Automatic Detection and Tracking of Coronal Mass Ejections Based on a Self-supervised Pretraining Model

Coronal mass ejections (CMEs) are violent solar eruptions that can trigger geomagnetic storms when they impact Earth’s magnetosphere, highlighting the critical need for accurate space weather forecasting. Manual CME catalogs are time-consuming, whereas existing automated methods still struggle to detect weak CMEs and to distinguish them from complex background structures. We propose an automated CME detection and tracking framework that leverages a self-supervised masked autoencoder (MAE). We pretrain the MAE on large-scale unlabeled LASCO C2 running-difference images so that it learns informative latent representations of CME structures without human annotation. The pretrained MAE encoder, with its weights frozen, then serves as a representation extractor for the downstream detection task, where a principal component analysis network (PCANet) module identifies candidate regions and a Markov random field (MRF) model refines their boundaries by enforcing spatial coherence and suppressing noise. Following segmentation, CME regions are associated across consecutive frames to enable tracking and the derivation of key CME parameters, including the start time ( T start ), central position angle (CPA), angular width, and propagation speed. On the 2012 LASCO C2 dataset, the framework shows good agreement with the Coordinated Data Analysis Workshop (CDAW) catalog and yields accurate CME-parameter estimates. A continuous three-month blind test on unseen LASCO C2 observations from 2012 September to November further shows stable event-level detection and parameter estimation under long-duration observational conditions. Ablation studies further support that the improvements in detection accuracy and temporal stability are associated with the self-supervised representations and the joint use of the PCANet and MRF modules.