Investigating Machine Learning Techniques for Estimating Line-of-Sight Gravity Differences from GRACE-FO Intersatellite Ranging Data

Kurzfassung

Time Variable Gravity (TVG) observations provide crucial insights into Earth’s dynamic processes, including ice sheet motion, ocean circulation, and tectonic activity - all key to understanding climate change and improving natural disaster forecasting. The GRACE-FO mission uses precise intersatellite ranging measurements to generate gravity field solutions on a monthly-mean basis. Though these solutions monitor seasonal and annual trends, they fail to capture fast-evolving climate events. By relating line-of-sight gravity difference (LGD) to ranging observations, it is possible to produce in-situ gravity estimates along the satellite orbit, offering a higher temporal resolution. However, difficulty lies in establishing a robust transfer function that converts geometric ranging observables into gravity field parameters, as conventional approaches often struggle with low-frequency signal recovery and are prone to noise amplification. This study explores the use of machine learning as an alternative transfer function. Three models were developed - a Random Forest Regressor (RF), a Multi-Layer Perceptron Neural Network (MLP), and a Long Short-Term Memory Neural Network LSTM) - and trained on synthetic GRACE-FO-like ranging data. The models were applied to two GRACE-FO test datasets: one synthetic and one real - corresponding to the July 2021 Ahr valley floods in western-Germany. Model performance was evaluated in both the time and frequency domains. On the simulated data set, the RF demonstrated the best performance, predicting LGDs with an mean squared error (MSE) of 3.15×10^-2 µGal^2 and excelling in lowfrequency (< 5 cycles-per-revolution (CPR)) signal recovery. The MLP generalised best to real data, producing results comparable to that of conventional transfer function techniques, whereas the RF and LSTM models showed signs of low-frequency aliasing. These findings indicate that machine learning offers a promising new approach to in-situ gravity estimation, but further development is required before operational implementation.