Machine learning for space situational awareness: a unified framework for initial orbit determination and tracklet association

Optical Space Situational Awareness (SSA) in Low Earth Orbit faces two major bottlenecks: Initial Orbit Determination (IOD) from very short observation arcs and the association of tracklets observed at different epochs. This thesis addresses both problems through Machine Learning, motivated by the growing debris density in LEO and the binding mitigation targets set by ESA's Zero Debris policies for 2030. The work first provides a critical review of the IOD landscape, benchmarking classical algorithms against recent learning-based approaches, establishing a quantitative baseline in the too-short-arc regime where classical methods become ill-conditioned. Building on this review, the thesis develops a neural architecture core for IOD from short arcs produced by low-cost wide-field staring telescopes. Validated via Monte Carlo simulation against classical and state-of-the-art baselines, the proposed approach achieves a median position error of the order of a few kilometers on arcs of 20 seconds or less, outperforming existing methods across the short-arc regime. Sensitivity analyses further characterize performance with respect to loss weighting, angular noise, and observing geometry. The thesis then presents a proof-of-concept matcher for tracklet-to-tracklet association, reframing the problem as feature matching between sets of tracklet descriptors. While intentionally embryonic in scope, this contribution demonstrates the viability of learned, neural-based matching and lays the groundwork for a fully developed matcher architecture. Together, the three pillars of the thesis (comparative review, orbit determination, and learned tracklet association) outline a scalable, learning-based pipeline supporting the densification of SSA infrastructure required to meet upcoming debris-mitigation goals.