Basin-wide groundwater level forecasting in Kumamoto, Japan: integrating transfer learning with Long Short-term Memory Network

Deep learning (DL) approaches have demonstrated the potential for accurate groundwater level (GWL) forecasting. However, the need to train the DL model independently for each station poses a significant challenge for large-scale applications. This study explored the potential of a transfer learning (TL) approach coupled with Long Short-Term Memory (LSTM) (TL-LSTM) networks for basin-wide GWL forecasting. The performance of TL-LSTM was compared against independently trained LSTM (IT-LSTM) models, which showed superior accuracy over several machine learning models for Kumamoto, Japan. The basin average values of performance indicators such as the coefficient of determination (R²) and Nash-Sutcliffe Efficiency (NSE) for IT-LSTM models were 0.93, 0.86, and 0.78 (R²); 0.91, 0.86, and 0.77 (NSE) for one-, two-, and three-month-ahead forecasting, respectively. Similar performances were achieved using the new model with significantly less training effort and computational time, making it suitable for large-scale GWL forecasting.