390 generating uplift, active faulting, and structural complexity. This geological framework, combined with intense precipitation and seismic activity, favors widespread landslide occurrence across the Andes (Figure 1). Figure 1 – Study area location and data classification map. Inset shows the distribution of testing, training, and new data (1995–2023) overlaid on a digital elevation model (DEM) of the Cusco province. Within this area, a total of 29,164 landslide events have been recorded by INGEMMET throughout the territory, mainly occurring in the high-prone areas of the Andes. 2.2. Methodology In this study, data from 1998-05-20 to 2021-11-01 and from 2021-11-01 to 2023-03-03 were used to assess the mapping of landslide susceptibility in Peru. A total of 29,164 landslides were considered, and this information was processed to create data for the algorithm. The methodology follows a six-step framework: (1) compilation and filtering of the national landslide inventory, (2) generation of non-landslide samples, (3) derivation of conditioning factors, (4) ANN training and validation, (5) basin-scale susceptibility mapping and (6) spatial downscaling assessment. Landslide evaluation was conducted at the hydrographic-basin scale, treated as a geologically closed system where topography governs the transformation of gravitational potential energy into kinetic energy, constraining landslide propagation within natural boundaries. A total of 229 hydrographic units (basins and inter-basins) were analysed (ANA, 2012). A national landslide inventory comprising 29,164 events (falls, slides, and flows) up to 2021-11-01 was compiled from INGEMMET via GEOCATMIN. Following (Zhu, 2024), 87,492 non-landslide points were generated for binary classification. An independent validation dataset of 1,320 landslide events (2021-11-01 to 2023-07-17), unavailable during model training (independent validation set), was used for performance assessment. Landslide conditioning factors included DEM, slope, aspect (8 classes), curvature, SPI, TWI, and land cover (7 classes), derived from MINAM DEM data. All variables underwent quality control using field verification, satellite imagery, and statistical optimization. Raster layers were resampled within each basin and merged with the target data, yielding 105,181 instances. A binary artificial neural network (ANN) was implemented using resilient backpropagation with weight backtracking (neuralnet, RStudio). Data were spatially split into training (76017-70%) and testing 29164-30%) subsets using random point selection (QGIS), and continuous variables were normalized and standardized. Multiple ANN configurations were evaluated, and the optimal model was selected based on performance metrics.
RkJQdWJsaXNoZXIy MTM0Mzk2