This study ANN (1 hidden layer, 5 neurons) 95.55% Note: Accuracy and additional metrics (e.g., sensitivity, specificity) were also reported in some studies, but only AUC is shown for comparability. Shahri et al. (2019) developed a landslide susceptibility map for southwestern Sweden using an artificial neural network (ANN) trained on 14 input variables, including topographic, land-use, and geological factors. Their model achieved an AUC of 82.5% and an overall classification accuracy of 80.1%. Kumar et al. (2023) applied multiple machine learning algorithms—namely k-nearest neighbors (kNN), ANN, and Rotation Forest—in southern Peru. Their best-performing ensemble model reached an AUC of 87% and an overall accuracy of 79%. They also reported additional performance metrics for individual algorithms, including sensitivity values up to 0.73 and specificity values up to 0.84. Rosi et al. (2023) comprehensively employed a Random Forest model to conduct a large-scale landslide susceptibility analysis in Central Asia. Their model exhibited high stability, achieving a mean AUC of 93.12% across five independent training iterations, although classification accuracy metrics were not reported. Table 3 places the results of the present study within this broader context. The proposed ANN model achieved the highest AUC value (95.55%) among the reviewed studies, despite its relatively simple architecture. Most landslide susceptibility studies are developed and validated at a fixed spatial resolution, with limited assessment of model behaviour under changes in DEM scale. Qualitative heuristic maps in Peru (Villacorta et al., 2012; Luque et al., 2021) are intrinsically tied to predefined cartographic scales and thematic generalization, restricting their transferability to local applications without recalibration, while recent machine learning approaches. Kumar et al., 2023; Rosi et al., 2023 typically evaluate performance at a single resolution without testing spatial scalability. The down-scaling experiment conducted in this study demonstrates that the proposed ANN framework preserves predictive consistency when applied to higher-resolution DEMs at basin scale. This robustness is theoretically supported by the self-affine nature of topographic surfaces, whereby morphometric variables such as slope, curvature, and hydrological indices maintain scale-consistent statistical relationships while increasing geometric detail with resolution. Because the ANN approximates a continuous non-linear mapping between terrain attributes and landslide probability, enhancing spatial resolution refines the representation of predictor variables without altering the learned functional structure. The improved spatial alignment observed at local scale therefore reflects increased topographic fidelity rather than structural model modification, confirming the multi-scale applicability and operational scalability of the proposed methodology. 5. CONCLUSION The evaluation of the Artificial Neural Network (ANN) model demonstrates a substantial improvement in landslide susceptibility mapping accuracy compared to existing Regional and Macro-Regional models in Peru. Receiver Operating Characteristic (ROC) analysis yielded an Area Under the Curve (AUC) of 95.50%, confirming excellent predictive capability. When tested against an independent validation dataset, the ANN-based map correctly classified 73.74% of
RkJQdWJsaXNoZXIy MTM0Mzk2