Wei et al. (2026) A constrained deep learning framework for spatiotemporal downscaling of GRACE-derived terrestrial water storage anomalies over China

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-10-12
• Authors: Min Wei, Hao Zhou, Yifeng Chen, Wenjing Ma, Siyou Xu, Zhicai Luo
• DOI: 10.1016/j.jhydrol.2026.136602

Research Groups

National Gravitation Laboratory, Institute of Geophysics, and School of Physics, Huazhong University of Science and Technology, Wuhan, 430074, China

Short Summary

This study developed a multi-scale constrained deep learning framework for spatiotemporal downscaling of GRACE/GRACE Follow-On (GFO) terrestrial water storage anomalies (TWSA) over China, integrating large-scale gravimetric constraints with fine-scale GLDAS hydrological variability to produce daily TWSA on a 0.25° grid. The framework significantly improved accuracy compared to GLDAS while preserving fine-scale spatiotemporal details, enabling better characterization of extreme hydrological events.

Objective

• To develop a multi-scale constrained deep learning framework for spatiotemporal downscaling of GRACE/GFO-derived terrestrial water storage anomalies (TWSA) over China, producing daily TWSA on a 0.25° grid by integrating GRACE/GFO gravimetric constraints with GLDAS fine-scale spatiotemporal variability.

Study Configuration

• Spatial Scale: China, 0.25° grid resolution.
• Temporal Scale: Daily.

Methodology and Data

• Models used: Multi-scale constrained deep learning model with a spatiotemporally constrained loss function (referred to as STDS framework).
• Data sources: GRACE/GRACE Follow-On (GFO) observations (for large-scale gravimetric constraints), Global Land Data Assimilation System (GLDAS) hydrological simulations (for fine-scale spatiotemporal variability).

Main Results

• The proposed STDS framework reduced the Root Mean Square Error (RMSE) relative to GRACE/GFO observations by 62.07% compared to GLDAS.
• It maintained a high temporal correlation (0.91) and spatial structural consistency (0.96) with GLDAS, indicating preservation of fine-scale spatiotemporal variability.
• The STDS framework successfully exhibited localized TWSA variations consistent with precipitation fields during extreme events like Typhoon Doksuri and the Zhengzhou flash flood.

Contributions

• Introduction of a novel multi-scale constrained deep learning framework for spatiotemporal downscaling of GRACE/GFO TWSA.
• Effective integration of large-scale gravimetric constraints from GRACE/GFO with fine-scale spatiotemporal variability from GLDAS, overcoming limitations of individual datasets.
• Generation of a high spatiotemporal resolution (daily, 0.25°) TWSA product for China, suitable for characterizing short-term hydrological variability and extreme events.

Funding

[No explicit funding information provided in the paper text.]

Citation

@article{Wei2026constrained,
  author = {Wei, Min and Zhou, Hao and Chen, Yifeng and Ma, Wenjing and Xu, Siyou and Luo, Zhicai},
  title = {A constrained deep learning framework for spatiotemporal downscaling of GRACE-derived terrestrial water storage anomalies over China},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136602},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136602}
}