Reanalysis and Forecasting of Total Water Storage and Hydrological States by Combining Machine Learning With CLM Model Simulations and GRACE Data Assimilation

Hydrological Models face limitations in simulating the water cycle due to deficiencies in process representation and such problems also weaken their forecasting skills. Here, we use Machine Learning (ML) to forecast the Gravity Recovery and Climate Experiment (GRACE) derived total water storage anomaly (TWSA) up to 1 year ahead over Europe with near real-time meteorological observations as predictors. Subsequently, we assimilate the forecasted and GRACE TWSA into the Community Land Model (CLM) to enhance its performance in both reanalysis and forecast. As found in five hindcast experiments, ML forecasted TWSA for the following year fits quite well to the actual GRACE observations over Europe, with an average correlation of 0.91, 0.92, and 0.94 in the Iberian peninsula, Danube, and Volga basins. Validation by observations and reanalysis data suggests that assimilating forecasted TWSA can improve CLM's capacity to forecast not only hydrological states but also hydrological droughts. Additionally, ML forecasted TWSA is a viable alternative to GRACE data in terms of enhancing hydrological forecasting on seasonal to annual scales through Data assimilation (DA). We also highlight the contribution of GRACE DA for generating a CLM based TWSA reanalysis that overcomes deficiencies of purely model-based TWSA. This study suggests that seasonal drought or water resource forecasting services might not only consider to integrate GRACE TWSA but would also benefit from constraining models with ML-forecasted TWSA. At shorter timescales, such forecasts could also be useful in the quick-look analysis of near real-time TWSA processing as is suggested for upcoming satellite gravity missions.