Multi-Scale Pixel and Slope-Unit Landslide Susceptibility Mapping with Applicability-Domain Validation: Evidence from the Karnali Highway Corridor, Western Nepal

Ayush Karki, Aakash Mandal, Amrit Bhusal , Akhilesh Kumar Karna

Landslides along Himalayan Road corridors cause repeated loss of life and prolonged disruption of access, yet susceptibility products for these settings rarely report how far their predictions can be trusted away from the training data. This study maps landslide susceptibility for the Karnali Highway (NH 58) corridor in western Nepal (about 2,155 km2) and pairs the map with an explicit reliability layer. A landslide inventory and an equal set of non-landslide sites were characterised by nine conditioning factors spanning topography, hydrology, geology, land cover, and proximity to roads and drainage. Two mapping units were compared: a pixel representation evaluated at three moving-window sizes (3x3, 15x15, 25x25 cells on a 12.5 m DEM) and a slope-unit representation. Random Forest, a radial-basis support vector machine, and XGBoost were tuned with Bayesian search, calibrated with Platt scaling, and combined as an equal-weight mean-probability ensemble. To avoid the optimistic scores that spatial autocorrelation produces under random splitting, all tuning and evaluation used spatially blocked cross-validation with 1,000 m centroid blocks for the pixel pipeline and stratified random cross-validation for the slope-unit pipeline. Under this blocking the pixel models reached a mean cross-validation AUC of 0.96 to 0.98 across models and scales, with XGBoost the strongest; the single blocked-holdout fold reached 0.99 at the 15x15 window and is reported as a secondary diagnostic. The slope-unit models, a more conservative mapping unit, scored 0.76 to 0.91. Performance peaked at the intermediate window and declined at 25x25. Aspect was the leading predictor across every model and scale, followed by NDVI, slope, and the topographic wetness index, a pattern that points to monsoon-driven moisture loading on south-to-southwest-facing slopes. About 37 percent of the corridor falls in the combined High and Very High pixel susceptibility classes. A Mahalanobis-distance applicability domain, blended with inter-model agreement and published as a binary trusted/untrusted mask, flags where the prediction extrapolates beyond the training feature space; about 98 percent of slope units, and a comparable share of pixels, fell inside the trusted domain, giving practitioners a spatial guide to where the map should and should not drive engineering decisions