Predicting Sub-Monthly Total Water Storage Variations Using a New 5-Day Mascon Product and Deep Learning

GRACE and GRACE Follow-On (GRACE-FO) missions provide unique information on the wetness state of a river basin with regard to its flood generation potential. However, the long latency of the standard monthly GRACE products has limited their direct applications in operational flood early warning.  The Center for Space Research (CSR) at The University of Texas is developing a new 5-day mascon product (CSR.5d) using GRACE measurements only. The 5-day solution represents the latest global mass changes and may thus be useful for detecting and predicting hydroclimate events (e.g., flooding) at the sub-monthly scale. In this work, we assessed the predictability of GRACE-like, short-term total water storage anomalies (TWSA) by using the experimental CSR.5d product as training samples. Specifically, a probabilistic deep learning model was used to learn the state-transition model underlying the CSR.5d TWSA, by using the antecedent TWSA and hydroclimatic variables (e.g., precipitation and air temperature) as predictors. By design, our method generates both predicted mean and uncertainty intervals at the same time. Performance metrics, obtained at the grid- and basin scales, suggest that the inherent dynamics of the TWSA variations can be learned well using the machine learning method, thus providing important insights on the operational use of the 5-day mascon product.