DisasterFlow AI: An AI-Powered Spatiotemporal Disaster Intelligence System for Flood Forecasting and Emergency Resource Allocation

Flood risk in dense urban environments such as Mumbai is spatially non-uniform and time-varying, making single-value, city-wide forecasts inadequate for emergency planning. This paper presents DisasterFlow AI, an end-to-end disaster-intelligence prototype that represents the city as a graph of 140 spatial monitoring nodes (944 edges) and combines a Spatiotemporal Graph Neural Network (ST-GNN) forecasting pipeline with an AI Priority Engine and a resource-allocation module. Hourly features (multi-window rainfall, rainfall anomaly, humidity, distance to water, elevation and slope) are engineered into a weighted flood-risk score and classified into LOW/MODERATE/HIGH/EXTREME categories with 6-, 12- and 24-hour-ahead targets. The forecast is fused with operational and resource-gap signals to rank emergency locations and compare required versus allocated resources. The system is served through a FastAPI backend and a Leaflet-based GIS dashboard. In the demonstrated state, it achieved 100% ST-GNN forecast coverage across 140 nodes, allocated 560 resources against 257 required with zero shortage locations, and classified 8 nodes as high, 28 as medium and 104 as low risk. The paper also documents the prototype’s limitations relative to a live-sensor production deployment.