Deep Neural Network Closed-loop with Raw Data for Optical Resident Space Object Detection

Abstract Optical survey is an important means for observing resident space objects and space situational awareness. With the application of astronomical techniques and reduction method, wide field of view telescopes have made significant contributions in discovering and identifying resident space objects. However, with the development of modern optical and electronic technology, the detection limit of instruments and infrastructure has been greatly extended, leading to an extensive number of raw images and many more sources in these images. Challenges arise when reducing these data in terms of traditional measurement and calibration. Based on the amount of data, it is particularly feasible and reliable to apply machine learning algorithms. Here an end-to-end deep learning framework is developed, it is trained with a priori information on raw detections and the automatic detection task is performed on the new data acquired. The closed-loop is evaluated based on consecutive CCD images obtained with a dedicated space debris survey telescope. It is demonstrated that our framework can achieve high performance compared with the traditional method, and with data fusion, the efficiency of the system can be improved without changing hardware or deploying new devices. The technique deserves a wider application in many fields of observational astronomy.