their statistical structure across spatial scales while allowing geometric detail to increase with resolution. Figure 4 – Local LSM downscaling experiment using ANN (Slope with road construction) Because slope, curvature, SPI, and TWI are deterministic derivatives of the DEM, enhancing spatial resolution refines the discretization of the same underlying morphodynamic surface rather than altering its governing relationships. From a modelling perspective, the ANN approximates a continuous non-linear function linking terrain attributes to landslide probability; therefore, improving input resolution enhances the representation of predictor variables without requiring modification of the learned functional mapping. The improved spatial alignment with documented landslide clusters under downscaling thus reflects increased topographic fidelity rather than changes in model structure. This demonstrates that the ANN framework operates as a resolution-adaptive classifier, capable of maintaining predictive consistency across scales. Consequently, the methodology is not constrained to a single mapping resolution and can be reliably transferred from national assessments to basin-scale and local hazard analyses without retraining, reinforcing its scalability and operational applicability. 4.2 Discussion of Results and Comparison with Previous Studies Numerous studies have applied machine learning techniques to landslide susceptibility mapping across diverse geographic contexts. Despite differences in model architecture, input variables, and spatial scale, the Area Under the Curve (AUC) remains the primary metric for performance evaluation. Table 3 presents a comparative summary of AUC values from selected recent studies. Table 3 - Comparison of AUC values in recent landslide susceptibility mapping studies Study Model Used AUC Shahri et al. (2019) ANN (14 input variables) 82.5% Kumar et al. (2023) Ensemble (kNN + ANN + RF) 87% Rosi et al. (2023) Random Forest 93.1%
RkJQdWJsaXNoZXIy MTM0Mzk2