Predicting soil water retention curve by artificial neural networks

Abstract

Measurement and determination of the hydraulic properties of soil with a direct method is time consuming and also costly. Today, experimental errors can be eliminated by using computerized methods of data processing. In this paper, an artificial neural network is described as an approach to predict the soil water retention curve from the readily available sets of soil properties. To predict the soil water retention curve, two basic neural network structures were considered. The first one is used to predict volumetric soil water content (cm3cm−3) according to the soil properties and soil matric head (SMH MPa) (water content predictor structures). The other structure is used to predict the SMH according to the soil properties and the volumetric soil water content (cm3cm−3) (matric head predictor structures). Two levels of input variables (five and six sets of inputs, i.e. sand, clay, silt, bulk density, SMH, and volumetric water content [cm3cm−3]) in combination with the usage of a sigmoid transfer function (log and tangent) with two training algorithms (Levenberg-Marquardt [trainlm] and Bayesian Regularization [trainbr] training algorithm) in hidden layers of networks were used. In the water target structures, another method for improving generalization called the early stopping method was used. In this technique the available data is divided into three subsets: Training set, validation set and test set. Comparison of results showed that applying five sets of inputs (sand, clay, bulk density and SMH, and volumetric water content [cm3cm−3]), and applying log sigmoid transfer function leads to better solutions. Furthermore, the Bayesian Regularization algorithm improved generalization and reduced the error of estimation. However, the early stopping method did not provide a suitable generalization performance. The best water content predictor network with five inputs is a two hidden layer network with eight and 14 nodes, respectively. Therefore, its architecture is 5-8-14-1. The best soil matric head predictor network is a two hidden layer network with five inputs and 5 and 6 nodes, respectively. Therefore, its architecture is 5-5-6-1. However, the water content predictor network was superior to the matric head predictor network. Acceptable results were obtained for prediction of the water retention by using artificial neural networks. Therefore, it is advisable to use such intelligence models to predict the soil hydraulic properties.