391 Model performance was assessed using confusion matrices, ROC curves, and AUC values. Susceptibility was computed independently for each hydrographic unit and integrated into a national landslide susceptibility map (LSM) at 30 m resolution, totalling approximately 3.15 billion pixels and corresponding to a 1:100,000 map scale. The final LSM was classified into very low, low, medium, high, and very high susceptibility classes following Peruvian standards (Villacorta et al., 2012), using second-order derivatives of the probability density function derived from ANN output probabilities (0–1). Finally, a downscaling experiment was performed using higher-resolution DEM data to evaluate model behaviour under spatial resolution changes. The objective was to test predictive consistency when transferring the nationally trained ANN to basin-scale and mine-scale applications without retraining. 2.2.1 Landslide inventory and data preparation The national landslide inventory compiled by INGEMMET was used as the primary event dataset. The inventory classifies landslides into slides, rockfalls, debris flows, and related types based on lithological, structural, geomorphological, and historical analyses. Records are point-based and concentrated along road corridors, introducing spatial bias and uneven territorial coverage. To reduce this effect, only points representing identified source areas were retained; records located in depositional zones or with unclear origin were excluded. All retained points were validated using DEM-derived terrain analysis and satellite imagery. 2.2.2 Non-Landslide Sampling Non-landslide (stable) points were generated following the methodology of (Zhu, 2024) , with adaptations to ensure representative sampling. Candidate areas were restricted to slopes below 10° and to “green zones” (very low susceptibility) defined in the national Peruvian LSM \cite{villacorta2012mapa}. Random point generation was performed in (QGIS), with the number of non-landslide points set to three times the number of landslide points and a fixed random seed (1000) to ensure reproducibility. No minimum distance constraints were applied. This procedure resulted in 87,492 non-landslide points. 2.2.3 Independent Validation Dataset An independent validation dataset comprising 1,320 landslide events collected by INGEMMET between 2021-11-01 and 2023-07-17 was used for model evaluation. These data were filtered using the same criteria applied to the training inventory and were not available during model development, enabling an unbiased assessment and comparison with existing qualitative susceptibility models. For clarity, Figure 3 illustrates the spatial distribution of landslide, non-landslide, and validation points for a representative sector in the Cusco region. 2.2.4 Topographic and Geomorphological Landslide Factors Topographic and geomorphological predictors were derived from a 30 m × 30 m DEM (ASTER, MINAM). The following variables were computed: slope, aspect (classified into eight directional classes; Althuwaynee et al., 2012), and planar curvature (McNab, 1993). Hydrological indices included the Topographic Wetness Index (TWI; Beven & Kirby, 1979) and Stream Power Index (SPI; Moore et al., 1991). Land cover from MINAM was reclassified into seven categories. Structural geological factors were not explicitly included due to the absence of nationwide structural datasets; morphometric variables were considered indirect proxies. Slope and aspect were derived using terra, curvature using spatialEco, and TWI and SPI using whitebox.
RkJQdWJsaXNoZXIy MTM0Mzk2