The main scientific goal of the NASA GRACE and GRACE Follow-On (GRACE-FO) missions, in collaboration with the German Aerospace Center (DLR), is to accurately monitor the Earth’s gravitational field and its temporal variations. The resulting gravity field solutions and derived terrestrial water storage changes (TWSC) are essential for understanding oceanography, calculating ice mass loss, monitoring hydrology, and modeling climate change. After GRACE’s final data release in May 2017, GRACE-FO began measurements in May 2018, resulting in an 11-month gap in the observational record. Various studies have sought to fill this data gap.

One approach is to use Swarm satellite data. The European Space Agency’s (ESA) Swarm mission, launched in 2013, primarily maps the geomagnetic and electric fields in the atmosphere. However, GPS data from these low-Earth orbit satellites can also detect long-wavelength changes in Earth’s gravity. Swarm provides independent information on TWSC, but with a much lower spatial resolution (1500 km compared to GRACE’s 350 km). Another approach is the use of data-driven techniques such as machine learning, which leverage hydro-climatological maps, including precipitation and soil moisture from models such as GLDAS-NOAH, to reproduce GRACE-like TWSC estimates. These datasets offer high temporal and spatial resolution but do not directly measure the gravitational field itself.

The aim of this thesis was to combine both approaches by using hydro-climatological data and Swarm-derived gravity data as inputs to neural networks to predict GRACE-like TWSC. The Amazon Basin, a highly hydrologically active region, was selected as the study area. The datasets included GRACE RL06 data, Swarm-derived gravity field models, and hydro-climatological variables such as precipitation, temperature, and soil moisture. Analysis showed that soil moisture was the most effective hydro-climatological predictor for TWSC and was used as the sole hydrological input in the final simulations.

Two neural network architectures were evaluated: a fully connected Multilayer Perceptron (MLP) with two hidden layers, and a convolutional neural network (CNN) with four convolutional layers followed by three fully connected layers. After optimizing the architecture and hyperparameters, the MLP slightly outperformed the CNN while requiring significantly less training time, making it suitable for further analysis.

Three experiments were performed: (1) using only soil moisture as input, (2) using only Swarm data, and (3) combining both datasets. Soil moisture alone led to underfitting (validation NSE 0.43), while Swarm data alone caused overfitting (validation NSE 0.6). The combination of both datasets produced the best performance, with a validation NSE of 0.68 and a high temporal correlation with GRACE data (CC 0.95). This demonstrates that integrating Swarm and hydro-climatological data provides the most reliable method for filling the GRACE/GRACE-FO data gap.

In conclusion, this study shows that data-driven techniques combined with complementary datasets offer a promising approach to mitigate the effects of missing GRACE data, although the performance remains limited by the availability of high-quality Swarm observations. ... Read more