Securing the Silent Reserve: Physics-Informed Deep Learning for Global Groundwater Storage Downscaling

Groundwater represents Earth’s largest accessible freshwater reserve, providing essential water resources for over 2 billion people worldwide and supporting approximately 40 percent of global irrigation. However, monitoring groundwater storage at actionable spatial resolutions remains a fundamental challenge in hydrology and water resource management. Current GRACE and GRACE Follow-On satellite missions provide groundwater storage anomaly data at approximately 0.5-degree spatial resolution, which proves insufficient for regional aquifer management, agricultural planning, and water policy enforcement. This study presents a novel physics-informed deep learning framework that enhances groundwater storage anomaly spatial resolution from 0.5 degrees to 0.125 degrees, achieving four-fold spatial refinement while maintaining hydrological consistency through soft multi-scale physical constraints. The framework integrates a temporal Convolutional Long Short-Term Memory encoder that captures six-month climate memory with a U-Net spatial decoder, constrained by water balance principles, soil-moisture-modulated infiltration efficiency, and regional mass conservation. Trained on 8,000 spatially and temporally distributed patches extracted from 57,503 globally valid locations spanning 225 monthly observations from April 2002 to September 2023, the model achieves a mean absolute error of 1.64 plus-minus 0.85 centimeters with a coefficient of determination of 0.9983, demonstrating 26 percent improvement over the nearest-neighbor interpolation baseline. Physics loss convergence at 3.4 plus-minus 0.1 confirms hydrological plausibility, while the framework provides spatially explicit uncertainty quantification essential for risk-aware water management decisions. This work advances sustainable groundwater governance by providing physically consistent high-resolution estimates that respect fundamental hydrological principles.