Comparison of artificial neural network and regression pedotransfer functions for prediction of soil saturated hydraulic conductivity

Soil saturated hydraulic conductivity (Ks) is among the most important soil hydraulic-physical properties that required for soil-water modeling. Due to high cost and time- consuming nature of Ks measurement, estimating Ks from basic, inexpensive and easily measured physical and chemical soil properties is becoming increasingly important. In the last two decades, the development of estimation methods called pedotransfer functions that use cheap auxiliary variables has been a sharpening focus of soil research. This study was conducted (i) to develop different pedotransfer functions and (ii) to evaluate and compare statistical regression and neural network based pedotransfer functions for estimating Ks in a sub- catchment of Zayanderood River, located in Chaharmahal-va-Backtiari province. The data set was divided in to subsets for modeling (n=86) and validation (n=25). Root-mean-square error (RMSE), mean error (ME) and percentage of relative improvement (RI) were used as the validation indices. The artificial neural network-based models provided more reliable estimation than the statistical regression-based pedotransfer functions.