Under review · Georisk
First author
Multi-Scale Pixel and Slope-Unit Landslide Susceptibility Mapping with Applicability-Domain Validation
Georisk (Taylor & Francis) · Karnali Highway corridor (~2,155 km²), western Nepal
Landslides along Himalayan road corridors cause repeated loss of life and long disruptions to access, yet most susceptibility maps never say how far their predictions can be trusted away from the training data. This study maps susceptibility for the Karnali Highway and pairs the map with an explicit reliability layer — a spatial guide to where the model should, and should not, drive engineering decisions.
Read the preprint (EarthArXiv) → doi.org/10.31223/X5C21WMethods & technical contribution
- Feature engineering: built nine topography, hydrology, land-cover and road-proximity conditioning factors, with raster workflows for NDVI/NDWI, resampling, masking, raster-stack preparation and prediction-map export (rasterio, GeoPandas, NumPy).
- Multi-scale design: compared pixel and slope-unit mapping units across moving-window scales of 3×3, 15×15 and 25×25 cells on a 12.5 m DEM.
- Modelling: trained and tuned Random Forest, RBF-SVM and XGBoost with Bayesian search, Platt-scaled probability calibration, and an equal-weight mean-probability ensemble (scikit-learn).
- Honest validation: used spatially blocked cross-validation with 1,000 m blocks to strip out spatial-autocorrelation optimism, then reported ROC/AUC, confusion matrices and SHAP feature-importance analysis.
- The reliability layer: built a Mahalanobis-distance applicability-domain mask that flags where predictions extrapolate beyond the training feature space, published as a trusted / untrusted binary.
- Deep learning: prototyped CNN workflows with Optuna-based tuning for patch-based hazard modelling (PyTorch).