Making GRACE and GRACE-FO more effective for climate research: reconstruction of terrestrial water storage anomalies over decades

The Gravity Recovery and Climate Experiment (GRACE) and its successor, the GRACE Follow-On (GRACE-FO) missions, have enabled the monitoring of Total Water storage anomalies (TWSA) from space. However, their combined observational record spans only two decades of monthly data, with a one-year gap between the two missions. This limited record constrains their application in climate research. To address this limitation, we developed two approaches to reconstruct GRACE TWSA: one using Machine Learning (ML) methods and the other using a novel Deep Learning (DL) approach.

In the ML approach, we integrated TWSA estimates from global hydrological models, land surface models, and ERA5 reanalysis data with Ensemble GRACE TWSA, enabling the reconstruction of TWSA records extending further back in time. For this purpose, various ML algorithms were employed, including Multivariate Linear Regression (MLR), Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), Gaussian Process Regression (GPR), and eXtreme Gradient Boosting (XGBoost).

Our DL approach combines Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks to capture the spatial and temporal dependencies in the TWSA data. As model inputs, we utilize multiple meteorological, hydrological, and vegetation-related variables from the ERA5 reanalysis. We also use the Oceanic Niño Index derived from NOAA’s Extended Reconstructed Sea Surface Temperature dataset to account for ocean variability. Additionally, land cover data (rain-fed and irrigated cropland, pastures, and urban areas) together with lake area fractions from ISIMIP are incorporated to represent anthropogenic influences.

We evaluated the reconstructed data against GRACE(-FO) observations and high-resolution Satellite Laser Ranging (SLR) TWSA data from the pre-GRACE period.  In this presentation, we show the validation results and compare the performance of ML- and DL-based approaches with each other and with other existing products. Our results, derived from both ML and DL methods, demonstrate improved accuracy compared to previous approaches, effectively capturing seasonality, trends, and human-induced variations. 

Our reconstructed data enhance the utility of GRACE and GRACE-FO for climate research by extending the temporal coverage of terrestrial water storage anomalies. This enables a deeper understanding of long-term hydrological trends, including the effects of climate variability and human activities on water storage.