Unravelling landslide susceptibility controls using machine learning and explainable AI in a tropical river basin of the Western Ghats, India

Landslide susceptibility assessment is essential for risk-informed land-use planning in humid tropical river basins, where intense monsoon rainfall, steep terrain and expanding human disturbance interact to destabilise slopes. This study develops an explainable Light Gradient Boosting Machine (LightGBM) framework for landslide susceptibility mapping in the Manimala River Basin, Kerala, India. A landslide inventory and twelve conditioning factors, including slope, rainfall, topographic wetness index, lineaments, streams, roads, curvature, lithology, soil, and land use/land cover (LU/LC), were used to develop the model. The dataset was divided into training and testing subsets using a 70:30 ratio, and model performance was evaluated using AUROC, precision–recall, and confusion-matrix-based metrics. The LightGBM model showed strong predictive performance, with an AUROC of 0.894 and an average precision of 0.878. Threshold-based validation yielded an accuracy of 0.833, precision of 0.864, sensitivity of 0.792, specificity of 0.875, and an F1-score of 0.826, indicating balanced classification of landslide and non-landslide samples. The susceptibility map revealed that high- and very high-susceptibility zones together occupy 5.69% of the basin, mainly concentrated in the eastern and north-eastern uplands along the Western Ghats, whereas 83.66% of the basin falls under the least susceptible class. SHAP-based explainability identified slope as the dominant control on landslide susceptibility, followed in relative importance by rainfall, topographic wetness index (TWI), and distance to lineaments. The proposed explainable framework provides both accurate susceptibility mapping and transparent geomorphic interpretation for climate-resilient local planning.