Enhancing landslide susceptibility predictions with XGBoost and SHAP: a data-driven explainable AI method

Abstract

Landslide susceptibility mapping is essential for disaster risk management, especially in geologically sensitive regions like the Himalayas, where steep slopes, heavy rainfall, and human activities intensify risks. This study integrates eXtreme Gradient Boosting (XGBoost) with SHapley Additive exPlanations (SHAP) to enhance both predictive accuracy and interpretability in Rudraprayag and Tehri Garhwal, Uttarakhand, India. Using high-resolution geospatial datasets, 20 conditioning factors were analyzed, with feature selection refined through Variable Inflation Factor (VIF) and Recursive Feature Elimination (RFE). The XGBoost model achieved 92.87% accuracy and an AUC of 0.96, outperforming traditional methods. SHAP analysis identified Distance from Roads, Elevation, Rainfall, and Slope as key influencing factors. The susceptibility map categorized 13.10% of the area as high or very high susceptibility, highlighting the need for targeted mitigation. This study underscores the role of explainable AI in landslide risk assessment, providing actionable insights for disaster preparedness and sustainable development.