Mountainous Areas Landslide Susceptibility Evaluation Based on Machine Learning

Abstract

Landslides pose a persistent threat in the geologically complex and ecologically fragile mountainous terrain of Yanbian Prefecture, Northeast China. To address this challenge, we propose an innovative susceptibility assessment framework that integrates ensemble machine learning with SHapley Additive exPlanations for interpretable prediction. Eleven critical environmental and anthropogenic factors—altitude, slope, aspect, curvature, lithology, land use, rainfall, TWI, NDVI, and distances to rivers and roads—were selected to build a comprehensive indicator system based on hydrological watershed units. A suite of advanced machine learning algorithms, including Random Forest, XGBoost, LightGBM, CatBoost, and AdaBoost, were employed and further optimized using ensemble strategies such as Stacking, Bagging, and Voting. Among them, the Stacking ensemble demonstrated superior predictive performance with the highest AUC value. More importantly, the integration of SHAP allowed for a transparent and quantitative interpretation of feature contributions, revealing that distance to roads, rainfall, and NDVI are the dominant drivers of landslide susceptibility in the region. This study not only advances the precision and interpretability of disaster prediction models but also offers practical insights for regional hazard mitigation, land-use planning, and sustainable ecological management.