Unsupervised Feature Learning for Temporal Segmentation of Auroral Image Sequences

Abstract

Optical observation of aurora can obtain extensive information about magnetosphere and solar-terrestrial electromagnetic activities. Temporal segmentation of auroral image sequences is the basis for detection and retrieval of auroral events and further analysis of statistical features (such as occurrence time, lifetime and periodic, etc.). Different from other classification tasks, there is no definite semantic labels in the temporal segmentation of auroral image sequences. In this study, we propose an unsupervised learning method to characterize the appearance and dynamic features of auroral image sequences. Then, the clustering algorithm is used to segment long sequences. Experimental results on auroral image sequences obtained from all-sky image observations at Yellow River Station (YRS), show that the segmentation result is consistent with that of human visual perception.