Comparative Analysis of Deep Learning Models for Reconstructing GRACE Terrestrial Water Storage Changes over India

Terrestrial water storage (TWS) anomalies from Gravity Recovery and Climate Experiment (GRACE)  plays a crucial role in identifying global groundwater usage patterns. However, the presence of 33 months of temporal gaps within the GRACE data poses challenges to spatiotemporal analysis. In this study, we explore the feasibility of employing two distinct deep learning models, namely CNN-LSTM and Multi-Layer Perceptron (MLP) based seasonal trend decomposition (STL), to reconstruct GRACE-TWS changes over India. To evaluate model performance, we utilize key metrics such as Pearson correlation (PR), Nash-Sutcliffe Efficiency (NSE), and Normalized Root Mean Squared Error (NRMSE). In a basin-wise comparison, the CNN-LSTM model outperforms the MLP-based STL method. Additionally, we tested the performance of deep learning models for various combinations of input variables, including precipitation, temperature, soil moisture, snow water equivalent, runoff, NDVI, evapotranspiration, and groundwater storage changes. Notably, the model with all input variables exhibited better performance during both training (Median PR = 0.982, Median NSE = 0.958, Median NRMSE = 0.051) and testing (Median PR = 0.956, Median NSE = 0.909, Median NRMSE = 0.086). Conversely, a combination with limited variables (Precipitation + Temperature + NOAH-TWS) showed relatively poor performance during training (Median PR = 0.933, Median NSE = 0.859, Median NRMSE = 0.106) and testing (Median PR = 0.907, Median NSE = 0.816, Median NRMSE = 0.136). Further, the inclusion of groundwater components as model input enhanced model performance, particularly in the arid regions of India.