Spatial differentiation and refined susceptibility mapping of flash floods in Chongqing, southwest China

Chongqing's environmental heterogeneity complicates flash-flood susceptibility assessment. Using 596 historical occurrence points and an equal number of pseudo-absence samples, this study combined K-means clustering, frequency ratio analysis, and random forest modelling to compare unified and partitioned approaches. Two locally interspersed environmental types were identified: a drainage-network convergence type in the western hills and valleys and a mountain-hillslope type in the northeastern and southeastern mountains. Spatial block cross-validation showed higher recall for the mountain-hillslope type under partitioned modelling (69.14% versus 49.47%), but no statistically established overall advantage. At a common score threshold of ≥0.6, partitioned modelling retained most high-susceptibility areas identified by the unified model and expanded their extent, increasing historical occurrence-point coverage from 48.49% to 64.60%. The findings support environmental classification as a useful approach to type-specific susceptibility assessment in Chongqing, while its transferability requires independent validation.