Using NASA Satellite Gravity Measurements to Detect Critical Groundwater Depletion Two Months Before Physical Well Networks

Groundwater depletion is a critical threat to global agricultural and drinking water systems, yet conventional groundwater monitoring relies on sparse physical well networks that typically detect shortages only after a crisis has already occurred. This study aims to pivot from reactive observation to proactive management by quantifying exactly how far in advance NASA Gravity Recovery and Climate Experiment (GRACE) and GRACE-FO satellite measurements can predict physical-well declines. Utilizing a 22-year data fusion spanning from 2002 to 2024, this research integrated GRACE satellite gravimetry, MODIS land surface temperature and vegetation indices, and 108,812 quality-filtered physical well measurements across California's heavily depleted Central Valley. A systematic lag correlation analysis identified the precise temporal gap between space-based signals and ground-based observations, demonstrating that changes in satellite gravity reliably predict physical well levels two months in advance (∣r∣=0.522, p=3.55×10−17). To operationalize this early warning window, an XGBoost machine-learning model was developed and tested against an unprecedented multi-year drought period (2020–2024). The XGBoost model predicted well depths with 44.9% higher accuracy than baseline linear methods (MAE = 38.7 feet). Subsequent SHAP explainability analysis validated the model's physical coherence, identifying 12-month rolling averages of GRACE liquid water equivalent (LWE) as the dominant predictor over instantaneous measurements or surface vegetation health. These findings conclude that satellite gravimetry offers a mathematically robust two-month head start for global water management. By actively monitoring 12-month rolling satellite averages, agencies can trigger managed aquifer recharge and targeted conservation efforts before ground-based systems confirm irreversible depletion. II. Table of Contents I. Abstract 2 II. Table of Contents 2 III. Key Words 3 IV. Abbreviations and Acronyms 3 V. Acknowledgements 3 VI. Biography 4 1. Introduction 4 2. Materials and Methods 6 2.1 Data Preprocessing and Feature Engineering 6 2.2 Lag Correlation Analysis 7 2.3 Model Training and Evaluation 7 2.4 SHAP Explainability Analysis 8 3. Results 8 3.1 Convergent Evidence from Space and Ground 8 3.2 GRACE as a Quantified Early Warning Signal 10 3.3 Model Performance 11 3.4 SHAP Feature Importance 12 3.5 Predicted vs. Actual Well Depth 13 4. Discussion 13 5. Conclusions 15 6. References 16