Uz et al. (2026) A dynamic soft-constrained deep learning paradigm for spatial downscaling of satellite gravimetry terrestrial water storage

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-01-24
• Authors: Metehan Uz, Kazım Gökhan Atman, Orhan Akyılmaz, C.K. Shum
• DOI: 10.1016/j.jhydrol.2026.135015

Research Groups

• GFZ Helmholtz Centre for Geosciences, Potsdam, Germany
• School of Mathematical Sciences, Queen Mary University of London, London, England, UK
• Department of Geomatics Engineering, Istanbul Technical University, Istanbul, Turkey
• Division of Geodetic Science, School of Earth Sciences, The Ohio State University, Columbus, OH, USA

Short Summary

This study develops a novel dynamic soft-constrained deep learning paradigm to spatially downscale GRACE/GRACE-FO terrestrial water storage anomalies (TWSA) from approximately 300 km to 50 km resolution. The method effectively preserves large-scale GRACE signals while incorporating high-resolution hydrological model patterns, significantly enhancing the spatial detail and localization of water storage changes, including groundwater depletion and glacier mass loss.

Objective

• To develop and apply a novel dynamic soft-constrained deep learning paradigm to spatially downscale GRACE/GRACE-FO terrestrial water storage anomalies (TWSA) from approximately 300 km to 50 km resolution, preserving large-scale observational signals while integrating high-resolution spatial patterns from hydrological models.

Study Configuration

• Spatial Scale: Downscaled from approximately 300 km (GRACE/GRACE-FO native resolution) to 50 km (0.5° x 0.5° grid).
• Temporal Scale: Monthly, from April 2002 through December 2022.

Methodology and Data

• Models used:
  • Deep Learning: Hybrid Variational U-Net architecture (combining U-Net and Variational Autoencoder (VAE)) with a novel dynamic soft-constrained loss function.
  • Hydrological Model: WaterGAP Hydrology Model (WGHM) version 2.2e.
  • Land Surface Model for Groundwater Storage Anomaly (GWSA) calculation: GLDAS NOAH model.
• Data sources:
  • Satellite Gravimetry: GRACE and GRACE Follow-On (GRACE-FO) Terrestrial Water Storage Anomalies (TWSA) from NASA’s Jet Propulsion Laboratory (JPLM) mascon solutions (RL06.1 v03).
  • Reanalysis Data: ERA5 climate reanalysis data (precipitation, evaporation, runoff, temperature, canopy water storage anomalies, snow water storage anomalies, soil moisture storage anomalies, cumulative water storage change).
  • Surface Soil Moisture: Soil Moisture Active Passive (SMAP) Level-4 (L4) data products (surface soil moisture anomalies).
  • Glacier Mass Change: Global Gravity-based Groundwater Product (G3P) glacier mass change time series.
  • In-situ Observations: Groundwater well observations (groundwater level anomaly time series) across the Contiguous United States (CONUS).
  • ENSO Index: Ni˜no 3.4 Sea Surface Temperature Anomaly (SSTA) index.

Main Results

• Successfully downscaled monthly JPLM TWSA from approximately 300 km to 50 km resolution from April 2002 to December 2022, with associated uncertainty estimates.
• Internal validation confirmed that the downscaled TWSA (DWSC) preserved basin-averaged temporal dynamics (trends, seasonality) from JPLM while effectively incorporating high-resolution spatial variability from WGHM.
• DWSC TWSA trends closely mirrored JPLM trends, capturing significant depletion in regions like the Middle East, India, and North America, and significantly outperforming WGHM in representing these trends.
• Spectral analysis confirmed DWSC retained the large-scale spectral power characteristic of JPLM TWSA (>600 km averaging radius) and effectively incorporated the high-frequency spatial variability of WGHM TWSA (<300 km averaging radius).
• DWSC captured El Ni˜no Southern Oscillation (ENSO)-driven interannual variability and extreme hydrological events (droughts, floods) with high fidelity to GRACE observations, outperforming WGHM in representing event magnitudes.
• DWSC realistically inherited glacier mass loss trends from JPLM, consistent with independent G3P data, despite WGHM lacking this information.
• DWSC showed significantly improved correlation and spectral consistency with high-resolution SMAP surface soil moisture compared to the original JPLM data.
• DWSC demonstrated superior predictive skill (higher Nash-Sutcliffe Efficiency) compared to alternative downscaled/assimilated products (GLWS, SeDA) relative to both JPLM and WGHM.
• Validation against in-situ groundwater well observations showed DWSC-derived groundwater storage anomalies (GWSA) effectively represented spatial patterns of long-term groundwater depletion, enhancing spatial localization compared to coarse-resolution JPLM TWSA.
• Predictive uncertainty for DWSC TWSA was quantified, yielding lower mean and higher standard deviation of uncertainties compared to JPLM, with more realistic spatial variations.

Contributions

• Introduces a novel dynamic soft-constrained deep learning paradigm using a Variational U-Net architecture for spatial downscaling of GRACE/GRACE-FO TWSA.
• Develops an innovative dynamic soft-constrained loss function that adaptively balances observational constraints from low-resolution GRACE data with high-resolution spatial patterns from hydrological models and reanalysis data.
• Generates a global, high-resolution (50 km), monthly TWSA dataset (DWSC TWSA) with associated uncertainty estimates, spanning April 2002 to December 2022.
• Demonstrates the ability to preserve large-scale GRACE signals (trends, seasonality, interannual variability, glacier mass loss) while effectively incorporating fine-scale spatial variability from hydrological models, addressing signal leakage and localization issues of coarse GRACE data.
• Provides robust validation against diverse internal and external datasets (ENSO, SMAP, independent glacier data, in-situ groundwater wells), showcasing enhanced spatial detail and robust performance across various hydro-climatic conditions and signal types.

Funding

No explicit funding projects, programs, or reference codes were provided in the paper. The acknowledgements section lists various data sources used in the study.

Citation

@article{Uz2026dynamic,
  author = {Uz, Metehan and Atman, Kazım Gökhan and Akyılmaz, Orhan and Shum, C.K.},
  title = {A dynamic soft-constrained deep learning paradigm for spatial downscaling of satellite gravimetry terrestrial water storage},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135015},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135015}
}