Improved estimation of global-scale changes in groundwater storage using machine learning.

The importance of groundwater as a reliable freshwater source is expected to increase amidst climate change, given its greater resilience to short- and mid-term fluctuations in the hydrological cycle. Over the last two decades, the Gravity Recovery and Climate Experiment (GRACE) has enabled large-scale satellite monitoring of changes in total terrestrial water storage globally. Changes in groundwater storage have, to date, been computed from total terrestrial water storage by deducting mass contributions from other terrestrial water stores: soil moisture, surface water, and snow and ice. Estimates of changes in these terrestrial water stores derive primarily from inadequately constrained land surface models (LSMs) and occasionally from in situ observations. Substantial uncertainty persists in the estimated changes in terrestrial water stores from uncalibrated, global-scale LSMs hindering robust monitoring of groundwater storage from GRACE. The project aims to devise a data-driven framework to calculate groundwater storage from total terrestrial water storage, leveraging newly accessible global-scale piezometric observations. Next, as the global process-based models oversimplify water storage components, a surrogate deep learning model for global ground water storage will be created to capture non-linear relationships in predictive variables and groundwater trends, and to transfer the learned knowledge to the regions with limited observations. Improved estimates of groundwater storage will help to inform basin-scale management for sustainable freshwater use.