Deep Learning-Based Forecasting of Groundwater Level Trends in India: Implications for Crop Production and Drinking Water Supply

Despite numerous studies in recent times, there is no consensus on the primary drivers for groundwater storage (GWS) changes over India. Thus, predicting future groundwater level trends seems remote. In this context, using Gravity Recovery and Climate Experiment (GRACE)-derived GWS, WaterGap model-based groundwater recharge (GWR), and groundwater withdrawal (GWW), we show that GWW exhibits a stronger dominance than GWR on GWS change over India. Furthermore, we developed feed-forward neural network (FNN), recurrent neural network (RNN), and deep learning-based long short-term memory network (LSTM) models using multidepth in situ observations from a dense network of monitoring wells (n = 5367, 1996-2018), to simulate and forecast groundwater levels (GWL) in India. The result demonstrates the better performance of LSTM (>84% of observation wells showing r > 0.6, RMSEn < 0.7) across India, outperforming both FNN and RNN in the testing period of 5 years (2014-2018). Our estimates also reveal that besides the prevailing long-term (1996-2018) statistically significant (p < 0.1) declining GWL trends in northwest India and the Ganges river basin, higher declining trends will potentially be observed in parts of north-central and south India in the forecasting period of 5 years (2019-2023). We envisage that the forecasting approach presented in the study can contribute toward an improved urban-rural drinking water supply and sustainable crop production for 1.3 billion people in India.

Automated summary

Despite the lack of consensus on the primary drivers for groundwater storage changes in India, this study utilized various models and data to analyze and predict future trends, demonstrating the superior performance of the LSTM model in forecasting groundwater levels.