The exchangeable sodium percentage (ESP) is the parameter that causes poorly developed soil structure and weak drainage over time, subsequently increasing the dispersion of soil particles, especially silt and clay, which cause soil erosion in Loess-derived soils. The number of 474 soil samples was collected within 0 to 30 cm of soil depth. Five non-linear models named Ridge regression (RidgeR), random forest (RF), support vector regression (SVR), cubist, and generalized boosted regression model (GBM), were considered to link several parameters with ESP in the region of approximately 2,730 ha. The mean square error (MSE), root mean square error (RMSE), and coefficient of determination (R2), were considered to assess the models. The RF and RidgeR algorithms were used to prioritize the variables impacting the ESP. The prediction model for mapping soil ESP had very high accuracy from RidgeR (RMSE = 0.039, R2 = 0.99), RF (RMSE = 0.257, R2 = 0.97), SVR (RMSE = 0.81, R2 = 0.544), cubist (RMSE = 0.006, R2 = 0.99), and GBM (RMSE = 0.426, R2 = 0.87). The sodium adsorption ratio (SAR) and slope steepness were the most important factors affecting ESP, and slope steepness was recognized as the most effective factor. We also identified that the applied models could simply assess ESP map. In addition, an action plan in different soil horizons, considering erosion and sediment processes, is efficient in controlling runoff at low ESP.

Keywords