Improving the accuracy of Kumamoto's groundwater level modeling using deep learning LSTM

This study tried to predict the groundwater level in Kumamoto City, Kumamoto Prefecture, Japan using deep learning. In Kumamoto prefecture, 80% of domestic water is obtained from groundwater. A large amount of groundwater is also used for irrigation and industrial water. Therefore, it is important to model the groundwater level with high accuracy. For the purpose, this study employed a deep learning method, Long Short-Term Memory (LSTM). LSTM is a kind of Recurrent Neural Network (RNN), which is suitable for making time series modeling among deep learning methods. As input, meteorological data such as daily mean temperature and precipitation were utilized. The daily average groundwater level at a gauging station in Kumamoto City was utilized as the target data. LSTM was trained with several combinations of hyperparameters. The results indicate that LSTM is enough capable to model the groundwater level in Kumamoto City with high accuracy, and then it would be a useful tool to solve various related issues such as climate change impacts assessments.