Unveiling environmental drivers of soil erosion in South Korea through SHAP-informed machine learning

Soil erosion remains a critical environmental issue in South Korea due to its mountainous terrain and variable climatic conditions, adversely affecting agriculture, water quality, and infrastructure. While the Revised Universal Soil Loss Equation (RUSLE) is widely used for large-scale assessments, its limitations in scalability and capturing complex environmental interactions in diverse landscapes persist. Machine learning (ML) techniques offer a promising alternative by handling large datasets and modeling non-linear relationships, yet their ‘black box’ nature often hinders interpretability, which is essential for practical applications in soil conservation. This study addresses these challenges by applying six ML algorithms—Random forests (RF), Support vector machines (SVM), AdaBoost (AB), Logistic regression (LR), K-nearest neighbors (KNN), and Multi-layer perceptron (MLP)—to develop soil erosion susceptibility maps (SESM) across South Korea. Using the RUSLE model for preliminary identification and combined with high-resolution satellite images, we assigned erosion-prone areas to serve as the training data for our analysis. Thirteen key environmental factors were considered. To enhance the interpretability of the ML models, we integrated SHapley Additive exPlanations (SHAP), a method to attribute predictions to individual features, to enhance ML model interpretability. Our results indicate that the RF and SVM models achieved the highest predictive accuracy. SHAP analysis revealed that slope and land cover were the most influential in predicting SESM. This approach bridges advanced modeling techniques and practical applicability, enabling targeted soil conservation strategies in South Korea and offering a scalable framework for other erosion-prone regions.