Type: Poster
Session: Virtual (Rubin Research Bytes)
Author: Jack Patterson

This work will also be presented as a talk in the Streaks, Flares, Glints, and Rubin Science session.

Abstract: We assess the effectiveness of deep neural networks for detecting satellite streaks in wide-field survey data acquired in 2022-24, comparable in depth to that of the upcoming Legacy Survey of Space and Time (LSST). Eleven distinct methodologies for detecting and localizing streaks in CFHT MegaCam images of on-ecliptic fields with limiting magnitudes m_r=26.5 were evaluated, ranging from traditional approaches such as Hough/Radon Transforms and ASTRiDE as a baseline, to several state-of-the-art deep neural network architectures. A standardized execution and testing framework was developed to ensure uniformity across evaluations, encompassing image manipulation and model output processing. Leveraging a labeled dataset of images containing streaks, the performance of each approach was quantitatively evaluated. Examination of the results revealed challenges stemming from significant variability in streak width, brightness levels, and time-varying morphology within individual satellite passes, which resulted in detected streaks of incorrect widths, streak angle misalignments, and high false positive rates. These challenges highlight the need for the development of more robust methodologies to address them effectively. Moving forward, we intend to refine the most promising approaches identified during our evaluation to improve the effectiveness of streak detection and localization.