depend on the site-specific frequency-magnitude relationship, commonly used in other natural hazards scenario definitions (Frey et al., 2018; GAPHAZ, 2017). As a result, the scenario definition is another primary source of uncertainty in GLOF modelling. Here we accounted for these uncertainties by considering multiple potential GLOF scenarios with varying magnitudes (3 from each lake) following the established trend in the existing literature (Frey et al., 2018; Sattar et al., 2021a; Sattar et al., 2021b). This allowed us to define 12 unique scenarios (3 for each lake), all flowing along the same trajectory. However, for disaster risk reduction worst-case GLOFs scenarios can be considered, although it could be argued that an inability to mitigate against the highest magnitude events should not remove the focus from smaller events that could, and should have mitigation schemes.
+
+The overtopping breach failure mode in the HEC-RAS dam breach model is a gradually progressive process of dam erosion due to an overflow of water, which contrasts with overtopping resulting from a large displacement wave induced by a rapid mass movement into the lake. These differences could have underestimated the peak discharge and flood arrival time in the current study (Allen et al. 2022). Here we do not model the process chain or the trigger processes, however, we assume breach scenarios of varied dimensions to assess the downstream impact. Moreover, GLOF is a highly complex process, involving several cascading events, usually initiated by mass movement and wave overtopping, followed by debris flow, hyper-concentrated flow, and secondary debris flow (GAPHAZ, 2017; Schneider et al., 2014; Worni et al., 2014). However, these complex modelling approaches are usually applied in the context where run-out distance is relatively short (Frey et al., 2018), which is not the case in Bhutan, especially in the Punatsangchu basin. Here, most simulated floods (8/12) can reach up to 150
+
+km downstream. The complex process chain modelling covering such a long flood reach distance is theoretically possible, it is highly challenging to model multiple scenarios using existing codes like r.avaflow (Mergili et al., 2017) due to high computation costs and numerical instabilities (Frey et al., 2018; Schneider et al., 2014). Therefore, since we modelled multiple scenarios over a very long run-out distance coupled with robust downstream risk assessment, the HEC-RAS 2D model is a reasonable alternative, especially for the first-order evaluation in this study. Nonetheless, through Google Earth high-resolution imagery, we observed that the Phochu river channel first starts with a relatively broad plain, especially in the Thanza area of Lunana, followed by a long stretch of deep valleys and gorges (20-30 km), and flood plains become wide as the river reaches the downstream site where settlements are most concentrated. This channel characteristic relates well to a typical feature that will allow erosion and entrainment of the enormous amount of debris from the upstream area that would multiply flood magnitude and thus damage caused to the lives and properties, as occurred during the 2016 GLOF event from Gongbatongsha Co in the central Himalaya (Sattar et al., 2022). Future studies addressing the critical challenges of GLOF modelling like long-runout distance, complex process chain, and high computational cost issues should be the priority.
+
+Because of their high discharge and entrainment capacity, GLOFs are typically known for carrying large boulders and mobilizing large amounts of sediment which can greatly modify channel morphology (Cook et al., 2018; Nativ et al., 2022)\*\*. For example, the 2016 GLOF from Gongbatongsha Co in central Himalaya, with a discharge volume about of  $1.1 \times 10^5$  m<sup>3</sup>, changed channel morphology up to 40 km downstream and widened channel width up to 12 m at some cross sections (Cook et al., 2018, Chen et al. 2023). Likewise, the 1994 Lugge Tsho GLOF also changed channel morphology for up to 25 km downstream (Maurer et al., 2020), while debris deposits were evident as far as 200 km downstream (Watanbe and Daniel, 1996). Huber et al. (2020) suggest that numerous present-day boulders along the river channels in the Himalaya have likely been transported by historical GLOFs and estimate that GLOFs of discharge  $10^3$ – $10^5$  m<sup>3</sup> are capable of transporting boulders of ca. 10 m diameter or more. Here, the majority of the modelled GLOF scenarios (8/12) exceed peak discharges of 10<sup>3</sup> m<sup>3</sup>/s indicating their potential to transport sediments of different grades including boulders. This implies that if future GLOF are of the same magnitude as the modelled scenarios, the impact might be far more than what we projected with our clear water flood rheology. Therefore, future studies should focus on assessing the impact of GLOF through a complex debris flow model or those allowing phase changes.
+
+Population and infrastructures are very dynamic in Bhutan as it is a rapidly developing country. According to ICIMOD's land cover and land use data, the built-up area within our 2D model domain has increased by ~194 % over the period from 2000 to 2020 with an annual rate of  $\sim$ 0.36 km<sup>2</sup>/yr. The agricultural land has increased from  $\sim$ 16.0 km<sup>2</sup> in 2000 to  $\sim$ 19.5 km<sup>2</sup> in 2015, but has dropped to  $\sim$ 17.5 km<sup>2</sup> in 2020, although the latter might be due to the conversion of farmland into builtup areas (Supplementary Fig. S4). Similarly, according to the Bhutan population and housing census 2017, the population in the Wangdue Phodrang and Punakha, the two districts most exposed to GLOF, have increased from 48,850 to 70,926 (+45 %) between 2005 and 2017 (NSB, 2018). This increasing downstream population and built-up area coupled with an increase in glacial lake area since the 1990s indicates that downstream GLOF risk would have also increased over the period. Most importantly, increasing GLOF risk in the downstream communities in the future means that continuous monitoring of GLOF in Punatsangchu basin is highly recommended.
+
+## 6. Conclusion
+
+This study draws upon multiple data sources, including traditional declassified Corona KH-4 and Hexagon KH-9, Landsat (TM, ETM, and OLI), DEM, LULC, OpenStreetMap, population and housing census data.
+
+Using these data and hydrodynamic modelling, we comprehensively assessed glacial lake outburst susceptibility, downstream exposure, hazard, vulnerability, and risk in the Bhutan Himalaya. The results showed that, although all the lakes in the Lunana glacier complex have one or more unfavourable GLOF susceptibility factors, Thorthormi Tsho and Lugge Tsho are relatively more susceptible to GLOF. A detailed hazard assessment for the potential GLOFs indicated that two hydropower projects and over 16,000 people dwelling along the Phochu and Punatsangchu banks in addition to numerous structures and agricultural land are exposed to these events. Our comprehensive GLOF risk assessment revealed 1 (Lunana) subdistrict block in the very high GLOF risk zone, 9 in the high-risk zone (mainly in Punakha and Wangdue Phodrang areas), 3 in medium, and 1 in a low-risk zone. The other 3 had zero GLOF risk under current conditions.
+
+We recommend continuous monitoring of the glacial lakes in the Lunana glacier complex (through field investigation and high-resolution satellite imagery) with a particular focus on Thorthormi Tsho and Lugge Tsho. Second, as for the downstream GLOF disaster management, we recommend that Lunana should be given the highest priority, followed by subdistrict blocks concentrated in the Wangdue Phodrang and Punakha districts like Guma and Thedtsho. The study highlights the importance of multi-source data in improving the knowledge of downstream GLOF risk and serves as a base for improving the GLOF risk reduction strategies in the region.

chatbot/rag/data/('1-s2.0-S2212420922001339-main', '.pdf')_extraction.md

@@ -0,0 +1,193 @@
+
+
+# ABSTRACT
+
+The Himalayas are home to many high-risk glacial lakes. Effective prevention of glacial lake outburst floods (GLOFs) in this region has recently become an urgent priority. As a major element of an integrated risk management strategy, the GLOF early warning system is a viable and promising tool for mitigating climate change-related risks. It prevents loss of life and reduces the economic and societal impacts of disasters. Within the framework of the Second Tibetan Plateau Scientific Expedition and Research program, we developed and implemented a monitoring and early warning system (EWS) for Cirenmaco, a transboundary high-risk glacial lake located in the central Himalayas. The EWS consists of monitoring lake-level change, end-moraine displacement, ice collapse, and downstream runoff. The monitoring data can be transmitted via the Beidou and Inmarsat satellites and a mobile network to the data center. The in-situ and real-time monitoring guarantee captures the precursors of ice collapse and glacial lake outburst and raises alarms in advance for downstream communities. In terms of data transmission, monitoring elements, and warning thresholds, the Cirenmaco EWS scheme is one of the most advanced among all similar GLOF early warning systems. Implementing an early warning system is the most practical strategy for mitigating potential threats to Himalayan high-risk glacial lakes. The EWS system is less expensive than lake dam immobilization and artificial drainage projects, and it provides more valuable environmental monitoring data in high mountain areas. The Cirenmaco EWS serves an effective demonstration for the construction of similar projects to prevent and mitigate GLOFs in the Himalayas.
+
+# 1. Introduction
+
+Due to climate warming and ongoing glacier recession, glacial lakes in the Tibetan Plateau and its surroundings (TPS) are in a state of rapid expansion [1–3], and numerous potential glacial lakes buried within the glacier beds are projected to be exposed in the future [4-7]. Over the past few decades, the number and area of glacial lakes on the plateau have increased by  $\sim$ 10.7% and  $\sim$ 15.2%, respectively [2]. The Himalayas have the highest concentration of contemporary glacial lakes among all subregions of the TPS [8]. Because of its longitudinal length, over 2000 km from east to west, and an altitude difference of up to 5000 m, great spatial heterogeneity is manifested across many aspects, such as climate composition, geographic resources, and ecological environment [9,10], prompting it to become one of the main study areas in cryospheric science, geology, and social humanities. To date, research on glacial
+
+https://doi.org/10.1016/j.ijdrr.2022.102914
+
+<sup>&</sup>lt;sup>b</sup> College of Earth and Environmental Sciences, Lanzhou University, Lanzhou, 730000, China
+
+<sup>\*</sup> Corresponding author.
+
+Corresponding author. Institute of Tibetan Plateau Research, Chinese Academy of Sciences, Beijing, 100101, China. E-mail addresses: weicaiwang@itpcas.ac.cn (W. Wang), zhangtg16@lzu.edu.cn (T. Zhang).
+
+lakes has focused on the distribution characteristics of contemporary or potential glacial lakes, and their accompanying hazards and/or risks [4,11–15], inventory of regional glacial lake outburst floods (GLOFs) [16–20], reconstruction of previous GLOFs and projection of GLOFs from high outburst potential lakes [21–26]. However, research on prevention and mitigation measures for GLOFs from high-hazard/risk glacial lakes remains lacking [27–29].
+
+Previous studies on identifying potentially dangerous glacial lakes selected the central Himalayas as subjects, including a large number of high-hazard/risk glacial lakes, over the other subregions of the TPS [12,15,30,31]. Zheng et al. [7] conducted a risk assessment for glacial lakes in the TPS, and another study focused on future changes in glacial lakes, as well as the accompanying enhanced hazards and risks. The results revealed that approximately 3200 potential glacial lakes are buried in the Himalayan glacier beds, accounting for 23.4% of the total number across the TPS. Approximately 32.1% of the Himalayan potential glacial lakes will be exposed by 2050 and 54.4% by 2100, at RCP8.5 scenario. The results of the risk assessment indicated that future GLOF susceptibility in this region will be at least 10 times higher than the current level. There have been several studies on glacial lake hazard/risk assessment for the entire Himalayan area, and some typical regional areas [32–35]. It is noteworthy that, although the evaluation schemes are varied among different studies, assessments of high-hazard/risk glacial lakes have comparatively high consistency and consensus [15]. Therefore, an in-depth understanding of glacial lakes with high outburst potential in this region provides a reliable basis for further studies.
+
+TPS regions have seen more than 100 GLOFs originating from glacial lakes (Fig. 1a) [18–20,36–38]. The Himalayas are a global hotspot of GLOFs in the TPS. Since the 1980s, the GLOF frequency has been  $\sim$ 1.3 times per year [39], and there remains a distinctively steady trend of outbursts compared to other glaciered areas of the world where the GLOF frequency is decreasing or remain constant [40]. Recent studies have suggested that the frequency of future GLOFs in some areas will not increase. One such area is the Peruvian Andes. These tropical glaciers have already retreated quite close to steep headwalls [41,42], and numerous posterior bedrock-dammed lakes have become exposed. A decrease in GLOF susceptibility has been observed, the potential for future glacial lake development is not high, as lakes will become progressively distant from their mother glacier, which is the main source of triggers. Additionally, several remedial studies have been conducted to prevent GLOFs.
+
+Unexpected Himalayan glacial lake outburst floods have caused thousands of deaths and enormous socio-economic losses [16]. For
+
+Image /page/1/Figure/6 description: A figure composed of four panels, labeled a, b, c, and d, illustrating the location and characteristics of the Cirenmaco glacial lake.
+
+Panel a is a map of a large mountainous region, showing the distribution of glacial lakes (blue), moraine-dammed GLOFs (yellow dots), and ice-dammed GLOFs (orange dots). Red stars indicate early warning systems. The location of Cirenmaco is marked. The map covers latitudes from 25°N to 45°N and longitudes from 70°E to 100°E.
+
+Panel b is a detailed topographic map of a specific area, showing the Poiqu River and its tributary, Zhangzangbo. Glaciers are highlighted in light blue with dashed outlines. The map's coordinates range from 28°2'N to 28°6'N and 86°0'E to 86°5'E.
+
+Panel c is a satellite image showing the Cirenmaco glacial lake, which is a distinct turquoise color, set within a rocky, mountainous landscape. The coordinates for this view are approximately 28°3'30"N to 28°4'30"N and 86°3'E to 86°5'E.
+
+Panel d consists of two bar charts showing changes in area over time from 1990 to 2020. The top chart, labeled 'Glacier', shows a general decrease in area from approximately 1.8 km² in 1990 to about 1.5 km² in 2020. The bottom chart, labeled 'Glacial lake', shows a general increase in area from about 0.12 km² in 1990 to a peak of over 0.3 km² in 2010, settling at around 0.28 km² in 2020.
+
+Fig. 1. Location of Cirenmaco in the TPS. (a) Since the 1980s, there have been many GLOFs originating from moraine and ice-dammed lakes in the TPS. Meanwhile, several EWS have been established in this region. (b, c) Cirenmaco is located in the Zhangzangbo Valley, central Himalaya. (d) The area changes of the lake and its parent glacier from 1988 to 2020.
+
+instance, a catastrophic GLOF originating from Luggye Tsho, Eastern Himalaya, on October 7, 1994, was triggered by the melting of buried ice beneath the moraine dam, and the resulting flood killed 21 people and destroyed local infrastructure [43,44]. The largest GLOF event recorded in the region occurred on June 16, 2013. Chorabari, a moraine-dammed lake in the central Himalayas, suddenly drained, and the outburst flood, accompanied by rainfall floods, washed away settlements downstream, causing more than 6000 deaths [45,46]. Therefore, determining how to prevent and mitigate these outburst flood disasters is urgent. It is necessary to assess potential GLOF hazards and risks and simulate possible outburst floods. Moreover, it is crucial to meaningfully implement a variety of remedial works, such as lake dam immobilization, artificial drainage, and construction of monitoring and early warning systems (EWS).
+
+Such engineering measures have been applied sporadically worldwide, even though numerous high-risk glacial lakes have been identified. The danger, high-altitude, and inaccessibility of high mountain areas make engineering measures several times more expensive than in densely populated plains. In the Himalayas, there are currently fewer than 10 engineering measures in place at present. For instance, Jialongco (Table 1), a moraine-dammed lake in the Himalayas, drained in 2002, resulting in severe downstream damage. Considering the serious threat posed, the local government artificially strengthened the moraine dam in 2020. In contrast, in the Peruvian Andes, at least 38 glacial lakes have been used for various remedial works, including 38 lakes that were artificially drained by open-cut excavation, tunnels, or siphon measures, 15 lake dams that were strengthened; and one lake that established an EWS [27,47]. These measures were taken mainly because many villages, towns, and cities are situated downstream of the high-risk glacial lakes in the surrounding region. An EWS generally consists of monitoring the glacial lake, its parent glaciers, and/or the downstream river, formulating a disaster mitigation plan, and training for evacuation. The latter two are non-engineering measures. However, they should be conducted based on an in-depth understanding of the original GLOF hazards. The difficulty lies in fostering strong support from local governments and residents [21,28].
+
+It is vital for the implementation of remedial work that high-risk glacial lakes, adequate economic and engineering reserves, and long-term development plans are prepared in advance. Cirenmaco is located in the Poiqu River Basin, central Himalaya, close to the China–Nepal border (Fig. 1a and b). Three GLOFs from this lake have occurred in 1964, 1981, and 1983, respectively. In the past few decades, the high outburst susceptibility of Cirenmaco and the accompanied transboundary threat of its GLOF have plagued residents in both regions [7,12,48,49]. An EWS was installed near Zhangmu Port in 2001 unilaterally by the Nepalese government, to protect downstream hydropower projects [50]. However, the absence of a glacial lake monitoring system caused the EWS to lose a considerable part of its effective functions, as the upper Poiqu River Basin is home to many high-risk glacial lakes located in China. Therefore, supported by the Second Tibetan Plateau Scientific Expedition and Research (STEP) Program, we have establish an EWS for Cirenmaco to monitor dynamic changes in the lake and its parent glacier, lake and downstream river water-level, meteorology, etc., to provide a timely warning signal to the downstream area.
+
+Table 1
+The categories of remedial work for glacial lakes with high risk, and corresponding measures. A typical case is given for the specific measure.
+
+| Remedial works                         | Measures                             | Typical cases                  |                                             |                                                                                                                                                                                                                                 | Reference     |
+|----------------------------------------|--------------------------------------|--------------------------------|---------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------|
+|                                        |                                      | Location                       | Name                                        | Explanation                                                                                                                                                                                                                     |               |
+| Artificial drainage                    | Open cut                             | Cordillera<br>Blanca, Peru     | Palcacocha                                  | This moraine-dammed lake failed in 1941, and then an<br>open cut was excavated in 1974, lowering the lake to a<br>level of 3 m.                                                                                                 | [27]          |
+|                                        | Tunnel                               | Cordillera<br>Blanca, Peru     | Parón                                       | The tunnel became functional in 1984 and the lake level<br>could be lowered by 52 m.                                                                                                                                            | [27]          |
+|                                        | Siphon                               | Cordillera<br>Blanca, Peru     | Lake 513                                    | This bedrock-dammed lake failed in 1991 and 2010, but<br>a siphon was already installed in 1989, lowering the lake<br>level of 8 m. Subsequently, a combination of tunnels was<br>used to lower the lake level by 20 m in 1993. | [51]          |
+| Lake dam<br>immobilization             | Earth dam                            | Central<br>Himalaya            | Jialongco                                   | This moraine-dammed lake had been drained in 2002<br>causing severe damage downstream, the lake is still a<br>high risk. Thus, in 2020 the local government<br>immobilized its dam.                                             | This<br>study |
+|                                        | Arch dam                             | Alps, Swiss                    | Trift glacier lake                          | The dam, built primarily to generate electricity, was<br>completed by 2007, but undoubtedly reduced the<br>incidence of glacier-related hazards.                                                                                | [52]          |
+| Monitoring and Early<br>warning system | Monitoring and/<br>or warning system | Central<br>Caucasus,<br>Russia | Bashkara Glacier<br>lakes                   | The two sister lakes failed in 1958 and 1959. To protect<br>the infrastructure and settlements downstream, an early<br>warning and monitoring system was installed in 2008.                                                     | [53]          |
+|                                        | Disaster<br>mitigation plan          | Cordillera<br>Blanca, Peru     | Lake 513                                    | When information of the GLOF arrived, alarm will be<br>issued to evacuate. At the same time, sirens were<br>installed in all important communities and major<br>transportation routes.                                          | [28]          |
+|                                        | Evacuation drills                    | Cordillera<br>Blanca, Peru     | Palcacocha,<br>Tullparaju,<br>Cuchillacocha | Detailed evacuation plans and evacuation briefs were<br>developed by relevant staff to protect the downstream<br>city of Huaraz, and were given extensive publicity.                                                            | [21]          |
+
+# 2. Previous studies on Cirenmaco
+
+Cirenmaco ( $86^{\circ}03'56''E$ ,  $28^{\circ}04'02''N$ ) glacial lake is located at 4612 m above sea level. Since the 1990s, the region in which it is situated (south slope of the central Himalayas) has been designated as a region of intensively developed of glacial lakes [2,8]. Cirenmaco's perennial mean annual precipitation can be estimated at ~1000 mm and the relevant mean annual temperature at approximately 0 °C by interpolating the nearest meteorological stations [24]. The lake, which was drained in 1981, had reached a disastrous and alarming level. Based on a post-disaster investigation conducted by Xu [54], approximately  $19 \times 10^6$  m³ of water was released, with a peak discharge of 15,920 m³/s at the breach and a total flood duration of 1 h. Because of the GLOF, substantial erosion was observed throughout the downstream valley, transporting approximately  $4 \times 10^6$  m³ of solid material. There was a possibility of debris flow within the flood spreading in the Zhangzangbo Valley. The GLOF, which moved over 50 km along the downstream Poi-qu/Sun Koshi River, destroyed two bridges (at the mouth of Zhangzangbo Valley and Zhangmu Port), 6 km of China-Nepal roads and buildings on both sides, and one hydropower facility, resulting in 200 fatalities [48,54].
+
+Cirenmaco has been identified as a high-hazard/risk glacial lake in recent studies [12,15,55] due to its steep surroundings, buried ice in the moraine dam, and the high exposure of downstream communities and infrastructure. Although the area of Cirenmaco underwent a rapid increase from 0.11 to 0.33 km<sup>2</sup> beginning in the 1980s, a stable trend in lake area has been maintained since 2010 (Fig. 1c and d). On the surface, the development potential of this lake has reached its limits. However, the dead ice in the lake bottom may continuously melt to increase the lake volume, and the buried ice melting in the moraine dam may continuously affect dam stability and trigger outburst floods [24]. A bathymetric survey of Cirenmaco was conducted in 2012. The results showed that approximately  $18 \times 10^6$  m<sup>3</sup> of water was held by this lake, with a maximum depth of 115 m [24]. Subsequently, Zhang et al. [25] simulated outburst floods at different drainage percentages and assessed the hazards and potential impacts of Cirenmaco GLOFs (Fig. 2a-c). Although their scheme for the outburst flood hydrograph as an input for the HEC-RAS model was proposed using an ideal drainage curve process, it has been demonstrated to be feasible in many other studies also [56,57]. In contrast, predicted GLOF propagation is transported entirely within a narrow valley. Because of DEM accuracy and simulation uncertainties, precisely identifying potentially damaged buildings, roads, etc., in such gullies is difficult. Therefore, prior simulation results for the Cirenmaco GLOF depth, velocity, and arrival time provide suitable reference points. The distance between Cirenmaco and the Zhangmu ports were found to be 23 km. The estimated travel time of the flood indicated that, if Cirenmaco released 100% of its volume, the GLOF would have at least 50 min to reach Zhangmu Port [25]. Moreover, studies of the 1981 GLOF reconstruction revealed that the flood time was more than an hour when it crossed this route [24]. However, we deduced that the actual flood migration time for this period may be
+
+Image /page/3/Figure/5 description: A multi-panel figure illustrating the simulation of a Glacial Lake Outburst Flood (GLOF). The figure is divided into four panels labeled a, b, c, and d.
+
+Panel (a) is a large topographic map showing a river valley in a mountainous region, spanning the border between China and Nepal. The Poiqu River flows from north to south through the valley. Several locations are marked with red dots, including Jialongco, Nyalam, Zhangzangbo Bridge, and Zhangmu Port. The map includes latitude and longitude grids, a scale bar from 0 to 4 km, and a north arrow. There are two circular insets with photographs. The top-left inset, pointing from Jialongco, shows a turquoise glacial lake surrounded by a rocky moraine. The bottom-right inset, pointing from an area near Zhangzangbo Bridge, shows a dry, rocky landscape.
+
+Panels (b), (c), and (d) are smaller maps showing the results of a flood simulation in a specific area downstream. 
+- Panel (b) displays the 'Flood depth (m)', with a color scale ranging from 0.001 (blue) to 56.7 (magenta).
+- Panel (c) shows the 'Flood velocity (m/s)', with a color scale from 0.1 (light blue) to 22.3 (red).
+- Panel (d) presents a 'Hazard' map with a legend indicating four levels: Low (dark green), Medium (light green), High (yellow), and Very high (red).
+
+Fig. 2. Simulation of the Circhmaco GLOF. (a) The outburst floods will cross the Zhangzangbo Bridge to Zhangmu Port and then enter Nepal at the largest drainage scenario (100%). (b) The Circhmaco GLOF depth and (c) velocity are used to composite the (d) hazard map. All data in this figure were provided by Zhang et al. [25]. The picture in the circle on the left shows the strengthened moraine dam of Jialongcuo, and on the right shows the melting of buried ice in the moraine dam of Circhmaco.
+
+Image /page/4/Figure/2 description: A composite image of five photographs, labeled a, b, c, d, and e, each showing a different scientific monitoring station in a remote, natural environment. Image 'a' displays a complex weather station on a rocky mountainside with a person in a red jacket nearby and solar panels on the ground. Image 'b' shows a station with a large solar panel and colorful prayer flags, set against a backdrop of green hills. Image 'c' features a station on the shore of a turquoise lake with snow-capped mountains in the background and a person standing next to it. Image 'd' depicts a station with a solar panel next to a rushing river. Image 'e' shows a simpler station on a steep, rocky slope.
+
+Fig. 3. Early warning system installed at Cirenmaco. A total of nine different services are contained in this project. These include (a) monitoring of the lake and its parent glacier, (b) monitoring of the environment of the downstream river, (c) water level monitoring of the glacial lake, (d) water level monitoring of the downstream river and (e) moraine dam dynamic monitoring.
+
+Image /page/4/Figure/4 description: A multi-panel figure displaying real-time monitoring data from a glacial valley. Panel (a) is a map of the area, showing the locations of various sensors including GNSS, cameras, meteorological stations, and water level sensors along a river and near a lake. The map includes latitude and longitude coordinates and a scale bar. Panels (b) through (j) are a series of photographs taken over time. Panels (b-d) show a glacier terminating in a turquoise lake. Panels (e-g) show the surface of the turquoise lake. Panels (h-j) show a river flowing through a rocky channel. The bottom section contains four graphs labeled (k), (l), (m), and (n), all plotted against dates from 2021/10/3 to 2021/10/19. Graph (k) shows a water level decreasing from approximately 1.8 m to 0.5 m. Graph (l) shows a water level that is stable around 0.5 m before spiking to a peak of about 0.85 m around 2021/10/19. Graph (m) shows temperature fluctuating daily between approximately -1°C and 5°C, with precipitation events shown as blue bars, including a significant event around 2021/10/18. Graph (n) shows temperature fluctuating between approximately 4°C and 12°C, also with a significant precipitation event around 2021/10/18.
+
+Fig. 4. Real-time monitoring data transmission via satellite and mobile networks. (a) Distribution of the main monitoring points in the Zhangzangbo Valley, and the monitoring data currently available, such as photographs taken by three visible light cameras, which are aligned (b–d) with the Cirenmaco inlet area and the snout of its parent glacier and (e–g) its outlet area, and (h–j) the terminus of the Zhangzangbo Valley. (k, l) Water level data and (m, n) meteorological data from the glacial lake and the mouth of the valley.
+
+~40 min, according to the observed average flow velocity (9.8 m/s) of the 1981 GLOF in Nepal [54]. These data suggest that approximately 40–60 min may be required to send a warning signal.
+
+In addition, the hazard map was composited by flood depth and velocity and translated into four intensity levels: low, medium, high, and very high (Fig. 2d). In the future, more refined and accurate simulation results can be relied upon to establish this hazard map, which has implications for determining evacuation locations in downstream areas, constructing GLOF prevention and mitigation facilities, and implementation of economic development measures. These studies provide an understanding of the possible flood damage from Cirenmaco and are the basis of this EWS construction [21,28,58].
+
+Image /page/5/Figure/2 description: A topographical map displaying water depth measurements of a lake situated in a mountainous region. The map is oriented with North at the top, as indicated by a north arrow. The y-axis shows latitude from 28°3'45"N to 28°4'15"N, and the x-axis shows longitude from 86°3'45"E to 86°4'15"E. A winding path of colored dots across the dark green lake surface represents the water depth. A legend in the bottom left corner, titled "Water depth (m)", explains the color scale. The depth ranges from 6.2 meters, represented by blue, to 115.3 meters, represented by red, with green, yellow, and orange indicating intermediate depths. The deepest area, shown in red, is located in the central part of the survey path. A scale bar indicates a length of 0.2 kilometers.
+
+Fig. 5. In 2012. Wang et al. [24] measured the water depth and estimated the volume of Circumaco using an unmanned boat.
+
+Image /page/5/Figure/4 description: A schematic diagram illustrates the major components of an early warning system (EWS) for a glacial environment. The diagram shows a mountainous landscape with a glacier labeled "Glacier ice" on a slope of bedrock. The glacier feeds into a large glacial lake, which is held in place by a moraine dam. Below the dam, there is buried ice and a downstream river system. Various monitoring systems are in place, indicated by labels and icons. On the glacier, there is "Glacier dynamic monitoring." Near the lake, there is "Meteorological monitoring" (with precipitation shown as snowflakes falling from clouds) and "Topographic mapping" (indicated by a drone). Within the lake, there is "GLacial lake dynamic monitoring" and a "Bathymetric survey." At the lake's edge, "Water level monitoring" is shown. On the moraine dam and downstream area, there is "GNSS displacement monitoring" and "Downstream river dynamic monitoring." Data from all these monitoring points is sent via blue lines to a central "Data transmission" tower. This tower transmits the data to a satellite, which then relays it to a "Data center." The data center sends information to "Observers," who can then issue an "Early warning," represented by a red icon with sound waves.
+
+Fig. 6. Schematic diagram illustrating major components of the Cirenmaco EWS, including the lake and its parent glacier monitoring, downstream river monitoring, water-level monitoring, moraine dam dynamic monitoring, meteorological monitoring, bathymetric survey, topographic mapping, data transmission, and transferring the early warning signal.
+
+# 3. Characteristics of EWS
+
+## 3.1. A typical case: an EWS established for 513 Lake, Cordillera Blanca, Peru, in 2012
+
+Two drainage events occurred in the Peruvian glacial Lake 513, in 1994 and 2010. An ice avalanche triggered the 1994 GLOF. However, because of the prior artificial drainage of the lake, the GLOF magnitude was comparatively low, so no damage was caused to the downstream area [51,59]. The 2010 GLOF was a catastrophic event, also triggered by an ice avalanche, that severely damaged the downstream city of Carhuaz, destroying 22 buildings and 100 km of roads, and causing 100 casualties [59]. In response to the serious threat posed by Lake 513, the local government implemented an EWS in 2011–2012 to prevent potential disaster [28]. Notably, the
+
+EWS of Lake 513 is the first glacial lake early warning system worldwide with a fully functional service system. It includes risk knowledge (hazard and risk assessment), monitoring and warning services (monitoring and warning of hazard occurrence), dissemination and communication (dissemination of warning signals), and response capability (disaster mitigation plans and evacuation drills) [60].
+
+Risk knowledge requires field investigation and analytical work to make a hazard and risk assessment of glacial lakes to clarify lake outburst probability and downstream exposure within the GLOF path. Monitoring and warning services consist of five components [28, 61]. The first is a glacier and lake monitoring station at the lake shore, which consists of four geophones and two cameras that record ground shaking during ice avalanches and flooding, as well as glacier–glacial lake dynamics. The second component is a hydrometeorological monitoring station, which is installed downstream of the lake outlet. The third component is the repeater station, whose main function is to transmit the monitored data to the data center. The fourth component is the data center in the city of Carhuaz, which is responsible for data evaluation and storage. The fifth component is the warning station, which is primarily responsible for the warning signal. The functionality of the dissemination and communication components, as well as the response capability, largely depends on active coordination with local authorities and residents to maximize the full functionality of the EWS and reduce potential losses. The EWS of Lake 513 provides a good reference for the construction of an EWS for Cirenmaco.
+
+## 3.2. EWSs on the TPS
+
+The early warning system essentially differs in predicting GLOFs versus rainfall-type floods. The EWS for GLOFs focuses on monitoring various elements and issuing timely warnings rather than providing specific flood features. This is due to GLOFs' high magnitude, short duration, and destructive properties. Forecasting rainfall-type floods, on the other hand, concentrates primarily on real-time hydrological processes based on dynamic courses of flow regime, streambed roughness, fluid density, etc., to establish effective standard curves to predict and warn of floods downstream with metrics generally on peak discharge, maximum water level, and arrival time [62].
+
+Five EWSs have been established over the past few decades in the TPS region, to prevent GLOFs (Fig. 1a, Table 2). EWSs can be divided into two categories, based on their technical structures. The first is a semi-automatic system that relies on staff assistance. For example, a manual EWS in Bhutan was installed, with six water-level sensors and two automatic weather stations, and included a control room with two workers in a village downstream of the glacial lakes. The workers are equipped with satellite phones for information transmission [63,64]. The second was a fully automatic EWS. These services generally consist of real-time monitoring of the glacial lake and its parent glacier, hydrometeorological monitoring, data transmission and storage, and early warning [28,61]. For example, in 2014, an EWS was installed in Kyagar, an ice-dammed lake in Karakoram, China. It consists of three observation stations situated the lakeshore, estuary, and far downstream. The parameters, including the water-level of the lake and river, monitored by automatic sensors and photographs of the lake and riverbed taken by automatic time-lapse cameras, are all transferred through a satellite link to a data center [29]. If a GLOF was identified, warning signals were sent to authorities. In addition, fully automatic EWSs require an indispensable threshold to determine GLOFs, which may be real-time dynamic warning indicators such as lake and river water levels, deformation of the lake and glacier, or meteorological data [65]. The water-level threshold is frequently used in EWSs [28]. The main difference between a semi-automatic EWS and a fully automatic EWS is data transmission. Due to the remote and inaccessible environment in high-altitude mountainous regions, data transmission is very difficult and expensive. However, long-term observations using satellite for data transmission incur high costs. The frequency and cost of equipment maintenance are also higher than those of semi-automatic systems.
+
+Overall, the quality of the current EWS applied to high-risk glacial lakes varies greatly around the world. In theory, a simple prediction scheme would require only the installation of water-level sensors for real-time analysis. Some activities in the Hindu Kush-Karakorum range rely on primitive beaconing methods to provide early warning [66]. If the purpose is not to investigate the mechanism of GLOF, but rather to provide early warning, it is a success as long as the effective alarm signal can reduce downstream exposure damage. Due to the limited data transmission scheme, fully automated and multidimensional monitoring of EWS is rare. Under the
+
+ Table 2
+
+ The GLOF early warning systems widely established in the TPS region, including the Tian Shan Mountains, Karakoram, and Himalayas.
+
+| ID | Lake name                                                          | Lake type          | Location                       | Longitude° | Latitude° | Explanation                                                                                                                         | References |
+|----|--------------------------------------------------------------------|--------------------|--------------------------------|------------|-----------|-------------------------------------------------------------------------------------------------------------------------------------|------------|
+| 1  | Teztor lake                                                        | Moraine-<br>dammed | Tien Shan,<br>Kyrgyzstan       | 42.54      | 74.43     | The EWS was installed in 2011, and successfully predicted the Teztor lake GLOF in 2012.                                             | [67]       |
+| 2  | Kyagar lake                                                        | Ice-<br>dammed     | Karakoram,<br>China            | 35.71      | 77.01     | The EWS combined three automatic terrestrial observation stations, and successfully predicted Kyagar GLOFs from 2015 to 2018.       | [29,38]    |
+| 3  | Ropal Tsho                                                         | Moraine-<br>dammed | Central<br>Himalaya,<br>Nepal  | 27.86      | 86.47     | A manual EWS was set up in 1997, and evolved into a fully automatic system in 1998.                                                 | [50]       |
+| 4  | Imja Tsho                                                          | Moraine-<br>dammed | Central<br>Himalaya,<br>Nepal  | 27.9       | 86.93     | The EWS combined two monitoring stations and two relay stations.                                                                    | [63]       |
+| 5  | Raphstreng Tsho;<br>Luggye Tsho;<br>Thorthormi Tsho;<br>Besta lake | Moraine-<br>dammed | Eastern<br>Himalaya,<br>Bhutan | 27.88      | 89.74     | The EWS was installed with six water level sensors and two weather stations, and successfully predicted Lemthang Tsho GLOF in 2015. | [63,64]    |
+
+premise of ensuring effective early warning efforts, a comprehensive monitoring system is crucial and valuable for understanding the GLOF generation mechanism, propagation process, and flow regime transformation. This is particularly true regarding the action of photo or video monitoring of glaciers and glacial lakes, which not only eliminates many false alarm signals but also motivates us to understand the potential glacier–glacial lake–moraine dam linkage process when GLOF occurs.
+
+# 4. Construction of the Cirenmaco EWS
+
+The Cirenmaco EWS project, which was established in 2020 and validated in 2021, is the first use of a comprehensive EWS application in a moraine-dammed lake in China. It is a fully automatic system that applies several new technologies and methods. The entire system complies with the current two basic mechanisms of moraine-dammed lake outbursts: overtopping and piping. Both mechanisms are likely to emerge in Cirenmaco drainage. Overtopping triggers mainly consist of the ice avalanche from its mother glacier, which has been detecting by photo and video monitoring and satellite remote sensing image analysis. A water-level sensor located at Cirenmaco Lake will record the potential wave displacement. This is the main surface dynamic process that we considered, because the Cirenmaco GLOF generated in 1981 was determined to have been caused by an ice avalanche based on field investigations. Piping is generally induced by the melting of buried ice; thus, surface displacement monitoring was employed to record the activity status of the Cirenmaco moraine dam. We also deployed three water-level sensors and two meteorological stations along the flood route in Zhangzangbo Valley to observe the response of the lake and river after the GLOF occurred, as well as the meteorological context of the disaster (see Table 3). We elaborates on this advanced EWS in terms of four major categories and 10 specific components.
+
+## 4.1. Monitoring system
+
+- (i). Glacial lake, related parent glacier, and downstream river monitoring. A 6 m observation tower was built on the right bank of the lake to equip various apparatuses (Fig. 3a). Three 360° rotatable high-definition dome cameras were mounted on the tower, to capture daytime video surveillance and high-frequency timing photographs (Fig. 4a). The first camera was pointed at the lake's inlet area and the snout of its parent glacier (filming every hour from 8 a.m. to 8 p.m.) (Fig. 4b–d). The second camera was pointed at the middle of Cirenmaco Lake (filming every 3 h from 8 a.m. to 8 p.m.), and the third was pointed at the outlet area of Cirenmaco (filming every 3 h from 8 a.m. to 8 p.m.) (Fig. 4e–g) and also monitors the status of the moraine dam. A thermal imaging dome camera (infrared night vision device) was installed at the same angle as the first camera to monitor the lake and glacier (filming every hour from 8 p.m. to 8 a.m.). The thermal camera has low power consumption and a fog-penetration function and can provide video and timed photos under nighttime or foggy conditions. We also installed a group of infrared and visible light cameras at the intersection of Zhangzangbo Valley and the Poiqu River to record the state of the downstream river environment (Figs. 3b and 4h–j). The filming frequency of both cameras was one frame per hour.
+- (ii). Water-level monitoring of glacial lakes and downstream rivers. This service consisted of three automatic water-level sensors that also monitored additional parameters, including water temperature and pressure. One was installed on the lake shore of Cirenmaco (Fig. 3c), recording the possible wave displacement induced by an ice avalanche. One was installed at the lake outlet, and the other was installed at the terminal of the Zhangzangbo Valley (Fig. 3d), near the location of the cameras. The recordings of these sensors were collected every 30 s (Fig. 4k and l).
+- (iii). Dynamic monitoring of moraine dam. This element includes thickness, displacement, and temperature monitoring of the moraine dam. The Cirenmaco moraine dam has accumulated an enormous quantity of rocks, debris, and buried dead ice (Fig. 2a). The stability of the moraine dam impacts the frequency and magnitude of outburst floods. Two global navigation satellite systems (GNSS) were used to detect the elements of the moraine dam (Fig. 3e). This technology was applied here for the first time to observe the moraine dam in a glacial lake. The main potential variations were the settlement of the moraine dam caused by the melting of buried ice, and the movement effect of the moraine dam when the GLOF occurred. Ground penetrating radar was used to detect the moraine dam and analyze its thickness.
+- (iv). Meteorological monitoring. Meteorological stations measured more than a dozen meteorological parameters, including wind speed, wind direction, precipitation, air temperature, air humidity, radiation, soil temperature, soil moisture, and atmospheric pressure (Figs. 3a and 4m–n). We selected three automatic meteorological stations at different altitudes in Zhangzangbo Valley. These were located at the lake shore of Cirenmaco, the middle altitude of the valley, and the intersection of the valley and the Poiqu River to monitor the three-dimensional and horizontal meteorological background of disasters within the study area. The total elevation difference between the three automatic meteorological stations was nearly 2500 m.
+- (v). Satellite Remote Sensing Image Monitoring. To trace the historical state of the entire sequence of Cirenmaco disasters, a series of satellite images, such as Landsat, Sentinel, and SPOT, were selected to illustrate the lake–glacier dynamics. Useful information such as glacier boundary, glacier velocity, and ice crevasse development should be extracted as much as possible. Regular updates and database maintenance are required to provide a broad contextual awareness for high-frequency water-level, meteorology, and displacement monitoring.
+- (vi). Bathymetric survey for the lake and topographic mapping for the lake and its surroundings. Lake volume is a critical metric in identifying potentially dangerous glacial lakes. Following the Cirenmaco bathymetric survey in 2012 (Fig. 5) [24], we conducted a more detailed bathymetric survey in 2020 to understand the lake's evolution. We also utilized drone images to create high-resolution topographic mapping around the lake.
+
+**Table 3**Primary equipment for the early warning system of Cirenmaco, as well as the involved monitoring elements, data acquisition intervals and accuracy.
+
+| Equipment                              | Number | Locations                                    | Primarily monitored elements                                                                                                                      | Accuracy                                                                 | Acquisition interval |
+|----------------------------------------|--------|----------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------|----------------------|
+| Pressure type water<br>level sensor    | 3      | lake, downstream<br>river                    | Water-level                                                                                                                                       | 0.1%FS                                                                   | 30 s                 |
+| Automatic<br>meteorological<br>station | 3      | lake, Zhangzangbo<br>Valley                  | air temperature; precipitation; wind speed and<br>direction; air humidity; radiation; soil<br>temperature; soil moisture; atmospheric<br>pressure | ±0.2 °C; ±0.4 mm; ±0.3 m/s; ±5°; ±3%; ±5 W/m2;<br>±0.5 °C; ±3%; ±0.3 hPa | 30 min               |
+| Camera                                 | 6      | glacier, glacial lake,<br>moraine dam, river | dynamic processes                                                                                                                                 | /                                                                        | 1–3 h                |
+| GNSS                                   | 2      | moraine dam                                  | surface displacement                                                                                                                              | <1 m                                                                     | 0.25 s               |
+
+## 4.2. Data-processing system
+
+- (vii). Data transmission. The data measured in this ESW were transmitted via the Beidou and Inmarsat satellites and a mobile network to a data center (Fig. 6). The devices situated around Cirenmaco need to employ satellites to transmit data due to the lack of mobile signal coverage. The water-level and meteorological and imaging data at the mouth (middle) of Zhangzangbo Valley can be transmitted through the mobile network. Furthermore, all these devices are powered by solar panels.
+- (viii). Data Center. It is located at the Institute of Tibetan Plateau Research Chinese Academy of Sciences, China. The center is responsible mainly for data storage, real-time surveillance, and transferring early warning signals. All data were moved or directly transferred to the Tibet Autonomous Region Disaster Cloud Platform (http://holo-earth.cn:9022/), established by the Institute of Tibetan Plateau Research, Chinese Academy of Sciences. There is a delay of approximately 2 min when data transmitted via satellite and mobile networks are loaded into the platform. Subsequently, the data can be immediately used for real-time alarm analysis. In addition, all the data measured from this EWS will be shared with Nepalese authorities, which is an important measure to increase cooperation and protect lives, livelihoods, and property on both sides downstream.
+
+## 4.3. Early warning system
+
+(ix). Warning system. In this study, a compound warning threshold was used. Those already successfully operated included the water-level and meteorological thresholds. In subsequent field investigations and indoor technical implementation studies, the deformation of the lake and glacier, moraine dam displacement threshold, and seismic threshold will be incorporated. It should be emphasized that we still lack experience with alarm techniques, and many novel strategies and thresholds are constantly being investigated and developed. The water-level threshold is a critical component of our early warning system, and the three water-level sensors scattered around Zhangzangbo Valley will be integrated. The warning level was separated into three grades, with the first level being the highest and the third being the lowest. The difference between the most recent measurement and the lowest value of the water-level over the past 10 min was used to calculate water-level variation. Eventually, three water levels with thresholds of 30, 25, and 20 cm were selected (Table 4), obtained based on the analysis of our monthly acquired monitoring data. The meteorological thresholds have an auxiliary role in this system and are simply intended to remind the observer to keep a closer eye on the Cirenmaco condition over time. A warning signal was automatically issued to the observer whenever the monitoring element exceeded a specified threshold. Subsequently, the observer checks the real time photos or video of the glacial lake and its parent glacier to determine the lake situation and informs the decision makers and authorities. Since the travel time of a GLOF from Circumaco to the cross-border is ~40 min, it was very important to make the warning process easy and efficient. Despite the EWS with a time delay of ~2 min for monitoring data transmitted to the data center, the data can be used for real-time warning analysis at the platform. This could benefit our valuable basic warning time.
+
+## 4.4. Response capacity development
+
+Disaster mitigation plans and evacuation drills for GLOFs are also critical in ensuring that potentially influenced people can be evacuated to a safe area and that their belongings are secured in a timely manner. This study examines the ability of authorities, who require a considerable degree of general awareness of the present and future challenges facing high mountain regions, specific knowledge of GLOF susceptibility, risks, and warning time, as well as considerable vigor and foresight. Prior to the establishment of the
+
+**Table 4**Constructed and in-process designs of early warning indicators and their threshold settings.
+
+| Equipment                           | Thresholds                         | Status      |
+|-------------------------------------|------------------------------------|-------------|
+| Water level sensor at lake shore    | I:30 cm; II:25 cm; III: 20 cm      | Run         |
+| Water level sensor at lake outlet   | I:30 cm; II:25 cm; III: 20 cm      | Run         |
+| Water level sensor at valley mouth  | I:30 cm; II:25 cm; III: 20 cm      | Run         |
+| Precipitation at lake shore         | I:16 mm/h; II:10 mm/h; III: 8 mm/h | Run         |
+| Deformation of the lake and glacier |                                    | To be built |
+| Moraine dam displacement            |                                    | To be built |
+| Seismic                             |                                    | To be built |
+
+Cirenmaco EWS, the relevant glacial lake research group, made up of researchers, had been working continuously in the area for five years and had played a key role in improving the knowledge of the local population and government regarding the glacial lake and its accompanying hazard/risk. Therefore, when the EWS was ready to be established, we gained strong support from the local government. This also facilitated many of our communication efforts for specific actions of GLOF prevention, and later maintenance of the EWS devices, and avoided some of the human damage caused by differences in perceptions, such as religious beliefs.
+
+Formulating disaster mitigation plans, implementing relief measures, and reserving relief supplies in advance are critical steps for minimizing casualties and losses [59]. Disseminating disaster prevention knowledge to residents, and allowing them to participate fully are also necessary. At present, relevant functionaries have been cooperating with the Disaster Emergency Management Department of the Tibet Autonomous Region and Nepal to fully utilize its function and help it become an effective, timeless, and representative EWS.
+
+# 5. Summary and perspectives
+
+In the Himalayas, where numerous high-risk glacial lakes are distributed, the ability to prevent and mitigate GLOFs remains insufficient, especially for transboundary threats. There are no reports that mention methods to effectively solve this problem. The construction cost of the Cirenmaco EWS was kept under 0.5 million dollars, which is much less than the cost used to strengthen the moraine dam of Jialongco (>\$10 million) in the same basin. Therefore, the EWS is relatively inexpensive to establish compared to other engineering measures on glacial lakes, and the availability of valuable field observations is important for furthering scientific research and development, making it suitable for large-scale replication in the Himalayas. The cost of the satellite transmission is also affordable. The pros and cons of each scenario must be weighed by the staff of each party, and any plan put in place must match the strength of regional social and economic development.
+
+The Cirenmaco EWS is advanced and applies many new technologies such as real-time video transmission. One of the main objectives of applying this feature is to record the entire process of a glacial lake outburst, from the initial triggering event to the flood breaking through the dam and causing damage downstream; as the dynamics of a GLOF, as it occurs, are yet directly observed. We describe in detail the Cirenmaco EWS implemented in our project in four main categories: monitoring systems, data-processing systems, early warning systems, and response capacity development. All monitoring data have been successfully transmitted in real-time, and the EWS has been functional since October 2021. Both China and Nepal could benefit from the EWS project. Our study is of great significance in adapting to the increasing disaster risks in high-altitude mountainous areas and in providing a demonstration for the prevention and mitigation of GLOFs in the Himalayas and TPS.
+
+# **Declaration of competing interest**
+
+The authors declare that they have no known competing financial interests or personal relationships that could have influenced the work reported in this study.
+
+# Acknowledgements
+
+This study was supported by the Second Tibetan Plateau Scientific Expedition and Research Program (STEP) (2019QZKK0208), Strategic Priority Research Program of the Chinese Academy of Sciences (XDA20100300), and National Natural Science Foundation of China (41771088). We are grateful to the two anonymous reviewers for their valuable comments that improved the quality of our manuscript. We would also like to thank Prof. Guoqing Zhang, Senior Engineer Zhongyan Wang, Mr. Pengsi Lei, Mr. Luosang Danba, Mr. Ciren Jiala, and Ms. Menger Peng for their assistance in the fieldwork.
+

chatbot/rag/data/('A glacial lake outburst flood risk assessment for the Phochhu river basin', '.pdf')_extraction.md

@@ -0,0 +1,379 @@
+
+
+Abstract. The melting of glaciers has led to an unprecedented increase in the number and size of glacial lakes, particularly in the Himalayan region. A glacial lake outburst flood (GLOF) is a natural hazard in which water from a glacial or glacier-fed lake is swiftly discharged. GLOFs can significantly harm life, infrastructure, and settlements located downstream and can have considerable ecological, economic, and social impacts. Based on a dam breach model, BREACH, and a hydrodynamic model, HEC-RAS (Hydrologic Engineering Centre's River Analysis System), we examined the potential consequences of a GLOF originating from the Thorthomi glacial lake, located within the Phochhu river basin, one of Bhutan's largest and rapidly expanding glacial lakes. Our analysis revealed that following a breach the Thorthomi glacial lake will likely discharge a peak flow of 16360 m<sup>3</sup> s<sup>-1</sup> within 4h. Such a discharge could potentially cause considerable damage, with an estimated 245 ha of agricultural land and over 1277 buildings at risk of inundation. To mitigate ecological, economic, and social impacts on downstream areas, our results emphasise an urgent need for understanding and preparing for the potential consequences of a GLOF from Thorthomi lake. Our findings provide valuable insights for policymakers and stakeholders involved in disaster management and preparedness.
+
+# 1 Introduction
+
+## 1.1 Glacial lakes and outburst floods
+
+Floods are one of the most common natural disasters worldwide and can cause extensive socio-economic damage. Globally, over the last 2 decades, floods have affected approximately 2.3 billion people and have caused an estimated USD 622 billion in damage (UNISDR, 2015). Glacial lake outburst floods (GLOFs) are floods caused by sudden water release from glacial or glacier-fed lakes and cause a rapid rise in water level over a short time in downstream areas, resulting in devastating consequences (Gurung et al., 2017; Komori et al., 2012; Taylor et al., 2023). GLOFs are infrequent but highly destructive natural disasters that are difficult to predict. Prior to their occurrence, the extent of damage is also difficult to predict. Over the past few decades, the acceleration of glacier melt and recession, primarily driven by climate change, has led to a significant increase in the number of moraine-dammed (natural dams formed by glacial processes) glacial lakes (Sattar et al., 2021; Westoby et al., 2014; Worni et al., 2014). Taylor et al. (2023) estimated that approximately 15 million people are exposed to risks associated with potential GLOFs and that most of these populations are concentrated within High Mountain Asian (HMA) areas. Due to climate warming (Gardelle et al., 2011), the eastern Himalayan area, in particular, has seen a significant increase in the number and area of glacial lakes, thereby increasing the vulnerability of nearby communities to potential GLOF impacts (Hagg et al., 2021).
+
+Although GLOF research and studies have gained global momentum in recent years, only a few studies have been per-formed in Bhutan. Numerous studies conducted in Nepal and
+
+<sup>&</sup>lt;sup>1</sup>National Center for Hydrology and Meteorology, Thimphu, Bhutan
+
+<sup>&</sup>lt;sup>2</sup>Department of Civil and Environmental Engineering, Nagoya University, Nagoya 464-8603, Aichi, Japan
+
+China have simulated and assessed GLOF risks, although detailed studies on Bhutan's exposure to GLOF-related hazards are scarce. Such scarcity can be attributed to a lack of required field data, as well as to Bhutan's limited exposure to the global scientific community.
+
+## 1.2 Past GLOF events in Bhutan
+
+In the past, Bhutan has faced several GLOF events; however, many of these events were either not reported or not documented. One of the most catastrophic GLOFs took place on 6 October 1994, when the moraine dam of Luggye lake partially collapsed, leading to the release of a massive amount of water and debris downstream, destruction to infrastructure and farmland, and the death of 21 people (Watanabe and Rothacher, 1996; Leber et al., 2000). Another significant GLOF occurred in 2009, when an outburst from Tshojo lake, located at the headwaters of the Phochhu river, caused downstream flooding. Based on satellite imagery and a sedimentological and geomorphological analysis, Komori et al. (2012) attributed an outburst from the supraglacial lake on the Tshojo glacier to the event. The most recent GLOF, the Lemthang Tsho outburst, took place on 28 July 2015. Gurung et al. (2017) reported that heavy rainfall triggered the event and that  $0.37 \times 10^6 \,\mathrm{m}^3$  of water was discharged.
+
+## 1.3 Potentially dangerous glacial lakes in Bhutan
+
+Based on the latest report from the National Center for Hydrology and Meteorology (NCHM) in Bhutan, 567 glacial lakes in the country span an area of 55.04 km<sup>2</sup>, accounting for 19.03% of total water bodies (NCHM, 2021). In 2001, the Department of Geology and Mines (DGM) in Bhutan and the International Centre for Integrated Mountain Development (ICIMOD) performed the first-ever inventory of glaciers, glacial lakes, and potentially dangerous glacial lakes (PDGLs), and they identified 24 glacial lakes that fit this category (Mool et al., 2001). However, in 2019, using field-verified data and the latest Sentinel-2 satellite images, NCHM reassessed the number of PDGLs and revised the number to 25, with eight lakes now considered to be safe based on lake morphology, surrounding features, bathymetry conditions, and associated feeding glaciers (NCHM, 2019). Figure 1 provides a map of rivers and the river basin system within Bhutan, together with the distribution of glaciers and glacial lakes. The Punatsangchhu river basin contains 11 PDGLs, which is the largest number of PDGLs per basin the in the country. The Phochhu sub-basin contains nine PDGLs, making it a hotspot for GLOFs and glacially related disasters.
+
+Warming climate exacerbates the hazards of GLOFs, so a comprehensive GLOF assessment is urgently needed since these risks will increase in the coming years. As such, a study assessing hazards associated with glacial lakes and GLOFs is crucial for understanding hazards, as well as their subsequent impacts on hydrological and socio-economic aspects within the Punatsangchhu river basin.
+
+## 1.4 Increasing concern regarding a Thorthomi GLOF
+
+Thorthomi lake is the largest of nine PDGLs within the Phochhu basin. Due to the significant potential risk posed to downstream settlements resulting from a GLOF, the Thorthomi glacial lake has become a serious concern due to the following factors: (1) rapid expansion of the Thorthomi supraglacial lake, (2) the size of glaciers and probable future lake size, (3) the weakened left lateral moraine of the lake due to the 1994 Luggye GLOF, (4) active sliding on the moraine wall separating the Thorthomi and Rapstreng lakes, (5) seepage from the lake, and (6) rock and snow avalanches, as summarised by Karma (2013). To address these factors, the government of Bhutan initiated a highpriority project, referred to as the National Adaptation Plan of Action (NAPA), under the United Nations Framework Convention on Climate Change (UNFCCC) funding scheme in 2006. The project sought to reduce the GLOF risk potential from Thorthomi lake and involved lowering the lake's water level over 4 years, resulting in a reduction of 3.68 m. However, due to challenging working conditions and health issues, the project fell 1.32 m short of its target, although approximately  $17 \times 10^6$  m<sup>3</sup> of lake water was artificially released. The project additionally included setting up a GLOF early warning system along the Punakha-Wangdue valley for alerting residents in the event of a GLOF.
+
+## 1.5 The focus of our study
+
+To contribute to risk management efforts, we evaluated the potential risk of a GLOF from Thorthomi lake. The physically based mathematical dam breach model, BREACH, was used to simulate a glacial lake dam breach and was coupled with the Hydrologic Engineering Centre's River Analysis System (HEC-RAS) to route the flood wave propagating downstream. We sought to simulate both the spatial extent and the lead time of flood wave arrival at several locations along the river. Prior to predicting Thorthomi GLOF hazards and potential risks, we reconstructed the 1994 Luggye GLOF event to validate the dam breach model and the flood wave routing model, which includes river topography and roughness
+
+Our study is one of a few studies that has simulated probable floods from Thorthomi lake and that has estimated inundation extent and flood arrival times within a scientific setting. Such studies form an essential basis for flood risk assessments, early warning system installation, economic planning, countermeasure planning, design, and stakeholder education and awareness programmes.
+
+The main components of our study are as follows:
+
+Image /page/2/Figure/2 description: A map of Bhutan showing its river and basin systems, along with the distribution of potentially dangerous glacial lakes (PDGLs). The map includes a north arrow, latitude and longitude lines from 27°0'0"N to 28°0'0"N and 89°0'0"E to 92°0'0"E, and a scale bar in kilometers. The legend details various features: Rivers (blue lines), Sub-basin boundaries (dashed lines), the Punatsangchhu basin (shaded gray), the Bhutan boundary (solid outline), Glaciers (light blue areas in the north), Main Roads (brown lines), and Urban Areas (yellow patches). The map also plots PDGLs as circles, categorized into two groups. The first group, 'PDGLs(sq.km)', is shown in purple with sizes corresponding to areas: 0.1 - 0.67, 0.67 - 1.46, and 1.46 - 2.9 sq.km. The second group, 'PDGLs(sq.km)-safe', is shown in green with sizes for areas: 0-0.06, 0.06-0.34, and 0.34-0.58 sq.km. Several lakes in the northern region are labeled, including Rapstreng Lake, Thorthomi Lake, and Luggye Lake.
+
+**Figure 1.** A map showing the rivers and basin system of Bhutan, as well as the distribution of potentially dangerous glacial lakes, the main road network, and major urban areas. Bubbles are scaled to total lake area and colour-coded for classification. PDGLs stand for potentially dangerous glacial lakes. Data source: National Center for Hydrology and Meteorology (NCHM).
+
+- estimating the geometry and water volume of Thorthomi lake using available glacial lake geometry data;
+- estimating the potential outburst flood hydrograph using a dam breach model, BREACH, with available physical parameters; and
+- 3. assessing Thorthomi GLOF hazards and potential risks using a 2D hydraulic model.
+
+## 1.6 Structure of the paper
+
+The paper contains six sections. Section 1 introduces the overall concept of a GLOF and provides information obtained from previous studies and information related to GLOFs in the context of Bhutan. Section 2 describes the study area and the GLOF event used for model validation and calibration. We explain materials and methods in Sect. 3. Obtained results are reported in Sect. 4. Based on the results, we discuss the consequences of a Thorthomi GLOF in Sect. 5. We conclude our study in Sect. 6.
+
+# 2 Study area and GLOF events
+
+We assessed flood risk for a catchment of under-construction hydropower plants, specifically the Punatsangchhu Hydroelectric Project Authority (PHPA-I and II), caused by a GLOF from Thorthomi lake. The catchment is the upper part of the Punatsangchhu river basin (PRB), located within the central portion of Bhutan and indicated by the grey area in Fig. 2a. PRB is one of the largest basins in Bhutan, spanning an approximate area of 9760 km², which covers approximately 25 % of the total area of the country (38 394 km²), and is drained by the Phochhu and Mochhu rivers (Fig. 2a) to the Indian plains. The annual averaged discharge of the basin ranges from 194 to 374 m³ s<sup>-1</sup>, with the highest recorded discharge of 2654 m³ s<sup>-1</sup>, observed at the WangdiRapid station (location shown in Fig. 2c), occurring in 2009 during Cyclone Aila.
+
+The Phochhu river, one of the main tributaries of the Punatsangchhu river, originates from the high mountains of Lunana, in northern Bhutan, and flows some 90 km downstream, where it joins Mochhu at Punakha Dzong (monastery) (Fig. 2c), and flows from this area as the Punatsangchhu river. The Thorthomi glacial lake, considered to be one of the most dynamic and dangerous glacial lakes within Bhutan, with an area of 4.3 km<sup>2</sup> (NCHM, 2019), is
+
+Image /page/3/Figure/2 description: The image displays three maps labeled (a), (b), and (c), detailing a study area in Bhutan. Map (a) is an inset showing the location of the Punatsangchu Basin (shaded gray) and a smaller study basin (outlined in red) within Bhutan, which borders China and India. Map (b) is a topographic map of the study basin, with an elevation scale ranging from 1153 m to 7207 m and a distance scale up to 40 kilometers. Map (c) is the main, detailed map of the area. It includes a legend defining symbols for River, Glacial lake, Study Basin, Punatsangchu\_Basin, Settlement Points (circle), Gauging Station (green square), and Power Plant (green triangle). This map shows the Phochhu river and its tributaries, the Mochhu and Dangchhu. Key locations identified include the glacial lakes Thorthomi Lake and Luggye Lake, and settlement points such as Lhedi, Thanza, Dangsa, Samdingkha, Punakha Dzong, Khuruthang, Jagathang, and Bajo. It also marks the WangdiRapid Station (a gauging station), two hydropower plants under construction (PHPA-I & II), and the confluence points of the Phochhu with the Mochhu and Dangchhu rivers. The map indicates the flow direction of the Phochhu river and labels the administrative regions of GASA, PUNAKHA, and WANGDUE.
+
+**Figure 2.** Map of the study area. (a) A Bhutan map showing rivers and river basins (the grey area shows the Punatsangchhu river basin). (b) The elevation distribution of the study area (the catchment area of PHPA-I and II excludes the Mochhu river basin). (c) Major settlement points within the study area.
+
+located at the headwater of the Phochhu sub-basin at over 4440 m above sea level. The Thorthomi glacial lake is widely recognised as the likely consequence of climate warming and, since feeding glaciers terminating in the lake rapidly melt, is expanding each year. Based on a comprehensive analysis of cryospheric, geotechnical, and geomorphological factors, Rinzin et al. (2023) concluded that Thorthomi lake is highly susceptible to GLOF events.
+
+The PRB consists of five administrative districts: Gasa, Punakha, Wangdue Phodrang, Dagana, and Tsirang. These districts constitute 16.6 % (735 533) of the total population of Bhutan (NSB, 2018). The Punakha and Wangdue Phodrang districts (Fig. 2) within the PRB are renowned as Bhutan's primary rice production regions, contributing 16 % and 11 %, respectively, to the nation's total rice output (NSB, 2021). The area is also rich in historical and cultural heritage, with notable landmarks such as Punakha Dzong, which served as the former capital of Bhutan. Fertile flood plains are located along the Phochhu and Punatsangchhu rivers, and the region encompasses settlements such as Samdingkha and Jagathang, together with major towns such as Khuruthang and Bajo. The floodplain of the Punatsangchhu river accommodates these settlements, while downstream, approximately 115 km away from Thorthomi lake, two significant
+
+Image /page/3/Figure/6 description: A topographic map showing four glacial lakes in a mountainous region. The lakes are labeled from left to right (west to east): Rapstreng Lake, Thorthomi Lake, Luggye Lake, and DrukChung Lake. A river, identified as 'Lake Outflow', flows along the southern side of the lakes. An arrow indicates the 'Flow Direction' is from east to west (right to left). A legend in the top right corner defines the symbols used: a blue shape for 'Glacial lake', a thin blue line for 'Lake Outflow', and a black dot for 'Lake outlet'. Each lake has a designated outlet marked on the map. In the bottom left, there is a scale bar indicating a length of 4 kilometers. A north arrow is located in the top left corner. The background of the map consists of gray contour lines representing the terrain.
+
+**Figure 3.** Four glacial lakes within the Lunana region (the Lunana complex).
+
+hydropower plants, PHPA-I and II, are currently under construction (see the bottom of Fig. 2c for locations of the two hydropower plants). Given the exposure of critical infrastructure and settlements to potential GLOFs from PDGLs, especially the Thorthomi glacial lake, an assessment of hazards within this area is of paramount importance.
+
+The Luggye glacial lake is one of the PDGLs in Lunana, Bhutan's northern region. The lake is one of four glacial lakes in an area that spans a few kilometres (Fig. 3) and had an outburst in 1994. Although there is no detailed official documentation on the GLOF at the Luggye glacial lake, reports and articles describing the event do exist (e.g. Koike and Takenaka, 2012; Meyer et al., 2006; Richardson and Reynolds, 2000; Watanabe and Rothacher, 1996). The event was also documented in a technical report (Leber et al., 2000), when the Royal Government of Bhutan launched a major investigative project in 2000 to study the cause of the event.
+
+The 1994 Luggye GLOF was a cascading phenomenon, where sudden drainage of the upstream Druk Chung glacial lake (Fig. 3) into Luggye lake increased hydrostatic pressure on the moraine dam of Luggye lake, releasing  $18 \times 10^6$  m<sup>3</sup> of flood water (Leber et al., 2000). The 1994 Luggye GLOF claimed the lives of 21 people and inflicted major damage to infrastructure and downstream settlements; notably, the Punakha Dzong (monastery) suffered significant damage, although it is located 93 km downstream from the lake (Richardson and Reynolds, 2000; Watanabe and Rothacher, 1996). During this period, a peak flow rate of  $2539 \,\mathrm{m}^3 \,\mathrm{s}^{-1}$ was observed at the WangdiRapid gauging station, located 15 km downstream of Punakha Dzong (Fig. 2) and approximately 108 km downstream of the flood source (data from NCHM). Contributions from the Mochhu (96  $\mathrm{m}^3$  s<sup>-1</sup> on 7 October 1994), Dangchhu basins (no gauging station in the basin), and other small tributaries to the peak flow rate should have been minor due to limited base flow during the season.
+
+To estimate the breach outflow hydrograph, several studies have attempted to reconstruct the 1994 Luggye GLOF event (e.g. JICA, 2001; Koike and Takenaka, 2012; Meyer et al., 2006). Koike and Takenaka (2012) estimated that peak discharge from the Luggye lake breach ranged from 1800 to
+
+ $2500 \,\mathrm{m}^3 \,\mathrm{s}^{-1}$ , depending on inflow conditions measured by Yamada et al. (2004).
+
+# 3 Materials and methods
+
+A schematic diagram showing input data, used models/methods, and outputs is provided in Fig. 4. We reconstructed the 1994 GLOF event to verify (1) a glacial lake geometry estimation model, (2) a dam breach model, (3) a digital elevation model and its error correction method, and (4) Manning's coefficient. Then, the same models were used to predict the potential risk caused by a Thorthomi GLOF.
+
+## 3.1 Uncertainty in GLOF bathymetry
+
+The volume and geometry of a glacial lake are regulating factors of the glacial lake outburst process. Glacial lakes are generally formed from a depression left behind by retreating glaciers, which, in most cases, are produced when a moraine is filled with meltwater. Depending on geomorphology, the presence of sediment, and glacial over-deepening capacity, formed glacial lakes can manifest specific lake bathymetry and influence glacial hydrology (Cook and Swift, 2012). Due to remote locations and high elevations, accessing and conducting field surveys to map glacial lake bathymetry is challenging.
+
+Despite challenges, measurements of lake bathymetry are crucial for determining a lake's volume and surface area and are necessary for assessing potential flood volumes and the risk of GLOFs. In 2019, the National Centre for Hydrology and Meteorology (NCHM) conducted bathymetric surveys in 14 of the 25 identified potentially dangerous glacial lakes and mapped their maximum depth and volume. However, even though the Thorthomi glacial lake is considered to be a critical lake that could burst in the future, due to the difficulty associated with conducting a survey, the bathymetry of the Thorthomi glacial lake remains unknown. Since lake geometry is a crucial parameter for dam breach modelling and subsequent hydraulic routing, lake depth and volume needed to be estimated.
+
+## 3.2 Estimating geometries of glacial lakes
+
+Estimating the potential flood volume of a glacial lake is critical for determining the magnitude of a GLOF. Bathymetry data are necessary for calculating lake volume as well as potential flood volume, but bathymetry for glacial lakes is scarce due to the challenging and inaccessible environments in which glacial lakes are often located, including for Thorthomi lake. Although several reports have estimated Thorthomi lake volume (Karma, 2013; Singh, 2009), no details on how volumes were estimated have been documented.
+
+To address data scarcity for glacial lake geometries, various studies have proposed methods for estimating glacial lake depth and volume based on other more accessible pa-
+
+rameters such as lake area (Cook and Quincey, 2015; Huggel et al., 2002; O'Connor et al., 2001; Sakai, 2012), as well as depression angle from the lakeshore (Fujita et al., 2013) and surrounding topography (Heathcote et al., 2015). Empirical relationships such as area–volume and area–depth are useful for estimating a lake's depth and potential flood volume. Cook and Quincey (2015) refined the area–volume relationship proposed by Huggel et al. (2002) by increasing sample size and removing duplicated samples. They also classified the predictability of lake volume and depth based on regions and lake types and determined that predictability is influenced by a lake's origin and evolution. The relationship proposed by Cook and Quincey (2015) takes the following form:
+
+$$D_{\text{mean}} = 0.1697 A^{0.3778},\tag{1}$$
+
+where  $D_{\text{mean}}$  is the mean depth (in m) and A is the area (in m<sup>2</sup>). The volume–area relationship (V, volume in m<sup>3</sup>) can simply be derived by multiplying the area of both sides, as follows:
+
+$$V = 0.1697A^{1.3778}. (2)$$
+
+Sakai (2012) used a similar approach and proposed a model for estimating maximum depth instead of mean depth. The bathymetric measurement data of 17 glacial lakes (15 moraine-dammed glacial lakes and 2 thermokarst lakes) from Bhutan, Nepal, and Tibet were used to derive an area—maximum depth—volume relationship, so estimations of depth and volume from the area of glacial lakes could be determined (Sakai, 2012). The regression equation took the following form:
+
+$$D_{\text{max}} = 95.665 A^{0.489},\tag{3}$$
+
+where  $D_{\text{max}}$  is maximum depth (in m) and A is area (in km<sup>2</sup>). The volume–area relationship (V, volume in  $10^6 \text{ m}^3$ ) takes the following form:
+
+$$V = 43.24A^{1.5307}. (4)$$
+
+For predicting the moraine dam breach process explained in the following section, lake geometries, especially the maximum depth, are crucial, so we employed the equations proposed by Sakai (2012). The equations allow the independent calculation of maximum depth and volume. As conceptualised by Cook and Quincey (2015), bathymetry of the lake was estimated based on idealised geometric shape. The lake bottom was also assumed to follow an elliptical shape, as commonly observed in most moraine-dammed glacial lakes in Bhutan.
+
+## 3.3 The moraine dam breach and its modelling
+
+### 3.3.1 Previous studies
+
+GLOFs are triggered by a breach of a moraine dam that holds the lake in place and are caused by an external triggering
+
+Image /page/5/Figure/2 description: A flowchart illustrating a methodology for creating an inundation map and hazard zonation. The chart uses a legend to define its components: gray rounded rectangles for 'Input Data', blue parallelograms for 'Model', and green rounded rectangles for 'Result'. The process starts with input data including 'Digital Elevation Model', 'Lake bathymetry data', and 'Lake moraine data'. These inputs, along with processed data like 'Field bathymetry data for GLOF reconstruction' and 'Regression analysis data for GLOF prediction', are fed into a 'Dam Breach Model'. The output of this model is a 'Breach flow hydrograph'. This result, along with the 'Digital Elevation Model' and other inputs like 'Land Use Land Cover Data' and 'Satellite imagery', goes into a 'Hydrodynamic Model'. The 'Hydrodynamic Model' produces 'Simulated flow Hydraulics'. This result, combined with 'Land Use Land Cover Data', 'Satellite imagery', and 'Settlement data', is processed by a 'GIS' model. The final output of the entire process is an 'Inundation Map Hazard Zonation'.
+
+Figure 4. A schematic diagram of the methodology employed in our study.
+
+event. While the structure of the dam itself is an important factor, destabilisation of a dam due to a trigger event is the primary cause of a breach. Since the overtopping of lake water is the major failure mode (Awal et al., 2010; Begam et al., 2018; Neupane et al., 2019), our study assumed that GLOFs are triggered by the overtopping of lake water.
+
+To estimate flood flow and associated hazards resulting from a dam breach, several studies have simulated dam breach floods using dam breach models (Bajracharya et al., 2007; Hagg et al., 2021; Huggel et al., 2002; Koike and Takenaka, 2012; Maskey et al., 2020; Meyer et al., 2006; Shahrim and Ros, 2020; Wang et al., 2008; Worni et al., 2014). BREACH is a numerical model describing the dam breach process and the resulting outflow hydrograph. The model is based on fundamental principles of hydraulics, sediment transport, soil mechanics, and the physical properties of dam materials and the reservoir. The model is physically based and was designed to predict the size, shape, and time of dam breach development, as well as the resulting flow rate and the volume of water released. Unlike parametric models, physically based breach models, including BREACH, consider the geotechnical aspects of dam materials, as well as hydraulic and sediment transport (Fread, 1988; Maskey et al., 2020; Worni et al., 2014), which increases the predictive accuracy of future GLOF processes. Due to this, the BREACH model has been widely used in studies of dam breach flood hazards and risk assessments (Fread, 1988).
+
+Koike and Takenaka (2012) used the BREACH model coupled with the flood flow model, FLO-2D, to perform a sce-
+
+nario analysis on the risks of a GLOF on the Mangdechhu river basin, due to an outburst flood of the Metatshota glacial lake in Bhutan. The study concluded that although the breaching potential of the lake is low due to the wide crest and gentle slope of the moraine dam, a GLOF would affect several houses and farmland located on the flood plain (Koike and Takenaka, 2012). Hagg et al. (2021) performed a GLOF hazard assessment within the Mochhu basin in Bhutan using the HEC-RAS dam break module, simulating a dam breach of the Shintaphu glacial lake, and concluded that risk is comparably small.
+
+Our study used BREACH for describing the dam breach process for target lakes due to its better predictive accuracy for a future extraordinary Thorthomi GLOF event. The dam is assumed to breach due to overtopping flow resulting from a trigger event, such as an ice calving/avalanche or a rock avalanche. Most of the geotechnical properties of dam materials required as an input parameter are available in the report published by the National Center for Hydrology and Meteorology (NCHM, 2020). A few properties were published by Koike and Takenaka (2012). Some unavailable data were estimated by referring to previous studies.
+
+### 3.3.2 Reconstruction of the 1994 Luggye GLOF dam breach
+
+For reconstruction of the 1994 Luggye GLOF, the dam breach outflow hydrograph was estimated using BREACH (Fread, 1988). The bathymetry of Luggye lake (Fig. 5) and
+
+Image /page/6/Figure/2 description: A bathymetric map showing the depth elevation of a lake. The lake is elongated and slightly curved. A north arrow is in the top left corner. The legend in the bottom left indicates the lake boundary with a thin black line and the lake outlet with a red-brown circle, which is located at the westernmost tip of the lake. A scale bar at the bottom shows a total length of 1.3 kilometers, with marks at 0, 0.325, and 0.65 kilometers. On the right, a vertical color scale labeled "Bathymetry Depth elevation (m)" ranges from 4353.5 (dark purple, deeper) to 4465.8 (dark green, shallower). The map shows the lake is shallowest around the edges and has its deepest point in the central-eastern area. The easternmost part of the lake is a white area labeled "No Data".
+
+Figure 5. The bathymetry of Luggye lake (data from NCHM).
+
+the material properties of the moraine dam (the middle row of Table 1) required for the model were based on various reports (NCHM, 2019, 2020). Topographic data of the moraine dam were derived from the digital surface model (DSM). Since wave overtopping is a more common failure mode for moraine-dammed glacial lakes as compared to a piping failure (Neupane et al., 2019), to estimate breach outflow from Luggye lake, overtopping failure of the moraine dam was assumed. The properties of moraine dam material have a significant effect on the growth of a breach (Maskey et al., 2020; Westoby et al., 2014). The mechanism in which the formation of a breach largely occurs determines the shape of the breach outflow hydrograph (Westoby et al., 2014). Therefore, gathering accurate in situ data for reliable breach process reproduction is essential. Based on the estimation by Fujita et al. (2008), deduced from a combination of field measurements and remote-sensing observations, the level of lake water was reduced 19 m during the event.
+
+### 3.3.3 The Thorthomi GLOF prediction
+
+Breach initiation is assumed to occur due to an overtopping wave at the existing outlet (Fig. 3) induced by any probable triggering event. Moraine material properties and topographic data for the Thorthomi lake BREACH model were either estimated from available terrain data or adopted from available reports and research documents, and this is also listed in the right row of Table 1.
+
+## 3.4 Flood routing
+
+### 3.4.1 The hydrodynamic model
+
+A hydrodynamic model is essential for understanding the characteristics of a flood wave caused by a GLOF propagating downstream, as well as for quantitatively evaluating the potential risks caused by a flood. To simulate the propagation of outflow from glacial lake breaches in Nepal, numerous studies, such as HEC-RAS used in Maskey et al. (2020), have employed various hydrodynamic models. Similar approaches that couple dam breach models to hydrodynamic models (e.g. Bajracharya et al., 2007; Koike and Takenaka, 2012; Westoby et al., 2015; Worni et al., 2014) have been
+
+performed for modelling the GLOF process chain in various regions.
+
+Worni et al. (2014) provided a summary of various hydrodynamic models that have been used to model GLOFs. Discussed models include HEC-RAS, FLO-2D, BASEMENT, and Delft3D. The choice of a hydrodynamic model depends on factors such as the end objective, data availability, and the available budget. Each model has its own level of accuracy; however, the accuracy of results is primarily dependent on the precision of the elevation model, including channel geometry and floodplain topography. Errors in the elevation model can lead to inaccuracies in results (Casas et al., 2006; Xu et al., 2021).
+
+HEC-RAS is a commonly used hydrodynamic model that allows users to perform 1D and 2D steady/unsteady flow simulations (Brunner, 2016b). We used the HEC-RAS to perform a 2D unsteady flow simulation for floods caused by a glacial lake dam breach. Since they represent spatially varied flood hydraulics (Horritt and Bates, 2001), the twodimensional models employed are standard in flood modelling. In a 2D unsteady simulation, flow varies in time, along two spatial dimensions, and processes are predicted by the laws of conservation of mass (continuity) and the conservation of momentum for two horizontal directions. We used a full set of momentum equations (the shallow water equations) to simulate flooding as clear water flow. Although high-viscosity and hyper-concentrated (sediment entrained) flows are inherent to the GLOF phenomenon (Clague and Evans, 2000; Vuichard and Zimmermann, 1987), to simplify modelling complexity and data requirements, most studies (Hagg et al., 2021; Koike and Takenaka, 2012; Maskey et al., 2020; Rinzin et al., 2023) have simulated GLOFs as clear water flow.
+
+Other important considerations in hydrodynamic modelling are Manning's roughness coefficient and channel geometry. Both have significant impacts in predicting inundation extent and flow characteristics (Mosquera-Machado and Ahmad, 2007; Ye et al., 2018; Zhu et al., 2019). Hagg et al. (2021) demonstrated the influence of Manning's roughness coefficient for glacial lake outburst floods from the Shintaphu glacial lake in Mochhu basin, Bhutan, and concluded that channel roughness is not essential for inundation extent but exerts a significant effect on flood velocity and flood arrival time.
+
+Since such information is needed to estimate the area needed for evacuation and the lead time for evacuation, flood travel time and peak flow are essential parameters for early warning purposes. In this study, flood travel time was calculated based on timing of the breach outflow hydrograph and the flow hydrograph at the point of interest, when there was significant inundation depth and extent. Peak flow is the maximum simulated flow resulting from a dam breach.
+
+| Moraine dam data                                 | 1994 Luggye GLOF | Thorthomi GLOF |
+|--------------------------------------------------|------------------|----------------|
+| Surface area of the lake (km2) (RSA)             | 1.46b            | 4.3b           |
+| Volume of water in the lake (106 m3), Eq. (4)    | 65.19b           | 400a           |
+| Maximum depth of the lake (m), Eq. (3)           | 96.93b           | 161a           |
+| Top elevation of the dam (m) (HU)                | 4465a            | 4446a          |
+| Toe elevation of the dam (m) (HL)                | 4370a            | 4345a          |
+| Slope of the upstream face of the dam (1 : ZU)   | 1 : 4.8a         | 1 : 6.2a       |
+| Slope of the downstream face of the dam (1 : ZU) | 1 : 6.5a         | 1 : 6.3a       |
+| Dam material properties                          |                  |                |
+| Grain size ( <i>D</i> 50) (mm)                   | 1.362b           | 2.01b          |
+| Porosity (%)                                     | 36.5c            | 36.5c          |
+| Cohesive strength (kN m-2)                       | 1.5b             | 1.5b           |
+| Internal friction (°)                            | 41b              | 39b            |
+| Unit weight (kN m-3)                             | 22.92b           | 22.43b         |
+| Manning's coefficient (s m-1/3)                  | 0.07b            | 0.07b          |
+
+Table 1. Input parameters for the 1994 Luggye GLOF reconstruction and the Thorthomi GLOF prediction.
+
+### 3.4.2 Ground elevation models and pre-conditioning
+
+The accuracy of hydrodynamic model results is heavily influenced by the quality of the elevation model used (Gyasi-Agyei et al., 1995; Yamazaki et al., 2014, 2017). Casas et al. (2006) demonstrated the effects of a topographic data source and resolution on flood peak discharge and the extent of inundation and then concluded that laser-based elevation data are a suitable source for hydraulic modelling. Similarly, the influence of grid size on inundation propagation and water depth under varied topographical settings in 2D modelling has been analysed (Tsubaki and Kawahara, 2013). Both fine grid size representing main topographic features of the floodplain and accurate elevations at each grid point are essential for simulating flood flow with less uncertainty. Therefore, using the best available elevation model for the hydrodynamic simulation of floods is essential.
+
+Since an accurate elevation model is essential for accurate hydrodynamic simulations, various methods for correcting generic noise errors and biases originating from topography measurements have been proposed and have been used in elevation models prior to running hydrodynamic/hydrological analyses.
+
+Below, we compare three different elevation models covering the study area. The first model, the Multi-Error-Removed Improved-Terrain (MERIT) Hydro DEM (Fig. 6a), was developed based on SRTM and AW3D DEM. Water layer data at a 3 arcsec resolution ( $\sim 90\,\mathrm{m}$ ) were developed for river hydrology analyses at global, as well as at local, scales (Yamazaki et al., 2017, 2019). Other bias-corrected elevation data are the Forest And Buildings removed Copernicus DEM (FABDEM) (Fig. 6b), developed from Copernicus DEM (COPDEM), where the height of trees and buildings
+
+are removed using machine learning, and are also a preferable source for terrain data (Hawker et al., 2022). The AW3D digital surface model (DSM) (Fig. 6c) was jointly developed by the Remote Sensing Technology Centre (RESTEC) and the NTT DATA Corporation, utilising PRISM data acquired by the Advanced Land Observing Satellite (ALOS) of the Japan Aerospace Exploration Agency (JAXA). Roughresolution DSMs are distributed by several organisations free of charge; however, for our study, we used the finer commercial DSM product, AW3D 2.5 m, as the primary source of topography. The cell size of the DSM for the focus area was approximately  $2.2 \,\mathrm{m}$  for both the X and Y directions and was projected to WGS84 UTM zone 45N. The AW3D 2.5 m DSM represents details of topography, especially in the river valley and in the developed flood plain, much better than other models.
+
+Since the AW3D DSM was obtained using satellite photogrammetry, representations of the river bottom, especially in forested and deep gorge areas, are sometimes inaccurate. If we directly used topography in hydrodynamic modelling, the DSM covered structures along the river, as well as bridges crossing the river, disturbing flood water flow. To avoid such anomalies in the elevation model and in the hydrodynamic simulation results, a river channel delineation was performed. The presence of spikes within the DSM, along the river's path, can obstruct flood water flow, resulting in the formation of non-existent deep pools. One way to improve topography surrounding a river is the use of bathymetric survey data. However, no such survey has been conducted within the study area. To improve representation for the river channel, our study utilised a rule-based correction method.
+
+The Agriculture Conservation Planning Framework (ACPF) is a GIS-based tool developed by the United States
+
+<sup>&</sup>lt;sup>a</sup> Estimated in this study. <sup>b</sup> NCHM (2020). <sup>c</sup> Koike and Takenaka (2012).
+
+Image /page/8/Figure/2 description: A figure with three panels labeled (a), (b), and (c), showing topographic maps of the same area at different resolutions. The area shown is around Punakha Dzong. Panel (a) is a low-resolution, pixelated map with a scale bar from 0 to 1.2 kilometers. Panel (b) shows the same area at a higher resolution, with less prominent pixelation. Panel (c) is a high-resolution, smooth map that includes a legend for Punakha Dzong and a north arrow in the upper right corner. All three maps use a color gradient from brown for higher elevations to blue-green for lower elevations and show the location of Punakha Dzong with a black outline.
+
+Figure 6. The Punakha Dzong region as represented by three different terrain models: (a) the MERIT HydroDEM 90 m, (b) the FABDEM-30 m, and (c) the AW3D 2.5 m DSM.
+
+Department of Agriculture (USDA) to identify areas with impeded water flow and to improve hydrologic flow using flow direction and an accumulation analysis (Porter et al., 2016). While the ACPF is a valuable tool for hydrologic flow and watershed planning, it has limited applicability for terrain correction in hydrodynamic modelling because the ACPF does not allow users to define the bathymetry of a river channel.
+
+Another widely used channel modification method is the inbuilt function of HEC-RAS. The Channel Design/Modification editor tool is a module used to modify an unrealistic cross-section or to introduce a user-defined channel crosssection (Brunner, 2016a). The tool effectively removes spikes in the elevation model along a river channel while maintaining the natural slope of the represented topography. The modified channel TIN (triangulated irregular network) can be overlain on the original DSM and exported as a single raster file with modified features. Rinzin et al. (2023) applied the method to modify terrain and to delineate river flow paths for a GLOF simulation. For our study, we used this tool to condition the DSM in the middle region of the model domain, a forested deep gorge, where huge spikes were included within the elevation model. The modification was only applied to the channel section, and the remaining portion was left as it was.
+
+### 3.4.3 Implementation for GLOF reconstruction and prediction
+
+Downstream propagation of the flood was simulated using the HEC-RAS model for the 1994 Luggye GLOF reconstruction. The calculation domain was defined by the 2D flow area. The overall size of the flow area was  $64 \,\mathrm{km^2}$ . The domain was modelled using a 20 m resolution computational grid, consisting of 157 188 computational cells, and solved with a time step of 1 s. The elevation of each grid cell was specified based on a 2.2 m, hydro-conditioned digital surface model, and Manning's n was set to 0.035, in the range provided by the HEC-RAS manual (Brunner, 2016a). The dam breach outflow hydrograph obtained from the BREACH
+
+model was used as the upstream boundary; and normal depth, calculated based on downstream slopes derived from the DSM, was used as the downstream boundary condition.
+
+For the Thorthomi GLOF prediction, we used a hydrodynamic model similar to the model used for the 1994 Luggye GLOF reconstruction. The domain was a bit shortened because the Thorthomi glacial lake is located 2 km downstream of the Luggye glacial lake. The overall size of the flow area was 62 km<sup>2</sup>. The domain was modelled using a 20 m resolution computational grid consisting of 153 790 computational cells with a temporal resolution of 1 s.
+
+# 4 Results
+
+## 4.1 The 1994 Luggye GLOF reconstruction
+
+### 4.1.1 Dam breach processes
+
+Simulated peak flow  $(Q_p)$  of the dam breach outflow hydrograph was  $6030\,\mathrm{m}^3\,\mathrm{s}^{-1}$ , and  $1.9\,\mathrm{h}$  ( $\sim 114\,\mathrm{min}$ ) was required to reach peak flow (the time to peak,  $T_p$ ) (Fig. 7a). The volume of the GLOF and the reduction of lake water level due to the event were  $21\times10^6\,\mathrm{m}^3$  and  $19\,\mathrm{m}$ , respectively, in agreement with the findings of Fujita et al. (2008), which was based on a combination of in situ observations and remotesensing data. Dimensions for the estimated breach of the dam are provided in Fig. 7b.
+
+### 4.1.2 Downstream peak flow and flood travel time
+
+Flow hydrographs at various locations along the flow path are provided in Fig. 8. Conforming to the findings of Meyer et al. (2006), after approximately 6 h, the GLOF had a peak discharge of 2897 m<sup>3</sup> s<sup>-1</sup> as it reached Punakha Dzong, located 93 km downstream of the lake. Peak flow at the WangdiRapid station (shown in Fig. 2c), 15 km downstream of Punakha Dzong, was 2455 m<sup>3</sup> s<sup>-1</sup>, close to the recorded value of 2539 m<sup>3</sup> s<sup>-1</sup>. Here, the recorded flow rate included the contribution of normal flow from tributaries, which was not ac-
+
+Image /page/9/Figure/2 description: The image contains two diagrams labeled (a) and (b). Diagram (a) is a line graph titled "Flow vs. Time". The x-axis is labeled "Time (h)" and ranges from 0 to 24. The y-axis is labeled "Flow (m³ s⁻¹)" and ranges from 0 to 6000. The graph shows a sharp peak in flow, reaching approximately 6000 m³ s⁻¹ at around 2 hours, and then gradually decreasing over the 24-hour period. Diagram (b) is a 3D representation of a trapezoidal channel cross-section. The top width of the channel is labeled as 102 m. There are two vertical height measurements indicated: one is 19 m from the bottom of the channel to an intermediate level, and the other is the total height of 22 m from the bottom to the top.
+
+Figure 7. (a) A breach outflow hydrograph from the BREACH model and (b) an illustration of breach parameters, breach width  $(\leftrightarrow)$ , breach depth  $(\updownarrow)$ , and the change in water surface elevation  $(\updownarrow)$ .
+
+Image /page/9/Figure/4 description: A line graph showing Flow (m³ s⁻¹) versus Time (h). The y-axis, labeled "Flow (m³ s⁻¹)", ranges from 0 to 4000. The x-axis, labeled "Time (h)", ranges from 0 to 24. There are two curves on the graph. The black curve, labeled "Punakha Dzong", starts rising sharply around 5.5 hours, peaks at approximately 2900 m³/s at 6 hours, and then gradually decreases. The blue curve, labeled "WangdiRapid Station", starts rising around 7 hours, peaks at about 2500 m³/s around 8 hours, and then gradually decreases. A red horizontal arrow labeled "~ 6 h" and identified in the legend as "Flood Travel Time" extends from time 0 to the peak of the Punakha Dzong curve at 6 hours.
+
+**Figure 8.** A simulated GLOF hydrograph at different locations along the flow path.
+
+counted for in our analysis. Good agreement of results for simulated flow and flood travel time with observed data, as well as previous studies, indicated that the performance of the employed models and the modelling approach were adequate and capable of yielding satisfactory results for predictive modelling of the target lake. The total inundated area along the basin was approximately 13.1 km<sup>2</sup>.
+
+## 4.2 Future Thorthomi GLOF prediction
+
+### 4.2.1 Lake bathymetry
+
+The estimated volume and maximum depth of Thorthomi lake based on Eqs. (3) and (4) were  $400 \times 10^6$  m<sup>3</sup> and 161 m, respectively (the top right of Table 1). The estimated volume and maximum depth of Thorthomi lake falls within the predicted band, considering a 95% confidence level. The utilised equations showed a good relationship between area and volume and between and area and maximum depth, with the prediction of  $400 \times 10^6$  m<sup>3</sup> for the mean and  $281 \times 10^6$  m<sup>3</sup> and  $560 \times 10^6$  m<sup>3</sup> for the lower bound and the upper bound, respectively. The prediction range for maximum depth was 161 m for the mean prediction (130 m and 270 m for the lower bound and the upper bound, respectively). Compared to other glacial lakes in Bhutan, the estimated parameters indicate that the Thorthomi glacial lake is one of the largest and deepest lakes. The bathymetry of Thorthomi lake, estimated based on the above parameters, is provided in Fig. 9.
+
+A recent study from Nepal proposed a glacial lake volume estimation equation by considering the width and length ratio
+
+Image /page/9/Figure/11 description: A bathymetric map of the Thorthomi Lake Area. The map includes a legend, a north arrow, a scale bar, and a color scale for elevation. The legend indicates symbols for Outlets (red circle), Outflow (wavy blue line), Thorthomi Lake Area (solid outline), PDGL area (dashed outline), and Glaciers (grey shaded area). The map shows the lake with a single outlet and outflow on its western side. Glaciers are located at the northern end of the lake, and a PDGL area is situated along the eastern side. The bathymetry, representing depth and elevation in meters, is shown with a color gradient from green (4446 m) to brown (4285 m), indicating the lake is deepest in its central and southern parts. A scale bar at the bottom is marked in increments up to 1.8 kilometers.
+
+**Figure 9.** The estimated bathymetry of Thorthomi lake.
+
+of the lake (Qi et al., 2022). Based on the equation provided in Qi et al. (2022), the water volume of Thorthomi lake can be estimated as  $227 \times 10^6 \, \text{m}^3$ . This volume is substantially small compared to the volume estimated by Eq. (4). The discrepancy may be related to the dataset used for each study, namely Eq. (4) is based on lakes in Bhutan, Nepal, and Tibet; and Qi et al. (2022) is based on lakes in the Peruvian Andes and other areas, including non-moraine-dammed lakes.
+
+### 4.2.2 Dam breach processes
+
+Different dam breach scenarios for maximum breach and a partial breach for a half breach width and depth (50% of maximum width and depth) were simulated to ascertain the potential risk under various breaching possibilities, including a partial dam breach, which occurred in 1994 (e.g. the 1994 Luggye GLOF). Simulated peak flow  $(Q_p)$  resulting from the Thorthomi dam breach under different breach scenarios ranged from  $9700 \,\mathrm{m}^3 \,\mathrm{s}^{-1}$  (for a 50 % breach depth) to 16 360 m<sup>3</sup> s<sup>-1</sup> (for maximum breach width and depth), with a time to peak  $(T_p)$  of 3.4 to 4 h, respectively (Fig. 10a). The bathymetry of the lake and the topography of the moraine dam dictate the total lake drawdown depth and the volume of the outburst flood. In this study, we estimated that 100 m of lake water depth will be lowered before the breach outflow channel becomes sufficiently stable, after sending  $283 \times 10^6 \,\mathrm{m}^3$  (approximately 70 % of estimated lake water) of flood water downstream (Fig. 10b). The breach outflow channel was assumed to be stable when its bottom elevation reached the natural bed level of the downstream channel and down-cutting ceased.
+
+Image /page/10/Figure/2 description: The image contains two panels, labeled (a) and (b). Panel (a) is a line graph that plots Flow in cubic meters per second (m³ s⁻¹) against Time in hours (h). The x-axis ranges from 0 to 24 hours, and the y-axis ranges from 0 to 15,000 m³ s⁻¹. There are three curves on the graph, each representing a different scenario: 'Maximum Breach' (dark red line), 'Half Breach Width' (blue line), and 'Half Breach Depth' (black line). The 'Maximum Breach' curve peaks at the highest flow, over 15,000 m³ s⁻¹ at approximately 4 hours. The 'Half Breach Depth' curve peaks at around 13,500 m³ s⁻¹ at about 3.5 hours. The 'Half Breach Width' curve has the lowest peak, just under 10,000 m³ s⁻¹ at about 4.5 hours. Panel (b) is a 3D diagram of a trapezoidal channel, illustrating its dimensions. The top width is indicated as 225 m, and the depth is shown as 100 m.
+
+Figure 10. (a) A dam breach outflow hydrograph obtained from the BREACH model for three different scenarios and (b) breach parameters, breach width  $(\leftarrow)$ , and breach depth  $(\cline{1})$  in metres for the maximum breach scenario.
+
+## 4.3 Downstream peak flow and flood travel time
+
+The simulated flow hydrographs for three different scenarios at eight major settlement areas are provided in Fig. 11. Peak flow of the GLOF gradually attenuated as it propagated downstream. Peak flow at Punakha Dzong ranged from 8900 to  $14\,130\,\mathrm{m}^3\,\mathrm{s}^{-1}$  and decreased from 8200 to  $11\,500\,\mathrm{m}^3\,\mathrm{s}^{-1}$  when it arrived at the hydropower plant (PHPA-I).
+
+A schematic representation of an approximate distance, peak flow, averaged channel slope, and the estimated flood travel time for a maximum breach condition is provided in Fig. 12 (refer to Sect. 3.4.1 for the definition of the peak flow and flood travel time in this study). The estimated peak flow at Punakha Dzong,  $14\,130\,\mathrm{m}^3\,\mathrm{s}^{-1}$ , is expected to be over 5 times higher than the 1994 Luggye GLOF (the recorded value is  $2539\,\mathrm{m}^3\,\mathrm{s}^{-1}$ , and the value estimated in this study is  $2455\,\mathrm{m}^3\,\mathrm{s}^{-1}$ ).
+
+# 5 Discussion
+
+## 5.1 Inundation hazard in five vulnerable areas under the maximum breach scenario
+
+The flood depth distribution, highlighting five vulnerable areas for a maximum breach scenario, is provided in Fig. 13. The villages of Thanza, Toncho, and Lhedi, located in the northernmost part of the study area (Fig. 2), are expected to be inundated under a Thorthomi GLOF scenario. The 1994 Luggye GLOF also caused major damage to these settlement areas, but a Thorthomi GLOF is expected to cause more severe damage due to larger flood volume and shorter lead time. Major settlements along the river basin lie in the lower valleys of the Punakha and Wangdue districts, where large areas are expected to be flooded. Major towns and settlements, such as Samdingkha, Khuruthang, and Bajo, are expected to be inundated. The Mochhu river converges with the Phochhu river at the top left of Fig. 13c. Substantial overflow surrounding the Mochhu river, around the confluence, has been predicted. This result is due to backwater flow from the Phochhu river. Water flow from the Mochhu river is not easy to accurately estimate in advance and was not accounted for in this study, so inundation surrounding Punakha Dzong may be underestimated. However, the contribution of water from the Mochhu river can be negligible because the base flow of the Mochhu river is approximately  $100\,\mathrm{m}^3\,\mathrm{s}^{-1}$ , which is substantially small compared to estimated peak flow in this area,  $14\,130\,\mathrm{m}^3\,\mathrm{s}^{-1}$ . Total inundated area due to a Thorthomi GLOF, with a maximum breach, was estimated to be approximately  $22\,\mathrm{km}^2$ , which is almost twice the area inundated under the Luggye GLOF simulation ( $13.1\,\mathrm{km}^2$ , estimated in this study).
+
+## 5.2 A comparison between three scenarios
+
+Figure 14 compares the maximum inundation depth and extent for three different scenarios for the town of Khuruthang. Simulation results for the three scenarios considered in this study revealed that the overall inundation extent and flood depths were higher for the maximum breach scenario. However, the depth and flood extent for the two other scenarios were comparable to the maximum breach scenario. The results indicate that even for a partial breach of the moraine dam, substantial damage within the downstream settlement areas (Punakha and Wangdue) is expected. The results imply that the difference in glacial lake bathymetry may also affect the maximum inundation in downstream areas but is not very sensitive because of the nature of the GLOF event (consisting of a rapid dam breach process and flood routing in steep valleys).
+
+## 5.3 Time series change of flood depth distribution surrounding Punakha Dzong
+
+The spatial distribution of flood depth for a maximum breach scenario, at different time steps, for Punakha Dzong and Khuruthang town are provided in Fig. 15. Due to higher peak flow and a longer flood duration, overall flood hazard potential for the inhabited area caused by the Thorthomi lake GLOF, as compared to damages during the 1994 Luggye GLOF, was significantly higher. Most of the flood path lies in the narrow V-shaped valley, where there are few to no settlements or infrastructure. We estimated that over 1277 houses, most in the lower region of the study area, will be inundated in a GLOF. Aside from this, infrastructure such as roads, bridges, and sand dredging equipment will be damaged.
+
+Notable damage during the 1994 GLOF occurred in Punakha Dzong. The area near the Punakha Dzong was completely inundated in the 1994 Luggye GLOF. The simulated future GLOF indicates that the Punakha Dzong area will be completely flooded, with a maximum depth of over 10 m (Figs. 13 and 15).
+
+## 5.4 Socio-economic impact
+
+The Punakha (the middle downstream of the domain; see Fig. 13) and the Wangdue districts (consisting of the Bajo and Jagathang settlements, as well as the downstream do-
+
+Image /page/11/Figure/2 description: An image displaying eight line graphs in a 2x4 grid, each showing simulated flood hydrographs for different locations. The locations are Lhedi, Samdingkha, Punakha Dzong, and Khuruthang in the top row, and Jagathang, Bajo, WangdiRapid station, and Hydropower plant in the bottom row. Each graph plots Flow in cubic meters per second (m³ s⁻¹) on the y-axis against Time in hours (h) on the x-axis, which ranges from 0 to 24. A legend in the top-left graph indicates three scenarios: 'Max' (red line), 'BRD-50%' (black line), and 'BRW-50%' (blue line). In all graphs, the 'Max' scenario shows the highest peak flow, followed by 'BRD-50%', and then 'BRW-50%'. The peak flows generally decrease and occur later in time for locations further down the grid. For example, in the 'Lhedi' graph, the 'Max' flow peaks at approximately 16,000 m³/s around 5 hours, while in the 'Hydropower plant' graph, the 'Max' flow peaks at approximately 11,000 m³/s around 9 hours.
+
+**Figure 11.** A simulated flow hydrograph at important locations, derived from the HEC-RAS; result for each scenario (max: maximum breach; BRD-50 %: half of maximum breach depth; BRW-50 %: half of maximum breach width).
+
+Image /page/11/Figure/4 description: A schematic diagram illustrating the characteristics of a flood along a river path, starting from Thorthomi Lake. The top part of the diagram shows a profile view of the river's elevation, which decreases as the distance from the lake increases. Key locations along the path are marked with their distances from the lake: Lhedi (19 km), Samdingkha (83 km), Punakha Dzong (90 km), Khuruthang (94 km), Bajo (103 km), and a Hydropower Plant (115 km). Below this profile, three horizontal bars provide quantitative data for different segments of the river. The first bar, labeled "Peak Flow (m³ s⁻¹)", shows a decreasing flow rate from 16,300 at Lhedi to 11,550 at the Hydropower Plant. The second bar, "Average Slope (dy/dx)", shows the slope for each segment, with values ranging from 0.001 to 0.038. The third bar, "Flood Travel Time (h)", indicates the time taken for the flood to traverse each segment, with values such as 2.6 hours for the first segment and 7.1 hours for the last.
+
+Figure 12. A schematic representation of flood parameters at six important locations along the flow path for the maximum breach scenario.
+
+main; see Figs. 2 and 13) are leading producers of rice, an essential crop for the country's GDP and food security. Any damage to agricultural land would have a devastating impact on farmers and the nation. Aside from potential damage to buildings and infrastructure, such as roads and bridges, agricultural land would also become submerged and destroyed by a flood. We estimated that approximately 193 to 245 ha of agricultural land will be inundated under different scenarios
+
+in a Thorthomi GLOF event. Figure 16 shows the potential extent of floods for different land use classes and highlights probable damage to agricultural land, particularly in the areas of Samdingkha and Jagathang.
+
+The overall hazard potential of a GLOF from Thorthomi lake under different scenarios is summarised in Table 2. Although the peak flow rate of each scenario is different (29 % to 37 % between the maximum and minimum for the result
+
+Image /page/12/Figure/2 description: A figure displaying a series of maps illustrating flood inundation in a study area. The figure is composed of a central legend, a study area map, and five detailed maps labeled (a) through (e).
+
+The legend defines symbols for Bridges, Education Centres, Health Care Centres, Historical Places, Road Network, Settlements, River, and Thorthomi\_Lake.
+
+The study area map shows a river originating from Thorthomi\_Lake with five locations marked along its path: (a), (b), (c), (d), and (e).
+
+Each of the five detailed maps shows a satellite view of one of these locations with a flood inundation overlay, a scale bar, and a legend for flood depth in meters (m).
+
+- Map (a) Thanza: Shows a wide flooded area. The flood depth legend is: 0 - 5 m, 5.1 - 10 m, and 10.1 - 15.6 m. The scale bar goes up to 50 kilometers.
+- Map (b) Samdingkha: Shows a flooded river channel. The flood depth legend is: 0 - 6 m, 6.1 - 12 m, and 12.1 - 18.7 m. The scale bar goes up to 1.6 kilometers.
+- Map (c) Punakha Dzong: Shows a flooded area at a river confluence. The flood depth legend is: 0 - 7 m, 7.1 - 14 m, and 14.1 - 23.2 m. The scale bar goes up to 2 kilometers.
+- Map (d) Khuruthang: Shows a flooded river bend. The flood depth legend is: 0 - 8 m, 8.1 - 15 m, and 15.1 - 25.8 m. The scale bar goes up to 3 kilometers.
+- Map (e) Jagathang & Bajo: Shows a flooded winding river. The flood depth legend is: 0 - 7 m, 7.1 - 15 m, and 15.1 - 22.6 m. The scale bar goes up to 3 kilometers.
+
+Each map uses shades of blue to represent flood depth, with darker shades indicating deeper water. Icons from the legend are placed on the maps to show the locations of infrastructure and settlements.
+
+Figure 13. A maximum GLOF inundation map of the study area under the maximum breach scenario. Map data: © Google Earth 2023, CNES/Airbus, Maxar Technologies.
+
+depicted in Fig. 11), the total inundation area, the number of submerged buildings, and the area of impacted cultivated land are not much different (12%, 22%, and 21%, respectively), implying that the estimated flood is significant even for the most minor flood scenario (BRW-50% scenario) for the Thorthomi GLOF. The scenarios indicate that most of the damage will occur for river properties and that farmland will be substantially damaged, even when a dam breach is not drastic. The soil in farmlands will also be eroded and covered by debris. Damage to irrigation is expected and may affect agriculture in farmland located behind flooded areas. Over the long term, damage to soil and irrigation would extensively reduce farmers' production. In advance, careful evacuation planning and business continuity planning (e.g. JICA,
+
+2015), including a plan for agriculture, are essential for mitigating damage caused by a future Thorthomi GLOF.
+
+## 5.5 Limitations of this study
+
+Due to a lack of actual surveyed data, volume and maximum depth were estimated based on the statistical relationships established by past studies; an uncertainty for the estimated bathymetry of Thorthomi lake is a major limitation of our study. As compared to the bathymetry presented, the use of actual, surveyed bathymetric data may yield a more accurate prediction. An additional limitation of our study is the clear water assumption. Compared to clear water, hyperconcentrated water has different dynamic properties. Debris in flood water may cause substantial damage to farmland, in-
+
+Image /page/13/Figure/2 description: A figure composed of four panels, labeled (a), (b), (c), and (d), illustrating a flood model for a river system. Panel (a) is a map showing the overall area. It includes a north arrow and a scale bar from 0 to 40 kilometers. The legend in panel (a) identifies symbols for Khuruthang Town, a River, the Model Domain, Thorthomi Lake, and the Maximum breach inundation extent. The map displays Thorthomi Lake at the top, feeding a river that flows downwards. A small rectangular area on the map is magnified in the other panels. Panels (b), (c), and (d) are three similar close-up aerial views of a town next to a river, showing flood inundation. Each of these panels includes a legend for water depth in meters (m), with three color-coded ranges: light gray for 0 - 10 m, medium blue for 10.1 - 15 m, and dark blue for 15.1 - 25.9 m. The blue overlay on the aerial images shows the extent and depth of the floodwaters, with the deepest water in the main river channel.
+
+**Figure 14.** A comparison of inundation depth and extent for three breach scenarios within the Khuruthang study area. (a) A model domain highlighting Khuruthang town. (b) The maximum breach scenario. (c) A 50 % breach depth scenario. (d) A 50 % breach width scenario. Map data: © Google Earth 2023, CNES/Airbus, Maxar Technologies.
+
+Image /page/13/Figure/4 description: A multi-panel figure, labeled Figure 15, illustrating the temporal change of the spatial extent of flood depth at Punakha Dzong and Khuruthang. Panel (a) is a map showing the model domain, which includes Thorthomi Lake in the north and a river flowing south. Two locations are marked: Punakha Dzong and Khuruthang Town. The legend indicates symbols for the river, model domain, and Thorthomi Lake. A scale bar shows distances up to 20 kilometers. Panels (b) through (g) show a time-series of the flood's progression in a zoomed-in area around Punakha Dzong and Khuruthang Town. Each of these panels has a legend for water depth in meters, categorized into three levels: 0-10 m (light blue), 10.1-15 m (medium blue), and 15.1-25.9 m (dark blue). The panels show the increasing extent and depth of the flood at different time steps: (b) 05:54, (c) 06:06, (d) 06:12, (e) 06:24, (f) 06:30, and (g) 08:00. The sequence shows the flood wave arriving, peaking in depth and extent around time step 06:30, and then beginning to recede by 08:00, with significant inundation affecting Khuruthang Town.
+
+Figure 15. The temporal change of the spatial extent of flood depth at Punakha Dzong and Khuruthang. (a) The hydrodynamic model domain. (b-g) The inundation depth at six time steps.
+
+frastructure, and human life. Research on glacial lakes and their outburst floods is an emerging field (e.g. Qi et al., 2022; Taylor et al., 2023). To obtain more accurate damage predictions, data and methods should be revised following research progress.
+
+The close proximity of glacial lakes within the Lunana region, especially the Thorthomi and Rapstreng lakes (Fig. 3), poses an even greater potential risk due to a possible cascad-
+
+ing GLOF event. Failure of the lateral moraine of Thorthomi lake would lead to lake water breaching into Rapstreng lake, which would consequently cause the failure of its moraine dam. Since our study considered failure of the terminal moraine in the direction of the existing outlet, it is highly unlikely for such an event to occur under the current scenario. Accordingly, a cascading GLOF was not assessed in our study, but such possibilities should also be explored to
+
+Image /page/14/Figure/2 description: A figure displaying three land use maps of a mountainous region, with a comprehensive legend. The main map on the left shows a large area with a scale bar from 0 to 20 kilometers. It depicts a valley with various land use classes, including large areas of green (Alpine Scrubs; Forests; Meadows; Shrubs) and light blue (Moraines; Snow and Glacier). A white line delineates the 'Simulation Domain' along the valley. Two smaller, zoomed-in maps are on the right, labeled 'Samdingkha' and 'Jagathang', each with a scale bar from 0 to 1.2 kilometers. These maps show a more detailed, pixelated view of sections of the river, with a red line indicating the 'Inundation Boundary'. The legend provides a key for the 'Land use class' colors: Black for 'Built up', Blue for 'Water Bodies', Brown for 'Cultivated Agriculture', Green for 'Alpine Scrubs; Forests; Meadows; Shrubs', Light Blue for 'Moraines; Snow and Glacier', and Grey for 'Non Built up; Rocky Outcrops'. All maps include a north arrow.
+
+Figure 16. The probable GLOF inundation extent on land use classes. (Land use data source: National Land Commission Secretariat, Bhutan.)
+
+**Table 2.** The damage potential of a GLOF from Thorthomi lake.
+
+| Hazards →<br>Scenarios↓                 | Total inundation area (km2) | Number of buildings inundated | Total cultivated agricultural land impacted (ha) |
+|-----------------------------------------|-----------------------------|-------------------------------|--------------------------------------------------|
+| Maximum breach                          | 22.7                        | 1277                          | 245.6                                            |
+| Half of maximum breach depth (BRD-50 %) | 20.8                        | 1044                          | 206.4                                            |
+| Half of maximum breach width (BRW-50 %) | 19.9                        | 1000                          | 193.4                                            |
+
+better understand the potential risk of cascading events which may cause more severe damage to society.
+
+# 6 Conclusion
+
+We explored future hazards and damages arising from a GLOF from Thorthomi lake, one of the potentially dangerous glacial lakes in Bhutan but not well investigated within the scientific literature to date. To validate the approach used in our study and to calibrate the model, we reconstructed the 1994 Luggye lake GLOF prior to assessing the hazards of a Thorthomi GLOF. The BREACH model was used to es-
+
+timate the outflow hydrograph emanating from a failure of moraine dams due to overtopping flow. Moraine materials and soil parameters used to parameterise the model were obtained from a report published by the National Center for Hydrology and Meteorology (NCHM), Bhutan. Propagation of the GLOF was simulated using a 2D routing module in HEC-RAS for modelling unsteady flow, which is an inherent characteristic of a GLOF where there is a sharp rise in the flow hydrograph.
+
+The bathymetry of Thorthomi lake was estimated based on a regression equation derived from the relationship between lake area—depth—volume found within moraine lakes. We estimated that the total volume of the lake is approximately  $400 \times 10^6 \,\mathrm{m}^3$ , with a maximum depth of 161 m. According to the maximum breach scenario, the Thorthomi GLOF may release  $283 \times 10^6 \,\mathrm{m}^3$  of water in under 12 h, with a peak flow rate of  $16\,360 \,\mathrm{m}^3 \,\mathrm{s}^{-1}$ , occurring approximately 4 h following initiation of the breaching process. Outflow hydrographs estimated by the model were used as the upstream boundary condition in hydrodynamic modelling.
+
+Flood routing was performed to reach a length of approximately 115 km, and then peak discharge, flood travel time, and flood depths at major downstream settlements were estimated. According to the maximum breach scenario, Punakha Dzong, which lies 90 km downstream of Thorthomi lake and at the beginning of major settlements, would witness a peak discharge of  $14\,128\,\mathrm{m}^3\,\mathrm{s}^{-1}$ , approximately 6 h following breach initiation. A potential GLOF from Thorthomi lake would cause extensive agricultural and infrastructural damage to 245 ha of agricultural lands, and, for a maximum breach scenario, 1277 buildings are expected to be inundated. Comparable damage is also expected for two minor flood scenarios, implying that such damage is inevitable for a future Thorthomi GLOF.
+
+A hazard assessment for a GLOF plays a crucial role for understanding and mitigating risks associated with these devastating natural events. Our study quantified the potential danger a GLOF from Thorthomi lake would pose to the downstream settlements and infrastructure. Such assessments will enable policymakers, local communities, and relevant stakeholders to make informed decisions regarding land use planning, disaster preparedness, and early warning systems.
+
+Since glacial environments are dynamic and subject to change due to climate variations, GLOF hazard assessments are not static. Continuous monitoring and regular reassessments of glacial lakes and associated hazards are essential to account for environmental shifts and to ensure the effectiveness of mitigation strategies. Furthermore, a multidisciplinary approach in GLOF hazard assessments is necessary. Collaborations between researchers, policymakers, local communities, and other stakeholders are essential for effective decision-making, disaster preparedness, and the implementation of mitigation measures. To essentially reduce GLOF risk, the development of methods to safely release dammed water to downstream areas is important.
+
+*Data availability.* Our study used open and commercial data (NSB, 2021, 2018; NCHM, 2021). Commercial data may be distributed under licence terms and conditions.
+
+Author contributions. TW: conceptualisation, data curation, method, and writing original draft; RT: conceptualisation, data curation, draft writing, and reviewing and editing.
+
+Competing interests. The contact author has declared that neither of the authors has any competing interests.
+
+*Disclaimer.* Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors.
+
+Acknowledgements. Tandin Wangchuk acknowledges a Human Resource Development Scholarship from the Japan International Cooperation Agency (JICA). We acknowledge Yuji Toda and Takashi Tashiro, as well as other lab members of the Hydraulic Research Laboratory at Nagoya University. We also express our gratitude to Shigeo Suizu, Tomoyuki Wada, and Toru Koike at Earth System Science Co., Ltd. We sincerely thank members of the National Center for Hydrology and Meteorology for their continued support of our study.
+
+Review statement. This paper was edited by Pascal Haegeli and reviewed by Stefan Ram and Rayees Ahmed.
+

chatbot/rag/data/('An ontology for Emergency Managing and', '.pdf')_extraction.md

@@ -0,0 +1,247 @@
+
+
+Abstract—In the domain of emergency management during hazard crises, having sufficient situational awareness information is critical. It requires capturing and integrating information from sources such as satellite images, local sensors and social media content generated by local people. A bold obstacle to capturing, representing and integrating such heterogeneous and diverse information is lack of a proper ontology which properly conceptualizes this domain, aggregates and unifies datasets. Thus, in this paper, we introduce empathi ontology which conceptualizes the core concepts concerning with the domain of emergency managing and planning of hazard crises. Although empathi has a coarse-grained view, it considers the necessary concepts and relations being essential in this domain. This ontology is available at https://w3id.org/empathi/.
+
+Index Terms—Ontology, Vocabularies, Crisis Management, Hazard Domain, Emergency, Ontology Quality, Knowledge Reuse, Disaster Management.
+
+# I. INTRODUCTION
+
+We can not change the nature. However, we can promote our planning, preparation, and response strategies about crises happening in the three phases of hazards, i.e. (i) pre-hazard, (ii) in-hazard and (iii) post-hazard. Currently, a substantial body of situational information collected from sources such as satellite images, sensors, social media content generated by people who are involved in crisis reporting and response, etc. Indeed, harnessing and exploiting this hazard-related Big Data is an essential means towards situational awareness which helps to manage crises, threads and risks of hazards in each phase. Despite the availability of such data, still there is a significant deficiency in representing, annotating and more importantly integrating this heterogeneous hazardrelated big data. This deficiency can be relieved by providing an ontology which conceptualizes and organizes situational and environmental awareness data (events, activities) subjected to hazards of any kind. Our investigation in the state-of-the-art hazard-related conceptualization (i.e., taxonomy, vocabulary and ontology) revealed existing of a few works which mainly conceptualized either hazard domain or crisis management domain from a limited perspective or for a particular type. For example, Humanitarian eXchange Language (HXL) [1] and EDXL-RESCUER Ontology [2] are mainly concerned with help and rescue aspect of crisis management domain. However, the broader and more diverse nature of this domain requires a comprehensive and abstract modeling and representation. Another deficiency is related to lack of incorporating relations into the conceptualization. Thus, they indeed should be called vocabulary or taxonomy rather than ontology (e.g., Management of Crisis vocabulary (MOAC)). These shortages motivated us to introduce an ontology that takes into consideration the prior conceptualizations (taxonomies and vocabularies) while relying on a promoted representation. Also, it brings new concepts and relations which were ignored previously although playing an essential role in capturing situational awareness (e.g., surveillance information, human sensing report, humanitarian event: prayer, concepts synonyms and instances). This paper presents our contribution in providing an ontology for Emergency Managing and PlAnning abouT Hazard crIses (empathi).
+
+The rest of the paper is organized as follows: Section II presents the relevant state-of-the-art hazard or crisis vocabularies and compare their main features. Section III lists the external vocabularies which are integrated empathi. Thereafter, Section IV introduces empathi along with its major top concepts. Section VII reviews the related work. We close with the conclusion and the future work.
+
+# II. STATE-OF-THE-ART VOCABULARIES
+
+In this section, we present an overview of the state-of-the-art vocabularies concerned with hazard domain as well as crisis management domain. With this respect, Table I represents a succinct comparative study of these vocabularies. The first column (i.e., Vocabulary) states the name of the state-of-theart vocabulary or ontology and empathi is included in the last row (based on recency). The second column (i.e., **Domain** Coverage) mentions the particular areas of the hazard-related domain or crisis management domain which it covers. The third column (i.e., URI) checks whether or not the URI of the vocabulary is dereferenceable. The fourth column (i.e., F for File) specifies the available format of the vocabulary (i.e., OWL, RDF, TTL). The fifth column (i.e., **D** for **Documenta**tion) shows whether the vocabulary has online documentation or not. Then, we list the significant publications utilizing this vocabulary within the sixth column (i.e., MC for Major Citations). The seventh and eight columns represent the number of classes #C) and relations #R) specified within the vocabulary. The last column represents the external resources (i.e., IV for Imported Vocabularies) imported by their respective vocabulary in the first column. In the following, we shortly describe each vocabulary.
+
+| Vocabulary | Domain Coverage                                                                          | U | F           | D | MC                                                | #C  | #R  | IV                                                                    |
+|------------|------------------------------------------------------------------------------------------|---|-------------|---|---------------------------------------------------|-----|-----|-----------------------------------------------------------------------|
+| HXL        | Disaster,Geography,Damage,<br>Organization, Humanitarian Response                        | ✓ | TTL         | ✓ | [3], [1], [4], [5]<br>[9], [10], [11], [12], [13] | 50  | 66  | [6], [7], [8]                                                         |
+| MOAC       | Impact of Crisis, Recovery<br>and Response Activities,<br>Geo-locations                  | ✗ | RDF         | ✓ | [3], [14]<br>[10], [16]<br>[17], [18]             | 70  | 30  | [7], [15]                                                             |
+| SMEM       | Social Media and Emergency<br>Management                                                 | ✗ | ✗           | ✓ | -                                                 | -   | -   | [19], [20], [21], [22], [7], [8]<br>[23], [24], [1], [25], [26], [27] |
+| DO         | Temporal and Spatial<br>Concepts, Impact,<br>Rehabilitation and Facilities<br>Facilities | ✓ | Web-<br>App | ✗ | -                                                 | 97  | -   | -                                                                     |
+| ERO        | Report Specification                                                                     | ✗ | ✗           | ✓ | [28], [29], [30], [31], [32]                      | -   | -   | [33], [34], [35]                                                      |
+| DoRES      | Events and Reports<br>Specification                                                      | ✓ | RDF         | ✓ | -                                                 | 96  | 261 | [7], [8]<br>[23], [36]                                                |
+| EF         | Fire Disaster Specification,<br>Protocol Design and Planning                             | ✗ | ✗           | ✓ | -                                                 | 37  | 90  | -                                                                     |
+| empathi    | Hazard Situational Awareness,<br>Crisis Management, Hazard Events                        | ✓ | OWL         | ✓ | -                                                 | 423 | 338 | [37], [19], [20], [8], [7]<br>[36], [21], [23], [38], [39]            |
+
+Table I: Comparison of the state-of-the-art Hazard-related Vocabularies, Taxonomy, and Ontologies. U, F, D respectively stand for referenceability of URI, availability of File and Documentation. Furthermore, MC, C, R and IV respectively stand for Major Citations, number of classes, number of relations, and imported vocabularies. ERO: EDXL-RESCUER Ontology, DO: Disaster Ontology, EF: Emergency Fire.
+
+- a) HXL: HXL stands for Humanitarian eXchange Language. HXL<sup>1</sup> is a standard aiming at information sharing during humanitarian calamity by overcoming the burden of interoperability. HXL ontology has a total of 50 classes and 66 relations. Main concepts contained in HXL are Place, Survey and assessment, Operation, Cash and Finance, Crisis. Furthermore, HXL provides links to UN OCHA vocabularies such as Global Coordination Groups<sup>2</sup>, Disaster Types<sup>3</sup>, Organization Types<sup>4</sup>, Vulnerable groups and Humanitarian themes<sup>5</sup>. Furthermore, HXL provides a hashtag schema containing related social media tags such as #channel, #crisis, #impact, #event, etc. [1].
+- b) **MOAC**: MOAC<sup>6</sup> which is concerned with management of crisis is a vocabulary [40] providing concepts mainly related to crisis management. It was created by the Inter-Agency Standing Committee (IASC)<sup>7</sup>, Emergency Shelter Cluster in Haiti<sup>8</sup>, UN-OCHA 3W Who What Where Contact Database<sup>9</sup> and Ushahidi Platform<sup>10</sup>.
+- c) **DoRES**: DoRES stands for DOcument-Report-Event-Situation Ontology. DoRES<sup>11</sup> is an ontology sharing information between individuals and organizations using situational reports for describing the situation [13]. It helps humanitarian organizations to structure their plans.
+- d) EDXL-RESCUER Ontology: EDXL-RESCUER<sup>12</sup> stands for Emergency Data Exchange Language Reliable and Smart Crowdsourcing Solution for Emergency and Crisis
+
+Image /page/1/Figure/18 description: A diagram illustrating the components that feed into a central concept called "empathi". The central concept is enclosed in a yellow oval. There are two main categories of components, each with its own label and color coding. On the left, under a blue rectangular label "Crisis/Hazard Related Taxonomy and Lexicon", there are six blue ovals: FEMA, MA-Ont, HXL, MOAC, EM-DAT, and Geo-Names. On the right, under a green rectangular label "Imported Vocabularies", there are six green ovals: DC-Terms, SKOS, LODE, SIOC, iContact, and FOAF. Arrows point from each of these twelve ovals towards the central "empathi" oval, indicating that it integrates information from all these sources.
+
+Figure 1: Importing existing ontologies in empathi
+
+Management. An ontology-based on EDXL developed for coordinating and interchanging information with the legacy system [32], [2].
+
+- e) Emergency Fire (EF): It is an ontology explicitly designed for fires in the building. It comprises of 131 terms along with definitions created after subjective research. It serves as a protocol for information sharing, analysis, evaluation and comprehension by an organization in the situation of disaster caused by fire [41].
+- f) Social Media Emergency Management (SMEM): During an unprecedented onset of a natural hazard, social media overflows with textual content about situational information, prayers, weather information, the impact of crisis and events. Of all the available information on the social media,
+
+<sup>&</sup>lt;sup>1</sup>https://github.com/hxl-team/HXL-Vocab/blob/master/Tools/hxl.ttl
+
+<sup>&</sup>lt;sup>2</sup>https://goo.gl/CD6vHY
+
+<sup>&</sup>lt;sup>3</sup>https://reliefweb.int/taxonomy-descriptions#disastertype
+
+<sup>4</sup>https://goo.gl/Uzy9UA
+
+<sup>&</sup>lt;sup>5</sup>http://vocabulary.unocha.org
+
+<sup>&</sup>lt;sup>6</sup>http://observedchange.com/moac/ns
+
+<sup>&</sup>lt;sup>7</sup>https://goo.gl/ESn99F
+
+<sup>8</sup>https://goo.gl/nDoa9F
+
+<sup>9</sup>https://goo.gl/jQLnYh
+
+<sup>10</sup>https://goo.gl/XPSnyG
+
+<sup>11</sup>https://goo.gl/Sw4XGt
+
+<sup>12</sup>http://www.rescuer-project.org
+
+which portion is contextually relevant and action-oriented to personnel in charge of a relief-giver organization. SMEM ontology provides concepts in a hierarchical structure which transforms high-volume of messy content to low-volume of action-oriented information [42].
+
+g) Disaster Ontology: It is one of the ontologies listed in Finnish Ontology Library Service ONKI<sup>13</sup>. Disaster Ontology (DO)<sup>14</sup> is comprised of 97 concepts (classes) concerning man-made and natural hazard. This ontology is useful for managing disaster situations but disregards concepts related to social media (e.g., news reports, modality of data, surveillance, prayer and monitoring the status of the services provided by organizations).
+
+# III. INTEGRATION OF EXTERNAL VOCABULARIES
+
+In this section, we list the external vocabularies which partially integrated into empathi. Not all of them are necessarily related to the domain of hazard or crisis management (we reuse generic concepts from well-known vocabularies, e.g., FOAF). Figure 1 concisely represents an integration aims at reusing the existing vocabularies following ontology design methodologies (Methontology [33] and NeOn [43]) or interlinking *empathi* to other vocabularies which enhance its visibility.
+
+- Federal Emergency Management Agency (FEMA)<sup>15</sup> provides a glossary of terms related to disaster preparation and management [37].
+- **Emergency Disasters Database** (**EM-DAT**)<sup>16</sup> provides a precise definition of concepts and furthermore a categorization of disturbance-related events [19].
+- MA-Ont<sup>17</sup> supports detailed properties describing media files and appropriate metadata mapping [20].
+- **iContact**<sup>18</sup> provides conceptual classes for defining international addresses. It is relevant for using GeoNames [36], [21] ontology for describing places.
+- **Friend Of A Friend (FOAF)**<sup>19</sup> was created for describing people, relations, and associated events. Coupled with SIOC [23] and disaster domain model [39], [38], it can describe social media communities formed during the disaster scenarios [7].
+- **GeoNames**<sup>20</sup> is a part of GeoNames Database providing information about 11 million places (toponyms) covering all the countries. Integration of GeoNames ontology to our ontology adds geospatial semantic information which is critical for actionable hazard response. The ontology contains 150 classes and two relations forming 758 axioms on location dereferencing. Mapping syntax provided by this ontology is compatible with schema.org, DBpedia
+
+- Linked Open Descriptions of Events (LODE)<sup>21</sup> defines event as an action which takes place at a certain time and has a specific location. It can be a historical action as well as a scheduled action. Thus, it provides the generic concept of Event along with locational (i.e., atPlace), temporal (i.e., atTime) aspects and people who play a role (i.e., involvedAgent).
+- **Simple Knowledge Organization System** (**SKOS**)<sup>22</sup>. We utilized this data model to describe the concepts of our domain. It provides a better organization of domain knowledge (i.e., Hazard Crisis) [8].
+- Semantically-Interlinked Online Communities (SIOC)<sup>23</sup> is a W3C ontological standard to describe information from online communities. It can support a volunteer or caregiver with actionable information in the realm of social media [23].
+
+# IV. CORE CONCEPTS OF EMPATHI
+
+Image /page/2/Figure/15 description: A concept map or ontology diagram showing relationships between different entities. At the top, an oval labeled 'empathi' has a 'has subclass' relationship pointing to 'Impact'. 'Impact' in turn has a 'has subclass' relationship with 'Affected Population'. 'Affected Population' is a central node with several outgoing relationships: 'has subclass' to 'Trapped People' and 'Missing People'; 'requireHumanSupport' to 'Volunteer Support'; 'needHelp' to 'Service'; and 'contactInfo' to 'iContact Thing'. Both 'Trapped People' and 'Missing People' have an 'ageRange' relationship pointing to 'Age Group'. The 'Service' node has three relationships: 'isLocatedAt' pointing to 'Place', 'currentServiceStatus' pointing to 'Status', and 'contactInfo' pointing to 'iContact Thing'. The arrows representing relationships are color-coded and have different styles (solid or dashed).
+
+Figure 2: Partial representation of the concept Affected Population in *empathi* 
+
+As Table I shows, *empathi* contains 423 classes and 338 relations. In figure 2, concepts linked to "Affected Population" via solid lines, forms structural concepts ("is-a"/ "subclass"), while the concepts linked via colored dotted line are semantically related concepts to "Affected Population". Furthermore, in figure 2, the concept "iContactThing" is taken from the vocabulary *iContact* shown in figure 1. In the following, we present only the super-classes which imply the generic coverage.
+
+- **Age Group:** this class groups people based on their age similarity by providing the following sub-classes (i) Adolescent, (ii) Adult, (iii) Child and (iv) Infant.
+- **Event:** this concept defines events along with spatial and temporal constraints happening in any phase of hazard. This concept embodies the following sub-classes
+
+ontology, LinkedGeoData ontology and INSEE ontology [36].
+
+<sup>13</sup> https://onki.fi/en/
+
+<sup>14</sup>http://onki.fi/en/browser/overview/disaster
+
+<sup>15</sup>https://goo.gl/QtKzev
+
+<sup>&</sup>lt;sup>16</sup>http://www.emdat.be/Glossary.Itisadatabase
+
+<sup>&</sup>lt;sup>17</sup>https://www.w3.org/ns/ma-ont
+
+<sup>18</sup> http://ontology.eil.utoronto.ca/icontact.html
+
+<sup>19</sup>http://xmlns.com/foaf/spec/
+
+<sup>&</sup>lt;sup>20</sup>http://www.geonames.org/ontology/documentation.html
+
+<sup>&</sup>lt;sup>21</sup>http://linkedevents.org/ontology/
+
+<sup>&</sup>lt;sup>22</sup>https://www.w3.org/2004/02/skos/
+
+<sup>&</sup>lt;sup>23</sup>http://rdfs.org/sioc/spec/
+
+| Hazard Type | Tweet                                                                                                                      | Hazard Concept              |
+|-------------|----------------------------------------------------------------------------------------------------------------------------|-----------------------------|
+| Flood       | 1. 188 killed, Airport closed they say that the runway came peeling off. it may take time to resume.                       | Impact (AP, ID)             |
+|             | 2. @? Hope people get adequate relief and no one is left out.Nation stands with Chennai.                                   | Event (HP)                  |
+|             | 3. @? in who can offer places to stay, pls fill the form for volunteers collating info #chennairains                       | Event (VS)                  |
+|             | 4. @? Chennai has been declared disaster zone. Army has been deployed. Army Helpline - +XXX XXX XXXX #ChennaiRains         | Service (H)                 |
+| Hurricane   | 5. Surveillance video captures several people looting a Houston store #Harvey in the wake of Hurricane                     | Event (CA)                  |
+|             | 6. The aftermath. 10,000 people now homeless because of Hurricane Irma. #WednesdayWisdom #climatechange                    | Impact (AP), Event(CC)      |
+|             | 7. #HurricaneIrmaRecovery Drive for #Homestead & #FloridaKeys today! Drop off supplies at @? #DJLMS #dontgivebackjustgive. | Service (S)                 |
+|             | 8. At least 56 of Florida's 639 nursing homes still have *no* electricity this morning, five days after #HurricaneIrma     | Impact (ID)                 |
+| Blizzard    | 9. @? "2-3 days, it could take before airlines begin to clear the backlog? @? at on #blizzard2016.                         | Impact (ID)                 |
+|             | 10. @?: I-75 in Kentucky closed due to large number of accidents, state patrol says #blizzard2016                          | Impact (AP,ID)              |
+|             | 11. @?: The baton is passed. Buoy 50 miles south of Wilmington #Jonas #blizzard2016                                        | Place (L)                   |
+|             | 12. @?: Blizzard with "life and death implications" hits Washington, Mid-Atlantic #blizzard2016                            | Place (L), Impact (AP)      |
+| Landslide   | 13. @?: Be sure to follow @BGSLandslides for lots of up to date information on landslides across the UK #StormFrank        | Report (ER)                 |
+|             | 14. #StormFrank landslide at Glasscarraig Norman Motte & Bailey in Co Wexford — @? #archaeology #floods                    | Place (L)                   |
+|             | 15. SRI LANKA: At least 73 dead after week of flooding, landslides; 243,000 in temp shelters — TorStar #ExtremeWeather     | Service (SH), Impact (AP)   |
+|             | 16. @?: #EcuadorEarthquake - landslides closing down roads & making it challenging for help to reach hardest hit towns     | Impact (ID), Hazard Type(*) |
+
+Table II: Sample of hazard-related tweets from different hazard types. AP: Affected Population, ID: Infrastructure Damage, HP: Human Prayer, VS: Volunteer Support, H: Helpline, CA: Criminal Activity, S: Supply, L: Location, ER: Expert Report, SH: Shelter, Hazard Type (\*): One hazard (Earthquake) causing another hazard (Landslide).
+
+- (i) Climate Change (ii) Criminal Activity (iii) Emergency Exercises (iv) Evacuation Plan (v) Humanitarian Event (vi) Recovery Plan (vii) Rescue Operation (viii) State Mitigation Plan (ix) Volunteer Support and (x) Early Warning.
+- **Facility:** defines an amenity made accessible for a specific purpose. It attributes to following sub-classes (i) Communication (ii) Electricity (iii) Gas Facility (iv) Water Facility and (v) Education Resource.
+- Hazard Type: lists different types of hazards that can affect human community. It is an entity type that embodies sub-classes (i) Airburst (ii) Coastal erosion (iii) Drought (iv) Earthquake (v) Explosion (vi) Fire (vii) Flood (viii) Hurricane (ix) Landslide (x) Sandstorm (xi) Storm (xii) Tornado (xiii) Toxic Radioactivity (xiv) Tsunami (xv) Volcano and (xvi) Winterstorm.
+- **Hazard Phase:** categorizes different activities carried out by various organizations before, during and after a catastrophic event into three sub-classes (i) During Hazard (ii) Pre-Hazard and (iii) Post-Hazard.
+- **Impact:** a forceful negative affect on someone or something in an unprecedented manner. This concept embodies sub-classes (i) Affected Population (ii) Animal Loss (iii) Health Issues (iv) Food Shortage (v) Financial Crisis (vi) Contamination (vii) Infrastructure Damage (viii) Severity.
+- **Involved Actors:** People or Organisation associated (negative or positive) with any catastrophic event. Sub-classes of this concept are (i) Organisation (ii) People.
+- Modality of Data: Information (raw, structured or semi-
+
+- structured) conveyed or represented by a particular arrangement or sequence of text, audio, video or photos. Sub-classes included by this concept are (i) Audio (ii) Photo (iii) Text and (iv) Video.
+- **Place:** a physical surrounding defined by longitude, latitude, and area, providing a relative position of the someone or something during a hazard situation. One sub-class of the Place is Location, described by longitude, latitude, and the area of the affected place.
+- Report: documents evidence of the destruction caused by the natural disaster. All the activities carried out by various governmental and non-governmental organizations (NGOs) are stated in the report. The report is a way to keep people vigilant. Sub-classes encompassed under the concept: Report is (i) Expert Report (ii) Human Sensing Report and (iii) Media Report.
+- **Service:** is an act of providing support to someone in a situation of distressing incidents. Core sub-classes of this concept are (i) Financial Care (ii) Healthcare Service (iii) Helpline (iv) Human Remains Management (v) Resource and Information Centre (vi) Supply (vii) Transportation and (viii) Prayer Location.
+- **Status:** defines the state of services that are planned during the pre/in/post hazard phases. Associated subclasses are (i) Available (ii) Offered (iii) Requested and (iv) Unavailable.
+- **Surveillance Information:** A systematic, ongoing collection, collation, and analysis of data and the timely dissemination of information to those who need to know
+
+Image /page/4/Figure/0 description: A diagram illustrating a hierarchical classification system, mapping concepts to a text example. At the top, a central oval labeled 'empathi' is the root of a tree structure. Solid arrows point downwards to create the hierarchy. From 'empathi', branches lead to 'Impact', 'Hazard phase', 'Place', and 'Hazard Type'. The 'Impact' node further branches into 'Population Affected' and 'Damage to Infrastructure'. 'Population Affected' leads to 'Dead and Toll', and 'Damage to Infrastructure' leads to 'Airport Damage'. The 'Hazard phase' node leads to 'Post Hazard', 'Place' leads to 'Location', and 'Hazard Type' leads to 'Flood'. At the bottom of the diagram, a sentence within a dashed box reads: 'Chennai Floods: 188 killed, Airport closed they say that the runway came peeling off. it may take time to resume'. Dashed arrows connect some of the lower-level ovals to highlighted phrases in the sentence: 'Location' points to 'Chennai' (highlighted in blue), 'Flood' points to 'Floods' (highlighted in orange), 'Dead and Toll' points to '188 killed' (highlighted in orange), and 'Airport Damage' points to 'Airport closed' (highlighted in green). An arrow from 'Post Hazard' also points towards the beginning of the sentence.
+
+Figure 3: Mapping tweet and its words to empathi concepts.
+
+so that action can be taken. The surveillance concept in the setting of natural disasters can help to identify the resulting health-related needs which in turn, will lead to the more rational and effective deployment of resources to affected populations.
+
+V. CASE STUDY
+
+A substantial impact of empathi ontology is empowering us to annotate text semantically (e.g., tweets posted during hazard). Figure 3 shows mapping of segments within a given tweet to empathi concepts. Thus, from an abstract level, this tweet is related to the flood occurred in Chennai and reports two associated impacts. With this respect, in a first experiment, we compiled 53 million tweets from the 30 significant hazards happened in the past. Such as Hurricane Sandy in 2012 and Irma in 2017, Oklahoma Wildfires in 2017, Chennai Flood in 2005, Alaska Earthquake in 2018, Florida Rains in 2000 and 2016, Houston Floods in 2017, New-Zealand Earthquake in 2016, Typhoon Haima in 2016, Winter Storm Kayla in 2016, and many more. After that, we identified the tweets related to the six central concepts of empathi i.e., (i) impact, (ii) modality of data, (iii) hazard type, (iv) place, (v) transportation and (vi) surveillance. Table III and IV show the statistics of identified tweets related to each of the chosen concepts and sub-concepts. Furthermore, Table II represents samples of these tweets (i.e., column two) along with the mapped *empathi* concepts (i.e., column three) from various hazard type (i.e., column one).
+
+| Sub-Concepts of empathi      | #Tweets |
+|------------------------------|---------|
+| Water Facility (Fac.)        | 218,968 |
+| Gas Facility (Fac.)          | 33,047  |
+| Involved Organization (Inv.) | 4,249   |
+| Severity (Imp.)              | 1,344   |
+
+Table III: Mapping sub-concepts of *empathi* to tweets of hazards. Fac.: Facility, Inv.: Involved, and Imp.: Impact are concepts
+
+Extensive coverage by *empathi* provides the capability of extracting structured information from unstructured and sparse content (e.g., Twitter) [44]. For identifying relevant information from unstructured social media text, it is essential to map
+
+| Concepts of empathi | #Tweets   |
+|---------------------|-----------|
+| Hazard Type         | 3,034,257 |
+| Impact              | 618,446   |
+| Modality of Data    | 509,645   |
+| Facility            | 258,117   |
+| Place               | 16,397    |
+| Transportation      | 6,694     |
+| Surveillance        | 1,588     |
+
+Table IV: Mapping concepts of *empathi* to tweets of hazards.
+
+the words to ontology classes enable efficient classification of tweets as relevant and irrelevant to crisis domain. For instance in figure 2, the tweet: Chennai Floods: 188 killed, Airport closed they say that the runway came peeling off. it may take time to resume is identified as a post-hazard tweet. Chennai links to concept Place, Floods links to concept Hazard type, 188 killed links to concept Affected population, and Airport closed links to concept Infrastructure Damage. Moreover, such a procedure is termed as semantic annotation using expanded concepts (a.k.a. hypernyms). It can improve understanding of social media messages which pose challenges like ill-formed sentences, ambiguous word senses, poor syntactic structure, and implicit referencing. Semantic features formed using empathi can enhance supervised and unsupervised learning in crisis domain [45].
+
+# VI. EVALUATING QUALITY OF empathi
+
+To build up a quality ontology, we followed the principles of ontology methodologies such as NeON [43] and Methontology [33] which encourage the reuse of existing ontologies. However, to quantitatively measure the quality of *empathi*, we designed a user evaluation survey. This survey contained 17 questions concerning with hierarchical, relational and lexical aspects of *empathi* inspired by [46]. Precisely, the participants in the survey have to evaluate the following criteria: (1) the correctness of structure (hierarchy) (2) the correctness of relations between concepts, and (3) lexical evaluation, i.e., quality of annotations associated with both concepts and relations. In the following, we elaborate on these criteria.
+
+- a) Structural evaluation: In this evaluation, the hierarchical structure is assessed concerning the correctness of "is-a" relationship as of whether the given concept A is-a particular type of the given concept B. For instance, "Animal Loss" is-a subclass of "Impact" in *empathi*. Such evaluation is necessary to confirm the utility of ontology for classification task [47]. We presented parts of hierarchy in the survey, and asked the participants, how far this hierarchy makes sense to them. They have to rate in the range 1 (fully disagree) -5 (fully agree).
+- b) Semantic relational evaluation: Ontology is evaluated for holding semantically correct relations between concepts. For instance, there is a relation between the concept "Affected Population" and the concept "Service" referring "need for help". Thus, Affected Population is the domain and Service is the range. Concerning w.r.t. the prior survey [47], having quality relations, higher capability for summarization, subgraph extraction, and contextualization tasks. We represented a number of relations of empathi to the participants
+
+Image /page/5/Figure/0 description: The image displays three bar charts stacked vertically. The top chart, labeled (a), is titled '7 Structure-related Questions (SQ)'. It has a y-axis for 'Agreement Percentage' from 0 to 100. The blue bars show the following percentages for each question: SQ1 is 92.3%, SQ2 is 92.3%, SQ3 is 100%, SQ4 is 92.3%, SQ5 is 53.8%, SQ6 is 84.6%, and SQ7 is 76.4%. The middle chart, labeled (b), is titled '7 Semantic Relations-related Questions (SRQ)'. It also has a y-axis for 'Agreement Percentage'. The purple bars show the following percentages: SRQ1 is 76.9%, SRQ2 is 92.3%, SRQ3 is 66.7%, SRQ4 is 84.6%, SRQ5 is 84.6%, SRQ6 is 84.6%, and SRQ7 is 53.8%. The bottom chart is a partially cropped grouped bar chart with a y-axis for 'Agreement Percentage'. It shows groups of three bars colored blue, orange, and green. The visible data points for the first group are 84.6%, 46.2%, and 76.9%. For the second group, they are 92.3%, 76.9%, and 92.3%. For the third and fourth groups, the values are identical: 84.6%, 69.2%, and 84.6%.
+
+Figure 4: Agreement of 13 evaluators on (a). 7 Questions that evaluate Structure of *empathi*, (b). 7 Questions that evaluate Semantic Relations of *empathi*, and (c). 3 Questions that evaluate Lexical characteristic of *empathi* for 4 concepts: Animal Loss, Health Issues, Food Shortage, Human Prayer. Agreement percentage is calculated as the percentage of evaluators responded "agree" or "yes" in the ontology evaluation form.
+
+LQ1 LQ2 LQ3 Health Issues
+
+LQ1 LQ2 LQ3 Animal Loss
+
+LQ1 LQ2 LQ3 Food Shortage uestions(LQ)
+
+LQ1 LQ2 LQ3 Human Prayer
+
+and asked them whether or not they confirm having such a relation or not.
+
+c) Lexical evaluation: This part examines expressiveness, completeness, and clarity of annotations of a given ontology. Expressiveness states the efficacy of the ontology to identify relevant information using natural language processing techniques. Completeness [48] evaluates whether an illustration of a concept using definition and labels adequately define various scenarios in crisis domain. For instance, w.r.t. the given concept "Human Prayer" the definition "prayer is a message to God from victim's relatives and family for protecting their lives and health" is provided along with the labels "send prayers", "heart prayers", "heart praying", "join pray-
+
+ing", "love prayers", "prayers affected", "temple", "church", "prayers city", "prayers families", "prayers involved", etc. Clarity evaluates whether or not the concept name in the ontology is meaningful and easily understandable to human and machine.
+
+d) **Results**: Our survey had 13 participants and the results have been illustrated in figure 4.
+
+The structural evaluation section comprised of seven questions expressed as follows; (SQ1) Concerning a hazard situation, is "Mental Stress" and "Physical Stress" two important concepts under "Health Issues"? (SQ2) Does "No Effect", "Minor", "Major", "Hazardous", "Catastrophic" represents sub-classes of "severity"? (SQ3) Are "Financial Crisis", "Food Shortage", and "Contamination" probable impacts of a Hazard? (SQ4) Are following triples meaningful: "Animal Loss isa Impact", "Communication Lines Failure is-a Infrastructure Damage", "Power Outage is-a Infrastructure Damage", and "Survived People is-a Affected Population"? (SQ5) Do you consider non-government organization's (NGO) report is an expert report? (SQ6) Do health report, service feedback, and weather report define human sensing? (SO7) Can "News Agencies Report" and "Social Media Report<sup>24</sup>" be categorized under "Media Report", a sub-class of "Report"? Questions SQ1-SQ6 were Yes/No questions, and SQ7 follows Likert Scale. The detailed results are represented in Figure 4a. We observe the average agreement rate above 84.5%.
+
+Regarding evaluation of semantic relations, we designed seven questions as follows: SRQ1: Do you think the following triples make sense; "Event occures in a Place", "Service is offered in a Place", "Each Hazard is associated to a couple (c) of Event" and "each Hazard leads to a couple of Services"? SRQ2: Can concepts, Event, and Service be linked to the concept "Place" using "isLocationAt" relation? SRQ3: Do you think two different types of Hazard can be related concerning Event, Service, and Place? SRQ4: Is the "currentStatus" relation between Facility and Status semantically correct? SRQ5: Is available, offered, requested and unavailable suitable categories for Status? SRQ6: Can the concepts "Service" and "Organization" be concerning the concept "Hazard Type"? SRQ7: Is the relation "needHelp" correctly links "Affected Population" with "Organization" and "Service"? Questions SRQ2-SRQ7 are Yes/No, and SRQ1 follow Likert Scale. From figure 4b, an average agreement of 75.4% was concluded in confirming above facts.
+
+The lexical evaluation of *empathi* was performed by representing definition and synonyms (or labels) describing each concept and asking participants to respond the following questions; (LQ1) Do labels appropriately represent the concept? (LQ2) Are labels complete? (LQ3) Are definitions and labels enough clear? Figure 4c shows the results with the total agreement rate 78.8%.
+
+# VII. RELATED WORK
+
+The aftermath of the disaster causes agencies/organization to structure a plan for the thoughtful recovery of the area. An
+
+<sup>&</sup>lt;sup>24</sup>http://www.aidforum.org/topics/disaster-relief/top-100-twitter-accounts
+
+opportunity to hasten this process is a need for a knowledge structure containing concepts, constraints, and links that provide before-hand information for disaster preparedness and act. In [49], the author designed a process model of the aftermath of the disasters using the Petri-Net containing inter-linked concepts for efficiency after-disaster execution and adaptation. Furthermore, the cost associated with the recovery of an area urges the need to have a structured source of events, concepts, and relations that define a distressing incident. In [50], the author defines the utility of an ontology providing seismic risk definition, prediction, and management to reduce damages. There is a recent study on crowdsourced emergency event detection in [51]. There, the authors propose the utilization of emerging knowledge from text using concepts and temporal information for events. In a recent work on crisis management [52], there has been a development of a suite of tools which can leverage our ontology for context-aware response generation. CrowdGeoKG [53] is a framework using the entities in the OpenStreetMap and is enriched by Wikidata.
+
+Events mainly characterize a disaster scenario, and it is essential to understand what is happening in an emergency situation. In [54], a given user can utilize GPS data to create a simulation model for predicting human mobility. We assume that identification of core disaster domain-specific concepts can help in annotating the GPS data similar to in OpenStreetMap [55]. In 2012, Hurrican Sandy brought in towering traffic on social sensing sites urging the need of an information filtering mechanism for assisting crisis coordination. A psycholinguistic driven domain-dependent lexicon was created in [56] for assistive crisis response. Moreover, tweets on Hurricane Sandy 2012 identified "power blackout" as one of the implications of the disaster. To identify its associated repercussions, one need a structured domain-model. For instance, crashing out of power affected the medical facility. Hence, there is a need for an ontology to bridge the facilities in emergency situation [39]. In [57], the author created a twitter stream of deceptive and peripheral messages using a knowledge source assisting Public Information Officers (PIOs) to make conscious decisions in an emergency situation.
+
+Table 1 motivates creating empathi. Prior ontologies related to crisis management and situational awareness fails to provide structural, lexical and relational benefits needed for extracting situation-specific information extraction from social media content [47]. It is either because these ontologies are diverse in their subject areas or are incomplete concerning referencing, documentation, and expressibility. On analyzing these state-of-the-art vocabularies, we identified subject areas, vocabularies concerning each subject area and used OWL to provide a formal representation. Since ontologies can be domain-specific (crisis domain) or generic (e.g., FOAF), we incorporate concepts relevant in addressing various issues in crisis management. For instance; concept Location was taken from Geonames and iContact, crisis-related concepts from HXL and MOAC. Hence, we addressed all the queries raised in [3]. Human social communication during an emergency event provides real-time insight into various domains such as facility, events, impact, report, surveillance, organization involved and activities carried out during and after the hazard. Extracting actionable information from active social channels is challenges because of 2 reasons: (1) absence of an ontology that map multiple concepts, (2) completeness and expressiveness of the ontology. We provide a utility based case-study (section V) where we used our ontology for mapping 53M tweets to the concepts and sub-concepts in *empathi*. Mapping of social media content to ontology concepts will improve classification and summarization task using state-of-the-art natural language processing and learning techniques.
+
+# VIII. CONCLUSION AND FUTURE WORK
+
+We propose *empathi* ontology as a crisis domain archetype that aids crisis management, hazard situational awareness, and hazard events during emergency scenarios. In our study, we demonstrated the prowess of ontology by its integration with relevant and state-of-the-art crisis vocabularies. Moreover, its efficacy was assessed through appropriate evaluation of its quality across three dominant criteria: structure, lexical and semantic relations using the human judge. Furthermore, we illustrated its effectiveness concerning social media domain by linking tweet content to ontology concepts. The ontology has been created using semantic web language (OWL) and links to FOAF, SIOC, DC-terms, and LODE. We aim to extend the ontology in following directions; the first direction is to introduce Internet of Things (IoT) for disaster management. A second possible direction is to improve the ontology with additional ontology quality tools as recommended by the PerfectO methodology<sup>25</sup>. For instance, use of LODE tool provides automatic documentation. WebVOWL tool provides an automated graph visualization. A third direction is to disseminate more the ontologies on ontology catalogs such as Linked Open Vocabularies (LOV) and LOV4IoT. LOV4IoT could be refined and extended to support the environment domain with various use cases such as flooding, fire, earthquake, tsunami.
+
+# ACKNOWLEDGMENT
+
+We are thankful to responders in providing their unbiased evaluation for ontology quality assessment. We acknowledge support from the National Science Foundation (NSF) award EAR 1520870: Hazards SEES: Social and Physical Sensing Enabled Decision Support for Disaster Management and Response. Any opinions, findings, and conclusions/recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the NSF.
+
+# REFERENCES
+

chatbot/rag/data/('An_Internet_of_Things_ontology_for_earthquake_emergency_evaluation_and_response', '.pdf')_extraction.md

