Liang et al. (2025) Disentangling and integrating spatiotemporal features: Deep learning-based downscaling of groundwater storage anomalies from GRACE and GRACE-FO satellites

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2025
• Date: 2025-11-27
• Authors: Qixiang Liang, Xingming Hao, Mengtao Ci, Mengqi Yuan, Yanfeng Di, Fan Sun, Chuan Wang, Jingjing Zhang, Xue Fan, Haibin Xiong
• DOI: 10.1016/j.ejrh.2025.102982

Research Groups

• State Key Laboratory of Desert and Oasis Ecology, Key Laboratory of Ecological Safety and Sustainable Development in Arid Lands, Xinjiang Institute of Ecology and Geography, Chinese Academy of Sciences, Urumqi, China
• University of Chinese Academy of Sciences, Beijing, China
• Akesu National Station of Observation and Research for Oasis Agro-ecosystem, Xinjiang, China
• Xinjiang Key Laboratory of Water Cycle and Utilization in Arid Zone, Xinjiang Institute of Ecology and Geography, Chinese Academy of Sciences, Urumqi, China

Short Summary

This study developed a deep learning downscaling framework to enhance GRACE/GRACE-FO derived Groundwater Storage Anomaly (GWSA) data from 0.5° to 0.1° resolution in Xinjiang, China, finding the Geographically and Temporally Weighted Neural Network Regression (GTNNWR) model most effective and revealing a significant groundwater depletion rate of 5.03 ± 9.42 mm/year from 2002 to 2023.

Objective

• To develop a multi-dimensional deep learning downscaling framework to enhance the spatial resolution of GRACE/GRACE-FO derived Groundwater Storage Anomaly (GWSA) data from 0.5° to 0.1° by decoupling and modeling spatiotemporal characteristics.
• To evaluate the performance of different downscaling models (Semi-supervised Variational Autoencoder Regression (SSVAER), Geographically Neural Network Weighted Regression (GNNWR), Geographically and Temporally Neural Network Weighted Regression (GTNNWR)) in terms of temporal continuity and spatial detail restoration.
• To investigate the spatiotemporal patterns and trends of groundwater storage change in Xinjiang, China, from 2002 to 2023 using the high-resolution GWSA data.

Study Configuration

• Spatial Scale: Xinjiang Uygur Autonomous Region, China (approximately 1.66 million km²). Downscaling from 0.5° to 0.1° spatial resolution.
• Temporal Scale: April 2002 to December 2023 (22 years), with monthly data.

Methodology and Data

• Models used:
  • Semi-supervised Variational Autoencoder Regression (SSVAER)
  • Geographically Neural Network Weighted Regression (GNNWR)
  • Geographically and Temporally Neural Network Weighted Regression (GTNNWR)
  • Three-cornered hat (TCH) method for uncertainty assessment.
  • LOESS-based seasonal trend decomposition method (STL) for time series analysis.
  • SHapley Additive exPlanations (SHAP) method for model interpretability.
• Data sources:
  • Satellite/Gravity Data: GRACE and GRACE-FO (Center for Space Research (CSR), Jet Propulsion Laboratory (JPL), Goddard Space Flight Center (GSFC) RL06 products, and reconstructed GRACE TWSA dataset).
  • Land Surface Model Data (GLDAS Noah L4 v2.1): Soil Moisture Anomaly (SMA), Snow Water Equivalent Anomaly (SWEA), Plant Canopy Water Anomaly (PCWA) (0.25° monthly, resampled to 0.5°).
  • Reanalysis Data (ERA5-Land): Evapotranspiration (ET), Precipitation (Pre), Surface Runoff (Runoff), Snowmelt (SMLT) (0.1° monthly).
  • MODIS Terra/Aqua Satellite Data: Normalized Difference Vegetation Index (NDVI), Land Surface Temperature (LST) (0.05° monthly).
  • Land Use/Cover Change (CLCD): 30 m annual.
  • Soil Type (HWSD): 1 km.
  • Digital Elevation Model (DEM): NASA DEM (30 m).
  • In-situ Observations: Groundwater level observations from 175 monitoring wells in Xinjiang (2005-2021, monthly average anomalies).

Main Results

• The combined CSR-JPL product exhibited the lowest uncertainty (13.80 mm absolute, 0.35 relative) for GRACE Terrestrial Water Storage Anomaly (TWSA) in Xinjiang.
• Xinjiang's TWSA showed a significant decreasing trend of 4.57 mm/year from 2002 to 2023.
• The GTNNWR model outperformed SSVAER and GNNWR for downscaling GWSA, increasing the correlation coefficient (CC) with in-situ well observations from 0.47 (original GRACE) to 0.57.
• From 2002 to 2023, Xinjiang's groundwater declined at a significant rate of 5.03 ± 9.42 mm/year.
• Groundwater depletion hotspots were identified in the Weigan River Basin (19.33–23.95 mm/year decline) and Ili River Basin (18.55–23.03 mm/year decline).
• Long-term GWSA trends were primarily driven by precipitation (CC = 0.47), while seasonal variability was strongly influenced by evapotranspiration (CC = 0.53).
• SHAP analysis indicated that DEM, ET, Soil type, and Precipitation were the most important factors for downscaled GWSA, while Land Use/Cover Change (LUCC) had the least importance, suggesting a limitation in capturing human-induced groundwater extraction.

Contributions

• Developed and validated a multi-dimensional deep learning downscaling framework (integrating SSVAER, GNNWR, and GTNNWR) for GRACE/GRACE-FO GWSA, effectively decoupling and fusing spatiotemporal features.
• Successfully enhanced the spatial resolution of GRACE-derived GWSA from 0.5° to 0.1°, providing higher precision for local-scale groundwater monitoring in arid regions.
• Demonstrated that the GTNNWR model significantly improved the correlation with in-situ groundwater well observations compared to other models and original GRACE data.
• Provided updated and high-resolution spatiotemporal patterns and trends of groundwater storage change in Xinjiang from 2002 to 2023, identifying key depletion hotspots.
• Disentangled the driving factors of groundwater changes across different timescales, showing precipitation's influence on long-term trends and evapotranspiration's impact on seasonal variability.
• Highlighted the critical need for high-resolution quantitative data on human activities (e.g., groundwater extraction) to further improve downscaling models in heavily managed arid regions.

Funding

• Strategy Priority Research Program (Category B) of Chinese Academy of Sciences [grant number XDB0720101]

Citation

@article{Liang2025Disentangling,
  author = {Liang, Qixiang and Hao, Xingming and Ci, Mengtao and Yuan, Mengqi and Di, Yanfeng and Sun, Fan and Wang, Chuan and Zhang, Jingjing and Fan, Xue and Xiong, Haibin},
  title = {Disentangling and integrating spatiotemporal features: Deep learning-based downscaling of groundwater storage anomalies from GRACE and GRACE-FO satellites},
  journal = {Journal of Hydrology Regional Studies},
  year = {2025},
  doi = {10.1016/j.ejrh.2025.102982},
  url = {https://doi.org/10.1016/j.ejrh.2025.102982}
}