Complementary insights from empirical and machine learning approaches for soil erosion susceptibility mapping in semi-arid mountainous regions

Soil water erosion poses a major environmental challenge in semi-arid regions, necessitating reliable susceptibility mapping approaches. This study compares the empirical Revised Universal Soil Loss Equation (RUSLE) with three machine learning (ML) models—Random Forest, Support Vector Machine (SVM), and XGBoost—for erosion susceptibility mapping in the Western High Atlas sub-basins (Morocco). RUSLE was applied to estimate sheet and rill erosion, whereas ML models were trained on gully erosion field inventory data and 22 environmental factors, enabling a complementary spatial comparison rather than a direct process-equivalence test. Spatial concordance analysis revealed >70% pixel-level agreement between RUSLE and ML outputs. Among the <30% discordant pixels, RUSLE overestimated ML-predicted susceptibility in approximately 14% of cases and underestimated it in about 15%. XGBoost achieved the highest predictive performance (AUC = 0.97; spatial cross-validation AUC = 0.890), outperforming RUSLE (AUC = 0.86 under a field-informed binary classification). SHAP analysis identified slope as the dominant predictor, followed by lithology, Topographic Position Index (TPI), and rainfall. We conclude that ML models provide robust and interpretable susceptibility maps, while RUSLE remains valuable for capturing broad spatial patterns. The <30% disagreement zones highlight priority areas for hybrid modeling approaches and targeted field validation, offering a pathway toward more integrated erosion risk assessment in data-scarce semi-arid environments.