Deep learning-based McIntosh classification of sunspot groups

Abstract

Different McIntosh classes of sunspot groups are associated with the occurrence of different levels flares. Thus, accurately classifying sunspot groups is of great significance for flare prediction. In this paper, a deep learning model named SungDC is proposed for the McIntosh classification of sunspot groups. The SungDC is designed as a single multi-classifier to simultaneously perform the classification of 60 McIntosh classes. An AGCM module is incorporated to enhance its feature extraction capability. An LCFPN neck is designed to mitigate the distortion of sunspot group features, thereby improving the quality of features. A deep learning dataset sourced from SDO/HMI continuous spectral full-disk solar images was built. In addition, a region-level data rotation augmentation technique (RLR) was improved to alleviate the problem of sample imbalance. The experimental results show that the AP, AR, and AF metrics of the SungDC are 0.645, 0.586, and 0.608, respectively. The precisions of the dki, eki, ehc, dkc, ekc, and fkc sunspot groups, which are tightly associated with M- and X-class flares, are 0.905, 0.828, 0.920, 0.710, 0.711, and 0.463, respectively. It demonstrates that the multi-classification challenge posed by sunspot groups can be feasibly addressed by deep learning methodologies. This method can also serve for research on flare prediction.