Evaluation of Regression and Artificial Neural Network Models to Estimate the Saturated Hydraulic Conductivity in Mazandaran Province

Soil saturated hydraulic conductivity (Ks) is one of the important factors involved in water, soil, and agricultural sciences. Ks measurement is important for solute and water movement modeling and, in turn, is costly and time consuming. It is also impractical to spatially and temporarily measure the Ks in large scale studies. Therefore, it would be wise to predict Ks using indirect methods such as pedotransfer functions (PTFs). The objective of this study was to use the regression and artificial neural networks methods as an alternative method to estimate the saturated hydraulic conductivity. Therefore, 80 undisturbed soil samples in three replications were collected in Mazandaran province, northern Iran, and analyzed by laboratory methods. Data was divided into two categories including the training (80%) and testing dataset (20%). In order to predict the soil saturated hydraulic conductivity, the multiple linear regression models (MLR), multilayer perceptron (MLP) and radial basis function (RBF) methods were used. To test the performance of the three methods, the correlation (R2), mean square error (RMSE) and consistent correlation coefficient (CCC) statistics between actual and predicted values were measured. The results showed that MLP with two hidden layers by sigmoid activation function was the best method for Ks estimation. R2, RMSE and CCC statistics were 0.871, 1.02 cm/h [M1] and 0.869, respectively, for the best predicted method. The sensitivity analysis showed that the soil bulk density, pH and porosity had the highest impact on Ks, while soil salinity affected the Ks slightly. Therefore, use of MLP with two hidden layers efficiently can predict Ks in the study area and could be introduced as a promising method for Ks estimation. Considering the slightly low sampling data, this research can be considered as a starting step for future comprehensive studies with high intensive sampling sites that would enhance the reliability of these results.