GRACE Downscaling and Machine Learning Models for Groundwater Prediction: A Systematic Review

Gravity Recovery and Climate Experiment (GRACE) satellites primarily monitor changes in land water storage, including groundwater, soil moisture, lake and river surface water, and canopy and snow water. However, its coarse spatial resolution of 0.25 degrees limits its ability to observe smaller basins. To assess aquifer depletion and evaluate a long-term water resource management framework, GRACE data are crucial. It remains rare for GRACE-focused studies to be conducted in great depth. A comprehensive review of 80 articles published between 2011 and 2025 was conducted using the Scopus and Web of Science databases. These articles focused on downscaling GRACE data using machine learning (ML) methods. The Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) reporting guidelines were used in this review. This study highlights the attributes of ML models, the input variables used, the evaluation metrics, and the output resolution. Based on the analysis of the articles, random forest (RF) methods were used in the majority of the papers. Gradient boosting (GB), artificial neural networks (ANN), support vector machines (SVM), support vector regression (SVR), and long short-term memory (LSTM) were the most widely used ML methods. As input variables, rainfall (Pr), soil moisture (SM), and runoff (Qs) are essential. In 2011, there were very few journal articles; since 2021, the number has increased. The number of published studies from China was the highest (24), followed by the USA (12) and Iran (9). A total of 38 journals published reviewed articles. In terms of articles, Remote Sensing generates 19%, Journal of Hydrology has 10%, and Journal of Hydrology: Regional Studies has 8%. The paper also discusses limitations, challenges, recommendations, and potential future directions for improving the accuracy of the GWS change prediction model.