Application of Multiple Machine Learning Models to Rainstorm Disaster Loss Assessment and Driving Mechanism Analysis in Zhejiang Province, China

Study region: This study targets 76 counties in Zhejiang Province, China, frequently affected by rainstorm disasters. Study focus: Rainstorm-induced losses arise from complex interactions between extreme weather and human-environment systems. To assess these losses, we propose a framework with five machine learning models—MLP, Random Forest, CatBoost, LightGBM, and XGBoost. Using 461 disaster records from 2001 to 2019, a multi-dimensional indicator system was developed, covering hazard, exposure, vulnerability, and environmental factors. Model performance was evaluated using accuracy, F1 score, ROC-AUC, and Cohen’s Kappa. Feature importance and SHAP analysis identified key drivers and explained behavior. New hydrological insights for the region: This study advances the understanding of compound mechanisms underlying rainstorm-related economic losses in Zhejiang Province. Among the five models, XGBoost exhibited relatively better performance than the other models in assessing rainstorm disaster losses. Short-duration extreme rainfall emerged as the dominant hazard driver, while socioeconomic indicators, particularly population density and per capita GDP—amplified disaster severity under high-exposure conditions. SHAP analysis revealed nonlinear threshold effects: losses escalate rapidly when extreme rainfall coincides with densely built environments. Meanwhile, environmental factors such as terrain slope and vegetation cover played a buffering role in mitigating low to moderate losses. These findings highlight the value of interpretable machine learning models in capturing the spatial heterogeneity and compound nature of rainstorm disaster impacts, providing scientific support for targeted risk zoning and mitigation planning.