@@ -0,0 +1,128 @@
+
+
+Abstract—Recent years have seen the fast-diffusion of internetconnected devices and the rise of the Internet of Things (IoT) research and application area. Research works are dealing with technologies that enable the so-called things to communicate among them and with users in order to provide data and/or accomplish tasks. This scenario is posing several challenges so that more and more researchers are dealing with them. Our work deals with the definition of both machine and human understandable descriptions of things by means of ontologies in order to enable the collaboration among physical objects and IT systems. Indeed, recent works highlighted how the IoT technology may be profitably used in several scenarios in order to accompish complex tasks where physical objects are active participants. First, we extend the Semantic Sensor Network ontology defined by the W3C Semantic Sensor Networks Incubator Group with concepts and roles that describe actuators. This leads to the definition of a comprehensive Internet of Things ontology. Then, with the support from domain experts, we analyze the earthquake emergency scenario in order to define for it a domain ontology by means of adding domain-related concepts to the IoT ontology. Furthermore, we compare our work with others that use ontologies to formally describes things.
+
+*Keywords*— Internet of Things, Ontology, Earthquake Emergency.
+
+## I. Introduction
+
+Since the term "Internet of Things" (IoT) was introduces by Kevin Ashton in 1999, the number of internet-connected devices had quickly increased. According to CISCO, the number of things connected to the internet exceeded the number of people on earth during 2008; furthermore, the company foresees that the number of internet-connected things will reach 50 billions in 2020. Nowadays, internet-connected devices are involved in many application domains [1]. Now, transportation and logistics, healthcare, smart environment (home, office, plant) can benefit from IoT technologies. Furthermore, the increasing number of new smart things and internet-connected devices developed by companies suggests that, in the near future, futuristic applications (such as robot taxi, city information model, and enhanced game room) will be available.
+
+1"The Internet of Things". http://share.cisco.com/internet-of-things.html. Accessed: 2014-3-4.
+
+Beside an exponential growing of the Internet of Things field, companies started to rethink their core activities including physical objects in their core processes [2], [3]. Indeed, whenever the Internet of Things enables the communication among physical objects and the business IT systems, such objects can be considered as active participants into the business processes of the companies [4]. For example, IoT-enabled objects such as sensors and/or actuators can provide data and/or accomplish operations may be used to execute business process tasks. This brings toward a collaboration among IT systems and physical objects.
+
+Although many efforts have been done in order to connect physical objects to IT systems and enable a collaboration among them, several core challenges arises [5], [6]: scalability, when millions of sensors and/or actuators can be reached to accomplish a given task the amount of information that a system has to manage could be unmanageable; deep heterogeneity, things are very heterogeneous, both regarding physical characteristics and data and/or operations that they provide and/or accomplish, this hinders the use of them in an IoT ecosystem; unknown topology, networks of things may be large and/or things may be unreachable; incomplete or innaccurate metadata, metadata may be used to enrich the description of things and their capabilities but they could be incomplete or inaccurate; and conflict resolution, regarding actuators, conflicts may arise, for instance, when multiple applications attempt to actuate the same device. Furthermore, it is noteworthy the fact that IoT end users are usually non-technical people [7], [3]. They are not experts of the IoT-related technologies, but they know very well the domain of application of the things. An IoT ecosystem should enable these users to use devices even if they do not know IoT implementation details. In order to deal with these challenges, several research works deal with the extension of things descriptions with semantic informations by means of ontologies. Indeed, an ontology is "an explicit specification of a conceptualization" [8] and it is suitable to formally describe a domain of interest, the Internet of Things domain in this case. Ontologies are key elements of the semantic web technologies [9] and were introduced to add semantic machine-understandable information to data on the web, enabling the so-called Web 3.0 [10]. Semantic enriched data can improve search and integration activities; furthermore, they also facilitate the machine capability to reason on data in order to achieve more knowledge of the domain of interest.
+
+Motivations that enabled the use of ontologies on the web lead to utilization of them in the Internet of Things domain on which both systems (mainly, sensors and actuators) and data may be semantically described. In particular, using ontologies things and their capabilities may be modeled at an abstract level which enables non-technical people in using them.
+
+Our work proposes an ontology for Internet of Things in order to deal with the deep heterogeneity challenge of IoT introduced before. Indeed, in order to enable the collaboration among several physical obtects and IT systems, it is needed a common meaning of the involved resources and this can be obtained using semantics [11].
+
+We designed an ontology whose goals are the description of an Internet of Things ecosystem. Starting from the work of the W3C Semantic Sensor Networks Incubator Group [11] on semantic sensor networks, we define an ontology that extends sensors description defined by the Incubator Group with a description suitable for actuators. A third part of the IoT ontology is domain related. In particular, we adopt the earthquake emergency evaluation and response domain as domain of interest of our ontology. The goal is to show how the developed ontology may be used to model the things that are involved in the earthquake emergencies.
+
+The rest of the paper is organized as follows. Section II gives an overview of the works that combine Internet of Things with semantics. Section III describes the Internet of Things ontology we propose, focusing on its three-modules architecture and the modeling of sensors and actuators. Next, Section IV completes the description of the ontology depicting a domain-specific ontology suitable for the earthquake emergency evaluation and response. Section V briefly gives information about the implementation of ontologies. Finally, Section VI closes the paper with a discussion on our approach to the semantic Internet of Things and gives a prevision of the future work.
+
+## II. RELATED WORK
+
+In the recent years, Internet of Things attracted many researchers. Indeed, due to its pervasiveness and the variety of involved technologies, Internet of Things interests several application areas, resulting in a great number of works in literature. Regarding collaboration, for example, researchers proposed standards that enable the syntactic interoperability among things [12], [13], [14]. Despite that, the formal modeling of things by means of semantics is still an emerging area and there not exists a de facto standard to do that. This obstructs the collaboration among things themselves and among things and IT systems [11]. This leads research groups
+
+to propose their own approaches to semantically describes the Internet of Things domain.
+
+De et al. [15] suggest a service modeling of things using ontologies extendig the Entity, Resource, Service, and Device core elements [16] of a thing model. The W3C Semantic Sensor Network Incubator Group [11] defines an ontology for sensor networks. Such ontology goes deep in the description of devices that are capable to sense features of interest. This work represents a comprehensive description of sensors and the approach is domain independent. Hachem et al. [6] aim at solving the IoT challenges by means of three ontologies that describe in detail devices and physics of the domain of interest. Their work proposes a three-layer view of the IoT-based real world. In detail, the authors identify the following layers: a physical layer, an information layer, and a functional layer. For each of them they design a specific ontology in order to give a proper description of the IoT domain. In this work, some aspects of the information and the functional layers are noteworthy. Indeed, in the respective ontologies there are concepts that rely on reasoning about measurement units and physical concept relations, and they are defined from a mathematical viewpoint. This simplify the searching for and composing of data activities. Wang et al. [7] propose an ontology-based framework to support the discovery and invocation of actuators in home care systems. The goal of their work is to simplify the customization of the home care system policies by non-technical people. A care professional may declare system rules by means of human-understandable concepts defined in a care domain ontology. Then, software handlers are responsible for the translation between ontology-based and device-specific descriptions of sensors and actuators used by the home care system. Discovery and invocation of actuators are supported by querying the system knowledge base on which each device is registered. In a further work, Wang et al. [17] define an ontology for the IoT. Such an ontology aims at giving a comprehensive view on the IoT domain, from sensing and actuating capabilities to the quality of services and information. For this purpose, and in order to keep distinct each area of IoT domain, the description ontology is designed with a modular approach. Each module is responsible for one of the IoT-related areas (e.g., resources, services, QoS) and, in some cases, it is developed starting by and extending third party works on IoT. In this work, a particular attention is given to IoT service modeling and testing, but it overlooks how this ontology may be applied in real world scenarios.
+
+Finally, some considerations have to be done about the domain of application we consider and the related work which deals with it. Emergency management involves many organizations and actors, from national agencies to volunteers. Furthermore, recent works also highlights that citizens may contribute in generating useful data in such mass casualty incidents [18]. It emerges a complex scenario where a large amount of data has to be processed and combined in order to
+
+better manage the emergency. For this purpose, several works propose the use of ontologies to add semantic information to data and processes involved in the emergency management activities [19], [20]. Despite that, the majority of them does not consider sensors and actuators in the domain ontologies. Thus, our work appear to consider another aspect of the emergency scenario: the things.
+
+## III. THE IOT ONTOLOGY
+
+The following section describes the architecture of the Internet of Things ontology we propose. Such architecture is composed by three modules, namely Sensors Module, Actuators Module, and Domain Module, divided in two layers (Figure 1). On one hand, the first layer contains Sensors and Actuators modules and formally describes both sensors and actuators in an IoT ecosystem by means of two ontologies, one for sensors and one for actuators. On the other hand the second layer contains the Domain Module. It details the generic descriptions of things given by the first layer adding ontology elements that describe specific-domain devices and features of interest involved in the application scenario, in this case the earthquake emergency evaluation and response scenario. This modular two-layers architecture provides both flexibility and compatibility of the whole ontology: all the specific-domain concepts and roles can be added to the Domain Module without requiring the modification of other modules; furthermore, in order to describe the domain scenario, other ontologies may be used in the domain-related layer without violating the whole IoT ontology design principles.
+
+Image /page/2/Picture/3 description: A block diagram illustrating the components of an "IoT Ontology". The diagram consists of a large, light gray, rounded rectangle labeled "IoT Ontology" at the top. Inside this main rectangle, there are three smaller, light blue, rounded rectangles. The top half of the main rectangle contains two of these smaller rectangles placed side-by-side. The one on the left is labeled "Sensor" and the one on the right is labeled "Actuator". Below these two, a single, wider, light blue rectangle spans the bottom half of the main container and is labeled "Domain".
+
+Figure 1. Internet of Things ontology architecture.
+
+In the following, we illustrate the three modules of the ontology.
+
+### A. Sensor Module
+
+This module is responsible for the description of sensors. Since our work aims at the compatibility with other ontology implementations, we decided to use the ontology designed by the W3C Semantic Sensor Network Incubator Group *et al.* [11] to semantically describe the sensors. This ontology uses the *Stimulus-Sensor-Observation pattern* [21] as starting point to define the description of sensors. According to such pattern: a stimulus is a change or a state in an environment that a sensor can detect and use to measure a property; a sensor is anything that senses and, usually, it consists in an object that observes, transforming incoming stimuli into
+
+another representation; then, an observation is the result of a measurement operation and connects a stimulus that a sensor measures and the sensor that observes such stimulus, it focuses on a single measurement that a sensor takes at a precise instant and in a precise location. By means of this pattern and the definition of the three related core concepts, the W3C Semantic Sensor Network Incubator Group designed the ontology for the semantic sensor networks. The above pattern may be analyzed through four different perspectives which are shown in the following.
+
+Sensor Perspective. It describes the property the sensor measures (ssn:Property) and how (ssn:Sensing and ssn:Stimulus) measurement taken. Furthermore, this perspective deals with sensor capabilities (ssn:MeasurementProperty) in terms of accuracy (ssn:Accuracy), resolution (ssn:Resolution), (ssn:Precision), and other similar sensor characteristics. These characteristics are related to particular conditions (ssn:Condition), consequently several measurement capabilities (ssn:MeasurementCapability) can be associated with each sensor.
+
+System Perspective. It focuses on the platform (ssn:Platform) description. In the semantic sensor network, a system (ssn:System) may be composed by several subsystems (ssn:hasSubSystem). Furthermore, each sensing device must be considered a system as well (ssn:SensingDevice ⊑ ssn:Device ⊑ ssn:System). This allows the description of complex platforms on which several sensors are aggregated in a single device. Finally, each system has its own operating range (ssn:OperatingRange) and survival range (ssn:SurvivalRange).
+
+Feature and Property Perspective. This perspective focuses on the properties (ssn:Property) that may be sensed and the related observations that are related to them. Properties are grouped in features of interest (ssn:FeatureOfInterest) that represents homogeneous collections.
+
+Observation Perspective. Finally, the fourth perspective proposed by W3C SSN Incubator Group deals with observations and related metadata. Observations are taken by means of a sensor (ssn:Sensor) at a certain moment, measuring a stimulus (ssn:Stimulus), using a sensing method (ssn:Sensing), and observing a property (ssn:Property) of a particular feature of interest (ssn:FeatureOfInterest). The result of the observation corresponds to the output of the sensor (ssn:SensorOutput).
+
+### B. Actuator Module
+
+At the best of our knowledge, there not exist generic purpose ontologies for actuators networks; as a consequence, we propose a new ontology of which goal is the modeling of actuators networks at a generic description level. leaving apart the domain-related details. Our ontology is based on the work of the W3C Semantic Sensor Network Incubator Group described above and uses several concepts defined in the SSN ontology. Reasons for this choice rely on the compatibility, the the high level of flexibility, and the level of standardization of that ontology. Taking advantage of the SSN ontology and extending it, we propose the Semantic Actuator Network ontology (SAN ontology) (Figure 2) that formally describes concepts regarding actuators.
+
+In the architecture of our Internet of Things ontology, the Semantic Actuator Network constitutes the actuator module. In conjunction with the SSN ontology it gives a comprehensive formalization for the representation of a thing network. Thus, by means of those two ontology, we can model any complex device composed by sensors and actuators.
+
+As SSN proposes the Sensor-Stimulus-Observation pattern in order to describe how a sensing activity works, we propose the *Actuator-Stimulus-Operation pattern* in order to describe the actuators. According to this pattern, an Actuator is an object capable to modify a property in the environment where it operates. To do that, it produces a Stimulus that is able to modify a property. Finally, an Operation represents the activity that is accomplished by the actuator.
+
+In order to better describe the proposed pattern, similarly to the the description of SSN ontology, we define two perspectives as follows.
+
+Actuator Perspective. This perspective gives a static description of the actuator and highlights the properties that the actuator may manage and/or modify. Similarly to the sensor perspective, this perspective also defines actuator specifications. An actuator (san:Actuator) is able to modify a property (ssn:Property) by means of generating a stimulus (ssn:Stimulus). In this case, a stimulus is the change that an actuator may actuate (san:models) in order to modify (san:modify) a property. It is worthy of note that a stimulus does not necessarily refers to the property that the actuator modifies. For instance, let us consider an electric motor as actuator. The property that we want to manage is its rotational speed. In order to modify it, we vary the rotor current that, in this case, is the stimulus on which the actuator acts. Furthermore, the actuator implements (san:implements) an acting process (san:Acting  $\square$  ssn:Process) that describes the process implemented by the actuator in order to manage the property of interest. As described for the sensors, in order to describe the actuator working specifications we define the change property concept (san:ChangePropery ≡ ssn:MeasurementProperty), related with properties (such as ssn:Accuracy, ssn:Resolution, etcetera) defined in the SSN ontology. Furthermore the concept san:ChangeCapability is defined. It links (ssn:inCondition) a set of working specifications with a specific condition (ssn:Condition).
+
+Operation Perspective. It provides the dynamic description of actuators. Indeed, it focuses on the operation (san:Operation) that a given actuator accomplishes in a given instant (san:operationResultTime). The operation is accomplished (san:operatedBy) by an actuator in order to modify (san:controlledProperty) a property of a specific feature of interest (san:featureOfInterest). Then, this perspective defines both the stimulus that the actuators receives and the method it uses (san:actingMethodUsed). Obviously, the operation is related with the result we want to obtain. Such result is the actuator input (san:ActuatorInput) since it consists on the information on which depends the property modification that is desired. Returning to the above example of the electric motor, the new value of the rotational speed is the input of the actuator.
+
+Both the *System perspective* and the *Feature and Property perspective* can be defined for the SAN ontology, as well. On the other hand, such perspectives have already been defined in the SSN ontology. Therefore, regarding the *System perspetive*, we simply introduced a concept to model an acting device (san:ActingDevice  $\sqsubseteq$  san:Actuator, ssn:Device) and define its relations with ssn:Device. As a consequence, according to our IoT ontology (obtained without modifying the SSN ontology), a system is a platform that may include both sensors and actuators. Regarding the *Feature and Property perspective*, no extensions are required since features of interest and properties are the same described for the SSN ontology. The only difference is that features and properties must be changed by actuators instead of being observed by sensors.
+
+### C. Domain Module
+
+The domain module is responsible for the description of domain-specific things. Since it uses the Sensors and Actuators modules in order to define its own concepts and relations, we choosed to create a second layer for it in the our IoT ontology architecture. By means of it, the resulting architecture is flexible and ontology designers may focus on the specific application domain without taking care of the complex structures of the ontologies for sensors and actuators, but only using their concepts to define the new ones required by the domain of interest. The next section illustrates an example for the earthquake emergency evaluation and response domain.
+
+## IV. THE IOT EARTHQUAKE ONTOLOGY
+
+Several works deal with the use of semantics in the emergency management fields (e.g., see [19], [20]). Nevertheless, they usually do not focus on "things" that are involved in the emergency domain. Here, we propose the *Earthquake Emergency (EEM) ontology* that formally models sensors and actuators, the former used to evaluate the earthquake entity and the latter used to respond to the emergency. In order to design and develop this ontology, we
+
+Image /page/4/Figure/0 description: A concept map or ontology diagram illustrating relationships between various entities, primarily centered around an 'Actuator'. The diagram uses rounded rectangles for concepts and labeled arrows for relationships. Concepts in green include 'ActingDevice', 'Actuator', 'Acting', 'ActuatorInput', 'OperationValue', 'Operation', 'ChangeCapability', and 'ChangeProperty'. Concepts in blue include 'Device', 'Stimulus', 'Process', 'Property', 'FeatureOfInterest', 'Condition', 'Accuracy', 'Resolution', and a box with ellipses. Solid lines indicate 'is a' relationships: 'ActingDevice' is a 'Device', 'Acting' is a 'Process', and 'ChangeProperty' is a 'Property'. 'Accuracy', 'Resolution', and others are types of 'ChangeProperty'. Dashed lines represent other relationships. An 'Actuator' is connected to a 'Device', 'implements some' 'Acting', 'models only' a 'Stimulus', is 'operatedBy only' an 'Operation', and 'hasChangeCapability only' a 'ChangeCapability'. An 'Operation' is linked to 'Actuator' ('operatedBy'), 'Acting' ('actingMethodUsed'), 'Stimulus' ('includesEvent'), 'Property' ('controlledProperty'), 'FeatureOfInterest', and 'ActuatorInput' ('operationDesiredResult'). A 'Property' is modified by 'Acting', is a proxy for 'Stimulus', and has a 'FeatureOfInterest'.
+
+Figure 2. The Semantic Actuator Network ontology (the light blue concepts belong to SSN).
+
+benefited from the help of earthquake domain experts: we consulted civil engineers that work in this kind of emergency and interviewed them to get a proper overview over the sensors and actuators involved in this scenario.
+
+In our domain analysis, we first identify five features of interest and related properties<sup>2</sup> (Figure 3):
+
+- emergency site (eem:SiteFeature) that has ground acceleration and the displacement the properties (eem:SiteAccelerationProperty and as eem:SiteDisplacementProperty);
+- earthquake (eem:EarthquakeFeature) magnitudo, the epicenter, and the distance the has (eem:EarthquakeMagnitudoProperty, eem:EarthquakeDistanceProperty, eem:EarthquakeEpicenterProperty);
+- the evaluation of cracks (eem:CrackFeature) that has their widths as property (eem:CrackWidthProperty);
+- the localization (eem:LocalizationFeature) that latitude-longitude pair as property eem:GPSLocalizationProperty);
+- the alarm (eem:AlarmFeature) that aims at notifying people of the earthquake entity and of additional information in order to reach a safe place (denoted by properties such as eem:AlarmNotificationProperty, eem:AlarmLightProperty, and eem:AlarmSoundProperty).
+
+Image /page/4/Figure/10 description: A line of black text on a white background, which appears to be a citation. The text reads: "²See http://earthquake.usgs.gov/learn/glossary/. Accessed: 2014-3-3." The number '2' at the beginning is a superscript.
+
+Image /page/4/Figure/11 description: A diagram illustrating a conceptual model or ontology, organized into three main columns connected by arrows indicating relationships. The first column, on the left, is titled 'FeatureOf Interest' in a light green box. Below it, a vertical line with 'is a' connects to a list of light blue boxes: 'SiteFeature', 'Earthquake Feature', 'CrackFeature', 'Localization Feature', and 'AlarmFeature'. The middle column has a light green box labeled 'Property'. Below it, a list of light blue boxes are all connected to 'Property' with arrows labeled 'is a'. These boxes are: 'Site Acceleration Property', 'Site Displacement Property', 'Earthquake Distance Property', 'Epicenter Property', 'Magnitudo Property', 'Crack Width Property', 'GPS Localization Property', 'AlarmNotification Property', 'AlarmLight Property', and 'AlarmSound Property'. The third column, on the right, is titled 'Quantity Kind' in a yellow box. Below it, a vertical line with 'is a' connects to a list of yellow boxes: 'Acceleration', 'Distance', 'Geographic Coordinates', 'Magnitudo', and 'Dimensionless'. Dashed arrows labeled 'hasProperty' connect features from the first column to their corresponding properties in the second column. For example, 'SiteFeature' has properties 'Site Acceleration Property' and 'Site Displacement Property'. Solid arrows connect properties from the second column to their quantity kinds in the third column. For example, 'Site Acceleration Property' points to 'Acceleration', and 'Epicenter Property' points to 'Geographic Coordinates'.
+
+The Earthquake Emergency Management ontology: Features of Interest and Properties (green concepts belong to SSN, yellow concepts belong to QU).
+
+We used the *Quantities Kinds and Units ontology* (QU)<sup>3</sup>
+
+<sup>&</sup>lt;sup>3</sup>"Library for Quantity Kinds and Units: schema, based on QUDV model OMG SysML(TM), Version 1.2", http://purl.org/NET/ssnx/qu/qu\#. Accessed: 2013-11-23.
+
+to represent units of measurement of such properties. This ontology defines two main concepts: qu:QuantityKind denotes the kind of the measurement, and qu:Unit denotes the unit of measurement. The two concepts are linked through the property qu:unitKind. In EEM, the properties we defined extend (is-a) both ssn:Property and qu:QuantityKind. In this manner, we can specify the unit of measurement for each property. Furthermore, as some units of measurement we need are not defined in QU, we directly added them in EEM. For instance, we introduced the magnitudo (eem:Magnitudo) and the geographic coordinates (eem:GeographicCoordinates).
+
+After property identification, we analyzed the devices involved in the earthquake emergency scenario (see Figure 4).
+
+We identify the following sensing devices: the fissurometer (eem:Fissurometer) that measures the width of cracks; the seismometer (eem:Seismometer), the accelerometer (eem:Accelerometer), and the GPS (eem:GPS), they all are needed to evaluate an earthquake; the seismograph station (eem:SeismographStation), that is composed by the previous three sensors; and the seismograph network (eem:SeismographNetwork), that is composed by three seismograph stations at least and is used to evaluate the earthquake epicenter. Then, we identify the following actuators: the alarm siren (eem:AlarmSiren) that informs people to evacuate a building; the signaling escape light (eem:SignalingEscapeLight) that highlights the escape path during a building evacuation; and, finally, the alarm message notifier (eem:AlarmMessageNotifier) that sends information about safe places after an earthquake to citizens' mobile devices.
+
+## V. IMPLEMENTATION
+
+As mentioned above, our IoT ontology is composed by three main ontologies: the Semantic Sensor Network (SSN) and the Semantic Actuator Network (SAN) ontologies describe the infrastructure of a network of things, whereas the Earthquake Emergency Management (EEM) ontology describes the domain of interest we evaluate in this work. We implemented the last two ontologies using the OWL 2 language [22]. It results in two files, san.owl and eem.owl, that define respectively the SAN ontology and the EEM ontology. These files are available at the following URL: https://bitbucket.org/gilbertotcc/iot-ontology.
+
+## VI. DISCUSSION AND CONCLUSION
+
+This work proposes an ontology for Internet of Things extending the ontology for semantic sensor networks developed by the W3C Semantic Sensor Network Incubator Group and applies it to the earthquake emergency domain.
+
+Although several works deal with semantic description of things, we focus on flexibility and compatibility proposing
+
+a three modules ontology in two layers: one dealing with a generic model for the things and one dealing with specific-domain things. Indeed, De et al. [15] gave a detailed description of things but their approach is oriented to the users of those devices and do not focus on modeling such sensors and actuators as, for example, the Semantic Sensor Network ontology [11] does about sensors. Instead, our work extends the SSN ontology and proposes a detailed description model for the actuators too. The work by Hachem et al. [6] shows an interesting analysis of the IoT domain and proposes an ontology composed by three modules. Their work focuses on the combination of things, not only on their description, up to include the mathematical laws involved in the domain. This goes beyond the goals of our work which deals only with devices and physics of the domain of interest. In our opinions, our approach results more easy to learn for end users of the ontology. In comparison with the work of Wang et al. [7], our approach is more flexible. Due to its modularity, the ontology we propose is not domain specific but is suitable for each scenario on which things are involved. The work proposed by Wang et al. [17] shows a comprehensive description of the IoT domain which includes also services and testing areas and it goes beyond the goal of our work, as well. Furthermore, their work does not show any example and/or case study in order to evaluate differences with our approach.
+
+In conclusion, our work represents a starting point to the development of an Internet of Things ontology also suitable for users, adding a semantic description for thing interfaces. By adding a semantic description to the physical objects, such ontology eases the collaboration among several physical objects such as sensors and actuators that use different communication protocols and data formats. At the moment, it is a comprehensive ontology to deeply describe both sensors and actuators. Furthermore, its architecture enables domain-specialists to build every domain ontology they need using the top layer of our IoT ontology offering a high level of flexibility.
+

chatbot/rag/data/('AntarcticGlaciers_Glacial_Lake_Outburst_Floods__GLOFs_', '.pdf')_extraction.md

@@ -0,0 +1,153 @@
+
+# **Glacial Lake Outburst Floods (GLOFs)**
+
+A **glacial lake outburst flood** (GLOF) is a release of meltwater from a moraine- or ice-dam glacial lake due to dam failure<sup>1,2</sup>. GLOFs often result in catastrophic flooding downstream, with major geomorphic and socioeconomic impacts<sup>3,4</sup>.
+
+GLOFs have three main features:
+
+- They involve **sudden** (and sometimes cyclic) releases of water.
+- They tend to be **rapid events**, lasting hours to days.
+- They result in **large downstream river discharges** (which often increase by an order of magnitude).
+
+Image /page/0/Picture/11 description: An aerial photograph from 2002 captures a dramatic natural event at the terminus of the Hubbard Glacier in Alaska. The image shows a massive, textured glacier with white and light blue ice on the left, meeting a steep, dark, rocky mountainside on the right. A channel of turbulent, gray water with white foam flows between the glacier and the mountain, indicating a glacial lake drainage. The caption at the bottom reads, "Glacial lake drainage at the terminus of the Hubbard Glacier, Alaska, 2002. Photo: USGS."
+
+Glacial lake drainage at the margin of the Hubbard Glacier, Alaska, 2002. Photo: USGS, Wikimedia Commons.
+
+# Why are GLOFs important?
+
+Some of the largest floods in Earth's history have been GLOFs<sup>5,6</sup>. They have caused large-scale landscape change<sup>5</sup>, and even altered regional climate by releasing huge quantities of freshwater to the oceans<sup>7,8</sup>.
+
+Today, GLOFs pose a risk downstream communities and infrastructure. In Peru alone, GLOFs were responsible for  $\sim$ 32,000 deaths in the 20<sup>th</sup> century<sup>3,4</sup>. They have killed hundreds to thousands of people in other mountain regions (e.g. the Himalayas), and destroyed roads, bridges, and hydroelectric developments<sup>3</sup>.
+
+# An increasing hazard
+
+Importantly, the general global trend of glacier shrinkage through the 20<sup>th</sup> and 21<sup>st</sup> centuries has seen the number and size of glacial lakes increase<sup>9-11</sup>, at the same time as human activities have expanded further into glaciated catchments. The study of how GLOFs occur and their impacts is therefore important for future hazard mitigation.
+
+Image /page/1/Picture/6 description: An aerial photograph of a vast, mountainous landscape dominated by glaciers and snowfields. The mountains are rugged and brown, contrasting with the white snow and the light blue and white glaciers that flow down through the valleys. Several bodies of water, identified as "glacial lakes" by white text and lines, are visible. These lakes, varying in color from deep blue to turquoise, are situated at the terminus of the glaciers. The labels appear in three locations: the upper left, the upper center, and the lower right of the image.
+
+Recent rapid growth of glacial lakes in the Bhutan-Himalaya in response to retreating glacier termini. Photo: NASA/USGS, <u>Wikimedia Commons</u>.
+
+# **Glacial lake settings**
+
+There are two main settings in which glacial lakes form<sup>1,2</sup>: (1) behind **moraine dams**, and (2) behind **ice dams**.
+
+## **Moraine-dammed lakes**
+
+Moraine-dammed lakes form during periods of glacier retreat from a moraine<sup>1,12</sup>. As a glacier margin retreats, water collects in the topographic low between the ice-front and the abandoned frontal and/or lateral moraine. Most existing moraine-dammed lakes (such as the Imja Tsho glacial lake in Nepal; see below) formed when mountain glaciers began to retreat from large moraine ridges constructed during the Little Ice Age<sup>12</sup>.
+
+Image /page/2/Picture/3 description: A photograph of a high-altitude mountain landscape illustrating various glacial features with labels. In the background, massive, jagged, snow-covered peaks rise against a clear blue sky. In the center of the image is a large, white, ice-covered body of water labeled 'lake'. To the upper left of the lake, a 'glacier' is shown descending from the mountains. To the right of the lake, a ridge of rock and debris is identified as a 'lateral moraine'. In the foreground, a vast, rugged area of rock, sediment, and small meltwater ponds is labeled as the 'terminal moraine complex'.
+
+Imja Tsho (or Imja Lake) in eastern Nepal, dammed by a terminal moraine complex. The lake has been growing rapidly since the 1960s as the Imja Glacier has retreated. Photo: Sharad Joshi, Wikimedia Commons, Edited by J.Bendle.
+
+## **Ice-dammed lakes**
+
+In contrast to moraine-dammed lakes, ice-dammed lakes form when drainage is blocked by a glacier that advances or becomes thicker<sup>13,14</sup>. Consequently, ice-dammed lake growth is closely related to glacier mass balance and climate. Ice-dammed lakes form wherever a glacier blocks the drainage of meltwater.
+
+The main settings for ice-dammed lakes are shown in the diagram below. These include: where a tributary valley is blocked by a trunk glacier; where a glacier from a tributary valley advances across the main trunk valley; in openings between the lateral glacier margin and ice-free valley sides; and at the point where two glaciers join.
+
+Image /page/3/Picture/3 description: A diagram illustrating various features in a glacial environment. The image shows large, light-blue glaciers surrounded by gray land. Dotted arrows indicate the direction of 'ice flow'. Several features are labeled, including a 'tributary glacier' merging with a larger glacier at an 'ice confluence', and a 'retreating tributary glacier'. Different types of lakes, shown in a darker blue, are also labeled: a 'marginal lake' at the edge of the ice, a 'supraglacial lake' on top of the ice, and a 'proglacial lake' in front of the glacier. A 'river' is shown flowing between two large glacier masses. 'Streams' and 'drainage' are depicted as water flowing away from the glaciers onto the land. The bottom right corner includes the credit '© J.Bendle'.
+
+Settings of ice-dammed lake growth. Potential routed of lake drainage, beneath or around glacier margins, are shown with dashed blue lines. © J.Bendle.
+
+# The failure of ice or moraine dams
+
+The failure of glacier and moraine dams depends on two main factors: (1) the **integrity of the dam**, and (2) the **nature of trigger mechanisms**<sup>1,2,12,13</sup>.
+
+## **Moraine-dam failures**
+
+Moraine dams tend to be narrow and sharp-crested<sup>12</sup>. As such, they are more likely to fail than broader dam types, such as ice-contact fans or landslides (see diagram below). Most moraines are made up of loose, poorly sorted, permeable sediment, and some contain ice cores<sup>12</sup>. Unconsolidated sediments are susceptible to failure, especially if saturated, while the melting of ice cores may cause moraines to subside over time. Despite these weaknesses, where a moraine is low, wide, and armoured by large rock material it may survive intact for hundreds or even thousands of years<sup>12</sup>.
+
+Image /page/4/Figure/3 description: A diagram comparing the cross-sectional profiles of various geological landforms and man-made structures. At the top, three brown shapes are labeled "moraines". Below that is a long, tapering, light-yellow shape labeled "outwash fan". Next are two long, low-profile brown landforms labeled "Daguangbao rock avalanche" and "Yaoyingyan landslide", with the former being longer than the latter. At the bottom left, two gray cross-sections of dams are labeled "Man-made gravity dams". A scale is provided at the bottom right, indicating a horizontal distance of 1000 m and a vertical distance of 200 m, illustrated with a generic shape labeled "width" and "height". The image is credited to "© J.Bendle" in the bottom left corner.
+
+Approximate width and height of typical moraine, outwash, landslide (see ref: 15), and man-made gravity dams. Compared to other common dam types, moraines are narrow, and contain a relatively small volume of material. © J.Bendle
+
+### **Displacement waves**
+
+Outburst floods in moraine-dammed settings are often caused by the sudden input of material into a lake causing displacement of water and overtopping of the dam<sup>12,16</sup>. Displacement (or *seiche*) waves are commonly triggered by avalanches or rockfalls, or calving of a lake-terminating glacier as shown below.
+
+Image /page/4/Picture/7 description: A diagram illustrating how a seiche wave is formed in a moraine-dammed lake. On the left, a large light-blue glacier is shown on top of a brown landmass. The edge of the glacier is breaking off into a body of water in a process labeled "calving". This creates a large wave, labeled "seiche wave", in the "moraine-dammed lake". The wave travels to the right and crashes over a mound of earth labeled "moraine dam" in a process called "overtopping".
+
+Calving of ice blocks at the glacier terminus is a common cause for lake water displacement, seiche wave development, and overtopping of moraine dams. © J.Bendle.
+
+Other triggers include, the rapid input of meltwater from an glacier upstream, heavy rainfall or snowmelt events that rapidly raise the lake level, or earthquakes that destabilise the moraine dam<sup>12,14</sup>.
+
+### **Breach incision**
+
+Once overtopped, a process called **breach incision** can occur. This occurs as water flowing across the dam surface erodes a channel into the moraine, starting a positive feedback process where channel incision allows more water to escape, and the higher discharges leaving the outlet encourage greater rates of erosion (see diagram below). This process allows a moraine dammed lake to drain
+
+very rapidly<sup>12,16</sup>. It also adds large volumes of unconsolidated sediment to the flood waters, which may then develop into highly destructive debris flows<sup>12</sup>.
+
+Image /page/5/Picture/4 description: A high-angle photograph of a mountainous, glacial landscape with several features labeled. In the center is a milky, light-green body of water labeled "lake." This lake is dammed by a large ridge of rock and sediment, which is labeled "moraine." A V-shaped gap in this ridge is identified as a "moraine breach." Higher up the slope from the lake, a patch of snow and ice is labeled "glacier." A narrow, winding trail crosses the steep, rocky terrain, and two hikers with large backpacks are visible on the path.
+
+### Breach incision process
+
+Image /page/5/Picture/6 description: A diagram illustrating the process of moraine incision over three time steps, labeled T1, T2, and T3. In T1, water is shown "overtopping" a moraine. In T2, this has led to a small "moraine incision," with water flowing through it. In T3, the incision has become a much larger, V-shaped channel. Arrows indicate that this process leads to an "increasing discharge" of water as the channel deepens.
+
+Example of moraine breach at Schoolroom Glacier lake, Grand Teton National Park, USA, with schematic to show breach incision process © J.Bendle. Photo: National Park Service, Wikimedia Commons.
+
+### Moraine degradation
+
+Piping (the seepage of water and sediment through a moraine weakening its internal structure) or the melting of ice cores can trigger moraine collapse, or lower the moraine to the point where a smaller displacement wave will cause overtopping and breach incision<sup>12</sup>.
+
+## Ice-dam failures
+
+Unlike moraine dam failures, the drainage of an ice-dammed lake does not necessarily end in dam destruction<sup>1,13,17</sup>. This is because water can drain through subglacial channels that close up when
+
+water discharge drops, or because an ice-margin may temporarily float before regaining contact with the bed<sup>18,19</sup>. The time-lapse video below (provided by <u>Joe Mallalieu</u>) gives an example of glacial lake drainage at Russell Glacier, Greenland, due to temporary glacier flotation<sup>19</sup>.
+
+http://www.antarcticglaciers.org/wp-content/uploads/2018/10/RG-GLOF-2015 JoeMallalieu.mp4
+
+### Ice dam flotation
+
+Ice-dam floatation occurs due to the difference in density between ice ( $\sim$ 0.9 g/cm³) and water ( $\sim$ 1.0 g/cm³)<sup>13,18,19</sup>. As shown in the diagram below, when lake level reaches 90% of the glacier dam height the ice will lift from the bed, and water can escape through subglacial conduits. However, as a lake empties, and its level falls below 90% of dam height, the glacier drops to the bed and subglacial conduits close up. This stops lake drainage and allows a lake to refill.
+
+Image /page/6/Picture/7 description: A scientific diagram, labeled T₁ in the upper right corner, illustrates a cross-section of a glacier or ice sheet interacting with a body of water. The diagram shows a large mass of light blue ice on the right and a body of darker blue water on the left, both resting on a sloping dark gray ground surface. A dashed black line labeled "flotation level" is shown above the water level. A section where the ice is grounded is marked with a bracket and the text "no flotation". An arrow points from this label to an inset box below, which provides a magnified view of the ice-ground interface. This inset is labeled "sealed conduits" and shows dashed, arch-shaped channels at the base of the ice.
+
+hydrostatic pressure < ice overburden pressure
+
+Image /page/6/Picture/9 description: A scientific diagram illustrating the process of lake drainage under a glacier, labeled "T₂" in the upper right corner. The main diagram shows a cross-section of a body of water on the left, dammed by a large light-blue ice mass on the right, both resting on a downward-sloping grey ground. A section where the ice begins to lift off the ground is labeled "ice-dam flotation." An arrow points from this area to a smaller, inset diagram below. The main diagram also shows water flowing from the lake under the glacier, indicated by an arrow and the label "lake drainage." The channel under the glacier is labeled "subglacial conduit." The inset diagram, connected by an arrow from "ice-dam flotation," is a magnified view of the base of the ice, showing several rounded channels labeled "open conduits."
+
+hydrostatic pressure > ice overburden pressure
+
+© J.Bendle
+
+Ice-dam flotation mechanism of glacial lake drainage. The ice dam lifts from the bed when hydrostatic (water) pressure is greater than ice overburden pressure. Lake water can then escape through subglacial conduits. © J.Bendle.
+
+### **Overspilling**
+
+Lake drainage can also occur by simple overspilling<sup>13,17</sup>, for example, during a large rainfall or snowmelt event that causes lake level to rise. This is most common at cold-based glaciers, which are frozen to the bed and less permeable than warm-based glaciers<sup>2</sup>.
+
+# Differences between moraine- and ice-dammed GLOFs
+
+For a given potential energy (think of this as lake volume) moraine dam outbursts typically produce higher magnitude floods (see diagram below)<sup>12</sup>. However, the threat of repeated GLOFs from a single lake is low because moraine dams are often destroyed during an outburst.
+
+Image /page/7/Figure/5 description: A log-log plot titled 'Flood size vs. potential energy of dammed lake water', which is noted as being 'Redrawn and modified from Costa & Schuster (1988)'. The x-axis is 'Potential energy (joules) of dammed lake water', with a scale from 10^7 to 10^16. The y-axis is 'Maximum flood discharge (m/sec)', with a scale from 10^1 to 10^5. The graph shows two positively correlated trend lines. The upper, green dashed line is labeled 'Moraine-dam outbursts (higher discharges for same potential energy)'. The lower, blue dashed line is labeled 'Ice-dam outbursts (lower discharges for same potential energy)'. A vertical dashed line at a potential energy of 10^13 joules highlights two points, one on each trend line, which are collectively labeled 'GLOF'. At this potential energy, the moraine-dam outburst has a discharge of approximately 8x10^3 m/sec, while the ice-dam outburst has a discharge of approximately 2x10^2 m/sec.
+
+Flood size vs. potential energy of lake water (plotted on logarithmic scale). Moraine-dam outbursts produce, on average, GLOFs with higher flood discharges, than ice-dam outbursts. For example, a moraine- and ice-dammed lake with the same potential energy of  $10^13$  joules (black dots), will lead to a peak flood discharge of  $10^4$  and  $10^2$  m/sec respectively. Diagram created based on data originally presented in refs 12 and 13. © J.Bendle.
+
+Ice-dammed lake drainage, by contrast, does not often result in dam destruction, meaning that ice-dammed lakes can drain and refill many times. A good example is Lago Catchet Dos in Patagonia (see image below), dammed by the Colonia glacier<sup>20</sup>. This lake has drained in subglacial tunnels beneath the ice at least 21 times in the last 10 years<sup>20</sup>.
+
+So, while in the modern-day ice-dammed lake outbursts typically produce lower discharge, and less destructive, floods<sup>12</sup>, they may have long-term impacts on downstream communities and infrastructure<sup>20</sup>.
+
+Image /page/8/Picture/3 description: A satellite image showing the Colonia Glacier and Lake Cachet Dos. The main image is in color and shows the large, light-blue glacier flowing through a valley. An area to the right of the glacier is labeled "lake Cachet Dos". In the lower left corner, there is a smaller, inset grayscale image showing a close-up of the glacier, which is labeled "pre-GLOF".
+
+Image /page/8/Picture/4 description: An annotated aerial photograph illustrating the aftermath of a glacial lake outburst flood (post-GLOF). The main image shows a large valley glacier flowing between dark, rocky mountains. Several features are labeled to explain the event. A label points to a small, dark body of water next to the glacier, identified as a "lake drained." A blue dashed line with an arrow, labeled "subglacial tunnel," traces the path of the water from the drained lake down the length of the glacier. In the top left, an inset image provides a magnified view of the glacier's terminus, showing numerous "stranded icebergs" in the water.
+
+Lago Cachet Dos in Patagonia, southern South America, has repeatedly drained and refilled in recent times (ref: 20). During a GLOF event, lake water drains via subglacial conduits at the bed of the Colonia Glacier. Note the large icebergs left stranded on the former lake bed after drainage. Imagery: GoogleEarth, Edited by J.Bendle.
+
+# Will GLOF hazards increase or decrease in the future?
+
+An important and interesting question! The threat from moraine-dammed GLOFs is typically greatest during periods of glacier retreat, whereas ice-dammed GLOFs are highest during periods of glacier growth.
+
+Therefore, we might expect the number of moraine-dammed GLOFs to increase as mountain glaciers continue to shrink worldwide. However, because moraine dams are normally destroyed in lake outbursts, the number of GLOFs will likely start to decrease over time, as the capacity for storing glacial meltwater is gradually lost.
+
+# **Further reading**
+
+- Explore this ArcGIS Story about Glacier Lake hazards in Alaska
+- Jokulhlaups
+- Glaciolacustrine landsystem
+- Valley glacier landsystem
+- Glacier lakes
+
+# **References**
+

chatbot/rag/data/('Characteristics and changes of glacial lakes and outburst floods', '.pdf')_extraction.md

@@ -0,0 +1,305 @@
+Image /page/0/Figure/3 description: A user interface element showing an icon next to text that reads 'Check for updates'. The icon is a square containing a circle with a color gradient from yellow to blue to red, with a red shape in the center.
+
+# Characteristics and changes of glacial lakes and outburst floods
+
+Guoqing Zhang <sup>1,14</sup> □, Jonathan L. Carrivick <sup>2,14</sup> □, Adam Emmer<sup>3,14</sup>, Dan H. Shugar <sup>4,14</sup>, Georg Veh<sup>5,14</sup>, Xue Wang<sup>1,6</sup>, Celeste Labedz<sup>4</sup>, Martin Mergili<sup>3</sup>, Nico Mölg<sup>7</sup>, Matthias Huss <sup>8,9,10</sup>, Simon Allen<sup>11,12</sup>, Shin Sugiyama <sup>13</sup> & Natalie Lützow <sup>5</sup>
+
+## Abstract
+
+Global glacier mass loss has accelerated, producing more and larger glacial lakes. Many of these glacial lakes are a source of glacial lake outburst floods (GLOFs), which pose threats to people and infrastructure. In this Review, we synthesize global changes in glacial lakes and GLOFs. More than 110,000 glacial lakes currently exist, covering a total area of ~15,000 km<sup>2</sup>, having increased in area by ~22% dec<sup>-1</sup> from 1990 to 2020. More than 10 million people are exposed to the impacts of GLOFs, commonly associated with dam failure or wave overtopping associated with mass movements. Although data limitations are substantial, more than 3,000 GLOFs have been recorded from 850 to 2022, particularly in Alaska (24%), High Mountain Asia (HMA; 18%) and Iceland (19%), the majority (64.8%) being from ice-dammed lakes. Recorded GLOFs have increased in most glaciated mountain regions of the world, with ongoing deglaciation and lake expansion expected to increase GLOF frequency further. In HMA, GLOF hazards are projected to triple by 2100, but changes in other regions will likely be lower given topographic constraints on lake evolution. Future research should prioritize acquiring field data on lake and dam properties, producing globally coordinated multitemporal lake mapping, and robust and efficient modelling of GLOFs for comprehensive hazard assessment and response planning.
+
+## Sections
+
+Introduction
+
+Global and regional changes of lake size and abundance
+
+Historical glacial lake outburst floods
+
+Future glacial lakes and outburst floods
+
+Summary and future perspectives
+
+A full list of affiliations appears at the end of the paper. ⊠e-mail: guoqing.zhang@itpcas.ac.cn; J.L.Carrivick@leeds.ac.uk
+
+## Introduction
+
+Glaciers have been losing mass since the end of the Little Ice Age<sup>1-3</sup>, and at high – and accelerating – rates since the 1960s (refs. 4,5). Indeed, overall mass loss from glaciers and ice caps (excluding the Greenland and Antarctic ice sheets) totalled  $267 \pm 16$  Gt year<sup>-1</sup> from 2000 to 2019 (ref. 4), much of which can be attributed to anthropogenic warming<sup>6-8</sup>. When the resultant meltwater is naturally dammed (by ice, a moraine or bedrock), glacial lakes can form on (supraglacial), within (englacial), under (subglacial), at the edge of (ice-marginal) or beyond the front of (proglacial) glaciers. There are an estimated 14,394 of these lakes (>0.05 km<sup>2</sup>) globally<sup>9</sup>, the majority of which are in Greenland, the Southern Andes, Antarctica and the Subantarctic, Alaska, Canada and High Mountain Asia (HMA) (Fig. 1a). The number and size of glacial lakes is generally increasing, especially for proglacial lakes, owing to close connections with contemporary glacier retreat and meltwater production at local<sup>10,11</sup>, regional<sup>12-15</sup> and global<sup>9</sup> scales. However, there is considerable regional variability<sup>9,16–19</sup>, including notable expansion in Alaska, Greenland periphery, Patagonia and HMA. Nevertheless, the overall growth pattern is expected to continue<sup>20-24</sup>, in line with a projected 26-41% decrease in glacier mass by 2100 under temperature increases of 1.5-4 °C25.
+
+The stability and longevity of glacial lakes are highly variable. Glacial lakes can drain suddenly and very quickly (hours to days), releasing a glacial lake outburst flood (GLOF). More than 3,000 GLOFs have been reported worldwide from 850 to 2022 (ref. 26) (Fig. 1b), each with characteristic rapid rise to peak discharge and often very large flood volumes. For example, the 1996 volcanic eruption beneath the Vatnajökull ice cap, Iceland, produced an outburst flood with an estimated peak discharge of ~55,000 m³ s⁻¹, thereby temporarily becoming the second largest river in the world after the Amazon²7,28. Indeed, ice-dammed lakes have produced some of the largest peak discharges recorded in human history²9-3¹, including peak discharges of >10⁵ m³ s⁻¹ in HMA³² and 10° m³ s⁻¹ in the Andes and Alaska³0,33,3⁴.
+
+Accordingly, GLOFs rank among the most prominent glacier hazards, killing thousands of people over the past century<sup>35</sup> and causing extensive damage tens to hundreds of kilometres downstream. For instance, the 1941 GLOF in Huaraz, Peru, destroyed a third of the city and killed about 5,000 people<sup>36</sup>, and the 2013 GLOF in Kedarnath, India, killed more than 6,000 people<sup>37</sup>. Given that millions of people are potentially at risk from GLOFs worldwide<sup>38</sup>, especially in a warming climate, understanding the location, timing of formation, evolution and physical characteristics of glacial lakes is critical to mitigating the downstream impacts to communities and infrastructure.
+
+In this Review, we synthesize glacial lake and GLOF hazard data sets from most glaciated regions of the world to assess their ongoing and projected changes. We begin by documenting the global and regional distribution of glacial lakes. We follow by assessing the characteristics and historical trends of GLOFs. We subsequently discuss future projections in glacial lakes and the corresponding GLOF hazard and risk, before ending with research priorities. Throughout, focus is on the physical aspects of glacial lakes and GLOFs in a warming world.
+
+## Global and regional changes of lake size and abundance
+
+Understanding glacial lakes is the first step in estimating the GLOF hazard and risk (Fig. 2). Methods used to map glacial lakes are now discussed, followed by quantification of their historical changes at global and regional scales.
+
+### Mapping glacial lakes
+
+A growing archive of satellite products and enhanced computing power<sup>39</sup> have improved mapping of glacial lakes (Fig. 2). In most cases, mapping relies on the use of the normalized difference water index (NDWI)<sup>40</sup> – a satellite image-derived calculation combining green and near-infrared wavelength bands to differentiate water bodies from all other landcover. Automated and semi-automated methods are used for lake mapping from an NDWI classification, although lake outlines have also been delineated manually with enhanced accuracy<sup>12,41</sup>. Landsat satellite imagery is often used as the basis for glacial lake mapping by the NDWI owing to its temporal coverage and a medium-resolution sensor 9,17,42. Nevertheless, Landsat imagery has limitations for NDWI-based lake mapping, including poor image contrast due to seasonal snow and  $ice\,cover\,on\,lakes, as\,well\,as\,confusion\,introduced\,by\,mountain\,shadows$ and cloud cover<sup>43</sup>. Cloud-free synthetic aperture radar (SAR) can overcome this limitation, but is of limited length and hence, generally, only suitable for validation of mapping algorithms or case studies  $^{44,45}\!$  . Cloud computing, artificial intelligence and machine learning are rapidly improving the ability to efficiently manage and automatically analyse remote sensing data<sup>46,47</sup>. However, interactive visual inspection and manual editing combined with original imagery are still essential to ensure a consistent data quality with time.
+
+In addition to mapping lake outlines, knowledge of glacial lake volume is important for quantifying hazards and for estimating downstream impacts of GLOFs<sup>48,49</sup> (Fig. 2). Empirical volume–area relationships for glacial lakes have been established in many instances (see Supplementary Table 1). Global compilations of lake depths, areas and volumes suggest that lake area and depths are moderately correlated, and lake area and volumes are well correlated<sup>50</sup>. These correlations allow for the estimation of lake depths and volumes from area, especially in regions where direct measurements are rare and difficult (see Supplementary Table 1). Bathymetric measurements of ice-marginal lakes have also been carried out in some locations, including the Southern Alps of New Zealand<sup>51,52</sup> and on the Tibetan Plateau<sup>53,54</sup>.
+
+### Global glacial lake characteristics
+
+After mapping the outline of glacial lakes, it is possible to determine their overarching characteristics, including size, area and volume. Various such inventories exist (see Supplementary Table 2), each compiled from satellite imagery but differing in maximum distance between glaciers and lakes (buffers of 1, 3 or 5 km) and the smallest lake size included  $(0.0002, 0.005 \, \text{or} \, 0.05 \, \text{km}^2)$ , introducing slight discrepancies at the regional and global scales. Here, several regional inventories  $^{9,12,18,19}$  are merged to produce a global contemporary glacial lake compilation; a size threshold of  $0.05 \, \text{km}^2$  is set for all Randolph Glacier Inventory (RGI) regions  $^{55}$  except Svalbard and Jan Mayen, Southern Andes and New Zealand (see Supplementary Table 2).
+
+This global compilation provides much insight into glacial lake characteristics. A total of 22,133 glacial lakes were revealed in 2015–2019, collectively covering -14,438 km² (see Supplementary Table 2). These totals are more than other reports of 14,394 lakes (>0.05 km²) situated within 1 km of glacier margins, covering a total area of -8,950 km² in 2015–2018 (ref. 9). Regional differences in the number, area and volume of global glacial lakes are evident (Fig. 1a). Glacial lakes are predominantly distributed in Greenland (31% of the total area), the Southern Andes (16%), Alaska and western Canada (16%), the eastern Canadian Arctic (15%) and HMA (6%). Glacial lakes in HMA are widespread, but with a smaller signal size (average area of 0.22 km² and volume of 0.002 km³ per lake). The remaining water mass storage in
+
+Image /page/2/Figure/1 description: This figure, titled "Global distribution of glacial lakes and historical GLOFs," is divided into two parts, 'a' and 'b'.
+
+Part 'a', labeled "Glacial lake characteristics," displays a world map with red dots indicating the distribution and size of glacial lakes. Superimposed on the map are several multi-ringed circular charts for different regions, showing the total volume of glacial lakes in Gigatons (Gt). The volumes for each region are as follows:
+- Alaska: 22.0 Gt
+- Western Canada and USA: 8.4 Gt
+- Arctic Canada, north: 1.9 Gt
+- Arctic Canada, south: 24.9 Gt
+- Greenland periphery: 43.6 Gt
+- Scandinavia: 5.6 Gt
+- Russian Arctic: 7.2 Gt
+- Iceland: 2.4 Gt
+- Central Europe: 0.2 Gt
+- Caucasus and Middle East: 0.03 Gt
+- Asia, North: 0.4 Gt
+- Asia, Central: 1.7 Gt
+- Asia, Southwest: 0.7 Gt
+- Asia, Southeast: 2.1 Gt
+- Low-latitude Andes: 0.7 Gt
+- Southern Andes: 38.3 Gt
+- New Zealand: 0.2 Gt
+A large chart at the bottom indicates a global total volume of 160.3 Gt. A legend explains that the concentric rings on these charts represent Lake size (km²), Volume ratio, Area ratio, and Number ratio.
+
+Part 'b', labeled "GLOF characteristics," shows a similar world map illustrating the locations of historical Glacial Lake Outburst Floods (GLOFs). Pie charts for various regions depict the proportion of different lake (dam) types that caused these floods. The legend for the dam types includes: Bedrock, Combined, Ice, Moraine, Supraglacial, Englacial, and Others. The pie charts show that moraine-dammed lakes (orange) and bedrock-dammed lakes (light beige) are common causes of GLOFs in many regions. A legend also provides a scale for the number of GLOFs.
+
+Fig. 1 | Global distribution of glacial lakes and historical GLOFs. a, Area and volume of present-day (-2018) glacial lakes within 1 km of glaciers, aggregated over 220 km hexagon grid cells (red shades), and regional statistics of lake volume, area and number according to Global Terrestrial Network for Glaciers (GTN-G) glacier regions of the Randolph Glacier Inventory (RGI) (doughnut plots); outer, middle and inner circles indicate the total volume, area and number percentage of glacial lakes in a given size interval, respectively, with the total lake area in each region in the centre of the circle. Two RGI subregions (Svalbard and Jan Mayen, Antarctic and Subantarctic) are not included as they are not part of
+
+the underlying global glacial lake data set\*. **b**, Global distribution of historical glacial lake outburst floods (GLOFs) aggregated over 220 km hexagon grid cells (red shading), and regional breakdown of the GLOF number by glacial lake (dam) types²6 (pie charts). Some RGI subregions are not presented as they are not covered by the global GLOF inventory²6. More than 80% of global glacial lakes are distributed predominantly in Greenland, the Southern Andes, Alaska and western Canada, the eastern Canadian Arctic and High Mountain Asia (HMA); >60% of historical GLOFs occurred in Alaska, HMA and Iceland.
+
+Image /page/3/Figure/1 description: A diagram illustrating the components of assessing glacial lake outburst flood (GLOF) risk. The diagram features three main circular sections arranged in a triangle around a central concept of 'Risk'.
+
+The top red section, 'Mapping of glacial lake', is divided into 'Manual' and 'Automatic or semi-automatic' methods. These methods utilize 'High-resolution satellite data', 'Cloud-free SAR imagery', 'Optical multispectral imagery', 'Cloud computing', 'Artificial intelligence', and 'Machine learning'.
+
+The bottom-left yellow section is labeled 'Estimates of volume and discharge'. It shows a process flow from 'Area → volume' to 'Volume → discharge', based on an 'Empirical relationship'.
+
+The bottom-right blue section is titled 'GLOF susceptibility and simulation'. The inner circle is split between 'GLOF susceptibility' and 'GLOF simulation'. The outer ring, labeled 'Hydrodynamic modelling', lists factors such as 'External triggers (landslide, ice/snow avalanche, extreme rainfall)', 'Lake expansion', 'Ice dam flotation', 'Dam stability', 'Downstream sediment transport and inundation', 'Moraine erosion and failure', 'Wave generation and overtopping', 'Hydrodynamics of lake', and 'Ice or rock avalanche into a lake'.
+
+Arrows show that all three sections feed into the central concept of 'Risk', which is defined as 'Socio-economic exposure and vulnerability'. There is also an arrow from the 'GLOF susceptibility and simulation' section to the 'Estimates of volume and discharge' section.
+
+Fig. 2 | Methodology for quantifying glacial lake changes and impacts. Methods to map glacial lakes (red), estimate lake volume and peak (yellow) and assess glacial lake outburst flood (GLOF) susceptibility and simulation (blue). Glacial lake mapping, estimates of GLOF susceptibility, volume and peak discharge, and hydrodynamic simulations provide important baseline integrated data for GLOF hazard assessment. SAR, synthetic aperture radar.
+
+the world's glacial lakes in 2015–2018 was about 160 Gt, equivalent to about 0.43 mm of sea-level rise<sup>9</sup>.
+
+### Glacial lake changes
+
+Glacial lakes are naturally transient features. Each tends to follow a common trajectory, as documented in many mountainous regions around the world<sup>16,56-59</sup>. Specifically, glacial lakes showed a relatively stable trend from 1850 to 1970, followed by a gradual expansion in most regions of the world from 1970 to 1990 and a robust expansion from 1990 to 2020 (see Supplementary Fig. 1). Various processes can lead to such changes in glacial lake evolution, including shifts in inflow or outflow; gradual infilling by sediment, particularly in deglaciating environments<sup>23,60,61</sup>; abrupt infilling following mass movements<sup>62,63</sup>; coalescence of several lakes; detachment of an ice-marginal lake from the parent glacier as it retreats; or partial or complete dam failure resulting in a GLOF<sup>23</sup>. Although global statistics are unavailable, evidence from the tropical Andes suggests that, on average, only one in ten GLOFs has resulted in complete lake drainage<sup>62</sup>. Indeed, most moraine-dammed and bedrock-dammed glacial lakes persist after a GLOF has occurred<sup>62</sup>. Accordingly, lakes can exist for vastly different timeframes: ice-dammed lakes can form and fail repeatedly as long as the ice dam is thick enough to impound a lake  $^{33,64}$ ; bedrock-dammed lakes within overdeepenings can persist for tens of thousands of years<sup>65-67</sup>; and moraine-dammed lakes can survive over centennial or millennial timescales<sup>68</sup>.
+
+The aforementioned inventories provide an opportunity to track these glacial lakes over time (Fig. 3a; see Supplementary Fig. 1). Overall, the area of global glacial lakes expanded steadily from 1990 to 2018. Specifically, the glacial lake (>0.05 km²) number and area increased by -53% (from 9,414 to 14,379) and 51% (from  $5.93 \times 10^3$  km² to
+
+ $8.95 \times 10^3$  km<sup>2</sup>) from 1990 to 2018 (ref. 9). However, this global increase is subject to uncertainties due to missing lakes when compared with regional multi-temporal lake inventories (see Supplementary Table 2).
+
+Regional data sets, which undergo extensive quality control during compilation, also indicate a general increase in glacial lake area. Between ~1990 and ~2020, glacial lake area increased by an average of ~22% dec<sup>-1</sup> (Fig. 3a), higher than the 17% dec<sup>-1</sup> suggested by the global data set<sup>9</sup>. Of course, much regional and temporal variability is embedded. For instance, overarching changes of <10% dec<sup>-1</sup> from ~1990 to ~2020 were observed in the Southern Andes (from 4.463.9 km<sup>2</sup> to 4,789.7 km<sup>2</sup>), Central Europe (from 30.9 km<sup>2</sup> to 31.5 km<sup>2</sup>) and Svalbard (from 110.0 km<sup>2</sup> to 111.0 km<sup>2</sup>) (Fig. 3a), potentially highlighting the importance of short-term temporal variability; lake area temporarily decreased at the edge of the Northern Patagonian Icefield from 1945 to 1976 (ref. 69) and in Greenland from 1987 to 1992 and from 2005 to 2010 (ref. 16), perhaps contributing to the smaller, yet still positive, long-term trends<sup>16,69</sup>. In contrast, much larger changes of >40% dec<sup>-1</sup> were observed in Iceland (from 54.8 km<sup>2</sup> to 132.4 km<sup>2</sup>), Scandinavia (from 258.2 km<sup>2</sup> to 596.9 km<sup>2</sup>) and the Russian Arctic (from 189.9 km<sup>2</sup> to 478.4 km<sup>2</sup>) (Fig. 3a).
+
+In HMA, local variability is observed, as evidenced by sub-regional records of glacial lake growth (see Supplementary Fig. 1). In the greater Himalaya, glacial lake area increased by ~3.6 km² year $^{-1}$  (0.6% year $^{-1}$ , from 666.4 km² to 774.7 km²) from 1990 to 2020. Yet changes in the Chinese Himalaya and Nepal Himalaya were ~1.28 km² year $^{-1}$  (0.8% year $^{-1}$ , from 166.5 km² to 215.3 km²) from 1970 to 2008 (ref. 70) and 0.6 km² year $^{-1}$  (1.1% year $^{-1}$ , from 55.5 km² to 81.0 km²) from 1977 to 2017 (ref. 71), respectively.
+
+In addition to such substantial local and regional changes, there is also evidence of changes by glacial lake type. Ice-marginal proglacial
+
+lakes are in direct contact with their parent glaciers and have strong interactions between them, which is mainly dammed by moraines, with more GLOFs recorded by lake type<sup>26,68</sup>. The imbalance between ice-marginal and glacier-detached proglacial lakes has been recorded worldwide (Fig. 3b-e). For example, in Southern Pamir, the number of proglacial lakes remained largely constant from 1968 to 1993 (the lower number in 1977 might be related to uncertainties due to poor image quality), increased to a peak of 44 in 2008 and then decreased to 26 by 2015 (ref. 72) (Fig. 3b).
+
+In the Andes, the number and area of proglacial lakes have progressively increased. From 1948 to 2017, three to four new proglacial lakes formed each year, with an average expansion rate of ~0.1 km² year¹ (ref. 73). Of these total lakes, the number of proglacial lakes was stable from 1948 to 1990 (71–74 lakes, with some fluctuations in between) and then decreased to 35 by 2018 (ref. 73). For the Peruvian Andes, an almost constant gain of detached proglacial lake area was observed throughout the study period (Fig. 3c). This phenomenon, together with the increasing average elevation of proglacial lakes, is a strong indication that large lakes dammed by Little Ice Age terminal moraines have lost contact with their parent glaciers³3.
+
+In the Southern Alps of New Zealand, an almost linear trend in proglacial lake area change has also been reported, with an average of -0.5 km² year⁻¹ in four time steps from 1990 to 2020 (ref. 74). Tasman Glacier, New Zealand's largest glacier, is retreating rapidly, and a proglacial lake formed in 1990 (ref. 52). The number of ice-marginal proglacial lakes in New Zealand was greatest in 2011, and thereafter has decreased slightly as lakes coalesced and as some became disconnected from the glaciers <sup>74</sup> (Fig. 3d).
+
+In the European Alps, particularly the Austrian Alps, the number and area of proglacial lakes increased from 1850 to 2015, with the mean change in area doubling from -0.038 km² year  $^{-1}$  during 1998–2006 to -0.079 km² year  $^{-1}$  during 2006–2015 (ref. 75). Similar changes have occurred in Switzerland. There, the number of proglacial lakes increased from 21 in 1900 to 47 in 2006, and from 47 in 2006 to 82 in 2016 (ref. 59) (Fig. 3e). The increase in the total area of proglacial lakes in the Swiss Alps has accelerated from -0.01 km² year  $^{-1}$  in 1850–1900 up to -0.15 km² year  $^{-1}$  in 2006–2016 (ref. 59). However, part of this variability is due to the artificial impoundment of some proglacial lakes, mainly for hydropower generation purposes.
+
+## Historical glacial lake outburst floods
+
+With the presence of glacial lakes comes the risk of a GLOF. GLOFs arise when the lake dam is breached (by various triggers depending on the dam type) and meltwater is suddenly and rapidly released from the lake, subsequently flowing downstream where it can cause substantial societal and environmental impacts (Fig. 4). Together, data from river gauges, satellite images, sediment stratigraphy, and eyewitness and media reports suggest that at least 3,151 GLOFs have occurred globally since 850 (ref. 26) (Fig. 1b), with most recorded in Alaska (24%, n = 768), HMA (18%, n = 569) and Iceland (19%, n = 590).
+
+However, understanding of GLOFs is severely limited by inadequate data. Triggers are often unknown or speculated owing to absent field data, necessitating reliance on climatic or geological-geomorphological evidence<sup>35,76-78</sup>. Numerical modelling is similarly challenged by absent knowledge of parameter values such as lake water temperature or time-varying lake level, limiting insight<sup>58,79,80</sup>. Moreover, GLOF frequency calculations suffer from selective reporting and changes in instrumentation and satellite coverage<sup>33,81,82</sup>. Nevertheless, it is possible to glean evidence of GLOF characteristics, as now
+
+discussed for ice-dammed, moraine-dammed and bedrock-dammed GLOFs. Resulting hazards and changes are also assessed.
+
+### Ice-dammed GLOFs
+
+GLOFs from ice-dammed lakes are the most common type. Their activity is largely controlled by the activity of the damming glaciers and resulting fracture or flexure of the ice dam (Fig. 4a). Flotation of the ice dam once a critical lake level is reached also acts as a potential breach mechanism<sup>33,83,84</sup>. When lake water suddenly drains through an ice dam, thermal and mechanical erosion (usually progressively enlarging a subglacial tunnel) leads to an exponentially rising discharge<sup>54,85,86</sup>.
+
+GLOFs from ice-dammed lakes constitute roughly two-thirds (2,039 of 3,151) of all reported GLOFs worldwide<sup>26,81</sup>. Their frequent occurrence is linked to the fact that they can drain repeatedly as an ice dam will 're-heal' temporarily once energetic egress of lake water has subsided. At the regional scale, Alaska is a hot spot of GLOFs from ice-dammed lakes (n = 764; 99% within Alaska, 37% of the global total in this category)<sup>19,26,33</sup> (Fig. 1b). Indeed, 750 ice-dammed lakes have formed in Alaska over the twentieth century<sup>87,88</sup>, with some lakes having a volume >10<sup>9</sup> m<sup>3</sup> when impounded by a thick outlet glacier<sup>87</sup>. Yet their GLOFs rarely cause damage, owing to low exposure of assets along river channels. Ice-dammed lakes are also the dominant source of reported GLOFs in Greenland periphery (n = 153; 100% of all GLOFs within Greenland periphery, 7.5% of global ice-dammed GLOFs)<sup>89,90</sup> the Karakoram (n = 190; 33.4% within HMA, 9.3% of global ice-dammed GLOFs) $^{29,91}$ , Scandinavia (n = 183; 96.8% within Scandinavia, 9.0% of globalice-dammed GLOFs) $^{92}$ , Central Europe (n = 258; 58.2% within Central Europe, 12.7% of global ice-dammed GLOFs) and Iceland (n = 237; 40.2% within Iceland, 11.6% of global ice-dammed GLOFs)<sup>86</sup> (Fig. 1b). In Central Europe, fatalities and damage from these occasional ice-dam failures were particularly high before the twentieth century, possibly related to limited warning and flood control measures<sup>93</sup>. In Scandinavia, and particularly in Norway, although the percentage of ice-dammed GLOFs was high in the mid-twentieth century, the total number and size of outbursts are small compared with other regions<sup>26</sup>.
+
+### Moraine-dammed GLOFs
+
+GLOFs from moraine-dammed lakes are the second most common type globally after ice-dammed lakes (Fig. 1b). In these situations, dam breaching can be caused by seepage of lake water through unconsolidated material, degradation of an ice core, or overtopping due to calving, avalanches or landslides into the lake<sup>76,83</sup> (Fig. 4). Moraine-dammed lakes in the ice-marginal phase of evolution are the most likely to produce GLOFs due to calving processes and proximity to ice avalanche release zones. In contrast, glacial lakes detached from glaciers become disconnected from the most frequent GLOF-triggering processes with increasing distance from glaciers<sup>56</sup>. The almost instantaneous, and sometimes complete, failure of a moraine dam produces a linearly rising discharge<sup>94</sup>. Furthermore, complete failure of the (moraine) dam restricts any future ability to impound meltwater<sup>95,96</sup>. Hence, moraine-dammed lakes usually fail only once.
+
+GLOFs from moraine-dammed lakes make up at least -13% (398 of 3,151) of the global total number of recorded GLOFs from 850 to 2022 (Fig. 1b). However, their occurrence is highly regionally variable. The majority of those reported are found in the low-latitude Andes (n=100; 61.3% of GLOFs within the low-latitude Andes, 25.1% of global moraine-dammed GLOFs), Central Asia (n=110; 46.2% within Central Asia, 27.6% of global moraine-dammed GLOFs) and the Himalaya (n=71; 77.2% within the Himalaya, 17.8% of global moraine-dammed GLOFs) (Fig. 1b).
+
+Image /page/5/Figure/1 description: A multi-part figure from a scientific publication detailing changes in glacial lakes. The figure is divided into five sections labeled 'a' through 'e'.
+
+Part 'a', titled 'Glacial lake area changes', displays a world map with bar charts showing the increase in glacial lake area (in km²) for various regions between two different years. Red dots on the map indicate the locations and the magnitude of area change per decade, with a legend showing dot sizes for percentage changes from 0-5% to 60-70%. The bar charts show the following approximate data:
+- Alaska: from ~650 km² in 1984 to ~1,300 km² in 2019.
+- Western Canada and USA: from ~250 km² in 1990 to ~400 km² in 2015.
+- Low-latitude Andes: from ~40 km² in 1990 to ~70 km² in 2015.
+- Arctic Canada, north: from ~110 km² in 1990 to ~230 km² in 2015.
+- Arctic Canada, south: from ~700 km² in 1990 to ~1,300 km² in 2015.
+- Greenland periphery: from ~1,800 km² in 1990 to ~3,000 km² in 2015.
+- Iceland: from ~40 km² in 1990 to ~140 km² in 2015.
+- Svalbard: from ~80 km² in 1990 to ~180 km² in 2020.
+- Scandinavia: from ~200 km² in 1990 to ~700 km² in 2015.
+- Russian Arctic: from ~150 km² in 1990 to ~550 km² in 2015.
+- Central Europe: from ~30 km² in 1990 to ~32 km² in 2015.
+- Caucasus and Middle East: from ~0.4 km² in 1990 to ~0.7 km² in 2015.
+- Asia, Southwest: from ~120 km² in 1990 to ~200 km² in 2018.
+- Asia, Southeast: from ~270 km² in 1990 to ~430 km² in 2018.
+- Asia, Central: from ~450 km² in 1990 to ~730 km² in 2018.
+- Asia, North: from ~10 km² in 1990 to ~19 km² in 2018.
+- Southern Andes: from ~3,500 km² in 1986 to ~5,100 km² in 2016.
+- New Zealand: from ~7 km² in 1990 to ~32 km² in 2020.
+- Global: from ~10,000 km² in 1990 to ~15,000 km² in 2020.
+
+Parts 'b' through 'e' are titled 'Proglacial lake changes' and show detailed charts for specific regions. These charts plot the number of new lakes (blue, above the x-axis) and detached lakes (green, below the x-axis) per year, as well as the area of these lakes. The total area (A) is noted for each time period.
+
+- Part 'b' (Southern Pamir): Data for 1968 (A=0.68 km²), 1977 (A=0.83 km²), 1993 (A=1.09 km²), 2002 (A=1.45 km²), 2008 (A=1.46 km²), and 2015 (A=0.90 km²).
+- Part 'c' (Peruvian Andes): Data for 1948 (A=6.27 km²), 1962 (A=5.84 km²), 1970 (A=0.55 km²), 1990 (A=4.98 km²), 2000 (A=4.26 km²), 2010 (A=3.16 km²), and 2018 (A=1.65 km²).
+- Part 'd' (New Zealand): Data for 1990 (A=9.60 km²), 2000 (A=1.26 km²), 2011 (A=10.34 km²), and 2020 (A=20.55 km²).
+- Part 'e' (Switzerland): Data for 1900 (A=0.20 km²), 1946 (A=0.49 km²), 1973 (A=11.13 km²), 1982 (A=0.40 km²), 2006 (A=0.52 km²), and 2016 (A=0.93 km²).
+
+**Fig. 3** | **Global and regional changes in glacial lake area. a**, Absolute regional lake area in -1990 and -2020 (bars) and the percentage changes from 1990 to 2020 (red circles) (see Supplementary Table 2); owing to data availability, years were selected as close to 1990 and 2020 as possible. **b-e**, Changes in the number (dark blue line) and area (light blue bar) of new lakes, and the number (dark green line) and area (light green bar) of detached proglacial lakes in Southern
+
+Pamir $^{72}$  (panel **b**), Peruvian Andes $^{73}$  (panel **c**), New Zealand $^{74}$  (panel **d**) and Switzerland $^{59}$  (panel **e**) (see panel **a** for locations); time blocks differ owing to contrasting data availability. Error bars indicate uncertainties and inaccuracies due to low image resolution or quality, or the use of images from different years for the same time step. Glacial lakes have expanded worldwide, especially for proglacial lakes.
+
+Indeed, 52% and 32% of glacial lakes in the Central Andes and Himalaya are moraine-dammed, respectively  $^{17,97}$ , and many others have associated geomorphological features suggesting possible past dam failures  $^{62,98,99}$ . The exact timing of these dam failures has yet to be determined, as the geomorphological evidence is largely limited to that visible in satellite images that were first available in the 1960s.
+
+### Bedrock-dammed and subglacial GLOFs
+
+Bedrock dams are usually stable and therefore these glacial lakes produce fewer GLOFs than ice-dammed and moraine-dammed lakes, with dam overtopping being the only mechanism for GLOF occurrence 100,101. GLOFs from bedrock-dammed lakes make up only ~1% (30 of 3,151) of the total number of GLOFs from 850 to 2022 (Fig. 1b). Indeed, overtopping of bedrock barriers has only rarely been reported<sup>81</sup>, in contrast to the high regional abundance of bedrock-dammed lakes<sup>62</sup>, especially in the Himalaya and Andes 17,97. Their limited occurrence might arise from the fact that unfractured bedrock dams can withstand increases in hydrostatic pressure, and hence are only overtopped by displacement waves from a mass movement entering the lake<sup>62</sup>. GLOFs from bedrockdammed lakes are only reported after the 1960s and those occurred mainly in the low-latitude Andes (n = 27; 16.6% within the low-latitude Andes, 90% of global bedrock-dammed GLOFs) (Fig. 1b). Although evidence is limited, GLOF counts from bedrock-dammed lakes might be underestimated given the short preservation time of any geomorphic evidence, especially compared with moraine dam failures<sup>102</sup>.
+
+Similarly, there are relatively few GLOFs attributed to englacial or subglacial lakes induced by geothermal or volcanic activity compared with other GLOF triggers and mechanisms<sup>103</sup>. GLOFs from englacial lakes produce poorly preserved geomorphological evidence and are difficult to detect with optical remote sensing. They are mainly reported in Central Europe (n = 112, 25.3% within Central Europe, 69.1% of global englacial GLOFs) (Fig. 1b). Floods caused by rapidly draining englacial lakes have also occasionally been reported in New Zealand, western Canada and the United States 104-106 (Fig. 1b). The majority of GLOFs from subglacial lakes are also found in Central Europe (n = 5; 1.13% within Central Europe, 71.4% of global subglacial GLOFs), but further evidence is found in Southeast Asia (n = 1; 1.18% within Southeast Asia, 14.3% of global subglacial GLOFs) and the low-latitude Andes (n = 1; 0.61%) within the low-latitude Andes, 14.3% of global subglacial GLOFs) (Fig. 1b). In Iceland, for example, at least 264 outbursts from a small number of subglacial lakes<sup>31</sup>, which repeatedly form and drain owing to geothermal activity beneath the ice caps, have been observed, with further evidence of glacier-volcano interactions both in historical and prehistoric times 107,108.
+
+## **GLOF** hazard assessment
+
+Understanding the characteristics and changes of glacial lakes and GLOF triggers and mechanisms is the basis for GLOF hazard assessment, which is critical for disaster prevention and damage mitigation. Globally, more than 10 million people are potentially exposed to the impacts of GLOFs, mainly in HMA<sup>38,109</sup>. Potentially dangerous glacial lakes have
+
+been identified using a range of assessment approaches, ranging from large-scale automated methods considering key determinants such as lake size, catchment area, ice/rock avalanche potential and dam steepness<sup>110,111</sup>, through to detailed catchment or lake-specific assessments including field investigations<sup>112-114</sup>. Identifying the dam type of the glacial lake is a critical first step in GLOF hazard assessments, as it determines the stability of the lake, triggering processes, the potential magnitude of the GLOF and the mechanisms of drainage<sup>115</sup>.
+
+For GLOF hazard assessment a distinction is typically made between those factors that are critical to the inherent stability of the lake dam (geometry and structure of the dam) and those that influence the potential for an external triggering event (a rock or ice avalanche)<sup>100</sup>. Additionally, the hydro-geomorphic characteristics of the lake catchment area, which can influence the susceptibility to precipitation or melt-triggered outburst events, have also been highlighted<sup>37,116</sup>. With high-resolution optical imagery (such as available from Google Earth) and corresponding high-quality digital terrain models, it has become possible to quantify various physical characteristics of the dam and catchment area remotely over large spatial scales<sup>111,117</sup>. However, precise geometric measurements (such as dam freeboard or dam height) and in situ characteristics (such as ice core, lithology) are lacking, and can only be confirmed through local site investigations.
+
+Typically, large glacial lakes are emphasized owing to the potential high-magnitude GLOFs that could originate from them, yet outbursts from small or even seasonal lakes have proven capable of eroding huge volumes of sediment and producing catastrophic downstream process chains 118. GLOFs can be both the trigger of a process chain or initiated as part of a larger process chain, as in the case of a mass movement causing a displacement wave, overtopping and a downstream flood 119. Several hydrodynamic numerical models 80,120-123 have been used to numerically propagate flood waves downstream, modelling not only a GLOF but also entire process chains (see Supplementary Table 3). However, in the absence of robust field data from past events, modelling is often undertaken with poorly constrained conditions and parameters. Some morphodynamic models that include sediment transport and bed elevation changes during a simulation have also been applied where independent pre-GLOF and post-GLOF topographic data are available 124,125. These models can estimate sediment entrainment and transport rates 126, and more generally this numerical modelling has the potential to interpolate between observations along a river reach, or to extend beyond them downstream.
+
+### Trends in GLOF frequency and magnitude
+
+Although suffering from known biases and limitations <sup>62,102</sup>, remote sensing observations and geomorphological evidence offer an opportunity to roughly estimate whether GLOFs have changed in frequency. The number of reported GLOFs increased from 1 in 850 to 3,151 in 2022, peaking from 2000 to 2020 (refs. 26,35). However, poor instrumentation and limited available satellite imagery might be responsible for fewer records of historical GLOF events before -1980, suggesting a bias in the estimation of the historical GLOF trend<sup>26,81,127</sup>. At the regional
+
+Image /page/7/Figure/1 description: A flowchart titled 'a GLOF initiation' illustrates the causes, processes, and impacts of a Glacial Lake Outburst Flood (GLOF). The chart is color-coded with a legend indicating that blue boxes represent 'Trigger or mechanism', yellow boxes represent 'Flood modelling', and red boxes represent 'Downstream impact'.
+
+The 'Trigger or mechanism' section shows several pathways. A 'Moraine-dammed' lake leads to a 'Moraine', which can be breached by an 'Ice or rock avalanche, landslide, overtopping, piping and dam breach'. Separately, 'Ice-dammed' and 'Supraglacial' lakes lead to 'Ice', which can be breached by an 'Ice avalanche, subaerial breaching, terminal retreat and dam breach'. Other triggers like 'Earthquake', 'Temperature, snowfall or rainfall', and 'Ice avalanche or collapse' also contribute to the breach mechanisms.
+
+These triggers lead to a 'GLOF', which is the start of the 'Flood modelling' section. The GLOF event influences its 'Discharge' and 'Extent'. The GLOF initiates the 'Process of flood flow downstream'.
+
+This downstream flow leads to 'Erosion and hazard', which is the start of the 'Downstream impact' section. The impacts listed are on 'Road', 'Bridge', 'Hydropower station', 'Electrical facility', 'Building', 'Livelihood', and 'Culture heritage'.
+
+### **b** GLOF hazard chain
+
+Image /page/7/Figure/3 description: An infographic titled "Hazard chain" illustrating the causes and effects of natural disasters in a mountainous, glacial environment. The top portion is a diagram showing a sequence of events. At the highest elevation, "Extreme snowfall and rainfall" occurs over snow-capped mountains. A "Glacier" flows down a valley, where an "Ice avalanche" (labeled 2) falls into a "Glacial lake" (1). A "Landslide" (3) is also shown falling into the lake. An "Earthquake" is indicated as another potential trigger. Below the lake, an "Ice core" (4) is depicted. The outflow from the lake creates a "Debris fan" (5), leading to "Erosion" (6) and eventually "Downstream damage and deposition" (7 and 8) in a village. An "Ice-dammed lake" is also shown in a separate valley. The bottom portion of the image consists of eight numbered photographs that correspond to the numbered locations in the diagram, providing real-world examples: 1. A muddy glacial lake. 2. An ice avalanche calving into a lake. 3. The scar of a landslide on a mountainside. 4. A glacier with a terminal lake. 5. A large debris fan in a valley. 6. A river valley showing significant erosion. 7. A damaged bridge and debris-filled riverbed. 8. A building damaged by large rocks and flood debris.
+
+**Fig. 4** | **The triggers and mechanisms of GLOFs. a**, The process of a glacial lake outburst flood (GLOF), from trigger of mechanisms (blue) and flood modelling (yellow) to downstream impact (red). **b**, The hazard chain, from triggering factors
+
+to glacial lake outburst, downstream impacts and damage, and corresponding photographs (1–8). Understanding the triggers and mechanisms of GLOFs and flood modelling practices can improve hazard mitigation.
+
+scale, GLOFs from ice-dammed lakes in Alaska revealed an unchanged frequency and decreasing magnitude from 1985 to 2020 (ref. 82). Furthermore, time-series analysis of satellite images in the Himalaya suggest an unchanged frequency of moraine-dammed GLOFs from 1988 to 2017 (ref. 128). Therefore, the increasing number of moraine-dammed and bedrock-dammed lakes in mountainous regions worldwide does not necessarily imply more GLOFs 912.
+
+Little is also known about how average or extreme GLOF flood discharges and volumes have changed, but some general themes have emerged. For instance, flood volumes of the largest reported GLOFs from ice-dammed lakes worldwide have decreased by an order of magnitude since 1900 (ref. 33). In contrast, the average peak discharge and volume of ice-dammed GLOFs remain unchanged 33. These changes align with the thinning of ice dams with atmospheric warming and glacier ablation, such that maximum lake levels gradually fall and drainage of ice-dammed lakes occurs earlier in the melt season 33,83. Moreover, as glaciers retreat, many glacier dams can no longer impound lakes at their termini at lower elevations, or the glaciers are lost altogether 82. Where ice-dammed lakes can form at higher elevations they tend to be smaller because the topography is steeper, and thus the water storage capacity is limited.
+
+## Future glacial lakes and outburst floods
+
+Whether the increasing number and size of glacial lakes will lead to higher-magnitude GLOF discharges and greater GLOF volumes in the future remains a controversial hypothesis. With such apparent historical changes in glacial lake number and corresponding GLOFs, predicting their changes in a warming climate is important for understanding future natural hazards <sup>129,130</sup>. The future of glacial lakes and potential hazard hot spots are now discussed.
+
+### Future glacial lake evolution
+
+The life cycle of glacial lakes is known to shift in a warming climate (Fig. 5). Supraglacial lakes can grow, coalesce and, eventually, form proglacial lakes. Simultaneously, existing proglacial lakes can expand in area and in volume (Fig. 5). Formation of new lakes and the growth of existing proglacial lakes can accelerate glacier terminus retreat, via increased frontal ablation <sup>131,132</sup>. Future increases in the volume of proglacial lakes will generally increase the downstream damage of any GLOFs <sup>38</sup>. In addition, some ice-dammed lakes can form following glacier surges, and episodic failures of ice-dammed lakes are highly destructive <sup>133,134</sup> (Fig. 5). Ice-dammed lakes will continue to breach in the near future, even as glaciers thin and retreat <sup>33,64</sup>.
+
+Beyond these conceptual changes, estimating subglacial topography and identifying potential overdeepenings are critical to project the location and size of future glacial lakes. As glaciers retreat, depressions can be exposed that have the potential to fill partly or fully with meltwater to develop into glacial lakes<sup>129</sup>. Identification of these overdeepenings requires knowledge (or an estimate) of subglacial bedrock morphology, which can be modelled by subtracting ice thickness<sup>20,135</sup> from a digital elevation model<sup>136</sup>, and a hydrological filling algorithm applied to fill the overdeepenings<sup>137</sup>. The difference between the filled and the original subglacial bed topography provides a bathymetric grid suggesting the extent and volume of the overdeepenings and, hence, of potential future glacial lakes<sup>24,136</sup>. Glacier evolution models<sup>20,138</sup> can also predict future rates of glacier retreat so as to estimate when glacial lakes might begin to form<sup>21,24,136</sup>. However, these approaches cannot currently determine the dam type nor confidently suggest the height of the future lake outlet<sup>24</sup>. Nonetheless, despite the high and as yet unassessed uncertainty, these ice thickness–overdeepening approaches have been used to project regional changes in glacial lakes, including  $HMA^{21,24}$  (Fig. 6; see Supplementary Table 4).
+
+In HMA, glacial lake area and volume are generally expected to increase. These changes arise owing to high projected glacier volume losses (60–75% by 2100 under unabated  $\rm CO_2$  emissions  $^{25,139}$ ), exposing subglacial overdeepenings and, thus, expanding existing lakes and developing new glacial lakes  $^{21,24}$ . For example, with complete glacier disappearance in HMA, -13,000 new glacial lakes (each larger than 0.01 km²) could potentially form, with a total area of -1,510 km² and volume of -52.3 km³ (ref. 24). Similar estimates of a maximum number of 10,068 [6,981, 13,538] new ice-dammed lakes are projected by 2100 under SSP585, with an area of 290 [204, 639] km² and volume of 2.9 [1.9, 4.9] km³ (ref. 64). However, these quantifications are uncertain. Indeed, a different ice thickness model and, hence, subglacial bed topography suggests -25,285 lakes (larger than 0.01 km²) could form by complete deglaciation, having an area of 2,683 ± 773.8 km² and volume of 99.1 ± 28.6 km³ (ref. 136).
+
+The magnitude of the projected glacial lake changes in HMA is regionally variable as well as strongly dependent on the emission scenario (hence, temperature rise). Generally, greater changes are expected in western regions (Fig. 6a-c). The total area of glacial lakes, for example, is expected to increase by  $7.2 \times 10^2$  km<sup>2</sup> in Central Asia compared with  $5.6 \times 10^2$  km<sup>2</sup> in Southwest Asia and  $2.3 \times 10^2$  km<sup>2</sup> in Southeast Asia (Fig. 6a,c). Glacial lake volume also reflects similar shifts (Fig. 6b,c). Future lake area and volume estimates show an increase of about 771.4 km<sup>2</sup> and 37.0 km<sup>3</sup>, respectively, from 2030 to 2100 under a high-emission scenario; generally, these have a continuous increase until about 2090 and a decrease thereafter (Fig. 6e,f). However, there are sub-regional differences, with larger increases in the Karakoram (82.5 km<sup>2</sup> and 4.2 km<sup>3</sup>) and Himalaya (187.0 km<sup>2</sup> and 10.4 km<sup>3</sup>) compared with smaller increases in the Hissar Alay (1.7 km<sup>2</sup>) and 0.03 km<sup>3</sup>) and the Hindu Kush (1.4 km<sup>2</sup> and 0.02 km<sup>3</sup>) over the same time period and emission scenario (Fig. 6e,f).
+
+Glacial lake area and volume are also expected to increase somewhat in the European Alps. At the bigger regional scale under complete deglaciation, lake area and volume are expected to increase by about 477.0 km<sup>2</sup> and 1.7 km<sup>3</sup>, respectively (Fig. 6a,b). However, at a national level, different climate scenarios and glacier evolution models are often used that make comparisons difficult. In the Swiss Alps, the number and area of lakes are 987 and  $6.22 \pm 0.25 \,\mathrm{km^2}$  in 2016, respectively (ref. 59). If all glaciers were to disappear, ~500 new glacial lakes (of at least 0.01 km<sup>2</sup>) could form, with a total area of about 50 km<sup>2</sup> and volume of 1.6 km<sup>3</sup> (ref. 129). These estimates are broadly in keeping with ice-free scenarios based on current ice thickness measurements; they suggest that 683 glacial lakes (larger than 0.005 km²) with a total area and volume of  $45.2 \pm 9.3 \text{ km}^2$  and  $1.16 [1.05, 1.32] \text{ km}^3$  (mean with 95% confidence interval), respectively, could develop<sup>23</sup>. Looking at specific emission scenarios rather than total ice elimination, glacial lake volumes of 0.35 [0.12, 0.49] km<sup>3</sup> and 0.94 [0.51, 1.04] km<sup>3</sup> could be realized by 2100 for SSP1-2.6 and SSP5-8.5, respectively<sup>23</sup>. In the upper Rhône catchment of the south-western Swiss Alps alone, 100 sites are predicted to have a high potential for future glacial lake formation<sup>140</sup>.
+
+Smaller changes are anticipated in other regions of the European Alps. In Austria, 42 potential new glacial lakes with a total area of about 2 km² could form with complete deglaciation <sup>141</sup>, an increase from the current 1,410 lakes with a total area of 17.1 km² in 2015 (ref. 75). A similar number of new lakes are expected under complete ice loss
+
+Image /page/9/Figure/1 description: A multi-panel figure illustrating the formation and growth of glacial lakes and the associated hazard of Glacial Lake Outburst Floods (GLOF). The figure is divided into four sections labeled a, b, c, and d. Sections a, b, and c each contain two diagrams comparing a 'Current climate' scenario with a 'Future climate' scenario. Section a, titled 'Supraglacial lake', shows small pools of water on a glacier's surface in the current climate, which increase in number and size in the future climate, a process labeled 'Supraglacial lake increase and proglacial lake formation'. Section b, 'Proglacial lake', depicts a lake at the glacier's terminus that grows larger as the glacier recedes in the future climate, labeled 'Proglacial lake growth'. Section c, 'Ice-dammed lake', shows the formation of a lake in a tributary valley dammed by the main glacier in the future climate, a process called 'Ice-dammed lake formation'. Section d, titled 'Increasing GLOF hazard potential', illustrates a GLOF event where a landslide falls into a glacial lake, causing a flood that damages a populated area downstream. This section also includes a Venn diagram where three overlapping circles for 'Hazard', 'Vulnerability', and 'Exposure' intersect to create 'Risk'. Arrows labeled 'Adaptation' and 'Countermeasure' are shown, linking the GLOF illustration to the risk concept.
+
+Fig. 5 | Future formation and growth of glacial lakes and GLOF hazard.  $\mathbf{a}$ - $\mathbf{c}$ , Evolution of supraglacial lakes (panel  $\mathbf{a}$ ), proglacial lakes (panel  $\mathbf{b}$ ) and ice-dammed lakes (panel  $\mathbf{c}$ ) from the current climate (top panel) to a future climate (middle panel).  $\mathbf{d}$ , The overarching change in glacial lake outburst flood
+
+(GLOF) hazard in a future climate ('+' and '-' indicate increases and decreases that influence the GLOF hazard), feeding into GLOF risk. The life cycle of different types of glacial lakes shifts in a warming climate and can contribute to increased GLOF risk, with more and larger lakes and greater exposure.
+
+in the Aosta Valley region, Italy; 46 potential new lakes larger than  $0.01\,\mathrm{km^2}$  might develop, with a total area of  $3.1\pm0.9\,\mathrm{km^2}$  and a volume of  $0.06\pm0.02\,\mathrm{km^3}$  (ref. 142), expanding the current 186 lakes with a total area of  $1.4\,\mathrm{km^2}$  (ref. 143). Likewise, 80 (48 with a high level of confidence) potential new lakes could form under full deglaciation in the Mont Blanc massif of the French Alps 144.
+
+Smaller changes are generally expected in glacial regions of the Southern Hemisphere, including the New Zealand Alps. Here, lake area is expected to increase by 42.3 km<sup>2</sup> by 2100 under RCP4.5 (Fig. 6a,b). More specifically, ice-marginal lakes across the Southern Alps could
+
+expand from 20.9 km² in 2020 to 40.4 km² in 2050 (ref. 74). However, from 2050 onwards, the rate of increase in total area is projected to decrease from 0.65 km² year $^{-1}$  in 2020–2050 to 0.04 km² year $^{-1}$  in 2050–2100 as lakes detach from glaciers and/or fill with sediment $^{74}$ .
+
+In the Peruvian Andes, glacial lake area is projected to increase by 65.5 km² and volume by 0.7 km³ by 2100 under RCP8.5 (refs. 73,145) (Fig. 6a,b). In the still glaciated Cordillera Blanca, 38 or 50 new lakes could emerge by 2100 under RCP2.6 or RCP8.5, respectively  $^{73}$ , increasing from the current 870 lakes; accordingly, the glacial lake area is expected to increase by -10% compared with 2018 values, from 35.2 km²
+
+### a Projected glacial lake area
+
+Image /page/10/Figure/2 description: A world map with several rectangular regions outlined. Some of these regions are color-coded according to a legend at the bottom. The legend is a color scale from dark blue to light blue, to pale yellow, to orange, corresponding to a numerical scale from 0 to 8, with the unit labeled as "10² km²". A large region in South America is colored dark blue. A region in Europe is colored pale yellow. A region in Central Asia is colored with orange, pale yellow, and light blue sections. An inset map of New Zealand in the bottom right is also colored dark blue. Other outlined regions in North America and northern Asia are not colored.
+
+### **b** Projected glacial lake volume
+
+Image /page/10/Figure/4 description: A world map showing continents in light gray. Several rectangular regions are outlined across the globe. Some of these regions are color-coded based on a legend at the bottom. The legend is a color bar with a scale from 0 to 3 in units of 10 km³. The colors range from dark blue (0-1), to a lighter blue-gray (1-2), to a pale yellow (2-3). On the map, a large portion of South America is colored dark blue. The Mediterranean region is colored light blue-gray. A region in East Asia is colored pale yellow, with its southern coastal area colored dark blue.
+
+### C Projected glacial lake area and volume in HMA
+
+Image /page/10/Figure/6 description: A map titled 'Projected glacial lake area and volume in HMA' shows a geographical region, likely High Mountain Asia, with data points represented by circles of varying sizes and colors. There are two legends. The first legend explains the size of the circles, which corresponds to the glacial lake area in square kilometers (km²). The categories are: a small dot for <3, and progressively larger circles for 3-10, 10-30, 30-60, and >60 km². The second legend is a color bar indicating the volume in 10⁻¹ km³, where colors range from blue (0.001) to red (1.5). The map shows a high concentration of large, red circles (indicating large area and high volume) in the central-western part of the region. Smaller circles, colored yellow and dark blue, are scattered more widely.
+
+### d Projected GLOF hazard and risk in HMA
+
+Image /page/10/Figure/8 description: A map showing projected GLOF (Glacial Lake Outburst Flood) hazard and risk in HMA (High Mountain Asia). The map displays a mountainous region divided into several sub-regions. Overlaid on these regions are groups of three vertical bar charts. A legend in the bottom-left corner explains that the bars represent 'Risk' (yellow), 'Exposure' (light orange), and 'Hazard' (darker orange). The legend also includes a scale bar labeled '600' corresponding to the height of the example 'Hazard' bar. The heights of the bars vary across the different regions on the map, indicating different levels of projected hazard, exposure, and risk.
+
+### e Scenario dependence of HMA glacial lake area
+
+Image /page/10/Figure/10 description: A set of four stacked area charts showing projections of glacial lake area from the year 2040 to 2100 under four different climate scenarios: SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5. The vertical axis is labeled "Glacial lake area (×10² km²)" and ranges from 0 to 8. The horizontal axis shows the years 2040, 2060, 2080, and 2100. Each chart shows a progressive increase in glacial lake area over time, with the area represented by stacked layers of different shades of blue. The total area increases with the severity of the climate scenario. For SSP1-2.6, the area reaches approximately 4.5 ×10² km² by 2100. For SSP2-4.5, it reaches about 5.8 ×10² km². For SSP3-7.0, it reaches about 6.8 ×10² km². For SSP5-8.5, the area increases most sharply, exceeding 8 ×10² km² by 2100.
+
+### **f** Scenario dependence of HMA glacial lake volume
+
+Image /page/10/Figure/12 description: A set of four stacked area charts illustrating the scenario dependence of HMA (High-Mountain Asia) glacial lake volume. The vertical axis represents "Glacial lake volume (×10 km³)" and ranges from 0 to 4. The horizontal axis for each chart shows the years from 2040 to 2100, with labels at 20-year intervals. Each chart depicts a different scenario, showing a projected increase in total glacial lake volume over time. The total volume is broken down by region, with each region represented by a different shade of blue or green in the stacked area chart. A legend at the bottom identifies the regions: Tibet, Himalaya, Karakoram, Kun Lun, Tien Shan, Qilian Shan, Pamir, Hissar Alay, Hindu Kush, and Hengduan Shan. In all four scenarios, the total volume increases from near zero in 2040. By 2100, the projected total volumes for the four scenarios are approximately 2.1, 2.8, 3.0, and 3.8 (×10 km³), respectively, from left to right.
+
+**Fig. 6 | Global potential future glacial lakes and GLOF hazard. a, b,** Published estimates of potential future glacial lake area (panel **a**) and potential future glacial lake volume (panel **b**) with complete deglaciation (see Supplementary Table 4). **c**, Potential future glacial lake area (circles) and volume (shading) in High Mountain Asia (HMA) by 2100 under RCP8.5 (ref. 24). Background shading depicts drainage basins using the colours in panels **e** and **f. d**, Future glacial lake outburst flood (GLOF) risk (light shading), exposure (medium shading) and
+
+hazard (dark shading) in HMA by 2100 under RCP8.5 (ref. 24). Physical or social vulnerability is not considered. **e,f**, Time series of potential future glacial lake area (panel **e**) and lake volume (panel **f**) in HMA under different SSP values, with shading representing different subregions<sup>21</sup>. Projected global glacial lakes will increase globally by 2100 under different emission scenarios, posing a potential hazard threat to downstream communities.
+
+to 37.0 km². In the deglaciating Vilcanota–Urubamba basin, the lake area in 2100 is projected to be 3.2% higher than 2016 values under RCP2.6 (from 26.9 km² to 27.8 km²) and 6.0% higher under RCP8.5 (from 26.9 km² to 28.5 km²)  $^{145}$ . The topographic setting means that the rate of increase is higher for 2016–2050 compared with 2050–2100 as glaciers are already retreating to very steep slopes by 2050 (ref. 145). In deglaciated areas of the Peruvian Andes, 201 sites (of at least 0.01 km²) with a total volume of 0.26 km³ have been compiled for a future glacial lake inventory  $^{146}$ .
+
+Often not included in these assessments is sediment deposition within newly exposed overdeepenings, infilling potential future lakes and complicating specific projections. Such infilling is particularly likely in tectonically active, high-relief regions with abundant sediment supply such as HMA and the Southern Alps<sup>61,147</sup>. Where lakes  $receive \, meltwater \, from \, GLOFs \, as \, well \, as \, from \, ablation\text{-}fed \, meltwater \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant \, and \, constant$ flows, and where GLOF-induced bank erosion widens river channels hundreds of kilometres downstream<sup>23,78</sup>, the sedimentation rates within future lakes will also be very high. Indeed, erosion rates and sediment transport in glacierized basins are generally expected to increase owing to cryosphere degradation<sup>61,147</sup>, but there has been almost no quantitative analysis of this process. One exception is in the Swiss Alps, where approximately half of newly formed glacial lakes and 11-25% of the total overdeepening volume are expected to disappear by 2100 owing to sediment infilling<sup>23</sup>. If sediment delivery is not considered, the potential future glacial lakes are likely to be overestimated in number and size.
+
+### Future GLOF hazard and risk
+
+With projected changes in glacial lakes come potential changes in GLOFs. Based on physical understanding and supported by models, increasing GLOFs are assumed for the future 24,64,136 (Fig. 5), albeit with much regional variability. In HMA, future GLOF hazards are expected to triple from current conditions due to lake development (Fig. 6d) owing to glacial and glacial lake changes, alongside a large and exposed downstream population. Even greater increases (greater than fivefold) are projected in the Karakoram and Pamir regions<sup>24</sup>, which are expected to become a hot spot of potential GLOF hazard<sup>136</sup>, although the timing of lake formation is uncertain 148. The characteristics of GLOFs in HMA are also expected to shift. Scenario modelling in the Poiqu River basin of China and Nepal, for example, suggests that permafrost degradation will enhance large mass movements (of tens of millions of cubic metres) into glacial lakes, resulting in larger flood magnitudes and, thereby, longer distances downstream of flood impact than historically observed149.
+
+Future GLOF hazards in other regions of the world are unlikely to increase as in HMA. For example, there is relatively low potential future for GLOFs in the Mont Blanc massif (Western European Alps) owing to the flat glacier bed topography<sup>144</sup>. In the Cordillera Blanca (Peru), glacier retreat and lake formation are already at an advanced stage, and new lakes still to form will be predominantly dammed by stable bedrock<sup>73</sup>. However, in Iceland, mega GLOFs induced by volcanic activity can occur in the future, irrespective and independent of climate change<sup>89</sup>.
+
+Of course, although GLOF hazard might increase, overall risk also necessitates consideration of exposure and vulnerability. To date, comprehensive risk assessments considering the changing flood hazard and societal vulnerability and exposure have been lacking. However, the expansion of communities and infrastructure, including hydropower facilities, has been and will continue to be a major driver
+
+of GLOF risk<sup>150–152</sup>. Thus, infrastructure expansion in glacial valleys, coupled with expanded glacial lakes, will likely result in enhanced GLOF risk, threatening downstream populations and infrastructure<sup>130,153–155</sup>. Accordingly, there are new challenges for warning systems or other disaster risk reduction measures<sup>156</sup>.
+
+## **Summary and future perspectives**
+
+Global glacier melt has accelerated owing to atmospheric warming, feeding rapid growth of glacial lakes in most world regions from the 1960s onwards. Although uncertain, it is estimated that the glacial lake number and area have increased -50% from 1990 to 2018 (ref. 9), such that more than 110,000 glacial lakes with a total area of -15,000 km² in 2015–2019 have been identified. This trend in glacial lakes is associated with a potential increase in GLOF hazards and risks. More than 3,000 GLOFs have occurred worldwide from 850 to 2022, with ~88% recorded since 1900. Indeed, recorded GLOFs have increased in frequency and magnitude in most glaciated mountain regions, especially after 2000 (refs. 26,35), a trend that is likely to continue with ongoing deglaciation, increasing GLOF risk and hazard when coupled with enhanced development in downstream river valleys. In HMA, in particular, GLOF risk is projected to increase threefold with complete deglaciation<sup>24,127</sup>.
+
+Although GLOF knowledge has advanced considerably, key challenges remain, particularly surrounding data. As a starting point, homogeneous and standardized multi-temporal glacial lake inventories are needed, so far limited by mapping efficiency. Visual inspection and manual editing of lake outlines have been the main methods used to date — time-consuming approaches that rely heavily on individual expertise. The use of cloud computing, machine learning and artificial intelligence technologies should overcome manual limitations. However, access to regional high-quality lake inventories to train machine learning classifiers is vitally important. With any such inventories, consistent thresholds for minimum lake area and distance from glaciers are needed, as well as consistent classifications. An updated multi-period inventory of global glacial lakes with a single criterion is expected to fill the gaps of missing lakes and regional deficiencies.
+
+Hydrodynamic modelling and reconstruction of GLOFs is also hindered by absent data. Field data of lake properties (including bathymetry, temperature and lake outflow peak discharge) and GLOF mechanisms are desperately needed in different glaciated environments and for different lake types to refine and test model parameters, as well as parametrize dam breach models. Yet such data are extremely difficult to acquire given accessibility and safety that make successful recovery of instrumentation and sensors challenging, especially in ice-proximal locations. Accordingly, local, regional and even global understanding of GLOF flow dynamics is limited, feeding into hazard assessment, early warning systems, infrastructure design and emergency response. In addition to the use of modern field observation instruments and data transmission, broader international cooperation and the establishment of data opening and sharing are important ways forward. Future research should focus on comparing model results with field observations, historical flood records and other independent data sets to improve model credibility.
+
+Assessments of future glacier lakes under-report the common problem of uncertainty about the location, size and timing of potential future lakes. The average error in measured ice thickness of  $10\pm24\%^{157}$  masks considerably higher uncertainties for specific glaciers, including an uncertainty range of  $\pm30\%$  suggested by the validation of ground-penetrating radar measurements in the Swiss Alps<sup>129</sup>. Thus, although
+
+common ice thickness models can estimate the total ice volume of all glaciers globally<sup>20,135</sup> or regionally<sup>158,159</sup>, differences between ice thickness models can lead to substantial local differences. Crucially, the potential future lake area is highly sensitive to small errors in, for example, the height of a subglacial rock barrier or the depth of an overdeepening<sup>23</sup>. Obtaining accurate glacier bed topography by ice thickness modelling remains a challenge, but it can be improved by increasing the accuracy of the underlying input data, adding measured thickness data and multi-model synthesis<sup>20,157,160,161</sup>. In the future, with the launch of P-band and L-band radar satellites, obtaining ice thickness using tomographic SAR technology will provide new opportunities and possibilities to improve the accuracy of future glacial lake prediction<sup>162,163</sup>.
+
+These improved data and models are vital to enhance GLOF hazard assessment, particularly in a warming world associated with enhanced GLOF risk. However, GLOF hazard assessments suffer from inconsistent approaches (indicators and their thresholds) and the lack of a common assessment framework (filed survey, hazard and risk assessment and management action)<sup>110,117</sup>. Technical guidelines have been established at the international level, but have yet to be adapted or widely consolidated into regional or national approaches, as the level and development of guidelines vary widely across regions<sup>115</sup>. Where existing assessments provide conflicting results, efforts should be undertaken to consolidate findings to determine which lakes should be prioritized for monitoring and other response actions. In particular, the need to consider small lakes is emphasized, as such lakes have often been overlooked but can have devastating downstream impacts<sup>77,118</sup>. Importantly, this knowledge needs to be used to anticipate the hazards of future GLOF events, including worst-case scenarios. Given the tendency for GLOFs to be triggered by or associated with other mass movements and cascading processes, multi-hazard perspectives are encouraged.
+
+Future efforts to move beyond hazards to comprehensive GLOF risk assessment must also involve a wider network of local experts and broader disciplines. Failure to adequately consider local vulnerabilities, needs and perceptions can ultimately limit the success of disaster risk reduction measures, such as early warning systems <sup>164</sup>. As a transboundary phenomenon, GLOFs pose particular challenges for disaster risk reduction where cooperation and sharing across political boundaries are required, and where societal and institutional capacities to manage GLOF risks can vary widely between upstream and downstream regions. Scientists, policymakers and stakeholders should work together towards broader international and regional transboundary cooperation to reduce the risk of GLOFs.
+
+## **Data availability**
+
+The global compilation of glacial lakes is available at https://doi.org/10.6084/m9.figshare.25656531.
+
+

chatbot/rag/data/('Climate change and the global pattern of moraine-dammed', '.pdf')_extraction.md

@@ -0,0 +1,256 @@
+
+Abstract. Despite recent research identifying a clear anthropogenic impact on glacier recession, the effect of recent climate change on glacier-related hazards is at present unclear. Here we present the first global spatio-temporal assessment of glacial lake outburst floods (GLOFs) focusing explicitly on lake drainage following moraine dam failure. These floods occur as mountain glaciers recede and downwaste. GLOFs can have an enormous impact on downstream communities and infrastructure. Our assessment of GLOFs associated with the rapid drainage of moraine-dammed lakes provides insights into the historical trends of GLOFs and their distributions under current and future global climate change. We observe a clear global increase in GLOF frequency and their regularity around 1930, which likely represents a lagged response to post-Little Ice Age warming. Notably, we also
+
+show that GLOF frequency and regularity – rather unexpectedly – have declined in recent decades even during a time of rapid glacier recession. Although previous studies have suggested that GLOFs will increase in response to climate warming and glacier recession, our global results demonstrate that this has not yet clearly happened. From an assessment of the timing of climate forcing, lag times in glacier recession, lake formation and moraine-dam failure, we predict increased GLOF frequencies during the next decades and into the 22nd century.
+
+<sup>&</sup>lt;sup>1</sup>College of Life and Environmental Sciences, Exeter University, Exeter, UK
+
+<sup>&</sup>lt;sup>2</sup>Planetary Science Institute, Tucson, AZ 85719, USA
+
+<sup>&</sup>lt;sup>3</sup>Department of Hydrology & Atmospheric Science, University of Arizona, Tucson, AZ 85742, USA
+
+<sup>&</sup>lt;sup>4</sup>Department of Geography, University of Zurich, 8057 Zurich, Switzerland
+
+<sup>&</sup>lt;sup>5</sup>Reynolds International Ltd, Suite 2, Broncoed House, Broncoed Business Park, Wrexham Road, Mold, Flintshire, UK
+
+<sup>&</sup>lt;sup>6</sup>Water, Sediment, Hazards, and Earth-surface Dynamics Laboratory, University of Washington Tacoma, WA, 98402, USA
+
+<sup>&</sup>lt;sup>7</sup>Met Office Hadley Centre, FitzRoy Road, Exeter, Devon, UK
+
+<sup>&</sup>lt;sup>8</sup>Department of Physical Geography and Geoecology, Charles University in Prague, Faculty of Science, Albertov 6, 128 43 Prague, Czech Republic
+
+<sup>&</sup>lt;sup>9</sup>Centre for Glaciology, Department of Geography and Earth Sciences, Aberystwyth University, Wales, SY23 3DB, Aberystwyth, UK
+
+<sup>&</sup>lt;sup>10</sup>Department of Geology, University of Dayton, 300 College Park, Dayton, OH 45469-2364, USA
+
+<sup>&</sup>lt;sup>11</sup>Department of Human Dimensions of Global Change, Global Change Research Institute, Czech Academy of Sciences, Bělidla 986/4a, 60300 Brno, Czech Republic
+
+<sup>&</sup>lt;sup>12</sup>Department of Engineering Geology, Institute of Rock Structure and Mechanics, Czech Academy of Sciences,
+
+V Holešovičkách 41, 182 09 Prague 8, Czech Republic
+
+<sup>&</sup>lt;sup>13</sup>Himalayan Research Center, Lainchaur, Kathmandu, Nepal
+
+## 1 Introduction
+
+There is increasing scientific and policy interest in detecting climate change impacts and assessing the extent to which these can be attributable to anthropogenic or natural causes. As a result, recent research demonstrating an anthropogenic fingerprint on a significant proportion of recent global glacier recession is an important step forward (Marzeion et al., 2014). The focus can now shift to glacier hazards but the complex nature of glacier–climate interactions (Roe et al., 2017) and their influence on hazards makes this a challenging task (Shugar et al., 2017).
+
+Mountain glaciers have continued to recede (Kargel et al., 2014; Cramer et al., 2014) and thin from their late Holocene (Little Ice Age, LIA) positions and, in many cases, the rate of recession and thinning has increased over recent decades largely as a consequence of global warming (Marzeion et al., 2014). Thinning, flow stagnation and recession of glacier tongues have resulted in the formation of moraine-dammed lakes (Richardson and Reynolds, 2000). These moraines, some of which contain a melting ice core, are built from rock debris transported by glaciers. When they fail, large volumes of stored water can be released, producing glacial lake outburst floods (GLOFs). These floods have caused thousands of fatalities and severe impacts on downstream communities, infrastructure and long-term economic development (Mool et al., 2011; Riaz et al., 2014; Carrivick and Tweed, 2016).
+
+Although much research has been carried out on the nature and characteristics of GLOFs and hazardous lakes from many of the world's mountain regions (e.g. Lliboutry et al., 1977; Evans, 1987; O'Connor et al., 2001; Huggel et al., 2002; Bajracharya and Mool, 2009; Ives et al., 2010; Iribarren et al., 2014; Lamsal et al., 2014; Vilímek et al., 2014; Westoby et al., 2014; Perov et al., 2017), there are significant gaps in our knowledge of these phenomena at the global scale and concerning their relationship to anthropogenic climate change. Detecting changes in the magnitude, timing and frequency of glacier-related hazards over time and assessing whether changes can be related to climate forcing and glacier dynamical responses is also of considerable scientific and economic interest (Oerlemans, 2005; Stone et al., 2013). Multiple case studies are insufficient to achieve a better understanding of the mechanisms leading to GLOF initiation so a more comprehensive understanding of the global frequency and timing of GLOFs is necessary. Testing such relationships at a global scale is also an important step toward assessment of the sensitivity of geomorphological systems to climate change.
+
+Despite numerous inventories of GLOFs at regional scales (see Emmer et al., 2016), no global database has been created which focuses specifically on GLOFs relating to the failure of moraine dams. A global database is required to place GLOFs in their wider climatic context (Richardson and Reynolds, 2000; Mool et al., 2011). This means that we are unable to answer some important questions concerning
+
+their historic behaviour and therefore the changing magnitude and frequency of GLOFs globally through time, and their likely evolution under future global climate change. This latter point is made even more difficult by the lack of long-term climate data from many mountain regions. Given the size and impacts of GLOFs in many mountain regions, better understanding their links to present and future climate change is of great interest to national and regional governments, infrastructure developers and other stakeholders. We argue that glacier hazard research needs to be increasingly seen through the lens of change adaptation.
+
+These issues and knowledge gaps can be addressed via a systematic, uniform database of GLOFs. Here we have compiled an unprecedented global GLOF inventory related to the failure of moraine dams. We discuss the problems involved in developing a robust attribution argument concerning GLOFs and climate change. This inventory covers only the subset of GLOFs that are linked to overtopping or failure of moraine dams. Our focus on moraine dams is motivated by (1) this type of event leaving clear diagnostic evidence of moraine-dam failures in the form of breached end moraines and lake basins, whereas ice-dammed lake failures commonly do not leave such clear and lasting geomorphological evidence and (2) the conventional hypothetical link between climate change, glacier response, moraine-dammed lake formation and GLOF production being more straightforward compared to the range of processes driving GLOFs from ice- and bedrock-dammed lakes.
+
+Such GLOF events are often triggered by ice and rock-falls, rockslides or moraine failures into lakes, creating seiche or displacement waves, but also by heavy precipitation or ice melt/snowmelt events (Richardson and Reynolds, 2000). While climate change plays a dominant role in the recession of glaciers, downwasting glacier surfaces debuttress valley rock walls, leading to catastrophic failure in the form of rock avalanches or other types of landslides (Ballantyne, 2002; Shugar and Clague, 2011; Vilímek et al., 2014). Other climatically induced triggers of moraine dam failures include increased permafrost and glacier temperatures leading to failure of ice and rock masses into lakes and the melting of ice cores in moraine dams, which leads to moraine failure and lake drainage.
+
+Attribution of climate change impacts is an emerging research field and no attribution studies on GLOFs are available so far. Even for glaciers only very few attribution studies have been published to date (Marzeion et al., 2014; Roe et al., 2017). Follow-up studies from the IPCC 5th Assessment Report (Cramer et al., 2014) proposed a methodological procedure to attribute impacts to climate change (Stone et al., 2013). Based on that, a methodologically sound detection and attribution study needs first to formulate a hypothesis of potential impacts of climate change. In our case physical process understanding supports the association between climate change and GLOFs associated with morainedam failure by climate warming, resulting in glacier reces-
+
+sion and glacial lake formation and evolution behind moraine dams which become unstable and fail catastrophically. The next step requires a climate trend to be detected, followed by the identification of the baseline behaviour of the system in the absence of climate change. The difficulty of identifying the baseline behaviour is related to several factors. The first is the existence of confounding factors, both natural and human related. For instance, the frequency of GLOFs from moraine dams also depends on factors such as the stability of the dam, including dam geometry and material or mitigation measures such as artificial lowering of the lake level (Portocarreo-Rodriguez and the Engility Corporation, 2014). Second, there are few long-term palaeo-GLOF records with which to assess baseline behaviour. Eventually, attribution includes the detection of an observed change that is consistent with the response to the climate trend, in our case a change in GLOF occurrence and the evaluation of the contribution of climate change to the observed change in relation to confounding factors. Our chief observational result is that there is an upsurge in GLOF frequency starting around 1930 and then a decline following roughly 1975 and persisting for decades (see also Carrivick and Tweed, 2014). At face value, when comparing this with the climate records, there seems to be no relationship between global GLOF frequency and concurrent climatic fluctuations, and a regional breakdown offers no solution; for example, strong climatic global (or Northern Hemisphere) warming during the period of declining GLOF frequency after 1975 appears to be counterintuitive. A simplistic inference would be that climate change does not influence GLOF incidence, but we reject this given our understanding of the physical drivers of glacier recession, lake development and drainage mechanisms. Although we know that GLOFs involve a complex set of dynamics, one of the important dynamical changes affecting GLOFs is the formation and growth of glacial lakes, and we know that there must be a relationship here to climatic warming. GLOF triggers also commonly involve extreme weather, such as extreme heat and extreme precipitation, which are intuitively linked to climate change as well, even if the attribution experiments have not yet been carried out. We thus have to dig deeper to see how GLOF frequency may be connected to climate change. The point arises that the conditions needed for a GLOF involve a long period of lake formation and growth, such that past climate changes are involved. In the Methods section we produce a model whereby the history of one climate variable and its time derivative – Northern Hemisphere mean temperature and warming rate – are linked to the GLOF record.
+
+## 2 Methods
+
+We produced a database of GLOFs developed from a collation of regional inventories and reviews (e.g. GAPHAZ<sup>1</sup>, WGMS<sup>2</sup> and GLACIORISK<sup>3</sup> databases and the GLOF Database provided under ICL database of glacier and permafrost disasters from the University of Oslo, Reynolds and Richardson, 2000; RGSL, 1997, 2002), regional overviews and reviews (e.g. Clague et al., 1985; Xu, 1987; Costa and Schuster, 1988; Reynolds, 1992; Ding and Liu, 1992; Clague and Evans, 2000; O'Connor et al., 2001; Zapata, 2002; Raymond et al., 2003; Jiang et al., 2004; Carey, 2005; Osti and Egashira, 2009; Narama et al., 2010; Ives et al., 2010; Wang et al., 2011; Carey et al., 2012; Mergili and Schneider, 2011; Fujita et al., 2012; Iribarren et al., 2014, Emmer, 2017) and case studies of individual GLOFs (e.g. Kershaw et al., 2005; Harrison et al., 2006; Worni et al., 2012). A complete list is available in the Supplement. The GLOF database was developed from a collation of regional inventories and reviews (see Supplement). Only GLOFs that could be dated to the year and to moraine failure were included. Past temperature trends from the glacier regions of interest were extracted from three independent global temperature reconstructions (CRUTEM4.2, Jones et al., 2012, NOAA NCDC (National Oceanic and Atmospheric Administration - National Climatic Data Center), Smith et al., 2008, and NASA GISTEMP (Goddard's Institute Surface Temperature Analysis), Hansen et al., 2010). These data sets provided temperature anomaly data relative to a modern baseline beginning in 1850 for CRUTEM4.2 and 1880 for NOAA NCDC and NASA GIS-TEMP.
+
+### 2.1 Test of direct linkage between GLOF rate and climate change
+
+We concentrate exclusively on the subset of GLOFs associated with the failure of moraine-dammed lakes as these are a major hazard in many mountain regions but also represent the best candidates of outburst floods for attribution to climate change. We differentiate these from other glacially sourced outburst floods, such as those resulting from the failure of an ice dam (Walder and Costa, 1996; Tweed and Russell, 1999; Roberts et al., 2003), dam overflow, volcanically triggered jökulhlaups (Carrivick et al., 2004; Russell et al., 2010; Dunning et al., 2013) or the sudden release of water from englacial or subglacial reservoirs (Korup and Tweed, 2007).
+
+The period over which climate data are available is dependent on the region but starts in 1850 in CRUTEM4.2 and 1880 in NOAA NCDC and NASA GISTEMP. The resolution of the data is generally 5°; however, NASA GISTEMP is provided at 1° resolution but it should be noted this does
+
+<sup>&</sup>lt;sup>1</sup>http://www.gaphaz.org/
+
+<sup>&</sup>lt;sup>2</sup>http://www.wgms.ch/
+
+<sup>&</sup>lt;sup>3</sup>https://cordis.europa.eu/project/rcn/54168\_en.html
+
+not imply there are more observational data in this analysis. For each region, we extract all grid points that contain a glacier as defined in the World Glacier Inventory – Extended Format (WGI-XF). With the exception of the European Alps no data set contains a complete continuous record for the period 1900–2012. We therefore take all available data points to form time series for each data set and derive a mean linear trend for the 1990–2012 period. Given large uncertainties and data gaps no attempt is made to statistically test these trends. The trends presented here are therefore considered illustrative of past changes in temperature for these regions.
+
+### Wavelet analysis of GLOF incidence
+
+Wavelets are a commonly used tool for analysing non-stationary time series because they allow the signal to be decomposed into both time and frequency (e.g. Lane, 2007). Here, we follow the methodology of Shugar et al. (2010), although we use the Daubechies (db1) continuous wavelet. The wavelet power shown here has been tested for significance at 95 % confidence limits, and a cone of influence is applied to reduce edge effects. We follow Lane (2007) in choosing an appropriate number of scales (S = 28, see his Eq. 28), which are related to the shape of the cone of influence.
+
+### 2.2 The Earth's recent climate record smoothed along glacier response timescales: development of the GLOF lag hypothesis
+
+A potentially destructive GLOF may elapse after a glacial lake grows to a volume where a sudden release of glacial lake water can exceed a normal year's peak instantaneous discharge. There are timescales associated with the period between a climatic (or other) perturbation and the occurrence of a GLOF. The following thought experiment demonstrates the concept of the lagging responses of GLOF activity to climate change: an initialized stable condition allows glacierclimate equilibrium, where neither the climate nor the glacier has fluctuated much for some lengthy period and where no other strongly perturbing conditions exist; e.g. there are no significant supraglacial or ice-marginal or moraine-dammed lakes, and a steady state exists in the supply and removal of surface debris. We then impose a perturbation (climatic or other) which favours eventual lake development and growth and eventually a GLOF. We describe two successive time periods which must pass before a significant GLOF can occur, and then a third period before a GLOF actually occurs: lakeinception time  $(\tau_i)$ , lake growth time  $(\tau_g)$  and trigger time  $(\tau_t)$ . The first two sum to the GLOF response time  $(\tau_{GLOF})$ as we define it:  $\tau_{GLOF} = \tau_i + \tau_g$ . The terms are for illustrative purposes: many supraglacial ponds initially go through a lengthy period where they fluctuate and drain annually and thus do not have a chance to grow beyond one season. Furthermore, lakes can grow to a point where limnological processes take over from climate; hence lake growth becomes
+
+detached from climate change. Even so, our set of definitions can be used to explain the lagging responses of glacier lakes and GLOFs to climatic history.
+
+A GLOF does not necessarily occur upon climate step change date  $+\tau_{GLOF}$ , which is the timescale over which the metastable system establishes a condition where a significant GLOF could occur. A trigger is needed (e.g. a large ice or rock avalanche into the lake or a moraine collapse as an ice core melts). After a sizeable glacial lake has developed, suitable GLOF triggers may occur with a typical random interval averaging  $\tau_t$ , which depends on the topographic setting of the glacier lake, valley-side geology, steepness, moraine dam properties and climate. As a result,  $\tau_t$  could range from years to centuries. Furthermore, as a lake usually continues to grow after  $\tau_{GLOF}$  has elapsed,  $\tau_{t}$  can in principle change, probably shortening as the lake lengthens and as the damming moraine degrades. The time elapsing between a climatic perturbation and a GLOF then is the sum of three characteristic sequential periods,  $\tau_i + \tau_g + \tau_t$ .
+
+The lake inception time  $\tau_i$  might be approximated by the glacier response time, which has been defined parametrically (Jóhannesson et al., 1989; Bahr et al., 1998) but in general describes a period of adjustment toward a new equilibrium following a perturbation. We take a simple parameterization (Jóhannesson et al., 1989) and equate  $\tau_1 = h/b$ , where h is the glacier thickness of the tongue near the terminus and b is the annual balance rate magnitude. The glacier response time approximating the lake inception time may be many decades for most temperate valley glaciers, but it can range between a few years and a few centuries. The glacier response time is a climate-change-forgetting timescale. After a few response times have elapsed, a glacier's state and dynamics no longer remember the climate change that induced the response to a new equilibrium. For illustration, we adopt  $\tau_1 = 60$  years, a value typical of many temperate valley glaciers.
+
+A supraglacial pond may drain and redevelop annually (posing no significant GLOF risk), but at some point, if there is a sustained long-term negative mass balance, supraglacial ponds commonly grow, coalesce and form a water body big enough that rapid partial drainage can result in a significant GLOF. That lake growth period is defined here as  $\tau_g$ , for which we adopt 20 years, a value typical of many temperate glacier lakes of the 20th century (e.g. Wilson et al., 2018; Emmer et al., 2015) Hence,  $\tau_{GLOF} = \tau_1 + \tau_g \approx 80$  years for the favoured values. Hence, a significant GLOF may occur at any time from 80 years following a large climatic perturbation: what the GLOF waits on is  $\tau_t$ , which could be years or a century. This concept can be extended to the lagging response of a whole population of glaciers following a perturbation in regional climate (Fig. 1).
+
+We distinguish between climate change, which may establish conditions needed for a GLOF to happen, and weather, which sometimes may be involved in a GLOF trigger. GLOF triggers are diverse; e.g. protracted warm summer weather may trigger an ice avalanche into the lake or moraine melt-
+
+Image /page/4/Figure/2 description: A figure composed of six graphs, labeled (a) through (f), analyzing Northern Hemisphere temperature anomalies and warming rates over various time scales, and their connection to glacial lake outburst floods (GLOFs).
+
+Graphs (a), (b), and (c) plot the 'Northern Hemisphere temperature anomaly' on the y-axis against 'Year' on the x-axis. They show the temperature anomaly data (blue line) along with smoothed versions for tau = 20 years (red), 40 years (gray), and 80 years (black). 
+- Graph (a) spans from year 600 to 2000, showing multi-proxy data until 1849 and instrumental data from 1850 onwards. The anomaly fluctuates around 0 until a sharp increase after 1850.
+- Graph (b) focuses on the period from 1850 to 2100, showing instrumental data up to 2014 and model projections from 2015 to 2100, depicting a continuous and steep rise in temperature anomaly.
+- Graph (c) is a zoomed-in view of the period from 1940 to 2100.
+
+Graphs (d) and (e) plot the 'Warming rate, degrees yr⁻¹' on the y-axis against 'Year' on the x-axis. They show the warming rate for tau = 20 years (red line) and smoothed versions for tau = 40 years (gray) and 80 years (black).
+- Graph (d) covers the years 500 to 2000, showing fluctuations in the warming rate that become more pronounced and positive in the later period.
+- Graph (e) covers the years 1850 to 2100, showing a significant increase in the warming rate, with the model projection area shaded in light red.
+
+Graph (f) is a dual-axis plot for the 'Northern Hemisphere land' from 1904 to 2097. 
+- The left y-axis represents 'Warming rate, mK yr⁻¹' (red line and dots), ranging from -10 to 15.
+- The right y-axis represents 'GLOFs/year' (blue line), ranging from 0 to 40.
+- The plot is divided into a 'Data' section and a 'Model' section around the year 2008. It shows a correlation between the warming rate and the frequency of GLOFs, noting a '45 year glacial lake response time'.
+
+Figure 1. Reconciliation of GLOF and climate records. (a) Blue curve: composite record of Northern Hemisphere land surface temperature (merged from multi-proxy data and instrumental records, as described in the main text), plus a model of land surface temperature during the period 2015–2100. Red, grey and black curves: moving historical averages of the blue curve, as described in the text, using  $\tau_{GLOF} = 20$ , 40 and 80 years. (b, c) Close-up of the more recent periods covered in (a). (d) Warming rate extracted from the moving historical averages using  $\tau_{GLOF} = 20$ , 40 and 80 years. Periods of cooling and warming are shown with blue and red tints, respectively, using the  $\tau_{GLOF} = 80$  years curve. (e) Close-up of (d) to a more recent period. (f) Comparison of a smoothed GLOF frequency curve (red line, GLOFs/year historical moving average) with the moving historical average Northern Hemisphere temperature (black curve) using  $\tau_{GLOF} = 80$  years and shifted +45 years, where the 45-year shift is considered to be reflective of  $\tau$ , the GLOF trigger timescale. See supplement text for more description and explanation.
+
+through, or heavy winter snow may trigger an ice avalanche into the lake.
+
+However, the relevant controlling climate in this example is that of the prior climatic history and the conditioning period defined by  $\tau_{GLOF}$  and the typical trigger interval  $\tau_t$ . Hence,  $\tau_{GLOF}$  is closely connected to climate, whereas  $\tau_t$  can be connected to weather for certain types of triggers.
+
+The assessment above is for a single step-function climate change. Considering that climate changes continuously and glacier characteristics vary, populations of glaciers must have full distributions of  $\tau_i$ ,  $\tau_g$  and  $\tau_{GLOF}$ . Even while glaciers are still adjusting to any big recent historical climate change, more climate change accrues; glacier and lake dynamics take all that into account, either increasing the likelihood and perhaps size of a GLOF or decreasing or delaying it. Hence, the
+
+overall GLOF frequency record cannot be synchronous with climatic fluctuations, and it also should not simply trace past climate change with a time lag; rather, the GLOF frequency record for any large population of glaciers should be definitely but complexly related to the recent climatic history.
+
+The functional dependence on climate history is not known for any glacier or population of glaciers, but to explore the concept of a lagged GLOF response to accrued climate changes, we assert that the integration function will tend to weight recent climatic shifts more strongly than progressively older climatic shifts, the memory of which is gradually lost as the glacier population adjusts. That is, because of glacier dynamics and the responses of a population of glaciers to climatic changes, the population eventually loses memory of sufficiently older climatic changes and adjusts
+
+asymptotically toward a new equilibrium. This should be true for any climate-sensitive glacier dynamics (Oerlemans, 2005). Though we do not know the functional form of the glacier responses (either for an individual glacier or a population), we nonetheless wish to illustrate our point while not driving fully quantitative conclusions. We propose that the integration of climate information into ongoing glacier dynamical adjustments occurs with exponentially declining weighting going backward in time from any given year. The exponential time-weighting constant may be similar to  $\tau_{GLOF}$ . We have computed a moving time-average Northern Hemisphere temperature with the weighting of the average specified by an assumed  $\tau_{GLOF} = 80$  years; the computed moving average pulls data, for any year, over the preceding period of  $\tau_{GLOF}$ ; i.e. it includes temperature information up to 240 years prior to any given year. The weighting of earlier years' temperatures within that  $\tau_{GLOF}$  is less than that of later years, according to the exponential. The cut-off at  $\tau_{GLOF}$  is arbitrary and was done for computational expediency, seeing that any climate fluctuation occurring before  $\tau_{GLOF}$  years earlier is inconsequential due to the exponential memory loss.
+
+We combined the Mann et al. (2008) multi-proxy Northern Hemisphere temperature anomaly from 501 AD to 1849, the Jones et al. (2012)<sup>4</sup> Northern Hemisphere land instrumental temperature record from 1850 to 2014 and a model of expected warming from 2015 to 2100. It is the recent climate history at each glacier lake or region that is strictly relevant, but lacking such records and needing here to only establish the concept, we settle for the treatment described above involving the Northern Hemisphere temperature anomaly.
+
+The model is a constant  $2.7\,^{\circ}\text{C}$  century<sup>-1</sup> warming; noise was added from a naturally noisy but overall non-trending instrumental record from 1850 to 1899, with some years repeated to append the 2015–2100 period (Fig. 1). The Mann et al. (2008) and Jones et al. (2012) data sets were brought into congruence in 1850. Then we smoothed the composite record and model results using the  $\tau_{\text{GLOF}}$  exponentially weighted filter, as described above, where the natural logarithmic "forgetting" timescale  $\tau_{\text{GLOF}} = 20$ , 40 or 80 years for three illustrative cases. Smoothing was computed for  $\tau_{\text{GLOF}}$ , i.e. 240 years if  $\tau_{\text{GLOF}} = 80$  years. Our favoured value  $\tau_{\text{GLOF}} = 80$  years is based on large Himalayan and other temperate glacier lakes. The shorter response times would likely apply to small glaciers or those occurring in steep valleys.
+
+Regardless of the functional form of the glacier response and lake dynamics, GLOF frequency in any given region or worldwide must lag the climate record. The historically filtered/smoothed temperature record and model incorporating  $\tau_{\rm GLOF} = 20$ , 40 and 80 years is shown in Fig. 1a–c together with the unsmoothed actual record and model temperature series. The temperature anomalies are plotted in panels (a), (b) and (c); and the warming rate in panels (d) and (e).
+
+The historically averaged/smoothed temperature record lags fluctuations in the unsmoothed record. The lag is most easily seen where temperatures start to rise rapidly in the 20th and 21st centuries. The high-frequency temperature anomaly fluctuations also show concordantly but in damped form in the smoothed moving average curves because the curves are historical moving averages with the heaviest weighting toward the more recent years. The lagging responses are also seen at several times when the running average curves variously show warming and cooling for the same year depending on the value of  $\tau_{\rm GLOF}$ .
+
+We posit that the historically filtered warming rate (more than the temperature anomaly) drives GLOF frequency. In Fig. 1 we show GLOF frequency (smoothed over 10-year moving averages) together with the warming rate extracted from the historically filtered temperature and model temperature time series. To get a better match with the temperature treated as such, we applied a further 45-year shift. From a glacier and lake dynamics perspective, this shift might relate to the trigger timescale,  $\tau_t$ . Singular values of  $\tau_{GLOF}$ and  $\tau_t$  should not pertain globally to all glaciers but should span wide ranges. The adopted values  $\tau_{GLOF} = 80$  years and  $\tau_t = 45$  years nonetheless make for a plausible match between the GLOF and climate records. These numbers make sense in terms of glacier and lake dynamics timescales, but we reiterate that our purpose with this climate-GLOF fitting exercise is illustrative. In sum, a notable shift in GLOF frequency does not connote a concordant shift in climate, though prior climate change may still underlie the cause.
+
+## 3 Results
+
+Our global analysis identifies 165 moraine-dam GLOFs, recorded since the beginning of the 19th century (Fig. 2a). The vast majority of these GLOFs (n = 160; 97 %) occurred since the beginning of the 20th century, at a time of climate warming and increasing glacier recession (Figs. 2 and 5). None of these GLOFs were associated with repeat events from the same lake. Around 65 % of GLOFs occurred between 1930 and 1990. Thirty-six GLOFs occurred in the mountains of western North America between 1929 and 2002 (Table S1 in the Supplement). Fifteen of these occurred in western Canada, 15 in the Cascade Range of the USA and four in Alaska. One occurred in Mexico and 1 in the Sierra Nevada. In the South American Andes we identified 40 GLOFs. Eleven occurred in Chile between 1913 and 2009 (including the large one in Patagonia at Laguna del Cerro Largo in 1989); one in Colombia in 1995 and 28 in Peru between 1702 and 1998. Fourteen GLOFs are listed from the European Alps. Three are from Austria between 1890 and 1940, five from Switzerland between 1958 and 1993 one from France in 1944 and five from Italy between 1870 and 1993. In the Pamir and Tien Shan mountains in central Asia, we identified 20 GLOFs, with most of these dating from the
+
+<sup>&</sup>lt;sup>4</sup>https://crudata.uea.ac.uk/cru/data/temperature/\T1\textbackslash#datdow
+
+Image /page/6/Figure/2 description: A scientific figure composed of multiple panels, labeled (a) through (f), analyzing moraine-dammed glacial lake outburst floods (GLOFs) across different regions. Each row represents a region: (a) Global, (b) European Alps, (c) Pamir & Tien Shan, (d) Hindu Kush Himalayala, (e) Alaska & Canadian Rockies, and (f) South American Andes.
+
+For each region, the left panel plots data from 1860 to 2000. It shows the cumulative sum of events as a black line (left y-axis), the number of events each year as blue bars (inner left y-axis), and the discharge in 10^6 m^3 as red dots (right y-axis).
+
+For regions (b) through (f), the right side contains three smaller plots showing regional temperature anomalies (°C) from 1850 to 2011 from three datasets: CRUTEM4.2 (blue), NOAA NCDC (green), and NASA GISTEMP (yellow). Each of these plots shows a linear trend line and its value in °C per 10 years. A grey shaded area marks the period from 1970 to 2011.
+
+Key data points for each region:
+- (a) Global: The left plot shows a cumulative sum of events approaching 200 by 2000, with a notable increase in event frequency after 1940. Discharge values reach up to 250 x 10^6 m^3.
+- (b) European Alps: Cumulative events reach about 17. Temperature trends are: CRUTEM4.2: 0.148 C/10 yr, NOAA NCDC: 0.108 C/10 yr, NASA GISTEMP: 0.125 C/10 yr.
+- (c) Pamir & Tien Shan: Cumulative events reach about 22. Temperature trends are: CRUTEM4.2: 0.149 C/10 yr, NOAA NCDC: 0.128 C/10 yr, NASA GISTEMP: 0.133 C/10 yr.
+- (d) Hindu Kush Himalayala: Cumulative events reach about 45. Temperature trends are: CRUTEM4.2: 0.056 C/10 yr, NOAA NCDC: 0.095 C/10 yr, NASA GISTEMP: 0.109 C/10 yr.
+- (e) Alaska & Canadian Rockies: Cumulative events reach about 70. Temperature trends are: CRUTEM4.2: 0.108 C/10 yr, NOAA NCDC: 0.101 C/10 yr, NASA GISTEMP: 0.099 C/10 yr.
+- (f) South American Andes: Cumulative events reach about 38. The discharge axis is logarithmic. Temperature trends are: CRUTEM4.2: 0.047 C/10 yr, NOAA NCDC: 0.122 C/10 yr, NASA GISTEMP: 0.158 C/10 yr.
+
+**Figure 2.** (a–f) Left: temporal distribution of regional GLOF frequency and magnitude. At all locations, the cumulative sum of events (black line) indicates an upsurge in the number of events per year. The timing of this upsurge differs by location and likely reflects an increase in reporting, especially in the early part of the record, rather than a change in GLOFs, at least until the 1970s–1990s after which the GLOF rate reduces. Right: global time series climate data from the five regions using CRUTEM 4.2, NOAA NCDC and NASA GISTEMP. Grey columns represent the baseline against which temperature is measured.
+
+Image /page/7/Figure/2 description: The image displays three vertically stacked plots, labeled (a), (b), and (c), all sharing a common horizontal axis representing the 'Year' from 1860 to 2010.
+
+Plot (a) is a bar chart with the vertical axis labeled 'Count', ranging from 0 to 8. The chart shows the frequency of an event over the years. Before 1930, the counts are very low, mostly 0 or 1. After 1930, the frequency increases significantly, with numerous years having counts between 1 and 4. The highest count of 8 occurs around the year 1960.
+
+Plot (b) is a wavelet power spectrum. The vertical axis, labeled 'Period (y)', is on a logarithmic scale, primarily showing values between 2 and 10^1 (10). A color bar on the left indicates the 'Wavelet power (10^3 W m^-2)', ranging from 0 (dark blue) to 6 (dark red). The plot shows areas of high power (red and yellow) concentrated in two main periods. The first is from approximately 1920 to 1960 at periods of 2 to 5 years. The second area of high power is from roughly 1955 to 1980 at periods of 5 to 8 years. Other smaller, less intense (blue and green) areas are scattered after 1960.
+
+Plot (c) is a line graph. The vertical axis, labeled 'Wavelet power (Wm^-2)', is on a logarithmic scale, ranging from below 10^0 to 2. The line shows a low power level before 1920, followed by a sharp increase to a peak around 1935-1940. The power remains high with significant fluctuations until about 2000, with another notable peak around 1990, before declining towards 2010.
+
+**Figure 3.** (a) Record of all precisely dated GLOFs from 1860 to 2011. (b) Wavelet power spectrum of global GLOF record, significant at 5 %. (c) Frequency-integrated wavelet power spectrum.
+
+late 1960s to the early 1980s. The largest number of GLOFs (55) is reported from the Hindu Kush Himalaya (HKH) including the mountains of Bhutan and Tibet, dated from the 20th and 21st century. Thirty are from Tibet (between 1902 and 2009), 12 from Nepal between 1964 and 2011(and one is reported to have occurred in 1543) and five in Pakistan between 1878 and 1974. There is uncertainty in reporting some of these GLOFs and we discuss this further in the Supplement.
+
+From around 1930 to about 1950, GLOFs occurred with regularity but a low frequency (Fig. 3). In other words, floods occurred with relatively long period variabil-
+
+ity (50–60 years). Starting around 1960, the frequency of these events increased (period decreased to approximately 20 years), remaining relatively high until about 1975, after which the statistically significant periodicities end, though GLOFs continue to occur.
+
+While incomplete data restrict a full analysis of GLOF triggers, precise date, magnitude and initiation at a global scale, many GLOFS triggered by ice avalanches and rockfalls occur during summer (see Fig. 4). The characteristics of GLOFs that could be influenced by climate change include changes in magnitude, frequency, timing (either changes in seasonality or changes over longer timescales) and trigger
+
+Image /page/8/Figure/2 description: A 3D bar chart showing the frequency of different types of failures across four seasons. The vertical axis is numbered from 0 to 35. The horizontal axis lists the categories: Total, Ice avalanche, Rock slope failure, Heavy rainfall, and Moraine failure. The legend indicates the seasons by color: blue for Spring, red for Summer, green for Autumn, and purple for Winter. The data presented is as follows: For the 'Total' category, Spring has a value of 2, Summer is 35, Autumn is 12, and Winter is 1. For 'Ice avalanche', Spring is 2, Summer is 27, Autumn is 5, and Winter is 1. For 'Rock slope failure', Spring is 0, Summer is 2, Autumn is 2, and Winter is 0. For 'Heavy rainfall', Spring is 0, Summer is 3, Autumn is 3, and Winter is 0. For 'Moraine failure', Spring is 0, Summer is 3, Autumn is 2, and Winter is 0.
+
+**Figure 4.** Seasonal variation in the occurrence of GLOF associated with the failure of moraine dams. Only a proportion of the GLOFs have seasonal data on timings.
+
+mechanisms. In addition, many rock avalanches into lakes triggering a GLOF may represent a paraglacial response to deglaciation from the LIA or earlier times (Knight and Harrison, 2013; Schaub et al., 2013) and this delayed response demonstrates the need to account for lags between changes in forcing and responses in attribution studies.
+
+## 4 Discussion
+
+From this analysis, we highlight three key observations: (1) GLOFs became more common around 1930 but then their incidence was maintained at a quasi-steady level for a few decades thereafter; (2) since about 1975, GLOF periodicity has decreased globally; and (3) the periodicities of GLOF occurrence have changed throughout the 20th century. These observations are discussed below.
+
+Our first main observation is that GLOF frequency increased dramatically and significantly around 1930 globally and between 1930 and 1960 regionally (Figs. 1 and 2). We find no obvious reason for an abrupt improvement of GLOF reporting in 1930. While acknowledging that incompleteness of the record must be a pervasive factor throughout the early period covered by the database, we discount reporting variations as the cause of the abrupt shift. For instance, this pattern is observed in the European Alps, a region with a long history of mountaineering, glacier research and valley-floor habitation and infrastructure development. Given that we record individual GLOFs in the 19th and early 20th centuries we argue that the increase in GLOF frequency in the 1930s represents a real increase rather than an observational artefact. Following the increase around 1930, we observe a similar rate of GLOFs for the subsequent years, typically 1 per year in the following decade, increasing to 2-3 per year during the 1940s (e.g. Figs. 1a, 2a). Again, there is no evidence that incompleteness of data is a main cause of the observed pattern. We therefore conclude that the incidence of global GLOFs has remained generally constant between about 1940 and about 1960. In the 1960s and early 1970s, several years
+
+Image /page/8/Figure/8 description: A figure displaying five maps that show temperature anomalies in different mountain ranges, based on CRUTEM4.2 data for the period 1991–2012 relative to 1901–1920. A color scale at the bottom indicates the temperature anomaly in degrees Celsius, ranging from light orange (0 °C) to dark red (2.1 °C).
+
+The five maps are:
+1. \*\*European Alps:\*\* Shows a high temperature anomaly, with the region colored in dark red (approximately 1.5 to 2.1 °C).
+2. \*\*Pamir and Tien Shan:\*\* Shows moderate to high anomalies, with colors ranging from orange to red (approximately 0.6 to 1.5 °C).
+3. \*\*Hindu Kush Himalaya:\*\* Shows lower temperature anomalies, with the region colored in light to medium orange (approximately 0.3 to 0.9 °C).
+4. \*\*Alaska and Canadian Rockies:\*\* Shows a high temperature anomaly, with the region colored in dark red (approximately 1.5 to 2.1 °C).
+5. \*\*South American Andes:\*\* Shows moderate anomalies, with the region colored in shades of orange (approximately 0.6 to 1.2 °C).
+
+Each map includes latitude and longitude lines, and the mountain ranges are indicated by clusters of black plus signs.
+
+**Figure 5.** Temperature anomalies in the CRUTEM4.2 data set for each mountain region. For each region we extract all the grid points that contain a glacier as defined in the World Glacier Inventory – Extended Format (WGI-XF) and these are shown as black crosses.
+
+saw more than five GLOFs. We argue below that the trend between 1940 and 1960 hides a more complex spatial and temporal pattern (Clague and Evans, 2000; Schneider et al., 2014).
+
+Our second main observation is that, while there is considerable variability between regions, GLOF incidence rates have decreased since about 1975 globally (Fig. 2). There are both more and larger GLOFs during the 1970s and early 1980s in the Pamir and Tien Shan, in the 1960s in the HKH and in the 1990s in Alaska, the Coast Mountains and Canadian Rockies; and then decreases in both magnitude and frequency follow these periods. In the Andes however, GLOF incidence decreased after the early 1950s. The latter observation may be at least partly attributable to considerable GLOF mitigation measures in Peru, such as engineeringbased lake drainage or dam stabilization (Carey et al., 2012; Portocarreo-Rodriguez, 2014). Carrivick and Tweed (2014) propose several reasons why "glacial floods" may have decreased in frequency in recent decades. These include successful efforts to stabilize moraine dams and changes in the ability of fluvial systems to transmit floods over time. We argue, conversely, that this reduction may represent a "lagged" response to glacier perturbations following climate change. More research is clearly needed on this question, and we believe that our analysis, along with that of Carrivick and Tweed's, will stimulate further work and discussion.
+
+Our third main observation is that for several decades in the 20th century, GLOF occurrence has been periodic, but that periodicity has varied. Since about 1975, and especially since 1990, the periodic nature of GLOF occurrence has diminished, even though GLOFs have continued. In other words, GLOFs since 1975 have become more irregular. We suspect that the switch to less periodic outburst floods in recent decades is related to an underlying mechanism such as topographic constraints and glacier hypsometries, with glaciers retreating into steeper slopes, implying a reduced rate of moraine-dammed lake formation – a phenomenon observed, for example, in the European Alps (Emmer et al., 2015).
+
+The statistics of small numbers affect these regional, timeresolved records, but the overall validity of a similar mid-20th century increase and then decrease in the frequency of GLOFs can be further detected in the global record and is statistically significant (Fig. 3). We argue that the reduction in global GLOF frequency after the 1970s (especially in central Asia, HKH and North America) is real, because the contemporary reporting is likely to be nearly complete given the scientific and policy interest in glacier hazards from the late 20th century. Hence, our conclusion is that globally and regionally there have been interdecadal variations in the frequency of GLOFs, and in general the most recent couple of decades have seen fewer GLOFs than the early 1950s to early 1990s. The record's (in)completeness is not able to explain a decreasing incidence rate. This temporal variation in GLOF frequency and recent decrease is therefore a robust and surprising result and has occurred despite the clear trend of con-tinued glacier recession and glacier lake development in re-cent decades.
+
+Our data allow us to test and refine the widespread assumption that GLOFs are a consequence of recent climate change (Bajracharya and Mool, 2009; Riaz et al., 2014). This is an important assumption because it implies that GLOF frequency will increase as the global climate continues to warm with potential major impacts for downstream regions.
+
+The global increase in GLOF frequency after 1930 must be a response to a global forcing, considering global glacier retreat (Zemp et al., 2015), and physical process understanding suggests that this is a lagged response to the warming marking the end of the LIA (Clague and Evans, 2000). Although the global response appears sudden, in 1930 the region-by-region assessment shows that the response was asynchronous regionally and temporally over three decades (Fig. 2). This is consistent with the fact that the end of the LIA was not globally synchronous (Mann et al., 2009) and we also argue that this reflects regional variations in glacier response times.
+
+We argue that as a climate shift occurs, after some period related to the glacier response time, previously stable or advancing glaciers start to thin and recede; after a further *limnological response time* proglacial ponds start to grow, coalesce and deepen into substantial moraine-dammed lakes. GLOFs typically occur after some additional period of time (the *GLOF response timescale*), but this time can be brief in glaciers with short response times, such as in the tropical Andes (Fig. 1).
+
+In the HKH and central Asia the near-concordant formation of many Himalayan glacier lakes and the abrupt increase in GLOF rates in the 1950s and 1960s suggest that the GLOF response time is much smaller than the limnological response time. The moraine evidence here indicates that a shift from mainly glacier advance to recession and/or thinning occurred widely, though regionally asynchronously, between 1860 and 1910. The HKH underwent this shift by around 1860 (Owen, 2009; Solomina et al., 2015) in response to warming following the regional LIA. The limnological response time in the Himalayan–Karakoram region is thus around 100 years, i.e. substantially longer than in the tropical Andes.
+
+We have arrived at a plausible explanation for the post-1930 (1930–1960) increases in GLOF rates. They are most likely heterogeneous, lagging responses to the termination of the LIA, with limnological response times of the order of decades to 100 years, depending on the region (e.g. Emmer et al., 2015). The limnological response times may be of a similar order to the glacier dynamical response times (Jóhannesson et al., 1989; Raper and Braithwaite, 2009) but are appended to them. Thus, measured from a climatic shift to increased GLOFs, the combined glaciological and limnological response times (plus GLOF response times, which may be the shortest of the three response times) may sum to roughly 45–200 years (Fig. 1). It cannot be much more
+
+than this, because then we would not see the multi-decadal oscillations in GLOF rates in some regions or globally.
+
+Some individual glaciers may have faster response times than estimated above (Roe et al., 2017), but taken on a broader statistical basis we infer that the most recent GLOFs are a delayed response to the end of the LIA. A fundamental implication is that anthropogenic climatic warming to date will likely manifest in increasing GLOFs in some regions of the world starting early this century and continuing into the 22nd century. In all the mountain regions considered here the available evidence indicates a warming trend over the last century around 0.1 °C decade<sup>-1</sup> (Figs. 2 and 5). The trend varies between data set and region, with the highest rates in the Pamir Tien Shan region and the lowest in the HKH. The most uncertain region is the Andes, where the sparseness of data prevents any meaningful assessment. The trends are consistent with the global mean land temperature trend  $0.95 \pm 0.02$  to  $0.11 \pm 0.02$  °C for 1901–2012, implying these regions have warmed at approximately the same rate as the global land surface.
+
+The baseline behaviour of glacial lake systems in the absence of climate change is not known in detail, but the low rate of GLOFs prior to 1930 may indicate that without warming the frequency would be low. The difficulty of attributing individual GLOF behaviour to climate change relates to the presence of non-climatic factors affecting GLOF behaviour, such as moraine dam geometry and sedimentology, climate-independent GLOF triggers (e.g. earthquakes) and the timescales related to destabilization of mountain slopes, producing mass movements into lakes (Haeberli et al., 2017). This represents the period of paraglaciation (e.g. Ballantyne, 2002; Holm et al., 2004; Knight and Harrison, 2013). These system characteristics may vary regionally and temporally within the evolutionary stage of a receding mountain glacier, and non-climatic factors such as lake mitigation measures additionally influence GLOF frequency and magnitude (Clague and Evans, 2000; Portocarreo-Rodriguez, 2014). We argue that while the original driver of lake development is likely to involve climate change (resulting in glacier downwasting and slowed meltwater flux through glaciers systems as glacier surfaces reduce in gradient) other mechanical and thermodynamic processes likely assume more importance as the lakes evolve, and these includes small-scale calving and insolationinduced melting of ice cliffs (e.g. Watson et al., 2017).
+
+We also recognize that contemporary mountain glaciers are dissimilar to those that existed in the LIA. They are, in the main, shorter, thinner and with prominent moraines. Assumptions that climate processes acted on similar glacial systems over time are therefore likely to be simplistic.
+
+Based on the analysis of our global GLOF database we have shown that a clear trend is detectable globally and regionally diversified in the 20th century with a sharp increase in GLOF occurrence around 1930. This trend is attributable to the observed climate trend, namely the warming since the end of the LIA. The delayed response of GLOF occurrence
+
+is an exemplar for the complexities of how natural systems respond to climate change, underlining the challenges of attribution of climate change impacts. We have shown here that attribution of GLOFs to climate change is possible, although the suite of factors influencing GLOF occurrence cannot be fully quantified.
+
+In addition, lake outbursts following moraine failures are likely to be quite different in different regions. This reflects differences in a number of factors including ground thermal conditions, presence or absence of ground ice and permafrost, influence of extreme weather and seismic processes, topography and glacial history. To assess these we would need to better understand the geomorphological timescales involved in lake evolution and failure to design a more robust statistical analysis and to understand each region's GLOF history. We thus recommend close attention by the Earth surface processes science community to various process timescales using field studies, satellite remote sensing and theoretical modelling.
+
+Our inventory and the global pattern of GLOFs that is derived from it lacks in many cases precise data on the processes responsible for GLOFs. This is a consequence of incomplete reporting of GLOFs in remote mountain regions, especially before the advent and wide use of remote sensing. In many cases the record is of a large flood being observed and then some time afterwards a collapsed moraine dam is seen and the flood is attributed to this collapse. Clearly the precise details of how the collapse occurred is not always available, and this uncertainty bedevils all similar detection and attribution studies, especially on those events associated with rapid geomorphological change. This intrinsic incompleteness in the record is problematic but should not prevent reasonable assertions on GLOF triggers to be made, especially if global-scale and consistent patterns in GLOF behaviour are observed.
+
+Future research should therefore more systematically study the factors influencing GLOF frequency and magnitude and lake formation where a distinction between GLOF conditioning and triggering factors will be helpful (e.g. Gardelle et al., 2011).
+
+If climate (such as temperature time series) influences GLOFs, as surely must be the case, long lag times are necessarily implied by the empirical data sets. With such lags as we have modelled, this brings the increase of GLOFs following 1930 into line with temperature increases at the end of the Little Ice Age. Subsequent changes in the GLOF rate (including a several-decades-long fall in GLOF rates) can similarly be attributed to fluctuations in global warming. If these conclusions are broadly correct, a further implication is that an acceleration in GLOF rates will probably occur in the 21st century, perhaps starting rather soon. Even though the actual global warming rate for the 21st century may be nearly constant, as modelled, the fitted warming rate as plotted in Fig. 1f accelerates because of the memory of a post-LIA, pre-anthropogenic quasi-stable climate. We are enter-
+
+ing a stage where anthropogenic warming will increasingly dominate GLOF activity and attribution of GLOFs to anthropogenic global warming will be confirmed. For now, this remains a hypothetical projection or expectation and is not yet borne out in the GLOF record.
+
+## 5 Conclusions
+
+We conclude that the global record of GLOF following the failure of moraine dams shows a dramatic increase in GLOF occurrences from 1930 to 1970, then a decline. We also observe that the GLOF frequency has not fluctuated directly in response to global climate. A reasonable premise is that climate, glaciers, glacier lakes and GLOFs are closely connected, but the connections between climate and GLOFs are hidden in response time dynamics. We argue that response times do not necessarily reflect linear processes and that lake growth may result in none, single or multiple GLOFs from the same lake systems. Accordingly, the response times must vary widely from region to region and glacier to glacier. From this we infer that the 1930 to 1970 increase in global GLOF activity is likely a delayed response following warming that ended the LIA and decreased the rate of morainedammed lake formation. We also infer that the decrease in GLOF frequency after 1970 is likely related to a delayed response to the stabilization of climate following the LIA. In addition, a minor cause (though important locally, for instance in Peru and Switzerland in particular), GLOF mitigation engineering, may have circumvented a few GLOFs, thus contributing to the downward trend in recent decades. We can expect a substantial increase in GLOF incidence throughout the 21st century as glaciers and lakes respond more dynamically to anthropogenic climate warming. This is corroborated by recent modelling studies projecting the location, number and dimension of new lakes in areas where glacier will recede over the coming decades in the Alps, the Himalayas or the Andes (Linsbauer et al., 2016).
+
+As a result, we argue that the sharply increased GLOF rates starting from 1930 followed by reduced GLOF frequency from high levels in the mid-20th century are both real and we speculate these trends may reflect the failure of sensitive glacial lake systems in a lagged response to initial glacier recession from LIA limits. The apparent robustness of contemporary lake systems suggests that only the most resilient moraine-dammed lakes have survived recent climate change. Predicting their future behaviour is of great importance for those living and working in mountain communities and those developing and planning infrastructure in such regions.
+
+Data availability. The database of GLOFs analysed in this paper is attached as a supplement.
+
+*Supplement.* The supplement related to this article is available online at: https://doi.org/10.5194/tc-12-1195-2018-supplement.
+
+Author contributions. The project was designed by SH following discussions with JK, CH and JR. Climate model data were provided by AW and RAB. Data analysis was carried out by SH, JK, DHS, LR and UH. JR, VV and AE provided inventory data. All authors helped to write and review the text.
+
+Competing interests. The authors declare that they have no conflict of interest.
+
+Acknowledgements. Stephan Harrison was funded by a Leverhulme Research Fellowship. Stephan Harrison, Richard A. Betts and Andy Wiltshire acknowledge funding under the HELIX (European Union Seventh Framework Programme FP7/2007-2013 under grant agreement no. 603864). Andy Wiltshire and Richard A. Betts acknowledge funding from the Joint UK DECC/Defra Met Office Hadley Centre Climate Programme (GA01101). John Kennedy of the Met Office Hadley Centre provided advice on handling the temperature observation data sets used in this project. Contributions by Jeffrey S. Kargel, Umesh K. Haritashya, Dan H. Shugar, and Dhananjay Regmi were supported by NASA's Understanding Changes in High Mountain Asia programme, the NASA/USAID SERVIR Applied Science Team programme, and by the United Nations Development Program. We thank C. Scott Watson and an anonymous reviewer for their detailed and incisive reviews of the paper. We also thank Georg Veh, Jonathan Carrivik and Sergey Chernomorets for further comments and clarifications of the inventory.
+
+Edited by: Chris R. Stokes
+
+Reviewed by: C. Scott Watson and one anonymous referee
+
+## References
+

chatbot/rag/data/('Crisis_and_Disaster_Situations_on_Social_Media_Str', '.pdf')_extraction.md

