Modelling the root zone soil moisture using artificial neural networks, a case study

Surface soil moisture constitutes a major component in the Earth's water cycle. In many cases, modelling and predicting soil moisture represent a serious problem in water resources field due to the problematic measurements or lack of measurements, etc. Data-driven models such as artificial neural networks (ANN) have been characterized as a robust tool to overcome these shortcomings. This study aims to identify the optimum ANNs to model the root zone soil moisture (up to 2 m depth) in the upper reach of the Spree River using the synthetic soil moisture data from SWAT model. Thus, three different approaches were developed and compared to determine the highest performing method. These networks can be broadly categorized into dynamic, static, and statistical neural networks, which are layer recurrent network (LRN), feedforward (FF), and radial basis networks, respectively. Data sets of precipitation and antecedent soil moisture were selected based on quantification of cross-, auto-, and partial auto-correlation coefficients to represent the best behaviour of root soil moisture. The time series data were subdivided into two subsets: one for network training and the second for network testing. The determination coefficient (R-2), root-mean-square error, and Nash-Sutcliffe efficiency were employed to test the goodness of fit between the actual and modelled data. Results show that, among the used methods, the LRN and FF networks have the top performance criteria, showing a reliable ability to be used as estimator for the soil moisture in this catchment.