Development of an LLM-driven Geospatial Mapping Service for Urban Disaster Decision Support

Abstract

Disaster information collected from news reports and emergency messages is often unstructured and cannot be directly used in a GIS environment. To support operational disaster monitoring at the Taipei Water Department (TWD), we developed an LLM-driven geospatial mapping service that converts multi-source disaster information into structured and spatially referenced data. The service extracts event type, location, and descriptions from disaster-related text, performs geocoding and spatial analysis, and provides web-based visualization. It was implemented using PostgreSQL/PostGIS, Flask, and Leaflet, with an end-to-end processing time of approximately 30 s per event. The service is currently in operational use at TWD for near-real-time disaster information monitoring, geospatial visualization, and decision support.