Landslide susceptibility assessment in the Altai Mountains, Xinjiang, China: integrating multi-source conditioning factors and interpretable machine learning

The cold alpine Altai Mountains in Xinjiang, China, are affected by strong topographic incision, complex tectonics, and valley-corridor engineering disturbance, yet regional landslide susceptibility studies integrating multi-source factor screening, robustness testing, and model interpretation remain limited. This study developed an interpretable machine-learning framework using 625 inventoried landslide points. Twenty factors from topography, geology, hydrology, environment, and human activity were screened using Spearman correlation, variance inflation factor (VIF), and information gain, retaining 12 factors. Logistic Regression (LR), Random Forest (RF), and Deep Forest (DF) models were evaluated using receiver operating characteristic (ROC) curves, confusion matrices, repeated resampling, Bootstrap confidence intervals, DeLong tests, spatial cross-validation, and threshold sensitivity analysis. SHapley Additive exPlanations (SHAP) and frequency ratio (FR) analysis were used for attribution and spatial consistency diagnosis. RF and DF outperformed LR, while their AUC difference was not significant. Considering robustness and computational efficiency, RF showed the best overall applicability, whereas DF provided auxiliary spatial discrimination in complex local backgrounds. SHAP and FR identified distance to roads, elevation, mean annual precipitation, and multi-scale surface deviation as key controls, indicating the combined influence of engineering disturbance, topographic zoning, and hydro-meteorological conditions. The results support landslide risk identification and zonal management in the Altai Mountains.