Abstract Landslides pose a severe threat to infrastructure, local economies, and human life in the mountainous regions of Northern Pakistan, particularly during the intense precipitation events of the summer monsoon season. However, traditional Landslide Susceptibility Mapping (LSM) is often hindered by the lack of comprehensive, up-to-date historical landslide inventories in data-scarce developing regions. This study presents a highly reproducible, cloud-based methodology using Google Earth Engine (GEE) to overcome this severe data scarcity. A multi-criterion heuristic approach leveraging Topographic Slope (SRTM 30m) and Vegetation Loss (Sentinel-2 NDVI, 10m) was used to automatically generate 1,000 pseudo-absence and pseudo-presence training points. To explicitly prevent target leakage and circular logic, a Random Forest classifier (100 decision trees) was trained to predict landslide probability using strictly geomorphological predictors (Elevation, Slope, Aspect, and Laplacian Curvature), hiding the spectral vegetation data from the learning algorithm. The model achieved an Overall Accuracy of 88.3% (Kappa = 0.76, RMSE = 0.312) against a blind spatial validation split. The resulting 4-tier landslide susceptibility map reveals that over 25% of the study area is under high to very high hazard risk. This research provides a critical, highly scalable spatial planning tool for the Abbottabad and Mansehra districts, demonstrating a robust framework for landslide prediction in data-scarce, high-relief geomorphic environments. Keywords: Landslide Susceptibility, Random Forest, Google Earth Engine, Sentinel-2, Geomorphology, Tectonic Hazard, Northern Pakistan
Gul Nawaz (Wed,) studied this question.