A high-resolution global flood susceptibility dataset derived from multi-source earth observation and geospatial data

Abstract

Floods are among the most damaging disasters worldwide, necessitating high-resolution modeling of where terrain and environmental conditions predispose societies to flooding. While geospatial data are invaluable for flood susceptibility mapping, comprehensive and globally consistent datasets remain scarce. We present the Global Flood Susceptibility Map (GFSM v1), a globally harmonized 30 m flood susceptibility dataset produced by incorporating multi-source flood conditioning factors accounting for topographic, hydrological, meteorological, and anthropogenic characteristics within a machine learning framework. A gradient-boosted tree model was trained on ~17,000 curated grid tiles with ~30.45 million training samples, using a global flood inundation dataset as training labels and nine conditioning factors (elevation, slope, aspect, wetness index, vegetation index, height above nearest drainage, water/roads proximity, and rainfall frequency). The model was deployed based on geopolitical and climate strata (units n = 192) to generate global pixel-level susceptibility. Rigorous cross-validation indicates strong model predictive accuracy (median area-under-curve ~0.95), and the output captures localized flood-prone regions worldwide. GFSM v1 offers a global open-source, high-resolution resource to support informed flood risk management, particularly in data-sparse regions.