Precise monitoring of global water cycle variations is crucial for comprehending Earth's climate system. Analyzing long-term trends in ice-sheet melting and freshwater availability reveals valuable insights, while short-term variations aid in monitoring natural hazards like floods and droughts. Total water storage (TWS), encompassing all water forms, serves as an essential climate variable for quantifying these variations. Traditionally, TWS has been modeled using global hydrological models, which provide spatial variance and short-term temporal variations but struggle to accurately represent long-term trends reflecting climate and human-induced changes. Since 2002, GRACE and GRACE-FO missions have offered a unique perspective by measuring gravity field variations to monitor global TWS anomalies (TWSAs) with unprecedented accuracy and global coverage. However, the coarse spatial resolution (approximately 3°) of GRACE products limits their application, particularly in understanding water storage changes in smaller catchments. This limitation stems from inherent design constraints of the satellite orbits and instruments, as well as from post-processing techniques necessary for signal extraction which often attenuate high-frequency signals. To address this, integrating higher-resolution information from hydrological models and measurements is crucial. Hydrological models directly simulate TWSAs, while parameters like precipitation provide valuable auxiliary information through water balance considerations. While data assimilation techniques have shown promise in downscaling GRACE products regionally, their global application remains challenging. This study tackles this challenge by developing a novel self-supervised deep learning approach capable of producing globally generalizable high-resolution TWSAs.

Previous research has explored the use of deep learning and machine learning for downscaling GRACE measurements. However, these studies often rely on supervised learning approaches, necessitating high-resolution ground truth data which are unavailable. To circumvent this, researchers have generated synthetic training data by downsampling high-resolution hydrological simulations or creating GRACE-like TWSAs through smoothing. These approaches rely on assumptions that the relationships between predictors and targets remain consistent across different resolution domains, which might not always hold true. Furthermore, many existing deep learning methods lack global generalizability, having been successfully applied only to continental or regional scales. This paper directly addresses these limitations by developing a self-supervised approach that eliminates the need for synthetic training data and achieves global-scale applicability.

The proposed model employs a self-supervised data assimilation framework to generate high-resolution TWSAs. The input features include GRACE mascon solutions (from NASA JPL), WaterGAP Hydrology Model (WGHM) simulations (at 0.5° resolution), and hydrological data from the Global Land Data Assimilation System (GLDAS) – specifically precipitation, evapotranspiration, and runoff. Latitude and longitude are also included as features. Data preprocessing involves normalization of features and splitting the global area into 16° × 16° patches (32 × 32 pixels at 0.5° resolution). The core of the model is a convolutional neural network (CNN) with an encoder-decoder architecture, incorporating residual blocks and batch normalization. The encoder downsamples the input features, extracting high-level features, while the decoder upsamples to reconstruct the high-resolution TWSAs. A novel loss function guides the self-supervised learning process. This loss function balances two objectives: minimizing the average absolute error (AE) between patch-averaged predicted TWSAs and GRACE TWSAs, and maximizing the Pearson correlation (R) between predicted and WGHM TWSAs while simultaneously minimizing the mean absolute error (MAE) between them. The final loss function combines these terms: L(Pc, Pw, P̂) = (1/B) Σ [AEc(Pc, P̂) + (1 - R(Pw, P̂)) × MAEw(Pw, P̂)], where Pc represents GRACE patches, Pw represents WGHM patches, P̂ represents predicted patches, and B is the batch size. Uncertainty estimation is accomplished using deep ensembles (five independent models trained from different random initial states) and Monte Carlo simulations (20 random samples of GRACE inputs per model), resulting in an ensemble mean and standard deviation for each pixel. The water balance equation closure is evaluated by comparing water storage changes derived from the downscaled TWSAs with those calculated from ERA5-Land water budget components (precipitation, evapotranspiration, runoff), using the Nash-Sutcliffe efficiency (NSE) as a metric. Finally, two widely used indices—the flooding potential index (FPI) and the drought severity index (DSI)—are derived from the downscaled TWSAs to demonstrate their applicability for environmental monitoring.

The study generated a global high-resolution (0.5°) TWSA product from April 2002 to December 2019, excluding Greenland and Antarctica. The product exhibits global-scale seasonal changes consistent with GRACE measurements. Regional maps reveal high-resolution details, inheriting refined structures from WGHM simulations while showing smoother results compared to WGHM alone. The global median uncertainty is 7.3 mm. Pixel-wise Pearson correlation between the downscaled TWSAs and WGHM simulations shows a median value of 0.80, a 51% improvement over GRACE TWSAs. Basin-wise RMSE comparisons with GRACE measurements yielded a global average of 21.9 mm, lower than the typical GRACE uncertainties (20–30 mm) and representing a 56% improvement over WGHM simulations. Time series analysis of six major river basins demonstrated better agreement between downscaled TWSAs and GRACE observations than WGHM simulations, particularly for basins with less stationary average TWSAs. The proposed method significantly improved the correlation between WGHM- and GRACE-derived long-term trends (from 0.47 to 0.94), annual amplitudes (from 0.83 to 0.97), and semi-annual amplitudes (from 0.83 to 0.95). The downscaled TWSAs also showed better ability to close the water balance equation in basins smaller than the GRACE-effective resolution, with a global median NSE of 83%. The improvement compared to GRACE was strongly correlated with basin size, while the improvement over WGHM was likely due to data assimilation. Analysis of long-term trends revealed high-resolution details not visible in GRACE data, such as groundwater depletion hotspots in the United States. Derived flood potential index (FPI) and drought severity index (DSI) maps from downscaled TWSAs showed better agreement with GRACE-derived indices than those from WGHM, indicating improved suitability for high-resolution environmental monitoring.

The high-resolution TWSAs generated by this study offer valuable insights for local-scale analyses of climate and anthropogenic impacts, enabling targeted strategies for sustainable water resource management. The findings address the research question by demonstrating the effectiveness of a self-supervised data assimilation model in improving the spatial resolution of global TWSA products without relying on unavailable high-resolution ground truth data. The significance lies in providing a globally applicable method for generating high-resolution TWSAs, enhancing the accuracy of water balance estimations, and improving the capabilities of environmental monitoring indices like FPI and DSI for more localized hazard assessments. The improved accuracy and spatial resolution of the TWSAs represent a substantial advancement over existing methods, enabling more detailed analysis of water cycle dynamics and their response to various factors.

This study successfully demonstrated a novel self-supervised data assimilation model using deep learning to generate high-resolution global TWSAs. The method effectively integrates GRACE(-FO) measurements and hydrological model simulations, leading to improved accuracy and spatial detail while maintaining global consistency. The resulting product enhances water balance closure, improves the representation of temporal signals, and provides valuable high-resolution information for environmental monitoring and hazard assessment. Future research could focus on improving glacier and human intervention modeling within the deep learning framework, exploring inter-water-form interactions, and incorporating additional data sources for further refinement.

The study acknowledges several limitations. The model's performance in glaciated regions is limited by the accuracy of the hydrological model in these areas. Human intervention effects are also not perfectly captured, requiring potential improvements by incorporating population density, land use, and water usage data. The uncertainties reported might be underestimated because uncertainties of the WGHM simulations and GLDAS inputs were unavailable. The reliance on the quality of the input hydrological simulations is another limiting factor, as imperfections in those simulations can affect the accuracy of the downscaled product.