teoaivalis/extreme-floods-kg

---

🌍 Flood Events Knowledge Graph Dataset

This dataset provides structured representations of flood disaster events, integrating event metadata, humanitarian reports, natural features from geospatial data, and knowledge graph triples extracted with Large Language Models (LLMs).
It supports research in Earth Observation (EO), disaster management, semantic reasoning, and interpretable AI.

---

📖 Dataset Summary

Floods are among the most frequent and devastating disasters worldwide, yet data describing them is often scattered across unstructured reports, geospatial sources, and satellite imagery.
This dataset unifies those heterogeneous data sources into a structured, ontology-aligned Knowledge Graph (KG) format.

Each event is represented with:

• Metadata: disaster type, location, date, country.
• Textual Descriptions: humanitarian situation reports from ReliefWeb.
• Geospatial Features: natural features extracted from OpenStreetMap (OSM).
• Knowledge Graph Triples: subject–predicate–object statements extracted by LLMs.

---

📂 Source Data

• ReliefWeb → primary source for event metadata and humanitarian impact reports.
• OpenStreetMap → environmental features around affected regions (e.g., rivers, forests, peaks).
• LLM-driven Triple Extraction → ontology-guided prompts applied to metadata and reports to generate KG triples.

Further details about data collection and preprocessing scripts are available in our GitHub repository:
👉 XtremeKG on GitHub

We also provide a demo where users can run queries over the triples.

---

📝 Example Record

Below is a full JSON representation of one event (Haiti Floods, May 2025):

{
    "event_id":"FL-2025-000045-COD",
    "name":"DR Congo: Floods - Apr 2025",
    "date":1743724800000,
    "country":"Democratic Republic of the Congo",
    "location":{
      "lat":-4.03833,
      "lon":21.7587
    },
    "disaster_type":"Flood",
    "description":"Heavy rainfall has been affecting Kinshasa, the capital of the Democratic Republic of the Congo, since 4 April, causing flooding and landslides that have resulted in casualties and damage. According to the International Federation of Red Cross and Red Crescent Societies (IFRC), as of 7 April, 33 people have died and 46 were injured due to the overflow of the Ndjili River, affecting 13 out of 24 municipalities in Kinshasa city area. Media reports further indicate that several roads and houses have been destroyed or damaged, power outages have been recorded, and the water supply has been cut off. ([ECHO, 8 Apr 2025](https:\/\/reliefweb.int\/node\/4144934))\n\nFollowing floods, the Ndjili river overflow, and landslides that occurred across 13 municipalities (out of 24) of the Kinshasa capital city (western Democratic Republic of the Congo) over the period from 4 to 11 April, the number of casualties has increased, and widespread damage was reported. The International Federation of Red Cross and Red Crescent Societies (IFRC) reports, as of 13 April, 165 fatalities, 28 injured persons, more than 7,000 displaced people (a number of these across four relief sites) and a total of over 60,000 affected people across the Kinshasa capital city area. In addition, the IFRC also reports several destroyed and damaged houses. The UNOSAT was activated on 11 April and the International Charter was activated on 10 April to provide the damage assessment. Over the next 96 hours, more rainfall is still forecast over the already affected Kinshasa area. ([ECHO, 14 Apr 2025](https:\/\/reliefweb.int\/node\/4146232))\n\nSevere flooding triggered by torrential rains in recent weeks has displaced almost 10,000 people in Tanganyika Province, the Democratic Republic of the Congo (DRC). This unfolding emergency reflects the double crisis facing the DRC, where extreme weather shocks such as flooding compound the suffering caused by ongoing conflict and mass displacement. The Rugumba River burst its banks, inundating large areas of Kalemie and Nyunzu territories. Homes, schools and agricultural land have been destroyed, leaving thousands without shelter or livelihoods. Stagnant and contaminated floodwaters are raising concerns over the risk of disease outbreaks, with reported cholera cases in the province already six times higher than during the same period last year. The flooding has hit a community already under severe strain. Since January, Tanganyika has received around 50,000 internally displaced people (IDPs) fleeing violence in South Kivu. Many had been sheltering in local homes, churches, and schools—now damaged or destroyed. The floods also wiped out key crops like cassava, maize, and peanuts, worsening an already serious food insecurity situation in the country. ([UNHCR, 15 Apr 2025](https:\/\/reliefweb.int\/node\/4146561))",
    "affected_regions":[
      {
        "region":"Kinshasa",
        "lat":-4.32758,
        "lon":15.31357
      },
      {
        "region":"Tanganyika",
        "lat":-5.93333,
        "lon":29.2
      }
    ],
    "osm_features":{
      "waters":1,
      "woods":5,
    },
    "triples":[
      ["FL-2025-000045-COD", "hasCountry", "Democratic Republic of the Congo"],
      ["FL-2025-000045-COD", "hasISO3", "cod"],
      ["FL-2025-000045-COD", "hasDisasterType", "Flood"],
      ["FL-2025-000045-COD", "hasStatus", "ongoing"],
      ["FL-2025-000045-COD", "hasDate", "2025-04-04T00:00:00+00:00"],
      ["FL-2025-000045-COD", "hasLatitude", "-4.03833"],
      ["FL-2025-000045-COD", "hasLongitude", "21.7587"],
      ["FL-2025-000045-COD", "hasGlide", "FL-2025-000045-COD"],
      ["FL-2025-000045-COD", "fatalitiesReported", "165"],
      ["FL-2025-000045-COD", "injuriesReported", "28"],
      ["FL-2025-000045-COD", "displacedPeople", "7000"],
      ["FL-2025-000045-COD", "peopleAffected", "60000"],
      ["FL-2025-000045-COD", "powerOutage", "true"],
      ["FL-2025-000045-COD", "waterSupplyCut", "true"],
      ["FL-2025-000045-COD", "hasWaters", "1"],
      ["FL-2025-000045-COD", "hasWoods", "5"],
      ["FL-2025-000045-COD", "causedBy", "Heavy rainfall"],
      ["FL-2025-000045-COD", "weatherCause", "Torrential rains"]
    ]
  }

---

⚡ How to Load

from datasets import load_dataset

# Login using e.g. `huggingface-cli login` to access this dataset
ds = load_dataset("teoaivalis/extreme-floods-kg")

# Print dataset info
print(ds)

# Inspect one record
print(ds["train"][0])

# Convert to pandas DataFrame
import pandas as pd
df = pd.DataFrame(ds["train"])
print(df.head())

---

📜 Citation

If you use this dataset, please cite:

@article{aivalis2025llm,
  title={LLM-Driven Knowledge Graph Construction from Earth Observation Data for Extreme Events},
  author={Aivalis, Theodoros and Klampanos, Iraklis A and Troumpoukis, Antonis},
  year={2025}
}

---