Forecasting Next Year's Global Land Water Storage Using GRACE Data

Existing approaches for predicting total water storage (TWS) rely on land surface or hydrological models using meteorological forcing data. Yet, such models are more adept at predicting specific water compartments, such as soil moisture, rather than others, which consequently impedes accurately forecasting of TWS. Here we show that machine learning can be used to uncover relations between nonseasonal terms of Gravity Recovery and Climate Experiment (GRACE) derived total water storage and the preceding hydrometeorological drivers, and these relations can subsequently be used to predict water storage up to 12 months ahead, and even exceptional droughts on the basis of near real-time observational forcing data. Validation by actual GRACE observations suggests that the method developed here has the capability to forecast trends in global land water storage for the following year. If applied in early warning systems, these predictions would better inform decision-makers to improve current drought and water resource management. Traditional methods for predicting short-term/seasonal variations in land total water storages rely on hydrological models. However, these models have a drawback-they are better at predicting water stored in specific parts of the land system like soil moisture than giving an accurate forecast for the overall integrated land total water storage. In this study, we demonstrate the applicability of machine learning in uncovering relationships between the de-season and de-linearized terms of global water storage variability as observed by the Gravity Recovery and Climate Experiment (GRACE) satellites, and the preceding hydrometeorological variables such as sea surface temperature. These relationships can then be utilized to forecast monthly changes in land total water storage up to 1 year ahead, and even to predict exceptional drought events based on near real-time observational forcing data. The validation by actual GRACE observations, lends further credence to the effectiveness of the method developed here, showcasing its potential to forecast trends in global land total water storage for the upcoming year. The potential applications of these predictions in early warning systems are highly promising, and we anticipate that they can assist decision-makers in enhancing current drought and water resource management practices. We identify a stable lag relationship between hydrometeorological variables and the GRACE derived total water storage change (TWSC) Using this stable lag relationship, we are able to forecast the global TWSC up to 1 year ahead with solely observational data as inputs Our forecasts exhibit high consistency with actual GRACE data in terms of global mean land water storage trends for the following year