Automated detection of coronal mass ejections with few-shot learning

Coronal mass ejections (CMEs) are a primary driver of severe space weather impacts, yet accurate delineation of CME structure in coronagraph images remains challenging. We present Segment Anything Model (SAM)–low-rank adaptation (LoRA), an efficient CME segmentation framework that employs the SAM to solar data via LoRA. The approach employs a multichannel preprocessing scheme that combines base- and running-difference coronagraph images to enhance CME morphology and localization. We compile a new, manually annotated Large Angle and Spectrometric Coronagraph dataset with CME-only and mixed CME/non-CME configurations. On CME-only data, SAM-LoRA delivers segmentation quality comparable to state-of-the-art methods while requiring substantially fewer labeled examples. On mixed data, image-level labels derived from the predicted masks (empty versus nonempty) enable high-quality CME presence detection. The parameter-efficient design reduces computational and annotation costs while retaining SAM’s strong priors. These results indicate that adapting foundation models is a promising path toward reliable CME detection and segmentation and provides a basis for extensible, multitask pipelines for tracking, parameter extraction, and forecasting in heliophysics.