Geospatial and Machine Learning Techniques for Landslide Risk Mapping in Sikkim, India

Identifying landslides and producing landslide susceptibility maps are essential components in supporting planners, local administrators, and decision-makers in effective disaster management strategies. The accuracy of these susceptibility maps plays a pivotal role in mitigating potential loss of life and property. Effective models for landslide susceptibility mapping require the integration of multiple factors that characterize both terrain features and meteorological conditions. Numerous algorithms have been developed and implemented in the literature to enhance the accuracy of these maps. This study employs a hybrid approach combining four machine learning techniques: Logistic Regression (LR), Random Forest (RF), Support Vector Classifier (SVC), and CatBoost, supplemented by a grid search to determine optimal hyperparameter settings. This methodological framework aims to achieve precise and reliable predictions for generating landslide susceptibility maps for Sikkim, India. In this study, eleven conditioning factors were considered, including aspect, slope, Land Use and Land Cover (LULC), elevation, distance to roads, distance to streams, the Normalized Difference Vegetation Index (NDVI), plan curvature, soil type, rainfall, and seismic activity. The performance of the models was assessed using several metrics, including training score, testing score, kappa, sensitivity, specificity, and Area Under the Curve (AUC). The results indicated that the random forest model outperformed the other models, achieving kappa and AUC values of 0.519 and 0.756, respectively, in developing susceptibility maps. Consequently, the random forest model emerges as the most reliable and effective tool for landslide susceptibility mapping within this study, making it an optimal choice for such predictive analyses.