Abstract
Direct measurement of unsaturated hydraulic parameters that are crucial inputs in any modeling of water flow or solute transport through the vadose zone is costly and time-consuming. Therefore, indirect methods like artificial neural networks (ANNs) can be used to estimate these parameters. Different ANNs conditions [two training algorithms (Trainlm and Traingd), two transfer functions (Tansig and Logsig), and different combinations of the input variables sand, silt, clay, bulk density (BD), soil organic matter (SOM) and initial (⊖i) and saturated (⊖s) volumetric water content] were used to predict the unsaturated hydraulic conductivity, K ψ [at six applied tensions (ψ) of 0–0.2 m] and sorptive number, α (at five ψ values of 0.03–0.2 m) that correspond to 138 soil samples from two neighboring study areas located in the Agricultural College, Shiraz University, Islamic Republic of Iran. A four-layer ANNs with three and four nodes in the hidden layers, performed the best in predicting K ψ and α. Traingd produced the best predictions over all input variables when Tansig and Logsig transfer functions were used for K ψ and α at different values of ψ, respectively. ‘Silt + clay + sand + BD + SOM’ was the most basic influential input variable for predictions of K ψ and α at almost all values of ψ. The accuracy of ANNs-predicted K ψ decreased with decreasing ψ from 0.2 to 0.1 m, followed by an increase to a ψ value of 0 m; a specific relationship was not observed for α. Therefore, ANNs can be used to predict K ψ with greater confidence at moderate values of ψ than at lower or higher values. The normalized root mean square error, NRMSE, confirmed that ANNs predictions for K ψ were more accurate than predictions for α. Because reliable predictions were obtained for K ψ, and to a lesser extent for α, it is recommended that such intelligence models are used to predict these vital soil hydraulic attributes.