Identification of Megaconstellations in Wide-field Astronomical Images with Machine Learning

Abstract Along with the development of megaconstellation, tens of thousands of artificial satellites are going to be launched and deployed, it poses a serious space sustainability risk and has a serious impact on astronomical observations. For wide-field optical astronomical surveys, the large number of satellites increases the probability that one will enter the field of view and streak-like image with distinct brightness and large size appears, data reduction efficiency is affected and image measurement pipeline may be interrupted. Automatic identification of the appearance of megaconstellation images in survey data can make contributions to optimize the data reduction pipeline and propose new mitigation standards and guidelines. Here an automatic identification pipeline based on machine learning model ShuffleNet V2 is developed, after trained with large amount of raw data, high efficiency is achieved. A trial survey was performed using an optical telescope with 4.8 square degrees field and raw images of 77 nights were obtained. With SExtractor and manual identification, the streak images of low-Earth orbital satellites are selected, and the efficiency of our method is investigated. It is demonstrated that an accuracy rate better than 98% and a recall rate better than 95% are achieved by our framework, and can be aware of the images with satellite signals effectively. Our method presented can be used as an auxiliary tool for reduction pipeline optimization and improving source measurements, and it deserves wide applications in similar tasks.