Automated detection of satellite trails in ground-based observations using U-Net and Hough transform

¹ Department of Astrophysics/IMAPP, Radboud University, PO Box 9010, 6500 GL Nijmegen, The Netherlands
² Department of Mathematics/IMAPP, Radboud University, PO Box 9010, 6500 GL Nijmegen, The Netherlands
³ Department of Astronomy and Inter-University Institute for Data Intensive Astronomy, University of Cape Town, Private Bag X3, Rondebosch 7701, South Africa
⁴ South African Astronomical Observatory, PO Box 9, Observatory, 7935, South Africa
⁵ Leiden Observatory, Leiden University, Postbus 9513, 2300 RA Leiden, The Netherlands
⁶ Department of Physics, University of Oxford, Denys Wilkinson Building, Keble Road, Oxford OX1 3RH, UK

^★ Corresponding author; fiorenzo.stoppa@physics.ox.ac.uk

Received: 25 July 2024
Accepted: 7 November 2024

Abstract

Aims. The expansion of satellite constellations poses a significant challenge to optical ground-based astronomical observations, as satellite trails degrade observational data and compromise research quality. Addressing these challenges requires developing robust detection methods to enhance data processing pipelines, creating a reliable approach for detecting and analyzing satellite trails that can be easily reproduced and applied by other observatories and data processing groups.

Methods. Our method, called ASTA (Automated Satellite Tracking for Astronomy), combined deep learning and computer vision techniques for effective satellite trail detection. It employed a U-Net based deep learning network to initially detect trails, followed by a probabilistic Hough transform to refine the output. ASTA’s U-Net model was trained on a dataset of manually labeled full-field MeerLICHT telescope images prepared using the user-friendly LABKIT annotation tool. This approach ensured high-quality and precise annotations while facilitating quick and efficient data refinements, which streamlined the overall model development process. The thorough annotation process was crucial for the model to effectively learn the characteristics of satellite trails and generalize its detection capabilities to new, unseen data.

Results. The U-Net performance was evaluated on a test set of 20 000 image patches, both with and without satellite trails, achieving approximately 0.94 precision and 0.94 recall at the selected threshold. For each detected satellite, ASTA demonstrated a high detection efficiency, recovering approximately 97% of the pixels in the trails, resulting in a False Negative Rate (FNR) of only 0.03. When applied to around 200 000 full-field MeerLICHT images focusing on Geostationary (GEO) and Geosynchronous (GES) satellites, ASTA identified 1742 trails −19.1% of the detected trails – that could not be matched to any objects in public satellite catalogs. This indicates the potential discovery of previously uncatalogued satellites or debris, confirming ASTA’s effectiveness in both identifying known satellites and uncovering new objects.