Machine Learning-Based Assessment of Current and Future Sheet Erosion Susceptibility in the Oued Zat Watershed, High Atlas, Morocco

This study aimed to model the current and future sheet erosion susceptibility in the Oued Zat watershed, located in the High Atlas of Morocco, using machine learning and climate projections. Sheet-erosion occurrence data were derived from field surveys conducted in November 2023, during which characteristic signs of sheet erosion were identified and georeferenced, resulting in 295 erosion-occurrence points and 400 non-erosion points. Five Support Vector Machine (SVM) models were developed and compared: SVM-Linear, SVM-Linear-RFE, SVM-RBF, SVM-RBF-RFE, and SVM-Polynomial. Sixteen environmental and anthropogenic conditioning factors were integrated, derived from remote sensing products, ground-based precipitation observations, soil and geological maps, and climate projections. Future precipitation was obtained from the CanESM5 Coupled Model Intercomparison Project Phase 6 (CMIP6) global climate model under the SSP2-4.5 and SSP5-8.5 scenarios. The model performance was evaluated using specificity, sensitivity, and area under the receiver operating characteristic curve (ROC-AUC). Among the tested models, SVM-Polynomial achieved the highest predictive performance (specificity = 0.92, sensitivity = 0.81, ROC-AUC = 0.91), followed by SVM-RBF (specificity = 0.81, sensitivity = 0.79, ROC-AUC = 0.88), whereas SVM-Linear showed the weakest performance (specificity = 0.64, sensitivity = 0.63, ROC-AUC = 0.68). The current susceptibility assessment based on SVM-Polynomial indicated that the watershed is predominantly characterized by moderate (28.89%) and high (27.79%) susceptibility classes. Very-high-susceptibility areas accounted for approximately 14% of the watershed and were mainly concentrated in its southern sector, whereas low- and very-low-susceptibility classes, primarily distributed in the northern part of the watershed, occupied less than 30% of the total area. Climate projections indicate contrasting changes in sheet-erosion susceptibility under SSP2-4.5 and SSP5-8.5, depending on the model and susceptibility class. Both SVM-RBF and SVM-Polynomial project an expansion of very-high-susceptibility areas under SSP5-8.5 relative to SSP2-4.5, with relative increases reaching 27.40% and 62.43%, respectively, by the end of the century. However, the models show contrasting responses in the lower susceptibility classes, particularly low-susceptibility areas, which increase under SVM-RBF but decrease under SVM-Polynomial. These differences highlight the uncertainty associated with future susceptibility projections. Overall, the results identify areas potentially requiring priority attention in soil conservation and watershed management, while emphasizing that the projections represent relative susceptibility patterns rather than direct estimates of soil-loss rates or future erosion occurrence.