Introduction: Gully erosion is a severe form of soil degradation, and although machine learning-based gully erosion susceptibility mapping (GESM) has shown strong potential, its application remains limited in tropical and subtropical environments such as gully-prone rural areas in Brazil. Materials and methods: This study evaluated random forest (RF), logistic regression (LR), support vector machine (SVM), and multilayer perceptron (MLP) algorithms for GESM in a gully-prone rural area in western São Paulo state. The methodology involved gully inventory mapping, environmental factor mapping, attribute selection, and a 70–30% train–test split, followed by susceptibility modeling and selection of the best-performing model, supported by systematic fieldwork. Results: Performance was excellent for MLP (AUC = 0.93) and RF (AUC = 0.92), and very good for LR and SVM (AUC = 0.89). Lower susceptibility was associated with forested areas, gentle slopes, and convex landforms, whereas higher susceptibility was associated with pasture areas, urban fringes, concave landforms, and hillslopes. High susceptibility was concentrated in the northern and northeastern sectors, where slopes are steeper and drainage density is higher. Conclusions: The results characterized key landscape features and enabled the production of a gully erosion susceptibility map, thereby improving understanding of local morphodynamics. The resulting map provides a valuable tool for identifying priority areas for soil conservation and erosion control, supporting environmental planning and land management in gully-prone rural landscapes.
Firmino et al. (Thu,) studied this question.