Spaceborne Reconstruction of Groundwater Level Using Deep Learning in NW India

Analysis and modelling of groundwater dynamics is crucial in India, where groundwater serves as a major resource. In this study, two deep learning algorithms, ANN (Artificial Neural Network) and LSTM (Long ShortTerm Memory), are leveraged to model and predict groundwater levels (GWL) of North-West India (NWI) using spaceborne hydro-meteorological datasets. Five variables precipitation, temperature, evapo-transpiration (ET), soil moisture, and GRACE (Gravity Recovery and Climate Experiment) derived TWSA (Terrestrial Water Storage Anomalies) are utilized for modelling GWL and its dynamics. The predictions of both models are validated using in-situ data collected from wells across the study area. Results show that LSTM performs better in modelling and predicting GWL of NWI with the error ranging in between $\mathbf{- 1 6}$ meter and 17 meter and 50% values falling in between -4 to 4 meter, whereas ANN performs relatively poor with the error range of -400 meter to $\mathbf{7 0 0}$ meter and $\mathbf{5 0 \%}$ predictions having an error ranging from -60 to 60 meter. The annual trend from soil moisture, TWSA and in-stu GWL highlights severe groundwater depletion in NWI, specifically in Punjab, Delhi and western Uttar Pradesh (UP).