Unsupervised Change Detection for Auroral Activity in Ground-based All-Sky Imager Observation

Aurora phenomenon is the most visible manifestation of space weather and play a crucial role in geoscientific studies of near-Earth space processes. Segmenting auroral image sequences from large vast observational data is vital for statistical study and determining the lifetime of auroral events. In particular, discovering auroral motion features from big data requires automatic classification and retrieval of auroral motion patterns. Using individual images or randomly segmented fixed-length subsequences as training samples is not appropriate. For untrimmed auroral videos of any length, this paper proposes an unsupervised framework to break down long sequences into shorter temporal units without predefined classes. Sparse and low-rank decomposition theory is used to process auroral morphological and motion information at both pixel level and sequence level, which ensures that both macroscopic variations in auroral sequences and local key structures in auroral images can be taken into account in calculating the difference between consecutive sequences. Experiments conducted on all-sky image sequences obtained at Yellow River Station show that our method achieves high performance for auroral activity change detection. The proposed approach provides an effective tool for monitoring auroral activities and parsing a continuous stream of auroral activity into meaningful events, which enables the statistical study of auroral behavior based on big data.