Machine learning-based landslide susceptibility assessment in the Uttarkashi District, Uttarakhand

Landslides are a major geological hazard that poses serious risks to infrastructure, transportation networks, and populations. This study develops a framework for evaluating landslide risk in the Uttarkashi District of Uttarakhand, India, using GIS and machine learning. With topographic, geological, hydrological, environmental, and anthropogenic characteristics, a landslide inventory comprising 1,994 landslide locations and the same number of non-landslide locations was prepared and evaluated using fourteen conditioning factors. Models like Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM) were trained and evaluated using spatial five-fold cross-validation and an independent test dataset. Elevation, distance from roads, NDVI, aspect, lithology, and land-use/land-cover were found to be among the most significant predictors using feature-importance analysis, albeit their relative significance differed between models. The mean AUC values for RF, XGBoost and LightGBM were 0.8805, 0.8846, and 0.8840, respectively, based on the spatial cross-validation data. LightGBM obtained an AUC of 0.8950 on the independent test dataset, whereas RF and XGBoost obtained an AUC of 0.8988. High- and very-high-susceptibility zones were frequently identified on the resulting susceptibility maps along the Yamuna and Bhagirathi River systems as well as important transportation routes like the Yamunotri and Gangotri highways. Additionally, this study shows how machine learning combined with remote sensing and GIS can effectively assess the susceptibility of a region to landslides. In the Himalayan context, it offers spatial data to support infrastructure management, slope stabilization, land-use planning, and disaster risk reduction.