.png){kind=icon}
.png){kind=icon}
Return
CMEGNets: A Self-Supervised Framework for Coronal Mass Ejection Detection & Region Segmentation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.asr.2026.01.061.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The detection and analysis of coronal mass ejections (CMEs) rely on robust segmentation of LASCO C2 and C3 coronagraph images. We introduce CMEGNets, a novel self-supervised framework that eliminates the need for manual annotations by leveraging contrastive pre-training and iterative pseudo-label generation. First, we pretrain a lightweight backbone using instance discrimination on large volumes of unlabelled LASCO data. Next, we derive statistical pseudo-masks via unsupervised clustering of feature activations, which in turn supervise fine-tuning of a U-Net segmentation head. Finally, a small set of expert-verified masks refines the network in a semi-supervised loop. On the LASCO C2 benchmark, CMEGNets achieves 99% classification accuracy between CME and non-CME frames and a 95% Dice coefficient for CME region segmentation. Moreover, our method reduces the annotation effort by more than 80%. These results demonstrate that CMEGNets effectively produce expert-quality coherent segmentations while avoiding costly manual labelling.
Journal
IF:
2.8
Papers:
1.3K
Citations:
2.0W