Abstract: Auroral appearance forms provide remarkable visual and distinguishable features for the study of solar-terrestrial physics. Proper classification is meaningful for studying the relationship between various types of auroral phenomena and the dynamics of the magnetosphere. The selection of auroral classification schemes has been criticized in supervised classification experiments, which require large amounts of labor. Additionally, the accuracy of manual labeling has been questioned. More importantly, the results of supervised classifications cannot be used to verify the correctness of classification schemes. Thus, we should investigate whether existing classification schemes are accurate, as well as identify better classification schemes. To address these issues, a clustering method was used to explore the feature space of aurora data based on an available auroral image characterization method. Nine cluster validation indices were used to select the optimal number of clusters. Six thousand all-sky images, which were randomly selected from observations acquired in the Arctic Yellow River Station in 2003—2004, were clustered using the Ncut algorithm. The results showed that schemes consisting of two and four classes were the most accurate. The two-class schemes had well-separated auroral types, and the distribution of pre-noon and post-noon occurrence peaks can be used to determine whether an aurora may be an arc. In the four-class schemes, although the naked eye failed to find a typical image that can represent all images in this class, the temporal distribution characteristics of these classes were very distinct, which proves that the auroral appearance is identifiable using an unsupervised method.

Key words: Aurora, Clustering, Auroral appearance, All-sky image