Evaluation of the introduction of a neural network into the objects detection process on astronomical images

Optical observations of in-orbit satellites and space debris have been rapidly increasing in relevance in the research field of Space Surveillance. Frequent observations allow for monitoring the objects' trajectories, addressing the concerning threats posed by the incessant growth of the in-orbit population. The Sapienza Space Systems and Space Surveillance Laboratory (S5Lab) research team has extensive experience in astronomical observations, including strategies development, scheduling, observation performance, and data analysis. The astrometric and photometric analysis of astronomical images is usually based on deterministic algorithms, but recent advances in computer vision have paved the way for the introduction of new methods that exploit Artificial Intelligence (AI). This approach can either substitute or support the analysis process in specific steps, improving and optimizing it even further. The S5Lab research team has recently developed an algorithm based on AI designed to perform object detection on the frames collected by the observatories managed directly by the S5Lab team. More specifically, a multilayer convolutional neural network has been trained and optimized to process images obtained by mini-SURGE: an observation system with a fixed, large Field of View (FoV) which constantly supports activities aimed at monitoring objects in the Geostationary Earth Orbit (GEO) ring. The data collected from a survey performed throughout the night represent a challenge in terms of processing time, especially for object detection. This paper proposes an object detection process performed via a trained neural network model that can detect and find the pixel coordinates of objects in an image without using any information about their orbital parameters. This method allows for the collection of several measurements per night for every object in the FoV, enabling the creation of an up-to-date catalogue of geostationary objects observable by the system. The outcomes will be presented through a direct comparison with detection methods based on deterministic algorithms, highlighting the potential of AI integration in the analysis of ground-based optical systems data.