Using Deep Learning to Identify Artificial Satellite Trails in Multi-band Photometric Astronomical Images

This research addresses the growing challenge of artificial satellite trail interference in ground-based astronomical observations by developing an efficient deep learning identification method. With the proliferation of satellite constellations in low Earth orbit, accurate detection of satellite trails has become crucial for preserving astronomical data quality. Using multi-band photometric survey observational data from the $Multi-channel$ $Photometric$ $Survey$ $Telescope$ (Mephisto) of Yunnan University , we constructed a specialized dataset of astronomical images containing satellite trails. We propose a novel ASA-U-Net model that integrates atrous spatial pyramid pooling with channel attention mechanisms into the U-Net architecture to effectively capture sparse satellite trail features that traditional semantic segmentation models often miss during downsampling. The model was implemented and validated on actual telescope data, demonstrating superior performance in end-to-end detection and marking of satellite trails compared to traditional methods. This approach significantly improves data processing precision without requiring manual parameter adjustments, making it suitable for processing massive nightly survey data and enhancing the quality of astronomical data products.