Debris flow susceptibility depends on complex factor interactions; traditional single models have limited accuracy and generalization in complex geological settings. This study develops a Stacking heterogeneous ensemble learning model based on watershed units, integrating 10 evaluation factors. It systematically compares five ensemble configurations and a single XGBoost model, analyzing driving mechanisms via Shapley Additive Explanations (SHAP) values. Results show that the optimal Stacking ensemble model, which uses linear SVM, XGBoost, and Gaussian Naive Bayes as base models and logistic regression as the meta-model, achieves an AUC of 0.927. Zones with extremely high susceptibility cover 24.24% of the area. Shapley Additive Explanations values reveal that intensive human engineering activities within 5 km of roads significantly enhance susceptibility by degrading vegetation and soil structure. Annual precipitation is the key water supply factor, while high-altitude snowmelt inhibits susceptibility. The Topographic Wetness Index in the range of 6.5∼10, Melton Ratio in the range of 700∼1000, and River Network Density in the range of 0.5∼1.5 synergistically regulate susceptibility via the coupled processes of water convergence, flow dynamics, and channel evolution. This study integrates the optimized Stacking model's high-precision prediction and mechanism interpretation, establishing a novel framework for debris flow susceptibility assessment in complex geological settings.
Ren et al. (Tue,) studied this question.