Integrating Satellite Retrievals, Numerical Models, and Machine Learning for Global Groundwater Recharge Estimation

Abstract Knowledge of the groundwater recharge rate determines whether aquifer use is sustainable. However, accurately measuring recharge globally presents significant challenges due to the complexity of subsurface processes and the lack of direct observational methods. This study addresses these challenges by developing a methodology that integrates satellite data, numerical models, and machine learning to estimate groundwater recharge globally. The methodology involves two steps. First, we run a numerical model, Hydrus‐1D, to simulate soil moisture fluxes in the unsaturated zone by solving the Richards equation in the vertical direction for 235 different points representing various climates and soil types across the globe. Second, using Hydrus‐1D inputs and outputs, we train a supervised ensemble machine‐learning model, specifically a Gaussian Process Regression model, as an emulator to mimic Hydrus‐1D. This enables us to process satellite observations efficiently to estimate annual recharge flux globally. Inputs for the model include NASA's SMAP soil moisture and GPM precipitation observations, ERA5 climate reanalysis data, and soil hydraulic properties. Rainfall, unsaturated hydraulic conductivity, and soil moisture are identified as the most significant predictors of groundwater recharge. The approach effectively captures global recharge patterns, particularly in regions with high rainfall, though it shows some limitations in arid areas with minimal recharge and heavily irrigated areas. We confirm the reasonableness of recharge estimates by comparing them with observed changes in subsurface water storage from the GRACE satellite mission. The method effectively captures the observed trends in water storage, demonstrating the model's capability to estimate recharge using large‐scale satellite and reanalysis data.