Spotting Satellite Streaks with Deep Learning

Hua-Jian Yu and colleagues developed ASA-U-Net, a deep learning model to detect satellite trails in telescope images. Using data from the Mephisto Telescope, the model combines channel attention and multi-scale pooling to spot faint and complex streaks. Tests show ASA-U-Net outperforms standard methods, reducing false detections and improving recall, making it a promising tool for cleaning astronomical survey data.