Landslide Susceptibility Mapping for Railway Slopes Using Comparative Models and InSAR-Based Validation

Abstract We evaluated landslide susceptibility along the Carajás Railway in Brazil using five approaches: Random Forest, Extreme Gradient Boosting, Multilayer Perceptron, Support Vector Machine, and Weight of Evidence. The models integrated an inventory of 148 landslide events with geomorphological, hydrological, and land-use conditioning factors. We assessed predictive performance using receiver operating characteristic curves and the area under the curve. Random Forest performed best (0.859), followed by Extreme Gradient Boosting (0.850), Multilayer Perceptron (0.687), Weight of Evidence (0.669), and Support Vector Machine (0.459). We independently evaluated susceptibility predictions using ground-deformation time series derived from Sentinel-1 interferometric synthetic aperture radar data processed with the Small Baseline Subset approach. Deformation rates increased progressively across susceptibility classes, with the highest magnitudes concentrated in areas classified as highly susceptible. Random Forest and Extreme Gradient Boosting showed the strongest agreement between susceptibility patterns and observed deformation. The results indicated that interferometric deformation provided an independent physical assessment of susceptibility predictions and strengthened the interpretation of model outputs. The integrated approach improved the reliability of susceptibility assessment and supported risk-informed management of railway slopes.