73.74%, demonstrating substantially superior predictive performance. Although the ANN exhibits a slight reduction in ROC–AUC compared to training results, it remains the most effective approach for national-scale landslide susceptibility mapping. As shown in Table 2, the ANN concentrates 55.25% and 18.49% of landslide points in the “High” and “Very High” classes, respectively, indicating stronger discrimination of high-risk zones. In contrast, the Macro-regional model classifies most events as “Medium” (46.56%) with only 8.25% in “Very High,” while the Regional model presents a more moderate distribution (36.41% “Medium” and 14.00% “Very High”). Table 2 - Distribution of Categories Across Different Models (values in %) Model Very Low Low Medium High Very High ANN 1.13 7.32 17.81 55.25 18.49 MacroRegional 0.38 15.82 46.56 28.99 8.25 Regional 3.41 12.04 36.41 34.37 14.00 Beyond numerical metrics, targeted spatial comparisons were conducted in representative regions selected based on population density, landslide frequency, and economic impact. Across all analysed areas, as an example Figure 3 is attached, the ANN-based LSM consistently showed stronger spatial coherence with observed landslide distributions and terrain morphology. In contrast, the Macro-Regional and Regional models frequently exhibited contiguity bias, characterized by homogeneous susceptibility zones unrelated to local slope, curvature, or hydrological controls. This effect is likely linked to their reliance on large polygon-based thematic inputs with limited internal variability. Figure 3 – LSM Comparison - Arequipa - Cuzco. (a) Terrain 3D Model, (b) ANN, (c) MacroRegional, (d) Regional. Blue dots represent landslide A local downscaling experiment, currently under development, provides additional theoretical and empirical support for the robustness of the proposed framework (see Figure 4). When the nationally trained ANN model was applied at basin scale (e.g., Rímac Basin) using higher-resolution DEM inputs (0.25 m × 0.25 m), susceptibility patterns exhibited increased sensitivity to fine-scale morphometric variability, including sharper slope gradients, localized curvature contrasts, and improved representation of hydrological convergence. This behaviour is consistent with the well-documented self-affine nature of topographic surfaces, which preserve
RkJQdWJsaXNoZXIy MTM0Mzk2