Reconstruction of GRACE Total Water Storage Through Automated Machine Learning

Abstract

AbstractThe Gravity Recovery and Climate Experiment (GRACE) satellite mission and its follow‐on, GRACE‐FO, have provided unprecedented opportunities to quantify the impact of climate extremes and human activities on total water storage at large scales. The ∼1‐year data gap between the two GRACE missions needs to be filled to maintain data continuity and maximize mission benefits. In this study, we applied an automated machine learning (AutoML) workflow to perform gridwise GRACE‐like data reconstruction. AutoML represents a new paradigm for optimal algorithm selection, model structure selection, and hyperparameter tuning, addressing some of the most challenging issues in machine learning applications. We demonstrated the workflow over the conterminous U.S. (CONUS) using six types of machine learning models and multiple groups of meteorological and climatic variables as predictors. Results indicate that the AutoML‐assisted gap filling achieved satisfactory performance over the CONUS. On the testing data, the mean gridwise Nash‐Sutcliffe efficiency is around 0.85, the mean correlation coefficient is around 0.95, and the mean normalized root‐mean‐square‐error is about 0.09. Trained models maintain good performance when extrapolating to the mission gap and to GRACE‐FO periods (after June 2017). Results further suggest that no single algorithm provides the best predictive performance over the entire CONUS, stressing the importance of using an end‐to‐end workflow to train, optimize, and combine multiple machine learning models to deliver robust performance, especially when building large‐scale hydrological prediction systems and when predictor importance exhibiting strong spatial variability.