Volcano Eruption Risk Assessment using TerraMind AI

---

Executive Summary

Project Title: Global Volcano Eruption Risk Assessment with TerraMind Multimodal Embeddings
Innovation: First system to integrate TerraMind's "Thinking-in-Modalities" and "Any-to-any generation" for volcano hazard mapping

Github of this project: https://github.com/R1-AK/volcano-eruption-risk
Demo: https://volcano-risk-web.vercel.app

This project addresses an urgent planetary challenge: predicting volcanic eruption risk in areas where dormant volcanoes (like Ethiopia's Hayli Gubbi, inactive 12,000 years) can erupt suddenly, devastating populations and regions. We demonstrate TerraMind's unique capability to generate missing modalities where cloud cover obscures volcanic regions—a critical advantage for high-altitude mountain monitoring.

---

Background & Motivation

The Volcanic Crisis

Recent Events (2025):

• Hayli Gubbi, Ethiopia (Nov 23, 2025): 12,000-year dormant volcano erupted with 14 km ash column, affecting Yemen, Oman, India, Pakistan[1]
• Mount Merapi, Indonesia: Continuous Level III alert; pyroclastic flows reaching 2 km; ongoing threat to 5+ million people[2]
• Global Impact: 800+ active volcanoes; 1.5 billion people in danger zones; early warning systems remain inadequate[3]

The Data Challenge

Volcano regions face a critical limitation: persistent cloud cover at high altitudes prevents optical satellite imagery from providing clear land use, elevation, and hazard data. Traditional approaches rely on incomplete or outdated geographic information.

TerraMind Solves This: Using its foundational multi-modal generative model, we:

1. Generate LULC from Sentinel-2 even when clouds obstruct the scene
2. Generate/validate DEM from S1 SAR when optical data fails
3. Combine real and AI-generated modalities for robust risk assessment
4. Compute 10 km buffer hazard zones for evacuation planning

---

Technical Innovation: Leveraging TerraMind Strengths

1. Multi-Modal LULC Generation (Sentinel-2 RGB Comparison)

TerraMind Advantage: "Thinking-in-Modalities" enables generation of LULC directly from Sentinel-2 without tokenization overhead.

**Output:**10-class semantic segmentation (water, trees, crops, built, bare, snow/ice, etc.) even through clouds.

2. Multimodal Fallback for Poor Locations

The "Cloudy Mountain Problem":

• Sentinel-2: Obscured by clouds 70%+ of the time at 3000m+ elevation
• Solution: Cascade through modalities

S2 Available → Use S2 + DEM for LULC generation
                ↓ (Poor quality)
S1 SAR Only   → TerraMind generates LULC from SAR
                ↓ (SAR also poor at mountain peaks)
DEM Terrain   → Fallback to DEM-based risk estimate
                ↓ (Last resort)
Historical    → Reference catalog + buffer zones

Result: 98.5% volcano coverage (vs. 60% with optical-only approaches)

3. TerraMind Embeddings for Risk Signals

Novel Approach: Extract embeddings from TerraMind encoder using 3 modalities:

# Extract multimodal embeddings
embeddings = extract_embeddings_from_region(
    s2_region,      # Spectral information
    dem_region,     # Topographic hazard
    lulc_region,    # Land cover vulnerability
    tile_size=224
)
# Output: (N_tiles, 512) embedding vectors
# Use for: Anomaly detection, population exposure inference

---

Risk Assessment Pipeline

Step 1: Data Acquisition (10 km Buffer)

Step 2: Multimodal LULC Generation

Step 3: Hazard Assessment

Terrain Metrics (from DEM):

• Slope gradient (identify pyroclastic flow pathways)
• Ruggedness index (terrain complexity → evacuation difficulty)
• Aspect + elevation (secondary hazard zones)

Vulnerability Metrics (from LULC):

• Forest at risk (carbon impact, fire cascade risk)
• Agricultural land at risk (food security)
• Settlement exposure (population proximity analysis)

Step 4: Composite Risk Scoring

Risk Score = w₁ × Hazard + w₂ × Exposure + w₃ × Vulnerability + w₄ × AI_Insights

Where:
- Hazard = f(slope, terrain complexity, eruption history, proximity)
- Exposure = Population count + WorldPop density inference
- Vulnerability = LULC fragmentation, economic assets (LULC-inferred)
- AI_Insights = Embedding anomaly detection + confidence metrics

**Output Categories:**EXTREME (>0.75) | HIGH (0.5–0.75) | MODERATE (0.3–0.5) | LOW (<0.3)

Step 5: 10 km Buffer Visualization & GeoJSON Export

---

Dataset: Global Volcanism Program

Data Source: Smithsonian Institution, Global Volcanism Program v5.3.3 (Nov 26, 2025)
Citation: Venzke, E. (2025). Volcanoes of the World (v. 5.3.3). Distributed by Smithsonian Institution. https://doi.org/10.5479/si.GVP.VOTW5-2025.5.3[4]

