Development of Long Short-Term Memory Model for Prediction of Water Table Depth in United Arab Emirates

Abstract

Water table depth is declining in most parts of the world, especially in those countries which have high temperatures almost throughout the year and receive very less precipitation throughout the year. Due to increasing population and intensive agricultural and industrial practices, the demand of freshwater is increasing and is predicted to increase in upcoming years. The countries which receive less rainfall throughout the year have limited groundwater recharge, resulting in declining of water table. United Arab Emirates belong to this category of countries where there is high temperature almost throughout the year and receives very less rainfall (less than 200 mm annually). Modeling groundwater in such an arid climate is of serious concern. This paper proposes LSTM models for prediction of water table depth at six different wells in different parts of United Arab Emirates. Data obtained for this study comprises of times series monthly water table depth data in meters from ground level from six different wells. Analysis of the data showed the drastic decline of water table depth between 1977 and 2011. These data were used to generate the input and target variables by adding three time-step lags in the given data. The time-step lag data was used as input to predict the current water table depth. In other words, the water table depth data of current target month was predicted using the previous three months water table depth data as input. Training of LSTM models was carried forward utilizing Python platform as a programming language library (TensorFlow). The trained models provided good accuracy in testing dataset. The training R2 values of all the six models were more than 0.96 and the testing R2 values of all the six models were more than 0.91.