Pedotransfer functions (PTFs) may turn to be an alternative to the direct measurement. There are, however, contradictory information about the accuracy and reliability of developed PTFs for soil hydraulic properties using various methods including artificial neural networks (ANNs), multi-objective group method of data handling (mGMDH) and multivariate linear regression (MLR). Comparing these methods is main purpose of the current study. Laboratory measurements in 188 undisturbed soil samples with wide range of properties were used to compute four moisture coefficients (θwp,θ fc,θ sr,θ afp) from which experimental LLWR (LLWRe) was calculated. Eleven various soil attributes were also measured in disturbed samples and employed as independent variables to predict the same moisture coefficients and LLWR (designated as LLWRi) by ANNs, mGMDH and MLR methods. LLWR was also predicted directly (indicated as LLWRd) from the soil attributes. Accuracy and reliability of the developed PTFs to predict LLWRd and LLWRi, as compared to the LLWRe, was evaluated using root mean square error, Akaike information criterion and relative improvement. ANNs appeared as the most accurate and reliable tool for LLWRd and LLWRi prediction; mGMDH and MLR ranked in descending order. Significant differences in the prediction accuracy or reliability between the developed PTFs were evaluated using AIC. Results showed that both were significantly improved from MLR to mGMDH and ANNs, but between mGMDH and ANNs they were only significant at the training step. For LLWRi it was significant for validation step, too. More over LLWRd was better correlated to LLWRe (as a reference) than LLWRi implying that predicting least limiting water range directly from the soil attributes led to more accurate and reliable prediction than from the moisture coefficients as obtained from the developed PTFs
