Abstract
<title>Abstract</title> <p>Landslides triggered by intense monsoon rainfall and active tectonics pose a persistent threat to transportation corridors and settlements in the Mandi district of the Northwestern Himalaya. Most existing susceptibility assessments for this region rely on static geomorphological and historical inventories, overlooking pre-failure slope deformation that may indicate emerging instability. This study presents the first landslide susceptibility assessment for Mandi district that integrates Interferometric Synthetic Aperture Radar (InSAR) derived ground deformation with machine learning to produce a more dynamic and reliable hazard map. Sentinel-1 SAR data from 2023–2024 were processed using Persistent Scatterer Interferometry (StaMPS) and Small Baseline Subset analysis (MintPy) to detect active deformation, which was used to refine the Geological Survey of India's Bhukosh landslide inventory, adding 36 previously unmapped landslides and increasing inventory completeness by 18%. Twelve uncorrelated topographic, hydrological, geological, climatic, and anthropogenic conditioning factors were selected using Pearson correlation and variance inflation factor analysis. Four machine learning algorithms, Extreme Gradient Boosting (XGB), Random Forest, Logistic Regression, and Multilayer Perceptron, were trained and compared using a nested spatial block cross-validation framework across block sizes of 1,000–10,000 m. XGB achieved the best performance at an optimal block size of 3,000 m, and the InSAR-refined inventory improved model AUC by 12–15% relative to a conventional inventory. SHAP, mean decrease accuracy, and accumulated local effects analyses identified slope, distance to roads, distance to faults, and rainfall as the dominant controls on landslide occurrence. The resulting five-class susceptibility map shows that high- and very high-susceptibility zones, covering about 25% of the district, contain the vast majority of mapped landslides, with frequency ratios of 1.55 and 6.5, respectively. These results demonstrate that coupling InSAR-derived deformation with machine learning substantially improves the accuracy and currency of landslide susceptibility mapping, offering a scalable approach for hazard mitigation and land-use planning in tectonically active Himalayan terrain.</p>