Models performance in predicting least limiting water range in northwest of Iran under semiarid and semi-humid climates

Performance of artificial neural networks (ANNs), multi-objective group method of data handling (mGMDH) and multivariate linear regression (MLR) was compared for estimating least limiting water range (LLWR). Eleven soil attributes of 188 soil samples (Inceptisols) were used as independent variables to estimate LLWR directly (indicated as LLWR d ) and indirectly via moisture coefficients (LLWR i ) by ANNs, mGMDH and MLR methods. ANNs appeared as the most accurate and reliable tool for LLWR d and LLWR i prediction, and mGMDH and MLR ranked in descending order, respectively. For LLWR d , root-mean-square error (RMSE) values decreased from 0.040 to 0.024 (for testing (validation) step), when the method shifted from MLR to ANNs. Accuracy and reliability were both significantly improved from MLR to mGMDH and ANNs, but between the two later methods, they were only significant at the training step. However, for LLWR i , it was significant for testing step, too. For testing step, the r value between LLWR d and LLWR i with the experimental LLWR (LLWR e ) was 0.91 and 0.89, respectively, showing the priority of direct estimation of the LLWR. Soil bulk density, organic carbon, calcium carbonate equivalent, dithionate bicarbonate extractable aluminum, clay and sand, respectively, were better predictors for LLWR d. .