Neural network-based framework for signal separation in spatio-temporal gravity data

Global, temporal gravity data such as those provided by the Gravity Recovery and Climate Experiment (GRACE) and GRACE-Follow on (GRACE-FO) satellite missions contain signals from many mass redistribution processes on Earth. These include hydrological, atmospheric, oceanic, cryospheric and solid Earth-related processes. As the measured gravity changes represent the sum of all signals, an optimal exploitation of these data for scientific applications requires strategies for separating the individual contained signals. We provide a neural network algorithm using a multi-channel U-Net architecture that translates the sum of several signals to the individual contained components based on their typical space–time patterns. The software contains strategies for transforming spatio-temporal gravity data depending on latitude, longitude, and time to 2-D “image” training samples. The software also includes implementations of strategies for introducing additional knowledge about the physical behavior of the individual signals as constraints to the training. In a closed-loop simulation example, simulated gravity signals induced by processes in the atmosphere and oceans, hydrosphere, cryosphere and solid Earth are successfully separated at relative RMS prediction errors between 19 and 67%. This shows that neural network-based methods can help solving geodetic tasks if the considered data is transformed into a suitable data format. To apply the framework to real observational data, we suggest training the network on representative, physical forward-modeled signals and subsequently applying the trained network to real data. The latter will additionally require external validation strategies. The software is freely available on GitHub under https://github.com/Betty-Heller/neural-gravity and is, in general, also applicable for signal separation in any other dataset depending on three variables.