Deep-GRASP: deep generative downscaling-assimilation-prediction for GRACE/GRACE-FO terrestrial water storage anomalies

The Gravity Recovery and Climate Experiment (GRACE) satellite mission and its successor, GRACE Follow-On (GRACE-FO), have revolutionized our understanding of global land hydrodynamics and terrestrial water storage anomalies (TWS). However, GRACE/GRACE-FO products have two main limitations: (i) their limited temporal coverage since approximately 2002, and (ii) their relatively low effective spatial and temporal resolutions (null and 1 month, respectively). To mitigate these limitations, we introduce Deep-GRASP, a deep generative machine learning methodology based on denoising diffusion probabilistic models with an architecture consisting of coupled U-Net, multilayer perceptrons, and convolutional LSTM. Deep-GRASP simultaneously performs three main tasks: (i) and (ii) the downscaling and assimilation of global GRACE/GRACE-FO TWS using hydrological models to reconstruct daily null TWS since 1980, and (iii) the prediction of TWS up to 1 year ahead. Focusing on the interval 2002-2026, we verify the downscaled, assimilated, and predicted TWS estimates against independent hydrometeorological observations, revealing an agreement of ~0.64 in the coefficient of determination (null) over the globe. In addition, by comparing with existing machine learning algorithms, we demonstrate that Deep-GRASP is approximately ~41% more accurate while being ~51% less computationally expensive. Hence, Deep-GRASP provides a novel and accurate machine learning framework for enhancing the GRACE/GRACE-FO products while mitigating their limitations. This is beneficial for studying the cryosphere, land hydrology, and climate dynamics.