Integrating Machine Learning With Geographic Information Systems and Remote Sensing for Erosion Risk Mapping in The Tamalate Watershed

Abstract

Soil erosion poses a significant threat to environmental sustainability, particularly in regions with complex topographic and hydrological characteristics. Accurate erosion risk mapping is essential for effective land management and mitigation strategies. This study aims to evaluate the performance of five machine learning models—Random Forest (RF), Gradient Boosting Tree (GBT), Decision Tree (DT), Generalized Linear Model (GLM), and Support Vector Machine (SVM)—in predicting erosion risk using remote sensing-derived indices. Eight environmental variables, including topographic, hydrological, and vegetation indicators, were analyzed after confirming no harmful multicollinearity (VIF < 3). Model performance was assessed using metrics such as accuracy, AUC, precision, recall, and F-measure. Results show that RF achieved the highest predictive accuracy (0.727) and AUC (0.772), with topographic wetness index (TWI) and normalized difference moisture index (NDMI) being the most influential variables. Conversely, DT tended to overestimate high-risk areas due to overfitting, while SVM and GBT provided more balanced classifications. The spatial classification outcomes revealed that model structure significantly impacts risk distribution, with ensemble models offering more reliable results. Although recall and sensitivity were high across models, specificity was generally low, particularly in GLM and DT, indicating difficulty in detecting non-risk areas. The study highlights the importance of selecting appropriate machine learning approaches and integrating diverse environmental indicators. Future research should address class imbalance and incorporate additional biophysical and socio-economic variables to enhance model robustness and policy relevance.