Forecasting Groundwater Level in Florida using Advanced Machine Learning Approaches

Mismanagement of groundwater resources leads to groundwater depletion and other environmental issues such as quality reduction and subsidence. Water table prediction is essential for optimal management of groundwater resources. Hence, artificial intelligence (AI) methods have been used widely to predict water tables in recent years. This paper adopted some new methods of machine learning, e.g., categorical boosting (CATBoost), extreme gradient boosting (XGBoost), and Convolutional neural network-Long Short-Term Memory (CNN-LSTM), to predict water table. The key input parameters were evaporation/transpiration, rainfall, temperature, and water table in the prior month. To better compare the models, simulations were executed in daily and monthly periods. DeLuca Preserve located in Florida was selected to test the proposed algorithms.  The results indicated that in general machine learning algorithms are appropriate approaches to predict water tables. CNN-LSTM algorithm with RMSE = 0.22 m and R2 = 0.96 showed better performance in predicting daily groundwater levels.  However, monthly water tables were predicted much better using CATBoost algorithm with RMSE = 0.11 m and R2 = 0.99.