Climate-to-Storage: Leakage-Controlled ConvLSTM Forecasting of GRACE-Derived Terrestrial Water Storage over Tamil Nadu, India

Abstract Deep networks are increasingly used to forecast satellite gravimetry, but the skill they show depends on how much memory of the storage record they are given. We examine this for Tamil Nadu using GRACE and GRACE Follow-On mascon terrestrial water storage (238 months, 2002-2024) and five TerraClimate variables. Each field was dated from its observation interval; a count-based labelling would have misdated 81 of 238 fields. All masks and statistics came from training months only, and 14 automated checks guarded against leakage. A convolutional long short-term memory network (ConvLSTM) predicted the next month from 18 months of climate alone (Model B) or with the latest storage field added (Model A). Over 19 withheld months, Model B had a root-mean-square error (RMSE) of 7.84 cm, against 12.76 cm for climatology and 12.26 cm for a Random Forest, but persistence was more accurate (5.12 cm) and the anomaly correlation was only 0.365. About 94% of Model B's gain over climatology came from correcting bias in an unusually wet period. Model A reached 4.52 cm yet was not significantly better than persistence (p = 0.374; low power). Climate forcing therefore adds limited month-ahead information, and persistence should always be reported as a baseline.