Magnetic Type Classification in Sunspot Group Based on Semi-supervised Learning and Knowledge Distillation

Abstract

Sunspot group, known as active solar regions, is the main sources of solar storms. The morphological and magnetic characteristics of solar active regions play a very important role in solar storm forecasting and is well described by Mount Wilson Sunspot Classification Scheme. The development of convolutional neural network methods in the field of image processing makes efficient magnetic type classification possible. In this paper, we propose a method based on semi-supervised learning and knowledge distillation for magnetic type classification in sunspot group. On the sunspot magnetic type classification dataset, our method achieves 95.14% total classification accuracy, 97.4%, 94.43% and 85.71% F1-scores of Alpha, Beta, and Beta-x types respectively.