Signal separation of temporal gravity signals for low-amplitude signal detection

Lithospheric uplift, once attributed mainly to plate tectonic and isostatic processes, is now recognized to be strongly influenced by convective processes in the Earth's mantle. Advances in satellite observations and data analysis have strengthened geodetic constraints on geodynamic models, specifically through satellite gravimetry. However, the superposition of mass change signals driven by different Earth processes requires robust signal separation to quantify the contributions of individual processes in the data.

Signal separation is a fundamental challenge in geodetic datasets, which commonly represent the superposition of multiple physical signals. Previous studies have explored isolating solid-Earth signals due to glacial isostatic adjustment (GIA) ^[1] applying a neural network–based signal separation method to simulated temporal gravity data. The neural network (NN) was trained to recognize and separate individual signal components by exploiting prior knowledge about their characteristic spatiotemporal behavior, derived from forward-modeled time-variable gravity data and additional constraints.

The employed NN architecture is a multi-channel U-Net designed to separate superimposed temporal gravity signals arising from mass redistribution in the atmosphere and oceans, continental hydrosphere, cryosphere, and solid Earth. The network separates these combined inputs into their constituent sub-components. The framework is generally applicable to signal separation in any three-axis dataset (e.g., latitude, longitude, and time), using a sampling strategy in which the data are partitioned along one axis to determine the optimal two-axis combination for training ^[2].

This work presents progress towards extracting signals originating from deep-Earth processes, particularly mantle convection signals, from time-variable gravity data such as observed by the Gravity Recovery and Climate Experiment (GRACE) and GRACE-Follow on (GRACE-FO) satellite missions. In this context, NN-based signal separation has been demonstrated primarily for signals with comparably large amplitudes. In contrast, time-variable gravity signals caused by processes in the Earth's mantle are approximately three orders of magnitude weaker than signals related to surface processes, rendering their detection and separation particularly challenging. The current study therefore focuses on enhancing sensitivity to low-amplitude mantle signals by leveraging the ability of machine learning methodologies to learn subtle spatiotemporal patterns.

For application to real data from the GRACE/-FO missions or the upcoming Mass-Change and Geosciences International Constellation (MAGIC), we propose training the framework on representative forward-modeled signals and simulated noise and subsequently applying the trained separation model to observational time-variable gravity data.

 

References:

• Heller-Kaikov B, Karimi H, Lekshmy Raj D, Pail R, Hugentobler U, Werner M. 2025 Signal separation in geodetic observations: satellite gravimetry. Proc. R. Soc. A 481: 20240820.
• Heller-Kaikov B, Pail R, Werner M. 2025, Neural network-based framework for signal separation in spatio-temporal gravity data Computers & Geosciences, Volume 207, 2026, 106057, ISSN 0098-3004.