Assessing machine learning algorithms for prediction of soil infiltration behavior in semi-arid agricultural soils

Accurate estimation of soil infiltration characteristics is fundamental for irrigation design, hydrological modeling, and sustainable soil and water management. However, the nonlinear interactions among soil physical, hydraulic, and chemical properties make reliable prediction of cumulative infiltration (It) and basic infiltration rate (Ib) challenging using conventional statistical approaches. This study evaluated the predictive performance of artificial neural networks (ANNs), random forest (RF), and multiple linear regression (MLR) for estimating It and Ib under three input scenarios of soil properties comprising (i) physical properties, (ii) physical and hydraulic properties, and (iii) physical, hydraulic, and chemical properties. Soil samples were collected from the Bajgah Plain, Fars Province, Iran, and infiltration characteristics were measured using a double-ring infiltrometer. The machine-learning models were trained and tested, and their performance was assessed using different statistical indicators. Among the evaluated models, the double-hidden-layer ANN achieved the highest predictive accuracy (NRMSE = 0.03–0.06, and R 2 = 0.98–0.99) for It using first and second scenarios and for Ib (NRMSE = 0.04–0.06, and R 2 = 0.99) using the first and third scenarios. The RF model also provided highly accurate predictions (NRMSE = 0.045–0.047, and R 2 = 0.96–0.97 for It, and NRMSE = 0.008–0.009, and R 2 = 0.999 for Ib). In contrast, the MLR exhibited the lowest predictive capability (NRMSE = 0.58–0.60, and R 2 = 0.63-066 for It, and NRMSE = 0.77–0.87, and R 2 = 0.27–0.42 for Ib) because of its inability to represent the nonlinear relationships governing infiltration processes. Incorporating saturated hydraulic conductivity improved the prediction of both cumulative infiltration and basic infiltration rate. Including soil chemical properties provided only marginal additional improvements, indicating that infiltration in the studied soils was primarily controlled by physical and hydraulic characteristics. The results demonstrate the superiority of random forest models for simulating soil infiltration characteristics and highlight the importance of selecting appropriate predictor variables according to the target infiltration parameter. The proposed models provide efficient tools for estimating infiltration where direct field measurements are limited and offer valuable support for irrigation planning, hydrological simulations, and sustainable soil and water management.