Abstract
Precise assessment of landslide potential in tectonically active mountainous areas like the Darjeeling Sikkim Himalaya (DSH) is a scientific challenge because of the complexity of the terrain and the factors controlling landslides. Despite several studies on landslide susceptibility mapping in the DSH, most conventional pixel-based approaches struggle to demarcate the geomorphological units and complex nonlinear relationships among the landslide conditioning factors. Furthermore, incomplete landslide inventories can also hamper landslide susceptibility assessment quality. Due to all these limitations, more robust methods integrated with geomorphologically based terrain segmentation and advanced machine learning methods are required, which will not only determine the landslide probability zones but also help to understand the landslide driving forces and interaction analysis among the geoenvironmental factors controlling landslides. In the present study, we have used slope unit-based landslide susceptibility mapping, using 4830 slope units, 17 conditioning factors and 8373 of the total updated landslide inventories with six models: random forest (RF), a generalized additive model (GAM), Categorical Boosting (CatBoost), Tabular Neural Network (TabNet), Bayesian additive regression trees (BARTs), and a convolutional neural network (CNN). The model hyperparameters were optimized using Bayesian optimization, except for the BART model. Based on the area under the receiver operating characteristic curve AUC value, RF (AUC = 0.848), BART (AUC = 0.847) and CatBoost (AUC = 0.846) achieved the highest predictive performance. Furthermore, SHAP based analysis shows that most of the landslides are controlled by elevation, aspect, slope, distance to faults, NDVI, proximity to roads and drainage networks. Whereas the SHAP-based interaction analysis shows that most of the landslides are controlled by non-linear, threshold-dependent relations, especially between slope rainfall, rainfall soil moisture, slope distance to roads and fault, which represents a complex interaction between the hydrological factors, geomorphic processes, tectonic activity and human interventions. This study can be used for disaster risk reduction and regional infrastructural planning in DSH.