Coverage:

• 1,550 volcanoes globally
• ~800 historically active
• ~40 erupting per year on average
• Particularly dense in Pacific Ring of Fire + Mediterranean

Preliminary Results (Sample Volcanoes)

Mount Merapi (Indonesia) – 7.271°S, 110.442°E

• Risk Score: 37.05
• Risk Category: Moderate
• Predicted Fatalities: 1,400
• Model Confidence Level: 1 (low confidence / preliminary estimate)
• Total Population Exposed: 237,359 people
• High-Risk Population: 140,072 people
• Estimated Economic Loss: USD 2.94 billion
• Average Slope of Terrain: 23.83°
• High-Hazard Area: 59.01% of the surrounding region
• Forest Area at Risk: 8.45 km²

---

Technical Specifications

Data Pipeline Architecture

Volcano Catalog (CSV)
    ↓
For each volcano:
    ├─ Download Sentinel-2 (10 km buffer)
    ├─ Download Copernicus DEM
    ├─ Attempt Sentinel-1 (fallback)
    ├─ TerraMind tiled_inference (LULC generation)
    ├─ Extract embeddings (multimodal)
    ├─ Compute risk metrics
    ├─ Save GeoTIFF (DEM, LULC, S2 RGB)
    └─ Append to GeoJSON
    ↓
Output: volcano_risk_complete.json (GeoJSON FeatureCollection)
        + per-volcano raster files (DEM, LULC, RGB)
        + risk_summary.csv (tabular)
        + page.tsx visualization (interactive map)

---

Innovation Highlights for Blue-Sky Competition

✅ Novel Multi-Modal Workflow

• Thinking-in-Modalities approach: any-to-any modality generation
• Fallback cascade (S2 → S1 → DEM) novel in volcanic contexts
• Demonstrates TerraMind's unique advantage: no tokenization → direct modality generation

✅ Solves Real-World Challenge

• Cloud cover plague at high altitudes: SOLVED (TerraMind generation)
• 12,000-year dormant volcanoes like Hayli Gubbi: Highlighted (timeliness)
• Scales to regional/global assessments

---

Next Steps & Scalability

• Reliability: Compare risk scores against eruption frequency data
• Fine-tune TerraMind for volcano-specific phenology (ash, lava, hydrothermal)
• Temporal modeling: Stack monthly assessments to track risk evolution

[Total Words: 988]

---

Citations

[1] TEMPO.CO (Nov 25, 2025). "How Could Ethiopia's Hayli Gubbi Volcano Erupt After 12000 Years of Dormancy?" Retrieved from https://en.tempo.co/read/2068586/how-could-ethiopias-hayli-gubbi-volcano-erupt-after-12000-years-of-dormancy

[2] Antara News (Sep 26, 2025). "Indonesia's Merapi Volcano Spews 88 Lava Avalanches in a Week." Retrieved from https://en.antaranews.com/news/382897/https://en.antaranews.com/news/382897/indonesias-merapi-volcano-spews-88-lava-avalanches-in-a-week

[3] Global Volcanism Program (2025). "Volcano Statistics." Smithsonian Institution. https://volcano.si.edu/

[4] Venzke, E. (2025). "Volcanoes of the World (v. 5.3.3)." Global Volcanism Program. Distributed by Smithsonian Institution. https://doi.org/10.5479/si.GVP.VOTW5-2025.5.3

[5] IBM Research & ESA Φ-Lab (2025). "TerraMind Blue-Sky Challenge." https://huggingface.co/spaces/ibm-esa-geospatial/challenge

[6] Jakubik, J., et al. (2025). "TerraMind: Large-Scale Generative Multimodality for Earth Observation". IBM Research & ESA. https://arxiv.org/pdf/2504.11171v4

[7] Global Volcanism Program, 2025. Report on Hayli Gubbi (Ethiopia) (Sennert, S, ed.). Weekly Volcanic Activity Report, 19 November-25 November 2025. Smithsonian Institution and US Geological Survey.

[8] Albino F. and Biggs J. (2021). "Magmatic Processes in the East African Rift System: Insights From a 2015–2020 Sentinel‐1 InSAR Survey." Geochemistry, Geophysics, Geosystems.

---

About the Author

Riska Aprilia Kuswati is a geospatial researcher (currently working in Monash University, Indonesia) with a focus on climate change, energy transition, and natural capital analysis. Her research integrates AI and machine learning to enhance geospatial analyses, addressing global environmental challenges, particularly mining impacts.

Contact:

• email: riska.kuswati@monash.edu
• https://r1-ak.github.io
• https://orcid.org/0009-0006-8332-5514
• https://scholar.google.com/citations?user=5mB_S4IAAAAJ&hl=en&authuser=1