@@ -0,0 +1,485 @@
+
+
+# ABSTRACT
+
+Aim/Purpose Vis-à-vis management of crisis and disaster situations, this paper focuses on
+
+important use cases of social media functions, such as information collection & dissemination, disaster event identification & monitoring, collaborative problem-solving mechanism, and decision-making process. With the prolific utilization of disaster-based ontological framework, a strong disambiguation system is realized, which further enhances the searching capabilities of the
+
+user request and provides a solution of unambiguous in nature.
+
+Background Even though social media is information-rich, it has created a challenge for
+
+deriving a decision in critical crisis-related cases. In order to make the whole process effective and avail quality decision making, sufficiently clear semantics of such information is necessary, which can be supplemented through
+
+employing semantic web technologies.
+
+Methodology This paper evolves a disaster ontology-based system availing a framework
+
+model for monitoring uses of social media during risk and crisis-related events. The proposed system monitors a discussion thread discovering whether it has reached its peak or decline after its root in the social forum like Twitter. The content in social media can be accessed through two typical ways: Search Application Program Interfaces (APIs) and Streaming APIs. These two kinds of API processes can be used interchangeably. News content may be filtered by time, geographical region, keyword occurrence and
+
+Accepting Editor Iris A Humala | Received: June 23, 2019 | Revised: August 19, August 21, 2019 | Accepted: August 22, 2019.
+
+Cite as: Narayanasamy, S., Muruganantham. D. & Elçi, A. (2019). Crisis and disaster situations on social media streams: An ontology-based knowledge harvesting approach. *Interdisciplinary Journal of Information, Knowledge, and Management*, 14, 343-366. https://doi.org/10.28945/4420
+
+(CC BY-NC 4.0) This article is licensed to you under a <u>Creative Commons Attribution-NonCommercial 4.0 International License</u>. When you copy and redistribute this paper in full or in part, you need to provide proper attribution to it to ensure that others can later locate this work (and to ensure that others do not accuse you of plagiarism). You may (and we encourage you to) adapt, remix, transform, and build upon the material for any non-commercial purposes. This license does not permit you to use this material for commercial purposes.
+
+<sup>\*</sup> Corresponding author
+
+availability ratio. With the support of disaster ontology, domain knowledge extraction and comparison against all possible concepts are availed. Besides, the proposed method makes use of SPARQL to disambiguate the query and yield the results which produce high precision.
+
+Contribution
+
+The model provides for the collection of crisis-related temporal data and decision making through semantic mapping of entities over concepts in a disaster ontology we developed, thereby disambiguating potential named entities. Results of empirical testing and analysis indicate that the proposed model outperforms similar other models.
+
+Findings
+
+Crucial findings of this research lie in three aspects: (1) Twitter streams and conventional news media tend to offer almost similar types of news coverage for a specified event, but the rate of distribution among topics/categories differs. (2) On specific events such as disaster, crisis or any emergency situations, the volume of information that has been accumulated between the two news media stands divergent and filtering the most potential information poses a challenging task. (3) Relational mapping/co-occurrence of terms has been well designed for conventional news media, but due to shortness and sparseness of tweets, there remains a bottleneck for researchers.
+
+Recommendations for Practitioners
+
+Though metadata avails collaborative details of news content and it has been conventionally used in many areas like information retrieval, natural language processing, and pattern recognition, there is still a lack of fulfillment in semantic aspects of data. Hence, the pervasive use of ontology is highly suggested that build semantic-oriented metadata for concept-based modeling, information flow searching and knowledge exchange.
+
+Recommendation for Researchers
+
+The strong recommendation for researchers is that instead of heavily relying on conventional Information Retrieval (IR) systems, one can focus more on ontology for improving the accuracy rate and thereby reducing ambiguous terms persisting in the result sets. In order to harness the potential information to derive the hidden facts, this research recommends clustering the information from diverse sources rather than pruning a single news source. It is advisable to use a domain ontology to segregate the entities which pose ambiguity over other candidate sets thus strengthening the outcome.
+
+Impact on Society
+
+The objective of this research is to provide informative summarization of happenings such as crisis, disaster, emergency and havoc-based situations in the real world. A system is proposed which provides the summarized views of such happenings and corroborates the news by interrelating with one another. Its major task is to monitor the events which are very booming and deemed important from a crowd's perspective.
+
+Future Research
+
+In the future, one shall strive to help to summarize and to visualize the potential information which is ranked high by the model.
+
+Keywords
+
+disaster management, social media, ontological support, semantic search,
+
+SPARQL, RDF
+
+# INTRODUCTION
+
+For long, there has been a huge demand to develop an efficient mechanism to effectively search and extract much-needed information from the social web. Manual annotation is effectively possible in information retrieval for a limited number of documents, but impractical for a large accumulation of content, particularly in social media. And moreover, automatic annotation processes are in an infant
+
+stage. As Ritter, Etzioni and Clark (2012) indicated, automatic annotation has not reached its complete stage. There would be deemed requirements to properly utilize ontologies to precisely govern the types of knowledge to harvest. With the support of ontology (Sakaki, Okazaki, & Matsuo, 2010), domain knowledge extraction is very relevant and relates all possible concepts. Ontology-based knowledge extraction is expected to provide a boost to the domain of this research, disaster management.
+
+Hence, pervasive use of ontology has been highly suggested (Ha-Thuc, Mejova, Harris, & Srinivasan, 2010; Luo, Osborne, & Wang, 2012) that builds semantic oriented metadata for concept-based modeling, information flow searching and knowledge exchanging. Once the semantic aspect of metadata is built for the content available over the Web on a domain of interest, then it will provide common grounds for understanding and sharing the information, as well as increasing relevancy and reducing, perhaps even minimizing inherent ambiguities. In the past, there were projects aimed at scavenging Web content with exemplary results for semantically annotating the metadata for domain-specific semantic searches. Several noteworthy examples may be mentioned; for instance, systems like PlanetOnto, ArtEquAKT, sparse kernel learning continuous relevance model for image annotation (Moran & Lavrenko, 2014), and integration of linked data in Knowledge-Based Systems for process planning (Rehage, Joppen, & Gausemeier, 2016). However, when it comes to working with the fast transient contents of social networks, a different paradigm is needed where this research comes in to contribute.
+
+The next potential task of the operation is accessing the news contents from social media platforms. Though the task seems simple, it is inherently complicated due to interoperability conundrum and accessibility disabilities. Most of the social media platforms have accorded privileges to the programmers accessing the content through its appropriate APIs, but they differ from one another in what they provide and are mostly resource-limited (i.e., in the sense of the number of allowed request for a unit time). The content in the social media can be accessed through two typical ways: permitting users to access and archive the past messages, it is called Search APIs; and allowing users to subscribe for the real-time data feeds, it is known as Streaming APIs. These two kinds of API processes can be used interchangeably and allow expressing information needs, such as filtering the news content by time, geographical region, keyword occurrence and availability ratio (Celik, Abel, & Houben, 2011; Raman, Kuppusamy, Dorasamy, & Nair, 2014). The harvested data however still requires further pre-processing.
+
+# DATA PRE-PROCESSING AND NORMALIZATION
+
+Though we have effective information extraction processes and well-established APIs to gather the news source content, there still is a pertinent need for preprocessing the extracted data. The news content extracted from social media can be pre-processed by natural language processing (NLP) toolkit. Common pre-processing operations are tokenization and labeling, part-of-speech tagging, semantic role labeling, dependency parsing and named entity linking (Kumar & Muruganantham, 2016; Lima, Espinasse, & Freitas, 2017). The next challenge on data pre-processing is to reduce the amount of data identifying and eliminating duplicate messages. It is not an easy task since every message posted in the social media can be valuable; de-duplicating the messages requires thorough clustering and then prioritizing them based on the event context. In order to prune this whole process, semantic-based technologies are used.
+
+Apart from that, there are various issues associated with handling social media messages. The most prominent issues are scalability and content. The scalability issue concerns Twitter stream size, volume, and velocity. Particularly during any large crisis or severe havoc (Otegi, Arregi, Ansa, & Agirre, 2015; Sakaki et al., 2010), a huge volume of tweets and millions of messages pertaining to that event may be posted. In these critical situations, the tweet velocity would never be at a constant rate. Instead, it grows drastically and records a huge response from people over the event. If it is observed at various times, it would be discovered that same/similar tweets were repeated and reposted again
+
+and again for the same event. This is the foremost challenge for the scalability issue; redundancy avoidance is the core factor for decision making and enhances the level of understanding over the specified event. Next, the content issue deals with tweets that are very brief and canonical (Abel, Gao, Houben, & Tao, 2011); most of the tweets posted in the social media are akin to normal speech and they pose a seminal challenge for the computational methods to deliver the correct form.
+
+# OBJECTIVES OF THIS RESEARCH
+
+The objective of this research is to provide informative summarization of social network content concerning happenings such as crisis, disaster, emergency and havoc based situations in the real world. A system is proposed that provides the summarized views of such happenings and corroborates the news by interrelating with one another. Its major task is to monitor the events which are very booming and deemed important from a crowd's perspective (Samuel & Sharma 2018; Sheth, Thomas, & Mehra, 2010). Important events cannot be adjudged as such. Instead, the suggested events must have the root on the social media such as Twitter after the specified news inception, and it must take the serious impact on the social media through series of takes by the social media users (Lei, Rao, Li, Quan, & Wenyin, 2014). The system proposed and evolved in this research monitors a discussion thread whether it has reached its peak or decline after its root in the social forum like Twitter. Eventually, it would give us insight into the evolutionary trends of the specified events over time.
+
+Thus this research aims to address the problems of entity ambiguity and its associated entity types for purposes of disaster management. We categorize the disaster based entity domains using ontology and enhance searching capability by incrementing the explicit connection which mutually exists between entity and ontology class. In order to achieve this task, we identify major issues to deal with and study them thoroughly for efficient processing of the following results: (1) Twitter streams and conventional news media tend to offer almost similar types of news coverage for a specified event but the rate of distribution among topics/categories differs. (2) On specific events such as disaster, crisis or any emergency situations, the volume of information accumulated between the two news media stands divergent and filtering the most potential information poses a challenging task. (3) Relational mapping/co-occurrence of terms suits well conventional news media but due to shortness and sparseness of tweets there remains a bottleneck for researchers. Therefore, we cover the details of the above-stated problems at length and propose algorithms and methods to exercise the cases conservatively.
+
+In the following sections, we present semantic filtering of entities from Twitter streams and identifying the potential meaning of tweets using domain ontology. Besides, we highlight a semantic model for disaster situations and detail at length different ontological tools available for effective filtering of semantic content. In that connection, we propose the model by which to analyze the semantic mapping of entities/terms over the concepts and disambiguate the potential named entities. Eventually, we detail the analysis and present the empirical results obtained from our proposed model that lead us to the conclusions.
+
+# **RELATED WORKS**
+
+In recent years, social media has gained momentum over collecting real-time events, and it has been proven that it has given the appropriate responses over the time of crisis when compared to other sources of decision-making systems. Celik et al. (2011) and Lei et al. (2014) did carry out timely searches over the crisis-related events even when the access to the online events was dropped consistently due to network latency and data traffic on the social media sites. Also, Abel et al. (2011) stated that extracting the relevant content over the crisis-related situations turned mostly ambiguous and many times redundant information. The key challenges that they addressed in the paper are how to overcome the difficulties in avoiding ambiguous results and removing redundancy. Sakaki et al. (2010) empirically discovered that during the course of extracting crisis-related information, compre-
+
+hending the phrases entered by social media users into the site is directly affecting the search results. Hence they used Wikipedia and other organized knowledge sources to map the canonical terms thus yielding better results through employing term dissimilarity.
+
+Grolinger, Capretz, Shypanski and Gill (2011) empirically showed event identification using statistical methods to observe an event's history and concluded on some proximity-based estimation of preciseness. They proposed a robust system that employs an algorithmic method based on 'Latent Dirichlet Allocation', which monitors the events and omits the content that is not relevant. They computed the average Euclidean distance between events, segregated abnormal changes present in the streams so that unknown distribution of data would be neglected, and got the Bag-of-Words. Another approach that they proposed in the algorithm is the log-likelihood rate by which the statistical ratio of data and TF-IDF difference present in the user-generated streams can be normalized based on term weighting and similarity score. In Wirtz, Kron, Löw and Steuer (2014), classification algorithms were used largely for event identification and classification. Many supervised learning algorithms were applied to divide the user-generated streams into the already-defined topic categories and carry out the detection process much easier than before. Approaches such as text classification and named entity recognition were extensively employed for exploring hidden events and disambiguating the selection process (using a vector space model to increase the viability of event detection).
+
+Liu, Brewster and Shaw (2013) explored the facts that covered the most common patterns on the disaster-related content and stored the details in a separate database. Weidong, Jidong, Jia, and Danni (2012) and Liu, Shaw, and Brewster, (2013) developed a system that detects online news events and searches related events on social media for the widespread collection of related event information. In contrast to our approach of ontology-based discrimination, all of the references mentioned above are statistical in nature.
+
+Lei Li and Tao Li (2013) developed a domain ontology for analyzing multi-document summarization pertaining to disaster or any crisis management. They provided many experimental models for developing an efficient ontological framework specifically on Hurricane Wilma in 2005. With that ontology, they precisely demonstrated the ontology-based multi-document summarization that performed well compared to other existing models. Jerman-Blažič, Matskanis and Bojanc (2017) discussed differences between man-made and natural calamities in detail and delineated the strategic measures, such as preparation, response, and recovery of any crisis or disaster-related situations. Their model was approved by the European Union and funded for the project REDIRNET. Later Selvam, Balakrishnan and Ramakrishnan (2018) constructed an ontology for Social Event Detection (SED) and applied it for Flickr (online photo management and sharing application) website by extracting the metadata features, such as geolocation, photoID, tags, description, title, timestamp, etc. On the other hand, they extensively utilized the Linked Open Data (LOD), such as Last.fm, Eventful, GeoNames, FourSquare, and so on, for productive discrimination of ontological properties. Although ontology/metadata-based, these works considered full documents or photo metadata in contrast to our study of short bursty messages, for example of Twitter.
+
+## SEMANTIC TECHNOLOGIES IN DISASTER MANAGEMENT
+
+The primary objective of Semantic Web technologies is to pave the way for users to easily find relevant information, navigate among diverse data sources and integrate heterogeneous information. For example, usage of semantic technologies will be highly important in getting relevant content and linking data elements to search concepts in the case of Twitter content during any heavy crisis or mass convergence event (Schulz, Ristoski, & Paulheim, 2013; Yates & Paquette, 2011). Nevertheless, the task is complicated because all such processes should be machine-readable and automated; this is where the semantic Web technologies come handy in affecting ontological enrichments (see Table 1).
+
+| Component<br>Name | Cached | Description                            |
+|-------------------|--------|----------------------------------------|
+| RDF, RDFS         | YES    | Vocabulary & Markup supported          |
+| MicroFormats      | YES    | Coupling Vocabulary & Markup Languages |
+| Data RSS Feeds    | YES    | Fetching Atom and Metadata             |
+| XSLT              | NO     | Define data prototyping                |
+| Web Services      | NO     | Interoperate with remote data objects  |
+
+Table 1. Semantic technologies for unambiguous data handling
+
+In the rest of this section, we expand on the need to semantically enrich social media content, survey semantic Web technologies to that end, and how to integrate data from various sources towards facilitating crisis management.
+
+## SEMANTIC ENRICHMENT OF SOCIAL MEDIA CONTENT
+
+As the semantic Web technologies play a seminal role in extracting meaningful information from social media and in the context of any crisis management (Heath & Bizer, 2011; Liu, Shaw, & Brewster, 2013), robust methods are required in dealing with different expressions of text that all in turn point to the central concepts. To get out this discrimination existing in the social media messages, the sole ways of solving this complexity is through making the Web machine-readable. To deal with the above-stated problems, semantic technologies provide an efficient technique called "named entity linking." Named entity linking is the process of detecting the entities prevailing in social media messages and associating the entities which are closely matched to the specified event (see Figure 1).
+
+Image /page/5/Figure/6 description: A flowchart illustrating a data processing pipeline. At the top, a component labeled 'Online News Extractor' is depicted as a stack of documents. This component feeds into two parallel processing streams. The left stream consists of two steps: 'Key Event Detection and Extraction' followed by 'Add Semantic Annotation to the Events'. The right stream also has two steps: 'Twitter Streams Entity Recognition & Filtering' followed by 'Semantic Labelling and Disambiguation of Events'. The outputs from both streams converge into a central process labeled 'Semantic Modelling & Ontology Mapping'. The final output of this process is directed into a cylindrical database symbol labeled 'Knowledge Repository' at the bottom of the flowchart.
+
+Figure 1. Semantic enrichment of news contents
+
+The named entity linking does two things. Firstly, it crawls the social media messages related to the crisis event detecting the potentially emerging entities like names, places, locations, organizations, category, etc. Table 2 displays ontological relations instrumental in extracting potential named entities of various genres using semantic technologies.
+
+Table 2. General domain-based ontological relations
+
+| Event                | General Relation                                                                                                              |
+|----------------------|-------------------------------------------------------------------------------------------------------------------------------|
+| Identify Relation    | is-a, isAbout, defines, occurs, exists, classifies, express, describes, isRelatedTo, sameAs.                                  |
+| Temporal Relation    | hasTime, timeInterval, time-span, timeStamp, during, eventDate, begin, end, since, nextTo.                                    |
+| Spatial Relation     | place, region, space, location, hasBoundary, nearTo, direction, overlap, placeName                                            |
+| Causal Relation      | cause, result, factor, agent, actor, action, activities, impact, consequence, result, participant, role, product, instrument. |
+| Exceptional Relation | isPartOf, hasSubEvent, hasComponent, hasMember, unifies, includes, involves, transitive, symmetric, negative, opposite        |
+
+Secondly, for every candidate event found, the named entity links searches for events in proximity to the named entity in that context. Unlike conventional search engines, it never searches based on a keyword. Instead, it makes use of the ontologies to augment the search process rendering the system to automate the process through enabling techniques such as RDF, RDFS, and OWL (Celik et al., 2011; Gruhl, Nagarajan, Pieper, Robson, & Sheth, 2009). Table 3 provides a sample of ontologies relevant in crisis management domain.
+
+Table 3. Domain specification of ontologies for crisis management
+
+| Domain Specification | Ontology Name                 | SW Representation  |
+|----------------------|-------------------------------|--------------------|
+| Resources            | SoKNOS                        | OWL-DL             |
+|                      | MOAC                          | RDF                |
+|                      | SIADEX                        | Unknown            |
+| People               | FOAF                          | RDF                |
+|                      | BIO                           | RDF                |
+| Organizations        | IntelLEO                      | RDF                |
+|                      | Organization OL               | RDF                |
+| Disaster             | EM-DAT                        | Online Query       |
+|                      | UNEP-DTIE                     | Online query       |
+|                      | Canadian Disaster<br>Database | Application System |
+| Damage               | HXL                           | RDF                |
+| Infrastructure       | OTN                           | RDF                |
+|                      | EPANET                        | Development        |
+| Geography            | GeoNames                      | RDF                |
+| Meteorology          | NEW weather<br>ontology       | OWL                |
+| Hydrology            | Ordnance Survey               | OWL                |
+|                      | Hydrology<br>Ontology         |                    |
+
+Once the named entity linking process is completed, thus having semantically enriched the messages present in the social media, then users can search for the information they want. This is called "faceted search". In faceted search, information is not crawled based on the keyword supplied but by associating the concepts for the term through proper ontological support (Grolinger et al., 2011; Malizia, Onorati, Díaz, Aedo, & Astorga-Paliza, 2010). For example, if we give the search word "virus", it will not search the information through using the keyword "virus"; instead, it finds relevant concepts for the keyword and associated terms in the ontology hierarchy and then related media messages based on the ontological concepts are returned.
+
+## PERTAINING CHALLENGES
+
+Tweets extracted from Twitter entails a huge amount of challenges to overcome in order to put them to use for a particular purpose. Wirtz et al. (2014) pointed out precisely that there would be no metrics stipulated about how much of information need to be monitored, extracted and evaluated. The following are major problems that present various challenges in yielding useful results:
+
+- a) Tweets extracted cover different aspects of the event and fail to make the distinction between the choices of the events.
+- b) Many tweets tend to be very noisy and sometimes irrelevant to the event thus causing unnecessary computational problems.
+- c) No measure to counteract rumors germinated during the events and it has been feared that it would spread vehemently over the short course of a period.
+- d) Dealing with misspelled tweets is a big task since there had been no apt use of a dictionary to makeover.
+
+# EVENT EXTRACTION
+
+Event extraction from Twitter is carried out through the Twitter Streaming API, which is the standard application fulfilling the filtration process effectively (El-Halees, & Al-Asmar, 2017; Sakaki et al., 2010). It extracts the tweets for the event by crawling through the hashtags created for the events by various NEWS sources on Twitter and monitors the user-generated posts subsequently (Silva, Wuwongse, & Sharma, 2013). The crawler is started to fetch the tweets that have been monitored and stored in the data store. Each and every tweet consists of the original post, author, timestamp, geographic information and hashtag from which it was obtained. All these properties are very useful in deriving the patterns and identifying the purpose of the posts.
+
+The formal definition for entity extraction of Twitter streams is expressed in the graph theory as G(E,V) where E and V represent some set of edges for the given set of vertices. To determine the potential social entities prevailed in the twitter streams and build the appropriate relationships between the entities, we link the set of edges E, in which  $v_i \in V$ , i = 1...N denotes the extracted entities in twitter streams and  $v_iv_j \in E$  denotes relationship between entities  $v_i, v_j \in V$ . In this connection, to estimate the candidate entity for the query q, the search engine would normally generate most ambiguous entity sets about the given candidate entity, and it is termed as
+
+$$a_i = q \leftarrow a_i \tag{1}$$
+
+However, in our proposed approach, we have introduced a novel method to tackle this ambiguity prevailed over the search results by incurring the semantic web ontology for the domain at which the entity is dealt with through the appropriate level of ontological weight, and it can drastically reduce with the addition of ontology candidate keyword KW, i.e. consequent of
+
+$$aw_i = q \leftarrow a_i, KW$$
+ (2)
+
+which reduced the entity ambiguity with which  $|aw_i| \le |a_i|$ ,  $|a_i| \in a_i$  is a cardinality of  $a_i$  and  $|aw_i| \in aw_i$  is a cardinality of  $a_i$ , KW. By utilizing a well-formed query for the candidate entity in the query, the named entity information would come as
+
+$$a_{"i"} = q \leftarrow "a_i" \tag{3}$$
+
+In some cases,  $|a_{il}| \le |a_i|$  and  $|a_{il}| \in a_{il}$  is a cardinality of "a;". As well as with
+
+$$aw_{i''} = q \leftarrow \text{"a}_i\text{"}, KW \tag{4}$$
+
+is about one of the information concentrations of a named entity. Then after pruning the entity cardinality, in the next process, the relationship between two named entities is based on the concept of co-occurrence. Thus,
+
+$$a_i \ a_i = q \ \leftarrow a_i, a_i \tag{5}$$
+
+which is a process to augment the semantic similarity between the two named entities and build the relationships between them, with which  $|a_i \cap a_j| \le |a_i|$  and  $|a_i \cap a_j| \le |a_j|$  and  $|a_i \cap a_j| \in a_i a_j$  is a cardinality of  $a_i a_j$ . Besides, with the supplement of a keyword towards the co-occurrence will usually subside the number of entities given, and that is
+
+$$aw_i aw_i = q \leftarrow qa_i, KW$$
+ (6)
+
+But it should satisfy that  $|aw_i \cap aw_j| \le |a_i \cap a_j|$ ,  $|aw_i \cap aw_j| \in aw_i aw_j$  is a cardinality of  $a_i a_j$ , KW. Similarly, effective utilization of the well-defined entity set for the query will yield the appropriate relationships between the two named entities.
+
+# ONTOLOGICAL INCLUSION FOR DISASTER MANAGEMENT
+
+Semantic technologies that support crisis events identification are often required for interacting between different developers and software applications operated by various agencies. In this context, affecting the semantically enabled system to communicate the information in a unified format is the critical challenge, and social media platforms behave differently to address the dimension of the problem to interrelate with one another. Interoperability shortcomings at the semantic level of concepts can be alleviated using common vocabularies as well as shared concepts in linking the whole processes (Liu, Brewster, & Shaw, 2013; Rajpathak & De, 2016). The best way of accomplishing this critical task is through the use of machine-understandable ontologies that can precisely define the concepts, categorize the events based on clustering approach and build an appropriate path between concepts and unified communication.
+
+The next aspect of the challenge in retrieving the crisis-related information is in extracting related events or messages from blogs, forums, and referring wikis. These are the places where vibrant information is present at a high rate and shared opinions and suggestion made by several authors. Now the critical challenge is interlinking not only social media platforms but also these blogs and forums. It is very difficult in repurposing the content and tough to identify the common events among these sites. For instance, take Wikipedia – it is a huge repository of publicly accessible knowledge source but reusing the same knowledge for other applications presents new challenges and difficulties (Kumar & Muruganantham, 2016; Yates & Paquette, 2011). Furthermore, a user can create accounts on many sites like in blogs, forums, wikis, and other social media platforms, but it is very complicated to inter-relate the candidate entities among these different social sites. The major problem pertaining to these media sources is that the information items in such sites are entirely disconnected and completely separated from one another. There is an absolute lack of exchanging the semantics of entities and unable to derive the facts from such information silos. Every site holds the information posted by their registered users independently and, at times, it has turned into stagnant information silos which are untapped by others.
+
+To meet the challenges rising from crisis or havoc situations, there would be a huge demand to decentralize the process and enable interactive processes to fetch hidden relevant facts from the content. Table 4 gives the details of handling the events from the time news originates to planning for future action taking against the crisis events (i.e., past to present).
+
+|                         | News Inception                              | Present State                 | Future Plan                |
+|-------------------------|---------------------------------------------|-------------------------------|----------------------------|
+| Goal                    | Inform & Publish the news                   | Share & Collect the news      | Engage & Prune             |
+| Main Activity           | Gather relevant event-<br>based information | Track & Monitor the<br>events | Prepare the action         |
+| Content                 | Initially, discrete data                    | Clustered Data                | Find Relationships         |
+| Information<br>Handling | Confidential                                | Privileged Access             | Absolute transpar-<br>ency |
+| Software Tools          | In-house Software                           | Commercial Software           | Open Source<br>Software    |
+
+Table 4. Information handling for crisis situations
+
+Semantic Web recently provided the necessary tools for effective information linking and interoperability. Moreover, many semantic Web vocabularies have successfully been deployed at various social platforms facilitating machine-understandable message processing (Benali, & Rahal, 2017; Ritter et al., 2012). Some of the semantic Web vocabularies are RSS, FOAF (Friend of a Friend), and SIOC (Semantically Interlinked Online Communities). With the help of these and other more refined semantic vocabularies, interlinking communities and social sites became effective and helped curtail down information redundancy.
+
+For example, let's consider the query for crisis-related content on Twitter such as "Was there a storm near the city?" In this query, the name of the city is not mentioned, but the tagging engine like DBPedia Spotlight (<a href="http://wiki.dbpedia.org/projects/dbpedia-spotlight">http://wiki.dbpedia.org/projects/dbpedia-spotlight</a>) and OpenCalais (<a href="http://www.opencalais.com/">http://www.opencalais.com/</a>) would annotate the given query and fix the appropriate entity for the annotated tokens based on the Agglomerative Clustering techniques applied on the collected tweets. Since there is a semantic link (i.e., rdf:type) between the query and DBPedia, it can be computed based on the similarity score and higher relevance of content. It is illustrated in the following:
+
+Image /page/9/Figure/6 description: A diagram illustrates a query and result relationship using DBpedia entities. The 'Query: Was there storm near the city?' points with a dashed arrow to a circle labeled 'Dbpedia-owl: City'. Below, the 'Result: There was heavy storm near New York City yesterday.' points with a dashed arrow to a circle labeled 'Dbpedia-city: New York City'. The two circles are connected by a double-headed arrow labeled 'rdf:type', indicating that 'Dbpedia-city: New York City' is a type of 'Dbpedia-owl: City'.
+
+## Ontology Languages
+
+The so-called 'semantic Web stack' comprising a number of semantic Web languages was suggested for effective utilization in information processing and that, by the way, leads to efficient semantic implementations on the retrieval systems. The first language that was suggested to use was the Resource Description Framework (RDF), by which the basic framework to represent the potential information on the Web content was availed. The basic structure of any RDF statements is just a triple (subject, predicate, and object), which further yields the hierarchical RDF graph to prune the data in a much effective way. In simpler terms, RDF statement is just denoting the relationship existing between the existing things called nodes and that node is interconnected with other nodes (i.e., semantically related nodes). RDF serialization uses XML syntax and terms such as element name, attributes, values, etc. The scope of using RDF is to make the system machine-readable and process the infor-
+
+mation semantically (see Table 5). Besides, it integrates the data without any serious glitches as it follows the well-formed logic which is universally acknowledged.
+
+Table 5. Semantic Web languages and their structural contents
+
+| SW Languages | Structural Contents                              |
+|--------------|--------------------------------------------------|
+| OWL-S        | Services in a machine-understandable format      |
+| SWSL         |                                                  |
+| WSML         |                                                  |
+| OWL          | Equality                                         |
+|              | Relation between Classes                         |
+|              | Cardinality Restrictions for Properties          |
+|              | Relation of Properties                           |
+| RDF Schema   | Classes                                          |
+|              | Objects                                          |
+|              | Properties                                       |
+| RDF          | Prepositions as Triplets                         |
+| WSDL         | Services in an exact human understandable format |
+
+A set of RDF statements form an RDF graph through interconnected nodes. As the RDF graph is conventionally followed in expressing the logical facts about the potential named entities, ontologies are used to give the domain and category of each thing (i.e., entity) and yield the appropriate relationship between them. Ontologies contain features to wholesomely express the rich relationship between entities and also set appropriate constraints on them (Grolinger et al., 2011; Wang et al., 2018). The language followed for effective creation of ontology is RDF Schema and OWL (see Tables 6 and 7). The RDF Schema (RDFS for short) employs the set of classes related to the entity and chooses the properties according to its domain.
+
+The basic objective of RDFS is to provide a well-structured description of entities properties. Some ontologies used to set classes and properties are called RDF vocabularies, and examples include FOAF, SKOS, MOAC, Dublin Core, etc., whereas OWL is facilitating the information interoperability and providing additional vocabularies to enhance the formal semantics of RDF Schema.
+
+Table 6. RDF and OWL Ontological Constructs
+
+| RDF/OWL  | Category   | Functions             |
+|----------|------------|-----------------------|
+| Class    | Definition | Class                 |
+|          |            | Enumerated Class      |
+|          |            | Restriction           |
+|          |            | IntersectionOf        |
+|          |            | UnionOf, ComplementOf |
+|          | Axiom      | subClassOf            |
+|          |            | Equality              |
+|          |            | disjointWith          |
+| Relation | Definition | Property              |
+|          |            | Domain, range         |
+|          |            | subPropertyOf         |
+|          | Axiom      | (Inverse)Functional   |
+
+| RDF/OWL  | Category   | Functions                                    |
+|----------|------------|----------------------------------------------|
+|          |            | Equality, inverseOf<br>Transitive, symmetric |
+| Instance | Definition | Type                                         |
+|          | Axiom      | (In)Equality                                 |
+
+## **ONTOLOGY EDITORS**
+
+In order to derive the meaning out of the collected information from Twitter streams and as we process the tweets into the respective semantic representations, there is a need to design an ontology that is well mapped to the information and facilitates fetching the content in the hierarchical format (Liu, Shaw, & Brewster, 2013; Malizia et al., 2010). In this connection, we are required to build ontology using standard text editors that are very simple in design and development. Besides, it can be formatted with the semantic Web languages called RDF-XML format. One such editor is called Protégé (<a href="https://protege.stanford.edu/">https://protege.stanford.edu/</a>) which is a tool that permits designing OWL ontologies and further helps to connect to the data interrelated with one another in the overall ontological framework. With Protégé, querying and reasoning help to disambiguate the information and assist in making the filtration absolutely error-free. Other popular ontology editors are SWOOP, OntoStudio, NeOn, Altova, WebODE, and so on. Among all the editors, Protégé was the first, freely available, and open-source ontology editor and framework for building intelligent systems thus tops the list and widely deployed in much recent research.
+
+Table 7. A notional mapping between RDF/OWL and relational concepts
+
+| RDF/OWL Terms | Relational Concepts                        |
+|---------------|--------------------------------------------|
+| rdf:class     | Table                                      |
+| rdf:property  | Column                                     |
+| rdfs:domain   | Table that the rdf:property is a column of |
+| rdfs:range    | Data Type of the column                    |
+| rdf:type      | Values of the Primary Key column           |
+
+## SEMANTIC MATCHING AND TRANSLATION
+
+As ontologies play a seminal role in semantic processing of information (Celik et al., 2011; Sheth et al., 2010), we should, therefore, try harnessing the potential meaning hidden in the collected information streams. Ontologies help process the meaning of different terms represented in the information and avail the system to understand the very basic structure of information in a precise way. The system processes the collected data automatically and does matching and translation with its rich vocabulary sets, which is fed as the dataset to the ontology while designing the complete framework of the domain ontology (Gruhl et al., 2009; Madani, Boussaid, & Zegour, 2015). Each term (i.e, a concept) in a tweet is mapped with corresponding vocabulary sets in the specific ontology domain, and sometimes it arches to other domains of vocabulary set to find the exact meaning of the concepts represented in the tweet. Precisely, the inclusion of mapping the terms over multiple ontologies is the biggest challenge in designing an application that is used to integrate the ontologies to disambiguate in parallel; that is a challenging research area to deal with. Further, to make the automation of semantic matching and translation effective, appropriate use of mapping rules over the information is necessary and should be defined using ontology matching tools. Ontology matching is variously called also as ontology aligning, mapping, and translation (for example, for Web services discovery: Fellah, Malki, & Elçi, 2016). Conventionally, ontology mapping tools come in two categories: element-based approach such as name similarity, entity similarity, concept similarity, etc., and structurebased approach such as sub/super -categories, -domains, -levels, etc. Besides, mapping requires infusing external knowledge such as Thesauri, WordNet, etc., to yield precision and high recall. Some of the ontology mapping tools available on the market are RiMOM, ASMOV, and AgreementMaker.
+
+## SEMANTIC SEARCH
+
+The next level in semantic utilization of social media harvested information is through semantic search, which should return results without any ambiguity and sparseness. As semantic mapping directly links information repositories, domain ontologies should facilitate the operation of effective semantic search (Kumar & Muruganantham, 2016) retrieving facts that are interconnected with one another in the ontological framework. In order to render the search process easier, appropriate use of indexing methods in the ontological inclusion over the concepts is deemed important. Ontologies follow the semantic indexing approach using its standard principle of "indexing the RDF triplets", thus smoothen the way of semantic search over the collected information. Several pieces of research have been carried out in this regard to fetch the precise and unambiguous results by semantically integrating information from diverse ontological frameworks in retrieving the results from multiple repositories.
+
+## INTEGRATION OF DATA
+
+Another critical issue faced in utilizing social media content is the integration of data, which may be considered from two different perspectives: data source (database/stream) reconciliation, and information integration. The basic objective of overcoming the problems of semantic heterogeneity between these two categories using the appropriate ontology framework is a challenging task to any research (Liu, Shaw, & Brewster, 2013). In considering database integration, the role of ontologies lies on the upper layer of the schema (i.e, semantic matching of information and table schema should be shared with the domain ontology and additionally, make use of the ontology to integrate the database schema rightly towards its order). While integrating information the core problem lies at integrating the terms from various sources and mapping the potential candidate terms relevant to its vocabulary sets and bring into the consolidated view called the new collection of a derived set. The challenge that lies here is to not change the original sense of the terms while mapping to the appropriate sense of terms in the vocabulary sets. Besides, in database integration, ontologies convert tables into respective classes in RDF triplet and columns in the table into data relation in the RDF Schema. The data models followed in the semantic conversion would always be 1:1 mapping cardinalities. In recent works (Kumar & Muruganantham, 2016; Kwak, Lee, Park, & Moon, 2010), the authors propose rules to dynamically map the data models into ontologies and consider mapping instances to class levels. Also, some language constructs are given to fetch data objects and, using the queries, they can be annotated dynamically representing in RDF.
+
+Whereas in information integration, accumulating the terms from various sources of datasets to bring them into a unified collection, several efforts were carried out in the recent past but failed to resolve it. Some early researchers have tried to apply the Description Logic (DL) as ontology language and observed few changes in the outcome. Later, the Prolog programming language is employed for expressing the information formally and integrating the terms using appropriate domain ontology.
+
+In the next section, we introduce our proposed information-centric model for management of crisis and disaster based situations through integration of many of the technologies mentioned above combined by our innovative approach. Our proposed model is introduced, followed by the empirical tests and discussion of findings.
+
+# THE PROPOSED MODEL
+
+The objective of this research is to potentially harness the information gathered from various social media platforms and render it relevantly interconnected with the selected news articles. In doing so,
+
+here we introduce some of the notation and problems that we define formally before presenting the Semantic Search for Events Algorithm.
+
+**Problem 1 (News stream):** For every news article related to disaster or crisis situation, content must be analyzed and scrutinized for a further level of comprehension. Let  $N = \{n_0, n_1, ..., n_i\}$  be the news posted on various sites and gathered from various news agencies. For every news article posted say  $n_i$ , we find the actually published time t ( $n_i$ ). Since the origin of news story gives the real arrival of news, it brings in the proximity among related news articles.
+
+**Problem 2 (Tweet stream):** Upon the arrival of every news article related to disaster or crisis, the next task of the system is to identify the equivalent social media content such as Twitter where the relevant news item is discussed and promulgated. Let  $S = \{s_0, s_1, ..., s_j\}$  be the Twitter Streams for the taken news articles and load the inter-related Twitter messages posted by various potential social users. For every tweet  $s_i$ , we find the actually posted time t ( $s_i$ ) and responses for the message.
+
+**Problem 3 (News recommendation problem):** Once the news items N and its associated social media contents (Twitter Streams) S are mapped, then the real task is to find the top-k most relevant news for the topic. Let's take the set of users interacted on the particular news topic  $U = \{u_0, u_1, u_2, ..., u_n\}$  in the social media platforms and explicitly categorize the social messages and news streams of general interest (i.e., for any social user  $u \in U$  at any point of time T, we recursively adopt the functional ranking which links the users interest among its neighbors).
+
+**Problem 4 (Social influence):** In order to find whether the news item  $n_i \in N$  has influenced the social media users  $U = \{u_0, u_1, u_2, ..., u_n\}$  effectively, we give the social influence model  $S = |U| \times |U|$  matrix where S (i, j) calculates the cumulative interest of the selected users  $u_i$  to the usergenerated content by  $u_j$ . This process states that each user in the context would pose an absolute interest to the user-generated content posted by the other user.
+
+**Problem 5 (Tweets-to-news model):** To merge the process, let N be the order of news collected and S be the streams of social media messages, we model the relationships between user-generated content and news items as  $M = |S| \cdot |N|$  matrix Z where S(i,j) is the closest proximity of user-generated content  $s_i$  to news item  $n_i$ .
+
+```
+Algorithm (Semantic Search for Events)
+
+Input: Seed words for each crisis event
+
+Output: Generation of semantic classes
+
+BaseTerms ← set of seed words given;
+
+for i: 1 to N (Number of Iterations) do
+
+BaseTerms ← ExpansionOf (Seed words, Corpus);
+
+BaseTerms ← Cluster (BaseTerms, Seed words, Corpus);
+
+end
+
+return BaseTerms
+```
+
+The algorithm carries two significant operations: (1) expand the seed words with the assistance of ontologies; and (2) cluster the events based on the similarities existing in the classification. For the given seed words for the event, it crawls for new terms that possess the similar distributional features to the seed word and assigned to the set of seed words (also called as candidate words). In fact, extracting the new terms for the seed word can be done based on the contextual features and top score similarity measure. For the clustering, the selection procedure would process the learned terms and
+
+seed terms based on the distributed similarity and set the minimum threshold value for estimating the exact precision of the terms.
+
+The sole plan of this algorithm is to get the patterns which are interlinked with a semantic relation and bring in the semantic class for the search terms (Kumar & Muruganantham, 2016; Wang et al., 2018). As listed in the algorithm, it has three core operations for finding the patterns existing in the disaster based situations:
+
+- 1. using the semantic class expansion algorithm, extract the candidate terms for the disaster event tags;
+- 2. find the patterns for the candidate terms selection and fix the semantic category of the events; and
+- 3. choose the cluster events which hold similar action terms and evaluate the patterns for further classification.
+
+It has been noted at several instances (Abel et al., 2011; Sheth et al., 2010; Wirtz et al., 2014) that news items and user-generated content at social media platforms (say, Twitter Streams) co-exist with one another with same news topic (see Figure 2). Sometimes, a published news story is pushed into social media platforms for further discussion and circulation. And, at many times, a news item first discussed vehemently on social networks then becomes a topic in news stories (Lei et al., 2014; Li, Liu, Li, Qin, 2016). In these two cases, the predominant factor is holding the current trending entities, which give the unflinching bond between social networks and news sites. There is an absolute relationship between user-generated content and news stories, which create an intermediary layer that paves the way to generalize the analysis. Hence, this would make our work equally applicable in deriving the ultimate decision for disaster management and assess the core patterns for decision making.
+
+During the analysis of the relationship between events and results, there would emerge a need to attain a similarity score for the ultimate decision process. The similarity score for the crisis/disaster management (Liu, Shaw, & Brewster, 2013; Malizia et al., 2010; Schulz et al., 2013) can be accepted and formulated based on the following assumptions:
+
+- 1. If there would be a high or low hazard during the disaster situations, then it requires the scientific or technical measure to be assured, and precautionary steps should be taken based on scientific or technical grounds.
+- 2. If there would be high or low outrage, then it is an emotive issue and should be tackled through proper negotiations or political balance.
+
+In these two cases, the analysis of the events played a crucial role in disseminating the user-generated content posted on social networks and determining the effective decision-making process (Heath & Bizer, 2011).
+
+## **DISASTER ONTOLOGY**
+
+To substantiate our proposed model, we constructed an ontology for disaster datasets with a glossary consisting of more than 150 definitions (i.e. mostly of recurring terms) and further accumulated terms related to disaster from books, papers, survey on seismic risk and other relevant disaster web sites. We constructed the ontology with associated concepts, its attributes and proper relationships between concepts. In constructing this ontology, we followed many dictionary terms (i.e. entities related to disaster) with their associated meanings (axioms) and connected the terms with taxonomic relationships. Relationship mapping of terms can be done in many ways such as Taxonomic (IS-A relationship), Meronomic (PART-OF relationship) and Telic (PURPOSE-OF relationship). The relationship mapping of terms can be achieved through inference rules to augment better reasoning and increase the credibility ratio of knowledge representations. On this proposed system, we used Protégé, an ontology tool which is more of an object-oriented paradigm and well suited for term inher-
+
+itances. The relationship IS-A is a generalization/specialization between the candidate entities: superclass entities publically generalize the subclass entities and the sub-class entities particularly employ specialization of superclass entities. Likewise, Protégé permits to formulate the disaster ontology by considering different instances to insert and able to accommodate a huge set of information for a digital archive.
+
+Image /page/15/Figure/2 description: A screenshot of a software interface displaying a flowchart related to risk assessment. The window has a menu bar with "File", "Tools", and "Help", and a toolbar with buttons like "Previous", "Next", and "Build graph". The main area is divided into two panes. The left pane contains a list of terms, with "Damage" highlighted in bold. The right pane shows a flowchart. At the center is a grey box labeled "Damage". An arrow flows from "Vulnerability" to "Damage", and another from "Damage" to "Exposure". The "Vulnerability" box is influenced by "Seismic Risk" and "Structural Vulnerability". The "Exposure" box is influenced by "Physics Elements Exposed" and in turn influences "Seismic Risk", "Functional Exposure", and "Strategic Exposure". Several boxes are marked with small numbered circles.
+
+Figure 2. Disaster Ontology using Protégé
+
+Our proposed system concerns mostly about urban risk with specific governance on seismic risk management. The effective building of this ontology paves the way for common knowledge, makes the concepts understandable, and prompts information into unambiguous semantics. This ontology construction has been performed in three steps:
+
+- 1. Fetch the core concepts of the domain (Seismic Risk) and relevant terms in the glossary.
+- 2. Extract the Super-Classes and Sub-Classes of the concepts using the IS-A relationship.
+- 3. Find other related types of relationships using inference rules (properties, slots, and roles associated with each concept).
+
+Relation mapping for the collected tweets can be performed and filtered using the relational properties displayed in Table 8). Entity resolution and disambiguation have been effectively dealt with in Disaster Ontology constructed above and resolve the term ambiguity persisting over the collected documents (Twitter streams).
+
+| Relation Name    | Source     | Target      | Description                                                                                                        |
+|------------------|------------|-------------|--------------------------------------------------------------------------------------------------------------------|
+| isResponsibleFor | Department | Process     | Identify which sector is responsible for the<br>event and map the relationship between de-<br>partment and process |
+| workIn           | Actor      | Department  | Map the relationship between the person and<br>the department. Identify the actor responsible<br>for the event.    |
+| isPartOf         | Task       | Process     | Find the task which is responsible for the pro-<br>cess and filter out the concepts related to the<br>event.       |
+| isA              | -          | -           | Relationship between super-class and sub-class                                                                     |
+| Perform          | Actor      | Task        | Group the actor performed the task on the<br>event.                                                                |
+| Produce          | Task       | Information | Filter the information for the task on the<br>event.                                                               |
+
+Table 8. Relationship mapping between concepts and classes
+
+## ENTITY RELATIONSHIP AND RANKING SCORE
+
+The disaster ontology has now become a knowledge source for our disambiguation effort. When we process each and every tweet, we find the exact match of those entities against the knowledge source such as DBpedia or YAGO. If it is not present, then it sends the NIL result. Now, by means of our proposed method, we can again cross-match with our own ontology created from news articles and find the exact match of those entities. In this method, the accuracy is relatively high because the created ontology is extracted from news articles related to the tweets and context of the news articles is highly relevant and appropriated match with the tweets. If we go for the entity-mention match with DBpedia, it lists out candidate mentions for the entity, and we need to probe for the context pertaining to the tweet. But if we match the same with our own ontology, it is exact and gives an appropriate match.
+
+Hence, in our approach, we take the link probability (Kumar & Muruganantham, 2016; Yates & Paquette, 2011) for the entity with DBpedia mention, and it can be defined as follows:
+
+$$F_{(e,m)} = \frac{Count(m,e)}{Count(m)} \tag{7}$$
+
+Here, we utilized an outlined ontology to arrange the mentions for the given named entities and appropriately estimate the similarity distance between them. Now the task is to estimate the distance between the entity and the suggested set of mentions from DBpedia. In this connection, we have taken the Cosine Similarity measure to access the similarity difference existing between the entity and candidate mentions as follows:
+
+$$CosSim(e,m) = \frac{Product(e,m)}{||e||*||m||}$$
+(8)
+
+By this method, we categorically filter the exact match of mention for the given entity and appropriately reference with DBpedia URI as stated in (Liu, Brewster, & Shaw, 2013; Malizia et al., 2010; Schulz et al., 2013). We utilized the DBpedia Spotlight to get the URI match of each entity and return the JSON results for our implementation.
+
+The result of the proposed approach would create a binary mapping of the entity and mentions, as seen in Table 9.
+
+Table 9. Identifying the relation between named entity and candidate mention
+
+| <i><b>Mention</b></i> | <i><b>NE<br/>Class</b></i> | <i><b>NE Link</b></i>          | <i><b>DBpedia Ontology<br/>Class</b></i> | <i><b>Score</b></i> |
+|-----------------------|----------------------------|--------------------------------|------------------------------------------|---------------------|
+| <i>Barack Obama</i>   | <i>Person</i>              | <i>Dbpedia: Obama, USA</i>     | <i>Dbpedia-owl: Person</i>               | <i>3</i>            |
+| <i>Chennai</i>        | <i>Location</i>            | <i>Dbpedia: Chennai, India</i> | <i>Dbpedia-owl: Place</i>                | <i>1</i>            |
+| <i>Cricket</i>        | <i>Sports</i>              | <i>Dbpedia: Cricket</i>        | <i>Dbpedia-owl: Sports</i>               | <i>2</i>            |
+
+Generally, entities in DBpedia have its name, label, type, etc. and, to fetch the entity name given in the DBpedia for the specified URI, it can be queried through the SPARQL query as follows. For example, searching for 'Sachin Tendulkar':
+
+```
+Select distinct *
+where {
+?URI rdf:label ?name
+?URI dbpprop:iupacname ?name
+filter(str(?name) = "Sachin Tendulkar")
+}
+```
+
+In order to get the category of a given entity from the DBpedia, we issue the following SPARQL query. For example, for 'Vehicle':
+
+```
+Select *
+where {
+    <http://dbpedia.org/resource/Vehicle>
+    <http://purl.org/dc/terms/subject>
+?categories.
+}
+```
+
+# EMPIRICAL TEST AND ANALYSIS
+
+We used Twitter4J API to gather disaster-related tweets from Twitter and utilized TextRazor API to effectively recognize the potential named entities present over the tweets and link them accordingly to its respective DBpedia URI. Additionally, we used the rich natural language processing tools of Stanford Core NLP Library to segregate tweet patterns and performed sentiment analysis for grasping the sense of the tweets. Tweets were collected on the month of August 2017 and, to witness the trust, we followed the leading news agencies on Twitter such as BBC World, CNN, New York Times, NDTV, and Breaking News. Tweets were crawled and stored only if they had at least one named entity that has its link on DBpedia URI. In our datasets, we were able to filter out 20 different topics and classified the tweets successively based on seismic risk by applying the classification rules. The algorithm proposed above is able to detect the factual information containing about 3 out of 5 tweets.
+
+| Event Category      | Total<br>Events | Potential Sub-Events by Relevance |         |         |
+|---------------------|-----------------|-----------------------------------|---------|---------|
+|                     |                 | R3                                | R3+R2   | R3-R1   |
+| Earthquake          | 75              | 35(46%)                           | 51(68%) | 59(78%) |
+| Tsunami             | 120             | 46(38%)                           | 79(65%) | 88(73%) |
+| Cyberattack         | 114             | 51(44%)                           | 87(76%) | 95(83%) |
+| Unrest in a Country | 150             | 77(51%)                           | 90(60%) | 97(64%) |
+| Celebrity Death     | 115             | 43(37%)                           | 66(57%) | 81(70%) |
+| Terror Attack       | 120             | 68(56%)                           | 79(65%) | 85(70%) |
+
+Table 10. Event relevance and categories
+
+We tested the DBpedia corpus to identify potential events on seismic risk, which provided the six complex event categories listed in Table 10. The entities were extracted based on the recommendations stated above and identified their relationship types in corresponding DBpedia URI. Besides, we again queried the DBpedia Knowledge Source for the sub-events correlated with the events extracted from the tweets. We substantially ranked the sub-events on the basis of frequency of occurrence and chose the best-matched event category to a tweet. After evaluating the event categories against DBpedia, we determined whether the event is of positive instance or not. Sometimes, the retrieved events would pose a challenging task such as if it is partially relevant but not exactly appropriate to the categorized concepts. During these anomalies, we assigned the following three relevance scores in order to fit the events into their appropriate decks:
+
+- Relevance (R1): Events with fuzzy relationship to the concept/category.
+- Relevance (R2): Events with positive occurrences of sub-events or subject-object mapping.
+- Relevance (R3): Events are positive instances and fit into the category for the posted query.
+- Otherwise, the relevance zero indicates the events with absolutely NIL relationship.
+
+Table 10 displays detected event categories and potential sub-events or co-occurrence of events with relevance scores. As was witnessed, the precision values varied considerably among the categories. The Stanford NLP Library was deemed fit to extract the potentially relevant tweets, and type filtering of events was absolutely effective at identifying the appropriately named entities. We obtained an accuracy of 74.13% and computed the Precision (0.641), Recall (0.716) and F-Measure (0.691) respectively for the given datasets.
+
+# **DISCUSSION**
+
+The dynamic change in the amount of information gathered at the various medium of platforms indicates the need for a rapid decision-making process in crisis events. It was observed that the information gotten from these sources rapidly varied. Statistics (Wirtz et al., 2014) showed that the frequency of report variation grows ten times greater than the previous day. Besides, to better account for the report variation of the information accumulation, the report dimensions were categorized into three crucial breakpoints, i.e., D+1, D+5, D+10. This elapsed gap fetches the detailed overview of the crisis or disaster based events and showed us the real potential of the event happenings (see Figure 3).
+
+Image /page/19/Figure/1 description: A line chart titled 'DAY WISE TWEETS' that plots the 'TWEET COUNTS' for four different events over a period of 10 days. The y-axis, 'TWEET COUNTS', ranges from 100 to 350,100. The x-axis shows the days, from DAY 1 to DAY 10. There are four lines representing four events: 'Event 1' (blue line with diamonds), 'Event 4' (yellow line with 'x' markers), 'Event 2' (orange line with squares), and 'Event 3' (gray line with triangles). Generally, Event 3 has the highest tweet count, followed by Event 2, Event 4, and finally Event 1 with the lowest count. All events show a significant peak in tweet counts on Day 4, with Event 3 reaching approximately 300,000, Event 2 reaching about 240,000, Event 4 at 180,000, and Event 1 at 90,000. Events 2 and 3 show another high peak on Day 10, reaching approximately 200,000 and 280,000 respectively, before all four events show a sharp drop in tweet counts at the very end of the period.
+
+Figure 3. Daily frequency of information on social media platforms
+
+Through the data obtained from various sources and on different days of report gathering, we can formulate deviance of patterns and get through the details of anomalies that exist in the report. By applying the pruning algorithm, we can sort the crisis events for the decision-making process and get to the core base of the events. In this research, the real task is to find the actual reason for the crisis event and get the substantiated evidence for its occurring. To augment this process, we classified the events into many chronological orders influenced by the usage of ontological background with semantic technologies. By mapping different day event reports, we scrutinize the process for discrimination (i.e., fetch the positive or negative or neutral feedback from the potential users on the social media) and allow filtering the facts based on cross-checking in tabulating the actual events of the situation.
+
+Our approach achieved the accuracy rate of 74.13% where other existing models succeeded getting 68.42% using Support Vector Machine (SVM), 67.93% using Maximum Entropy Model (MEM), and 64.71% using Conditional Random Fields (CRF) based on the analysis successfully performed with the help of Table 9. Since our proposed model extensively uses the dedicated ontology of Crisis and Disaster, instead of employing the Bag-of-Words (BoW) method, we employed Bag-of-Concepts (BoC) and Relevance of Concepts, as well as calculating the semantic similarity score between ambiguous terms. Deep proliferation of the ontological network paved the way to yield the subcategories of a topic and skimmed the words that are completely unambiguous. The relevance R of the concepts were derived with other three relevances R1, R2, and R3 as shown in Table 10, whereas the other existing methods mostly used only a single relevance score and restricted the research scope to Bag-of-Words model.
+
+The major contribution of this research is in collecting crisis-related temporal data from multiple bursty short-message sources and decision making through semantic mapping of entities over concepts disambiguating potential named entities. The problems persisting over entity ambiguity and its associated entity types were addressed as well. We categorized the disaster-based entity domains using ontology and enhanced the searching capability of the system by incrementing the explicit connection mutually existing between entity and an ontology class.
+
+# **CONCLUSIONS**
+
+In this paper, we proposed a novel solution to harvest and compare the content of Twitter streams and conventional news sources such as CNN, New York Times, BBC World, NDTV, and Breaking News in the cases of havoc situations. We developed a semantic filter that can map the concepts correlated between Twitter streams and traditional news sources, and can disambiguate the candidate entities based on the ontological framework particularly loaded with disaster/crisis events.
+
+The major advantage of our work is that, instead of pruning a single news source, it paves the way for clustering the information from diverse sources and harnessing the potential information to derive the hidden facts in it. We also developed a disaster ontology for this research and used it to segregate the entities which pose ambiguity over other candidate sets.
+
+Empirical results show that the approach based on our model outperforms other models available in the literature to solve this research gap by various other approaches. In the future, we shall strive to extend the model in order to help summarize and visualize the potential information ranked high by the model.
+
+# **REFERENCES**
+

chatbot/rag/data/('D_record__Disaster_Response_and_Relief_Coordination_Pipeline', '.pdf')_extraction.md

@@ -0,0 +1,140 @@
+
+## **ABSTRACT**
+
+We employ multi-modal data (i.e., unstructured text, gazetteers, and imagery) for location-centric demand/request matching in the context of disaster relief. After classifying the *Need* expressed in a tweet (the WHAT), we leverage OpenStreetMap to geolocate that *Need* on a computationally accessible map of the local terrain (the WHERE) populated with location features such as hospitals and housing. Further, our novel use of flood mapping based on satellite images of the affected area supports the elimination of candidate resources that are not accessible by road transportation. The resulting map-based visualization combines disaster-related tweets, imagery and pre-existing knowledge-base resources (gazetteers) to reduce decision-making latency and enhance resiliency by assisting individual decision-makers and first responders for relief effort coordination.
+
+## CCS CONCEPTS
+
+• **Information systems** → *Decision support systems*;
+
+## **KEYWORDS**
+
+disaster relief, location-centric processing, flood mapping, need matching
+
+## **ACM Reference Format:**
+
+Shruti Kar<sup>1\*</sup>, Hussein S. Al-Olimat<sup>1\*</sup>, Krishnaprasad Thirunarayan<sup>1</sup>, Valerie L. Shalin<sup>1</sup>, Amit Sheth<sup>1</sup>, Srinivasan Parthasarathy<sup>2</sup>. 2018. D-record: Disaster Response and Relief Coordination Pipeline. In *Proceedings of SIGSPATIAL Workshop (ARIC 2018)*, Jennifer B. Sartor, Theo D'Hondt, and Wolfgang De Meuter (Eds.). ACM, New York, NY, USA, Article 4, 4 pages. https://doi.org/10.475/123\_4
+
+## 1 INTRODUCTION
+
+Recent catastrophic events, e.g., resulting from floods and hurricanes, combined with increased urbanization and interdependent infrastructure reveal an increasing vulnerability to natural and human-made hazards. A resilience framework must accommodate both precise information for action and for various levels of analysis, posing substantial challenges [4]. These include characterizing hazards with an integrated ontology across voluminous, multi-modal
+
+Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for third-party components of this work must be honored. For all other uses, contact the owner/author(s).
+
+ARIC 2018, Nov. 2018, Seattle, Washington
+© 2018 Copyright held by the owner/author(s).
+ACM ISBN 123-4567-24-567/08/06.
+https://doi.org/10.475/123\_4
+
+data sources, facilitating rescue/response actions, and assessing the effects of pre- and post-disaster actions on individual needs.
+
+The dire real-time need for integrated, coherent multi-modal data from different sources has motivated substantial research [8]. Social media is a significant source of such data. Twitter users, for example, post short texts as updates. These messages (a.k.a. tweets) reveal *Needs* and requires matching with possible available help. These messages can provide critical information to first responders and individual decision-makers to contribute effectively to relief efforts [18]. Tweets authored about crisis events may contain crucial but unstructured location-centric information (such as incident reports) even when explicit geo-tagging is absent. The ability to extract specific location information from this unstructured terminology and the other attached information is critical for timely assistance by the first responders.
+
+Crowdsourced data (such as OpenStreetMap) that covers the geographical area of a disaster event can assist in the extraction of location content from tweets. To achieve this, we developed D-record: **D**isaster **re**sponse and **re**lief **coord**ination, a pipeline that utilizes multi-modal data (i.e., unstructured text, gazetteers, and satellite imagery). D-record allows first responders to visualize and match location-centric *Needs* and options for available help while taking into account transportation constraints and available options in the local area of a disaster.<sup>1</sup>
+
+## 2 RELATED WORK
+
+The following related work informs the D-record pipeline and clarifies its contributions
+
+Disaster-Centric Ontologies. Lightweight lexicons or vocabularies like the Management of Crisis Vocabulary (MOAC)<sup>2</sup>, CrisisNLP<sup>3</sup>, and CrisisLex [21] provide the different concepts for disaster management but lack the granularity required for describing response and miss crucial spatial and thematic details. Other efforts like UNOCHA's Humanitarian eXchange Language (HXL)<sup>4</sup> improve information sharing during disasters but lacks specific crisis-related concepts. The Social Media and Emergency Management (SMEM) ontology [13] combines emergency domain knowledge with social media information for Incident Classification. However, SMEM focuses on major coarse-grained categories of disasters such as meteorology or biology. As stated above, state-of-the-art techniques in event representation for disaster response is critically limited with respect to the usability and dissemination of recommended action and response (intentional intervention) required for relief
+
+<sup>\*</sup>Authors contributed equally.
+
+<sup>&</sup>lt;sup>1</sup>The source code of the pipeline is available at https://github.com/shrutikar/d-record.
+
+<sup>&</sup>lt;sup>2</sup>http://observedchange.com/moac/ns/
+
+<sup>3</sup>http://crisisnlp.qcri.org/
+
+<sup>4</sup>http://hxlstandard.org/
+
+efforts coordination. Given the above limitations, D-record uses a location-centric ontology to assist in matching relevant needs with available help to support a disaster response use-case as discussed in Section 3.
+
+**Need Classification.** For extracting the type of *Needs* (e.g, food, shelter, or medical) from text data, state-of-the-art Need classification techniques such as [16] focus on the identification of information type. However, these techniques do not provide a comprehensive solution for matching classification output. [5] merely summarizes disaster information at a certain geographical level without providing any actionable decisions for response and relief. Instead, they provide summaries of topics to decision makers. Our previous classification technique [3] uses an ontology and location detection to enhance situational awareness through recognizing the interdependencies between disaster events, which helps in the early identification of Needs. Inspired by these techniques, D-record employs a coarser Need classifier to make its classifications compatible with the affordances of locations (e.g., a shelter can support a variety of Needs such as food and water whereas a hospital can provide medical and rescue support *Needs*). In other words, we resort to coarser needs (as opposed to fine-grained needs) for effective matching because available help centers satisfy multiple needs.
+
+**Matching.** Our previous technique on Seekers and Suppliers matching [18] uses hand-crafted rules to match the *Need* requested with a possible provider, e.g., matching someone offering shelter with someone requesting shelter. Other approaches to *Need* matching such as [6] uses structured input from users and utilizes moderators in the loop to match those *Needs* with organizations presently available to help. The *Need*-capacity matching technique by [17] quantifies the degree of impact of a disaster and prioritizes the matching accordingly. Finally, [14] provides both online and offline solutions for requesting, providing and coordinating resources.
+
+Limitations of the above work include: (a) the inadequate assumption that all routes are available for matching by assuming the full accessibility of roads during a disaster [19], and (b) the matching problem being solved for a different level of granularity than for the required individual level of *Needs*.
+
+**Flood Mapping.** Some logistical matching methods assume the full accessibility of routes. Hu et al. [7], for example, provides an algorithm for optimal path selection in a logistics supply-chain management system. In contrast, [15] provides a portal for UN staff to upload spatial data including satellite images and crowdsourced data that can identify blocked routes. We use our customized state-of-the-art human-guided flood mapping technique [11] to prune the flooded routes to reflect the situation on the ground relevant for better-informed transportation.
+
+**Data Visualization.** The majority of the visualization techniques in the literature place geo-tagged tweets on a map [12] or display users' structured input such as in [20]. However, these techniques fail for social media-based systems where geo-tagged tweets are relatively infrequent, and the information is buried inside unstructured texts. [10] showed the effectiveness of layered visualization systems, where each layer can contain a different source of information. In D-Record, each layer can include a different *Need* class or OpenStreetMap location features. To visualize the kinds of
+
+Image /page/2/Figure/8 description: A diagram illustrating a system for disaster relief coordination and response, which processes data from multiple sources to create a layered map for matching seekers and providers. The process is shown in three main parts. The first part details the data sources on the left: 'Classify and Geoparse Tweets' using Twitter and LNEx; 'Satellite Image for Flood Mapping' using satellite imagery to identify flooded locations; and 'OpenStreetMap and Crowd-sourced Location Features'. The outputs from these sources are processed to 'Prune Flooded Locations' and are then combined into a stack of map layers shown in the center. The layers, from top to bottom, are: Flood Areas, Twitter Medical/Rescue Needs, CS Medical/Rescue Available, OSM Medical/Rescue Available, Twitter Shelter/Food/Supplies Needs, CS Shelter/Food/Supplies Available, OSM Shelter/Food/Supplies Available, and a base layer from MapBox. Each layer is associated with a specific icon. The final part, at the bottom, is labeled 'Disaster Relief Coordination and Response' and contains a box for 'Map Directions & Seekers/Providers Matching', illustrated with icons of a person, a map, and another person, signifying the matching of those in need with those providing aid.
+
+Figure 1: D-record map layers and pipeline functions
+
+*Needs* effectively they are bound to their spatial footprints extracted using our tool LNEx [2].
+
+## 3 METHOD
+
+The following section describes the different data sources, the functionality of the pipeline, and the map layers we employ that support the visualization of information. Figure 1 contains a summary of map layers and their purpose.
+
+**Data Sources.** D-Record utilizes three data sources: Twitter, OpenStreetMap, and Satellite Images. Twitter data consists of two groups. The first group contains a set of two targeted streams from two disaster events, namely, the Chennai Flood in 2015 and Houston Flood in 2016 [2]. The second group of crisis tweets is from CrisisNLP [9] and CrisisLexT26[16] datasets labeled for their information type (e.g., affected individuals, donations, and volunteering). The first group drives the pipeline. The second is used to train a binary classifier to recognize the expression of *Need* in tweets and expand the concepts in the underlying ontology (see Section 3). Crowd-sourced data from volunteers (contributed as Excel sheets) contain information about shelters and help requests/offers (e.g., Chennai Flood - List of Corporation and Relief Centres<sup>5</sup>).
+
+The OpenStreetMap gazetteer provides location names and metadata (such as geo-coordinates) that produce the affordances of each location with respect to the needed help (e.g., a hospital can provide rescue and medical support) allowing for matching a help offer with a *Need* request from tweets. The Output of our flood mapping tool [11], which uses satellite images, determines the flooded geo-points and overlays this information on top of the map.
+
+**Location-Centric Ontological Modeling.** Determining spatially and temporally specific instances for coordinating resources *Needs* benefit from a domain-specific location-centric event ontology for social media data. Event location is key to both aggregation and analysis of related event instances. Therefore, we created an event ontology that uses information from past disaster and risk reports<sup>6</sup> to enrich existing disaster event vocabularies.
+
+The D-Record ontology<sup>7</sup> represents the temporal distinction between the response and relief phases of a disaster. The class "Concepts" includes two kinds of concepts: *Needs* and Availability. The class OSMFeatures covers the affordances of various location types (i.e., OSM map features<sup>8</sup>) in relation to the *Need* classes.
+
+<sup>&</sup>lt;sup>5</sup>https://rebrand.ly/reliefspreadsheet
+
+<sup>6</sup> such as by https://www.acaps.org/
+
+<sup>&</sup>lt;sup>7</sup>The ontology can be found at https://github.com/shrutikar/d-record.
+
+<sup>&</sup>lt;sup>8</sup>http://wiki.openstreetmap.org/wiki/Map\_Features
+
+Image /page/3/Figure/2 description: A flowchart detailing a data processing pipeline for disaster response and scenario dissemination. The process begins on the left with a 'Data' section, which includes three sources: 'Stream and Crowd Sourced Data' (with icons for Twitter and Excel sheets), 'Pre-Disaster Data' (with the OpenStreetMap logo), and 'Satellite Imagery' (with the NOAA logo). The data flows through a series of processing steps. Stream and crowd-sourced data undergo 'Semantic Filtering' and then 'Information Retrieval/Extraction'. This extraction process, also fed by pre-disaster data, branches into 'Geoparsing' (using LNEx), 'Seekers / Providers need classification', and 'OSM location classification with map feature types'. The 'Seekers / Providers' data is further processed by 'Event Ontology & Situational Data' using 'Lexicon NLP'. Satellite imagery is used for 'Flood Mapping'. The processed data streams converge: 'Geoparsing' and 'Event Ontology' data become 'Typed geolocated data', while 'OSM location classification' and 'Flood Mapping' data are used to 'Prune/Filter Locations'. Both of these outputs are then stored via 'Caching in ElasticSearch'. From the cache, 'map layers' are created using 'mapbox'. These layers are used for 'Seekers / Providers Matching'. This matching process leads to 'Provide turn-by-turn directions with contact information', which is also informed by 'Prune non-available routes' from the flood mapping. There is a feedback loop from providing directions back to the matching process. The final step is 'Scenario Dissemination', illustrated by a map with colored risk zones and a route between points X and A.
+
+Figure 2: D-record system architecture
+
+| <b><i>Need</i> Class</b>              | <b>CrisisNLP and CrisisLexT26 Classes</b>                                                     |
+|---------------------------------------|-----------------------------------------------------------------------------------------------|
+| Shelter/Food/<br>Supplies <i>Need</i> | donation_needs_or_offers_or_volunteering_services,<br>displaced_people_and_evacuation         |
+| Medical/Rescue<br>Help <i>Need</i>    | missing_trapped_or_found_people, deaths_reports, in-<br>jured_or_dead_people, affected_people |
+
+Table 1: Our mapping of CrisisNLP and CrisisLexT26 classes
+
+The set of keywords for a "Needs" concept was expanded using topic modeling learned from the labeled data for the two classes: Shelter/Food/Supplies and Medical/Rescue Help. The relevance probability of each word to the Need class/topic functioned as a feature while training our Need classifier as shown in Section 3. Finally, the sub-classes of the Needs concept are mapped to the OSMFeatures sub-classes representing the options for available help at each location type (i.e., the affordances). For example, hospital maps to the Medical/Rescue Help class and pharmacy maps to the Shelter/Food/Supplies Need class.
+
+**Text Classification.** To leverage the existing labeled data from CrisisNLP and CrisisLexT26, we combined the overlapping classes to form the two *Need* classes (see Table 1). We used tweets from these datasets to train an SVM-based classifier which allows Drecord to categorize any given crisis related text into one of the two classes: Shelter/Food/Supplies *Need* or Medical/Rescue Help *Need*. The *Need* class Shelter/Food/Supplies includes donations, volunteer services, and *Needs* for food, water, shelter, or clothes. The *Need* class Medical/Rescue Help comprises affected people, death reports, injured or dead people, missing or trapped people, and other medical and rescue-related information. The classified streaming text appears on the map for matching.
+
+The classifier had to capture tweet semantics. A simple Bag-of-Words model lacks context. Instead, feature engineering [22] vectorizes the text sentences to capture their semantics adequately. Before featurizing, the text was preprocessed by stemming, case folding and removing "noisy" lexical elements (such as URLs, non-ASCII characters, mentions, punctuations, dataset-specific stopwords, and hashtags). Finally, we designed an SVM classifier with lexicon-based features, TF-IDF vectors, and gensim's word2vec embeddings<sup>9</sup>.
+
+Using the relevance probabilities from topic modeling described earlier, we form lexicon-based two-tuple feature vectors (i.e., Shelter/Food/Supplies and Medical/Rescue Help). Each element in the vector represents the word frequency multiplied by its relevance score. This vector was concatenated with the other features from TF-IDF and gensim's word2vec embeddings, which captures the semantics of a word by looking at the context where it was mentioned. To address the class imbalance, SMOTE oversampled the minority class synthetically.
+
+**System Architecture.** D-record uses three major forms of data (see Figure 2): Streaming and Crowdsourced, pre-disaster, and satellite imagery data. The knowledge extracted from text streams (filtered using hashtags of the disaster event) and crowdsourced excel sheets provides situational awareness and the various kinds of help available at each location. The pre-disaster information available from OpenStreetMap (sliced using a bounding box of the disaster event) represented the available help confirmed or pruned using satellite images. The output from our flood mapping method [11] helped prune out routes that were unavailable during the matching process for the location seeking help.
+
+Our tool LNEx extracts and attaches each location to their latlong information which locates tweet information on the map, following *Need* classification. Since OpenStreetMap provides the metadata for locations (including names, types, and geo-coordinates), the flood map information is used to prune out the locations that are flooded and not accessible for help. D-record caches all of this information in Elasticsearch. The *Need* information and location of a tweet indicate where the help is needed while OSM location features identify available help. We create the map layers on top of MapBox<sup>10</sup>. The system then matches a *Need* in the ontology with the locations that can fulfill that *Need*. D-record ultimately prunes all routes containing a flooded section according to the flood map.
+
+## 4 RESULTS
+
+SVM with the SMOTE (to partially overcome class imbalance) and the Gradient Boosting Algorithm performed the best and achieved a 0.8 F-Score in the *Need* classification task. We tested it using the leave-one-out technique (i.e., testing on one dataset and training using the rest). D-record extracted locations from tweets to support plotting these tweets on the map. The experiments used our Chennai and Houston datasets, containing 169,838 and 415,057 tweets, respectively. From these tweets, LNEx extracted 85,564 locations from the Chennai dataset and 241,684 locations from Houston dataset. As for the number of OpenStreetMap location features, 1,103 and 2,826 locations were retrieved for the affected areas of Chennai and Houston, respectively. Other crowdsource data received from the excel sheets were relatively few, around 41 locations.
+
+## 5 DEMONSTRATION
+
+Figure 3 shows D-record where users can choose the time range to filter the data using ⓐ, the dataset using ⓑ, and select the map layers using ⓒ. To match a *Need*, the user clicks on the orange icons as in ① which shows a textbox with the tweet text and the extracted location. Users can click on a "Match Need" button to get the closest and non-flooded nearby location which can provide help for the given *Need*, as in ②. Users obtain the matched location information by hovering the mouse over the green icons as in ③. The tool will provide the full information of the matched location and the contact number if available with the turn-by-turn directions obtained using MapBox directions API<sup>11</sup> as shown in ④.
+
+<sup>9</sup>https://radimrehurek.com/gensim/
+
+<sup>10</sup> https://www.mapbox.com/
+
+<sup>11</sup>https://www.mapbox.com/help/define-directions-api/
+
+Image /page/4/Figure/2 description: A screenshot of a disaster management user interface called "HazardsSEES". The interface is split into a left sidebar and a main map view. The left sidebar has a logo, a time range selector set to "12-01-2015 1:30 PM - 01-30-2016 2:00 PM", a location selector with "Chennai" selected, and a legend for map icons such as "Flooded Areas", "Medical/Rescue Help Available", and "Shelter/Food/Supplies Needed". The main map area displays a map of Chennai with purple patches indicating flooded areas. On the map, there is a pop-up window showing a message: "@Uber\_Chennai - Need help in rescuing a group of girls from Perumalagaram Salai Chennai! #ChennaiMicro #ChennaiFloods" with a "Match Need" button. A blue line indicates a route on the map. Another pop-up identifies the "ACS Medical College And Hospital". In the bottom right, a box shows "Matched Location Info" for the hospital, including its address and turn-by-turn directions.
+
+Figure 3: Example screenshot of the D-record tool.
+
+## 6 CONCLUSIONS AND FUTURE WORK
+
+A domain-specific location-centric event ontology is crucial for situation awareness and disaster response. We demonstrated our D-record pipeline for *Need*-offer matching and discussed the functionality of the system and the multi-modal data used to fire its engine. The pipeline can be used to match coarse-grained *Needs* with possible suppliers meaningfully using location information available on the map. In the future, a finer-grained classifier can be designed to do more flexible or specific matching. Additionally, we plan to develop a custom entity extractor building on our prior work [1], to extract emerging entities during the onset of a disaster for more advanced spatiotemporal reasoning. We also plan to use weather data and background knowledge to mark flood-prone areas to help in preparedness in addition to response. To empower first and local responders, we intend to bring this pipeline for broader use by the disaster response community and port it to smartphones.
+
+## **ACKNOWLEDGMENTS**
+
+We would like to thank our collaborators Jiongqian (Albert) Liang, Jiayong Liang (Jay), Desheng Liu, and Nikhita Vedula from the Ohio State University for providing the flood mapping data.
+
+This research was supported by the NSF award EAR-1520870 "Hazards SEES: Social and Physical Sensing Enabled Decision Support for Disaster Management and Response". All views are those of the authors and do not necessarily reflect the views of the sponsor.
+
+

chatbot/rag/data/('Deep learning-based GLOF modelling for hazard assessment and risk management', '.pdf')_extraction.md

@@ -0,0 +1,427 @@
+
+
+## **ABSTRACT**
+
+Glacial Lake Outburst Flood (GLOF) has become a crucial aspect as the increase in the meltdown of glaciers results in the breach of unstable debris dams. Hence, it is essential to understand the nature of the glacial lakes for proper planning and development of the region in the long term. In this paper, a deep learning network is developed for GLOF hazard and risk assessment. The Shepard Convolutional Neural Network Fused Deep Maxout Network (ShCNNFDMN) is developed by fusing the Shepard Convolutional Neural Networks (ShCNN) and the Deep Maxout Network (DMN) based on regression analysis. Here, various data and feature attributes, like geometric properties, location properties, lake-based properties, and global properties are determined from the glacial lake data. Afterthat, hazard assessment is carried out based on these parameters by the ShCNNFDMN. Then, risk assessment is performed based on the hazard levels and the feature attributes. The ShCNNFDMN is analyzed based on metrics, such as Hazard modelling error, Risk prediction error, Mean Average Error (MAE), and R-Squared are found to produce values of 0.462, 0.423, 0.358, and 0.288, respectively. The proposed method is useful in applications, like infrastructure planning, taking preventive and mitigative actions in downstream areas of glacier lakes.
+
+### **ARTICLE HISTORY**
+
+Received 30 October 2023 Accepted 10 June 2024
+
+## **KEYWORDS**
+
+GLOF; risk assessment; hazard assessment; Shepard Convolutional Neural Networks; Deep Maxout Network
+
+## 1. Introduction
+
+Global warming and increased human activities have caused major changes in the drylands, and terrestrial water resources all over the world (Zhang et al. 2023). The high voluminous melting glacier water runoff, and accumulates in front of receding glaciers or depressions over sagging and thinning glacier surfaces creating moraine-dammed lakes (Sattar et al. 2021). Generally, glacier shrinking results in the expansion and creation of a huge number of glacial lakes on/in front of the glaciers (Wang et al. 2018). A huge volume of water is released because of certain failures in the moraine or sudden emptying of these lakes due to the overflow of dams, where the destructive events caused by released water sediment are termed Glacial Lake Outburst Floods (GLOF) (Liu et al. 2020; Worni, Huggel, and Stoffel 2013). The GLOF is characterised by its potential for erosion and high transport (Breien et al. 2008) and it can be effectively converted into debris
+
+flows of about 1.5 tm<sup>-3</sup> densities. It is also due to dynamic slope movements of the glaciers over the lakes, namely slopes (Wang and Goh 2021), landslides, rock falls, or icefalls (Awal et al. 2010; Emmer and Vilímek 2013). The incidence of GLOF has been high, especially in the Himalayan area, and has posed a tremendous risk to the local inhabitants (Li et al. 2018). Further, the quantity of water flow in the river affects various activities, such as the allocation of water for hydropower projects, irrigation systems, ecosystems, etc. (Ali et al. 2020).
+
+An in-targeted framework based on the adaption of climate change as well as glacier hazard management is performed from the lessons of the devastating GLOF events (Wang et al. 2018). Over the past decades, the GLOF hazard has increased the risk level due to the significant growth of the existing glacial lakes (Sattar, Goswami, and Kulkarni 2019). The impact of GLOF is transboundary, where the floods destroy most of the downstream infrastructure and kill hundreds of people
+
+thus causing huge damage (Khanal et al. 2015; Liu et al. 2020). Based on a regional scale, the GLOF hazard evaluation is often divided into three parts. The first stage involves gathering the fundamental data about glacial lakes, including their size, position, and surrounding terrain. In the second stage, glacial lakes that are potentially dangerous are identified using the variables that have been chosen. Finally, by using model simulation, field investigation, and high-resolution remote sensing images the study of dangerous glacial lakes is performed (Huggel et al. 2004; McKillop and Clague 2007; Osti, Egashira, and Adikari 2013; Wang et al. 2018). The implementation of different GLOF-related hazard remediations is carried out to develop a significant hazard management system (Sattar, Goswami, and Kulkarni 2019). The GLOF hazard assessment models are developed by considering the glacial lake outburst debris flow scenario (Worni et al. 2012) after earthquakes and small glacial lakes in the areas, which are familiar for landslides and collapses along the channels (Liu et al. 2020). Numerous approaches have been proposed in the past for performing GLOF risk assessment and these techniques vary from each other depending on the subjectivity, input data, choice of examined characteristics, quantity, structure, and so on (Washakh et al. 2019).
+
+In general, because of the potentially hazardous conditions in high mountains, the assessment of glacier hazards cannot be based solely on historical records and past events. Hence, it is necessary to implement various modelling approaches for the determination of present and future GLOF risks and hazards (Allen et al. 2016; Frey et al. 2018; Schaub et al. 2013; Schneider et al. 2014). Generally speaking, identifying exposure, vulnerability, and convergence of hazards are the fundamental elements utilised in risk assessment and management. The subset of machine learning techniques based on representation learning and artificial neural networks is known as deep learning (Wang et al. 2023). Images are processed using a convolutional neural network (Wu et al. 2023). for object detection and classification. The relative GLOF risk towards the downstream region is caused by the cumulative and peak discharge of the hazard components. This leads to overtopping, potential GLOF triggers, damming moraine conditions, and drainable volume functions of the lake. It is also helpful for the determination of hazards based on the vulnerability and exposure of elements downstream of the lake (Emmer 2018; Frey et al. 2018). Moreover, for responsive disaster preparedness and mitigation implementing detailed models and assessments are essential to address extreme-magnitude scenarios (Sattar et al. 2021).
+
+### 1.1. Motivation
+
+In the Himalayan region, glacial lakes have emerged and expanded quickly as a result of glacier recession brought on by climate change. The region is now more vulnerable to Glacial Lake Outburst Floods (GLOFs) as a result of the increased melting. The infrastructure and way of life in the nearby low-lying communities could suffer if potentially hazardous glacier lakes collapsed catastrophically. GLOF has become a crucial aspect in the economic and social stability of the downstream areas as the increase in the meltdown of glaciers results in the breach of unstable debris dams. Hence, it is essential to understand the nature of the glacial lakes for proper planning and development of the region in the long term. Also, the prevailing works carried out with a focus on GLOF risk assessment with their merits and problems confronted that motivated the development of the ShCNNFDMN.
+
+The major intention of this research is to implement a GLOF risk assessment model using the proposed ShCNNFDMN. Here, the GLOF risk assessment is carried out by considering the various properties of the glacial lake. Initially, properties, like geometric, location, lake-based, and global properties are excerpted from the input data. Thereafter, these features are applied to the ShCNNFDMN for assessing the hazard and risk. Here, the ShCNNFDMN is modelled by fusing the ShCNN with the DMN based on regression modelling. At first, the hazard assessment is carried out by the ShCNNFDMN based on the extracted features, and thereafter, the ShCNNFDMN carries out the risk assessment using the features as well as the hazard levels.
+
+The key contribution of this work is as follows:
+
+• Proposed ShCNNFDMN for risk/hazard assessment: In this work, GLOF risk and hazard assessment is carried out by using the ShCNNFDMN, which is formulated by combining ShCNN with DMN, based on regression modelling to improve the efficiency of assessment. Here, fusion is accomplished by applying the concept of Fractional Calculus (FC) to produce a regression model.
+
+The remaining part of the work is arranged as given: section 2 depicts the related works, section 3 portrays the ShCNNFDMN proposed in this work for GLOF risk assessment. The experimental results are shown in Section 4, and Section 5 concludes the study with recommendations for enhancements.
+
+## 2. Literature review
+
+Sattar et al. (2021) proposed a physical hydrodynamic model for modelling the lake outburst and hazard
+
+assessment of the glacial lakes. This method was developed for assessing the GLOF risks across the Barun-Arun river valley in Nepal. Here, the hydraulic flow of the river at six downriver encampments was analyzed, and the possible impact at every location was estimated by evaluating two low-level, two moderate-level, two high-level, and two extreme-level magnitudes were investigated for the current lake size and using this, the future dimension of the lake was modelled. Liu et al. (2020) developed a GLOF hazard assessment model for analyzing the risk of GLOF in Bhote Koshi Basin (BKB). This model assessed the GLOF hazard levels by regarding the situation that the numerous landslides caused by earthquakes evolved into debris in the outburst flood thereby enlarging the volume and discharge of the debris flow. This approach was effective in comprehending the data required for GLOF hazard analysis. However, the method was not appropriate for very high-hazard glacial lakes. Saifullah et al. (2020) devised a Remote sensing technique for GLOF risk assessment along the China-Pakistan Economic Corridor (CPEC). This research was aimed at investigating the formation and impacts of barrier lakes, supra-glacial, and end moraine based onsitu and remote sensing approaches in the Hunza River basin along CPEC considering the peak discharges and volume of the lakes. Emmer and Vilímek (2013) studied the Lake and breach hazard assessment model for assessing the risks of GLOF of moraine-dammed lakes. In this work, various methods for assessing the lake and breach hazards were outlined and the approaches were applied for assessing the hazard levels of the moraine-dammed lakes in Cordillera Blanca (Peru).
+
+Wang et al. (2018) introduced a Hydrologic Engineering Center-River Analysis System (HEC-RAS) for carrying out the integrated hazard assessment of the Cirenmaco glacial lake. This method combined various approaches, like 2D hydraulic modelling, bathymetric survey, and remote sensing for assessing the hazards presented by the Cirenmaco glacial lake. The results produced by the approach were shown to be highly effective in mitigating the risks, but specific care was needed when examining the outcomes of the model because of the complicated nature of GLOFs. Sattar, Goswami, and Kulkarni (2019) developed a two-dimensional hydrodynamic modelling approach for assessing the GLOF hazards of South LhonakLake. Here, a realistic bathymetric framework was constructed for depicting the various GLOF scenarios, and the hazard probability was assessed by utilising one and two-dimensional hydrodynamic modelling techniques. This method made it easier to build structures in the middle of the flow channel and may help manage the risk that extreme flow occurrences bring to areas downstream. Unfortunately, the approach did not provide in-
+
+situ measurements of the geomorphic characteristics over the flow channel or an analysis of engineering properties. Frey et al. (2018) proposed a Scenario-based multisource GLOF hazard mapping technique for assessing the risks of multisource GLOFs. Here, the cascaded mass movement of debris was simulated by utilising a chain of interrelating numerical models. Later, vulnerability evaluations for breach formation and rock-ice avalanches were employed for defining the events of related probability and varying magnitude. However, it failed to enhance the hazard mapping in urban areas that includes a detailed computation of the impacts of urban infrastructure on the flow behaviour of GLOFs. Li et al. (2018) used an Unmanned Surface Vessel (USV) and remote sensing to assess the hazards of GLOF in the Jialong Co glacial lake. Geomorphological analysis, volume assessment, and area dynamics were used for estimating the hazard potential of Jialong Co. The evolution of the Jialong Co in the last years was analyzed using Sentinel 2 satellite and Landsat series image data. Then, the volume, and underwater topography of the lake were detected utilising USVs for bathymetric measurements. Later, the hazard potential was determined by combining the geomorphic conditions, volume data, and area dynamics obtained from in-situ measurements. This plan assisted in determining the volume and underwater topography of the lake, but it did not improve accuracy or carry out longterm geographic monitoring of the moraine dam. Ahmed et al. (2022). uses a combination of remote sensing, GIS, and dam break modelling to evaluate the GLOF danger of Gangabal Lake, which is situated in the Upper Jhelum basin of the Kashmir Himalaya. Glab-Top-2, multitemporal satellite data, and the Cosi - Corr model were also used to evaluate the parameters of the Harmukh glacier, which feeds Gangabal Lake. Wangchuk and Tsubaki (2024) studied the possible effects of a GLOF coming from one of Bhutan's biggest and fastest-growing glacial lakes, Thorthomi, which is a glacial lake in the Phochhu River Basin. The findings here highlight the critical necessity to comprehend and get ready for the possible aftermath of a GLOF from Thorthomi Lake in order to lessen the effects on downstream ecosystems, economies, and societies.
+
+## 3. Proposed ShCNNFDMN-based glof modelling for hazard assessment and risk management
+
+In the current era, the dramatic rise in temperature worldwide due to global warming has resulted in an increase in the number of glacier lakes as well as the dimension of the existing ones. These phenomena have led to a potential risk of GLOF that can wreak havoc on man-made infrastructure and the natural environment. Hence, it is necessary to assess the risk of GLOF to avoid and mitigate any undue casualties. This paper proposes a hybrid ShCNNFDMN-based GLOF risk assessment technique, which is implemented as follows. Gathering information from the dataset is the primary step in the risk assessment process. After that, a number of features, including geometric, locationbased, lake-based, and global properties, are ascertained by applying the input data to the data and feature attribute extraction phase. The location properties that are taken into consideration are indicated by the latitude, longitude, and altitude, with the geometric property being the shape of the lake. Additionally, global characteristics like the Global Terrestrial Network for Glaciers (GTN-G) region and the Universal Transverse Mercator (UTM) grid zone, as well as lake-based parameters like lake area, perimeter, water volume, type, and uncertainty, are also extracted from the input data. Following the mining of the data and feature properties, the suggested hybrid deep learning method ShCNNFDMN is used to assess the hazards. The ShCNN (Ren et al. 2015) and DMN (Sun, Su, and Wang 2018) are combined to create the suggested ShCNNFDMN. The GLOF hazard is evaluated in this instance and assigned to one of the following five categories: Very High (VH), High (H), Medium (M), Low (L), and Very Low (VL). Lastly, the suggested ShCNNFDMN is used to execute the risk assessment based on the data, feature properties, and hazard levels. The GLOF risk is also divided into five categories, which include VH, H, M, L, and VL. The suggested ShCNNFDMN's structural view for the GLOF risk assessment is displayed in Figure 1.
+
+### 3.1. Data acquisition
+
+The initial step in GLOF assessment is to aggregate the data from the GLOF inventory dataset (https://zenodo. org/record/4477945#.ZGXf7nZBxPZ), which can be formulated as.
+
+$$C = \{C_1, C_2, \dots, C_i, \dots, C_c\}$$
+ (1)
+
+where,  $C_i$  refers to the *i*th record contained in the dataset  $C_i$ which is considered for GLOF risk prediction, and c specifies the total number of records contained in the dataset.
+
+### 3.2. Feature attributes
+
+Once the data  $C_i$  is retrieved, it is subjected to the extraction of various feature attributes, such as geometric properties, location properties, lake-based properties, and global
+
+properties are determined from the input data. The main intention of this task is to identify the key features of the lake, thereby enabling efficient hazard and risk assessment.
+
+#### 3.2.1. Geometric properties
+
+The geometric property of the glacial lake is considered an important aspect as it determines the stability of the lake. Here, the shape of the lake is taken into account and is found from the input lake image. At first, the boundary values of the lake are computed and then the image is split up into multiple grids. The grids that are contained inside the boundary of the lake are assigned a value of "1", and those falling outside the boundary are allocated a value of "0". Thus, a shape feature matrix comprising of the binary value is obtained from the input lake image, and the shape index matrix thus generated is represented as N.
+
+#### 3.2.2. Location properties
+
+Location is another significant feature that influences the severity of calamities caused by GLOF. The glacial lakes are generally located in uninhabited and remote mountain valleys, however, wide-ranging GLOFs can cause significant damage to properties as well as multiple casualties downstream up to tens of kilometres. Here, various location attributes, such as longitude, latitude, and altitude are computed from the input data, as the location of the glacial lake can be accurately found using these attributes.
+
+- (i) *Longitude*: Longitude is a reference with respect to the prime meridian and it refers to the angular distance (decimal degree) to the west or east of the prime meridian. The longitude feature is represented as  $J_1$ .
+- (ii) Latitude: Latitude gives the location of any place with respect to the equator, and is given as a measure of angular distance (decimal degree) to the south or north of the equator. Most of the glaciers are located in higher-latitude regions, such as the Arctic and Antarctic. But the low-latitude regions with high mountain ranges, like the Himalayas and Andes also contain glaciers. The latitude of the glacial lake measured is indicated as  $J_2$ .
+- (iii) Altitude: This attribute refers to the elevation of the glacier lake above sea level and is measured in metres/feet. Glacial lakes are more commonly found in high-altitude regions, and their numbers have increased with the retreat of glaciers. Let  $J_3$  designate the altitude feature.
+
+#### 3.2.3. Lake-based properties
+
+The devastating power of GLOF depends on the characteristics of the glacial lake and hence, these parameters
+
+Image /page/5/Figure/3 description: A flowchart titled 'Figure 1. Structural view of the proposed ShCNNFDM N for GLOF risk assessment.' The diagram illustrates a process for risk assessment. At the top, under 'Data and feature attributes', four categories of properties are listed with their corresponding data points: 'Geometric properties' lead to 'Shape of the lake'; 'Location properties' lead to 'Latitude, longitude, and altitude'; 'Lake-based properties' lead to 'Lake area, perimeter, water volume, lake type, and uncertainty'; and 'Global properties' lead to 'UTM grid zone and GTN-G region'. An arrow points down from 'Data and feature attributes' to two parallel processes: 'Risk assessment' on the left and 'Hazard assessment' on the right. The 'Hazard assessment' box contains a 'Proposed ShCNNFDM N' model, which is composed of 'Shepard Convolutional Neural Networks (ShCNN)' and 'Deep Maxout Network (DMN)'. The output of this assessment is a 'Hazard class'. The 'Risk assessment' box also contains a 'Proposed ShCNNFDM N' model, which receives inputs from 'ShCNN' and 'DMN'. This risk assessment process also takes the 'Hazard class' as an input. The final output from the 'Risk assessment' box is the 'Risk class'.
+
+Figure 1. Structural view of the proposed ShCNNFDMN for GLOF risk assessment.
+
+have to be considered while evaluating the risk assessment. The lake-based properties, such as lake area, perimeter, water volume, lake type, and uncertainty (Zheng et al. 2021) are considered here.
+
+- (i) Lake area: Lake area indicates the surface area of the lake and is obtained by multiplying the length and width of the lake and is expressed as  $J_4$ . Lake area is measured in square metres and is measured based on the UTM.
+- (ii) *Perimeter*: The total length of the shoreline line of the lake gives the perimeter value and is indicated
+- (iii) Water volume: The volume of the lake is measured by finding the product of the surface area of the
+
+- lake, and is expressed in cubic metres. The water volume feature of the glacial lake is characterised as  $J_6$ . The destructive potential of GLOF depends on the volume of water contained in the lake.
+- (iv) Lake type: The impact of GLOF depends on the type of the glacial lakes, which can be either Ice (I), Moraine (M), or others (O). Moraine-dammed glacial lakes are found to be a more prominent cause of GLOF than other glacial lakes. The laketype feature is symbolised as  $J_7$ .
+- (v) Uncertainty: This parameter deals with the uncertainty of the measurements of the glacial topology obtained using remote sensing techniques, and is considered so that the values can be analyzed as per their relevance. Uncertainty of the lake area is
+
+jan
+
+expressed as,
+
+$$J_8 = J_6/A \times A^2/2 \times 0.6872 \tag{2}$$
+
+where A refers to the image's spatial resolution, I<sub>6</sub> designates the perimeter, and  $J_8$  represents the uncertainty feature.
+
+#### 3.2.4. Global properties
+
+In addition to the above features, the global properties, such as the UTM grid zone, and GTN-G region of the glacial lake are considered to improve the risk assessment efficiency.
+
+- (i) UTM grid zone: This refers to the UTM zone in which the glacial lake is present. The UTM coordinate system partitions the globe into 60 zones, and in all zones, the coordinates are measured in northings and eastings in metres, and each zone has a width of 6° longitude. The UTM grid zone is designated as  $J_9$ .
+- (ii) *GTN-G region*: The GTN-G region in which the glacial lake is located is termed the GTN-G region feature, and is characterised by  $J_{10}$ . These regions are essential in analyzing the variations in glaciers and other attributes in a region. Once the various features are excerpted from the database, they are combined to obtain a feature vector as given below,
+
+$$J = \{J_1, J_2, \dots, J_{10}\}$$
+ (3)
+
+where,  $J_1$  is the latitude,  $J_2$  specifies the longitude,  $J_3$ symbolises the altitude,  $J_4$  refers to the lake area,  $J_5$ characterises the perimeter of the lake, J<sub>6</sub> denotes the water volume,  $J_7$  represents the lake type,  $J_8$  is the uncertainty feature,  $J_9$  exemplifies the UTM grid zone, and  $J_{10}$  terms the GTN-G region. The feature vector is generated J and the shape index matrix N is applied to the proposed ShCNNFDMN for hazard assessment.
+
+### 3.3. Hazard assessment
+
+A hazard assessment is carried out to find the probability that the water will be released from a glacial lake. This process is carried out by identifying the potentially hazardous lake first and then assessing the probability of water release based on the hazard index of the lake from the features extracted J. Here, hazard assessment is carried out by using the proposed ShCNNFDMN, which is created by combining the ShCNN (Ren et al. 2015) with DMN (Sun, Su, and Wang 2018) based on regression analysis. A statistical
+
+method for determining the relationships between variables in a given set of data is regression analysis. It can evaluate the statistical significance of the association, or the probability that the correlation is the product of chance, as well as show how strong the relationship is. One of the two main purposes of regression analysis is to either predict the value of the dependent variable for those for whom some information about the explanatory factors is available, or to assess the effect of an explanatory variable on the dependent variable. The ShCNNFDMN is fed with the shape index matrix N and a feature is generated J for assessing the GLOF hazard. The ShCNNFDMN comprises three parts, such as the ShCNN model, the DMN model, and the ShCNNFDMN layer. Initially, the shape index matrix N is subjected to the ShCNN model, which classifies the hazard into either of the five categories, such as VH, H, M, L, and VL. Thereafter, the output of the ShCNN is forwarded to the ShCNNFDMN layer, which fuses the output of the ShCNN with the feature vector *J* based on regression modelling, which is applied to determining the relationship between the two inputs, thereby enhancing the effectiveness of hazard assessment. Later, the fused output generated by the ShCNNFDMN layer is subjected to the DMN model for hazard assessments along with the shape index matrix N. Here, regression modelling is accomplished by applying the concept of Fractional Calculus (FC) (Bhaladhare and Jinwala 2014). Figure 2 demonstrates the architectural view of the proposed ShCNNFDMN for hazard assessment, and the various processes undertaken are expounded on in the ensuing subsections.
+
+#### 3.3.1. ShCNNmodel
+
+Initially, the hazard assessment is accomplished by using the ShCNN (Ren et al. 2015) by applying the shape index matrix N. The ShCNN is a type of CNN that is developed by incorporating the Shepard technique for altering the basic CNN to attain Translation Variant Interpolation (TVI). The key advantage of using the ShCNN is its low computational cost and its ability in realising end-to-end TVI operations for intermittently spaced data. The network attains superior results by adding a smaller number of feature maps and optimisation of the TVI process. The Shepard structure computes the weights of the familiar pixels based on the spatial distance between them and the target pixels differently, and its convolution form is formulated
+
+$$H_1 k = \begin{cases} (Z*N)_k / (Z*L)_k & \text{if } L_k = 0\\ N_k & \text{if } L_k = 1 \end{cases}$$
+ (4)
+
+Image /page/7/Figure/3 description: An architectural diagram illustrating the proposed ShCNNFDMN model for hazard assessment. The diagram shows a multi-stream process. In the top stream, a 'Shape index matrix N' is fed into an 'ShCNN model', which is represented by a linear sequence of seven blocks, producing 'Output H1'. In the middle stream, a 'Feature vector J' is input to the 'ShCNNFDMN layer'. This layer, which also receives 'Output H1', contains two components: 'Regression Modeling' and 'Fusion'. The output of this layer is 'Output H2'. In the bottom stream, the 'Shape index matrix N' is also input into a 'DMN model'. This model is depicted as a more complex structure with three rows of interconnected blocks. The final output of the entire system, originating from the DMN model, is the 'Risk level R'.
+
+Figure 2. Architectural view of the proposed ShCNNFDMN for hazard assessment.
+
+wherein, N represents the shape index matrix to the ShCNN, k signifies the image coordinate,  $H_{1k}$  refers to the output, \* implies the convolution function, L characterises the binary operator, which has a value  $L_k=0$  in case the pixel values are not known, and Z refers to the kernel function that has a weight inversely proportional to the separation among the pixel under consideration and  $Z_k=1$ .
+
+The characterisation of the convolutional kernel is the major factor that affects the interpolation result in the Shepard framework, and so a Shepard interpolation layer is employed to allow a kernel design that has good flexibility and is data-driven.
+
+##### Shepard interpolation layer:
+
+The expression given below describes the Shepard interpolation layer's feed-forward pass,
+
+$$G_{j}^{d}(G^{d-1}, L^{d}) = \alpha \left( \sum_{l} \frac{Z_{jl}^{d} * G_{l}^{d-1}}{Z_{jl}^{d} * L_{l}^{d}} + p^{d} \right), \quad d = 1, 2, 3, \dots \tag{5}$$
+
+Here, d represents the index of layer,  $G^{d-1}$  indicates the input of the present layer, whose mask is characterised as  $L^d$ , the term j in  $G_j^{d-1}$  and  $G_j^d$  denotes the feature map index in the (d-1) th and dth layer, respectively,  $Z_{il}$  signifies the trainable parameters and convolution of
+
+[Icon of a ship in a circle]
+
+ $Z_{il}$  is performed with the mask of the present layer  $L^d$  in the denominator and activations of the final layer in the numerator.  $G^{d-1}$  can signify the feature maps of the convolution or pooling layers of the CNN, and this can also be considered as the preceding Shepard interpolation layer. The term  $\alpha$  is used to impose non-linearity in the network and *p* symbolises bias.
+
+Figure 3 explicates the structure of the ShCNN, where it is applied with the shape index matrix N. The various inputs, such as feature maps/images, and masks representing the position where interpolation must happen are fed as input to the Shepard interpolation layer. Any complex interpolation operator can be constructed by applying the interpolation layer in a repeated manner with several non-linear layers. The mask refers to the binary map which has a value "0" for missing areas and "1" for familiar areas, and the same kernel is used with the mask as well as the image. The insignificant values of the preceding convolved mask  $Z^d*L^d$  are zeroed and a threshold is applied to generate the mask of the (d+1)th layer. Shepard interpolation layer with multiple stages is required to learn the sophisticated manner of propagation in processes where relatively bigger missing regions, like inpainting, occur. The output generated by the ShCNN is represented as  $H_1$ .
+
+#### 3.3.2. ShCNNFDMN layer
+
+Once the output of the ShCNNH<sub>1</sub> is obtained, it is fed to the ShCNNFDMN layer together with the location properties, lake-based properties, and global properties of the lake. One of the main advantage of ShCNNFDMN layer is simple and robust. The location properties are given by  $P_1 = \{J_1, J_2, J_3\}$  which include the latitude  $J_1$ , longitude  $J_2$ , and altitude  $J_3$ . Further, the lake-based properties are given by  $P_2 = \{J_4, J_5, J_6, J_7, J_8\}$  that comprise the lake area  $J_4$ , perimeter  $J_5$ , water volume  $J_6$ , lake type  $J_7$ , and uncertainty  $J_8$ . The global properties of the lake  $P_3 = \{J_9, J_{10}\}$  incorporate the UTM grid zone  $J_9$  and GTN-G region  $J_{10}$ . Here, the features are fused with the classified output of the ShCNN by using the concept of regression modelling. The relationship between the features and the classified output can be determined by using regression modelling, which is carried out here by utilising FC (Bhaladhare and Jinwala 2014), which is a subdivision of applied mathematics that solves the integral and derivative equations by using the Laplace transform. Here, the problem is first solved by converting it into the Laplace domain and later, the actual solution is determined by applying inverse transform. Here, the features excerpted as considered to be taken at various time intervals, and the
+
+weighted features are combined with the classified output of the ShCNN. This process is formulated as follows,
+
+At the interval t, the output of the ShCNNFDMN layer is expressed as,
+
+$$g = \sum_{q=1}^{n_1} P_{1q} \omega_q \tag{6}$$
+
+Here,  $P_1$  refers to the location properties of the lake comprising latitude, longitude, and altitude,  $n_1$  refers to the number of location properties, and  $\omega$  refers to the weight coefficients.
+
+When the time interval t-1 is considered, then the output of the ShCNNFDMN layer is obtained by considering the lake-based properties, and this is mathematically modelled as,
+
+$$g_1 = \sum_{q=1}^{n_2} P_{2q} \omega_q \tag{7}$$
+
+where,  $P_2$  symbolises the lake-based properties and  $n_2$ represents the number of lake-based properties. Further, the output of the ShCNNFDMN layer at time instance t-2 is formulated based on the global properties of the glacial lake, and is modelled as,
+
+$$g_2 = \sum_{q=1}^{n_3} P_{3q} \omega_q \tag{8}$$
+
+Here,  $P_3$  represents the global properties of the lake, such as the UTM grid zone, and GTN-G region, and  $n_3$  characterises the number of global properties.
+
+Moreover, the output of the ShCNNFDMN at  $(t-3)^{th}$  intervals is considered to be the classified layer of the  $ShCNNH_1$ , and these are combined by using FC. Applying the concept of FC (Bhaladhare and Jinwala 2014),
+
+$$\begin{aligned} y(t+1) &= u \cdot y(t) + \frac{1}{2}u \cdot y(t-1) \\ &\qquad + \frac{1}{6}(1-u)y(t-2) \\ &\qquad + \frac{1}{24}u(1-u)(2-u) \cdot y(t-3) \end{aligned}\tag{9}$$
+
+Substituting the respective values ShCNNFDMN layer, the equation depicted above can be written as,
+
+$$H_2 = u \cdot g + \frac{1}{2}u \cdot g_1 + \frac{1}{6}(1 - u)g_2 + \frac{1}{24}u(1 - u)(2 - u) \cdot H_1$$
+ (10)
+
+Shape index matrix N
+
+Image /page/9/Figure/4 description: A diagram illustrating the architecture of a neural network. The data flows from left to right through a series of layers. The sequence of layers and the dimensions of the data after each layer are as follows: two light blue 'Conv layers' resulting in dimensions of 1x26x32 and 1x26x64 respectively; a light yellow 'Leaky ReLU layer' with an output of 1x26x64; a light blue 'Maxpooling layer' resulting in 1x13x64; an olive green 'Flatten layer' with an output of 1x832; and two light yellow 'Dense layers' with outputs of 1x32 and 1x5. The final output of the network is labeled 'Output H₁'.
+
+Figure 3. Structure of the ShCNN.
+
+Applying the values of g,  $g_1$ , and  $g_2$  from equations (6-8),
+
+$$\begin{aligned}H_{2} ={}& u \cdot \sum_{q=1}^{n_{1}} P_{1q} \omega_{q} + \frac{1}{2} u \cdot \sum_{q=1}^{n_{2}} P_{2q} \omega_{q} \\ & + \frac{1}{6} (1 - u) \sum_{q=1}^{n_{3}} P_{3q} \omega_{q} \\ & + \frac{1}{24} u (1 - u) (2 - u) \cdot H_{1}\end{aligned}\qquad (11)$$
+
+Here,  $H_1$  characterises the output of the ShCNN and is found using equation (4),  $H_2$  which signifies the output generated by the ShCNNFDMN layer, and u symbolises a constant that signifies the derivative order in FC.
+
+#### 3.3.3. DMN model
+
+The DMN is applied with the shape index matrix N, and the output  $H_2$  produced by the ShCNNFDMN layer for performing hazard assessment. The output produced by the ShCNN is also a hazard assessment level, and the output produced by the ShCNN is combined with the features and given to the DMN, thereby boosting the efficiency of hazard assessment. The ShCNN and DMN perform classification at various stages, and thus in the DMN, more refined features are used along with the shape-indexed matrix, which minimises the classification cost as well. Here, the DMN (Sun, Su, and Wang 2018) is mainly applied for hazard assessment due to its ability to produce superior outcomes in resource-constrained scenarios. DMN predict more consistently and accurately. The DMN is applied with an inputS;  $(S = \{N, H_2\})$  created by considering the shape indexed matrix N and the output of the
+
+ShCNNFDMN layer  $H_2$ , and the various processes taking place in the DMN are explicated as given below,
+
+$$b_{x,y}^{1} = \max_{y \in [1,v_1]} S^T V_{\dots xy} + w_{xy}$$
+ (12)
+
+$$b_{x,y}^{2}T = \max_{y \in [1,v_{2}]} b_{x,y}^{1}TV_{...xy} + w_{xy}$$
+ (13)
+
+$$b_{x,y}^{e}T = \max_{v \in [1,v_e]} b_{x,y}^{e-1}TV_{...xy} + w_{xy}$$
+ (14)
+
+$$b_{x,y}^{f}T = \max_{y \in [1,y_f]} b_{x,y}^{f-1}TV_{\dots xy} + w_{xy}$$
+ (15)
+
+$$R_{x} = \max_{y \in [1, y_{f}]} b_{x, y}^{f} \tag{16}$$
+
+Here, x refers to the total count of layers in DMN,  $A_x$  refers to the hazard assessed by the DMN,  $v_e$  designates the overall count of units in the  $e^{th}$  layer,  $w_{xy}$  characterises the bias,  $V_{xy}$  symbolises the weight, and  $b_{x,y}^e$  designates the output produced by the eth layer. As depicted in the above equation, a max-pooling function is used by the DMN, and the maximal value of the output generated in each layer is fed to the consecutive layers. When v > 2 the DMN can estimate any conventional nonlinear activation functions. The architectural view of the DMN is depicted in Figure 4 and the output produced is indicated as R.
+
+### 3.4. Risk management using the proposed ShCNNFDMN
+
+After the hazard level *R* is identified by the ShCNNFDMN, risk assessment is carried out based on the hazard levels *R* 
+
+Image /page/10/Figure/1 description: The number 106 is shown in black text against a white background.
+
+Image /page/10/Figure/2 description: A diagram illustrating the architecture of a convolutional neural network. The model processes an input labeled 'S = {N, H₂}' through a series of layers to output a 'Risk level R'. The architecture is laid out in three rows. The first row consists of an Input layer (64x64x3), a Convolution layer (62x62x64), a Lamda layer (62x62x32), a Batch Normalization layer (62x62x32), and a Max pooling layer (31x31x32). The second row continues with a Convolution layer (29x29x128), a Lamda layer (29x29x64), a Batch Normalization layer (29x29x64), a Max pooling layer (14x14x64), a Dropout layer (14x14x64), and another Convolution layer (12x12x256). The third row includes a Lamda layer (12x12x64), a Batch Normalization layer (12x12x64), a Dropout layer (6x6x64), a Flatten layer (1x2304), and a Dense layer (1x1) which produces the final output. A legend at the bottom clarifies the layer types corresponding to different colors: Input layer, Lamda, Max pooling layer, Flatten, Convolution layer, Batch Normalization layer, Dropout layer, and Dense.
+
+Figure 4. Architectural view of the DMN.
+
+and various feature attributes N, and J. Risk assessment is extremely essential as it gives a clear picture of the aftereffects of a GLOF, thereby helping in taking effective risk reduction strategies. Glacial lakes that are assessed to be under HVH hazard levels pose a greater risk to the communities residing downstream (Zheng et al. 2021). Here, risk assessment is performed using the ShCNN based on the risk index, which can be obtained considering the exposure index and the hazard index. The exposure index refers to the potency of GLOF causing devastation to the infrastructures as well as life and hazard index terms the combined magnitude and likelihood of the GLOF. The ShCNNFDMN is already elucidated in section 3.3., and combines the ShCNN (Ren et al. 2015) and the DMN (Sun, Su, and Wang 2018) using regression analysis.
+
+## 4. Results and Discussion
+
+The results obtained during the experimentation of the ShCNNFDMN for GLOF risk assessment are detailed in this section. Further, the experimental set-up, evaluation measures, dataset, and analysis of the approach are also demonstrated.
+
+### 4.1. Experimental set-up
+
+The ShCNNFDMN for GLOF risk assessment proposed in this work is implemented on a system with the Python language using the glacial lake inventories dataset (https://zenodo.org/record/4477945#. ZGXf7nZBxPZ). Table 1 shows the experimental parameters of the proposed method.
+
+**Table 1.** Experimental parameters of the proposed method.
+
+| Methods          | Physical<br>hydrodynamic<br>model | GLOF Hazard<br>assessment | Lake and breach<br>Hazard assessment | HEC-RAS | Two-dimensional<br>hydrodynamic model | Scenario-based<br>multi-source GLOF | Proposed<br>ShCNNFDMN |
+|------------------|-----------------------------------|---------------------------|--------------------------------------|---------|---------------------------------------|-------------------------------------|-----------------------|
+| Epochs           | 20                                | 20                        | 25                                   | 20      | 20                                    | 20                                  | 30                    |
+| Batch size       | 32                                | 32                        | 32                                   | 32      | 32                                    | 32                                  | 64                    |
+| Learning<br>rate | 0.001                             | 0.001                     | 0.001                                | 0.001   | 0.001                                 | 0.001                               | 0.001                 |
+
+The bold values represent the best performance.
+
+Image /page/11/Figure/3 description: A geographical map showing the distribution of glacial lakes over a Digital Elevation Model (DEM) of a mountainous region. The map is bounded by coordinates from 70°0'0"E to 100°0'0"E longitude and from 20°0'0"N to 40°0'0"N latitude. A compass rose is present in the upper right corner. The legend indicates that the blue areas represent 'Glacial Lakes'. The DEM values, shown in grayscale, range from a 'Low' of 232 to a 'High' of 8233. The map includes a scale bar marked up to 1,240 Kilometers and a textual scale of '1 cm = 210 km'.
+
+Figure 5. Glacial Lake Dataset.
+
+### 4.2. Dataset description
+
+The data used for the experimentation of the ShCNNFDMN is acquired from the glacial lake inventories dataset and has been mapped in Figure 5 (https://zenodo.org/record/4477945#.ZGXf7nZBxPZ). This dataset comprises the data of the glacial lakes present in the Third Pole region. The data encompassed are acquired over a period of 15 years from 1990 to 2015, Further, the data is modelled for considering future conditions for 2050 and 2100 on the Third Pole under an ice-free scenario, Representative Concentration Pathway (RCP) 8.5, RCP 4.5, and RCP 2.6. It contains various data regarding the glacial lake shape, location, uncertainty, lake water volume, hazard value, risk value, exposure value, etc.
+
+### 4.3. Evaluation measures
+
+Two parameters, such as Hazard modelling error, and Risk prediction error are considered to evaluate the supremacy of the ShCNNFDMN.
+
+(i) *Hazard modelling error*: The hazard modelling error is measured by finding the deviation of the predicted hazard value by the ShCNNFDMN from the expected values.
+
+$$HME = \frac{\sum_{i=1}^{n} |y_i - x_i|}{n}.$$
+ (17)
+
+where, HME refers to the Hazard Modelling Error,
+
+ $y_i$  refers to the prediction, $x_i$  refers to the true value, and n designates the total number of data.
+
+(ii) *Risk prediction error:* The risk prediction error is also found similar to the hazard modelling error by determining the difference between the predicted risk value from the anticipated risk value.
+
+$$RPE = \frac{\sum_{i=1}^{n} |y_i - x_i|}{n}.$$
+ (18)
+
+where, *RPE* refers to the Risk Prediction Error.
+
+(iii) *Mean Average Error (MAE):* The MAE, is a statistical measure of inaccuracies between paired observations that represent the same occurrence.
+
+$$MAE = \frac{\sum_{i=1}^{n} |y_i - x_i|}{n}.$$
+ (19)
+
+where, MAE refers to the Mean Average Error.
+
+(iv) *R-Squared:* The percentage of a dependent variable's variance that can be accounted for by an independent variable is expressed statistically as R-squared.
+
+$$R^2 = 1 - \frac{RSS}{TSS}. (20)$$
+
+where, RSS refers to the sum of squares of residuals, and TSS refers to the total sum of squares.
+
+### 4.4. Performance analysis
+
+The performance of the ShCNNFDMN is examined by considering the values of Hazard modelling error, and Risk
+
+prediction error determined when different k-fold values and training data percentages are used for various iterations.
+
+#### 4.4.1. Based on k-fold
+
+Figure 6 demonstrates the performance assessment of the ShCNNFDMN while varying the k-fold values. In Figure 6 (a), the analysis of the ShCNNFDMN based on hazard modelling error is exhibited. With a k-fold of 9, the value of hazard prediction error measured by the ShCNNFDMN is 0.558, 0.526, 0.518, 0.505, and 0.487, corresponding to 20, 40, 60, 80, and 100 iterations. Likewise, the assessment of the ShCNNFDMN in terms of risk prediction error is illustrated in Figure 6(b). The ShCNNFDMN is observed to have attained a risk prediction error of 0.495 for 20 iterations, 0.507 for 40 iterations, 0.497 for 60 iterations, 0.484 for 80 iterations, and 0.428 for 100 iterations with k-fold of 9. In Figure 6(c), the analysis of the ShCNNFDMN based on MAE is showed. With a k-fold of 9, the value of MAE measured by the ShCNNFDMN is 0.450, 0.414, 0.410, 0.395, and 0.370,
+
+corresponding to 20, 40, 60, 80, and 100 iterations. Figure 6 (d) explicates the assessment of the ShCNNFDMN on the basis of R-Squared. For k-fold of 9, the ShCNNFDMN recorded R-Squared of 0.346 for 20 iterations, 0.331 for 40 iterations, 0.324 for 60 iterations, 0.312 for 80 iterations, and 0.292 for 100 iterations.
+
+#### 4.4.2. Considering training data
+
+The evaluation of the ShCNNFDMN considers the hazard modelling error and risk prediction error while taking into account various training data percentages. Figure 7(a) explicates the assessment of the ShCNNFDMN on the basis of hazard modelling error. For 90% of training data, the ShCNNFDMN recorded hazard modelling error of 0.644 for 20 iterations, 0.582 for 40 iterations, 0.561 for 60 iterations, 0.540 for 80 iterations, and 0.484 for 100 iterations. Further, the assessment of the ShCNNFDMN concerning the risk prediction error is exhibited in Figure 7(b). The risk prediction error measured by the ShCNNFDMN
+
+Image /page/12/Figure/9 description: The image displays four line graphs, labeled (a), (b), (c), and (d), showing the performance assessment of a model named ShCNNFDM. Each graph plots a different performance metric against 'K-Fold' on the x-axis, which ranges from 5 to 9. Each graph includes five lines representing the model run with a different number of iterations: 20 (blue with 'x' markers), 40 (pink with '\*' markers), 60 (green with right-pointing triangle markers), 80 (teal with left-pointing triangle markers), and 100 (red with circle markers).
+
+(a) The top-left graph shows 'Hazard modelling error' on the y-axis, from 0.50 to 0.70. For all lines, the error generally decreases as K-Fold increases. The error is lowest for the model with 100 iterations (red line, from ~0.53 to ~0.48) and highest for the model with 20 iterations (blue line, from ~0.71 to ~0.56).
+
+(b) The top-right graph shows 'Risk prediction error' on the y-axis, from 0.45 to 0.70. Similar to the first graph, the error decreases as K-Fold increases and as the number of iterations increases. The 100-iteration model has the lowest error (red line, from ~0.54 to ~0.43), while the 20-iteration model has the highest (blue line, from ~0.72 to ~0.51).
+
+(c) The bottom-left graph shows 'MAE' (Mean Absolute Error) on the y-axis, from 0.38 to 0.52. The MAE decreases as K-Fold increases and as the number of iterations increases. The 100-iteration model performs best with the lowest MAE (red line, from ~0.45 to ~0.37), and the 20-iteration model performs worst with the highest MAE (blue line, from ~0.52 to ~0.45).
+
+(d) The bottom-right graph shows 'R-squared' on the y-axis, from 0.30 to 0.44. In this graph, the R-squared value decreases as K-Fold increases for all models. Contrary to the error metrics, a higher R-squared value is better, and the model with fewer iterations performs better. The 20-iteration model has the highest R-squared (blue line, from ~0.44 to ~0.34), while the 100-iteration model has the lowest (red line, from ~0.36 to ~0.29).
+
+Figure 6. Performance assessment of the ShCNNFDMN based on (a) Hazard modelling error, (b) Risk prediction error, (c) MAE, and (d) R-Squared.
+
+Image /page/13/Figure/3 description: The image contains four line graphs, labeled (a), (b), (c), and (d), which show the performance assessment of a model called ShCNNFDMN. Each graph plots a different performance metric against the percentage of training data, which ranges from 50% to 90% on the x-axis. All graphs compare five versions of the model based on the number of iterations: 20, 40, 60, 80, and 100.
+
+(a) The first graph plots 'Hazard modelling error' on the y-axis from 0.50 to 0.75. For all lines, the error decreases as the training data increases. The error is lowest for 100 iterations (red line, from ~0.56 to ~0.48) and highest for 20 iterations (blue line, from ~0.75 to ~0.64).
+
+(b) The second graph plots 'Risk prediction error' on the y-axis from 0.45 to 0.75. The trend is similar to the first graph, with the error decreasing with more training data and more iterations. The 100-iteration line drops from ~0.57 to ~0.44, while the 20-iteration line drops from ~0.74 to ~0.62.
+
+(c) The third graph plots 'MAE' (Mean Absolute Error) on the y-axis from 0.375 to 0.525. Again, the error decreases as training data and the number of iterations increase. The 100-iteration line goes from ~0.43 to ~0.36, and the 20-iteration line goes from ~0.53 to ~0.42.
+
+(d) The fourth graph plots 'R-squared' on the y-axis from 0.30 to 0.42. In this case, the R-squared value decreases as the training data percentage increases. The model with 20 iterations has the highest R-squared values (from ~0.425 to ~0.335), while the model with 100 iterations has the lowest (from ~0.39 to ~0.295).
+
+The caption below reads: 'Figure 7. Performance assessment of the ShCNNFDMN based on (a) Hazard modelling error, (b) Risk prediction error, (c) MAE, and (d) R-squared.'
+
+Figure 7. Performance assessment of the ShCNNFDMN based on (a) Hazard modelling error, (b) Risk prediction error, (c) MAE, and (d) R-Squared based on training data.
+
+with 20, 40, 60, 80, and 100 iterations is 0.620, 0.569, 0.524, 0.463, and 0.436, respectively, for 90% of training data. In Figure 7(c), the analysis of the ShCNNFDMN based on MAE is exhibited. With 90% of training data, the value of MAE measured by the ShCNNFDMN is 0.423, 0.422, 0.417, 0.387, and 0.364, corresponding to 20, 40, 60, 80, and 100 iterations. Likewise, the assessment of the ShCNNFDMN in terms of R-Squared is illustrated in Figure 7(d). The ShCNNFDMN is observed to have attained a R-Squared of 0.338 for 20 iterations, 0.318 for 40 iterations, 0.318 for 60 iterations, 0.311 for 80 iterations, and 0.296 for 100 iterations with 90% of training data.
+
+### 4.5. Comparative techniques
+
+The efficiency of the ShCNNFDMN is investigated by comparing the proposed technique with other prevailing
+
+GLOF risk assessment methods, such as the Physical hydrodynamic model (Sattar et al. 2021), GLOF Hazard assessment (Liu et al. 2020), Lake and breach Hazard assessment (Emmer and Vilímek 2013), HEC-RAS (Wang et al. 2018), Two-dimensional hydrodynamic model, and Scenario-based multi-source GLOF.
+
+### 4.6. Comparative analyis
+
+The analysis of the ShCNNFDMN is carried out to examine its superiority in performing GLOF risk/hazard assessment by taking into consideration different values of training data and k-fold.
+
+#### 4.6.1. Using K-fold
+
+The examination of the ShCNNFDMN based on k-fold is with respect to hazard modelling error and risk prediction error is displayed in Figure 8. The hazard
+
+Image /page/14/Figure/1 description: Two line graphs, labeled (a) and (b), compare the performance of seven different models. Both graphs plot an error metric on the y-axis against 'K-Fold' on the x-axis, which ranges from 5 to 9. Graph (a) shows 'Hazard modelling error' on the y-axis, ranging from 0.50 to 0.75. Graph (b) shows 'Risk prediction error' on the y-axis, ranging from 0.45 to 0.75. The seven models compared are: 'Physical hydrodynamic model' (blue), 'GLOF Hazard assessment' (pink), 'Lake and breach Hazard assessment' (green), 'HEC-RAS' (cyan), 'Two-dimensional hydrodynamic model' (red), 'Scenario-based multi-source GLOF' (yellow), and 'Proposed ShCNNFDMN' (dark gray). In both graphs, the 'Proposed ShCNNFDMN' model consistently has the lowest error across all K-Fold values. In graph (a), its error decreases from approximately 0.52 at K-Fold 5 to 0.47 at K-Fold 9. In graph (b), its error decreases from approximately 0.56 at K-Fold 5 to 0.43 at K-Fold 9. The other models generally show higher and more variable error rates.
+
+Image /page/14/Figure/2 description: A line graph, labeled (c), plots the Mean Absolute Error (MAE) on the y-axis against the K-Fold value on the x-axis. The x-axis ranges from 5 to 9, while the y-axis ranges from 0.375 to 0.525. The graph compares the performance of seven different models, each represented by a colored line with unique markers. The legend identifies six of the models: 'GLOF Hazard assessment' (dark blue line with 'x' markers), 'Lake and breach Hazard assessment' (pink line with star markers), 'HEC-RAS' (cyan line with triangle markers), 'Two-dimensional hydrodynamic model' (red line with circle markers), 'Scenario-based multi-source GLOF' (yellow line with diamond markers), and 'Proposed ShCNNFDMN' (gray line with pentagon markers). A seventh line, green with circular markers, is also present but not identified in the legend. The 'Proposed ShCNNFDMN' model consistently shows the lowest MAE across all K-Fold values, decreasing from approximately 0.445 at K-Fold 5 to 0.362 at K-Fold 9. Most other models also show a general downward trend in MAE as the K-Fold value increases.
+
+Image /page/14/Figure/3 description: A line graph, labeled (d), showing the relationship between R-squared values and K-Fold for seven different models. The x-axis is labeled 'K-Fold' and ranges from 5 to 9. The y-axis is labeled 'R-squared' and ranges from 0.30 to 0.44. The graph displays seven lines, each representing a different model, and all lines show a downward trend, indicating that the R-squared value decreases as the K-Fold value increases. The models and their approximate data points are as follows: 'Physical hydrodynamic model' (blue line with x markers) starts at approximately R-squared=0.445 at K-Fold=5 and ends at 0.348 at K-Fold=9. 'GLOF Hazard assessment' (pink line with star markers) starts at 0.443 and ends at 0.345. 'Lake and breach Hazard assessment' (green line with circle markers) starts at 0.435 and ends at 0.33. 'HEC-RAS' (cyan line with triangle markers) starts at 0.42 and ends at 0.33. 'Two-dimensional hydrodynamic model' (red line with circle markers) starts at 0.422 and ends at 0.325. 'Scenario-based multi-source GLOF' (yellow line with diamond markers) starts at 0.412 and ends at 0.312. 'Proposed ShCNNFDMN' (gray line with pentagon markers) consistently has the lowest R-squared values, starting at 0.398 and ending at 0.30.
+
+Figure 8. Comparative assessment of the ShCNNFDMN based on (a) Hazard modelling error, (b) Risk prediction error, (c) MAE, and (d) R-Squared considering k-fold.
+
+modelling error-based assessment of the ShCNNFDMN is demonstrated in Figure 8(a). The ShCNNFDMN is found to have produced a hazard modelling error of 0.483with k-fold of 8, while the prevailing risk assessment approaches, such as the Physical hydrodynamic model, GLOF Hazard assessment, Lake and breach assessment, HEC-RAS, Two-dimensional hydrodynamic model, and Scenario-based multi-source GLOF, measured hazard modelling error of 0.620, 0.594, 0.619, 0.612, 0.583, and 0.516, correspondingly. In Figure 8(b), the investigation of the ShCNNFDMN with respect to risk prediction error is presented, For a k-fold value of 8, the risk prediction error recorded by the Physical hydrodynamic model is 0.639, GLOF Hazard assessment is 0.554, Lake and breach Hazard assessment is 0.526, HEC-RAS is 0.544, Two-dimensional hydrodynamic model is 0.527, Scenario-based multi-source GLOF is 0.496, and ShCNNFDMN is 0.454. The MAE-based assessment ShCNNFDMN is demonstrated in Figure 8(c). The ShCNNFDMN is found to have produced a MAE of 0.379 with k-fold of 8, while the prevailing risk assessment approaches, such as the Physical hydrodynamic model, GLOF Hazard assessment, Lake and breach Hazard assessment, HEC-RAS, Two-dimensional hydrodynamic model, and Scenario-based multi-source GLOF, measured hazard modelling error of 0.470, 0.450, 0.448, 0.425, 0.416, and 0.415, correspondingly. In Figure 8(d), the investigation of the ShCNNFDMN with respect to R-Squared is presented, For a k-fold value of 8, the R-Squared recorded by the Physical hydrodynamic model is 0.366, GLOF Hazard assessment is 0.356, Lake and breach Hazard assessment is 0.355, HEC-RAS is 0.350, Two-dimensional hydrodynamic model is 0.344, Scenario-based multi-source GLOF is 0.330, and ShCNNFDMN is 0.313.
+
+#### 4.6.2. Based on training data
+
+Figure 9 represents the analysis of the ShCNNFDMN for GLOF risk/hazard assessment considering different percentages of training data. The ShCNNFDMN is examined for its efficacy on the basis of hazard
+
+Image /page/15/Figure/3 description: A set of four line graphs, labeled (a), (b), (c), and (d), comparing the performance of seven different models against varying amounts of training data. The x-axis for all graphs is 'Training data (%)', ranging from 50 to 90. The seven models compared are: 'Physical hydrodynamic model', 'GLOF Hazard assessment', 'Lake and breach Hazard assessment', 'HEC-RAS', 'Two-dimensional hydrodynamic model', 'Scenario-based multi-source GLOF', and 'Proposed ShCNNFDMN'.
+
+Graph (a) plots 'Hazard modelling error' on the y-axis (from 0.45 to 0.75). Generally, the error for all models decreases as training data increases. The 'Proposed ShCNNFDMN' model consistently has the lowest error, decreasing from approximately 0.54 to 0.48. The 'Physical hydrodynamic model' starts with the highest error at about 0.74.
+
+Graph (b) plots 'Risk prediction error' on the y-axis (from 0.45 to 0.75). The error trends downwards for most models with more training data. The 'Proposed ShCNNFDMN' model again shows the best performance with the lowest error, dropping from about 0.57 to 0.42.
+
+Graph (c) plots 'MAE' (Mean Absolute Error) on the y-axis (from 0.350 to 0.550). All models show a decrease in MAE as training data increases. The 'Proposed ShCNNFDMN' model has the lowest MAE, decreasing from around 0.435 to 0.36. The 'Physical hydrodynamic model' has the highest MAE across the range.
+
+Graph (d) plots 'R-squared' on the y-axis (from 0.30 to 0.44). In this graph, the R-squared values for all models generally decrease as the percentage of training data increases. The 'Physical hydrodynamic model' has the highest R-squared values, while the 'Proposed ShCNNFDMN' has the lowest.
+
+Figure 9. Comparative assessment of the ShCNNFDMN based on (a) Hazard modelling error, (b) Risk prediction error, (c) MAE, and (d) R-Squared based on training data.
+
+modelling error and this is portrayed in Figure 9(a). For training data of 80%, the hazard modelling error produced by the Physical hydrodynamic model, GLOF Hazard assessment, Lake and breach Hazard assessment, HEC-RAS, Two-dimensional hydrodynamic model, Scenario-based multi-source GLOF, and ShCNNFDMN is 0.601, 0.613, 0.627, 0.669, 0.592, 0.542, and 0.484, respectively. Likewise, Figure 9(b) illustrates the examination of the ShCNNFDMN on the basis of risk prediction error. The risk prediction error measured by ShCNNFDMNis 0.463, for 80% of training data, whereas the other techniques produced risk prediction error of 0.563 for the Physical hydrodynamic model, 0.553 for GLOF Hazard assessment, 0.517 for Lake and breach Hazard assessment, 0.541 for HEC-RAS, 0.516 for Two-dimensional hydrodynamic model, 0.495 for Scenario-based multi-source GLOF. The ShCNNFDMN is examined for its efficacy on the basis of MAE and this is portrayed in Figure 9(c). For training data of 80%, the MAE produced by the Physical hydrodynamic model, GLOF Hazard assessment, Lake and breach Hazard assessment, HEC-RAS, Two-dimensional hydrodynamic model, Scenario-based multisource GLOF, and ShCNNFDMN is 0.473, 0.452, 0.446, 0.443, 0.440, 0.423, and 0.380, respectively. The R-Squared-based assessment of the ShCNNFDMN is demonstrated in Figure 9(d). The ShCNNFDMN is found to have produced a R-Squared of 0.307 with training data 80%, while the prevailing risk assessment approaches, such as the Physical hydrodynamic model, GLOF Hazard assessment, Lake and breach Hazard assessment, HEC-RAS, Two-dimensional hydrodynamic model, and Scenario-based multi-source GLOF, measured hazard modelling error of 0.370, 0.350, 0.349, 0.339, 0.320, and 0.318, correspondingly.
+
+### 4.7. Comparative discussion
+
+The GLOF risk/hazard assessment technique proposed in this work using ShCNNFDMN is examined for its
+
+Table 2. Comparative discussion of the ShCNNFDMN.
+
+| Variations       | Metrics                      | Physical<br>hydrodynamic<br>model | GLOF Hazard<br>assessment | Lake and breach<br>Hazard<br>assessment | HEC-<br>RAS | Two-dimensional<br>hydrodynamic<br>model | Scenario-based<br>multi-source<br>GLOF | Proposed<br>ShCNNFDMM |
+|------------------|------------------------------|-----------------------------------|---------------------------|-----------------------------------------|-------------|------------------------------------------|----------------------------------------|-----------------------|
+| K-fold           | Hazard<br>modelling<br>error | 0.613                             | 0.634                     | 0.603                                   | 0.563       | 0.557                                    | 0.507                                  | 0.473                 |
+|                  | Risk<br>prediction<br>error  | 0.554                             | 0.604                     | 0.551                                   | 0.517       | 0.483                                    | 0.467                                  | 0.427                 |
+|                  | MAE                          | 0.450                             | 0.442                     | 0.437                                   | 0.436       | 0.415                                    | 0.398                                  | 0.363                 |
+|                  | R-squared                    | 0.348                             | 0.343                     | 0.330                                   | 0.328       | 0.325                                    | 0.313                                  | 0.297                 |
+| Training<br>data | Hazard<br>modelling<br>error | 0.565                             | 0.601                     | 0.578                                   | 0.638       | 0.573                                    | 0.527                                  | <b>0.462</b>          |
+|                  | Risk<br>prediction<br>error  | 0.592                             | 0.545                     | 0.592                                   | 0.466       | 0.453                                    | 0.435                                  | <b>0.423</b>          |
+|                  | MAE                          | 0.439                             | 0.427                     | 0.416                                   | 0.390       | 0.382                                    | 0.371                                  | <b>0.358</b>          |
+|                  | R-squared                    | 0.353                             | 0.340                     | 0.328                                   | 0.313       | 0.305                                    | 0.303                                  | <b>0.288</b>          |
+
+efficiency considering hazard modelling error and risk prediction error with respect to other related works and this is explicated in Table 2. The values of hazard modelling error, risk prediction error, MAE, and Rsquared portrayed are recorded when k-fold value of 9 and training data of 90% are considered. The superior values of 0.462, 0.423, 0.358, and 0.288 are obtained by the ShCNNFDMN assessment based on hazard modelling error, risk prediction error, MAE, and R-Squared for 90% training data. The usage of multistage classifier developed with ShCNN and DMN based on regression analysis effectively modelled the hazard and risk values, thus minimising the risk prediction error.
+
+##### Below are some of the factors that contribute to the proposed approach's superior performance:
+
+The proposed ShCNNFDMN is formulated by combining the ShCNN and DMN. The key advantage of using the ShCNN is its low computational cost and its ability in realising end-to-end TVI operations for intermittently spaced data. The network attains superior results by adding a smaller number of feature maps and optimisation of the TVI process. DMN perform classification at various stages, and thus in the DMN, more refined features are used along with the shapeindexed matrix, which minimises the classification cost. Here, the DMN is mainly applied for hazard assessment due to its ability to produce superior outcomes in resource-constrained scenarios. DMN predict
+
+Table 3. ANOVA test for the proposed method.
+
+|             | Sum of squares | Degrees of<br>freedom | Mean squares | F    | P value   |
+|-------------|----------------|-----------------------|--------------|------|-----------|
+| Factor<br>1 | 14.35          | 3                     | 7.845        | 18.2 | 7.823e-05 |
+| Error       | 6.15           | 15                    | 0.571254     | -    | -         |
+| Total       | 19             | 17                    | -            | -    | -         |
+
+more consistently and accurately. Thus, the proposed method ShCNNFDMN produce better performance.
+
+### 4.8. Analysis of ANOVA Test
+
+The ANOVA test analysis is shown in Table 3. It's known as the statistical approach. It is the procedure for breaking down the data variables into distinct parts in order to conduct further tests. To get information about the independent and dependent variables, the data are grouped using ANOVA. The F test is the another name for the ANOVA test. The ANOVA formula is written as follows:
+
+$$F = \frac{MST}{MSE}$$
+
+Where, F = ANOVA coefficient, MST = Mean sum of squares due to treatment, MSE = Mean sum of squares due to error.
+
+## 5. Conclusion
+
+This paper uses a unique deep-learning network called ShCNNFDMN to propose a fresh framework for assessing GLOF risks and hazards. In this case, evaluation is done based on a number of glacial lake attributes, including geometric, geographical, lake-based, and global attributes. First, the features are taken from the input data and evaluated by the ShCNNFDMN. Regression analysis is performed using the FC idea, and the resulting ShCNNFDMN is the result of the fusion of ShCNN and DMN. Initially, the ShCNN is trained with the geometric feature of the lake that was retrieved in the form of a shape-indexed matrix for classification. The final output and features are combined using the ShCNNFDMN layer, and for fine risk assessment, the DMN and shape-indexed matrix are applied. Superior values of 0.462, 0.423, 0.358, and 0.288 are noted in the ShCNNFDMN assessment based on hazard modelling error, risk prediction error, MAE, and R-Squared. The proposed method helps to prevent major hazards and risks. The proposed method is useful in various applications, like proper infrastructure planning, and taking preventive and mitigative actions in downstream areas of glacier lakes. The limitation of the proposed method is that it does not consider more evalution metrics to evaluate the model's performance. In future work, parameters, like the geomorphic parameters of the river channel, such as slope, and moraine can be considered to improve the risk assessment efficiency. Also, some more evaluation metrics will be considered to evaluate the model's performance.
+
+## Disclosure statement
+
+No potential conflict of interest was reported by the author(s).
+
+## **Funding**
+
+This work was supported by The International Talent Program of the Chinese Academy of Sciences [grant number 2021PC0045].
+

chatbot/rag/data/('Definition and classification system of glacial lake for inventory and hazards study', '.pdf')_extraction.md

@@ -0,0 +1,154 @@
+
+
+Abstract: Glacial lakes are not only the important refresh water resources in alpine region, but also act as a trigger of many glacial hazards such as glacial lake outburst flood (GLOF) and debris flow. Therefore, glacial lakes play an important role on the cryosphere, climate change and alpine hazards. In this paper, the issues of glacial lake were systematically discussed, then from the view of glacial lake inventory and glacial lake hazards study, the glacial lake was defined as natural water mainly supplied by modern glacial meltwater or formed in glacier moraine's depression. Furthermore, a complete classification system of glacial lake was proposed based on its formation mechanism, topographic feature and geographical position. Glacial lakes were classified as 6 classes and 8 subclasses, i.e., glacial erosion lake (including cirque lake, glacial valley lake and other glacial erosion lake), moraine-dammed lake (including end moraine-dammed lake, lateral moraine-dammed lake and moraine thaw lake), ice-blocked lake (including advancing glacier-blocked lake and other glacier-blocked lake), supraglacial lake, subglacial lake and other glacial lake. Meanwhile, some corresponding features exhibiting on remote sensing image and quantitative indices for identifying different glacial lake types were proposed in order to build a universal and operational classification system of glacial lake.
+
+Keywords: glacial lake; glacier; definition; classification; inventory; hazard
+
+# Introduction
+
+Lakes refer to water bodies which are formed in land surface basin or waterlogged depression, having certain water area and water exchange being relatively slow (Ma et al., 2011). Lakes are not only the important component of the Earth's hydrosphere, but also can faithfully record the regional climate change in different time scales and human activities around
+
+**Received:** 2017-04-23 **Accepted:** 2017-06-05
+
+Foundation: National Natural Science Foundation of China, No.41261016, No.41561016; Opening Foundation Projection of State Key Laboratory of Cryosphere Sciences, CAS, No.SKLCS-OP-2016-10; Youth Scholar Scientific Capability Promoting Project of Northwest Normal University, No.NWNU-LKQN-14-4; Geological Survey Project of China Geological Survey, No.DD2016034206
+
+Author: Yao Xiaojun (1980-), PhD and Associate Professor, specialized in the research of GIS and cryospheric change. E-mail: xj\_yao@nwnu.edu.cn
+
+lake. So lake is often considered as an important information carrier to reveal global climate change and the regional response (Cenderelli and Wohl, 2001; Ding *et al.*, 2006; Yao *et al.*, 2014; Liu *et al.*, 2014). Glacial lakes, as one of the lake types, are precious fresh water resources and the natural landscape in alpine region, and yet act as a trigger of many glacial hazards (Richardson and Reynolds, 2000; Wang *et al.*, 2010). Under the background of global warming, glacier-related hazards such as glacial lake outburst flood (GLOF) and debris flow presented an increase tendency in amount and harmful intensity (Cui *et al.*, 2014). In the Tibetan Plateau, for example, there had been at least 28 GLOF events since the 1930s (Sun *et al.*, 2014; Yao *et al.*, 2014), and had characteristics of frequency increase and temporal extension after 2000. Therefore, glacial lakes have attracted widespread attention from the academia and local government (Cui *et al.*, 2014).
+
+Globally, glacial lakes are mainly located in North America (e.g. Rocky Mountains and Coastal Mountains of Alaska, USA) (O'Connor and Costa, 1993; Clague and Evans, 2000), South America (e.g. Andes Mountains) (Carey, 2005), Europe (e.g. Iceland, the Alps and Caucasus) (Huggel et al., 2002; Björnsson, 2003; Stokes et al., 2007; Emmer et al., 2015), and Asia (e.g. Altay Mountains, Tianshan Mountains, Karakoram Range, the Himalayas) (Fujita et al., 2008; Chen et al., 2010; Janský et al., 2010; Li et al., 2011; Engel et al., 2012; Wang et al., 2012; Wang et al., 2013; Wang et al., 2016; Song et al., 2016; Song et al., 2017). The interest area of glacial lakes in China was mainly concentrated in the Himalayas (Gao et al., 2015; Liao et al., 2015), Tianshan Mountains (Wang et al., 2013), Karakoram Range (Wortmann et al., 2014), Altay Mountains (Chen et al., 2015) and southeastern Tibet (Wang et al., 2012; Song et al., 2016). And these studies focused on the change of glacial lakes (Liu et al., 2011; Wang et al., 2013; Gao et al., 2015; Song et al., 2016; Wang et al., 2016), the identification of potential dangerous glacial lakes (Cao et al., 2016; Liu et al., 2016), and the simulation of GLOFs (Le et al., 2014). In recent years, the inventory of glacial lakes in some mountainous regions had been carried out based on multi-source remote sensing images by scholars and institutions. For instance, Wang et al. (2010) and Pradeep et al. (2001) digitalized the glacial lakes in the Hindu Kush-Himalaya Range, respectively. These datasets had been an important basis for recognizing the spatial-temporal characteristics of glacial lakes change and understanding the response of glacial lakes to the climate change in this region. However, there is still great controversy about the definition and classification system of glacial lakes, which will directly lead to some difficulties for comparing the result from different researches. Some conclusions of glacial lakes change were even opposite. For example, Gao et al. (2015) and Wang et al. (2014) analyzed the glacial lakes change in the Koshi River basin from 2000 to 2010, the former believed that both number and area of glacial lakes were increasing, but the latter thought that the number was decreasing and yet the area was increasing. Even more puzzling was the fact that, the remote sensing images in 2000 used in the two studies above were the same, with the number of glacial lakes being 1228 and 1680, respectively. Therefore, it is necessary and urgent to accurately define glacial lakes, to build a complete classification system of glacial lakes and to provide corresponding features for remote sensing identification.
+
+# 2 Glacial lake
+
+In the Glossary of Cryospheric Science (Qin et al., 2016), glacial lake is defined as lake
+
+formed by glaciation. In the Wikipedia Encyclopedia (https://en.wikipedia.org), a glacial lake is a lake with origins in a melted glacier. They are formed when a glacier erodes the land, and then melts, filling the hole or space that it has created. Lü et al. (1999) thought that a glacial lake was a natural water body similar to artificial reservoir formed by ancient or modern glaciers. Additionally, there are many similar definitions of glacial lake: (1) glacial lake is one kind of lake formed by glaciation or supplied by glacial meltwater (Cao et al., 2016); (2) glacial lake is plateau lake located at the terminus or the lateral part of one glacier and glacier provides water resource when it retreats or melts (Tang et al., 2014); (3) glacial lakes were located in a basin formed by alpine glacier movement since the Last Glaciation Maximum (LGM), and their water were mainly from modern glacial meltwater or atmospheric precipitation (Chen et al., 2015). In these definitions of glacial lake above, it is emphasized that glacial lakes are formed by glaciation. The main differences are whether the time information being given or not and the material source of glacial lakes. The discrepancy of accepted concept of glacial lake caused that it was difficult to separate glacial lakes from natural lakes, which was particularly obvious in the inventory of glacial lakes and the change researches afterwards. Instead, the lakes in a specified distance of glaciers were treated as glacial lakes by scholars. For instance, Wang et al. (2013) and Zhang et al. (2015) selected lakes in a 10 km buffer of glaciers as glacial lakes in the Tianshan Mountains and Tibetan Plateau. However, it is debatable that these lakes selected are whether supplied by glacial meltwater or there are other lakes formed by glaciation beyond the specified extent. Undoubtedly, glacial meltwater is the main material source of glacial lakes as these definitions above mentioned. But it is very difficult to quantify the glacial meltwater and calculate its proportion in lake water volume in reality. If one lake receives glacial meltwater but is distant from the glacier, or the glacier supplied lake disappears, it is also one problem to judge whether these lakes are glacial lakes or not. These challenges had been causing that some research results of glacial lakes were not comparable; meanwhile, it was also difficult to share dataset of glacial lakes among different organizations.
+
+From the mechanism of glacial lake formation, glacial lakes are generated in erosion depressions caused by glaciers advance and retreat, and then receive glacial meltwater and precipitation. That is to say, glaciation is the dominant factor in the formation of glacial lakes. The LGM is the latest glaciation period when glacier coverage was the largest on the Earth. Theoretically, a glacial lake can be defined as the natural water body formed in the depression by glaciation since the LGM. However, there are still many problems and controversies in this definition of glacial lake. One thorny issue is the determination of the coverage extent of glaciers in the LGM. The other is whether the great lakes over the Tibetan Plateau such as Namco, Yamzho Yumco Lake are glacial lakes or not. Recently, the studies of glacial lake were mainly focused on the following fields: the inventory of glacial lakes based on remote sensing images, the response of glacial lakes to modern climate change, the coupling relationship between glacier variation and glacial lake evolution, the identification of dangerous glacial lakes and glacial lake outburst flood/debris flow disasters, etc. On the one hand, the focus of these studies is modern evolution processes of glacial lakes in the temporal dimension, and on the other hand they all emphasized the role of modern glaciers on the formation and evolution of glacial lakes. Therefore, in view of the modern process and practical application, glacial lakes can be defined as natural water mainly supplied by
+
+modern glacial meltwater or formed in glacier moraine's depression. In the latter definition of glacial lake, it is emphasized on the role of modern glaciers in the formation and change of glacial lakes. For the inventory or digitalization of glacial lakes in one region, it is suggested that the glacier dataset including the WGI, the First and Second Chinese Glacier Inventory can be selected as the foundation of modern glaciers. To be sure, the classification system of glacial lakes in the next section is based on the latter definition of glacial lake.
+
+# 3 Classification system of glacial lake
+
+Internationally, there has been so far no accepted standard for the classification system of glacial lakes. Some organizations and scholars proposed the different classification systems of glacial lakes according to their own research purposes. In the inventory of glacial lakes in the Hindu Kush-Himalayan region, ICIMOD (Pradeep et al., 2001) divided the glacial lakes into 5 classes: glacial erosion lake, moraine-dammed lake, ice-blocked lake, supraglacial lake and subglacial lake. Based on this classification schema, Wu et al. (2011) proposed a detailed classification system including 3 classes and 10 subclasses. In the study of glacial lakes in the Altay Mountains, Yi and Cui (1994) suggested a multiple classification schema. According to the mechanical and thermal differences in the formation of glacial lakes, they were classified as glacial erosion lake, ice-blocked lake, moraine-dammed lake, glacial thaw lake and glacial composite lake. According to the water supply of glacial lakes, they were divided into ice-water lake and non-ice-water lake; the former were mainly supplied by glacial meltwater and the latter were dominated by surface runoff from atmospheric precipitation. In the inventory of glacial lakes in the Chinese Himalayas, Wang et al. (2010) presented a classification system including moraine-dammed lake, ice-blocked lake, cirque lake, glacial erosion lake, landslide-dammed lake, supraglacial lake and glacial valley lake. Cao et al. (2016) adopted a classification system including 3 classes and 6 subclasses: glacial erosion lake (including cirque lake and other glacial erosion lake), moraine-dammed lake (including end moraine-dammed lake, lateral moraine-dammed lake and other moraine-dammed lake) and supraglacial lake. Additionally, Wang et al. (2016) proposed an integrated classification schema of glacial lakes according to the formation age of glacial lakes, the properties of dam, the shape of lake basin, the water supply, the area change, the spatial relation between glacial lake and its supply glacier, the risk level, etc. But there was a defect of disagreement between the actual glacial lakes classification and the theoretical classification system. On the basis of previous researches and the principles of systematic, normalization, operability and scalability, we proposed a complete classification schema of glacial lakes including 6 classes: glacial erosion lake, moraine-dammed lake, ice-blocked lake, supraglacial lake, subglacial lake and other glacial lake (Table 1). According to the position or geomorphologic characteristics of glacial lakes, the first 3 classes can be divided into 8 subclasses; cirque lake, glacial valley lake, other glacial erosion lake, end moraine-dammed lake, lateral moraine-dammed lake, moraine thaw lake, advancing glacier-blocked lake and other glacier-blocked lake.
+
+Due to the wide application of satellite remote sensing images in the inventory of glacial lakes and studies of their changes, the interpretation characteristics of different types of glacial lakes are given in the following. Before the interpretation of glacial lake types, it is noted that researchers should firstly determine the approximate distribution of glacial lakes based on the inventory of glaciers and digital elevation model (DEM) data; that is to say,
+
+lakes are only considered in the basin of modern glaciers. Meanwhile, although the lake is easily identified because of the obvious differences in the spectral characteristics of water body and its surrounding features in remote sensing images, it is very difficult to distinguish the types of glacial lakes scarcely by a single sensor or mono-phase remote sensing image. So other data sources and technologies should be needed, too. For instance, DEM data can be used to produce the mountain shadow so as to remove the wrong interpretation of glacial lakes; 3D visualization technology such as the Google Earth and high-resolution remote sensing images can be integrated to distinguish the type of glacial lakes.
+
+| Table 1 | The classification | system of | glacial lakes |
+|---------|--------------------|-----------|---------------|
+|---------|--------------------|-----------|---------------|
+
+| Class                   | Subclass                       | Description                                                                          |
+|-------------------------|--------------------------------|--------------------------------------------------------------------------------------|
+| Glacial erosion<br>lake | Cirque lake                    | The lake in one cirque                                                               |
+|                         | Glacial valley lake            | The lake in U-shaped valley by glaciation                                            |
+|                         | Other glacial erosion lake     | The lake formed by glacier erosion but not belonged to other classes of glacial lake |
+| Moraine-dammed<br>lake  | End moraine-dammed lake        | The lake between the end moraine ridge and glacier terminus                          |
+|                         | Lateral moraine-dammed lake    | The lake beside the lateral moraine ridge                                            |
+|                         | Moraine thaw lake              | The lake on the moraine ridge                                                        |
+| Ice-blocked lake        | Advancing glacier-blocked lake | The lake blocked by advancing glacier                                                |
+|                         | Other glacier-blocked lake     | The lake with the dam being glacier ice                                              |
+| Supraglacial lake       |                                | The lake on the surface of glacier                                                   |
+| Subglacial lake         |                                | The lake within the glacier or over the glacier bed                                  |
+| Other glacial lake      |                                | The lake blocked by landslide, avalanche, debris flow, etc.                          |
+
+## 3.1 Glacial erosion lake
+
+Glacial erosion lake refers to the water body in the depression formed by erosion and abrasion of glacier in the process of glacier movement. At present, the common glacial erosion lakes were mostly formed by erosion of Quaternary glaciers. These glacial lakes were mainly supplied by atmospheric precipitation and had few modern glaciers in their upper part. As mentioned above, if no modern glaciers exist in the upper lakes, these lakes will not belong to glacial lakes. That is to say, ancient glacial erosion lakes should not be considered. According to the location and geomorphologic characteristics of glacial erosion lakes, they can be divided into the following 3 subclasses.
+
+### **3.1.1** Cirque lake
+
+The outstanding feature of the cirque lake is that it is located in the cirque. The typical cirque is like an arm-chair surrounded by three steep rock walls and a high reverse rock sill (see Figure 1, the centroid coordinates of the lake are 29°42'18"N, 96°18'32"E). Cirque lakes are usually small and near the equilibrium-line altitude (ELA) of glaciers. In the identification of cirque lake based on remote sensing images, it is necessary to use slope data derived from DEM data or Google Earth software to check the consistency between terrain features of glacial lakes and the cirque. If its terrain features satisfy the shape of the cirque, glacial lakes can be classified into cirque lake.
+
+### **3.1.2** Glacial valley lake
+
+Large glaciers usually formed a U-shaped valley being steep and straight on both sides and
+
+flat at the bottom by laterally eroding in tributary and downward eroding on the glacier bed. Glacial meltwater and atmospheric precipitation flowed into the valley and then constituted a lake, namely glacial valley lake. This kind of glacial lakes were usually larger and far from modern glaciers. In order to expediently distinguish with moraine-dammed lakes since the Little Ice Age (LIA) and consider their potential danger on the downstream areas, some lakes in alpine region could been classified into glacial valley lake. According to the distance from the Akkol Lake to its
+
+Image /page/5/Picture/3 description: A Google Earth satellite image displays a rugged, snow-covered mountain range. In the center of the image, a dark blue body of water is outlined in yellow and identified by a label in the bottom-left corner as a "Cirque lake." The lake is situated in a large, bowl-shaped valley, or cirque, surrounded by steep, icy mountain peaks and glaciers.
+
+Figure 1 Cirque lake
+
+supply glacier (Kanas Glacier) in the Altay Mountains and the area of Kanas Glacier (Figure 2), it is suggested that the identification of glacial valley lake can be based on the following criteria: (1) the morphology of lake should be U shaped; (2) there should be large glacier with an area of above 20 km<sup>2</sup> in the upper lake; (3) the distance from lake to its supply glacier should be less than 15 km so as to distinguish with tectonic lakes.
+
+Image /page/5/Figure/6 description: A Landsat TM, 5-4-3 band combination satellite image of a mountainous, glaciated region. The image is framed by latitude and longitude coordinates, with latitude ranging from 49°0'N to 49°10'N and longitude from 87°30'E to 87°50'E. The landscape is dominated by light blue glaciers and snow, with exposed mountain ridges appearing in reddish-brown and green hues. Two key features are labeled: 'Akkol Lake', a dark blue body of water on the left, and 'Kanas Glacier', a large glacier on the right. A scale bar in the bottom right corner indicates a length of 4 km.
+
+Figure 2 Glacial valley lake
+
+### **3.1.3** Other glacial erosion lake
+
+Glacial lakes formed by glacier erosion, which are difficult to be identified by morphological features and do not belong to the other types describing in the flowing, can be classified as other glacial erosion lake.
+
+## 3.2 Moraine-dammed lake
+
+Moraine-dammed lake is a water body between moraine ridge and glacier due to the ob-
+
+struction of moraine ridge. According to the location, moraine-dammed lakes can be divided into 3 subclasses: end moraine-dammed lake, lateral moraine-dammed lake and moraine thaw lake
+
+### 3.2.1 End moraine-dammed lake
+
+When glacier retreats, glacial meltwater accumulated in the space between glacier terminus and end moraine ridge, then formed a lake, namely end moraine-dammed lake. This kind of moraine-dammed lake is the main type causing glacial lake outburst flood or debris flow in the Himalayas and Nyainqentanglha Mountains. End moraine-dammed lakes are usually easy to distinguish in the remote sensing image because they are mostly linked to glacier terminus or the distance between them is very small (Figure 3a). In addition, the end moraine ridge uplifting is clear in the Google Earth software (Figure 3b), which can help to identify the end moraine-dammed lake, too. In the downstream of larger glacier terminus, there are usually several end moraine ridges and some lakes are formed among them. These lakes are mainly supplied by glacial meltwater and are formed by glaciation, so they can be classified as end moraine-dammed lake. Based on the measurement among these lakes in the Himalayas, the distance between the upper and the lower end moraine-dammed lakes is less than 2.0 km. In other words, the lakes with a distance being far from the upper end moraine-dammed lake than 2.0 km should be not identified as end moraine-dammed lake.
+
+Image /page/6/Figure/5 description: A side-by-side comparison of two satellite images of a glacial region, labeled (a) and (b). Both images show the same area, with latitude coordinates from 27°54'N to 27°58'N and longitude from approximately 88°3'E to 88°6'E. Panel (a) is a 'Landsat OLI, 5-4-3 band combination' image. It displays a mountainous landscape with white glaciers and two blue lakes outlined and identified by a legend as 'End moraine-dammed lake'. A scale bar indicates a length of 1 km. Panel (b) is a 'Google Earth' image, providing a more detailed, photographic view. In this image, the lakes appear icy and light blue. Arrows point to the 'End moraine', a ridge of glacial debris that dams the lower lake.
+
+Figure 3 End moraine-dammed lake
+
+### **3.2.2** Lateral moraine-dammed lake
+
+Lateral moraine-dammed lake refers to the water body beside the lateral moraine ridge of glacier. These glacial lakes usually appear around the larger valley glacier and are formed by main valley glacier blocking the meltwater of tributary valley glacier. For instance, there are 5 lateral moraine-dammed lakes beside the Ngozumpa Glacier which is the largest glacier in the southern Mount Everest (Figure 4), and the third Gokyo Lake with an area of 0.59 km<sup>2</sup> is the biggest one.
+
+### **3.2.3** Moraine thaw lake
+
+In the end/lateral moraine ridge of glacier, there are usually many but small ponds (Figure 5),
+
+which are named as glacial thaw lake or moraine-dammed lake (Yi and Cui. 1994; Wang et al., 2016). The principal agent controlling these lakes is freeze-thaw process. For example, dead ice within the moraine ridge melts due to the temperature rising, and then the surface sinks and forms the depression retaining water. Although these ponds do not receive glacial meltwater but are formed by the dead ice meltwater of glacier, they can be classified as glacial lakes, namely moraine thaw lake. It should be pointed out that thaw lakes widely distributed in rock glacier and frozen ground region do not belong to moraine thaw lake, and they should be
+
+Image /page/7/Picture/3 description: A Landsat TM satellite image, using a 4-3-2 band combination, shows the Ngozumpa Glacier flowing through a rugged, snow-covered mountain range. The glacier is a large, light-colored river of ice running vertically through the center of the image. Several dark blue lakes are visible in the surrounding brown and white terrain. Some of these lakes, located alongside the glacier, are outlined in purple to indicate they are "Lateral moraine-dammed lakes," as explained by a legend in the bottom left corner.
+
+Figure 4 Lateral moraine-dammed lake
+
+eliminated in the inventory of glacial lakes.
+
+## 3.3 Ice-blocked lake
+
+Whether the glacier advanced and blocked the valley or the branch glacier rapidly retreated and was separated from the main glacier, they all resulted in the water accumulation, and then formed a lake. The common feature of these lakes is that their dams are composed of glacier ice, so they can be named as ice-blocked lake. According to their forming mechanism, ice-blocked lakes can be divided into advancing glacier-blocked lake and other glacier-blocked lake. The representation of the former is Kyagar Thso Lake located in the Yarkant River basin in China (Zhang et al., 1989; Wang et al., 2009). An example of the latter is Merzbacher Lake near the South Inylchek Glacier in Kyrgyzstan (Liu et al., 1998; Shen et al., 2009). Other glacier-blocked lakes also exist in the interior of the Tibetan
+
+Image /page/7/Figure/8 description: A figure with two panels, (a) and (b), showing satellite imagery of glacial lakes. Panel (a) is a Landsat OLI image displaying a wide view of a mountainous, glaciated area with latitude and longitude markings from 28°10'N to 28°14'N and 86°36'E to 86°40'E. A scale bar indicates a length of 2 km. A red box highlights a specific region containing several lakes. Panel (b) is a zoomed-in Google Earth image of the highlighted area from panel (a). It provides a detailed view of different types of lakes, which are labeled with arrows. These include a large 'end moraine-dammed lake', a smaller 'lateral moraine-dammed lake', and a cluster of small lakes within a yellow dashed outline identified as 'moraine thaw lake'.
+
+Figure 5 Moraine thaw lake (water body in blue-green color within the yellow-color polygon in (b)
+
+Image /page/8/Figure/2 description: A satellite image from Landsat OLI, using a 5-4-3 band combination, displays a mountainous region with glaciers. The map is marked with longitude lines from 80°48'E to 80°54'E and latitude lines from 34°18'N to 34°20'N. A specific body of water is outlined in yellow and identified in the legend as an 'Ice-blocked lake'. A scale bar in the lower-left corner indicates a distance of 2 kilometers.
+
+Figure 6 Ice-blocked lake over the Tibetan Plateau
+
+Plateau (Figure 6), which has not been previously reported. For advancing glacier-blocked lakes, they usually survived for a few months to one year affected by temperature increase in low altitude and lake water erosion (Wang *et al.*, 2009). For other glacier-blocked lakes, their evolution presented different characteristics. For example, Merzbacher Lake outburst floods mostly occurred in July – September and caused 50 GLOF events from 1932 to 1997 (Liu *et al.*, 1998), showing a life cycle of glacier-blocked lake to some extent. But for glacier-blocked lakes in the interior of the Tibetan Plateau, they are usually stable.
+
+## 3.4 Supraglacial lake
+
+Supraglacial lake refers to the water body on the surface of glacier due to different ablation. These lakes usually appear on the surface of ablation zone of debris-covered glaciers. As seen in Figure 7, there are many supraglacial lakes in Rongbuk Glacier on the northern Mount Everest, the largest of which had an area of 0.47 km² in 2016. In Tomur region of Tianshan Mountains, there are also many supraglacial lakes in some debris-covered glaciers, such as Tomur Glacier, Tugaibieliqi Glacier, Qong Terang Glacier, Koxkar Baxi Glacier, etc. When supraglacial lakes are connected with the drainage system inside the glacier, the lake water can be quickly exhausted. So supraglacial lakes change rapidly in annual and inter-annual scales. In view of the spatial resolution of Landsat TM/ETM+/OLI remote sensing images widely used, the quick variation of supraglacial lake, and being the origin of moraine-dammed lake (Richardson and Reynolds, 2000), it is suggested that the supraglacial lake with an area of above 0.02 km² can be collected in the inventory of glacial lakes. The main difference between supraglacial lake and moraine thaw lake is their locations: the former is located on the surface of glacier; the latter is located on the moraine ridge of glacier.
+
+## 3.5 Subglacial lake
+
+The water body within the glacier can be named as subglacial lake. It had been found that there were more than 140 subglacial lakes under the Antartic ice sheet, the largest one was the Vostok Lake (Wingham *et al.*, 2006). For alpine glaciers, lakes like subglacial lake had not been reported until now. However, small supraglacial lakes possibly exist in the distribution region of maritime glaciers such as southeastern Tibetan Plateau, because the drainage
+
+Image /page/9/Figure/2 description: A two-panel figure displaying satellite imagery of the Rongbuk Glacier. The left panel is a wide view from a Landsat OLI, 5-4-3 band combination, showing a vast area of snow-covered mountains and glaciers. The coordinates range from 86°45'E to 87°0'E longitude and 28°0'N to 28°10'N latitude. A scale bar at the bottom indicates a length of 4 km. A red box highlights a specific area on the glacier. The right panel is a magnified view of the highlighted area, showing a close-up of a supraglacial lake. The lake is colored light blue and outlined in magenta. A legend identifies the outlined shape as a 'Supraglacial lake'. The latitude for this panel ranges from 28°5'N to 28°10'N, and a scale bar at the bottom is labeled from 0 to 500 km.
+
+Figure 7 Supraglacial lake
+
+channels under the glacier are well developed. It is impossible to identify subglacial lake only relying on satellite remote sensing image, other instrument like ground penetrating radar (GPR) is needed. For building a complete classification system of glacial lake, subglacial lake is listed, but it can be removed in the inventory of glacial lakes based on satellite remote sensing images.
+
+## 3.6 Other glacial lake
+
+Except for 5 classes of glacial lake mentioned above, the water body formed by landslide, rock collapse, avalanche and debris flow blocking the valley in the glaciation area is named as other glacial lake. This kind of lake belongs to dammed lake and has a great potential harm to the downstream settlements, roads, hydropower stations and other infrastructures. Therefore, it should be taken into account in the inventory of glacial lakes. The first basis for identifying the other glacial lake is its dam which can be examined from the high-resolution remote sensing images. The second is that the appearance of the other glacial lake is usually related to the earthquake, debris flow and other geologic activities, and there is an obvious change of lake size around the time. For instance, the debris flow caused by Ranzeria Co Lake outburst formed two glacial lakes in Jiali County (Sun *et al.*, 2014). The quantitative index of identifying the other glacial lake is suggested that the distance from the glacier terminus is below 10 km (Wang *et al.*, 2013; Zhang *et al.*, 2015).
+
+# 4 Conclusions
+
+As an important object of the cryospheric science, glacial lakes are not only closely related to the climate change and glacier movement, but also play a role on mountain disaster chain. Therefore, glacial lakes had been paid more and more attentions by many scientists and governments. In this paper, the definition of glacial lake was comprehensively discussed. It was noted that the definition was based on two perspectives of glaciation and glacial melt-
+
+water. But the fuzzy of spatial and temporal information made the poor operability in the inventory of glacial lakes. In short, it is difficult to determine whether one lake is a glacial lake or not. Theoretically, a glacial lake can be defined as the natural water body formed in the depression by glaciation since the LGM. However, the controversy of glacial coverage area in the LGM and researches' knowledge deficiency of Quaternary glaciology also caused difficulties of identifying glacial lakes based on theoretical concept of glacial lake. On the consideration of the interest of glacial lakes' studies, an alternative definition of glacial lake was proposed, i.e. glacial lake could be defined as the natural water mainly supplied by modern glacial meltwater or formed in glacier moraine's depression. Meanwhile, a complete classification system of glacial lake was proposed based on its formation mechanism, topographic feature and geographical position. Glacial lakes were classified as 6 classes and 8 subclasses: glacial erosion lake (including cirque lake, glacial valley lake and other glacial erosion lake), moraine-dammed lake (including end moraine-dammed lake, lateral moraine-dammed lake and moraine thaw lake), ice-blocked lake (including advancing glacier-blocked lake and other glacier-blocked lake), supraglacial lake, subglacial lake and other glacial lake.
+
+Although this paper attempted to provide the features or quantitative indices for identifying different types of glacial lakes using satellite remote sensing images, and emphasized that the existence of modern glaciers was the primary basis for distinguishing a glacial lake, there were still some problems in the inventory of glacial lakes. When glaciers previously existed in the First Chinese Glacier Inventory and there were glacial lakes in the downstream, if these glaciers disappeared in the Second Chinese Glacier Inventory, did these lakes belong to glacial lakes? So it is very important to ensure the consistency of subject. In addition, the quantitative indices proposed for identifying some types of glacial lakes were derived from the typical glacial lakes, and their representativeness needs to be further studied.
+

chatbot/rag/data/('Design and Implementation of an Advanced Glacier2Lake Outburst Flood Early Warning System(GLOF 3EWS) Using LoRa IoT devices', '.pdf')_extraction.md

