The GRACE and GRACE Follow-On missions are separated by an 11-month gap between 2017 and 2018 and contain 22 more missing months. These gaps in the time series lead to a difficult recovery of gravity variation signals with pluri-annual temporal scales. In this context, various studies proposed machine learning approaches and decomposition techniques to predict the missing values.

This study summarizes the different approaches that we have implemented and compares their results. We consider both grid and spherical harmonics at global scales. Some gap-filling solutions use an extrapolation of the GRACE products and some others propose to use Swarm gravity field products to reduce the missing data. We tested several methods in terms of their capacity to predict signals on monthly or annual periods, randomly chosen between 2005 and 2010. The Root-Mean Square Error between the predictions and the original solution gives an estimation of the uncertainty associated with each method.

We show that simple methods like « Constant, Trend, Annual and Semi-annual fit » do not deliver the complexity of the original signal. We finally conclude that the Singular Spectrum Analysis (SSA) and Multivariate SSA produce the best results at large spatial scales.