Dynamic Monthly Population-Exposure-Based Flash-Flood Risk Mapping Using an Explainable XGBoost Framework

Flash floods are sudden and highly destructive, posing serious threats to human life. Conventional annual-scale or static susceptibility assessments cannot adequately capture intra-annual dynamics, while susceptibility alone does not directly represent potential population risk. This study integrated a 1990–2016 flash-flood inventory, monthly precipitation, multisource topographic and environmental factors, and population exposure to develop an XGBoost- and SHAP-based monthly flash-flood susceptibility (FFS) model for Hunan Province, China. As a retrospective application, population exposure was then incorporated to map population-exposure-based flash-flood risk (FFR) from January to December 2024. Independent temporal testing for 2013–2016 showed good model performance (AUC = 0.82). SHAP identified monthly maximum 1-day precipitation (M1P) as the most important factor (23.0%), followed by elevation (21.1%) and the topographic wetness index (14.7%). FFR increased markedly in June–July, peaking in July, when high- and very-high-risk areas covered 26.86% of the province, and declined sharply in August, remaining relatively stable thereafter. High-risk areas were mainly concentrated in eastern, southeastern, and parts of central Hunan. Jaccard similarity between high-to-very-high FFS and FFR ranged from 0.478 to 0.687, indicating that population exposure reshaped the spatial priority of susceptibility hotspots. The framework provides insights into seasonal variations in population-exposure-based flash-flood risk and supports improved understanding of spatial risk patterns.