Comparison of Stepwise Multilinear Regressions, Artificial Neural Network, and Genetic Algorithm-Based Neural Network for Prediction the Plant Available Water of Unsaturated Soils in a Semi-arid Region of Iran (Case Study: Chaharmahal Bakhtiari Province)

Abstract

ABSTRACT Plant available water (PAW) is one of the physical parameters of soils and the basic data of irrigation plans. Although various theoretical or empirical approaches have been proposed to describe this phenomenon, it is still possible to investigate and evaluate the relevance and applicability of new sciences such as artificial neural network method in predicting this phenomenon. In existing methods for determination of PAW, time-consuming tests are required. Nowadays, the capabilities of artificial neural network (ANN) methods in modeling have led to the use of ANN in parallel with the application of conventional approaches in various engineering sciences. In this study, artificial neural networks have been used as a new method to predict the PAW of soils. The study area is Khanimirza plain in Chaharmahal va Bakhtiari province. Soil sampling was performed randomly from 0 to 20 cm depth. The measured property in this study was the amount of plant available water (PAW). Readily available parameters including sand, silt and clay percentage, organic carbon, bulk density (BD), pH, Electrical conductivity (EC), calcium carbonate equivalent (CCE), and calcium carbonate are considered as model inputs. Modeling was performed using Stepwise multilinear regressions (SMLR), artificial neural network (ANN) and genetic algorithm-based neural network (ANN-GA). The results of PAW modeling showed that ANN-GA model with 0.90 coefficient is better than the other two methods. In general, ANN and ANN-GA showed better performance than SMLR. In fact, ANN and ANN-GA do not use a special type of equations and the network can achieve satisfactory results by establishing a proper relationship between input and output data.