Ibrahim et al. (2026) Securing the Silent Reserve: Physics-Informed Deep Learning for Global Groundwater Storage Downscaling

Identification

• Journal: Iconic Research and Engineering Journals
• Year: 2026
• Date: 2026-03-06
• Authors: Kunle Adefarati Ibrahim, Ologun Sodiq Babatunde, Chiagoziem C. Ukwuoma, Richard Joshua Akeredolu, Victoria Chioma Ayozie-Samuel; Maryam Saleem
• DOI: 10.64388/irev9i9-1714801

Research Groups

• College of Environment and Civil Engineering, Chengdu University of Technology, Chengdu, Sichuan, China
• University of Electronic Science and Technology of China, Chengdu, China
• Chengdu University of Technology Oxford Brookes College, Chengdu, Sichuan, China
• College of Ecology and Environmental Science, Chengdu University of Technology, Chengdu, Sichuan, China
• College of Mathematics and Statistics, Chengdu University of Technology, Chengdu, Sichuan, China

Short Summary

This study introduces a novel physics-informed deep learning framework to enhance the spatial resolution of global groundwater storage anomaly data from 0.5 degrees to 0.125 degrees. The framework achieves a 26% improvement in accuracy over baselines while maintaining hydrological consistency through soft multi-scale physical constraints.

Objective

• To develop a physics-informed deep learning framework that enhances the spatial resolution of global groundwater storage anomaly data from 0.5 degrees to 0.125 degrees, addressing the insufficiency of current satellite data for regional aquifer management and providing physically consistent, high-resolution estimates.

Study Configuration

• Spatial Scale: Global land coverage, downscaling from 0.5-degree resolution (approximately 55 kilometers) to 0.125-degree resolution (approximately 14 kilometers), using 32x32 coarse-resolution patches and 128x128 fine-resolution patches.
• Temporal Scale: Monthly observations from April 2002 to September 2023 (225 months), with the ConvLSTM encoder capturing a six-month climate memory.

Methodology and Data

• Models used: Physics-informed deep learning framework integrating:
  • Temporal Convolutional Long Short-Term Memory (ConvLSTM) encoder (2 layers, 64 channels, 3x3 kernel) for capturing climate memory and recharge lag dynamics.
  • U-Net spatial decoder (3 downsampling/upsampling levels with 64, 128, 256 channels) for spatial refinement.
  • Dual-head neural network architecture for predicting mean groundwater storage anomaly and spatially varying uncertainty variance.
  • Multi-scale soft physics constraints: sign agreement, magnitude correlation with soil moisture modulation (α(θ) = 0.1 + 0.4θ), and regional mass balance (aggregated over 32x32 pixel patches).
  • Spatial smoothness regularization (total variation).
• Data sources:
  • Global Groundwater Product version 1.12 from GRACE and GRACE Follow-On satellite missions (monthly groundwater storage anomaly at 0.5-degree resolution).
  • ERA5-Land reanalysis (monthly total precipitation, volumetric soil moisture in the surface layer, two-meter air temperature, and surface pressure).

Main Results

• Achieved a four-fold spatial resolution enhancement of groundwater storage anomaly data, from 0.5 degrees to 0.125 degrees.
• The model achieved a mean absolute error (MAE) of 1.64 ± 0.85 centimeters and a coefficient of determination (R²) of 0.9983 on independent test patches.
• Demonstrated a 26% improvement in MAE compared to the nearest-neighbor interpolation baseline (2.21 ± 1.25 centimeters).
• Physics loss converged to 3.4 ± 0.1 (total physics loss of 2.99 at epoch 30), confirming the satisfaction of soft hydrological constraints and hydrological plausibility.
• The framework successfully captures complex spatial patterns with smooth transitions and accurately tracks seasonal cycles in groundwater storage, with optimal performance achieved using a six-month climate memory window.
• Uncertainty quantification, while present, exhibited calibration challenges, indicating that predicted uncertainties do not reliably reflect actual error magnitudes.

Contributions

• First application of temporal memory architectures (ConvLSTM) to capture recharge lag effects in groundwater storage anomaly downscaling.
• Introduction of novel multi-scale soft physics constraints that respect regional (32x32 pixel patches) rather than strict pixel-level water balance.
• Integration of dynamic soil-moisture-modulated infiltration efficiency to capture fundamental hydrological non-linearity.
• Implementation of aleatoric uncertainty quantification through a dual-head neural network architecture, providing confidence estimates for risk-aware water management decisions.
• Provides physically consistent, temporally coherent, and uncertainty-aware spatial refinements that advance sustainable groundwater governance.

Funding

• Computational resources provided by institutional facilities at Chengdu University of Technology and University of Electronic Science and Technology of China.
• Acknowledgment of NASA and GRACE/GRACE-FO mission teams for satellite gravimetry observations.
• Acknowledgment of the European Centre for Medium-Range Weather Forecasts for ERA5-Land reanalysis data.
• Acknowledgment of the GFZ German Research Centre for Geosciences for processing the Global Groundwater Product.

Citation

@article{Ibrahim2026Securing,
  author = {Ibrahim, Kunle Adefarati and Babatunde, Ologun Sodiq and Ukwuoma, Chiagoziem C. and Akeredolu, Richard Joshua and Saleem, Victoria Chioma Ayozie-Samuel; Maryam},
  title = {Securing the Silent Reserve: Physics-Informed Deep Learning for Global Groundwater Storage Downscaling},
  journal = {Iconic Research and Engineering Journals},
  year = {2026},
  doi = {10.64388/irev9i9-1714801},
  url = {https://doi.org/10.64388/irev9i9-1714801}
}