An LLM-Constructed Database of Natural Disasters' Impacts

The consequences of natural disasters are described inconsistently across journals, media, and institutional reports, varying in precision, terminology, and credibility. For climate-impact assessment and adaptation planning, these fragmented observations hold essential information. We have developed a structured, quality-assured database of disaster impacts, constructed through a large-language-model (LLM)–assisted extraction pipeline that transforms heterogeneous publicly available sources into comparable, traceable indicators. Full-page documents from open-access scientific journals , multilingual news outlets, encyclopedic archives were ingested and filtered, preserving publication metadata and contextual cues (“who said what, when, and why”). A uniform event schema captures hazard type, timing, location, fatalities, displacements, losses, and institutional responses (evacuations, closures, power restoration). Automated routines handle deduplication, georeferencing, unit normalization, and conflict resolution, while provenance is retained at paragraph level. Four models—ChatGPT-4o, ChatGPT-4o mini, Gemini 2.5 Pro, and Gemini 2.5 Flash—were benchmarked on a 250-event ground-truth database (from different sources) using field-level similarity metrics for dates, locations, hazards, and numerical impacts. Gemini 2.5 Pro achieved the highest weighted accuracy, but ChatGPT-4o mini delivered near-parity with four- to ten-fold lower token cost and was adopted for production extraction. The curated ledger currently covers 185 events (1570–2025) across European coastal regions, encoded in FAIR-compliant formats (JSON, GeoJSON, NetCDF) and published via the OCEANIDS Zenodo community and GeoServer (WMS/WFS). Each record carries credibility scores, versioned values, and geocoding confidence, forming a reproducible, machine-readable foundation for climate-impact analysis and decision support.