This study comprehensively implemented three machine learning (ML) algorithms, namely XGBoost, RF and SVM, to assess the spatial distribution of soil erosion vulnerability within the sub-watersheds of the Western High Atlas (Morocco). To ensure reliable results, a robust methodological approach was adopted, incorporating an inventory of eroded/non-eroded areas, Sentinel satellite observations, and a digital surface model. After a multicollinearity analysis, the five factors of the Revised Universal Soil Loss Equation (RUSLE) model were selected for modeling to assess their reliability in predicting erosion. The results obtained from the erosion susceptibility analysis across all sub-watersheds indicate that the high to very high risk categories account for an average of 30.24% and 39.06%, respectively. In contrast, 10.24% and 15.68% of the areas were classified in the low and very low risk categories, respectively. The effectiveness of the models applied in this study was evaluated using several performance metrics, including accuracy, precision, recall/sensitivity, and specificity. The three tested models demonstrated a promising ability to predict and map the most erosion-prone areas, with AUC-ROC values reaching 0.95 for XGBoost, 0.94 for RF, and 0.92 for SVM, respectively. The innovative contribution of this study lies in improving the results of the RUSLE model by combining it with machine learning algorithms, with a particular focus on identifying sensitive areas and quantifying soil loss percentages in a semi-arid context. It provides practical solutions for the sustainable management of watersheds, as well as the preservation of land and water resources in the face of soil degradation processes.