Machine-learning-based two-dimensional building exposure assessment to rainfall-triggered landslides in the South Sikkim Himalaya
Landslide risk in the Sikkim Himalaya is increasingly shaped by the building stock accumulating on steep, weathered slopes, yet quantitative information on the exposure of buildings to rainfall-triggered slope failure remains scarce at the settlement scale. This study couples a Random Forest (RF) landslide susceptibility model with a settlement-scale building inventory to quantify building exposure in South Sikkim, Indian Himalaya. Ten conditioning factors—slope, aspect, curvature, distance to rivers, distance to roads, distance to lineaments, Topographic Wetness Index (TWI), Land Use/Land Cover (LULC), Normalized Difference Vegetation Index (NDVI) and Stream Power Index (SPI)—were combined with a multi-source landslide inventory to train the classifier and produce a five-class susceptibility surface, which was then intersected with the building layer. The RF model attained an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.802, indicating good discriminative capability. High-susceptibility zones cover approximately 24% of the study area but contain 38% of the buildings, revealing a systematic over-representation of the built environment in the most hazardous terrain and a strong potential for cascading disruption of settlement services. The framework is data-parsimonious, reproducible, and directly transferable to comparable Himalayan settlements, providing an operational basis for hazard-informed regulation of new construction and for integrating exposure into landslide risk reduction in mountain regions.