@@ -0,0 +1,164 @@
+
+Abstract
+
+This paper describes the creation and deployment of a Disaster Early Warning System (DEWS) designed to detect Glacial Lake Outburst Floods (GLOFs). This system is a Comprehensive Integrated dIsaster pRevention Configuration (CIIRC)that has been built to withstand sub-zero temperatures and harsh weather conditions while displaying durability and resilience. GLOFs pose a significant risk to communities living near glacial lakes, necessitating the use of advanced monitoring equipment. The proposed system may transfer crucial data in real time from remote glacier lake monitoring stations to a central hub via radio frequency/ long-range communication, or satellite communication technology. Key properties such as lake water level, temperature, hydro pressure, and seismic activity are continuously recorded and transmitted from remote areas to a central monitoring station (CMS). The CMS examines incoming data using risk assessment algorithms (thresholds), which provide warnings and notifications as soon as potential GLOF issues are detected. The integration of this DEWS with local emergency response plans allows for more prompt and informed decisionmaking, which helps to mitigate GLOF-related tragedies. This comprehensive system intends to improve community resilience by delivering early warnings and facilitating proactive reaction measures.
+
+Keywords: GLOF, EWS, Disaster, Pressure sensor, LoRa.
+
+## 1 Introduction
+
+A GLOF is defined as a rapid and catastrophic overflow due to the breakage of the moraine dam which dams a glacial lake. Normally, the moraine dams of a glacial lake comprise unconsolidated rocks and sediments, which are sensitive to climatic changes and inputs generated by seismic activity as well as anthropogenic factors [1]. However, GLOF disasters may lead to destruction in the lower areas. This results in severe destruction and loss [2], [3] Climate change is resulting in increased incidence and intensity of GLOFs. Hence, the current need to increase early monitoring and risk mitigation of GLOF. Effective early prediction of GLOF It will update you on dangers of GLOF in advance. This can be done with various methods like
+
+- Remote Sensing: Besides the satellite image, there is other remote sensing-based data that could be used in monitoring the changes in the size of glacial lakes, water level, among other indicators of GLOF risk.
+- Ground-based monitoring: Ground-based stations are also used in gathering data on several parameters, including water levels, turbidity, and many more.
+  - •Numerical modeling: Numerical models can simulate how glacial lakes and moraine dams behave under different conditions.
+
+- GLOF mitigation aims at reducing the chances of a GLOF happening or its impacts if one happens. Some of the possible strategies adopted include:
+- Water level in the glacial lake: It can be drained with a controlled release or even by developing tunnels or spillways.
+- The moraine dams may be reinforced by either adding more material within the dam or constructing support structures.
+- •Relocation of communities and infrastructure outside hazard zones of GLOFs is considered the most effective measure of mitigation taken in the long term.
+  - Early Warning Systems (EWS): EWS may give the communities an opportunity to evacuate before the occurrence of a GLOF event. Most EWS depend on the combination of several early detection means with a communication system that issues warnings to affected communities.
+
+
+The work on GLOF early detection and mitigation has been ongoing, and the new methods and technologies are in the pipeline very continuously. The most important thrust in such work is on making more appropriate and reliable techniques for early detection. The other key areas of research include developing cost-effective sustainable mitigation. In the current focus, this research relates to the design of DEWS for GLOF with real-time measurement and communication of data, analysis, and action.
+
+## 2 Literature review
+
+GLOFs are one of the foremost natural hazards occurring in mountainous regions, especially in the Himalayas and Andes. This is following the breach of the glacial lake dam, releasing a massive amount of water and debris that travels downstream. GLOFs also trigger incredible damage to structures, properties, and human life. Of late, there has been more frequency and intensity along with climate and glacier retreat. This requires GLOF detection with adequate EWS in place to mitigate the risks associated with such floods.
+
+There are several studies that have been conducted on GLOF detection and EWS. For example, Wang W (2022) et.al [4] designed an early warning system for the Cirenmaco glacial lake located in the central Himalayas. The system is more interested in monitoring and providing various features with timely warnings rather than specific flood features. Another study about the computational challenges and solutions involved in the development of GLOF early warning system by Kumar B (2022) et.al [5]. It believes that an early warning system is required to monitor vulnerable glacier
+
+lakes and provide robust tools to the disaster managers to plan mitigation, thus saving human life and property. International Centre for Integrated Mountain Development (ICIMOD) has also made some studies on GLOF detection and EWS. In the report of Mitigation (2008) [6], ICIMOD speaks about EWS, monitoring, and GLOF mitigation measures. The GLOF warning systems were based on 'extended line of sight' (ELOS) VHF radio technology. The warning signal would be transmitted via ELOS ground wave
+
+An early warning system on glacial lake outbursts to save lives and livelihoods of Nepal Himalaya communities was developed through a joint initiative of the Government of Nepal, UNDP in Nepal, and the Global Environment Facility [7]. The system encompasses hydro-met stations; GLOF sensors automated early warning sirens and a linked dynamic mass SMS alert system polygon. The system benefits more than 71,000 vulnerable people; both the local and the tourists visiting the Everest Region of Nepal. The system has the potential for replication and upscaling, with 21 critical lakes in Nepal and 25 in Tibet, China. Integration of climate actions in the routine planning and implementation holds the key for resilient and sustainable development. Among these, Kumar B (2020) et.al [5] proposed an ultrasonic sensor-based design architecture of a GLOF early warning system. The study majorly depicts the need to have an early warning system, especially one that detects GLOFs in real-time with timely warnings to the communities at risk.
+
+The monitoring of GLOF and EWS is crucial for risk management associated with GLOF. This includes the deployment of remote sensing, ground-based sensors, and machine learning algorithms that incorporate data collection, processing, and analysis methods in the detection and prediction of GLOFs. These systems have been evaluated based on real-world data compared with existing GLOF detection and EWS. Besides this, new insight coupled with enhanced process understanding facilitated from reconstructing past GLOFs has contributed significantly to effective detection of GLOFs and EWS [8].
+
+## 3 Insights for CIIRC design
+
+### • Measuring Water Pressure:
+
+Measured at the installation point, the system measures the pressure of water within streams or rivers and in glacial lakes. Sudden changes in pressure may indicate a rising level of water-a warning sign for possible GLOF. A rise in hydro pressure could initiate subglacial conduits or tunnels, which would further weaken the glacier dam. By monitoring hydro pressure, the likelihood and triggering of a GLOF can be reflected and a prognosis drawn on the potential for dam failure.
+
+119
+
+89
+
+121
+
+### Measuring Water Level:
+
+The system also monitors the fluctuations of the water level in the lake, which will trigger alarms with quick increases. Rapid rises in water level might be indicative of
+
+sudden water inflow, perhaps due to a breach in a glacier dam or icefall collapse. Ongoing monitoring of water level conditions can hence provide advance warnings to communities downstream.
+
+127 128
+
+### Measuring Temperature:
+
+Continuous monitoring of temperature offers valuable information relating melt rates with changes in lake, river, and stream dynamics. Changes in lake water temperatures may indicate changes in the amount of glacial meltwater input, which can be due to increased ice melting or changes in the environment of glacial regions. Monitoring this change identifies the actual stability of the glacier lake. Statistical anomalies in lake water temperature, among other monitoring parameters, can generate alarms and warnings. A quick and steady rise in temperature may be an initial sign of the changes in the glacial environment and may give an early prelude to a GLOF.
+
+137 138
+
+### • Measuring Seismic activities:
+
+It can capture sudden movements or changes in the glacier structure that may be a precursor to a GLOF event, providing more detail about the evolving dynamics. Recording acceleration or changes in velocity can add valuable data in many ways. Abrupt changes in seismic activity near the glacier dam may indicate structural weaknesses or failure points. Continuous monitoring will be able to deliver an estimation of probable loss due to dam failure and subsequent water overflow. Seismic signals that occur with a GLOF can be studied to determine more about the dynamics of the events, including the time of occurrence and the magnitude. This information will be used in understanding the characteristics of the flood and increasing the ability to analyze future risks. Pressure sensors combined with seismic sensors will enable subglacial processes to be monitored. For instance, even the process of creating subglacial conduits or tunnels may affect pressure dynamics and cause seismic signals. Integration of pressure and seismic data could thus offer a more defined view of subglacial activity.
+
+4 The DEWS Design flow
+
+- The proposed PSA based DEWS has been presented in a block diagram in figure 1. The proposed DEWS, therefore, caters to a specific requirement such as glacial lake monitoring and assessing the risk of GLOFs. The system, thus, is provided with Pressure Sensing Apparatus (PSA) along with a communication system, well suited to provide real-time data, such as hydro pressure and water level measurements and temperature measurements. The system can also incorporate a seismic sensor for
+- instance accelerometers for monitoring the seismic activity.
+
+162
+
+It has advanced communication interfaces such as RF/long range (LoRa)/ SATCOM modules (transmitter module at GLOF site and receiver module at central monitoring station) for early warning capabilities. Integrated data loggers of the pressure sensor (DCX-22) can be programmed to record and transmit the data to a central monitoring
+
+188 189
+
+192
+
+195 196
+
+station at regular intervals (minimum interval of 1min). Real-time data transmission can be derived from satellite communication, with suitable uplink and downlink facilities or through any other form of wireless RF communication means thus ensuring real-time monitoring at the location. The moment remote-sensing-based data from the DEWS is available, a DAS comes into play as a fundamental part of the early warning process.
+
+The DAS is under constant surveillance, checking the real-time data against set thresholds and patterns. Such thresholds are mostly set by topographical data and are carefully fine-tuned by the insights gained from historical and researched contexts. The data acquisition system instantly sends an alarm signal whenever deviations cross a set threshold. The risk is thus passed to the concerned authorities, usually present at site in the locality; in this case, the local community in-charge / National Disaster Management Authority, NDMA through the signal and hence causes an urgency call for action. Advanced communication technology hastens the alarm signal and disaster management becomes urgent with a current emphasis. With the NDMA being the custodian authority, it thus coordinates quick and prompt disaster response to manage or lessen the impending disasters. Such responses by the NDMA will include various emergency measures that involve the inclusion of immediate alerts to the communities at risk. Such alerts will include such actions as emergency evacuation messages to people living in the vicinity of the disaster, thus reducing possible risks and saving lives and livelihoods.
+
+Image /page/5/Figure/4 description: A block diagram illustrating a system for monitoring Glacial Lake Outburst Floods (GLOF). The diagram is divided into two main parts. The left part, labeled "Glacial Lake at a Remote Place," shows a process flow starting from a "GLOF Site" to a "Pressure Sensor," then to "PSA," then to "Glacial lake parameters," and finally to an "RF / SATCOM / LoRa Transmitter" which sends a signal via an antenna. The right part, labeled "Central Monitoring Station," shows a signal being received by an antenna connected to an "RF / SATCOM / LoRa Receiver." The data then goes to a "Data Acquisition System (DAS)." A decision block checks for the "Threshold for GLOF." If the condition is not met ("No"), the process loops back to the DAS. If the condition is met ("YES"), an "Alarm to Local community / NDMA" is issued.
+
+F
+
+## 5 Proposed Pressure Sensing Apparatus (PSA)
+
+Figure 2 Shows the graphic layout of the PSA. A high strength hose 10m long and 26mm in diameter is used, where accurate measurements of the pressure are ensured by the pressure sensor contained in this hose. This sensor is fixed on one end of an iron rod larger than the internal diameter of the hose at about 2cm in length. The rod is inserted in the hose using silicone sealant, but the lower end is held inside with hose clamps. It served the purpose by giving a waterproof seal and provides weight to stabilize the sensor's position. At the top of the hose, a T-junction is fitted so that a logger cable can be connected to connect up a pressure sensor to the data logger. The second leg of this T-junction feeds out to a 2-liter rubber expansion bladder. This bladder is normally half full; this way, volume changes in the antifreeze solution, such as those due to solar heating, build only enough pressure in the assembly that it can easily be sealed. The hose is filled with an appropriate quantity of antifreeze and water; the illustration here is 50:50 mix of ethylene glycol with water. It serves as the antifreeze solution. In establishing the installation of the PSA, a pole with a length of 5 meters is installed in the central region/ shallow side/ shallow bank of the stream/river. The length of the pole depends on the demand. The guides attached to the Teflon pulley hold the PSA on the pole. The pulley system makes sure that water level fluctuations do not pull the pressure sensor off of the river or streambed [9]. In case the PSA is deployed on an ice floor of a frozen pond/lake, this pulley arrangement allows the pressure sensor to dip as the ice melts, thereby ensuring that it remains in contact with the underlying surface of the pond or lake where the meltwater collects. The installation process involves using drilling equipment with extensions to a depth of 3 meters in the riverbed or streambed or frozen pond/lake.
+
+Image /page/6/Picture/4 description: A diagram illustrates a monitoring system set up in a glacial lake. The main part of the diagram shows a cross-section of a body of water labeled "Glacial Lake," with a device submerged in it. This device is connected by a hose to equipment on the shore. Two inset diagrams provide detailed views of the components. A red arrow points from the submerged device to a close-up diagram showing its internal structure. This structure is a long tube containing a "Metal sinker" at the bottom, a "Pressure sensor" above it, a "Cable," and "Antifreeze liquid," with a "Hose" connected at the top. Another red arrow points from the equipment on the shore to a second close-up diagram. This diagram shows a green, spherical "Bladder" connected via a "Hose" and "Fittings" to a "Cable to logger."
+
+Fig. 2. Design flow
+
+### Pressure Sensing device
+
+The DEWS can usually contain several sensors and equipment to measure water level, temperature, pressure, or a single pressure sensor like DCX-22 (which can work up to -10-degree celcius without the PSA setup / with PSA setup can work up to -80 celcius). The KELLER Druckmesstechnik AG, Switzerland sensor DCX-22 of figure 3 has a logger for data and can measure the water level, hydro pressure, and temperature. The pressure sensor used in the PSA is DCX-22 and is flexible to measure the water level or pressure of water and temperature. The sensor is also able to give accurate readings for submerged applications. With the data logger, which is equipped with a long-lasting battery, measurements can be recorded for long periods of time, from minutes to hourly intervals. Also, the pressure sensors are able to be calibrated by the users using the calibration program provided by the vendor. Sensors must be calibrated in an environment exactly the same as that in which they will be used. Sensors are often calibrated upright on a level surface. Sensor recalibration may be needed in several cases, such as when the sensor is serving under new or changed maintenance conditions, its measuring setup is altered, or its station has been in service for many years.
+
+The pressure sensor, in combination with the data logger, can be programmed to acquire data on hydro pressure and temperatures for set time intervals. The result ensures that the collection of measurements during the monitoring period will be accurate. Nevertheless, it could be such that there will be cases wherein data gaps may have occurred through the exposure of the sensor to either freezing or low water levels, mostly. The densities of the surrounding environment should be considered when determining the height of a water level using the measurements of hydrostatic pressure acquired with the help of the pressure sensing device. Equation 1 supports the conversion of pressure readings in N/m2 to the corresponding height of the water column in meters. Another important consideration in the calculations is that the length of an iron rod needs to be added to the depth/water level for appropriate estimates.
+
+$$d = \frac{P}{(0*9)} \tag{1}$$
+
+where P = hydrostatic pressure (Nm-2),  $\rho = water$  density (kgm-3), g = gravitational acceleration (ms-2), d = height of water column (m).
+
+Image /page/7/Picture/7 description: A studio photograph of an industrial sensor on a white background. The device consists of a long, cylindrical stainless steel probe connected to a long, coiled black cable. The other end of the cable terminates in a multi-part stainless steel connector. The tip of the probe is a black cap with small holes, and the connector also has a visible sensor element at its end.
+
+Fig. 3. DCX-22 KELLER Pressure Transducer
+
+## 7 Installed PSA at field
+
+The configuration in figures 4 and 5 here constitutes the working scheme of the PSA: the RF transmitter is located at a remote position, while the RF receiver mounted with a Data Acquisition System (DAS) is located at the central monitoring station. This design, primarily aimed at the purpose of observing depth and water levels in lakes in Antarctica, promises much towards the detection of GLOFs. In fact, by serving as a DEWS, it can be fundamentally important in the premature observation and alerting authorities about possible GLOF catastrophes. As a result of unfavorable climatic conditions and inaccessibility of regular transportation for frequent visits, the system has been engineered to possess wireless communication capabilities. This allows real-time data transmission from the installation site to a central monitoring station about 2 kilometers off the lake site. Long Range (LoRa) transmitter and receiver modules are chosen in the wireless communications technology that will be made use of for this system since this will allow the transcription of critical lake parameters, like hydro pressure, temperature, and water level.
+
+The system comprises reliable backup mechanisms in terms of power to continuously and uninterruptedly transmit and receive data. It employs the use of solar panels and insulation for the proper protection of the parts constituting the transmitter and receiver from probable disruptions through power and ensures a smooth, steady flow of data. The pressure sensor is programmed to become operational to serially measure the lake parameters at regular intervals. The acquired data is then sent to the central monitoring station, as exemplified in figure 5. This systematic approach, of course, allows comprehensive data to be collected but also ensures timely updates at the monitoring station-thus making the whole system of monitoring effective in the challenging polar environment.
+
+Image /page/8/Picture/5 description: A scientific instrument is set up on the edge of a body of water in a snowy, icy landscape. The instrument consists of a metal pole with a grey box labeled "CRANE" in red letters at the top, and a reel with tubing attached to the side. The pole is mounted on a wooden plank that extends from a rocky bank over the partially frozen water, with the bottom of the pole submerged. The background is a vast expanse of snow and ice under a blue sky.
+
+Fig. 4. Working model of PSA setup installed in a lake.
+
+331
+
+333
+
+Image /page/9/Figure/2 description: A composite image, labeled Fig. 5, containing five panels labeled (a) through (e), illustrating a data acquisition system and its output.
+
+Panel (a) shows a small solar panel and an electronic device in a blue case placed on rocks outdoors.
+
+Panel (b) is a close-up of the electronics inside a box, including a yellow battery pack labeled "LEMON 14.8V", a blue circuit board, and various wires.
+
+Panel (c) depicts a person sitting at a desk, working on a laptop with other computer equipment nearby.
+
+Panel (d) is a line graph plotting Temperature (°C) versus Time. The y-axis ranges from -6 to 12. The x-axis shows time over a 24-hour period. There are two lines: a blue line for "Water Temperature" and an orange line for "Surface Temperature". The water temperature fluctuates more widely, peaking around 11°C, while the surface temperature remains lower, peaking around 2°C.
+
+Panel (e) is a screenshot of a spreadsheet with data columns for Time, Timestamp, Date, CH1 Pressure(mbar), and CH2 Temperature(°C). The data shown is for 16-Jan-23, with timestamps starting at 12:44:29. Sample data includes a pressure of 984.68 mbar and a temperature of 5.6°C.
+
+The caption below reads: "Fig. 5. (a) & (b) Transmitter module with solar panel and power backup (c) Sample..." with the rest of the text cut off.
+
+**Fig. 5.** (a) & (b) Transmitter module with solar panel and power backup (c) Sample data measured at 15min interval by PSA (d) Central monitoring station with receiver and (e) Sample of real time data received from PSA at 1 minute interval using LoRa receiver
+
+## Features of proposed DEWS
+
+- The proposed Pressure Sensing Assembly can come out as an important monitoring device for Glacial Lake Outburst Floods and therefore can help in the prevention or reduction of its impact. With an interfacing of an appropriate wireless communication
+- system with the PSA, it is likely to provide real-time data to the NDMA via early warnings through PSAs in the following ways:
+- 1. An Early Warning System: Place the PSA strategically around/at the glacial lakes. It will constantly observe the change in water pressure and lake levels. If the water pressure increases rapidly, along with an increase in the water level, that might depict a GLOF event. The sooner an event can be observed, the better it can alert downstream
+- 2. Real-time Data: The PSA can also forward in real time data concerning pressure levels and the temperature of the water. These data then forwarded to a central monitoring station for close scrutiny by experts to determine emerging trends that could characterise an impending GLOF.
+- 3. Trend analysis: So, with this data for very long periods, the scientists and researchers get to understand how this glacial lake acts historically, thus they will be in a better position to identify some trends and patterns which may have been happening before these GLOFs. It means better predictions are made, and pro-active mitigations can be developed.
+- 400 4. Risk Assessment: By these continuous updates of the lake conditions, PSA facilitates risk assessment towards GLOF. Information collected in this context may help policymakers decide about evacuations, infrastructure development, and so on.
+- 5. Remote Access: The PSAs that are equipped with remote data transmission, can prove very useful in remote and inaccessible glacial regions. This will enable experts to monitor conditions and collect data even from challenging environments.
+
+Regular maintenance and calibration of the PSA is essential to ensure that accurate data would be collected. Proper maintenance also prevents one from false alarms hence ensuring that one's PSA would still be reliable for GLOF monitoring. To minimize data loss from a low battery, a solar panel can be interfaced to charge the system. In a nutshell, PSA has become an essential tool in early warnings, assessments, and monitoring of glacial lakes, which contribute to preventing and mitigating the impacts of GLOFs, finally serving to safeguard communities downstream from ice melts that trigger outburst floods.
+
+
+The development and deployment of a dedicated Disasters Early Warning System for Glacial Lake Outburst Floods are hence among the very important advances into mitigating impact involvement with this kind of catastrophic happening. Integration of satellite communication technology makes it possible to monitor remote glacial lake conditions in real time, thus enabling the early detection of potential GLOF risks. Combining data from sensors measuring the levels of water, temperature, and hydro-pressure activity, the system uses complex algorithms to riskassess and generate accurate alerts. The addition of a seismic sensor in this system will enable monitoring of seismic activity. Successful integration of this DEWS with local emergency response plans helps communities get warnings in time, or even permit an efficient evacuation and response strategy. Because climate change is increasingly influencing glacial dynamics, this system would be proactive in protecting the most vulnerable populations against the rising threat of GLOFs, thus supporting sustainable disaster risk reduction efforts. In general, DEWS the CIIRC structure is very robust and likely to be with us for a long time. This assurance originates from its launching and rigid testing under the extreme conditions of Antarctica's polar region for nearly three months.
+
+## 438 Acknowledgment
+
+- The authors gratefully acknowledge the support rendered by Centre for Incubation,
+- 440 Innovation, Research and Consultancy (CIIRC), Jyothy Institute of Technology,
+- 441 Bengaluru and Sri Sringeri Sharada Peetham.
+
+## 442 References
+

chatbot/rag/data/('Disaster Prediction Knowledge Graph Based on Multi-Source Spatio-Temporal Information', '.pdf')_extraction.md

@@ -0,0 +1,456 @@
+
+
+Abstract: Natural disasters have frequently occurred and caused great harm. Although the remote sensing technology can effectively provide disaster data, it still needs to consider the relevant information from multiple aspects for disaster analysis. It is hard to build an analysis model that can integrate the remote sensing and the large-scale relevant information, particularly at the sematic level. This paper proposes a disaster prediction knowledge graph for disaster prediction by integrating remote sensing information, relevant geographic information, with the expert knowledge in the field of disaster analysis. This paper constructs the conceptual layer and instance layer of the knowledge graph by building a common semantic ontology of disasters and a unified spatio-temporal framework benchmark. Moreover, this paper represents the disaster prediction model in the forms of knowledge of disaster prediction. This paper demonstrates experiments and cases studies regarding the forest fire and geological landslide risk. These investigations show that the proposed method is beneficial to multi-source spatio-temporal information integration and disaster prediction.
+
+**Keywords:** disaster prediction knowledge graph; spatio-temporal; disaster dynamic prediction; multi-source data fusion; forest fire risk prediction; geological landslide risk prediction
+
+Image /page/0/Picture/14 description: A graphic on a white background features a yellow circular icon with a white checkmark inside. To the right of the icon, the text 'check for updates' is displayed in a dark sans-serif font. The words 'check for' are on the top line, and the word 'updates' is on the bottom line in a bolded font.
+
+Citation: Ge, X.; Yang, Y.; Chen, J.; Li, W.; Huang, Z.; Zhang, W.; Peng, L. Disaster Prediction Knowledge Graph Based on Multi-Source Spatio-Temporal Information. *Remote Sens.* 2022, 14, 1214. https://doi.org/10.3390/rs14051214
+
+Academic Editors: Hideomi Gokon, Yudai Honma and Shunichi Koshimura
+
+Received: 31 January 2022 Accepted: 26 February 2022 Published: 1 March 2022
+
+**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.
+
+Image /page/0/Picture/19 description: A Creative Commons Attribution (CC BY) license icon. The icon is a rectangle divided into two horizontal sections. The top section is grey and contains two white circles with black outlines. The left circle has the letters 'CC' in black, and the right circle has a black stick figure icon. The bottom section is black with the letters 'BY' in white.
+
+Copyright: © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
+
+# 1. Introduction
+
+Disaster is a general term for events that can have a destructive impact on humans and the environments on which they depend. Although natural disasters are inevitable, it must take efforts to greatly reduce losses. However, due to the wide distribution of natural disasters, time, location, and scale of natural disasters have great uncertainty. It greatly increases the difficulty to resist natural disasters. Therefore, the critical issue is the prediction of disasters with multi-source data of disaster scenarios for formulating emergency plans.
+
+Remote sensing technology can obtain a large amount of strongly dynamic information in large observation ranges with high speed in real-time [1]. It can identify the fine features of ground objects, and extract information such as buildings, roads, and cultivated land by combining remote sensing technology and artificial intelligence technology. It plays an important role in the fields of disaster monitoring, early warning, and emergency decision-making. The critical problem is the fusion of different data types, different data structures, and different types of expert knowledge when the remote sensing information extraction for the multi-scale, multi-temporal, multi-type, multi-precision remote sensing
+
+Remote Sens. 2022, 14, 1214 2 of 21
+
+extraction information, other geographic information, and other spatio-temporal data in disaster scenes. There are studies for addressing the problem of spatio-temporal disaster information fusion, including knowledge representation of spatio-temporal disaster data, attribute analysis based on data spatio-temporal relations, and spatio-temporal dynamic geographic information of and expert knowledge.
+
+In the real-world disaster scene, there exists large-scale multi-source heterogeneous data. There are complex associations between the data, as well as between the data and the disaster knowledge. It can effectively represent and model the data as well as the in-between relations using a knowledge graph. This is an artificial intelligence technology proposed by Google in 2012. Knowledge graph is a data modeling method that represents knowledge as concept, entity, and semantic relations with each other in the form of a graph [2]. Knowledge graphs can connect multi-source spatio-temporal disaster data and expert knowledge by a graph in order to support the analysis of the multi-source heterogeneous data under dynamic and static situations in natural disaster scenarios.
+
+The spatio-temporal knowledge graph is a knowledge graph that can realize the semantic association of spatio-temporal data, which is the basis of disaster prediction. Sun et al. proposed a spatio-temporal knowledge graph ontology model which can manage and query spatio-temporal data in specific fields by adding spatio-temporal attribute information to entities [3]. Geographic information is a type of spatio-temporal data with spatial and temporal characteristics. Therefore, the construction technology of a geographic knowledge graph is very meaningful for this study. Zhang et al. proposed a method for constructing a geographic knowledge graph that takes the spatio-temporal characteristics into account. They studied the formal representation of geographic knowledge and laid a foundation for the acquisition, fusion, reasoning, and application of the spatio-temporal data [4]. Based on ontology, Liu et al. proposed a fast retrieval method that takes the semantic knowledge of geospatial data into account for reducing users' dependence on data storage methods and database grammar rules [5]. Jiang et al. proposed the construction process of a geographic knowledge graph, which helps to realize the knowledge of geographic information by reviewing the geographic knowledge graph [6].
+
+At present, there exists research work regarding the construction of ontology and knowledge graph for specific disasters. Qiu et al. constructed model ontology and data ontology for flood management based on the four phases of disaster management [7]. Xu et al. constructed a conceptual model for earthquake disasters based on geo-ontology and relevant rules of earthquake emergency decision-making [8]. Scheuer et al. constructed flood risk assessment ontology by integrating SWEET ontology and MONITOR ontology [9]. Tao et al. summarized the knowledge graph construction process and key methods for the integrated comprehensive disaster reduction, and took the Jiuzhaigou earthquake as an example to demonstrate the knowledge graph construction process and construction results for Jiuzhaigou earthquake disaster reduction [10]. However, the above studies only focused on the knowledge representation and ontology construction for a single disaster type when building a knowledge model without considering the unified knowledge graph for different disaster types.
+
+Furthermore, the existing knowledge graph models usually do not focus on strategies and rules based on common semantics so that it is hard to provide specific disaster-handling strategies in disaster emergency scenarios for supporting decision-making. The disaster prediction knowledge graph developed by Qiu et al. [7] visualizes the disaster situation. It can more autonomously determine the environmental models and disaster-related data used in different disaster phases using the semantic constraints defined in the ontology [11]. Li et al. constructed a knowledge graph for flow disasters with the personalized visualization of different scenes for multilevel users [12]. Xie et al. combined the complex semantic relationship between earthquake prevention and control entities by extracting various earthquake disaster information, and established a semantic network by constructing a knowledge graph of earthquake disaster prevention and mitigation as the core [13]. Du et al. proposed a top-down and bottom-up method for constructing a
+
+Remote Sens. 2022, 14, 1214 3 of 21
+
+natural disaster emergency knowledge graph [14]. They realized the transformation from multi-source data to interconnected knowledge by taking the flood disaster emergency knowledge graph as an example for experimental verification. At present, the most related research on the disaster knowledge graph respectively works on the perspectives of disaster ontology, disaster knowledge, and disaster events, particularly paying attention to construction methods. In applications, there are only domain knowledge-driven analyses. This rarely integrates dynamic disaster data and model methods deeply for the analysis based on the fused multi-feature fusion. It leads to the problem that the emergency plan cannot be flexibly adapted to the actual situations.
+
+Affected by the attribute characteristics of surface elements, the development of disasters will show certain patterns when the disasters spatially migrate and change over time. This study selects forest fires and geological landslides as typical research objects. This paper proposes the Disaster Prediction Knowledge Graph (DPKG) that facilitates multi-source spatio-temporal data management and knowledge association. The proposed method supports semantic association of spatio-temporal data, efficient query, and quantitative analysis in order to implement information analysis with multiple types of disasters. The proposed method can be used for intelligent deduction of disaster situation and intelligent identification of disaster risk. This paper demonstrates experiments for verifying the benefits of the proposed method with respect to forest fire prediction and geological landslide risk prediction.
+
+The main contributions of this paper are as follows: (1) This paper proposes a disaster prediction method which realizes dynamic data-driven disaster prediction based on the DPKG. (2) This paper proposes a fusion method for the dynamically changing spatio-temporal data and knowledge model. The proposed method is beneficial to improve the ability of natural disaster monitoring, early warning, and emergency response. It can provide a quantitative reference for disaster prediction and play a practical role to a certain extent.
+
+# 2. Materials and Methods
+
+## 2.1. DPKG Architecture
+
+This section describes the theoretical way for mapping disaster prediction scenarios to knowledge graphs, mainly including the hierarchical semantic model of disaster prediction, and the mapping relationship between disaster prediction scenarios and DPKG architecture.
+
+### 2.1.1. Hierarchical Semantic Model of Disaster Prediction Scenarios
+
+The occurrence of disasters is related to a variety of factors. It facilities disaster prediction and decision-making by the analysis of disaster-inducing factors and their interrelationships.
+
+The affect factors of occurrence of natural disasters are composed of common factors and characteristic factors. The common factors are closely related to different types of disasters. The characteristic factors are closely related to the occurrence of a certain natural disaster and weakly related to other types of natural disasters. This study establishes a hierarchical semantic model of disaster prediction considering the affect factors, which provided a theoretical basis for the construction of a conceptual layer for forest fire risk prediction and geological landslide risk prediction, as shown in Figure 1. Aiming at the prediction of forest fires and geological landslides, this paper extracts the common factors and characteristic factors of disaster occurrence by considering remote sensing information (vegetation, farmland, roads, buildings, and water bodies), terrain, meteorology, and human factors closely related to the occurrence of disasters.
+
+Remote Sens. 2022, 14, 1214 4 of 21
+
+Image /page/3/Figure/1 description: A flowchart illustrating the factors used for Forest Fire Risk Prediction and Geological Landslide Risk Prediction. The chart is divided into three color-coded sections. On the left, in a pink section, are "Common factors in disasters," which include Terrain (Slope, Aspect, Height), Meteorological (Wind speed, Wind direction, Precipitation, Temperature, Humidity), and Ground cover (Arable land, Woodland, Water, ......). These common factors feed into a central box labeled "Disaster Occurrence Characteristics." From this central box, the chart branches into two specific disaster types. The top right, in a blue section, details factors for "Forest Fire Risk Prediction," specifically the "Combustibles factor," which includes Vegetation types, Vegetation Coverage, Species, and Tree age. The bottom right, in a green section, details factors for "Geological Landslide Risk Prediction," specifically "Geological factors," which include Stratigraphic lithology, Fault structure, Surface roughness, Surface relief, Plane curvature, and Section curvature. Brackets indicate that the common factors combined with the combustibles factor lead to Forest Fire Risk Prediction, and the common factors combined with geological factors lead to Geological Landslide Risk Prediction.
+
+**Figure 1.** A hierarchical semantic model for disaster prediction scenarios regarding forest fires and geological landslides.
+
+### 2.1.2. Mapping between Disaster Prediction Scenarios and DPKG Architecture
+
+For modeling disaster prediction using a knowledge graph, it is necessary to represent the disaster prediction scenario as the hierarchical semantic model. It maps the data with spatio-temporal characteristics in the disaster prediction scenario into the DPKG according to the structure of the semantic model. The disaster analysis model can be transformed into knowledge. The DPKG also needs to integrate the transformed disaster analysis model and expert knowledge for disaster prediction.
+
+In addition, the architecture design of the DPKG for disaster prediction needs to consider the fusion of multi-source heterogeneous spatio-temporal data including remote sensing extraction information, terrain, meteorology, vegetation, and human factors, as well as the fusion of spatio-temporal data with disaster analysis models and expert knowledge for linkage analysis. This study designs an architecture of DPKG for disaster prediction, as shown in Figure 2, in order to support the modeling of disaster prediction and the collaborative analysis with fused data.
+
+Image /page/3/Figure/6 description: A flowchart diagram illustrating the 'DPKG architecture'. The architecture is divided into several components within nested boxes. The main section is split into two parts: 'Set of rules (TBox)' and 'Set of facts (ATox)'. The 'Set of rules (TBox)' contains a 'Conceptual layer' (with 'Space ontology', 'Time ontology', and 'Common semantic ontology of disaster prediction') and 'Inference rules' (with 'First-order logical reasoning rules', 'Production inference rules', and 'Spatio-temporal semantic reasoning rules'). The 'Set of facts (ATox)' shows a flow where 'Unstructured data', 'Semi-structured data', 'Structured data', and 'Disaster prediction reasoning criterion' feed into a 'Knowledge extraction' process, which results in a 'Set of facts (ABox)'. An arrow points from this main section to two modules at the bottom: 'Disaster fusion data spatio-temporal semantic query module' and 'Disaster prediction rule reasoning module'.
+
+Figure 2. DPKG architecture.
+
+Remote Sens. 2022, 14, 1214 5 of 21
+
+The DPKG for disaster prediction consists of two parts, namely the rule set (TBox) and the fact set (ABox). The TBox consists of the conceptual layer and the inference rules. The conceptual layer is the common semantic basis for describing the hierarchical relationship of various factors in disaster prediction scenarios. The inference rules are the common logic basis for multi-source spatio-temporal data that supports semantic reasoning. The ABox constitutes the instance layer of the knowledge graph, which contains the time, space, status, and other attribute information of various types of ground objects related to disaster emergency extracted from multi-source heterogeneous data. The ABox and the TBox constitute the reasoning basis of the DPKG. Through the rapid retrieval of spatio-temporal data in disaster areas, multi-source spatio-temporal data analysis can be realized. The architecture diagram will be described in detail later.
+
+## 2.2. Construction of Disaster Prediction Knowledge Graph
+
+This section introduces the knowledge representation language, the design of the conceptual layer and instance layer, and the knowledge representation of the disaster prediction model.
+
+### 2.2.1. Knowledge Representation Language
+
+Knowledge graph representation refers to the use of computer symbols to scientifically mark the objectively existing knowledge for facilitating the semantic reasoning [15].
+
+OWL (Web Ontology Language) is a semantic description standard including three sub-languages: OWL-Lite, OWL-DL, and OWL-Full. At present, OWL is the most standardized, rigorous, and expressive language for a knowledge graph. The commonly used knowledge representation form is RDF (Resource description framework) triples. Each piece of knowledge can be represented as the following triple format: (subject, predicate, object), such as (vegetation, yes, combustible). This study chooses OWL as the knowledge representation language.
+
+### 2.2.2. Design of Conceptual Layer
+
+The conceptual layer of the DPKG is a logical structure for multi-source spatiotemporal data. It contains semantic concepts and their interrelationships. Based on the semantic associations between different concepts, such as the subordinate relationship, attribute subject-object relationship, a tree-like hierarchical concept triple interconnection network is constructed for ensuring the consistence of the intrinsic semantic concepts of multi-source spatio-temporal data. The conceptual layer of disaster prediction DPKG is composed of time ontology, space ontology, and disaster prediction common semantic ontology.
+
+#### 1. Common Semantic Ontology of Disaster Prediction
+
+The dynamic prediction of forest fires and geological landslide disasters needs the semantic reasoning and collaborative computing of multi-source spatio-temporal data.
+
+For this goal, this paper proposes a tree-like taxonomy of multi-source concepts of geographic entity associated with multi-source spatio-temporal data. According to the taxonomy, concepts of geographic entity are divided into five fields. This paper defines the domain attribute predicates related to the concepts of each geographic entity.
+
+The concepts of geographic entity constitute a common semantic ontology for the prediction of forest fire and geological landslide disaster.
+
+The storage model of common semantic ontology is hierarchical and expandable. It builds the foundation for the predicate logic reasoning based on the common hierarchical semantic relationship of disaster prediction, as shown in Figure 3.
+
+Remote Sens. 2022, 14, 1214 6 of 21
+
+Image /page/5/Figure/1 description: A black and white flowchart titled 'Conceptual Framework of Disaster Prediction KG'. The main concept branches into three categories: 'Time object', 'Space object', and 'Role object'. 'Time object' further divides into 'Effective time' and 'Time granularity'. 'Space object' divides into 'Geographic description' and 'Features'. 'Role object' divides into 'Mechanism' and 'Post'. The 'Features' category from 'Space object' is connected to a lower level of five boxes: 'Ground cover', 'Meteorological', 'Topography', 'Geology and lithology', and 'Historical disaster'.
+
+Figure 3. Common semantic ontology of disaster prediction.
+
+#### 2. Time Ontology
+
+Time ontology provides a specification of unified time semantic representation in order to ensure that the time information of entities is comparable and computable. This paper leverages the Semantic Web Rule Language (SWRL) time ontology proposed by Stanford University for representing the common time concepts of the DPKG [16]. The logical structure of the SWRL time ontology is shown in Figure 4.
+
+Image /page/5/Figure/5 description: A diagram illustrating a data model or ontology for time-related concepts. It shows several colored rectangular boxes connected by labeled arrows. A pink box labeled "ExtendProposition" connects to a central blue box "ValidTime" via two arrows: a solid one labeled "hasValidTime" and a dotted one labeled "hasPredictTime". The "ValidTime" box connects to a yellow box "Granularity" with an arrow labeled "hasGranularity". Below "ValidTime", a light green box "ValidInstant" and a yellow box "ValidPeriod" are shown as subclasses, indicated by arrows labeled "SubClassOf" pointing to "ValidTime". At the bottom, a large red box is labeled "xsd::DateTime". "ValidInstant" has an arrow labeled "hasTime" pointing to "xsd::DateTime". "ValidPeriod" has two arrows pointing to "xsd::DateTime", one labeled "hasStartTime" and the other "hasFinishTime".
+
+Figure 4. SWRL temporal ontology.
+
+#### Space Ontology
+
+The expression of spatial ontology applies A Geographic Query Language for RDF Data (GeoSPARQL) [17] that is a geographic semantic query specification proposed by OGC (Open Geospatial Consortium) [18]. The geographic semantic query mainly includes the following three common contents.
+
+#### OWL Ontology Vocabulary
+
+It represents all feature entities as subclasses of Spatial Objects. All geometric objects can be subdivided into points, linestrings, and polygons. Geometric objects are represented using features and geometric objects (Geometry) and their defined relationships, as shown in Figure 5.
+
+Image /page/5/Figure/11 description: A class diagram illustrates the relationships between different spatial object types, prefixed with 'ogc:'. At the top is the superclass 'ogc:SpatialObject'. Two classes, 'ogc:Feature' and 'ogc:Geometry', inherit from 'ogc:SpatialObject'. There is a directed association from 'ogc:Feature' to 'ogc:Geometry' labeled 'ogc:hasGeometry'. Three other classes, 'ogc:Point', 'ogc:LineString', and 'ogc:Polygon', all have arrows pointing to the 'ogc:Geometry' class. The 'ogc:Geometry' class is detailed in a larger box with two sections: 'Metadata' and 'Serialization'. Under 'Metadata', the properties are listed as: 'ogc:dimension : xsd:int', 'ogc:coordinateDimension : xsd:int', 'ogc:spatialDimension : xsd:int', 'ogc:isEmpty : xsd:boolean', 'ogc:isSimple : xsd:boolean', and 'ogc:is3D : xsd:boolean'. Under 'Serialization', the properties are 'ogc:asWKT' and 'ogc:WKTLiteral'.
+
+Figure 5. GeoSPARQL part of the ontology table.
+
+Remote Sens. 2022, 14, 1214 7 of 21
+
+#### Geometry extension expression
+
+The GeoSPARQL specification follows the OGC standard for the expression format of point, line, and surface. For example, the coordinate information of a surface data will be recorded in the attribute of *ogc:asWKT*, where *WKT* stands for Well Known Text, which is the ASCII code representation method of spatial objects.
+
+#### Topology extension query
+
+GeoSPARQL specifies eight query relationships for all spatial objects, as shown in Table 1. In addition, it provides many commonly used distance query functions (*geof:distance*), buffer query functions (*geof:buffer*), and convex Package constructor (*geof:convexHull*), to enhance spatial query.
+
+Table 1. GeoSPARQL spatial topological relationship.
+
+| geof:sfEquals  | geof:sfDisjoint | geof:sfIntersects | geof:sfTouches  |
+|----------------|-----------------|-------------------|-----------------|
+| geof:sfCrosses | geof:sfWithin   | geof:sfContains   | geof:sfOverlaps |
+
+### 2.2.3. Construction of Disaster Prediction Inference Rules
+
+#### Construction of First-Order Logical Inference rules
+
+First-order logic (FOL) is a formal system used in mathematics, philosophy, linguistics, and computer science. Specifically, hierarchical relationships and object attribute relationships can be described as first-order predicate logic, as shown in Figure 6. The common ontology, time ontology, and space ontology of disaster prediction are formally modeled with OWL language. The hierarchical relationship and attribute relationship implies a series of first-order logical reasoning.
+
+```
+1
+```
+
+Figure 6. Hierarchical relationship and object attribute relationship.
+
+#### 2. Construction of Production Inference rule
+
+Due to the limitation of the knowledge representation of first-order logic, this paper applies SWRL (Semantic Web Rule Language). It extends the set of OWL axioms to include Horn-like rules. In this study, SWRL is used to further enhance the expression of rules for DPKG. A rule has the form, where both antecedent and subsequent are conjunctions of atoms written. Using this syntax, a rule asserting that the combination of the tree species' properties and the tree species' forest fire risk-level properties implies the tree's forest fire risk-level properties, as shown in Figure 7.
+
+```
+\langle species(?x,?y) \land forestFireRiskLevel(?y,?z) \Rightarrow forestFireRiskLevel(?x,?z) \rangle
+```
+
+**Figure 7.** SWRL inference rule example.
+
+#### 3. Construction of Spatio-Temporal Semantic Inference Rule
+
+Although SWRL can support basic quantitative calculations, it cannot support quantitative analysis of space and time including spatio-temporal semantics. This study leverages the construction method of production inference rules to build the spatio-temporal semantic inference rules. The spatio-temporal semantic inference rule *RuleObject* is a set of rules for the automatic execution of reasoning programs by the DPKG. Each rule is composed
+
+Remote Sens. 2022, 14, 1214 8 of 21
+
+of an event object TriggerObject and an action object ActionObject, which are expressed as RuleObject = (Tr, Ac), where Tr represents the event contained in the RuleObject, Ac represents the action object ActionObject contained in the RuleObject, and R is the reasoning result. The event TriggerObject in this article is defined as a triple that is represented as TriggerObject = (O, T, S), where O represents the geographic entity set contained in the event object, T and S respectively represent the intersection of the geographic entity set in the time dimension and the space dimension. A spatio-temporal co-occurrence scene with a set of geographic entities can be described as an event object that is a definition of the applicable conditions of an inference rule.
+
+The event (or action) object concept is further subdivided into independent events (or actions) and event (or action) combinations, which together form the concept of knowledge inference rules. The logical structure is shown in Figure 8.
+
+Image /page/7/Figure/3 description: A hierarchical diagram illustrating the structure of a 'RuleObject'. The 'RuleObject' is at the top and branches into two main components: 'TriggerObject' on the left and 'ActionObject' on the right. The 'TriggerObject' branch is further subdivided. It points to 'IndependentTrigger' and 'TriggerCombination'. 'IndependentTrigger' then points to 'NormalTrigger' and 'AbnormalTrigger'. 'TriggerCombination' points to 'AndTrigger Combination' and 'OrTrigger Combination'. The 'ActionObject' branch is similarly subdivided. It points to 'IndependentAction' and 'ActionCombination'. 'ActionCombination' then points to 'AndAction Combination' and 'OrAction Combination'. There are also arrows pointing upwards, indicating relationships back to parent objects. For instance, on the right side, arrows point from 'IndependentAction' and 'ActionCombination' back to 'ActionObject', and from 'AndAction Combination' and 'OrAction Combination' back to 'ActionCombination'. A similar pattern of upward-pointing arrows exists on the 'TriggerObject' side of the diagram.
+
+Figure 8. Schematic diagram of the conceptual hierarchy of knowledge inference rules.
+
+### 2.2.4. Design of Instance Layer
+
+#### Knowledge Extraction from Unstructured Data
+
+The dynamic disaster prediction of forest fires and geological landslides requires high spatio-temporal resolution of land cover data. This paper uses Gaofen-2 satellite remote sensing images with the resolution of 0.8 m as the data source. The spatial distribution of disaster-bearing bodies such as buildings and roads is extracted by deep learning methods. The spatial distribution of surface vegetation is obtained by NDVI numerical calculation method [1,19]. It converts the spatial information into triples according to the specification of spatio-temporal and professional attribute representation defined by the concept layer. The triples are stored in the graph database *GraphDB*.
+
+#### 2. Knowledge Extraction from Semi-Structured Data
+
+The terrain data for dynamic disaster prediction of forest fires and geological landslide is the raster geographic data with GeoTIFF format. It converts all types of raster data into vector data of surface elements. It converts geological and lithological data including stratigraphic age, fault, and lithology distribution data into vector geographic information with SHP format. For all types of vector geographic information, it converts the spatial information and feature attributes into triple that are stored in *GraphDB*.
+
+#### 3. Knowledge Extraction from Structured Data
+
+The meteorological data for the dynamic disaster prediction of forest fires and geological landslide is multi-field structured data, which has a direct mapping relationship with the spatio-temporal and professional attributes. Since the vast majority of meteorological data is normal non-hazardous data, this paper only converts the spatial information element attributes of potential disaster-causing meteorological indicators into a triple, so as to avoid the low reasoning speed caused by a large amount of irrelevant data to the reasoning. The triple is stored in *GraphDB*.
+
+Remote Sens. 2022, 14, 1214 9 of 21
+
+#### 4. Knowledge Extraction from Disaster Prediction Reasoning Criterion
+
+This paper introduces the knowledge extraction method by taking the geological landslide prediction model as an example. It is hard to build a general prediction model for a geological landslide because a single model with idiosyncratic conditions cannot deal with the rich types of geological landslide prediction models and the involved data. This paper defines the inference rules for the geological landslide disaster prediction according to the geological landslide risk probability calculation using machine learning and the effective precipitation models. Based on the general law of geological landslide risk analysis [20], this paper selects the H-index shown in Formula (1) as the metric for the quantitative evaluation of the geological landslide risk:
+
+$$H = P_{(T)} \times P_{(S)} \times P_{(I)}, \tag{1}$$
+
+where H is the risk probability of geological landslide disasters within a specific time and space range;  $P_{(T)}$  is the time probability, i.e., the probability of geological landslide disasters occurring within a specific time range;  $P_{(s)}$  is spatial probability, i.e., the probability of geological landslide disasters occurring within a specific spatial range;  $P_{(I)}$  is the intensity probability, i.e., the intensity of possible geological landslide disasters.
+
+This paper introduces a time probability model based on dynamic precipitation data. The model is mainly composed of two parts. One is the effective precipitation; another is the fitting relationship between the effective precipitation and the frequency of geological landslides. The effective precipitation R in the region considers the respective precipitation contributions of the previous days to the cumulative precipitation that causes the geological landslide. The later the time, the greater the contribution rate to the geological landslide.
+
+$$R = R_0 + \alpha_1 R_1 + \ldots + \alpha_n R_n, \tag{2}$$
+
+where  $R_0$  is the daily precipitation;  $\alpha$  is the contribution rate; n is the number of precipitation days, and  $R_n$  is the precipitation on the n-th day before. A linear fitting relationship between the effective precipitation and the frequency of geological landslides is defined for the time probability of geological landslides associated with dynamic precipitation data.
+
+The geological landslide strength is considered as a parameter representing the destructive force of the geological landslide. It includes the volume of the geological landslide, and the sliding velocity. At present, there is no unified set of indicators to describe them. In this paper, the geological landslide volume parameter is used to obtain the geological landslide strength, and the formula is as follows:
+
+$$m_L = log V_{L}, \tag{3}$$
+
+where  $V_L$  is the volume of a single geological landslide with the unit of m<sup>3</sup>; which is calculated based on DEM, and  $m_L$  represents the strength of a single geological landslide. By calculating the strength of each geological landslide in the study area by the above formula, it can obtain the statistics of geological landslide strength according to the frequency.
+
+There are many common factors that lead to forest fires and geological landslides. Certain triggering factors can be monitored when forest fires and geological landslides occur. The development of the two types of disasters will develop and spread within a certain time and space range. Affected by the attribute characteristics of surface elements, both disasters will show certain change patterns when they dynamically migrate and change in space over time. Therefore, this paper extends the geological landslide risk calculation formula to forest fires by considering the above-mentioned commonalities of the two disasters. It builds the unified prediction model for forest fire according to Formula (1). Based on Analytic Hierarchy Process, it builds the forest fire-driven comprehensive index for forest fire risk prediction model.
+
+Remote Sens. 2022, 14, 1214 10 of 21
+
+## 2.3. Query and Reasoning of Disaster Prediction DPKG
+
+### 2.3.1. Spatio-Temporal Semantic Query for Disaster Fusion Data
+
+When a specific attribute of the external dynamic data satisfies the numerical conditions defined by the rule, it is regarded as an abnormal event candidate. Its time range and space range are treated as the basic condition of semantic query. It queries with SPARQL Protocol and RDF Query Language (SPARQL) [21] statements for a geographic entity that has an intersection with the abnormal event candidate and satisfies the reasoning conditions. When querying the spatio-temporal relationship in the graph database, it can quickly locate the affected area of the abnormal event candidate because the graph database is good at efficient depth-first query, as shown in Figure 9. It plays an important role in the emergency field with high response speed requirements.
+
+#### Disaster prediction reasoning flow chart
+
+Image /page/9/Figure/5 description: A flowchart illustrating a process for handling dynamic data to identify and query abnormal events. The process begins at a 'Start' node and proceeds to 'Real-time input of dynamic data (such as meteorological data)'. Next, a decision is made in an oval labeled 'Judging the dynamics Whether the data attribute satisfies the condition'. If the condition is not met ('No'), the process goes directly to the 'End' node. If the condition is met ('Yes'), the flow continues to 'Generate abnormal events to be judged based on the spatiotemporal scope of external dynamic data'. This is followed by 'Based on the abnormal events to be determined and the geographic location affected by the abnormal events Entity Type Generate Query Statement'. The next step is 'With the help of efficient in-depth retrieval of graph database, the area affected by abnormal events can be quickly locked'. This leads to the final process step, 'Query to get the affected geographic entities within the scope of the abnormal event', which then connects to the 'End' node.
+
+Figure 9. Disaster prediction query flow chart.
+
+With the help of the spatio-temporal semantic query of the disaster fusion data, it can quickly find geographic entities located within the temporal and spatial scope of the AbnormalTrigger based on the temporal-spatial intersection relationship in order to filter the buildings and roads affected by the AbnormalTrigger, as shown in Figure 10.
+
+Image /page/9/Figure/8 description: A conceptual diagram on a 2D coordinate system. The horizontal axis is labeled "spatial dimension" and the vertical axis is labeled "time dimension". The diagram features two large, overlapping ovals. The oval on the left is light green and is associated with the green text "Spatio-temporal distribution of buildings and roads", with a green arrow pointing up from the oval to the text. The oval on the right is light orange and is associated with the orange text "Spatial and temporal distribution of AbnormalTrigger", with an orange arrow pointing up from the oval to the text. The overlapping area of the two ovals is colored reddish-pink. A pink arrow points from this intersection to the text below, which reads "Buildings and roads affected by AbnormalTrigger" in pink.
+
+**Figure 10.** Schematic diagram of the spatio-temporal intersection of anomalous events and geographic entities.
+
+In order to ensure the efficiency of spatial information query in disaster emergency scenarios, this paper builds multi-scale geocoding indexes for multi-source geographic entities in GeoSPARQL. It significantly improves the depth-first query speed of random geospatial information in comparison with the frequent joint query of multiple tables in the relational database.
+
+Remote Sens. 2022, 14, 1214 11 of 21
+
+### 2.3.2. Disaster Prediction Rule Reasoning
+
+#### 1. First-order logical reasoning
+
+As defined in the Common Semantic Ontology of Disaster Prediction, surface combustibles include Woodland, Farmland, Building, and Grassland, and Coniferous Forest belongs to Woodland. Based on first-order logical reasoning, it concludes that Coniferous Forest belongs to Surface Combustibles. With the prefix of *hazard*, the above process can be formally defined as shown in Figure 11. Based on these expressions, *(hazard : Coniferous Forest rdf : subClassOf hazard : Surface Combustibles)* can be inferred.
+
+```
+1
+```
+
+Figure 11. First-order logical reasoning process.
+
+#### 2. Production reasoning
+
+This paper takes SWRL rule reasoning as an example. The feasible reasoning on the DPKG includes reasoning of geographic entity from one attribute to other attributes; reasoning from one geographic entity and its attributes to geographic entities and their attributes.
+
+For example, given the forest fire risk entity and its attribute, i.e., the comprehensive forest fire-driven index *fireDangerValue*, the goal is to build an inference rule for the forest fire risk-level prediction. For this task, the domain analysis rule "when fireDangerValue > 3.8, fireDangerLevel is level 3" is formalized as shown as line 1 in Figure 12.
+
+```
+1
+```
+
+Figure 12. Production reasoning process.
+
+One more example, knowing the value of aspect entity and aspect attribute, the goal is to construct an inference rule to infer the upslope aspect value. For this task, the domain analysis rule "When aspect is  $<180^{\circ}$ , upAspect = aspect +  $180^{\circ}$ " is formalized as shown as line 2 in Figure 12.
+
+The prediction can be carried out using the SWRL reasoning engine with the inference rules.
+
+#### 3. Spatio-temporal semantic reasoning
+
+This paper builds an automatic real-time monitoring mechanism for emergency information for disaster prediction. It extracts the dynamic information related to the occurrence and development of disasters, i.e., attributes such as space, time, and status of target objects, from structured, semi-structured, and unstructured data sources. It encapsulates the extracted information into real-time event message objects with GeoJSON format. The update of dynamic data will trigger the automatic judgment by the DPKG as shown in Figure 13. The results drive the chained reasoning workflow defined by the disaster prediction inference rules.
+
+Remote Sens. 2022, 14, 1214 12 of 21
+
+#### **Calculation process**
+
+Image /page/11/Figure/2 description: A flowchart titled 'Calculation process' illustrates a system for predicting geological and forest fire risks. The process begins with two inputs: 'Weather forecast data for the next 7 days (updated at 8:00 and 20:00 daily)' and 'Weather monitoring data (lag about 10-20 minutes, updated every 5 minutes)'. Both are fed as 'input' into a process called 'Instance layer dynamic update'. The output of this stage is then used as 'input' for the central calculation step, 'The triplet disaster prediction model starts the calculation'. This model also utilizes 'Spatio-temporal knowledge graph data'. The model produces two 'output' streams: 'Prediction results of geological landslide risk in the next 7 days' and 'Forecast results of forest fire risk in the next 7 days'. The flowchart also includes two dashed lines indicating dependencies. The top line states, 'Forecast days depend on the availability of weather forecast data', connecting the forecast data to the final predictions. The bottom line states, 'Actual real-time performance depends on the lag of meteorological monitoring data', connecting the monitoring data to the final predictions.
+
+Figure 13. Workflow of disaster dynamic prediction.
+
+This paper designs a chain-type automatic reasoning method based on the reasoning criterion. The method divides the calculation logic of the geological landslide prediction model into a series of antecedents and consequences. The antecedents are execution conditions that contain a series of geographic entities. When there exists an intersection of the geographic entities and the conditions set in the antecedent are met, the subsequent associated with the antecedent, i.e., the action object, will be triggered.
+
+It establishes the automatic flow logic between independent criteria such as time probability, space probability, intensity probability, vulnerability of hazard-affected body, and value amount by defining for RuleObject. In this way, it can analyze the relations between time information and space information, which is not feasible using the production reasoning with SWRL.
+
+# 3. Results
+
+This section demonstrates experiments and cases studies of the forest fire and geological landslide risk based on DPKG. These investigations show that the proposed method is beneficial to multi-source spatio-temporal information integration and disaster predictions.
+
+## 3.1. Case Study: Forest Fire Risk Prediction
+
+In this case study, Yanyuan County, in the Liangshan Yi Autonomous Prefecture of China's Sichuan Province, was selected to carry out a dynamic prediction experiment of forest fire disasters. Yanyuan County is located between  $100^{\circ}42'09''-102^{\circ}03'44''$  east longitude and  $27^{\circ}06'31-28^{\circ}16'31''$  north latitude, with a total area of 8398.6 square kilometers. There are rich vegetation and many miscellaneous irrigation and thatch grass in the area. The forest fire disasters are easily caused by the inducing factors such as low rainfall. Therefore, Yanyuan County strongly needs to forecast the situation in advance.
+
+In this case, multi-source heterogeneous data related to forest fire risk prediction are collected, such as vegetation data, terrain data, meteorological data, and land cover data. In order to fully represent the spatio-temporal features of the above data in the DPKG knowledge graph, this paper uses Protégé to construct a time ontology and a space ontology. The concepts related to forest fire prediction form the conceptual layer of DPKG. For the above-mentioned multi-source heterogeneous data, we built a diversified knowledge extraction method for transferring the data into triples according to the semantics of the conceptual layer.
+
+The triples form an instance layer of the DPKG of disaster prediction. An example of the extraction process of the slope in the terrain data is shown in Figure 14.
+
+Remote Sens. 2022, 14, 1214 13 of 21
+
+Image /page/12/Figure/1 description: A diagram illustrates a data processing workflow for geographic information. The flow starts with a TIFF file, a raster image of a geographical area. An arrow labeled 'Raster to Vector' points to a Shapefile, which is a colored vector map of the same area. From the Shapefile, an arrow labeled 'Vector to GeoJSON' points to a code snippet of a GeoJSON file. The GeoJSON code shows a 'FeatureCollection' with properties for a 'Slope'. An arrow labeled 'GeoJSON to Triple' points from the GeoJSON to a knowledge graph representation. This graph has a central node 'SlopeEntity' connected to four other nodes via labeled edges: 'hasSlope' points to '21', 'hasTileCode' points to 'z20\_x322\_y786', 'hasTime' points to '2021-10-27 01:17', and 'hasCenterLonLat' points to '102.98438°E, 26.502315°N'.
+
+Figure 14. Knowledge extraction process using slope as an example.
+
+Forest fire risk prediction needs the theoretical support because the above spatiotemporal data is only a prerequisite for forest fire risk prediction. In this case, the AHP (Analytic Hierarchy Process) is used to evaluate forest fire risk indicators. This paper builds an indicator system for four types of forest fire risk factors: meteorology, terrain, vegetation, and man-made. The weights are shown in Figure 15.
+
+Image /page/12/Figure/4 description: A flowchart diagram illustrating the 'Comprehensive Index of Forest Fire Drivers'. This main index branches into four primary factors, each with a numerical weight. The 'Vegetation factor' has a weight of 0.24 and is subdivided into 'Vegetation Types' (0.18) and 'Canopy' (0.06). The 'Meteorological factor' has the highest weight of 0.40 and is broken down into four sub-factors: 'Wind speed' (0.10), 'Precipitation' (0.10), 'Temperature' (0.10), and 'Humidity' (0.10). The 'Terrain factor' has a weight of 0.21 and includes 'Slope' (0.10), 'Aspect' (0.06), and 'Height' (0.05). Finally, the 'Human factor' has a weight of 0.15 and is represented by a single sub-factor, 'Festival' (0.15).
+
+Figure 15. Forest fire risk factor indicator system.
+
+The comprehensive indicator of forest fire driving factors is defined as Formula (4):
+
+$$CIFFD = \sum v_i \times w_i \ i \in [1, 10], \tag{4}$$
+
+where *CIFFD* represents a composite index of forest fire drivers;  $v_i$  indicates the level corresponding to the original value of the factor, and  $w_i$  indicates the weight corresponding to the factor. The larger the CIFFD, the greater the risk of forest fires.
+
+In this case, 10 factors in four categories of terrain, vegetation, meteorology, and human factors in the experimental area are used as the query input for the DPKG instance layer. The definition in Formula (4) is transformed into spatio-temporal semantic reasoning rules. The reference output is a comprehensive index of forest fire driving factors.
+
+Define spatio-temporal semantic inference rule 1: If the update of meteorological data is detected, it divides the spatial range of the meteorological data update into several rectangular areas. The rectangular areas automatically triggers the query of 10 factors of
+
+Remote Sens. 2022, 14, 1214 14 of 21
+
+terrain, vegetation, meteorology, and human factors in the rectangular areas for obtaining the original value of the factor.
+
+Define spatio-temporal semantic inference rule 2: It obtains the corresponding level value based on the original value of the factor. Based on Formula (4), the comprehensive index of forest fire driving factors is obtained.
+
+Define spatio-temporal semantic inference rule 3: Based on the natural discontinuity method, the calculated comprehensive index of forest fire driving factors is divided into 7 intervals for representing different degrees from low to high. The forest fire risk warning information of the experimental area is obtained, as shown in Figure 16.
+
+Image /page/13/Figure/4 description: A flowchart diagram titled "Semantic inference of forest fire disaster". The diagram is divided into three vertical sections by dashed lines, labeled "Rule one", "Rule two", and "Rule three".
+
+In "Rule one", the process starts with "Meteorological data update", which leads to "Divide rectangular area". An arrow points from this to a box labeled "Query regional forest fire-related factors".
+
+An arrow from "Rule one" points to "Rule two". In this section, there are two icons: a square with a diagonal line and a circle inside, and a stacked triangle. Below them, the text reads "Gets the corresponding level value". This leads to a box labeled "Forest fire driver composite index".
+
+An arrow from "Rule two" points to "Rule three". In this final section, there is a bar chart with bars of increasing height, labeled "Forest fire driving factor composite index divided into intervals". The process concludes in a box labeled "Get rank".
+
+Figure 16. Forest fire risk factor index system.
+
+The antecedent rules in the above spatio-temporal semantic inference rules can trigger the subsequent rules when satisfying the conditions. Therefore, the dynamic changes of meteorological data can automatically trigger the forest fire risk calculation. It realizes the automatic completion of the chain reasoning of the whole process of DPKG without manual effort.
+
+In this case, the actual meteorological record data of Yanyuan County Meteorological Observatory on 3 January 2022 is used as the query input of the instance layer of the DPKG. The output is calculated based on the reasoning of the DPKG.
+
+For the forest fire risk index of the whole county of Yanyuan, the larger the index value, the higher the forest fire risk. The experimental results are shown in Figure 17.
+
+Image /page/13/Figure/9 description: A map displaying the Forest Fire Risk Index for an irregularly shaped geographical area. The map uses a color scale to represent different levels of risk. A legend on the left side details the index values corresponding to each color: dark green for 0.85-2.90, medium green for 2.91-3.60, light green for 3.61-4.00, yellow for 4.01-4.10, light orange for 4.11-4.20, orange for 4.21-4.60, and red for 4.61-5.29. Scattered across the map are several purple dots, which the legend identifies as 'January 3, 2022 06:35 fire point'. In the bottom right corner, there is a scale bar marked in miles, with increments shown at 0, 3, 6, 12, 18, and 24 miles.
+
+**Figure 17.** Map of landslide hazards in Yanyuan County based on multi-Graded Cascade Random Forest.
+
+Remote Sens. 2022, 14, 1214 15 of 21
+
+NASA's Fire Information for Resource Management System shows the actual fire point in Yanyuan County at 6:35 on 3 January 2022, as shown in Figure 17. The results show that the proposed method successfully hits the fire point of the forest fire species within the time and space range of the predicted forest fire risk index greater than 4.21.
+
+The case uses the traditional spatial analysis tool ArcGIS as a baseline method to calculate the forest fire risk in Yanyuan County for the comparative experiments. It is necessary to load and query Yanyuan County data from 14 types of raster datasets. Steps such as uniform coordinate system, uniform pixel size, grid cropping, and stitching are performed on each raster data, and then the prediction calculation can be performed. There are many steps in the baseline operation, and the whole process takes more than 190 min. When using the grid calculator, the performance is low because it needs complex numerical computations in multiple steps.
+
+In addition, the boundaries of various data are not completely coincident, which will lead to missing or abnormal results of the boundary area analysis, and wrong predictions. Although it is possible to forcibly align the pixel positions by means of translation, it will cause errors by an operation that lacks a realistic basis, and changes the spatial distribution of the original data. In addition, it has a low prediction efficiency when using the traditional spatial analysis tools because they cannot process other target regions in parallel.
+
+## 3.2. Case Study: Geological Landslide Risk Prediction
+
+In another case study, Xiji County in Guyuan City of China's Ningxia Hui Autonomous Region is selected to carry out an experiment regarding the dynamic prediction of geological landslide hazards. Xiji County locates between  $105^{\circ}20'-106^{\circ}04'$  east longitude and  $35^{\circ}35'-36^{\circ}14'$  north latitude, which has a total area of 3130 square kilometers. In this area, the loess landslide disaster caused by the Haiyuan earthquake is particularly serious because of the special geological environmental conditions. It has formed 765 geological landslides [22]. It is very dangerous that new geological landslide disasters are easily generated by the inducing factors, e.g., rainfall. Therefore, it is urgent to take an effective method to predict disaster in advance.
+
+In this case, the basic geological data, basic geographic data, hydrological data, human activities, and land is collected for the prediction and calculation of geological landslide risk. In order to fully represent the spatio-temporal features of the above data in the knowledge graph, this paper uses Protégé [23] to construct a time ontology and a space ontology, and concepts related to geological landslide prediction form the conceptual layer of spatial-temporal knowledge graph for disaster prediction. For the above-mentioned multi-source heterogeneous data, a diversified knowledge extraction method is adopted to convert the data into triples according to the semantics of the conceptual layer. It forms the instance layer of the DPKG for disaster prediction using the triples.
+
+The above spatio-temporal data is only a prerequisite for geological landslide risk prediction. It needs the theoretical support of the geological landslide risk prediction. The landslide risk assessment is defined as Formula (2). Geological landslide risk prediction involves time, spatial, and intensity probability. In this case, there are 13 factors belonging to four categories of topography, geology, lithology, meteorology and hydrology, land cover, and human activities in the experimental area. The information of the factors is queried from the knowledge graph. It takes the DEM factors in the topographic data as the query input of the instance layer of the knowledge graph. The linear fitting model between historical landslides and DEM data is transformed into spatio-temporal semantic reasoning rules, as shown in Figure 18. In this way, the spatial probability of landslide disasters in the experimental area can be calculated.
+
+Remote Sens. 2022, 14, 1214 16 of 21
+
+Image /page/15/Figure/1 description: A flowchart illustrating a model for calculating Geological Landslide Risk Probability. At the top, a computer monitor icon is labeled "Geological Landslide Risk Probability," which points to a formula: H = P(T) × P(S) × P(I). The flowchart has two main input branches. The first branch combines three inputs: "The local precipitation data for the past 10 days is an empirical function of the independent variable," "Deep Random Forest Model," and "Linear fit model between historical landslides and DEM data." These are processed through a step called "Knowledge Representation" to generate "Spatiotemporal Semantic Reasoning Rules." The second branch combines three other inputs: "Precipitation rate," "Topography, Geology and lithology, Meteorology and hydrology, Land cover and human activities," and "DEM." These are processed through a step called "Instantiate" to create a "Disaster prediction spatiotemporal knowledge graph instance layer." This instance layer provides "Data Support" to the "Spatiotemporal Semantic Reasoning Rules." Finally, the "Spatiotemporal Semantic Reasoning Rules" are used to derive the three probabilities P(T), P(S), and P(I) for the final risk calculation.
+
+Figure 18. Geological landslide risk prediction calculation formula.
+
+The time probability is a dynamic element. In this case, the time probability is defined as the empirical function with the independent variable of the local precipitation data in the past 10 days, as shown in Formula (3). It transforms Formula (3) into spatio-temporal semantic inference rules, and the time probability of landslide disaster occurrence in the experimental area.
+
+Define spatio-temporal semantic inference rule 1: If the time probability satisfies specific conditions, it is defined as a dangerous time probability.
+
+Define spatio-temporal semantic reasoning rule 2: For a new dangerous time probability, the calculation of the spatial probability and the intensity probability will be automatically triggered. The spatial distribution of the geological landslide risk value in the test area will be calculated according to Formula (1).
+
+Define spatio-temporal semantic inference rule 3: Based on the natural discontinuity method, the geological landslide risk probability value calculated by the previous rule is divided into 10 intervals in ascending order, representing different degrees, every two intervals are combined into a grade for a total of five grades, as shown in Figure 19.
+
+Image /page/15/Figure/7 description: A flowchart diagram titled "Semantic reasoning of landslide disaster" is divided into three sections labeled "Rule one," "Rule two," and "Rule three." In "Rule one," two inputs, "Day time probability" and "Nearest time interval time probability," feed into a box labeled "Time probability of dangerous geological landslide." An arrow from this box leads to "Rule two." In "Rule two," the process is described as "Geographic grid computing," represented by an icon of a map with a grid overlay. This process results in the "Spatial distribution of geological landslide risk values." An arrow from this result leads to "Rule three." In "Rule three," the "Natural discontinuity method" is applied to the risk values. The outcome is visualized as five vertical bars of increasing height, representing different risk levels labeled below as "VL, L, M, H, VH," which likely stand for Very Low, Low, Medium, High, and Very High.
+
+**Figure 19.** Space-time semantic reasoning logic with criteria (VL: very low, L: low, M: moderate, H: high, VH: very high).
+
+The antecedent rules in the above spatio-temporal semantic inference rules can trigger the subsequent rules when the conditions are satisfied. Therefore, the dynamic change of precipitation data can automatically trigger the geological landslide risk calculation for
+
+Remote Sens. 2022, 14, 1214 17 of 21
+
+realizing the automatic completion of the whole process of DPKG chain reasoning without manual effort.
+
+In this case, the actual precipitation record data of Xiji County Meteorological Observatory on 18 September 2018 is used as the query input of the instance layer of the DPKG. Based on the reasoning of spatial-temporal knowledge graph, the geological landslide time probability was daily calculated for the whole county of Xiji from 8 to 28 September 2018. From 18 to 19 September 2018, the geological landslide time probability in the county met the above geological landslide risk time probability conditions for the two consecutive days. The experimental results are shown in Figure 20. Due to the setting of reasoning rule 2, the calculation of geological landslide risk will be triggered automatically. The calculation results of geological landslide risk probability on 18 September 2018 are shown in Figure 21.
+
+The graph of the time probability of landslide
+
+Image /page/16/Figure/4 description: A bar chart titled 'The graph of the time probability of landslide'. The vertical axis is labeled '%0' and ranges from 0 to 20 in increments of 2. The horizontal axis displays dates from 9/8 to 9/28. The chart consists of blue vertical bars for each day, showing the time probability. There are notable peaks in probability on 9/15, with a value of approximately 17.8, and on 9/19, with a value of approximately 18.5. Another smaller peak occurs on 9/18 with a value of about 16. For most other days, the probability fluctuates between approximately 6 and 10. A dashed blue line is overlaid on the chart, starting at a value of about 14, dipping to around 13, and then leveling off at approximately 12.5 for the remainder of the dates.
+
+Bar: The graph of the time probability of landslide in Xiji County from September 8 to 28, 2018
+
+Broken line: The graph of dangerous time probability of landslide in Xiji Countyfrom September 8 to 28,2018
+
+**Figure 20.** The graph of the time probability of landslide. The bar is the graph of the time probability of landslide in Xiji County from 8–28 September 2018. The broken line is the graph of dangerous time probability of landslide in Xiji County from 8–28 September 2018.
+
+Image /page/16/Figure/8 description: A map displaying the geological landslide risk probability for a specific region. The map is color-coded, with a legend indicating the probability levels. A north arrow is in the top left corner. The legend, titled 'Geological Landslide Risk Probability', shows a scale from low risk (dark green) to high risk (red). The probability ranges are as follows: 0-0.024 (dark green), 0.025-0.078 (green), 0.079-0.149 (light green), 0.15-0.231 (yellow-green), 0.232-0.329 (yellow), 0.33-0.439 (light orange), 0.44-0.553 (orange), 0.554-0.675 (dark orange), 0.676-0.812 (orange-red), and 0.813-1 (red). The map is predominantly green, with patches of yellow, orange, and red indicating higher-risk areas. A single yellow dot on the map is labeled 'September 18, 2018 Landslide section'. A scale bar at the bottom right is marked in meters, with intervals at 0, 6250, 12500, and 25000.
+
+Figure 21. Map of landslide hazards in Xiji County based on multi-Graded Cascade Random Forest.
+
+Remote Sens. 2022, 14, 1214 18 of 21
+
+In this case, since the distribution data of buildings and roads in Xiji County is included in the DPKG, it further filtered the buildings and roads in the high-risk areas of geological landslides on 18 September 2018, as shown in Figure 22. From Internet news, we learned that a landslide occurred on National Highway 309 on 18 September 2018, which caused road blockage. The actual geological landslide section is located in the dangerous road area that has been correctly predicted by our method, as shown in Figure 22. Experiments show that it successfully hit the geological landslide event within the time and space range of the warning.
+
+Image /page/17/Figure/2 description: A diagram illustrating the identification and data representation of geological hazards. On the left is a map of a region with a legend indicating different features: grey for Road, pink for Building, red for Risk road, orange for Risk Building, and purple for the September 18, 2018 Landslide section. The map includes a north arrow and a scale bar from 0 to 24000 meters. A small section of the map is magnified and shown in the center as a satellite image of a winding road in a mountainous area. This image has colored overlays corresponding to the legend. To the right, two entity relationship diagrams show how these features are structured as data. An arrow from the purple landslide section points to a 'Landslide Entity' diagram with the following attributes: 'hasCenterLonLat' is '102.98438°E, 26.502315°N', 'hasTime' is '2018-09-18', and 'hasTileCode' is 'z20\_x321\_y765'. An arrow from the red road section points to a 'RiskRoad Entity' diagram with attributes: 'hasCenterLonLat' is '102.98438°E, 26.502315°N', 'hasTime' is '2018-09-18', and 'hasTileCode' is 'z20\_x321\_y766'.
+
+**Figure 22.** Road distribution map of Xiji County included in the landslide disaster warning area on 18 September 2018.
+
+Prediction was carried out in Xiji Country respectively for the area of 3130.00 km<sup>2</sup>, the area of 1509.54 km<sup>2</sup>, and the area of 800.88 km<sup>2</sup>. The experimental results are shown in Figure 23. With the increase of the experimental range, the time of experimental calculation also increases. The performance is basically in a satisfied performance range. Therefore, our method is applicable for large-scale disaster prediction.
+
+Image /page/17/Figure/5 description: A horizontal bar chart with three bars. The vertical axis has three labels: 3130.00, 1509.54, and 800.88. The horizontal axis is labeled in minutes (min) and ranges from 0.00 to 10.00, with major gridlines at intervals of 2.00. The top bar, corresponding to 3130.00, extends to a value of 7.81. The middle bar, for 1509.54, has a value of 3.09. The bottom bar, for 800.88, has a value of 2.70. The bars are blue, with the top and bottom bars having a gradient from light to dark blue.
+
+Figure 23. Comparison of forecast time for different areas in Xiji Country.
+
+# 4. Discussion
+
+Compared with the traditional prediction methods, the disaster prediction method based on the DPKG has the advantages of more accurate spatial overlay analysis, high degree of automation, and fast calculation speed, as shown in Table 2. However, it cannot be ignored that the proposed method in this paper still has certain limitations: (1) Before forecasting the target area, the data required for disaster prediction needs to be preprocessed and stored in advance, and data storage requires a certain amount of time. The larger the area, the more data dimensions required for prediction, and the longer the time. (2) The data used in this case is not accurate enough. For example, the spatial resolution of terrain
+
+Remote Sens. 2022, 14, 1214 19 of 21
+
+data is 30 m, and the spatial resolution of surface coverage data for disaster prediction is 10 m. The accuracy still needs to be improved. If higher-resolution data is obtained, and more detailed tree species, tree age, and other data can be involved in disaster prediction, the accuracy of disaster prediction can be further improved. (3) The comparison with traditional disaster forecasting methods is not sufficient in the experiment that only considers the most well-known tool ArcGIS.
+
+
+
+# 5. Conclusions
+
+Compared with plain text data and structured databases, DPKG has the advantages of richer semantic representation, more accurate data content, and better query performance. The knowledge graph provides a new means for the organization, management, fusion, and analysis of multi-source heterogeneous data in complex disaster environments. It facilitates prediction, prevention, and mitigation of the natural disasters. This paper proposes the DPKG that integrates multi-source remote sensing information in disaster scenes with other multi-source heterogeneous information. It considers not only dynamic data describing spatio-temporal facts but also domain analysis models. Based on the analysis of disaster influencing factors, this paper proposes a common disaster prediction method based on the DPKG for the dynamic data-driven disaster prediction. This paper introduces new ideas by taking forest fires and geological landslides as examples. The proposed method can serve the research and application of disaster emergency response with the strong help of artificial intelligence technology.
+
+- (1) From the perspective of cross-domain knowledge integration, knowledge graph integrates remote sensing knowledge and expert knowledge through semantic technology. It effectively connects multi-source heterogeneous data (GIS, meteorology, terrain, ground sensors, etc.) with expert knowledge in the field of disasters;
+- (2) The DPKG contains dynamically updated spatio-temporal facts that reflect changes in the real world. Through knowledge graph query, the query performance in disaster emergency scenarios can be improved, and dynamic data updates can automatically drive prediction;
+- (3) It provides efficient data storage and management methods for practitioners in the fields of remote sensing and geo-information, which helps to improve the efficiency
+
+Remote Sens. 2022, 14, 1214 20 of 21
+
+of spatio-temporal data query. It reduces the manual effort by using reasoning of the knowledge graph.
+
+The DPKG proposed in this paper aims at the integration of remote sensing information, geographic information, and the correlation between dynamic data and static knowledge in complex disaster environments. It lays the foundation for future in-depth research, particularly for knowledge extraction and knowledge discovery for disaster prediction. In future work, indexing and query performance need to be further verified in the case of more types and larger amounts of data. The depth of knowledge reasoning needs to be further improved. In future research, we will enhance the DPKG by leveraging graph neural networks and deep learning models, so as to integrate a large number of spatio-temporal facts, and realize disaster emergency decision-making, historical data verification, and new model construction; it needs to build a knowledge service system for natural disaster monitoring and prediction as well as the emergency decision-making. In addition, this article basically focuses on verifying the availability and usability of our model through case studies. We will compare with more traditional prediction methods in the future
+
+**Author Contributions:** Conceptualization, X.G. and Y.Y.; methodology, X.G. and W.L.; validation, X.G., W.L. and W.Z.; resources, W.L.; data curation, X.G. and W.L.; writing—original draft preparation, X.G. and J.C.; writing—review and editing, X.G., Y.Y., L.P. and Z.H.; funding acquisition, L.P. All authors have read and agreed to the published version of the manuscript.
+
+**Funding:** This work was supported by the Beijing Municipal Science and Technology Project (Z191100001419002), and Ningxia Key R&D Program (2020BFG02013).
+
+**Institutional Review Board Statement:** Not applicable.
+
+Informed Consent Statement: Not applicable.
+
+Data Availability Statement: Data sharing not applicable.
+
+Conflicts of Interest: The authors declare no conflict of interest.
+
+# References
+

chatbot/rag/data/('Early Prediction System for Glacier Lake Outburst Floods (using ARIMA GRU LSTM)', '.pdf')_extraction.md

@@ -0,0 +1,69 @@
+
+
+Abstract: "Glacial Lake Outburst Floods (GLOFs)" are sudden, high-magnitude floods resulting from the breach of natural dams containing glacial lakes. With climate change accelerating glacial lake formation, GLOFs present growing threats to downstream communities and infrastructure. This study introduces an AIpowered chatbot system designed for real-time GLOF risk prediction and assessment. By leveraging predictive models such as ARIMA, GRU, and LSTM, the system analyses user-provided environmental data, including temperature and water levels, to generate actionable insights. A simple and intuitive chatbot interface allows users to input key parameters, which are processed using predefined thresholds to categorize risk levels as high, moderate, or low. The system enhances disaster preparedness and management through accessible and timely decision support. Future work aims to integrate IoT-enabled realtime data streams to improve prediction accuracy further, making this solution a scalable and effective tool for mitigating GLOF risks in vulnerable regions.
+
+Keywords: GLOFs, AI-Powered chat bot, Predictive models.
+
+## 1. Introduction
+
+In recent years, advancements in artificial intelligence (AI) and predictive analytics have significantly transformed disaster risk management. Among these, the prediction of Glacial Lake Outburst Floods (GLOFs) has garnered attention due to their catastrophic impact on downstream communities and infrastructure. With climate change accelerating glacial lake formation, GLOF prediction has become increasingly vital. However, the task is challenging due to the dynamic nature of glacial systems, limited real-time data, and the complex interplay of environmental factors.
+
+To address these challenges, AI-powered chatbot systems have emerged as an innovative solution. By leveraging pretrained predictive models like ARIMA, GRU, and LSTM, these systems analyze environmental data such as temperature and water levels to predict GLOF risks. This approach reduces computational complexity, ensures real-time responsiveness, and provides accurate risk assessments. Additionally, the chatbot interface enhances accessibility, allowing non-experts to receive actionable insights for disaster preparedness and management.
+
+## 2. Literature Review
+
+### A. Traditional Methods
+
+Traditional methods for GLOF prediction primarily relied on field surveys, remote sensing, and hydrological modelling. While effective for historical assessments, these methods lacked real-time responsiveness and required significant expertise and resources. Additionally, they were unable to account for rapidly changing environmental conditions, limiting their predictive accuracy in dynamic glacial systems.
+
+### B. Environmental Data Analysis
+
+Environmental data analysis plays a crucial role in GLOF prediction. Early approaches focused on analyzing individual factors, such as glacial lake volumes or temperature trends, but failed to integrate multiple parameters effectively. Recent advancements in machine learning models like ARIMA and GRU have enabled the analysis of complex, multivariate datasets. These models can identify temporal patterns and interactions between variables, significantly improving prediction accuracy and reliability.
+
+### C. AI and Chatbot Systems
+
+The integration of AI-powered chatbots has revolutionized the accessibility and usability of early-warning systems. Predictive models, including LSTM, have been incorporated to analyze real-time data and provide accurate risk categorizations. Chatbots offer a user-friendly interface, allowing stakeholders to input environmental parameters and receive actionable insights instantly. This eliminates the need for specialized knowledge and reduces reliance on computationally expensive systems. Future advancements include integrating IoT sensors for continuous data collection and enhancing prediction accuracy through real-time feedback loops.
+
+## 3. Methodology Data Collection and Preprocessing
+
+The dataset utilized for this project includes historical and real-time environmental data sourced from public repositories, including satellite imagery, hydrological datasets, and meteorological records. Key parameters such as temperature, precipitation, and water levels were analyzed to ensure
+
+<sup>\*</sup>Corresponding author: harrini.d.s@gmail.com
+
+accuracy and consistency. Preprocessing steps involved normalizing data to handle varying units and filling missing values using interpolation techniques. Data augmentation, such as synthetic generation of extreme scenarios, was employed to improve the robustness of predictive models. The dataset was split into training, validation, and testing sets in a 70:20:10 ratio for optimal model evaluation.
+
+### A. Model Development
+
+Predictive models, including ARIMA, GRU, and LSTM, were employed for analyzing temporal and multivariate data. These models were selected for their ability to capture timeseries patterns and long-term dependencies. The final predictive layers were tailored to classify GLOF risk into high, moderate, or low categories. A threshold-based logic system was integrated to interpret model outputs into actionable risk levels.
+
+### B. Training and Validation
+
+The models were trained using the Adam optimizer and mean squared error (MSE) loss function. Early stopping was implemented to avoid overfitting, and learning rate scheduling was applied to enhance convergence. The training process was monitored using validation metrics, including root mean square error (RMSE) and accuracy. Augmented scenarios in the training data ensured better generalization for unseen conditions.
+
+### C. Testing and Evaluation
+
+The trained models were tested on a holdout test set comprising unseen environmental data. Evaluation metrics such as RMSE, accuracy, precision, and recall were computed to assess the system's predictive performance. Misclassified or ambiguous cases were analyzed to refine model thresholds and improve overall reliability.
+
+### D. Deployment
+
+The predictive system was deployed as an AI-powered chatbot application. The chatbot interface, built using Python's Flask or FastAPI frameworks, allows users to input environmental data such as temperature and water levels. Based on these inputs, the chatbot provides real-time GLOF risk categorization and actionable recommendations. The frontend offers an intuitive chat container to ensure accessibility for users without technical expertise.
+
+Image /page/1/Figure/12 description: A flowchart, labeled "Fig. 1. Flow diagram," illustrating a machine learning workflow. The process flows downwards through several stages, each represented in a separate horizontal section on a light blue background. The stages are: 1. Data Collection & Preprocessing, 2. Model Development, 3. Training & Validation, 4. Testing & Evaluation, and 5. Deployment. The final stage, Deployment, branches out to two components: "User Input (Temperature, Water Level)" and "Backend Prediction (Flood Risk)."
+
+Fig. 1. Flow diagram
+
+## 4. Results
+
+The early prediction system for Glacial Lake Outburst Floods (GLOFs), utilizing advanced machine learning models like LSTM and GRU, demonstrated significant predictive accuracy, achieving an RMSE of approximately 0.85 and an accuracy of 88% on the test dataset. The models effectively captured temporal patterns in environmental data, including temperature, precipitation, and water levels, generalizing well to unseen scenarios. Training metrics, including validation loss and accuracy, showed consistent improvement, indicating robust model learning.
+
+## 5. Conclusion
+
+In this project, we successfully developed an early prediction system for Glacial Lake Outburst Floods (GLOFs) using advanced machine learning models integrated into a chatbot interface. The predictive models, including LSTM and GRU, were trained on environmental datasets comprising parameters like temperature, precipitation, and water levels. By leveraging pre-trained models and fine-tuning them for GLOF prediction, the system achieved high accuracy and reduced computational overhead compared to traditional approaches.
+
+The chatbot provided a user-friendly platform for real-time risk assessment, allowing non-expert users to input data and receive actionable insights. While the system demonstrated strong performance in categorizing risk levels, slight inaccuracies were observed in cases with overlapping environmental thresholds, reflecting the inherent complexity of GLOF prediction. Future enhancements, such as incorporating IoT-based real-time data, are planned to further improve prediction accuracy and system reliability.
+
+## Acknowledgement
+
+We extend our heartfelt gratitude to our Guide Ms. Narmatha M, for her invaluable guidance and continuous support throughout this project. We also thank the Department of Artificial Intelligence and Machine Learning faculty members and staff at Sri Shakthi Institute of Engineering and Technology for providing essential resources and facilities. Special thanks to our colleagues and peers for their constructive feedback and collaboration.
+
+## References

chatbot/rag/data/('Early_Warning_System_for_Glacial_Lake_Outburst_Using_Multi_Source_Data', '.pdf')_extraction.md

@@ -0,0 +1,276 @@
+
+
+Abstract—Glacial Lake Outburst Floods (GLOFs) pose serious risks to ecosystems, infrastructure, and communities downstream. By combining several data sources, such as satellite imaging, remote sensing, hydrometeorological data, and real-time sensor networks, this study presents a unique early warning system (EWS) intended to mitigate these hazards. The EWS can monitor and forecast GLOF incidents more accurately and quickly by using machine learning algorithms and predictive models, which will improve preparedness and mitigation activities in areas that are at risk. The system continuously monitors glacial lakes using satellite imaging and remote sensing, gathering vital information on the dynamics of the lakes and the surrounding ecosystem. This data is paired with hydrometeorological data gathered from real-time sensor networks and ground-based stations, including temperature variations and precipitation levels.
+
+Index Terms—Glacial Lake Outburst, Early Warning System, climate change, Multi source data
+
+# I. INTRODUCTION
+
+Global glaciers are rapidly retreating as a result of the accelerating speed of climate change. As a result of this process, glacial lakes have formed and grown, frequently being blocked by loosely consolidated moraines or unstable ice. Glacial lake outburst floods, or GLOFs, are caused by abrupt breaches of these natural dams. GLOFs can release enormous amounts of water in a matter of minutes and are distinguished
+
+by their great size and abrupt onset. These floods can have disastrous aftereffects, resulting in a large number of fatalities, extensive property damage, and extensive harm to downstream communities' infrastructure and natural environments.
+
+The current techniques for tracking glacial lakes and fore-casting GLOFs are frequently constrained by their need for recurring field surveys. The world's glaciers are melting at previously unheard-of speeds as a result of climate change. The creation and growth of glacial lakes as a result of this melting presents a serious risk of glacial lake outburst floods (GLOFs). When the water in a glacial lake is released, usually as a result of the lake's natural dam collapsing or other precipitating events like landslides or ice avalanches, GLOFs—sudden, high-magnitude floods—occur. These floods can have disastrous effects, including downstream agricultural and natural landscape degradation, infrastructure damage, and fatalities.
+
+Due to traditional monitoring methods not being able to provide real-time data or adequate warning signals, anticipating GLOFs in glacial lakes has proven to be incredibly difficult.
+
+The retreat and melt of glaciers across the globe is at its peak, and all this is mainly due to an increase in global temperature owing to climate change. Unstable glacial lakes have started to appear in various mountainous regions due to this change. These lakes are often obstructed by loose rocks
+
+or ice moraines, making them incredibly prone to surprising breaches. Such breaches can lead to glacial lake outburst floods (GLOFs) which unleash tremendous amounts of water and debris in a short period of time. Such phenomena put a lot of downstream communities at risk along with causing a lot of destruction to several ecosystems, infrastructural buildings, and even agricultural lands. These GLOFs are expected to occur more frequently than they already do, putting the atrisk populations in danger.
+
+# II. BACKGROUND
+
+Glacial lake outburst floods (GLOFs) are characterized by the rapid and sudden release of moisture from glacial lakes, formed as glaciers melt and retreat. Usually, ice or moraines, which are naturally unstable and prone to fracturing, act as barriers to keep these lakes back. including breaches can be caused by a wide range of events such as landslides, seismic activity, extreme rainfall, or the gradual warming of ice dams brought on by warming global temperatures. Particularly in high-altitude areas like the Himalayas, the Andes, and the Alps, the number of glacial lakes has increased as a result of the glaciers' accelerated retreat due to ongoing climate change.
+
+GLOF prediction and prevention has always been a major challenge due to the flexible and changing complexity of glacial lake systems at a multi-level throughout history. While infrequent field surveys and remote sensing are some of the lake monitoring techniques, they are not capable of collecting data in real time and capturing the fast changes in the environment rendering them ineffective. The alarming systems set up to signal GLOF has been mostly late or ineffective which catastrophically impacts the communities located downstream. The advancements in machine learning, remote sensing, sensor networks and satellite technology have provided an opportunity to develop more efficient real time GLOF monitoring systems. By combining data from different sources such as satellite images, hydrometeorological data and groundbased sensors, an integrated system enables efficient monitoring of glacial lakes. This approach is enhanced even further through the use of machine learning algorithms which are able to analyze the data and detect early indicators, making the alarms timelier.
+
+# III. LITERATURE SURVEY
+
+## A. Glacial Lake Formation and GLOF Events
+
+The impact and frequency of Glacial Lake Outburst Floods (GLOFs) has been a concerning factor on a global scale especially in the higher altitudinal belts which are witnessing accelerated global warming which in turn is causing drastic changes in glaciers. Places like the Himalayas, Andes, Alps and Rockies are particularly vulnerable to such events. Various researchers have studied the processes involved such as J.S. Kargel, U.K. Haritashya [1] highlighted the reasons behind the formation of glacial lakes and the various reasons that act as triggers for GLOFs. Their investigations
+
+reveal the glacier boundary destabilization factors such water in the form of ice avalanches, precipitation, earthquakes and other volcanic activities. Such factors would create a certain level of chaos which would lead to terrifying bursts endangering the regions prone downstream.
+
+## B. Traditional Monitoring Methods and Limitations
+
+Over the last few years, satellite monitoring of glacial lakes has increased due to the ability to produce higher resolution data and cover wider areas compared to traditional groundbased methods. For instance, X. Wang, X. Guo, C. Yang [2] examined the accuracy of monitoring glacier lakes with satellite imagery. One major disadvantage of satellite monitoring is that it is not as frequent as the events it is trying to capture, thus it may not capture the rapid and highly dynamic changes in the lake glacier's conditions at a time, despite the advantages that come with it. These weaknesses show that there is a requirement for an ongoing, more holistic monitoring system that can offer real time information and early warnings. A more sophisticated early warning system would significantly enhance the identification and forecasting of GLOF events with the help of advancements in satellite technology, remote sensing, and sensor networks, as well as the analytical capabilities of machine learning algorithms.
+
+## C. Advances in Remote sensing for Glacial Monitoring
+
+Better and continuous monitoring of glacier lakes is now possible due to recent improvements in satellite technology and remote sensing capabilities. In their study, for instance, G. Zhang, T. Yao, H. Xie [3] employed high- resolution satellite data from MODIS, Sentinel-2, and Landsat to inform their work. Glacial lakes' growth and volume changes have been effectively monitored and charted in satellite data studies, proving the effectiveness of remote sensing technology in pattern and trend identification. This paper highlights how data from these cutting-edge satellite images can be used by Glacial Lake Outburst Flood (GLOF) Early Warning Systems (EWS). Through the collection of wide-ranging, high-quality lake images, these platforms deliver valuable information on the shifting processes and ecological status of glacier lakes.
+
+## D. Multisource Data Integration for Real-time Monitoring
+
+Ground-based sensors provide timely measurements of physical attributes such as temperature, precipitation and seis mic activity, while high resolution satellite imagery allows for continuous monitoring of glacial lake dynamics [4]. This multifaceted approach, in addition to enhancing the capacity to detect early signs of potential GLOF events improves the overall reliability of forecasts. However, these benefits do not come without a number of logistical and technical challenges. The two major obstacles in this regard are merging numerous data sources, and it demands clever data processing
+
+# knowledge to ensure seamless workflow [5].
+
+## E. Case studies and Early warning system Implementations
+
+Several case studies have successfully implemented earlywarning systems for Glacial Lake Outburst Floods (GLOFs), which have shown the potential benefits of GLOFs early warning systems. The International Centre for Integrated Mountain Development (ICIMOD) has piloted early warning systems that leverage satellite data, remote sensing and community- based monitoring systems to issue timely alarms, in the Hi- malayas. This integrated approach not only improves the accurate prediction but also ensures active participation of the community in monitoring the system, which makes people more prepared and provides a sense of ownership. Work like this has been done in the Fan, X., Xu, Q., van Westen, C.J., Huang. (2008) state.
+
+# IV. METHODOLOGY
+
+## A. Data Processing
+
+Make the satellite images into shapes that help objects and Track changes. This may include the use of techniques such as twitching images specific to highlight Hydrometeorological data: Mix data from weather stations and Censorship in the usual time slot to create a clear time Shows patterns over time. You can stick together and sort data from ground -based sensor to create a complete overview of the environmental conditions, Similar to the hydrometeorological data.
+
+## B. Feature Extraction and Analysis
+
+### Ground -based full sensor data recovery:
+
+To detect the symptoms of potential volatility in the lake, the sensors gathered on the ground to extract the environment and physical symptoms from the records. Earthquake hobbies, temperature races and downs are important variables required to maintain a look. Initiative warnings of instability inside Glacial Lake -Gadget may be located using journal aggregation and evaluation methods, including non -conformity detection and time collection analysis.
+
+### Image processing:
+
+Satellite analysis for PC images for important features for the use of refined photo processing techniques. This involves monitoring the shift within the boundaries of the lake and detecting the variation inside the floor space of the lake, and comparing ice and ice conditions on site. To draw attention to significant differences and patterns in SNAP shots, techniques can be used with image partitions, part detection and alternative identity algorithms.
+
+## C. Machine Learning Model Development
+
+### Exercise and testing:
+
+To create a reliable model, data must be divided into training and test kits. The model can learn and recognize high -risk signals by practicing historical GLOF events. The model is fed data during the training phase so that it can identify and absorb patterns related to previous GLOF events. To confirm the accuracy of the model after training, unseen data must be used for testing. The model's ability to extract the prophecies of the prophecies, the ability to extract for real world conditions is guaranteed by this stage. To maximize accuracy and performance, the hyperparameter may need to be replaced in this phase.
+
+### Future modeling:
+
+The model can evaluate the current data in real time so that it is trained and verified to find a gloff risk pattern. Unusual temperature spikes, sudden changes in the surface of the lake, or closer to seismic activity to the lake, can be all important signs. The sustained monitoring of these parameters is performed by the model, which immediately releases the alarm, when possible, dangers are identified.
+
+## D. Early Warning System Implementations
+
+Real -time warning: Install an exact intersection points for each risk category, when the alarm is immediately triggered to warn of the importance of a potential gloff cross the significant values. For high -risk conditions, the first warning system (EWS) creates automatically alerts, which later expanded the right authorities and population nearby. This guarantees that interested parties are properly informed so that they can take the right protection measures.
+
+Data Virtue and Dashboard: Create an interactive dashboard showing model conditions and real -time data. For each lake during observation, this gives the dashboard a variety of visualization, including time series charts, maps and risk level indicators. Straits can quickly evaluate GLOFF risk for these visual aids and evaluate them, who inform and inform decisions to reduce any effect.
+
+Risk assessment: Determine the risk level of each monitored lake using the production from the approximate model. The model uses a mixture of historical trends and uses variable power variables to classify lakes in low, medium, or high-risk groups. The greatest risk of lakes that pose the greatest risk, it helps to determine priorities for classification resources and monitoring of activities.
+
+## E. Testing and Calibration
+
+Field Test: Conduct field experiments of the initial warning system (EWS) by modeling multiple risk scenarios for glacier flooding (GLOF). These tests are required to confirm that
+
+the system can predict GLoft events and provide timely alerts. Real-time simulation exercises can assess the accuracy and reaction of the system; simulation exercises can include controlled liberation, false stress tests at lake boundaries, or artificially elevated water levels.
+
+Adjustment and calibration of the model: model: Change the parameters of the model to improve its accuracy and dependence in light of this field test results. Some important steps, including calibration, are
+
+Refining sensor box: The sensor is sufficient to identify gloves without installing the false alarm to adjust the threshold. Change the Parameters to Machine Learning Models: Adjust the parameters of machine learning algorithms in EWS. This can change the complexity of the model, learning speed, or decision limit to increase the performance of prediction.
+
+Changing data preparation methods: In order to guarantee the best possible processing of input, you can visit the data to pre -process processes again. To better prepare the data, they can adjust the generalization area, upgrade the noise filtering strategies, or refine the data change processes.
+
+## F. Continuous Monitoring and Improvement
+
+Feedback loop: Establish a method of ongoing monitoring to evaluate the first warning method (EWS) efficiency. To determine the user inputs to get the improvement needed. The accuracy and effect of the system is maintained by adjusting the future model in response to fresh data and user response. EWS can be adjusted under changing conditions and may be better over time due to this relapse method.
+
+System upgrade: To use the latest development in technology, update regular data sources, machine learning algorithms and alarm systems. EWS skills can be improved by integrating new techniques and approaches when they become available. To guarantee reliable function, inspect and maintain regular sensors and data collection equipment. Reorganization of the sensor, updates software and confirms that each part is involved in this process is all included in this process.
+
+EWS can be effective and adapted to the changing conditions of the glacial lake isos system by implementing a strong feedback loop and system improvement.
+
+# V. CONSTRUCTION OF CIRENMACO EWS
+
+To take advantage of the latest technology developments, update data sources, machine learning algorithms and
+
+alarm systems regularly. EWS capabilities can be improved by integrating new technologies and approaches as they are available. To ensure reliable operation, inspect and maintain data collection sensors and equipment regularly. Recalibration sensors, updating the software and confirming that all parties are operating as intended are all included in this process.
+
+EWS can remain efficient and adaptable to the conditions of change of lake's glacial ecosystems, implementing a strong feedback cycle and system improvements. This strategy helps to reduce hazards related to Glacial Lake Outburst Floods (GLOFs) by guaranteeing that the system will continue to issue precise and timely warnings. After an assessment of the vulnerability of the lake, the EWS design phase focuses on building the required infrastructure. This involves setting up a communication network around high-risk regions and strategically placing sensors. Among the key sensors used are: water level sensors: These provide vital information on growing or declining water volumes by tracking variations in the lake's water levels. seismic sensors: They identify soil movement that can indicate potential triggers such as earthquakes or landslides.
+
+Meteorological stations: By monitoring changes in temperature and rainfall, these stations help determine environmental factors that can cause gloff.
+
+Flow Sensor: These downstream sensors keep an eye on the discharge of unusual water that may indicate a adjacent explosion. To ensure real-time data transfer in a central control center, all these sensors are associated with a reliable communication network, which often uses satellite technologies or GSM. Continuous monitoring and quick notifications are possible by this configuration, which is necessary to reduce the risks associated with the glof and improve the security of downstream communities. Data processing and integrations come after the infrastructure setup. Important information about lake levels, seismic activity and weather is shown on a real -time dashboard that tracks the sensor data in a centralized database. While the risk thresholds are set for critical matrix, machine learning algorithms examine the data to find the pattern to find the risk to explore that gloff events. If sensor data exceeds specified levels, EWS automatically triggers alerts and notifications. These signs enable emergency services, local communities, and relevant officers to prepare and react quickly. This method helps communities to be informed and flexible against potential flood hazards in the Cirenmaco region.
+
+# VI. IMPLEMENTATION
+
+## A. Software Implementation
+
+The main goal of real -time monitoring, data processing, predictive modeling, and alert generation is the main goal of a reliable, scalable and user -friendly platform software
+
+Image /page/4/Figure/1 description: A figure containing a map labeled 'A' and an inset bar chart labeled 'B'. The map shows the GLOF (Glacial Lake Outburst Flood) susceptibility of a mountainous region. A legend in the bottom left defines the susceptibility levels with colored dots: Very low (white), Low (light green), Medium (yellow), High (blue), and Very high (red). These dots are plotted across the topographical map, which displays snow-covered peaks in light blue and lower elevations in brown and green. Several specific locations are labeled, such as M20, M24, M47, and M51. The map includes a 10 km scale bar and coordinate markings. The inset bar chart 'B' shows the count of lakes versus their size-interval in square kilometers. The x-axis shows size intervals from 0.05-0.1 to >0.5 km². The y-axis shows the count, up to 30. The bars are stacked and colored according to the susceptibility legend, indicating that the most numerous lakes are in the smallest size category (0.05-0.1 km²) and have low susceptibility, while a large portion of the biggest lakes (>0.5 km²) have very high susceptibility.
+
+Fig. 1. —(A) spatial distribution of 64 glacial lakes assessed with their GLOF susceptibility classes across the Mahalangur Himalaya. (B) The bar graph shows the GLOF susceptibility of glacial lakes in various size interval [13].
+
+project that applies CIRENMACO Early Warning System (EWS) for Glascable Lake Outburst Flood (Glofs). Use a trusted cloud-based solution, such as Google Cloud platform or AWS (Amazon web services), to effectively manage the processing and storage of large-scale real-time data. For managing sensors data and historical records, cloud solution provides scalability, safety and reliable performance in Fig 1. Set the risk threshold for important matrix. The system should automatically sound alarm and inform the proper authorities, nearby communities and emergency services when the threshold is crossed. For automated alert delivery, use services such as AWS SNS (simple notification service) or comparable solution.
+
+## B. Frontend Development
+
+Design of user interface (UI): Provide a simple, easy-to- use interface that shows major data visualizations, such as historical trends, danger conditions and changes in the level of lake. Users should be able to interact with various charts and map-based visualizations, view real-time data and track the level of risk through user interfaces.
+
+Dashboard features:
+
+- Create a dashboard that allows users to see real -time sensor data classified by seismic activity and water level.
+- Get historical data to analyze trends.
+- Check the color-coded risk map of the item lake area.
+- When the thresholds cross, get information and alert.
+
+## C. Data Processing and Preprocessing
+
+Data normalization and cleaning: As the system comes into the system, put the data in procedures to validate, clean and standardize. This involves managing outlier or missing
+
+Image /page/4/Figure/13 description: The image displays two stacked bar charts, labeled A and B, analyzing different types of lakes based on their size. The legend identifies four types of lakes by color: green for Supraglacial lakes, orange for Proglacial lakes, purple for Unconnected glacier-fed lakes, and yellow for Non-glacier-fed lakes. The x-axis for both charts is labeled "Size-interval (km²)" and has six categories: 0.001-0.002, 0.002-0.01, 0.01-0.02, 0.02-0.1, 0.1-0.5, and >0.5.
+
+Chart A, on the left, shows the "Number" of lakes on its y-axis, which ranges from 0 to 140. The highest number of lakes, approximately 125, is found in the 0.002-0.01 km² size interval. The number of lakes generally decreases for larger size intervals, with the >0.5 km² category having fewer than 10 lakes.
+
+Chart B, on the right, shows the total "Area (km²)" of the lakes on its y-axis, which ranges from 0 to 8. In contrast to the number of lakes, the total area generally increases with the size interval. The largest total area, approximately 7.5 km², is contributed by lakes in the >0.5 km² size category, which are predominantly proglacial and unconnected glacier-fed lakes. Proglacial lakes contribute the most to the total area in the larger size intervals (0.1-0.5 and >0.5 km²).
+
+Fig. 2. The frequency and area of glacial lakes according to different size interval and types [14].
+
+values, ensuring that the measurement units are consistent, and convert data into a format that is compatible.
+
+Adapt the storage and recovery by Fig.2 structuring the database with effective sequencing to provide quick data recovery, which is necessary for real -time, mass dataset analysis.
+
+To install the necessary equipment in high -risk areas, the process begins with the initial fieldwork. During this phase, Cirenmaco Lake is surrounded with carefully downstream flow sensors, meteorological stations, seismic sensors and water level sensors. It asks for reliable equipment that can avoid challenging environmental conditions, as well as trained workers to guarantee ideal sensor placements and accuracy. Each sensor is connected to a reliable communication network and is set to collect data continuously. Typically, satellite or GSM technology is employed, providing a reliable link to send data to the Central Monitoring Center in real time.
+
+## D. Co-ordination and Integration of data
+
+A centralized data receives the sections of repository data that collect from the sensor. To maintain stability between sources, data is processed, cleaned and normalized here. After that, the dashboard has been configured for real -time monitoring, showing important data metrics and visualizations that help analysts track signs of lake stability, such as earthquake activity and variations in the water level. At this point, the system is calibrated to identify abnormalities that can point to potential glof events, and baseline data is collected to understand specific trends.
+
+## E. Using Machine Learning
+
+A comprehensive strategy that mixes machine learning for future stating analysis with real -time environmental data collection is required to apply Cirenmaco Early Warning System (EWS) for Glassel Lake Outbst Floods (Glofs) as a software project. Designing system architecture is the first stage, and data processing and storage are controlled by cloud-based services such as AWS or Google Cloud. A streaming pipeline (such as Apache Kafka) is used to manage the import of real -time data. This pipeline collects data from the sensor that monitors weather, seismic activity and water
+
+level. After that, this data is pre -evacted using procedures to clean, normally and format it to guarantee accuracy. Other time-series features are developed to catch temporary patterns, and major features-as temperature fluctuations and water level trends-are chosen based on their importance for glof risk.
+
+Algorithm Selection: Classify risk levels using past GLOF data by using supervised learning techniques like Random Forest or Support Vector Machine (SVM). Potential trends in water levels or other indicators can be predicted using timeseries analysis (e.g., LSTM neural networks).
+
+Training and Testing Models: Create and train a machine learning model to identify risk indicators using historical data, then assess the model's accuracy using a testing dataset. On the basis of incoming data, the trained model may then forecast future GLOF hazards.
+
+Backend integration: Put the learned model into practice in the backend, where it may process sensor data in real time and produce a prediction or danger level.
+
+## F. Alert and Notification System
+
+The Cirenmaco Early Warning System's (EWS) warning and notification system is made to give stakeholders quick and accurate alerts so they can respond to the hazards of glacial lake outburst floods (GLOFs). When crucial thresholds are crossed, this system sends out alarms. It does this by continuously monitoring environmental data, including temperature, water levels, and seismic activity. In order to provide a progressive reaction that matches the intensity of the detected risk, the alert system is set up with risk levels—typically low, moderate, high, and emergency—that correlate to various alert circumstances.
+
+The system automatically sends out messages across a variety of channels to important stakeholders when sensor data shows that risk thresholds are reached. Local authorities, emergency responders, and community leaders receive notifications through push notifications on mobile apps, email, and SMS. The solution guarantees prompt and effective notification distribution by utilizing APIs such as Firebase Cloud Messaging for mobile app alerts, SendGrid for emails, and Twilio for SMS. In the event of a severe emergency, the warning system also enables automatic voice calls to high-priority receivers, guaranteeing that vital information reaches them right away.
+
+# VII. SYSTEM ARCHITECTURE
+
+The Early Warning System (EWS) for Glacial Lake Outburst Floods (GLOFs) combines multi-channel alerting mechanisms, machine learning-based risk prediction, and real-time data collection. By utilizing cloud computing, IoT devices, and powerful analytics, the design guarantees dependable and scalable operations. The architecture is
+
+broken down in detail below:
+
+## A. Data Collection Layer:
+
+IoT devices and sensors:
+
+- Real-time data is gathered by sensors positioned throughout the glacial lake, including
+- Water levels and flow rates are examples of hydrological data
+- Weather information: humidity, precipitation, and temperature.
+- Seismic Information: Possible landslide activity and ground vibrations.
+- Wireless communication technologies, such as LoRa, Zigbee, or GSM, are used to transfer data.
+- Remote sensing and satellite data: lake surface area, changes over time, and snow/ice melt patterns are all provided by satellite imagery (e.g., from Landsat, Sentinel).
+
+## B. Data Transmission and Integration Layer:
+
+### Edge Technology:
+
+Sensor data preprocessing and short-term storage are managed by IoT gateways at the data collection location. This guarantees that there is less data loss when being transmitted.
+
+### Cloud-Based Communication:
+
+Protocols like HTTPS and MQTT are used to safely transfer data to a central cloud-based system. Redundancy measures are put in place to guarantee dependability and uninterrupted functioning. This configuration increases the early warning system's overall efficacy by enabling safe and effective real-time data transmission.
+
+## C. Data Storage Layer:
+
+To ensure minimal loss during transmission, edge devices—like IoT gateways at the data collection site—preprocess and temporarily store sensor data. Protocols like HTTPS and MQTT are used to securely transfer the data to a central cloud-based system, with redundancy mechanisms in place for dependability. Real-time sensor data and historical datasets for analysis and machine learning model training are stored on centralized storage platforms such as AWS, Azure, or Google Cloud. Structured data is managed by SQL databases (PostgreSQL), whereas unstructured or semi-structured data is handled by NoSQL databases (MongoDB). In order to prevent data loss and guarantee data durability and availability for long-term storage, periodic backups are kept in cold storage systems such as Amazon S3 Glacier. This configuration facilitates secure, scalable, and effective data handling.
+
+## D. Data Processing and Analytical Layer:
+
+The preprocessing engine handles missing values, eliminates outliers, and arranges data for analysis by cleaning, validating, and normalizing raw sensor data to guarantee accuracy and consistency. In order to identify anomalous patterns in water levels or seismic activity and forecast the probability of a Glacial Lake Outburst Flood (GLOF) catastrophe, machine learning models such as LSTM and Random Forest examine both historical and current data. The risk assessment module uses time-series analysis and dynamic updates to classify risk levels—low, medium, high, and critical—based on machine learning outputs and predetermined thresholds. Vulnerable communities' safety and readiness are improved by this configuration, which guarantees effective and precise detection and prediction of GLOF events, facilitating prompt risk assessment and response.
+
+## E. User Interface and Visualization Layer:
+
+The dashboard interface provides real-time data visualizations to stakeholders, including academics and authorities. This comprises geographic maps showing the locations of lakes and the corresponding danger ratings, as well as trends from sensor data, machine learning forecasts, and alarms. Easy access is ensured by the interface's compatibility with mobile applications and online browsers.
+
+An end-user-friendly interface for seeing alerts and recommendations is offered by the mobile app and notifications. For important alerts, users receive push notifications, guaranteeing prompt information and action. With this configuration, end users and stakeholders are always informed and able to take prompt action to reduce risks.
+
+## F. Alert and Notification Layer:
+
+Notifications are sent by the automatic alarm system when risk thresholds are surpassed. It makes use of many notification channels, such as push notifications (using Firebase Cloud Messaging), emails (using SendGrid or AWS SES), and SMS (using Twilio or comparable APIs). Alerts give information about the location, level of risk, and suggested courses of action. The system initiates extra actions, such as automated voice calls or direct integration with disaster management systems, in response to high-priority alarms, including impending GLOFs. This all-inclusive alarm system guarantees prompt and efficient communication to the appropriate authorities and local communities, improving their capacity to react quickly and reduce hazards.
+
+## G. Monitoring and Maintenance Layer:
+
+The system tracks health, uptime, and error rates using cloud monitoring technologies such as Google Cloud Monitoring and AWS CloudWatch. To maintain openness and facilitate postevent study, it records all information, forecasts, and warnings. There is a feedback loop in place, whereby new data is frequently added to machine learning models to increase their accuracy. Additionally, user input is incorporated to improve system usability and alert thresholds. In order to successfully enable real-time monitoring, data processing, predictive modeling, and warning generation for glacial lake outburst floods (GLOFs), this all-encompassing methodology guarantees the system's dependability, transparency, and continual improvement.
+
+## H. Key Features of the Architecture:
+
+As more lakes and sensors are added, the system's scalability is built to manage growing data volumes, guaranteeing futureproofing as the project expands. Redundant systems that provide minimal downtime and continuous operation help to preserve their resilience. Access control, data encryption (SSL/TLS), and frequent security audits are used to safeguard private data from breaches and unwanted access. Because of its interoperability, the system may easily interact with other systems, such as disaster response platforms and GIS tools, to offer complete risk management. Glacial lake outburst flood (GLOF) risk monitoring and mitigation are successfully supported by the system's strong, effective, and flexible modular design in Fig 3.
+
+Image /page/6/Figure/12 description: A flowchart illustrating the methodology for assessing Glacial Lake Outburst Flood (GLOF) susceptibility. The process begins with 'Satellite imagery', which is used for 'Digitization' to create a 'Glacial Lake Inventory'. A decision diamond then checks if the 'Lake area' is greater than or equal to 0.05 km². If 'No', the lake is 'Excluded'. If 'Yes', it is 'Included for further analysis'. This analysis involves 'Factors computation & setting criteria (Index value)', which uses 'Lake Glacier Watershed DEM' as an input. Parallel to this, factors are identified, selected, and defined (F1, F2, F3, F4, F5, F6). Another input, 'AHP (determination of factor weights)', provides factor weights. The next step is the 'Multiplication of index value with factor weight', followed by the 'Sum of Scores'. This sum leads to the final 'GLOF Susceptibility' assessment. The susceptibility is categorized into five levels, each with a corresponding score range and color: 'Very Low (0.25-0.40)' in green, 'Low (0.40-0.55)' in light green, 'Medium (0.55-0.70)' in yellow, 'High (0.70-0.85)' in blue, and 'Very High (≥0.85)' in red.
+
+Fig. 3. Flow chart for GLOF [15]
+
+# VIII. EXPLANATION OF FLOW CHART
+
+Satellite photos: The first step in the procedure is gathering high-resolution satellite photos of glacier lakes and the areas around them.
+
+Digitization: A Glacial Lake Inventory is produced by digitizing these satellite photos. In this process, the visual data is transformed into an analyze-able digital format.
+
+Glacial Lake Inventory: Using the digitized data, a list of all the glacial lakes found in the satellite photos is created.
+
+Lake Area 0.05 km 2: The surface area of the lakes is used to filter them out of the inventory. For additional analysis, only lakes having an area of at least 0.05 km<sup>2</sup> are taken into account.
+
+Factor Identification: Choose pertinent geological and environmental elements.
+
+Factor Computation: To calculate factor values and establish criteria, use data from Digital Elevation Models (DEMs).
+
+AHP Analysis: Apply the Analytic Hierarchy Process (AHP) to ascertain factor weights. Index values are multiplied by weights, and scores are added up.
+
+Sort lakes according to their GLOF susceptibility levels: very low, moderate, medium, high, or very high.
+
+# IX. DISCUSSION
+
+The deployment of an Early Warning System (EWS) for Glacial Lake Outburst Floods (GLOFs) powered by machine learning greatly improves catastrophe risk management in glacial areas impacted by climate change. By combining automated alarms, predictive analytics, and real-time data collection, the system enhances outburst forecasts by analyzing hydrological and environmental data using models like LSTM networks. Scalability and dependability for huge sensor and satellite image collections are guaranteed by cloud-based storage. Generalizing models to different terrains and ensuring data quality in remote places are challenges. In spite of this, the EWS has enormous potential to help sustainable development, safeguard vulnerable people, and reduce GLOF risks. With further improvements, it might serve as a template for other dangerous regions around the world.
+
+# X. CONCLUSION
+
+Disaster risk management in glacial regions is revolutionized by the deployment of an Early Warning System (EWS) for Glacial Lake Outburst Floods (GLOFs) based on machine learning. The system offers a proactive way to lessen the effects of GLOF by combining automatic alerting systems, real-time data collection, and sophisticated predictive analytics. For a thorough risk assessment, it examines a variety of datasets, such as hydrological, meteorological, and seismic inputs, enabling authorities and communities to take preventative measures.
+
+The project's approach is flexible and scalable for different glacier situations, even in the face of obstacles like data availability, model generality, and infrastructure constraints.
+
+Multi-channel notifications and cloud computing improve reach and dependability. In addition to providing a foundation for upcoming developments in environmental monitoring and early warning systems, this initiative emphasizes the role that technology may play in mitigating climate threats.
+
+# XI. ACKNOWLEDGEMENT
+
+With great appreciation, we would like to thank everyone who helped us finish this project on the Cirenmaco Early Warning System (EWS) for glacial lake outburst floods (GLOFs). We are really appreciative of the direction and assistance given by our mentors and advisors, whose knowledge and experience were crucial throughout this project.
+
+We sincerely thank the research institutions and organizations that gave us access to vital data and necessary resources. We are particularly appreciative of the meteorological and environmental organizations' technical assistance, whose data allowed us to create a more dependable and strong early warning system. We also want to express our gratitude to our team members for their commitment, diligence, and cooperation.
+
+# REFERENCES
+

chatbot/rag/data/('Extreme Precipitation in a Warming Climate _ Climate Central', '.pdf')_extraction.md

@@ -0,0 +1,199 @@
+Climate Matters • May 1, 2024
+
+# Extreme Precipitation in a Warming Climate
+
+Image /page/0/Picture/4 description: An infographic on a dark blue background titled "WARMER AIR HOLDS MORE MOISTURE". A text box states, "1°F increase = 4% more water vapor". A graph shows "Temperature" on the x-axis and "Moisture" on the y-axis, with a blue line curving upwards to illustrate that as temperature rises, the air holds more moisture. To the right, a cloud is shown above a temperature bar that is blue on the left and transitions to yellow on the right, with "32°F" marked at the transition point. The cloud is shown producing "Heavier snow" on the colder side and "Heavier rain" on the warmer side. The Climate Central logo is in the bottom right corner.
+
+Click the downloadable graphic: Warmer air holds more moisture
+
+Image /page/0/Picture/6 description: A map of the United States titled 'HEAVIER DOWNPOURS' with the subtitle 'Change in precipitation on heaviest 1% of days'. The map is color-coded to show the percentage change in precipitation from 1958 to 2021. A color key on the right shows a scale from 0% to 40%, with colors ranging from light green to dark blue. The map displays the following regional percentage changes: Northeast +60%, Midwest +45%, Southeast +37%, a southern region +21%, a northern plains region +24%, Southwest +17%, Northwest +1%, Alaska +21%, Hawaii -19%, and the Caribbean -24%. The source is cited at the bottom as 'USGCRP, 2023: Fifth National Climate Assessment'. The Climate Central logo is in the bottom right corner.
+
+Click the downloadable graphic: Extreme Precipitation Change 1958 to 2021
+
+Image /page/0/Picture/8 description: A map of the contiguous United States titled "WARMER FUTURE, MORE EXTREME PRECIPITATION," which shows the "Projected change in precipitation on heaviest 1% of days." The map is color-coded by county to represent the "% change at 2°C (3.6°F) of global warming." A vertical color bar legend on the right indicates the percentage change, ranging from 0% (light beige) to 40% (dark green) in increments of 10%. Most of the country is shaded in green, indicating an increase in extreme precipitation, with the darkest shades concentrated in the Northeast, Great Lakes region, and parts of the Southeast. A small area in the Southwest, including parts of Texas and New Mexico, is shaded in light beige, indicating little to no change. The source is cited at the bottom as "USGCRP, 2023: Fifth National Climate Assessment," and the map is branded with the "CLIMATE CENTRAL" logo.
+
+Click the downloadable graphic: Future Extreme Precipitation 2024
+
+## KFY CONCEPTS
+
+Climate change is supercharging the water cycle, bringing heavier precipitation extremes — and related flood risks — across the U.S.
+
+As the climate has warmed from 1958 to 2021, the most extreme precipitation days have intensified in every major U.S. region, led by the Northeast (+60%) and Midwest (+45%).
+
+This hazardous intensification is expected to continue with future warming.
+
+With 2°C (3.6°F) of warming, 85% of 3,111 total U.S. counties are likely to experience a 10% or higher increase in precipitation falling on the heaviest 1% of days.
+
+High future levels of extreme precipitation intensification are concentrated in: Alaska, Hawaii, Tennessee, Alabama, Georgia, Mississippi, Maine, North Carolina, and Kentucky.
+
+People, ecosystems, and infrastructure in both wet and dry locations are facing the risks that come with short bursts of extreme rainfall.
+
+## Download KML map and XLSX
+
+**data file:** projected change in precipitation extremes under 2°C of warming for over 3,000 U.S. counties
+
+## Climate change is supercharging the water cycle
+
+Climate change is bringing heavier rainfall extremes and increased, inequitable flood risk to many parts of the U.S.
+
+For every 1°F of warming the air can hold an extra 4% of moisture, increasing the chances of heavier downpours that contribute to the risk of flash floods.
+
+Image /page/2/Picture/7 description: An infographic on a dark blue background titled "WARMER AIR HOLDS MORE MOISTURE". The infographic explains the relationship between air temperature and moisture content. A graph shows "Temperature" on the x-axis, increasing from left to right, and "Moisture" on the y-axis. A blue line curves upward, indicating that as temperature rises, the air's capacity to hold moisture increases. A text box states, "1°F increase = 4% more water vapor". An arrow points from the upward-curving line to a large grey cloud. Below the cloud, precipitation is shown. On the left, under colder conditions, snowflakes fall, labeled "Heavier snow". On the right, under warmer conditions, raindrops fall, labeled "Heavier rain". A temperature bar below the precipitation is blue on the left and transitions to yellow on the right, with a marker at "32°F" indicating the freezing point. The logo for Climate Central is in the bottom right corner.
+
+Click the downloadable graphic: Warmer air holds more moisture
+
+Heavy downpours bring more rain, faster — causing flash flooding and landslides that can displace families, drown crops, damage infrastructure, and expose people to hazardous debris, contaminants, and waterborne disease.
+
+The rapid onset of flash floods limits time to get people out of harm's way. In the U.S., racial minorities and those living in mobile homes are disproportionately exposed to flood risk, especially in the South and in rural areas.
+
+In the U.S., extreme daily rainfall has become more frequent since the 1980s. Hourly rainfall intensity has also increased since 1970 — by 13% on average across 150 U.S. locations analyzed by Climate Central.
+
+Risky rainfall extremes are the focus of analysis of past and projected future change in extreme rainfall across the U.S. featured in the Fifth National Climate Assessment.
+
+### Heaviest precipitation days have become heavier
+
+As the climate has warmed over recent decades, the most extreme precipitation days have become more intense across the U.S.
+
+As precipitation extremes intensify, the wettest days each year bring increasing flood hazards. And this intensification trend has been widespread.
+
+In the Northeast and Midwest, the amount of precipitation falling on the heaviest 1% of days has increased 60% and 45%, respectively, from 1958 to 2021 according to the Fifth National Climate Assessment.
+
+Image /page/4/Picture/2 description: A map of the United States titled "HEAVIER DOWNPOURS" shows the "Change in precipitation on heaviest 1% of days" from 1958 to 2021. The map is color-coded to represent the percentage change, with a legend on the right showing a scale from 0% (light yellow-green) to 40% (dark blue). The eastern half of the country is shaded in shades of blue, indicating a significant increase in precipitation, while the western half is shaded in greens, indicating a smaller increase. Specific percentage changes are noted for different regions: Northeast +60%, Midwest +45%, Southeast +37%, Northern Plains +24%, Southern Plains +21%, Southwest +17%, and Northwest +1%. Hawaii shows a decrease of -19%, and a shape representing Puerto Rico/U.S. Virgin Islands shows a decrease of -24%. The source is cited as "USGCRP, 2023: Fifth National Climate Assessment," and the Climate Central logo is in the bottom right corner.
+
+Click the downloadable graphic: Extreme Precipitation Change 1958 to 2021
+
+The heaviest rainfall events have become wetter across all other major regions of the continental U.S. from 1958 to 2021 as well, led by: the Southeast (+37%); the Northern Rockies and Plains (+24%); and the South (+21%).
+
+### Warmer future, heavier precipitation extremes
+
+With continued warming, precipitation extremes are likely to increase globally — even in regions with decreasing average precipitation, according to the latest reports from the Intergovernmental Panel on Climate Change.
+
+Most of the U.S. is projected to see increases in precipitation extremes with 2°C (3.6°F) of global warming, according to analysis in the Fifth National Climate Assessment. The planet has already warmed more than 1.1°C (2°F). Only by rapidly expanding and accelerating efforts to reduce heat-trapping emissions can we ensure that clobal warming levels remain below 2°C.
+
+Image /page/5/Picture/2 description: A choropleth map of the contiguous United States titled "WARMER FUTURE, MORE EXTREME PRECIPITATION," which shows the "Projected change in precipitation on heaviest 1% of days." The map is color-coded by county to indicate the percentage change at 2°C (3.6°F) of global warming. A legend on the right shows a color scale from light tan to dark green, representing a range from 0% to 40% change. The darkest green, indicating the highest increase in precipitation (30-40%), is concentrated in the Northeast, the Upper Midwest, and parts of the Pacific Northwest. Most of the rest of the country is shaded in lighter greens, indicating increases of 10-30%. The Southwest, particularly Arizona and parts of California and New Mexico, shows the lowest change, with colors in the light tan and light green range (0-10%). The source is cited as "USGCRP, 2023 Fifth National Climate Assessment," and the logo for "CLIMATE CENTRAL" is in the bottom right corner.
+
+Click the downloadable graphic: Future Extreme Precipitation 2024
+
+With 2°C (3.6°F) of warming, the majority (85% or 2,645) of the 3,111 total U.S. counties are likely to experience a 10% or higher increase in precipitation falling on the heaviest 1% of days.
+
+On average, U.S. counties are likely to experience a 17% increase in precipitation falling on the heaviest 1% of days.
+
+Counties likely to experience at least a 30% increase in extreme precipitation are concentrated in: Tennessee, Alabama, Georgia, Mississippi, Maine, North Carolina, and Kentucky.
+
+Alaska and Hawaii are likely to experience some of the highest levels of extreme precipitation intensification (43% and 30%, respectively) at 2°C (3.6°F) of global warming.
+
+A recent study suggests that, with high future levels of heat-trapping emissions (RCP8 5), U.S. flash floods could also intensify ��� especially in
+
+the Southwest — underscoring that the risks posed by short bursts of extreme rainfall affect both wet and dry regions.
+
+These rising rainfall intensity risks are unlikely to be equally shared. A recent study suggests that the burdens of an estimated 26% increase in overall U.S. flood risk by 2050 could disproportionately impact Black communities along the Atlantic and Gulf coasts.
+
+## Three key points: extreme precipitation in a warming climate
+
+1. For every 1°F of warming, the air can hold an extra 4% of moisture.
+
+This relationship between air temperature and water vapor pressure is governed by the laws of thermodynamics and represented in the Clausius-Clapeyron equation.
+
+The U.S. has already warmed 2.6°F since 1970 — meaning our atmosphere can already hold about 10% more moisture.
+
+### 2. More moisture in warmer air increases heavier downpours.
+
+Theory suggests that rainfall becomes more intense with warming. Both rain gauge data and modeling experiments support this theory.
+
+Globally, the frequency and intensity of heavy rainfall events have increased since the 1950s, largely due to humancaused climate change, according to the latest reports from the Intergovernmental Panel on Climate Change.
+
+In the U.S., extreme daily rainfall events have been on the rise since the 1980s.
+
+3. Intensifying rainfall due to climate change has cost the U.S. billions in inland flood damages over the last three decades.
+
+Inland flooding in the U.S. caused \$230 billion in damages from 1988 to 2021. Over one-third (37%) of those damages are attributed to precipitation changes due to climate warming.
+
+That's \$84 billion in past flood damage due to warming-induced rainfall intensification. And the most intense downpours have caused the largest damages.
+
+Learn more: Extreme Weather
+Toolkit: Heavy Rain and Flooding
+
+LOCAL STORY ANGLES
+
+### How vulnerable is your area to flooding?
+
+Factors like local topography, age of infrastructure, watershed health, and
+
+precipitation trends influence flood vulnerability. See current flood risks in your neighborhood with FloodFactor's ZIP code-level risk identifier. A 2021 report found that one-fourth of critical infrastructure is put at risk of failure by flooding. The American Society of Civil Engineers grades the integrity of infrastructures by state and category.
+
+## Who is most vulnerable to heavy rain-related hazards?
+
+Racial minorities and mobile homes are disproportionately exposed to flooding, especially in rural areas and in the southern U.S. The American Flood Coalition has resources on adaptation, equitable disaster recovery, and funding opportunities for flood-resilient communities.
+
+## What are the health and safety risks from flooding?
+
+Climate Central's report, *After the Storm: Health risks from damp, moldy homes*, summarizes the potential health risks from poor indoor air quality following heavy precipitation and flooding. Mold is just one of many potential hazards in a storm-damaged home. Other safety hazards include debris, structural damage, or downed electrical wires. The CDC provides tips to protect health and safety following a storm or flooding.
+
+## What can you do to protect against flooding and extreme precipitation?
+
+Firstly, know your own risk using the Federal Emergency Management Agency's flood maps. Individual homeowners can insure their homes, maintain rainwater systems, reduce impervious surfaces on their property, elevate important utilities and equipment, and take on low-cost indoor projects recommended by FEMA. Communities can invest in flood management systems, green infrastructure and watershed restoration.
+
+## CONTACT EXPERTS
+
+### **Greg Carbin**
+
+Branch Chief, Forecast Operations National Weather Service
+
+**Contact:** gregory.carbin@noaa.gov **Related expertise:** short-term forecasting of extreme rainfall
+
+## FIND EXPERTS
+
+Submit a request to SciLine from the American Association for the Advancement of Science or to the Climate Data Concierge from Columbia University. These free services rapidly connect journalists to relevant scientific experts.
+
+**Browse maps** of climate experts and services at regional NOAA, USDA,
+
+and Department of the Interior offices.
+
+**Explore databases** such as 500 Women Scientists, BIPOC Climate and Energy Justice PhDs, and Diverse Sources to find and amplify diverse expert voices.
+
+**Reach out** to your State Climate
+Office or the nearest Land-Grant
+University to connect with scientists,
+educators, and extension staff in your
+local area.
+
+## **METHODOLOGY**
+
+Data on past (1958–2021) and projected future (under 2°C of global warming) change in total precipitation (including rain and snow) falling on the heaviest 1% of days across the U.S. are from: USGCRP, 2023: *Fifth National Climate Assessment*. Crimmins, A.R., C.W. Avery, D.R. Easterling, K.E. Kunkel, B.C. Stewart, and T.K. Maycock, Eds. U.S. Global Change Research Program, Washington, DC, USA.
+
+https://doi.org/10.7930/NCA5.2023. Data were accessed via Figure 2.8a and Figure 2.12a metadata and the Fifth National Climate Assessment Interactive Atlas.
+
+(
+
+## Custom email signup
+
+Get updates, media alerts, and climate reporting resources
+
+Sign up  $\rightarrow$ 
+
+### Support our work
+
+There are lots of ways to help fund Climate Central today
+
+Learn more
+
+Contact Donate Financials
+Us Funders
+
+Careers
+
+Instagram Facebook X
+
+YouTube
+
+(
+
+#### Extreme Precipitation in a Warming Climate | Climate Central
+
+Terms of Privacy Editorial use policy independence
+
+Image /page/12/Picture/3 description: A logo on a white background. At the top, the words 'POWERED BY' are written in a light gray, sans-serif font, with a short horizontal line on either side. Below and to the left is a stylized letter 'C' logo, composed of three colored segments: blue at the top, yellow on the left, and red at the bottom.
+
+Copyright © 2025 Climate Central Registered 501(c)(3). EIN: 26-1797336
+
+EN ES

chatbot/rag/data/('Flood susceptibility modelling using advanced ensemble machine learning models', '.pdf')_extraction.md

