Evaluating hybrid and machine learning models for landslide susceptibility mapping in a high-risk region of the Western Ghats, India

A critical component of managing landslide hazard is the development of landslide susceptibility maps (LSM). This study evaluates and compares the performance of five machine learning models; Random Forest, XGBoost, LightGBM, CatBoost, and a novel hybrid classifier for LSM across 20 high-risk districts identified from Landslide Atlas of India (2023). Fifteen landslide conditioning factors were considered and a three-stage selection process (VIF, Pearson correlation, ReliefF) reduced these to eleven factors. A landslide inventory (2015–2020, Bhukosh) was used for calibration and validation. All models demonstrated excellent performance, achieving accuracies >0.9. Among them, the hybrid model outperformed others, showing lower errors, higher stability and reliability. The spatial agreement of LSM zones was evaluated using kappa statistics, user accuracy, and producer accuracy. While very high and very low zones showed strong consistency across models, intermediate zones posed greater uncertainty. Approximately 21% of the study area lies within high to very high susceptibility, with the Idukki, Nilgiris, Wayanad, Kozhikode, and Kodagu being the most vulnerable area. The SHapley Additive exPlanations (SHAP) analysis revealed that elevation, slope and rainfall are the most influential factors. These results provide a robust framework for landslide mitigation, disaster risk reduction, and sustainable planning in the Western Ghats.