Reconstructing GRACE-Terrestrial Water Storage Anomalies Using CNN-LSTM Neural Networks for Effective Drought Characterization in Major Indian River Basins

The temporal gaps in the GRACE and GRACE-FO satellite missions present significant challenges for analyzing Terrestrial Water Storage (TWS) anomalies. This study employs Convolutional-Long Short-Term Memory (CNN-LSTM) neural networks to bridge these gaps, enabling effective drought characterization in major Indian river basins. This approach integrates GRACE-TWS data with meteorological and climatic factors such as precipitation, temperature, runoff, evapotranspiration, and vegetation to generate reconstructed TWS anomalies. The reconstructed groundwater storage anomalies (GWSA) were validated using 4,431 in-situ observation wells, with Pearson’s correlation coefficient (PR) and Normalized Root Mean Square Error (NRMSE) as metrics. Non-perennial river basins like the Mahanadi, Godavari, and Krishna exhibited the best validation results (PR = 0.6–0.86; NRMSE = 0.1–0.2) compared to the perennial basins such as Ganga, Brahmaputra, and Indus demonstrated relatively weaker validation performance (PR = 0.2–0.7; NRMSE = 0.1–0.5). Further, the reconstructed GRACE-GWS helped in identifying new drought events alongside previously documented occurrences, corroborating findings from existing literature. Notably, new hydrological droughts were detected during the gap period between GRACE and GRACE-FO (2017–2018) in the Krishna, Cauvery, and Pennar basins, influencing drought characteristics such as severity and frequency. The gap period droughts are of D0 to D2 category based on United States Drought Monitor (USDM), and occurrences can be justified by decline of NE monsoon. Overall, the reconstructed TWS dataset provided continuous data, particularly for the gap periods, significantly enhancing the identification and analysis of drought occurrences.