Identifying Moving Objects in Astronomical Surveys Using Artificial Intelligence

Moving objects appear as streaks of light in astronomical images due to their apparent motion relative to background stars. As surveys increase their sky coverage and depth, finding and characterising these streaks has become an increasingly important challenge to the astronomical community. Among natural objects that produce streaks, asteroids and comets are central to understanding the formation and evolution of our Solar System. If their orbits pass close to Earth, they can also become a social concern by posing a collision threat. To mitigate this scenario, one of the priorities of planetary defence is to obtain new observations of both known and unknown near-Earth asteroids (NEAs). Additional detections can refine our understanding of the orbits of these objects and improve our assessment of impact risk. This thesis presents a suite of methods based on artificial intelligence to detect and characterise streaks across multiple wide-field surveys. In the first part of my PhD, I develop a pipeline that integrates a convolutional neural network (CNN) to find asteroid streaks in data from the VLT Survey Telescope (VST). My pipeline represents the first dedicated tool for systematic asteroid discovery in VST data and covers the signal-to-noise ratio range of 3â 20. It achieves performance comparable to other state-of-the-art pipelines and shows a strong potential to make new VST discoveries and precoveries, especially in the fainter end of the distribution, which remains largely unexplored. The second part of the thesis builds on the methods developed in the first part to minimise reliance on human intervention in streak-detection processes. I apply a version of my pipeline for NEA detection to data from the Zwicky Transient Facility (ZTF), confirming that my algorithms can be used across surveys. I demonstrate that my pipeline achieves precision comparable to that of the human scanners who usually vet ZTF detections and that it can find additional objects not flagged by the existing approach. I also prove that using synthetic streaks during CNN training does not degrade performance. This result is key to validating the future use of synthetic asteroid streaks, which are needed when insufficient real streaks are available or to explore streak morphologies that have not yet been detected. The final part of my thesis focuses on extracting physical information from those streaks. I develop a method for photometric analysis of asteroid streaks using data from the Euclid space telescope. My work is the first to ever use Euclid photometry to determine asteroid rotation periods, which can provide insights into the collisional evolution and physical characteristics of these objects. I analyse 23 000 appeareances of 2321 known asteroids and extract extremely dense light curves, filling in a gap in the literature: at Euclid magnitudes, between 20 and 24, most of the available photometry is sparse. I combine a Lombâ Scargle approach with a Markov chain Mon