DisasterChain-MM: A Multimodal Dataset of Grounded Causal Impact Chains for Global Disasters

Dataset Summary

DisasterChain-MM is a high-fidelity, multimodal dataset designed to advance research in causal reasoning, environmental monitoring, and disaster impact assessment. Building upon the EM-DAT database, it provides a comprehensive reconstruction of over 1,400 global extreme environmental events from 2014 onwards.

Unlike traditional disaster catalogs that only provide static metadata, this dataset introduces Grounded Causal Impact Chains. Using advanced Large Language Models (Qwen 2.5 72B and Llama 3 70B), we extract step-by-step causal sequences of physical and socio-economic impacts (e.g., Extreme Precipitation -> River Overflow -> Flood -> Infrastructure Damage -> Displacement). Crucially, every step in the causal chain is strictly grounded to real-world news reports via verified supporting quotes.

Furthermore, each event is enriched with extensive multimodal data, including 21-day weather time series and multispectral satellite imagery availability, making it a unique resource for bridging Natural Language Processing (NLP) and Earth Observation (EO).

Supported Tasks and Applications

• Causal Reasoning & Event Extraction: Training models to identify cascading effects in unstructured text;
• Multimodal Disaster Assessment: Combining satellite imagery (Sentinel-1, Sentinel-2), weather data, and text to predict physical damage or human casualties;
• Climate Impact Analysis: Analyzing the socio-economic downstream impacts of specific weather triggers across different global regions.

Dataset Structure

The dataset is provided in JSON format. The core JSON structure contains the following nested objects for each disaster event:

1. Root Metadata & Impact Data

Standardized identifiers, geographical coordinates, dates, and quantitative human impact metrics (deaths, affected population, financial damage) sourced from EM-DAT.

2. weather_data

Retrieved via the Open-Meteo API, containing:

• Pre-event and post-event statistical summaries (mean rainfall, snowfall, min/max temperatures);
• A 21-day daily time series capturing the meteorological evolution before, during, and after the event.

3. satellite_data

Sourced via Copernicus Data Space Ecosystem (CDSE) / Sentinel Hub:

• Precise bounding boxes (bbox) for the affected areas;
• Temporal availability arrays for Sentinel-1 (SAR), Sentinel-2 (Optical/False Color), and Sentinel-3 (Thermal);
• Integration with ESA WorldCover land use data.

4. news_data

Metadata regarding the information retrieval process:

• Sources queried (Google News, ReliefWeb, IFRC, Wikipedia);
• Relevance scores and filtering penalties ensuring only high-quality, event-specific journalistic text is retained.

5. summary & causal_chain (The LLM Reasoning Layer)

• Summary: A concise, factual narrative of the event generated by Llama 3 70B.
• Causal Chain: Generated by Qwen 2.5 72B-Instruct. A normalized JSON array containing the chronological steps of the disaster. Each step includes:
  • type_event: A semantically normalized category (e.g., Flood, Casualties, Infrastructure Damage);
  • description: The specific manifestation of the impact;
  • supporting_quote: An exact or fuzzy-matched quote from the retrieved news context, ensuring a zero-hallucination rate in the causal extraction;
  • token_usage: Computational footprint tracking.

Methodology and Data Pipeline

The construction of this dataset followed a rigorous, multi-stage pipeline designed to eliminate LLM hallucinations and ensure physical consistency:

1. Event Anchoring: Sourcing base events from EM-DAT (post-2014) and geocoding locations via Nominatim;
2. Multimodal Retrieval: Automated fetching of historical weather data and satellite imagery footprints;
3. Textual Context Gathering: Scraping, parsing, and algorithmically scoring news articles for relevance;
4. Causal Extraction: Using Qwen-72B to extract causal graphs. A strict Python-based fuzzy-matching algorithm (threshold ≥ 0.80) was applied to cross-reference every LLM-generated quote against the raw news text. Any unverified causal step was systematically dropped;
5. Semantic Normalization: Over 500 raw event types were algorithmically normalized into a clean taxonomy of ~390 standardized categories.

Bias, Risks, and Limitations

• Media Coverage Bias: The richness of the causal chains and the availability of news data are inherently biased toward events that received significant international or English-language media coverage. Events in remote areas may have shorter causal chains.
• Satellite Data Availability: Cloud cover heavily impacts the usability of Sentinel-2 optical imagery during severe weather events (e.g., hurricanes). Sentinel-1 SAR is provided to mitigate this.

Citation

If you use this dataset in your research, please cite our paper:

Coming soon...