Decision tree-based modeling of soil erosion susceptibility in periglacial landforms of Eastern Anatolian Mountains, Türkiye

Periglacial soils developed under the influence of permafrost and periglacial processes provide a unique environment for investigating climate change and soil formation processes at the microtopographic scale. However, the susceptibility of these soils to erosion in high-altitude environments remains poorly understood. The study area encompasses four different mountainous regions in northeastern Anatolia, Türkiye (Mount Kısır-3197 m a.s.l., Mount Keldağ-3033 m a.s.l., Mount Ilgar-2918 m a.s.l., Mount Cin-2957 m a.s.l.). A total of 110 soil samples were collected and analyzed from different periglacial landforms (stony earth circle, non-sorted step, and mud circle) developed in these mountainous areas. A total of 110 soil samples were collected from different periglacial landforms (stony earth circle, non-sorted step and mud circle). In the present study, the predictability of aggregate stability (AS), dispersion ratio (DR), and the erodibility K factor of soils formed on periglacial landforms was evaluated using a the decision tree regression algorithm. The predictability of selected soil erosion factors was evaluated using the CHAID decision tree algorithm in a regression framework. According to the results, the coefficients of determination (R²) between observed and predicted values were 0.26, 0.49, and 0.65 for aggregate stability, dispersion ratio, and erodibility K factor, respectively, while the normalized Root Mean Square Error (nRMSE) values were found to be 19.81%, 32.64%, and 98.78%. In the study, in which nRMSE values were examined to eliminate the dependency of error metrics on data units, The model showed the most acceptable predictive performance for AS, although its R² value was relatively low (0.26), while the higher nRMSE values for DR and K indicated substantially greater prediction errors. Clay and electrical conductivity (EC) were identified as the most influential predictors of AS, while clay, silt, and microbial biomass carbon (MBC) were the key predictors of DR. For the soil erodibility K factor, organic carbon (OC), exchangeable cations, and CO₂ levels emerged as the dominant predictors. Overall, the findings indicate that AS should be selected as the primary erosion susceptibility parameter when applying decision tree models to soils developed on periglacial landforms. Moreover, the current study demonstrates that relying solely on R²as a measure of model performance may be misleading, and that MAE and nRMSE should be jointly considered to adequately evaluate both the magnitude and percentage of prediction errors. The dispersion ratio and erodibility K factor showed relatively limited predictive performance in the present dataset, indicating variability in their explanatory power across site conditions. The USLE-K calculation produced some negative values, potentially due to the high organic matter content of the studied soils, suggesting that the empirical equation may have limitations when applied to these periglacial soils. These results should be interpreted considering model uncertainty and data-driven limitations for the present dataset.