Analysis of gap filling techniques for GRACE/GRACE-FO terrestrial water storage anomalies in Canada

Abstract

Terrestrial water storage anomaly (TWSA) data derived from the Gravity Recovery and Climate Experiment (GRACE/GRACE-FO) have become indispensable for hydrological monitoring since the first mission launched in 2002. Evaluation of basin dynamics is enabled through TWSA integration with independent hydrological and climate datasets such as precipitation, evapotranspiration, and surface runoff. With over 20 years of observations, long-term statistical analysis reveals water storage trends, provided the 11-month gap between the two missions is filled. This study compares four methods for reconstructing GRACE-like TWSA over Canada, namely: extreme gradient boosting, artificial neural networks, automated machine learning, and projection onto convex sets. The GRACE mascon product released by the Jet Propulsion Laboratory is used over the study region. Nine sets of hydrological and climate parameters are used as predictors for the machine learning models, namely GRACE average seasonal signals, TWSA from the GLDAS Catchment Land Surface Model, precipitation, air temperature, glacier surface mass balance, ocean tides, the North Atlantic Oscillation, the Multivariate ENSO Index, and sea surface temperature. For each of the four methods, 20 % of the GRACE TWSA data (the ‘testing data’) is omitted from the model training, and the model is used to fill the artificial gap. Root mean square error (RMSE) is computed using the difference between testing data and model predictions. Normalized RMSE expresses the RMSE as a proportion of the range of the GRACE TWSA data. The automated machine learning algorithm performs best in terms of mean normalized root mean square error, with 6 % (2.4 cm equivalent water height RMSE), followed by projection onto convex sets, extreme gradient boosting, and artificial neural networks. Inclusion of glacier surface mass balance models derived from the GMAO Modern-Era Retrospective Analysis for Research and Applications Version 2 reanalysis and the Randolph Glacier Inventory improved the automated machine learning average root mean square error by 70 %. Results indicate that filling the gap between missions allows for more comprehensive analysis of terrestrial water storage changes, including basin-by-basin analysis, which is ongoing.