Signal separation in global, temporal gravity data using a multi-channel U-Net

One big challenge in the analysis and interpretation of geodetic data is the separation of the individual signal and noise components contained in the data. Specifically, the global, temporal gravity data obtained by the GRACE and GRACE Follow-On satellite missions contain spatial-temporal gravity signals caused by all kinds of mass variations in the Earth’s system. While only the sum of all signals can be measured, for geophysical interpretation, an extraction of individual signal contributions is necessary.

Therefore, our aim is to develop an algorithm solving the signal separation task in global, temporal gravity data. Since the individual signal components are characterized by specific patterns in space and time, the algorithm to be found needs to be able to extract patterns in the 3-dimensional latitude-longitude-time space.

We propose to exploit the pattern recognition abilities of deep neural networks for solving the signal separation task. Our method uses a multi-channel U-Net architecture which is able to translate the sum of various signals as single-channel input to the individual signal components as multi-channel output. The loss function is a weighted sum of the L2 losses of the individual signals.

We perform a supervised training using synthetic data derived from the updated Earth System Model of ESA. The latter consists of separate datasets for temporal gravity variations caused by mass redistribution processes in the atmosphere, the oceans, the continental hydrosphere, the cryosphere and the solid Earth.

In our study, we use different parts of this dataset to form training and test datasets. In this fully-synthetic framework, the ground truth of the individual signal components is also known in the testing stage, allowing a direct computation of the separation errors of the trained separation model.

In our contribution, we present results on optimizing our algorithm by tuning various hyperparameters of the neural network. Moreover, we demonstrate the impact of the number of superimposed signals and the definition of the loss function on the signal separation performance of our algorithm.