An aurora sequence classification method based on deep learning is proposed. It combines the rich spatial domain information and the sequence information corresponding to the advantages of convolutional neural network (CNN) features and long short-term memory (LSTM) network. In addition, aurora attributes employed as feedback constraints to the CNN make features more suitable for aurora images. Supervised aurora sequence classification and unsupervised aurora event detection are performed on the Chinese Yellow River Station All-Sky Imager (ASI) dataset. The experiment shows that our method can characterize aurora sequences effectively and can be able to implement automatic classification for massive aurora sequences.