PASSIT-CME: A Pixel-annotation-free System for Automatic Segmentation, Inpainting, and Tracking of Coronal Mass Ejections

Coronal mass ejections (CMEs), one of the most significant and intense solar eruptive activities, exert profound impacts on Earth and the interplanetary space environment. Consequently, automatic detection and tracking of CMEs have become crucial tasks for mitigating their impacts. Considering the complexity of manually annotating regions of CME on coronagraph images and the presence of anomalous data, we have developed a new automatic CME tracking system that does not rely on pixel-level annotations and can handle obvious data errors. The proposed system consists of three stages: error area segmentation and inpainting, CME segmentation, and CME tracking. First, we apply a weakly supervised semantic segmentation framework, Token Contrast, to the tasks of CME and error area segmentation. Subsequently, we propose a CME inpainting model, Mask-Aware Recurrent Feature Reasoning, to restore erroneous regions in running-difference coronagraph images. Then, we employ Kalman filters to track the evolution of CMEs and extract their kinematic parameters. All deep learning models in our system are trained on the dataset without pixel-level labels, which can be easily constructed from publicly available CME catalogs. Moreover, by comparison with existing catalogs and methods, we demonstrate that the proposed system is reliable in providing CME initial kinematics, facilitating studies on the origin and propagation of CMEs.