Soil erosion represents a major environmental issue that threatens ecosystem integrity and land sustainability, making the development of reliable susceptibility models crucial for supporting mitigation and management policies. This study evaluates the potential of three machine learning algorithms Weighted Subspace Random Forest (WSRF), Regularized Random Forest (RRF), and Naive Bayes (NB) for soil erosion susceptibility mapping in the Pardo River watershed, situated between the states of São Paulo and Minas Gerais, Brazil. A total of 120 sampling locations, including erosion and non-erosion occurrences, were identified through field surveys and high-resolution imagery obtained from Google Earth Pro. Initially, fifteen conditioning factors related to erosion processes were considered; however, after applying multicollinearity and relevance analyses, thirteen variables were retained for the final modeling framework. To evaluate model robustness, the dataset was randomly partitioned into training (70%) and testing (30%) subsets. Model performance was assessed using statistical indicators, including accuracy and AUC-ROC metrics. The NB, RRF, and WSRF models achieved accuracy values of 0.87, 0.89, and 0.88, respectively, while the corresponding AUC-ROC values reached 0.93, 0.96, and 0.95. Among the evaluated approaches, RRF yielded the highest predictive performance, demonstrating the effectiveness of machine learning techniques for supporting sustainable land management and erosion-prone area conservation. In addition, the proposed methodological framework offers a transferable approach for future susceptibility studies and contributes to expanding geospatial modeling applications across different environmental settings.