Machine-Learning Pedotransfer Functions for Modelling Saturated Hydraulic Conductivity in a Semi-Arid Region

Accurate values of saturated hydraulic conductivity (Ks) underpin irrigation scheduling, hydrological modelling, and the broader soil-water-plant balance, yet measuring Ks directly in the field is slow, labour-intensive, and rarely practical at large spatial scales. This study therefore examined whether pedotransfer functions (PTFs) built with machine-learning (ML) algorithms can substitute for direct measurement in the alluvial soils of Sohag Governorate, Egypt. A total of 206 surface samples were characterized for particle size distribution, soil organic matter (SOM), bulk density (Bd), pH, electrical conductivity (EC), and Ks, with sand, silt, clay, SOM and Db used as predictors, while pH and EC were measured for soil characterization and were not included as model predictors. After a random calibration-validation split, twenty regression and ML techniques were fitted and compared in R-Studio. Linear formulations captured only a moderate share of the variability, reaching a validation R2 of 0.56, while most nonlinear algorithms performed distinctly better. The radial-basis-function support vector machine (SVM-RBF) gave the strongest overall validation result (R2 = 0.827, RMSE = 12.505 mm h−1, MAE = 6.954 mm h−1, RPD = 2.445). The stacked ensemble reached the highest calibration R2 (0.943), but its validation R2 fell to 0.696, a gap that points to some degree of overfitting. These results indicate that kernel-based and ensemble PTFs offer a practical, time-saving route to estimating Ks and can support irrigation and water-management planning across semi-arid alluvial landscapes without the burden of exhaustive field measurement.