Abstract: Widespread rainfall-induced shallow landslides occur frequently in the southeastern hilly region of China, and slope soil thickness is an important controlling parameter of landslide susceptibility. However, soil thickness shows strong spatial variability, unclear distribution patterns along slopes, and relatively low assessment accuracy in existing models. This study aims to reveal the spatial distribution characteristics of soil thickness in the southeastern hilly region and to develop a high-accuracy prediction model, thereby improving risk prevention and control of rainfall-induced widespread landslides. The study was conducted in a representative landslide-prone headwater catchment in Majian Town, Zhuji City, Zhejiang Province, China (area≈2.88 km²). A high-quality soil-thickness database was established using direct measurements, ground-penetrating radar (GPR), and UAV photogrammetry. The Boruta algorithm and recursive feature elimination (RFE) were applied to identify the main factors influencing the spatial distribution of soil thickness. Seven soil-thickness prediction models were developed using machine-learning methods, including random forest (RF), extreme gradient boosting (XGBoost), support vector regression (SVR), regression kriging (RK), and stacking ensemble approaches. The results show that the XGBoost model achieved the highest predictive performance, with soil-thickness prediction accuracy exceeding 90%. SHAP analysis was further applied to investigate model interpretability. The findings provide a reference for slope stability assessment and regional risk prevention and control of rainfall-induced landslides in the southeastern hilly region of China.