Landslide Susceptibility Assessment Using an Explainable Stacking Learning Framework (ESLF) in Xinyuan County, Xinjiang, China

To address the lack of transparency in machine learning methods for landslide susceptibility assessment (LSA), this study proposes an Explainable Stacking Learning Framework (ESLF), taking Xinyuan County, Xinjiang, China as the study area. The framework effectively integrates Deep Neural Networks (DNN), Random Forests (RF), and Support Vector Machines (SVM), leveraging their respective strengths in pattern recognition, decision analysis, and hyperplane-based classification. A comprehensive landslide inventory and 11 predisposing factors were compiled to generate susceptibility maps through the application of Stacking, alongside individual DNN, RF, and SVM models. The results indicate that the stacking model outperforms single models, achieving AUC of 0.907, accuracy of 0.930, F1-score of 0.897, and a Kappa coefficient of 0.866. Very high and high susceptibility zones are mainly distributed in southern Talede Town, Biesituode Township, Xinyuan Town, Alemale Town, western Nalati Town, and northeastern parts of Zeketai, Areletuobie, Kansu Towns, and Tuergen Township. SHAP (SHapley Additive exPlanations) permutation importance analysis identifies elevation (995–2,253 m), distance to rivers (