Deep learning–based detection and classification of small bright sources on ISS images from the Cassini mission and application to Saturn’s outer magnetosphere

ABSTRACT Recent advancements in artificial intelligence, particularly in deep neural networks, have opened new horizons for space exploration, especially in the field of astrometry. In this study, we address the large-scale categorization of bright astronomical sources in the Cassini Imaging Science Subsystem (ISS) image data set (Porco et al.; NASA Planetary Data System, PDS). Specifically, we train the state-of-the-art YOLO (You Only Look Once) network to detect and classify small sources such as stars, satellites, and cosmic rays in ISS images. To compensate for the lack of labelled data, we first develop an automated ground truth generation process using a custom detection and classification algorithm based on traditional image processing. The conducted quantitative and qualitative analyses of the network’s predictions demonstrate that it can reliably identify key features for the detection and classification of bright sources. Moreover, it is able to correct some errors in the automatically generated ground truth data such as missed detections or misclassifications. Finally, we analyse the temporal variations of detected cosmic rays in Saturn’s outer magnetosphere and compare them with Neutron monitor and sunspot number count rates. The good agreement between the expected cosmic ray behaviour outside Saturn’s magnetosphere and the particles detected by our network supports the validity of our approach. This result highlights the potential of our method to support deeper scientific investigations aiming for a better understanding of the Saturnian system.