The accurate prediction of satellite orbits is essential for the proper functioning of space-based services such as navigation, communication, and Earth observation. However, atmospheric drag is a significant source of error in satellite orbit prediction, especially in Low Earth Orbit (LEO), where the majority of space objects operate. The thermosphere, the outermost layer of Earth's atmosphere, plays a crucial role in determining atmospheric drag. However, the thermospheric density is subject to high levels of uncertainty, up to 30%, when computed from atmospheric models. This uncertainty is particularly relevant in LEO, where it can affect the operational lifespan of satellites.

To increase the accuracy of thermospheric density models, this paper presents an assessment of the accuracy of orbit predictions using empirical Thermospheric Mass Density (TMD) observations obtained from the Swarm C satellite. The study uses a Principal Component Analysis (PCA) to decompose a fine grid of density into the main temporal and spatial modes. Then, each of the modes is calibrated with Swarm C observations using a Least Squares Estimation (LSE) algorithm. The calibration is validated with the observation using unbiased metrics. These observations were assimilated into the NRLMSISE-00 atmospheric model, and the resulting calibrated density model was used to predict the orbits of Swarm C, GRACE-FO and Sentinel 1-A satellites. To analyse the consistency of the results, a slicing window analysis was performed, and the median evolution of the windows was computed.

The results of the study indicate that a calibrated density model with a several satellite geometries, such as the cannonball model, a panel model, or a scaled panel model, can significantly improve the accuracy of orbit predictions in LEO. During March 2022, a period of medium solar activity, the median accuracy of orbit predictions for Swarm C was reduced from 20.67 km with NRLMSISE-00 to 13.75 km with the calibrated model and a scaled panel model geometry. During the same period, the median accuracy of orbit predictions for GRACE-FO C was reduced from 17.98 km with NRLMSISE-00 to 2.89 km with the calibrated model and an (unscaled) panel model geometry. These findings have important implications for the sustainable operation of satellites in the increasingly crowded space environment.