A Neural Network Framework to Estimate Groundwater Levels in the Mississippi River Valley Shallow Alluvial Aquifer

Abstract

Abstract. Groundwater is an important source of irrigation water in the Mississippi Delta region where pumping rates have steadily increased over the past two decades. Groundwater withdrawals have exceeded aquifer recharge rates causing a decline in groundwater levels in the Mississippi River Valley Shallow Alluvial (MRVA) aquifer. In this study, we present the results of an artificial neural network (ANN) technique used to estimate groundwater levels of two wells in Sunflower County and Leflore County, Mississippi that are within the MRVA aquifer. The performance of two neural network learning algorithms, Levenberg-Marquardt and Bayesian Regularization, was examined in order to identify an optimal ANN architecture that can simulate the changes in daily groundwater level and provide acceptable predictions up to three months ahead. The two algorithms were trained using different hidden layer combinations and delays (5, 25, 50, 75, and 100) in order to establish the best network model. We discuss the modeling process and accuracy of these two methods for training and prediction stages based on mean square error (MSE). A recurrent neural network trained with the Bayesian Regularization algorithm, with 2 hidden layers and a delay of 100, had the lowest MSE (0.0005612) and produced the best forecast of daily groundwater levels for a lead time of up to three months. Groundwater level forecasting with adequate lead time will help resource managers identify areas where groundwater levels may reach a critical threshold and implement appropriate groundwater conservation policies and practices.