GeoRisk Intelligence

Landslides in mountainous regions require advanced predictive frameworks to mitigate escalating risks to communities and infrastructure. This study has proposed a novel hybrid ensemble Machine Learning (ML) models integrating Extreme Gradient Boosting (XGB), Stochastic Gradient Boosting (SGB), and Rotated Random Forest (RRF) with wrapper-based Recursive Feature Elimination (RFE) algorithm for landslide susceptibility mapping (LSM) in Uttarakhand, India, a Himalayan region chronically vulnerable to slope failures. Addressing critical gaps in conventional approaches, the framework incorporates landslide typology-specific modeling (soil, debris, rock) and infrastructure vulnerability quantification. A geospatial database of 5,200 historical landslides and 16 conditioning factors (CgFs) was optimized through hyper-parameter tuning via the Random Search (RS) method, enhancing model generalizability. The XGB-RFE model achieved superior predictive accuracy, validated through repeated cross-validation, with a peak area under the curve (AUC) of 0.996 for total landslides and 0.917–0.990 for typology-specific assessments, identifying slope, land use/land cover (LULC), topographic wetness index (TWI), and road proximity as dominant predictors. Geospatial analysis classified 38%–51% of the study area as Very High susceptibility, concentrated in the northern and northwestern zones of the study area characterized by steep slopes and dense infrastructure. Integration of Google Open Buildings data with landslide hazard assessments enabled the development of Uttarakhand first landslide vulnerability-building map, showing that 30.06% of 372,412 structures (112,000+ buildings) are located in high-risk zones. These results offer practical insights for disaster risk reduction and infrastructure planning, supporting policymakers in formulating proactive, data-driven strategies to enhance resilience in landslide-prone mountain regions.