Investigating Machine Learning Techniques for Estimating Line-of-sight Gravity Differences from GRACE-FO data

Kurzfassung

Time Variable Gravity (TVG) observations provide crucial insights into Earth’s dynamic processes, including phenomena such as ice sheet dynamics, ocean circulation, tectonic activity and changes in mass distribution, which are key to understanding climate change and natural disaster forecasting. The GRACE-FO system with its precise inter-satellite distance measurements is used to obtain gravity field solutions, which are computed monthly and thus impede the observation of short-term processes such as hydrological phenomena. By linking between Line-of-Sight (LOS) gravity differences and intersatellite ranging data, in-situ gravity estimates along the orbit can be generated. While these point-wise estimates yield a higher temporal resolution, a key challenge in this method is the derivation of an effective transfer function which translates geometric ranging observations into gravity field quantities. Traditional derivation techniques have various disadvantages, such as limitation to high frequency regimes and noise sensitivity. Using Machine Learning techniques may address some of those limitations by deriving an adaptive, data-driven transfer function. This work explores the implementation of deep neural networks to derive LOS gravity differences from intersatellite ranging data. The approach is evaluated against standard methods, such as monthly spherical harmonic solutions and daily Kalman-filtered solutions. By utilising machine learning, this study seeks to refine TVG observations to enable both improved monitoring of short-term geophysical processes and understanding of the GRACE-FO LRI measurement system.