Enhancing landslide susceptibility mapping in the Himalayas: geospatial and machine learning with explainable AI (XAI)

Abstract

Landslides present a critical hazard in the Himalayas, where steep topography, intense rainfall, and tectonic activity converge to destabilize slopes. Accurate delineation of high-susceptibility zones is essential to safeguard lives, infrastructure, and ecosystems. Here, we construct a comprehensive Landslide Susceptibility Map (LSM) for Uttarakhand, a landslide-prone state in northern India, by integrating advanced ensemble machine learning (ML) with explainable AI. Our analysis comprises 35 geo-environmental variables, ranging from historical landslide inventories and remote sensing data to GIS-based geomorphological, hydrological, and anthropogenic layers. We evaluate six ML models (Logistic Regression, Support Vector Machine, Random Forest, Extra Trees, Gradient Boosting, and eXtreme Gradient Boosting) before consolidating them into a stacking ensemble (SE), achieving an Area Under the Curve (AUC) of 0.987 on the training set and 0.979 on the test set. Across models, false-negative rates were low; Extra Trees minimized missed events (FNR = 3.5 %) but with a high false-positive rate (23.6 %), whereas XGBoost and the SE achieved a better sensitivity–specificity balance (FNR = 5.6 and 5.5 %, respectively) with comparatively lower false positives, favoring operational use. Spatial transferability to Sikkim was strong (Uttarakhand test accuracies 0.864–0.917; Sikkim 0.905–0.971), with XGBoost yielding the highest Sikkim test accuracy (0.971) and ensemble approaches (GB, XGBoost, SE) all exceeding 0.96, highlighting robust generalization across different Himalayan regions. Our ensemble model surpasses all individual models and classifies the study area into five susceptibility zones (very low to very high), with 18.20 % of Uttarakhand, particularly in Pithoragarh, Chamoli, and Rudraprayag districts, falling under high-susceptibility zones. Further interpretability is provided by SHapley Additive exPlanations (SHAP), which highlight key drivers of slope failure, including slope angle, fault proximity, and rainfall. Our findings highlight the value of combining robust ML techniques with geoscientific data, thereby enhancing hazard assessments and informing disaster risk reduction across the Himalayas and similarly vulnerable terrains worldwide.