Multi-label Learning for Detection of CME-Associated Phenomena

Coronal mass ejections (CMEs) are considered as one of the driving sources of space weather. They are usually associated with many physical phenomena, e.g. flares, coronal dimmings, and sigmoids. To detect these phenomena, traditional supervised-learning methods assumed that at most one event occurred in a CME; therefore each CME instance is associated with a single label and the phenomenon is processed in isolation. This simplifying assumption does not fit well, as CMEs might have multiple events simultaneously. We propose to detect multiple CME-associated events by multi-label learning methods. With the data available from the Atmospheric Imaging Assembly (AIA) and the Large Angle and Spectrometric Coronagraph (LASCO), texture features representing the events are extracted from all of the associated and not-associated CMEs and converted into feature vectors for multi-label learning use. Then a function is learned to predict the proper label sets for CMEs, such that eight events, i.e. coronal dimming, coronal hole, coronal jet, coronal wave, filament, filament eruption, flare, and sigmoid, are detected explicitly. To test the proposed detection algorithm, we adopt the four-fold cross-validation strategy on a set of 551 labeled CMEs from AIA. Experimental results demonstrate the good performance of the multi-label classification methods in terms of test error.