1,2,2,1- 1ETH Zurich, Institute of Geodesy and Photogrammetry, Department of Civil, Environmental and Geomatic Engineering, Zürich, Switzerland (beheller@ethz.ch)
- 2Technical University of Munich, Chair for Astronomical and Physical Geodesy, TUM School of Engineering and Design, Munich, Germany
Time-variable gravity data from the Gravity Recovery and Climate Experiment (GRACE) and GRACE-Follow on (GRACE-FO) satellite missions are widely used in the Earth science community. GRACE/-FO monthly gravity data products contain strong correlated noise and the superposition of various geophysical mass change signals. Therefore, several processing steps are applied to derive user-friendly Level-3/Level-4 data products used for most applications of GRACE/-FO data. These processing steps include the application of de-noising filters and the reduction of signals besides the considered target signal, such as terrestrial water storage. Both steps introduce errors to the data, which finally propagate to the considered application domain: The de-noising step usually does not only suppress noise but also removes parts of the signal. The signal separation step, often performed using physical reduction models, is affected by model errors.
We consider a separation method using machine learning techniques, replacing both the filtering and signal separation steps. The original method was published by Heller-Kaikov et al. in 2026 and showed promising results in a closed-loop simulation setup. Building upon that, we train the neural network-based pattern recognition algorithm on the task of decomposing a sum of time-variable gravity signals and a GRACE-type noise component into the individual components. After training the network on simulated signal and noise components, we test the resulting separation algorithm on real GRACE/-FO Level-2 data. To evaluate the de-noising and signal separation capabilities of our framework, we validate our results against alternative data products such as the GFZ GravIS terrestrial water storage or ice mass change products.
Our results demonstrate how machine learning algorithms can help solve the signal-noise and signal-signal separation problems in spatio-temporal data, therefore representing an alternative to state-of-the-art de-striping filters and reduction model-based signal separation strategies.
Heller-Kaikov, B., Pail, R., Werner, M. (2026): Neural network-based framework for signal separation in spatio-temporal gravity data, Computers & Geosciences, 207. doi: 10.1016/j.cageo.2025.106057
How to cite: Heller-Kaikov, B., Schlaak, M., Pail, R., and Soja, B.: Separation of time-variable gravity signals in GRACE/-FO data with Machine Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7248, https://doi.org/10.5194/egusphere-egu26-7248, 2026.
