Integrating Kernel-Based Machine Learning Techniques with Geospatial Analysis for Landslide Susceptibility Mapping in Sikkim, India

Abstract

Landslides are among the most destructive natural hazards, causing substantial loss of life and infrastructure, particularly in mountainous regions. Accurate identification of landslide-prone areas remains challenging due to complex and nonlinear interactions among geological, topographical, and environmental factors. This study evaluates kernel-based machine learning approaches, namely, kernel logistic regression (KLR) and support vector machine (SVM), for landslide susceptibility mapping (LSM) in Sikkim, India. A balanced data set comprising 693 landslide and 695 nonlandslide samples was constructed using 15 conditioning factors representing topographic, hydrological, environmental, geological, and seismic conditions. The grid search method was utilized to identify optimal hyperparameter configurations. Model performance was quantitatively assessed using accuracy, kappa coefficient, specificity, area under curve (AUC), landslide density, and success rate curve (SRC). The results demonstrate that the KLR with a polynomial kernel (KLR_poly) achieved the highest predictive performance, with an accuracy of 0.741 and an area under curve (AUC) of 0.829, outperforming other KLR variants, while the SVM with an rbf kernel (SVM_rbf) achieved the highest predictive performance, with an accuracy of 0.749 and an AUC of 0.764, outperforming other SVM variants. Furthermore, the landslide density analysis showed a systematic increase in landslide concentration from low to high susceptibility classes for nonlinear models, while SRC results indicated moderate to good predictive capability across all models, confirming effective spatial prioritization of landslide-prone areas. Additionally, a feature importance analysis was conducted to identify the most significant factors influencing landslide occurrence.