Solar Active Regions Detection and Tracking Based on Deep Learning

Abstract Solar active regions serve as the primary energy sources of various solar activities, directly impacting the terrestrial environment. Therefore, precise detection and tracking of active regions are crucial for space weather monitoring and forecasting. In this study, a total of 4,532 HMI and DMI longitudinal magnetograms are selected for building the dataset, including the training set, validating set, and eight testing sets. They represent different observation instruments, different numbers of activity regions, and different time intervals. A new deep learning method, ReDetGraphTracker, is proposed for detecting and tracking the active regions in full-disk magnetograms. The cooperative modules, especially the re-detection module, NSA Kalman filter, and the splitter module, better solve the problems of missing detection, discontinuous trajectory, drifting tracking bounding box, and ID change. The evaluation metrics, IDF1, MOTA, MOTP, IDs, and FPS, for the testing sets with 24h interval on average are 74.3%, 75.0%, 0.130, 14.5, and 13.7, respectively. With the decreasing intervals, the metrics become better and better. The experimental results show that ReDetGraphTracker has a good performance in detecting and tracking active regions, especially capturing an active region as early as possible and terminating tracking in real-time. It can well deal with the active regions whatever evolve drastically or with weak magnetic field strengths.