Artificial Neural Networks for Predicting the Water Retention Curve of Sicilian Agricultural Soils

Modeling soil-water regime and solute transport in the vadose zone is strategic for estimating agricultural productivity and optimizing irrigation water management. Direct measurements of soil hydraulic properties, i.e., the water retention curve and the hydraulic conductivity function, are often expensive and time-consuming, and represent a major obstacle to the application of simulation models. As a result, there is a great interest in developing pedotransfer functions (PTFs) that predict the soil hydraulic properties from more easily measured and/or routinely surveyed soil data, such as particle size distribution, bulk density (rho(b)), and soil organic carbon content (OC). In this study, application of PTFs was carried out for 359 Sicilian soils by implementing five different artificial neural networks (ANNs) to estimate the parameter of the van Genuchten (vG) model for water retention curves. The raw data used to train the ANNs were soil texture, rho(b), OC, and porosity. The ANNs were evaluated in their ability to predict both the vG parameters, on the basis of the normalized root-mean-square errors (NRMSE) and normalized mean absolute errors (NMAE), and the water retention data. The Akaike's information criterion (AIC) test was also used to assess the most efficient network. Results confirmed the high predictive performance of ANNs with four input parameters (clay, sand, and silt fractions, and OC) in simulating soil water retention data, with a prediction accuracy characterized by MAE = 0.026 and RMSE = 0.069. The AIC efficiency criterion indicated that the most efficient ANN model was trained with a relatively low number of input nodes.