Mapping landslide susceptibility and assessing population exposure in Bhojpur District, Nepal, using machine learning

Landslides pose a major hazard to human life, livelihoods, and infrastructure in the Himalayan region of Nepal. There is a general understanding regarding the intensification of landslides due to environmental dynamics and human activities. This study develops a landslide susceptibility map for Bhojpur District, Nepal, using a Random Forest classifier, and conducts a population exposure analysis to identify human vulnerability to potential landslides. Eight critical factors such as elevation, slope, aspect, land cover, distance from roads, distance from streams, distance from rivers, and sediment transport index were incorporated for landslide analysis. These factors, together with a landslide inventory, were used to develop the Random Forest (RF) model. The class imbalance inherent in such studies was addressed through inverse-frequency class weighting, and model performance was evaluated using spatial block cross-validation, ensuring independent evaluation and reducing bias due to spatial autocorrelation. The RF model achieved a pooled Area Under the Receiver Operating Characteristic (ROC-AUC)of 0.798, an Average Precision (AP) score of 0.378 (3.6 times the random baseline), a Balanced Accuracy of 0.669, and an optimal-threshold F1-score of 0.424. The resulting susceptibility map indicates that 93.40% of the district falls within the Very Low Risk category, whereas only 0.61% is classified as High or Very High-Risk. Population exposure analysis based on 2026 projections shows that 140,244 residents (96.99%) occupy Very Low-Risk areas, while only 176 residents (0.12%) live in in High or Very High-Risk zones. This approach offers a spatially reliable method for landslide risk management and risk exposure assessment in resource-limited areas in the Himalayas.