Author(s): Sul-Min Yun; Seong-Sim Yoon; Hyunjun Ahn; Il-Moon Chung
Linked Author(s):
Keywords: Groundwater recharge; Precipitation-based deep learning; LSTM; RNN; Time-series forecasting
Abstract: This study builds deep learning-based time-series models to predict monthly groundwater recharge using only monthly precipitation data and compares the performance of a recurrent neural network (RNN) and a long short-term memory network (LSTM). Groundwater recharge for 2001-2015 was obtained from observations and model outputs; 2001-2012 and 2013-2015 were used as the training and validation periods, respectively. Input features consisted of precipitation-based variables, including current-month precipitation, 1-6-month lagged precipitation, 3/6/12-month moving averages, and harmonic terms representing monthly seasonality. All inputs were normalized to the range 0-1. Both models were implemented in PyTorch, and mean squared error loss, the Adam optimizer, and early stopping were used to reduce overfitting. During training, the LSTM achieved a coefficient of determination (R²) of about 0.95 and a Kling-Gupta Efficiency (KGE) of about 0.97, while the RNN yielded R² of about 0.93 and KGE of about 0.89, indicating slightly lower but still high goodness of fit. For the independent validation period (2013-2015), both models reproduced the seasonal variability of monthly recharge and the peaks and low-recharge periods under wet and dry conditions in a stable manner. The validation results showed R² ≈ 0.84 and KGE ≈ 0.89 for the LSTM and R² ≈ 0.84 and KGE ≈ 0.92 for the RNN, indicating overall similar statistical performance, with the RNN performing slightly better in terms of hydrological consistency (KGE). These results demonstrate that monthly groundwater recharge can be effectively estimated from precipitation data alone when appropriate feature engineering and deep learning architectures are applied. In particular, precipitation-based LSTM and RNN models can serve as practical tools for long-term recharge prediction and for assessing climate-change and drought scenarios in regions with sparse hydrological and groundwater observations. Future work should integrate additional hydrological variables such as soil moisture and evapotranspiration, as well as spatial information, to further improve model generalization.
Year: 2026