The Swarm satellite mission, in operation since November of 2013, has been collecting high-accuracy GPS tracking data for over a decade. Kinematic orbit solutions derived from this data have been used as input to gravity field estimation techniques for some time, with research initiatives such as the 'Multi-approach gravity field models from Swarm GPS data' dedicated to the production of highly accurate time-varying models. The production of these models is of great value to the geodetic community, as they provide an alternative and independent source of gravity field information which can be used to bridge gaps in the service of dedicated gravity field missions such as the Gravity Recovery and Climate Experiment, and provide additional temporal and spatial coverage of Earth's gravity field. The determination of this time-varying component is of great scientific value, as these models capture the complex motion of mass about the Earth occurring over short periods of time; such as the melting of polar icecaps and drought. The focus of this thesis is the improvement of the Swarm kinematic orbits used as input for such models.

We begin with an introduction to the research, including background of the topic, the scope of the research, the current state of the art and the gap in knowledge to be addressed culminating in the presentation of research questions. Input kinematic orbits from the Technical University of Delft, the Astronomical Institute of the University of Bern and the Institute of Geodesy Graz are retrieved. We apply input consolidation using interpolation to bring each input to consistent 1-second frequency sampling. We then apply a 0.3m outlier-filtering using the residuals of each input with respect to a reduced dynamic reference orbit. The consolidated and pre-screened input orbits are then combined using the following techniques: arithmetic mean weighting, inverse-variance weighting, iterative variance component estimation, weighting with the residuals with respect to a \ac{RDO} reference, and optimisation using an \ac{RDO} reference.

The resulting combined orbits are then assessed using \ac{SLR} residuals retrieved using the 'GNSS High Precision Orbit Determination Software Tools' of the German Aerospace Center. An additional analysis of the uncertainty of these orbits is performed using the projection of the orbit uncertainty along the line-of-sight vector from the satellite laser ranging station to the satellite. Results demonstrate that the combined orbit solutions outperform the input orbits across a range of metrics, with the best-performing combined orbits produced using the arithmetic mean method achieving an \ac{SLR} residual root mean-square error of 2 cm in 2022, and 4 cm in 2023, improving upon the next-best input orbit by 20 cm of error. This improvement is achieved without the loss of any volume of data. The combined orbits are additionally less sensitive to poor tracking data, and are therefore more robust than the inputs. These combined orbits show promise as use in gravity field estimation from Swarm kinematic orbits, as an improvement in the input orbits corresponds to an improvement in the gravity field model.

Finally, we explore the advantage of using SLR residuals to re-scale the input orbit noise such that each input is statistically consistent. We find that this technique further improves the quality of the combined orbits created using the inverse variance and variance component estimation techniques.