Predicting Forest Trail Degradation Susceptibility Using GIS-Based Explainable Machine Learning

Abstract

Effective trail management is essential for preventing environmental degradation and promoting sustainable recreational use. This study proposes a GIS-based, explainable machine learning framework for predicting forest trail degradation using exclusively environmental variables, eliminating the need for costly visitor monitoring data that remains unavailable in most operational forest settings. Field surveys conducted in Geumjeongsan, South Korea, classified trail segments as degraded or non-degraded based on physical indicators such as erosion depth, trail width, and soil hardness. Environmental predictors-including elevation, slope, trail slope alignment (TSA), topographic wetness index (TWI), vegetation type, and soil texture-were derived from spatial analysis. Three machine learning algorithms (Binary Logistic Regression, Random Forest, and Gradient Boosting) were systematically compared using confusion matrix metrics and AUC-ROC (Area Under the Receiver Operating Characteristic Curve). Random Forest (RF) was selected for its strong performance (AUC-ROC = 0.812) and seamless integration with SHAP (SHapley Additive exPlanations) for transparent interpretation. Spatial block cross-validation achieved an AUC-ROC of 0.729, confirming robust spatial generalization. SHAP analysis revealed vegetation type as the most significant predictor, with hardwood forests showing higher degradation susceptibility than mixed forests. A susceptibility map generated from the RF model indicated that 40.7% of the study area faces high to very high degradation risk. This environmental-only approach enables proactive trail management across data-limited forest systems globally, providing actionable insights for sustainable trail maintenance without requiring visitor use data.