@@ -0,0 +1,393 @@
+
+
+## ABSTRACT
+
+Floods are one of nature's most destructive disasters because of the immense damage to land, buildings, and human fatalities. It is difficult to forecast the areas that are vulnerable to flash flooding due to the dynamic and complex nature of the flash floods. Therefore, earlier identification of flash flood susceptible sites can be performed using advanced machine learning models for managing flood disasters. In this study, we applied and assessed two new hybrid ensemble models, namely Dagging and Random Subspace (RS) coupled with Artificial Neural Network (ANN), Random Forest (RF), and Support Vector Machine (SVM) which are the other three stateof-the-art machine learning models for modelling flood susceptibility maps at the Teesta River basin, the northern region of Bangladesh. The application of these models includes twelve flood influencing factors with 413 current and former flooding points, which were transferred in a GIS environment. The information gain ratio, the multicollinearity diagnostics tests were employed to determine the association between the occurrences and flood influential factors. For the validation and the comparison of these models, for the ability to predict the statistical appraisal measures such as Freidman, Wilcoxon signed-rank, and t-paired tests and Receiver Operating Characteristic Curve (ROC) were employed. The value of the Area Under the Curve (AUC) of ROC was above 0.80 for all models. For flood susceptibility modelling, the Dagging model performs superior, followed by RF, the ANN, the SVM, and the RS, then the several benchmark models. The approach and solution-oriented outcomes outlined in this paper will assist state and local authorities as well as policy makers in reducing flood-related threats and will also assist in the implementation of effective mitigation strategies to mitigate future damage.
+
+© 2021 China University of Geosciences (Beijing) and Peking University. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
+
+## 1. Introduction
+
+Among various hydro-meteorological hazards, floods are identified as the most devastating hazards in the world (Prăvălie and Costache, 2013; Mishra and Sinha, 2020; Sarkar and Mondal, 2020). Flood often causes severe loss of human lives and socio-economic damage (Hirabayashi et al., 2013; Costache, 2019). As per the United Nations Office for Disaster Risk Reduction (UNISDR) statistical data, about 150,061 flood events happened worldwide, and around 157,000 people died by
+
+E-mail address: nguyentthuylinh58@duytan.edu.vn (N.T.T. Linh).
+
+floods from 1995 to 2015, which were liable for 11.1% of the global disaster casualties (Wahlstrom and Guha-Sapir, 2015; Hong et al., 2018a). Several investigations report that floods affected around 200 million people each year on a global scale (Tien Bui et al., 2019a). Moreover, considering climate change predictions, land use pattern changes and the increase of population also, it is anticipated that the occurrence rates and intensity of floods are likely to be exacerbated by 2050 which may cause huge loss ( $\approx$ US\$ 1 trillion) (Bubeck and Thieken, 2018; Ali et al., 2019; Alexander et al., 2019; Huang et al., 2019; Jadandideh-Tehrani et al., 2019; Xu et al., 2019; Serraj and Pingali, 2019). Thus, studies on flood modelling spatially occurring at basin-scale at the regional level are urgently needed to alleviate or take adaptation and mitigation measures against the devastating flood's effects (Cao et al., 2019; Su et al., 2019; Costache et al., 2020).
+
+ $<sup>^{\</sup>ast}\,$  Corresponding author at: Institute of Research and Development, Duy Tan University, Danang 550000, Vietnam.
+
+Generally, based on their occurrences, floods are of four types, i.e., flash floods, riverine floods, coastal floods, and urban floods (Tien Bui et al., 2019b; Costache et al., 2020). Flash floods are most devastating among these types of floods, which may cause huge losses of life and properties because of its rapid onset attributes along with grate velocity of flow (Bui et al., 2020a). Moreover, a massive amount of debris has moved along the river basin; thus, floods bring severe destruction for the infrastructure and fatalities (Siegel, 2020; Brinkmann, 2020; Luino et al., 2020; Kuriqi et al., 2016). In the under-developing and developed countries, also the severe destruction by the effect of flash floods are reported (Kuriqi and Ardiçlioğlu, 2018; Bui et al., 2020b; Munir et al., 2020; Okaka, 2020). On the other side, the developing countries are more affected by flash floods compared to developed countries due to the lack of proper infrastructures, essential financial assets, and technological advancement to predict the flood events or to mitigate the consequences of the floods. It is, thus, of immense importance to build up a high-performance prediction model for flood events likelihood and the susceptibility mapping of inundation regions.
+
+Bangladesh, a highly flood-prone country, is positioned in the junction of three mighty Ganges-Brahmaputra-Meghna river systems in the Southeast Asian region (Rahman, 2019). Because of its unique hydro-geographic setting and low-lying flood plain area, Bangladesh experiences floods every year from June to September during the rainy season. Flash floods cause tremendous economic loss, which may be exceeded by US\$20 million and a significant number of fatalities, damage to housing, and infrastructure, including roads and bridges (Fao and Unicef, 2017). Flash floods are frequent events in northern Bangladesh that often occur in the downstream regions, especially in the lower Teesta River basin. For example, in August 2017, a recent flash flood occurred in the lower Teesta River basin, especially in the districts of Kurigram, Lalmanirhat, and Nilphamari, resulting in a significant landmass being submerged and five individuals swept away. According to the Network for Information, Response and Preparedness Activities on disaster (NIRAPAD), a flash flood adversely affected about 6.8 million people and more than 560,000 hectares of cropland in August 2017; the total damage is up to US\$ 10 million (Fao and Unicef, 2017). Flood Susceptibility Modeling (FSM) should be undertaken in the flood danger areas to support decision makers and emergency management in mitigating floods and maintaining risk-prone natural resources. Therefore, FSM is a valuable method to classify at-risk regions and to secure these high-risk regions and natural resources (Masks et al., 2019).
+
+Prediction and development of flood's risk maps are challenging due to several factors, such as roughness heterogeneity (Ardıclıoğlu and Kuriqi, 2019) and the complexity of the river geomorphology (Kuriqi et al., 2019). Nevertheless, in recent decades, regional data were obtained from the satellite by using Remote Sensing (RS) methods coupled with Synthetic Aperture Radar (SAR) data (Li et al., 2019; Nikolaos et al., 2019; Pourghasemi et al., 2020a). Also, optical sensor images (Arora et al., 2019; Talha et al., 2019; Bui et al., 2020a) have been applied to predict flood-prone susceptible regions. The RS data, along with GIS technology, are utilized in a combined fashion to build the list of flood occurrence and several affecting parameters. Remote sensing databases combined with several statistical and mathematical techniques have been used in different research works on basin-scale FSM and flood hazard appraisals (Pradhan, 2010; Youssef et al., 2011; Pourghasemi et al., 2020b). These state-of-art techniques can handle spatial datasets and produce high resolution and prediction performances (Ma et al., 2020; Abba et al., 2020; Uthayakumar et al., 2020).
+
+Many scholars have applied several kinds of methods for developing flood susceptible models (Sahana et al., 2020; Costache et al., 2020; Talukdar et al., 2020). Therefore, many forms of flood-susceptible models can be (Chen et al., 2019a; Bui et al., 2019a; Mahmood et al., 2019; Hong et al. 2018a; Termeh et al., 2018; Siahkamari et al., 2018), like (1) analytical hierarchy process (expert based model) (Souissi et al., 2019; Dano et al., 2020; Nachappa et al., 2020), (2) statistical and bi-variate based models like frequency ratio (FR) (Chen et al.,
+
+2020a; Moghaddam et al., 2019; Khosravi et al., 2019a, Sahana et al., 2020), information value (IV model) (Xu, 2013; Chen et al., 2014), certainty factor, logistic regression (LR) (Tien Bui et al., 2019a; Shafapour Tehrany et al., 2019; Pham et al., 2020a; Ali et al., 2020), weights-of-evidence (WoE model) (Chen et al., 2019b; Paul et al., 2019), fuzzy logic (FL) (Wang et al., 2019a; Sahana and Patel, 2019), neuro fuzzy logic (ANFIS) (Termeh et al., 2018; Hong et al., 2018b) (3) machine learning algorithms (MLAs) (Bui et al., 2019b; Wang et al., 2019b; Wang et al., 2020; Shahabi et al., 2020; Dodangeh et al., 2020), and (4) hydrological models like Soil Water Assessment Tool (SWAT) (Oeurng et al., 2011) and Hydraulic Engineering Centre-River Analysis System among others (Getahun and Gebre, 2015). Machine learning methods, which many researchers use in FSMs, have recently gained more and more interest (Bui et al., 2020a; Chen et al., 2019a; Hong et al., 2018a; Wang et al., 2020). The most popular machine learning techniques are random forest (RF) (Avand et al., 2019; Paul et al., 2019; Chen et al., 2020b; Vafakhah et al., 2020; Nhu et al., 2020), decision trees (DT) (Choubin et al., 2019; Moghaddam et al., 2019; Chen et al., 2020c), support vector machine (SVM) (Termeh et al., 2018; Khosravi et al., 2019a), artificial neural networks (ANN) (Falah et al., 2019; Moghaddam et al., 2019; Pham et al., 2020b; Bui et al., 2020b), and that can forecast flash flooding in the areas at risk, FSM, however, faces numerous obstacles, such as choosing the best modeling methods from a large range of methods, and each approach yields distinct outcomes (Shafizadeh-Moghadam et al., 2018). Also, there are several drawbacks to modeling or forecasting the FSMs for any of these approaches. Consequently, these drawbacks were solved by hybrid ensemble machine learning algorithms, which demonstrated greater results than standard and single models (Pham et al., 2016, 2017a). Of late, researchers have implemented algorithms of ensemble machine learning and techniques of data mining quite extensively, such as Reptree (Chen et al., 2019a; Ghasemain et al., 2020), naive Bayes (Ali et al., 2020; Pham et al., 2020c; Tang et al., 2020), bagging (Shahabi et al., 2020; Chen et al., 2019b; Yariyan et al., 2020), random subspace (Pham et al., 2020b; Chen et al., 2019a) for high precision flood susceptibility models. The excellent results of hybrid machine learning algorithms inspire the implementation and creation of hybrid machine learning algorithms for many natural hazard models. There was, however, no universal consensus on the selection of the best tool for various forms of simulation of natural hazards, such as landslides or flood susceptibility (Chen et al., 2019b). Researchers advocate designing and evaluating new methods for mapping flood susceptibility and other forms of simulation of natural hazards (Chen et al., 2019a). Therefore, in the present study, we also developed the ensemble machine learning algorithms like alternating decision tree-based random subspace and Dagging models for flood susceptibility modelling in the Teesta river basin. However, Dagging ensemble machine learning algorithms have been applied very rarely in different branches of environmental science and disaster management. Thus, we attempted to compare the performance of conventional machine learning and ensemble machine learning algorithms for building flood susceptibility models. The hybrid ensembles of the Dagging, and the RS, coupled with Artificial Neural Network (ANN), Support Vector Machine (SVM) and Random Forest (RF) models have been applied scarcely for natural hazards modelling.
+
+Thus based on concerns mentioned above, the main aim of the present study is (i) to investigate and compare the ANN, SVM, RF, RS, and Dagging model's ability to predict flood susceptibility maps, and (ii) to assess flood influencing factors for generating flood susceptibility maps in the Teesta River basin, Bangladesh. In this study, for generating the flash flood susceptibility map, the Teesta River basin, which is prone to regular flash floods was chosen as a study case. The novelty of this study lies on (i) for the first time, the mapping of flood susceptibility in the Teesta River basin of northern Bangladesh was generated more rigorously; and (ii) for the first time a new hybrid ensemble Dagging model that to best of the author's knowledge has never been utilized before was employed in modelling flood susceptibility maps. The goal of
+
+this contribution is to apply four state-of-art models and ensemble models (RF, ANN, SVM, and RS) along with the proposed Dagging hybrid model to map flood susceptibility and distinguish flood hazard regions in the studied basin. The outcome of the study will assist the regional as well as local authorities and the policy-makers for mitigating the risks related to floods and also help in developing appropriate mitigation measures to avoid potential damages.
+
+## 2. Material and methods
+
+### 2.1. Study area description and materials
+
+The basin considered in this study is positioned in the Teesta subcatchment of the northern region of Bangladesh, occupying about 2284 km² (Fig. 1), spanning over the five major districts, including Nilphamary, Lalmanirhat, Kurigram, Rangpur, and Gaibandha districts. In Bangladesh part, this basin is situated between the geographical locations of latitudes of 25°30′02′′N–26°18′37′′N, and 88°52′58′′E–89°45′ 34′′E longitudes. From the geomorphologic point of view, the floodplain region is the highest geomorphic unit in Bangladesh, and the drainage system comprises several small rivers, which are spread over elevation varying from 5 m to 110 m. For the occurrence of floods, the general slope of the river ranges of 0.47–0.55 m/km (Rahman et al., 2011), which indicates a quite flat terrain. Morphologically, the depression of the basin is shallow and in the moribund river valley, which generates changes of long and morphologically in the pathways of the river basin. This basin is susceptible to flooding. Every year, the flash floods
+
+are a common occurrence; the situation of the basin in the rainy season is worst concerning flood conditions. The hydrological characteristics of the study area are relatively complex, where the distribution of the river network is very dense in association with six rivers, including Buri-Teesta, Naotora, Ghagot, Dharla, Jamuna, Old-Brahmaputra.
+
+In terms of geology, the Teesta basin is situated under the Bengal basin, on the Rangpur saddle of the Indian platform. The new floodplain deposits, including silt, clay, fine to medium sand, are associated with this region (Islam et al., 2014).
+
+The subtropical monsoonal climate in association with two predominant seasons, namely, monsoon (June to September) and dry season (October to May), are the characteristics of this basin. The average annual rainfall of this basin is higher than in 1900 mm. (Akter et al., 2019), and accounts for more than 80% of the total annual rainfall mostly occurred in the monsoon season. In August 2017, the basin area was recorded to have heavy and unusual rainfall. The tropical depression occurred from 6 to 10 August 2017, which was considered as the most destructive phenomenon that happened in the study area.
+
+Consequently, the flash flood occurred, which leads to destroying a tremendous amount of farmland and more than 20,000 houses (Akter et al., 2019). The average temperature in this basin during monsoon and dry seasons is around 15°C and 35°C, respectively. Evapotranspiration has reached its maxima in April, when other climatic parameters including sunshine, wind speed are close to their maximum as well (Zaharia et al., 2020; Hampf et al., 2020). As per the population census 2011 (BBS, 2016), approximately 10.42 million inhabitants are living in this study area. The study area comprised 14% of the
+
+Image /page/2/Figure/10 description: A map of the Teesta River Basin, showing its elevation and the location of training and validation points. The map is titled "Teesta River Basin" and includes a scale bar from 0 to 20 km and a compass rose. The main map displays the river basin with elevations ranging from a low of 18 meters (light green) to a high of 69 meters (dark brown), as indicated by the legend. A river, represented by a light green line, meanders through the basin. Scattered across the map are numerous black triangles representing "Training Points" and white pentagons representing "Validation Points". The map is gridded with longitude lines from 88°53'E to 89°44'E and latitude lines from 25°36'N to 26°10'N. In the bottom left corner, an inset map shows the location of the study area, highlighted by a red box, within the context of surrounding countries including India, Bangladesh, Nepal, Bhutan, China, and Myanmar, with the Bay of Bengal to the south. The caption below the map reads: "Fig. 1. The location of the study area having the training and validation flood points."
+
+Fig. 1. The location of the study area having the training and validation flood points.
+
+agricultural land to the total agricultural land of the Northern part of Bangladesh. Thus, this basin has become a lifetime for Bangladesh. The basin area has been covered with sub-urban types of settlements which are dominated by poor people. Thus, the present study area has been exposed to frequent flooding because of inadequate resources and lack of awareness.
+
+As the research region has undergone regular flooding each year, the historical flooding inventories were also prepared based on the field survey and understanding of local residents. For flood susceptibility modeling, we used multiple data forms in the current analysis. To prepare land use land cover maps, we used Landsat and Operational Land Imager (OLI) (Path/Row: 138/42, Spatial resolution: 30 m., date:19/03/2019), obtained from the website of the United States Geological Survey (USGS). We used ASTER GDEM (Version 2) to extract topographical variables and hydrological variables (Spatial Resolution: 30 m). We collected rainfall from Bangladesh Meteorological Department (BMC), Dhaka, Bangladesh. We utilized a soil taxonomy map from the Natural Resources Conservation Service (NRCS) of the Department of Agriculture of the United States (USDA). The drainage map was generated using topographic maps (scale-1:250,00) collected from the Bangladesh Water Development Board.
+
+### 2.2. Methodology
+
+The methodological flow chart was prepared to show the summary of the whole work that includes flood inventory map, generation of flood conditioning factors, evaluation of the flood conditioning factors using information gain ratio and multi-collinearity test and flood susceptibility models using machine learning methods (ANN, SVM, RF, RS, and Dagging) were utilized for the present study (Fig. 2).
+
+#### 2.2.1. Flood inventory map
+
+The preparation of flood inventory map of the research area is the preliminary stage for creating the flood susceptibility map (Bui et al.,
+
+2020a; Sarkar and Mondal, 2020). The accessible technologies of GIS was used to achieve and investigate the normal state at the flooded sites. Since flood prediction relies on predictive analysis, where the near-future flood likelihood would be influenced by the recent past factor affecting flood, and that is why this is a critical step. 2.283 km<sup>2</sup> of the Teesta River Basin for the preparation of the flood inventory map, where 167 flood points (Fig. 1) of the Teesta River Basin were collected in this report (Plate 1). Hence, the flood points were checked from historical data sources, fieldwork, perception of local residents, and Google Earth ®. The obtained 167 flood points were then randomly divided into 80 percent (134 points) and 20 percent (33 points) categories to train and test flood-susceptible models. Flood vulnerability mapping, however, is known to be a binary classification in which the flood inventory has been categorized into two groups, such as flood points and nonflood points. Therefore, binary values such as 1 as flood points and 0 as non-flood points are needed to prepare the training flood inventory, which is considered to be the dependent variable for prediction. The flood points were the exact points where recurrent floods have been observed, while the non-flood points were the points where floods have not been registered in the last 12 years. We are expected to obtain negative samples or non-flood points, close to flood points. Several researchers suggested choosing the same number of non-flood points as the positive or flood samples to circumvent bias (Tang et al., 2020). Therefore, based on the topographic map, historical flood info, field survey and NDWI 12-year maps, we randomly selected 167 non-flood points. Non-flood points have been allocated to 0. Subsequently, we split the non-flood points arbitrarily into 80% (134 points) and 20% (33 points) categories. Thus, as training data sets, we prepared a dependent factor that included 134 flood points as 1 and 134 non-flood points as 0.
+
+Similarly, validation datasets were also designed to test the final models, including both 33 flood and non-flood points as 1 and 0. Both the datasets for training and validation are seen in Fig. 1. Depend on the training datasets, by using the 'extract values to point' tools in
+
+Image /page/3/Figure/10 description: A flowchart illustrating a methodological approach for flood susceptibility analysis. The process is divided into three main sections: Data collection and analysis, Statistical test, and Flood susceptibility modeling.
+
+1. \*\*Data collection and analysis\*\*: This section starts with identifying 'Flood influencing factors'. Data is collected from four sources: DEM, Satellite image, Meteorological data, and River. These sources are used to derive several factors:
+ \* From DEM: Slope, Aspect, Soil type, TRI, Curvature, TWI, SPI, STI.
+ \* From Satellite image: LULC.
+ \* From Meteorological data: Rainfall.
+ \* From River: Distance to road.
+
+2. \*\*Statistical test\*\*: This section, labeled 'Selection of Flood conditioning factors', involves the 'Selection of influencing factors' using three methods: Information gain ratio, Multicollinearity diagnostics tests, and Pearson's correlation.
+
+Both the data from the analysis and the selected factors from the statistical tests feed into a central 'Training and validation data set'.
+
+3. \*\*Flood susceptibility modeling\*\*: This section uses the training and validation data set for two parallel processes:
+ \* \*\*Building of flood models\*\*: The models listed are Artificial Neural Network, Support Vector Machine, Random forest, Random subspace, and Dagging.
+ \* \*\*Validation of flood susceptibility models\*\*: The validation metrics are ROC curves, RMSE, MAE, and R-square.
+
+The outputs from both the model building and validation processes contribute to the final step, which is 'Flood Susceptibility Mapping'.
+
+Fig. 2. Methodological approach applied for flood susceptibility analysis.
+
+ArcGIS 10.5 software, we derived data from twelve flood conditioning parameters (spatial datasets). We then imported these datasets into the WEKA package (version 3.9.3), and all the modeling was performed over there.
+
+In plate 1. field photographs of the flooding situation in the Teesta river basin representing (a,c,d) flooded road, (b) damaged houses due to flooding, (e) flooded village, (f, h) destroyed road due to devastating flooding, (g) camping on the national road by the people affected by flooding, and (i) over flow on the culvert.
+
+#### 2.2.2. Methods for preparing flood conditioning factors
+
+The construction of the susceptible spatial flood model is typically very complex and precise, as many geospatial topographical and hydrological variables are needed. Recognition of variables driving the flood is thus a critical activity, and the scientifically selected parameters will validate the accuracy of the flood susceptibility maps. Based on the available literature on flood susceptibility, in the current study area, the twelve flood influencing parameters were selected, such as elevation, curvature, aspect, slope, topographic roughness index (TRI), topographic wetness index (TWI), stream power index (SPI), sediment
+
+transport index (STI), land use/land cover (LULC), distance to the river, soil type, and rainfall (Janizadeh et al., 2019; Paul et al., 2019; Bui et al., 2019b; Chen et al., 2019b; Sturzenegger et al., 2019; Moghaddam et al., 2019; Arabameri et al., 2019). The all influencing factors were transformed into the format of raster with 30m spatial resolution.
+
+Topographic factors are of paramount importance for modelling flood studies, which directly and indirectly influence the hydrological characteristics of the study area (Bui et al., 2020b; Arabameri et al., 2020; Lei et al., 2020). At first, a Digital Elevation Model (DEM) was prepared from the ASTER GDEM (Version 2) with 30m spatial resolution for the study basin in ArcGIS 10.2 environment. We have derived topographic factors from DEM, such as slope, curvature, aspect, topographic wetness index (TWI), stream power index (SPI), sediment transport index (STI), and topographic roughness index (TRI) in ArcGIS environment.
+
+*2.2.2.1. Elevation.* The elevation is the most vital factor affecting modelling floods (Dodangeh et al., 2020; Bui et al., 2020b). The flooding is inversely related to the elevation, Higher the elevation, lower the chances
+
+Image /page/4/Figure/9 description: A figure, labeled Fig. 3, displays four maps of the same geographical region, illustrating data layers for flood susceptibility mapping. The maps are labeled a, b, c, and d.
+
+Map (a) shows Elevation in meters (m). The color scale ranges from red for high elevation (69 m) to green for low elevation (18 m). A legend also indicates that a black dot represents a flooding point.
+
+Map (b) shows Curvature. The color scale ranges from green for high curvature (0.824) to red for low curvature (0.327).
+
+Map (c) shows Aspect (Direction). A detailed legend provides color codes for different directions: Flat (-1) is grey, North (0-22.5 and 337.5-360) is red, Northeast (22.5-67.5) is orange, East (67.5-112.5) is yellow, Southeast (112.5-157.5) is light green, South (157.5-202.5) is light blue, Southwest (202.5-247.5) is medium blue, West (247.5-292.5) is purple, and Northwest (292.5-337.5) is dark purple.
+
+Map (d) shows Slope in degrees (°). The color scale ranges from blue for high slope (5.75°) to red for low slope (0°).
+
+A scale bar at the bottom of the figure indicates a length of 80 kilometers.
+
+Fig. 3. Data layers for flood susceptibility mapping (a) elevation, (b) curvature, (c) aspect and (d) slope
+
+of occurrences of flooding, and vice versa (Chen et al., 2020d). Hence, the Teesta river basin is situated in a flat and lower elevated region. Consequently, the chances of flooding are frequent in the entire region (Basin, 2017) (Fig. 3a).
+
+2.2.2.2. Curvature. As the curvature affects the flooding water budget, it is thus, used for the modelling of flood susceptibility (Ahmadlou et al., 2019). Consequently, The curvature distinguishes the divergent and convergent runoff regions (Torcivia and López, 2020). The runoff convergence operation is related to the negative value regions (Costache and Bui, 2020; Rau et al., 2019). These areas are highly vulnerable to flash-flooding (Fig. 3b).
+
+2.2.2.3. Aspect. Aspect, another factor, affects the directions of flooded water flows, along with which maintains the soil humidity (Chu et al., 2020). The aspect is, thus, indirectly affecting the flooding. For example, the shaded slope area, where the humidity of soil is relatively high and with extreme runoff (Fig. 3c).
+
+2.2.2.4. Slope. The other important factor for influencing the flood is the slope, which controls the flowing water's (flooded) speed (Stevaux et al., 2020). The higher the slope angle, the lower the chances of water stagnation, a lower rate of infiltration, and the higher the flow
+
+velocity. Therefore, the lower and flat regions, like the Teesta river basin, will have more chances of flooding (Fig. 3d).
+
+2.2.2.5. Topographic roughness index. Topographic roughness index (TRI) is one of the major influencing factors for flood events. It depends on the local topographic of the study basin. The occurrence of higher flood probability is associated with the lower TRI value. In the present study, the TRI map was kept in stretch format, and the value ranges between 0-27 (Fig. 4a).
+
+2.2.2.6. Topographic wetness index. For the occurrence of flood, Topographic Wetness Index (TWI), is an effective factor (Abdel Hamid et al., 2020), by which the difference in the wetness of a basin was spatially expressed (Meles et al., 2020). This index shows the amount of water contained in every pixel size of the region. (Zhang et al., 2020), and that is computed using the Eq. (1):
+
+$$TWI = \frac{\ln{(A_s)}}{\tan{\beta}} \tag{1}$$
+
+The specific catchment area  $(m^2m^{-1})$  and the slope gradient (in degrees) are indicated by  $A_s$  and  $\beta$ , respectively. In general, the high TWI values and flooding are strongly correlated with each other (Shit et al.,
+
+Image /page/5/Figure/11 description: A figure displaying four maps of the same geographical region, labeled a, b, c, and d, each illustrating a different topographical index. A scale bar at the bottom indicates distances from 0 to 100 km.
+
+Map (a) shows the Topographic Roughness Index (TRI). It is a pixelated map with colors ranging from green for low values to red for high values. The legend indicates a range from a low of 0 to a high of 27.
+
+Map (b) shows the Topographic Wetness Index (TWI). This is a higher resolution map with colors from green (low) to red (high). The legend shows a range from a low of -1.54 to a high of 7.72.
+
+Map (c) shows the Stream Power Index (SPI). The entire region is colored a uniform brownish-yellow. The legend has a color gradient from green to yellow, labeled 'Low' and 'High' respectively, without numerical values.
+
+Map (d) shows the Stream Transport Index (STI). This map highlights a network of streams in shades of yellow and red against a dark green background. The legend indicates a range from a low of 0 to a high of 140.64.
+
+Fig. 4. Data layers for flood susceptibility mapping (e) topographic roughness index, (f) topographic wetness index, (g) Stream power index, (h) stream transport index.
+
+2020). The TWI value in the study area ranges from -1.54 to 7.72 (Fig. 4b).
+
+2.2.2.7. Stream power index. Stream power index (SPI) has an essential influence on the fluvial system (Knighton, 1999). The SPI is computed using Eq. (2).
+
+$$SPI = A_s tan\beta \tag{2}$$
+
+Where,  $A_s$  and the slope gradient represent the specific catchment area is indicated by  $\beta$  (radians) (Wu et al., 2020a). The sediment transport capacity and the erodibility of its bed is referred to as total SPI (Chen et al., 2020c).
+
+2.2.2.8. Sediment transport index. Sediment transport index (STI), another factor for causing flooding, can increase the frequency of flood, which upshots in the damage of foundation (Wang et al., 2019c). The STI is derived from DEM using Eq. (3).
+
+$$STI = \left(\frac{A_s}{22.13}\right)^{0.6} \left(\frac{sin\beta}{0.0896}\right)^{1.3} \tag{3}$$
+
+where, each pixel of the slope is represented by  $\beta$ , and the area of upstream is indicated as  $A_s$  (Wang et al., 2019c). Based on hydro-climatic and geomorphologic attributes of the basin area, the STI is calculated
+
+(Shukla and Gedam, 2019). The channel's bed changes with the deposition of sediment, which reduced the capability of storing water within the channel and caused flooding (Antoniazza et al., 2019). In the present study, the STI value ranges from 0 to 140.64 (Fig. 4d).
+
+2.2.2.9. Land use/Land cover. LULC influences the runoff of the surface and sediment transportation and, consequently, directly affects the frequency of floods (Benito et al., 2010). Because the LULC directly controls the surface runoff generation and infiltration. The flooding is more frequent in the built-up area as these regions do not allow to infiltrate water and generate surface water. In contrast, the forest area allows infiltrating water that causes less flooding (Abdel Hamid et al., 2020). The relationship between flood events and the density of vegetation is inverse when it compares hydrological responses at various temporal scales (García-Ruiz et al., 2008; Dodangeh et al., 2020). In the present study, we have prepared a LULC map from Landsat and OLI image (path/row: 138/43; spatial resolution: 30 m) using an Artificial neural network in ENVI software (version 5.3). We classified six classes in the LULC map, such as vegetation, bare land, built-up area, sand bar, agricultural land, and water body (Fig. 5a).
+
+*2.2.2.10. Distance to the river.* Most affected flood-inundation areas are usually located adjacent to the river. The distance from the river is an essential conditioning factor for identifying the regions which are
+
+Image /page/6/Figure/12 description: The image displays four maps of the same geographical region, labeled a, b, c, and d, each illustrating a different environmental factor. A scale bar at the bottom indicates a range from 0 to 80 km.
+
+Map (a) is titled "Land use land cover" and shows the distribution of different land types. The legend includes: Vegetation (dark green), Bare land (light blue), Built up (red), Sand bar (gray), Agricultural land (orange), and Water body (dark blue). The map shows a river system running through the region, surrounded by agricultural land, sand bars, and built-up areas, with large expanses of vegetation further away.
+
+Map (b) is titled "Distance to river (m)" and uses a grayscale gradient to show proximity to the river. Black represents a low distance of 0 m, while white represents a high distance of 1503 m. The river channels are clearly delineated in black.
+
+Map (c) is titled "Soil types (USDA)" and uses a color-coded system to display different soil classifications. The legend includes: Water (dark green), Usterts (medium green), Aquults (light green), Humults (lighter green), Udults (even lighter green), Ustults (yellow-green), Aqualfs (yellow), Ustalfs (light orange), Ochrepts (medium orange), Aquepts (darker orange), and Psamments (darkest orange).
+
+Map (d) is titled "Rainfall (mm)" and shows rainfall distribution using a color gradient. The legend indicates that red represents high rainfall (550.411 mm) and green represents low rainfall (361.066 mm). The map shows higher rainfall in the northwest, gradually decreasing towards the southeast.
+
+Fig. 5. Data layers for flood susceptibility mapping (a) land use, land cover, (b) distance to river, (c) soil types, and (d) rainfall.
+
+vulnerable to flooding of a basin (Tehrany et al., 2015) because the distance controls the incident of flooding and the river flow to river factor (Gupta, 2020). The higher the distance to the river is, the lower the likelihood of flood events will be (Binh et al., 2020). At the basin scale, the regional level flooding is related to the storage of terrestrial water (Deng et al., 2020; Frappart et al., 2019). In the present study, we have prepared the distance to the river map from the topographic map having a scale of 1: 50,000 and Google Earth (Fig. 5b).
+
+2.2.2.11. Soil types. One of the vital influencing factors which control the mechanism of rainfall-runoff is soil types (Xie et al., 2019). The soil properties directly control the water infiltration that influences the rainfall-runoff generation, although the other factors like local weather conditions, erosion process control rainfall-runoff (Flügel, 1995; Phillips et al., 2019). The higher the rate of infiltration, the lower the chances of flooding occurrences. In the present study, we have used the United States Geological Survey (USDA) soil map and classified it into 12 classes based on USDA soil taxonomy (Fig. 5c).
+
+2.2.2.12. Rainfall. Rainfall has been chosen as one of the major influencing factors for the occurrence of flooding (Pourghasemi et al., 2020a). Because the high-intensity rainfall within a short period can cause flooding (Aristizábal et al., 2020; Lu et al., 2020; Peptenatu et al., 2020). We have obtained rainfall data from four meteorological stations of Bangladesh and employed the kriging interpolation method for preparing rainfall maps in ArcGIS 10.3 environment. We used the kriging method for interpolation because we have data only for four points, and this method is highly suggested when the amount of data is very less (Kourgialas and Karatzas, 2011). However, the annual rainfall ranges from 361 to 550mm in the study area (Fig. 5d).
+
+### 2.3. Method for flood influencing factors using Information gain ratio and multicollinearity test
+
+Primarily, the evaluation of the flood influencing parameter's importance or the ability for flooding is critical before the training and validation checking of the model (Sameen et al., 2020). The quantification of the importance of each collected parameter has been derived based on their statistical properties and correlation with the flooding. To identify the influential parameters for the prediction of FSMs, the Information Gain Ratio (InGR) technique (Bui et al., 2020a) has been utilized. For measuring the importance of every influencing factor, an InGR value is allocated. The higher InGR values are associated with, the more significant influencing factor. The InGR model is chosen to be employed in the present study due to its simplicity and efficacy and calculated using Eq. (4).
+
+$$\begin{aligned} \text{Gain ratio}(x,Z) = & \frac{\text{Entropy }(Z) - \sum_{1}^{n} \sum_{\substack{i=1 \ |Z|}}^{n} \frac{|Z_i|}{|Z|} \text{ Entropy }(Z_i)}{-\sum_{i=1}^{n} \frac{|Z_i|}{|Z|} \log \frac{|Z_i|}{|Z|}} \end{aligned} \tag{4}$$
+
+where from the training point Z with  $Z_i 1 = 1, 2, 3, .... n$  subsets, the attribute x is belonging.
+
+Several multi-collinearity tests, such as correlation coefficients of Pearson (Xu and Li, 2020), variance decomposition proportions (Zhang et al., 2019), conditional index (Al-Quraishi et al., 2020), and Variance Inflation Factors (VIF) and tolerances (Javidan et al., 2020) have been used to quantify influencing factors for all probability models. In this study, we have utilized the Pearson's correlation coefficient and VIF for quantifying the importance of twelve flood conditioning factors. The VIF >9 and very low correlation shows the problem of multicollinearity in the factors (Duque and Aquino, 2020). It is, thus, highly recommended to exclude the conditioning factor for modelling, which has the VIF >9.
+
+### 2.4. Methods for flood susceptibility modelling
+
+#### 2.4.1. Artificial Neural Networks (ANN)
+
+In the present study, with a Back-Propagation (BP) and errorcorrection learning algorithm, the ANN model having three layers (input, hidden and output layers) has been used and successfully employed in the flood susceptibility modelling (Bui et al., 2019b; Zhao et al., 2020). In this study, for the input layers and ten hidden nodes, which are identical with the numbers of the critical parameters were set (Table 1). On the other hand, one node has been employed for the output layer, which is coded as 1 for flood events and 0 for non-flood. Many algorithms are available for the training of an ANN model; however, the most popular one is BP. Thus, for estimating the nonlinear relationship between the critical parameters and the events of the flood, the BP based ANN model has been applied (Dodangeh et al., 2020). Firstly, the initial weights are chosen randomly by BP. The comparison between computed and observed values have taken place. The differences between computed and observed value can be termed as an error. It has been evaluated by several error measures techniques like mean squared error (MSE), root mean square error (RMSE) (Hou et al., 2020). Based on a rule of generalized delta (Rumelhart et al., 1986), the adaptation of initial weights occurs to allocate the total error among the neurons in the network (Pijanowski et al., 2002). The iteration of this process occurs, while the level of error is at a lower level.
+
+#### 2.4.2. Support Vector Machine (SVM)
+
+A widely popular machine learning model is the SVM, along with a set of linear indicator functions, which has been employed for the issues of function determination, is introduced by Vapnik (2013). For the data transformation in the SVM model, the function of the kernel mathematical system has been used. By using the training datasets, a hyper-plane has been generated, when the conversion from actual SVM datasets to high dimensional feature space has occurred (Choi et al. 2020). For the differentiation of the actual output space, the best linear hyper-plane has been applied. It also used to categorize the data into two classes, such as the susceptibility of non-flood and flood {0, 1}. The capability of the SVM model mainly depends on the suitable kernel functions, e.g., polynomial kernel (PL), sigmoid kernel (SIG), radial basis function (RBF) and linear kernel (LN). Several studies (Tien Bui et al., 2012; Tehrany et al., 2015) reported that the RBF outweighs among different kernel functions for flood susceptibility modelling, which is selected as a benchmark kernel function. Due to the flexibility of radial basis kernel to deal with various dimensionalities of the dataset and its better generalization ability, it has been applied mainly for modelling the flood susceptibility (Chen et al., 2020a; Yang and Cervone, 2019). The significant limitations of modelling with SVM are commonly related to its difficulty in capturing vital parameters (Choubin et al., 2019).
+
+**Table 1**The parameters of the machine learning algorithm used for flood susceptibility modelling
+
+| Model name                   | Description of parameters                                                                                                                                                        |
+|------------------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| Artificial neural<br>network | Hidden layer-6, learning rate-0.28, momentum-0.17, seed-4, training time-500, validation threshold-20, Normal to binary filter-TRUE                                              |
+| Random forest                | Batch size-100, seed-5, number of iteration-150, max depth-2, calc out of bag-TRUE, Compute attribute importance-TRUE                                                            |
+| Random subspace              | Classifier-J48, max depth- -1, minimum number-2, minimum proportion of the variance-0.002, seed-5, numbers of executions slots-1, number of iteration-20, subspace size-0.45     |
+| Support vector<br>machine    | Complexity parameter-1, epsilon-1.0E-12, kernel- radial basis function, random seed-5, number of fold- -1, tolerance parameter-0.001, check turned off-True, Calibrator-Logistic |
+| Dagging                      | Classifier-M5P, max depth- -1, minimum number-2, a minimum proportion of the variance-0.001, seed-3, numbers of fold-10, verbose-TRUE                                            |
+
+#### 2.4.3. Random forest
+
+The Random forest (RF) approach, is the aggregate of decision trees for classification or regression to predict, was pioneered by Breiman (2001). It is a popular ensemble-learning algorithm (Rahman et al., 2019). The RF is divided into two subsets; Breiman's "random subspace random subspace random subspace random subspace" idea is the first one, and the second is Ho's "random selection features. The Random subspace is an ensemble machine learning process, which utilizes and improves the accurate prediction of a weak classifier by the creation of a set of classifiers. The RF generates various sets of samples by sampling with replacement and creates multiple matching regression tree training phases and then estimates the data classification based on the voting outcomes of multiple classifiers. In the end, a classification with a considerable number of votes over all the trees in the forest has been found and selected by RF.
+
+In the training phase of the regression tree, the rules based on the response parameters were developed to categorize the observation datasets until the prediction has the lowest possible node deviation. One of the significant problems with regression trees is that it tends to outfit the training detests and, thus, performs inadequately when given an uncertain dataset (Criminisi and Shotton, 2013). RF is a feasible way to overcome this weakness. When each regression tree has been trained in the RF algorithm, a portion of the input records and predictor parameters have been randomly chosen as input to the training phase. A set of regression trees has been built after multiple sampling practices, and each set of regression trees is only a training outcome for a randomly chosen subset. It is not advisable to apply a full sample to train the decision trees, because the full sample training neglects the laws of local samples. In the case of flood susceptibility to make the comparison with the outcomes of a new hybrid model, the model which is applied as one of the benchmark models is the RF model.
+
+#### 2.4.4. Random subspace
+
+Ho (1989) introduced the Random subspace model as a new ensemble machine learning algorithm for solving environmental problems. On modified feature space, the several classifiers of this model are integrated and trained. For the classifiers, which are training base, this model creates several training subsets (Ho, 1998). Unlike other ensemble models, it uses several samples on feature space, rather than the instance space (Havlíček et al., 2019). This method gets the benefit from the bootstrapping and the aggregation. The training set x, the base classifier w, the number of subspaces L, and the number of subspaces is the four inputs of RS (Luo et al., 2019). The L subsets of ds features have been produced randomly and gathered in subspace, in RS. The base classifier has been trained for generating L different classifiers in subspace on each of the subsets. For producing an ensemble classifier E, these classifiers are combined. In Kuncheva et al. (2010), a detailed description of the algorithm is given. For avoiding the problems regarding over-fitting and to deal with the datasets with the most unnecessary feature, it is highly recommended to use this method (Pham et al., 2017b).
+
+#### 2.4.5. Dagging
+
+Dagging was first pioneered by Ting and Witten (1977), which is a renowned re-sampling ensemble model applying most votes for the integration of different classifiers to get better precision in prediction for the base classifiers (Zhao et al., 2013; Walia and Kumar, 2019). The application of Dagging ensemble technique in the field of flood susceptibility is rare. However, this model has been broadly employed in the fields of medicine (Hosni et al., 2019), banking (Sawant and Bhurchandi, 2019), and recently, in landslide susceptibility modelling (Pham et al., 2017a). Instead of creating the bootstrap samples to obtain the base classifier, it creates several separate samples. It is considered a
+
+promising machine learning algorithm for classification in recent decades. The Dagging ensemble model has been applied for solving numerous problems regarding classification in the case of the real world. For a particular training dataset including N samples, the creation of an M dataset can occur and come from the actual training datasets. Every dataset has several n (n<N) samples that are unique from each other. In Dagging, the technique of sampling which is utilized in this. The samples are not replaced in the actual training datasets and can be selected as a part of a specific dataset where the number of sample datasets is increased. Accordingly, on every sample dataset, a base classifier was built. In the end, several classifiers will be gained based on the training dataset.
+
+### 2.5. Precision appraisal
+
+Either a positive or negative pixel has been included in each class of flood susceptibility mapping classification. The real positives and true negatives are referred to as those positive (flood) and negative (nonflood) pixels, which are divided accurately into positive or negative classes. In contrast to this, false positives and false negatives are referring to those positive and negative points which have been mistakenly divided (Schumann et al., 2014). The area has explained the precision of two classes in a receiver operating curve (ROC) graph under the ROC curve (AUC) value. This method has been employed for validating the five flood susceptible models. Hence, by presenting the 1-specificity on X-axis against sensitivity on Y-axis, the ROC curve has been plotted. The sensitivity indicates the sum of the pixels accurately divided as the pixels of the flood. Also, the 1-specificity represents the several non-flood pixels accurately divided, by utilizing the following Eq. (5) the value of AUC can be calculated (Tien Bui et al., 2018).
+
+$$AUROC = \frac{\dot{a} TP + \dot{a} TN}{P + N}$$
+ (5)
+
+Where the AUC has been calculated to appraise the model's overall performance when categorizing indefinite data, in general, AUC varying from 0.5 to 0.6 donates an incompetent model. In contrast, the AUC value between 0.6 and 0.7 indicates a poor performance model. Along with it, a model with 0.7 to 0.8 AUC value demonstrates moderate performance. When the AUC value greater than 0.8 indicates a high fitness model with the dataset (Sajedi-Hosseini et al., 2018). In addition to the AUC value, the root means square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) was applied for evaluation and comparison of model outcomes in this study.
+
+To appraise the accuracy of flood susceptibility modelling, several non-parametric statistics have been applied in this study. Any statistical inference is not required for non-parametric models (Derrac et al., 2011). The non-parametric tests, such as the Friedman test and paired t-test (Friedman, 1937) were adopted when the distribution of the data was normal (Martínez-Álvarez et al., 2013). These tests have been employed to compare the significant differences between two or more models (Beasley and Zumbo, 2003). At the 5% (p<0.5) (null hypothesis) significance level, the flood model's performances are similar, and this is the first inference for running this test. The outcomes were not applied, in the case of true p-values for all the models in the Friedman test were >0.5 (Bui et al., 2018).
+
+The Wilcoxon signed-rank test was conducted to verify the flood models by evaluating the statistical significance of systematic pairwise differences among the models. In the case of the five susceptibility models, the p-value and the z-value were utilized to evaluate the significance of differences among them to make sure of the statistical significance. The null hypothesis will be rejected if the value of p is < 5%, and the value of z is more than the critical values of z (+1.96 and -1.96).
+
+Image /page/9/Figure/2 description: A vertical bar chart showing the 'Average Merit' for different 'Flood conditioning parameters'. The y-axis, labeled 'Average Merit', ranges from -0.1 to 0.6. The x-axis, labeled 'Flood conditioning parameters', lists ten categories: LULC, Distance to road, Elevation, Slope, TRI, SPI, STI, Curvature, TWI, and Aspect. Each bar is blue and includes a black error bar. The bar for LULC is the tallest, with an average merit of approximately 0.52. The next three bars, 'Distance to road', 'Elevation', and 'Slope', have similar heights, with average merits of about 0.12, 0.11, and 0.11, respectively. The remaining parameters have much lower average merits, all below 0.05: TRI (approx. 0.03), SPI (approx. 0.03), STI (approx. 0.015), Curvature (approx. 0.01), TWI (approx. 0.01), and Aspect (approx. 0.005).
+
+Fig. 6. The importance of flood conditioning factors using the information gain ratio, the standard error bars show the accuracy of the results.
+
+## 3. Results
+
+### 3.1. The influential factors in modelling flood susceptibility by InGR and multicollinearity tests
+
+The analysis result is presented in Fig. 6, where the values of the InGR of each parameter are calculated using a 10-fold cross-validation technique. The findings of InGR showed that the LULC (0.52), DR (0.12), elevation (0.11), and slope (0.495) are the more significant with the high InGR values as compared to other factors of influence. Total roughness index TRI (0.03), stream power index SPI (0.03), and sediment transport index STI (0.02) demonstrate moderate relevancy. Similarly, the total wetness index, TWI (0.01), and curvature (0.01) show low relevance. It is noteworthy; the aspect factor exhibits the value of InGR= 0.00; therefore, it is a less influencing factor for predicting FSMs.
+
+The studied river basin is predominated by agricultural land near the main river. In contrast, the vegetation occupies 55% of the total area features. For the flash floods incident, these features are most favourable (lcyimpaye, 2018). Also, for the occurrence of floods, the most recognized significant features are the elevation and slope. This is much found and explained in the works of Tien Bui et al. (2019b). The uneven topography with lower elevation and curvatures in convergent pattern and low slope angle causes a higher chance of flooding in the study area. However, for the occurrence of flood, there is no predictive value in case of an aspect due to the flooding point distribution in the directions of the slope is flat and smooth.
+
+The outcomes of multicollinearity diagnostics tests (Table 2) reveal that elevation with the largest VIF (2.71), followed by rainfall, STI, SPI, slope, soil type, distance to the river, TRI, TWI, LULC, curvature, and aspect respectively. The results also indicate that there is no collinearity problem among 12 potential factors for involving flood susceptible models. Additionally, Pearson's correlation between flooding and its
+
+potential influencing factors are presented in Table 3. According to the prior research (Douglas et al., 2000), the probability of flood occurrence is increasing with higher correlation values. The critical potential factors for the occurrence of the flood were distance to the river, followed by elevation, soil type, rainfall, aspect, curvature, TWI, LULC, slope, SPI, STI, TWI. In this immediate area of study, the majority of floods occurred where the distance to the river is low (0-100) m, the soil type is aquafssol, the elevation is 18–69 m, the curvature is >0.327, land cover class have bare land, SPI is high, STI is high, rainfall ranges from 361 to 550 mm, the slope belonging from 0–5°, TWI belonging from – 1.54 to 7.72, TRI is high, and the aspect is flat.
+
+### 3.2. The analysis of flood susceptible models
+
+The five machine learning models, such as ANN, SVM, RF, subspace random, and Dagging, were employed to compute the maps of flood susceptibility for each pixel in the basin. The Dagging model has been proved as the high-performance model of prediction among all the benchmark models for the geospatial datasets obtained based on the above results of the experiment. Several techniques, such as natural break, equal interval, quantile, regular interval, standard deviation, and manual technique, are available for reclassifying the flood susceptible models in the ArcGIS 10.3. Among these techniques, quantile and natural break methods have been widely reported in the available literature of flood susceptibility studies (Tehrany et al., 2015; Tien Bui et al., 2019b). The quantile is one of the famous reclassifying techniques which is able to provide better outcomes in comparison to other reclassifying techniques, and it has been mostly used for reclassifying the flood susceptible maps (Tehrany et al., 2015). For this reason, in this present study, the quantile reclassifying approach was selected. The flood susceptibility models were reclassified into five classes such as very low, low, moderate, high and very high (Fig. 7).
+
+**Table 2**The evaluation of the flood conditioning factors using a multi-collinearity test (VIF).
+
+| Parameters                      | Elevation | Aspect | Curvature | Slope   | Topographic<br>roughness<br>index | Topographic<br>wetness<br>index | Stream<br>power<br>index | Sediment<br>transport<br>index | Soil types | Distance<br>to river | Rainfall | Land use<br>land<br>cover |
+|---------------------------------|-----------|--------|-----------|---------|-----------------------------------|---------------------------------|--------------------------|--------------------------------|------------|----------------------|----------|---------------------------|
+| Multicollinearity test<br>(VIF) | 2.7128    | 1.0786 | 1.08072   | 1.33623 | 1.196                             | 1.1935                          | 1.56453                  | 1.6207                         | 1.31539    | 1.29923              | 2.67087  | 1.10026                   |
+
+**Table 3**The Pearson's correlation coefficient between flooding and its potential factors.
+
+| Influential<br>factors | Distance to<br>river | Elevation | Soil<br>types | Rainfall | Aspect | Curvature | Topographic<br>wetness index | LULC   | Slope  | Stream<br>power index | Sediment<br>transport index | Topographic<br>roughness index |
+|------------------------|----------------------|-----------|---------------|----------|--------|-----------|------------------------------|--------|--------|-----------------------|-----------------------------|--------------------------------|
+| Correlation (Pearson)  | 0.6024               | 0.4519    | 0.4429        | 0.3463   | 0.2325 | 0.1417    | 0.0947                       | 0.0662 | 0.0435 | 0.0359                | 0.0289                      | 0.0169                         |
+
+The five flood susceptible models having five classes have been represented in Fig. 7. The findings showed that, in the case of Dagging model, the smallest area percentage (14.03%) comes under the very high class, followed by low (14.89%), high (15.61%), moderate (23.78%), and very low (31.70.42%) classes. The percentages of area are 30.20%, 16.22%, 13.32%, 16.88%, and 23.39 % for the very low, low, moderate, high, and very high classes, respectively, in the case of the SVM model. In the case of RS, RF, and ANN models, the area percentages for very high flood susceptible classes are 21.70%, 29.93%, and 33% (Fig. 8). The very high flood susceptibility class has occupied the highest area (1160 km<sup>2</sup>; indicated by red colour), and the very low flood susceptible class has occupied the smallest area (510 km<sup>2</sup>; indicated by dark green) in case of Dagging hybrid model. In the case of the SVM model, the maximum area (1120 km<sup>2</sup>) has come under the very high flood susceptible class, while the moderate flood susceptible class has covered the minimum area (500 km<sup>2</sup>). Out of the total area 1020, 1090, and 90 km<sup>2</sup> areas have come under the very high flood susceptibility class of the RS, RF, and ANN models respectively. In the class of low flood susceptibility, the areas are 780, 1130 and 1230 km<sup>2</sup>.
+
+It is obtained in all maps, in the areas of a flat plain which is close to the main river, which has very high and high flood susceptible areas. The identical results were found in the works of Tehrany et al. (2015). Also, in the eastern and south-eastern parts of the river basin, the very high and high susceptible classes were mainly located. Almost the same spatial distribution pattern of flood susceptible areas has been found in the case of all five models in terms of area under high and very high class, though in the case of the Dagging model, the percentages (14.03%) is comparatively smaller than the other models. This
+
+indicates a more practical outcome related to the Dagging hybrid model than other models, with greater accuracy and dependability. This outcome may be time decreasing and costing of a capable mollification planning in the Teesta river basin targeted for land-use plan and policy implication.
+
+The region surrounding the river basin and the main channel of the river were majorly susceptible to flood. They were in the class of very high susceptibility as depicted by the map of flood susceptibility. This outcome was also according to the results of Yang et al. (2015) in Hainan Province, China. The settlement area of the study areas, indicated by red colour (Fig. 5a), was found in the same region (eastern part of the study area) where very high and high flood susceptible areas were located (Fig. 7). Yang et al. (2015) also reported that the identical spatial pattern of flood susceptible zones and urban areas were found in their working area. Yang et al. (2015) stated that the flooding has mostly harmed the world's approximately 0.18 million people at annual scale in those regions, which are nearer the rivers and mostly dominated by urban and semi-urban areas.
+
+### 3.3. Validation and comparison
+
+We have validated the five flood susceptible models using RoC based on the validation and training points (Fig. 9). The validation for the flood susceptible maps has analysed, which were generated utilizing the models through the datasets of training point (for flood existing the success rate curve) and validation point (for anticipated flood event the prediction rate curve). That is why, on the maps of flood susceptibility, the points of flood validation and training were positioned. The Area
+
+Image /page/10/Figure/10 description: A figure displaying five maps, labeled (a), (b), (c), (d), and (e), each illustrating 'Flood Susceptibility' for the same geographical region. The maps use a color gradient to indicate the level of risk. A legend on the right side of the figure defines the color scheme: dark green for 'Very low', light green for 'Low', yellow for 'Moderate', orange for 'High', and dark red for 'Very high'. All five maps show a similar pattern with a central, elongated area of high to very high susceptibility (orange and red), surrounded by areas of low and very low susceptibility (green) along the irregular borders. The extent and intensity of the high-risk areas vary slightly across the five maps. Below the legend, a scale bar indicates distances from 0 to 80 km, with markings at 0, 10, 20, 40, 60, and 80 km.
+
+Fig. 7. Flood susceptibility models using (a) Artificial neural network, (b) Support Vector machine, (c) Random forest, (d) Random subspace, and (e) Dagging,
+
+Image /page/11/Figure/2 description: A clustered bar chart compares the area distribution across different flood susceptible zones as predicted by five different models. The y-axis is labeled "Area (km²)" and ranges from 0 to 1600. The x-axis is labeled "Flood susceptible zones" and has five categories: Very Low, Low, Moderate, High, and Very high. The legend indicates the five models: ANN (blue), RF (red), Random subspace (green), SMO (purple), and Dagging (teal). The approximate data for each zone is as follows: For the 'Very Low' zone, the areas are: ANN ~80, RF ~1100, Random subspace ~1020, SMO ~1120, and Dagging ~1160 km². For the 'Low' zone, the areas are: ANN ~1350, RF ~580, Random subspace ~540, SMO ~600, and Dagging ~550 km². For the 'Moderate' zone, the areas are: ANN ~520, RF ~360, Random subspace ~600, SMO ~490, and Dagging ~880 km². For the 'High' zone, the areas are: ANN ~500, RF ~510, Random subspace ~720, SMO ~620, and Dagging ~580 km². For the 'Very high' zone, the areas are: ANN ~1220, RF ~1100, Random subspace ~800, SMO ~860, and Dagging ~510 km².
+
+Fig. 8. The flood susceptible area under different classes of five machine learning models.
+
+can explain the accuracy of the models Under Curve (AUC). The higher the AUC, the higher the model's performance to predict. According to the success rate curve's (training dataset) outcome, the Dagging model (AUC = 0.87) outperformed the RF (AUC = 0.86), RS (AUC = 0.85), SVM (AUC = 0.84), and ANN (AUC = 0.83) models (Fig. 9a). Similarly, the outcomes of the curve of prediction rate by utilizing validation dataset (Fig. 7b) the AUC values for the Dagging model is 0.873, which outweighs the SVM (AUC = 0.86), RS (AUC = 0.83), ANN (AUC = 0.83), and RF (AUC=0.81). The results suggest that though the five FSMs had high prediction accuracy (AUC >0.80), the Dagging model is the best for the flood susceptibility modelling in the case of this study area, as the accuracy of prediction from 0.86 to 0.87 was enhanced by this model
+
+The analytical performance of five flood susceptible models based on the training and testing data using several error measures like root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination  $(R^2)$  are presented in Table 4. The finding clearly showed that the proposed Dagging model had the lowest RMSE= 0.19, lowest MAE=0.08, and very high  $R^2 = 0.84$  compared to other benchmarked models for the training stages (Table 4). In the training stage, the result of statistical performance for the models (five) follows the decreasing order: Dagging>ANN>FR>SVM>RS. Likewise, in the training stage, in the validation phase, the Dagging ensemble model showed the lowest RMSE = 0.18, lowest MAE = 0.08. Very high  $R^2$  = 0.85, which represents a very high level of satisfaction in case of FSMs along with fewer errors. The performance analysis of the five models in the validation stage follows the descending order: Dagging> RF>ANN>SVM> RS. Overall, the Dagging hybrid model had the best performance, while the least achievement is in the case of random subspace for the mapping of flood susceptibility regarding the spatial prediction by the use of both the datasets of training and validation. In fact, for this study, the Dagging model has been found as the most dominant ensemble hybrid model that performed with the higher accuracy of prediction among all models. For the evaluation of the flood susceptibility mapping, this outcome is substantially agreed, and the method of the hybrid ensemble model is appropriate, logical, and reliable, which was confirmed by Tehrany and Jones (2017). By using the Friedman test (Ipe, 1987) and Wilcoxon signed-rank tests (Hong et al., 2019a) at the level of 95% confidence (Tables 5 and 6), the considerable differences among the FSMs of five models were computed. In Tables 5 and 6, the outcomes of the Freidman test and the Wilcoxon signed-rank test are revealed. The findings of Friedman's test showed that the confidence level (or p-value) was < 0.05, which suggested to accept the primary assumption and reject the null hypothesis because, at the level of 95% confidence, the considerable difference among the benchmark models was absent. The values of chi-square ( $\chi^2$ ) and p-value of five
+
+FSMs are far-away from 0.05 and 3.841, respectively, which are the standard values (Table 5). Whether any significant differences are present or not among the five FSMs, cannot be detected by the Friedman test. To overcome this disadvantage, many statistical tests are available. However, we have employed the Wilcoxon signed-rank test to investigate whether statistical differences are present or not between any pairs of FSMs. If the p-value and z value were lesser than the level of significance (<0.05) and higher than the critical values of z (+1.96 and -1.96) respectively, the null hypothesis must be discarded. There were significant differences in the performance of the FSMs. Also, the results of the Wilcoxon signed-rank test (Table 5) indicates that p-value and z value are both distant from the standard values of 0.05 and (from -1.96 to +1.96), respectively for all pairs except for the pairs of Dagging vs. ANN, Dagging vs. SVM, digging vs. RF, Dagging vs. RS, and RF vs. SVM. This implies statistically significant differences among all FSMs. On the other hand, the p values were > 0.05, and the z values exceeded the critical values: therefore, the outcome undoubtedly stated that there are no statistically significant differences in the case of the Dagging model's performance as compared to the other FSMs except for the ANN.
+
+Additionally, to confirm the superiority of the Dagging model over other models, the paired t-test (Altman, 1990) has been performed in this study. It is noted that a significant level of  $\alpha$  is chosen to be p<0.05. If the estimated p-value is less than the significant level  $\alpha$ , the null hypothesis will be rejected that there is no statistical difference between the two paired models. The Dagging model has notably outperformed ANN (p-value =0.0001) and RS (p-value =0.017), except for RF (p-value =0.107) and SVM (p-value=0.263). Thus, based on the statistical t-test, the Dagging model is considered the best fitted for modelling of FSMs in the Teesta River basin.
+
+For investigating the Dagging ensemble model's performance, the ANN, RF, RS, and SVM were regarded in this study. To increase the best performance, the accurate measurement of the parameters of these state-of-arts models should be set carefully. In Table 1, the parameters for these models are shown. By the utilization of the best flood influencing factor, SVM, ANN, RS, RF, Dagging hybrid models were used for modelling flood susceptibility prediction. The criteria of appraisal (RMSE, MAE,  $R^2$ - and AUC) were computed for the comparison among the Dagging model and other benchmarked methods based on training and validation datasets.
+
+## 4. Discussion
+
+The accurate estimation of flood susceptibility is an essential process for safeguarding people 's and developing feasible and effective mitigation measures (Sahana et al., 2020). Therefore, flooding management to
+
+Image /page/12/Figure/2 description: The image displays two ROC (Receiver Operating Characteristic) curve plots, labeled A and B. Both plots graph Sensitivity on the y-axis against 1 - Specificity on the x-axis, with scales from 0.0 to 1.0. Each plot compares the performance of five different models and includes a diagonal red reference line.
+
+Plot A shows five curves that are closely grouped, indicating similar performance. The legend, titled "Source of the Curve", provides the Area Under the Curve (AUC) values for each model:
+- ANN (blue line): AUC = 0.834
+- RF (green line): AUC = 0.863
+- Random subspace (light green line): AUC = 0.859
+- Dagging (purple line): AUC = 0.871
+- SVM (yellow line): AUC = 0.845
+
+Plot B also shows five curves, but they are more spread out than in plot A. The legend provides the following AUC values:
+- ANN (blue line): AUC = 0.835
+- RF (green line): AUC = 0.811
+- Random Subspace (light green line): AUC = 0.839
+- SVM (purple line): AUC = 0.865
+- Dagging (yellow line): AUC = 0.873
+
+Fig. 9. The validation of five flood susceptible models using RoC based on training point (a) and validation point (b).
+
+prevent and decrease the damages caused by flooding is the primary requirement. To do so, the FSM has become the foundation of management across the world (Sahana et al., 2020; Wang et al., 2020). Therefore, researchers always try to use new and robust techniques to obtain very high precision and accurate results, which will be very
+
+helpful to propose flood management plans (Wang et al., 2019a, 2020; Bui et al., 2020a; Nachappa et al., 2020). This is the reason; we also attempted to apply some advanced models, like ANN, SVM, RF, random subspace and newly proposed Dagging algorithms to prepare flood susceptibility maps for Teesta river basin of Bangladesh. It is indispensable
+
+**Table 4**The accuracy assessment of five flood susceptible models for training and testing data using different error measures.
+
+| Methods | ANN      |            | RF       |            | Random subspace |            | SMO      |            | Dagging  |            |
+|---------|----------|------------|----------|------------|-----------------|------------|----------|------------|----------|------------|
+|         | Training | Validation | Training | Validation | Training        | Validation | Training | Validation | Training | Validation |
+| RMSE    | 0.22     | 0.27       | 0.252    | 0.254      | 0.3             | 0.31       | 0.253    | 0.28       | 0.19     | 0.189      |
+| MAE     | 0.121    | 0.16       | 0.172    | 0.173      | 0.2             | 0.203      | 0.152    | 0.17       | 0.081    | 0.084      |
+| R2      | 0.774    | 0.718      | 0.712    | 0.738      | 0.67            | 0.599      | 0.696    | 0.677      | 0.839    | 0.852      |
+
+**Table 5** The results of the Friedman test for five flood susceptible models with the p value = 0.05.
+
+| Models  | Mean ranks | Chi-square | Significance |
+|---------|------------|------------|--------------|
+| ANN     | 1.91       | 218.82     | 0            |
+| RF      | 3.2        |            |              |
+| Rsp     | 3.58       |            |              |
+| SMO     | 3.03       |            |              |
+| Dagging | 3.28       |            |              |
+
+to model the flood susceptibility in this region as the whole country has experienced frequent flooding for a very long time.
+
+Among the applied approaches, ANN, SVM, and RF have been widely used for landslide susceptibility modelling (Hong et al., 2019b; Fang et al., 2020; Wu et al., 2020b), flood susceptibility modelling (Chen et al., 2019b; Chen et al., 2020c; Ali et al., 2020; Khosravi et al., 2019b), gully erosion (Amiri et al., 2019; Tien Bui et al., 2019a; Amiri et al., 2019) and forest fire modelling (Gigović et al., 2019; Tehrany et al., 2019; Pourghasemi et al., 2020b), but scarce studies applied random subspace and Dagging model for different natural hazards prediction including flood susceptibility. Therefore, the performance of the conventional three models and two new models was evaluated and compared by using the AUC value of the ROC curve at training and testing scale and a series of non-parametric statistics. Also, the influences of the flood conditioning parameters were assessed by using the information gain ratio.
+
+The findings showed that the proposed Dagging models outperformed (AUC for training datasets: 0.87, and AUC for testing data sets: 0.873) the other four models (Pham et al., 2017a; Hong, 2018; Yariyan et al., 2020). However, the other four models also performed better as the AUC value achieved more than 0.8 at both training and testing datasets. Therefore, it is reasonable to conclude that Dagging performed better than other models because it is an ensemble machine learning technique which has been efficiently used for obtaining accurate results for predicting any natural hazards or other environmental components (Pham et al., 2017b). Several scholars already applied Dagging models along with other models for predicting the landslide susceptibility and found highly accurate findings (Pham et al., 2017a, 2020a; Nguyen et al., 2019; Kalantar et al., 2020; Nhu et al., 2020). Nguyen et al. (2020) reported that the Dagging model outperformed the other four ensemble machine learning models like random subspace, random subspace, and cascade generalization for groundwater potentiality modelling. Nguyen et al. (2020) accounted that Dagging was more capable than other algorithms for reducing the variance, bias, and noise of the groundwater potential modelling. In the present study, Dagging performed better for training and testing datasets; therefore, it can be stated that the Dagging model is also highly capable of modelling the flood susceptibility. Therefore, it
+
+**Table 6**The comparison of the performances of the ensemble machine learning models for flood susceptibility using the Wilcoxon signed-rank test.
+
+| Pairwise Comparison            | Positive | Negative | Z value | P-value | Significance |
+|--------------------------------|----------|----------|---------|---------|--------------|
+| ANN vs. RF                     | 296      | 37       | -12.27  | 0.05    | Yes          |
+| ANN vs. Ransdom Subspace       | 269      | 64       | -9.53   | 0.05    | Yes          |
+| ANN vs. SMO                    | 225      | 108      | -7.75   | 0.05    | Yes          |
+| ANN vs. Dagging                | 239      | 94       | -6.64   | 0.05    | Yes          |
+| RF vs. Random subspace         | 247      | 86       | -5.93   | 0.05    | Yes          |
+| RF vs. SMO                     | 152      | 181      | 1.83    | 0.01    | No           |
+| RF vs. Dagging                 | 198      | 135      | -0.63   | 0.01    | No           |
+| Random subspace vs. SMO        | 145      | 188      | 2.73    | 0.05    | Yes          |
+| Random subspace vs.<br>Dagging | 143      | 190      | 1.91    | 0.05    | No           |
+| SMO vs. Dagging                | 180      | 153      | -1.67   | 0.05    | No           |
+
+is highly recommended to apply the Dagging model for other natural hazards prediction for obtaining high precision results.
+
+## 5. Conclusion
+
+Two new hybrid ensemble models, namely Dagging and the RS, were used for the first time in the present analysis, along with three machine learning models, e.g., the ANN, the RF, and the SVM, to model the flood susceptibility mapping in the Teesta River basin, Northern Bangladesh. This analysis primarily selected a total of 413 flooding points with twelve variables such as elevation, slope, curvature, aspect, SPI, TWI, STI, LULC, rainfall, river width, TWI, and soil types affecting the flooding. Multi-collinearity diagnostic tests (VIF), IGR, and the correlation matrix of Pearson were utilized to measure the relevance of flood conditioning parameters. The effect of the parameters of flood conditioning was measured by the approaches of feature selection, such as the information gain ratio. To verify the flood prone models based on the training and validation datasets, the Friedman test, Wilcoxon signed-rank test, t-paired test, and the RoC curve have been used. In the case of the Dagging model, the highest versatility and predictive capabilities were found, followed by the RF, the ANN, the SVM and the RS model. In the result of validation step, the implementation of the ROC Curve shows that, relative to other methods, the Dagging model has the maximum performance (AUC = 0.863-training stage; AUC = 0.873-validating stage). The standard of all models for the mapping of flood susceptibility, however, was acceptable and reliable. Dagging is one of the most capable methods for flood-susceptible modeling, the results of the study showed. A total area of 29.62 percent was described as extremely vulnerable to flooding as an ideal model. The biggest downside, however, is that for certain variables, like SPI and land use, the implementation of these models does not take into consideration the modifications over time, since these variables are very complex. Future analysis on the temporal scale will be done depending on the availability of temporal databases of these variables. In addition, by conducting sensitivity analysis on different influential variables, these models can be upgraded. Compared to the other models, the dagging model has advantages, including less candidate parameters, high optimization power, and rapid convergence to prepare susceptibility maps for flush
+
+Flood management strategies have recently been considered a top priority, particularly in Bangladesh, where flash floods occur every year. Other basins and areas, however, have not been tested for flood control measures yet. The present research, therefore, was conducted in the Teesta River basin using some sophisticated machine learning algorithms to provide useful knowledge on the methods to be implemented to help the implementation of successful flash flood mitigation strategies and land-use policy planning by local authorities and other parties.
+
+## **Declaration of Competing Interest**
+
+The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
+
+## Acknowledgements
+
+The authors would like to thank the USGS and other agencies of Bangladesh for providing the data for conducting this research. The authors also would like to acknowledge the Department of Geography, University of Gour Banga, India, for providing necessary facilities to carry out this research. The authors are grateful to the anonymous reviewers for their valuable comments and suggestions to improve this manuscript further. Alban Kuriqi was supported by a PhD scholarship granted by Fundação para a Ciência e a Tecnologia, I.P. (FCT), Portugal,
+
+under the PhD Programme FLUVIO–River Restoration and Management, grant number: PD/BD/114558/2016.
+
+## Appendix A. Supplementary data
+
+Supplementary data to this article can be found online at https://doi.org/10.1016/j.gsf.2020.09.006.
+
+## References
+

chatbot/rag/data/('GLOF Early Warning System Factors_gemini', '.pdf')_extraction.md

@@ -0,0 +1,251 @@
+# A Comprehensive Analysis of Glacial Lake Outburst Flood (GLOF) Triggers for the Development of Advanced Early Warning Systems
+
+The GLOF Hazard Landscape: Formation and Dam Mechanics
+
+## **Defining the GLOF Phenomenon**
+
+A Glacial Lake Outburst Flood (GLOF) is the sudden and catastrophic release of a significant volume of water from a lake fed by glacier melt. These lakes can form at the side, in front, within, beneath, or on the surface of a glacier, impounded by natural dams composed of moraine complexes, glacial ice, or bedrock. GLOFs are distinguished from more common hydrometeorological floods by their defining characteristics: extreme peak discharges that can be several times greater than those of rainfall-induced floods, and an exceptional capacity for erosion and sediment transport. This high energy allows them to mobilize vast quantities of rock, ice, and mud, often transforming into highly destructive debris flows as they travel downstream. These events represent one of the most severe and unpredictable hazards in high-mountain regions, posing a significant threat to human lives, livelihoods, and critical infrastructure such as roads, bridges, and hydropower facilities. The rapid onset and immense destructive power of GLOFs necessitate the development of specialized and highly reliable Early Warning Systems (EWS) to mitigate their devastating impacts.
+
+Glacial Lake Genesis in a Warming Climate
+
+The formation and proliferation of glacial lakes are a direct and highly visible consequence of global climate change.<sup>5</sup> As atmospheric temperatures rise, glaciers worldwide are retreating at an accelerated rate.<sup>7</sup> This process of deglaciation uncovers over-deepened basins and depressions in the landscape, often located between the retreating glacier terminus and a proglacial moraine—a ridge of unconsolidated debris deposited by the glacier at its furthest extent.<sup>1</sup> These depressions subsequently fill with meltwater, giving birth to new glacial lakes.<sup>1</sup>
+
+This process is not static; it is a dynamic feedback loop. The formation of a glacial lake can itself accelerate further ice loss from the parent glacier through calving (the breaking off of ice chunks from the glacier front) and subaqueous melting at the ice-water interface. Consequently, as the climate continues to warm, existing lakes expand and new lakes form, leading to a rapid global increase in the number, total area, and stored volume of glacial water. This escalates the potential energy available for release, inherently increasing the potential magnitude and frequency of future GLOF events. In High Mountain Asia, for instance, a warming rate of 0.32 °C per decade is driving this widespread expansion of glacial lakes, amplifying the regional GLOF hazard.
+
+## **Dam Typology and Associated Failure Characteristics**
+
+The nature of a GLOF event—its likelihood, trigger mechanism, and recurrence potential—is fundamentally dictated by the type of dam impounding the glacial lake. Understanding the distinct mechanical properties and failure modes of the two primary dam types, moraine and ice, is therefore a prerequisite for designing an effective and appropriately targeted EWS.
+
+### **Moraine-Dammed Lakes**
+
+Moraine-dammed lakes are the most common type in many of the world's most vulnerable mountain ranges, including the Hindu Kush Himalaya. The dams are composed of moraine: a heterogeneous and unconsolidated mixture of boulders, gravel, sand, and fine sediment left behind by a retreating glacier. These structures are inherently weak, poorly sorted, and often highly permeable, making them susceptible to various failure mechanisms. Furthermore, many moraine dams contain a core of buried glacial ice, which can melt over time, compromising the dam's structural integrity from within.
+
+The failure of a moraine dam is typically a singular, catastrophic event. Mechanisms such as overtopping by a large wave or excessive inflow can initiate a process of "breach incision," where flowing water rapidly erodes a channel through the loose dam material.<sup>4</sup> This process creates a positive feedback loop—a deeper channel allows a higher discharge, which in turn accelerates erosion—leading to a complete and often explosive failure of the dam structure.<sup>4</sup> Because the dam is often destroyed in the process, the threat of repeated GLOFs from the same lake is generally low.<sup>4</sup> The primary hazard is therefore a single, high-magnitude flood.
+
+### Ice-Dammed Lakes
+
+Ice-dammed lakes form when a glacier itself blocks the natural drainage of a valley.<sup>17</sup> These lakes are responsible for a large number of GLOF events globally, particularly in regions like Alaska and Patagonia.<sup>4</sup> Unlike moraine dams, ice dams do not typically fail through erosional breaching. Instead, they are subject to unique failure mechanisms driven by the physical properties of ice. The most common mechanism is flotation, which occurs when the water level in the lake reaches approximately 90% of the ice dam's height. Due to the density difference between water () and ice (), the buoyant force of the water can physically lift the glacier off its bed, allowing water to escape through subglacial tunnels.<sup>4</sup>
+
+A critical characteristic of ice-dammed GLOFs is their potential for recurrence. As the lake drains and the water level drops below the 90% threshold, the glacier settles back onto its bed, and the subglacial conduits can close up, resealing the dam.<sup>4</sup> This allows the lake to refill, setting the stage for future outburst events.<sup>17</sup> Ice-dammed lakes can therefore produce cyclic GLOFs, posing a persistent, long-term hazard that may recur annually or over several years.<sup>4</sup>
+
+The fundamental distinction between the failure modes of moraine and ice dams necessitates tailored EWS philosophies. An EWS designed for a moraine-dammed lake, which is prone to a single, catastrophic failure, must prioritize the detection of immediate, high-energy triggers like landslides or extreme rainfall that can initiate a breach. The focus is on capturing the precursors to a singular, high-consequence event. Conversely, an EWS for an ice-dammed lake, which can fail and reform cyclically, requires a design focused on long-term, continuous monitoring of the recurring hazard cycle. This system must track the slower buildup of hydrostatic pressure and lake level over multiple seasons to predict the next drainage event. This critical difference impacts the required sensor longevity, power management strategies, data analysis algorithms, and the nature of community preparedness and response planning.
+
+## **Meteorological Drivers as Primary GLOF Triggers**
+
+Meteorological conditions are a fundamental driver of GLOF activity, acting as both long-term conditioning factors that increase overall hazard levels and as acute, short-term triggers that can initiate a dam failure. An effective EWS must be capable of monitoring and interpreting these atmospheric influences across different timescales.
+
+### The Role of Temperature
+
+Rising atmospheric temperature is the primary climate-driven force increasing GLOF risk globally. Its influence is multifaceted, affecting the entire glacial environment from the ice mass itself to the stability of the surrounding landscape.
+
+Long-Term Conditioning: Sustained warming trends are the principal cause of glacier mass loss, which directly leads to the formation of new glacial lakes and the expansion of existing ones. This relentless increase in the number and volume of lakes inherently elevates the potential GLOF hazard across a region. Beyond the lakes themselves, rising temperatures have a profound impact on the geotechnical stability of high-mountain terrain. As temperatures increase, permafrost—perennially frozen ground that acts as a cement binding rock and soil together—begins to thaw and degrade. This process significantly reduces the stability of rockwalls and moraine slopes surrounding glacial lakes, "pre-conditioning" the landscape and increasing the probability of rockfalls, avalanches, and landslides. These mass movements are a primary GLOF trigger, as detailed in Section 4.
+
+**Short-Term Triggering:** Superimposed on the long-term warming trend are short-term extreme temperature events or heatwaves. These periods of intense warmth can cause a rapid and dramatic acceleration of snow and ice melt.<sup>7</sup> The resulting surge in meltwater runoff can overwhelm a glacial lake's natural outflow capacity, causing the water level to rise rapidly over a period of hours to days.<sup>2</sup> This rapid increase in lake volume can directly trigger an outburst flood through the mechanism of dam overtopping, particularly in moraine-dammed lakes.<sup>7</sup> Analysis of historical GLOF events indicates a strong correlation, with outbursts being significantly more likely to occur during anomalously warm periods or during abrupt climatic transitions from cold to warm conditions.<sup>18</sup>
+
+### **Intense Precipitation and Snowmelt Dynamics**
+
+While temperature sets the long-term stage, intense precipitation events are a common and powerful proximate trigger for GLOFs. The link between climate change and such events is well-established; a warmer atmosphere can hold more moisture (approximately 7% more for every 1°C of warming), which increases the potential for heavier downpours and more extreme rainfall events.<sup>20</sup>
+
+**Mechanism of Failure:** An episode of intense, heavy rainfall or an event combining rain with rapid snowmelt can deliver a massive volume of water into a glacial lake's catchment area in a very short time.<sup>2</sup> This sudden influx causes the lake level to rise swiftly, potentially leading to the overtopping of the dam crest.<sup>4</sup> This mechanism is particularly critical for moraine dams due to their unconsolidated nature.
+
+The Positive Feedback Loop of Breach Incision: The process does not stop at simple overtopping. Once water begins to flow over the crest of a moraine dam, its erosive power starts to cut a channel into the loose material.<sup>4</sup> This initial incision creates a pathway for a greater volume of water to be discharged. This increased flow possesses even more erosive energy, which deepens and widens the breach channel further. This self-reinforcing, or "positive feedback," cycle can cause the dam to fail catastrophically in a matter of hours, releasing the entire volume of the lake in a devastating flood wave.<sup>4</sup>
+
+**Link to Weather Systems:** In many high-mountain regions, these extreme precipitation events are associated with large-scale meteorological patterns. In the Himalayas, for example, the summer monsoon can bring vigorous, moisture-laden systems that result in orographic lifting—where air is forced upward by mountain slopes, causing it to cool and release its moisture as extremely heavy rainfall. The 2013 Kedarnath disaster in the Indian Himalaya was a tragic example of a GLOF from Lake Chorabari triggered by exceptionally heavy rainfall.
+
+The interplay between these meteorological factors creates a classic cascading hazard chain that a sophisticated EWS must be designed to monitor and interrupt. The causal sequence is not a simple linear progression from rain to flood. It begins with anthropogenic climate change, which leads to a warmer atmosphere capable of holding and delivering more moisture. This increases the frequency and intensity of extreme precipitation events. When such an event occurs over a glacial catchment, it causes a rapid rise in the lake level, leading to dam overtopping. This, in turn, initiates the critical positive feedback loop of breach incision, culminating in a GLOF. Simultaneously, the same long-term warming that fuels these storms is also degrading permafrost on adjacent slopes, increasing the likelihood of a landslide trigger. Therefore, an effective EWS cannot rely on a single parameter like a rainfall gauge. It must adopt a holistic approach, integrating meteorological forecasts to predict intense precipitation, temperature sensors to monitor melt rates and long-term warming trends, and geotechnical sensors to assess slope stability, recognizing that these factors are deeply interconnected and can amplify one another's hazardous effects.
+
+### **Atmospheric Pressure: An Assessment of Influence**
+
+The user query specifically requested information on the role of atmospheric pressure as a GLOF trigger. A thorough review of the scientific literature indicates that while atmospheric pressure is an integral component of weather systems, it is not considered a direct physical trigger for the failure of a glacial dam.<sup>1</sup>
+
+The hazardous weather conditions associated with GLOFs, such as intense storms, are indeed characterized by low-pressure systems.<sup>19</sup> However, the dam failure is caused by the physical impacts of the associated phenomena—namely, the mass of water from heavy precipitation or the rapid influx from temperature-driven melt—rather than the change in barometric pressure itself. The direct force exerted by atmospheric pressure changes on the lake surface or dam structure is negligible compared to the immense forces of hydrostatic pressure from the water column or the kinetic energy of an avalanche.
+
+For the purpose of designing an EWS, monitoring atmospheric pressure is highly valuable, but its role is predictive rather than causative. As a key parameter measured by any standard Automatic Weather Station (AWS), barometric pressure trends are crucial for forecasting the approach of storm systems that could bring triggering conditions (i.e., heavy precipitation). Therefore, it serves as an important input for meteorological forecasting models that can provide advance warning of hazardous weather, but it should not be mistaken for a direct physical trigger of dam failure. The critical pressure that must be monitored for direct risk assessment is the **hydrostatic pressure** exerted by the lake's water, which is a function of water depth.
+
+## Lake-Centric Triggers: Water Level, Pressure, and Dam Integrity
+
+While external meteorological and geotechnical events often act as the final trigger, the inherent state of the glacial lake and its dam determines the system's underlying susceptibility to an outburst. Factors such as the water level, the resulting hydrostatic pressure, and the long-term structural integrity of the dam are critical parameters that an EWS must continuously monitor.
+
+### **Hydrostatic Pressure and Lake Level Thresholds**
+
+The most direct and quantifiable threat posed by a growing glacial lake is the increase in hydrostatic pressure. As a lake fills with meltwater, its volume and depth increase, exerting a progressively greater force on the inner face of the impounding dam.<sup>1</sup> A GLOF can occur through "dam self-destruction" when this hydrostatic pressure—the force exerted by the weight of the water column—exceeds the lithostatic pressure and shear strength of the dam material, causing a structural failure without a dynamic external trigger.<sup>27</sup> This risk is amplified in deep lakes and where the dam has been weakened by other long-term degradation processes.<sup>27</sup>
+
+Crucially for the design of a quantitative EWS, empirical research has identified specific, monitorable thresholds that correlate with a high probability of an outburst flood, particularly for ice-dammed lakes. Studies of recurring GLOFs have shown that an outburst becomes highly likely when the lake's **normalized depth (n')**—the ratio of the water depth to the dam height—exceeds a value of **0.60**.<sup>28</sup> This critical depth corresponds to a typical water pressure on the dam face of approximately **510 kPa**.<sup>28</sup>
+
+These empirically derived values are of paramount importance for EWS development. They provide a concrete, physics-based foundation for moving beyond qualitative hazard assessment to a quantitative, automated alarm system. A monitoring system equipped with water level and pressure sensors can be programmed with these thresholds to automatically escalate alert levels. For example, a normalized depth approaching 0.60 or a pressure nearing 510 kPa could trigger a "Yellow Alert" or "Warning" status, prompting heightened surveillance and preparedness activities. A continued rise beyond these thresholds could trigger a "Red Alert" or "Evacuation" order. These quantifiable parameters represent the most direct and reliable indicators of imminent risk from pressure-induced dam failure.
+
+### **Long-Term Dam Degradation Processes**
+
+The structural integrity of a moraine dam is not static; it can degrade over time through slow, often invisible processes. These long-term mechanisms weaken the dam, reducing its ability to withstand hydrostatic pressure and making it susceptible to failure from a much smaller or less intense trigger event.
+
+**Melting of Buried Ice:** Many moraine dams, particularly those formed during the Little Ice Age, contain significant cores of buried glacial ice. This ice, insulated by the overlying debris, can persist for centuries. However, as regional atmospheric temperatures rise due to climate
+
+change, this buried ice begins to melt.<sup>2</sup> The melting of the ice core causes a loss of volume and structural support within the dam, leading to subsidence, the formation of internal voids, and a general weakening of the entire structure.<sup>4</sup> This internal degradation is a critical long-term destabilization mechanism that can ultimately lead to the dam's collapse.<sup>12</sup>
+
+**Piping and Seepage:** Moraine dams are composed of unconsolidated materials with varying permeability. This allows water from the lake to seep into and through the dam structure. This seepage can create preferential flow paths within the dam. If the velocity of this internal flow is sufficient, it can begin to erode and transport fine-grained sediment from the dam's core, a process known as "piping". Over time, this internal erosion can create large conduits or "pipes" through the dam, progressively weakening its structure until it can no longer support the pressure of the impounded water, leading to a sudden and catastrophic failure.
+
+The existence of these quantifiable pressure and depth thresholds provides a powerful tool for building an automated EWS. However, the "hidden" nature of long-term dam degradation processes like ice core melt and internal piping reveals a potential vulnerability in a monitoring strategy that relies solely on surface-level measurements. A dam that has been significantly weakened from within might fail at a hydrostatic pressure *below* the established 510 kPa threshold, meaning an alarm based on that value alone might come too late or not at all. This implies that a truly comprehensive EWS for a high-risk, moraine-dammed lake must be informed by a deeper understanding of the dam's internal composition and integrity. This requires a preliminary geotechnical assessment, potentially using techniques like Ground Penetrating Radar (GPR) to map the presence and extent of any buried ice core.<sup>32</sup> The results of such a survey can provide a more accurate picture of the true risk level and allow for the calibration of alarm thresholds that are tailored to the specific structural condition of the dam, closing a critical blind spot in the monitoring strategy.
+
+## Dynamic and Geotechnical Triggers: Mass Movements and Seismicity
+
+While meteorological and lake-centric factors create and condition the GLOF hazard, the most common and often most violent triggers are dynamic geotechnical events. These triggers are characterized by their rapid onset, typically occurring over seconds to minutes, and their immense energy, which can overwhelm even a relatively stable dam.
+
+Mass Movement Impacts: Avalanches, Landslides, and Displacement
+
+### Waves
+
+The most frequently documented trigger for GLOFs, particularly from moraine-dammed lakes, is the sudden impact of a large mass movement into the lake basin.<sup>15</sup> These events include ice avalanches from hanging glaciers, rockfalls from destabilized valley walls, or debris-laden landslides.<sup>4</sup>
+
+**Mechanism of Displacement Wave Generation:** When a substantial volume of ice, rock, or debris rapidly enters a lake, it displaces a massive volume of water, generating a high-energy impulse wave, also referred to as a seiche or a tsunami-like wave. The size of this wave is a function of the volume, velocity, and geometry of the incoming mass. These waves can be exceptionally large, with observed or modeled heights reaching tens of meters. <sup>35</sup>
+
+Frequency and Dam Failure: This trigger mechanism is remarkably common. In the Himalayas, displacement waves from ice or rock avalanches are responsible for nearly half of all recorded moraine-dam failures. <sup>12</sup> Globally, it is the most common dynamic cause of GLOFs. <sup>34</sup> The resulting displacement wave can easily overtop the freeboard (the height of the dam crest above the water level). This initial overtopping can be powerful enough to cause immediate dam rupture, but more commonly, it initiates the catastrophic breach incision process. The initial surge of water over the dam crest begins to erode the unconsolidated moraine, and the subsequent reflected waves continue to overtop the dam, rapidly carving a breach that leads to the complete and violent drainage of the lake. <sup>4</sup>
+
+The prevalence of mass movements as the primary GLOF trigger has a profound implication for EWS design: a GLOF EWS cannot be merely lake-focused; it must be *catchment-focused*. A monitoring system that only includes sensors at the lake itself, such as water level gauges, would only detect the *consequence* of the trigger—the arrival of the displacement wave at the dam. This would provide, at best, only seconds to minutes of warning time, which is insufficient for any meaningful downstream evacuation. To provide a true *early* warning, the system must be capable of detecting the precursor to the trigger: the instability of the slope *before* it fails. This fundamentally expands the spatial scope and technological complexity of the required sensor network, necessitating the monitoring of the stability of surrounding slopes, hanging glaciers, and permafrost zones with instruments like geophones, extensometers, or remote sensing techniques like InSAR.
+
+### Seismic Activity: Direct and Indirect Triggering Mechanisms
+
+In tectonically active mountain ranges like the Himalayas and the Andes, earthquakes are
+
+another significant GLOF trigger. Seismic events can initiate dam failure through both direct and indirect pathways.
+
+**Direct Impact:** Intense ground shaking during an earthquake can directly compromise the structural integrity of a moraine dam. The seismic waves can cause liquefaction of saturated sediments within the dam or induce cracking and slumping, leading to a rapid structural failure.<sup>2</sup>
+
+**Indirect Impact (More Common):** More frequently, earthquakes act as an indirect trigger by initiating the very mass movements described in the preceding section.<sup>4</sup> The seismic energy can be the final catalyst that destabilizes an already precarious slope, causing a massive ice or rock avalanche to plunge into the lake, thereby generating a displacement wave.<sup>39</sup>
+
+Post-Earthquake Hazard Amplification: The GLOF hazard in a region can be dramatically amplified for months or even years following a major earthquake. A large seismic event can trigger thousands of landslides throughout a mountain range, depositing vast quantities of loose, unconsolidated sediment onto valley floors and slopes. This material then becomes readily available for entrainment by any subsequent flood. A GLOF that might have otherwise been a relatively clean water flood can scour and incorporate this loose debris, transforming into a much larger, denser, and more destructive debris flow. This means that even a small glacial lake, previously considered low-risk, could produce a catastrophic debris flow if its outburst path is now loaded with earthquake-induced landslide deposits. Hazard assessments must therefore be re-evaluated in the aftermath of any significant regional earthquake.
+
+A second, paradigm-shifting realization for EWS design comes from the analysis of the seismic signals generated by GLOFs themselves. Detailed studies of past events, such as the 1994 GLOF from Lugge Tsho in Bhutan, have revealed that the flood's highly turbulent flow and the transport of boulders and sediment along the riverbed generate a continuous, high-frequency seismic signal. This GLOF-induced ground vibration is distinct from the short, sharp signal of an earthquake and can be detected by seismometers located up to 100 kilometers away. By using arrays of seismometers and applying signal processing techniques like cross-correlation, it is possible to not only detect the onset of a GLOF but also to track the location of the flood front as it propagates down the valley in real-time. This transforms the role of seismometers from merely being detectors of a potential trigger (an earthquake) to being a powerful, remote tool for detecting and tracking the hazard itself. This provides a robust, wide-area monitoring capability that is less vulnerable to being destroyed by the flood and can provide hours of warning time to downstream communities.
+
+The following table synthesizes the primary GLOF triggers, their physical mechanisms, and the key parameters that can be monitored by an EWS.
+
+| Trigger<br>Category           | Specific<br>Trigger                                                                                                   | Physical<br>Mechanism                                                                                                               | Primary<br>Monitorable<br>Parameters                                                         | Secondary/<br>Proxy<br>Parameters                                                  | Typical<br>Onset Time                                                       |
+|-------------------------------|-----------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------|-----------------------------------------------------------------------------|
+| Meteorological                | Intense<br>Rainfall /<br>Rapid<br>Snowmelt                                                                            | Dam overtoppin<br>g leading to<br>positive<br>feedback of<br>breach<br>incision.4                                                   | Precipitatio<br>n intensity<br>(); Rate of<br>lake level<br>rise ()                          | Air<br>temperatur<br>e; Snow line<br>elevation;<br>Barometric<br>pressure<br>trend | Hours to<br>Days                                                            |
+|                               | Sustained<br>High<br>Temperatures                                                                                     | Accelerated<br>glacier and<br>snow melt<br>increasing<br>lake<br>volume;<br>Permafrost<br>degradatio<br>n<br>weakening<br>slopes.10 | Air<br>temperatur<br>e<br>anomalies;<br>Cumulative<br>positive<br>degree<br>days             | Solar<br>radiation;<br>Glacier<br>surface<br>velocity                              | Days to<br>Weeks<br>(trigger);<br>Years to<br>Decades<br>(conditionin<br>g) |
+| Lake-Centric                  | Hydrostatic<br>Pressure                                                                                               | Water<br>pressure<br>exceeds<br>the<br>structural<br>strength<br>(lithostatic<br>pressure)<br>of the dam,<br>causing<br>failure.27  | Lake water<br>level/depth;<br>Sub-surfac<br>e water<br>pressure<br>(hydrostatic<br>pressure) | Lake<br>volume;<br>Normalized<br>depth ()                                          | Weeks to<br>Months                                                          |
+|                               | Dam<br>Degradation                                                                                                    | Melting of<br>buried ice<br>core or<br>internal<br>erosion                                                                          | Dam<br>surface<br>temperatur<br>e; Seepage<br>rate/turbidit<br>y                             | Ground-pe<br>netrating<br>radar (GPR)<br>survey<br>results (for                    | Years to<br>Decades                                                         |
+|                               |                                                                                                                       | (piping) weakens the dam structure, lowering its failure threshold. 4                                                               | y; Surface<br>subsidence<br>(via InSAR)                                                      | ice core<br>presence)                                                              |                                                                             |
+| Dynamic /<br>Geotechni<br>cal | Ice/Rock<br>Avalanche<br>or<br>Landslide                                                                              | Rapid mass entry into lake generates a large displaceme nt (seiche) wave, causing dam overtoppin g and breach. 1                    | Ground vibration (geophone/ seismomete r); Acoustic signal                                   | Slope deformatio n rate (InSAR, extensomet ers); Rock/ice temperatur e             | Seconds to<br>Minutes                                                       |
+|                               | Earthquake  Ground shaking directly compromis es dam stability or indirectly triggers mass movements into the lake. 7 |                                                                                                                                     | Seismic P-<br>and S-wave<br>arrival; Peak<br>Ground<br>Acceleratio<br>n (PGA)                | Regional<br>seismic<br>hazard<br>maps                                              | Seconds                                                                     |
+
+## Framework for a GLOF Early Warning System (EWS)
+
+The complex, multi-causal nature of GLOFs dictates that a reliable EWS cannot be based on a single sensor or parameter. Instead, a robust framework must be built upon an integrated,
+
+multi-sensor approach that provides detection, verification, and redundancy.<sup>36</sup> This framework should be structured around the four internationally recognized components of a modern EWS: Risk Knowledge, Monitoring and Warning, Dissemination and Communication, and Response Capability.<sup>36</sup> The following sections focus on the technological aspects of the Monitoring and Warning component.
+
+### A Multi-Sensor, Integrated Approach
+
+The optimal EWS design is a hierarchical, multi-scale system that intelligently allocates monitoring resources based on risk. This approach avoids the financially and logistically prohibitive task of installing dense in-situ sensor networks at every one of the thousands of glacial lakes in a region.<sup>1</sup>
+
+The framework begins with wide-area **surveillance** using freely available, moderate-resolution satellite data (e.g., Landsat, Sentinel) to create and regularly update comprehensive glacial lake inventories. This regional-scale analysis identifies "hotspots"—lakes that are expanding rapidly, are situated beneath potentially unstable slopes, or exhibit other characteristics associated with high risk. These identified hotspots are then prioritized for more focused and resource-intensive **intensive care** monitoring. This may involve tasking high-resolution commercial satellites for more detailed analysis or, for the most critical and potentially dangerous glacial lakes, deploying a ground-based network of in-situ sensors. This tiered strategy ensures that the most sophisticated monitoring efforts are directed where the risk is greatest, creating a scalable and economically viable national or regional EWS.
+
+### **In-Situ Monitoring Technologies and Parameters**
+
+In-situ, or ground-based, sensors are the cornerstone of a real-time GLOF alarm system. They provide high-frequency, continuous data directly from the hazard source, enabling the detection of rapid-onset triggers and the tracking of critical parameters against predefined alarm thresholds.
+
+- Automatic Weather Stations (AWS): An AWS is fundamental for monitoring meteorological triggers. It should be equipped to measure air temperature, precipitation (type and intensity), solar radiation (a key input for melt models), wind speed, and barometric pressure (for weather system forecasting).<sup>50</sup>
+- Water Level and Pressure Sensors: These are the most direct way to monitor the state
+
+of the lake. Technologies include non-contact sensors like ultrasonic or radar gauges, which are mounted above the water surface, and submersible pressure transducers placed in the lake.<sup>32</sup> They provide continuous data on the lake's water level, allowing for the calculation of the rate of rise and the real-time tracking of hydrostatic pressure against critical thresholds.<sup>53</sup>
+
+- Geophones and Acoustic Sensors: Geophones are sensitive instruments that detect
+  ground vibrations. Deployed on or near the moraine dam and on surrounding slopes, they
+  can detect the high-frequency seismic signals generated by rockfalls, ice avalanches, or
+  the initial dam breach, providing an immediate warning of a dynamic trigger event.<sup>36</sup>
+  Acoustic sensors can similarly detect the audible sound of these events.
+- Seismic Sensors (Seismometers and Accelerometers): These instruments serve a dual purpose. A local network can detect regional earthquakes that may act as triggers.<sup>32</sup>
+   More significantly, a regional network of seismometers can be used to detect the distinct, continuous ground vibrations generated by the GLOF itself as it moves downstream, providing a powerful tool for remote detection and tracking.<sup>41</sup>
+- Ground-Based Cameras and Radar: Automatic, solar-powered cameras can provide invaluable visual information, capturing images of the glacier terminus to monitor calving activity, the dam face to identify seepage or slumping, and the lake surface to detect waves.<sup>36</sup> Ground-based interferometric radar can be used for high-precision monitoring of movement on the dam or critical slopes.
+
+### Remote Sensing and Data Integration for Comprehensive Monitoring
+
+Remote sensing from satellite platforms provides the essential broad-area context and long-term monitoring capability that is impossible to achieve with in-situ sensors alone. Much of this data is freely and publicly available, providing a cost-effective foundation for regional risk assessment.
+
+- Optical Satellite Imagery: Platforms like the Landsat series (USGS) and Sentinel-2 (ESA) provide multispectral imagery that is crucial for creating lake inventories and monitoring long-term changes in lake area, number, and color (which can indicate turbidity and sediment load).<sup>3</sup> The increasing application of machine learning and deep learning algorithms to this data allows for the automated and rapid delineation of thousands of lakes, making large-scale analysis feasible.<sup>54</sup>
+- Synthetic Aperture Radar (SAR): SAR satellites, such as Sentinel-1, are indispensable for GLOF monitoring due to their ability to acquire images through clouds and at night.<sup>7</sup> This overcomes the primary limitation of optical imagery in persistently cloudy mountain regions. SAR data is used to monitor changes in lake area and can detect the presence of floating ice.<sup>56</sup> An advanced technique, Interferometric SAR (InSAR), uses pairs of SAR
+
+- images to detect and measure ground surface deformation with millimeter-level precision, making it an exceptionally powerful tool for monitoring the stability of moraine dams and surrounding slopes for precursor movements.<sup>55</sup>
+- Digital Elevation Models (DEMs): DEMs provide the 3D topographic data necessary for GLOF hazard assessment. They are used to characterize the geometry of the lake basin and dam, estimate lake volume, and, crucially, to run hydrodynamic models that simulate potential flood paths, inundation depths, and travel times for downstream communities.<sup>46</sup>
+   Several global and regional DEMs are freely available for this purpose.
+
+The following table provides a comparative overview of key sensor technologies, detailing the parameters they measure, their operational characteristics, and their availability. This information is intended to guide the selection of an appropriate suite of technologies for a comprehensive EWS.
+
+| Technology Type                       | Specific Instrument                                           | Parameter(s) Measured                               | Relevance to GLOF Trigger                                                                         | Advantages                                                                        | Disadvantages/Constraints                                                              | Data Availability                                                                   |
+|---------------------------------------|---------------------------------------------------------------|-----------------------------------------------------|---------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------|----------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------|
+| <b>In-Situ<br/>Sensor</b>             | Radar<br>Water<br>Level<br>Sensor                             | Lake<br>water<br>level, rate<br>of rise             | Monitors<br>hydrostatic<br>pressure,<br>overtopping risk52                                        | High<br>temporal<br>resolution,<br>real-time<br>data,<br>non-contact              | High<br>cost,<br>difficult<br>installation/maintenance,<br>power<br>requirements       | Commercial<br>Purchase                                                              |
+|                                       | Submersible<br>Pressure<br>Transducer                         | Hydrostatic<br>pressure,<br>water<br>temperature    | Direct<br>measurement of<br>pressure<br>on dam53                                                  | High<br>accuracy,<br>continuous data                                              | In-water<br>deployment risk<br>(icebergs,<br>currents),<br>cable<br>damage             | Commercial<br>Purchase                                                              |
+|                                       | Geophone                                                      | High-frequency                                      | Detects<br>rock/ice                                                                               | Highly<br>sensitive                                                               | Can be<br>sensitive                                                                    | Commercial                                                                          |
+|                                       |                                                               | ground<br>vibration                                 | avalanches, dam<br>breach<br>onset36                                                              | to impact<br>events,<br>provides<br>immediat<br>e trigger<br>alert                | to noise<br>(wind,<br>animals),<br>requires<br>careful<br>placemen<br>t                | Purchase                                                                            |
+|                                       | Seismom<br>eter                                               | Broad-sp<br>ectrum<br>ground<br>motion              | Detects<br>earthqua<br>ke<br>triggers<br>and<br>tracks<br>GLOF<br>propagati<br>on41               | Can<br>detect<br>events<br>remotely<br>(100+<br>km),<br>wide-are<br>a<br>coverage | Requires<br>specializ<br>ed<br>analysis,<br>high<br>initial<br>cost                    | National/<br>Global<br>Networks<br>(often<br>public);<br>Commerc<br>ial<br>Purchase |
+|                                       | Automati<br>c<br>Weather<br>Station<br>(AWS)                  | Temperat<br>ure,<br>precipitat<br>ion,<br>radiation | Monitors<br>meteorol<br>ogical<br>triggers<br>and melt<br>condition<br>s50                        | Provides<br>essential<br>data for<br>forecasti<br>ng and<br>melt<br>modeling      | Requires<br>regular<br>maintena<br>nce/calibr<br>ation in<br>harsh<br>environm<br>ents | Commerc<br>ial<br>Purchase                                                          |
+| Remote<br>Sensing                     | Optical<br>Satellite<br>(e.g.,<br>Sentinel-<br>2,<br>Landsat) | Lake<br>area,<br>color,<br>snow<br>cover            | Long-ter<br>m<br>monitorin<br>g of lake<br>expansio<br>n and<br>catchme<br>nt<br>condition<br>s45 | Wide<br>area<br>coverage<br>, long<br>historical<br>archive,<br>free data         | Obscure<br>d by<br>clouds,<br>limited<br>use in<br>winter/da<br>rkness                 | Freely<br>available<br>(ESA/US<br>GS)                                               |
+|                                       | SAR<br>Satellite<br>(e.g.,                                    | Lake<br>area,<br>surface                            | All-weath<br>er<br>lake/slope                                                                     | All-weath<br>er,<br>day/night                                                     | Complex<br>data<br>processin                                                           | Freely<br>available<br>(ESA)                                                        |
+| Sentinel-1)                           | deformation (InSAR), surface texture                          | monitoring7                                         | capability, detects subtle ground movement                                                        | geometric distortions in steep terrain                                            |                                                                                        |                                                                                     |
+| High-Resolution Commercial Satellites | Detailed imagery of dam, slopes, glacier terminus             | Detailed site inspection, change detection          | Very high spatial detail (<1 m)                                                                   | High cost per image, requires specific tasking                                    | Commercial Purchase                                                                    |                                                                                     |
+
+## **EWS in Practice: Insights from Global Case Studies**
+
+The theoretical principles of EWS design are best understood through their application in real-world, high-risk environments. Case studies from around the globe provide invaluable lessons on successful strategies, technological challenges, and the critical importance of integrating technology with institutional and community frameworks.
+
+### Kyagar Glacier, China
+
+The EWS implemented at the ice-dammed Kyagar Lake in the Karakoram Mountains represents a model of an effective, integrated system designed for a remote and recurring hazard. <sup>50</sup> The system combines remote sensing with a network of terrestrial stations. SAR satellite imagery is used to monitor the seasonal filling of the lake and estimate its volume. <sup>50</sup> This is complemented by a solar-powered, automatic observation station located near the ice dam, which records meteorological data and takes daily images of the glacier terminus. <sup>50</sup> Crucially, the system includes two additional monitoring stations with radar water level sensors located far downstream on the Yarkant River. This distributed network provides exceptional warning lead times: an outburst detected at the glacier station provides 34–36 hours of warning, while the downstream stations provide 22 hours and 7.5 hours of warning, respectively. <sup>50</sup> When a GLOF is detected, an automated alarm is sent via satellite to the mobile phones of authorities, enabling timely emergency response. The Kyagar system demonstrates
+
+the power of a layered monitoring network to maximize warning time.
+
+### Laguna 513, Peru
+
+The EWS for Laguna 513 in the Cordillera Blanca was specifically designed to address the region's primary trigger: ice and rock avalanches from the steep slopes of Mount Hualcán.<sup>36</sup> The core of the monitoring component consists of four geophones installed on the moraine dam to detect the ground vibrations from an avalanche impact, and two cameras providing a constant visual feed of the glacier face and the dam.<sup>36</sup> This data is transmitted in near real-time to a data center in the downstream city of Carhuaz. The system is built on a clear, tiered warning protocol with defined thresholds for geophone measurements that trigger different alert levels (yellow, orange, red).<sup>36</sup> This case highlights a system tailored to a specific, dominant trigger mechanism. It also underscores the importance of a well-defined institutional framework; by law, only the mayor can authorize an evacuation, so the EWS is designed to provide actionable information to a specific decision-maker.<sup>36</sup>
+
+### Thorthormi Lake, Bhutan
+
+The experience at Thorthormi Lake in the Bhutan Himalayas serves as a critical lesson in the operational challenges of maintaining an EWS in one of the world's most extreme environments. During a GLOF event in October 2023, the Automatic Water Level Station (AWLS) at the lake was completely damaged and washed away by the flood. <sup>59</sup> This event highlights the extreme vulnerability of in-situ equipment and underscores the absolute necessity of designing hardened, robust systems and ensuring redundancy. If a single point of failure (the destruction of the primary sensor) can disable the entire warning capability, the system is flawed. Furthermore, a study of a 2019 GLOF event at the same location revealed that the first warning for many residents came not from the official EWS, but from natural indicators—the unusually loud sound of the river and ground vibrations. <sup>60</sup> This demonstrates that community-based knowledge and observation are vital components of a resilient warning system and must be formally integrated with technical monitoring. The recommendations following the 2023 event—to relocate the AWLS, add more sensors for redundancy, and enhance community planning—point to a continuous process of adaptation and improvement. <sup>59</sup>
+
+### Imja Tsho and Tsho Rolpa, Nepal
+
+The cases of Imja Tsho and Tsho Rolpa in Nepal illustrate the long-term evolution and challenges of GLOF risk management. Tsho Rolpa was one of the first lakes in the region to have a technical EWS installed in the late 1990s, which included water level sensors and a series of 19 downstream siren towers. However, over two decades later, this system has reportedly fallen into disrepair due to a lack of sustained funding for maintenance, leaving the community vulnerable once again. This is a stark reminder that the financial commitment for an EWS cannot end with its installation; a long-term operational and maintenance budget is non-negotiable. In contrast, a pilot project at Imja Lake has demonstrated the potential of modern, lower-cost technologies. A Wi-Fi-based network was established to connect monitoring devices, including a field server collecting meteorological data and water levels, and to provide internet connectivity to remote villages. This approach not only facilitates data transmission for the EWS but also provides co-benefits to the community, which can help foster local ownership and support for the system's maintenance.
+
+Synthesizing the lessons from these diverse case studies reveals a critical conclusion: technology alone is insufficient to create a successful EWS. The most sophisticated sensor network is rendered useless if it is not supported by a robust institutional framework, a sustainable long-term maintenance plan, and deep engagement with the at-risk community. The Kyagar and Laguna 513 systems succeed because their technology is embedded within a clear operational protocol involving specific authorities. The Thorthormi case demonstrates that technology is fragile and that local, traditional knowledge can be a life-saving form of redundancy. The Nepal cases starkly illustrate that an EWS is a long-term commitment, not a one-time project; without sustained funding for upkeep, the initial investment can be entirely lost. Therefore, the design process for an EWS must extend beyond hardware and software specifications. It must include a multi-year budget for operations and maintenance, a plan for co-designing warning and response protocols with local emergency managers, and a comprehensive strategy for community training, education, and drills to build trust and ensure that a warning, when issued, leads to effective, life-saving action.
+
+## Recommendations for EWS Design and Implementation
+
+Based on the comprehensive analysis of GLOF triggers, monitoring technologies, and practical case studies, the following recommendations are provided to guide the design and
+
+implementation of a robust and effective GLOF Early Warning System.
+
+### **Prioritizing Key Monitoring Parameters for a Tiered Warning System**
+
+A scalable and resource-efficient strategy should be adopted, focusing monitoring efforts in a tiered approach based on assessed risk levels.
+
+- Tier 1 (Regional Surveillance): This foundational layer should involve the systematic, continuous monitoring of all glacial lakes within a given region using freely available satellite data. Annual or bi-annual inventories should be created using optical (Landsat, Sentinel-2) and SAR (Sentinel-1) imagery to track changes in lake area and number. This wide-area surveillance identifies emerging threats and allows for the prioritization of resources.
+- Tier 2 (Targeted Monitoring): For lakes identified in Tier 1 as high-risk (e.g., rapid expansion, proximity to critical infrastructure, location beneath unstable slopes), a more intensive monitoring schedule should be implemented. This includes more frequent analysis of high-resolution commercial satellite imagery and the application of advanced techniques like InSAR to detect and quantify slope deformation and dam subsidence, which are precursors to failure.
+- Tier 3 (In-Situ Alarm System): For the most critical, high-consequence lakes designated as "Potentially Dangerous Glacial Lakes" (PDGLs), a permanent, real-time in-situ sensor network is warranted. This network must be designed with redundancy and should prioritize the continuous monitoring of the most immediate trigger parameters:
+  - 1. Rate of Lake Level Rise: The most direct indicator of rapid inflow from melt or rainfall.
+  - 2. **Ground Vibration:** Geophones placed on the dam and surrounding slopes to provide an instantaneous alert for an avalanche trigger.
+  - 3. **Real-time Meteorological Data:** An on-site AWS to correlate lake level changes with temperature and precipitation events.
+
+### Integrating Data Streams for Predictive Modeling and Anomaly Detection
+
+An advanced EWS should move beyond simple, single-parameter threshold alarms. All data streams—from remote sensing platforms and in-situ sensors—should be transmitted to a central data processing hub. This integrated dataset enables more sophisticated analysis. Machine learning algorithms can be developed and trained on this multi-parameter data to
+
+identify complex correlations and anomalies that may signal an impending GLOF. For example, a model could learn to recognize a hazardous signature composed of a combination of factors—such as several days of above-average temperatures, the onset of intense rainfall, and a subtle increase in microseismic activity on a nearby slope—that would be missed by individual sensor alarms. This data fusion approach enhances predictive capability and can increase warning lead times.
+
+### **Establishing Actionable Thresholds and Robust Warning Protocols**
+
+The ultimate purpose of the technology is to trigger a timely and effective human response. This requires a clear, unambiguous warning protocol.
+
+- Calibrate Actionable Thresholds: The system's alarm logic should be based on
+  empirically derived, physically meaningful thresholds. The normalized depth of >0.60 and
+  hydrostatic pressure of >510 kPa provide an excellent starting point for calibrating alarms
+  related to dam stability failure.<sup>28</sup> Thresholds for other parameters, such as rainfall
+  intensity () or the magnitude of a seismic signal from a geophone, must also be
+  established through site-specific analysis and modeling.
+- Develop a Multi-Level Warning Protocol: A tiered warning system, similar to the one
+  used in Peru (Green for normal, Yellow for alert, Orange for prepare to evacuate, Red for
+  evacuate), should be developed in close collaboration with local and national disaster
+  management authorities.<sup>36</sup> Each level must be tied to specific sensor readings or model
+  outputs and must have a corresponding, pre-agreed set of actions for authorities and
+  the public.
+- Ensure Redundancy in Monitoring and Communication: The system design must assume that components will fail. Redundancy is critical. This means deploying multiple sensors to measure the same parameter at critical locations and, equally important, establishing multiple channels for disseminating warnings. A robust communication plan should utilize sirens, mobile phone alerts (SMS and cell broadcast), radio broadcasts, and designated community messengers to ensure the warning reaches everyone at risk, even if one communication pathway fails.<sup>36</sup> Given that some GLOF triggers provide only minutes of warning, the entire system from detection to dissemination must be automated and optimized for speed.<sup>44</sup>
+

chatbot/rag/data/('Glacial Lake Inventory and Lake Outburst Flood Debris Flow Hazard Assessment', '.pdf')_extraction.md

@@ -0,0 +1,298 @@
+
+
+Abstract: Glacial lake outburst floods (GLOF) evolve into debris flows by erosion and sediment entrainment while propagating down a valley, which highly increases peak discharge and volume and causes destructive damage downstream. This study focuses on GLOF hazard assessment in the Bhote Koshi Basin (BKB), where was highly developed glacial lakes and was intensely affected by the Gorkha earthquake. A new 2016 glacial lake inventory was established, and six unreported GLOF events were identified with geomorphic outburst evidence from GaoFen-1 satellite images and Google Earth. A new method was proposed to assess GLOF hazard, in which large numbers of landslides triggered by earthquake were considered to enter into outburst floods enlarge the discharge and volume of debris flow in the downstream. Four GLOF hazard classes were derived according to glacial lake outburst potential and a flow magnitude assessment matrix, in which 11 glacial lakes were identified to have very high hazard and 24 to have high hazard. The GLOF hazard in BKB increased after the earthquake due to landslide deposits, which increased by  $216.03 \times 10^6$  m<sup>3</sup>, and provides abundant deposits for outburst floods to evolve into debris flows. We suggest that in regional GLOF hazard assessment, small glacial lakes should not be overlooked for landslide deposit entrainment along a flood route that would increase the peak discharge, especially in earthquake-affected areas where large numbers of landslides were triggered.
+
+**Keywords:** glacial lake outburst flood (GLOF); debris flow; Bhote Koshi; landslides; Gorkha earthquake; hazard assessment
+
+# 1. Introduction
+
+Many studies have demonstrated that most glaciers are retreating because of global warming and that the meltwater makes an important contribution to the development of glacial lakes in the Himalayas [1–5]. The sudden emptying of these lakes due to dam overflow and moraine or ice dam failure releases large volumes of water and sediment in destructive events called glacial lake outburst floods (GLOF) [6,7]. In the Himalaya region, at least 62 GLOF have been reported, which caused catastrophic destruction and fatalities in downstream regions [8–12]. The peak flood discharge can easily attain tens of thousands of m³/s and travel more than 100 km away [8,13]. Given their high magnitude discharge and long runout distance characteristics, the GLOF impact is sometimes transboundary, especially in the Himalayas. More than 10 GLOF events originated in Tibet, and the catastrophic floods killed hundreds of people and destroyed much infrastructure downstream, causing enormous damage in Nepal and India [14,15]. As a result, GLOF hazard assessment is receiving increased attention from researchers and governments.
+
+Water 2020, 12, 464 2 of 20
+
+In previous GLOF hazard assessment studies, only glacial lakes with an area  $>10^5$  m<sup>2</sup> or volume  $>10^6$  m<sup>3</sup> were considered to be risky of outburst [15,16]. However, some small outbursts occurred in the high mountain regions, but are often ignored due to the limited scale of the events or difficult access [17]. Veh [18] detected 10 previously unreported GLOF from Landsat time series in a study area covering only 10% of the Hindu Kush Himalayan region. In addition, glacial lake outburst floods are highly unsteady flows characterized by pronounced changes as they propagate down to the valley [13]. The outburst flood can change from a normal flood to a hyperconcentrated flow or debris flow [19,20], and the volumes and peak discharges can increase several to ten times owing to erosion, slides from lateral slopes, sediment entrainment and bulking process along the flow path [21,22]. As an example, in Norway, a glacial lake outburst flood developed into a debris flow due to erosion and blockage, and the volume increased nearly ten times from 25,000 to 240,000 m<sup>3</sup> [23]. Sediment can be entrained by scouring unconsolidated deposits of the channel bed, or eroding landslide and collapse of lateral slopes [24,25]. In the seismic belts, large numbers of weak structures and broken rocks are developed along the active fracture zone, and the soils become looser after an earthquake [26]. Studies showed that large earthquakes, such as the Chi-chi earthquake and Wenchuan earthquake, trigger many collapses and landslides, resulting in an increase in loose deposits [27,28]. Although rare reported GLOF events in the Himalaya are directly triggered by earthquakes [29], the loose deposits and landslides induced by earthquakes may affect the magnitude and impact of GLOF. Therefore, it is necessary to build a GLOF hazard assessment model, considering small glacial lakes and the scenario of glacial lake outburst debris flows after earthquakes, especially in areas where many collapses and landslides have developed along the channels.
+
+The Bhote Koshi Basin across China and Nepal, is a highspot area of glacial lakes and GLOF events (Wang and Jiao, 2015). Four glacial lakes have experienced six GLOF events since 1935 (Figure 1). Taraco Lake failed on 28 August 1935, and the GLOF damaged more than 10 hm<sup>2</sup> of wheat fields (Lv, 1999); The Jialongco GLOF occurred on 23 May and 29 June 2002, which caused 7.5 million yuanin economic losses in Nyalam County (Chen et al., 2007). The Cinrenmaco Lake experienced two GLOF events, first in 1964 and second on 10 July 1981. The GLOF in 1981 had the most destructive effects, in which more than 200 people were killed in Nepal, and the total losses were estimated at approximately three million dollars [14,30]. The latest GLOF event occurred on the night of 5 July 2016, which was caused by Gongbatongshaco (GBTSC) Lake in the Zhangzangbo Valley. GBTSC is a small moraine-dammed lake, with a surface area of  $1.7 \times 10^4$  m<sup>2</sup> and it was almost empty after the outburst. Although it only released  $1.1 \times 10^5$  m<sup>3</sup> of water, the peak discharge reached 2400 m<sup>3</sup>/s at Khukundol, 30 km downstream from the lake, due to severe erosion and sediment entrainment [31]. The GLOF caused severe damage downstream of Bhote Koshi, damaging 77 houses, 3 bridges and the Araniko Highway, and destroying the intake dam of the Upper Bhote Koshi Hydropower Project in Nepal (Figure 1). The 2016 GLOF damage sits within the area affected by the Gorkha earthquake (magnitude M 7.8 in 2015), where extensive landslides and rockfalls were triggered on the slopes, and some landslide deposits even blocked the river [32,33]. Therefore, the small glacial lake GBTSC GLOF caused a serious disaster, which caused us to reconsider the small-lake induced GLOF hazard after
+
+The aims of this study are: (1) to establish a detailed glacial lake inventory of BKB after the Gorkha earthquake, based on high resolution remote sensing satellite images; (2) to evaluate the GLOF hazard of BKB considering the scenario that outburst floods evolve into debris flows due to erosion and entrainment of loose solids.
+
+Water 2020, 12, 464 3 of 20
+
+Image /page/2/Figure/1 description: A figure composed of four photographs labeled a, b, c, and d, illustrating the effects of a glacial lake outburst flood. Panel 'a' shows a rocky basin with a small lake, with a blue dashed line indicating the much higher 'Lake surface before outburst' and a person circled and labeled 'Man' for scale. Panel 'b' displays a wide mountain landscape with snow-capped peaks, showing the 'Breach' in the moraine from which the lake burst and the resulting large 'Debris fan' of rocks and sediment below. Panel 'c' depicts a river valley downstream, where a village is situated next to the river, with signs of landslides and erosion on the steep hillsides. Panel 'd' shows an aerial view of a concrete dam structure being overwhelmed and damaged by a powerful, muddy flood carrying significant debris.
+
+**Figure 1.** Photographs of Gongbatongshaco (GBTSC) Lake and the glacial lake outburst floods (GLOF) damage caused downstream: (a) GBTSC Lake after outburst, (b) the breach and debris fan in front of the lake, (c) landslides and river bank collapse triggered by GLOF near Friendship Bridge and (d) the destroyed dam of Upper Bhote Koshi Hydroelectric Project in Nepal.
+
+# 2. Study Area
+
+The study site is located in the central Himalayas and covers latitudes  $27^{\circ}37'-28^{\circ}31'$  N and longitudes  $85^{\circ}40'-86^{\circ}20'$  E with an area of 3406 km² (Figure 2). Bhote Koshi, which is also called Poiqu in China, is a transboundary river with a length of 143 km. It originates in the Bangbulei Mountains in northern Nyalam County, China, flows into Nepal, and at last feeds into the Ganges River. The Araniko Highway, built along the Bhote Koshi Valley, is a key trade and transport route between China and Nepal. Approximately 200,000 people live in the watershed, among which only 2.7% of them live in China, and the infrastructure in this region is particularly vulnerable [34].
+
+The Bhote Koshi Basin stretches across the Higher Himalaya and Lower Himalaya, and the South Tibetan Detachment System (STDS) and the Main Central Thrust (MCT) pass through it. The basin is strongly affected by seismic activity. According to the statistics of the United States Geological Survey (USGS) earthquake records (http://earthquake.usgs.gov/earthquakes), there were 213 earthquakes (magnitude larger than M 4.5) in the area of 150 km² around BKB from 1983 to 2016, including a M 8.3 earthquake, three earthquakes larger than M 7.0 and 79 earthquakes equal or larger than M 5.0. The latest large earthquake, M 7.8, on 25 April 2015 and its largest aftershock (M 7.3) on 12 May 2015, produced severe impact in the study area. The epicenter of the major aftershock was only 19 km southeast of Kodari. The lower part from Zhangzangbo Valley to Dolalghat, which is approximately half of the region, was located in seismic intensity zones VIII and VII, and the upper part was in the VI zone according to the seismic intensity of the Gorkha earthquake provided by the USGS National Earthquake Information Center (Figure 2).
+
+The elevation ranges from the highest peak of Mt. Shishapangma at 8012 m to the lowest point of 591 m in Dolalghat, Nepal. Given the large relief, the landforms are different from north to south. In the north and central parts of the basin are alpine regions and gorges, while the valley becomes broader in the south. The climate also varies considerably from south to north. The Himalayan southern slope region of the basin is affected by the Indian monsoon and experiences high precipitation levels. Meanwhile, due to blockage by the Himalayan range, the warm, moist air from the Indian monsoon
+
+Water 2020, 12, 464 4 of 20
+
+can hardly reach the northern part of the basin. According to monthly data obtained from the Nyalam meteorological station (3810 m a.s.l.) and the Zhangmu meteorological station (2250 m a.s.l.), the mean annual temperature ranges from 3.8  $^{\circ}$ C to 10.1  $^{\circ}$ C, and the mean annual precipitation ranges from 643.4 mm in the north to 2820.6 mm in the south.
+
+Image /page/3/Figure/2 description: A detailed topographical and geological map titled "Figure 2. Study area and GLOF events in Bhotekoshi Basin." The map covers a region from 85°40'E to 86°20'E longitude and 27°40'N to 28°30'N latitude. It displays elevation using a color scale, ranging from 591 to 8012 meters above sea level (m a.s.l.). Key features are identified in a comprehensive legend. These include an earthquake epicenter marked with a yellow star (M7.3), GLOF (Glacial Lake Outburst Flood) events marked with orange circles at locations like Jialongco and Taraco, settlements like Kodari and Barhabise, rivers such as Bhotekoshi and Poiqu, faults shown as red lines, and the national boundary. The map also delineates glaciers, glacial lakes, the Araniko Highway, and zones of seismic intensity labeled with Roman numerals V, VI, VII, and VIII. A north arrow and a scale bar for 0 to 20 km are included. An inset map in the bottom right corner places the study area within the Central Himalaya, showing its relation to India, Nepal, and China, and marking additional earthquake epicenters.
+
+Figure 2. Study area and GLOF events in Bhote Koshi Basin.
+
+Water 2020, 12, 464 5 of 20
+
+# 3. Methods
+
+## 3.1. Glacial Lake and Landslide Inventory Mapping
+
+Glacial lake and landslide identification were based on GaoFen-1 (GF-1) satellite images. Twenty-two GF-1 images obtained from the China Center for Resources Satellite Data and Application (http://www.cresda.com/CN/) were Level-1A products, with cloud coverage less than 20% (Table 1). Seven images were used to map glacial lakes and fifteen images were used to map landslides preand post-earthquake. Geometric correction and image sharpening were conducted in ENVI 5.2 before mapping in ArcGIS 10.2. The resolution of pan sharpened images was 2 m. This high-quality imagery available allowed us to recognize glacial lakes as small as 0.01 km<sup>2</sup>. Manual visual interpretation was used to identify glacial lakes and landslides.
+
+| Data Usage                           | Sensor | Product ID | Date              | Cloud (%) |
+|--------------------------------------|--------|------------|-------------------|-----------|
+| Glacial lake<br>mapping              | PMS1   | 2056413    | 20 December 2016  | 13        |
+|                                      | PMS1   | 2056415    | 20 December 2016  | 4         |
+|                                      | PMS1   | 2056414    | 20 December 2016  | 3         |
+|                                      | PMS1   | 1929663    | 1 November 2016   | 13        |
+|                                      | PMS1   | 1929664    | 1 November 2016   | 16        |
+|                                      | PMS1   | 1929665    | 1 November 2016   | 6         |
+|                                      | PMS2   | 1524197    | 14 April 2016     | 15        |
+| Landslide mapping<br>post-earthquake | PMS2   | 1242505    | 13 December 2015  | 11        |
+|                                      | PMS2   | 1242506    | 13 December 2015  | 1         |
+|                                      | PMS2   | 820531     | 22 May 2015       | 1         |
+|                                      | PMS1   | 827786     | 23 May 2015       | 3         |
+|                                      | PMS2   | 1062062    | 26 May 2015       | 14        |
+|                                      | PMS2   | 1062061    | 26 September 2015 | 0         |
+|                                      | PMS2   | 1242505    | 13 December 2015  | 11        |
+|                                      | PMS1   | 1251892    | 17 December 2015  | 3         |
+| Landslide mapping<br>pre-earthquake  | PMS2   | 751296     | 11 April 2015     | 0         |
+|                                      | PMS2   | 598009     | 19 January 2015   | 12        |
+|                                      | PMS2   | 507470     | 9 December 2014   | 1         |
+|                                      | PMS3   | 507469     | 9 December 2014   | 19        |
+|                                      | PMS1   | 646048     | 22 September 2014 | 16        |
+|                                      | PMS1   | 232717     | 22 May 2014       | 2         |
+|                                      | PMS1   | 142225     | 30 December 2013  | 9         |
+
+Table 1. GaoFen-1 images used in this study.
+
+All glacial lakes were verified and modified against Google Earth to see if there are some misinterpretations of the results due to the effect of terrain shadow. The characteristics and surrounding information of all lakes (larger than 0.01 km²) were measured or estimated, aided by Topography Mission digital elevation model (SRTM DEM) (30 m) and Google Earth. These data compose a complete inventory and provide a basis for identifying dangerous glacial lakes. The inventory of the database consisted of 17 parameters, and some important attributes are explained as follows:
+
+- (a) Name: some glacial lakes were annotated according to the topographical map of 1978.
+- (b) Longitude and latitude: the central location of a glacial lake was calculated automatically in ArcGIS based on WGS84 coordinates.
+- (c) Elevation (m a.s.l.): the central elevation of a glacial lake was derived from the DEM.
+- (d) Dam type: moraine dam, ice dam and bedrock dam, which was specified based on remote sensing images and the topography map (1:100,000; produced in 1978).
+- (e) Area (km<sup>2</sup>): the glacial lake surface area was calculated automatically in ArcGIS 10.2, based on UTM projection zone 48 on a WGS84 ellipsoid.
+- (f) Dam width (m): these values were estimated using Google Earth.
+
+Water 2020, 12, 464 6 of 20
+
+(g) Volume (m³): each glacial lake's volume was estimated using Equation (1), which was established between lake areas and volumes of lake water based on data from 33 Himalayan glacial lakes measured in the field [34],
+
+$$V_{\rm gl} = 0.0578 A_{\rm gl}^{1.4683} \tag{1}$$
+
+where,  $A_{gl}$  is glacial lake area.
+
+- (h) Estimated freeboard values (1, or 0): the height of the freeboard is difficult to measure by remote sensing but is a crucial parameter that influences dam failure. Here, we estimated whether the height was larger than only a few meters (the value was 1) or indeed close to zero (the value was 0) [35], so it is a semiquantitative parameter.
+- (i) Potential triggering impacts: whether the mass movement around a glacial lake can enter into the lake, such as rockfalls (R), landslides (L), ice and glacier avalanches (IGA), debris flows (DF) or flood from a lake situated upstream (ULF). If there is no mass movement, the value was null. This was identified based on Google Earth and the slope maps derived from the DEM, so it is also a semiquantitative parameter.
+- (j) Distance to mother glacier (m): the distance between the back edge of a glacial lake to the mother glacier. If they are in contact, the value was 0; if there is no glacier around the lake, the value was set to null.
+- (k) Distance to the nearest settlement (m): the drainage distance from the glacial lake dam to the nearest major settlement was measured using ArcGIS 10.2.
+- (l) Drainage gradient (°): the average drainage gradient was estimated by a DEM-derived drainage map.
+
+In this study, the term landslide refers to mass movement of a slope, including rockfalls, slope failure and soil slides. Most landslides can be easily identified by visual inspection for vegetation loss or deposits. If there is no vegetation in some areas, the morphology needs careful attention. Landslides are classified as pre-earthquake and post-earthquake landslides. The landslide volume (V) was estimated using a power-law landslide area–volume empirical formula (Equation (2)):
+
+$$V_{\rm s} = \alpha A_{\rm s}^{\gamma} \tag{2}$$
+
+where  $A_s$  is landslide area,  $\alpha$  and  $\gamma$  are empirically calibrated scaling parameters derived from mixed soil and bedrock landslides in the Himalayas;  $\alpha$  is 0.257, and  $\gamma$  is 1.36 [36].
+
+## 3.2. Glacial Lake Outburst Hazard Assessment
+
+Glacial lake outburst hazard assessment includes two steps, glacial lake outburst potential assessment and flow magnitude assessment. First, a qualitative method was used to identify glacial lake outburst potential; then, the outburst flow characteristics were determined, flood or debris flow according to loose matter along the flow path and the channel gradient, and then the magnitude at the nearest settlement was calculated. Finally, the GLOF hazard was derived by the glacial lake outburst probability and flow magnitude based on a matrix diagram, which has been widely used in flood, landslide and rock fall hazard assessments [7,37]. GLOF hazards in BKB were divided into four classes: "Very High", "High", "Medium" and "Low". The process of GLOF hazard assessment is summarized in Figure 3.
+
+Water 2020, 12, 464 7 of 20
+
+Image /page/6/Figure/1 description: A flowchart illustrating the process for a glacier lake outburst hazard assessment. The process starts with two main inputs: 'GF-1 images Google Earth' and 'SRTM DEM'. The flowchart is divided into two main parallel paths that converge at the end. The left path determines the 'Qualitative outburst potential assessment'. It starts with 'Mapping glacial lakes' from the GF-1 images, leading to a 'Glacial lake inventory' and then 'Identifying glacial lake outburst potential'. This identification is based on five criteria: 'Dam type' (Bedrock dam vs. Moraine/ice-dam), 'Distance to glacial lake' (>1000, 500-1000, <500), 'Dam width crest' (>60, <60), 'Freeboard' (>1, ~0), and 'Rockfalls/landslide/debris flow/flood from upstream' (No, Yes). Based on these, a 'Qualitative outburst potential assessment' is made, categorized as Low, Medium, or High. The right path determines the 'Flow magnitude assessment'. It uses 'Mapping landslides' from GF-1 images to create a 'Landslides inventory'. It also uses the 'SRTM DEM' to determine 'Slope' and a 'Stream map', which are combined to find the 'Drainage gradient'. Two decision points follow: 'Landslide in downstream?' and 'Gradient > 8°?'. If the answer to both is 'No', it results in 'Flood discharge'. If the answer to either is 'Yes', it results in 'Debris flow discharge'. Both discharge types lead to a 'Flow magnitude assessment', categorized as <200, 200-500, or >500. Finally, the 'Qualitative outburst potential assessment' (relabeled as 'Outburst probability' with levels Low, Medium, High) and the 'Flow magnitude assessment' (with levels Low, Medium, High) are combined in a hazard matrix to produce the final 'Glacier lake outburst hazard assessment'. The matrix shows the resulting hazard level: Low, Medium, High, or Very High. For example, a High outburst probability and a High flow magnitude result in a Very High hazard. The entire flowchart uses a color code: blue for low risk, yellow for medium risk, and red/dark red for high/very high risk.
+
+**Figure 3.** Flow chart illustrating the process used for creating the glacial lake inventory of Bhote Koshi Basin in 2016, classifying glacial lake outburst potential and outburst risk assessment.
+
+### 3.2.1. Glacial Lake Outburst Potential Assessment
+
+Many criteria and schemes, derived from GLOF experiences all around the world, have been proposed to assess and identify potentially dangerous glacial lakes based on GLOF [38–42]. Here, a qualitative assessment method is proposed to identify the potentially dangerous glacial lakes from three criteria. The first one is potential triggering impacts, such as rockfalls, landslides, snow and ice avalanches, debris flows and flood from a lake situated upstream [43]. Such mass movement entering the lakes trigger displacement waves that subsequently overtop and erode the dams is the most common cause of dam failure in the Himalaya [8,17,44–46]. Steep glacier surfaces that are in contact with or close to a lake are prone to ice avalanches [47–49]. In addition, steep topography is also likely to cause rockfalls and landslides, and as a glacier retreats, much glacial debris remains, which may start a debris flow under heavy rainfall or intense glacier melting [8,38,50]. The second one is dam stability. Studies show most of the GLOF events in the Himalayas are caused by moraine-or ice dam failures, and the bedrock dams are the more stable with low outburst probability [8,51]. The dam width crest is an indicator for the susceptibility of a dam to fail [52]. The third one is freeboard, which is considered a crucial parameter that influences whether a potential impact wave overtops the dam [7,49].
+
+Water 2020, 12, 464 8 of 20
+
+Five key indicators were selected to identify glacial lake outburst potential according to the three criteria. These parameters of each glacial lake were easily obtained from the 2016 glacial lake inventory database. The critical values for assessment are given in Figure 3. The key indicator was defined with qualitative probabilities high, medium and low, and considered independently. The overall potential is not the mean of the individual indicators. A high-potential outburst glacial lake must satisfy three criteria that are high, a low-potential lake has two or three low criteria and no high criteria and the rest are medium-potential lakes. Finally, three potential degrees were classified as high, medium and low.
+
+### 3.2.2. Flow Magnitude Assessment
+
+The flow magnitude is highly dependent on the peak discharge at the breach and the channel condition [53]. The peak discharge depends on the lake volume and the breach geometry [38]. For a rapid hazard assessment, complex breach processes and flow behavior are beyond the scope. In this paper, the worst breach scenario was assumed, i.e., a full breach that empties the glacial lake water completely. The maximum discharge ( $Q_p$ ) was estimated using the empirical formula (Equation (3)):
+
+$$Q_p = 2V_{ql}/t \tag{3}$$
+
+where  $V_{\rm gl}$  is the glacial lake volume, and t is the drainage duration in seconds, which is assumed to be 1000 s [54].
+
+The outburst flood peak discharge increases due to erosion and entrained sediments. Thus, we first needed to judge whether an outburst flood would develop into a debris flow. The average channel gradient and unconsolidated deposits along the channel are key factors that affect whether an outburst flood evolves into a debris flow [13,55]. Erosion is found to occur where the channel gradients exceed 8° [38] and abundant unconsolidated deposits are distributed in the channel and on the slopes [56]. Channel gradients were calculated based on drainage maps derived from the DEM. The unconsolidated deposits include moraine deposits, fluvial and glaciofluvial sediments and landslide deposits. The maximum eroded sediment volume per unit channel length varies from ten to hundreds of cubic meters due to local and regional differences in geology, topography and hydrology of torrent catchments [38,54]. Therefore, it was hard to set a value certain of sediment depth or volume eroded by flood in different channels or basins. A rough assessment was used to estimate the flow magnitude to the nearest settlement. Flood peak discharge was estimated using an empirical equation (Equation (4)) [57,58]:
+
+$$Q_{\rm pl} = \frac{W}{\frac{W}{Q_p} + \frac{L}{VK}} \tag{4}$$
+
+where  $Q_p$  is the flood peak discharge m<sup>3</sup>/s; W is the capacity of the lake, m<sup>3</sup>;  $Q_p$  is the peak discharge at the breach, m<sup>3</sup>/s; L is the distance from the glacial lake dam, m; and VK is an empirical coefficient equal to 3.13 for rivers on plains, 7.15 for mountain rivers and 4.76 for rivers flowing through terrain with intermediate relief [59], which here we set the value as 7.15.
+
+For an outburst debris flow, the water source is the outburst flood. Therefore, the peak discharge of the debris flow consists of outburst flood discharge and soil particle flow. Blocking was not considered here, so the debris flow peak discharge ( $Q_{df}$ ) can be calculated [60]:
+
+$$Q_{\rm df} = (1 + \varphi)Q_{vl} \tag{5}$$
+
+where  $\varphi$  is the increase coefficient of debris flow peak discharge, which can be calculated by:
+
+$$\varphi = (\gamma_s - \gamma_w)/(\gamma_s - \gamma_c) \tag{6}$$
+
+where  $\gamma_s$  is the specific gravity of the solid material, g/cm<sup>3</sup>, and usually determined as 2.65 g/cm<sup>3</sup>;  $\gamma_w$  is the unit weight of water,  $\gamma_w = 1$  g/cm<sup>3</sup>;  $\gamma_c$  is the unit weight of the debris flow, g/cm<sup>3</sup>. Studies show
+
+Water 2020, 12, 464 9 of 20
+
+glacial lake outburst debris flow in Tibet is usually diluted flow [52], and the density is 1.3–1.8 g/cm<sup>3</sup>. For the convenience of calculation, here we set the average value of  $\gamma_c$  as the density of GBTSC outburst debris flow, 1.55 g/cm<sup>3</sup>.
+
+According to Chinese debris flow prevention and control standards (DZT-0220-2006), a peak flow discharge of more than 200 m $^3$ /s is defined as a large hazard. However, the scale of a glacial outburst flood/debris flow is usually larger than that of a rainfall-triggered debris flow [16]. Therefore, in this paper, three flow magnitude classes were established: flow discharge <200 m $^3$ /s (low), 200–500 m $^3$ /s (medium) and >500 m $^3$ /s (high). Finally, the GLOF hazard was derived by the glacial lake outburst probability and flow magnitude based on a matrix diagram, which has been widely used in flood, landslide and rock falls hazard assessments [7,37]. GLOF hazards in BKB are divided into four classes: "Very High", "High", "Medium" and "Low".
+
+# 4. Results
+
+## 4.1. Glacial Lake Inventory
+
+A total of 122 glacial lakes larger than 0.01 km² with an area of 20.38 km² were identified based on the GF-1 images from 2016 (see Supplementary Materials). According to the dam type, 84 moraine-dammed lakes with a total area of 16.87 km² accounts for the largest number and area of all lakes. These moraine-dammed lakes are mainly distributed at 5100–5400 m a.s.l. There are 25 bedrock-dammed lakes that account for 15.3% of the area of all lakes. The average area of a bedrock-dammed lake is 0.12 km², and are mainly distributed at 4100–4700 m a.s.l. The ice-dammed lakes are the least and smallest, occupying 10.7% and 1.9% of the total number and area. The ice-dammed lakes consist of tiny and small lakes, with a mean area of 0.03 km², mainly distributed at 5000–5200 m a.s.l. (Figure 4a).
+
+Image /page/8/Figure/6 description: The image displays characteristics of a glacial lake basin in 2016 through three panels labeled a, b, and c.
+
+Panel a is a topographical map showing the distribution of glacial lakes. A legend indicates the dam type and size class of the lakes. Dam types are bedrock dam (light green circles), ice dam (yellow circles), and moraine dam (dark green circles). The size classes, in square kilometers, are represented by circles of increasing size: ≤ 0.02, 0.02 - 0.1, 0.1 - 0.5, 0.5 - 1, and > 1. The map includes rivers and place names such as Tajilin, Nyalam, and Zhangmu.
+
+Panel b is a dual-axis bar chart showing the number and area of glacial lakes by size class. The x-axis shows size classes: <0.02, 0.02 - 0.1, 0.1 - 0.5, 0.5 - 1, and >1 km². The left y-axis represents the number of lakes (grey bars), and the right y-axis represents the area in km² (teal bars). The data shows that the number of lakes is highest in the smallest size class (<0.02 km²) with approximately 56 lakes, and decreases as size increases. Conversely, the total area is greatest for the largest size class (>1 km²), at about 12 km², despite having the fewest lakes (approximately 2).
+
+Panel c is a bar chart showing the percentage of glacial lakes by area (teal bars) and by number (grey bars) at different elevations. The x-axis shows elevation in meters, from 4000 to 5800. The y-axis shows the percentage. The distribution is concentrated at higher elevations, with the largest percentages found between 5100 m and 5400 m. The peak percentage by area is at 5200 m (around 31%), while the peak percentage by number is at 5300 m (around 14%).
+
+**Figure 4.** Characteristics of glacial lakes in 2016. (a) Glacial lakes distribution of different dam type and size class, (b) the number and area variation in size classes and (c) the percentage of number and area at different elevations.
+
+As the area of glacial lakes vary greatly, from 0.01 to  $5.29 \, \text{km}^2$ , we classify them into five size classes: tiny (A  $\leq 0.02 \, \text{km}^2$ ), small ( $0.02 < A \leq 0.1 \, \text{km}^2$ ), medium ( $0.1 < A \leq 0.5 \, \text{km}^2$ ), large ( $0.5 < A \leq 1 \, \text{km}^2$ ) and giant (A  $> 1 \, \text{km}^2$ ). The percentage of numbers and areas for each size class are shown in Figure 4b.
+
+Water 2020, 12, 464 10 of 20
+
+The main size class of a glacial lake is tiny, accounting for 45.9% (n = 56) of the total number, and small glacial lakes account for approximately 36.1% (n = 44). The total area of tiny and small lakes is 12.4%. Four glacial lakes are large, and three are giant lakes, occupying 69.7% of the total area. Tha mean value of lake area is  $0.17 \text{ km}^2$ , and the largest glacial lake is Galongco, with a surface area of  $5.29 \text{ km}^2$ .
+
+The glacial lakes are distributed at elevations ranging from 4100 to 5750 m a.s.l. and are separated into different elevation classes every 100 m (Figure 4c). Most glacial lakes are located at elevations of 5000–5600 m a.s.l., accounting for 66.4% and 86.6% of the total number and area, respectively. Approximately 27.9% of glacial lakes are located below 5000 m a.s.l. and are evenly distributed in each elevation class with an average 3.5% by number. Approximately 13.9% of glacial lakes are distributed from 5200–5300 m a.s.l. and account for 28.1% of the area of all lakes. It is noticeable that the largest percentage by area is distributed at 5000–5100, which accounts for 30.55%.
+
+## 4.2. Glacial Lake Outburst Flood Hazard
+
+The glacial lake outburst potential assessment results show that 19 glacial lakes have high outburst potential, in which all of these lakes are moraine-dammed and ice/glacier avalanche is the main potential triggering impact; 51 are medium risk, in which two are ice-dammed and nine are bedrock-dammed, and 42 lakes with an area less than 0.1 km²; 52 glacial lakes are low, in which 11 are ice-dammed and 16 are bedrock-dammed (Figure 5a). It is noticeable that 11 out of 19 high outburst potential lakes have an area less than 0.1 km², and the one bedrock-dammed lake, Gongco, with an area of 2.9 km², has low outburst potential.
+
+Image /page/9/Figure/5 description: The image displays two maps, labeled 'a' and 'b', of the same mountainous river basin. Both maps show elevation, rivers, and settlements, with coordinates ranging from 85°40'E to 86°20'E and 27°40'N to 28°30'N. A scale bar indicates a length of 20 km. The elevation ranges from 591 m to 8012 m. Map 'a' shows the 'Outburst potential' categorized as Low (small blue dots), Medium (yellow dots), and High (large red dots). Map 'b' shows the 'Hazard rank' categorized as Low (blue dots), Medium (orange dots), High (red dots), and Very High (large red dots). Map 'b' also includes an overlay of landslides, distinguishing between 'pre-earthquake' (yellow areas) and 'post-earthquake' (pink areas), with the latter being more extensive. Several settlements are named, including Dolalghat, Barhabise, Gumthang, Zhangmu, Quxiang, Nyalam, Gangga, and Tajilin.
+
+**Figure 5.** (a) Glacial lake outburst potential and (b) glacial lake outburst flow/debris flow hazard assessed in this study, considering landslide deposit distribution.
+
+In this study, 1670 landslides with a total area of 18.70 km<sup>2</sup> were identified, in which 1183, with an area of 12.18 km<sup>2</sup>, were triggered by the Gorkha earthquake (Figure 5b). These post-earthquake landslides vary in size ranging from 230 m<sup>2</sup> to 254,474 m<sup>2</sup>. Most of the landslides are distributed in the middle and southern parts of the basin, and a large number of Gorkha earthquake-triggered landslides are concentrated in the VII region of seismic intensity. A lot of landslides were distributed in the sub-basins such as Gumthang, Deqingdang, Chongduipu, Zhangzangbo and Dianchangchanggou. Large landslides reach channels, and some small landslides are mostly located on steep slopes
+
+disconnected from a river channel. The other 487 landslides occurred before the Gorkha earthquake and have an area of 6.51 km<sup>2</sup>. The largest mapped landslide that occurred before the earthquake is  $0.81 \text{ km}^2$ . The total landslide deposit volume is estimated at  $91.67 \times 10^6 \text{ m}^3$  before the earthquake, and the volume increased to  $216.03 \times 10^6 \text{ m}^3$  after the earthquake. Considering landslide distribution, 73 glacial lake outburst floods are highly prone to debris flow, which will increase the magnitude.
+
+According to the glacial lake outburst potential and flow magnitude, GLOF hazard assessment results are shown in Table 2 and Figure 5. Eleven glacial lakes are identified with very high hazard, among which seven could evolve into debris flows. Twenty-four glacial lakes are high hazard, among which 11 could evolve into debris flows. Thirty-two glacial lakes are identified with medium hazard; the other 55 glacial lakes are considered to have low to no hazard to downstream areas. Four very high-hazard glacial lakes are located in the Chongduipu gully, which presents a large threat to Nyalam County, especially the giant glacial lake Galongcuo that could generate a peak flow discharge of about 224,449 m<sup>3</sup>/s. Both Jialongcuo and Cirenmacuo have burst out twice before, and are also identified as very high hazard due to high freeboards and hanging glaciers behind the lakes. Eight glacial lakes with low-outburst probability but high-magnitude flow are considered to have high hazard. Among these lakes, Gongcuo and Darecuo are bedrock-dammed lakes and have no potential triggering impacts around the lakes, so they are considered to have a low probability of outburst. However, because of their large volumes, the outburst flows are assumed to be high. 63% (n = 22) of the very high and high glacial lakes' areas are larger than 0.1 km<sup>2</sup>, and their peak flow discharges were larger than 1000 m<sup>3</sup>/s. Ten small glacial lakes (area <0.1 km<sup>2</sup>) identified as high hazard. Three small glacial lakes, No. 16 (area 0.09 km<sup>2</sup>), No. 18 (area 0.05 km<sup>2</sup>) and No. 81 (Nongjue, area 0.07 km<sup>2</sup>), are considered very high hazard for they may cause peak debris flow discharges of 1118 m<sup>3</sup>/s, 597 m<sup>3</sup>/s and 994 m<sup>3</sup>/s at the nearest settlement, respectively.
+
+Table 2. Very high and high hazard glacial lakes.
+
+| Id  | Name          | Longitude<br>(°) | Latitude<br>(°) | Elevation<br>(m) | Dam Type | Area (km²) | Qpl (m3) | Qdf (m3) | Probability of<br>Outburst | Flow<br>Magnitude | Hazard    |
+|-----|---------------|------------------|-----------------|------------------|----------|------------|----------|----------|----------------------------|-------------------|-----------|
+| 1   | Qiezelaco     | 86.26            | 28.37           | 5532             | moraine  | 0.26       | 1967     |          | High                       | High              | Very High |
+| 3   | Cawuqudenco   | 86.19            | 28.34           | 5423             | moraine  | 0.55       | 6666     |          | High                       | High              | Very High |
+| 7   | Paquco        | 86.16            | 28.30           | 5307             | moraine  | 0.58       | 7950     |          | High                       | High              | Very High |
+| 16  |               | 86.09            | 28.22           | 5178             | moraine  | 0.09       | 726      | 1814     | High                       | High              | Very High |
+| 18  |               | 86.06            | 28.17           | 5194             | moraine  | 0.05       | 239      | 597      | High                       | High              | Very High |
+| 22  | Cirenmaco     | 86.07            | 28.07           | 4633             | moraine  | 0.34       | 5087     | 12,717   | High                       | High              | Very High |
+| 34  | Gangxico      | 85.87            | 28.36           | 5212             | moraine  | 4.52       | 172,879  |          | High                       | High              | Very High |
+| 61  | Galongco      | 85.84            | 28.32           | 5077             | moraine  | 5.29       | 145,746  | 364,365  | High                       | High              | Very High |
+| 62  |               | 85.82            | 28.30           | 5093             | moraine  | 0.27       | 1832     | 4580     | High                       | High              | Very High |
+| 80  | Jialongco     | 85.85            | 28.21           | 4380             | moraine  | 0.63       | 8336     | 20,840   | High                       | High              | Very High |
+| 81  | Nongjue       | 85.87            | 28.19           | 4628             | moraine  | 0.07       | 398      | 994      | High                       | High              | Very High |
+| 2   | Youmojiaco    | 86.23            | 28.35           | 5337             | moraine  | 0.55       | 4881     |          | Medium                     | High              | High      |
+| 6   | Gangpuco      | 86.16            | 28.32           | 5543             | moraine  | 0.22       | 2355     |          | Medium                     | High              | High      |
+| 8   | Southhu       | 86.15            | 28.30           | 5343             | moraine  | 0.17       | 1227     |          | Medium                     | High              | High      |
+| 9   | Taracuo       | 86.13            | 28.29           | 5257             | moraine  | 0.23       | 2186     |          | Medium                     | High              | High      |
+| 10  | Tuzhuocuo     | 86.10            | 28.25           | 5201             | moraine  | 0.15       | 1309     | 3272     | Low                        | High              | High      |
+| 23  |               | 86.03            | 28.07           | 4486             | bedrock  | 0.03       | 257      | 642      | Medium                     | High              | High      |
+| 32  | Yinreco       | 85.89            | 28.37           | 5245             | moraine  | 0.28       | 2878     |          | Low                        | High              | High      |
+| 40  | Mabiya        | 85.91            | 28.32           | 5384             | moraine  | 0.14       | 931      |          | Medium                     | High              | High      |
+| 42  |               | 85.92            | 28.32           | 5345             | moraine  | 0.08       | 504      |          | Medium                     | High              | High      |
+| 43  | Mulaco        | 85.93            | 28.32           | 5306             | moraine  | 0.11       | 760      |          | Medium                     | High              | High      |
+| 44  | Xiahu         | 85.95            | 28.31           | 5232             | moraine  | 0.31       | 3352     |          | Medium                     | High              | High      |
+| 51  | Cuonongjue    | 85.92            | 28.26           | 5095             | moraine  | 0.23       | 2353     |          | Low                        | High              | High      |
+| 63  |               | 85.83            | 28.29           | 5013             | moraine  | 0.26       | 1863     | 4658     | Medium                     | High              | High      |
+| 64  |               | 85.83            | 28.29           | 5050             | moraine  | 0.06       | 204      | 511      | Medium                     | High              | High      |
+| 70  |               | 85.78            | 28.29           | 5418             | moraine  | 0.05       | 130      | 324      | High                       | Medium            | High      |
+| 72  |               | 85.78            | 28.27           | 5309             | moraine  | 0.07       | 184      | 459      | High                       | Medium            | High      |
+| 83  |               | 85.87            | 28.17           | 4712             | moraine  | 0.04       | 125      | 312      | High                       | Medium            | High      |
+| 84  | Daroco        | 85.92            | 28.18           | 4366             | bedrock  | 0.48       | 10,966   | 27,414   | Low                        | High              | High      |
+| 85  |               | 85.91            | 28.15           | 4486             | ice      | 0.20       | 2468     |          | Medium                     | High              | High      |
+| 86  |               | 85.92            | 28.14           | 4871             | moraine  | 0.09       | 597      |          | Medium                     | High              | High      |
+| 88  |               | 85.94            | 28.07           | 4524             | bedrock  | 0.06       | 391      | 977      | Low                        | High              | High      |
+| 89  | Bhairab Kunda | 85.88            | 27.99           | 4102             | bedrock  | 0.06       | 304      | 760      | Low                        | High              | High      |
+| 102 |               | 85.83            | 28.05           | 4250             | bedrock  | 0.07       | 210      | 524      | Low                        | High              | High      |
+| 103 | Congco        | 85.87            | 28.33           | 5113             | bedrock  | 2.09       | 28936    |          | Low                        | High              | High      |
+
+Note:  $Q_{\rm pl}$  is the flood peak discharge at the nearest settlement and  $Q_{\rm df}$  is the debris flow peak discharge at the nearest settlement.
+
+# 5. Discussion
+
+## 5.1. Glacial Lake Inventory
+
+The new 2016 glacial lake inventory indicates that BKB is highly developed glacial lakes. Glacial lake inventory studies have also been conducted in other regions along the Himalayas [61–64]. Glacial lake studies in the Himalayas show that the greatest numbers and areas of glacial lakes are distributed in the central Himalaya [2,7,9,10,62,65]. To compare the glacial lakes and GLOF of BKB with other regions throughout the central Himalayas, the Gyirong River Basin (GRB), which is next to BKB with a similar area was selected and the glacial lake density (glacial lake number/basin area) and lake area per basin area (total glacial lake area/basin area) were calculated (Table 3). The results show that the glacial lake density of BKB is four times that of the central Himalayas and the lake area per basin area is four times that of GRB. The basin area of GRB is larger than BKB, and the glacial lakes density is similar, while the lake per basin area varies greatly. This is due to more large and giant lakes in BKB. According to the statistics, the largest lake is less than 0.5 km² in GRB, while there are seven lakes larger than 0.5 km² with the largest being 5.29 km² in BKB. Studies show the glacial lake expansion rate reaching 0.26 km²/year in Poiqu [66], while the rate of GRB is 0.09 km²/year [63]. Glacial lake expansion is the result of glacier retreating response to climate change. That means BKB is more sensitive to climate change than GRB.
+
+**Table 3.** Comparison of glacial lake and glacial lake outburst flood among Bhote Koshi Basin, Gyirong River Basin and Central Himalaya.
+
+| Region                   | Basin Area<br>(km2) | Number of<br>Glacial Lakes | Glacial Lake<br>Area (km2) | Glacial Lake<br>Density | Lake per Basin<br>Area |
+|--------------------------|---------------------|----------------------------|----------------------------|-------------------------|------------------------|
+| Bhote Koshi Basin        | 3406                | 122                        | 20.38                      | 0.04                    | 0.0060                 |
+| Gyirong River Basin [63] | 4640                | 148                        | 7.12                       | 0.03                    | 0.0015                 |
+| Central Himalaya [2]     | 280,000             | 1943                       | 203.7                      | 0.01                    | 0.0007                 |
+
+The analysis of multitemporal and high-resolution remote sensing images during the compilation of the glacial lake inventories provided a good opportunity to identify previously unreported GLOF events [67,68]. Six unreported glacial lake outburst events were found when we mapped glacial lakes from GF-1 and Google Earth. These glacial lakes have retained typical outburst geomorphic and sedimentological features, such as V-shaped breaches, debris fans and subsequent devastated channels (Figure 6). All of them were moraine-dammed lakes, and their surface areas are 0.01–0.11 km<sup>2</sup>. Two glacial lakes (Figure 6a,b) are located in Keyapu Valley and the other four (Figure 6c–f) are in Chongduipu Valley. All glacial lakes except No. 31 are fed by glaciers, and the distances to the glaciers are less than 500 m. The V-shaped breach and debris fan of glacial lake No. 86 is the largest, and its mother glacier is thick and hangs behind the lake. The surface area of the glacial lake is 0.09 km<sup>2</sup> and the freeboard is much more than one. The rest of the outburst events were small scale and seemed to cause no downstream damage since no erosion was observed in downstream channels. The outburst flood formed a deposition fan at the intersection with the main channel, such as glacial lake No. 31, where the outburst deposition blocked the channel and formed a small lake. Some vegetation has covered the debris fan (glacial lake No. 86) and deposition fan (glacial lake No. 81). It shows that the glacial lake outburst occurred a long time ago. However, the outburst time (year) cannot be determined because of the lack of long-term and high-quality (low cloud cover or high-resolution) data. We traced theses lakes on Google Earth images, and it shows that outburst signs have existed since 1984. As we documented in the literature, we found there was a GLOF event in 1955 in Nyalam, but the record did not mention which lake burst out [30]. Four glacial lakes (Figure 6c-f) are located in the Chongduipu Basin upstream of Nyalam County. As we cannot be sure of the exact outburst time, the outburst magnitudes can be estimated only through debris fans. Glacial lake No. 86 had the largest outburst magnitude with a debris fan area of approximately 252,808 m<sup>2</sup>, and the gully downstream is highly eroded. This lake may have caused damage in downstream areas. The other three glacial lake deposits
+
+are found at the intersections with the main channel (Chongduipu), which means their outburst floods did not propagate downstream. Therefore, we conclude that the GLOF event in 1955 was caused by glacial lake No. 86.
+
+Image /page/13/Figure/2 description: A composite image of three satellite photographs, labeled a, b, and c, showing different glacial landscapes with annotations. Each panel includes coordinates and a date. Panel a, dated June 08, 2015, at 85.88°E, 28.39°N, shows a light-colored glacial lake with a breach leading to a deposition area that has formed a smaller, green 'deposition blocked lake'. Panel b, dated November 01, 2016, at 85.93°E, 28.32°N, displays a large, green glacial lake fed by a glacier, with a breach at its outlet and a 'debrirs fan' (likely debris fan) below it. Panel c, dated October 10, 2017, at 85.78°E, 28.27°N, shows two dark glacial lakes near a glacier, with a 'V-shaped breach' and a prominent 'eroded channel' downstream.
+
+Image /page/13/Figure/3 description: A composite image of three satellite views, labeled d, e, and f, showing different glacial landscapes, all dated November 01, 2016. Each panel includes a north arrow. Panel d, on the left, shows a glacier feeding into a dark glacial lake. Below the lake, labels point to a 'breach', a 'debris fan', and a 'deposition fan'. The location is given as No.81: 85.87°E, 28.19°N. Panel e, in the middle, displays a glacier, a brownish-yellow glacial lake, a 'debris fan', and a 'V-shaped breach'. The location is No.83: 85.87°E, 28.17°N. Panel f, on the right, shows a glacier, a green glacial lake, a 'V-shaped breach', and a 'debris fan'. The location is No.86: 85.92°E, 28.14°N.
+
+**Figure 6.** Unreported glacial lake outburst flood events were identified based on Google Earth images (**a**,**c**) and GaoFen-1 (GF-1) images (**b**,**d**-**f**) in the Bhote Koshi Basin.
+
+The six undocumented outburst events found show that high-resolution remote sensing images make it possible to trace minor GLOF events that were unreported because of difficult access or few people living in high mountain regions. It also proves that the GLOF frequency is high in BKB. Other unknown and unpublished GLOF have also been found in Bhutan, Nepal and other parts of the Himalayas, based on long-time-series remote sensing data that show glacial lake changes (disappearance or abrupt shrinkage) and typical topographic features, such as exposed debris fans and sediment tails in downstream river channels [9,17,18]. A database of past GLOF events as complete as possible is essential for robust and reliable GLOF hazard assessment [69]. The gradually improving GLOF inventory helps us better understand the mechanism of GLOF and to do hazard assessment.
+
+Water 2020, 12, 464 15 of 20
+
+## 5.2. Glacial Lake Outburst Hazard
+
+A GLOF is a complex process, and the hazard magnitude is determined by the outburst water volume and flood routing [8,22,38]. The outburst water volume is mainly related to the glacial lake volume [54,70]. Since the depth of a glacial lake is hard to acquire by remote sensing data and investigation, the surface area of the lake becomes an important indicator for assessing lake volume and hazard. In previous GLOF hazard assessment studies, glacial lakes smaller than 0.1 km² were assumed not to pose a hazard potential relevant to downstream locations [47]. Khanal et al. [34] identified 10 critical lakes in the BKB, and all of them are larger than 0.2 km². Indeed, small glacial lake outbursts can cause damage to downstream locations. According to the inventory of historical glacial lake outburst floods in the Himalayas, small GLOF, such as Zanaco, Geiqu and Choradari Lakes with areas smaller than 0.1 km², damaged downstream roads and villages [9]. On one hand, a small outburst can create a much larger outburst from another lake located downstream, for examples, the reach of GLOF from lakes Artesoncocha and Chacrucocha [43,67]. On the other hand, a small glacial lake outburst flood can transform into debris flows due to downstream sediment entrainment. For example, a small ice-dammed lake outburst in 2009 (area of 34,000 m²) at Keara in the Andes caused damage 10 km downstream [41].
+
+During the flood routing, landslides on the slopes enter the flood, transforming the flood into debris flows and greatly increasing the discharge, volume and impact. Ignoring the earthquake-induced landslides would underestimate the basin's GLOF hazard. If we do not consider landslides triggered by earthquakes transforming the glacial lake outburst floods into debris flows in the BKB, only nine glacial lakes are identified as having very high hazard, 16 are at high hazard, 12 are at medium hazard and 85 are at low hazard. The hazards rank of two very high hazard small glacial lakes, eight high hazard lakes and 20 medium hazard lakes, accounting for 24.6% of the total lakes, would be decreased. It leads to the GLOF hazard of BKB greatly underestimated.
+
+In this study, the GLOF caused by GBTSC on 25 July 2016 is a good example. GBTSC is located in the Zhangzangbo Valley on the right bank of the Poiqu River, and the average gradient is 182‰ (Figure 7b). This lake was tiny; the surface area before the outburst was 0.01 km<sup>2</sup>. After the outburst, the lake was almost empty as shown in Figure 7b. The width of the breach was 27 m, and the depth was 9 m. The peak discharge was 618 m<sup>3</sup>/s, and increased to 4019 m<sup>3</sup>/s at the section of the Zhangzangbo Valley mouth (approximately 7.2 km from the breach), according to the investigation and assessment report written by the Institute of Mountain Hazards and Environment (http://www.imde.ac.cn). The discharge increased by almost eight times because the outburst flood changed to a debris flow. The Zhangzangbo was in the seismic intensity zone VII and intensely impacted by the Gorkha earthquake. Many landslides were triggered along the river, and some landslide deposits blocked the river (Figure 7c,d). The loose mass volume increased to  $8.74 \times 10^6$  m<sup>3</sup> in Zhangzangbo according to the landslide distribution. These landslide deposits provided rich masses for the debris flow. Once the GLOF occurred, these deposits were easily eroded and entrained, leading the flood to change to a debris flow and amplifying the discharge. Tens of thousands of landslides were triggered by the M 7.8 (Gorkha) and M 7.3(Dolakha) earthquakes [71,72]. It will take some time to transport these landslide deposits, which accumulated on the slope or in the channel. Thus, in the region affected by strong earthquakes, we must strengthen the monitoring of high-hazard glacial lakes and pay special attention to glacial lake outburst debris flows after an earthquake.
+
+Image /page/15/Figure/1 description: A composite figure with four panels labeled a, b, c, and d, showing satellite images that compare channel morphology in a mountainous region before and after a geological event. 
+
+Panel a is a large overview image dated Oct. 24, 2016. It shows the Zhangzangbo and Poiqu rivers, and two lakes named Gongbatongshaco and Cirenmaco. Three areas along the Zhangzangbo river are highlighted with red dashed boxes and labeled B, C, and D. The image includes a scale bar from 0 to 2,200 meters and a north arrow.
+
+Panel b shows two images of area B. The top image, from Nov. 10, 2015, shows a river channel. The bottom image, from Oct. 24, 2016, shows significant changes, with labels pointing to "Moraine terrace erosion" and a newly formed "debris fan" in the widened river channel.
+
+Panel c displays two images of area C. The left image, from Jun. 01, 2015, shows a river bend. The right image, from Oct. 24, 2016, shows the same bend with significant widening, labeled to indicate "Lateral erosion" and "Bank collapse," which is highlighted with a yellow dashed line.
+
+Panel d presents two images of area D. The left image, from Jun. 01, 2015, shows a "Blockage" and two areas of "Landslide deposits" circled in yellow. The right image, from Oct. 24, 2016, shows that the previous landslide deposits have been eroded, and new features are labeled, including a "Landslide" and a "Boulder" in the debris-filled channel.
+
+**Figure 7.** Comparison of channel changes before and after GBTSC GLOF along the Zhangzangbo Valley. (a) The flow path of GBTSC GLOF; (b) GF-1 images showing that the flood left a large debris fan in the front of the lake, and eroded the moraine terrace; (c) lateral erosion and bank collapse in the moraine terrace and the width of channel increased; (d) landslide deposits distributed along the channel and blocked the channel before bursting, while GLOF eroded landslide deposits and triggered bank slump and landslide afterward, causing an increase in channel width. (Note: the 1 June 2015 images of c and d are from Google Earth, others are GF-1 images).
+
+# 6. Conclusions
+
+In this study a new detailed 2016 inventory of glacial lakes in the BKB was established and six unreported GLOF that occurred before 1984 have been detected with geomorphic outburst evidence based on high-resolution remote sensing images. The BKB is one of the most hotspot small river basins for glacial lakes and GLOF in the central Himalayas. High-resolution remote sensing images are useful for detecting unreported GLOF events in high mountainous regions and sparsely populated regions, which is conducive to improving the GLOF inventory and better assessing GLOF hazard. A rough but more comprehensive method was proposed to assess GLOF hazard, which considers the probability for a flood to develop into a debris flow in the downstream, where large numbers of landslides triggered by earthquake are distributed. The GLOF hazard in BKB increases due to landslide deposits volume, which increased approximately  $124.36 \times 10^6$  m<sup>3</sup> after the Gorkha earthquake, and 11 glacial lakes are identified as very high hazard, nine are high hazard, 32 are medium hazard and 55 are low hazard. However, about 24.6% of the all lakes' hazards would be underestimated without earthquake-induced landslides, in which most of them are small glacial lakes. Therefore, for regional GLOF hazard assessment, small glacial lakes should not be overlooked for landslide deposit entrainment along a flood route and flood eroding channel bed would increase the peak discharge, especially in earthquake affected areas where large numbers of landslides were triggered. We suggest
+
+Water 2020, 12, 464 17 of 20
+
+that more attention should be paid to the very high and high-hazard glacial lakes and to improving the engineering security standard for defending against flood hazards downstream of BKB.
+
+**Supplementary Materials:** The following are available online at http://www.mdpi.com/2073-4441/12/2/464/s1. Table S1: glacial lake inventory of Bhote Koshi.
+
+**Author Contributions:** N.C. and M.L. conceived the original ideas and drafted the original manuscript. Y.Z. and M.D. carried out field investigation and data collection. N.C. and M.L. revised the original manuscript. All authors have read and agreed to the published version of the manuscript.
+
+**Funding:** This research was funded by National Natural Science Foundation of China (Grant NOs. 41671112 and 41861134008) and the 135 Strategic Program of the IMHE, CAS (Grant NO. SDS-135-1705).
+
+Conflicts of Interest: The authors declare there no conflict of interest.
+
+# References
+

chatbot/rag/data/('Glacial Lake Outburst Flood (GLOF) Hazard and Risk Management Strategies', '.pdf')_extraction.md

