Water erosion is one of the most widespread land degradation processes in arid and semi-arid mountainous regions, causing significant soil loss and severely impacting natural resources. This study aims to assess water erosion susceptibility in the Upper Tassaoute watershed (High Atlas, Morocco) using two machine learning models: Random Forest (RF) and Support Vector Machine (SVM). An inventory of approximately 200 eroded sites, established through the integration of field observations and satellite imagery, was used for model training (70%) and validation (30%). Twenty environmental conditioning factors were selected, encompassing topographic, geological, climatic, soil, and vegetation variables. The performance of both models was evaluated using the Receiver Operating Characteristic (ROC) curve and the Area Under the Curve (AUC), showing satisfactory predictive accuracy for both RF and SVM. The analysis of variable importance revealed that NDVI, slope, curvature, soil properties, and lithology were among the most influential factors. The results confirm the effectiveness of machine learning approaches for mapping water erosion vulnerability and provide a robust scientific basis to support sustainable land management strategies in sensitive mountainous environments.
Nait-taleb et al. (Thu,) studied this question.