Machine Learning using Graph Neural Networks for Star Identification in Celestial Navigation

Contemporary maritime, aviation and space operations depend heavily on Global Navigation Satellite Systems (GNSS) such that deliberate spoofing or jamming, natural phenomena, and even simple occlusion may constitute strategic and safety-critical risks. Celestial navigation offers a resilient alternative for position and orientation determination when satellite-based systems are unavailable, operating independently of external signals and it’s associated vulnerabilities. Although the historical use of sextants is slow and incompatible with modern needs, advances in star-trackers, computer vision, and machine learning present opportunities for developing more sophisticated, data-driven star-identification methods suitable for modern navigation requirements. In this context our research investigates the current state of star-identification methods, wherein an automated system must reliably recognise observed stars, match them to catalogued identities, in order to enable celestial navigation. Although immune to external signal interference associated with GNSS systems, celestial navigation still suffers risks from a degraded input caused by various types of sensor and atmospheric noise. This motivates the search for a star-identification solution that is not only timely but increasingly resilient to the varied interference captured in celestial imagery. Firstly, we address an issue of data-scarcity. Genuine annotated night-sky imagery is extremely limited, often insufficient for training or evaluation of robust algorithms and is prohibitively difficult to acquire in large quantities. This problem is compounded when we con-sider the significant variation in methods and intended use, which would result in these static-datasets not being widely applicable. A method trained to recognise stars up to an apparent magnitude of seven may be ineffective when evaluated on a dataset that only captured stars to a threshold of six. To overcome this, we first develop a synthetic star-field simulation pipeline using open-source planetarium software. This generates photorealistic night-sky images with known ground-truth star identities, annotations and meta-data. These simulated images are configurable by various observation and design parameters and incorporate realistic distortion and atmospheric effects. The end result being a widely configurable tool for generating la-belled night sky datasets that enables more rigorous and large scale training and evaluation of star identification methods. This approach helps bridge the gap between simplified analytical simulations and actual observational data. We then propose SIGNN (Star Identification using Graph Neural Networks), a novel machine-learning (ML) approach for automated star recognition. SIGNN represents the celestial sphere of observable stars via a graph. Each node represents a star, encoding its relative brightness, whilst edges encode and link stars based on their angular separations. This graph representation mirrors the intuitive patterns used in manual celestial navigation (constellations and star angles) but allows a data-driven learning of more informative and discriminative pat-terns. Graph attention layers are able to weight the most informative star relationships while diminishing the contribution of unreliable and corrupted observations. We show that SIGNN outperforms both traditional geometric and modern machine learning baselines, including grid-pattern matching, and convolutional neural network approaches RPNet and Spider-CNN under false-star, dropped-star, magnitude and positional noise. Finally, we examine the difficulty posed by stars at the periphery of captured images, which experience systematically lower identification accuracy due to incomplete local star patterns. To address this, we introduce an enhanced model variant, SIGNN (EOI), which incorporates specialised training with artificially clipped star patterns. Both SIGNN and SIGNN (EOI) are evaluated using the photo-realistic data pipeline, as well as under controlled noise exposures. The aim of this research is to enable reliable, GNSS-independent celestial navigation by developing an automated, noise-tolerant star identification pipeline that remains accurate under realistic sensor, atmospheric and scene degradation. This thesis advances star identification by introducing a configurable, photo-realistic para-metric data generation and noise simulation pipeline that enables fair and reproducible benchmarking under matched observing conditions, and makes robustness to false, dropped, positional and magnitude noise an explicit design objective. We demonstrate that graph-based representations of stellar point patterns improve robustness to realistic noise and edge-of-image degradation, moving away from hand-designed rules to learned, graph-structured features. The approach is broadly transferable to point-pattern identification tasks, and in particular to satellite, UAV and spacecraft localisation and attitude determination. Beyond astronomy, the findings show how graph structures can represent complex spatial relationships and how synthetic training data with realistic noise can be used to create models that remain reliable under varied and severe degradation. By applying graph attention networks to a real-world navigational context, this work expands their use beyond more typical domains such as social networks and chemistry, and, through the use of open-source software and publicly available code and data-generation tools, enables adaptation of the methodology to other critical systems that depend on identifying noisy point patterns. In particular, SIGNN is designed with GNSS and GPS vulnerability in mind, supporting navigation in degraded or denied satellite-signal environments that arise in space, defence and disaster recovery operations.