Filling the data gap between grace and grace follow-on missions using deep learning algorithms

Abstract

The main purpose of GRACE (Gravity Recovery and Climate Experiment) satellites, which was launched on 17 March, 2002, is to monitor the changing gravity field of the Earth due to mass changes in the hydrosphere, cryosphere and ocean then record all the data on a monthly basis. GRACE satellites provide the dynamics of terestrial water storage anomalies (TWSA) on a global scale with very high accuracy, both spatially and temporally. These data are used in sustainable water resources management, hydrologic and climatic studies. The mission of the GRACE satellites ended on 17 October, 2017 due to battery problems. Thereupon, the GRACE-Follow on (GRACE-FO) was launched with the same mission. To that end, there was an 11-months data gap between GRACE and GRACE-FO. It is predicted that such a data gap in the TWSA time series will lead to significant bias and uncertainties in the model predictions. This study covers the period between December 2013 and December 2020. In the study process, besides the data gap between two missions, there is also 12-months data gap due to technical reasons. As a result of this study, a total of 23 months of data was filled by using deep learning (DL) algorithms. As input data, ERA5-Land driver data which include monthly temperature, precipitation, cumulative water storage changes and ERA5L-derived TWSA were used. As an additional input, long wavelet of the gravitational signal derived from SWARM TWSA were used. All data used in the study were temporally rescaled monthly and spatially at 1°x1° resolution before deep learning process. The L2 data obtained from the SWARM satellite was not used directly, but was first subjected to corrections applied as in CSR mascon solutions (degree 1 correction, C20/C30 and GIA corrections). SWARM Sh (Spherical Harmonic) models truncated at degree and order 12. Then, Gaussian smoothing filter with a radius of 1000 km was applied to reduce the noise in the recovered SWARM-derived TWSA. In this study, it is aimed to fill the data gap by using three different Deep Learning (DL) algorithms, namely, Convolutional Nural Network (CNN), Deep Convolutional Autoencoders (DCAE) and Bayesian Convolutional Neural Network (BCNN). The time period covering the study is 84 months, including gaps (December, 2013 - December, 2020). The mascon solutions are available for 61 months within the study period. 13 months randomly selected from the available mascon solutions were used for the test and the remaining 48 months were used for training the DL models. Finally, totally 23 months of TWSA data gap was successfully filled using DL algorithms.