Transformation of Geospatial Modelling of Soil Erosion Susceptibility Using Machine Learning

Soil erosion presents substantial environmental and economic challenges, especially in areas prone to land degradation. This study assesses the use of Machine Learning (ML) methods—Support Vector Machines (SVM) and Generalized Linear Models (GLM)—to model Soil Erosion Susceptibility (SES) in the Saddang Watershed, Indonesia. It incorporates environmental, hydrological, and topographical factors to improve prediction accuracy. The approach includes multi-source geospatial data collection, erosion inventory mapping, and relevant factor selection. SVM and GLM were applied to classify SES, with performance evaluated using accuracy, AUC, and precision metrics. Results show SVM classified 40.59% of the area as moderately susceptible and 38.50% as low susceptibility. GLM identified 24.55% as very low and 38.59% as low susceptibility. Both models demonstrated high accuracy (SVM: 87.4%, GLM: 87.2%) and strong AUC values (SVM: 0.916, GLM: 0.939), though GLM showed better specificity and recall. Feature importance analysis highlights that GLM favors hydrological factors like river proximity and drainage density, while SVM balances across various environmental inputs. These findings affirm the value of ML-based geospatial modeling for SES assessment, supporting interventions such as reforestation and erosion control. SVM is suitable for localized planning, whereas GLM offers strategic-level insights. This research contributes to advancing environmental modeling by embedding domain knowledge into ML frameworks, and suggests future work integrate real-time remote sensing and more sophisticated models for broader SES prediction.