@@ -0,0 +1,184 @@
+
+## **Abstract**
+
+GLOFs, driven by climate change-induced glacier melt, present a serious threat to downstream communities in mountainous regions. The present study addresses this pressing issue by employing a systematic methodology to analyze scientific literature on GLOF risk management and hazard mitigation. The research synthesizes key insights through thematic analysis, focusing on GLOF hazard mitigation, hazard and risk mapping, and the development of a methodological framework for managing GLOF hazards and risks. It evaluates structural measures such as spillways and diversion canals, and non-structural measures, including strategic land-use planning and community-based strategies. The study highlights the importance of Early Warning Systems (EWS), GLOF risk management skills, and knowledge transfer, emphasizing integration with climate change adaptation. It also discusses facilitating factors like policy legislation, institutional support, and addressing knowledge disparities. While the physical aspects of GLOFs have been widely explored, research on their social impacts is relatively limited, leaving a crucial gap in understanding the full extent of these disasters. The findings of the study offer a comprehensive resource for policymakers, planners, and disaster management professionals, providing a holistic approach to mitigating GLOF risks and enhancing resilience in vulnerable regions. This research is poised to significantly contribute to effective GLOF hazard mitigation and climate change adaptation strategies.
+
+**Keywords** GLOFs; Hazard and Risk; Management · Methodological Framework; EWS · Knowledge disparity · Policy legislation
+
+## 1 Introduction
+
+Glacial lakes, a prominent feature of the glacial landscape, form as a result of the dynamic interaction between the glacial processes and surrounding topography (Steffen et al. 2022; Peng et al. 2023). These glacial lakes originate from several processes and mechanisms, primarily involving the accumulation of glacial meltwater behind the ice dams, moraines
+
+Department of Geography and Disaster Management, School of Earth and Environmental Sciences, University of Kashmir, Srinagar 190 006, India
+
+Image /page/0/Picture/12 description: The Springer logo, featuring a black outline of a chess knight piece on the left and the word 'Springer' in a black serif font on the right, all on a white background.
+
+Rayees Ahmed
+Rayeesrashid84@gmail.com
+
+or bedrock barriers that obstruct the normal flow of the water (Zhang et al. 2022; Hu et al. 2022). The formation of the glacial lakes is closely linked to the glacier retreat during the deglaciation period (Molg et al. 2021; Rather et al. 2024a, b; Knight and Harrison 2018). These lakes vary in size, volume, depth, and morphology and can be broadly classified into four main types, such as moraine-dammed lake (formed when glacial debris creates a natural dam obstructing the meltwater flow), ice-dammed lake (forms when a glacier temporarily blocks a river impounding meltwater), glacier erosion lake (found in depressions on mountain slopes that have been glacially sculpted out by erosion) and other glacial lakes (Lakes formed within glaciated valleys primarily fed by the direct glacial melt) (Gupta et al. 2023; Ahmed et al. 2021). The glacial lake expansion is considered as one of the important indicators of the climate change (Nie et al. 2013; Zheng et al. 2021). The ongoing lake formation and expansion of different types of glacial lakes is directly linked to the warming climate and rapid melting of the glaciers (Pandey et al. 2021; Ahmed et al. 2022; Mir et al. 2022, 2024).
+
+GLOFs are serious natural hazards that could have devastating impacts on communities that inhabit downstream of glacial lakes. The chance of GLOF incidence is rising as a result of climate change (Harrison et al. 2018). The melting of glaciers leads to the formation of new lakes and increasing water levels, which raises the possibility of severe flooding (Benn et al. 2012; Zheng et al. 2021; Munoz et al. 2024). As such, the drafting of GLOF hazard and disaster management strategies has become a key issue for planners, policymakers and disaster management professionals. A multidisciplinary approach is necessary to design efficient GLOF hazard and disaster management strategies, drawing on knowledge from a variety of domains, such as hydrology, geography, disaster management, meteorology, glaciology, and engineering. These approaches need to take into account the different needs and vulnerabilities of different communities or social groups, as well as the complex interactions between physical, social, and economic factors.
+
+Installing EWS in glacial lake regions requires robust planning, accurate data transfer, and careful assessment of site-specific variables. The Laguna 513 EWS in the Peruvian Andes serves as a successful model (Huggel et al. 2020), highlighting the need for location-specific adaptation. Wireless sensor networks (WSNs) offer cost-effective, real-time monitoring for GLOF risk reduction (Singhal et al. 2022). In Sikkim, India, the GLOF EWS integrates GIS and real-time sensors for effective flood simulation and evacuation planning (Kumar et al. 2022). Effective risk management also demands community engagement and addressing institutional gaps to enhance preparedness and mitigation efforts in vulnerable regions. Recent studies emphasize the importance of risk communication, capacity building, and community involvement in GLOF disaster management. For instance, Pandey et al. (2021) identified key challenges in developing community-centered strategies in Nepal. Effective GLOF management requires cross-disciplinary collaboration and careful execution to protect vulnerable communities. Internationally, GLOFs are prioritized in the Sendai Framework for Disaster Risk Reduction, urging the development of regional and national mitigation plans (UNDRR 2015)
+
+Awareness of GLOFs has grown, leading to efforts like the NDMA's collaboration with Swiss agencies to create management guidelines and a global risk reduction program. These initiatives focus on hazard mapping, early warning systems, and community-based strategies. Effective GLOF management must consider the political and social challenges of remote, marginalized regions, incorporating local knowledge for sustainability. Hazard and disaster management plans need to take into account the distinct social, economic, and environmental conditions of the impacted communities in order to reduce the hazards associated with GLOFs. Comprehensive risk assessments, early warning systems,
+
+Image /page/1/Picture/5 description: The logo for Springer, featuring a black and white stylized chess knight icon to the left of the word 'Springer' in a black serif font, all on a white background.
+
+and emergency preparedness are all essential components of successful plans. Since local communities are frequently the first to respond to GLOF events, the solutions also need to empower and involve them. The efficacy and sustainability of GLOF hazard and disaster management systems can be improved by incorporating traditional knowledge and practices.
+
+## 2 Methodology
+
+This review-based study employed a methodical approach to thoroughly analyse the scientific literature on risk management, hazard mitigation, and GLOFs, with an emphasis on the consequences of glacial retreat brought on by climate change. The evaluation process was guided by well-crafted research questions that focused on important issues such as risk management techniques, GLOF hazard mitigation strategies, and the relationship between these factors and the effects of climate change in high-mountain areas.
+
+Explicit inclusion and exclusion criteria were developed in order to guarantee a strict selection procedure. Systematic search strings using keywords like "Glacial Lake Outburst Floods," "Hazard Mitigation," "Hazard Assessment," "GLOF modeling," "Risk Management," "Structural and Non-Structural measures," and "Climate Change" were conducted to find pertinent literature on reliable scholarly databases such as Scopus, google scholar, PubMed and Web of Science (Figure S1). Titles and abstracts were used in the first screening process to find research that met the predetermined criteria. This process eventually resulted in the selection of peer-reviewed articles, academic papers, and reports that precisely addressed the themes that were being investigated (Fig. 1).
+
+A comprehensive full-text analysis of the articles selected was carried out, with a focus on rigorous methodology and relevance assessment (Gao et al. 2024). During this procedure, important
+
+Image /page/2/Figure/7 description: A flowchart illustrating a six-step methodology for a literature review. The process starts with a box labeled "Literature Review" for the "Extraction of relevant research articles for thematic analysis". Step 1 is the "Selection of Keywords," which are categorized into GLOFs, Hazard and Risk Assessment, GLOF mitigation measures, and GLOF management and mitigation. Step 2 is the "Selection of relevant database," with a word cloud of terms like "glacial lake(s)" pointing to four databases: Web of Science, Scopus, Google Scholar, and PubMed. Step 3, "Extraction of research items," shows the number of articles found in each database: WOS (n=612), Scopus (n=590), Google Scholar (n=557), and PubMed (n=334). Step 4 is the "Selection of time frame," which is 1990-2023. Step 5 shows the result after filtering, with "Records after removing duplicate items n=823". Finally, Step 6 shows the final count after further filtering: "Records after removing conference papers, books, reports and book chapters n=685".
+
+**Fig. 1** Methodology flow chart for the extraction of research articles from 1990 to 2023. The red boxes show steps for generating the database for the Review analysis
+
+Image /page/2/Picture/9 description: The logo for Springer, featuring a black line drawing of a chess knight piece on the left, followed by the word "Springer" in a black serif font on a white background.
+
+data was extracted, such as the approaches used in each study, the primary conclusions, and any limitations noted by the authors. After the data were retrieved, a thematic analysis was conducted on the data to classify the studies according to common themes such as risk mapping, early warning systems, structural and non-structural strategies, and climate change adaptation. The synthesis of results was a critical step in developing a thorough understanding of GLOF hazard reduction and risk management measures. Based on the collective insights gleaned from the scholarly literature, the synthesis of knowledge from many investigations allowed for the formulation of clear conclusions and suggestions. The methodology relied heavily on critical reflection, which involved an analysis of the strengths and limitations inherent in the evaluated literature. This critical viewpoint aided in the construction of a well-structured review study by allowing for a sophisticated assessment of the state of knowledge on GLOF hazard reduction. The documentation of the whole review process, including search strategies, inclusion/exclusion decisions, and data synthesis methodologies, guaranteed the study's findings were transparent and reproducible.
+
+## 3 GLOF Hazard Mitigation
+
+GLOFs are a significant natural hazard in regions where glaciers are retreating due to climate change. Glacial lakes are formed when glaciers melt and are contained by unstable moraine dams composed of loose rock and debris (see Fig. 2c). A GLOF could happen if the dam's foundational stability is compromised by the pressure of the water behind it, causing a sudden and disastrous release of water and debris downstream. GLOFs have the potential to result in significant financial losses, infrastructure destruction, and fatalities. As a result, populations residing downstream of glacial lakes must implement efficient hazard mitigation measures. Effective GLOF hazard mitigation consists of a number of elements, such as hazard mapping, early warning systems, and structural and non-structural mitigation. The schematic representation of the GLOF triggering mechanism and downstream impacts is illustrated in Fig. 3.
+
+A crucial part of GLOF hazard mitigation plans is early warning systems. EWS enables the population downstream to warn in advance of potential GLOF events so they can take the necessary precautions and safety measures to reduce damage and casualties. Monitoring systems, communication networks, and dissemination systems are the parts of a typical early warning system. Physical measurements like lake level, water temperature, and seismic activity can be included in monitoring systems. The state of the moraine dam and the extent and volume of the glacial lake can both be usefully inferred from remote sensing techniques like satellite imagery and aerial photography. To ensure that warnings are efficiently transmitted to populations downstream, communication networks are essential. Radios, TVs, cell phones, and sirens are examples of communication networks. All community members must be able to access and rely on the communication networks, regardless of their socioeconomic level or place of residence. Traditional media, like radio and newspapers, as well as new media, like social media and smartphone apps, can be used as dissemination platforms. Reaching a wider audience through the usage of new media is especially beneficial for younger generations who are more inclined to use social media and mobile apps. A number of countries have successfully put in place EWS for GLOFs. The Department of Hydrology and Meteorology in Nepal, for instance, has reduced the effects of GLOFs in the country by implementing the EWS for GLOFs (Pandey et al. 2021). To offer early warning of possible GLOF incidents, the system combines monitoring systems, communication networks, and dissemination systems.
+
+Image /page/3/Picture/5 description: The image displays the logo for the publisher Springer. The logo consists of a black icon of a knight chess piece on the left, followed by the word "Springer" in a black serif font on the right, all set against a white background.
+
+Image /page/4/Figure/2 description: A figure titled 'Glacial Lake Outburst Flood (GLOF) Hazard and Risk Management...' is presented in three panels labeled a, b, and c. Panel (a) is a schematic diagram illustrating a GLOF monitoring and warning system. It shows a satellite using remote sensing to monitor a glacial lake, with data on 'Hazard Potential' sent to a server. Downstream, gauge and warning stations along a river send 'Alarm' signals to the server. The server then alerts 'End Users'. The system also includes regular monitoring via remote sensing and GIS to assess risk, ultimately leading to a 'GLOF Warning' for communities. Panel (b) is a circular diagram detailing the 'GLOF hazard and Risk Management' cycle. The cycle includes 'Risk Analysis', 'Disaster Risk Reduction' (with measures like lowering lake water level and engineering solutions), 'Response' (including Early Warning Systems and emergency measures), and 'Recovery' (including relief, rehabilitation, and reconstruction). Panel (c) is a visual depiction of a GLOF event, showing a 'Moraine dammed Lake' at a 'Breaching point'. The resulting 'Flood water' and 'Debris flow' travel down through forests and valleys, impacting 'Agriculture fields', a 'Bridge', and a 'Settlement' downstream.
+
+Fig. 2 a) A Schematic diagram illustrating the location of monitoring stations and the monitoring/warning processes, b) Structural and Non-structural measures for the GLOF Hazard and Risk management, c) Illustration of GLOF event and downstream impacts
+
+## 4 Hazard and Risk Mapping
+
+Hazard mapping is another key element for GLOF hazard mitigation. The mapping of various parameters related to GLOF hazard and Risk in the downstream region such as location of the glacial lake and its feeding glacier, moraine dams, ice cores, mass movements, evolution of the lake, downstream area, population, infrastructure, potential flood propagation
+
+Image /page/4/Picture/6 description: The logo for Springer, featuring a black and white icon of a chess knight's head facing left, positioned above two horizontal lines. To the right of the icon, the word "Springer" is written in a black serif font. The background is white.
+
+Image /page/5/Picture/1 description: A diagram illustrating the causes and effects of a Glacial Lake Outburst Flood (GLOF). The diagram is split into two main sections: 'GLOF Triggering Mechanism' at the top and 'Downstream Impacts' at the bottom. The top section shows a mountainous, glacial environment with a large blue glacial lake contained by a moraine. Several triggering mechanisms are shown with red arrows pointing into the lake, including a rock fall, a snow avalanche, ice calving from a glacier, and a cloud burst. The bottom section, labeled 'Downstream Impacts,' shows the floodwater rushing from the breached lake into a valley below. This flood impacts a community, showing houses, agricultural fields, forests, a bridge, and a more developed area with a factory and a road. People are depicted in an orange raft on the floodwaters. A legend on the right side of the diagram explains the various symbols used: Agricultural field, Forest, Bridge, House, Glacial lake, Moraine, Glacier Snout, Snow avalanche, Rock fall, Potential flood from lake in the upstream, Ice calving, Ice cored moraine, and Cloud burst.
+
+Fig. 3 GLOF triggering mechanism and downstream impacts
+
+parameters etc is called as Hazard and Risk mapping. The mapping can be used to create risk evaluations and provide guidance for creating plans to mitigate hazards. The process of GLOF hazard and risk can be carried out using a combination of Remote sensing techniques and hydrodynamic models such as satellite imagery, aerial photographs, field investigations, etc. Besides physical parameters, the mapping process should also take into consideration the social and economic factors that increase the vulnerability of the downstream population. For instance, the National Centre for Hydrology and Meteorology in Bhutan has created an inventory of glacier lakes that contains details about the lakes' locations, sizes, and potential hazards (Bajracharya et al. 2020). The hazard maps created from the inventory have influenced the creation of GLOF hazard mitigation plans in Bhutan. Several studies have been carried out in different parts of the world on the GLOF hazard and Risk assessment. For example, Ives et al., 2010 analysed the evolution of glacial lakes and risk assessment in the Hindu Kush Himalayas, Wang et al. 2022 have done an integrated assessment of GLOF disaster in Qinghai Tibetan Plateau. Similarly, in Nepal GLOF risk assessment was carried out by various researchers (Shrestha et al. 2010; Khanal et al. 2015; Rounce et al. 2016; Hu et al. 2022; Khadka et al. 2019; Sattar et al. 2021; Khanal et al. 2015), Bhutan (Ives et al. 2010; Rinzin et al. 2023; Gurung and Khanal 2017; Hagg et al. 2021; Scapozza et al. 2019; Rinzin et al. 2021; Rawat et al. 2023b; Negi et al. 2021; Ahmed et al. 2023; Wangchuk et al. 2023). A boom of studies on the GLOF risk assessment and Inventorization of glacial lakes have been conducted in several catchments of the Indian Himalayan Region (Sattar et al. 2021; Ahmed et al. 2021, 2022; Mir et al. 2018; Majeed et al. 2021; Rawat et al. 2023; Aggarwal et al. 2016; Mal et al. 2021; Dubey and Goyal 2020; Deswal et al. 2021; Gupta et al. 2022; Rinzin et al. 2023; Thakur et al. 2016). In the Peruvian Andes GLOF hazard and risk have been studied extensively for instance Frey et al. (2018) have proposed a robust GLOF management strategy for a data-scarce region like Santa Teresa in Peru. Emmer and Vilmek (2014) have provided a new method to evaluate the potential danger of glacial lakes in Peru. Similarly, a huge number of studies have been carried out in the region which have focused on the preparation of lake and
+
+Image /page/5/Picture/4 description: The black and white logo for Springer, featuring a stylized chess knight icon to the left of the word 'Springer' in a serif font.
+
+GLOF inventories, hazard level, risk assessment as well as structural and non-structural measures for the GLOF hazard and risk management (Drenkhan et al. 2019; Ahmed et al. 2022; Frey et al. 2018; Huggel et al. 2020; Emmer and Vilmek 2014; Munoz et al. 2024; Hagen et al. 2023; Batka et al. 2020; Huggel et al. 2020; Carey et al. 2012; Wood et al. 2021). From the above-mentioned studies, it is quite clear that researchers are long been working on the glacial lake hazard and risk mapping, however, only findings from a few studies have been implemented for the planning on the ground. These risk maps along with other useful findings should be taken into consideration while planning and developmental activity in the ecologically fragile areas to lessen the impacts of the potential GLOFs.
+
+## 5 Methodological Framework for GLOF hazard and Risk Mapping
+
+The development of a comprehensive methodological framework is one of the key elements for evaluating and reducing the possible risks associated with GLOFs. An extensive literature survey was conducted to build this framework, ensuring a strong foundation based on current knowledge and techniques. The initial phase includes the detection and Inventorization of glacial lakes at different scales, employing field surveys and data from remote sensing to create an extensive database. The dynamics of glacial lakes, such as changes in volume, size, and possible triggers for outburst events, are then evaluated using modeling approaches and remote sensing technology. The comprehensive methodological framework for GLOF hazard and risk mapping is illustrated in Fig. 4. Framework for the evaluation of hazard level, physical vulnerability and social vulnerability of GLOFs is illustrated in Supplementary Figure S3.
+
+The following stage is dedicated to hazard assessment and includes identifying susceptible downstream locations, estimating the volume of the flood, and analysing possible flood routes. Simulating possible GLOF scenarios requires the use of sophisticated hydrological and hydraulic modelling techniques in conjunction with GIS. After that, risk mapping combines vulnerability and exposure analyses with hazard data, taking into account downstream areas' infrastructure, population density, and land use. This stage helps to prioritise risk reduction efforts by giving a spatial representation of the potential impact zones.
+
+Stakeholder participation is crucial to this process; local communities, governmental organisations, and scientific specialists should all be included. Their contributions improve data quality, evaluate models, and add useful local knowledge. The developed methodological framework serves as a crucial resource for planners, policymakers and disaster management authorities. Targeted policy interventions, such as land-use planning, early warning systems, and infrastructure development in high-risk zones, are made possible by the spatial visualisation of prospective risk areas. Decision-makers can use the framework's results to help create and carry out policies that lessen communities' susceptibility to GLOF disasters. Additionally, the methodological framework provides a systematic and standardized approach to GLOF hazard and risk assessment, ensuring consistency and comparability across different regions and contexts. Adaptive management strategies can be improved by regular updates that are based on continuous observation and investigation. In conclusion, strengthening policy implementation and protecting vulnerable communities from the potentially disastrous effects of GLOFs depends on the development of a strong GLOF hazard and risk mapping framework that is backed by scientific data and stakeholder cooperation.
+
+Image /page/6/Picture/7 description: The logo for the publisher Springer, featuring a black outline of a chess knight's head facing left, positioned above two horizontal lines. To the right of the icon is the word 'Springer' in a black serif font, all on a white background.
+
+Image /page/7/Figure/1 description: A flowchart illustrating a framework for the hazard and risk mapping of glacial lakes. The process begins with inputs like Satellite imagery, Maps, Reports, Past inventories, and Toposheets, which are used for 'Mapping glacial lakes at different time points' with the help of 'Geospatial tools (RS and GIS)'. This leads to the creation of a 'Glacial lake data base'. The next step is the 'Identification of PDGLs' (Potentially Dangerous Glacial Lakes) using 'Integrated criteria' and 'Geospatial tools'. This generates a 'List of PDGLs'. These are then ranked for hazard level using various 'Ranking Criteria' such as 'Lake area, volume and Peak discharge', 'Moraine condition', 'Past GLOF history', and others. The output is a 'Hazard Level' categorized as High, Medium, or Low, determined using the 'Natural Jenks Classification method in ArcGIS'. Following this, 'high priority lakes' are selected for further assessment. This involves an 'In-depth analysis through field investigation' covering eight 'Thematic areas of study' including 'Moraine dam condition' and 'Analysis of the downstream agriculture and infrastructure'. This detailed study, including field investigation and modeling, results in 'Inundation maps, risk maps for the planners and policy makers'. After 'Analysis and interpretation of data', the process culminates in a 'Final Report'. The flowchart includes small map images as examples of inundation maps.
+
+Fig. 4 Framework for the hazard and risk mapping of glacial lakes
+
+## 6 Structural Measures
+
+The structural measures involve the construction and development of physical infrastructure to minimise the risk associated with the potentially dangerous glacial lakes. Structural measures comprise of the construction of spillways, diversion channels, siphoning as well as other engineering approaches to manage the rapid flow of water and debris material in the event of a GLOF. The effective implementation of the structural measures depends on the location and characteristics of the moraine dam. The effectiveness is also determined by the availability of funds and technical expertise. The following steps outline the key components of this framework.
+
+Image /page/7/Picture/5 description: The black and white logo for Springer. On the left is an icon of a knight chess piece, a horse's head facing left, with two horizontal lines underneath. To the right of the icon is the word "Springer" in a serif font.
+
+### 6.1 Site-Specific Vulnerability Assessment
+
+Conducting a comprehensive site-specific vulnerability assessment is the first step towards identifying regions that are vulnerable to GLOF events. This entails identifying the most vulnerable areas by carefully examining glacial lakes, possible flood routes, and downstream areas. Essential information is gathered by field surveys and data from remote sensing.
+
+### 6.2 Infrastructure Development and Engineering Solutions
+
+The framework presents engineering solutions that are customised to the specific characteristics of every risk zone that has been identified, taking into account the results of the vulnerability assessment. To prevent flooding, structural measures could involve the construction of retaining walls, debris dams, and other barriers. Floodwaters can also be redirected away from inhabited areas by channelizing rivers and creating diversion channels.
+
+### 6.3 Early Warning Systems and Monitoring Infrastructure
+
+To give at-risk communities timely warnings, the framework places a strong emphasis on the integration of early warning systems and monitoring infrastructure. Real-time data collection is facilitated by the installation of sensors, weather monitoring equipment, and gauging stations close to glacial lakes. When combined with communication networks, these methods allow authorities to alert people in good time, which facilitates preparation and efficient evacuation (Fig. 2a).
+
+### 6.4 Community Involvement and Capacity Building
+
+Local communities must be actively involved for structural measures to be implemented successfully. The framework emphasises how crucial community involvement and capacity-building initiatives are. These programmes inform locals about possible dangers, how to evacuate, and how crucial it is to keep an eye on and maintain the structural safeguards put in place.
+
+## 7 Non-Structural Measures
+
+Non-structural measures emphasise on community-based strategies for hazard mitigation rather than involving physical infrastructure. Land-use planning, public awareness campaigns, and community readiness and response planning are examples of non-structural measures. Planning for community readiness and response entails collaborating with nearby communities to create strategies for tackling GLOF events. Plans for evacuation routes, emergency shelters, and protocols for community member collaboration and communication can all be included in these plans. Campaigns to raise public awareness are essential to ensuring that communities are aware of the dangers posed by GLOFs and are equipped to respond to emergencies. Media campaigns, instructional materials, and community gatherings are a few examples of public awareness initiatives. By making sure that future development considers the possibility of GLOF risks land-use planning may
+
+Image /page/8/Picture/12 description: The image displays the logo for Springer, a publishing company. The logo consists of a black icon and black text on a white background. On the left, there is an icon of a chess knight's head facing left, positioned above two short horizontal lines. To the right of the icon, the word "Springer" is written in a serif font.
+
+contribute to reducing the risk of GLOFs. Zoning regulations, construction restrictions, and other measures to lessen a community's susceptibility to GLOFs are examples of landuse planning (Fig. 2b).
+
+Although GLOF hazard mitigation approaches are important, there are a number of obstacles that prevent them from being implemented effectively. The isolation and political marginalisation of numerous communities at danger from GLOFs is one of the primary obstacles. The lack of finances, technical know-how, and political clout in these communities can make it challenging to put into practice sensible hazard mitigation plans. The lack of funds for GLOF hazard mitigation measures is another issue. Low-income nations with little funding for hazard mitigation and infrastructure development make up a large number of those at risk from GLOFs. Notwithstanding these obstacles, there exist prospects for enhancing GLOF hazard reduction tactics.
+
+GLOF risk reduction involves several strategies to mitigate the catastrophic impacts of GLOFs, which are becoming more frequent due to climate change and glacier retreat (Shijin et al. 2022; Taylor et al. 2023). Key measures include infrastructure development (e.g., spillways, dams), hazard mapping, early warning systems, and land use planning (Ahmed et al. 2024). Effective risk reduction also requires educating and involving local communities, as well as addressing the underlying causes of GLOFs, such as rising temperatures and glacial retreat. International cooperation is essential, particularly in transboundary regions, with organizations like ICIMOD facilitating collaborative efforts to develop and implement regional risk reduction strategies.
+
+## 8 Early Warning System (EWS)
+
+An EWS is a vital component of GLOF risk reduction. It is intended to identify and monitor changes in glacial lakes and to send alerts to people downstream in the event of a potential GLOF. EWS can give fast and accurate information to decision-makers and local populations, allowing them to take necessary action to mitigate the impact of a GLOF. EWS usually involves a range of monitoring devices, like water level gauges, weather stations, and seismometers, that are installed in and around glacial lakes to track changes in water level, temperature, and seismic activity (Fig. 2a). This monitoring data is sent to a central control center, where it undergoes analysis and processing to determine the likelihood of a GLOF. When a GLOF is detected, the EWS sends an alarm to downstream populations through different means of communication such as sirens, text messaging, and social media. The alert informs individuals about the location and timing of the anticipated flood and urges them to escape to safe areas. In addition to alarms, EWS can give real-time information regarding the extent and magnitude of a GLOF, allowing emergency personnel to better organize their response and deploy resources (Fig. 2a). The efficacy of an EWS, on the other hand, is dependent upon aspects, including the reliability and quality of monitoring data, the timeliness and accuracy of alerts, and the ability of local people to respond to alerts and escape to safe locations. To reduce the impact of GLOFs, EWS should be supplemented with other risk reduction measures such as risk mapping, land use planning, and infrastructure development.
+
+Image /page/9/Picture/6 description: The logo for the publisher Springer, featuring a black line drawing of a chess knight icon to the left of the word 'Springer' in a black serif font, all on a white background.
+
+## 9 GLOF Risk Management Skills and Knowledge Transfer
+
+Effective GLOF risk reduction hinges on knowledge transfer and skill development. This involves sharing expertise and best practices among stakeholders—scholars, policymakers, emergency responders, and local communities—to enhance local capabilities in recognizing GLOF indicators, responding to alerts, and evacuating safely. Knowledge transfer improves the accuracy of risk assessments through data sharing on glacial dynamics and climate effects, and by incorporating lessons from past GLOF events. It also fosters global collaboration, crucial for managing risks in transboundary regions. By focusing on capacity building, data exchange, and international cooperation, communities can build resilience, refine risk assessments, and implement effective mitigation strategies.
+
+## 10 GLOF Hazard Mitigation and Climate Change Adaptation
+
+Climate change impacts, such as rising temperatures and changing patterns of precipitation, are increasingly being attributed to GLOFs, which may lead to the development and instability of glacial lakes (Zheng et al. 2021). As a result, measures to lower the danger of future GLOFs should be closely integrated with climate change adaptation initiatives. One important strategy for GLOF hazard prevention and climate change adaptation is to design and build infrastructure, such as dams and spillways. These engineering structures can aid in regulating the flow of water from glacial lakes, lowering the risk of massive flooding. However, such infrastructural solutions must be carefully developed and planned to ensure long-term effectiveness and sustainability.
+
+Another strategy is to encourage ecosystem-based adaptation (EbA) initiatives, which can assist in lowering the risk of GLOFs and increase community resilience to the effects of climate change. Restoration of deteriorated habitats, such as wetlands, marshes, bogs, and forests, may assist with managing the flow of water and lessen the risk of floods as part of EbA efforts. Furthermore, EbA strategies can help to strengthen local populations' capacity to adapt to climate change impacts and resilience to GLOFs. Furthermore, successful GLOF hazard reduction and climate change adaptation necessitate the development of integrated risk management strategies that address the root causes and drivers of GLOFs, such as climate change impacts, glacier retreat, and human activities. To develop and implement appropriate risk reduction strategies, different stakeholders, including researchers, policymakers, emergency responders, and local communities, must work together.
+
+## 11 Facilitating Factors for GLOF Hazard Mitigation and Climate Change Adaptation
+
+GLOFs are a severe threat throughout mountainous areas, particularly where glacial lakes exist. Glaciers are melting as a result of climate change, increasing the likelihood of GLOFs and posing a severe threat to people, infrastructure, and the environment. Several supportive factors can promote GLOF hazard mitigation and climate change adaptation activities to decrease the risks associated with GLOFs and adapt to the impacts of climate change. Several
+
+Image /page/10/Picture/9 description: The logo for Springer, featuring a black line drawing of a chess knight piece to the left of the word 'Springer' in a black serif font, all on a white background.
+
+actions can help to mitigate GLOF hazards and adapt to climate change. A few examples are given below:
+
+- Strong political will
+- Collaboration and partnerships
+- Engagement and the empowerment of local people
+- Adequate funding
+- Integration with other development goals
+- Access to technology and data
+- Awareness and education
+
+## 12 Policy Legislation and Institutional Support
+
+Effective GLOF hazard mitigation and climate change adaptation require strong policy, law, and institutional support. Governments and other stakeholders have to develop and enforce laws and regulations to support risk management plans, provide needed resources, and encourage transparency and accountability. Furthermore, organizations must be built or enhanced to allow policy implementation and risk reduction measures. Climate litigation is being informed by interdisciplinary research on GLOF hazards, particularly concerning anthropogenic climate change. Notably, a Peruvian resident is suing a German energy firm for possible GLOF dangers; the case is presently pending before a German court. Research relates GLOF dangers and glacier changes to greenhouse gas emissions. Even though these incidents have an impact, it's important to take into account larger viewpoints on accountability and justice. An analysis of the Palcacocha case (Figure S2) highlights the different duties at different levels and highlights the impact of governance, socioeconomic conditions, cultural features, and climatic and non-climatic elements in creating exposure and vulnerability to GLOF threats.
+
+- Policy and legislation
+- Institutional support
+- Financial support
+- Capacity building
+- Monitoring and assessment
+
+Effective GLOF and climate change management relies on robust policies, institutional support, financial resources, capacity building, and monitoring. Policies should reduce emissions and promote sustainable practices, while institutions coordinate efforts and provide technical support. Financial backing is essential for infrastructure and early warning systems, sourced from various channels. Capacity building enhances stakeholder skills through targeted training. Comprehensive monitoring ensures progress and informs improvements in risk management strategies.
+
+## 13 Knowledge Disparity
+
+Knowledge disparity, the uneven distribution of information and expertise, significantly impacts GLOF hazard mitigation and climate change adaptation. Communities lacking access to scientific data, technical skills, and early warning systems face challenges
+
+Image /page/11/Picture/19 description: The Springer logo, featuring a black and white icon of a chess knight piece to the left of the word 'Springer' in a black serif font, all on a white background.
+
+in addressing GLOF risks. Financial constraints further exacerbate this issue, limiting the ability to implement protective measures and adapt strategies. Institutional capacity gaps, such as inadequate specialized agencies or poor coordination, also hinder effective responses. Additional factors like language barriers, traditional knowledge, and gender roles can contribute to knowledge imbalances, affecting the success of risk reduction efforts. Addressing these disparities requires a collaborative approach, including information sharing, institutional strengthening, and social inclusivity, to enhance community understanding and resilience against GLOFs and climate change.
+
+## 14 Conclusion
+
+This study underscores the critical need for a comprehensive framework to mitigate GLOF hazards, particularly in the face of accelerating glacier retreat driven by climate change. By integrating both structural and non-structural measures, the research highlights the importance of proactive land-use planning, community-based strategies, and early warning systems (EWS) in managing GLOF risks effectively. The study's emphasis on detailed risk mapping and spatial analysis equips decision-makers with the tools necessary for targeted interventions, ensuring resources are allocated efficiently in highrisk areas. The implications of this study extend to policymakers, practitioners, and researchers, providing a roadmap for developing sustainable and adaptable solutions to the complex challenges posed by climate change in high-mountain regions. Legislative support, institutional capacity, and knowledge transfer are identified as crucial elements in enhancing the effectiveness of GLOF risk mitigation efforts. Addressing knowledge disparities and fostering interdisciplinary collaboration are essential for building resilient communities. Looking forward, the study advocates for ongoing research, particularly in understanding the socio-economic impacts of GLOFs, to ensure a holistic approach to disaster risk reduction. By integrating climate change adaptation into GLOF mitigation strategies, the findings offer a pathway for safeguarding vulnerable communities and preserving the stability of high-mountain environments.
+
+**Supplementary information** The online version contains supplementary material available at https://doi.org/10.1007/s11269-024-03958-x.
+
+**Funding** No funding was received to conduct this study.
+
+Data Availability Not applicable.
+

chatbot/rag/data/('Glacial lake outburst flood risk assessment using remote sensing and hydrodynamic modeling', '.pdf')_extraction.md

@@ -0,0 +1,309 @@
+
+
+# Abstract
+
+Glacier-associated hazards are becoming a common and serious challenge to the high mountainous regions of the world. Glacial lake outburst floods (GLOFs) are one of the most serious unanticipated glacier hazards, with the potential to release a huge amount of water and debris in a short span of time, resulting in the loss of lives, property, and severe damage to downstream valleys. The present study used multi-temporal Landsat and Google earth imageries to analyze the spatio-temporal dynamism of the selected glacial lake (moraine-dammed) in the Satluj basin of Western Himalayas. Furthermore, GLOF susceptibility of the lake was assessed using a multi-criteria decision-based method. The results show that the lake area has increased from 0.11 to 0.26 km<sup>2</sup> over the past 28 years from 1990 to 2018. The susceptibility index value for the lake was calculated as 0.75, which indicates that the lake is highly susceptible to the GLOF. The depth and volume of the lake were estimated to be 16 m and 57×10<sup>5</sup> m<sup>3</sup>, respectively, using an empirical formula. HEC-RAS, HECGeo-RAS, and Arc-GIS software were utilized in this study to perform unsteady flow analysis and to determine the GLOF impact on the downstream area. The worst-case GLOF scenario (breach width of 75 m) was revealed during an overtopping failure of the moraine dam, resulted in a peak discharge of  $4060 \text{ m}^3$ /s and releasing a total water volume of  $57 \times 10^5 \text{ m}^3$ . The breach hydrograph has been routed to calculate the spatial and temporal distribution of peak flood, inundation depth, velocity, water surface elevation, and flood peak arrival time along the river channel. The analysis further reveals that the routed flood waves reach the nearest settlement, i.e., Rajpur town, situated at a distance of 102 km in the downstream valley of the lake at 6 h after the beginning of the lake breach event with a peak discharge/flood of 1757 m<sup>3</sup>/s and maximum flow velocity of 1.5 m/s. With the ongoing climate warming and glacier retreat, moraine-dammed lakes are becoming more hazardous and thus increasing the total threat. Therefore, it is mandatory to monitor and assess such lakes at regular intervals of time to lessen the disastrous impacts of GLOFs on the livelihood and infrastructure in the downstream valleys. The findings of this study will aid in the creation of risk management plans, preparatory tactics, and risk reduction techniques for GLOF hazards in the region.
+
+ $\textbf{Keywords} \ \ Glacier \ recession \cdot Glacial \ lake \ expansion \cdot GLOF \ susceptibility \cdot Glacial \ lake \ outburst \ flooding \ (GLOFs) \cdot HEC-RAS \cdot Western \ Himalaya-India$ 
+
+# Introduction
+
+The glacierized basins of the Himalayan region have witnessed a serious effect of the changing climate, and even little changes in climate have a wide range of consequences/implications for the Himalayas and its ecology (Ahmad
+
+Responsible Editor: Philippe Garrigues
+
+Rayees Ahmed rayeesrashid84@gmail.com
+
+Extended author information available on the last page of the article
+
+et al. 2022; Gurung et al. 2021; Wani et al. 2022). The ongoing climate change/variability has resulted in the significant mass loss of the glaciers throughout the high terrains of the world and several among them are under threat (Kumar et al. 2020; Ahmed et al. 2021a; Sarkar et al. 2020); consequent to that, a large number of glacial lakes have shown a discernible increase in terms of number and area (Bolch et al. 2012). The meltwater from these retreating glaciers not only increases the water level of the existing lakes, but it has also led to the formation of thousands of new lakes (Mir et al. 2018; Ahmad et al. 2022). For example, over 2070 lakes were formed in the Nepal Himalayas alone during the second
+
+Image /page/0/Picture/15 description: The logo for Springer, featuring a black line drawing of a chess knight piece facing left, positioned on two horizontal lines. To the right of the icon, the word "Springer" is written in a black serif font. The background is white.
+
+half of the twentieth century as a result of changing climatic regimes (Maharjan et al. 2018). Some of these lakes are potentially dangerous/disastrous and pose a serious threat to the life and property in the downstream region.
+
+Snow and glacier melt in the Himalayas is caused by a local weather system driven by mountain-induced dynamic and thermodynamic processes that affect precipitation pattern, cloudiness, and other factors, forming glacial lakes near the terminus (proglacial lakes), on the surface (supraglacial lakes), or within the glaciers (subglacial lakes). Due to the fast accumulation of glacier meltwater in these lakes, the effects of rapid flash flooding in riparian areas downstream have been disastrous (Richardson and Reynolds 2000; Jain et al. 2012). The melting of mountain glaciers and the growth of massive glacial lakes are two of the most visible and dramatic effects of climate change in this environment (Bolch et al. 2012). The inventory of glacial lakes has shown a total of around 2000 glaciers (Prasad et al. 2019) and 769 glacial lakes in the Satluj basin (Randhawa and Sharma 2013). Similar to other parts of the Himalayas, glaciers in the Satluj basin are retreating and glacial lakes are expanding at a greater pace (Gaddam et al. 2016; Mandal and Sharma 2020; Singh et al. 2021) and are expected to lose 53% and 81% of the area at the end of the century under the RCP 8.5 scenarios of CNRMCM5 and GDFDL-CM3 models (Prasad et al. 2019). As a result, the size and number of supraglacial and proglacial lakes that merge to form larger glacier lakes behind the loose moraine may grow faster (IPCC 2012; Worni et al. 2013; Emmer 2017; Harrison et al. 2018). Moraine-dammed glacial lakes have a higher risk of breaching, making them potentially dangerous due to the massive volume of water accumulated. Large outburst flood (GLOF) and risks can be produced in downstream areas during any sort of breaching and rapid discharge of water and debris from such lakes (Richardson and Reynolds 2000; Westoby et al. 2014; Worni et al. 2014). GLOF represents a catastrophic event in the Himalayas and is common in the glacierized areas of the region, thus posing a potential risk to the downstream hydro-power plant projects due to the extreme flow of a river to cause great damage. GLOF phenomenon is related to complex flow characteristics in both spatial and temporal scales (Westoby et al. 2014).
+
+From the last few decades, the Indian Himalayas have witnessed a large number of glaciers associated events, for example, Chog Kumdan GLOF event (1929) in Shyok River basin (Gunn 1930), Chorabari GLOF event (2013) in Mandakini basin (Allen et al. 2016), and recent being Chamoli disaster (2021) in Uttarakhand (Shugar et al. 2021) are some of the events well documented in the scholarly literature. These disastrous GLOF incidents are an example of calamitous GLOF events that have caused havoc on people's lives and livelihoods (Carrivick and Tweed 2016). According to the study carried out by Dubey and Goyal
+
+Image /page/1/Picture/5 description: A block of black text on a white background. The text reads: "2020, in Indian Himalayas, there are around 23 glacial lakes that are at very high risk and 56 in the high-risk category. Nonetheless, the risk posed by multiple triggering mechanisms should be estimated, particularly in the Himalayan region, where 68 percent of hydropower projects are located on probable GLOF tracks (Schwanghart et al. 2016)." The numbers "2020" and "2016" are in blue text.
+
+In the year 2000, the Satluj valley witnessed a devastating flood caused by cloud bursts and the resultant GLOF events, causing extensive damage in downstream areas. Thus, the region needs continuous monitoring and assessment of lakes that have the potential to cause a GLOF event (Randhawa et al. 2021). Preparation of an inventory and identification of lakes as potentially dangerous is the initial step for the hazard assessment of the lakes, followed by the GLOF modeling. To prevent and monitor the GLOF hazard and analyze the damage that may occur in the future, the integration of remote sensing and GIS could play an important role in identifying dangerous lakes and monitoring GLOF events in real-time. Geographic information system (GIS) and remote sensing (RS) technology have been successfully used globally for the monitoring and assessment of glaciers, glacial lakes, and associated hazards. The GIS and RS-based tools and techniques allow rapid investigation of huge glaciated areas in a very inexpensive way and are therefore considered very suitable for glacial lake and GLOF studies (Imdad et al. 2022; Ahmed et al. 2021b). Keeping in view the above discussion on how climate change is inducing glacier retreat and thereby glacial lake expansion and increasing hazard potentiality of the GLOFs, the aim of the present study is to achieve the following objectives.
+
+- To analyze the spatio-temporal changes in a selected lake for the GLOFs study using multi-temporal satellite imageries from 1990 to 2018.
+- To assess the GLOF susceptibility of the lake using a multi-criteria decision-based method.
+- To perform flood routing assessment of glacial lake hazard through hydrodynamic model (HEC-RAS), the study incorporates the following: (i) estimate the volume of a lake using empirical equation, (ii) evaluate multiple dam breach scenarios on the basis of varying breach width and breach formation time, producing breach outflow hydrograph, (iii) one dimensional (1-D) hydrodynamic mathematical modeling to evaluate the hazard potential of the selected lake, and (iv) a detailed study on flow hydraulic of potential GLOF and its downstream impact at different sites along the river.
+- The main objective of this research is to create a flood inundation map by employing the HEC-RAS model, RS and GIS, including predictions of water spread, maximum flood depth, travel duration, and peak discharge owing to the GLOF hazard.
+
+Image /page/1/Picture/11 description: The Springer logo, featuring a black outline of a chess knight piece to the left of the word 'Springer' in a black serif font, all on a white background.
+
+# Study area
+
+The present study has selected a potentially dangerous glacial lake from the Satluj basin of the Indian Himalayas to carry out a GLOF modeling using HEC-RAS. The lake is located between 31° 39′ 36.47″ N Latitude and 78° 09′ 59.85″ E longitude at an elevation of 4267 m above sea level (Fig. 1). The lake is dammed by an unconsolidated moraine and connected to the snout of the feeding glacier (compound basin). The nearest settlement which is vulnerable to the probable GLOF event is located at a distance of approximately 8.5 km in the downstream region. The HPPCL Kashang hydropower project is also located in the downstream region of the lake at a distance of 39.06 km from the outlet. The snout of the glacier (feeding lake)
+
+has shown a retreat of 153 m at the rate of 7.6 m/year from 2000 to 2020. The current area of the lake and its feeding glacier is estimated to be 0.26 km<sup>2</sup> and 5.72 km<sup>2</sup>, respectively. The glacier meltwater from this lake is supplied to the Satluj River. Satluj River is one of the key rivers in the Indus system. The Satluj River rises in Tibet at Manasarovar and ends in India near Namigia and it runs through the two states of India, i.e., Punjab and Himachal Pradesh. Satluj reaches the Anandpur Dun, a valley area between Sivaliks and Himalayan outer range near Nangal Town. The earlier studies have focused on the mapping, evolution, identification, and hazard assessment of glacial lakes. We have tried to evaluate the impacts of GLOF in the downstream valley and GLOF was routed from the lake outlet to the last selected location (Rajpur town), covering a distance of 102 km (Fig. 1).
+
+Image /page/2/Figure/5 description: A figure composed of four maps, labeled a, b, c, and d, showing the location of a selected lake for GLOF (Glacial Lake Outburst Flood) hazard assessment. 
+
+Map 'a' displays a map of India with the state of Himachal Pradesh highlighted in red. The map includes latitude and longitude lines and a scale bar in kilometers.
+
+Map 'b' is an elevation map of Himachal Pradesh, showing the Sutlej basin outlined in black. The legend indicates elevation ranges from a low of 243 meters to a high of 6619 meters. Major rivers like the Chenab, Sutlej, Yamuna, and Ganges are labeled.
+
+Map 'c' provides a more detailed elevation map of the Sutlej basin, outlined in red. The elevation ranges from a low of 454 meters to a high of 6446 meters. It shows the location of a specific lake, Rajpur town, and HPPCL Karcham. Rivers such as the Sutlej, Chenab, and Indus are also shown.
+
+Map 'd' is a close-up satellite image of the lake. The lake is outlined in red and is fed by a glacier, shown in cyan. Key features are labeled, including the 'Glacier', its 'Snout', the 'Lake' itself, the 'Moraine-dam' that contains the lake, the 'Lake outlet', and the resulting 'Stream'.
+
+Fig. 1 Location map of the selected lake for GLOF hazard assessment, (a) Himachal Pradesh on map of India, (b) Satluj basin on the map of Himachal Pradesh, (c) Lake location on the Satluj basin map and (d) Glacio-geomorphological features around the lake
+
+Image /page/2/Picture/7 description: The logo for Springer, featuring a black and white line drawing of a chess knight icon to the left of the word "Springer" in a black serif font, all on a white background. The knight icon is facing left and has two horizontal lines underneath it.
+
+# **Datasets and methodology**
+
+Satellite data have been widely used for the change detection purpose of various parameters of glaciers and glacial lakes (Paul and Andreassen 2009). Traditional field-based surveys, which are highly recommended for measuring and monitoring changes in glaciers and glacial lakes, are often limited by the ruggedness and harshness of Himalayan terrain (Pratap et al. 2016). To overcome this problem multi-temporal satellite data offers a wide range of opportunities to monitor such glaciers and glacial lakes. In this study, we used satellite data of four decades to analyze the spatio-temporal dynamism of the selected lake. Landsat imagery with its free availability and fine spatial resolution of 30 m hosted by the Global Land Survey (GLS) data system of the United States Geological Survey (USGS) was downloaded for the years, namely 1990 (4–5 Thematic Mapper (TM)), 2000 (4–5 Thematic Mapper (TM)), 2010 (Enhanced Thematic Mapper (ETM)), and 2018 (Operational Land Imager (OLI)). Change detection of the glacial lake on the multi-temporal Landsat images is depicted in Fig. 2. The other data set which is required to
+
+carry out GLOF modeling is digital elevation model (DEM) (Table 1).
+
+In this study, we used Satellite Radar Topography Mission (SRTM) DEM with a spatial resolution of 30 m downloaded from the web portal www.earthexplorer.usgs.gov. The DEM was used to extract the catchment/watershed of the glacial lake and its feeding glacier. It was also used to generate a cross-section at 1 km intervals as well as elevation data. The cross-sections were created downstream from the lake to the final selected location. The cross-sections from
+
+Table 1 Satellite data used in the study
+
+| Date of acquisition | Sensor        | Spatial resolution (m) |
+|---------------------|---------------|------------------------|
+| 3 December 1990     | Landsat 4-5TM | 30                     |
+| 19 December 2000    | Landsat 4-5TM | 30                     |
+| 22 August 2010      | Landsat 4-5TM | 30                     |
+| 6 October 2018      | Landsat 8OLI  | 30                     |
+| 9 October 2018      | Google earth  | 1.5–6                  |
+| 12 November 2014    | SRTM-DEM      | 30                     |
+
+Image /page/3/Figure/8 description: A figure displaying four satellite images in a 2x2 grid, illustrating the growth of Satluj lake over time. Each panel is labeled with a letter and includes a title, a north arrow, coordinate grids, and a scale bar.
+
+- Panel a, titled "Satluj lake - 1990", shows a small, light-blue lake outlined in green.
+- Panel b, titled "Satluj lake - 2000", shows the lake has grown in size and is a darker blue, outlined in magenta.
+- Panel c, titled "Satluj lake - 2010", depicts a further enlarged, dark blue lake outlined in red.
+- Panel d, titled "Satluj lake - 2018", shows the lake at its largest extent in the series, appearing dark teal and outlined in yellow.
+
+All images show the surrounding terrain in reddish-brown hues. The longitude ranges from approximately 78°9'45"E to 78°10'15"E, and the latitude ranges from 31°39'30"N to 31°39'45"N. A scale bar in each image indicates distances up to 0.3 km.
+
+Fig. 2 Time series Landsat images showing growth of glacial lake area from 1990 to 2018, **a** lake area 1990 **b** lake area 2000, **c** lake area 2010, and **d** lake area 2018
+
+Image /page/3/Picture/10 description: The Springer logo, featuring a black and white icon of a knight chess piece to the left of the word 'Springer' in a black serif font, all on a white background.
+
+the DEM were generated because the field-measured crosssections were not available for this area due to the rugged terrain and harsh weather conditions. Besides that, highresolution Google Earth imagery was used for the crosschecking and validation of the lake extents. Chow, 1959, has suggested a Manning coefficient range between 0.03 to 0.07 for the hilly terrain with a steep slope and no vegetation, gravel, cobbles, boulders, and bushes on the banks. Considering this, we have used a Manning coefficient range of 0.04 to 0.06 as an input parameter for the GLOF routing of the selected lake of Satluj basin in HEC-RAS software. The overall methodology is given in Fig. 3.
+
+## **GLOF susceptibility and AHP**
+
+In order to assess the GLOF susceptibility, different parameters have been employed in the past studies pertaining to various parts of the Himalayan region (Che et al. 2014; Jain et al. 2012; Aggarwal et al. 2014; Kougkoulos et al. 2018;
+
+Allen et al. 2016). However, as of now, there is no particular set of parameters that can be used to analyze the GLOF susceptibility. In this study, we have used 5 parameters or potential factors that govern the mechanism of potential GLOF to assess the GLOF susceptibility of the selected lake. The parameters were identified and compiled from the literature survey which include lake area, lake expansion, distance between lake and feeding glacier, type of dam, and potential of ice or rock avalanche (Khadka et al. 2018; Gurung et al. 2021). These parameters have been calculated using the satellite imagery (Landsat TM/ETM+/OLI) and digital elevation model. The parameters and their source are given in Table 2.
+
+After the identification and calculation of parameters weights were allocated to each factor using analytical hierarchy process (AHP). AHP model propounded by Satty, 1980, has been widely used for the hazard and risk assessment such as land degradation, landslides, snow avalanche, debris flow (Himan et al. 2014), and GLOF hazard assessment (Khadka
+
+Image /page/4/Figure/7 description: A methodology flowchart illustrating the steps for a GLOF (Glacial Lake Outburst Flood) analysis using the HEC-RAS model. The process begins with two primary inputs in yellow boxes: 'SRTM 30m DEM' and 'Landsat data series 1990-2018'. The SRTM data is used to 'Delineate river reach' and then 'Creation of geometry database along the river in Hec-GeoRAS'. This involves creating 'Bank lines', 'Flow path centerlines of river', and 'cross-section generation across the river', which leads to 'Topology creation' and then 'Export RAS data into HECRAS'. This data is then adjusted for 'cross section', 'bank station', and 'Manning coefficient'. The Landsat data is used for 'Mapping of glacial lake', which informs both the geometry creation and the 'Lake area and volume estimation from Huggle's equation'. This leads to generating a 'Lake breach hydrograph' using Froehlich's equation. The core of the model is the '1 D hydrodynamic modeling' which uses 'Boundary condition' inputs of 'Breach Hydrograph' and 'Friction slope'. The adjusted RAS data and the modeling results feed into 'Unsteady flow estimation'. This is followed by 'Flood routing of breach hydrograph along the river channel'. The final outputs of the analysis are 'GLOF inundation mapping' and 'Downstream impact assessment'.
+
+Fig. 3 Methodology flow chart depicting the steps followed in HEC-RAS model for the GLOF analysis
+
+Image /page/4/Picture/9 description: The image displays the logo for the publisher Springer. On the left is a black outline of a chess knight piece, specifically the horse's head, facing left, with two horizontal lines beneath it. To the right of the icon, the word "Springer" is written in a black serif font. The background is white.
+
+Table 2 Index values assigned to factors/parameters
+
+| Factors (f) | Factor/parameter                           | Class                                           | Rank                  | Index<br>value (Ci) | Rank | Selected lake | Factor<br>weight<br>(Wi) | Data source                            | Reference                                                                  |
+|-------------|--------------------------------------------|-------------------------------------------------|-----------------------|---------------------|------|---------------|--------------------------|----------------------------------------|----------------------------------------------------------------------------|
+| f1          | Lake area (km2)                            | > 0.5<br>0.5–0.1<br>< 0.1                       | High<br>Medium<br>Low | 1<br>0.50<br>0.25   | 5    | 0.262         | 0.04                     | Landsat data<br>(2020)                 | Campbell (2005);<br>Aggarwal<br>et al. (2014);<br>Washakh et al.<br>(2019) |
+| f2          | Lake expansion (%)                         | < 50%<br>50–20%<br>< 20%                        | High<br>Medium<br>Low | 1<br>0.50<br>0.25   | 3    | 56.42         | 0.16                     | Multi-temporal<br>satellite data       | Bolch et al.<br>(2012); Che<br>et al. (2014);<br>Jain et al.<br>(2012)     |
+| f3          | Distance<br>between lake<br>and glacier    | At snout<br>> 500<br>< 500                      | High<br>Medium<br>Low | 1<br>0.50<br>0.25   | 4    | At the snout  | 0.09                     | Landsat and<br>Google Earth<br>imagery | Che et al. (2014);<br>ICIMOD<br>(2011)                                     |
+| f4          | Type of dam                                | Ice dam<br>Moraine dam<br>Other types of<br>dam | High<br>Medium<br>Low | 1<br>0.50<br>0.25   | 2    | Moraine dam   | 0.24                     | Google Earth<br>imagery and<br>DEM     | Huggel et al.<br>(2002)                                                    |
+| f5          | Potential of<br>ice or rock ava-<br>lanche | Susceptible<br>Not susceptible                  | High<br>Low           | 1<br>0.25           | 1    | High          | 0.22                     | Google Earth<br>imagery                | Allen et al.<br>(2016); Islam<br>and Patel<br>(2020)                       |
+
+et al. 2018; Gurung et al. 2021). Each parameter/factor were ranked into three classes, viz., high, medium, and low with a corresponding index value of 1, 0.25, and 0.5 (Prakash and Nagrajan 2017; Gurung et al. 2021). The final weights for each of the parameters/factors were determined by multiplying the factor weight (Wi) by the class index values (Ci) based on the measurement of lake factors. At the last stage, final weights of each factor were added to assess the GLOF susceptibility of the lake (Eq. (1)).
+
+GLOF Susceptibility = 
+$$\sum_{i=0}^{n} (Ci) * (Wi)$$
+ (1)
+
+The GLOF susceptibility index value ranges from 0.25 to 1. Using an equal interval class system, the values are divided into five categories: very low (0.25-0.40), low (0.40-0.55), medium (0.55-0.70), high (0.70-0.85), and very high (>8.5).
+
+## **Hydrodynamic modeling using HEC-RAS**
+
+Modeling of GLOF hazard and risk has been carried out in several locations throughout the world, using a wide range of approaches and methodologies. In this study, we have used HEC-RAS hydrodynamic model to perform GLOF modeling and to evaluate impacts of flood caused due to the potential GLOF in the downstream region of the selected lake located in the Satluj basin. HEC-RAS model has been widely used for this purpose across the world (Dortch et al. 2011; Klimeš et al. 2014b; Watson et al. 2015; Hussain
+
+Image /page/5/Picture/9 description: A close-up of a paragraph of text from a document, printed in a black serif font on a white background. The text reads: et al. 2020; Kougkoulos et al. 2018; Wang et al. 2018). It is an integrated system of software that is designed to use in a multi-tasking environment (Hussain et al. 2020) and is most significantly used for glacial hazard studies (Klimeš et al. 2016). It is a user-friendly reliable model that has the dynamic capability of performing complex flow simulations in design, management, and operation of river systems. The model is based on one-dimensional St. Venant equations that are used to simulate flood scenarios generated by a glacial lake outburst. In hydraulic model, two key input data sets (geometry data and flow data) are required to compute Unsteady Flow Analysis (UFA). Cross-section data derived from the field survey is recommended, but due to the rugged terrain and harsh weather conditions, it is not easy to carry out the field survey.
+
+Geometry data (river cross-sections, center lines, bank lines, flow lines, and elevation data) was derived from the satellite imagery in a plugin of HEC-Geo-RAS in ArcGIS environment. Subsequently, the geometric data was imported into HEC-RAS software for the unsteady flow simulation. The flow data requires upstream and downstream boundary conditions. The dam breach outflow hydrograph is commonly used for upstream and downstream boundary conditions as normal depth, which is calculated using the channel bed slope. The present study considered a glacial lake as a dam failure structure with a specific crest level and crest length. Various parameters such as simulation time series, elevation above the invert level and its breach dimensions were specified to the corresponding lake. An
+
+Image /page/5/Picture/11 description: The logo for the publisher Springer, presented in black on a white background. On the left is an icon of a chess knight's head facing left, positioned above two short, parallel horizontal lines. To the right of the icon, the word "Springer" is written in a serif font.
+
+inline structure that characterizes the lateral-frontal moraine is entered at the lower elevation of the lake, and the failure catastrophe is modeled and evaluated by breaching the inline structure before performing a dam break analysis. For an upper boundary condition, glacial lake acts as a storage area by its elevation storage relationship. The total volume and maximum depth acquired from the empirical equations and DEM were used to generate the elevation-storage volume. DEM was also used to get the Triangular Irregular Network (TIN) terrain to set up the 1-D (hydrodynamic) modeling and subsequently minimum elevation of the lake was derived from the TIN network. The storage capacity at the minimum elevation of the lake (last elected location) is zero, whereas it was obtained for the maximum elevation (topmost elevation of lake). On the basis of these two parameters, elevation storage capacity for the lake was generated. Such parameters are required as an input to the empirical equations that estimate the potential peak discharge and the failure time of a GLOF event (moraine dam breach event). For a dam break analysis, the Froehlich model uses fewer input parameters to estimate peak breach outflow hydrograph. At the final stage, unsteady flow analysis was computed to derive important results like maximum flood depth, maximum flood velocity, peak flood hydrographs, and water surface elevation at different crosssections in the downstream region.
+
+## Moraine dam breach modeling
+
+Glacial lake area, maximum depth, and volume of water in the lake are the important parameters required to perform GLOF modeling. As discussed above, field-based observation was not possible for the current studied lake due to the rugged terrain and harsh weather conditions; thus, we relied on the area-based empirical relationship to obtain the depth and volume of the glacial lake. There are several empirical formulas available in the scholarly literature but the empirical relation developed by Huggel is most extensively employed in past GLOF hazard studies, where bathymetric surveying data are not available. However, this equation was developed for the Swiss Alps region, but due to the unavailability of bathymetry data, we used this equation to derive the volume of the glacial lake. The equation given by Huggel et al. (2002) is given as:
+
+$$V = 0.104A^{1.42} \tag{2}$$
+
+V (m<sup>3</sup>) and A (m<sup>2</sup>) are the volume and area of the lake, respectively. By employing Eq. (2), we get a total volume of  $5749 \times 10^3$  m<sup>3</sup>.
+
+Important parameters such as failure time of dam, extent of overtopping prior to the failure, shape, size, and breach formation time should be considered for the evaluation of the dam failure, because the dam-break flood estimation is dependent upon these parameters. Factors such as initial and final breach width and level, breach shape, breach development formation time, and reservoir level at the time of initial breach are all important breach features that need to be included into existing dam break models (Jain et al. 2012).
+
+We have estimated these essential breach characteristics (breach width and breach formation time) that are required as inputs to carry out dam break analysis have been approximated using Froehlich's empirical relations:
+
+$$B_{w} = 0.1803K_{o}(V_{w})^{0.32}(h_{b})^{0.19}$$
+(3)
+
+$$T_{f} = 0.00254 (V_{w})^{0.53} (h_{b})^{0.9}$$
+(4)
+
+In the above equations,  $B_w$  and  $V_w$  are the breach width (m) and volume of lake (m³), whereas  $h_b$  and  $T_f$  are the breach height (m) and breach formation time (hr). Equation 1 mentioned in the above section was used to compute the volume of the lake.  $B_w$  is the function of  $V_w$  and  $h_b$ . In the present study, a total of nine dam breach scenarios based on varied breach width and breach formation time are designed for a mechanism of a glacial lake outburst flood. The failure mechanism as an overtopping is considered in breach scenarios to evaluate a potential worst-case GLOF event of a lake. These scenarios produce peak flood hydrographs at the downstream cross-sections generated at various locations.
+
+# **Results and discussion**
+
+## Change in glacial lake area
+
+Using Landsat imagery in combination with high-resolution google earth imagery, we have identified and mapped the selected glacial lake, its feeding glacier, and other geomorphological features such as moraines, crevasses, snout, supraglacial channel, and streams. The lake extents derived from the multi-temporal satellite images have shown a considerable increase of the lake area from 0.112 km<sup>2</sup> in 1990 to 0.26 km<sup>2</sup> in 2018. Thus, showing an overall increase of 0.12 km<sup>2</sup> (56.4%) over the period of 28 years indicates a rapid expansion at the rate of 0.01 km<sup>2</sup>/year. The snout of the glacier (feeding lake) has shown a retreat of 153 m at the rate of 7.56 m/year from 2000 to 2020. The lake has been identified as potentially dangerous by employing the various parameters derived from the scholarly literature. Previously, it has been identified as potentially dangerous glacial lake by Prakash et al. (2013). The present glacial lake is situated at the retreating glacier snout dammed by loose and unconsolidated material. Furthermore, it has been identified that there are probable rockfall/avalanche zones in the
+
+Image /page/6/Picture/16 description: The logo for Springer, featuring a black and white line drawing of a chess knight piece to the left of the word "Springer" in a black serif font. The background is white.
+
+upstream region of the lake that can be disastrous at some point of time when they hit the lake. The nearest settlement and Kashnag project power house are located just 17 km from the lake outlet. Keeping in view these factors, we have carried out the GLOF modeling of the lake for a distance of 102 km in the downstream region. In this study, we have calculated the breach width and breach formation time by using Eqs. (2) and (3). Breach width is a function of the total volume of water in upstream and the breach height. A breach width of 75 m was obtained by using Eq. (2) to calculate the initial dam-breach hydrograph and subsequently to perform an unsteady flow analysis.
+
+## **GLOF** susceptibility
+
+In the present study, utilizing satellite data such as Landsat (TM, ETM + and OLI), Google Earth imagery, and DEM, the characteristics of the glacial lake and its feeding glacier were investigated in detail. The major conclusions drawn from the GLOF susceptibility index based on the essential parameters mentioned in Table 2 are as follows. The investigation reveals that the lake has an area of 0.26 km<sup>2</sup> which is greater than the threshold area mentioned for the lake to be potentially dangerous (Campbell 2005; Aggarwal et al. 2014; Washakh et al. 2019). The lake area has shown a significant increase in area over a period of 28 years from 1990 to 2018, indicating a growth of 56.42% with a mean increase of 0.003 km<sup>2</sup>, thus depicting the high index value in terms of GLOF susceptibility. Lakes that are in direct contact with the glacier (at the snout) have a high potential to cause a GLOF event due the sudden calving of the retreating snout (Che et al. 2014; ICIMOD 2011). In our case, the selected lake is located at the snout of the glacier with little glacier calving visible on the Google Earth imagery (Fig. 1). The type of dam plays an important role in determining the hazard level of a particular lake. Lake dammed by ice or moraines are considered potentially most dangerous as compared to the landslide or bedrock dammed lakes. The selected lake has been classified as a moraine dammed lake through the visual interpretation technique (Huggel et al. 2002). It is also crucial to note that certain triggering factors, including as heavy rain, landslides, rockfall, mass movements, avalanches, and cloud bursts, may cause these glacial lakes to suddenly burst. From the Google Earth imagery, potential ice and rock avalanche zones are clearly visible. In case if a lake is hit by such triggering factors, it may cause sudden breaching of the moraine dam and the resultant flood may wash away the agricultural land and infrastructure in the downstream region.
+
+It is quite clear from the Table 2 that all of the selected parameters have shown high index value and their corresponding rank value is also high. The lake is area is greater than 0.2 km<sup>2</sup>, with an expansion rate of more than 56% and it
+
+is present at the retreating snout of the glacier. Furthermore, lake is dammed by moraine and susceptible to potential rock and ice avalanche. In terms of susceptibility index, the total value of the lake was calculated as 0.75, which indicates that the lake is highly susceptible to GLOF and therefore needs continuous monitoring and assessment with the help of remote sensing GIS and hydrodynamic modeling.
+
+## Assessment and simulation of GLOF
+
+The GLOF simulation has been carried out for the selected lake dammed by an unconsolidated moraine as visible on the Google Earth imagery. The lake is present at the snout of the glacier which is retreating rapidly from the past few decades and adding more and more influx of the glacier melt into the lake. With the increasing temperatures and resultant glacier melting, more expansion is expected from this lake in the coming years. Considering all these factors, the flood routing from a glacial lake up to a distance of 102 km to the basin outlet has been characterized in the model by adding various river cross-sections derived from SRTM DEM. Each cross-section is identified with a different river station and subsequently given as an input in the HEC-RAS model as depicted in Fig. 4.
+
+The downstream effects of the GLOF caused by the dam failure are modelled using the unsteady flow analysis in HEC-RAS model. The one-dimensional moraine breach simulations were used to evaluate the unsteady flow hydrodynamic models understanding and sensitivity to the input dam-breach parameter (breach width and breach formation time). Therefore, to estimate and evaluate its effect on initial breach hydrograph, possible varied breach width  $(B_w)$  and breach failure time  $(T_f)$  of the lake were considered for different worst-case scenarios, in which breach width has been taken as 40, 60, and 75 m and breach formation time has been taken as 40, 30, and 20 min, respectively (Fig. 5).
+
+Figure 5a to f shows the one-dimensional breach hydrographs scenarios, executed with varied breach width (Bw) and breach formation time (Tf). We have considered three scenarios with varied Bw = 75 m, Bw = 60 m, and Bw = 40 mwith a constant breach formation time (Tf). Similarly, keeping breach width (Bw) constant, model sensitivity to breach formation time was also analyzed by considering three scenarios with varied Tf (20, 30, and 40 min). Based on the various breach parameters, overtopping failure mechanism have been used to analyze the various scenarios. The results derived from these scenarios are given in Table 3. It is inferred from this study that, among the nine different breach scenarios, GLOF breach hydrograph for 75 m breach width (Bw) and 30 min breach formation time (Tf), which produced peak discharge of 4507 m<sup>3</sup>/s, has been considered an input for upstream boundary condition and generates flood routing at different sites along the river flow channel,
+
+Image /page/7/Picture/11 description: The logo for the publisher Springer, featuring a black outline of a knight chess piece to the left of the word "Springer" in a black serif font, all on a white background.
+
+**Fig. 4** Cross-section network generated along the river channel from the lake outlet
+
+Image /page/8/Figure/3 description: A map showing a river system with its source at a glacial lake. The map is gridded with longitude and latitude lines, ranging from 77°45'0"E to 78°15'0"E and 31°25'0"N to 31°40'30"N. The river flows from the 'Glacial lake' in the northeast to an 'Outlet point (Rajpur town)' in the southwest. The river's path is marked with numerous red cross-sections, which are numbered sequentially from 1000 near the outlet to 104000 near the lake. A legend in the bottom right corner defines the symbols used: a light blue area for the glacial lake, a blue line for the river, a purple line for flowpaths, a red line for a cross section, and a green dot for the outlet point. A scale bar at the bottom indicates distances in kilometers, up to 24 km. A north arrow is present in the top right corner.
+
+which gives different peak flood hydrograph and other significant flood wave parameters like water depth, flow velocity, flood peak arrival time, and the corresponding water surface elevations in the downstream area and so this was used for detailed studies on downstream impact.
+
+The assessment of the different magnitude of flood peak hydrographs at just downstream of the lake and various locations across the river is shown in Fig. 6. In this study, attenuated flood peak hydrographs have been evaluated at five different locations at a distance of 29.3 km, 39.1 km, 44.4 km, 47.7 km, and 102 km (Rajpur town) downstream from the lake. The estimated flood peak 4507 m<sup>3</sup>/s at GLOF site and the same get mitigated to 464 m<sup>3</sup>/s at the downstream outlet site, respectively. The flood peak arrival time from lake to the outlet at a distance of 102 km is about 6 h and 5 min. This reflects that the downstream river reach is relatively wide with a gentle slope compared to that narrow valley of upstream glacial area; as a result, flood water spreads across a flood plain resulting in gradually decreasing of peak discharge as it proceeds downstream from the lake as shown in Fig. 6 and Table 4. In addition to this, flood routing along the river reach allowed calculation of other flood hydraulic parameters such as flow depth, flow velocity, and flood peak arrival time which are given in Table 4 for each of the affected locations.
+
+The resultant flood at the nearest site (bridge) located 29.3 km from the lake site will arrive at 01:40 with a maximum discharge of 3229  $\rm m^3~s^{-1}$  and maximum flood depth and velocity of 8.13 m and 8.3 m/s, respectively. At Karcham barrage located 39 km from the lake outlet, a peak flood with
+
+2692 m³ s<sup>-1</sup> will arrive at 2:20 with a maximum depth of 3.8 m and a maximum flood velocity of 4.4 m/s. At the settlement 1 situated at a distance of 44 km from the lake site, the maximum discharge was estimated as 2672.12 m³ s<sup>-1</sup> with maximum flood depth and velocity of 4.1 m and 4.4, respectively. Similarly, at settlement 2 located at a distance of 48 km, the maximum peak discharge was estimated as 2510 m³ s<sup>-1</sup>. The maximum flood depth and velocity for the same was estimated by the model as 4.4 m and 5.4 m/s, respectively. The peak flood hydrograph and other flood wave parameters were also calculated for the last selected site, namely, Rajpur town located at a distance of 102 km from the lake site. The results show a maximum peak discharge of 481 m³ s<sup>-1</sup> with a maximum flood depth and velocity of 3.3 m and 1.5 m/s (Fig. 6).
+
+After the analysis of Table 4 and Fig. 6, it was observed that the intensity of peak flood hydrographs and other significant flood wave parameters like water depth, flow velocity, and water surface elevation are decreasing from the lake site to the last selected location, i.e., Rajpur town. This is because of the elevation differences that decrease from the lake site to the downstream areas. The water spread area increase with the decrease in elevation and thereby decreasing the peak discharge and other associated parameters such as maximum depth, velocity, and water surface elevation.
+
+### Flood inundation and simulation
+
+In the present study, unsteady flow routing of simulated GLOF hydrograph was computed from the selected lake up
+
+Image /page/8/Picture/11 description: The image displays the logo for Springer, a publishing company. The logo consists of a black-and-white graphic and text on a white background. On the left, there is a stylized icon of a chess knight piece, which is a horse's head facing left, positioned above two short, parallel horizontal lines. To the right of the icon, the word "Springer" is written in a black serif font.
+
+Image /page/9/Figure/2 description: The image displays six line graphs, labeled a through f, illustrating GLOF (Glacial Lake Outburst Flood) hydrographs under different scenarios. All graphs plot discharge, Q (m³/s), on the y-axis against Time (HH:MM) on the x-axis, which ranges from 00:00 to 14:24. Each graph contains three curves representing different conditions, with their peak discharge values labeled.
+
+Graphs a, b, and c show hydrographs for breach formation time scenarios:
+- Graph a (constant breach width = 75m): Shows peak discharges of 5174.62, 4507.01, and 4068.17 m³/s for breach formation times of 20, 30, and 40 minutes, respectively.
+- Graph b (constant breach width = 60m): Shows peak discharges of 4817.09, 4151.72, and 3761.88 m³/s for breach formation times of 20, 30, and 40 minutes, respectively.
+- Graph c (constant breach width = 40m): Shows peak discharges of 4159.71, 3609.08, and 3362.44 m³/s for breach formation times of 20, 30, and 40 minutes, respectively.
+
+Graphs d, e, and f show hydrographs for breach width scenarios:
+- Graph d (Constant Tf = 30 min): Shows peak discharges of 4507.01, 4151.73, and 3609.07 m³/s for breach widths of 75m, 60m, and 40m, respectively.
+- Graph e (Constant Tf = 20 min): Shows peak discharges of 5174.62, 4817.09, and 4159.71 m³/s for breach widths of 75m, 60m, and 40m, respectively.
+- Graph f (Constant Tf = 40 min): Shows peak discharges of 4068.17, 3761.88, and 3362.44 m³/s for breach widths of 75m, 60m, and 40m, respectively.
+
+Fig. 5 Lake breach outflow hydrographs showing peak discharge obtained for varied (a-c) breach formation time (Tf) keeping breach width constant, similarly (d-e) varied breach width (Bw) keeping the breach formation time constant
+
+Table 3 Breach scenarios with varied breach width and breach formation time
+
+| Failure mode- overtopping   |            |            |            |            |            |            |            |            |            |
+|-----------------------------|------------|------------|------------|------------|------------|------------|------------|------------|------------|
+|                             | Scenario 1 | Scenario 2 | Scenario 3 | Scenario 4 | Scenario 5 | Scenario 6 | Scenario 7 | Scenario 8 | Scenario 9 |
+| Breach width (m)            | 75         | 60         | 40         | 75         | 60         | 40         | 75         | 60         | 40         |
+| Breach formation time (min) | 20         | 20         | 20         | 30         | 30         | 30         | 40         | 40         | 40         |
+| Peak discharge (m3/s)       | 5174.62    | 4817.09    | 4159.71    | 4507.01    | 4151.72    | 3609.08    | 4068.17    | 3761.88    | 3362.44    |
+
+Image /page/9/Picture/6 description: The Springer logo, featuring a black and white line drawing of a chess knight piece to the left of the word 'Springer' in a black serif font, all on a white background.
+
+**Fig. 6** Routed flood hydrographs at selected downstream locations along the flow path
+
+Image /page/10/Figure/3 description: A line graph plots Discharge in cubic meters per second (m³/s) on the y-axis versus Time in hours and minutes (HH:MM) on the x-axis. The y-axis ranges from 0 to 5000, and the x-axis ranges from 0:00 to 14:24. The graph shows several colored lines representing discharge at different locations and distances. A legend in the top right corner identifies five data series: GLOF at 0 Km (dark blue), Bridge at 29.3 Km (red), Karcham barrage at 39.06 Km (light blue), Settlement 1 at 44.39 Km (purple), and Settlement 2 at 47.69 Km (red). The GLOF line shows a sharp peak at 4500 m³/s shortly after 0:00. The subsequent lines show peaks that are progressively later in time and lower in magnitude: the Bridge line peaks at about 3250 m³/s around 2:00; the Karcham barrage and Settlement 1 lines peak around 2700 m³/s between 2:24 and 2:45; and the Settlement 2 line peaks at 2500 m³/s around 3:00. There is also a green line on the graph, not listed in the legend, which shows a later, smaller peak of about 1800 m³/s around 6:00.
+
+**Table 4** Peak flood and time of peak at different sites along the flow routing
+
+| Sites            | Distance from lake (km) | Peak flood (m3/s) | Maximum depth (m) | Maximum velocity (m/s) | Flood peak<br>arrival time<br>(HH:MM) |
+|------------------|-------------------------|-------------------|-------------------|------------------------|---------------------------------------|
+| Just d/s of lake | 0                       | 4507              | 8.1               | 8.3                    | 00:20                                 |
+| Bridge           | 29.3                    | 3229              | 9.7               | 4.3                    | 01:40                                 |
+| Karcham barrage  | 39.1                    | 2692              | 3.8               | 4.4                    | 02:20                                 |
+| Settlement 1     | 44.3                    | 2672              | 4.2               | 4.4                    | 02:30                                 |
+| Settlement 2     | 47.6                    | 2510              | 3.9               | 5.5                    | 02:40                                 |
+| Rajpur town      | 102                     | 481               | 3.3               | 1.5                    | 06:05                                 |
+
+to the Rajpur town, which is located at a distance of 102 km downstream from lake. The flood wave routed the Rajpur town at 6 h and 5 min after the beginning of the GLOF event. Maximum water depth and maximum flow velocity along the flow channel were estimated as 27.4 m and 13.3 m/s, respectively. Figures 7 and 10 depict the spatial plots of inundation depth and flow velocity along the main flow channel from the Satluj lake to the Rajpur town and also provide a visualization of the three main inundated areas downstream of the lake.
+
+The resultant flooding may cause damage to the infrastructure such as roads and bridges along the river's flow path downstream of the GLOF lake. The flood wave inundates a bridge located at 30 km downstream of the lake shown in Fig. 8a. Within 8 min of the flood wave arriving, flow velocities reach a high of 13.6 m/s (Fig. 9) and a water depth of 10.2 m (Fig. 7) with an arrival timing of 8 min is recorded, potentially causing significant inundation of the bridge and roads. GLOF wave reaches the first bridge at location 27°39′22.79″ N, and 88°36′16.26″ E located at 29.3 km downstream of lake shown in Fig. 8a, where flow depth is estimated up to 9 m and flow velocity of 4.28 m/s.
+
+The GLOF potentially inundates three settlements, one barrage along the river flow path. The flood wave encounters the first settlement at locations 27°36′22.58″ N, and 88°38′31.76″ E along the bank of the channel at 44.4 km downstream, with a maximum flow velocity of 4.5 m/s and maximum flow depth of 4.2 m (Figs. 7 and 10). Figure 8d shows that the maximum flow depth and velocity recorded are 3.932 m and 5.498 m/s in the second settlement.
+
+The analysis shows that the flood routing varies and gradually decreases along with the downstream of lakes. In this flood routing, peak flood decreases gradually from the lake up to 5 km then sharply declined as it proceeds downstream as shown in Fig. 10. We have observed that flood height have shown rise and fall continuously along the downstream of the river reach covering all the various sites located at few kilometers distance. The maximum flood height would be around 16 m at 20 km downstream of the lake and the peak flood at this location was found to be 3126 m³/s. The food peak reached Karcham barrage at 39 km downstream would be around 2692 m³/s.
+
+The estimated flood depth in the downstream area varies depending upon the morphology of the river channel. The
+
+Image /page/10/Picture/11 description: The image displays the logo for Springer, a publishing company. The logo consists of a black-and-white graphic and text on a white background. On the left, there is a stylized icon of a chess knight piece, which is a horse's head facing left, positioned above two short, parallel horizontal lines. To the right of the icon, the word "Springer" is written in a black serif font.
+
+Image /page/11/Figure/2 description: A figure from a scientific paper showing a map of a mountainous region with a river valley, illustrating maximum flood depth. The map is bounded by longitudes 77°40'0"E to 78°10'0"E and latitudes 31°27'0"N to 31°37'30"N. A main legend in the bottom right indicates that the maximum flood depth ranges from a low of 0.001 m (blue) to a high of 27.444 m (magenta). There are three inset maps that provide detailed views of specific locations along the river. Each inset includes a line graph plotting flood depth in meters against flood wave arrival time in hours. The top-left inset shows a graph with a peak flood depth of about 8 m. The top-middle inset shows a peak flood depth of about 10 m. The bottom-middle inset shows a peak flood depth of about 4.5 m. Each inset has its own legend for flood depth, showing different ranges. A scale bar at the bottom left of the main map indicates distances from 0 to 32 km.
+
+Fig. 7 Spatial distribution of Maximum Flood inundation
+
+Image /page/11/Figure/4 description: A composite image of four maps, labeled a, b, c, and d, each showing a satellite view of a river with a color overlay indicating the maximum flood depth. Each map includes a legend, a scale bar in kilometers, and geographic coordinates. Map (a) highlights a bridge crossing the river, with a legend for Max Flood Depth (m) showing ranges: 0.02-2.26, 2.26-4.55, 4.55-6.95, 6.95-9.3, and 9.30-13.61. Map (b) shows an area of settlements along a river bend, with a legend showing flood depth ranges: 0.02-2.48, 2.48-4.47, 4.47-6.23, 6.23-8.36, and 8.36-11.64. Map (c) shows a settlement along a wider section of the river, with a legend showing flood depth ranges: 0.04-2.16, 2.16-3.93, 3.93-5.46, 5.46-7.19, and 7.19-10.83. Map (d) shows a dense settlement on a sharp river bend, with a legend showing flood depth ranges: 0.12-2.41, 2.41-4.95, 4.95-7.81, 7.81-11.31, and 11.31-16.32. In all maps, the colors range from light blue for shallower depths to dark purple for the deepest floodwaters.
+
+Fig. 8 Flood inundation maps: a focused view at inundate bridge, b zoomed view of inundated settlement 1, c zoomed view at inundate settlement 1, and d zoomed view of Rajpur town
+
+Image /page/11/Picture/6 description: The logo for Springer, featuring a black and white line drawing of a chess knight piece on the left, followed by the word "Springer" in a black serif font. The knight is facing left and is positioned above two horizontal lines. The background is white.
+
+Image /page/12/Figure/2 description: A figure from 'Environmental Science and Pollution Research' showing a map of a mountainous region, illustrating flood data. The main map, with coordinates from 77°40'0"E to 78°10'0"E and 31°27'0"N to 31°37'30"N, displays a yellow outline of a flood path. A legend in the bottom right indicates the 'Max Flow Velocity (m/s)' with a color scale from green (Low: 0.480) to red (High: 13.280). There are three inset maps providing detailed views of specific locations. Each inset includes a color-coded representation of flow velocity and a line graph plotting 'Flow velocity (m/s)' against 'Flood wave arrival time, hr' from 00:00 to 14:24. The top-left inset shows a settlement with flow velocities up to 8.66 m/s and a graph peaking around 1.2 m/s. The top-middle inset shows a bridge with flow velocities up to 9.08 m/s and a graph peaking just below 4 m/s. The bottom-middle inset shows settlements with flow velocities up to 13.28 m/s and a graph also peaking just below 4 m/s. A scale bar at the bottom left of the main map indicates a distance of 32 km.
+
+Fig. 9 Spatial distribution of maximum velocity inundation
+
+**Fig. 10** Peak flood and flood height impact on the downstream of glacial lake
+
+Image /page/12/Figure/5 description: A line graph plots two variables, Peak Flood and Flood Height, against the Distance from Lake outlet. The x-axis represents the 'Distance from Lake outlet (km)' and ranges from 0 to 95. There are two y-axes. The left y-axis represents 'Peak Flood (m³/s)' and ranges from 0 to 5000. The right y-axis represents 'Flood Height (m)' and ranges from 0 to 18. The legend indicates that the blue line with circular markers represents 'Peak flood', and the red line with triangular markers represents 'Flood Height'. The 'Peak flood' line shows a general downward trend, starting at approximately 4500 m³/s near the outlet and decreasing to about 1750 m³/s at 95 km. The 'Flood Height' line fluctuates significantly. It starts around 8 m, peaks at nearly 12 m, drops, then spikes to a maximum of about 14.5 m at a distance of 20 km. It then drops to its lowest point of about 5 m at 35 km, followed by several other peaks and troughs, including a double peak of about 11 m between 75 and 80 km, and ends with a sharp rise to about 12.5 m at 95 km.
+
+flood heights and peak flood discharges were estimated using a one-dimensional hydrodynamic model and to analyze the potential impact of GLOFs on downstream areas. The obtained results of flood simulation were used to generate a flood inundation map. Regular monitoring is required to analyze the behavior of the glacial lake and to prevent and monitor the GLOF hazard and assess the damage that may occur in the future. Over the past few decades, GLOFs are
+
+Image /page/12/Picture/8 description: The logo for the publisher Springer, shown in black on a white background. On the left is an icon of a knight chess piece facing left, and to the right is the word "Springer" in a serif font.
+
+becoming common in the glacierized basins of the world, resulting in thousands of deaths and severe damage to the property in the downstream regions. Given the frequency of earthquakes and ongoing warming, the likelihood of GLOFs and related disasters is significant. It is not only the likelihood of a glacial lake outburst that comes under the category of risk but vulnerability and adaptability of various elements such as population, roads, and infrastructure are also part of it.
+
+# **Discussions**
+
+In the current study, we investigated the dynamics of the highly susceptible glacial lake (dammed by moraine) in the Satluj basin of Western Himalayas for a period of 28 years from 1980 to 2018. The results show a clear growth in the glacial lake area from 0.112 km<sup>2</sup> in 1990 to 0.262 km<sup>2</sup> in 2018 derived from the multi-temporal satellite (Landsat TM, ETM and OLI) with a spatial resolution of 30 m. The glacial lake studies using multi-temporal satellite data, especially Landsat series, have been successfully employed by various researchers in different parts of the Himalayan region. The investigations reveal that glacial lakes are increasing in both number and area; for example, in Jhelum basin of Kashmir Himalayas, the number has increased from 253 in 1990 to 324 in 2018, with a growth rate of 21.4% and the area on the other hand has increased from  $18.84 \pm 0.1 \text{ km}^2$  in 1990 to  $22.13 \pm 0.12 \text{ km}^2$  in 2018 with a growth rate of 14.7% (Ahmed et al. 2021a, 2021b). Similarly in Koshi basin, the glacial lakes have shown an increase of 31.5% between 1975 and 2010 (Shresth et al. 2017), Central Himalayas by 17 and 24% (Nie et al. 2013; Shukla et al. 2018), Nepal Himalayas by 25% during 1977 to 2015 (Khadka et al. 2018), Nepal and Bhutan Himalayas by 20–60% during 1990–2009 (Gardelle et al. 2011), Sikkim Himalayas by 8.3% during 1988–2014 (Debnath et al. 2018), Entire third pole by 23% during 1990-2010 (Zhang et al. 2019), and Eastern Himalayas by (Begam and Sen 2019; Mal et al. 2021). Rapid glacier recession and glacial lake expansion in the high terrains of the Himalayan region are a clear indicator of the changing climate and global warming. Temperature plays a key and dominant role in influencing the sensitivity of the glaciers. Glaciers in the Himalayan region are under the influence of the western disturbances, and Indian and Asian monsoons. Over the last few decades, glaciers in the Himalayan region are retreating and losing mass at a faster rate under the influence of the climate change (increasing temperatures and decreasing precipitation) resulting in the formation and development of various types of glacial lakes especially supraglacial lakes, moraine-dammed lakes, icedammed lakes, etc. Some of these lakes are potentially dangerous, which can burst and may cause huge damage to the life and property in the downstream valleys as in case of the Chorabari lake outburst (Allen et al. 2016) which results in the number of deaths and destruction to the property in the downstream regions. Therefore, monitoring and assessment of the potentially dangerous glacial lakes and evaluating their probable GLOF impacts are a key step to reduce the damage and loss associated with such events.
+
+Apart from the glacial lake dynamics, we have also evaluated the probable GLOF impacts from the high susceptible glacial lake on various important locations in the downstream region using combined approaches of remote sensing GIS and one-dimensional hydrodynamic model (HEC-RAS). GLOF event with a breach width of 75 m was revealed during the overtopping failure of the moraine dam, which resulted a peak discharge of 4063 m<sup>3</sup>/s and releasing a total volume of  $57 \times 10^5$  m<sup>3</sup> water. The GLOF hydrograph has been routed to calculate the spatial and temporal distribution of peak flood, inundation depth, velocity, water surface elevation, and flood peak arrival time along the river. The findings of our modeling in Satluj basin revealed that channel roughness is not necessary for the extent of the inundation, but it is critical for the flood's velocity and the flood arrival time. Our results are in line with the findings from other studies pertaining to the HEC-RAS modeling (Sattar et al. 2020; Hagg et al. 2021).
+
+HEC-RAS has been widely used in the past by various researchers for the GLOF hazard assessment and it is considered an important tool to analyze the likelihood of GLOFs resulted due to the breaching of moraine dams. The model output can be used for better planning and preventive measures. In the Himalayan region, a number of hydropower projects are situated on the various rivers and tributaries and a large number are being planned to be constructed on these rivers. There is a considerable potential that these dams will be breached, putting the hydropower plants and projects safety in danger (AHEC 2011). As a result, taking into account GLOF scenarios and estimating hydrologic parameters are now required for effective planning and design of hydroelectric power projects viz a viz their safety. Since field survey was not possible for the current studied lake due to the rough terrain, harsh weather conditions, and lack of instrumentations and funding, therefore, we relied on the satellite data and hydrodynamic modeling (HEC-RAS). But we suggest an in-depth analysis of the selected susceptible glacial lake through field-based investigations.
+
+HEC-RAS has been effectively applied to model GLOF simulations in different parts of the Himalayan region. For example, Sattar et al. (2020) have used the HEC-RAS model for the hazard assessment of the potentially dangerous glacial lake (Safed Lake) in central Himalayas. Thakur et al. (2016) have used the one-dimensional hydrodynamic model (HEC-RS) to evaluate the impacts of GLOFs on the hydropower projects in Dhauliganga River basin. Similarly,
+
+Image /page/13/Picture/9 description: The logo for Springer, shown in black on a white background. On the left is an icon of a knight chess piece, and to the right is the word "Springer" in a serif font.
+
+Khanal et al. (2015) have applied the HEC-RAS model for the GLOF risk assessment in Nepal transboundary region. Maskey et al. (2020) have also used HEC-RAS for the GLOF modeling of Thulagi and Barun glacial lakes in Nepal Himalayas. Besides that, there are various studies that have successfully this model to evaluate the GLOF impacts on the downstream region in the Himalayas (Bajrachraya et al. 2007; Gilany and Iqbal 2020; Zaidi et al. 2013; Hussain et al. 2020; Amin et al. 2020; Majeed et al. 2021; Sattar et al. 2019) and outside the Himalayan region (Klimeš et al. 2014a; Kougkoulos et al. 2018; Anacona et al. 2015; Somos-Valenzuela and McKinney 2011; Klimes et al. 2015; Ahmed et al. 2022a, b).
+
+## Limitations and recommendations
+
+With the ongoing climate change-induced continuous glacier retreat, new glacial lakes have formed and enlarged the existing ones; therefore, the frequency of glacier lake outburst floods (GLOFs) is anticipated to grow. We have investigated the potentially dangerous glacial lake in the Satluj basin by employing the 1-D hydrodynamic modeling. Glacial lake outburst flow modeling is nothing more than an approximation of a physical phenomenon that can be used to study the physical phenomenon and its impacts on water resource structure design and flood management. In GLOF modeling, assumptions are mostly related to breach parameters, particularly breach breadth and depth, which affect flood peak and flood arrival times. Due to the lack of fieldbased bathymetry data and several other related parameters associated with the dam breach as well as routing of the flood hydrograph are the rough estimates/assumptions that cause errors and uncertainties in the results. The use of DEM with a coarse spatial resolution (30 m) also overestimates the inundation and the magnitude of the flood; therefore, we suggest the use of fine or high-resolution terrain data (DEM) to reduce the uncertainties in the results. Generally, in case if the volume of the storage area (lake) is unknown, empirical equations developed in different parts of the world are used to calculate the volume of the lake. But different equations yield different volumes, thus showing a huge variation in the results. Thus, in this backdrop, we suggest that an average of all these empirical equations should be taken into account while computing the volume of the lake also suggested by Emmer et al. (2021) and Ahmed et al. (2021a). In general, the mechanism of glacial lake bursting and the genesis of glacial lake breaches are not well understood. Irrespective of the limitations and assumptions discussed above, hydrodynamic modeling serves a valuable purpose as it provides a reasonable estimate of glacial lake outburst flood and allows for an effective estimation of the design flood.
+
+# **Conclusions**
+
+Glacial lake outburst floods are one of the major concerns in the Himalayan region. With the expansion and formation of new glacial lakes, the risk of GLOF is further increasing in the region. This study used multi-temporal satellite data, SRTM DEM, and HEC-RAS software to analyze the changes in the selected GLOF-susceptible lake and to evaluate the potential impacts of GLOF in the downstream area. Using a multi-criteria decision-based method, the selected lake was found to be highly susceptible to GLOF. The lake area has shown a discernible increase from 0.12 km<sup>2</sup> in 1990 to 0.26 km<sup>2</sup> in 2018 with an overall increase of 0.11 km<sup>2</sup> (56.4%) over the period of 28 years, indicating a rapid expansion at the rate of 0.01 km<sup>2</sup>/year. Several hydrodynamic simulations were performed for different scenario models with varied breach width, breach formation time and breach depth. The worst case GLOF scenario (breach width=75 m) is revealed during an overtopping failure of the moraine dam, which resulted a peak discharge of 4063 m<sup>3</sup>/s and releasing a total water volume of  $5749 \times 10^3$  m<sup>3</sup>. The GLOF hydrograph was routed to calculate the spatial and temporal distribution of peak flood, inundation depth, velocity, water surface elevation, and flood peak arrival time along the river. The findings of this study show that the routed flood waves reach the nearest town, i.e., Rajpur located at a distance of 102 km downstream of a lake at 6 h after the initiation of the breach event, with a peak flood of 1756 m<sup>3</sup>/s and maximum flow velocity of 1.5 m/s. If a GLOF occurs, it can have severe consequences for downstream livelihood and infrastructure. As a result, such lakes must be examined regularly, and mitigating efforts to reduce GLOF vulnerability should be prioritized. With the increasing and apparent GLOF risk across the Himalayan region, the findings of this study highlight the urgent need of forward-looking, collaborative, and long-term initiatives to prevent future impacts of GLOFs as well as to boost the sustainable development in the region.
+
+**Acknowledgements** The authors are thankful to the United States Geological Survey (USGS) for freely providing the satellite data used in this study.
+
+Author contribution Every author has contributed to the successful compilation of this study. MR and RA: conceptualization, methodology, software, writing—original draft, formal analysis. MR, RA, SKJ and AKL: data curation, formal analysis, writing—review and editing. SKJ and AKL: writing—review, editing, supervision. All authors read and approved the final manuscript.
+
+Data availability Not applicable.
+

chatbot/rag/data/('Glacial lake outburst flood risk assessment', '.pdf')_extraction.md

@@ -0,0 +1,368 @@
+
+#### **ARSTRACT**
+
+The ongoing trend of warming climate has made Glacial Lake Outburst Floods (GLOFs) a major cryospheric hazard worldwide, especially in the Himalayas. GLOFs in the Himalayan region are mostly caused by moraine-dammed proglacial lakes and icedammed lakes. These sporadic disasters have resulted in significant loss of life and property. This study offers a comprehensive analysis of the GLOF hazard potential of a potentially dangerous proglacial lake (PDGL) in the Ladakh region. This research explores the GLOF threat from the lake using multi-criteria analysis and advanced 2D hydrodynamic modeling approaches. The mass balance response of the mother glacier, its flow dynamics, and glacier-lake interactions were examined for the past 22 years. The findings show that over this period, the PDGL has had a notable expansion of 78.7%, accompanied by a significant recession of 13.2% in its feeding glacier. The glacier has witnessed an average thickness loss of  $\sim$ 7 m at the rate of 0.32 m  $a^{-1}$  during this period. The average, lowest, and maximum depth of the glacier were found to be 30.95, 14.30, and 50.57 m, respectively and the average velocity of the glacier was estimated as 3.38 m a<sup>-1</sup>. Because of the lake's rapid expansion and steep surrounding slopes, it was classified as a high-hazard lake. The risk to the downstream community was assessed through 2D hydrodynamic modeling using the HEC-RAS tool. The maximum discharge under the worst-case scenario for the piping and overtopping failures was estimated as 3890.99 m<sup>3</sup> s<sup>-1</sup> and 5111.39 m<sup>3</sup> s<sup>-1</sup>, respectively. The area potentially under the threat of inundation was calculated to be 4.74 and 5.38 km<sup>2</sup> for the moderate and worst-case scenarios respectively. The expected maximum flood velocities range from 18.26 to 23.78 meters, respectively for the moderate and worst-case scenarios. At several locations in the downstream area, routed hydrographs representing the GLOF propagation were generated. The findings show that the flood wave in the worst-case scenario
+
+#### **ARTICLE HISTORY**
+
+Received 5 June 2024 Accepted 3 October 2024
+
+#### **KEYWORDS**
+
+GLOF; glacial lake expansion; risk assessment; HECRAS, Ladakh region; Western Himalayas; Panikhar, Kargil would arrive at the first settlement in  $50\,\text{min}$ , with a peak velocity of  $12.36\,\text{m s}^{-1}$ . The potentially inundated area includes critical infrastructure such as bridges, residential houses, and roads. To mitigate the potential risk associated with this lake, a more detailed and on-site study is highly recommended.
+
+# 1. Introduction
+
+In most glaciated parts of the world, the effects of climate change have resulted in ongoing glacier recession and the formation of glacial lakes behind recently exposed unstable moraines (Zemp et al. 2009; Bolch et al. 2012; Nie et al. 2017). Mountain settlements in the Himalayas, where numerous such glacial lakes are created, are seriously at risk from glacial lake outburst floods (Clague and Mathews 1973; Ding and Liu 1992; Ahmed 2024). Glacial lake outburst floods (GLOFs) are catastrophic occurrences that take place when a glacial lake, impounded by a glacier or moraine, breaks, releasing a significant amount of water and debris downstream causing flash floods in the downstream area, ruining crops, damaging infrastructure like bridges and hydropower plants, and resulting in fatalities. The majority of these flood episodes are linked to processes like dam breach creation and subsequent damming moraine failure (Richardson and Reynolds 2000; Emmer and Vilímek 2013; Worni et al. 2014). The possible triggers of a GLOF event may be glacial calving activity at the lake terminus (Zhang et al. 2020), snow or ice avalanches (Schaub et al. 2016), landslides, extreme weather events (Schmidt et al. 2020) like cloudbursts (Bhambri et al. 2016) or seismic activity (Kargel et al. 2016). These events are potent threats whose effects can reach distances of more than 100 km in the downstream region (Richardson and Reynolds 2000).
+
+GLOFs pose a threat to about 15 million people globally and the most prominent high-vulnerability zone is the Himalayan region followed by the Andes (Taylor et al. 2023). Due to the Himalayan terrain's steep slopes and constrained flow pathways, GLOF episodes have typically been described as being particularly catastrophic (Mool 1995; Reynolds 1998; Richardson and Reynolds 2000; Worni et al. 2013). The recent Kedarnath disaster in the Central Himalayas in June 2013, which was brought on by unrelenting rain and subsequent eruption of the Chorabari glacial lake, serves as an illustration of the degree of fatality these occurrences can bring about (Das et al. 2015; Allen et al. 2016). The increased GLOF hazard in the region requires rigorous research initiatives for the identification and monitoring of hazardous glacial lakes in this region. Given that most of the glacial lakes in the Himalayas are inaccessible because of its tough terrain and harsh climate, remote sensing methods are used to monitor these lakes through time and space. Numerous studies have been conducted in the past to evaluate the distribution and GLOF hazard of glacial lakes throughout the world, especially in the Himalaya, which includes countries like Nepal, Tibet, Bhutan, and India (Bajracharya et al. 2007; Bolch et al. 2008; Fujita et al. 2009; Ives et al. 2010; Shrestha et al. 2010; Raj et al. 2013; Rounce et al. 2017; Emmer et al. 2018; Jiang et al. 2018; Nie et al. 2018; Byers et al. 2018; Emmer et al. 2020; Ashraf et al. 2021; Ahmed et al. 2022a, 2022b; Banerjee and Bhuiyan 2023).
+
+The Himalayan region, hence, has been the most well-known case for GLOF disaster risk studies. In the central Himalayas, a GLOF hazard vulnerability study by the International Centre for Integrated Mountain Development (ICIMOD), highlighted some priority sites for GLOF risk mitigation. The study mapped 1,947 glacial lakes using data from the Global Land Ice Measurement from Space (GLIMS) program, of which 835 were deemed dangerous because of their locations upstream of populated areas and major roadways. The study showed that Nepal urgently needs to reduce GLOF danger, especially in the regions downstream of the 835 little glacial lakes (Bajracharya et al. 2007). Rounce et al. (2017) evaluated the risk presented by GLOFs in Nepal using satellite data from 2000 to 2015 and the previous GLOF data. The risk connected to each lake was estimated by combining the hazard and projected downstream effects. The results reported 31 high-risk lakes and 11 very high-risk lakes.
+
+For the Upper Indus River basin, Gupta et al. (2021) conducted an assessment of potentially hazardous glacial lakes identifying 20 and 48 high-risk lakes for two different GLOF scenarios respectively. The evolving status and potential consequences of glacier lakes and the GLOF threat in the Hindu Kush Himalaya region were studied by Ashraf et al. (2021). An overall increase of 26% in the number of glacial lakes was found from 2001 to 2013 and 36 lakes were characterized as potentially dangerous. The increased risk posed by GLOFs in the context of a warming climate is highlighted, along with the necessity for proactive monitoring, risk assessment, and adaptation measures. From an Indian perspective, Jammu and Kashmir has the highest cumulative degree of GLOF hazard, followed by Uttarakhand, Arunachal Pradesh, Sikkim, Himachal Pradesh, and Arunachal Pradesh (Mal et al. 2021). Dubey and Goyal (2020) also placed J&K in second place, only after Sikkim, concerning the GLOF threat. Ahmed et al. (2022) analyzed the GLOF hazard scenario in the Jhelum basin using remote sensing and GIS and identified 21 potentially dangerous and 7 highly hazardous lakes.
+
+Studies focusing on individual glacial lakes have also been conducted in all parts of the Himalayan region. Majeed et al. (2021) reconstructed the 2014 Gya glacial lake outburst flood in Ladakh using the hydrodynamic model HEC-RAS and the results revealed the possible cause of the GLOF to be the piping failure. The study also predicted the possibility of 5 times more severe GLOF in the future from the lake. Mir et al. (2018) evaluated the GLOF risk of Dalung and Padam lakes in the Zanskar Valley using the Mike 11 model. The results declared both the lakes as potentially dangerous and susceptible to GLOFs of significant intensities. Sattar et al. (2021) analyzed the GLOF potential of Lower Barun glacial lake in Sikkim Himalayas and the lake was found to be susceptible to GLOF in case of a strong triggering event. Ahmed et al. (2023) estimated the risk associated with the Gangabal lake in Kashmir Himalayas using the HEC-RAS model. The results indicate that the lake has the potential to generate a peak discharge of 16,601.03 m<sup>3</sup>/s in the worst-case scenario at the time of a GLOF event. Rawat et al. (2023a) assessed a PDGL in the Satluj River basin in India and reported an areal expansion from 0.11 to 0.26 km<sup>2</sup> over the past 28 years from 1990 to 2018 and high GLOF risk. Sattar et al. (2019) conducted the moraine breach modeling of the South Lhonak lake in Sikkim and declared the lake to be highly vulnerable to GLOF, owing to its alarming rate of expansion over the
+
+past few decades and a loosely consolidated moraine dam. This very lake witnessed a severe GLOF disaster on October 4, 2023, killing at least 40 people and destroying public infrastructure in its downstream area. Similar studies have been conducted far and wide in the Himalayas that focus on predictive GLOF modeling of hazardous glacial lakes (Emmer and Vilímek 2013; Haemmig et al. 2014; Klimeš et al. 2014; Round et al. 2017; Wang et al. 2018, Sattar et al. 2019; Hussain et al. 2020; Majeed et al. 2021; Sattar et al. 2021; Ahmed et al. 2022; Wang et al. 2022; Gouli et al. 2023; Das et al. 2024; Schmidt et al., 2020) etc.
+
+The present study is one such attempt at understanding the GLOF hazard risk associated with a rapidly expanding moraine-dammed glacial lake in the Ladakh region of the Western Himalayas using multi-criteria decision analysis and 2D GLOF modeling. The identified PDGL is situated above the Panikhar and Pranti villages in the Kargil district of Ladakh. This study focused on the following objectives. (1) Studying the glacier-lake interactions of the Panikhar glacier and glacial lake for the past 22 years (2) Estimating the mass balance, ice thickness, and surface velocity of the Panikhar glacier, (3) Assessing the GLOF hazard level of the lake and associated risk to the downstream community using geospatial analysis and hydrodynamic modeling.
+
+# 2. Study area
+
+The glacial lake under investigation in this study is situated above the Panikhar and Pranti villages of the Kargil district in the Suru River catchment of the Western Himalayas. This lake is located between 34°2′2.154″N latitude and 75°46′59.172″E longitude, 4062 meters above sea level while its mother glacier lies, in close proximity, to its West (Figure 1). The Suru catchment is drained by the Suru River which flows through Kargil district and subsequently joins the Indus as its left bank tributary. The total catchment area is 4408 km² and the elevation range is from 2588 masl to 7047 masl. The mother glacier of the lake (GLIMS ID: G075817E34034N) is situated in the southwest of the Suru River catchment at a distance of about 15 kilometers from the well-known Parkachik glacier. The glacial lake or its mother glacier has not been identified with any specific names in previous literature. Because of this, we have given them the names of the village that is closest to them: Panikhar Glacier and Glacial Lake. The Panikhar village is situated about 13 kilometers downstream in the North-East of the glacier-lake complex and is vulnerable to potential GLOF events from the lake.
+
+## 2.1. Glacial geomorphology
+
+Panikhar glacier is oriented in the West-East direction and the lake is proglacially located close to its snout at its eastern side. The ablation zone of the glacier exhibits a bend towards the northeast according to the local geomorphology. The elevation of the glacial valley ranges from 5500 to 4100 meters above sea level (masl) and the glacial lake is situated at a height of 4100 masl. The contour lines clearly show that the steep slope to the northwest of the glacial lake is prone to landslides and snow
+
+Image /page/5/Figure/3 description: A figure titled "Figure 1. Location map of the Panikhar glacier, its proglacial lake and Panikhar village." The figure is composed of several maps and photographs. In the top left, two maps show the location of the study area within India and the region of Jammu & Kashmir. The main map in the top center is a topographic map of the Suru Catchment, with elevations ranging from a low of 2588m to a high of 7047m. This map includes a legend for the Glacial lake, Panikhar Glacier, Suru Catchment, and River Streams. To the right of this map are three photographs: one showing icebergs floating on the lake's surface, a full view of the lake, and a view of the Moraine Dam. The bottom half of the figure is a large Landsat 8 satellite image from October 2022. This image details the Panikhar Glacier, the glacial lake, the downstream flowpath, and Panikhar Village, which are outlined and labeled. A legend for this satellite image indicates that the RGB channels correspond to Band\_5 (Red), Band\_4 (Green), and Band\_3 (Blue).
+
+Figure 1. Location map of the Panikhar glacier, its proglacial lake and Panikhar village.
+
+avalanches, which might catalyze a GLOF. The glacial lake is impounded by the glacier's terminal moraine and the Panikhar settlement lies 13 km northeast of the glacial lake (Figure 2).
+
+# 3. Datasets
+
+This study used Sentinel-2A (10 m), PlanetScope (3 m), Landsat TM (30 m), Landsat ETM+ (30 m), and Landsat OLI (30 m) for the mapping of glacial lake and its mother glacier and estimating area changes in the glacial lake and its feeding glacier from 2000 to 2022. Moreover, high-resolution Google imagery was also used for verification and validation purposes. Advanced Land Observing Satellite (ALOS) Phased Array Synthetic-Aperture Radar (PALSAR) radiometrically terrain corrected (RTC) Digital Elevation Model (DEM) was downloaded online from https://asf.alaska.edu (accessed on 23rd May 2023), from which topographic information such as elevation, slope, aspect, hill shade, etc. was obtained. Furthermore, ASTER DEM (30 m) was
+
+Image /page/6/Figure/1 description: A topographical map showing the glacial geomorphology of the Panikhar glacier. The map includes latitude and longitude lines, with longitude ranging from 75°47'0"E to 75°51'0"E and latitude from 34°1'0"N to 34°3'0"N. A scale bar indicates a range up to 2.25 km. The map displays contour lines with elevations from 4100 to 5800 meters. A legend defines the symbols used on the map: a black outline for Glacial Valley, a thin line for Contour line, light blue shading for Panikhar glacier, darker blue for Glacial lake, a red triangle for U shaped valley, a green asterisk for Cirque, a purple diamond for Roche moutainee, a red dashed line for Lateral moraines, and a thick black-and-white dashed line for Terminal moraine. The map illustrates the glacier, a glacial lake, and the locations of these various geomorphological features. An inset map in the bottom right corner shows a satellite image of the same area.
+
+Figure 2. Glacial geomorphology of the study area.
+
+Table 1. Datasets used in the study.
+
+| S. No | Data/Tool               | Resolution | Year                   | Source                                                    |
+|-------|-------------------------|------------|------------------------|-----------------------------------------------------------|
+| 1.    | Planet CubeSat          | 3m         | 2016, 2018, 2020, 2022 | Planet<br>https://www.planet.com/                         |
+| 2.    | Landsat 7 ETM+          | 30 m       | 2000, 2010             | USGS Earth Explorer<br>http://www.earthexplorer.usgs.gov  |
+| 3.    | Landsat 8 OLI           | 30 m       | 2022                   | USGS Earth Explorer<br>http://www.earthexplorer.usgs.gov  |
+| 4.    | Sentinel 2 A            | 10 m       | 2016, 2018, 2020, 2022 | USGS Earth Explorer<br>http://www.earthexplorer.usgs.gov/ |
+| 5.    | ALOS-PALSAR<br>DEM      | 12.5 m     | 2015                   | https://asf.alaska.edu/                                   |
+| 6.    | SRTM DEM                | 90 m       | 2000                   | USGS Earth Explorer<br>http://www.earthexplorer.usgs.gov/ |
+| 7.    | ASTER DEM               | 30 m       | 2022                   | https://www.earthdata.nasa.gov/                           |
+| 8.    | Copernicus (GLO 30) DEM | 30 m       | 2023                   | https://spacedata.copernicus.eu/collection                |
+| 9.    | Google Earth Imagery    | <1 m       | 2000–2022              | Google Earth Pro Tool                                     |
+| 10.   | CRU gridded data        | 0.5°       | 1990–2002              | https://crudata.uea.ac.uk/cru/data/hrg                    |
+
+also used for glacier mass balance estimation through the geodetic method. PlanetScope imagery (3 m resolution) was acquired from the Planet Labs website www.planet.com. Landsat data was downloaded *via* the USGS Earth Explorer site, https://earthexplorer.usgs.gov, and Sentinel-2A images were obtained from the "Copernicus internet hub of European Space Agency" (ESA). Table 1 gives a summary of the datasets utilized in this study. These datasets have been extensively used in previous studies to analyze changes occurring on the Earth's surface (Ahmad et al. 2022a; Imdad et al. 2023; Malik et al. 2024; Mushtaq et al. 2024; Saleem et al. 2024).
+
+Image /page/7/Picture/2 description: A black and white circular icon showing a stylized boat on water. The boat is a simple silhouette with a small mast, floating on a wavy line, all enclosed within a circle.
+
+# 4. Methods
+
+## 4.1. Glacial lake mapping and change detection
+
+We have employed Planet CubeSat RGB images to delineate the glacial lake outlines for the years 2016, 2018, 2020, and 2022. However, the Planet CubeSat data is available from the year 2016 onwards. Hence glacial lake outline delineation for 2000 and 2010 was carried out from Landsat 7 ETM+ images. Out of several techniques and approaches previously used for glacial lake mapping, we have employed an integrated approach using the normalized difference water index (NDWI) method coupled with manual digitization for the removal of any possible errors. NDWI is a widely employed method for the extraction of water pixels from a satellite image, using the difference in the reflectance values of Blue and Green bands (McFeeters 1996; Ahmed et al. 2022; Zhang et al. 2022; Rather et al. 2024a). However, the NDWI method faces certain issues while classifying pixels containing muddy patches, shadows, cloud cover, or ice (Kaplan and Avdan 2017; Wang et al. 2020; Ahmad et al. 2022b, Mir et al. 2022; Rather et al. 2024b). These issues were tackled through the generation of relief, slope, aspect, and hill shade maps from the ALOS Pulsar DEM accompanied by manual correction using high-resolution Google Earth imagery. Cloud cover acts as a major impediment in the process of mapping different features from satellite imagery, therefore, images were selected from the post-monsoon months. Wherever a cloud-free image was not available, images from the adjacent months were taken to address this problem. The high resolution of the Planet CubeSat images (3 m) proved helpful in the minimization of any errors in the lake boundary delineation.
+
+### 4.1.1. Estimation of glacial lake depth, volume and peak discharge
+
+A glacier lake's depth, volume, and peak flow are crucial factors in determining the GLOF threat connected to that specific glacial lake. For calculating the glacial lake depth and volume from satellite datasets, we have to rely upon certain empirical equations which are based on the glacial lake area, developed by researchers across the world. Field surveys of glacial lakes require enormous effort and resources owing to the harsh terrain of the Himalayan cryosphere; therefore, empirical equations are widely employed for this purpose (Evans 1986; Wang et al. 2011; Emmer and Vilímek 2013; Fujita et al. 2013). Several equations for the calculation of glacial lake depth, volume, and potential peak discharge have been developed which include those by Huggel et al. (2002), Wang et al. (2011), Qi et al. (2022), Patel et al. (2017), Fujita et al. (2013), etc. A summary of these equations is given in Table 2. To reduce bias, we computed the lake volume using these empirical equations and took the average of all the equations into account for the final result. For lake volume estimation the following equations: (Equations (1) and (2)), given by Qi et al. (2022), showed results closest to the average values. These equations use the geometry of the lake beside the lake area to calculate the lake volume.
+
+For lakes larger than 0.1 km<sup>2</sup>
+
+$$V = 40.67^{1.184} - 3.218 \frac{mxw}{mxl} \tag{1}$$
+
+Table 2. Empirical equations used for depth and volume calculations.
+
+| Empirical Equation                           | Parameter Calculated | Reference            |
+|----------------------------------------------|----------------------|----------------------|
+| $D = 0.104 \times A0.42$                     | Lake Depth           | Huggel et al. (2002) |
+| $D_{m} = 55 \times A0.25$                    | Lake Depth           | Fujita et al. (2013) |
+| $Dm = 4 \times 10^{-5} \times A + 5.0564$    | Lake Depth           | Patel et al. (2017)  |
+| $V = 0.104 \times A1.42$                     | Lake Volume          | Huggel et al. (2002) |
+| $V = 0.035 \times A1.5$                      | Lake Volume          | Evans (1986)         |
+| $V = 0.0578 \times A1.4683$                  | Lake Volume          | (Liu et al., 2020)   |
+| $V = 40.67^{1.184} - 3.218 \text{ mxw/mxl}$  | Lake Volume          | Qi et al. (2022)     |
+| $V = 557.4^{2.455} + 0.2005 \text{ mxw/mxl}$ | Lake Volume          | Qi et al. (2022)     |
+| $V = A \times D_{m}$                         | Lake Volume          | Patel et al. (2017)  |
+| $Q_{max} = 0.00077 \times V^{1.017}$         | Peak discharge       | Huggel et al. (2002) |
+| $Q_{max} = 0.72 \times V0.53$                | Peak discharge       | Evans (1986)         |
+
+For lakes smaller than 0.1 km<sup>2</sup>
+
+$$V = 557.4^{2.455} + 0.2005 \frac{mxw}{mxl}$$
+ (2)
+
+## 4.2. Mapping the feeding glacier
+
+We employed Planet CubeSat images along with Landsat 7 ETM+ and Sentinel 2 A imagery for mapping the feeding glacier of the lake. The Normalized difference Snow Index (NDSI) combined with visual interpretation was used for glacier area change detection analysis. NDSI method has been extensively used in the Himalayan cryosphere by several studies and is considered to be a reliable method for glacier mapping and change detection (Bolch et al. 2012; Kulkarni et al. 2012; (Paul et al. 2015)).
+
+The error encountered in the estimation of glacier and lake areas is directly proportional to the resolution of the satellite images and it was evaluated using Equation (3).
+
+$$A_{er} = 100 (n^{0.5} \times m) / A_{gl}$$
+ (3)
+
+Where n is the ratio of the glacial lake's perimeter to the spatial resolution, m is the unit pixel area of the image (m<sup>2</sup>), and  $A_{ol}$  is the glacial lake area (m<sup>2</sup>).
+
+### 4.2.1. Mass balance estimation of the glacier
+
+Estimating the mass balance of glaciers can reveal important information about how glaciers behave in response to climate change and how glaciers interact with glacial lakes over a certain period. Various studies have employed different techniques to estimate the mass balance of glaciers globally (Soruco et al. (2009), Zemp et al. (2009), Barrand et al. (2010), Pratap et al. (2016) and Kumar et al. (2019) such as the glaciological method ((Kulkarni, 1992)), geodetic method (Kumar et al. 2017; Majeed et al. 2021), Hydrological method (Bhutiyani 1999; Pratap et al. 2016), etc. However, in this study, we conducted the mass balance estimation of the Panikhar glacier using the geodetic method as this is a convenient method to estimate mass balance keeping in view the harsh terrain of the region ((Rashid et al. 2020)). The basis of the geodetic approach is the calculation of the glacier's surface elevation difference over
+
+time. This method computes the surface elevation difference or thickness change of the glacier using digital elevation models from different time periods. The glacier area and ice density assumptions are then used to translate the elevation change into mass loss. (Braithwaite 2002; Muhammad et al. 2019; Romshoo et al. 2023). SRTM DEM and ASTER DEM (30 m resolution), acquired from https://www.earthdata.nasa.gov/ for the years 2000 and 2022 respectively, were used to calculate the elevation change of the glacier during this period using the DEM differencing technique. The DEMs were pre-processed in ArcMap 10.3 and the elevation difference was computed by comparing the elevation values of corresponding pixels from the two DEMs, using the raster calculator tool. The elevation difference was then translated into volume change using the cell size of the DEM. Assuming the ice density to be constant at 850 kg/m<sup>3</sup>, the mass loss from the glacier was calculated in kg/m<sup>3</sup> which was subsequently converted into meter water equivalent (m.w.e.).
+
+### 4.2.2. Glacier surface velocity and ice thickness estimation
+
+Surface velocities for the Panikhar Glacier were determined using the Glacier Image Velocimetry (GIV), a feature-tracking-based technique. GIV is a software package specifically designed for computing glacier velocity fields at high spatial resolution, utilizing unique points or characteristics on the glacier surface visible in consecutive satellite images. This meticulous process involves filtering individual velocity maps based on user-specified criteria, detecting and eliminating anomalous values using outlier detection functions, and computing statistical metrics for each grid cell over time. Sentinel-2 and Landsat true color composites were analyzed using GIV, with 20 multitemporal images with 30-meter spatial resolution from 2017 to 2023. Ice-thickness modeling during the same period utilized Copernicus DEM data at the same spatial resolution. Sentinel-2 images were used as input data for estimating glacier velocity, spanning from 2017 to 2023, with GIV configured to derive surface velocity considering a minimum temporal baseline gap of approximately 11 d and a maximum of three years. The displacement for each pair of images was computed using a multi-pass frequency domain cross-correlation algorithm, with subsequent filtering to remove pixels exceeding maximum velocity thresholds and enhancing result accuracy through spatial smoothing and rectification of systematic georeferencing errors using stable grounds.
+
+The estimation of glacier thickness was conducted in MATLAB. For ice-thickness modeling, we employed the Copernicus DEM at the same spatial resolution. Our study adapted codes originally given by van Wyk de Vries et al. to suit our specific research requirements in our investigation of the Panikhar Glacier. We estimated ice thickness using an ice-velocity-based method. The equation follows the methodology proposed by (Gantayat et al., 2014), with further refinements by van Wyk de Vries et al., particularly focusing on the mean slope and surface velocity computations.
+
+$$H = \frac{4\sqrt{1 \cdot 5U_s}}{Af^3 \ (\rho f sin\alpha)3} \tag{4}$$
+
+In this equation, H depicts the ice thickness, Us denotes the surface velocity, f is the slope factor (often assumed as 0.8, especially suitable for Himalayan glaciers as noted by (Paterson, 1991)), g stands for the acceleration due to gravity (9.8 m/s<sup>-2</sup>),  $\alpha$  signifies the glacier slope, and  $\rho$  indicates the ice density (typically taken as 900 kg/m<sup>3</sup> according to (Farinotti et al., 2009)).
+
+Glacier volume estimation adopted a piece-wise method, involving the summation of the product of the squared ground pixel size and its respective ice thickness. In Equation (5),  $H_{j,k}$  represents the mean value of each pixel in the ice-thickness map, while r depicts the ground resolution of the pixel, assuming square pixels applicable to optical remote sensing data.
+
+$$Vp = \sum_{i=1}^{n_j} \sum_{k=1}^{n_k} H_{j,k} \times r^2$$
+ (5)
+
+## 4.3. GLOF hazard analysis
+
+In the Himalayan region, a widely used method for determining GLOF vulnerability is predictive analysis utilizing a multicriteria framework (Bolch et al. 2008; Wang et al. 2011; Worni et al. 2013; Rounce et al. 2017; Wang et al. 2018; Dubey and Goyal 2020; Khadka et al. 2021; Ahmed et al. 2022). Despite the extensive use of this method in GLOF investigations, this method involves some amount of subjective expert judgment. This is because the GLOF susceptibility analysis is a complex process, and local variations in the influencing criteria make it nearly difficult to develop a uniform framework for this purpose globally (Allen et al. 2019; Taylor et al. 2023). In this study, we adopted the framework given by Che et al. (2014) as found to be suitable in our study area. This framework includes 10 relevant criteria pertaining to the glacial lake and its surrounding conditions that affect GLOF susceptibility and classifies the lakes into 5 levels of hazard from level 1 to level 5 using the weighted sum method (Table 3).
+
+## 4.4. GLOF modeling
+
+The two-dimensional (2D) GLOF modeling of the lake was done using HEC-RAS 2D (Version 6.3.1) to simulate two distinct scenarios: piping and overtopping. The aim was to assess the potential flood impacts from the glacial lake on the nearby downstream settlements. This approach has been extensively used to model the GLOF events globally as well as in the Himalayan region (Klimeš et al. 2014; Worni et al.
+
+Table 3. Integrated criteria used for assessing GLOF hazard, adopted from Che et al. (2014).
+
+| Index                                        | Criteria                     | Weight | Criteria Presence |
+|----------------------------------------------|------------------------------|--------|-------------------|
+| Type of glacial lake                         | Terminal moraine dammed lake | 0.15   | Yes               |
+| Area of lake                                 | >0.2 km2                     | 0.15   | Yes               |
+| Distance between lake and its mother glacier | <500 m                       | 0.15   | Yes               |
+| Average slope of the glacier                 | >7°                          | 0.10   | Yes               |
+| Slope of the downstream                      | >20°                         | 0.10   | No                |
+| Top width of the dam                         | <60 m                        | 0.10   | No                |
+| Area of the glacier                          | >2 km2                       | 0.09   | Yes               |
+| Slope between lake and its mother glacier    | >8°                          | 0.07   | No                |
+| Change of lake area                          | >10% in a decade             | 0.06   | Yes               |
+| Elevation of the lake                        | >5000 m                      | 0.03   | No                |
+
+Image /page/11/Picture/2 description: A close-up, black and white line drawing of a circle. Inside the circle, there is a horizontal line with a small, dark rectangle resting on top of it. Below the horizontal line is a dark, filled, semi-circular shape. There are faint markings that appear to be letters 'd' and 'f' above the rectangle.
+
+2014; Sattar et al. 2019; Sattar et al. 2021; Majeed et al. 2021; Ahmed et al. 2022; Rawat et al. 2023b).
+
+### 4.4.1. Model input and setup
+
+HEC-RAS requires a digital elevation model (DEM), for providing detailed terrain data for the study area. This raster-based DEM serves as the foundation for defining the flow area of interest, spanning from the glacial lake to the vulnerable settlements downstream. Within this defined flow area, a 2D computational mesh is constructed using cells of uniform dimensions. Each cell in the mesh is characterized by: Manning's roughness coefficient (N value), representing the flow resistance and the elevation data extracted from the DEM, specifying the terrain height at each cell location. ALOS PALSAR DEM was employed in this study given its better spatial resolution of 12.5 meters.
+
+### 4.4.2. Simulating two different GLOF failures
+
+- Piping Failure: This scenario simulates the failure of the lake dam or natural barrier due to internal erosion (piping). The modeling process involves steps like defining an initial breach hydrograph corresponding to the piping failure, setting the boundary conditions to initiate breach formation, and simulating flood propagation downstream. The 2D hydrodynamic modeling approach captures the breach evolution and flood wave propagation characteristics along the flow path. The characteristics of the potential flood wave were simulated for two different case scenarios presuming 50% and 100% drainage of the lake respectively.
+- b. Overtopping Failure: In overtopping failure, flood generation occurs when water levels exceed the capacity of the dam or barrier, causing water to spill over the top. The modeling procedure includes defining an appropriate initial breach hydrograph corresponding to the overtopping event and configuring the boundary conditions to simulate breach initiation and the subsequent flood wave propagation downstream. This is followed by employing the 2D hydraulic model to capture the breach dynamics and the flood wave behavior, considering terrain elevations and flow resistances. This was also done keeping in view the moderate and worst-case scenarios with 50% and 100% volume of the lake.
+
+### 4.4.3. Model outputs and analysis
+
+The 2D modeling process generates spatially distributed outputs including water depth distribution, flow velocity profiles indicating flood wave dynamics, and inundation extents depicting areas affected by the flood waters. These outputs enable detailed analysis of flood impacts along the route from the Lake to the settlements for both piping and overtopping scenarios. By integrating terrain data, hydraulic properties, and the breach conditions, the modeling study provides critical insights into potential flood hazards and helps in the development of mitigation frameworks for disaster preparedness and response. This comprehensive approach leverages advanced hydraulic modeling techniques to simulate complex flood scenarios, facilitating informed decision-making and risk assessment in vulnerable regions prone to GLOF events.
+
+# 5. Results
+
+## 5.1. Depth, volume, and peak discharge of the lake
+
+The glacial lake depth and volume estimations were carried out using the various regression equations stated in Table 4. The results of these equations vary considerably and hence an average value of all of these equations was considered as the final estimate of lake depth and lake volume to minimize the bias in the final estimates as suggested by Emmer (2018). Using this approach, the average depth of the lake was estimated as  $31.50\,\mathrm{m}$  while the average volume was estimated as  $14.04\times10^6\,\mathrm{m}^3$ . The equation developed by Fujita showed exaggerated estimates of lake depth in comparison to the other equations and in calculating the lake volume, the maximum value was obtained through Liu's equation while Qi's equation showed results closest to the average value of all equations.
+
+The estimation of the Peak discharge of the lake was done using the equations given by Huggel et al. (2002) and Evans (1986), and the values obtained from these equations are recorded in Table 4. To minimize the bias, the average value of these two equations, which came out to be 7954.19 m<sup>3</sup>s<sup>-</sup>1, was taken as the final estimate of the peak discharge of the lake as suggested by Emmer et al. (2018).
+
+## 5.2. Glacial lake area change
+
+The change detection analysis of the lake, using multi-temporal datasets, revealed an overall increase of 78.72% in the lake's area at a rate of  $0.017\,\mathrm{km^2}$  per annum. The lake area has gone up from  $0.10\,\mathrm{km^2}\,\pm\,75.9~\mathrm{m^2}$  in 2000 to  $0.22\,\mathrm{km^2}\,\pm\,0.82~\mathrm{m^2}$  in 2010,  $0.33\,\mathrm{km^2}\,\pm\,0.82~\mathrm{m^2}$  in 2016,  $0.37\,\mathrm{km^2}\,\pm\,0.72\,\mathrm{m^2}$  in 2018,  $0.41\,\mathrm{km^2}\,\pm\,0.67\,\mathrm{m^2}$  in 2020 and  $0.47\,\mathrm{km^2}\,\pm\,0.62~\mathrm{m^2}$  in 2022. The lake area has witnessed an overall increase of  $\sim\!0.37\,\mathrm{km^2}$  in the past 22 years. It has almost doubled in each decade from 2000 to 2020 which depicts a very rapid expansion scenario, much likely to accelerate further in the future decade of 2020–2030 keeping in view the increasing temperature trends in the region. Figure 3 depicts the expansion of the lake through lake outline maps.
+
+## 5.3. Changes in the feeding glacier of the lake
+
+### 5.3.1. Area changes
+
+The feeding glacier of Panikhar Lake was also subjected to spatio-temporal change detection analysis. The results revealed a significant decrease of  $\sim 0.61 \, \mathrm{km}^2$  in the glacier area accounting for a 13.2% overall decrease in the glacier area for the past
+
+|  | Table 4. A summary of | f various equations used | d for depth and vo | lume measurements. |
+|--|-----------------------|--------------------------|--------------------|--------------------|
+|--|-----------------------|--------------------------|--------------------|--------------------|
+
+| Emp. Equation | Depth (m) | Volume (106 m3) | Peak Discharge (m3/s) |
+|---------------|-----------|-----------------|-----------------------|
+| Huggel        | 25.08     | 11.78           | 11,963.14             |
+| Fujita        | 45.58     | –               | –                     |
+| Evans         | –         | 11.33           | 3945.24               |
+| Liu           | –         | 19.17           | –                     |
+| Qi            | –         | 14.43           | –                     |
+| Patel         | 23.86     | 11.21           | –                     |
+| Average Value | 31.50     | 14.04           | 7954.19               |
+
+Image /page/13/Figure/3 description: A figure displaying six satellite images in a 2x3 grid, illustrating the expansion of a lake from the year 2000 to 2022. The images are for the years 2000, 2010, and 2016 in the top row, and 2018, 2020, and 2022 in the bottom row. Each image shows the lake outlined and colored in dark blue against a terrestrial background. Each panel includes a scale bar from 0 to 0.4 kilometers, a north arrow, and coordinate axes for latitude (from 4°2'0"N to 34°2'20"N) and longitude (from 75°50'20"E to 75°51'0"E). The lake appears as two small, separate bodies of water in 2000, which merge and grow significantly in size through 2010, 2016, 2018, 2020, and 2022. The images for 2016, 2018, and 2022 show small white patches on the lake's surface. The caption below the images reads: 'Figure 3. Expansion of Pari lake from 2000 to 2022 depicted on Planet Cube Sat (RGB)'.
+
+Figure 3. Expansion of Panikhar glacial lake from 2000 to 2022 depicted on Planet CubeSat (RGB) and Landsat 7 ETM+ images.
+
+Image /page/13/Figure/5 description: A line graph showing the relationship between Lake Area and Glacier Area from the year 2000 to 2022. The x-axis represents the Year, from 1998 to 2024. There are two y-axes: the left y-axis represents Lake Area in square kilometers (km²), ranging from 0 to 0.5, and the right y-axis represents Glacier Area in square kilometers (km²), ranging from 3.8 to 4.8. The graph contains four lines: a solid green line for Lake Area, a solid red line for Glacier Area, a dashed green line for the linear trend of Lake Area, and a dashed red line for the linear trend of Glacier Area. The Lake Area shows a steady increase from 0.1 km² in 2000 to 0.47 km² in 2022. Specific data points for Lake Area are (2000, 0.1), (2010, 0.22), (2016, 0.33), (2018, 0.37), (2020, 0.41), and (2022, 0.47). The Glacier Area shows a general decrease from 4.63 km² in 2000 to 4.02 km² in 2022, with a sharp drop between 2018 and 2020. Specific data points for Glacier Area are (2000, 4.63), (2010, 4.58), (2016, 4.4), (2018, 4.39), (2020, 4.04), and (2022, 4.02). The trend lines confirm the inverse relationship: as the glacier area decreases, the lake area increases.
+
+Figure 4. Glacier and glacial lake area trends.
+
+22 years. The overall area of the glacier has decreased from  $4.63\,\mathrm{km}^2\pm139~\mathrm{m}^2$  in 2000 to  $4.58\,\mathrm{km}^2\pm138.4~\mathrm{m}^2$  in 2010,  $4.50\,\mathrm{km}^2\pm0.01\,\mathrm{m}^2$  in 2016,  $4.39\,\mathrm{km}^2\pm0.01\,\mathrm{m}^2$  in 2018,  $4.04\,\mathrm{km}^2\pm0.02\,\mathrm{m}^2$  in 2020 and  $4.02\,\mathrm{km}^2\pm0.02\,\mathrm{m}^2$  in 2022 with an annual decrease rate of  $\sim\!0.03\,\mathrm{km}^2$  (Figures 4 and 5). The glacier has witnessed a snout retreat of 348 m in this period. The glacier and the glacial lake areas behold a reciprocal relationship with each other which means that the glacial lake is expanding at the expense of its feeding glacier (Figure 4).
+
+### 5.3.2. Mass balance of the glacier
+
+The glacier mass balance was estimated through the geodetic method and the results reveal that the glacier has, on average, undergone a thickness loss of 7.04 m from 2000
+
+Image /page/14/Figure/1 description: A map titled "Figure 5. Spatio-temporal evolution of the Penikhar glacier for the last 22 years." The main image is a satellite view of a glacier, shown in light blue, surrounded by mountainous terrain. The map includes latitude and longitude coordinates, a scale bar from 0 to 1.7 Kms, and a north arrow. Overlaid on the glacier are several colored outlines indicating its extent in different years. A legend titled "Glacier Outlines" shows the following: a light purple line for 2022, a light green line for 2020, a blue line for 2018, a black line for 2016, a tan line for 2010, and a red line for 2000. The satellite data is from Sentinel 2A (2022), using Band 11 (red), Band 8 (green), and Band 3 (blue). In the bottom right corner, there is an inset map titled "Glacier Snout," which is a close-up of the glacier's terminus. This inset has its own scale bar from 0 to 0.4 Kms and a legend showing the glacier snout outlines for 2016 (dark green), 2010 (light blue), 2000 (black), 2022 (red), 2020 (pink), and 2018 (light green). The outlines on both maps generally show the glacier receding over time.
+
+Figure 5. Spatio-temporal evolution of the Panikhar glacier for the last 22 years.
+
+to 2022 at the rate of  $0.32\,\mathrm{m~a}^{-1}$ . The total mass balance of the glacier was estimated to be  $-5.922\,\mathrm{m}$  w.e. which amounts to a total mass loss of  $\sim 5.92$  metric tons per square metre during this period at the rate of 0.27 metric tons per square metre per year (Figure 6). The mass balance values across the glacier fall from  $-0.05\,\mathrm{m}$  w.e. to  $0.06\,\mathrm{m}$  w.e. with a mean negative value. This depicts that despite certain accumulation regions on the glacier, the glacier has an overall negative mean mass balance. These results are in proximity to other studies in the region. For instance, Kumar et al. (2017) reported an annual mass balance of  $-0.53\pm0.16\,\mathrm{m}$  we. per year for the glaciers in the Western Himalayas. Majeed et al. (2021) reported a mass balance of  $-0.24\,\mathrm{m}$  w.e. per year for clean glaciers and  $-0.37\,\mathrm{m}$  w.e. per year for debris-covered ones in the Pangong region.
+
+### 5.3.3. Surface velocity, ice thickness, and volume of the glacier
+
+According to the velocity estimations, the glacier's mean velocity was estimated to be  $3.38 \,\mathrm{m}$  a-1, while its surface velocity varied from  $0.87 \,\mathrm{m/y}$  to  $9.76 \,\mathrm{m}$  a<sup>-1</sup>. (Figure 7a). The glacier is characterized by a slowly moving snout with slightly increasing velocities in the upper accumulation zone. Previous studies in the region also report the existence of an increasing trend in the surface velocity of the glaciers with increasing elevation until a maximum is reached towards the transient snow lines (Bhushan et al. 2018; Ahmed et al. 2022).
+
+The ice thickness measurement of the glacier, conducted in MATLAB, reveals that the thickness of the glacier varies from 14.39 m to 50.87 m (Figure 7b) while the mean thickness was estimated to be 30.95 m. The thickness of the glacier increases
+
+Image /page/15/Figure/3 description: A map illustrating the spatial variation of mass balance on a glacier. The map is oriented with a north arrow pointing up. The glacier is outlined with a blue 'Glacier Boundary' line and is overlaid with a color gradient representing the 'Mass Balance (m.w.e)'. The color scale in the legend ranges from red (0.06) to yellow (-0.01) to green (-0.05), indicating areas of mass gain and loss. The western, higher-elevation part of the glacier shows a small red area of positive mass balance, while the majority of the glacier is yellow and green, indicating negative mass balance. Brown 'Contour (100m)' lines show the topography, with elevations ranging from 4100 meters at the eastern tip to over 5100 meters in the west. The map includes a scale bar from 0 to 2 kilometers. The area is located between approximately 75°47'0"E and 75°50'0"E longitude and 34°1'0"N and 34°3'0"N latitude.
+
+Figure 6. Spatial variation of geodetic mass balance across Panikhar glacier in meter water equivalent (m.w.e).
+
+Image /page/15/Figure/5 description: The image displays two maps, labeled A and B, showing characteristics of a glacier. Both maps are georeferenced with latitude and longitude coordinates and include a 1 km scale bar and a north arrow. Map A, titled "Glacier Velocity (m/y)", uses a color scale from blue (Low: 0.87 m/y) to red (High: 9.76 m/y) to show glacier surface velocity. The highest velocities are concentrated in the central, wider part of the glacier, indicated by yellow and red patches. Map B, titled "Glacier Thickness (m)", uses a color scale from blue (Low: 14.39 m) to magenta (High: 50.87 m). This map shows the glacier is thickest in the central area, depicted in magenta, and thins out towards the edges, shown in blue.
+
+Figure 7. Figure depicting (A) glacier surface velocity and (B) Ice thickness distribution.
+
+Image /page/16/Figure/1 description: A line graph displays Mean Temperature and Total annual precipitation from 1990 to 2020. The x-axis represents the Year, from 1990 to 2020. The left y-axis represents Temperature in degrees Celsius, ranging from -1 to 8. The right y-axis represents Precipitation in millimeters (mm), ranging from 50 to 400. There are two data series plotted. The first, in red with square markers, shows the Mean Temperature (T\_mean). It exhibits a clear upward trend over the period, starting near 0 degrees Celsius in 1990 and rising to about 1 degree Celsius by 2020. A red dashed trendline with a shaded confidence interval is shown, with an R-squared value of R²=0.275. The second data series, in green with square markers, shows the Total annual precipitation. This series shows significant year-to-year fluctuation, with values ranging from approximately 150 mm to over 300 mm. A green dashed trendline with a shaded confidence interval is nearly horizontal, indicating no significant trend, which is supported by a very low R-squared value of R²=0.0138.
+
+Figure 8. Temperature and precipitation trends in Ladakh.
+
+towards the center and decreases as we move towards the edges. The total glacier volume as estimated from the thickness estimates and the glacier pixel areas was found to be 45,783 m<sup>3</sup>.
+
+## 5.4. Temperature and precipitation variability over the region
+
+The Ladakh region is overall a data-scarce region with only a few observatories for the generation of observed temperature and precipitation data. Keeping this in mind, we projected the trends in the mean annual temperature (T mean) and Total annual precipitation in the region using a gridded climatic dataset given by the Climate Research Unit (CRU) to peek into the climatic controls of the glacial lake expansion. It was found that the mean annual temperature has witnessed a slight positive trend ( $R^2 = 0.275$ ) during this period while the total annual precipitation in the region has witnessed a slight negative trend with an  $R^2$  value of 0.0138 (Figure 8).
+
+The warming temperature trends in this region and declining precipitation trends are also confirmed by previous studies that use various gridded climate data products to analyze the climatic variability in Ladakh (Shafiq et al. 2016; Ahmad et al. 2019; Mann et al. 2022). This may lead to further glacier recession and subsequent glacial lake expansion in the near future making the local populations vulnerable to cryospheric disasters such as the GLOF.
+
+# 6. GLOF risk
+
+## 6.1. GLOF susceptibility analysis
+
+We carried out the GLOF susceptibility analysis of the glacial lake based on the approach used by Che et al. (2014) which includes criteria pertaining to both the lake and the surrounding conditions that affect GLOF susceptibility (Table 3) and classifies the lakes into 5 levels of hazard from level 1 to level 5. The lake was found to be a High-hazard lake associated with a level 4 hazard. Some of the major reasons that account for the high hazard index of the lake are its large size, unstable moraine dam, direct connection to its mother glacier, marked rate of expansion, and a large-
+
+Image /page/17/Figure/3 description: A geographical map illustrating potential zones of mass movement in a mountainous region. The map is set against a satellite image background and includes longitude and latitude markings, a scale bar, and a compass rose. An inset map in the top-left corner provides a closer satellite view of the 'Glacier Snout' and 'Panikhar Lake', with a steep slope outlined in red. Red lines connect this outlined area to the main map, which highlights the 'Potential Zones of Mass Movement'. These zones are color-coded based on slope steepness. A legend in the bottom-right corner defines the symbols and colors used. 'Glacial Lake' is shown in solid light blue, and 'Glacier' is a white area with blue hatching. The 'Slope (degrees)' is categorized by color: green for 0-8, light green for 8-19, yellow for 19-31, orange for 31-45, and red for >45. The map shows that the areas with the highest potential for mass movement (red and orange) are on the steep slopes adjacent to the glacial lake and glacier.
+
+**Figure 9.** The local topographic setting of the Panikhar glacier-lake complex and zones of possible mass movement.
+
+sized mother glacier (4.02 km<sup>2</sup>) with a slope of 16.5°. Moreover, the lake is prone to rock fall/debris fall from its adjoining steep slopes as evident from the slope profile of the lake surroundings (Figure 9).
+
+## 6.2. Hydrodynamic modelling
+
+We analyzed the potential impacts of an outburst event originating from the Panikhar Lake by analyzing the temporal and spatial properties of the flood wave using two-dimensional hydraulic modeling for two different case scenarios. In scenario 1, we assumed that 50% of the total volume of the lake is drained whereas scenario 2 was the worst-case scenario assuming a 100% drainage of the lake.
+
+### 6.2.1. Scenario 1: assuming 50% water volume of the lake is released
+
+Given the dynamic nature of high-magnitude flood events, we focused on understanding the changing hydraulic properties over time, particularly examining discharge, flow velocity, and depths across the flow channel. Assuming drainage of 50% of the lake's total water volume, we found that the maximum discharge at the point of the breach was 1908.25 m³/s and 3271.50 m³/s for the piping and overtopping failures respectively (Figure 10a). The routed hydrographs depict that the flood wave in this scenario will reach the first settlement in 1 h and 4 min after the dam breach in both the piping and the overtopping failures with a peak discharge of 2701.1 m³/s and 2714.7 m³/s respectively (Figure 11a). The flood depths along the flow path are
+
+Image /page/18/Figure/1 description: A close-up, black and white image showing the number 18 followed by a plus sign enclosed in a circle.
+
+Image /page/18/Figure/2 description: The image contains two line graphs, labeled A and B, which are breach hydrographs comparing 'Piping Failure' and 'Overtopping failure' scenarios. Both graphs plot Discharge (m³/s) on the y-axis against Time (hh:mm:ss) on the x-axis.
+
+Graph A has a y-axis ranging from 0 to 3500 m³/s. The blue line represents 'Piping Failure' and shows a peak discharge of 1908.25 m³/s. The orange line represents 'Overtopping failure' and shows a higher peak discharge of 3271.50 m³/s. The x-axis ranges from 1:00:00 to 4:22:55.
+
+Graph B has a y-axis ranging from 0 to 5000 m³/s. The green line represents 'Piping Failure' with a peak discharge of 3890.99 m³/s. The red line represents 'Overtopping Failure' with a higher peak discharge of 5111.39 m³/s. The x-axis ranges from 1:00:00 to 3:25:42.
+
+The caption below the graphs reads: 'Figure 10. Breach hydrographs depicting peak discharge at the breach in piping and overtopping'.
+
+Figure 10. Breach hydrographs depicting peak discharge at the breach in piping and overtopping failures (A) 50% volume of the lake (B) 100% volume of the lake (worst case scenario).
+
+depicted in Figure 12. The maximum flood velocity was estimated to be 18.26 m/s while the maximum depth of the flood wave was found to be 14.87 m.
+
+The maximum area under the threat of inundation along the valley was calculated as 4.74 km<sup>2</sup> for both the overtopping and pipping failure situations. The mean flood depth across the flow channel was estimated to be 3.41 meters and the mean flood velocity was found to be 3.84 m/s. We found that the potential flood wave in this scenario inundates bridges, roads, and a few residential houses along the channel in settlement 2 and near the Suru Valley bridge (Figure 13).
+
+### 6.2.2. Scenario 2: assuming 100% water volume of the lake is released
+
+The maximum discharge, in the worst-case scenario with 100% water release, at the point of breach was calculated to be 3890.99 m<sup>3</sup>/s and 5111.39 m<sup>3</sup>/s for the piping and overtopping scenarios respectively (Figure 10b). A routed flood wave in the simulation reached the closest settlement (settlement 1) of the Panikhar village in 50 min after the lake-breach event with a peak discharge of 6345.06 m<sup>3</sup>/s and 6453.9 m<sup>3</sup>/s for the piping and overtopping scenarios respectively (Figure 11b).
+
+Image /page/19/Figure/3 description: The image displays four line graphs arranged in a 2x2 grid, labeled A and B on the right side for each row. The graphs show discharge over time for different locations under two failure scenarios: "Piping Failure" (left column) and "Overtopping Failure" (right column).
+
+Row A contains two graphs with the y-axis representing Discharge (m³/s) from 0 to 3000. The x-axis represents Time in hh:mm:ss format.
+- The top-left graph, "Piping Failure," shows five hydrographs for different locations: Settlement 1, Surrur Valley Bridge, Settlement 2, Bridge 2, and Picnic Spot Damasna. The peak discharge for Settlement 1 is approximately 2600 m³/s at around 2:10:36. The other locations have lower peaks that occur later in time.
+- The top-right graph, "Overtopping Failure," shows similar hydrographs. The peak discharge for Settlement 1 is also around 2600 m³/s but occurs earlier, at approximately 2:00:04.
+
+Row B contains two graphs with the y-axis representing Discharge (m³/s) from 0 to 8000. The x-axis represents Time in hh:mm format.
+- The bottom-left graph, "Piping Failure," shows five hydrographs for: Settlement 1, Suru Valley Bridge, Settlement 2, Bridge 2, and Picnic Spot Damasna. The peak discharge for Settlement 1 and Suru Valley Bridge is approximately 6400 m³/s at around 02:00. The other locations have lower peaks that occur later.
+- The bottom-right graph, "Overtopping Failure," shows similar hydrographs. The peak discharge for Settlement 1 and Suru Valley Bridge is also around 6400 m³/s but occurs slightly earlier, at approximately 01:55.
+
+**Figure 11.** Routed hydrographs at various locations along the downstream flow channel in piping and overtopping failures (A) Assuming release of 50% of the water volume of the lake (B) Assuming release of 100% of the water volumeof the lake.
+
+The two-dimensional routing of the breach hydrograph is simulated over a distance of 25.6 km along the flow channel, from the lake to the Damasna picnic spot in Panikhar village. The maximum area under the threat of inundation along the valley was calculated as 5.38 km² for both the overtopping and piping failures (Figure 16). Across the entire channel, the mean flow discharge and mean flow depth are recorded at 588 m³/s and 5.31 m for pipping failure 755 m³/s, and 5.33 m for overtopping failure with a mean flow velocity of 5.37 m/s. The flood depth and velocity vary along the flow channel as per the terrain profile of the flow channel (Figure 13). The flood depth and velocity values decline towards the lower reaches of the flow channel, as a result of the increase in the channel width and decrease in the slope.
+
+The potential flood wave inundates infrastructure such as roads, bridges, and settlements along the channel with varying depths and velocities for both the piping and overtopping failures (Figures 14 and 15, Table 5). We evaluated the flood depths and flood velocities along the flow channel at settlements 1 and 2 to determine the maximum flood depths, peak flood velocities, and the time of the peak. Settlements 1 and 2 witness the highest potential flood depths of 7.9 and 7.65 m in the overtopping scenario whereas the lowest flood depths are witnessed at the Damasna picnic spot. Settlement 1 witnesses the highest peak velocities of 12.36 m/s and 11.35 m/s in the Piping and the overtopping scenarios (Figure 15).
+
+# 7. Discussion
+
+The Himalayan region is a hotspot of GLOF risk (Taylor et al. 2023) and this risk is projected to increase three times by the year 2100 (Zhang et al. 2024). 18% of all the
+
+Image /page/20/Figure/1 description: A close-up, black and white image showing the number 20 next to a circular icon. The icon contains a simple line drawing of an oil lamp with a flame, and a short line extends downwards from the bottom of the circle.
+
+Image /page/20/Figure/2 description: This image, labeled Figure 12, displays two sets of maps illustrating flood depth and velocity profiles in a downstream flow channel. The top set of three maps, labeled (A), (B), and (C), shows the 'Flood Depth (m)'. A color scale indicates that the depth ranges from a low of 0.001 m (blue/green) to a high of 14.87 m (red). The maps depict a river channel winding through a mountainous terrain, with varying depths indicated by the color coding. The bottom set of three maps, also labeled (A), (B), and (C) for the same geographical areas, shows the 'Flood Velocity (m)'. The corresponding color scale ranges from a low of 0.001 m (green) to a high of 18.26 m (red). Each map includes latitude and longitude coordinates and a 2 km scale bar.
+
+Figure 12. Flood depth and velocity profiles in the downstream flow channel in the first case scenario (50% volume). Parts A, B and C represent the 1st, 2nd and 3rd sections along the flow channel.
+
+Image /page/21/Figure/3 description: A composite image showing several annotated satellite views of a river valley. The river and its floodplain are highlighted in a translucent green. The main panel on the left shows a wide view of the river with a white arrow indicating the "Flow Direction." Several points are marked with red pins and labels along the river: "Picnic Spot Damasna," "Bridge 2," "Settlement 2," "Settlement 1," and "Suru Valley Bridge." Red boxes highlight the areas of Bridge 2 and Settlement 2. The middle panel is a closer view of the "Suru Valley Bridge" area, with white arrows pointing to structures labeled "Houses," one of which is circled. The right side of the image contains two smaller, zoomed-in panels. The top right panel shows a close-up of "Settlement 2," with a red box around it and an arrow pointing to "Houses." The bottom right panel is a close-up view of "Bridge 2" spanning the river.
+
+Figure 13. Buildings, roads, bridges and other infrastructure under potential threat of GLOF inundation in scenario (1) in the downstream flow channel of Panikhar lake.
+
+previous GLOF events have occurred in the High Mountain Asia (HMA) which includes the Himalayan region. Most of the GLOF events in Himalayan region have been reported from the moraine and ice-dammed glacial lakes (Shrestha et al. 2010). The key triggers of GLOF in the region are extreme precipitation events and mass movements such as ice avalanches (Shrestha et al. 2010; Zhang et al. 2024). The rapid expansion of mountainous populations into high-risk zones is one of the major contributors to increasing GLOF risk in this region (Taylor et al. 2023). A large body of literature has been produced by researchers to generate baseline data about the GLOF phenomenon in the data-scarce Himalayan region (Emmer et al. 2022). These studies have focused on the preparation of glacial lake inventories, identification of potentially dangerous glacial lakes, and estimation of GLOF risk associated with hazardous lakes.
+
+Categorizing a lake as potentially dangerous is very difficult given the multitude of factors influencing a GLOF event such as lake size, lake dam, lake surroundings, landslide, avalanche susceptibility, or extreme weather event. Therefore, under particular circumstances, an otherwise stable lake can also produce a GLOF event. Extreme precipitation events have caused notable GLOF events in the past through dam overtopping or collapse. This adds to the complexity of understanding and predicting the GLOF disaster. The same problem arises with the categorization of glacial lakes as potentially dangerous or otherwise based on size. Generally, some area threshold (e.g.  $A > 0.1 \text{ km}^2$ ) is defined to exclude smaller lakes from the list of potentially dangerous lakes. For instance, Worni et al. (2013) did not classify the Gya Lake in Ladakh as potentially hazardous, maybe because it was ice-covered or because its size was below the mapping threshold of the study (0.01 km<sup>2</sup>), leading to the assumption that it did not pose any significant threat to the settlements downstream. It can be challenging to leave out lakes that may be dangerous, as the same was demonstrated by the 2013 Chorabari lake outburst event (Das et al. 2015; Bhambri et al.
+
+Image /page/22/Figure/1 description: This image is Figure 14, which displays flood depth and velocity profiles for a downstream flow channel. The figure is divided into two main horizontal sections. The top section shows flood depth, and the bottom section shows flood velocity. Each section contains three maps labeled (A), (B), and (C), which depict different segments of a river flowing through a mountainous terrain. For the top section, a color scale represents 'Flood Depth (m)', ranging from a low of 0.001 (blue) to a high of 17.60 (red). For the bottom section, a color scale represents 'Flood Velocity (m/s)', ranging from a low of 0.001 (green) to a high of 23.78 (red). The maps include geographical coordinates, scale bars in kilometers, and a north arrow on the 'C' panels.
+
+Figure 14. Flood depth and velocity profiles in the downstream flow channel in the worst case scenario (100% volume). Parts A, B and C represent 1st, 2nd and 3rd sections along the flow channel.
+
+2016). According to several studies, small glacial lakes have the potential to seriously harm local livelihoods and cause major problems, as demonstrated by the Domkhar food in Ladakh (Ikeda et al. 2016) and other minor foods in Hunza region (Ashraf et al. 2021). The triggering factors of the GLOF are often dynamic and may change from time to time both quantitatively and qualitatively and hence require regular monitoring. e.g. Lakes that are stable as such, may get exposed to unstable
+
+Image /page/23/Figure/3 description: A figure with two side-by-side panels, each showing a map with a flood depth overlay and an inset graph of water velocity over time. The left panel is located at 75°57'0"E and shows a map with a settlement labeled "Settlement 1". The flood depth is represented by a color scale from blue (Low: 0.001 m) to red (High: 17.60 m). The inset graph shows velocity in m/s on the y-axis and time in hh:mm on the x-axis, from 01:26 to 03:50. The velocity peaks at over 10 m/s around 02:00. The right panel is located at 75°58'0"E and 34°8'0"N and shows a map with a settlement labeled "Settlement 2". It uses the same depth color scale. The inset graph shows velocity peaking at over 5 m/s around 02:00 over the same time period. Both maps include a north arrow.
+
+Figure 15. Flood depth and flood velocity curves at settlements 1 and 2 in Panikhar village.
+
+Table 5. Flood parameters at various locations of the flow path downstream of Panikhar lake for the worst case scenario i.e.100% water release.
+
+| Location               | Distance from the Lake (km) | Overtopping Failure |                  |                        | Pipping Failure   |                  |                        |
+|------------------------|-----------------------------|---------------------|------------------|------------------------|-------------------|------------------|------------------------|
+|                        |                             | Mean<br>Depth (m)   | Max<br>Depth (m) | Peak Velocity<br>(m/s) | Mean<br>Depth (m) | Max<br>Depth (m) | Peak Velocity<br>(m/s) |
+| Settlement 1           | 15.6                        | 4.35                | 7.97             | 11.35                  | 3.05              | 6.61             | 12.36                  |
+| Suru valley<br>Bridge  | 16.11                       | 2.25                | 4.16             | 9.05                   | 1.96              | 4.29             | 8.66                   |
+| Settlement 2           | 18.01                       | 4.63                | 7.65             | 5.27                   | 4.71              | 7.67             | 3.70                   |
+| Bridge 2               | 19.08                       | 2.73                | 5.33             | 5.82                   | 1.97              | 3.99             | 5.36                   |
+| Picnic Spot<br>Damasna | 22.50                       | 1.70                | 6.96             | 6.615                  | 5.13              | 6.70             | 7.187                  |
+
+Image /page/23/Picture/7 description: A composite figure displaying satellite maps of a river valley to highlight specific locations and features. The main panel on the left shows a winding river in a mountainous region, with a green shaded area along its banks. The river's flow direction is indicated by a white arrow pointing upwards. Several locations are marked with red pins and labeled, including "Settlement 1", "Suru Valley Bridge", "Settlement 2", "Bridge 2", and "Picnic Spot Damasna". On the right, there are three zoomed-in panels. The top panel shows a close-up of "Settlement 2", with red arrows pointing to "Houses". The middle panel is a close-up of "Bridge 2", with red arrows indicating "Bridge 2 and buildings". The bottom panel provides another view of the area, with a red oval highlighting "Roads and houses in Settlement 2".
+
+Figure 16. Inundation area in scenario 2 (100% water release) with some important areas with buildings, roads, bridges and other infrastructure under potential threat of GLOF inundation.
+
+surrounding slopes or ice avalanche zones after expanding into the glacial valleys thus enhancing their susceptibility (Haeberli et al., 2017).
+
+As moraine and ice-dammed lakes are more vulnerable to GLOF, they should receive greater attention when it comes to GLOF prioritization. Also, lakes that are in close proximity to the parent glacier are more likely to produce a GLOF as these lakes are more at threat of Icefalls and glacial calving activity. The likelihood of piping activity leading to a dam collapse is higher in lakes with underground drainage (Narama et al. 2018). Additionally, because thermokarst processes may block the intra-moraine channels, which can result in overflow and subsequent lake outburst through breaching, glacial lakes with subterranean drainage are more prone to GLOFs unlike the ones with stable surface drainage (Narama et al. 2018; Petrakov et al. 2020). Lakes with smaller dam width-to-height ratios and lower freeboard levels should be the focus of the GLOF study as they are important indicators of the dam's stability and can be estimated using high-resolution satellite imagery.
+
+Studies similar to the present study had been previously conducted for the South-Lhonak Lake in Sikkim, which, later on, witnessed a severe GLOF event this year killing at least 40 people. However, effective measures were not taken to prevent this disaster which could have been easily done through certain practical measures taken on time. To avoid such tragedies in the future, a closer examination of the GLOF hazard is necessary, aided by more precise tools and dependable datasets. However, scientific knowledge of the environmental processes at play in the Himalayan region is often hampered by the sparse network of monitoring sites required to generate an ample amount of reliable meteorological and glaciological data. Regularly updated weather data coupled with high-resolution imagery, DEMs and detailed inventories can help in understanding the processes of the Himalayan cryosphere much more closely to prevent any future disasters. In addition to fine satellite datasets, bathymetric surveys are needed to estimate lake volumes more precisely because, even in similar geographic locations, lakes' sizes are not always equally predictable (Cook and Quincey 2015). Although there are numerous empirical equations devised for the estimation of the depth and volume of the lakes, there is always a chance of over or underestimation of these attributes. Hence bathymetric datasets can be extremely helpful in the estimation of the lake volume and the potential maximum discharge from a glacial lake in case of a GLOF event.
+
+Mapping the societal elements of risk associated with the GLOF disaster is a largely ignored domain in the Himalayan region. At the national and state levels, GLOF risk profiles, that map the exposure and vulnerability of human life and property to this specific cryospheric danger must be produced and regularly updated. India lacks an early warning system infrastructure for the GLOF disaster even when the northern and the northeastern states of the country are under serious GLOF danger. Hence effective early warning systems need to be installed at critically hazardous glacial lakes that will help in evacuating the vulnerable population at the time of a future GLOF disaster and evading major risks associated with human life and key infrastructure such as the hydroelectric projects.
+
+# 8. Conclusion
+
+The glacier-lake interactions were studied for the Panikhar glacial lake and its feeding glacier for the past 22 years. We found that the glacial lake has undergone a remarkable expansion during this period. The area of the lake has increased by 0.37 km<sup>2</sup>, from
+
+~0.10 km<sup>2</sup> in 2000 to ~0.47 km<sup>2</sup> in 2022, showing a net increase of 78.7%, at the rate of 0.017 km<sup>2</sup> per year while as its mother glacier has retreated significantly from ~4.63 km<sup>2</sup> in 2000 to ~4.02 km<sup>2</sup> in 2022 showing a total decrease of 13.2% in its total area i.e. 0.03 km<sup>2</sup> per annum. The glacier and glacial lake area plots show an inverse relationship which shows that the glacial lake is rapidly expanding at the expense of its mother glacier. The mass balance estimate of the Panikhar glacier also reveals a mass loss of 5.92 metric tons per square metre and an average thickness loss of 7.04 m over the past 22 years. The glacial lake is impounded by the terminal moraine of the glacier, rendering it hazardous, due to its swift rate of expansion and various other contextual factors. Notably, the presence of zones susceptible to potential mass movement in the surrounding areas also contributes to the heightened risk associated with this lake. The mean surface velocity and thickness of the glacier were estimated to be 3.38 m s<sup>-1</sup> and 30.95 m respectively. The potential outburst flood was modeled for two different GLOF scenarios using the HEC-RAS tool to generate flood depth and velocity profiles along the flow path for a distance of 25 km. The results reveal that the glacial lake is a potential threat to the downstream villages, especially the nearby Panikhar village. A total area of 5.38 km<sup>2</sup> which includes key infrastructure such as settlements, bridges, and roads is vulnerable to a potential GLOF event in the worst-case scenario. The routed hydrographs indicate that the flood wave generated at the time of an overtopping failure will arrive at the nearest settlement in 50 min with a peak velocity of 12.36 m/s. This study suggests regular monitoring of the lake and the installation of an early warning system (EWS) to mitigate and manage the risk from the lake in the future, especially at the time of an extreme weather event.
+

chatbot/rag/data/('GlacialLakeOutburstFloods-AReviewofEventsCausesandImpact', '.pdf')_extraction.md

