Integrating machine learning and numerical methods for enhanced landslide susceptibility and hazard mapping in the Bhotekoshi watershed, central Nepal

Landslides pose a significant threat in the mountainous regions of Nepal. Landslide susceptibility maps are commonly used to identify potential landslide zones by statistically analyzing geological, topographical, and hydrological factors, assuming that similar conditions may trigger future failures. While such maps provide valuable insights into landslide-triggering conditions, they are limited in assessing risk to settlements and infrastructure located downslope or in valley bottoms. This study integrates machine learning based landslide susceptibility with numerical runout modeling to provide a comprehensive landslide hazard assessment in the Bhotekoshi watershed, overcoming the limitations of traditional models that focus solely on statistical susceptibility. To conduct the susceptibility analysis, a total of 439 landslides were mapped from 2012 to 2021 using satellite images. Of these, 70% were used for training two machine learning (ML) models: random forest and Xtreme Gradient Boosting (XGBoost), and the remaining 30% were used for validation. Among the two ML models, Random Forest model demonstrated slightly superior performance, achieving higher predictive accuracy. After the machine learning susceptibility analysis, the study transitions into a regional-scale landslide runout analysis. First, a back analysis of the past landslide event was conducted to fine-tune the model parameters (internal angle of friction and basal friction angle) and validate performance of the runout model. Following the back analysis, the regional-scale numerical modeling of landslide runout was conducted by designating areas classified as the highest susceptibility class in the Random Forest susceptibility map as potential release zones. This approach allows for a detailed examination of landslide propagation and potential impacts along the downslope settlements and infrastructures. The analysis clearly demonstrates that integrating both machine learning and numerical runout methods significantly increases the estimated exposure of population, buildings, and roads within the very high hazard class compared to relying solely on susceptibility methods. Specifically, population exposure rises from 360 to 7743, buildings increase from 97 to 2771, and road exposure expands from 41 to 251 km. This result highlights the significant risk of underestimating exposure in the analyses that solely rely on landslide susceptibility models. Integration of susceptibility and runout analysis improves landslide risk assessment, aiding in land-use planning and disaster mitigation strategies.