Landslide Susceptibility Evaluation Based on Deep Learning and Imbalanced Sampling at Multi-Scale

The main objective of the present study was to conduct landslide susceptibility assessment and zoning analysis based on multi-scale imbalanced sampling and deep learning methods. Jiangkou Town, China, was selected as the study area. Grid cells with resolutions of 12.5 m, 25 m, and 50 m were chosen. Imbalanced sampling was applied using landslide/non-landslide ratios of 1:1, 1:2, and 1:3 to construct multi-scale modeling datasets. Susceptibility conditioning factors were screened using the frequency ratio (FR), Pearson correlation coefficient, and multicollinearity diagnostics and nine factors were obtained: slope, aspect, plane curvature, profile curvature, lithology, distance to river, distance to fault, annual rainfall, and land use. Six models—Logistic Model Tree (LMT), Kernel Logistic Regression (KLR), EfficientNet, ResNet, Transformer, and U-Net—were selected to establish 54 susceptibility evaluation models under various combinations of resolution and sampling ratios. The predictive reliability of the models was evaluated using receiver operating characteristic (ROC) curves and Kappa coefficients. Among the evaluated configurations, ResNet at a 12.5 m resolution with a 1:3 sampling ratio was retained as the preferred overall mapping configuration. It achieved a validation AUC of 0.963 and a Kappa coefficient of 0.778, together with strong susceptibility-zonation selectivity. The highest individual Kappa coefficient (0.819) was obtained by ResNet at a 25 m resolution with a 1:3 sampling ratio. Thus, the preferred configuration was identified through an integrated interpretation of the validation AUC, Kappa agreement, and susceptibility-zonation performance rather than by maximizing a single metric. The landslide susceptibility maps produced were classified into five levels and validated using the landslide distribution, landslide density and frequency ratio within each susceptibility zone. Most models showed good predictive performance in areas characterized by very high and very low susceptibility. According to the results of the comparison of the different susceptibility levels, it appears that ResNet and EfficientNet produced the most similar spatial predictions, while ResNet and Transformer presented the largest deviations. The deviations are mainly located near river valleys and areas with intense human activity. The proposed methodological framework and results can support disaster prevention, land-use planning, and regional risk management, particularly in mountainous areas with complex geological and topographic conditions.