A Deep Learning Method for Automatic Coronal Mass Ejection Feature Identification

Abstract Coronal mass ejections (CMEs), intense solar eruptive phenomena, are the primary drivers of extreme space weather storms on Earth. As space activities become increasingly frequent and infrastructure more reliant on space-based systems, the rapid and accurate detection and tracking of CMEs is critical. Here, we present a deep learning–based algorithm for automated CME feature extraction, comprising four key stages: image preprocessing, segmentation, tracking, and feature extraction. Initially, Solar and Heliospheric Observatory/Large Angle and Spectrometric Coronagraph C2 images undergo basic preprocessing. A segmentation model that integrates convolutional neural networks with transformer architectures is then employed to achieve pixel-level structural segmentation of CMEs. Subsequently, CME events in time series are tracked using Intersection over Union (IoU), Kalman filtering, and the Hungarian algorithm. Finally, critical physical parameters, including onset time, linear velocity, central position angle, and angular width, are extracted. Quantitative evaluations demonstrate that the proposed method performs well in metrics such as pixel accuracy, recall, F 1 score, and IoU. Qualitative analyses further reveal the capability of the algorithm to rapidly and precisely identify and track CMEs, including faint early-stage features. This approach is well suited for real-time CME warning and dynamic forecasting, offering robust data support for assessing the potential impact of CMEs on Earth.