Predicting Groundwater Level Changes in Northern India from GRACE and GRACE Follow-On Data using Deep Learning Algorithms.

Predicting and monitoring groundwater levels and their usage is essential for the sustainable use of groundwater resources. In this study, two deep learning-based models, Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) were developed to predict groundwater level (GWL) changes by combining the Gravity Recovery and Climate Experiment (GRACE) and GRACE Follow-On (GRACE-FO) data with publicly available hydro-meteorological data sets (precipitation and temperature). To test the feasibility of the developed models, we have selected twelve in-situ wells across various parts of the Ganges basin, as it is one of the most significant basins in India and allegedly suffered heavy depletion of groundwater. The performance of the models was evaluated with the help of Pearson's correlation coefficient (PR), Normalized Root Mean Square Error (NRMSE), and Goodness of fit (R2 score). It is noticed that the CNN model show relatively higher PR (0.81 on training and 0.82 on testing) and lower NRMSE (0.13 on training and 0.16 on testing) values compared to the LSTM model. The CNN-based model also achieves high R2 score both on training and testing datasets than the LSTM model. All these observations suggest that GWL predicted from the CNN model better matches the in-situ GWL compared to the LSTM model. Further sensitivity analysis of the deep learning models suggests that GRACE ΔTWS played a significant role in estimating GWL prediction compared to precipitation and temperature.