@@ -0,0 +1,250 @@
+
+
+# 1 Introduction
+
+Glaciers are formed by accumulation of snow at higher altitudes where the average temperature lies in the freezing zone. As the accumulation increases, deeply buried snowflakes get tightly packed and crystallized, resulting in glacial ice formation. As the average global temperature has increased due to heightened anthropogenic activities, the glaciers have receded significantly, resulting in formation of a greater number of moraine-impounded glacier lakes [1-4]. The surface area and the depth of these glacial lakes are growing as a result of glacial retreat, raising the possibility of breaching of moraine dams. Breaching of these dams can further be triggered by extreme precipitation, permafrost deterioration, and dead-ice melting [5]. This will result in instant release of large quantities of sediment-laden glacial water that can ignite vast damage to the downstream parts of the dams. These events, known as Glacial Lake Outburst Floods (GLOFs), can get aggravated if erodible materials are present [6]. Due to its high velocity, continued entrainment of sediment in its flow path, and tendency to flow along existing river channels for very long runout distances, these debris floods/flow events pose significant threat to the lives and property located along its flow channel [6].
+
+In India, the Kedarnath disaster in 2013 [7], and the Chamoli disaster in 2021 [6], caused by cloud outbursts and ice/rock avalanches, respectively, claimed thousands of lives and caused substantial socio-economic losses. Two series of occurrences struck Kedarnath in 2013: the first resulted in landslides, breaching, blockades, flooding, and failures of river banks; the second was outburst flooding in Chorabari Tal Lake in conjunction with erosion of the terrain and concomitant landslides. In the Chamoli Disaster in 2021, a significant amount of rubble blocks fell into the glacial pocket,
+
+Department of Civil Engineering, Indian Institute of Technology Jammu, Jagti 181221, India e-mail: 2021pce1008@iitjammu.ac.in
+
+P. Mahajan ( $\boxtimes$ ) · R. Kumar · R. Bhowmik
+
+forcing the lake to burst. This event had endangered lives and inflicted significant financial losses. Nie et al. [8] inventoried and studied over 60 GLOF events in the Himalayan region alone and reported the estimated losses to be over hundreds of billion dollars. Thus, implementation of adequate mitigation measures and land-use and development plans is vital while developing infrastructure in the mountainous regions.
+
+This paper aims to inventory the reported GLOF events, regarding their causes, the methodologies employed for their study, and predictive dam-breach models reported in the literature. The outcome of this study will aid in consolidating the knowledge developed regarding the effect of climate change, especially on the young and fragile Himalayan region.
+
+# 2 Glacial Lakes and GLOFs
+
+Classification of Glacial lakes into different categories plays a vital role in understanding their origin and evolution. However, there is presently no globally accepted classification system for glacial lakes. Numerous researchers assigned classifications to the glacial lakes in accordance with their knowledge of the region's morphology. In 2018, ICIMOD [9] classified glacial lakes into four groups and seven sub-groups, as shown in Table 1.
+
+An attempt has also been made to summarize the reported cases of moraine-impounded glacial lakes with the objective of understanding the causes of their formation and to study the areas likely to be affected if failure occurs. This summary is given in Table 2. Table 2 further summarizes the volume, depth, and peak discharge of the lakes, probable triggering factors, and methodologies employed to analyze the probability of the GLOF events.
+
+## 2.1 Methodologies
+
+### 2.1.1 Remote Sensing Tools
+
+In conventional times, data was collected from the topographic maps available by reputed agencies to identify the aerial extent of the glacial lake. With the increase in advancement of technology, various remote sensing (RS) techniques such as multispectral scanner (MSS) [4], linear imaging self-scanning (LISS) [12] satellite data, and thematic mapper (TM) [4] tools are used, provided a minimum cover of snow and cloud is present in the sites. Data elevation models (DEMs) such as SRTM [1] and ASTER [5] are also used to recreate the digital elevation models of the sites and conduct in-depth studies of the concerned areas.
+
+In Zanskar Basin [4], the conventional methodology of using topographic maps is utilized in conjunction with remote sensing methodologies like multi-spectral
+
+**Table 1** ICIMOD classification of Glacial lakes (ICIMOD [9])
+
+| Sl.<br>No | Glacial lak                   | e type                                   | Description                                                                                                                                                                                                         |
+|-----------|-------------------------------|------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| 1         | Moraine<br>dammed<br>lake (M) | End moraine dammed lake                  | The lake's water usually touches the walls of the side moraines, but the water is held back by the end moraine (dam), but not necessarily, in contact with the glacier, and may have glacier ice at the lake bottom |
+| 2         |                               | Lateral moraine dammed lake              | Lake formed in the tributary valley, trunk valley, or<br>between the lateral moraine and the valley wall, or at the<br>junction of two moraines. Lake is held back by the<br>outside wall of a lateral moraine      |
+| 3         |                               | Other moraine dammed lake                | Lake dammed by other moraines (including kettle lakes and thermokarst lakes)                                                                                                                                        |
+| 4         | Ice<br>dammed                 | Supra-glacial lake                       | Body of water (ponds or lakes) on the surface of a glacier                                                                                                                                                          |
+| 5         | lake (I)                      | Dammed by<br>tributary valley<br>glacier | Lake dammed by glacier ice with no lateral moraines. Can be at the side of a glacier between the glacier's margin and valley wall                                                                                   |
+| 6         | Bedrock                       | Cirque lake                              | A small pond occupying a cirque                                                                                                                                                                                     |
+| 7         | dammed<br>lake (B)            | Other glacier erosion lake               | Body of water occupying depressions formed by glacial erosion. These are usually located on the mid-slope of hills, but not necessarily in a cirque                                                                 |
+| 8         | Other dam                     | med lakes                                | Lakes formed in a glaciated valley and fed by glacier<br>melt, but the damming material is not directly part of the<br>glacial process, for example, debris flow, alluvial, or<br>landslide-blocked lakes           |
+
+scanner, linear imaging self-scanning, advanced space-borne thermal emission and reflection radiometer to determine the characteristics of the glacial lakes. In Chamlang Lake Tsho [5], images obtained from RS tools and photogrammetric techniques, and field-based methodologies, like sonar-sounding machine, GPS, and fishing line method were utilized to study the area. Figure 1 depicts the various steps involved in processing the data to recreate the profile of the glacial lakes.
+
+### 2.1.2 Numerical Modeling
+
+For the study of glacial lakes, empirical formulations and numerical modeling are often used. One of the widely used numerical tools is Hydrologic Engineering Centre's River Analysis System (HEC-RAS) [6]. This tool simulates GLOF phenomenon from the source to the downstream flow path locations. Using the model, cross-sectional measurements of the main river section at regular intervals can be determined, and the same can be utilized for routing floods. Hydrographs can be generated along different sections of the flow channel which can be further used to estimate the peak discharge intensity [1, 6]. However, there are some limitations of
+
+(continued)
+
+| e         |
+|-----------|
+| ă         |
+| 급         |
+| .2        |
+| gla       |
+| Į.        |
+| 0.        |
+| Sis       |
+| Ly.       |
+| Ωï        |
+| 9         |
+| Ŧ         |
+| ö         |
+| ψ         |
+| ior       |
+| at        |
+| Ę         |
+| sic       |
+| cons      |
+| <u>ت</u>  |
+| Ö         |
+| n T       |
+| <u>s</u>  |
+| ta        |
+| C         |
+| ž         |
+| Ĕ         |
+| Гa        |
+| pa        |
+| ō         |
+| а         |
+| g         |
+| <u>S</u>  |
+| 9         |
+| ğ         |
+| et;       |
+| me        |
+| ŝ         |
+| se        |
+| an        |
+| дc        |
+| Ę         |
+| ō         |
+| də.       |
+| Į.        |
+| 0 /       |
+| ar,       |
+| Ξ         |
+| шш        |
+| Su        |
+|           |
+| e 2       |
+| Table 2   |
+| <u>Ta</u> |
+| F .       |
+
+| Table 2 Summary of reported causes, methodology and parameters taken for consideration for the analysis of glacial lake |                                                                  |                                                                               |                                                                                 |                                                                                                             |                                                                                     |                                      |                                      |                               |                            |
+|-------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------|-------------------------------------------------------------------------------|---------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------|--------------------------------------|--------------------------------------|-------------------------------|----------------------------|
+| Sl. No                                                                                                                  | Events                                                           | Area                                                                          | Triggering<br>factors                                                           | Tools                                                                                                       | Methodology                                                                         | Volume of<br>lake<br>(million<br>m³) | Lake<br>average<br>depth<br>(meters) | Peak<br>discharge<br>(cumecs) | References                 |
+| 1                                                                                                                       | Glacial lake<br>(33° 09′<br>32.56″ N<br>and 76° 59′<br>05.38″ E) | Zanskar Basin,<br>J&K, India                                                  | Collapse of<br>unstable natural<br>dams                                         | Topographical maps in<br>early days. Replaced by<br>remote sensing using<br>MSS, TM, LISS III,<br>ASTER DEM | Monitoring of<br>extent of glacial<br>lake is observed                              | 4.2                                  | 10                                   | 1.7–196                       | Babu<br>Govinda<br>Raj [4] |
+| 2                                                                                                                       | Glacial Lake<br>L2                                               | Dhauliganga<br>River Basin,<br>Pithoragarh<br>Distt,<br>Uttarakhand           | Lake area<br>expansion,<br>dead ice<br>melting,<br>seepage,<br>seismic activity | HEC-RAS (hydrologic<br>engineering centre’s<br>river analysis system)                                       | Analysis of the<br>glacial lake using<br>numerical<br>modeling                      | 2                                    | 30                                   | 4272                          | Jha and<br>Khare [6]       |
+| 3                                                                                                                       | Lumi Chimi<br>Lake                                               | Sun-Koshi<br>Basin,<br>Transboundary<br>between Tibet<br>(China) and<br>Nepal | Retreat of<br>glaciers                                                          | Dambreak and<br>hydrodynamic modeling                                                                       | Risk assessment<br>of the event<br>before it occurs                                 | 307.2                                | 80                                   | 5040–8380                     | Shrestha<br>et al. [1]     |
+| 4                                                                                                                       | Chamlang<br>South Tsho                                           | Hongu valley,<br>eastern Nepal<br>Himalaya                                    | Lake area<br>expansion, ice<br>avalanches,<br>rock falls,<br>landslides         | Observational data,<br>photographs,<br>ground-based surveys,<br>field and remote sensing<br>data            | GLOF hazards,<br>bathymetry,<br>surrounding<br>material decide<br>the vulnerability | 34.9–35.6                            | 40.2                                 | Not available                 | Lamsal<br>et al. [5]       |
+
+References Gurung et al. [11] Aggarwal et al. [12] Liu et al. [3] et al. [10] Goodsell Not available discharge (cumecs) 2611.136 Peak 1270 1198 average (meters) depth Lake 6.29 14.1 25 37 Volume of available (million 42.93 0.37 0.97  $m^3$ Not position due to ice 51 houses swept away, 18 bridges channel shifted to 0.976 Nu. Million Five people died, washed away and outburst flooding using MIKE-11 Dam Break agricultural land more than CNY Analysis of the dangerous lake economic loss estimated loss Methodology Four bridges washed away, supra-glacial collapse and affected and 100 million Sub-glacial potential Model from Landsat and Indian Remote Sensing, MIKE-11 Dam Break Model observation and remote Satellite images, field Satellite images, field investigation and past studies Visual representation Satellite data, images data was utilized sensing data Tools composition of dam expansion, ice avalanches blackhole due to collapse of ice Formation of supra-glacial observations Earthquake, draining of temperature and rainfall, Triggering factors Lake area rainfall, Climatic material puod Westland Tai Poutini National Park North-western Bhutan Bomê County, South-eastern Tibet (China) New Zealand, Gasa District, Sikkim Area 
+ Table 2
+ (continued)
+ Guangxieco Lake Teesta basin Lemthang Events Joseph Glacier Franz Tsho SI. No S 9 \_  $\infty$ 
+
+(continued)
+
+References
+
+Khan et al. [13]
+
+Schneider et al. [2]
+
+9000-40,000 Peak discharge (cumecs) 4500 Volume of lakeLakelakeaverage(million depthdepthm³)(meters) Not available 21 0.2 - 0.4 $\infty$ Santa Valley highway affected, people are prone to flood, Karakoram Highway Damaged several bridges swept away and agricultural land affected More than 2000 Methodology Images from Landsat and Sentinel, Global Digital Elevation Model (GDEM), ArcMap 10.5 Remote sensing images, digital elevation data, RAMMS and IBER model—numerical and physically based avalanche and debris Tools flow Blockage of melt water originating from Mochuwar Triggering factors Rock-ice avalanche glacier Gilgit-Baltistan, Pakistan Carhuaz, Peru Area Table 2 (continued) Shishper Glacier lake Lake 513 Sl. Events
+No 10 6
+
+Image /page/7/Figure/2 description: A flowchart illustrating a three-step process, with each step represented by a blue, right-pointing chevron arrow. The first step is labeled "Image Processing to Delineate Glacial Lakes." The second step is "Stereo-data Processing for 3D Topographic Maps." The third and final step is "Lake Bathymetry." The text within each chevron is white.
+
+Fig. 1 Steps involved for the interpretation for the data
+
+this model, one of which is its inability to consider debris and sediment-laden water in the simulations. In the study on the Dhauliganga Basin [6], HEC-RAS was used to create the glacial lake representation. HEC-RAS was also used to generate the hydrological model for the Mandakini River [7].
+
+## 2.2 Empirical Equations for Discharge Estimation and Stability of Moraine Dams
+
+The simplest and fastest methods for determining the various parameters required for vulnerability assessment of moraine dams are empirical relationships. Empirical relationships, on the other hand, do not incorporate fundamental rules of fluid mechanics and hydraulics as they are derived through statistical study of historical events. Various researchers have proposed empirical equations to determine peak discharge values based on estimated average breach width and time of failure in moraine dams [3].
+
+Peak discharge is an important consideration for the assessment because it helps to visualize the size of the flood downstream. The findings of the multiple regression analysis showed that compared to other factors like embankment width and embankment length  $(E_1)$ , empirical relationships with parameters like volume of water  $(V_w)$  and height of water  $(H_w)$  gave more precise forecasts of maximum flow discharge  $(Q_{\text{max}})$  [14]. The models that predict peak outflow using only  $H_w$  and  $V_w$  do not considerably benefit from the addition of  $E_1$  and  $Q_{\text{max}}$  to the relationship, making  $H_w$  and  $V_w$  more significant features to define peak discharge compared to other parameters. Table 3 shows the summary of the empirical relationships developed by various researchers.
+
+# 3 Estimation of Peak Discharge of Reported Cases
+
+Using the empirical relationships mentioned in Table 3, the average breach width B, time of failure  $t_f$ , and peak discharge values for 15 different cases of moraine dam are estimated and tabulated in Table 4. The results showed that when the glacial lake's volumetric capacity increases, the peak discharge of the lake also consistently rises.
+
+**Table 3** Summary of empirical relationships
+
+| <b>Table 3</b> Summary of empirical relationships |                             |                                                                                                   |                                                      |
+|---------------------------------------------------|-----------------------------|---------------------------------------------------------------------------------------------------|------------------------------------------------------|
+| Sl. No                                            | Parameter                   | Lake volume (m3)**                                                                                | References                                           |
+| 1                                                 | Lake volume<br>(m3)         | $V = 0.104A^{1.42}$                                                                               | Huggel et al. [15]                                   |
+| 2                                                 |                             | $V = 0.035A^{1.5}$                                                                                |                                                      |
+| 3                                                 |                             | $V = 0.104A^{1.42}$                                                                               | Evans [16]                                           |
+| 4                                                 |                             | $V = A * MD$                                                                                      | Wang et al. [17]                                     |
+| 5                                                 | Lake depth (m)              | $D = 0.104A^{0.42}$                                                                               | Huggel et al. [15]                                   |
+| 6                                                 |                             | $MD = 4A * 10^{-5} + 5.0564$                                                                      | Wang et al. [17]                                     |
+| 7                                                 |                             | $D = 0.087A^{0.434}$                                                                              | Patel et al. [18]                                    |
+| 8                                                 | Average breach<br>width (m) | $B = 0.1803V_{W}^{0.32}h_{b}^{0.19}$                                                              | Froehlich [19]                                       |
+| 9                                                 | Time of failure<br>(h)      | $t_{f} = 0.00254V_{W}^{0.53}h_{b}^{-0.90}$                                                        |                                                      |
+| 10                                                | Peak discharge<br>(m3s-1)   | $Q_{max} = 75V^{0.67}$                                                                            | Clague and Mathews [20]                              |
+| 11                                                |                             | $Q_{max} = 0.72V^{0.53}$                                                                          | Evans [16]                                           |
+| 12                                                |                             | $Q_{max} = 0.0048V^{0.896}$                                                                       | Popov [21]                                           |
+| 13                                                |                             | $Q_{max} = 0.00077V^{1.017}$                                                                      | Huggel et al. [15]                                   |
+| 14                                                |                             | $Q_{max} =$<br>$1.165(\frac{L}{B})^{\frac{1}{10}}(\frac{B}{b})^{\frac{1}{3}}b(H-h)^{\frac{3}{2}}$ | Lue et al. [22]                                      |
+| 15                                                |                             | $Q_{max} = 0.00013(PE)^{0.60}$                                                                    | Huggel et al. [15]                                   |
+| 16                                                |                             | $Q_{max} = \frac{2V}{T_{w}}$                                                                      | Huggel et al. [15]; Popov [21], and<br>Haeberli [23] |
+
+<sup>\*\*</sup>Where V is the volume of glacier lake; A is an area of glacier lake in  $m^2$ ; MD is the average depth of glacial lake in m;  $V_w =$  reservoir volume in  $m^3$ ;  $h_b$  is the height of water above breach invert level; L is the lake's length (m), B is the breach's maximum width (m), b is its avg width (m), H is the lake's maximum depth (m), and h is height of remaining dam (m)
+
+However, it has been observed that the average breach width, B, and time of failure,  $t_f$ , change in proportion to the mean bathymetric depth, as shown in Table 5.
+
+For the Safed glacial lake located in Goriganga Basin, Uttrakhand, India [34], Table 5 shows the parametric effect of the breach depth and volume released on the average breach width and time of the failure. When the breach depth is reduced to half, as in Scenario 2, it is shown that the breach breadth decreases by 30%. The average breach width further decreases by 30% if the drop continues to 15 m, as shown in Scenario 2. When the time of failure is taken into account, it can be seen that the moraine dam's failure time increases by 29 and 30% in comparison to scenarios 1 and 2, respectively, based on the equations proposed by Froehlich, 1995 [12]. This shows that the time of failure is directly connected to volume released but indirectly related to breach width, whereas the breach width is directly related to volume released but indirectly related to breach depth. Due to the inverse relationship between average breach width and time of failure, the breach depth has a significantly
+
+(continued)
+
+ Table 4
+ Summary of physical information about glacial lakes including average breach width and time of failure
+
+| <b>Table 4</b> Summary of physical information about glacial lakes including average breach width and time of failure |                  |                  |                    |                |                |                                                                                    |                                                                                |
+|-----------------------------------------------------------------------------------------------------------------------|------------------|------------------|--------------------|----------------|----------------|------------------------------------------------------------------------------------|--------------------------------------------------------------------------------|
+| Sl. No                                                                                                                | Lake             | Area (m2, × 106) | Volume (m3, × 106) | Mean depth (m) | Type of lake   | Average breach width<br><i>B</i> =<br>0.1803*1.4* <i>V</i> w0.32<br><i>h</i> b0.19 | Time of failure<br><i>t</i> f =<br>0.00254 <i>V</i> w0.53<br><i>h</i> b(−0.90) |
+| 1                                                                                                                     | Ice Cave Lake    | 0.0035           | 0.01               | 2.9            | Ice-dammed     | 5.9                                                                                | 0.1                                                                            |
+| 2                                                                                                                     | Gruben Lake 5    | 0.01             | 0.05               | 5              | Thermokarst    | 10.9                                                                               | 0.2                                                                            |
+| 3                                                                                                                     | Crusoe-Baby Lake | 0.017            | 0.08               | 4.7            | Ice-dammed     | 12.6                                                                               | 0.3                                                                            |
+| 4                                                                                                                     | Gruben Lake 3    | 0.021            | 0.15               | 7.1            | Ice-dammed     | 16.6                                                                               | 0.2                                                                            |
+| 5                                                                                                                     | Gruben Lake 1    | 0.023            | 0.24               | 10.4           | Moraine-dammed | 20.8                                                                               | 0.2                                                                            |
+| 6                                                                                                                     | MT' Lake         | 0.0416           | 0.5                | 12             | Ice-dammed     | 27                                                                                 | 0.3                                                                            |
+| 7                                                                                                                     | Lac d'Arsine     | 0.059            | 0.8                | 13.6           | Moraine-dammed | 32.1                                                                               | 0.3                                                                            |
+| 8                                                                                                                     | Nostetuko lake   | 0.2622           | 7.5                | 28.6           | Moraine-dammed | 75.7                                                                               | 0.5                                                                            |
+| 9                                                                                                                     | Between Lake     | 0.4              | 7.5                | 18.8           | Ice-dammed     | 69.9                                                                               | 0.8                                                                            |
+| 10                                                                                                                    | Abmachimai Co    | 0.565            | 19.4               | 34.3           | Moraine-dammed | 106.2                                                                              | 0.8                                                                            |
+| 11                                                                                                                    | Gjanupsvatn      | 0.6              | 20                 | 33.3           | Ice-dammed     | 106.6                                                                              | 0.8                                                                            |
+| 12                                                                                                                    | Quongzonk Co     | 0.753            | 21                 | 27.9           | Moraine-dammed | 104.7                                                                              | 1                                                                              |
+| 13                                                                                                                    | Laguna Parón     | 1.6              | 75                 | 46.9           | Moraine-dammed | 173.6                                                                              | 1.2                                                                            |
+| 14                                                                                                                    | Summit Lake      | 5                | 250                | 50             | Ice-dammed     | 258.4                                                                              | 2.1                                                                            |
+| 15                                                                                                                    | Phantom Lake     | 6                | 500                | 83.3           | Ice-dammed     | 355.4                                                                              | 1.9                                                                            |
+
+| 5 |
+|---|
+|---|
+
+| Discharge                  |                          |                                    |                                   |                          |                         | Reference                     |                               |
+|----------------------------|--------------------------|------------------------------------|-----------------------------------|--------------------------|-------------------------|-------------------------------|-------------------------------|
+| Glacier, tunnel events     |                          | Glacier, non-tunnel events         |                                   | Moraines                 | Earth- and<br>rock-fill |                               |                               |
+| Clague and Mathews<br>[20] | Walder and<br>Costa [24] | Haeberli [23]                      | Walder and Costa<br>[24]          | Popov [21]               | Evans [16]              |                               |                               |
+| $Qmax = 75(V/106)0.67$     | $Qmax = 46(V/106)0.66$   | $Qmax = 2 V/tw$ where $tw =1000 s$ | $Qmax = 1100(V/106)0.44$          | $Qmax = 0.0048 V0.896$   | $Qmax = 0.72 V0.53$     |                               |                               |
+| 3.43                       | 2.2                      | 20                                 | 145.01                            | 18.42                    | 94.91                   | Maag and<br>Janischowsky [25] |                               |
+| 10.08                      | 6.37                     | 100                                | 294.4                             | 77.9                     | 222.73                  | Clague and Mathews<br>[20]    |                               |
+| 13.81                      | 8.69                     | 160                                | 362.04                            | 118.69                   | 285.74                  | Maag and<br>Janischowsky [25] |                               |
+| 21.04                      | 13.15                    | 300                                | 477.39                            | 208.46                   | 398.71                  | Kääb and Haeberli<br>[26]     |                               |
+| 28.83                      | 17.93                    | 480                                | 587.06                            | 317.62                   | 511.5                   | Kääb and Haeberli<br>[26]     |                               |
+| 47.14                      | 29.11                    | 1000                               | 810.85                            | 613.08                   | 754.72                  | Blown and Church<br>[27]      |                               |
+| 64.58                      | 39.7                     | 1600                               | 997.13                            | 934.14                   | 968.21                  | Vallon [28]                   |                               |
+| 289.3                      | 173.9                    | 15,000                             | 2669.44                           | 6938.99                  | 3170.41                 | Evans and Clague<br>[29]      |                               |
+| 289.3                      | 173.9                    | 15,000                             | 2669.44                           | 6938.99                  | 3170.41                 | Maag and<br>Janischowsky [25] |                               |
+| 546.88                     | 325.62                   | 38,800                             | 4055.32                           | 16,259.66                | 5246.47                 | Meon and Schwarz<br>[30]      |                               |
+| Table 4 (continued)        |                          |                                    |                                   |                          |                         |                               |                               |
+| Sl. No                     | Discharge                |                                    |                                   |                          |                         |                               | Reference                     |
+|                            | Glacier, tunnel events   |                                    | Glacier, non-tunnel events        |                          | Moraines                | Earth- and rock-fill          |                               |
+|                            | Clague and Mathews [20]  | Walder and Costa [24]              | Haeberli [23]                     | Walder and Costa [24]    | Popov [21]              | Evans [16]                    |                               |
+|                            | $Qmax = 75(V/106)0.67$   | $Qmax = 46(V/106)0.66$             | $Qmax = 2 V/tw where tw = 1000 s$ | $Qmax = 1100(V/106)0.44$ | $Qmax = 0.0048 V0.896$  | $Qmax = 0.72 V0.53$           |                               |
+| 11                         | 558.15                   | 332.23                             | 40,000                            | 4110.04                  | 16,709.53               | 5331.86                       | Costa and Schuster [31]       |
+| 12                         | 576.7                    | 343.1                              | 42,000                            | 4199.23                  | 17,456.2                | 5471.53                       | Meon and Schwarz [30]         |
+| 13                         | 1353.17                  | 794.87                             | 150,000                           | 7352.24                  | 54,613.01               | 10,742.74                     | Lliboutry et al. [32]         |
+| 14                         | 3031.66                  | 1759.54                            | 500,000                           | 12,487.81                | 160,618.47              | 20,334.84                     | Mathews and Clague [33]       |
+| 15                         | 4823.6                   | 2780.21                            | 1,000,000                         | 16,941.02                | 298,894.8               | 29,362.07                     | Maag and<br>Janischowsky [25] |
+
+| Sl.<br>No | GLOF scenarios | Breach depth<br>(hb) | Volume released<br>(106 × m3) | Average breach width<br>B = 0.1803*1.4*Vw0.32<br>hb0.19 | Time of failure<br>tf = 0.00254Vw0.53<br>hb(-0.90) |
+|-----------|----------------|----------------------|-------------------------------|---------------------------------------------------------|----------------------------------------------------|
+| 1         | Scenario 1     | 60                   | 4.34                          | 73.10                                                   | 0.21                                               |
+| 2         | Scenario 2     | 30                   | 2.17                          | 51.30                                                   | 0.27                                               |
+| 3         | Scenario 3     | 15                   | 1.08                          | 36.00                                                   | 0.35                                               |
+
+Table 5 Parametric effect of the breach depth and volume released
+
+more noticeable effect even when the effect of volume released is assumed to be the same in all circumstances.
+
+# 4 Concluding Remarks
+
+The high-altitude glaciers have suffered greatly as a result of extensive anthropogenic activities that led to climate change. In areas where glacial lakes are growing in lateral extent and bathymetry, increasing possibility of higher peak discharges and GLOFs are risking the lives of the local inhabitants, local flora, fauna, and other infrastructure.
+
+In this paper, reported studies on glacial lakes are examined in order to ascertain their formational behavior, its after-effects, and potential corrective actions. The approaches that are utilized globally to anticipate the size of these glaciers include remote sensing and numerical simulation utilizing HEC-RAS. These lakes have been categorized by ICIMOD based on the failure behavior that is anticipated to occur in the future. In order to predict the peak discharge, breach width, and time of failure of various structures that result in the abrupt discharge of enormous amounts of water in the downstream region, numerous empirical equations have been developed and reviewed in this paper. The established empirical relationships are then used to estimate the average breach width, B, time of failure,  $t_f$ , and peak discharge for 15 identified cases of glacial lakes.
+
+The peak discharge values aid in identifying the glacial lake's susceptibility in the downstream sections of the impacted area. The results show that as the depth of the breach grows, so does the likelihood of failure. If the breach depth is reduced by 50%, the time of failure increases by 28–29%, and the average breach width drops by about 30%.
+
+Reviewing all the various glacial lakes enables us to spot certain significant problems that, if resolved, can reduce the impact of GLOF to some level: (a) Quantification of the climatic conditions in space—time and further determination of projections for the future, (b) establishment of a database that compiles data on glacier surface areas and their pace of retreat across time, (c) development of an algorithm that aids in projecting the size of lakes based on present-day spatial features, (d) development of hydrological models that account for potential changes in the regional availability of water.
+

chatbot/rag/data/('Glacier Lake 513 Peru_ Lessons for early warning service development', '.pdf')_extraction.md

@@ -0,0 +1,128 @@
+
+
+## Author(s):
+
+By Christian Huggel, Alejo Cochachin, Fabian Drenkhan, Javier Fluixá-Sanmartín, Holger Frey, Javier García Hernández, Christine Jurt, Randy Muñoz, Karen Price, Luis Vicuña
+
+Image /page/0/Picture/16 description: A wide-angle photograph of a majestic mountain landscape. In the background, a large, snow-covered mountain peak with glaciers reaches up into a blue sky with white clouds. Below the mountain, a vibrant turquoise alpine lake sits in a rocky basin, surrounded by steep, rugged cliffs. The foreground is a rocky slope with numerous large boulders and some sparse vegetation, offering a high vantage point looking down towards the lake.
+
+Image /page/0/Picture/17 description: An icon of an image file, depicting a green landscape, is shown to the left of the text 'Site location' in a dark sans-serif font on a white background.
+
+Site location: a) Peru, b) Cordillera Blanca and c) the Hualcán-Carhuaz area (in dashed line).
+
+Glacier shrinkage – accelerated over the last decades due to climate change – is exposing large areas in mountain regions worldwide. But an even dire consequence of the melting ice is the
+
+forming of more glacier lakes, which are increasing in size. Glacial lakes have caused some of the world's most devastating floods, for example, in the Andes, Himalayas and Alps, where thousands of human lives were lost and huge infrastructure damages reported (Carrivick and Tweed, 2013; Bajracharya et al., 2007; Carey 2005). Climate change is rapidly reshaping living conditions in high mountains – altering flood patterns and creating new flood hazards – leaving populations at imminent risk in several regions (Cook et al., 2016; Emmer et al., 2015; Frey et al., 2016; Drenkhan e al., 2019).
+
+Climatic, glaciological and hydrological information and services can play an essential role for early detection of potential hazards and risks, and for effectively reducing risks. However, infrastructure for climate and associated services are poorly developed in many high mountain areas and need to be substantially strengthened. This report on the design, implementation, operation and circumstances around the setting up of an early warning system for glacier lake outburst floods (GLOFs) in the Peruvian Andes highlights the challenges related to accessing and installing equipment in many high mountains regions.
+
+## Laguna 513 disaster
+
+Image /page/1/Picture/5 description: An overview map displaying a mountainous region with a red line tracing the path of an event. The map features a large aerial view and a smaller, inset close-up view. The inset, outlined in a blue box, shows a snow-covered mountain labeled "Mt. Hualcán" and a turquoise lake labeled "Laguna 513". A red oval on the mountain peak is marked as the "Detachment area (11 Apr 2010)". The red line on the main map originates from this detachment area, flows down the mountain, and follows a valley path towards a town labeled "Carhuaz" and the "Santa River". A scale bar at the bottom right indicates a distance of 2 kilometers, and a north arrow is present for orientation.
+
+Overview map of Lake 513 and Carhuaz, indicating the 2010 ice avalanche source zone producing the GLOF that reached down to Carhuaz.
+
+The Cordillera Blanca in the tropical Andes of Peru is a glaciated mountain range with a long history of disastrous GLOF events (Carey 2005; 2010). GLOF risks result from considerable physical hazard levels and the high levels of vulnerability and exposure of downstream populations (Frey et al., 2018). The Laguna 513 glacier lake (9°12'45"S, 77°33'00"W) is located at 4 428 metres (m) above sea level at the foot of Mount Hualcán (6 104 m) in the Santa River basin (Figs. 1, 2). The lake, which formed in the late 1960s as a result of glacier shrinkage, was declared highly dangerous in 1988 and subjected to exhaustive security works to artificially lower its level by some 20 m until 1994. This did not reduce GLOFs risk to zero, though the probability of occurrence and magnitude of GLOFs was substantially lowered. In 2004, authorities and specialists produced a report indicating that the lake could be considered safe due to the infrastructure in place (INDECI 2004; Muñoz et al. 2016).
+
+However, Laguna 513 was heavily impacted when a 450 000 cubic metres (m3) rock-ice avalanche detached from the southwest slope of Mt. Hualcán (Carey et al., 2012) (Fig. 2) on 11 April 2010, at about 8 a.m. local time. The avalanche caused a tsunami-like push-wave on the lake, resulting in a dam spillover despite the over 20 m freeboard. Traces of the wave indicate an overtopping of the dam by about 5 m – corresponding to a wave height of about 24 to 25 m – with a peak discharge of rate of several tens of thousands m3 per second (Schneider et al., 2014). The resulting GLOF damaged several bridges, water service infrastructure along its trajectory and eventually reached the debris fan of the city of Carhuaz (about 20 000 inhabitants), where the coarse material of the GLOF was deposited. A total of 0.7 km2 of agricultural land was buried and the Santa Valley highway was affected, but no lives were lost.
+
+Local and national authorities, as well as Peruvian and international experts, met in the weeks following the disaster to discuss ways to better protect people and their assets in future incidents. As a result, plans for a GLOF Early Warning System (EWS) were initiated in 2011 and implemented within three years. The GLOF EWS, the first in the Andean region, was established in the framework of the Glacier Project (<a href="www.proyectoglaciares.pe">www.proyectoglaciares.pe</a>) with financial support from the Swiss Agency for Development and Cooperation (SDC). CARE Peru and the University of Zurich jointly implemented the EWS in close collaboration with the municipality of Carhuaz and the National Water Authority of Peru (ANA) and its Office for Glacier and Lake Evaluations (former Glaciology and Water Resources Unit - UGRH) in Huaraz.
+
+## **Development and implementation of a GLOF Early Warning System**
+
+The design, organization and operation of the GLOF was structured to adhere to internationally recognized EWS components (cf. Fluixá-Sanmartín et al., 2018):
+
+- risk knowledge
+- monitoring and warning
+- · dissemination and communication
+- response capability.
+
+### Risk knowledge
+
+Understanding the risks encountered at a particular location is fundamental for the design of an EWS. Risks can be assessed using established methods that analyze physical hazards by means of critical indicators and thresholds (in this case related to the different components involved in the GLOF process), exposure of people and assets (e.g. infrastructure) and the vulnerability (e.g. social, economic) of the elements at risk. Comprehensive risk assessments for GLOFs are rare (Allen et al., 2016) and complex because GLOFs are typically the results of a cascade of triggering and propagating mass flow processes (Schneider et al., 2014; Westoby et al., 2014).
+
+The 2010 GLOF served as a reference to analyze the physical hazards by simulating the process cascade with an iterative approach of coupled, physically-based numerical mass movement and hydrodynamic models (RAMMS and IBER). This model chain was then used to simulate three potential future scenarios of different magnitudes (small, medium, large) and corresponding probabilities of occurrence (high, medium, low, respectively). This hazard assessment procedure followed international standards and was in line with the recently established guidelines of the International Commission on Glacier and Permafrost Hazards in Mountains (www.gaphaz.org), a joint commission of the International Association of Cryopsheric Sciences (IACS) and the International Permafrost Association (IPA). The modelling, together with field surveys, resulted in a GLOF hazard map for local communities and city of Carhuaz (Schneider et al., 2014) (Fig. 3). Exposure and vulnerability were assessed using publicly available data sources (such as census data) and additional surveys in the catchments.
+
+However, risks are perceived in very different ways by different actors. It is a challenge to understand and take these differences in perception into consideration but this is essential for wide user acceptance and the long term success of risk reduction measures. Repeated workshops were conducted in the different communities of the catchment to learn about the risk perceptions and priorities of local leaders and people. At a later stage, ethnographic studies, which included longer (several months) research visits with the local communities, were also conducted in the catchment. These were important to gain more in-depth understanding of how local people perceive their natural environment, and the relationships they maintain with mountains, glaciers and lakes that surround them. This led to a deeper understanding of how they perceived the diverse types of risks and how they understood the connections between those risks and those involved in the GLOF EWS project. Thereby cultural and political aspects were highlighted and people's concerns about water came to the fore, that is water availability, mostly in terms of access to water (e.g. water rights, allocation).
+
+### Monitoring and warning
+
+Image /page/3/Picture/5 description: A generic image file icon, which depicts a landscape with a green hill and a cloud, is shown next to the text "Glacier Lake" in a dark gray sans-serif font on a white background.
+
+Monitoring and warning are central elements of an EWS. Monitoring instruments and technical measuring must be set up to detect hazards so that timely warnings can be issued. The challenge is to identify the environmental processes and variables that are critical to the early detection of an extreme event (such as a GLOF) – and that are measurable by sensors. The insights and improved understanding of the processes gained from the reconstruction of the 2010 outburst and the modelling of potential future scenarios (risk knowledge element) served to identify where and what to measure and monitor. Knowledge about GLOF travel times – from triggering to reaching population centres – are, for instance, critical for the design of an EWS and for later visualization and planning with local authorities and communities (see "Dissemination and communication").
+
+Reference projects for GLOF EWS were rare at the time of the design for Laguna 513 (2011/2012), and completely non-existent for the Latin American region. The "art" in such a design is in taking account of all possible flood trigger processes while measuring the point that still allows for timely warning. The different types of flood trigger processes – ice avalanches, moraine instabilities, rock slope failures – strongly depend on local conditions. It is critical to adequately understand the physical environment and interplay of processes that can result in different GLOF scenarios.
+
+The harsh, extreme physical environment in which glacier lakes (as origin of GLOFs) form are often the biggest challenge. At high altitudes, such as that of Laguna 513, there are large daily temperature fluctuations, long periods of cloudiness, heavy precipitation and high solar radiation as well as a steep topography in what is a remote environment. All of these factors need to be considered in the design and implementation. The GLOF EWS adventurer/scientists had to make
+
+provisions for reduced energy, for complications with data transmission from the sensors and the limited access for sensor installation and frequent following-up and maintenance.
+
+Another crucial element for monitoring and warning, in particular for an EWS in extreme environments such as in this case, is the redundancy in the system. Even in a well-calibrated and tested EWS, a sensor or data transmission failure is likely to occur at some point, sufficient redundancy is indispensable to avoid that sensor failure results in EWS failure as a whole.
+
+An addition problem is funding the long term maintenance of the EWS. Small municipalities with limited budget have other priorities like investing in health and education services.
+
+With the knowledge and information gathered, the local and international team worked together, each contributing their expertise, to design an EWS that would overcome the challenges of Laguna 513. The design comprised two stations – a main station at the Laguna 513 dam and a station in the Pampa Shonquil, which would include meteorological measuring instruments – a data centre in the municipality of Carhuaz, a warning station in the community of Pariacaca, and a repeater station for transferring the signal from the lake to the data centre (Fig. 3).
+
+The stations were equipped with the following instruments:
+
+- Data centre Carhuaz (2 640 m a.s.l.): receiving antenna, screen with real-time data access, server for data storage, infrastructure for launching alarms.
+- Repeater station (3 189 m a.s.l.): receiving and sending antenna.
+- Station Laguna 513 (4 491 m a.s.l.): 2 cameras taking photos every 5 seconds during daylight times, one looking at the face of Mt. Hualcán, one observing the dam. 4 geophones located close to the station, continuously measuring and sending data in 5 second intervals, in order to detect potential mass movements (e.g. ice avalanche) impacting the lake. Receiving and sending antenna and data logger.
+- Station Pampa Shonquil (3 600 m a.s.l.): river discharge station (using pressure sensor), meteorological station with sensors for measuring air temperature and humidity, precipitation, wind speed, and solar radiation. Sending antenna and data logger.
+- Information receiving and warning station at Pariacaca (3 138 m a.s.l.): the monitoring system informs locals about events at Laguna 513 and sirens activated from the Data Centre Carhuaz to facilitate evacuation.
+
+All stations were equipped with solar panels and batteries for energy generation and storage, however energy availability remained a limiting factor, in particular at the glacier lake station as the peaks of the Cordillera Blanca experience a high frequency of cloud coverage. Each station had a mast where most of the instruments were fixed, a concrete lockable box for the electronic equipment, and a protective fence. Emergency power aggregates were available in the municipal building to prevent data losses and interrupted access during blackouts.
+
+The geophones (devices recording ground movements and converting them into voltage) were the principle instruments used for registering a potential GLOF trigger. The back-up cameras could be used to get an overview of the current situation and, particularly during the test phase of the system, for relating geophone measurements to the magnitude of (avalanche) events. The pressure sensor in the riverbed at the Pampa Shonquil station added redundancy to the system and, if calibration measurements were taken, could be used to continuously record runoff. Later, it was planned to install wire sensors in the river channel bed below Laguna 513 that would detect unusually high and dangerous river flow discharge, which could be applied in debris flow warning systems.
+
+A permanently manned hut with wardens next to the station at Pampa Shonquil was an important element of the EWS, especially for consideration of redundancy. Its main purpose was to control the freshwater intake for the municipality of Carhuaz but the location gave the wardens a perfect view towards Laguna 513, they could radio warnings to the authorities in case of an event (as was the case in the 2010 event).
+
+For security, all data is first stored in the data logger at each respective station, then transmitted at 5-second intervals to the data centre server, which has a cloud back-up. All data are directly transferred to a website that permits real-time remote access. In the data centre itself – a separate office in the municipality of Carhuaz – a screen continuously displays the data from this webpage.
+
+Warning protocols represent essential elements of an EWS. The protocol documents and defines warning procedures, typically differentiating a number of warning levels and associated actions as well as the responsible institutions, organizations or committees and people. Local, regional and national laws, rules and guidelines had to be taken into consideration in the Laguna 513 warning protocol. The members of the Local Emergency Operation Centre, civil defense, selected government officials and the mayor, whose authority it is to launch the evacuation alarm, had to be involved. Accordingly, the protocol was accompanied by a list of responsible persons and their phone numbers. It defined three warning levels – yellow, orange and red – plus a normal green baseline level and how these warning levels are reached and what type of actions need to be taken. To this purpose, thresholds of physical variables and processes had to be determined based on sensor measurements. Definition of these thresholds is critical and involves an extended period of calibration and testing, typically of many months, especially if no prior measurements are available as in the case of Laguna 513.
+
+### Dissemination and communication
+
+If measurements on a geophone surpasses a defined threshold, a short message to immediately check EWS data and information is automatically sent to the mobile phones of all the responsible personnel identified in the warning protocol. The subsequent steps to be taken are based on the action plan and on the available data. Alarms cannot be automatically launched by the EWS because, under Peruvian law, only the mayor can authorize an evacuation.
+
+Carhuaz's alarm module has two long-range acoustic sirens and the capacity to send predefined text messages to community and district leaders and stakeholders such as school principals, hospital chiefs, the police and firefighters. Communities upstream of Carhuaz receive warnings and associated information through the Local Emergency Operation Centre and the central authorities of Carhuaz. Pariacaca, which is on the flood pathway, has a warning station with sirens. Furthermore, the EWS protocols were adapted to fit with the Peruvian protocols for risk assessment, allowing communication with the National Emergency Operations Centre at Lima to ask for help (Muñoz et al. 2016).
+
+### Response capability
+
+The ability of people at risk to appropriately respond to the levels of warning issued is possibly the most critical element of an EWS – it is also the most susceptible to failure as the last element of the EWS chain. Failures or errors along the monitoring and warning chain have to be accommodated such that this last element is not adversely affected or threatened.
+
+For the Laguna 513 EWS, information sessions were held with the population at risk. During these, the concept and functionality of the EWS, as well as its potentials and limitations, and clear instructions on actions that have to be taken in case of an alarm were explained and discussed. The instructions include the directive to immediately escape the endangered zones, and a clear indication of the evacuation routes and safety zones. A detailed map with all evacuation routes
+
+was prepared by the civil defense of Carhuaz on the basis of the hazard map developed in the risk knowledge phase of the EWS GLOF design. Emergency simulations are scheduled several times a year for the entire country as Peru's seismic risk is very high. Such simulations, some of them taking place at nighttime, have been used to expose both the population as well as the responsible authorities to a test evacuation und near-realistic conditions, and to familiarize them with the EWS of Laguna 513.
+
+## **Operational aspects and lessons learned**
+
+In 2010, when discussions and activities related to the GLOF EWS started, Carhuaz was the main local actor and the centre for data and information reception. However, the technical, operational and social dimensions of the EWS were beyond the capacity of such a small city. Long-standing national and international expertise – necessitating regular presence on site and permanent joint capacity building and exchange with local people and authorities – was indispensable to address the challenges. In July 2015, full responsibility of the EWS was handed over to the local authorities in a ceremony attended by representatives from the local, provincial and national governments of Peru, from the Swiss government, and from local communities and schools as well as national and international experts. By that time, the EWS had made headlines in the Peruvian, Swiss and international media.
+
+In 2016, much of the central tropical Andes, including the Cordillera Blanca region, was affected by a strong drought. In normal years, after a long dry Austral winter season, farmers count on the start of the rainfall season in October. In 2016, no rainfall was recorded in October and November.
+
+After first requests to remove the EWS by some local inhabitants (cf. Fraser, 2017), farmers got desperate and rumors started to spread that the rain gauges and antennas of the EWS at Laguna 513 were responsible for the lack of rainfall. In a rather dramatic turn of events – driven by community-level power policy factors and weak communications from authorities on the extraordinary meteorological events – a large number of locals gathered at Laguna 513 on 24 November and decided to dismantle the EWS station at the lake. Reactions at the local, national and international levels were vigorous. There was animated disorderly mix of supporters of the EWS on social media. Others expressed incomprehension, disgust, shame and critiqued the voluntarily exposure of lives at risk by this destructive action.
+
+The destruction of the station affected the monitoring and warning components of the EWS from a technical and operational point of view. But service could be maintained thanks to the wardens located at the intermediate site (Pampa Shonquil). Response capabilities and institutional mechanisms were not affected. However, it was crucial to understand the root causes of this action.
+
+The results of intensive research into the incident, which centred on social sciences, are summarized below. The lessons learned are relevant for the development of climate and warning services beyond Peru.
+
+The dismantlement of meteorological and EWS stations by local people is not unique to this site, nor to Peru. Similar incidents have occurred in other regions, such as Himalayas, Andes and Alps of Europe, though those experiences were poorly documented.
+
+Local intra and inter-community conflicts, as well as distrust and biases against the participation of, and installations from, external institutions, can have a strong, but invisible, impact on acceptance.
+
+The relation of local (risk exposed) people to their natural environment and their perceptions of different risks strongly determines their attitude towards risk reduction efforts. Local perspectives
+
+may differ substantially from government or technical and scientific perspectives. For instance, local people can have intimate relations with mountains, glaciers and lakes as places of spirituality and the origin of life. Hence, a GLOF may be understood as a reaction of, for example, a glacier (as a mountain spirit) and a lake (as a being) to human disturbance or inappropriate human behavior. Traditional knowledge and narratives have to be recognized and acknowledged as part of a constructive dialogue and in finding acceptable solutions.
+
+As a consequence, the acquisition of a profound understanding of the social, political and cultural conditions, particularly in terms of power dynamics, is a prerequisite for early warning as well as more generally for climate adaptation service development. It is necessary for collaboration among diverse people, actors and experts, including local populations, physical and social scientists, engineers, local governments, technical governmental institutions, and nongovernmental organization (NGOs). It is encouraged to give the social sciences a more prominent role.
+
+Authorities often believe that an EWS is primarily a technical measuring and data transmission system. The recognition that an EWS also consists of institutional, social, cultural and political components is fundamental because an EWS can only be operational if all components fulfill their function. Furthermore, it is critical that the local authorities and people understand that an EWS cannot reduce risks to zero – its main objective is to avoid harm to human lives. Therefore, it needs to be accompanied by other risk reduction measures, in particular appropriate land-use planning.
+
+## Conclusion
+
+EWS in extreme environments, such as glacier lakes, bring many challenges. The system needs to be carefully designed to achieve robust energy provision, smooth and reliable data transmission, measurement of critical physical variables and the required degree of redundancy. Many months of system calibration are indispensable. Local authorities must clearly understand this. In addition, maintenance of the EWS needs to be budgeted annually by the local authorities to guarantee the sustainability of the system.
+
+The Laguna 513 EWS has become the model for several other EWS in the Peruvian Andes (e.g. Huaraz-Palcacocha, Urubamba-Chicón) and beyond. While the experience and capacity development can be replicated, it is also crucial to acknowledge that every location is an individual case with special characteristics that need appropriate attention.
+
+#

chatbot/rag/data/('Glacier Lake Management_ GLOF Early Warning System', '.pdf')_extraction.md

@@ -0,0 +1,130 @@
+
+
+## **Glacier Lakes & Floods**
+
+Glacier lakes are indirectly a result of climatic changes expressed by glacier fluctuations. During the earth history glacier lakes have survived only a short time period, but glacier lake outbursts had a profound impact in shaping the landscape. The dominant type of glacial lakes may shift over time from glacier-dammed lakes to moraine-dammed lakes in different glaciation situations in dependence of the topographic setting. In the course of deglaciation, the types of glacier will change and therefore the type of glacial lakes. Glacier-dammed lakes occur mainly in times of glacier advance, but there are also some constellations in which they are the result of glacier retreat. Moraine-dammed lakes are mainly linked to glacier retreat regimes.
+
+Glacier lake outbursts from these lakes can attain extremely high peak discharges by sudden dam failure. As a consequence of topographical and climatic conditions, a characteristic distribution pattern of glacier lake types can be recognized in individual mountain areas. In terms of natural hazards, sudden outbursts from small lakes with high peak discharges may have a more severe impact on human settlement than the gradual drainage of large ice-dammed lakes. Not all glacier lake outbursts have to be necessarily released as water floods. They may also occur as debris flow with a high hazard potential.
+
+## Glacial Lake Outburst Flood (GLOF)
+
+The acronym GLOF is used for glacier floods caused by the drainage of naturally dammed lakes in the glacier, on or at the margin of glaciers. "Glacier floods represent in general the highest and most far reaching glacial risk with the highest potential of disaster and damages". A lake outburst can be triggered by several factors: ice or rock avalanches, the collapse of the moraine dams due to the melting of ice buried within, the washing out of fine material by springs flowing through the dam (piping), earthquakes or sudden inputs of water into the lake e.g. through heavy rains or drainage from lakes further up-glacier. The glacier lake may breach at the end-moraine or at the lateral moraine.
+
+# GLOF - Monitoring & Early Warning System C-DAC's GLOF Early Warning System for Sikkim State
+
+Through extensive research, CDAC has deployed India's first GLOF Early Warning System for Sikkim to predict GLOFs, which will help the Government authorities in case of an impending GLOF event. The GLOF Early Warning System has the following capabilities:
+
+Image /page/1/Picture/10 description: A screenshot of a software application titled "GLOF Information System". The main window features a scenic background image of snow-capped mountains rising above a layer of clouds under a clear blue sky, with a dark mountain silhouette on the right. Overlaid on this image in large white text is the title "Glacial Lake Outburst Flood Management System". The application has a menu bar with the options: Home, GLOF Analysis, Settings, Help, and About Us. At the bottom, it displays "Powered by ICIMOD 2012", "Version 1.3", and a red button for "Licensing Policy".
+
+Image /page/1/Picture/11 description: A simple, green icon of a person standing with their arms raised and spread out to the sides, and their legs apart, against a white background. The figure is slightly blurry.
+
+Water Level Sensors developed indigenously by C-DAC, has been deployed at Shakho Chho Glacial Lake, North Sikkim.
+
+The near-real time sensors transmit data through INSAT satellite to control centre at Gangtok (SSCST).
+
+The flood simulation model runs the simulation for various GLOF scenarios. The model is capable of presenting Flood Simulation, Inundation information and Flood Arrival Time, in the event of a glacier lake outburst flood.
+
+Image /page/2/Picture/4 description: A screenshot of a software application titled 'GLOF Information System'. The 'Lake Information' tab is selected, displaying data for the year 2010. The selected lake is '78A013A1084027 (South Lhonak Chhu)'. The 'Lake Details' section provides the following information: Lake ID: 78A013A1084027, Lake Name: South Lhonak Chhu, alternate Name: South Bonak chhu, Lake Type: C, Lake Class: M, Lake Condition: Overtopping, Past GLOF occurrence: Yes, Dam Type: --NA--, Dam Condition: --NA--, Perimeter [mtr.]: 4957.20, Area [sq. mtr.]: 1009353.73, and Sensor Installed: --NA--. To the right, there is a small satellite image of the lake and a green box indicating 'Total Lakes in Sikkim: 203'.
+
+# The Glacial Lake Outburst Flood Management System - Interface & functionalities This is the main interface of the GLOF Management System. It features tools for GLOF
+
+Analysis, Glacial Lake Management & Flood Simulation Statistics.
+
+The GLOF Analysis Tool has three sub-tools: Lake Information, Simulation & the Sensor Information. The "Lake Information" tool gives basic information about the glacial lake selected by the user from the drop-down menu, such as, Lake Name, Lake Condition, etc. The satellite image of the glacial lake is shown on the right hand side.
+
+The Simulate tool gives information regarding the GLOF simulation progress and the outputs of the simulation. The user will be able to see the time at which the GLOF at a particular glacial lake occurred and the period of GLOF simulation.
+
+Image /page/2/Picture/9 description: A screenshot of a software application titled "GLOF Information System". The window displays the "Sensor Information" tab. On the left, there are buttons for "View", "Add", "Edit", and "Delete". The main area shows details for a sensor. Under "View Sensor Information", a dropdown for "Sensor Installed on Lake" shows "70A0934184040 [ ]". Under "Sensor Details", the "Sensor ID" is "SM516", the "Sensor Location" Easting is "27.8709000000" and Northing is "86.6103400000". The fields for "Sensor Model", "Sensor Type", and "Sensor Accuracy" are blank. A green button indicates "Number of Lakes with sensor: 1".
+
+The user can choose to see the simulation results such as Simulation video, Inundation map, Inundation report & Hydrographs.
+
+The Sensor Information tool gives basic information about the sensors deployed at the glacial lakes for monitoring water-levels. Each sensor has a Sensor ID linked to the Lake ID. The user may add information of new sensors when sensors are deployed on new glacial lakes in future or delete existing ones, in case if they are decommissioned.
+
+The Settings tool gives the user the flexibility to load the input data from any location on the computer. The data will be copied to an in-built database system for computations.
+
+Image /page/3/Picture/3 description: A screenshot of a software application titled "GLOF Information System". The window has a menu bar with options for "Home", "GLOF Analysis", "Settings", "Help", and "About Us". The main content area features buttons to "Connect to Database" and "Activate the Model". Below these are several sections, including "Input Settings", "Path Settings", "Database Settings", and "Simulation Settings". The "Input Settings" section is expanded, showing a field for "Lake Shapefile" with the "Year" set to "1990" and an "Upload" text box. A lower section, "Summary of Settings Made", displays a list of configurations:
+1. GIS Path: C:/Program Files/Quantum GIS Wroclaw/bin/qgis.bat
+2. Lakes in Year 1962: F:/All\_Glacier\_Data\_for\_QC/Lake Layers/Other Lakes/TA\_29Oct1962.shp
+3. DB Host: localhost
+4. Lakes in Year 1990: F:/All\_Glacier\_Data\_for\_QC/Lake Layers/Other Lakes/TA\_12Oct1990.shp
+At the bottom of the window are buttons for "Save current Settings", "Restore last saved Settings", and "Delete".
+
+Image /page/3/Picture/4 description: A low-resolution screenshot of a mapping or GIS software interface. The main window displays a map with a large, irregularly shaped yellow polygon in the center. Inside the polygon, the text "Kanchipuram Taluk" is visible. On the left side of the screen, there is a control panel with various options and a data table. The top part of the panel is titled "Route and Locate" and has fields for "View" and "Layer," with "Taluk Name" selected in the Layer field. Below this, a table shows attribute data with columns such as "OBJECTID" and "Taluk Name".
+
+The Query tool gives the user three options based on his expertise or authority in using the system. The "Ready-made Queries" option is for user who wishes to know basic information about the glacial lake of interest on the GIS interface. The queries are tailor-made for users who do not have expertise in GIS. In this mode, only the basic functionalities such as zoom in/out, info tool, select feature, etc. are only available on the GIS interface.
+
+The Custom Query option invokes the "Advanced Query" tool for users who are familiar with the GIS query system. The needs to select the query of interest from the list of attributes displayed on the left panel. The output will be displayed as a map.
+
+The Advanced Mode option invokes all the tools and features available on the GIS interface for the advanced users familiar with GIS.
+
+## **Contact Us**
+
+### Centre for Development of Advanced Computing (C-DAC),
+
+Emerging Solutions & e-Governance Group (ES&EG)
+
+4th Floor, C-DAC Innovation Park,
+
+Panchawati,
+
+Pune - 411008, Maharashtra, India
+
+Phone No. +91-20-2550 3233.
+
+Fax: +91-20-2550 3231
+
+Email: geomatics[at]cdac[dot]in / sandeepk[at]cdac[dot]in
+
+Image /page/3/Picture/17 description: A low-contrast image showing a white logo on a white background. The logo consists of the word "MEITY" in a large, bold, sans-serif font. Below it, in a smaller font, are the words "MINISTRY OF ELECTRONICS &" on the first line and "INFORMATION TECHNOLOGY" on the second line.
+
+Image /page/3/Picture/18 description: The image displays the logo for 'Azadi Ka Amrit Mahotsav' on a white background. The logo features a stylized, waving Indian flag with saffron on top and green on the bottom, separated by a white space. Below the flag, the word 'Azadi' is written in saffron, and directly underneath it, the word 'Mahotsav' is written in green.
+
+Image /page/3/Picture/19 description: A low-resolution digital graphic on a white background, featuring a green stick figure of a person with their arms raised to the sides. To the right of the figure is a small, vertical, stylized shape that is green on the top and red on the bottom.
+
+## **CENTERS**
+
+- Bengaluru
+- Chennai
+- Delhi
+- Hyderabad
+- Kolkata
+- Mohali
+- Mumbai
+- Noida
+- North East
+- Patna
+- Pune
+- Thiruvananthapuram
+
+## **LINKS**
+
+- About Us
+- Products & Services
+- R&D
+- Careers
+- Tenders
+- Press Kit
+- Video Gallery
+- Events
+- Awards
+- Downloads
+- Achievements
+- Alliance
+
+## **CONTACT US**
+
+Image /page/4/Picture/29 description: A cropped screenshot of a digital map. The map displays an area labeled "MANSAROVAR" in all caps, with "मानसरोवर" written in Hindi script directly below it. Other visible labels include "Overhead Water Tank" and "माराडा कालना" in Hindi script at the top right. A partially visible label at the top left reads "ite of ology". The map shows roads as thin gray lines and green areas as light green patches.
+
+## Centre for Development of Advanced Computing C-DAC Innovation Park,
+
+Panchavati, Pashan, Pune - 411 008, Maharashtra (India)
+
+Phone: +91-20-25503100 Fax: +91-20-25503131
+
+Image /page/4/Picture/33 description: A simple, stylized icon of a person in a vibrant green color against a white background. The figure is depicted standing with its arms and legs spread out, resembling a star shape or someone doing a jumping jack. The image is slightly blurry or pixelated.
+
+Help | Website Policies | Copyright Policy | Terms & Conditions | Reach Us | Sitemap
+
+Website owned & maintained by: Centre for Development of Advanced Computing (C-DAC) © 2025 C-DAC. All rights reserved.
+
+Image /page/5/Picture/4 description: A simple, green stick figure icon is shown against a white background. The figure has its arms raised and spread out to the sides, and its legs are also spread apart, resembling the pose of a jumping jack.

chatbot/rag/data/('Glacier Retreat and Glacial Lake Outburst Floods (GLOFs)', '.pdf')_extraction.md

@@ -0,0 +1,14 @@
+
+
+## **Summary**
+
+Glacier retreat is considered to be one of the most obvious manifestations of recent and ongoing climate change in the majority of glacierized alpine and high-latitude regions throughout the world. Glacier retreat itself is both directly and indirectly connected to the various interrelated geomorphological/hydrological processes and changes in hydrological regimes. Various types of slope movements and the formation and evolution of lakes are observed in recently deglaciated areas. These are most commonly glacial lakes (ice-dammed, bedrock-dammed, or moraine-dammed lakes).
+
+"Glacial lake outburst flood" (GLOF) is a phrase used to describe a sudden release of a significant amount of water retained in a glacial lake, irrespective of the cause. GLOFs are characterized by extreme peak discharges, often several times in excess of the maximum discharges of hydrometeorologically induced floods, with an exceptional erosion/transport potential; therefore, they can turn into flow-type movements (e.g., GLOF-induced debris flows). Some of the Late Pleistocene lake outburst floods are ranked among the largest reconstructed floods, with peak discharges of up to 107 m3/s and significant continental-scale geomorphic impacts. They are also considered capable of influencing global climate by releasing extremely high amounts of cold freshwater into the ocean. Lake outburst floods associated with recent (i.e., post-Little Ice Age) glacier retreat have become a widely studied topic from the perspective of the hazards and risks they pose to human society, and the possibility that they are driven by anthropogenic climate change.
+
+Despite apparent regional differences in triggers (causes) and subsequent mechanisms of lake outburst floods, rapid slope movement into lakes, producing displacement waves leading to dam overtopping and eventually dam failure, is documented most frequently, being directly (ice avalanche) and indirectly (slope movement in recently deglaciated areas) related to glacial activity and glacier retreat. Glacier retreat and the occurrence of GLOFs are, therefore, closely tied, because glacier retreat is connected to: (a) the formation of new, and the evolution of existing, lakes; and (b) triggers of lake outburst floods (slope movements).
+
+**Keywords:** glacier retreat, GLOFs, natural hazards, slope movement, climate change
+
+**Subjects:** Floods, Glacial Lake Outburst (GLOFs)
+

chatbot/rag/data/('Google_Flood_Forecasting', '.pdf')_extraction.md

@@ -0,0 +1,176 @@
+# The Google Flood Forecasting Initiative
+
+Sella Nevo $^{1[0000-0002-4743-3634]}$ 
+
+Google Research, Tel Aviv
+
+Abstract. The Google Flood Forecasting Initiative is the world's first large-scale machine-learning-based operational flood forecasting system. It currently covers more than 360 million people, and provides flood warnings to governments, disaster management organizations, and individuals at risk. It combines two separate models - a hydrologic model and an inundation model - which together can provide high accuracy, actionable forecasts for flooding in upcoming days.
+
+**Keywords:** Flood forecasting · Hydrology · LSTM.
+
+# 1 Introduction
+
+Floods are responsible for 5,000 to 60,000 fatalities [5, 18, 4, 17, 12, 11], affect between 95 million and 250 million people [5, 13, 12, 17, 4, 22], and cause between \$21 and \$33 billion in economic damages annually [3–5]. The majority of these harms are caused by riverine floods [17] - where a river overflows its banks and inundates the floodplain around it. The frequency and severity of floods has been increasing in recent decades [21, 10], and are expected to rise significantly in the future due to climate change, land use changes and other long term processes [16, 15]. On a brighter note, reliable early warning systems have been shown to prevent up to 43% of fatalities [2, 26], and up to 50% of economic damages [23, 1]. They have also been shown to have an impressive cost-benefit ratio of 1:9 [6], and much higher in low and middle-income countries (LMICs) [25], making them a top climate change adaptation policy recommendation by the United Nations and the World Bank.
+
+Human costs of flooding are heavily concentrated in LMICs - for example, about half of all flood-related deaths globally occur in India and Bangladesh [17]. Despite the promise of early warning systems, many of these countries lack the resources for the comprehensive data collection, ongoing recalibration and maintenance, and computational resources that classic flood forecasting systems demand.
+
+The Google Flood Forecasting Initiative aims to provide accurate and actionable flood warnings globally. It utilizes machine learning to create forecasting systems that are both more accurate and more scalable than classic flood forecasting systems, and then disseminates the resulting warnings via Google's various interfaces (Search, Maps, and smartphone notifications) as well as specialized interfaces to support governments and NGOs engaged in disaster management. At the time of writing, the system provides flood forecasting systems to regions covering more than 360 million people.
+
+# 2 Problem Statement
+
+Flood forecasting systems aim to provide governments, NGOs and individuals in the affected region reliable and actionable information about upcoming floods. The modeling task of riverine flood forecasting systems can often be divided into two main sub-tasks.
+
+The first is forecasting conditions within the river, sometimes described as hydrologic modeling. In this task one traditionally uses inputs such as precipitation, temperature, soil moisture and others throughout the river basin, and produces as an output either the river discharge or the water level at specific points within the river at a specific time, usually up to two weeks into the future. Traditional hydrologic models are "conceptual" models - they are inspired by the actual physical processes involved, though usually are vastly simpler than the actual processes they describe. They will have somewhere between several to several dozen parameters, which are calibrated to produce outputs that fit the historical measurements in the specific river.
+
+The second task is forecasting behavior across the floodplain, sometimes described as inundation modeling. Here we assume we already know the discharge or water level in the river (provided by a hydrologic model or a real-time measurement), and model the movement of water across the floodplain. The goal is to produce a spatially accurate map of flood extent (which areas are flooded) or flood depth (how deep is the water in each point on the map). The most common practice for these types of models are called hydraulic models - these are physics-based models which find finite-element solutions to a set of differential equations (specifically the St. Venant equations [8]) on a grid. These models can describe water behavior across the floodplain very accurately, when all their input data is accurate and they are run at sufficient resolution.
+
+The case for incorporating AI into these two types of models is not identical, but similar principles apply. Both traditional models (hydrologic and hydraulic models) require significant manual calibration, and often continuous recalibration based on new information. This leads to high costs of deployment and operation, up to hundreds of thousands of dollars for a single basin - which significantly limits their availability in the regions that need warning systems the most, and prevents scale-up of high-quality systems to truly global scales. Relatedly, both have extremely limited spatial transferability - researchers have been largely unsuccessful in utilizing models trained in one location to improve performance in other locations, e.g. ones with less data available. Computational costs can also act as a critical limitation - especially for the hydraulic model, which is incredibly sensitive to resolution yet requires computation proportional to one over the resolution cubed. Finally, both models share a limitation that is common in physics-based and conceptual models - they can be overly rigid. Even when repeated historical data shows some assumption of the model to be incorrect, it is often difficult (and sometimes impossible) to change the model to correct its errors.
+
+# 3 Method
+
+Our flood forecasting system uses ML-based methods to tackle each of the tasked described above. We will describe our models for each task separately. For each task, we currently have two different architectures in production - used in different circumstances, e.g. depending on data availability and performance. See figure 1 for an overview of the full system including all components described below, and other components beyond the scope of this chapter.
+
+Image /page/2/Figure/4 description: A flowchart illustrating a four-stage process for a flood alert system. The stages are labeled from left to right: Data management, Hydrologic modeling, Inundation modeling, and Alerts. 
+
+1. \*\*Data management\*\*: This stage takes three inputs: "Stream gauge measurements," "Precipitation measurements," and "Precipitation forecast." These inputs feed into a process labeled "Data: Ingestion, Quality Control, Correction."
+
+2. \*\*Hydrologic modeling\*\*: The output from Data management flows into this stage. It contains two models in blue boxes: "Linear" and "LSTM." A separate input, "External stream gauge stage forecasts," points to a pink vertical box labeled "Warning thresholds," which is positioned next to the Linear and LSTM models.
+
+3. \*\*Inundation modeling\*\*: The output from Hydrologic modeling flows into this stage. It contains two methods in green boxes: "Thresholding" and "Manifold."
+
+4. \*\*Alerts\*\*: The output from Inundation modeling flows into the final stage. This stage lists three types of alerts that are sent out: "Alerts to responsible authorities," "Alerts to emergency units," and "Alerts to population."
+
+Each stage title is accompanied by a small illustrative icon: a green checkmark for Data management, a line graph for Hydrologic modeling, a map of a river delta for Inundation modeling, and three smartphones displaying notifications for Alerts.
+
+Fig. 1. A high-level overview of the Google flood forecasting system.
+
+## 3.1 Hydrologic Model
+
+The two architectures used for hydrologic modeling are linear models, and LSTMs. All our models currently forecast water levels (as opposed to discharge), and provide forecasts in gauged locations (i.e. locations where there is a stream gauge installed) - though work is ongoing to expand our reach beyond this. The training and validation scheme for both models is described in a sub-section after the models themselves.
+
+Linear Model The model input includes past river stages from the target gauge (the gauge being forecast) and its upstream gauges (typically 2-5 gauges) for each time step (usually hourly). The output is a future river stage at the target gauge for a given lead time. A multiple linear regression model is trained with historical records of the above inputs and outputs. The model is optimized
+
+# 4 S. Nevo
+
+using the mean square error (MSE) loss function with L2 regularization. Linear models are trained separately for each target gauge, with a separate linear model trained for all lead times up to the gauge's maximal lead time. For example, a target gauge with a selected maximal lead time of 24 hours and hourly resolution implies 24 trained linear models. Figure 2 illustrates the schema of the linear model.
+
+Image /page/3/Figure/2 description: A diagram titled 'Linear Model' illustrates a predictive model's workflow. On the left, a series of inputs are shown, represented by red parallelograms. These inputs consist of 'upstream stage' and 'current stage' data for multiple past time steps, starting from (t-u) and continuing through to (t-1), indicated by vertical dots. These inputs are fed into a central blue vertical rectangle labeled 'linear regression'. This block also specifies 'u = 72 hour lookback'. The linear regression model produces an output, a red parallelogram labeled 'prediction (t)'. This prediction is then connected by an arrow to a smaller blue rectangle labeled 'MSE' (Mean Squared Error). Finally, an arrow from the MSE block points to another red parallelogram labeled 'target gauge stage (t)', representing the actual value to which the prediction is compared.
+
+Fig. 2. Schema of the linear model.
+
+Long Short-Term Memory Network (LSTM) Model Building on the work described in [19, 20, 14], the LSTM model consists of two LSTMs: a sequence-to-one hindcast model and a sequence-to-sequence forecast model, where the output of the hindcast model is used as the initial state of the forecast model. The hindcast model processes data from past days sequentially, taking the following variables as inputs at each timestep: (i) Precipitation measurements through the basin, (ii) near-real-time stage, and (iii) a linear combination of stage measurements from upstream gauges over some upstream "lookback" period (typically a few days). The linear combination of stage measurements is produced by a separate linear layer with gauge-specific weights that combines a (variable) number of upstream inputs per gauge into five features that are fed as inputs into the hindcast LSTM (similar to the linear model described above).
+
+The hindcast LSTM runs until the current time, defined to be the time of the last available measurement. The final cell state and hidden state of the hindcast LSTM are passed through a fully connected layer and the output of this "state handoff" layer is used as the initial cell state and hidden state of the forecast
+
+LSTM. The forecast LSTM advances one step for every lead time, producing the relevant forecasts for all lead times up to the maximal one. The main goal of using a state handoff between two LSTMs is to distinguish between the different inputs that are available in real-time during hindcast and forecast - and specifically the availability of precipitation measurements.
+
+All weights of the LSTMs (hindcast and forecast) are shared between all target gauges, i.e. they are regionally calibrated. The only gauge-specific weights are those of the linear upstream combiner layer which allows the architecture to support varying numbers of upstream gauges, travel times, sizes of tributaries, etc.
+
+The system estimates the uncertainty of the water stage, or the time dependent distribution over the predicted stage, using a countable mixture of asymmetric Laplacians (CMAL). The parameters of this distribution are generated by feeding the hidden state of the forecast LSTM into a dedicated head layer for each forecasted time step. At each time step, the loss is calculated as the negative log-likelihood of the observed stages given the LSTM forecasts using the next maximal lead time values. It is important to note that the likelihood-based loss function is calculated only over the outputs of the forecast LSTM, and the hindcast LSTM is used only to initialize the forecast state. Since training is shared for all target gauges, the maximal lead time in the training phase is taken as the maximum of the gauge-specific values. Figure 3 illustrates the schema of the LSTM model.
+
+Image /page/4/Figure/5 description: A diagram illustrating the architecture of an LSTM (Long Short-Term Memory) model for time-series forecasting. The model processes input data through several layers to predict future values. The flowchart begins on the left with inputs for different time steps, from (t-h) to (t-0). For each time step, the inputs are 'IMERG', 'current stage', and a stack of 'upstream stage' data. These inputs are fed into a 'fully connected (linear)' layer. The outputs from these layers are then passed to a 'hindcast LSTM (seq2one)' layer. The hidden state from the hindcast LSTM is processed by another 'fully connected' layer and then used to initialize a 'forecast LSTM (seq2seq)' layer. This forecast LSTM produces a sequence of outputs, each of which goes through a 'CMAL' block to generate a probability distribution. These distributions are compared against the 'target gauge stage' for future time steps from (t+0) to (t+k). The model's performance is evaluated using a 'Σ negative log-likelihood loss'.
+
+Fig. 3. Schema of the LSTM model.
+
+Training and Validation The stage forecast models are trained and validated with historical data using a cross-validation scheme, in which each fold uses one year's worth of data for validation and the rest for training (i.e., 1-year leave-out). For operational use, the models were retrained on the full data set to produce the best real-time forecasts possible.
+
+Operational Use For Alerting For each target gauge, if the maximal fore-casted river stage between the forecast's "current time" and the gauge's maximal lead time is above the predefined gauge-specific warning threshold, an alert is issued, and this maximal stage is used for inundation mapping (see below).
+
+## 3.2 Inundation Model
+
+The two architectures used for inundation modeling are the Thresholding model, and the Manifold model, both described below. The Thresholding model produces an inundation extent map but no forecast for flood depths. The Manifold model additional produces the water depth at each pixel, but requires a digital elevation model (DEM) and requires more effort to implement and deploy at scale. The training and validation scheme for both models is described in a sub-section after the models themselves.
+
+Thresholding Model The model assumes that each pixel in the area of interest (AOI) becomes inundated when the target gauge exceeds a (pixel-specific) threshold water stage. These thresholds are learned from the series of historic stage data at the target gauge and the corresponding state of the pixel (dry/wet) during these events. Each pixel in the inundation map is treated as a separate classification task, predicting whether the pixel will be inundated or not. We refer to the "wet" class as the positive class.
+
+The algorithm described below identifies pixel-specific thresholds and is aimed at maximizing some F-score using an optimized global parameter called minimal ratio. An F-score [7] refers to a weighted geometric mean of precision and recall, with  $F_1$  referring to the simple harmonic mean between them and  $F_3$ giving more weight to precision over recall, for example. The algorithm below can be optimized for any choice of F-score. To achieve this, an iterative process is applied to each pixel. In each iteration, we find the threshold that maximizes the ratio of true wet events (where the water stage at the gauge is above the threshold and the pixel was wet) to false wet events (where the water stage at the gauge is above the threshold and the pixel is dry). The threshold that maximizes this ratio is the most cost-effective threshold in the sense that it provides the most true wets per false wet instance. At the first iteration all training events are considered; then, after each selection of a threshold and its respective truefalse ratio, events with stage measurements above the threshold are discarded and a new iteration starts with the remaining events. If the new true-false ratio calculated is lower than the minimal ratio parameter value, the process stops and the final threshold for the pixel is the one found in the previous iteration.
+
+It can be shown that for every minimal ratio parameter value, no other set of pixel-specific thresholds achieves simultaneously better precision (i.e., fraction of all flooded pixels that are predicted as being flooded) and recall (i.e., fraction of all pixels that are predicted to be flooded and are really flooded); implying it is Pareto optimal. Therefore, for any F-score there exists some value of the minimal threshold parameter which finds the thresholds that optimize this F-score.
+
+In cases where the river stage input is higher than all past stage data, the Thresholding model's output inundation map is initialized from the most severe inundation extent seen in the historical events and expanded in all directions. The expansion distance is a linear function of the difference between the forecasted stage and the stage of the highest historical event.
+
+This Thresholding model requires almost no site-specific data like DEMs, and no manual work, making it appealing for large-scale deployment across many areas of interest in a short amount of time.
+
+Figure 4 illustrates the schema of the Thresholding model.
+
+Image /page/6/Figure/6 description: A diagram illustrating a 'Thresholding model' for creating flood inundation maps. The process starts with 'Historic inundation maps and gauge stage data' on the left. An arrow points from a specific pixel on these maps to a timeline representing 'Historic pixel state for each flood event'. This timeline shows a sequence of 'Dry' (orange boxes) and 'Wet' (blue boxes) states. Below this, a corresponding timeline shows the 'Historic gauge stage for each flood event' with values: 2.1m, 2.8m, 3.1m, ..., 4.5m, 5.7m, 6.2m, ..., 7.3m, 8.5m. An upward arrow labeled 'Optimized stage threshold for the pixel' points from this historical data. This threshold is then used with a new 'Gauge stage' of '4.2m' to produce an 'Output flood inundation map', which is displayed at the top left of the diagram.
+
+Fig. 4. Schema of the Thresholding model.
+
+Manifold Model The Manifold inundation model provides a machine-learning alternative to hydraulic models, by computing physically reasonable flood inundation. Its inputs are a DEM for the AOI and a target water stage. It outputs both the flood inundation extent and the inundation depth at each AOI pixel. The model is divided into two major parts, described below.
+
+Flood Extent To Water Height Algorithm
+
+The flood extent to water height algorithm converts a DEM and an inundation
+
+extent map (i.e., wet/dry state for each pixel) into a water height map, which is a per-pixel water height in meters above sea level. The algorithm tries to find a physically reasonable water height map that best matches the input inundation map, where the physically reasonable requirement is defined as: (1) the water height surface must be smooth, i.e. we aim to find a water height map that does not change significantly between neighboring pixels; and, (2) the water height surface should not have a minimum or a maximum at the interior of flooded regions. This optimization problem is not differentiable, and thus cannot be easily solved directly. Instead, the following heuristic can be shown to produce an optimal solution to the above optimization problem. The algorithm identifies the boundaries of the inundated areas of the input inundation map. The water height at these boundaries is extracted from the corresponding DEM. In between these boundaries, the algorithm uses the Laplace differential equation to interpolate the water heights. The water height map is defined as a low-resolution image, where every pixel is set to be of 32x32 DEM pixels. This assures that the output map is smooth and does not contain high frequency changes, while also reducing the computational complexity of the process. In addition, outlier DEM pixels, which are pixels that cause high Laplace tension, are removed to assure that the overall function is smooth.
+
+### Gauge Stage To Flood Depth Algorithm
+
+When inferring inundation depth for a real-time gauge water stage forecast, we do not have access to the full current flood extent (as the extent-to-height algorithm above assumes). To be able to provide depth in this more challenging setting, we first perform some precomputation. We first apply the Thresholding model described above to all past events in our training data, producing an inundation extent map for each gauge stage measurement. We then apply the flood extent to water height algorithm described above to produce a water height map for each past event.
+
+In real-time, when we receive an input water stage at the target gauge, the model simply performs per-pixel piecewise-linear interpolation between the water height maps of the training dataset to generate a new water height map corresponding to this input stage. The resulting water height map and the DEM are then used to generate: (1) an updated inundation extent map, by assigning a dry state to a pixel if its water height is lower than the DEM and a wet state otherwise; and, (2) an inundation depth map, as the difference between the water height and the DEM height for wet pixels. The use of the Thresholding model (as opposed to directly using satellite-based imagery of actual historical flood extents) as the input flood extent maps ensures that higher gauge stages always yield larger inundation extents, and thus removes unnecessary noise. When the model infers an inundation depth map for a gauge's water stage higher than all the events in the training set, it extrapolates the water height map by adding the gauge level difference to every pixel in the highest water height map computed from observed gauge stages, and uses the extrapolated water height and DEM to compute the water depth map.
+
+Figure 5 illustrates the schema of the Manifold model.
+
+Image /page/8/Figure/2 description: A diagram illustrating a workflow, labeled the "Manifold model", for generating an "Output flood inundation map and water depth map". The process starts at the bottom with a horizontal axis representing gauge stage, with points for historic flood events at 2.1m, 3.1m, 5.7m, and 7.3m. For each historic event, a "Computed inundation map (Thresholding model)" is shown. These maps are then used to create "Physically-constrained water height maps" in the middle row, a process that also incorporates Digital Elevation Model (DEM) data. The model can then take a new input, shown as a "Gauge stage" of 4.2m, to generate an "Interpolated and smoothed water height map". This map, along with DEM data, is used to produce the final two output maps: a flood inundation map showing the extent of the water in teal, and a water depth map showing varying depths in shades of blue.
+
+Fig. 5. Schema of the Manifold model.
+
+Training and Validation The inundation models are trained and validated based on historical flood events, where flood inundation extent maps from satellite data, along with the corresponding gauge water stage measurements, are available. Similar to the stage forecast models, a 1-year leave out cross validation scheme is used for training and validation. For operational use, the models are retrained with all historical data. It should be noted that, contrary to flood inundation extent, it is much more difficult to obtain gound truth data for flood inundation depth at scale. The Manifold model is, therefore, trained and validated only on the inundation extent. However, since the Manifold model is constrained to only produce physically reasonable water height maps, accurate inundation extent metrics on the test dataset imply reasonably reliable inundation depth results (with further validation ongoing).
+
+# 4 Resource Requirements
+
+## 4.1 Data sets
+
+The system utilizes the following data sets.
+
+**Precipitation Measurements** We currently use IMERG satellite data, though other sources can be used (as long as the models are re-trained for the new dataset).
+
+**Precipitation Forecasts** We currently use Global Forecast System (GFS) precipitation forecasts, though other sources can be used (as long as the models are re-trained for the new dataset)
+
+Stream Gauges Stream gauges are measurement devices that measure the water level (and sometimes estimate discharge) of the river. We use these as labels for the hydrologic model, and inputs for both the hydrologic model and the inundation model. We use both open stream gauge datasets such as GRDC, CAMELS, and Caravan - as well as proprietary datasets from the governmental hydrometeorology agencies we work with.
+
+Elevation Maps Accurate elevation maps are crucial for the inundation model, especially when producing depth estimates. The publicly available global DEMs (e.g., NASA SRTM and MERIT) unfortunately lack the required spatial accuracy and resolution for detailed flood inundation simulation. Furthermore, they are based on data from over a decade ago, thus failing to capture the frequent topography changes caused by past floods. Consequently, we currently construct higher-resolution, up-to-date DEMs for each AOI from high resolution satellite optical imagery data in a process that is based on stereographic imaging. To keep the DEM up to date, the model is retrained annually based on fresh imagery in locations where flooding causes frequent topography changes. We currently use proprietary Google-generated elevation maps, yet one can use any reliable hydrologically-conditioned elevation maps of sufficient resolution, which depending on the basin could be anywhere between 0.5 to 30 meters.
+
+Historical Flood Extent Maps We use the synthetic aperture radar ground range detected (SAR GRD) data from the Sentinel-1 satellite constellation to determine flood inundation maps at known timepoints and locations [24]. At any AOI, a SAR image is available once every several days, from which an inundation map was inferred using a binary classifier. Every pixel within a SAR image is classified as wet/dry via a Gaussian-mixture based classification algorithm. In order to calibrate and evaluate the classification algorithm, we have collected a dataset of Sentinel-2 multispectral images of flood events that coincide with the SAR image dates and locations. Reference Sentinel-2 flood maps were created by calculating per-pixel Normalized Difference Water Index (NDWI=(B3-B8)/(B3+B8), where B3 and B8 are green and near infrared bands, respectively) and applied a threshold of 0.
+
+## 4.2 Computational Resources
+
+Training and using these models requires computational resources as well. How-ever, significant effort has gone into minimizing the computational requirements.
+
+Training the hydrologic models takes seconds for the linear model and approximately 3 GPU hours for the LSTM (per location). Inference time is negligible for both.
+
+Training the inundation models takes between 10 CPU minutes to 10 CPU hours depending on the AOI for both the lightweight model and the manifold model. For comparison, calibrating hydraulic models takes about 100-1200 CPU years for the same areas of interest [9].
+
+# 5 Field Evaluation
+
+At the time of writing (early 2022), these models have been incorporated in operational systems for several years, currently covering about 360 million people across India and Bangladesh. So far, the models' performance in real-time operational systems is consistent with their cross-validated performance on historical data.
+
+## 5.1 Hydrologic Models
+
+Arguably the most popular metric for evaluating the accuracy of hydrologic models is the Nash-Sutcliffe Efficiency coefficient. This metric is the  $R^2$  score of the predicted discharge, where the  $R^2$  score (or coefficient of determination) of a prediction is defined as  $1 - \frac{\sum (\hat{x}_i - x_i)^2}{var(x)}$  (with  $x_i$  representing the true value of the i'th example, and  $\hat{x}_i$  representing the prediction for the i'th example). This can be thought of as describing the percentage of target's variance explained. An  $R^2$  score of 1 means a perfect prediction (and no variance remaining), while a score of 0 means the prediction is no better than guessing the average. This trait makes this score easy to understand, regardless of the scale and behavior of the target variable. Since most of our models predict water level rather than discharge, we use the closest parallel - an  $R^2$  score over the water level predictions. Our hydrologic models have achieved an  $R^2$  score of about 0.98 in our operational systems.
+
+A metric which is less common in the professional literature but is perhaps easier for non-hydrologists to parse is simply the average error of our predictions, in centimeters. Our hydrologic models have so far seen an average error of 8.5 centimeters, across all our operational systems.
+
+## 5.2 Inundation Models
+
+All of our models are evaluated at a 16x16 meter resolution. The precision of an inundation forecast is defined as the number of pixels that were both forecasted as being inundated and were indeed inundated divided by the number of pixels that were forecasted as being inundated. The recall of an inundation forecast is defined as the number of pixels that were both forecasted as being inundated and were indeed inundated over the number of pixels that were actually inundated. The f1 score of a forecast is the geometric mean of its precision and its recall.
+
+Our inundation models have been run and evaluated at a resolution of 16x16 meters per pixel, and achieve 0.69 per-pixel f1 score on average across our operational systems.
+
+## 5.3 Impact Evaluation
+
+In addition to evaluating our systems' accuracy, we also aim to evaluate their effectiveness in driving protective action. In 2021, a randomized controlled trial
+
+run by the Yale Economic Growth Center found that our collaboration with local NGOs to generate and distribute flood warnings in India increased the portion of people who received warnings prior to flooding events by 3x, and increased the portion of people who took protective action ahead of a flooding event by 4.4x.
+
+# 6 Lessons Learned
+
+In the five years since we began this project, we have learned many lessons along the way. Here are a few highlights.
+
+Focusing on the model is often not enough to drive real-world change. he real world is complicated, and the effectiveness of systems depends on only on the quality of your model. You have to ensure you have good (real-time) data collection systems, build a good product around your model outputs, and build partnerships with the key stakeholders that would use your systems. Especially in underserved communities, many components of the service pipeline may be flawed - and so it's critical to monitor and ensure that the full process from model to on-the-ground value works properly.
+
+Relatedly, missing data and data errors drive the majority of operational inaccuracy. Academic benchmarks and analyses will focus exclusively on model quality, yet in the vast majority of operational systems errors in model inputs can completely dominate over model mistakes. As a result, investing in error detection, error correction, uncertainty estimation, out-of-distribution detection, review and oversight, and other similar tools is incredibly important.
+
+Interaction with floods is incredibly diverse and requires different communication in different locations. People have different needs and expectations from flood forecasting systems. For example, in areas that rarely experience flooding, a flood with a depth of 15 centimeters can be a critical event - and people will expect a severe warning for it. However, in areas that experience frequent flooding, shallow flood waters are considered a non-event and people may be frustrated or surprised by even an informational update for such events. In extreme cases, like in some of the most flood-affected regions in Bihar, India, people ask to be warned only if and when flood waters are expected to reach waist height. Another example of such diversity is how people communicate with others about floods. In some regions, such as in much of South Asia, communities will share with their networks any information they have about upcoming floods. In others, such as parts of South America, there is stigma associated with being affected by floods, and so people won't discuss it - including critical safety information.
+

chatbot/rag/data/('Increasing risk of glacial lake outburst floods from future Third Pole deglaciation', '.pdf')_extraction.md

@@ -0,0 +1,143 @@
+
+Warming on Earth's Third Pole is leading to rapid loss of ice and the formation and expansion of glacial lakes, posing a severe threat to downstream communities. Here we provide a holistic assessment of past evolution, present state and modelled future change of glacial lakes and related glacial lake outburst flood (GLOF) risk across the Third Pole. We show that the highest GLOF risk is at present centred in the eastern Himalaya, where the current risk level is at least twice that in adjacent regions. In the future, GLOF risk will potentially almost triple as a consequence of further lake development, and additional hotspots will emerge to the west, including within transboundary regions. With apparent increases in GLOF risk already anticipated by the mid-twenty-first century in some regions, the results highlight the urgent need for forward-looking, collaborative, long-term approaches to mitigate future impacts and enhance sustainable development across the Third Pole.
+
+he Hindu Kush-Himalaya, Tibetan Plateau and surrounding areas are widely known as the Third Pole of the Earth as it is home to the largest number of glaciers outside the polar regions<sup>1</sup>. Widespread retreat of glaciers is taking place over most of its territory and has accelerated in recent decades as one of the consequences of global warming<sup>2-4</sup>. This glacier wasting is associated with the rapid expansion and new formation of glacial lakes<sup>5-8</sup>, bringing both large opportunities and risks<sup>9,10</sup>. Particularly, when water is suddenly released, glacial lake outburst floods (GLOFs)11 can devastate lives and livelihoods up to hundreds of kilometres downstream of their source<sup>12,13</sup>. This threat is most apparent in the Third Pole<sup>14,15</sup> where the warming rates are distinctly higher than the Northern Hemisphere and global mean<sup>16</sup>, and numerous GLOFs have been recorded, originating from both moraine-dammed and ice-dammed glacial lakes<sup>15,17</sup>. While outbursts from ice-dammed glacial lakes have been concentrated in the Karakoram and Pamir regions<sup>18,19</sup>, outbursts from moraine-dammed glacial lakes are observed across the Third Pole, and most frequently along the main Himalayan arc<sup>20</sup> where glacial lakes are increasing rapidly in both size and number<sup>5,6</sup>. The impact of a GLOF can extend across international boundaries<sup>21</sup>, creating severe challenges for early warning and other risk reduction strategies, particularly in critical transboundary areas<sup>22</sup>. Despite the severe threat that such large extreme events pose for sustainable mountain development over the Third Pole<sup>4</sup>, there remains a lack of understanding regarding how and where GLOF risk will evolve in the future. GLOF assessment over the Third Pole has typically been focused on moraine-dammed proglacial lakes owing to their potentially large flood volumes<sup>23,24</sup>, weak dam composition<sup>25</sup> and clear link to climate change<sup>17</sup>. Outburst
+
+floods from moraine-dammed glacial lakes can be triggered by various mechanisms, including intense precipitation and snowmelt<sup>26,27</sup>, and most commonly, from the impact of ice and/or rock avalanches into a lake<sup>20,28</sup>. While robust long-term trends in GLOF frequency are not evident<sup>17,29</sup>, the GLOF threat is expected to increase in response to future warming as lakes expand towards steep and destabilizing mountain cliffs<sup>30,31</sup>. With the expansion of communities, tourism, hydropower, transportation and other crucial sectors into exposed areas, substantial risk reduction and adaptation strategies will be required to avoid increased impacts on sustainable development in vulnerable mountain regions. Focusing here on moraine-dammed proglacial lakes, we draw on a comprehensive inventory of glacial lakes and past GLOF events across the Third Pole, to model and evaluate how and where GLOF hazard and risk will change in response to future deglaciation, using a robust and unified evidence-based approach.
+
+## Past evolution and present state of glacial lakes
+
+To understand the present state of glacial lakes across the Third Pole and changes over the past decades, all glacial lakes  $\geq$ 0.01 km² were mapped based on archival Landsat satellite images from 1990 to 2015 (Methods). Our inventory reveals the existence of 26,633 glacial lakes in the Third Pole, with a total area of 1,968.8  $\pm$  1.2 km² and approximately one-third of them being dammed by moraines (in total 7,650 with a combined area of 535.1  $\pm$ 0.7 km²; 6,958 of these lakes were formed in proglacial areas or are in a transitional state). The dataset suggests a clear heterogeneity in their spatial distribution (Fig. 1 and Supplementary Table 1) as today's glacial lakes are principally distributed along the Hindu Kush–Himalaya–Hengduan
+
+<sup>1</sup>State Key Laboratory of Desert and Oasis Ecology, Xinjiang Institute of Ecology and Geography, Chinese Academy of Sciences, Urumqi, China. <sup>2</sup>Climatic Change Impacts and Risks in the Anthropocene (C-CIA), Institute for Environmental Sciences, University of Geneva, Geneva, Switzerland. <sup>3</sup>University of Chinese Academy of Sciences, Beijing, China. <sup>4</sup>Department of Geography, University of Zurich, Zurich, Switzerland. <sup>5</sup>China-Pakistan Joint Research Center on Earth Sciences, CAS-HEC, Islamabad, Pakistan. <sup>6</sup>Dendrolab.ch, Department of Earth Sciences, University of Geneva, Geneva, Switzerland. <sup>7</sup>Laboratory of Hydraulics, Hydrology and Glaciology (VAW), ETH Zurich, Zurich, Switzerland. <sup>8</sup>Swiss Federal Institute for Forest Snow and Landscape Research (WSL), Birmensdorf, Switzerland. <sup>9</sup>Department of Geosciences, University of Fribourg, Fribourg, Switzerland. <sup>10</sup>Institute of Tibetan Plateau Research, Chinese Academy of Sciences, Beijing, China. <sup>11</sup>CAS Center for Excellence in Tibetan Plateau Earth Sciences, Beijing, China. <sup>12</sup>Department of F.A. Forel for Environmental and Aquatic Sciences, University of Geneva, Geneva, Switzerland. <sup>13</sup>These authors contributed equally: Guoxiong Zheng, Simon Keith Allen. <sup>188</sup>e-mail: baoam@ms.xjb.ac.cn; markus.stoffel@unige.ch
+
+Image /page/1/Figure/1 description: A map of High Mountain Asia, showing the distribution of glacial lakes and projections for aggregated lake area under different climate scenarios. The map spans from approximately 70°E to 110°E longitude and 25°N to 45°N latitude, covering regions like the Tibetan Plateau, Himalayas, Karakoram, and Hindu Kush. The map is divided into several labeled regions, including West Tien Shan, Pamir, West Himalaya, Central Himalaya, East Himalaya, Hengduan Shan, South and East Tibet, Inner Tibet, East Kun Lun, West Kun Lun, Qilian Shan, and Hissar Alay. Bar charts are overlaid on each region, illustrating the aggregated lake area for five conditions: 'Present' (cyan), 'RCP2.6' (light blue), 'RCP4.5' (light purple), 'RCP8.5' (purple), and 'Ice free' (magenta). A scale bar indicates that a specific height corresponds to 250 km². The charts generally show an increase in lake area from present conditions to the 'Ice free' scenario. A dashed line within the bars distinguishes the 'Moraine dammed' lake area. An inset map in the top right corner shows the geographical context of the main map, outlining the area within China and India.
+
+**Fig. 1** | Region-wide present and projected glacial lakes to 2050, 2100 and on an ice-free Third Pole. Map showing the geographical extent of the Third Pole and the spatial distribution of its present glacial lakes. Bar charts in different colours denote the present and potential future glacial lake areas (present plus projected results under three RCPs, and under the ice-free scenario) that were aggregated into the Global Terrestrial Network for Glaciers (GTN-G) regions<sup>46</sup>, respectively. The proportional area of present moraine-dammed glacial lakes to all present glacial lakes is shown in grey. Dashed lines show estimated changes in 2050.
+
+Shan mountains and ranges of inner and southeast Tibet as well as the Tien Shan mountains. The Hengduan Shan holds the largest number of glacial lakes  $(6,571;\ 436.2\pm0.5\ km^2)$ , whereas the eastern Himalaya has the most moraine-dammed glacial lakes  $(1,533;\ 183.7\pm0.4\ km^2)$ . At the basin scale, glacial lakes are most abundant in the Brahmaputra River Basin  $(9,665;\ 778.8\pm0.8\ km^2)$ , followed by the Indus, Yangtze, Ganges, Amu Darya and Salween river basins (all with >1,000 lakes, Supplementary Fig. 3 and Supplementary Table 2). The Brahmaputra River Basin is also home to the largest number of moraine-dammed glacial lakes  $(2,331;\ 212.0\pm0.4\ km^2)$ .
+
+Glacial lakes on the Third Pole have undergone a notable increase in total area and number since the 1990s<sup>5-7</sup>. Our analyses show an overall increase of  $6.8 \pm 0.1\%$  in glacial lake area and 5.9% in lake number across the Third Pole between 1990 and 2015 (Supplementary Fig. 4 and Supplementary Table 1). We propose that this recent increase has principally been driven by an expansion of moraine-dammed glacial lakes that charactize many of the long, flat, debris-covered glacier tongues in the region. To test this hypothesis, we split the glacial lakes into two categories: moraine-dammed and other-dammed glacial lakes (Methods). We find that if only the moraine-dammed glacial lakes are considered, the area gain is substantially larger at  $31.3 \pm 0.3\%$ , with an increase in number of 26.7%. In contrast, the size of other-dammed glacial lakes remains virtually unchanged  $(-0.02 \pm 0.2\%)$ , with a slight decrease in number of -1.4%. Still, changes show clear heterogeneity both at regional and basin scales (Supplementary Figs. 4 and 5).
+
+## **Historical GLOF activity**
+
+As GLOF events across the Third Pole have been heavily reported in previous studies<sup>18-20,29</sup>, we systematically collected and collated these records to better understand their spatial distributions and the characteristics of different source lakes, which were categorized into moraine, ice, bedrock and complex ones depending on the dam type (Fig. 2a; see Supplementary Table 3 for all records assembled). On the basis of this dataset, we found that at least 296 GLOF events from 109 glacial lakes occurred over the Third Pole since 1560 CE. GLOFs in the eastern Himalaya are primarily from the failures of moraine-dammed glacial lakes (Fig. 2c and Supplementary Table 3) and account for ~45% of the past GLOF sources alone. GLOFs resulting from ice-dammed glacial lakes are all reported in the Pamir-Karakoram regions (Fig. 2a), where the prevalent glacier surges have led to repetitive blockage and eventual release of glacial lakes<sup>19</sup>. Although most past GLOF sources were related to moraine-dammed glacial lakes (74 out of 109, Fig. 2b), ice- or complex-dammed flood sources (34 out of 109, Fig. 2b) contributed to more than half of the past outbursts (187 out of 296, Fig. 2b) due to the repetitive and high-frequency nature of these events. This is exemplified in the case of Kyagar Tsho (see Fig. 2a for its location and Supplementary Table 3), situated in the headwaters of Yarkant River, which produced at least 33 outburst floods in the past century with an accelerating trend during recent decades<sup>32</sup>. Practical experience has proven that real-time monitoring and early warning systems are effective in forecasting such active and high-frequency floods resulting from ice dam breaches<sup>32,33</sup>. The complex-dammed
+
+Image /page/2/Figure/1 description: A multi-panel figure, labeled a through e, presenting data on Glacial Lake Outburst Floods (GLOFs) in the Third Pole region of Asia.
+
+Panel a is a map showing the spatial distribution of reported GLOFs across Central and South Asia. GLOFs are color-coded by dam type: Moraine (cyan), Ice (orange), Bedrock (magenta), and Complex (green). The map also indicates transboundary GLOF threats, marking present (red dots) and future (hollow red circles) potentially dangerous glacial lakes, along with modeled flow paths and cross-border points.
+
+Panel b is a donut chart illustrating GLOF frequency by dam type. The counts are: Moraine (182 total from segments of 108 and 74), Ice (187 total from segments of 113 and 74), Complex (34 total from segments of 26 and 8), and Bedrock (2 total from segments of 1 and 1).
+
+Panel c is a donut chart showing GLOF frequency by source region. The regions and their respective counts are: Hindu Kush (112, 26, 3, 1), Karakoram (96, 11, 8, 6, 4), East Himalaya (62, 49), Central Himalaya (7), South and East Tibet (5), and Hengduan Shan (3, 1). There is also a segment with a value of 2.
+
+Panel d is a bar chart comparing the number of GLOF sources. For Moraine dams, there are 74 total sources, with 19 having disappeared after a burst. For Bedrock dams, there is 1 source, with 0 disappearing. For Complex dams, there are 8 sources, with 6 disappearing.
+
+Panel e is a detailed map of the Himalayan region, showing towns like Kathmandu and Zhangmu, and a high concentration of potentially dangerous glacial lakes and their modeled flow paths, with five specific points numbered along one such path.
+
+**Fig. 2 | Reported historical GLOFs and potential transboundary threats on the Third Pole. a**, Map showing the spatial distribution of recorded GLOF sources by lake dam type as well as present and projected (ice-free scenario) glacial lakes with possible transboundary GLOF threats across the Third Pole. For ice-dammed GLOF sources, only those with known geographic coordinates are shown (see Supplementary Table 3 for complete records). The black box near the bottom of the panel is the location of **e. b**, Double doughnut chart representing the number of past GLOF sources and flood frequency by lake dam type. **c**, Double doughnut chart showing the number of past GLOF sources and flood frequency per region. **d**, Statistics of GLOF sources by lake dam type. Ice-dammed cases were not counted owing to their repetitive nature. **e**, Amplified map showing a hotspot of potential transboundary GLOF threat between China and Nepal, and a historical GLOF hotspot in the eastern Himalaya. The circled numbers represent five concentrated regions with potential transboundary GLOF threats. The capital cities of Nepal and Bhutan are indicated with yellow squares. The left black box is the location of Supplementary Fig. 13 and the right black box is the location of Supplementary Fig. 14. Base maps: Google, Europa Technologies.
+
+glacial lakes are usually positioned on the surface of debris-covered glaciers and are held back by a mix of ice and moraine. These lakes often vanish fully after an outburst (Fig. 2d), but some exceptions exist. For instance, Merzbacher Lake (see Fig. 2a for the specific site and Supplementary Table 3), located upstream of Aksu River, has produced at least 66 recorded outbursts over the last century because of its peculiar geographic situation and water supply modes 14,34. The occurrence of floods from such lakes is sudden and difficult to monitor and mitigate, as drainage can occur within or underneath the glacier 34.
+
+## Present hotspots of GLOF hazard and risk
+
+Based on a conceptual model of GLOF hazard and risk optimized for large-scale automated application (Methods), we implemented
+
+the first quantitative assessment of potential GLOF hazard and risk for a total of 6,958 existing moraine-dammed glacial lakes across the Third Pole and provide hierarchical classifications. This study compliments numerous national- to regional-scale GLOF hazard and risk inventories that have previously been undertaken (Supplementary Table 4), allowing a comparison of GLOF hazard and risk across the entire Third Pole. The model considers GLOF hazard on the basis of well-established conditioning and triggering factors<sup>30</sup> and identifies infrastructure and communities at risk within downstream simulated flood paths. In contrast to previous work, we evaluate model reliability by validating classification results against those glacial lakes that have generated outburst floods in the past. The model successfully classified GLOF hazard as high or very high for 46 out of the 48 existing glacial lakes (that is,
+
+Image /page/3/Figure/1 description: A figure with three panels (a, b, c) illustrating Glacial Lake Outburst Flood (GLOF) hazard and risk. Panels a and b are maps with pie charts, and panel c is a histogram.
+
+Panel a, labeled "GLOF hazard", is a map showing various regions with overlaid pie charts. The size of each pie chart corresponds to the "Aggregated hazard value", with a scale showing concentric circles for values <10, 50, 100, 250, and 500. The slices of the pie charts are colored to represent different hazard levels: dark green for Very low (VL), light green for Low (L), yellow for Medium (M), orange for High (H), and red for Very high (VH).
+
+Panel b, labeled "GLOF risk", is a similar map. The size of its pie charts represents the "Aggregated risk value", with a scale for values <5, 25, 50, 125, and 250. The color coding for risk levels is the same as in panel a.
+
+Panel c is a histogram with the x-axis labeled "Normalized hazard value" (from 0 to 1.0) and the y-axis labeled "Frequency (n = 6,958)" (from 0 to 400). The bars are colored according to the "Hazard class": dark green for "Very low", light green for "Low", yellow for "Medium", orange for "High", and red for "Very high". The distribution is roughly bell-shaped, peaking around a normalized hazard value of 0.4. Above the histogram, brown dots represent "Historical GLOFs (n = 48)", with dashed lines connecting them to the corresponding hazard values. The number of historical GLOFs is indicated for some ranges, showing 1, 1, 9, and 37, with the majority occurring at high and very high hazard values.
+
+**Fig. 3** | Region-wide present GLOF hazard and risk across the Third Pole. a,b, Pie charts showing the proportion of different GLOF hazard (a) and risk (b) levels per region. VL, very low; L, low; M, medium; H, high; VH, very high. Aggregated hazard and risk values are the total sums of normalized values within each region. **c**, Histogram presenting the distribution of GLOF hazard levels across all moraine-dammed glacial lakes assessed. The vertical dashed lines indicate the distribution of assessed GLOF hazard values for glacial lakes where historical GLOFs have been recorded, which were further used to validate the overall model performance. The dots refer to the number of GLOF sources for a given hazard value.
+
+95.8% accuracy, Fig. 3c and Supplementary Table 6), from which the past GLOFs originated. For example, Jialong Co and Cirenma Co, located in eastern Himalaya, have been the origin of several GLOFs since 1960, and are both ranked within the 99th percentile of the most hazardous lakes. Overall, our assessment indicates that one-third (2,323) of all glacial lakes assessed have high to very high hazard levels, and one in six (1,203) glacial lakes would pose a potentially high to very high risk to downstream communities. More than one-third (2,619) of glacial lakes assessed appear not to be dangerous as no downstream infrastructures are evident within the flood paths (see Supplementary Table 5 for all infrastructures considered). On a regional scale, the eastern Himalaya has the highest value in terms of GLOF hazard, which is twice that of the adjacent Hengduan Shan, inner Tibet, central and western Himalaya, and southeast Tibet; three to four times that of the Tien Shan areas; and seven to eight times that of the Pamir, Hindu Kush and Karakoram (Fig. 3a). For GLOF risk, the eastern Himalaya likewise has the highest value, being two to three times that of the central Himalaya, Hengduan Shan and western Himalaya; four times that of southeast Tibet and western Tien Shan; and six to eight times that of inner Tibet, Hindu Kush, eastern Tien Shan and Pamir (Fig. 3b). Notably, some regions such as inner and southeast Tibet present high levels of GLOF hazard (Fig. 3a), but lower levels of GLOF risk (Fig. 3b) owing to much lower exposure values compared with the central and eastern Himalaya (Supplementary Fig. 6). While
+
+mapping of roads and other exposed elements over inner and southeast Tibet is considered less complete than elsewhere<sup>35</sup>, these results correspond well with other approaches based on gridded data including population density and cropland area that have likewise identified highest risk levels in central and eastern Himalaya<sup>36</sup>. At the basin scale, the Brahmaputra River Basin has the highest level in terms of GLOF hazard, followed by Ganges, Indus, Salween, Ili, Amu Darya river basins and Tibetan Plateau interior (Supplementary Fig. 7a). However, with respect to GLOF risk, the Ganges River Basin has the highest level, followed by Brahmaputra, Indus, Amu Darya, Ili, Salween and Syr Darya river basins (Supplementary Fig. 7b).
+
+## **Anticipation of future changes**
+
+We expand our assessment to consider how and where GLOF hazard and risk will evolve under continued global warming and glacier shrinking. Future glacial lake formation and associated hazard and risk are considered under three different CO<sub>2</sub> emission scenarios (Representative Concentration Pathways, RCP2.6, RCP4.5 and RCP8.5)<sup>37</sup> and projections of a glacier model<sup>38</sup> for the middle (2050) and end (2100) of the twenty-first century (Methods). These results are compared with an ice-free scenario which assumes that the Third Pole's glaciers have totally vanished. Our modelling suggests that >13,000 new glacial lakes with a combined maximum area of ~1,510 km² and a combined maximum volume of ~50 km³ could potentially emerge in an ice-free environment (Supplementary
+
+Table 1). This includes the expansion of existing glacial lakes to reach their topographically constrained extent, and potential new glacial lakes forming in depressions within the former bed of glaciers. In total, ~47% of this additional lake area would have already emerged by 2050, and 86% by 2100 under RCP8.5 (or 44% and 56%, respectively, under RCP2.6) (Supplementary Table 1). This future glacial lake development is mostly clustered along the Pamir-Karakoram-Kunlun Shan-Himalaya ranges as well as the western Tien Shan mountains (Fig. 1 and Supplementary Table 1), where the existing glaciers contribute ~74% of the total Third Pole glacier area (Supplementary Table 1). There is notable regional variation in projected glacial lake development from east to west, with regions such as southeast Tibet and eastern Himalaya already revealing close to their maximum lake area by 2050, under all RCP scenarios (Fig. 1). In contrast, in Karakoram, Pamir and western Himalaya, lake area will continue to increase strikingly into the late twenty-first century and beyond. On a basin scale, the future glacial lakes will be expected chiefly at the headwaters of the Indus and Tarim river basins, followed by Amu Darya, Ganges, Brahmaputra river basins and the interior ranges of the Tibetan Plateau (~89% of all projected glacial lakes, Supplementary Fig. 3 and Supplementary Table 2). The Brahmaputra, Ganges and Yangtze river basins almost approach their maximum lake area in 2050 under the three CO<sub>2</sub> emission scenarios, whereas lakes in the Indus and Tarim river basins as well as the Tibetan Plateau interior will continue to expand beyond 2100.
+
+To assess the possible implications of this future glacial lake development on GLOF hazard and risk, we applied the same model as was used for the present glacial lakes, with some slight modifications to account for future unknowns (Methods). For the whole Third Pole region, results indicate that future GLOF hazard and risk will increase by almost threefold relative to current conditions, with obvious regional variations (Fig. 4 and Supplementary Fig. 8). This is considered a conservative maximum estimate given that glacial lakes will appear gradually, and an unknown quantity of moraine dams could breach, and lakes thereby drain, before other lakes have emerged. Nonetheless, under the high-emission scenario RCP8.5, much of the Third Pole could already be approaching a state of 'peak risk' by the end of the twenty-first century, or even mid-century in some regions, with risk levels comparable to even the ice-free scenario (Fig. 4). In an ice-free environment, average glacial lake volumes will be ~80% larger than today, while topographic potential for ice/rock avalanche triggering will increase by ~230%, as lakes will be positioned closer to steep mountain headwalls (Supplementary Fig. 9). This suggests that the increase in potential GLOF frequency will generally be more important than the future change in magnitude. Meanwhile, in terms of risk, although overall exposure levels will increase due to the larger number of lakes, future lakes will on average have lower exposure levels (-40% compared with today) as they are located higher, and farther away from existing communities. As a consequence, average risk values per lake will be 14% lower in the future, while the overall number of lakes classified to have high or very high risk levels will increase from 1,203 to 2,963 lakes (that is, by a factor of 2.5). This analysis does not consider future changes in the distribution of population and infrastructure, particularly hydropower, that is expanding higher into alpine valleys<sup>39</sup>. Likewise, tourism expansion into remote mountain valleys could further enhance risks, while outmigration from rural areas could reduce risk in some cases<sup>40,41</sup>. Collectively, the Karakoram, Pamir and western Himalaya will be the regions with the most substantial increase in GLOF hazard, while in terms of risk, the largest increases will be in central Himalaya, Karakoram and western Himalaya (Fig. 4 and Supplementary Fig. 8). The eastern Himalaya will remain the primary GLOF hotspot under all assessed future scenarios, while the emergence of the Karakoram as an additional major hotspot of GLOF hazard and risk is most notable over longer timescales under RCP4.5 and RCP8.5, and under the ice-free
+
+scenario. While lake volumes will on average be larger than today, it is clear that even in the future, it is not necessarily the largest lakes that will be the biggest threat to communities (Supplementary Fig. 10). At the basin scale, the future increase in GLOF risk will be greatest in the Indus, Ganges and Amu Darya river basins, with the Ganges River Basin maintaining the highest risk levels under all future scenarios (Supplementary Figs. 11 and 12). By 2100, the Indus River Basin will exhibit the highest GLOF hazard levels (currently exhibited by the Brahmaputra River Basin), with close to a fourfold increase from today under the RCP4.5 scenario.
+
+## Potential transboundary threats
+
+The mountain ranges of the Tibetan Plateau and surrounding areas span 11 nations (Fig. 2a), giving rise to potential transboundary natural disasters, especially GLOFs. In fact, several relevant cases have been reported, for instance, in the transboundary Poiqu River Basin between China and Nepal<sup>22</sup> and the Shakhimardan catchment between Kyrgyzstan and Uzbekistan<sup>42</sup> (Fig. 2a). A well-known outburst from Cirenma Co (Fig. 2a) in 1981 caused serious devastation, including but not limited to the destruction of the Sino-Nepal Friendship bridge and part of the Nepal Sun Koshi hydropower station, and the loss of 200 lives in Nepal<sup>43</sup>. Recognizing the broader transboundary threat across the Third Pole, we identified where simulated flood flow paths from existing moraine-dammed glacial lakes intersected with national borders. Consequently, a total of 438 glacial lakes were found to have the potential to produce a transboundary flood, of which 191 were classified to be in the high or very high risk classes (Fig. 2a). Further analysis revealed that ~86% of these potentially dangerous lakes are located between China and Nepal, forming five hotspots of transboundary floods (Fig. 2e). It can be inferred from the illustrated example of one of these hotspots shown in Supplementary Fig. 13 that a possible outburst will have serious effects on downstream transboundary settlements and infrastructures.
+
+The transboundary GLOF threat will amplify in the future. Analysis suggests that the number of future potential transboundary flood sources under the ice-free scenario will roughly double (an additional 464 lakes), with 211 of these lakes classified in the high or very high risk classes (Fig. 2a). The border region between China and Nepal will remain a major hotspot (~42% of all these future dangerous glacial lakes), while the border region between Tajikistan and Afghanistan emerges clearly as an additional potential transboundary hotspot (~36%, up from ~5% at present).
+
+## **Summary and implications**
+
+We conclude that while the most dangerous hotspots of GLOF hazard and risk are now located within the eastern Himalaya, consistent with previous studies<sup>21,23,36,44</sup> and recorded GLOF events (Supplementary Fig. 10; see Supplementary Fig. 14 for a visual example in the eastern Himalaya), over the course of the twenty-first century, additional hotspots will emerge in the regions to the west, particularly under high-emission scenarios. For Karakoram and Pamir, where glaciers have recently exhibited a stable or slightly positive mass balance45, repetitive, high-frequency outburst events associated with ice-dammed glacial lakes will probably remain the dominant GLOF threat for several decades to come. However, the eventual depletion of glaciers in these regions projected during the twenty-first century will see conditions becoming less conducive for glacier surging and ice-dam formation<sup>19</sup>, and the threat of moraine- or bedrock-dammed glacial lakes substantially increasing. It is also important to emphasize that climate and related glacial changes as assessed here are just one driver of future GLOF risk; in some regions, future changes in exposure and societal vulnerability could potentially lead to different risk scenarios, especially where political, economic and social conditions facilitate or impede effective disaster risk reduction. Ultimately, determining exactly which lakes are the highest priority for response actions should be guided by consensus across multiple lines of evidence,
+
+Image /page/5/Figure/1 description: A map of the 'Third Pole' region in Asia, showing projected aggregated risk values for various mountain ranges under different climate scenarios. The map displays longitude from 70°E to 110°E and latitude from approximately 30°N to 40°N. Bar charts are overlaid on specific regions, including West Tien Shan, Hindu Kush, Karakoram, West Himalaya, Central Himalaya, and East Himalaya, among others. Each bar chart consists of five bars representing different conditions, as explained in a legend titled 'Aggregated risk value'. The conditions are: 'Present' (light yellow), 'RCP2.6' (light pink), 'RCP4.5' (pink), 'RCP8.5' (magenta), and 'Ice free' (dark purple). A reference bar indicates a value of 600. In all regions, the aggregated risk value increases progressively from the 'Present' condition to the 'Ice free' condition. A separate legend in the top right corner shows the 'Total risk value' for the entire 'Third Pole' region. The values for the five scenarios are: 1,510, 2,705, 3,197, 3,597, and 3,861, corresponding to the increasing severity of the climate scenarios.
+
+**Fig. 4 | Region-wide future changes in GLOF risk to 2050, 2100 and on an ice-free Third Pole.** Bar plots indicate projected changes in GLOF risk per region from the present to 2050 and 2100 (under three RCPs) as well as under the ice-free scenario. Dashed lines show estimated changes to 2050. The future GLOF risk was estimated based on currently known infrastructures that are exposed to modelled flood flow paths from potential future lakes and includes the assessed risk from present moraine-dammed glacial lakes (that is, these lakes are assumed to remain in the future). Inset: changes in GLOF risk over the whole Third Pole; note that the scale differs from the regional risk values.
+
+considering not only the wealth of information on existing GLOF hazard and risk (Supplementary Table 4), but also future emerging threats. Particularly where existing and emerging transboundary risks have been identified, risk reduction strategies will require high levels of international planning and cooperation. These transboundary regions represent vital economic corridors of activity, but some suffer from political tensions and challenges that can negatively affect timely data sharing, communication and coordination of early warning and disaster preparedness. We appeal to the relevant nations and international research communities to work together to prevent and mitigate the damages and losses that GLOFs bring to the Third Pole region, and to remain flexible and alert to the new challenges that could emerge under a future warmer climate.
+
+## Online content
+
+Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41558-021-01028-3.
+
+Received: 2 June 2020; Accepted: 16 March 2021; Published online: 6 May 2021
+
+## References
+
+
+
+© The Author(s), under exclusive licence to Springer Nature Limited 2021
+
+## Methods
+
+Glacial lake delineation and lake changes. We used Landsat satellite imagery to outline the boundaries of glacial lakes that are defined as water bodies formed primarily by glaciation and fed by glacial/snow melts, and are situated on the surface, adjacent or downstream of a glacier, or on palaeoglaciation landforms<sup>47</sup> across the Third Pole. Two time windows,  $1990 \pm 4$  and  $2015 \pm 1$ , were chosen to reveal the present state of glacial lakes and their changes over the last ~25 years. The selection was determined largely by the availability of Landsat datasets in such a large area. To minimize the effects of ice or seasonal snow cover on glacial lake mapping in these high elevations, we picked images mostly from the warm season (June to November), which corresponds roughly to the highest and most stable water areas/levels of the lakes in a hydrological year<sup>48</sup>. Overall, we used a total of 403 Landsat 4-5 Thematic Mapper (TM) scenes (97.8% from the warm season) for the 1990 epoch and 294 Landsat 8 Operational Land Imager (OLI) scenes (100% in the warm season) for the 2015 period (Supplementary Figs. 1 and 2a,b). We also gave priority to images with no or little cloud coverage, that is, an average cloud cover percentage of 7.7% and 4.2% for 1990 and 2015, respectively (Supplementary Figs. 1 and 2c). Landsat images were obtained from the US Geological Survey (USGS) data portal with a spatial resolution of 30 m. Images were first converted from the digital numbers to the top of atmosphere. Landsat 8 OLI images were further fused using a nearest neighbour diffusion-based pan-sharpening algorithm<sup>49</sup> to improve their spatial resolution to 15 m.
+
+An automated water body classification algorithm<sup>50</sup> based on global-local threshold segmentation for water indices greyscale images was applied to extract all water bodies from the Landsat images. A slope threshold of <20° as well as a shaded relief threshold of >0.25 were employed to alleviate the disturbances from mountain shadows<sup>50</sup>. Subsequently, careful visual inspection and manual re-editing were carried out to remove other water bodies (for example, reservoirs, rivers and small streams) and to correct the wrongly mapped glacial lakes based on corresponding Landsat scenes, online maps and other glacial lake datasets<sup>5,4</sup> available for the Third Pole. A 10 km buffer<sup>5,52</sup> around the Randolph Glacier Inventory (RGI 6.0)54 was used to preliminarily determine the distribution of what could be considered glacial lakes. However, to ensure the integrity and reliability of our inventory, we also included additional glacial lakes in areas where glaciers are missing from the RGI 6.0 dataset<sup>55</sup>. Both the glacial lake datasets for 1990 and 2015 were then cross-checked and attributed. The area of all glacial lakes was calculated based on the universal transverse mercator (UTM) projected coordinate system and their uncertainty ( $\delta$ ) was estimated using the formula<sup>56</sup>:  $\delta = P/G \times G^2/2 \times 0.6872$ , where P is the perimeter of the glacial lake and G is the spatial resolution of the images used. Lastly, we selected all glacial lakes ≥0.01 km<sup>2</sup> and superimposed them in a high-resolution Google Earth 3D view mode together with the corresponding Landsat scenes to distinguish the moraine-dammed glacial lakes separately. The remaining lakes were grouped and termed 'other-dammed glacial lakes'. Lakes dammed with ice or lying on the surface of glaciers were not included in the temporal change analyses due to their transient character and large seasonal fluctuations.
+
+The relative change in glacial lake area between 1990 and 2015 and its uncertainty were estimated using the following equations:
+
+$$R = \frac{A_2 - A_1}{A_1} \times 100\%$$
+
+$$\delta R = \sqrt{\left(\frac{\delta A}{A_2 - A_1}\right)^2 + \left(\frac{\delta A_1}{A_1}\right)^2} \times R$$
+
+where R is the relative change in glacial lake area and  $\delta R$  is its uncertainty;  $A_1$ ,  $A_2$  are the lake areas at the beginning and end of the period, respectively; and  $\delta A_1$ ,  $\delta A_2$  are their uncertainties,  $\delta A = \sqrt{\delta A_1^2 + \delta A_2^2}$ .
+
+Modelling of future glacial lakes. Glacier-bed topography, that is, ice-free digital elevation model (DEM) can be estimated using model approaches that determine local ice thickness based on the characteristics of surface topography and considerations of ice-flow dynamics8. The potential sites of future glacial lake formation under scenarios of complete deglaciation can then be projected by detecting overdeepenings on the glacier-bed. Here we used the latest global glacier ice thickness dataset<sup>57</sup> consisting of a consensus of five different ice thickness models as well as the 30 m Advanced Land Observing Satellite Global Digital Surface Model (ALOS AW3D30 v2.2) to extract the bed topography of all Third Pole glaciers. Three analytical steps were performed to determine the possible sites of future glacial lake formation: (1) glacier areas with slopes <20° were mapped first as potential locations of glacier-bed overdeepenings as lakes normally form in flat topography; (2) depressions in the glacier bed were detected by filling them using a classic filling algorithm<sup>58</sup>, while a slope raster was generated based on the filled glacier-bed topographies; and (3) depressions with a slope <1° were identified8 and intersected with the layer of possible overdeepenings obtained in the first step to generate the final layer of the possible future lakes. Ultimately, we selected all modelled overdeepenings ≥0.01 km<sup>2</sup> as the potential sites of future lake formation, to ensure comparability to the glacial lake inventories in 1990 and 2015. The area and volume of all modelled future glacial lakes were estimated based on a bathymetry raster that was derived from the difference between the
+
+overdeepenings-filled DEM and the original glacier-bed topography. The mean and maximum depth of each glacial lake was also estimated using the detected depression extent and its bathymetry raster. In general, ice-thickness models have proven robust in detecting the locations of overdeepenings  $^{\rm sp}$ , although local uncertainties in ice-thickness estimates (and thereby volume of the overdeepenings and associated future lake) could be on the order of  $\pm 100\%$  for individual cases  $^{\rm co}$ . Likewise, dam composition (moraine or bedrock) cannot be distinguished for future lakes, and lake volumes will be ultimately determined by dam geometry and the presence of an outlet channel.
+
+To estimate the time horizon in which potential future glacial lakes could emerge, we simulated glacier recession based on an ensemble of 14 global circulation models of the Climate Model Intercomparison Project Phase 5 (ref. 61) under three different CO<sub>2</sub> emission scenarios — RCP2.6, RCP4.5 and RCP8.5, in addition to the hypothetical end case with a completely ice-free Third Pole. Two major time horizons over the course of the twenty-first century, 2050 and 2100, were considered. The Global Glacier Evolution Model (GloGEM)38 was employed to model the evolution of all individual RGI 6.0 glaciers based on projected climate forcing until 2100. GloGEM resolves all components of the glacier surface mass balance and computes changes in glacier geometry (thickness, length) in annual time steps. Results for the ensemble of global circulation models have been averaged for the three considered RCPs. The minimum elevation for each glacier at both 2050 and 2100 under each RCP scenario was calculated and used to determine the exposed overdeepenings at each time step. We assume that the potential glacial lake is forming or has formed if the modelled minimum elevation of its parent glacier is greater than or equal to the elevation of the modelled overdeepening that was retrieved based on the filled glacier-bed topographies.
+
+GLOF hazard assessment and validation. Moraine-dammed glacial lakes dominate past GLOFs recorded in most of the world17, especially in the Himalaya<sup>20,29</sup>, and their evolution can be directly linked to atmospheric warming and glacial recession17. Hence, for the current GLOF hazard and risk assessment, we focused on such lakes. A conceptual model<sup>21</sup> was used to assess hazard and risk for 6,958 existing moraine-dammed proglacial lakes across the Third Pole. The model does not apply to some glacial lakes situated on the surface or flanks of debris-covered glaciers, as predisposing and triggering factors in these cases are different. Furthermore, we recognize the transition and complexity that occurs between different lake types, meaning that some degree of subjectivity in lake classification cannot be avoided. The model defines three indices21: the GLOF risk index as a function of the hazard index (combining likelihood and potential magnitude of a GLOF) and exposure index (potential for people and infrastructure to be affected). Here we improved the model for large-scale implementation so that it can run on a single glacial lake object iteratively. The large-scale assessment was based on a single UTM grid zone. All calculations and analyses described below were based on the 90 m Multi-Error-Removed Improved-Terrain (MERIT) DEM62, which is proven to have the best performance on the Third Pole<sup>63</sup>
+
+The potential hazard of a GLOF was defined based on four key factors<sup>21</sup>, which encompass well-established processes that drive GLOF triggering and magnitude30. (1) Lake volume was calculated using a lake depth-area-volume empirical relationship<sup>64</sup> and used as a proxy for possible maximum flood magnitude. (2) Total watershed area upstream of each glacial lake was used as an indicator of the potential for heavy rain and glacier/snow melt runoff to flow into the lake or cause it to overflow<sup>20,26</sup>. (3) The likelihood of ice and/or rock avalanches triggering an outburst was calculated based on the concept of topographic potential which includes: the potential for ice and/or rock to detach (parameterized by slope angle); and the potential for the ice and/or rock avalanche to reach a glacial lake (parameterized by the overall trajectory slope or angle of reach). Due to the quality of available glacier datasets at such a large scale, we did not distinguish whether the surrounding slopes were rock or ice-covered. Thus, we assumed that an avalanche was possible from any slope >30°, and the resulting avalanche could feasibly impact the lake where the overall slope trajectory is >14° (tangent angle = 0.25)65. (4) Downstream slope of the lake dam was considered to be an important factor controlling dam stability and the potential for self-destruction. It is also relevant for the erosional capacity of a GLOF event in downstream areas. This factor was calculated as the average horizontal downstream slope of each lake in three different buffer zones depending on different lake areas (A) intervals ( $A \ge 0.1 \text{ km}^2$ , 900 m;  $0.1 > A \ge 0.05 \text{ km}^2$ , 600 m;  $0.05 > A \ge 0.01 \text{ km}^2$ , 300 m). These intervals were derived empirically based on careful analyses of a large number of lakes using high-resolution satellite images.
+
+All factors were then normalized to 0 to 1 using the percent ranking method and assumed to have equal weight to the hazard of a GLOF event. For display and comparative purposes, hazards of all glacial lakes were eventually classified into five levels (very low, low, medium, high, very high) using the natural breaks (Jenks) method and then were aggregated to each region and river basin based on the sum of all lake hazard levels.
+
+The verification of model outcomes is highly dependent on empirical information<sup>21</sup>. Here, to verify the results from the GLOF hazard assessment and provide the historical context of GLOF activity, we compiled the most complete inventory of past GLOFs over the Third Pole. Thus, we collected and re-evaluated almost all of the historical GLOF events (Supplementary Table 3) that have been
+
+reported on the Third Pole and found a total of 296 GLOF events from 109 glacial lakes. Finally, 48 moraine-dammed glacial lakes that did not disappear after previous GLOFs and remained  $\geq$ 0.01 km² in the 2015 epoch were used to validate the hazard assessment.
+
+GLOF exposure and risk. Exposure represents the presence of people, livestock, infrastructure and other assets that could be threatened by potential GLOF hazard<sup>21</sup>. Here we used infrastructures (such as buildings, highways, railways and historic places, see Supplementary Table 5 for all features used) to characterize the communities and populations that might be affected within a potential GLOF flow path, which was simulated using the GIS-based modified single-flow hydrological model<sup>66</sup>. All infrastructure features were downloaded from OpenStreetMap (accessed 1 August 2019) and were converted to a raster grid. The maximum downstream flow distance for each lake path was determined based on an empirically derived worst-case scenario defined by the angle of reach from the source lake, with a tangent angle = 0.05 (3° angle of reach)<sup>67</sup>. Beyond this flow distance, damages from a GLOF are not expected. The sum of the angle of reach for each OpenStreetMap raster pixel that is exposed to the lake flow path was aggregated as the quantified measure of exposure for each glacial lake. As of 2016, the OpenStreetMap was considered to be 83% globally complete and increasing rapidly, although generally lower (yet relatively homogeneous) levels of completeness are seen over the Third Pole<sup>35</sup>. The exception is Nepal where intense humanitarian mapping efforts have led to near total completeness.
+
+The exposure values for each glacial lake were also normalized to 0 to 1 using the percent ranking method and then multiplied with the normalized hazard value to give the final risk level associated with each lake. The risk values of all glacial lakes (except those with an exposure value of zero indicating no risk to downstream communities) were divided into five classes using the natural breaks (Jenks) method and aggregated to each region and river basin to obtain their overall risk levels. In the absence of reasonable proxy variables at such a large scale, the vulnerability component of risk was not considered here.
+
+Future GLOF hazard and risk assessments. The same modelling approach was also applied to assess future GLOF hazards and risks for all modelled glacial lakes formed at 2050 and 2100 under three different RCPs and the ice-free scenario. The difference was that the current topography of the glacier areas was substituted by the modelled glacier-bed topography and then the selected factors were calculated for these future scenarios. However, because we cannot judge whether these potential glacial lakes will be dammed by moraine or bedrock, and resulting dam geometries are highly uncertain, we used the average dam slope of present glacial lakes in each region or river basin to characterize all modelled lakes in the same region or basin. Likewise, no robust basis exists to predict the future development of downstream infrastructure and communities across all of the Third Pole; thus, the future risk assessment considers only the change in exposure of existing infrastructures and communities.
+
+## Data availability
+
+The Landsat datasets are freely available from the USGS data portal (https://glovis.usgs.gov/). The Randolph Glacier Inventory 6.0 (ref. <sup>54</sup>) data are freely available at http://www.glims.org/RGI/. The ALOS Global Digital Surface Model (AW3D30 v2.2) is freely available at https://www.eorc.jaxa.jp/ALOS/en/aw3d30/index.htm. The Multi-Error-Removed Improved-Terrain (MERIT) DEM is freely available at http://hydro.iis.u-tokyo.ac.jp/~yamadai/MERIT\_DEM/ on reasonable request. The OpenStreetMap data are freely obtained from http://www.openstreetmap.org/ and available under the Open Database License (http://www.openstreetmap.org/ copyright). The composite glacier ice thickness data are available at https://www.research-collection.ethz.ch/handle/20.500.11850/315707. The produced glacial lake inventories and modelled future lakes as well as relevant assessment results are available at Zenodo under the identifier https://doi.org/10.5281/zenodo.4477945.
+
+## Code availability
+
+The GLOF hazard and risk assessment models are available at Zenodo under the identifier https://doi.org/10.5281/zenodo.4477947. Additional model or code used in this study are available from the corresponding authors on request.
+

chatbot/rag/data/('Monitoring Glacier Lake Outburst Flood (GLOF) of Lake Merzbacher Using Dense Chinese High-Resolution Satellite Images', '.pdf')_extraction.md

@@ -0,0 +1,285 @@
+
+
+# Monitoring Glacier Lake Outburst Flood (GLOF) of Lake Merzbacher Using Dense Chinese High-Resolution Satellite Images
+
+Changjun Gu, Suju Li\*, Ming Liu, Kailong Hu and Ping Wang
+
+National Disaster Reduction Center, Ministry of Emergency Management, Beijing 100124, China \* Correspondence: lisuju@ndrcc.org.cn
+
+Abstract: Establishing an effective real-time monitoring and early warning system for glacier lake outburst floods (GLOFs) requires a full understanding of their occurrence mechanism. However, the harsh conditions and hard-to-reach locations of these glacial lakes limit detailed fieldwork, making satellite imagery a critical tool for monitoring. Lake Mercbacher, an ice-dammed lake in the central Tian Shan mountain range, poses a significant threat downstream due to its relatively high frequency of outbursts. In this study, we first monitored the daily changes in the lake area before the 2022 Lake Mercbacher outburst. Additionally, based on historical satellite images from 2014 to 2021, we calculated the maximum lake area (MLA) and its changes before the outburst. Furthermore, we extracted the proportion of floating ice and water area during the period. The results show that the lake area of Lake Mercbacher would first increase at a relatively low speed (0.01 km<sup>2</sup>/day) for about one month, followed by a relatively high-speed increase (0.04 km<sup>2</sup>/day) until reaching the maximum, which would last for about twenty days. Then, the lake area would decrease slowly until the outburst, which would last five days and is significant for early warning. Moreover, the floating ice and water proportion provides more information about the outburst signals. In 2022, we found that the floating ice area increased rapidly during the early warning stage, especially one day before the outburst, accounting for about 50% of the total lake area. Historical evidence indicates that the MLA shows a decreasing trend, and combining it with the outburst date and climate data, we found that the outburst date shows an obvious advance trend (6 days per decade) since 1902, caused by climate warming. Earlier melting results in an earlier outburst. This study provides essential references for monitoring Lake Mercbacher GLOFs and building an effective early warning system.
+
+Keywords: Lake Mercbacher; GLOF; remote sensing; early warning; climate change
+
+Image /page/0/Picture/8 description: A graphic on a white background featuring a yellow circular icon with a white checkmark inside. To the right of the icon, the text 'check for updates' is displayed in two lines. The top line reads 'check for' in a regular dark gray font, and the bottom line reads 'updates' in a bold dark gray font.
+
+Citation: Gu, C.; Li, S.; Liu, M.; Hu, K.; Wang, P. Monitoring Glacier Lake Outburst Flood (GLOF) of Lake Merzbacher Using Dense Chinese High-Resolution Satellite Images. *Remote Sens.* **2023**, *15*, 1941. https://doi.org/10.3390/rs15071941
+
+Academic Editor: Yi Luo
+
+Received: 25 February 2023 Revised: 3 April 2023 Accepted: 4 April 2023 Published: 5 April 2023
+
+Image /page/0/Picture/12 description: A Creative Commons Attribution (CC BY) license icon. The icon is a rectangle with a grey top section and a black bottom section. The grey section contains two symbols in white circles with black outlines. On the left is the Creative Commons 'CC' symbol, and on the right is the attribution symbol, which is a stylized icon of a person. In the black section below the attribution symbol, the letters 'BY' are written in white.
+
+Copyright: © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
+
+# 1. Introduction
+
+Glacial deglaciation is becoming a worldwide phenomenon due to accelerating global warming [1–4]. As glaciers recede, their stability decreases, and they release stored water, which can lead to glacial hazards [5–7]. The melting water of glaciers, combined with overdeepenings in former glacier beds, has caused the global glacier lake volume and quantity to increase rapidly [8–10]. In fact, it has been suggested that the lakes formed from melting glaciers have increased by 50% in just 30 years (1990–2018). Dammed by abandoned moraines and ice, glacier lakes are susceptible to outbursts and can cause severe damage to human settlements and infrastructure downstream [7,11–15]. Statistics show more than 6300 people have died of glacier lake outburst floods (GLOFs) since the 1500s in High Mountain Asia alone [14]. Moreover, 15 million people globally are exposed to impacts from potential GLOFs, especially in High Mountains Asia (HMA) [13]. Real-time monitoring and early warning systems can effectively forecast active and high-frequency floods resulting from ice dam breaches [9]. However, GLOFs dynamics are challenging to measure and quantify due to unpredictable triggers and hard-to-reach locations [16].
+
+Remote Sens. 2023, 15, 1941 2 of 18
+
+Fortunately, increasingly detailed digital topographic data and satellite imagery provide more possibilities in GLOF monitoring [17–19].
+
+Located in the upper reaches of the Kumarik River in Kyrgyzstan, a tributary of the Aksu River in China, Lake Merzbacher is the largest glacial lake in the central Tien Shan mountain range of northwestern HMA [19]. The lake derives its water from the melt of the Inylchek Glacier and is divided into two parts: an upper lake and a lower lake [20]. The lower lake is ice-dammed and much larger than the upper one during the warm season when it is filled with water. Studies and reports have shown that there were 80 recorded outbursts of lower lake from 1932 to 2021 [21]. The average elevation of the lower lake surface is about 3380 m, and the drop from the lake to 200 km far away downstream communities is approximately 2000 km. This substantial drop in elevation makes Lake Merzbacher outburst floods to be more destructive and pose a significant threat to human settlements and infrastructure downstream.
+
+Lake Merzbacher, first discovered by Gottfried Merzbacher in 1903 [22], has been the focus of research on its formation and outburst mechanism since recording its first outburst flood in 1931. The mystery of the outburst mechanism was not solved until Ajrapet demonstrated the process in 1971, revealing that the northern Inylchek Glacier dammed the lake, and a hidden stream existed under the ice dam. Lake Merzbacher exhibits a unique outburst mechanism that differs from other glacial lakes. Research has shown that the northern Inylchek Glacier dams the lake, and a hidden stream flows beneath the ice dam. As the lake fills with water, the dam gradually rises, and water flows downstream through the hidden stream [23]. This process provides a short window of opportunity for early warning. Previous studies have explored the potential methods to capture this signal. An understanding of the outburst probability is central to early warning, and thus predicting the outburst probability of GLOFs draws much academic attention. By selecting the main factors which link to dam failure, the qualitative probability of dam failure was derived and mapped over a large region [24,25]. However, identifying potentially dangerous lakes is crucial to early warning efforts. Xie et al. [19] first build an index to monitor the outburst of Lake Merzbacher based on time series remote sensing images before and after the flood. Nonetheless, the temporal-spatial resolution of available satellite images can hardly capture all the details before the outburst of Lake Merzbacher. One possible scenario is that there would not be enough satellite shreds of evidence before the outburst, especially considering the frequent clouds in this region. Therefore, continuous monitoring of the lake before an outburst is vital for the timely and accurate forecast of disasters.
+
+Research has identified four distinct periods in the outburst process of Lake Merzbacher during a GLOF: icefall, quick water storage, early warning, and post-drainage. However, the characteristics and duration of the four periods are still unclear. One chief reason for this is that there are not enough available images. Glacial lakes are often located in alpine regions, where the cloudy weather decreased the availability of the time series satellite images. Furthermore, widely used satellite images, including MODIS, Landsat, and Sentinel, are deficient for the long revisit period and thus can hardly support GLOF monitoring. During the past twenty years, China's Earth observation activities have rapidly developed, creating a spaceship fleet resembling those from NASA, ESA, or JAXA, forming Chinese Earth Observation Satellite (CEOS) constellations [26,27]. Five major CEOS constellations are ready for use, including GaoFen satellites and HuanjingJianzai satellites, which have been widely used in land cover mapping [28–31]. The high resolution and short revisit period of (CEOS) constellations provide more opportunities in time series remote sensing monitoring. This study aims to obtain a comprehensive record of the Lake Merzbacher outburst using dense Chinese satellite images, and to analyze key elements during the outburst flood process, ultimately quantifying warning signals.
+
+Remote Sens. 2023, 15, 1941 3 of 18
+
+# 2. Materials and Methods
+
+## 2.1. Study Area
+
+Lake Merzbacher, located in the Inylchek Region at 79°52′E and 42°13′N, is situated in the central Tien Shan mountain range in Kyrgyzstan and close to the Chinese border [19,20]. It is considered the largest moraine lake in Central Asia, which was formed by Inylchek Glacier and first discovered and named by the German explorer Gottfried Merzbacher in 1903 [32]. Specifically, the northern branch of the Inylchek Glacier, with a length of 41 km and an area of 247 km², retreat and the southern branch of the Inylchek Glacier, with a length of 61 km and an area of 574 km², expansion together formed this glacier lake (Figure 1). This glacial lake plays a vital role in preventing the direct runoff of North Inylchek Glacier drainage water into the valley [33]. Lake Merzbacher comprises two small lakes connected by the Merzbacher River [32]: the upper one at 3400 m and the lower one at 3300 m. Lower Lake Merzbacher, dammed by the Southern Inylchek Glacier and larger than Upper Lake Merzbacher, is responsible for the glacier lake outburst flood (GLOF) that occurs regularly and has shown a decreasing trend since the beginning of the 20th century [34].
+
+Image /page/2/Figure/4 description: A side-by-side comparison of two maps of a mountainous region. The left panel is a topographic map showing the border between Kyrgyzstan and China. It labels Merzbacher Lake with a red dot, the Kumarik River, and a hydropower station with a red triangle. A legend explains the symbols and provides an elevation scale from 1500 to 5000 meters. The coordinates range from 79°0'E to 79°50'E longitude and 41°20'N to 42°20'N latitude. The right panel is a satellite image focusing on the glacial area. It labels the Northern Inylcheck Glacier, the Southern Inylcheck Glacier, an Upper Lake, and a Lower Lake. The glaciers are shown in white and light blue. The coordinates for this panel range from 79°49'E to 79°56'E longitude and 42°9'N to 42°15'N latitude. Each map includes a scale bar.
+
+Figure 1. The location of lake Merzbacher.
+
+Located in the upper reaches of Kunmaalik River in Aksu Prefecture, Xinjiang Uygur Autonomous Region, Lake Merzbacher's GLOF poses a significant threat to the downstream inhabitants. The lake's primary water source is from the melting of glaciers and snow during the ablation season, starting in early March and ending in September due to the semi-continental climate. This is precisely why glacial lake outburst floods (GLOFs) occur more frequently during this season [35]. To monitor the Lake Merzbacher GLOF, Xiehela hydrographic station was established in 1956, which is downstream of the lake. Since then,
+
+Remote Sens. 2023, 15, 1941 4 of 18
+
+outburst records have occurred almost every year. After extensive research by predecessors, the mechanism behind GLOFs is no longer a secret [36]. When the ice dam of a glacial lake lifts due to a continuous increase in water, the lake enters the drainage process. The concealed water channels beneath the dam open up, connecting the lake water with the outside world, and the lake water is rapidly drained.
+
+## 2.2. Data Sources
+
+Satellite data: All the satellite images were obtained from China Centre for Resources Satellite Data and Application, Beijing, China, (https://data.cresda.cn/#/home, accessed on 24 February 2023), except for Beijing-2 (BJ-2), which was obtained from Twenty First Century Aerospace Technology Co., Ltd., Beijing, China (https://www.21at.com.cn/, accessed on 24 February 2023) (Table 1).
+
+| <b>Table 1.</b> Satellite ima | ges from 2014 to 2022 |
+|-------------------------------|-----------------------|
+|-------------------------------|-----------------------|
+
+| Date           | Sensor    | Resolution | Date                      | Sensor     | Resolution |
+|----------------|-----------|------------|---------------------------|------------|------------|
+| 1 August 2014  | GF-1 WFV1 | 16 m       | 29 July 2021              | GF-1B_PMS  | 2/8 m      |
+| 10 August 2014 | GF-1 WFV3 | 16 m       | 18 May 2022               | HJ-2A_CCD2 | 16 m       |
+| 12 July 2015   | GF-1 WFV3 | 16 m       | 28 May 2022               | HJ-2A_CCD3 | 16 m       |
+| 9 August 2015  | GF-1 WFV1 | 16 m       | 3 June 2022               | HJ-2B_CCD3 | 16 m       |
+| 25 June 2016   | GF-1 WFV4 | 16 m       | 11 June 2022              | GF-1C_PMS  | 2/8 m      |
+| 2 July 2016    | GF-1 WFV1 | 16 m       | 17 June 2022              | HJ-2A_CCD4 | 16 m       |
+| 6 June 2017    | GF-1 WFV2 | 16 m       | 21 June 2022/ 8 July 2022 | HJ-2A_CCD4 | 16 m       |
+| 28 July 2017   | GF-1 WFV4 | 16 m       | 12 July 2022              | GF-1B_PMS  | 2/8 m      |
+| 6 July 2018    | GF-1 WFV2 | 16 m       | 13 July 2022              | GF-1B_PMS  | 2/8 m      |
+| 10 July 2018   | GF-1 WFV1 | 16 m       | 14 July 2022              | GF-1C_PMS  | 2/8 m      |
+| 19 July 2018   | GF-1 WFV3 | 16 m       | 14 July 2022              | HJ-2B_CCD3 | 16 m       |
+| 4 August 2018  | GF-1 WFV2 | 16 m       | 15 July 2022              | GF-1D_PMS  | 2/8 m      |
+| 13 August 2018 | GF-1 WFV4 | 16 m       | 16 July 2022              | HJ2A_CCD3  | 16 m       |
+| 5 July 2019    | GF-1D PMS | 2/8 m      | 17 July 2022              | GF1_WFV2   | 2 m        |
+| 23 July 2019   | GF-1 WFV3 | 16 m       | 18 July 2022              | BJ-2       | 0.8 m      |
+| 5 August 2020  | GF-6 PMS  | 2/8 m      | 20 July 2022              | HJ2A_CCD4  | 16 m       |
+| 3 July 2021    | GF-6 WFV  | 16 m       |                           |            |            |
+
+Gaofen-1 (GF-1): The GF-1 satellite, the first satellite in the China high-resolution earth observation system, was launched on 26 April 2013, and was put into use on 30 December 2013 [37]. Two high-resolution (HR) cameras (2 m resolution panchromatic/8 m resolution multispectral cameras) and four wide field of view (WFV) cameras (16 m resolution multispectral cameras) were equipped with GF-1, which have four spectral channels (blue, green, red) distributed in the visible and NIR spectral domain ranging from 450 to 890 nm (Table 2) [38]. The repetition cycle of GF-1 WFV data can reach four days by combining the four cameras.
+
+**Table 2.** Technical parameter of GaoFen-1 (GF-1) (wide field of view (WFV) and panchromatic/multispectral high resolution (PMS)).
+
+| Payload | Spectral<br>Type                                      | Spectral<br>Range (nm)                              | Spatial<br>Resolution<br>(m) | Swath<br>Width (km)       | Revisit<br>Interval       |
+|---------|-------------------------------------------------------|-----------------------------------------------------|------------------------------|---------------------------|---------------------------|
+| WFV     | Blue<br>Green<br>Red<br>Near infrared                 | 450~520<br>520~590<br>630~690<br>770~890            | 16                           | 800 (Set of<br>4 cameras) | Four days<br>(side-swing) |
+| PMS     | Panchromatic<br>Blue<br>Green<br>Red<br>Near infrared | 450~900<br>450~520<br>520~590<br>630~690<br>770~890 | 2<br>8                       | 60 (Set of<br>2 cameras)  |                           |
+
+Remote Sens. 2023, 15, 1941 5 of 18
+
+Gaofen-6 (GF-6): The GF-6 satellite was launched on 2 June 2018, and was officially put into use on 21 March 2019. GF-6 is a low-orbit optical satellite equipped with a panchromatic/multispectral high-resolution (PMS) camera and a wide field view (WFV) camera (Table 3) [39,40]. This study used the GF-6 WFV and PMS images obtained on 5 August 2020 and 13 July 2021. GF-6/WFV with 16 m resolution has 8 bands, except for one coastal band (400~450 nm) and three visible bands (450~520 nm, 520~590 nm, and 630~690 nm). Same as GF-1, the red-edge I band (0.69~0.73  $\mu$ m), red-edge II band (0.73~0.77  $\mu$ m), a purple band (0.40~0.45  $\mu$ m), and a yellow band (0.59~0.63  $\mu$ m) are newly added bands. GF-6/PMS contains one panchromatic band with 2 m resolution and four multispectral bands with 8 m resolution.
+
+**Table 3.** Technical parameters of GaoFen-6 (GF-6) (wide field of view (WFV) and panchromatic/multispectral high-resolution (PMS)).
+
+| Payload | Spectral<br>Number | Spectral Type | Spectral<br>Range<br>(nm) | Spatial<br>Resolution<br>(m) | Swath<br>Width<br>(km) | Revisit<br>Interval |  |
+|---------|--------------------|---------------|---------------------------|------------------------------|------------------------|---------------------|--|
+|         | 1                  | Blue          | 450~520                   |                              |                        |                     |  |
+|         | 2                  | Green         | Green 520~590             |                              |                        |                     |  |
+|         | 3                  | Red           | 630~690                   |                              | 800                    | Four days           |  |
+| WFV     | 4                  | Near infrared | 770~890                   | 17                           |                        |                     |  |
+|         | 5                  | Costal        | 400~450                   | 16                           |                        |                     |  |
+|         | 6                  | Yellow        | 590~630                   |                              |                        |                     |  |
+|         | 7                  | RedEdge1      | 690~730                   |                              |                        |                     |  |
+|         | 8                  | Rededge2      | 730~770                   |                              |                        |                     |  |
+|         | 1                  | Panchromatic  | 450~900                   | 2                            |                        | _                   |  |
+|         | 2                  | Blue          | 450~520                   |                              |                        |                     |  |
+| PMS     | 3                  | Green         | 520~590                   | 0                            | 90                     |                     |  |
+|         | 4                  | Red           | 630~690                   | 8                            |                        |                     |  |
+|         | 5                  | Near infrared | 770~890                   |                              |                        |                     |  |
+
+Gaofen-7 (GF-7): The GaoFen-7 (GF-7) satellite, launched on 3 November 2019, carries two laser altimeters with the full waveform for global stereo mapping and is the first submeter-level resolution mapping satellite in the Medium- and Long-Term Development Plan for China's Civil Space Infrastructure (2015–2025) [41–43]. The GF-7 satellite operates in a sun-synchronous orbit with a designed lifespan of 8 years. Equipped with two linearray stereo cameras, it can effectively obtain full-color stereo images with a width of 20 km and a resolution better than 0.8 m and multispectral images with a resolution of 3.2 m. The two-beam laser altimeter on board conducts ground observation with a 3 Hz observation frequency, and the ground footprint diameter is less than 30 m. It acquires full-waveform data with a sampling frequency higher than 1 GHz. The laser altimeter system is designed to improve the elevation accuracy without ground control points of the two line-array stereo mapping cameras [44].
+
+HuanjingJianzai-2 A/B (HJ-2A/B): As the successors of HuanjingJianzai-1 (HJ-1) A/B satellites, HuanjingJianzai-2 A/B (HJ-2A/B) are in a sun-synchronous orbit, with both satellites having the same technical status and a design life of 5 years [45,46]. Each satellite has 4 optical payloads, including a 16 m camera, a hyperspectral imager, an infrared camera, and an atmospheric correction instrument. The 16 m camera payload consists of four visible-light CCD cameras, which can provide a multispectral image with a width of 800 km through field-of-view stitching. The hyperspectral imager has a width of 96 km, while the infrared camera has a width of 720 km. For atmospheric radiation correction, the atmospheric correction instrument can synchronously acquire atmospheric multispectral information with the same field of view as the 16 m camera in orbit.
+
+Beijing-2 (BJ-2): The Beijing-2 satellite (BJ-2) was successfully launched from the Satish Dhawan Space Center in Sriharikota, India, at 0:28 Beijing time on 11 July 2015 (16:28 UTC on 10 July 2015), carried by the PSLV-XL carrier rocket to a 651-km sun-synchronous orbit. The Beijing-2 constellation consists of three 0.8 m resolution optical remote sensing
+
+Remote Sens. 2023, 15, 1941 6 of 18
+
+satellites [47]. The satellite, manufactured by Surrey Satellite Technology Ltd. (SSTL) in the UK, has a 7-sided structure, weighs about 450 kg, and is about 2.5 m high. The SSTL-300S1 agile platform can provide a fast-swinging ability of 45 degrees and in-orbit realization of 5 imaging modes, including multi-scene, strip, along-track stereo, cross-track stereo, and regional imaging. The VHRI-100 imaging instrument provides in-orbit panoramic and multispectral images with a swath width of about 24 km and a resolution of 0.8 m (ground sampling distance (GSD)) for panchromatic and 3.2 m for blue, green, red, and near-infrared multispectral images.
+
+2. Climate data: The land surface temperature data used in the article come from the ERA5-LAND reanalysis dataset provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) [48]. ERA5-LAND uses the simulated land-atmosphere variables from the ECMWF's fifth-generation reanalysis product ERA5 as forcing and is obtained using the modified land surface hydrology model HT-ESSEL and CY45R1. Although it has not undergone data assimilation, the observational data indirectly affect its simulation results. Compared with ERA5, ERA5-LAND has a higher spatial resolution, with a horizontal resolution of up to 0.1° (9 km) and a temporal resolution of 1 h. Due to the limitation of currently available data (https://developers.google.com/earth-engine/datasets/catalog/ECMWF\_ERA5\_LAND\_MONTHLY\_BY\_HOUR#description, accessed on 24 February 2023), the hourly land surface temperature data of ERA5-LAND from 1 January 1981 to 1 December 2022 in the study area was used in the study.
+
+## 2.3. Methods
+
+- 1. Processing of the satellite images and climate data: Preprocessing methods are shared in multispectral data of GF-1, GF-6, HJ-2 A/B, and BJ-2, including five steps: (1) Radiometric calibration; (2) Atmospheric correction; (3) Orthorectification; (4) Image fusion; (5) Image registration. The radiometric calibration and atmospheric correction were performed with the ENVI FLAASH module [44]. After finishing the preprocessing, combining 2 m resolution panchromatic and 8 m resolution multispectral data by the ENVI Gram-Schmidt Pan-Sharpening module to enhance the spatial resolution of GF-1, GF-6, and BJ-2 images [45]. DSM with a 2 m spatial resolution was derived from GF-7 by using the rational polynomial coefficient (RPC) model [41]. Time series temperature data were extracted by Google Earth Engine (GEE) cloud platform (https://earthengine.google.com/, accessed on 24 February 2023).
+- 2. Lake area extraction and changes analysis. The visual interpretation method was applied in this study. To improve the interpretation accuracy, the normalized difference water index (NDWI) [47] was introduced to better distinguish water bodies (Figure 2). The NDWI can enhance the identification of water bodies and effectively distinguish water bodies from floating ice. By using the green channel (maximum reflectance of water) and the near-infrared (NIR) channel (minimum reflectance of water), the NDWI was calculated as below:
+
+$$NDWI = \frac{\rho Green - \rho NIR}{\rho Green + \rho NIR}$$
+ (1)
+
+Lake area changes, including water and ice area changes, were calculated in ArcGIS 10.8. To characterize the direction of water surface changes, the centroid shift method was used to describe the expansion process of the lake surface (Figures S1 and S2).
+
+Remote Sens. 2023, 15, 1941 7 of 18
+
+Image /page/6/Figure/1 description: A side-by-side comparison of two satellite images, labeled 'a' and 'b', of a glacial lake in a mountainous region, both dated 8 July 2022. Image 'a' shows the natural view of the dark blue lake with floating ice, surrounded by brown mountains and glaciers. Image 'b' overlays a Normalized Difference Water Index (NDWI) map on the same view. A legend for the NDWI shows a color scale from blue (Low: 0.31) to red (High: 0.58). The map displays lower NDWI values in blue over the main body of the lake and higher values in yellow and orange where the glacier meets the water. Both images include a compass rose and a scale bar indicating 1 km.
+
+Figure 2. Satellite image (a) and image after calculating the NDWI (b).
+
+3. Uncertainty assessment of glacier lake area. Visual interpretation was employed in extracting the lake area, and errors are unavoidable. Research has shown that mixed pixels caused by spatial resolution are a key factor in error sources. Using an error of one pixel on either side of the defined lake boundary is more appropriate [49]. Therefore, the uncertainty of the individual lake area can be calculated as follows:
+
+$$e = n^{1/2} \times m \tag{2}$$
+
+$$R = \frac{e}{A} \times 100\% \tag{3}$$
+
+where e is the absolute area error (m<sup>2</sup>) of each glacier lake, n represents the number of pixels on the lake boundary (approximately the ratio of lake circumference to spatial resolution), m is the area of each pixel in the remote sensing product, R is the relative error of a single lake, and R is the area of the lake. The area error obtained from the above equation shows that the total absolute area error of Lake Merzbacher is 0.08 km<sup>2</sup>, the average relative error is 0.22%, and the relative area error is between 0.2% and 0.68% (Table 4).
+
+| Date | Area<br>(km2) | Perimeter<br>(m) | Absolute<br>Area Error | Relative<br>Area Error | Date | Area<br>(km2) | Perimeter<br>(m) | Absolute<br>Area Error | Relative<br>Area Error |
+|------|---------------|------------------|------------------------|------------------------|------|---------------|------------------|------------------------|------------------------|
+|------|---------------|------------------|------------------------|------------------------|------|---------------|------------------|------------------------|------------------------|
+
+| Date          | Area<br>(km2) | Perimeter (m) | Absolute<br>Area Error<br>(km2) | Relative<br>Area Error<br>(%) | Date         | Area<br>(km2) | Perimeter (m) | Absolute<br>Area Error<br>(km2) | Relative<br>Area Error<br>(%) |
+|---------------|---------------|---------------|---------------------------------|-------------------------------|--------------|---------------|---------------|---------------------------------|-------------------------------|
+| 1 August 2014 | 2.99          | 9.94          | 0.01                            | 0.21                          | 11 June 2022 | 1.25          | 8.52          | 0                               | 0.02                          |
+| 12 July 2015  | 2.28          | 9.57          | 0.01                            | 0.27                          | 17 June 2022 | 1.45          | 7.92          | 0.01                            | 0.39                          |
+| 25 June 2016  | 1.94          | 8.82          | 0.01                            | 0.31                          | 21 June 2022 | 1.52          | 7.54          | 0.01                            | 0.37                          |
+| 6 July 2017   | 1.8           | 8.74          | 0                               | 0.01                          | 8 July 2022  | 2.07          | 8.91          | 0.01                            | 0.29                          |
+| 3 August 2018 | 2.5           | 10.15         | 0.01                            | 0.26                          | 12 July 2022 | 2.15          | 10.03         | 0                               | 0.01                          |
+| 5 July 2019   | 1.86          | 8.61          | 0                               | 0.01                          | 13 July 2022 | 2.1           | 9.91          | 0                               | 0.01                          |
+| 5 August 2020 | 2.05          | 8.25          | 0                               | 0.01                          | 14 July 2022 | 2.09          | 9.76          | 0.01                            | 0.30                          |
+| 3 July 2021   | 1.96          | 8.70          | 0.01                            | 0.30                          | 15 July 2022 | 2.08          | 9.85          | 0                               | 0.01                          |
+| 18 May 2022   | 0.72          | 5.84          | 0                               | 0.68                          | 16 July 2022 | 2.03          | 9.27          | 0.01                            | 0.30                          |
+| 28 May 2022   | 0.95          | 6.71          | 0.01                            | 0.55                          | 17 July 2022 | 2.01          | 9.12          | 0                               | 0.01                          |
+| 3 June 2022   | 1.06          | 6.82          | 0.01                            | 0.50                          | 18 July 2022 | 1.28          | 5.26          | 0                               | 0                             |
+
+# 3. Results
+
+## 3.1. Maximum Lake Area (MLA) Change before Outburst from 2014 to 2022
+
+Previous research has shown that the area of the lake continues to expand until the outburst occurs, and once the lake surface area starts to decrease, the GLOF event is imminent. Therefore, monitoring changes in the lake area prior to an outburst is crucial. Analysis of satellite images during the research period revealed a decreasing trend in the
+
+Remote Sens. 2023, 15, 1941 8 of 18
+
+maximum lake area (MLA) from 2014 to 2021 (Figure 3b). Over the past nine years, the average MLA was 2.07 km<sup>2</sup>. The largest MLA was 2.99 km<sup>2</sup> on 1 August 2014 (Figure 3a), and the smallest MLA was 1.67 km<sup>2</sup> on 12 July 2016 (Figure 3d). From 2014 to 2022, the MLA occurred in July, except in 2014 and 2018 two years, which occurred in August. From the perspective of the average elevation of the lake, which shows a similar change trend as the MLA change from 2014 to 2022 (Figure 3b). The highest and the lowest average elevation of the MLA were observed in 2014 (3318.55 m) and 2019 (3289.97 m), respectively.
+
+Image /page/7/Figure/2 description: A multi-panel figure showing the evolution of a glacial lake from 2014 to 2022. The figure is composed of ten panels labeled (a) through (j). Panel (a) is a large satellite image from October 1, 2014, showing the lake with colored outlines indicating its extent for each year from 2014 to 2022. A legend details the colors for each year: W2022 (orange), W2021 (dark blue), W2020 (green), W2019 (cyan), W2018 (blue), W2017 (magenta), W2016 (light green), W2015 (yellow), and W2014 (red). Panel (b) is a dual-axis line graph for the years 2014-2022, plotting Elevation (m) on the left y-axis (ranging from 3270 to 3330) and Area (km²) on the right y-axis (ranging from 0 to 3.5). The orange line represents elevation, and the blue line represents area. Panels (c) through (j) are smaller satellite images of the lake on different dates: (c) 2015 12 July, (d) 2016 25 June, (e) 2017 06 July, (f) 2018 04 October, (g) 2019 09 July, (h) 2020 05 October, (i) 2021 05 July, and (j) 2022 15 July, showing variations in the lake's color, ice cover, and clarity over the years.
+
+**Figure 3.** Maximum lake area (MLA) change before outburst since 2014. (a) MLA extent in each year and overlay on the image obtained on 1 August 2014; (b) MLA changes from 2014 to 2022; (c–j) MLA overlay on images obtained from 2015 to 2022.
+
+Based on the observation of Lake Merzbacher GLOF in 2022, which started at least from 18 May and lasted for about two months (Figure 4). Initially, water coalesces into the southern part of the lake by gravity with an area of  $0.72~\rm km^2$ . With the lake increasingly receiving meltwater from snow and ice, the lake area continues to increase. From 18 May to 28 May, the lake area increased at a rate of  $0.023~\rm km^2/d$ . During the next 11 days (28 May~8 June), the lake area increased at a relatively low speed compared to the first ten days, with a rate of  $0.01~\rm km^2/d$ . After that, the lake expanded rapidly from 8 June to 17 June, with a rate of  $0.04~\rm km^2/d$ . For some time after that (17 June~8 July), the lake area increased rapidly, with a speed of  $0.03~\rm km^2/d$ . The lake area reached its maximum in 12 July with a rate of  $0.03~\rm km^2/d$  from 08 July to 12 July. However, the lake area decreased from 2.19 km² on 12 July to 2.12 km² on 14 July. In the next four days (15 July~18 July), the lake area continued to shrink with an average rate of  $0.21~\rm km^2/d$ , especially during the period from 17 July to 18 July, and the lake area decreased rapidly from 2.01 km² to  $1.28~\rm km^2$ .
+
+Remote Sens. 2023, 15, 1941 9 of 18
+
+Image /page/8/Figure/1 description: This figure, titled "Figure 4. Lake extent change in 2022 before and after the outburst," displays a series of satellite images, a map, and a line graph to illustrate the changing size of a lake. The main part of the figure is a grid of 14 satellite images dated from May 18, 2022, to July 20, 2022. These images show a lake in a mountainous, snowy area, with its extent outlined. The lake's size is observed to increase progressively until mid-July, after which it appears to have drained. In the bottom right, there is a map with a legend showing the superimposed outlines of the lake on various dates, along with a scale bar. Next to the map is a line graph plotting the lake's area in square kilometers against the date. The y-axis, labeled "Area(km²)," ranges from 0 to 2.5. The x-axis, labeled "Date," shows dates from May 18 (0518) to July 18 (0718). The graph shows the area increasing from approximately 0.7 km² on May 18 to a peak of about 2.2 km² on July 14, before sharply declining to around 1.2 km² by July 18, visually representing the outburst event.
+
+Figure 4. Lake extent change in 2022 before and after the outburst.
+
+The centroid changes of the lake area from 2014 to 2022 are presented in Figure 5. The centroid of the maximum lake area in the nine-year period was mostly located along the central line in a direction from the northeast to the southwest. When using the southwestern boundary as a reference frame, the farthest centroid of the maximum lake area was observed in 2018, measuring 1.2 km from the centroid point to the southernmost boundary. Second to 2018, the centroid in 2014 was the second farthest centroid of the maximum lake area during the research period. Simultaneously, the two largest lake areas occurred in the above two years. Notably, even in the years with the largest lake area, the centroid of the maximum lake area remained within a certain distance from the southernmost boundary. Based on the dense records in 2022 prior to the outburst, from 18 May to 28 May, the lake area expanded 0.2 km toward the northeast. In the following 14 days (28 May to 11 June), it continued to expand approximately 0.2 km to the northeast, and from 11 June to 21 June, it continued to expand northeastward with a distance of about 0.1 km. Before reaching its maximum, the lake area expanded approximately 0.3 km northeastward. Subsequently, the lake area began to shrink back toward the southwestern boundary. For the first few days (12 July to 15 July), the centroid of the lake area had little change. However, since 15 July, a significant retreat of the centroid of the lake area was observed. The most significant retreat occurred in a 24 h period from 17 July to 18 July.
+
+Remote Sens. 2023, 15, 1941 10 of 18
+
+Image /page/9/Figure/1 description: A figure with two side-by-side maps showing changes in a geographical feature, likely a glacier, over time. Each map has an inset scatter plot. Both maps feature a scale bar from 0 to 1 km and a compass rose.
+
+The left panel displays annual changes from 2014 to 2022. The map shows the glacier with colored outlines and corresponding dots for each year. The legend indicates the colors: 2014 (red), 2015 (yellow), 2016 (green), 2017 (purple), 2018 (blue), 2019 (cyan), 2020 (pink), 2021 (dark blue), and 2022 (brown). The inset scatter plot shows the position of a point over these years, with the X-axis from 79.842 to 79.847 and the Y-axis from 42.2015 to 42.2055.
+
+The right panel displays changes over specific dates, likely within a single year. The map shows the glacier with colored outlines and dots for dates like 0518, 0528, 0608, 0611, 0617, 0621, 0708, 0712, 0714, 0715, 0716, 0717, and 0718, each with a unique color as defined in the legend. The inset scatter plot shows the position of a point on these dates, with the X-axis from 79.84 to 79.845 and the Y-axis from 42.197 to 42.205.
+
+Figure 5. Centroid changes of lake area in historical and 2022 before the outburst.
+
+3.2. Using Lake Area and Ice Cover to Monitor Hazard of Glacier Lake Outburst Flood of Lake Merzbacher
+
+Previous research has divided the entire outburst process of Lake Merzbacher into four stages: icefall (Stage I), quick water storage (Stage II), early warning (Stage III), and post-drainage (Stage IV). Therefore, the ice cover can be an essential index to predict the outburst. By setting the NDWI threshold value, we classified the lake as water and an ice area, two classes. As depicted in Figure 6, the ice area accounted for 53.3% of the total lake area on average from 2014 to 2022. The ice area proportion is the maximum (69.87%) in 2021, while that of 2016 is the minimum (23.83%). According to the records of the outburst date from 2014 to 2021, Lake Merzbacher outburst on 10 October, 9 July 2015, 25 June 2016~2 July 2016, 7 July 2017, 12 October 2018, 9 July 2019~12 July 2019, 7 October 2020, and 9 July 2021. During the first stage, the ice area is generally higher than the water area. In the second stage, the water area will expand rapidly, and the ice at the bottom of the lake will emerge from the water. Thus, the ice area will expand simultaneously before the outburst. When the dam body is raised, the lake area will shrink, and more floating ice will emerge, which indicates the lake has entered an early warning stage. As shown in Figure 6, water areas in 2014, 2015, 2018, 2019, and 2021 are smaller than ice areas, indicating that these images may have been captured during Stage II or Stage III.
+
+Remote Sens. 2023, 15, 1941 11 of 18
+
+Image /page/10/Figure/1 description: A stacked bar chart, labeled (a), showing the ice area and water area in square kilometers from 2014 to 2022. The x-axis represents the year, and the y-axis represents the area in square kilometers (km²), ranging from 0 to 3.5. The legend indicates that light blue bars represent 'Ice area' and dark blue bars represent 'Water area'. The data for each year is approximately as follows: 2014: Ice area ~1.7 km², Water area ~1.3 km², Total ~3.0 km². 2015: Ice area ~1.45 km², Water area ~0.8 km², Total ~2.25 km². 2016: Ice area ~0.45 km², Water area ~1.5 km², Total ~1.95 km². 2017: Ice area ~0.85 km², Water area ~0.95 km², Total ~1.8 km². 2018: Ice area ~1.55 km², Water area ~0.95 km², Total ~2.5 km². 2019: Ice area ~1.25 km², Water area ~0.6 km², Total ~1.85 km². 2020: Ice area ~0.9 km², Water area ~1.15 km², Total ~2.05 km². 2021: Ice area ~1.35 km², Water area ~0.6 km², Total ~1.95 km². 2022: Ice area ~1.0 km², Water area ~1.25 km², Total ~2.25 km².
+
+Image /page/10/Figure/2 description: A combination chart, labeled (b), showing changes in water and ice area over time. The horizontal axis is labeled 'Date' and shows dates from 0518 to 0717. The chart has two vertical axes, both labeled 'Area(km²)'. The left y-axis ranges from 0 to 2.5 and corresponds to a stacked bar chart. The right y-axis ranges from -0.4 to 0.6 and corresponds to a line chart. The legend indicates that dark blue bars represent 'Water area', light blue bars represent 'Ice area', and a red line represents 'W-I' (Water area minus Ice area). The stacked bars show a general increase in total area over the period, from approximately 0.7 km² on 0518 to a peak of about 2.1 km² on 0714. The red line fluctuates, starting near -0.2, dropping to a low of about -0.35 on 0621, and reaching a peak of approximately 0.55 on 0714.
+
+**Figure 6.** Water area and Ice area of Lake Merzbacher before the outburst from 2014 to 2022 (a); Water area and Ice area daily changes before the outburst (b).
+
+Time series images captured in 2022 before and after the outburst provide more details. At the beginning of the observation period, the lake area was mainly covered by ice on 18 May. As the ice melted, the water area gradually increased and exceeded the ice area on 28 May and 3 June. During the storage process of the glacial lake, the ice originally deposited at the bottom of the lake basin was lifted by the water body and floated on the surface of the lake, resulting in ice dominating the lake area again on 17 June and 21 June. As the meltwater continued to merge, the water area surpassed the ice area again until reaching the maximum lake area on 8 July to 12 July. As the ice dam body rose from 12 July to 17 July, the gap between the water and ice area gradually increased and narrowed. During this period, ice deposited at the bottom of the ice dam was lifted by the water body, and the lake was dominated by ice again before the outburst.
+
+## 3.3. The Relationship between the Temperature and the Outburst Date since 1980
+
+According to the relevant records of the published studies and news reports, Lake Merzbacher has experienced nearly yearly outbursts since 1902 (Figure 7). From the perspective of the outburst date, it can be found that over 80% of outburst events occurred on 25 June, 28 July, and 14 October. From 1981 to 2022, almost every outburst event occurred on 23 July and 16 October, except in 1988 and 1996. Based on ERA5 climate data, the hottest month in every year since 1981 was extracted (Figure 7). As depicted in Figure 7, the hottest month occurred on 24 July or 20 October since 1981. From 1981 to 2022, the hottest month and the outburst date were the same in 25 years. Especially from 2015 to 2022, Lake Merzbacher outburst occurred in the hottest month every year. Lake Merzbacher outburst occurred in the second hottest month in 11 years, of which 5 years is one month ahead of the hottest month, while the rest of the 6 years are one month later than the hottest month. Only two outburst events occurred in the cold season. The main water storage source of Lake Merzbacher is the meltwater of North Inylchek Glacier, so the water storage of the glacial lake strongly depends on air temperature.
+
+Further analysis of the outburst time reveals that since the 1930s, the outbursts of Lake Merzbacher have tended to occur earlier in the year (Figure 8). Over the period from 1902 to 2022, the Lake outburst earlier on average 6 days per decade ( $R^2 = 0.21$ ). Since the water storage of Lake Merzbacher comes from glacial meltwater, one potential reason for this fact can be climate warming, which is melting glaciers earlier. Climate change has raised average temperatures in the surrounding area of Lake Merzbacher (Figure 8). From 1981 to 2021, the average temperature increased from  $-9.59\,^{\circ}\text{C}$  to  $-9.06\,^{\circ}\text{C}$ , increasing at a rate of 0.18  $^{\circ}\text{C}$  per decade. From the perspective of April, May, June, July, and October, the average temperature increased during the same period. Among them, the fastest growth in
+
+Remote Sens. 2023, 15, 1941 12 of 18
+
+average temperature has been seen in April, which increased at a rate of  $0.45\,^\circ\text{C}$  per decade. In other words, Lake Merzbacher has shown an early warming trend.
+
+Image /page/11/Figure/2 description: The image displays three charts, labeled (a), (b), and (c), analyzing outburst data, likely related to Lake Merzbacher as mentioned in the text above the charts. 
+
+Chart (a) is a histogram showing the frequency of outbursts by month. The x-axis is labeled 'Month' with ranges: [5, 6], (6, 7], (7, 8], (8, 9], (9, 10], (10, 11], and (11, 12]. The y-axis is 'Frequency', from 0 to 30. The frequencies for each month range are as follows: [5, 6] is 3, (6, 7] is 25, (7, 8] is 28, (8, 9] is 14, (9, 10] is 7, (10, 11] is 1, and (11, 12] is 3. The highest frequency of outbursts occurs between months 6 and 8.
+
+Chart (b) is a bar chart showing the 'Outburst date' (month) over time. The x-axis is 'Year', from 1981 to 2020. The y-axis is 'Outburst date', representing the month from 0 to 14. Most outbursts occur in month 7 or 8. There are notable peaks where the outburst occurred in month 12, specifically around 1994 and 2016.
+
+Chart (c) is a grouped bar chart comparing the 'Outburst Date' (blue bars) with the 'Hottest Month' (orange bars) for each year from 1981 to 2021. The x-axis is 'Year', and the y-axis is 'Month', from 0 to 14. In most years, the outburst date and the hottest month are closely aligned, typically around month 7 or 8. However, in some years, such as 1988 and 1996, the outburst date (month 12) is significantly later than the hottest month (month 7).
+
+**Figure 7.** Date of Lake Merzbacher outburst since 1902. (a) Lake Merzbacher outburst Frequency in different months since 1902; (b) Lake Merzbacher outburst date since 1902; (c) The relationship between the outburst date and the hottest date since 1902.
+
+Image /page/11/Figure/4 description: The image displays two graphs, labeled (a) and (b).
+
+Graph (a) is a scatter plot titled "Outburst date since 1902". The x-axis is labeled "Year" and ranges from 1900 to beyond 2000. The y-axis is labeled "Days" and ranges from 100 to 350. The plot contains numerous blue circular data points. A dotted blue trendline is drawn through the data, showing a negative slope. The equation for the trendline is given as "y = -0.6397x + 1493.3", and the R-squared value is "R² = 0.21".
+
+Graph (b) is a line graph showing temperature data over time. The x-axis is labeled "Year" and ranges from 1981 to 2020. There are two y-axes, both labeled "Temperature (°C)". The left axis ranges from 0 to 5, and the right axis ranges from -15 to 5. The graph contains five colored lines representing different months: July (yellow), April (dark blue), June (gray), October (light blue), and May (orange). A dotted gray line represents the overall "Trendline". The equation for this trendline is "y = 0.0448x - 10.191", and the R-squared value is "R² = 0.1343".
+
+**Figure 8.** Detailed records of outburst date from Lake Merzbacher from 1902 to 2022 (a). Temperature changes in Lake Merzbacher from 1981 to 2022 (b).
+
+Remote Sens. 2023, 15, 1941 13 of 18
+
+# 4. Discussion
+
+The glacier lake outburst flood (GLOF) significantly threatens downstream communities, infrastructure, and livelihoods [50–52]. Establishing an early warning system is crucial in mitigating disasters and reducing damage [53]. GLOFs release large amounts of water stored in glacier lakes rapidly downstream [54]. Therefore, capturing the signals of an imminent GLOF is essential for early warning. Previous studies have shown Lake Merzbacher outbursts at a high frequency, especially after entering the twenty century. Additionally, there is a certain regularity in Lake Merzbacher's outburst mechanism [19,20]. The lake area expands before reaching its maximum, after which it rapidly shrinks until the outburst. This study analyzed the lake's daily changes in area leading up to the 2022 outburst. To the best of our knowledge, this is the first study to observe these changes daily before a Lake Merzbacher outburst.
+
+When analyzing the maximum lake area (MLA) before the outburst of Lake Merzbacher from 2014 to 2022, it was found that the MLA did not exceed 3 km<sup>2</sup> and showed a decreasing trend, which is consistent with the findings of Li et al. [30]. In fact, the MLA has shown a decreasing trend, at least after entering the new century. On the contrary, glacier lakes showed an increasing trend in the Tienshan region reported by Wang et al. [55]. The upper lake area also showed an increasing trend [30]. The lake area expansion or shrinkage are all the consequences of climate change. On the one hand, climate change can lead to glacial ablation and thus provide more melting water for the glacier lakes. Therefore, the quantities and volume of glacier lakes showed an increasing trend. However, Lake Merzbacher was dammed by the southern branch of the Inylchek Glacier. The increase in temperature will also accelerate the piping inside the ice dam [56]. As a result, the outbursts tended to occur earlier (Figure 8). However, this does not mean Lake Merzbacher is not dangerous anymore. Under the effects of climate change, glaciers in the Tien Shan of Central Asia have generally retreated over the past half-century [57]. A general increasing trend has prevailed in the temporal variation of peak and total discharges of the jokulhlaup floods of the lake [49,58]. Furthermore, the southern branch of the Inylchek Glacier proceeds northward and flows into Lake Merzbacher, which may break the stability of the ice dam and thus significantly increase the GLOF risk for Lake Merzbacher. Therefore, the outburst of Lake Merzbacher requires our ongoing attention.
+
+A previous study divided the entire process of the Lake Merzbacher outburst into four periods: icefall, quick water storage, early warning, and post-drainage [18]. According to satellite images in 2022, since at least 4 May, water starts to be injected into the lake (Figure 9). As the temperature keeps increasing, more and more glacier meltwater flows into Lake Merzbacher. The water area would expand at a relatively low speed during this period. This process continues for up to about one month, the same as in previous records. In the next month, the lake area would expand at a relatively high speed until reaching the maximum lake area. As the dam body gradually rises under the water pressure [20], a system of englacial channels hidden in the dam opens, and the lake water discharges through these channels. During this period, the lake area showed a relatively low decreasing speed and can be detected from some corners (S1–S3). This period lasts about 5 days and is significant to early warning. Furthermore, the entire outburst process of Lake Merzbacher will repeat nearly every year [16]. It means that Lake Merzbacher, because of its predictability, would be a promising candidate for an early warning system (EWS). Compared to other engineering measures on glacial lakes, establishing the EWS is a relatively inexpensive option. Generally, the GLOF EWS has been widely established in the HMA [58]. For example, a manual EWS was set up in 1997 by Ropal Tsho, a moraine-dammed located in the Rolwaling Valley of Nepal [49]. The manual EWS by Ropal Tsho evolved into a fully automatic system one year after its establishment, which constituted two components: a sensing system and a warning system [59]. Imja Tsho is another glacial lake located in the Solukhumbu District of Nepal, near the base of Mount Everest [60]. It is a glacial lake that formed as a result of the retreat of the Imja Glacier. The lake has grown rapidly in recent decades due to climate change, and there are
+
+Remote Sens. 2023, 15, 1941 14 of 18
+
+concerns that it could breach its natural dam and cause flooding downstream. The EWS of Imja Tsho combined two monitoring stations (web camera) and two relay stations [56]. EWSs can be complicated or easy. In theory, only the installation of water-level sensors can achieve a simple prediction scheme. However, a comprehensive monitoring system, usually composed of glacial lake monitoring, parent glaciers monitoring, the downstream river environment change detection, formulating a disaster mitigation plan, releasing early-warning information, and training for evacuation, is of great significance to reveal the GLOF generation mechanism [53]. Therefore, though the advantage of dense satellite images has been demonstrated, field-based EWS components are still required.
+
+Image /page/13/Picture/2 description: A low-resolution satellite image of a mountainous, snowy landscape. The image features a mix of white and light blue areas, representing snow or ice, and darker, rocky terrain. A small, dark blue body of water is visible near the center. In the top-left corner, there is a compass rose. In the bottom-left corner, a scale bar indicates a distance from 0 to 1 km. The date "4 May 2022" is written in the bottom-right corner.
+
+Figure 9. Satellite image of Lake Merzbacher captured in 4 May 2022.
+
+The difficulty in using satellite images to provide warnings of outburst floods from glacier-dammed lakes is to identify the drainage date [61]. The proportion of ice and water could also provide important information during the process. Xie et al. [16] proposed an index by calculating the ratio of floating ice and total lake area change rates. They found that when this index is less than 0.5 and the lake area is larger than 3 km<sup>2</sup>, the outburst will occur in the next 5–8 days. It is still a rough calculation. This study also calculates the proportion of ice and water before the outburst. As depicted in Figure 6, W-I shows regular fluctuations. Especially in the early warning stage, we found that the W-I showed an obvious decreasing trend, which indicates that more and more ice arose over the lake's surface. The ice and water areas are nearly the same size one day before the outburst. From the view of the satellite image, the lake area was almost dominated by ice. With the channel opening, water storage in the lake begins to flow downstream, and the floating ice over the lake's surface moves to the dam. Therefore, we can also speculate on the status of the lake according to the proportion of floating ice. Historical images from 2014 to 2021 can also help verify the assumption. According to the accurate records of outburst dates in 2020 and 2021, only two or four days are left before the outburst. As is shown in Figure 3, the floating ice almost bestrews the lake's surface in 2020 and 2021. It is speculated that the corresponding date of images in 2018 and 2019 are also near the outburst date. On the other hand, field-based EWS components may improve the timeliness and accuracy of early warning greatly.
+
+The uncertainty and limitations in this research can be summarized below. As mentioned before, this is the first research to fully record the whole outburst process using
+
+Remote Sens. 2023, 15, 1941 15 of 18
+
+high-resolution images. However, it is a pity that we failed to obtain an effective image on 19 July 2022 because of the cloudy weather. It is on this day the outburst occurred. Synthetic aperture radar (SAR) is not vulnerable to weather conditions and can be applied as an effective reinforcement [62]. Even though we frequently monitor the lake area, hydrographic station data are significant for early warning. As the lake water flows out, the downstream hydrographic station would also witness the process of rapid increase followed by a rapid decrease of water [17]. Therefore, combining hydrographic station records or other field-based EWS components and dense satellite images can achieve accurate surveillance.
+
+# 5. Conclusions
+
+This study demonstrates the potential of using dense Chinese high-resolution satellite images to monitor glacier lake outburst flood (GLOF) and release early warning. Based on the dense high-resolution satellite images, this study monitored the whole process of the Lake Merzbacher GLOF in 2022 and further analyzed historical GLOF events in Lake Merzbacher. Our results show that the lake area would expand slowly (0.01 km<sup>2</sup>/d) and then rapidly expand  $(0.04 \text{ km}^2/\text{d})$  until it reaches the maximum lake area. After entering the early warning stage, which would last five days, the lake area would decrease slowly until the outburst. Further analysis shows that the GLOF is not far away when the floating ice area accounts for 50% or larger than 50% of the total lake area. The Lake Merzbacher GLOF is closely related to the temperature, and the outburst mainly occurred in the hottest months (July and October), especially in the last ten years. With the average temperature increase under climate change, Lake Merzbacher outbursted earlier and earlier (6 days earlier per decade since 1902). Our study provides essential references for monitoring Lake Merzbacher and establishing the early warning system (EWS). We also suggest that adding some field observation equipment, including a meteorological station and a hydrologic station, can improve the timeliness and accuracy of early warning.
+
+**Supplementary Materials:** The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/rs15071941/s1, Figure S1: Changes in the lake area were observed at a regional scale; Figure S2: Changes in the connection between the upper and lower parts of the lake.
+
+**Author Contributions:** Conceptualization, C.G. and M.L.; methodology, C.G.; software, C.G.; validation, M.L., K.H. and P.W.; formal analysis, C.G.; investigation, C.G.; resources, S.L.; data curation, C.G.; writing—original draft preparation, C.G.; writing—review and editing, S.L.; visualization, K.H.; supervision, S.L.; project administration, S.L.; funding acquisition, S.L. All authors have read and agreed to the published version of the manuscript.
+
+**Funding:** This research was funded by the National Key Research and Development Program of China, grant number 2019YFE0127700; the National Key Research and Development Program of China, grant number 2021YFB3901200; and the China High-Resolution Earth Observation System, grant number 03-Y30F03-9001-20/22.
+
+**Data Availability Statement:** Data used in this study will be available upon request from the first author.
+
+Conflicts of Interest: The authors declare no conflict of interest.
+

chatbot/rag/data/('Predicting glacial lake outburst susceptibility on the southern Tibetan Plateau with historical events and machine learning methods', '.pdf')_extraction.md

