Evaluation of Intelligence Models to Estimate the Least Limiting Water Range Using Conveniently Measurable Soil Properties

Abstract

Direct measurement of the least limiting water range (LLWR) is costly and time-consuming. In this study, genetic algorithm-based neural network (ANN-GA), artificial neural network (ANN) and stepwise multivariate regression (SMR) were used to estimate the LLWR of soil using easily measurable soil properties in the Khanmirza Plain. Then, depending on the location of each area, a total of 250 points were randomly identified as approximate sampling sites. Results showed that the accuracy of the SMR model with the percentage of clay, organic carbon and fine sand had a coefficient of determination of 0.42. The ANN-GA and ANN models with the highest coefficient of determination (R2 = 0.98) and mean square error (MAE = 0.0538) were suitable for estimating the least limiting water range. Therefore, the efficiency of models showed that the ANN and ANN-GA predicted the LLWR more accurate compared to the SMR and their results were close to the measured ones.