estimation of saturated hydraulic conductivity in some soils of Ilam province using artificial neural networks and regression methods

Saturated hydraulic conductivity (Ks) is one of the essential input for water flow and solute transport modelling, irrigation and drainage design, groundwater modeling and environmental processes. Direct measurement of Ks is possible, but that is usually time consuming, tedious, expensive and impractical for larger scale studies. Also, these methods are partly unreliable because of soil heterogeneity and experimental errors. One solution to govern this problem is using indirect methods such as pedo-transfer functions (PTFs). Since PTFs have not yet been developed to soils in the study area, this study evaluates and describes neural network and statistical regression PTFs to predict Ks from limited or more extended sets of the readily available soil properties. For this purpose, Ks from 95 points of Sirwan-Chardawel sub-basins in Ilam province were measured using Guelph permeameter. Also, some of the readily available soil parameters were obtained. The accuracy and reliability of the derived PTFs were evaluated using root mean square error (RMSE), mean error (ME) and Pearson correlation coefficient (r). The highest correlation coefficients of 0.58 and 0.56 were found between Ks and geometric mean particle diameter and sand content, respectively. The results indicated that artificial neural network and regression PTFs can predict Ks with relatively good accuracy even if a few readily available soil properties are measured (rR-val= 0.85, RMSER-val= 6.81 mm/hr and rANN-test= 0.87, RMSEANN-test= 10.80 mm/hr). However, based upon results, the prediction accuracy of ANN model at both training and testing stages increased if more readily available soil properties are used (rtrain= 0.92, RMSEtrain= 4.36 mm/hr and rtest= 0.89, RMSEtest= 7.17 mm/hr). In general, it was found that ANNs method had better performance than linear regression model in predicting Ks.