@@ -0,0 +1,413 @@
+#### **ORIGINAL PAPER**
+
+
+# Abstract
+
+Glacial Lake Outburst Floods (GLOFs) have inflicted varying degrees of damage on the ecological environment, infrastructure, and human life in the alpine regions. Consequently, effectively predicting GLOFs has emerged as a critical research focus for disaster prevention and mitigation. This study focuses on the southern Tibetan Plateau and systematically examines the distribution characteristics of glaciers and glacial lakes, which are key contributors to GLOFs, and reviews historical GLOF events and their developmental patterns using remote sensing imagery, geographic information systems (GIS) and machine learning techniques. Based on historical GLOF data, a susceptibility evaluation system was developed by integrating Support Vector Machine (SVM) and Multi-Layer Perceptron (MLP) algorithms. This system quantitatively assesses the susceptibility of each glacial lake within the study area. The findings indicate that the majority of historical GLOF events occurred between June and September, with triggering factors including icefalls, moraine dam destabilization, and heavy precipitation. Climate warming, particularly during the 1960s and from 1990 to 2020, has significantly influenced the frequency of GLOF events, revealing substantial spatial heterogeneity, glacial dependence, and climate sensitivity of GLOFs. The evaluation results show that approximately 17.4% of the glacial lakes are situated in high and very high susceptibility classes, with the highest susceptibility observed in the central and eastern sub-regions. The MLP model demonstrated slightly higher accuracy than the SVM model, with AUC (Area Under the Receiver Operating Characteristic Curve) values of 0.96 and 0.90, respectively. This study offers a novel methodology and technical support for assessing the risk of glacial lake outbursts in the Tibetan Plateau and similar alpine mountain regions, providing a scientific basis for the development of disaster prevention and mitigation strategies.
+
+**Keywords** Glacial lake outburst flood · Spatial pattern · Susceptibility evaluation · Machine learning method · Tibetan plateau
+
+Image /page/0/Picture/10 description: The Springer logo, featuring a black and white icon of a chess knight to the left of the word 'Springer' in a black serif font, all on a white background.
+
+Extended author information available on the last page of the article
+
+# 1 Introduction
+
+Glacial Lake Outburst Floods (GLOFs) are catastrophic natural hazards caused by the sudden release of water from glacial lakes, often resulting in large-scale floods or debris flows. These events pose significant threats to downstream ecosystems, infrastructure, economic development, and human life (Ahmed et al. 2021; Chen et al. 2024). The Tibetan Plateau is one of the world's most densely populated regions with glaciers and glacial lakes, especially in its southern part. The complex geological structures and extreme climatic conditions in this area contribute to a high frequency of GLOFs (Harrison et al. 2018; Liu et al. 2014). In the context of global warming, the Tibetan Plateau and other alpine mountain regions have experienced severe GLOFs that have caused extensive damage to infrastructure and the ecological environment (Tang et al. 2023). Notable historical events include the 1954 Sangwang Glacial Lake outburst in Kangma County, Tibet, which resulted in approximately 400 fatalities and affected over 20,000 people downstream (Liu et al. 2019); the 1988 Guangxie Glacial Lake outburst in Bomi County, destroying 51 houses, causing five deaths and the closure of the Sichuan-Tibet Highway for over 200 days (Liu et al. 2019); the 2016 Gongbashatong Glacial Lake outburst in Nyalamu County, which flooded Zhangmu Port and caused extensive damage to infrastructure in downstream areas of Nepal, including the towns of Kodari and Tatopani; and the 2020 Jinweng Glacial Lake outburst in Nagchu City, causing significant damage to water conservancy and transportation infrastructure downstream (Yang et al. 2022). In recent years, as major engineering projects expand into alpine regions of the Tibetan Plateau, the potential threats posed by GLOFs to infrastructure, as well as their mutual feedback effects, have become increasingly prominent. Therefore, accurately predicting the likelihood of future GLOFs is crucial for developing effective disaster prevention and mitigation strategies.
+
+The formation of GLOFs is influenced by multiple factors, including glacier retreat, the morphological characteristics of glacial lakes, climate change, dynamic processes of glacier-glacial lake interactions, and geotectonic activities (Cook et al. 2018; Mou et al. 2024; Pandey et al. 2022). The interplay of these factors increases the complexity and uncertainty associated with predicting GLOFs (Shi et al. 2024; Wang et al. 2025; Westoby et al. 2014; Worni et al. 2014). Wangchuk et al. (2022) utilized Sentinel-1 SAR radar backscatter intensity data to monitor changes in glacier and lake areas and analyze the susceptibility of GLOFs. Aggarwal et al. (2016) utilized multispectral remote sensing imagery in conjunction with parameters such as lake area, distance to the parent glacier, slope, and distance from the basin outlet to identify 18 lakes with potential outburst risk. Emmer and Vilímek (2014) used decision trees and numerical calculations based on remote sensing imagery and digital terrain models to assess the susceptibility of GLOFs, providing effective monitoring tools for different GLOF scenarios. However, spatial and temporal prediction of GLOFs remains a significant challenge, particularly in the Tibetan Plateau region, characterized by high altitudes, variable climate, and complex topography (Zhang et al. 2021a; Zheng et al. 2021). Current approaches to evaluating GLOF susceptibility can be broadly categorized into statistical models and physical models. Statistical models focus on predicting the likelihood of future GLOFs by quantifying the relationships between historical events and contributing factors through statistical analysis (Chen et al. 2024; Das et al. 2024). Physical models rely on detailed topographic, hydrological, and glacial lake parameter data, employing numerical simulations to estimate the potential extent and magnitude of outburst
+
+Image /page/1/Picture/5 description: The logo for Springer, featuring a black outline of a chess knight piece to the left of the word "Springer" in a black serif font, all on a white background.
+
+floods with a mechanistic understanding of the processes involved (Zhang et al. 2024). These traditional risk assessment methods are often insufficient for addressing this complex environment and handling numerous evaluation units (Ahmed 2024; Niggli et al. 2024). Analyzing historical outburst events is crucial for studying GLOFs. Through the analysis of historical GLOF events, potential driving factors of glacial lake outbursts can be identified, providing valuable insights for future risk assessment and management (Ahmed 2024; Taylor et al. 2023). Moreover, data-driven, machine learning-based methods have emerged as powerful tools in geohazard prediction and have been widely applied to predict the susceptibility of landslides, avalanches, and debris flow (Chen et al. 2020; Viallon-Galinier et al. 2023; Wen et al. 2022; Zhou et al. 2024). These methods can extract hazard-related features from complex data and construct prediction models that incorporate multi-dimensional and multi-factor considerations. Compared to traditional statistical models, machine learning techniques offer significant advantages in handling large-scale data and capturing nonlinear relationships (Wen et al. 2022). Therefore, machine learning methods hold great potential for enhancing risk prediction of GLOFs.
+
+The aim of this paper is to quantitatively predict the susceptibility of GLOFs in the southern Tibetan Plateau by integrating historical GLOF events with machine learning methods. First, we systematically analyzed the development patterns of glaciers and glacial lakes in the study area and compiled historical records of GLOFs to examine their spatial and temporal patterns. Subsequently, a system of conditioning factors was constructed, and machine learning algorithms were applied to learn from historical events and develop a predictive model for the outburst susceptibility of each glacial lake in the region. The model's prediction accuracy and the relative importance of each conditioning factor were assessed. Finally, the paper discusses the response of GLOFs to climate change, identifies highly susceptible glacial lakes, and explores their dominant control factors. The results provide methodological support for predicting GLOF risks in similar regions and offer a theoretical basis for the prevention and management of GLOF disasters in the southern Tibetan Plateau.
+
+# 2 Study area
+
+The study area is situated on the southern Tibetan Plateau, encompassing the Himalayas and portions of the Hengduan Mountains, spanning a total area of 674,000 km². Based on geographical features, the study area is subdivided into western, central, and eastern subzones (Fig. 1a). The topography of the Himalayas varies significantly from north to south: the northern region consists of a high plateau and wide valley basins, with altitudes ranging from 4500 to 5200 m, while the southern part descends abruptly to below 3000 m, showcasing dramatic landscapes of deep valleys and towering mountains. The southern slopes of the Himalayas are lower in elevation, with the lowest point reaching no more than 500 m near the Ganges Plain. The southern Tibetan Plateau is primarily influenced by the Indian monsoon, the southern branch of the westerly circulation, and local climatic factors (Wen et al. 2024). Each year, around late February or early March, the Indian Ocean monsoon begins to impact the study area(Sun et al. 2022). The towering Himalayas block the warm, humid air brought by the monsoon, leading to substantial precipitation and snowfall on the southern side (Xu et al. 2022). In contrast, the southern branches of the westerly wind belt dominate the winter climate, which is relatively dry, with limited rain and snowfall. Most
+
+Image /page/2/Picture/6 description: The image displays the logo for the publisher Springer. On the left is a black outline of a chess knight piece, specifically a horse's head facing left, positioned above two short horizontal lines. To the right of this icon, the word "Springer" is written in a black serif font. The entire logo is set against a plain white background.
+
+Image /page/3/Figure/2 description: A multi-panel figure illustrating a study on glacial lakes in the Tibetan Plateau. The figure is composed of four panels labeled a, b, c, and d. Panel a is a large map of the Tibetan Plateau and surrounding areas, spanning from 75°E to 95°E longitude and approximately 27°N to 35°N latitude. The map shows the locations of glaciers (in light blue), glacial lakes (in pink), and historical Glacial Lake Outburst Floods (GLOFs, marked with purple circles). The study area is divided into Western, Middle, and Eastern subzones. Major cities like New Delhi, Kathmandu, Thimphu, and Lasa are labeled. Panel b is a Landsat 8 satellite image from 2020-05-10 of a glacial lake named Jinwengcuo. The lake is shown in dark blue, surrounded by reddish and cyan terrain. Panel c is another Landsat 8 image of the same lake, Jinwengcuo, from 2020-11-18, showing a change in the lake's appearance and surrounding snow/ice cover. Panel d is a photograph dated 2019-07-13, showing a person in a blue jacket standing on the shore of a lake next to a monitoring station. A small boat or buoy is visible on the water in the background.
+
+Fig. 1 Glacial Lake inventory map and location of the study area. (a) Glacial Lake inventory map; (b) & (c) Remote sensing images before and after Jinwucuo outburst; (d) Field survey photo of glacial lake
+
+of the winter snow is driven by oscillations in the north-south branch flow or by the intrusion of the strong Siberian cold front (Sun et al. 2022). Locally, the climate is influenced by the foehn wind effect, where the windward slopes of the southern Himalayas receive heavy precipitation, while the leeward side forms hot, dry valleys. This creates a stark contrast in climate between the northern and southern sides of the Himalayas. For example, the northern part of Gyirong County has a semi-arid plateau valley monsoon climate, characterized by an annual mean temperature of 2°C, with the warmest month averaging 10–18 °C and the coldest month averaging -10 °C. Annual precipitation in this region is approximately 300–600 mm. In contrast, the southern part of Jilong County experiences a subtropical mountain monsoon climate, with an average annual temperature of 10–13 °C, the warmest month exceeding 18 °C, and annual precipitation of around 1000 mm.
+
+# 3 Data and methodology
+
+## 3.1 Glacial lake inventory and historical GLOF events
+
+The glacial lake data utilized in this study are sourced from the Cataloging Data Set of High Asian Ice Lakes (Wang et al. 2020). This dataset integrates glacier cataloging data and 668 Landsat series images, processed using ArcGIS and ENVI software, with boundaries extracted through manual visual interpretation. The percentage difference between the absolute lake area and GPS-measured boundaries ranges from 5.5 to 25.5%. The dataset covers two time periods: 1990 and 2018. In 1990, the study area contained 9093 glacial lakes covering a total area of 711.2 km², while in 2018, the number of glacial lakes
+
+Image /page/3/Picture/8 description: The logo for Springer, featuring a black line drawing of a chess knight piece to the left of the word 'Springer' in a black serif font, all on a white background.
+
+increased to 10,173, with a total area of 819.3 km<sup>2</sup>. The 2018 glacial lakes are illustrated in Fig. 1a. Additionally, the glacier data were derived from the best integrated quality product generated through data fusion in the Assessment of Glacier Inventories for the Third Pole Region (1990–2015) (Xia and Shiqiao 2022). This dataset employs hierarchical analysis to filter eight contributing factors and evaluates the comprehensive quality of eight existing glacier cataloging datasets for the Third Pole Region by classifying them. A new glacier catalog was created by fusing the highest-quality data within each evaluation unit (He and Zhou 2022). The eight glacier datasets used are: Randolph Glacier Inventory (RGI) updated Glacier Area Mapping for Discharge in Asian Mountains (GAMDAM), Second Chinese Glacier Inventory (CGI-2), Hindu Kush-Himalayan Glacier Inventory (HKHGI), Western Himalayas Glacier Inventory (WHGI), Karakoram-Pamir Glacier Inventory (KPGI), Second Pakistan Glacier Inventory (PGI-2), and Southeast Tibetan Plateau Glacier Inventory (SETPGI). Historical GLOFs are primarily reviewed through the literature, focusing on the timing and localization of each event, as shown in Fig. 1a. To ensure the accuracy and reliability of the reported events, all GLOF occurrences since 1986 were cross-validated using historical remote sensing imagery, primarily from Landsat series, to verify their timing, spatial extent, and geomorphic signatures (Fig. 1b&c). Furthermore, ground-truthing was carried out at several key sites. Since 2017, we have performed multiple field surveys focusing on glacial lakes with documented historical GLOF events. During these surveys, UAV (Unmanned Aerial Vehicle) were used to identify potential triggering factors, such as ice and rock avalanches, around these lakes. Additionally, unmanned boats equipped with multibeam sonar systems were employed to map the lakebed topography, providing precise measurements of the glacial lake reservoirs (Fig. 1d). The information on these historical GLOF events is detailed in Table 1. On the basis of these data, statistical analyses were conducted to examine the elevation, flow direction, and slope of glaciers, as well as the elevation, area, type, and expansion trend of glacial lakes. Additionally, temporal and spatial clustering analyses of historical GLOF events were performed using GIS tools.
+
+## 3.2 Learning samples and conditioning factors
+
+Sample selection plays a critical role in evaluating the GLOF susceptibility, as it directly influences the model's training effectiveness and prediction accuracy. In this study, 66 documented GLOF events within the study area were used as positive samples, excluding the Lumu Lake outburst due to unavailable localization data. An equal number of glacial lakes are randomly selected as negative samples using GIS tools to ensure an appropriate balance between positive and negative samples. Evaluation units are defined as the watersheds of individual glacial lakes within the study area. The conditioning factors must be representative, measurable, and based on readily available data. In this study, eight conditioning factors were selected: Glacial Lake Area, Glacial Lake Replenishment Area, Distance to Glacier, Elevation, Elevation Coefficient of Variation (ECV), Topographic Wetness Index (TWI), Peak Ground Acceleration (PGA), and Distance to Potential Icefall (Fig. 2). Among them, the Glacial Lake Area reflects the storage capacity of the glacial lake. The areas of glacial lakes used as a learning sample are reconstructed as comprehensively as possible using remote sensing imagery and geomorphologic evidence. Glacial Lake Replenishment Area determines the lake's ability to receive meltwater and precipitation. Distance to Glacier characterizes the extent of direct glacial activity impact. Those three factors can be
+
+Image /page/4/Picture/5 description: The logo for Springer, featuring a black line drawing of a chess knight piece on the left, facing left, next to the word "Springer" in a black serif font. The background is white.
+
+calculated and obtained using GIS. Additionally, elevation directly influences the local climate around the glacial lake, hydrological processes, and the extent and intensity of flood impacts following an outburst. Elevation data are extracted from the ASTER GDEM with 30 m resolution (https://www.gscloud.cn/). The ECV reflects the degree of terrain relief and the magnitude of elevation change within the region. The TWI is a physical indicator that captures the influence of regional topography on runoff flow direction and water accumulation. Both the ECV and TWI are derived from GIS calculations. Strong earthquakes can trigger landslides, avalanches, or ground deformation in nearby mountainous regions, leading to significant changes in glacial lake levels or even catastrophic outbursts. The seismic factor is quantitatively represented by the PGA, with data sourced from the Global Earthquake Model Global Seismic Hazard Map (Johnson et al. 2023). Icefalls are among the primary natural triggers for GLOFs. In this study, the distance to potential icefall is used to quantitatively assess their impact on GLOFs. The potential icefall data are derived from the research team's previous findings. Relative to the evaluation factors in existing studies, this study places particular emphasis on ECV and TWI to better characterize the impacts of heterogeneous mountainous terrain. These two topographic variables—ECV and TWI—are critical for understanding water routing and flood dynamics, especially within the complex geomorphological settings of the southeastern Tibetan Plateau. Furthermore, we introduce 'Distance to potential icefall' as a novel factor, informed by recent findings on icefall-triggered GLOFs in the Himalaya.
+
+After preparing the conditioning factors, descriptive statistics should be conducted on the values of all positive samples, along with an analysis of their degree of dispersion, to evaluate their effectiveness in differentiating GLOF susceptibility. Specifically, measures such as mean, standard deviation, and coefficient of variation will be used to quantify the variability and discriminatory power of each factor (Wen et al. 2022). A higher coefficient of variation indicates greater dispersion of the factor values, which, to some extent, suggests that the factor has weaker discriminatory power in evaluating GLOF susceptibility. In addition, the combination of conditioning factors must undergo a multicollinearity test. Multicollinearity refers to the reduced objectivity and accuracy of results caused by high correlations between explanatory variables in the model (Thompson et al. 2017). To assess the correlation between factors, the variance inflation factor (VIF) is used. The formula for calculating the VIF index is as follows (Thompson et al. 2017):
+
+$$VIF = \frac{1}{1 - R_i^2} \tag{1}$$
+
+where,  $R_j^2$  is the multiple linear regression coefficient. The VIF indicates the severity of multicollinearity. A VIF value closer to 1 signifies weaker multicollinearity, while a higher VIF value indicates stronger multicollinearity. Typically, a VIF value of 10 is used as the threshold for assessing multicollinearity. When VIF<10, there is no significant multicollinearity, which meets the requirement for model stability.
+
+To further evaluate the relative importance of each conditioning factor and enhance model interpretability, a feature importance analysis is performed in the Results section. This analysis elucidates which factors dominate susceptibility modeling and how they interact with others to contribute to GLOF occurrence, thereby improving the transparency and physical interpretability of the model outputs. Additionally, the Discussion section provides
+
+Image /page/5/Picture/7 description: The Springer logo, presented in black on a white background. On the left is a stylized icon of a chess knight facing left. To the right of the icon is the word 'Springer' written in a serif font.
+
+|                     | ID Name                      | Longitude           | Latitude        | Date                                | Death<br>toll | Trigger                     | Reference                                | Verified by remote sensing    |                       |
+|---------------------|------------------------------|---------------------|-----------------|-------------------------------------|---------------|-----------------------------|------------------------------------------|-------------------------------|-----------------------|
+|                     |                              |                     |                 |                                     |               |                             |                                          | Data set Acquisi-             |                       |
+| _                   | Lumu Lake                    | N/A                 | N/A             | 1931/06/08<br>(Tibetan<br>calendar) | 72            | Icefall                     | (Yao et al. 2014)                        |                               |                       |
+| 7                   | Taraco                       | 86.13°              | 28.29°          | 1935/08/28                          |               | Moraine dam destabilization | (Cui et al. 2010; Veh et al. 2020)       |                               |                       |
+| 3                   | Qiongbixiamaco               | 88.92°              | 27.85°          | 1940/07/10                          |               | Icefall                     | (Cui et al. 2010; Veh et al. 2020)       |                               |                       |
+| 4                   | Sangwangco                   | 90.01°              | 28.24°          | 1954/07/16                          | 400           | Icefall                     | (Liu et al. 2019)                        |                               |                       |
+| 2                   | Lureco                       | 90.59°              | 28.27°          | $\sim\!1950s$                       |               | Icefall                     | (Yao et al. 2014)                        |                               |                       |
+| 9                   | Dacangco                     | 86.35°              | 28.19°          | 1956/06/01                          |               | Heavy rainfall              | (Tong et al. 2019)                       |                               |                       |
+| _                   | Cuoalong Glacial<br>lake     | 90.59°              | 28.05°          | ~1956–1966                          |               | Icefall                     | (Komori et al. 2012; Zhang et al. 2021b) |                               |                       |
+| ∞                   | Cirenmaco                    | .90 <sup>.</sup> 98 | 28.07°          | 1964/07                             | 200           | Icefall                     | (Veh et al. 2020)                        |                               |                       |
+| 6                   |                              | .90 <sup>.</sup> 98 | 28.07°          | 1981/07/11                          |               | Icefall                     | (Veh et al. 2020; Xu 1988)               |                               |                       |
+| 10                  |                              | .90 <sup>.</sup> 98 | $28.07^{\circ}$ | 1983                                |               | Icefall                     | (Wang et al. 2018)                       |                               |                       |
+| 11                  | Longdaco                     | 85.35°              | $28.62^{\circ}$ | 1964/08/25                          |               | Icefall                     | (Cui et al. 2010; Veh et al. 2020)       |                               |                       |
+| 12                  | Jilaico<br>(Gehaipuco)       | 87.81°              | 27.96°          | 1964/09/21                          |               | Icefall                     | (Cui et al. 2010; Veh et al. 2020)       |                               |                       |
+| 13                  | Damenlahaico                 | 93.04°              | 29.87°          | 1964/09/26                          |               | Icefall                     | (Cui et al. 2010; Veh et al. 2020)       |                               |                       |
+| 4                   | Jichudrake North 1           | 89.29°              | 27.87°          | $\sim\!1960s$                       |               | Icefall                     | (Komori et al. 2012; Zhang et al. 2021b) |                               |                       |
+| 15                  | Upper Jiejiu Tsho            | 90.72°              | 28.27°          | $\sim\!1960s$                       |               | Icefall                     | (Komori et al. 2012; Zhang et al. 2021b) |                               |                       |
+| 16                  | Upper Shegong<br>Tsho        | 90.74°              | 28.30°          | $\sim\!1960s$                       |               | Icefall                     | (Komori et al. 2012; Zhang et al. 2021b) |                               |                       |
+| 17                  | 17 Jichudrake North 2 89.31° | 89.31°              | 27.87°          | $\sim\!1960s$                       |               | Icefall                     | (Komori et al. 2012; Zhang et al.        |                               |                       |
+| Table 1 (continued) |                              |                     |                 |                                     |               |                             |                                          |                               |                       |
+| ID                  | Name                         | Longitude           | Latitude        | Date                                | Death<br>toll | Trigger                     | Reference                                | Verified by remote<br>sensing |                       |
+|                     |                              |                     |                 |                                     |               |                             |                                          | Data set                      | Acquisi-<br>tion time |
+| 18                  | Ayaco                        | 86.49°              | 28.35°          | 1968/08/15                          |               | Moraine dam destabilization | (Cui et al. 2010; Veh et al. 2020)       |                               |                       |
+| 19                  |                              | 86.49°              | 28.35°          | 1969/08/17                          |               | Moraine dam destabilization | (Yao et al. 2014)                        |                               |                       |
+| 20                  |                              | 86.49°              | 28.35°          | 1970/08/15                          |               | Moraine dam destabilization | (Yao et al. 2014)                        |                               |                       |
+| 21                  | Zhari                        | 90.61°              | 28.30°          | 1981/06/24                          |               | Icefall                     | (Yao et al. 2014)                        |                               |                       |
+| 22                  | Yindapuco (Jinco)            | 87.91°              | 27.95°          | 1982/08/27                          |               | Moraine dam destabilization | (Reynolds 2023)                          |                               |                       |
+| 23                  | Dig Tsho                     | 86.58               | 27.87           | 1985/08/4                           |               | Icefall                     | (Cenderelli and Wohl 2001)               |                               |                       |
+| 24                  |                              |                     |                 | 2015/04/25                          |               | Earthquake                  | (Byers et al. 2017)                      | Landsat-8                     | 2015/05/25            |
+| 25                  | Sanga North                  | 93.88°              | 30.12°          | ~1976–1987                          |               | N/A                         | (Veh et al. 2019)                        | Landsat-5                     | 1988/10/09            |
+| 26                  | Guangxieco                   | 96.50°              | 29.46°          | 1988/07/15                          | 5             | Icefall                     | (Liu et al. 2019)                        | Landsat-5                     | 1988/10/27            |
+| 27                  | Gebumaco Upper               | 96.54°              | 29.75°          | 1991/06/12                          |               | Icefall                     | (Nie et al. 2018)                        | Landsat-5                     | 1993/10/09            |
+| 28                  | Gebumaco                     | 96.55°              | 29.75°          | 1991/06/12                          |               | Moraine dam destabilization | (Wang et al. 2011)                       | Landsat-5                     | 1993/10/09            |
+| 29                  | Chubung                      | 86.46°              | 27.87°          | 1991/07/12                          |               | Landslide                   | (Reynolds John 1999; Veh et al. 2020)    | Landsat-5                     | 1991/11/15            |
+| 30                  | Upper Luggye                 | 90.32°              | 28.08°          | 1991/09/30                          |               | N/A                         | (Veh et al. 2019)                        | Landsat-5                     | 1991/09/30            |
+| 31                  |                              |                     |                 | 2010/10/04                          |               | N/A                         | (Veh et al. 2019)                        | Landsat-5                     | 2010/10/04            |
+| 32                  | Baitang Weat                 | 92.78°              | 29.55°          | 1991/10/09                          |               | N/A                         | (Veh et al. 2019)                        | Landsat-5                     | 1991/10/09            |
+| 33                  | Langbuco Upper               | 86.42°              | 27.92°          | 1992/08                             |               | Icefall                     | (Nie et al. 2018)                        | Landsat-4                     | 1992/09/22            |
+| 34                  | Rejieco                      | 88.90°              | 27.97°          | 1992/09/10                          |               | Icefall                     | (Zhang et al. 2021b)                     | Landsat-5                     | 1992/11/10            |
+| 35                  | Zangla Tsho                  | 82.11°              | 30.35°          | 1994/08                             |               | N/A                         | (Zhang et al. 2021b)                     | Landsat-5                     | 1994/09/09            |
+| 36                  | Luggye Tsho                  | 90.30°              | 28.09°          | 1994/10/07                          |               | Moraine dam destabilization | (Veh et al. 2020; Wang et al. 2012)      | Landsat-5                     | 1994/11/09            |
+| 37                  | Xiaga                        | 91.94°              | 28.80°          | 1995/05/26                          |               | Icefall                     | (Yao et al. 2014)                        | Landsat-5                     | 1995/09/02            |
+| 38                  | Zhanabo                      | 85.37°              | 28.66°          | 1995/06/07                          |               | Moraine dam destabilization | (Yao et al. 2014)                        | Landsat-5                     | 1995/11/01            |
+| 39                  | Zanglaco West                | 81.99°              | 30.37°          | 1995/09                             |               | N/A                         | (Zhang et al. 2021b)                     | Landsat-5                     | 1995/10/30            |
+| 40                  | Kongyangmi La<br>Tsho        | 88.78°              | 27.90°          | 1996/05/13                          |               | Icefall                     | (Nie et al. 2018)                        | Landsat-5                     | 1996/05/13            |
+| 41                  | Zhalonggabu                  | 85.48°              | 28.66°          | ~1996                               |               | N/A                         | (Veh et al. 2019)                        | Landsat-5                     | 1996/10/18            |
+| 42                  | Unnamed Lake                 | 92.38°              | 27.70°          | 1997/10/18                          |               | N/A                         | (Zhang et al. 2021b)                     | Landsat-5                     | 1997/10/18            |
+| ID                  | Name                         | Longitude           | Latitude        | Date                                | Death<br>toll | Trigger                     | Reference                                | Verified by remote<br>sensing |                       |
+|                     |                              |                     |                 |                                     |               |                             |                                          | Data set                      | Acquisi-<br>tion time |
+| 43                  | Tam Pokhari                  | 86.84°              | 27.74°          | 1998/09/03                          |               | Icefall                     | (Lamsal et al. 2016; Veh et al. 2020)    | Landsat-5                     | 1998/09/15            |
+| 44                  | Upper Gangri<br>Tsho1        | 90.81°              | 27.88°          | 1998/09/10                          |               | Icefall                     | (Komori et al. 2012)                     | Landsat-5                     | 1998/09/10            |
+| 45                  | Upper Gangri<br>Tsho2        | 90.81°              | 27.89°          | 1998/09/10                          |               | Upstream inflow             | (Zhang et al. 2021b)                     | Landsat-5                     | 1998/09/10            |
+| 46                  | Dareco South                 | 85.92°              | 28.13°          | ~2001–2003                          |               | Icefall                     | (Liu et al. 2019)                        | Landsat-5                     | 2003/12/25            |
+| 47                  | Longjiuco                    | 89.74°              | 28.21°          | ~2002–2003                          |               | Heavy rainfall              | (Zhang et al. 2021b)                     | Landsat-5                     | 2003/11/18            |
+| 48                  | Jialongco                    | 85.85°              | 28.21°          | 2002/05/23                          |               | Icefall                     | (Chen et al. 2006)                       | Landsat-5                     | 2004/05/17            |
+| 49                  | Jialongco                    | 85.85°              | 28.21°          | 2002/06/29                          |               | Icefall                     | (Chen et al. 2006)                       | Landsat-5                     | 2004/05/17            |
+| 50                  | Degaco                       | 90.67°              | 28.33°          | 2002/09/18                          |               | Snow avalanche              | (Liu et al. 2013)                        | Landsat-5                     | 2003/11/18            |
+| 51                  | Jitang Sourth                | 94.32°              | 30.66°          | 2002–2003                           | 9             | N/A                         | (Liu et al. 2019)                        | Landsat-5                     | 2003/10/19            |
+| 52                  | Ouguchongguco<br>Upper       | 93.54°              | 29.63°          | 2004/10/21                          |               | Icefall                     | (Yao et al. 2014)                        | Landsat-5                     | 2004/10/21            |
+| 53                  | Dagonglongba<br>Lake         | 96.46°              | 29.75°          | ~2007–2008                          |               | Moraine dam destabilization | (Zhang et al. 2021b)                     | Landsat-5                     | 2008/06/28            |
+| 54                  | Sangwangco East              | 90.22°              | 28.27°          | 2008/10/31                          |               | Moraine dam destabilization | (Liu et al. 2019)                        | Landsat-5                     | 2008/12/01            |
+| 55                  | Langco                       | 91.81°              | 27.83°          | 2007/08/10                          |               | Heavy rainfall              | (Yao et al. 2014)                        | Landsat-5                     | 2008/05/16            |
+| 56                  | Zhemaico                     | 92.34°              | 28.01°          | 2009/07/03                          |               | Moraine dam destabilization | (Liu et al. 2013)                        | Landsat-5                     | 2009/10/10            |
+| 57                  | Cuoga                        | 94.00°              | 30.83°          | 2009/07/29                          | 2             | Icefall                     | (Yao et al. 2014)                        | Landsat-5                     | 2009/09/17            |
+| 58                  | Geiqu Lake                   | 87.99°              | 27.95°          | 2010/07                             |               | Heavy rainfall              | (Yao et al. 2014)                        | Landsat-5                     | 2010/11/12            |
+| 59                  | Nalongzangbu<br>Lake         | 94.95°              | 30.53°          | 2014/06/30                          |               | Icefall                     | (Zhang et al. 2021b)                     | Landsat-8                     | 2014/08/23            |
+| 60                  | Ranzeranco<br>(Ranzeriaco)   | 93.53°              | 30.47°          | 2013/07/05                          |               | Moraine dam destabilization | (Wang et al. 2021)                       | Landsat-8                     | 2013/09/12            |
+| 61                  | Lemthang Tsho                | 89.58°              | 28.06°          | 2015/06/28                          |               | Landslide                   | (Gurung et al. 2017; Veh et al. 2020)    | Landsat-8                     | 2015/10/18            |
+| 62                  | Yindapuco Upper              | 87.88°              | 27.93°          | 2015/10/31                          |               | N/A                         | (Reynolds 2023)                          | Landsat-8                     | 2015/11/10            |
+| 63                  | Zangbuco East                | 82.11°              | 30.16°          | ~2016                               |               | N/A                         | (Liu et al. 2019)                        | Landsat-8                     | 2016/07/10            |
+| Table 1 (continued) |                              |                     |                 |                                     |               |                             |                                          |                               |                       |
+| ID                  | Name                         | Longitude           | Latitude        | Date                                | Death<br>toll | Trigger                     | Reference                                | Verified by remote sensing    |                       |
+|                     |                              |                     |                 |                                     |               |                             |                                          | Data set                      | Acquisi-<br>tion time |
+| 64                  | Gongbatongshaco              | 86.05°              | 28.07°          | 2016/07/05                          |               | Icefall                     | (Cook et al. 2018)                       | Landsat-8                     | 2016/07/14            |
+| 65                  | Langmale Lake                | 87.14°              | 27.81°          | 2017/04/20                          |               | Rock avalanche              | (Byers et al. 2019)                      | Landsat-8                     | 2017/11/06            |
+| 66                  | Jinwengco                    | 93.63°              | 30.35°          | 2020/06/26                          |               | Landslide                   | (Yang et al. 2022)                       | Landsat-8                     | 2020/08/30            |
+
+Image /page/6/Picture/4 description: The logo for the publisher Springer, featuring a black outline of a chess knight piece to the left of the word "Springer" in a black serif font, all on a white background.
+
+Image /page/7/Picture/3 description: The Springer logo, featuring a black outline of a chess knight piece to the left of the word 'Springer' in a black serif font, all on a white background.
+
+Table 1 (continued)
+
+Image /page/8/Picture/4 description: The logo for Springer, featuring a black and white line drawing of a chess knight's head facing left, positioned above two horizontal lines. To the right of the icon is the word "Springer" in a black serif font, all on a white background.
+
+Image /page/9/Picture/3 description: The logo for Springer, featuring a black outline of a knight chess piece to the left of the word "Springer" in a black serif font, all on a white background.
+
+Image /page/10/Figure/2 description: A figure from a scientific paper titled 'Fig. 2 GLOF conditioning factors', displaying eight different maps (labeled a through h) of a mountainous region. The figure includes a scale bar from 0 to 500 km and a north arrow. The maps are as follows:
+
+a. 'Glacial lake area': Shows the distribution of glacial lakes, represented by small pink dots within a dashed outline of the region.
+
+b. 'Glacial lake replenishment area (m²)': A map colored mostly dark blue, with a color scale ranging from 0 to 2.88303e+10.
+
+c. 'Glaciers': Displays the location of glaciers in light blue.
+
+d. 'Elevation (m)': A topographic map with a color scale from green (low elevation) to brown (high elevation), ranging from 0 to 8848 meters.
+
+e. 'Elevation variation coefficient': A map with colors from blue to yellow, with a scale from 0 to 1.628.
+
+f. 'Topographic wetness index': A map primarily in shades of green and yellow, with a scale from 1.96 to 26.5.
+
+g. 'Peak ground acceleration (m/s²)': A map with a color gradient from blue to yellow and orange, with a scale from 0.095 to 0.589.
+
+h. 'Potential icefall': Shows the location of potential icefall areas as black dots and a labeled pink shape within the region's outline.
+
+Fig. 2 GLOF conditioning factors
+
+Image /page/10/Picture/4 description: The logo for the publisher Springer, featuring a black outline of a chess knight's head facing left, followed by the word "Springer" in a black serif font, all on a white background.
+
+a comparison of our selected factors with those in previous studies, emphasizing both commonalities and region-specific innovations.
+
+## 3.3 GLOF modeling and model evaluation
+
+### 3.3.1 Support vector machine (SVM)
+
+SVM is a machine learning algorithm primarily employed for classification and regression tasks. The primary objective of SVM is to identify an optimal hyperplane that effectively separates data points of different classes in an N-dimensional feature space, where N denotes the number of features (Yousefi et al. 2020). Simultaneously, SVM aims to maximize the margin, defined as the distance between the hyperplane and the nearest data points from each class. Mathematically, a hyperplane is characterized by its normal vector w and offset b, and can be expressed as:
+
+$$f(x) = w^T x + b (2)$$
+
+where, *x* is the feature vector, serving as the condition factor in the susceptibility evaluation. SVMs can address nonlinear classification problems by employing a kernel function. The kernel function maps the data into a higher-dimensional space, enabling the identification of a linear hyperplane that separates the data. Commonly used kernel functions include the linear kernel, polynomial kernel, and radial basis function (RBF) kernel. In this context, the kernel function computes the similarity between each sample in the high-dimensional space and the support vectors, which can be interpreted as weights. By analyzing the weights assigned to each feature, we can evaluate the relative importance that the model attributes to different condition factors.
+
+SVM was selected in this study due to its effectiveness in handling high-dimensional and small-sample datasets, which are characteristic of the compiled GLOF inventory. SVM often demonstrates higher generalization capacity in sparse datasets and requires fewer hyperparameters to tune (Ragab et al. 2021). Additionally, SVM is less prone to overfitting when appropriately regularized (Yousefi et al. 2020), which is critical given the imbalanced nature of GLOF event distribution.
+
+### 3.3.2 Multilayer perceptron (MLP)
+
+MLP is a feed-forward artificial neural network model primarily employed for supervised learning tasks. In classification tasks, MLP transforms raw input features through multiple fully connected layers, and the predicted probability for each class is obtained using the Softmax function (Ramchoun et al. 2016). MLP can be conceptualized as a directed graph consisting of multiple layers of nodes, where each node, except those in the input layer, represents a neuron with a nonlinear activation function. The architecture of an MLP comprises three types of layers: the input layer, hidden layers, and output layer. The input layer receives the input feature vector, with each feature corresponding to an individual input neuron. The hidden layers, which follow the input layer, may consist of one or more layers. Neurons in the hidden layers process the outputs from the previous layer by applying weights and activation functions. The output layer generates the final prediction. During
+
+Image /page/11/Picture/11 description: The logo for Springer, featuring a black outline of a chess knight piece to the left of the word 'Springer' in a black serif font, all on a white background.
+
+training, the MLP adjusts its weights via back-propagation and gradient descent, thereby determining the relative importance of each condition factor.
+
+The choice of MLP was motivated by its ability to model complex nonlinear interactions among conditioning factors, which is essential in capturing the multifactorial nature of GLOF susceptibility. MLP can be more effectively adapted to address data imbalance through architecture design and loss function adjustments (Castro and Braga 2013). Furthermore, integrating two fundamentally different model architectures (kernel-based SVM and network-based MLP) enables robust cross-validation and allows us to evaluate the consistency of spatial susceptibility patterns across distinct algorithmic paradigms.
+
+### 3.3.3 Model evaluation
+
+The Receiver Operating Characteristic (ROC) curve is a graphical representation that plots the true positive rate (sensitivity) on the Y-axis and the false positive rate (1-specificity) on the X-axis at various classification thresholds (Mandrekar 2010). This plot serves as a visual tool for evaluating model performance. Sensitivity and specificity reflect the model's ability to correctly identify positive and negative samples, respectively. However, these two metrics alone do not fully capture the overall accuracy of the model. Therefore, the Area Under the Curve (AUC) value is often used to comprehensively evaluate model performance. The AUC represents the area under the ROC curve and is a widely accepted metric for assessing the performance of binary classifiers. The AUC can be calculated using the following formula (Mandrekar 2010):
+
+$$AUC = \frac{\sum_{i=1}^{n_0} r_i - n_0 (n_0 + 1) / 2}{n_0 n_1}$$
+ (3)
+
+where,  $n_0$  represents the number of negative samples,  $n_1$  denotes the number of positive samples, and  $r_i$  indicates the positional rank of the i-th negative sample within the entire test dataset. The AUC value ranges from 0 to 1, with values closer to 1 indicating higher prediction accuracy of the model.
+
+Finally, the outburst susceptibility index of all glacial lakes was classified into five categories using the natural breakpoint method: very high susceptibility (VHS), high susceptibility (HS), medium susceptibility (MS), low susceptibility (LS), and very low susceptibility (VLS). Glacial lakes with VHS and HS were analyzed in detail, along with an examination of the significance of each conditioning factor. The complete methodological process of this study is illustrated in the Fig. 3 and can be summarized as follows: (1) Historical GLOF events and randomly selected non-outburst glacial lakes were used as learning samples, combined with multiple conditioning factors to construct a comprehensive dataset; (2) The dataset was randomly partitioned into training (80%) and validation subsets (20%); (3) Two machine learning algorithms—SVM and MLP—were trained using the training subset to develop susceptibility evaluation models for GLOFs; (4) The trained models were assessed through ROC curves, with their performance quantified by AUC values based on the validation subset; (5) The susceptibility index derived from the optimal models was classified using the natural breakpoint method to identify lakes prone to GLOF events. This systematic integration ensures that SVM and MLP effectively capture complementary patterns of
+
+Image /page/12/Picture/9 description: The logo for the publisher Springer, featuring a black outline of a chess knight piece on the left, next to the word "Springer" in a black serif font. The background is white.
+
+Image /page/13/Figure/2 description: A flowchart detailing a methodology for assessing Glacial Lake Outburst Flood (GLOF) susceptibility. The flowchart has two main pathways. The first pathway takes "Historical GLOF events", "Glacier", and "Glacial lake" data, performs a "Statistical analysis" to determine "Spatiotemporal differentiation characteristics", and results in an understanding of "GLOFs in response to climate change". The second pathway uses machine learning. It starts with inputs of "Randomly selected glacial lakes" and "Conditioning factors" (which include Glacial lake area, Glacial lake replenishment area, Distance to glaciers, Elevation, Elevation variation coefficient, Topographic Wetness Index, Peak ground acceleration, and Distance to potential icefall). This data is split into a "Training sample 80%" and a "Validation sample 20%". The training sample is fed into a "Machine learning algorithm" to create an "Evaluation model of GLOF susceptibility (SVM)" and an "Evaluation model of GLOF susceptibility (MLP)". These models help determine the "Importance of conditioning factors". The validation sample is used for an "ROC test". Both the machine learning algorithm and the ROC test contribute to a "GLOF susceptibility index". This index is then processed by a "Natural breakpoint method" to identify "GLOF-prone lakes".
+
+Fig. 3 Flowchart of the methodology employed for GLOF susceptibility analysis
+
+GLOF susceptibility, while ROC and AUC provide robust criteria for model selection and reliability assessment.
+
+# 4 Results
+
+## 4.1 Characteristics of glacial lake and historical GLOF events
+
+### 4.1.1 Glacier and glacial lake
+
+The study area contains a total of 34,151 glaciers, distributed as follows: 18,460 glaciers in the western subzone, 6759 in the central subzone, and 8932 in the eastern subzone. The elevation range of glacier spans from 2741 m to 6716 m, as shown in Fig. 4a. In this figure, the horizontal axis represents glaciers sequentially numbered from west to east. Overall, glaciers at the eastern and western ends of the southern Tibetan Plateau are found at lower elevations, while those in the central section are distributed at higher elevations. The interquartile range (25–75%) of glacier elevations is between 5110 m and 5601 m, with an average elevation of 5339 m (Fig. 4b). The primary glacier flow direction across the study area is northward, followed by northeast and northwest, while southwestward flow is the least common, followed by southward flow. The three subzones exhibit distinct glacier flow patterns: in the western subzone, the dominant direction is northward, consistent with the overall trend, while southwestward flow is the least common; in the central subzone, the dominant flow direction is westward, followed by southeastward, with northwestward flow being the least frequent; in the eastern subzone, the primary flow direction is northeastward, followed by southwestward (Fig. 4c). The overall glacier slope in the study area predominantly falls within the range of 10°-35°, with the highest concentration (22%) in the 20°-25° range. The slope distribution pattern is similar across all three subzones, with the 20°-25° interval being the most common (Fig. 4d).
+
+In 2018, a total of 10,173 glacial lakes were identified in the study area. The western subzone, despite having the highest concentration of glaciers, contained the fewest glacial lakes, with only 1861. In contrast, the central subzone had 3573 glacial lakes, while the
+
+Image /page/13/Picture/10 description: The image displays the Springer logo on a white background. On the left side of the logo is a black line drawing of a chess knight piece, with the horse's head facing to the left. Below the horse's head are two horizontal lines representing the base of the piece. To the right of the chess piece, the word 'Springer' is written in a black serif font.
+
+Image /page/14/Figure/2 description: A figure from a scientific paper, labeled "Natural Hazards (2025) 121:17677–17705", containing four graphs labeled a, b, c, and d that analyze elevation and contribution ratios across different subzones.
+
+Graph a is a scatter plot showing Elevation (m) on the y-axis (from 3000 to 7000) versus ID on the x-axis (from 0 to over 30000). The plot is divided into a Western subzone, a Middle subzone, and an Eastern subzone. The elevation points in the Western subzone generally increase from west to east. The Middle subzone shows a high concentration of points at high elevations, mostly between 5500 m and 6500 m. The Eastern subzone shows a general decrease in elevation from west to east.
+
+Graph b is a box plot summarizing the total elevation data. The y-axis is the same as in graph a. The box, representing the 25%-75% interquartile range (IQR), is approximately between 5000 m and 5500 m. The median is around 5300 m, and the mean is slightly higher. Whiskers extend up to about 6500 m and down to 4000 m, with numerous outliers shown as dots.
+
+Graph c is a radar chart showing the Contribution ratio (%) on the radial axis (from 0 to 20) versus direction in degrees (0 to 360). It compares the total data (black line) with the Western (blue), Middle (purple), and Eastern (magenta) subzones. The Western subzone peaks around 330 degrees, the Middle subzone peaks around 270 degrees, and the Eastern subzone has peaks around 30 and 150 degrees.
+
+Graph d is a polar plot showing the Contribution ratio (%) on the radial axis (from 0 to 25) versus an angle from 0 to 90 degrees. It uses the same color scheme as graph c. All four lines show an increasing contribution ratio with the angle, peaking between 20 and 30 degrees before decreasing. The Eastern subzone shows the highest peak contribution ratio, over 10%.
+
+**Fig. 4** Elevation, orientation and slope of glaciers. (a) Elevation of glaciers from west to east; (b) Box chart of elevation distribution of all glaciers; (c) Radar map of glacier flow direction; (d) Slope map of glaciers
+
+eastern subzone had the highest number, with 4739. The elevation of these glacial lakes ranged from 2174 m to 6078 m, as shown in Fig. 5a. The horizontal axis represents the sequence of glacial lakes numbered from west to east. Generally, the elevation of the glacial lakes decreases linearly from west to east. The elevation distribution of the glacial lakes predominantly fell within the interquartile range (25–75%) between 4568 m and 5232 m, with an average elevation of 4891 m (Fig. 5b), which is slightly lower than the overall elevation distribution of glaciers. The areas of the glacial lakes mainly ranged from 5400 m<sup>2</sup> to 58,137,579 m<sup>2</sup>. The western subzone featured smaller glacial lake areas, while the central subzone had the largest average glacial lake area, at 92,880 m<sup>2</sup>. The median area of glacial lakes in the eastern subzone was nearly the same as in the central subzone, with values of 28,934 m<sup>2</sup> and 28,258 m<sup>2</sup>, respectively (Fig. 5c). In terms of glacial lake types, Iceuncontacted Lakes represented the largest proportion at 46.39%, while Supraglacial Lakes accounted for the smallest proportion at 3.88%. Non-Glacier-Fed Lakes made up approximately 40.59% (Fig. 5d). From 1990 to 2018, both the number and area of glacial lakes in the three subzones showed an upward trend. The central subzone had the highest rate of increase, with a growth of 11.41% in number and 5.33% in area. The western subzone exhibited a 9.19% increase in the number of glacial lakes and a 1.38% increase in area. The eastern subzone showed an 8.73% increase in the number of glacial lakes and a 2.82% increase in area. Additionally, glacial lakes in the study area generally begin to thaw gradually at the end of April and refreeze by October.
+
+Image /page/14/Picture/5 description: The Springer logo, featuring a black and white icon of a chess knight to the left of the word 'Springer' in a black serif font, all on a white background.
+
+Image /page/15/Figure/2 description: A composite figure with five charts labeled a, b, c, d, and e, analyzing data on glacial lakes. 
+
+Chart (a) is a scatter plot showing the elevation of glacial lakes. The y-axis, 'Elevation (m)', ranges from 3000 to 6500. The x-axis, 'Glacial Lake ID', ranges from 0 to over 10000 and is divided into 'Western subzone', 'Middle subzone', and 'Eastern subzone'. The data points are concentrated between 4000 and 6000 meters.
+
+Chart (b) is a box plot of the elevation data from chart (a). The y-axis is 'Elevation (m)' from 3000 to 6500. The box shows an interquartile range from approximately 4700 to 5200 m, with a median around 5000 m.
+
+Chart (c) consists of three box plots comparing the 'Area (m²)' of lakes in the 'Western subzone', 'Middle subzone', and 'Eastern subzone'. The y-axis ranges from 0 to 200000. All three subzones show a median area below 50000 m² but have numerous outliers with larger areas. A legend explains the components of the box plot: 25%-75%, 1.5IQR, Median, Mean, and Outliers.
+
+Chart (d) is a pie chart showing the distribution of lake types by percentage. The categories are: 'Ice-uncontacted Lake' (46.39%), 'Non-Glacier-Fed Lake' (40.59%), 'Ice-contacted' (9.14%), and 'Supraglacial Lake' (3.88%).
+
+Chart (e) is a grouped bar chart showing the percentage of the total number and area of lakes in each subzone. The y-axis is 'Percentage (%)' from 0 to 12. For the 'Western subzone', the number is about 9.2% and the area is about 1.5%. For the 'Middle subzone', the number is about 11.2% and the area is about 5.2%. For the 'Eastern subzone', the number is about 8.8% and the area is about 2.8%.
+
+**Fig. 5** Elevation, area, type and expansion trend of glacial lake. (a) Elevation of glacial lakes from west to east; (b) Box chart of elevation distribution of all glacial lakes; (c) Box chart of glacial lake areas; (d) Pie chart of glacial lake types; (e) The number and area changes of glacial lakes from 1990 to 2018
+
+### 4.1.2 Historical GLOF events
+
+The earliest documented GLOF in the study area occurred on August 28, 1935 (Tibetan calendar), at Lumu Lake in Bomi County. As of 2024, a total of 66 GLOF events have been recorded in the area, as detailed in Table 1. Among these, Cirenmacuo Glacial Lake and Ayacuo Glacial Lake each experienced three outbursts, while Dig Tsho Glacial Lake and Upper Luggye Glacial Lake each experienced two outbursts. Geographically, these historical GLOF events are primarily concentrated in the central and eastern subzones of the study area. Among the three subzones, the western subzone recorded the fewest GLOF events, with only three events, while the central subzone experienced the most, with 44 events, and the eastern subzone had 19. To further evaluate the spatial distribution characteristics of these events, a spatial autocorrelation analysis using the Global Moran's I index was conducted. The results showed a Moran's I value of 0.239, a Z-score of 1.69, and a p-value of 0.091, indicating a weak spatial clustering pattern of GLOF events. However, the clustering was not statistically significant at the 95% confidence level, although marginal significance was observed at the 90% confidence threshold. Regarding temporal trends, a higher number of GLOF events occurred during the 1960s and from 1990 to 2020 (Fig. 6a). The majority of these GLOFs occurred between June and October, coinciding with the rainy season on the Tibetan Plateau (Fig. 6b). The primary causes of these events (Fig. 6c) include icefalls (48.48%), moraine dam destabilization (18.18%), and heavy rainfall (6.06%). With global warming, the frequency and variety of GLOFs on the Tibetan Plateau have increased. These
+
+Image /page/15/Picture/6 description: The image displays the Springer logo, which features a black line drawing of a chess knight piece to the left of the word "Springer" written in a black serif font, all on a white background.
+
+Image /page/16/Figure/2 description: An image displaying three charts labeled a, b, and c, analyzing the frequency and causes of certain events. Chart a is a line graph showing the number of events per decade from the 1930s to the 2020s. The y-axis, labeled 'Number', ranges from 0 to 18. The data points are: 1930s (2), 1940s (1), 1950s (4), 1960s (10), 1970s (1), 1980s (7), 1990s (18), 2000s (12), 2010s (10), and 2020s (approximately 1). The peak occurred in the 1990s. Chart b is a bar chart showing the number of events per month. The y-axis, labeled 'Number', ranges from 0 to 14. The values for each month are: April (2), May (3), June (9), July (11), August (11), September (9), October (7), and N/A (14). The highest frequency is in July and August. Chart c is a pie chart illustrating the percentage distribution of event causes. The largest portion is 'Icefall' at 48.48%. Other causes include 'Moraine dam destabilization' (18.18%), 'N/A' (16.67%), 'Heavy rainfall' (6.06%), 'Landslide' (4.55%), 'Rock/Snow avalanche' (3.03%), 'Upstream inflow' (1.52%), and 'Earthquake' (1.52%). A legend below the pie chart provides the key for each colored segment.
+
+**Fig. 6** The timing and triggering factors of historical GLOF events. (a) Chronological distribution map of historical GLOF events; (b) Monthly distribution map of historical GLOF events; (c) Pie chart of triggering factors of historical GLOF events
+
+events mainly result in debris flows or floods, causing extensive damage to infrastructure such as highways, bridges, and power stations. They also lead to the silting of farmland, forests, and homes, disrupting transportation and local production, and inflicting severe economic losses. For example, the 1988 Guangxie Glacial Lake outburst destroyed 23 km of the Sichuan-Tibet highway, causing a road closure that lasted more than 200 days. Some GLOFs resulted in significant casualties. For example, the 1954 Sangwang Glacial Lake outburst claimed the lives of 400 people.
+
+## 4.2 Generating GLOF susceptibility maps
+
+### 4.2.1 Multicollinearity and importance of GLOF conditioning factors
+
+The descriptive statistics for each conditioning factor in all positive samples (historical GLOF events) are presented in Table 2. Among these factors, elevation has the smallest coefficient of variation, approximately 0.089, while glacial lake area exhibits the largest coefficient of variation at 2.32. Other factors with relatively smaller coefficients of variation include the ECV, TWI, and PGA, indicating that historical GLOF events are more concentrated around these factors. For both positive and negative samples, the conditioning factors were used as independent variables, with the occurrence of GLOF serving as the dependent variable. These variables underwent multicollinearity diagnostics, and the results are shown in Table 2. The highest VIF value is 1.768, which meets the requirement for multicollinearity diagnostics in susceptibility evaluation. Therefore, these eight factors were ultimately selected as the conditioning factors for GLOF susceptibility.
+
+The importance of each conditioning factor in the modeling process using machine learning algorithms for training samples is shown in Fig. 7. Due to the differing principles of each algorithm, the importance values for the conditioning factors vary. Among these factors, Distance to Glaciers, TWI, and Distance to Potential Icefall have importance greater than 0.1 in two models. Additionally, the MLP model identified Glacial Lake Area, Glacial Lake Replenishment Area, and Elevation as the most important factors, with the ECV also surpassing an importance of 0.1. The SVM model assigned the highest importance to Distance
+
+Image /page/16/Picture/9 description: The logo for the publisher Springer, featuring a black outline of a chess knight piece on the left, followed by the word "Springer" in a black serif font. The background is white.
+
+Table 2 Descriptive statistics and VIF of prepared conditioning factors
+
+| Conditioning factors                 | Descriptive statistics |                    |             |         |           |             |                          | VIF   |
+|--------------------------------------|------------------------|--------------------|-------------|---------|-----------|-------------|--------------------------|-------|
+|                                      | Mean                   | Standard deviation | Total       | Minimum | Median    | Maximum     | Coefficient of Variation |       |
+| Glacial lake area (m2)               | 354278.06              | 801444.42          | 19,485,300  | 6188.69 | 144929.57 | 5,813,520   | 2.32                     | 1.435 |
+| Glacial lake replenishment area (m2) | 16,644,600             | 27,457,000         | 915,452,000 | 5690.98 | 6,553,840 | 160,366,000 | 1.73                     | 1.768 |
+| Distance to glaciers (m)             | 728.68                 | 1519.38            | 40077.44    | 0       | 200.74    | 9039.55     | 2.02                     | 1.239 |
+| Elevation (m)                        | 4928.64                | 429.48             | 271075.68   | 3829.14 | 5009.42   | 5708.42     | 0.089                    | 1.578 |
+| ECV                                  | 0.0135                 | 0.00677            | 0.743       | 0.00111 | 0.013     | 0.0416      | 0.48                     | 1.580 |
+| TWI                                  | 11.25                  | 2.26               | 618.96      | 6.159   | 11.37     | 17.263      | 0.22                     | 1.584 |
+| PGA (m/s2)                           | 0.239                  | 0.0778             | 13.14       | 0.0974  | 0.246     | 0.415       | 0.34                     | 1.278 |
+| Distance to potential icefall (m)    | 14300.79               | 22711.11           | 786543.62   | 523.74  | 6219.83   | 111382.88   | 1.63                     | 1.459 |
+
+Image /page/17/Picture/4 description: The logo for Springer, featuring a black outline of a knight chess piece to the left of the word 'Springer' in a black serif font, all on a white background.
+
+to Potential Icefall, with a value of 0.39. The mean importance value of Distance to Potential Icefall across both models was notably high, reaching 0.26.
+
+### 4.2.2 GLOF susceptibility maps
+
+The SVM and MLP algorithms from the Scikit-learn library were invoked through Jupyter Notebook to calculate the conditioning factor values for the training and validation samples. Additionally, the GridSearchCV function was used for parameter optimization. After executing the models, the results were imported into the GIS platform. In the SVM model, optimal hyperparameters were obtained through GridSearchCV, with C=1000 and  $\gamma$ =0.003. The conditioning factors for each glacial lake unit in the study area were then input into the model to calculate the GLOF susceptibility index. The final GLOF susceptibility evaluation results are shown in Fig. 8a, with a value range of [0.026, 1]. Using the natural breakpoint method, the susceptibility index map was classified into five categories: very low, low, medium, high, and very high susceptibility. In the susceptibility index classification results (Table 3), 474 glacial lakes were identified as VHS, representing 4.7% of the total number of glacial lakes in the study area, with susceptibility index values ranging from 0.74 to 1.0. Additionally, 1297 glacial lakes were classified as HS, accounting for 12.7% of the total number of glacial lakes, with index values ranging from 0.528 to 0.740. In the MLP model, a three-layer neural network was constructed, consisting of an input layer, a hidden layer, and an output layer. The optimal configuration was achieved through parameter optimization, resulting in 30 neurons in the hidden layer with the Tanh activation function. The GLOF susceptibility evaluation results based on this model are illustrated in Fig. 8b, with a value range of [0.12, 0.937]. In the susceptibility index classification results (Table 3), 1286 glacial lakes were classified as VHS, accounting for 12.6% of the entire study area, with susceptibility index values ranging from 0.616 to 0.937. Additionally, 2548 glacial lakes were identified as HS, representing 25% of the study area, with index values ranging from 0.505 to 0.616. In terms of modeling accuracy, the MLP algorithm exhibits slightly higher accuracy, with an AUC value of 0.96 compared to the SVM's AUC of 0.90.
+
+# 5 Discussion
+
+## 5.1 GLOFs in response to climate change
+
+The formation of glacial lakes is closely tied to glacial movements (Liu et al. 2019). These lakes typically form at the leading edges of glaciers or in other low-lying areas, with their dams primarily composed of moraine, ice, debris, and occasionally bedrock. Over time, moraine dams can become unstable, leading to outburst events when they can no longer withstand rising water levels within the lake. This phenomenon is particularly common during warmer seasons or periods of climate warming. Global warming has led to increased temperatures in alpine and glacial regions, accelerating the melting of ice. Glaciers worldwide are retreating at an alarming rate, with many regions experiencing significant thinning and retreat (Ding et al. 2021; Jobanpreet and Chitradevi 2024). The meltwater from these glaciers replenishes glacial lakes, resulting in substantial lake expansion and destabilizing the surrounding glaciers. The debris-free glaciers on the Tibetan Plateau have retreated at
+
+Image /page/18/Picture/8 description: The logo for Springer, featuring a black and white line drawing of a chess knight piece to the left of the word "Springer" in a black serif font, all on a white background.
+
+Image /page/19/Figure/2 description: A horizontal bar chart titled "Conditioning factors" shows the importance of various factors based on three different metrics: SVM (purple), MLP (green), and Mean (orange). The x-axis represents "Importance (%)" and ranges from 0.00 to 0.40. The y-axis lists the conditioning factors. The data presented is as follows:
+- Glacial lake area: SVM is approximately 0.07, MLP is 0.16, and Mean is 0.12.
+- Glacial lake replenishment area: SVM is approximately 0.08, MLP is 0.13, and Mean is 0.11.
+- Distance to glaciers: SVM is approximately 0.17, MLP is 0.13, and Mean is 0.14.
+- Elevation: SVM is approximately 0.07, MLP is 0.13, and Mean is 0.10.
+- ECV: SVM is approximately 0.04, MLP is 0.12, and Mean is 0.08.
+- TWI: SVM is approximately 0.15, MLP is 0.12, and Mean is 0.13.
+- PGA: SVM is approximately 0.06, MLP is 0.06, and Mean is 0.06.
+- Distance to potential icefall: SVM is approximately 0.39, MLP is 0.14, and Mean is 0.26.
+The chart indicates that "Distance to potential icefall" is the most important factor, especially for the SVM model. The top of the image includes a citation: "Natural Hazards (2023) 121:17677–17705".
+
+Fig. 7 Importance of GLOF susceptibility conditioning factors
+
+Image /page/19/Figure/4 description: A figure displaying two maps, labeled 'a' and 'b', that show the Glacial Lake Outburst Flood (GLOF) susceptibility for a mountainous region. The maps cover a geographical area from 90°0'0"E to 90°30'0"E longitude and 28°0'0"N to 28°10'0"N latitude. Both maps feature a grayscale elevation background, with light blue areas indicating glaciers. A north arrow is in the top right, and a scale bar indicating 0 to 10 km is in the bottom right.
+
+Map 'a' is titled "GLOF susceptibility (SVM)". Its legend indicates an elevation range from a low of 1406 m to a high of 7452 m. The susceptibility is categorized into five levels:
+- VLS (Very Low Susceptibility): 0.026 - 0.256 (Green)
+- LS (Low Susceptibility): 0.256 - 0.392 (Light Green)
+- MS (Medium Susceptibility): 0.392 - 0.528 (Yellow)
+- HS (High Susceptibility): 0.528 - 0.740 (Orange)
+- VHS (Very High Susceptibility): 0.740 - 1 (Red)
+
+Map 'b' is titled "GLOF susceptibility (MLP)". It shares the same glacier and elevation legend as map 'a'. Its susceptibility categories are:
+- VLS (Very Low Susceptibility): 0.120 - 0.283 (Green)
+- LS (Low Susceptibility): 0.283 - 0.397 (Light Green)
+- MS (Medium Susceptibility): 0.397 - 0.505 (Yellow)
+- HS (High Susceptibility): 0.505 - 0.616 (Orange)
+- VHS (Very High Susceptibility): 0.616 - 0.937 (Red)
+
+Both maps show various colored polygons representing the susceptibility levels of different glacial lakes, with red and orange areas indicating high to very high susceptibility.
+
+Fig. 8 Partial Map of GLOF Susceptibility Evaluation Results. (a) GLOF Susceptibility Evaluation Result based on SVM. (b) GLOF Susceptibility Evaluation Result based on MLP
+
+an accelerated pace due to climate warming, with a reduction in area from  $94.59 \times 10^3$  km<sup>2</sup> to  $61.16 \times 10^3$  km<sup>2</sup> from 1988 to 2022 (Zhao et al. 2024). In the Himalayas, glaciers have shrunk by 20–30% (Nie et al. 2021). The retreat of glaciers increases the volume of glacial lakes at the glacier terminus and raises water levels, thereby increasing the risk of dam
+
+Image /page/19/Picture/7 description: The Springer logo, featuring a black and white line drawing of a chess knight piece to the left of the word 'Springer' in a black serif font, all on a white background.
+
+| Susceptibility<br>classification | SVM         |                            |                   | MLP         |                               |                             |
+|----------------------------------|-------------|----------------------------|-------------------|-------------|-------------------------------|-----------------------------|
+|                                  | Interval    | Number of<br>glacial lakes | Percentage<br>(%) | Interval    | Number<br>of glacial<br>lakes | Per-<br>cent-<br>age<br>(%) |
+| VHS                              | 0.740~1     | 474                        | 4.7               | 0.616~0.937 | 1286                          | 12.6                        |
+| HS                               | 0.528~0.740 | 1297                       | 12.7              | 0.505~0.616 | 2548                          | 25.0                        |
+| MS                               | 0.392~0.528 | 5047                       | 49.6              | 0.397~0.505 | 2602                          | 25.6                        |
+| LS                               | 0.256~0.392 | 1349                       | 13.3              | 0.283~0.397 | 2142                          | 21.1                        |
+| VLS                              | 0.026~0.256 | 2006                       | 19.7              | 0.120~0.283 | 1595                          | 15.7                        |
+
+failure. Marine glaciers, in particular, are highly sensitive to global warming, undergoing intense ablation, thinning, and retreat (Christian et al. 2022). This leads to the expansion of glacial lakes and the weakening of their dams, ultimately triggering GLOFs. This underscores the critical dependence of GLOF events on glacial dynamics.
+
+Climate change can significantly impact precipitation patterns, leading to increased intensity and frequency of rainfall events (Maity and Maity 2022). Such changes may result in a rapid rise in glacial lake levels, elevating the risk of outbursts (Liu et al. 2014). Additionally, climate change influences glacier dynamics, accelerating both forward and lateral glacier movements, which in turn increases the frequency of glacial lake formation and the likelihood of outburst events. GLOFs exhibit pronounced cyclicality and trends under varying climatic conditions. The 1960s, regarded as one of the coldest decades in Tibet (Gong et al. 2022), provided favorable climatic conditions for glacier accumulation and advance. However, extreme weather events and unusual climatic fluctuations during this transitional period increased the risk of GLOFs. This was particularly evident during warmer transitional years when icefalls into glacial lakes became a significant trigger for multiple outburst events. As the climate warmed in the 1970s and 1980s, the frequency of GLOFs declined considerably due to diminished conditions for glacier accumulation. Icefalls and associated factors remained the primary triggers during this time. By the 1990s, intensified glacier ablation, thinning, and retreat—driven by mean annual temperatures exceeding multi-year averages—resulted in new dynamics. The period from 2000 to 2016 marked the warmest 16 years in Tibet's recorded history (Chen et al. 2022; Gong et al. 2022). During this era of accelerated global warming, GLOF events surged dramatically. In addition to icefalls, triggers expanded to include landslides, heavy rainfall, and the melting of buried ice. These non-icefall-related factors accounted for more than half of the GLOF events during this period, underscoring the increasing climate sensitivity of GLOFs.
+
+The spatial differentiation of glaciers, glacial lakes, and historical GLOFs across the three subzones of the study area reveals distinct regional characteristics. The western subzone hosts the highest number of glaciers, reflecting its lower temperatures and drier climatic conditions. Its greater distance from the Indian Ocean and reduced water vapor transport result in less snowfall; however, the cooler climate supports longer glacier survival. In contrast, the eastern and central subzones, being closer to the Indian Ocean (Fig. 9), benefit from stronger water vapor transport and higher precipitation levels, creating favorable conditions for the formation of glacial lakes. The relatively lower elevation of glacial lakes in the eastern subzone is likely due to the higher water vapor content and warmer climate, which promote greater ice melt at glacier termini, facilitating glacial lake forma-
+
+Image /page/20/Picture/7 description: The logo for Springer, featuring a black line drawing of a knight chess piece on the left, followed by the word "Springer" in a black serif font, all on a white background.
+
+Image /page/21/Figure/2 description: A topographical map of the Tibetan plateau and surrounding regions, illustrating elevation and major atmospheric circulation patterns. The elevation is represented by a color scale in the bottom left, ranging from low (green, 0 m) to high (brown and white, 8848 m). The central feature is the snow-covered Tibetan plateau, with its southern edge divided into a 'Western subzone', 'Middle subzone', and 'Eastern subzone'. Several cities are marked, including New Delhi, Lasa, Dacca, Wulumuqi, Xining, Chengdu, and Guangzhou. Arrows indicate key weather systems: blue arrows from the northwest show the 'Westerlies circulation', purple arrows from the southwest represent the 'Indian Ocean monsoon', and blue arrows from the southeast depict the 'Southeast Asian monsoon'. A pink arrow points to the Bay of Bengal, labeled 'Study area'. The map also includes latitude lines for N20°, N30°, and N40°, and distance markers such as '~1200km' and '~545km' related to the monsoon paths.
+
+Fig. 9 Schematic diagram of climate background of the study area
+
+tion. The occurrence of GLOF events also varies spatially. The western subzone experiences the fewest GLOFs. In contrast, the eastern and central subzones have a higher frequency of GLOFs, driven by the greater number and larger size of glacial lakes and the higher rates of temperature increase. The varying distances from the Indian Ocean among the three subzones play a crucial role in shaping the spatial variability of water vapor transport and precipitation. The eastern and central subzones, being closer to the Indian Ocean, are more sensitive to climate change. Conversely, the western subzone, farther from the Indian Ocean and located at relatively higher latitude, experiences a drier and colder climate. Its climate is more influenced by the southern branch of the westerly circulation, resulting in a delayed response of glacial lakes to climate change. This highlights the climate sensitivity of GLOFs from another perspective.
+
+## 5.2 GLOF-prone lakes and model performance
+
+In the GLOF susceptibility evaluation, various machine learning models were employed to predict GLOF susceptibility and to calculate the weights of each evaluation factor. The accuracy of the models and their ability to effectively identify areas with high susceptibility demonstrated that both SVM and MLP are suitable for training with small datasets. Notably, all glacial lakes classified as HS and VHS by the SVM model were also classified as such in the MLP model. The MLP model, however, provided more conservative results, identifying a broader range of HS and VHS glacial lakes, accounting for 37.6% of the total. In contrast, the SVM model identified 17.4% of glacial lakes as HS and VHS. Figure 10 presents histograms of the susceptibility indices for all glacial lakes derived from both models. Although neither dataset passed the normal distribution test, the figure clearly shows that the
+
+Image /page/21/Picture/7 description: The Springer logo, featuring a black and white line drawing of a chess knight piece to the left of the word 'Springer' in a black serif font, all on a white background.
+
+Image /page/22/Figure/2 description: The image contains two histograms, labeled 'a' and 'b', presented side-by-side. Both plots have a y-axis labeled 'Count' and an x-axis ranging from 0.0 to 1.0. Plot 'a' is a histogram for 'SVM' with orange, diagonally hatched bars. Its y-axis ranges from 0 to 1600. The data distribution is skewed to the right, with the highest count reaching approximately 1500 in the bin corresponding to an x-value of about 0.35. There are vertical dashed lines at x-values of approximately 0.25, 0.35, 0.5, and 0.7. Plot 'b' is a histogram for 'MLP' with purple, diagonally hatched bars. Its y-axis ranges from 0 to a value above 1200. The data distribution is more symmetric, resembling a bell curve, with its peak count of approximately 1300 in the bin corresponding to an x-value of about 0.5. There are vertical dashed lines at x-values of approximately 0.3, 0.5, and 0.6.
+
+Fig. 10 Histogram of outburst susceptibility indices for each glacial lake (The vertical dashed lines represent the threshold values for susceptibility classification based on the natural break method). (a) Outburst susceptibility indices based SVM. (b) Outburst susceptibility indices based MLP
+
+overall mean of the MLP results is higher, further confirming the conservatism of the MLP model's outcomes. Overall, this analysis reveals that these glacial lakes with HS and VHS ratings are predominantly located in the central and eastern subregions of the study area, highlighting spatial patterns of GLOF risk. Field validation further supports the reliability of the model predictions. In particular, Rongxia Zangbu Glacial Lake, which was investigated during field surveys, was identified as highly susceptible to GLOF. The SVM model assigned it a susceptibility score of 0.992, while the MLP model yielded a score of 0.78, both indicating a high risk of outburst. This moraine-dammed lake, formed by glacial deposition at an elevation of approximately 5050 m, has exhibited significant changes in recent years. Field observations reveal that its surface area expanded from 1,714,497 m<sup>2</sup> in 2019 to 1,887,616 m<sup>2</sup> in 2023. Concurrently, the width of the terminal moraine dam decreased dramatically from 939 m in 2019 to 166 m in 2023. Given the continuous expansion of the lake area and the substantial narrowing of the moraine dam, Rongxia Zangbu Glacial Lake shows clear signs of destabilization and poses a substantial potential risk of future outburst. Similarly, although the Gebumaco Glacial Lake has been historically outburst (Wang et al. 2011), it still covers an area of 445,722 m<sup>2</sup>, and the glacier which it originates has developed ice crevasses that are at risk of triggering ice avalanches. In this assessment, the SVM result is 0.545 and the MLP result is 0.517, both indicating moderate susceptibility. These findings are largely consistent with field observations, underscoring the practical applicability of the model-based susceptibility evaluation.
+
+The most critical factors influencing GLOFs are Distance to Glaciers, TWI, and Distance to Potential Icefall, highlighting their pivotal role in GLOF dynamics. These factors are physically meaningful, as they directly reflect glacier—lake interactions and the likelihood of sudden mass movements impacting lake stability. Subsequently, factors such as Glacial Lake Area, Glacial Lake Replenishment Area, Elevation, and ECV further underscore the combined influence of geographic and climatic conditions on GLOF susceptibility. This aligns with the physical mechanisms underlying GLOF formation, which commonly involve rapid meltwater influx, direct calving of glaciers into lakes, or impacts from icefalls on moraine-dammed basins (Ahmed et al. 2021). Among all factors, Distance to Glaciers and Distance to Potential Icefall are especially significant, emphasizing the intrinsic link between glacial activity and lake instability. Glacial lakes predominantly form in close proximity to glaciers,
+
+Image /page/22/Picture/6 description: The Springer logo, featuring a black and white icon of a chess knight's head to the left of the word 'Springer' in a black serif font, all on a white background.
+
+as their genesis and hydrological input are tied to glacier melt and retreat (Ahmed et al. 2021). Moreover, historical analyses confirm that icefalls frequently act as immediate triggers for GLOF events, especially in high-relief regions such as the Himalayas and Andes. Despite relatively high coefficients of variation revealed in statistical analyses for these two proximity-related factors, their physical relevance to GLOF initiation remains indisputable, marking them as essential indicators of hazard potential. Topographic factors, including TWI, Elevation, and the ECV, also play crucial roles. These factors exhibit smaller coefficients of variation in statistical analyses, indicating that historical GLOF events tend to occur within specific ranges of TWI, elevation, and ECV, enhancing their discriminative power. TWI represents the hydrological accumulation potential of a landscape, being influenced by slope, catchment water accumulation, and terrain (Chowdhury 2023). Both TWI and ECV, as topographic derivatives, reflect terrain wetness and relief, respectively, capturing the hydrological and geomorphological context in which glacial lakes evolve. Similarly, Glacial Lake Area and Replenishment Area jointly influence the lake's hydrological budget, with their effects on stability modulated by glacier proximity and elevation. Specific elevation ranges are more conducive to the formation and persistence of glacial lakes, often coinciding with climatic and glaciological thresholds relevant to GLOF risk. The Glacial Lake Replenishment Area significantly impacts the rate and volume of water accumulation in a glacial lake, with larger areas corresponding to higher potential GLOF risks. In conclusion, GLOFs are typically triggered by a combination of interrelated factors. These critical variables not only shape the formation mechanisms of GLOFs but also profoundly influence the likelihood and scale of such disasters.
+
+This study primarily focuses on analyzing historical records to identify the characteristics and susceptibility of GLOFs. Although the study is centered on the Southern Tibetan Plateau, the proposed methodological framework—integrating historical GLOF inventories, geomorphological indicators, and machine learning classifiers—can be adapted to other regions, provided that appropriate regional datasets are available. However, the uncertainty of the current analysis is also constrained by several data- and model-related constraints. First, it does not incorporate projections or scenario modeling to assess potential trends in glacial lake expansion and GLOFs under future climate change scenarios. Given that climate change can profoundly influence glacier retreat rates, glacial lake expansion, and the frequency of GLOFs, the absence of such dynamic considerations may lead to an underestimation of future GLOF risks. Second, the spatial coverage and accuracy of historical GLOF event records are uneven, often biased toward more accessible regions with longer observation histories, introducing potential geographic bias into model training. Furthermore, the limited number of known GLOF cases creates challenges for machine learning classifiers. Although SVM and MLP demonstrate some robustness in dealing with fewer samples, their generalization ability remains limited due to the risk of overfitting. Consequently, future research should consider integrating global or regional climate models, such as the CMIP6 dataset, to project the evolution of glaciers and glacial lakes under various greenhouse gas emission scenarios. Combining these modeling results with remotely sensed monitoring and in situ observational data could facilitate the construction of a multi-source data-driven dynamic risk assessment system.
+
+Image /page/23/Picture/4 description: The logo for the publisher Springer, featuring a black and white line drawing of a chess knight piece to the left of the word 'Springer' in a black serif font. The background is white.
+
+# 6 Conclusions
+
+This study presents a comprehensive analysis of glaciers, glacial lakes, and historical GLOF events within the southern Tibetan Plateau. The findings reveal that glaciers in the study area are abundant and widely distributed, with elevations displaying a distinct pattern: lower levels in the eastern and western extremities. Glacier flow predominantly occurs northward, with slopes most commonly ranging between 20° and 25°. Accompanying the retreat of glaciers in the region, the number of glacial lakes has shown a steady year-on-year increase, particularly in the central and eastern subzones, where both the number and area of glacial lakes have grown significantly. Glacial lake areas vary considerably, with the central subzone exhibiting the largest average glacial lake area. Historically, 66 documented GLOF events have occurred within the study area, most frequently between June and October. These events were primarily triggered by icefall (48.48%), moraine dam destabilization (18.18%), and heavy precipitation (6.06%). The impact of climate warming on GLOFs has been particularly pronounced, notably during the 1960s and from 1990 to 2020, when rising temperatures and shifting precipitation patterns intensified the frequency of these events. The study underscores the significant dependence of GLOF events on glacial dynamics and their sensitivity to climatic variations, emphasizing the interplay between glacier changes and climate shifts in shaping the risk and occurrence of these hazardous events.
+
+The study employed two machine learning algorithms, SVM and MLP, to predict the GLOF susceptibility based on historical events. The evaluation results indicated that approximately 17.4% of glacial lakes were classified into HS and VHS, with higher susceptibility predominantly observed in the central and eastern subzones. The MLP model demonstrated slightly better accuracy than the SVM model, achieving AUC values of 0.96 and 0.90, respectively. Additionally, further analysis highlighted that the MLP model produced more conservative predictions regarding GLOF susceptibility.
+
+Overall, climate warming and glacier retreat have intensified the expansion and instability of glacial lakes, leading to an increased frequency and severity of GLOF events. The findings highlight the spatial and temporal patterns of GLOF events and their connection to climate change, providing valuable insights for guiding the prevention and mitigation efforts. The integration of machine learning technology has deepened our understanding of the complex nonlinear relationships between environmental factors and GLOF occurrences, providing an advanced and effective method for risk assessment in the region. Moving forward, future research should focus on integrating real-time remote sensing data, in situ observations, and advanced forecasting techniques to establish dynamic monitoring frameworks and early warning systems. These efforts will provide a robust scientific foundation for disaster prevention and management in alpine glacier regions.
+
+**Acknowledgements** This study was supported by the Central Guidance for Local Science and Technology Development Foundation of China (Grant No.2024ZYD0159), Open Project of Major Hazard Measurement and Control Key Laboratory of Sichuan Province (Grant No.KFKT-202306) and the Talent Introduction Project of Xihua University (Grant No.Z231013).
+
+**Funding** Sichuan Science and Technology Program (2025YFNH0008). Open Project of Major Hazard Measurement and Control Key Laboratory of Sichuan Province (KFKT-202306). Talent Introduction Project of Xihua University (Z231013).
+
+Image /page/24/Picture/8 description: The logo for the publisher Springer, featuring a black outline of a chess knight's head facing left, positioned to the left of the word "Springer" in a black serif font. The entire logo is set against a plain white background.
+
+# **Declarations**
+
+**Competing interests** The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
+
+# References

chatbot/rag/data/('Predicting the Risk of Glacial Lake Outburst Floods in Karakorum', '.pdf')_extraction.md

@@ -0,0 +1,388 @@
+
+**Abstract.** Glacier snouts respond to climate change by forming proglacial meltwater lakes, thereby influencing glacier mass balance and leading to advancements and surges. The positive feedback of climate change results in more frequent ice-dammed glacial lake outburst floods (GLOFs) in the Karakorum and surrounding regions, often facilitated by englacial conduits. However, the complex and multi-factor processes of conduit development are challenging to measure. Determining the lake depths that might trigger GLOFs and the numerical model specifications for breaching are still being determined. Empirical estimates of lake volumes, along with field-based monitoring of lake levels and depths and the assessment of GLOF risks, enable warnings and damage mitigation. Using historical data, remote sensing techniques, high-resolution imagery, crosscorrelation feature-tracking, and field-based data, we identified the processes of lake formation, drainage timing, and triggering depth. We developed empirical approaches to determine lake volume and trigger water pressure leading to a GLOF. The correlation of glacier surge and lake volume reveals that glacier surge velocity plays a crucial role in lake formation and controlling the size and volume of the lake. Lake volume estimation involves geometric considerations of the lake basin shape. A GLOF becomes likely when the lake's non-dimensional depth (n') exceeds 0.60, correlating with a typical water pressure on the dam face of 510 kPa. Additionally, we identified the critical risk zone of lakes, where all lake outburst floods occur, as the point where the lake volume reaches or exceeds 60% of its capacity. These field-based and empirical findings not only offer valuable insights for early warning procedures in the Karakorum but also suggest that similar approaches can be effectively applied to other mountain environments worldwide where GLOFs pose a hazard.
+
+## 30 1. Introduction
+
+Globally, glacier shrinkage is a strikingly visible sign of climate change. However, within High Mountain Asia (HMA), particularly the Karakorum, Kunlun Shan, and Eastern Pamirs, the glaciers have been gaining mass since 1970 (Berthier and Brun, 2019; Gardelle et al., 2012; Kääb et al., 2015; Minora et al., 2013; Yao et al., 2012). This positive response to climate change consequently influences glacier dynamic behaviours, with the HMA glaciers thickening, increasing glacier surges, and
+
+Image /page/1/Picture/1 description: A Creative Commons Attribution (CC BY) license icon. The icon is a grey rectangle with a black bar at the bottom. On the grey section, there are two white circles with black symbols inside. The left circle contains the letters 'CC', and the right circle contains the attribution symbol, which is a stylized person. On the black bar at the bottom, the letters 'BY' are written in white.
+
+65
+
+Image /page/1/Picture/2 description: The image shows the logo for EGUsphere, a preprint repository. The logo consists of the word "EGUsphere" in blue. The letters "EGU" are capitalized and enclosed within a blue circular gear-like shape. The word "sphere" is in lowercase. Above the word "sphere," there are two swooshes, a blue one on top and a gray one underneath. Below the main text, the words "Preprint repository" are written in a smaller, gray, sans-serif font. The background is white.
+
+advancing glacier termini throughout the region (Bhambri et al., 2013; Bolch et al., 2017). This behaviour contrasts neighbouring regions with negative glacier mass budgets, such as the Himalayas, Hindukush, and Tibet (Bazai et al., 2021; Bolch et al., 2011; Frey et al., 2014). In the latter areas, glaciers continue to shrink, thinning and reducing volume, showing no significant glacier advance (Dehecq et al., 2019; Farinotti et al., 2020; Yao et al., 2012). As a result, the Moraine Lake formation has increased the number of GLOFs in the glacier-retreating regions (Yong et al., 2017) as well as the number of glacier avalanches as increased (Byers et al., 2023; Kääb and Girod, 2023; You and Xu, 2022). However, in the glacier advance regions, the positive variation in regional climate feedback has prompted the rapid formation of ice-dammed lakes accompanied by sudden releases of meltwater originating from these lakes (Carling, 2013; Hewitt, 1982; Hewitt, 1998; Hewitt and Liu, 2010).
+
+The mechanisms and frequency of ice dam glacial lake outburst floods (GLOFs) still need to be better understood, hindering accurate prediction (Bazai et al., 2021; Cook et al., 2016; Harrison et al., 2018; Richardson and Reynolds, 2000). Recent studies have investigated changes in frequency due to climate change (Rick et al., 2023; Veh et al., 2023) and conducted global assessments of flood volume and risk (Rick et al., 2023). Despite these efforts, understanding the drainage and predicting flood events from ice-dammed lakes remain challenging. Predicting these events is crucial due to their potential to cause devastating impacts on human lives and livelihoods, ecosystems, infrastructure (*e.g.*, roads, bridges, hydropower systems), river channel stability, and effects on agriculture and fisheries (Carrivick and Tweed, 2016; Cook et al., 2016; Emmer, 2017; John et al., 2000; Neupane et al., 2019; Zhang et al., 2022).
+
+Ice-dammed Lake floods represent the dominant hazard in cryospheric regions, comprising 70% of GLOFs. In contrast, moraine-dammed lakes contribute only 9% (with the remaining 16%, 3%, and 2 % triggered by unknown dam types, volcanic activity, and bedrock, respectively) (Carrivick and Tweed, 2016). GLOFs have been recorded causing damage up to 120 km from moraine-dammed lakes (Richardson and Reynolds, 2000) and around 500 km from ice-dammed lakes (Hewitt and Liu, 2010). The resulting impacts include hundreds of human fatalities and the other impacts noted above (Carrivick and Tweed, 2016; Cui et al., 2014; Cui et al., 2015; Kreutzmann, 1994; Mason, 1929; Stuart-Smith et al., 2021; Zhang, 1990; Zheng et al., 2021). While attempts have been made to explore the breaching mechanisms of moraine lake outburst floods, triggered by ice or debris falls, strong earthquake shaking, internal piping, or overtopping waves that exceed the shear resistance of the dam (Emmer and Vilímek, 2013; Richardson and Reynolds, 2000), the understanding of the complex and unclear mechanisms of ice dam lake outburst floods remains a challenge (Werder et al., 2010), making a prediction using complex modelling currently impossible. Therefore, there is an urgent need for simplified approaches to GLOF prediction to mitigate downstream impacts.
+
+Despite the uncertainty related to the detail of GLOF initiation, sudden glacier advances during surge cycles have a prominent role in the formation of ice-dammed lakes by creating an ice barrier in the valleys, particularly at narrow valley floor sections and at confluences (Bazai et al., 2021; Bhambri et al., 2019), damming minor and major rivers. Glacier surges are the main
+
+Image /page/2/Picture/1 description: A rectangular Creative Commons license icon. The icon has a grey background on the top and a black bar on the bottom. On the left side of the grey area is the Creative Commons 'CC' logo inside a white circle. To its right is the Attribution symbol, a black stick figure of a person inside a white circle. In the black bar below the Attribution symbol, the letters 'BY' are written in white.
+
+70
+
+Image /page/2/Picture/2 description: The image shows the logo for EGUsphere. The logo consists of the text "EGUsphere" in blue. The letters "EGU" are bold and enclosed within a blue circular, gear-like shape. The word "sphere" follows, with a capital "U". Above the word "sphere," there are two swooshes, a blue one on top and a gray one below. Underneath the main logo text, the words "Preprint repository" are written in a smaller, gray font. The entire logo is on a white background.
+
+agents related to the formation of ice-dammed lakes in the Swiss Alps (Haeberli, 1983), Northern Norway (Xu et al., 2015), Argentinian Patagonia (Vandekerkhove, 2021), Alaska (Trabant et al., 2003), Karakorum, and in the Pamir (Bazai et al., 2021; Hewitt and Liu, 2010) and Tianshan regions (Ng, 2007; Shangguan et al., 2017). Recent studies reveal that the draining processes of ice-dammed lakes potentially involve subglacial breaching, overspill, rapid ice mass instability, and slow deformation of subglacial cavities (Björnsson, 2003; Haemmig et al., 2014; Round et al., 2017). Several attempts have been made to explore the drainage behaviour of ice-dammed lake outburst floods (Hewitt and Liu, 2010). However, due to the remoteness, danger, and inhospitable terrain where such lakes can be found, real-time data are few, and significant gaps remain in our knowledge of these processes.
+
+In the Karakorum, ice-dammed lakes are found in five major valleys, three of which are densely populated and highly vulnerable to unexpected GLOF events. Recent advances in understanding have been made (Bazai et al., 2021) concerning the formation of episodic ice-dammed lakes, which, due to ice mass transfer variations, are linked to the changes in the glacier surface velocity and fluctuations in the crevasse density during the surge cycle (Rea and Evans, 2011; Sharp, 1985). Despite these advancements in knowledge, globally, the techniques for measuring and estimating lake volume (the volume of the lake before an outburst), the critical depth (for GLOF release), and timely prediction of ice-dammed lakes remain largely unexplored and unidentified (Round et al., 2017; Shangguan et al., 2016; Steiner et al., 2018). Very limited ice-dammed lake volume data are available. Still, these ice-dam lake volumes were measured either while the lake was empty (after the GLOF events) or partially filled and thus shallow (Round et al., 2017; Shangguan et al., 2016; Steiner et al., 2018). The downstream threat from those lakes is extremely high, while their full lake volumes are unknown. To measure the flood volume and flood magnitude for a deep and potentially full lake, the lake volume measurement is recognized as a critical variable that needs to be accurately calculated or at least well-estimated (Bazai et al., 2021; Bazai et al., 2022). An accurate estimate of lake volume will also help explore the timing, triggering depth of the lake, and frequency of ice-dammed lake outburst floods in relation to surge cycles. Timing information can be approximated by correlating glacier velocities and GLOF occurrences (Bazai et al., 2021; Bazai et al., 2022). The current research focuses of Bazai and colleagues aims to leverage historical records, regular field investigations, and remote sensing techniques to develop a comprehensive understanding, which should lead to better prediction of GLOFs through modelling dam failure mechanisms. Herein, the primary objective is to enhance predictive capabilities regarding GLOF event timing by refining empirical lake volume estimation and identifying critical depths for future risk reduction.
+
+## 2. Methodology:
+
+100
+
+### 2.1. Data acquisition
+
+The identification and detection of the Shishper, Khurdopin, and Kyager ice-dammed lakes was accomplished using all available open and commercial satellite imagery sources from 1970 to 2022. The datasets include 590 images of Landsat 2-5, 7-9 and 45 images from Sentinel-2, downloaded from the United States Geological Survey (USGS) website
+
+Image /page/3/Picture/1 description: A rectangular Creative Commons license icon, specifically the CC BY (Attribution) license. The icon is divided horizontally into a grey upper section and a black lower section. The grey section contains two white circles with black outlines. The left circle has the letters 'CC' in black inside it. The right circle contains a black stick figure icon, representing attribution. The black lower section has the letters 'BY' in white.
+
+105
+
+110
+
+115
+
+120
+
+125
+
+Image /page/3/Picture/2 description: The logo for EGUsphere, a preprint repository. The logo features the text "EGUsphere" in a blue sans-serif font. The letters "EG" are inside a blue circular gear-like shape. Above the "sphere" part of the text, there are two swooshes, one blue and one gray. Below the main text, the words "Preprint repository" are written in a smaller, gray font.
+
+(https://earthexplorer.usgs.gov/) (Table S1). The commercial high-resolution images consisted of 35 images from Gaofen-1 (GF-1) and Gaofen-2 (GF-2), 11 images from SPOT-6 and SPOT-7 and five images from Global Planet (https://data.cresda.cn/#/2dMap, https://earth.esa.int/eogateway and https://www.planet.com/products/planet-imagery, respectively). The following DEM datasets have been used for measuring lake volume, depth and dam height: the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) and the Phased Array type L-band Synthetic Aperture Radar (PALSAR)-DEM data scenes from the National Aeronautics and Space Administration (NASA) Earth Science Data Center website (https://search.earthdata.nasa.gov/). KH-9 and Shuttle Radar Topography Mission (SRTM) data were downloaded from http://earthexplorer.usgs.gov/ (Table S2). For more precise and high-resolution data, a field survey of the Shishper and Khurdopin glacier lakes was conducted in 2018, 2019, 2021, and 2022 and used Unmanned Aerial Vehicles (UAV) to determine annual lake extents, lake depths, glacier heights, and thickness, termini positions, and glacier surface displacements. The glacier outlines were obtained from the Randolph Glacier Inventory (RGI 6.0) (Consortium, 2017) and modified according to surge movements with time (https://www.planet.com/products/planet-imagery/).
+
+### 2.2. Glacier Lake mapping and glacier surface velocity
+
+Satellite imagery had a spatial resolution of 0.8 to 30 m (Table S1). The images were selected based on the visibility of the glacier surface and lake areas, and overall, 23 ice-dammed lakes from eight surge events were identified through optical images related to the Khurdopin, Kyager, and Shishper glaciers (Table 1). The presence of lakes was determined based on the Normalized Difference Water Index (NDWI) (McFeeters, 1996), and the outlines of all 23 lakes were digitized manually using Landsat false-color composites (near-infrared, red, and green bands) to distinguish water bodies from other objects (Huggel et al., 2002). The extent of six lakes from Shishper and Khurdopin that occurred after 2017 were obtained in the field using GPS (G639; Accuracy: Single: 1 ~ 3m; SBAS: 0.6m) survey points along the lake shorelines (Fig.1a-d), as well as from Unmanned Aerial Vehicle (UAV) generated Digital Surface Models (DSMs). Field trips were conducted for Khurdopin in 2017 and 2018 and for Shishper in 2019, 2021 and 2022. Alternatively, high-resolution satellite imagery from Planet (3 m) and GF-1 and 2 (0.8 m to 4 m resolution, respectively) and SPOT-6 and SPOT-7 (1.5 m) were used to extract the lake boundaries. The use of high-resolution imagery aims to obtain accurate lake surface levels, and the latter is used for measuring volume and lake depth. The coupled lake extent and polygon help reduce the uncertainty of the lake extent obtained for Landsat 2-5 images. The above method was used to extract the extent of the previous lakes of Khurdopin and Kyager glacier, previously published (Bazai et al., 2021; Bazai et al., 2022), which data are incorporated into the current analysis, and the same method was used to extract the extent of Shishper lakes. These lake polygons and extents were further used to measure lake volume and depth mediated by field derived lake levels as appropriate (Bazai et al., 2022).
+
+Image /page/4/Picture/1 description: A Creative Commons Attribution (CC BY) license icon. The icon is a rectangle with a grey top half and a black bottom half. On the left side of the grey section is the Creative Commons logo, which consists of the letters "CC" in black inside a white circle. To the right of that is the attribution symbol, a black stick figure inside a white circle. On the black bottom section, below the attribution symbol, the letters "BY" are written in white.
+
+Image /page/4/Picture/2 description: The logo for EGUsphere, a preprint repository. The logo features the text "EGUsphere" in blue, with the letters "EGU" enclosed in a blue, gear-like circle. Above the word "sphere," there are two swooshes, one blue and one gray. Below the main text, the words "Preprint repository" are written in a smaller, gray font. The background is white.
+
+Image /page/4/Figure/3 description: A figure composed of five photographic panels, labeled (a) through (e), showing observations of Shishper Lake and Khurdopin Lake.
+
+Panel (a) is titled "Shishper Lake March 2019" and shows an aerial view of a murky green glacial lake surrounded by rocky, snow-dusted slopes. A blue arrow points across the water, and a dotted rectangle labeled "Panel b" highlights a section of the shoreline.
+
+Panel (b), titled "Shishper Lake Shoreline," shows a close-up of the rocky shoreline. Four colored dashed lines indicate the receding water levels from 2019 (blue, highest), 2020 (red), 2021 (green), to 2022 (light green, lowest).
+
+Panel (c), titled "Khurdopin Lake February 2018," shows a river flowing through a valley filled with jagged glacial ice. A blue wavy arrow indicates the direction of water flow.
+
+Panel (d), titled "Khurdopin Lake March 2018," shows the same area as panel (c) a month later, where the river has formed a lake with a surface of broken ice chunks.
+
+Panel (e), titled "Khurdopin Lake Shoreline," shows the shoreline of the lake against the glacier. Two dashed lines indicate the water levels for 2017 (black, lower) and 2018 (red, higher), with a blue wavy arrow pointing towards the lake.
+
+**Fig. 01** Shishper and Khurdopin glacial lake view in the field. a) Oblique view from a helicopter in 2019 (image captured in March 2019 during lake monitoring (Image by Gilgit-Baltistan Disaster Management Authority (GBDMA)); b) Shishper shoreline elevations of four lakes that outburst in the given years (image captured in June 2022); c and d are the regular monitoring views of the lake in the field: e) are the lake elevations in the given years. The Khurdopin Lake elevations were identified in the field in April 2018.
+
+The surface velocities of the Khurdopin, Shishper, and Kyager glaciers from 1990 to 2022, highlighting the quiescent and surge phases, were obtained from published data (Bazai et al., 2021; Bazai et al., 2022) using image-to-image correlation open-source software COSI-Corr (Leprince et al., 2012; Leprince et al., 2007). The software effectively assesses the glacier surface velocity (Leprince et al., 2012; Steiner et al., 2018). Utilizing a displacement calculation, this technique was used to co-register and correlate surface features (Bazai et al., 2021; Steiner et al., 2018). The surface velocity and overall movement during the surge were monitored by observing changes in the GPS-registered glacier front positions every three months from March 2019 and measured for three years for the Shishper glacier in the field. When coupled with COSI-Corr measure velocities, these latter procedures gave precise results. The velocity estimation procedure generally yields an accuracy of ¼ of a pixel (Sattar et al., 2019). We estimated velocity root-mean-square errors (RMSE) to justify the image processing accuracy.
+
+145
+
+140
+
+Image /page/5/Picture/1 description: The Creative Commons Attribution license icon, also known as CC BY. The icon is a rectangle divided horizontally. The top part is gray and features two white circles. The left circle contains the letters 'CC' in black, and the right circle contains a black stick figure icon, representing attribution. The bottom part of the rectangle is black with the white letters 'BY'.
+
+150
+
+Image /page/5/Picture/2 description: The logo for EGUsphere, a preprint repository. The logo features the text "EGUsphere" in blue. To the left, the letters "EGU" are enclosed in a blue circular gear-like symbol. Above the "sphere" portion of the text, there are two swooshes, one blue and one gray. Below the main text, the words "Preprint repository" are written in a smaller, gray font.
+
+The Shishper, Khurdopin and Kyager glaciers are surge-type glaciers (Copland et al., 2011; Hewitt, 1998). Since 1972, eight surge events have been occurred from Khurdopin (three surges), Kyager (three surges), and Shishper (two surges) (Table 1) at an interval of 17–20 years for three glaciers. The Landsat 2-4 images from 1970 to 1990 have errors in the selected glacier area. Therefore, the initial surges for Khurdopin and Kyager between 1970 and 1989 were not considered for estimating the annual velocity. Orthorectified Landsat scenes from TM to OLI–2 and sentinel 2 were used to estimate the yearly and event-based velocities of all three glaciers from 1989 to 2022 to obtain information about the surge events and glacier front changes. Within this period (1989 to 2022), some satellite images were absent, and for precise results, cloud-free images were chosen each year.
+
+### 2.3. Field observation and lake volume measurement
+
+Field observations were performed in the Shimshal and Shishper Valley for the Khurdopin glacier in 2017, 2018, and for 155 Shishper in 2019, 2021, and 2022 to investigate glacier surges, glacier front dynamics, and ice-dammed lakes: lake level and lake depth. The latter value is from shoreline elevations and the elevation of the ice dam at its lowest point (Table 1 and Figure 1). Six lakes were regularly monitored: four from the Shishper Glacier and two from the Khurdopin Glacier. 23 GLOF events from eight surge cycles that occurred during the first year of the surge, or following, are presented in Table 1, with lakes resealing after GLOFs. The field data for six events from the Khurdopin and Shishper glaciers helped to reduce the uncertainty 160 and error for the data related to 17 lake extractions through remote sensing techniques noted above. All the previously recorded lakes from Khurdopin and Shishper were drained via single subglacial conduits with stable inlets and varying outlet positions and lengths. As closely as possible, we identified the inlet and outlet positions of the drainage conduits. The inlet position of the conduit in the ice-dammed lake basin is always in the deepest position, an important factor in determining the lake's depth. 165 The lowest ice dam height also tended to be in the vicinity of the conduit. The inlet positions were geolocated in empty lake basins using GPS, and the lake depth was calculated for these locations. From the field survey, we identified the approximate position of conduits from the presence of surface depressions in the glacier. We estimated the curvilinear conduit lengths between the inlet and outlet.
+
+170
+
+Image /page/6/Picture/1 description: A Creative Commons Attribution (CC BY) license icon. The icon is a rectangle with a gray upper section and a black lower section. On the left side of the gray section is a white circle containing the letters 'CC' in black. On the right side is another white circle containing a black icon of a person. The black lower section contains the letters 'BY' in white.
+
+Image /page/6/Picture/2 description: The image shows the logo for EGUsphere. The text "EGUsphere" is written in a blue sans-serif font. The letters "EGU" are enclosed within a blue circular gear-like shape. Above the word "sphere", there are two curved lines, one blue and one gray. Below the main text, the words "Preprint repository" are written in a smaller, gray font.
+
+180 **Table 1:** GLOF and Surge date and Lake volume measurement since 1970 using remote sensing technique. The average surge velocity of each of the 23 GLOFs from eight surge cycles is presented.
+
+| No.<br>s | Glacier<br>Name | Date       | Sensor | Lake<br>Area<br>km2 | Aster/UAV<br>Lake<br>Volume<br>Estimate<br>(106 m3) | Lake Vol<br>Uncertainty<br>(+/-106 m3) | Vol. after<br>[106 m3] | Average<br>velocity<br>(m/d) | Date of<br>next clear<br>image after<br>GLOF | Type of<br>drainage<br>Complete<br>(C)<br>Partial (P) | Surge cycle<br>and<br>Resealed<br>GLOF | Surge<br>Duration in<br>months            |
+|----------|-----------------|------------|--------|---------------------|-----------------------------------------------------|----------------------------------------|------------------------|------------------------------|----------------------------------------------|-------------------------------------------------------|----------------------------------------|-------------------------------------------|
+|          | Khurdopin       | 20/08/1977 | LM02   |                     |                                                     |                                        |                        |                              |                                              | P                                                     | 1977–1979                              | May 1977–<br>Aug 1979;<br>27 Months       |
+| 1        | Khurdopin       | 15/8/1999  |        |                     |                                                     |                                        |                        | 0.33                         |                                              | C                                                     | 1998-1999                              | Jan 1995 to<br>Sep 2002;<br>92 Months     |
+| 2        | Khurdopin       | 05/30/2000 | L5 TM  | 1.87                | 186                                                 | 2.1                                    | x                      | 0.33                         | 08/26/2000                                   | C                                                     | Resealed                               |                                           |
+|          | Khurdopin       | 04/07/2001 | LE07   | 0.295               | 19.5                                                | 1.5                                    | x                      | 0.44                         | 06/26/2001                                   | C                                                     | Resealed                               |                                           |
+| 4        | Khurdopin       | 07/15/2002 | LE07   | 0.60                | 52.1                                                | 1.6                                    | 2.2                    | 0.87                         | 08/16/2002                                   | P                                                     | Resealed                               |                                           |
+| 5        | Khurdopin       | 07/28/2017 | LE07   | 0.180               | 16.2                                                | 1.4                                    | x                      | 1.41                         | LC 08<br>08/01/2017                          | C                                                     | 2016-2018                              | June 2006<br>to Aug<br>2009: 38<br>months |
+| 6        | Khurdopin       | 03/18/2018 | LC08   | 0.402               | 19.8                                                | 0.9                                    | x                      | 0.53                         | 02/25/2018                                   | C                                                     | Resealed                               |                                           |
+| 7        | Kyager          | 08/01/1977 | LM02   | 1.181               | 40.73                                               | 5.8                                    | x                      |                              | 10/14/1977                                   | C                                                     | 1976-1977                              | Jan 1975 to<br>Aug 1978;<br>43 Months     |
+| 8        | Kyager          | 07/18/1978 | LM02   | 2.17                | 82.12                                               | 15.6                                   | x                      |                              | 06/07/1979                                   | C                                                     | Resealed                               |                                           |
+| 9        | Kyager          | 03/08/1997 | L5 TM  | 3.30                | 127.3                                               | 2.9                                    | x                      | 0.4                          | 04/09/1997                                   | C                                                     | 1994-1996                              | Jan 1995 to<br>Sep 2002;<br>92 Months     |
+| 10       | Kyager          | 09/10/1998 | L5 TM  | 3.32                | 133.5                                               | 3.5                                    | x                      | 0.3                          | 10/11/1998                                   | C                                                     | Resealed                               |                                           |
+| 11       | Kyager          | 09/07/1999 | L7ETM+ | 2.19                | 86.12                                               | 1.23                                   | x                      | 0.46                         | 08/17/1999                                   | C                                                     | Resealed                               |                                           |
+| 12       | Kyager          | 06/25/2000 | L7ETM+ | 0.91                | 23.48                                               | 1.12                                   | x                      | 0.49                         | 08/03/2000                                   | C                                                     | Resealed                               |                                           |
+| 13       | Kyager          | 09/08/2002 | L7ETM+ | 2.93                | 115.19                                              | 1.09                                   | x                      | 1.29                         | 10/09/2002                                   | C                                                     | Resealed                               |                                           |
+| 14       | Kyager          | 06/14/2008 | L5 TM  | 1.45                | 94.95                                               | 1.65                                   | x                      | 0.61                         | 23/06/2008                                   | C                                                     | Resealed                               | June 2006<br>to Aug<br>2009: 38<br>months |
+| 15       | Kyager          | 07/28/2009 | L5 TM  | 1.39                | 91.35                                               | 1.56                                   | x                      | 0.56                         | 04/08/2009                                   | C                                                     | Resealed                               |                                           |
+| 16       | Kyager          | 07/16/2015 | L8 OLI | 1.56                | 53.5                                                | 0.87                                   | x                      | 1.14                         | 05/08/2015                                   | C                                                     | Resealed                               | Jan 2013 to<br>Aug 2018:<br>67 months     |
+| 17       | Kyager          | 07/14/2016 | L7ETM+ | 1.63                | 45.89                                               | 1.49                                   | 2.9                    | 0.53                         | 30/07/2016                                   | P                                                     | 2014-2016                              |                                           |
+| 18       | Kyager          | 07/31/2016 | L7ETM+ | 1.48                | 44.32                                               | 1.23                                   | x                      | 0.38                         | 08/09/2016                                   | C                                                     | Resealed                               |                                           |
+| 19       | Kyager          | 08/10/2017 | L8 OLI | 2.91                | 113.99                                              | 0.73                                   | 11.9                   | 0.38                         | 26/08/2017                                   | P                                                     | Resealed                               |                                           |
+| 20       | Kyager          | 08/06/2018 | L8 OLI | 2.38                | 87.98                                               | 0.53                                   | x                      | 0.29                         | 08/29/2018                                   | C                                                     | Resealed                               |                                           |
+| 21       | Shishper        | 06/23/2019 | L8 OLI | 0.37                | 24.10                                               | 2.1                                    | x                      | 0.95                         | 07/13/2019                                   | C                                                     | 2017-2019                              | Dec 2018 to<br>June 2022:<br>41 Months    |
+| 22       | Shishper        | 05/29/2020 | L8 OLI | 0.50                | 24.90                                               | 1.5                                    | x                      | 0.46                         | 06/22/2020                                   | C                                                     | Resealed                               |                                           |
+| 23       | Shishper        | 05/16/2021 | L8 OLI | 0.52                | 25.77                                               | 1.4                                    | x                      | 0.29                         | 07/15/2021                                   | C                                                     | Resealed                               |                                           |
+|          | Shishper        | 05/07/2022 | L8 OLI | 0.41                | 27.66                                               | 1.1                                    | x                      | 0.24                         | 05/10/2022                                   | C                                                     | Resealed                               |                                           |
+
+Image /page/7/Picture/1 description: A rectangular icon for the Creative Commons Attribution license. The icon has a grey upper section and a black lower section. The grey section contains two white circles: the left one has the letters 'CC' in black, and the right one has a black stick figure symbol for attribution. The black lower section contains the letters 'BY' in white.
+
+185
+
+190
+
+195
+
+200
+
+205
+
+210
+
+Image /page/7/Picture/2 description: The logo for EGUsphere, a preprint repository. The logo features a blue circular gear-like symbol with the letters "EG" inside. To the right, the name "EGUsphere" is written in blue, with "EGU" in a bold font and "sphere" in a lighter font. Above the word "sphere," there are two swooshes, one blue and one grey. Below the main name, the words "Preprint repository" are written in a smaller, grey font.
+
+In addition, we used a UAV (DJI Mavic 2 Pro) equipped with a high-resolution camera (4000 pixels × 2250 pixels) to obtain multiple aerial photographs with a minimum of 85% image overlap (Entwistle and Heritage, 2017; Entwistle and Heritage, 2019; Tonkin and Midgley, 2016). The UAV flew at a low uniform height (500 m - to reduce the image distortion) to generate high-resolution orthomosaics and DSMs of the glacier lake surfaces, empty lake basins, and glacier termini. The integral of the difference between the elevation of the lake surface and the elevation of the bottom in the extent of the lake obtains the volume. The lake volume data for the glacial lakes of Khurdopin and Kyager were presented previously (Bazai et al., 2021; Bazai et al., 2022), and the same approach was applied to obtain the Shishper glacier lake volume and depth data (Table 1). In addition to UAV data, we utilized data from KH-9 (1974), ASTER (2000-2019), PALSAR-DEM (June 2008), and SRTM (February 2000) for the computation of lake volumes (Table S2). The SRTM DEM without voids serves as the reference dataset, and the vertical uncertainties of the SRTM DEM are reported to be ±10 m (Rodriguez et al., 2006). The corrected DEMs from Karakorum regions were used by Bazai et al. (2021) and Gardelle et al. (2013).
+
+### 2.4. Geometry of lake basin
+
+Considering that the lake area in satellite images often exhibits a triangular planform (e.g., Fig 1a), we explored the possibility of using a geometric shape to approximate the volume of the lake basin. It is imperative to accurately extract boundary parameters, including lake elevation, length, width, and depth, and this information was manually acquired. Employing NDWI, we identify lake outliers through Landsat false-color composites, which use near-infrared, red, and green bands to distinguish water bodies from other features. We employ standard connected component analysis to (Dillencourt et al., 1992) to manually calculate the area and the perimeter and determine other dimensions of each lake. Initial calculations are pixel-based and later converted to metric units by multiplying pixel counts with their respective pixel sizes. The pixel size for high-resolution images varied from 0.8 to 3m and was cross-validated with Khurdopin and Shishper glacier lakes UAV data having a pixel size of 0.063m as well as with field survey evidence. Trials demonstrated that the known volume of the lakes determined using DEMs of the lake basins once drained could be approximated if the length of the lake from the upstream inlet to the ice-dam face (Z) and the breadth of the lake at the ice dam (C) are known. Improved estimates are obtained if the lake's depth (h) is known at the deepest point close to the dam face, which value can be obtained from a DEM of the drained basin. Alternatively, where a lake is present, this latter parameter can be obtained by plumbing the depth from a boat.
+
+The first consideration was whether the valley sides might be considered to provide a V-shaped lake cross-section. Each lake volume (V) was approximated by an irregular tetrahedron (Fig. 2a left-hand panel) where the depth (h) is unknown, but the distance from A to B (X in Fig. 2a) and the length C are known values. Assuming the lake surface is an isosceles triangle, and the vertical face at the dam wall is an equilateral triangle, the volume can be obtained from:
+
+$$V = \sqrt{V^2}$$
+ (1)
+
+$$V^{2} = \frac{1}{144} \left[ Y_{1}^{2} D^{2} (Z_{2}^{2} + X^{2} + C^{2} + E^{2} - Y_{1}^{2} - D^{2}) + Y_{2}^{2} E^{2} (Y_{1}^{2} + X^{2} + C^{2} + D^{2} - Y - E^{2}) + Z^{2} C^{2} (Y_{1}^{2} + Y + D^{2} + E^{2} - Z^{2} - C^{2}) - Y_{1}^{2} Y_{2}^{2} C^{2} - Y_{2}^{2} Z^{2} D^{2} - Y_{1}^{2} Z^{2} E^{2} - C^{2} D^{2} E^{2} \right]$$
+
+Image /page/8/Picture/1 description: A Creative Commons Attribution (CC BY) license icon. The icon is a grey rectangle with a black bar at the bottom. On the grey part, there is a white circle with 'CC' inside on the left, and a white circle with a person icon on the right. The black bar at the bottom contains the letters 'BY' in white.
+
+220
+
+225
+
+230
+
+235
+
+Image /page/8/Picture/2 description: The logo for EGUsphere, a preprint repository. The logo features the text "EGUsphere" in blue. The letters "EGU" are inside a blue, gear-like circle. Above the word "sphere," there are two swooshes, one blue and one gray. Below the main text, the words "Preprint repository" are written in a smaller, gray font.
+
+Where the values for lakeside lengths  $Y_1$  and  $Y_2$ , the main length Z, lakeside length D, and lakeside length E are defined in Fig. 2a and obtained from geometry.
+
+Image /page/8/Figure/4 description: The image contains two panels, labeled (a) and (b).
+
+Panel (a) displays two diagrams of a three-dimensional shape, likely representing a lake basin formed by an ice dam. The diagrams show an irregular tetrahedron with vertices labeled A, B, C, D, and E. The shape is divided horizontally by a dashed line, separating the 'OPEN VALLEY' from the 'ICE DAM'. Various dimensions are labeled, including Y1, Z, X, and h. An arrow indicates 'Inflow' at the top vertex A.
+
+Panel (b) is a scatter plot comparing two methods of volume calculation. The y-axis is labeled 'Lake Volume from DEM, V\_DEM (M m^3)' and ranges from 0 to 200. The x-axis is labeled 'Volume of Irregular Tetrahedron/10, V\_Tet (M m^3)' and ranges from 0 to 100. The plot contains data points for three locations: Shishper (black circles), Khurdopin (blue circles), and Kyager (red circles). A light blue dotted line represents a 1:1 ratio. A red dotted trend line is fitted to the data with the equation V\_DEM = 1.3525V\_Tet and an R-squared value of R^2 = 0.88.
+
+**Fig. 02** (a) Definition diagram for calculating the volume of the lakes assuming (left) an irregular tetrahedral shape and (right) an irregular pentahedral shape. The blue shading represents the horizontal surface of the lake, and the white area represents the vertical ice wall. Panel (b) Relationship between the volumes of irregular tetrahedrons/10, derived from Eq. 1, and the volumes of the lakes determined using DEMs.
+
+In passing, it can be noted that the volumes of the irregular tetrahedrons in most cases were not dissimilar to the volumes of regular tetrahedrons (*i.e.*, triangular-based pyramids), the equation for which is simple, in contrast to Equation 1. Nevertheless, using equation 1, the volume of the 'tetrahedron' lakes was around 10 times greater than the volume of the lakes determined using the DEMs (Fig. 2b). This result indicates that the actual depth of the lake (*h*) must be much less than that value associated with an equilateral triangle of side length C (Fig. 2a left-hand side).
+
+## 3. Result:
+
+The relationship between the glacier surge velocity of (Khurdopin, Kyager, and Shishper) and 23 GLOFs is presented in Figure 3a-d and its triggering time or month in the year. In panels a to c, the GLOF occurred after the peak of the glacier surge, and the resealed lake formed while the surge velocity declined. These responses to slowing of the glacier velocity lasted 2-4 years after the surge peak. The relationship between the timing of glacier surges and the timing of GLOFs is shown in Fig. 3d. The three Karakorum glaciers can be used as regional exemplars of surge behavior controlling GLOF occurrence as there is a temporal relationship between the occurrence of periods of glacier surging and the occurrence of GLOFs (Fig. 3a-d). GLOFs occur towards the end of a surge period or immediately afterward; the detail is presented in Table 1. This pattern of behaviour prompted the hypothesis that glacier thickening and thinning during surging might control the development of ice-dammed
+
+Image /page/9/Picture/1 description: A Creative Commons Attribution (CC BY) license icon. The icon is a rectangle with a grey top section and a black bottom section. On the left side of the grey section is the Creative Commons logo, which consists of the letters 'CC' inside a white circle. To the right of that is the attribution symbol, a stick figure inside a white circle. In the black section at the bottom, the letters 'BY' are written in white.
+
+Image /page/9/Picture/2 description: The logo for EGUsphere, a preprint repository. The logo features the text "EGUsphere" in blue. The letters "EGU" are bold and enclosed in a blue circular gear-like shape. The word "sphere" follows, with a blue arch and a gray arch curving over it. Below the main text, the words "Preprint repository" are written in a smaller, gray font. The background is white.
+
+lakes (Bazai et al., 2022). Lake volumes would increase when the speed of the ice was low such that ice mass would be conserved or increased, and fracturing of the ice would reduce. The corollary pertains to when the ice speed increases, the glacier thins, and the fracturing of the ice mass increases, providing hydraulic drainage conduits.
+
+Image /page/9/Figure/4 description: The image contains four charts, labeled (a), (b), (c), and (d), which analyze glacier velocity and related events over time.
+
+Charts (a), (b), and (c) are line graphs plotting 'Average Velocity m y⁻¹' against 'Years'. Each chart has a secondary y-axis on the right labeled 'GLOF Triggered' with months from March to November. These charts show periods of increased velocity called 'Surge periods' highlighted with a grey background.
+
+- Chart (a) displays data from 1986 to 2020. The average velocity is shown by a red line, with a dashed blue line at an average of 79 m y⁻¹. There are two surge periods: 1998-2002, with an average peak velocity of 343 m y⁻¹, and 2016-2020. GLOF events are marked with blue circles.
+
+- Chart (b) covers the years 1989 to 2020. The average velocity is 29.9 m/y. It shows two main surge periods: 1994-1998 (average peak velocity 289 m/y) and 2013-2016 (average peak velocity 423 m/y). A smaller 'Mini Surge' is noted around 2008. GLOF events are marked with orange circles.
+
+- Chart (c) spans from 1989 to 2022. The average velocity is 87 m/y. It highlights two surge periods: 1997-2002 (average peak velocity 310 m/y) and 2017-2020. GLOF events are marked with grey circles.
+
+- Chart (d) is a timeline starting from 01/01/1990, with a numerical x-axis from 2000 to 12000. It plots events for different locations: Khurdopin (blue dots), Kayager (orange dots), and Shishper (grey dots). Horizontal lines indicate the season: blue for Winter and orange for Summer.
+
+Fig. 03 Relationship between glacier surge and GLOF, with annual glacier velocity analysis during the surge and quiescent phases for three glaciers: (a) Khurdopin, (b) Kyager, and (c) Shishper glacier. GLOFs for these glaciers occurred between the months of March to November. The combined analysis is presented in (d), illustrating the occurrences of GLOFs (dots) in relation to periods of glacier surging (bars). Some points are plotted below the timeline to avoid coincident positions. The blue and red lines show the winter (from October to April) and summer (May to September) seasons in which the GLOF occurred dominantly in the summer months.
+
+As a first attempt to relate glacier behaviour in a predictive sense to lake formation, we sought to determine the resultant lake volume from the prior surge speed. In Fig. 03, the relationship between surge and GLOF was developed using annual velocity data for ease of trend comprehension. However, daily measurements of glacier velocity were conducted to assess its impact on lake volume, given the velocity sensitivity to the triggering time of GLOFs. Consequently, the glacier's daily velocity was specifically recorded on the day when the GLOF was initiated, as detailed in Table 01. Despite a broad data spread, a least-squares regression analysis, with four outliers excluded from the analysis, defines a statistically significant negative power (-1) relationship between surge speed and lake volume (Fig. 4). Raising the glacier surge speed to the power of -2.0 and multiplying by 10 reproduced well the data trend for low glacier surge speeds. Similarly, -0.8 and a -3.0 power functions define the lower and upper limits of the data spreads. These results, although clearly not definitive, were a spur to investigate if there
+
+Image /page/10/Picture/1 description: The image shows the Creative Commons Attribution (CC BY) license icon. It is a rectangular icon with a grey upper section and a black lower section. On the grey section, there are two white circles. The left circle contains the letters "CC" in black, and the right circle contains a black icon of a person. The black section at the bottom has the letters "BY" in white.
+
+Image /page/10/Picture/2 description: The image shows the logo for EGUsphere, a preprint repository. The logo consists of the text "EGUsphere" in blue. The letters "EGU" are enclosed within a blue, gear-like circle. Above the word "sphere," there are two swooshes, one blue and one gray. Below the main text, the words "Preprint repository" are written in a smaller, gray font. The background is white.
+
+might be a theoretical geometric relationship between the volume of the lake and the control on lake water level exerted by the surge speed. The physical basis for such a relationship is that an increase in surge speed should cause the glacier to thin and fracture after a period of ice thickening and lake volume increase. During surging, the height of the ice barrier impounding the lake would reduce, and fracturing in the ice would increase, both lowering the lake level and reducing the lake surface area. Therefore, surge speed should control lake depth, volume, and potential GLOF volumes.
+
+Image /page/10/Figure/4 description: A scatter plot showing the relationship between Glacier surge speed (U\_s) on the x-axis and Lake Volume (V\_DEM) on the y-axis. The x-axis, labeled "Glacier surge speed, U\_s (m d⁻¹)", ranges from 0 to 1.6. The y-axis, labeled "Lake Volume, V\_DEM (M m³)", ranges from 0.0 to 200.0.
+
+The plot displays data points for different glaciers, indicated by a legend:
+- Shishper: black circles
+- Khurdopin: blue circles
+- Kyager: red circles
+- Khurdopin drained: blue triangles
+- Kyager drained: red triangles
+
+Most of the circular data points fall within a light blue shaded area, suggesting an inverse relationship where lake volume decreases as glacier surge speed increases. Several dotted curves representing different power-law models are overlaid on the data:
+- V\_DEM = 30U\_s⁻³ (brown dotted line)
+- V\_DEM = 25U\_s⁻¹ R² = 0.33 (black dotted line)
+- V\_DEM = 10U\_s⁻²·⁰ (green dotted line)
+- V\_DEM = 10U\_s⁻⁰·⁸ (light blue dotted line)
+
+The text above the plot reads, "therefore, surge speed should control lake depth, volume, and potential GLOF volumes."
+
+**Fig. 04** Variation in glacial lake volume as a function of the glacier surge speed. Data from three glaciers. Most lakes drained completely, but three had residual volumes (triangles). Outliers are shown as black-ringed symbols. A -0.8-power function defines the lower limit to the data spread, whilst a -2.0-power function defines the central tendency of the data trend, and a -3-power function defines the upper limit to the data spread (see text for explanation).
+
+Most of the lakes investigated occur in steep-walled side valleys, the ice dam forming one side of a roughly triangular water surface (Fig. 5a-b). Projecting the valley walls downwards, the geometry of the lake basin can be considered to be either: 1) an irregular tetrahedron, with the lake surface defined as an isosceles triangle (Fig. 5a-b); and the vertical cross-section at the ice dam defined by an equilateral three-sided triangle; or, 2) an irregular pentahedron whereby the vertical cross-section at the ice dam is defined by a quadrilateral (Fig. 2a). In the former case, only the length of the lake (X) and the breadth of the lake
+
+Image /page/11/Picture/1 description: A Creative Commons Attribution (CC BY) license icon. The icon is a rectangle with a grey top section and a black bottom section. The grey section contains two white circles. The left circle has the letters 'CC' in black, and the right circle has a black stick figure icon, representing attribution. The black bottom section has the letters 'BY' in white.
+
+280
+
+285
+
+Image /page/11/Picture/2 description: The logo for EGUsphere, a preprint repository. The logo features the text "EGUsphere" in blue. The letters "EGU" are inside a blue circular gear-like symbol. Above the word "sphere", there are two swooshes, one blue and one gray. Below the main text, the words "Preprint repository" are written in a smaller, gray font. The background is white.
+
+270 (C) at the ice dam are needed to estimate the lake volume, whereas for the second case, the depth (h) of the lake at the ice dam is required as well. The values for X and C are obtained from remote sensing imagery of the lake surface. Before a GLOF, the depth h can be obtained by deploying a plumb-line or echo-sounder in the lake or estimated through a topographic survey of the bounding slopes. Subsequent to GLOF drainage, h can be measured from a DEM or field survey. Assuming an irregular tetrahedron (Equation 1) implies the valley walls form a V-shaped cross-section to the lake at the ice dam face. In contrast, 275 assuming an irregular pentahedron implies the valley walls form a square-shaped section, which is more akin to typical Ushaped glacial valleys.
+
+Given the idealized geometric properties of the water mass, and assuming C and X vary little in contrast to h, the lake's volume increases as a function of  $h^3$  (tetrahedron) or  $h^2$  (pentahedron). Given the assumption that depth reduces as surge speed increases provides an explanation for the potential -2 to -3 power function relationships between surge speeds and lake volumes. For example, concerning a tetrahedron, assuming the equilateral geometry is maintained as water depth (h) in the lake increases, then, considering Pythagoras' theorem, the volume of the lake increases as the cube of the depth and by corollary as the -3 power of the surge speed. This later function is shown in Fig. 4, and the trend closely defines the upper limit of the empirical data, whilst the -2 relationship, which should pertain to pentahedral volumes, follows the approximate trend of the central tendency of the data spread. Thus, the geometrical properties assumed in the analysis delimit both an upper limit to the relationship between surge velocity and lake volume as well as the main trend.
+
+The appropriateness of assuming geometric shapes for the lake volume was checked by comparing the calculated geometric volumes with the measured DEM lake volumes. Utilizing Eq. 1 and only the values X and C, the volume of the 'tetrahedron' lakes was around 10 times greater than the volume of the lakes determined using the DEMs (Fig. 2a). This result indicates that the actual depth of the lake (h) must be much less than that value associated with an equilateral triangle of side length C.
+
+290 Nevertheless, this procedure provides a means to estimate lake volume from plan-view data alone.
+
+Image /page/12/Picture/1 description: A rectangular Creative Commons license icon, divided horizontally. The top section is gray and contains two white circles. The left circle has the letters 'CC' in black, and the right circle has a black icon of a person. The bottom section is black and has the letters 'BY' in white.
+
+Image /page/12/Figure/3 description: The image contains three panels labeled (a), (b), and (c). Panels (a) and (b) are photographs of glacial dammed lakes. Panel (a) is an aerial view of a large, light-blue glacial lake surrounded by dark, rocky terrain. Panel (b) is a closer view of a similar lake, showing chunks of ice or rock floating in the water within a steep valley. Panel (c) is a scatter plot comparing two volume calculation methods. The y-axis is labeled 'Lake Volume from DEM, V\_DEM (M m³)' and the x-axis is labeled 'Volume of Pentrahedron, V\_Pen (M m³)'. Both axes range from 0.0 to 250.0. The plot shows about 20 grey circular data points that form a strong positive linear trend. A grey dotted line represents the linear regression, with the equation 'V\_DEM = 0.8828 V\_Pen' and 'R² = 0.93' written next to it. A blue dotted line labeled '1:1' is also shown for reference. The caption below the figure reads: 'Figure 5: Examples of Shishper glacial dammed lakes exhibiting roughly triangular surface 2D shapes and overall pyra...'
+
+**Fig. 05:** Examples of Shishper glacial-dammed lakes exhibiting roughly triangular surface 2D shapes and overall pyramidal-shaped volumes (see Fig. 2a); (a) oblique aerial view towards the ice dam; (b) oblique aerial view from the ice dam axially up the impounded valley (image captured in March 2019 during lake monitoring (Image by (Image by Gilgit-Baltistan Disaster Management Authority (GBDMA)). Panel (c) Relationship between the volumes of irregular tetrahedrons/10, derived from Eq. 1, and the volumes of the lakes determined using DEMs.
+
+In contrast to the assumption of a tetrahedron, improving lake volume estimates were obtained considering the measured DEM-derived values of *h* along with the values of *X* and *C*. Once again, assuming an irregular tetrahedron as in Fig. 2a, the analysis demonstrated that the tetrahedral lake volume was roughly half that of the DEM volume (not illustrated). Rather, considering
+
+Image /page/13/Picture/1 description: A Creative Commons Attribution (CC BY) license icon. The icon is a rectangle with a light gray top half and a black bottom half. On the gray half, there are two white circles with black symbols inside. The left circle contains the letters 'CC', and the right circle contains a stick figure person icon. On the black half, below the person icon, the letters 'BY' are written in white.
+
+Image /page/13/Picture/2 description: The logo for EGUsphere, a preprint repository. The logo features the text "EGUsphere" in a blue sans-serif font. The letters "EG" are inside a blue circular gear-like shape. Above the "sphere" part of the text, there are two swooshes, one blue and one gray. Below the main text, the words "Preprint repository" are written in a smaller, gray font.
+
+the valley-cross section to be U-shaped (roughly quadrilateral) rather than V-shaped means that doubling the area of the triangular dam face section to form a quadrilateral gives a lake volume defined as an irregular pentahedron (square-based pyramid) (Fig. 2a right-hand side) in contrast to an irregular tetrahedron (Fig. 2a left-hand side). The volume of an irregular pentahedron is:
+
+$$V = \frac{1}{3} (C h) X \tag{2}$$
+
+which is more amenable to solving than Equation 1.
+
+The relationship shown in Fig. 5c is preferable to that shown in Fig. 2a as it lies close to the 1:1 relationship between  $V_{Pen}$  and  $V_{DEM}$  but requires knowledge of the parameters as height or depth: h, length; X, and width; C. In contrast, the less exact relationship in Fig. 2a requires only X and C. If assuming a rectangular base to the pentahedron exactly matched the DEM volume, then the coefficient value would be unity. Thus, the coefficient of 0.8986 reflects the deviation of the cross-sectional shape of the lake at the dam face from a rectangle. Note that the relationships between both determinations of lake volume progressively deviate from a 1:1 relationship as lake volume increases. This trend might indicate that larger lakes are less well-defined as tetrahedrons or pentahedrons as the volumes increase.
+
+### 3.1. Predicting the Timing of GLOF Events
+
+The timing of a GLOF remains difficult to determine, but the main driver is the critical depth. The critical depth is that depth that exerts sufficient pressure at the ice dam wall to induce completed connectivity within the sub-glacial GLOF drainage conduit. For the cases of the glacial lakes (Shishper, Khurdopin, and Kyager), the glacial lake depths (h) have been normalized by dividing by the minimum value of the ice dam height to give values (n') that range between 0 for a fully-drained lake to a hypothetic value of 1 if the lake level reached the height of the ice barrier. At the approximate time of GLOF occurrence, the resultant values of n' range between 0.32 and 0.95 (Fig. 6a-c). Most GLOFs occur for a range of n' values between 0.61 and 0.95 (Fig. 6d). Thus, n' = 0.60 can be regarded as a warning level value with the risk of a GLOF occurring imminently increasing as n' approaches unity.
+
+As the water pressure (P) at the dam face increases linearly with water depth in each lake, any variation in the pressure with n' that deviates from the linear trend reflects changes in the height of the ice dam (Fig. 6d). Thus, for example, the values of pressure for the Shishper lakes around an n' value of 0.8 reflect greater overall ice dam heights in contrast to the Kyager lakes for similar values of n'. The two values of low pressure for n' < 0.6 reflect low ice dam heights and presumably low structural integrity within the ice mass, allowing ready conduit development. Low values of n' (< 0.6) likely are associated with shallow lakes of low hazard potential. Overall, the Kyager data (Fig. 6d) indicate that a minimum water pressure of around 500 kPa should be regarded as a threshold for general concern for GLOF occurrence in the region.
+
+330
+
+325
+
+Image /page/14/Picture/1 description: A Creative Commons Attribution (CC BY) license icon. The icon is a rectangle with a grey top section and a black bottom section. On the left side of the grey section is the Creative Commons logo, which consists of the letters 'CC' inside a white circle. To the right of that is the attribution symbol, a stick figure inside a white circle. In the black section at the bottom, the letters 'BY' are written in white.
+
+Image /page/14/Figure/3 description: A four-panel figure, labeled (a) through (d), analyzing glacier lake characteristics. Panels (a), (b), and (c) show cross-sectional profiles of lakebeds, while panel (d) shows a scatter plot of water pressure versus lake depth.
+
+Panel (a) is a cross-section of the Shishper Glacier lakebed. The x-axis represents the cross-section in meters (0-2100), and the y-axis shows elevation in meters (2520-2660). A secondary top x-axis shows volume in 10^6 m^3 (0-32). The plot displays the lakebed profile, the glacier, and lake levels for years 2019 (E=2650, V=24.10), 2022 (E=2641, V=27.66), 2021 (E=2638, V=25.77), and 2020 (E=2636, V=24.90). An inset map shows the glacier and lake outlines.
+
+Panel (b) is a cross-section of the Khurdopin Glacier lakebed. The x-axis is cross-section in meters (0-3400), and the y-axis is elevation in meters (3360-3440). The top x-axis shows volume in 10^6 m^3 (0-180). It shows lake levels for years 2000 (E=3440, V=186), 2002 (E=3420, V=52.1), 2018 (E=3418, V=19.80), 2001 (E=3416, V=19.5), and 2017 (E=3415, V=16.2). An inset map is included.
+
+Panel (c) is a cross-section of the Kyager Glacier lakebed. The x-axis is cross-section in meters (0-6500), and the y-axis is elevation in meters (4750-4830). The top x-axis shows volume in 10^6 m^3 (0-130). It displays numerous lake levels from 1977 to 2018, with a detailed legend providing the year, elevation (E), and volume (V) for each.
+
+Panel (d) is a scatter plot of Water Pressure (P) in kPa versus Normalized glacier lake depth (n'). The x-axis ranges from 0.0 to 1.0, and the y-axis from 0 to 1600. Data points are shown for Shishper (black circles), Khurdopin (blue circles), and Kyager (red circles). Linear trendlines are fitted to each dataset with the equations: P = 1514.9 n' for Shishper, P = 977.1 n' for Khurdopin, and P = 835.5 n' for Kyager. A light blue shaded area from n' ≈ 0.6 to 1.0 is labeled 'Critical range'.
+
+**Fig. 06:** The relationship between lake volume and elevation and the critical depth for GLOFs. a) Shishper; b) Khurdopin; c) Kyager; d) Water pressure at dam face as a function of n'. The cross sections are denoted by (A' and A). The date of each lake flood event is shown as L, E is the lake's elevation, and V is the lake volume. The straight, solid red lines relate specific lake elevations to volumes. The UAV photographs captured in the field were used for panel a, and the images captured by the GF-2 were used for panel b.
+
+## 4. Discussion:
+
+335
+
+Although predictive models related to ice-dam lake development and subsequent GLOF risk would best be based on modelling the physics of the systems, the controlling parameters are numerous and complex. For example, the mechanisms of glacial sliding, over-burden pressure, tensile and driving stresses require consideration, as do flexure and ice fracture mechanics, thermal erosion, and water pressure, amongst other controls (Carrivick et al., 2020). Few of these controls are well-understood, and, importantly, even where there is an adequate theory, the field data required to inform modelling are absent for specific potential GLOF locations. Consequently, there is an urgent need for simpler methods to predict the likely volume of ice-dammed lakes and the probable triggering water levels that lead to GLOFs. Given that requirement, it is acknowledged that
+
+Image /page/15/Picture/1 description: A Creative Commons Attribution (CC BY) license icon. The icon is a rectangle with a grey top half and a black bottom half. The grey section contains two white circles with black outlines. The left circle has the letters 'CC' in black, and the right circle has a black icon of a person. The black section has the letters 'BY' in white.
+
+355
+
+Image /page/15/Picture/2 description: The logo for EGUsphere, a preprint repository. The logo features the text "EGUsphere" in blue. The letters "EGU" are bold and enclosed in a blue circular gear-like shape. The word "sphere" follows, with a blue arch and a gray arch curving over it. Below the main text, the words "Preprint repository" are written in a smaller, gray font. The background is white.
+
+the relationships proposed herein are empirical and apply specifically to glaciers within the Karakorum region. However, there is no reason to suppose that similar functions based on geometric considerations and critical depth (Zhao et al., 2017) might not be developed elsewhere, including for moraine-dammed lakes (Yao et al., 2010). Below, the approach is explored for glacially dammed lakes worldwide.
+
+Image /page/15/Figure/4 description: The image contains three plots, labeled (a), (b), and (c), which analyze data related to ice-dammed glacier lakes and Glacial Lake Outburst Flood (GLOF) events.
+
+Plot (a) displays the maximum volume level of several glacier lakes against their elevation. The y-axis represents 'Elevation (m.a.s.l)' and the x-axis represents 'Volume M m³'. Each lake (Kyager, Khurdopin, Merzbacher, Chilinji, Medvezhiy, Shishper) is shown as a horizontal bar divided into a 'Safe lake level' (light blue) and a 'Triggering lake level' (red). Black dots indicate the volume and elevation of past 'GLOF Event's. An inset shows similar data for the Russel and Rio Colonia lakes. Each lake has an associated n' value: Kyager n'=0.67, Khurdopin n'=0.76, Merzbacher n'=0.83, Chilinji n'=0.60, Medvezhiy n'=0.71, Shishper n'=0.67, Russel n'=0.80, and Rio Colonia n'=0.79.
+
+Plot (b) is a vertical bar representing the 'Normalized Glacier Lake depth (n')'. The bar is divided into a 'Safe Zone' (light blue, for n' values below 0.6) and a 'High Risk and Critical Zone' (red, for n' values above 0.6). Dotted lines connect the n' values of various lakes to their corresponding position on this risk scale. Merzbacher, Russell, Rio Colonia, Khurdopin, and Medvezhiy fall into the high-risk zone, while Kyager, Shishper, and Chilinji are in the safe zone.
+
+Plot (c) is a scatter plot comparing 'Reference Volume (M m³)' on the y-axis to the 'Volume of Irregular Tetrahedron (M m³)' on the x-axis for various glaciers around the world. The data points, color-coded by location (e.g., Medvezhiy-Pamir, Grænalón-Iceland, Merzbacher-TianShan), show a strong positive linear correlation. A dashed line of best fit is shown with the equation y = 1.0689x and an R² value of 0.9865.
+
+**Fig. 07:** (a) Represents the maximum volume level of ice-dammed glacier lakes that experience outbursts based on specific volume and elevation criteria. In this representation, the red signifies each lake's triggering level, depth, or zone, where all GLOF events were initiated. Among these lakes, n' values represent the non-dimensional relationship between lake depth and glacier height; panels (b) illustrate each lake's normalized critical lake depth and the high-risk and critical zones for Glacial Lake Outburst Floods (GLOFs). It also highlights the high-risk and critical zones for each lake, where values exceeding n' 0.6
+
+Image /page/16/Picture/1 description: A rectangular icon representing the Creative Commons Attribution license, also known as CC BY. The icon is divided horizontally into two sections. The top, larger section is grey and contains two white circles with black outlines. The circle on the left has the letters 'CC' in black, and the circle on the right has a black pictogram of a person. The bottom, smaller section is black and has the letters 'BY' in white.
+
+360
+
+365
+
+370
+
+375
+
+380
+
+385
+
+390
+
+Image /page/16/Picture/2 description: The image shows the logo for EGUsphere, a preprint repository. The logo consists of the text "EGUsphere" in blue. The letters "EGU" are inside a blue, gear-like circle. Above the word "sphere," there are two swooshes, one blue and one gray. Below the main text, the words "Preprint repository" are written in a smaller, gray font. The background is white.
+
+indicate a critical threshold for the potential occurrence of an ice-dammed lake outburst flood. Panel (c) shows the relationship between the referenced volume and measured volume through a geometric approach to the location of glaciers presented in Fig. S1.
+
+Glacially dammed lakes worldwide exhibit consistent behavior in terms of lake formation, filling, and volume gain in response to low glacier velocity (Bazai et al., 2021). Additionally, specific pressure, volume, and critical lake depth values for initiating outburst floods are evident (see Figure 6). Building upon the information presented in Figure 6 for the Karakorum, Figure 7, based on 50 GLOF events from glacial dammed lakes, offers a depiction of the conditions under which glacier lake volume measurements are successfully estimated with high accuracy. Among the 50 GLOF events, 27 examples are given from Pamir, Tianshan, Greenland, and northern Patagonia (see Figure 7a-c). These relevant data are limited, often with only estimates or measurements of lake volume available. Figures 7a and b serve as key components and summaries for this discussion. identifying the lake volumes, elevations, and the critical depth values for GLOF outbursts. The measured lake volumes and the related water surface elevations are shown in Fig. 7a, together with critical depth values (n'). Critical depth values (n')exceed 0.60 in all cases of GLOFs (Fig. 7b), from which we infer that a safe lake level can be defined as  $\leq$  0.60, and the trigger level is  $\geq 0.60$ . Values of values n' < were associated only with slow, non-catastrophic, lake drainage. Therefore, in the case of future ice-dammed lakes, values of critical depth (n') exceeding 0.60 should be a cause for concern and would serve as a warning level. The geometric procedure outlined herein to estimate lake volume produced an excellent correlation with measured lake volumes (R<sup>2</sup> value of 0.986; Fig. 7c). These findings suggest that future exploration should concentrate on specific volume and depth parameters to determine critical thresholds associated with depth, elevation, or volume for future predictive purposes.
+
+Estimating the volume of an undrained ice-dammed lake from a field survey is dangerous due to floating ice (Fig. 05b), rugged terrain, and sudden drawdowns. The utilization of DEM measurements for lake volume estimation may also introduce high uncertainties or errors due to the difficulty in defining the lake depths (Carrivick et al., 2020; Emmer, 2018). However, for rapid response or mitigation policy purposes, the empirical model (Eq. 1) used in the current study proves to be quite efficient when applied to measure the lake volume before a GLOF.
+
+Understanding glacier surges, lake formation, and the interactions between lakes and glaciers is crucial for advancing knowledge and developing empirical or numerical GLOF models in mountainous regions (Carrivick et al., 2020; Quincey and Luckman, 2014). In this study, the glacier and lake interactions and their relationship have been explored, and their effect on lake volume and draining processes has been examined. Glacier surge speed is routinely determined using remote-sensing imagery (Paul, 2015), as is lake surface area (Quincey and Luckman, 2014). Thus, remote sensing provides a means to develop images similar to Fig. 4 for specific locations around the globe where ice-dammed lakes form due to glacier surging. Although the data within Fig. 4 are scattered, a negative relationship between surge velocity and lake volume is strongly implied. Specifically, data points scatter around a median trend according to a theoretical -2 power function. Most values fall below a theoretical -3 power function (Fig. 4). Clearly, more data points within Fig. 4 would be desirable. Although GLOFs cannot be predicted from this approach, the likely volume of water that might be released catastrophically can be determined. For sites
+
+Image /page/17/Picture/1 description: A Creative Commons Attribution (CC BY) license icon. The icon is a grey rectangle with a black bar at the bottom. On the grey section, there are two white circles with black symbols inside. The left circle contains the letters 'CC', and the right circle contains the attribution symbol, which is a stylized person. On the black bar at the bottom, the letters 'BY' are written in white.
+
+395
+
+400
+
+405
+
+415
+
+420
+
+Image /page/17/Picture/2 description: The image shows the logo for EGUsphere, a preprint repository. The logo consists of the word "EGUsphere" in blue. The letters "EGU" are enclosed within a blue, gear-like circle. Above the "sphere" part of the word, there are two swooshes, a blue one on top and a gray one below it. Underneath the main logo, the text "Preprint repository" is written in a smaller, gray font. The background is white.
+
+that are deemed to pose a threat to human life and infrastructure, once the lake volume is better constrained, either through DEM analysis or geometric considerations, the value of n for any specific lake provides a ready indicator of the probability of an imminent GLOF. In contrast to the lower trend for water pressures associated with the Kyager lakes, the higher water pressures required to cause the Khurdopin and Shishper lakes to empty may reflect greater structural integrity, possibly related to a greater downstream extent of the glacier dams. These structural issues can be examined in the future. Still, at this stage, if n' exceeds 0.6, an initial general warning could be issued to communities downstream of the ice dam. In principle, the estimated volume of a potential GLOF can then be routed downstream using standard hydrodynamic flood routing procedures to determine the timing, depth, and extent of flooding at locations where inundation is forecast. Thus, the severity of the likely impact on humankind can be determined, and specific warning times can be derived from the rate of travel of the GLOFs. These results represent a step forward from the observations made by Carrivick et al. (2020), who proposed the exploration of the interaction between lake water and glaciers to understand the lake formation process and identify lake depth, level, and volume. Based on this understanding, empirical models can be generated to predict GLOFs in a timely manner. The current study offers a comprehensive explanation and advanced knowledge building upon the findings of Carrivick et al. (2020). The study revealed that glacier surge and GLOF development relationship indicates that glacier dynamics control lake volume. Furthermore, the study explores the lake-draining process and identifies the relative depth above which an ice-dammed lake's sudden draining may occur. Future research should further investigate the range of n' values and the nature of ice-failure mechanisms connecting lake drawdown and GLOF progression.
+
+## **410 5. Conclusion:**
+
+The hazards associated with glacier lake outburst floods (GLOFs) in cryospheric regions are increasing worldwide in response to climate change, posing a significant threat to inhabitants in densely populated areas, particularly in the Himalayan Karakoram and Hindu Kush regions. Year on year there is a rise in human casualties and losses to residences, infrastructure, the energy sector, and local and international trade. Despite the escalating risk, the understanding of these hazards remains limited. It is imperative to determine the causes of these hazards, make timely predictions, and formulate new mitigation policies to minimize losses. This research identified the critical depth, lake volume, pressure, and elevation of ice-dammed lakes worldwide associated with GLOFs. An inverse relationship between lake volume and glacier surge speed was observed, indicating that surge speed controls lake depth, volume, and potential GLOF volumes. Comparing surveyed lake volumes with geometric estimates for 23 GLOF events from the Karakoram and 27 events from around the world, a linear regression ( $R^2 = 0.95$ ) demonstrated that geometric estimates can be robust in the absence of detailed field or remote-sensing surveys. A GLOF will likely occur when the lake's non-dimensional depth (n') exceeds 0.60, corresponding to a typical water pressure on the dam face of 510 kPa.
+
+## **Competing interests**
+
+The authors declare no conflict of interest.
+
+Image /page/18/Picture/1 description: A Creative Commons Attribution (CC BY) license icon. The icon is a rectangle with a grey top half and a black bottom half. The grey half contains two white circles. The left circle has the letters 'CC' in black, and the right circle has a black icon of a person. The black bottom half has the letters 'BY' in white.
+
+Image /page/18/Picture/2 description: The image displays the logo for EGUsphere. The logo consists of blue and gray elements on a white background. On the left, there is a blue circular gear-like icon containing the letters 'EG'. This icon flows into the main text, 'EGUsphere', which is written in a bold blue font. Above the 'sphere' part of the word, there are two swooshes, a blue one on top and a gray one underneath. Below the main text, the words 'Preprint repository' are written in a smaller, gray font.
+
+## 425 Acknowledgements
+
+430
+
+435
+
+440
+
+445
+
+450
+
+This study was supported by the Second Tibet Plateau Scientific Expedition and Research Program (STEP) (Grant No. 2019QZKK0906) and the National Natural Science Foundation of China (Grant no. 42350410445). Special thanks to the monitoring team of Gilgit-Baltistan Disaster Management Authority (GBDMA), Quaid-i-Azam University, and Karakoram International University for their support and data sharing, and thanks to the Special Research Assistant program of the Chinese Academy of Sciences.