A dynamic soft-constrained deep learning paradigm for spatial downscaling of satellite gravimetry terrestrial water storage

• Efficiency of deep learning aided dynamic soft constrained paradigm for spatial downscaling. • Validation of higher resolution TWSA to multi-satellite derived products. • Proposed paradigm localizes high resolution TWSA signal both in spatial and spectral domains. • Downscaled TWSA preserved higher resolution adhering the balance of water cycle. The Gravity Recovery and Climate Experiment (GRACE) and GRACE Follow-On (GRACE-FO) satellite gravimetry missions have contributed significantly to our knowledge of variations in Earth’s Terrestrial Water Storage anomalies (TWSA) throughout the last two decades. However, the ability to quantifying hydrometeorological and other climate/weather episodes is hindered by limitations in the current TWSA spatiotemporal resolutions at monthly sampling and approximately coarser than 300 km. In this study, we used Deep Learning (DL) approach that is specifically developed for accurate and effective spatial downscaling of TWSA time series from NASA’s Jet Propulsion Laboratory (JPLM). Each TWSA maps of JPLM are downscaled from 300 km to 50 km spatial resolution spanning from April 2002 through December 2022 by using inherent spatiotemporal correlations of WaterGAP Hydrology Model (WGHM) TWSA. For this purpose, a novel dynamic soft-constrained loss function is introduced and applied that adaptively balances while optimizing the TWSA signal with low-resolution JPLM observations against high-resolution spatial patterns derived from the WGHM hydrological model and ERA5 inputs. Internal validation shows that while the downscaled TWSA preserves basin-averaged temporal dynamics (trends, seasonality) from JPLM, the correlations and spectral analyses show it successfully incorporates WGHM TWSA’s high-resolution spatial variability. External validation of downscaled TWSA products also demonstrates their ability to capture El Niño Southern Oscillation (ENSO)-driven interannual variability, glacial mass loss trends, spectral consistency with Soil Moisture Active Passive (SMAP) satellite-derived surface soil moisture at high-resolution band and similar predictive skill against to previous studies. Furthermore, the validation against to groundwater well observations indicates that the downscaled TWSA effectively represents the spatial patterns of long-term groundwater depletion in heavily stressed aquifers and significantly enhancing the spatial localization of depletion or recharging signals relative to the coarse-resolution JPLM TWSA