Reconstructing pre-GRACE terrestrial water storage anomalies using deep learning

Abstract

Terrestrial water storage (TWS), the summation of all water components above and below the ground, plays an important role in describing the Earth’s climate, as water availability is decisive for ecosystems and human development. Since 2002, the Gravity Recovery And Climate Experiment (GRACE) and its successor mission GRACE-FO measure TWS anomalies with unprecedented accuracy, which enabled a leap in hydrological research. However, the use of the GRACE dataset is limited due to its relatively short time span and the one-year gap between GRACE and GRACE-FO. Here, we develop a deep learning model to reconstruct GRACE-like TWS anomalies from 1941 to 2021. The reconstruction covers global land area except Greenland and Antarctica. Our approach combines convolutional neural networks (CNNs) with long short-term memories (LSTMs) to take into account the spatial and temporal structure of the input. By a wide selection of input datasets, we aim at reconstructing climate and human-induced TWS variability, including long-term trends. For our final model, we train an ensemble of models to increase the prediction accuracy and obtain uncertainty estimates. Two products are provided: One product is based on the input of ERA5, lake fraction, land use and an ENSO index, while the other additionally incorporates simulated TWS from the WaterGAP global hydrology model. Both products show a higher agreement with GRACE than previous reconstructions. TWS changes show a high correlation with the ERA5 basin-scale water balance during the pre-GRACE era, suggesting an accurate reconstruction of the seasonal cycle. To evaluate the predicted trend on a global level, we compare the TWS reconstructions with global mean sea level (GMSL) contributor estimates. The model trained only with climate reanalysis and human-influence input shows a higher sea level budget closure in the near past (1980-2002), while the model with additional hydrology model input shows a higher closure in the distant past (1941-1980). This work suggests that the application of powerful deep learning techniques is a promising method for the reconstruction of TWS anomalies.