<p>Evaluating landslide susceptibility is crucial for reducing landslide risks and improving early warning systems, thereby enhancing disaster preparedness and safeguarding vulnerable communities. This study aims to apply machine learning methods for the spatiotemporal analysis of landslide susceptibility in the Hindu Kush Himalaya (HKH) region, integrating geospatial data and advanced modeling techniques to improve the accuracy of risk area identification. Four models—Gradient Boosting Machine (GBM), Random Forest (RF), Support Vector Machine (SVM), and k-Nearest Neighbor (kNN)—were used for predictive analytics. The GBM model demonstrated the highest performance with an Area Under the Curve (AUC) of 0.93, followed by RF (AUC of 0.92). The analysis indicates that 35–40% of the HKH region is highly susceptible to landslides, with hotspots located in Nepal, northern India (Uttarakhand and Himachal Pradesh), the western Himalayan region, and parts of Myanmar. Most landslides occur during the monsoon season, with key contributing factors including rainfall intensity, seismic activity, slope gradient, and anthropogenic activities such as deforestation and illegal mining. Model validation using cross-validation techniques confirmed strong predictive performance, despite limitations related to data resolution and potential biases. These findings provide a framework for landslide risk management, supporting disaster preparedness, early warning systems, and land-use planning in the HKH region and other vulnerable areas globally.</p> Graphical Abstract <p>The graphical abstract displays an integrated machine learning-based spatiotemporal framework for landslide susceptibility assessment in the Hindu Kush Himalayan (HKH) region. The process integrates five thematic components: study area overview, conditioning factors analysis, data processing and modeling pipeline, spatiotemporal pattern investigation, and model-based outputs. The HKH region and prominent conditioning factors such as slope, elevation, curvature, rainfall, NDVI, geology, land use/land cover, proximity to roads and rivers, and anthropogenic stressors like illegal mining and deforestation are depicted in the left panel. The methodology, such as landslide inventory compilation, stratified random sampling, factor selection, and application of four supervised machine learning algorithms—Gradient Boosting Machine (GBM), Random Forest (RF), Support Vector Machine (SVM), and k-Nearest Neighbor (kNN)—are emphasized in the middle panel. Feature importance estimation and model performance using accuracy metrics like precision, recall, F1-score, and the Area Under the ROC Curve (AUC) are also represented. The right panel displays the spatiotemporal distribution of 2004–2017 landslides, which mirrors seasonally clustered activity during the monsoon season. It also shows triggering factor attribution (e.g., construction, seismicity, freeze–thaw, rainfall), model output reflecting susceptibility classes (very low to very high), and performance comparisons among all four models. The GBM model performed better with an AUC of 0.93, followed closely by RF (AUC = 0.92), demonstrating high predictive ability. Results depict high-susceptibility clusters in Nepal, northern India (Uttarakhand and Himachal Pradesh), the western Himalayas, and northeastern Myanmar. This graphical integration shows the advantage of combining geospatial analytics with advanced machine learning to inform landslide risk assessment, early warning systems, and policy-level land-use planning in exposed mountain terrain.</p> <p></p>