Semi-supervised Classification And Unsupervised Clustering Of All-sky Aurora Images Based On Deep Learning

Aurora is the typical visible atmospheric phenomenon in the high latitude area of the north and south poles of the earth.It can directly reflect the relationship between the change of the earth's magnetic field and solar activity.Through the effective classification of auroral images and the study of the occurrence law of different types of aurora,it can help people to study the scheme of aurora occurrence and the dynamic process of the magnetosphere boundary layer.And it is also of great significance to further study the Earth's magnetosphere structure and the energy coupling process between the Earth and the sun.The long-term aurora observation has captured a large number of auroral images,and the number continues to increase by tens of millions every year.The existing auroral image classification research mainly adopts manual analysis or supervised classification mode,while the temporal and spatial resolution of auroral image collected by different auroral stations and auroral observation equipment are different.Both manual analysis and supervised classification need to label a lot of data by human visual inspection and can not meet the needs of massive auroral image processing.Consequently,how to classify the auroral image efficiently,accurately and reasonably is an important research topic in the field of aurora research.Based on the characteristics of auroral images,this paper introduces a semi-supervised classification of auroral images with only a few labels and an unsupervised clustering of auroral images without labels.For the first time,the deep learning technology is applied to semi-supervised classification and clustering of auroral images.Semi-Supervised Ladder Network is used in this paper to classify all-sky auroral images in a semi-supervised way.The network has good classification accuracy after trained by a large number of unlabeled samples and a small number of labeled samples.In this paper,two groups of experiments are carried out to classify the Chinese Yellow River Station's dayside all-sky auroral images by using the Semi-supervised Ladder Network.Firstly,no auroral images,cloudy images,and auroral images were semi-supervised classified.Secondly group,arc auroral images,drapery auroral images,hot-spot auroral images and radial auroral images are semi-supervised classified.Two groups of experiments in the training data samples with 10%,20%and 30%of the label have high classification accuracy.Through the analysis of the classification accuracy and results of different proportions of labeled samples,the validity of the classification method is proved.An unsupervised all-sky auroral image clustering network(AICNet)based on the deep learning technology is proposed for unsupervised auroral image clustering.After specifying the number of clusters,auroral image features are extracted by deep convolutional auto-encoder.Auroral images with similar features are automatically clustered into one group through the clustering network.AICNet not only avoids the artificial determination of classification scheme,but also does not need to manually label data,which makes the auroral image classification more objective and efficient.4000 dayside all-sky auroral images captured at the Chinese Yellow River Station from 2003 to 2008 were used in this work.By calculating three clustering validity indexes,we can get the optimal classification of aurora,which is to aggregate them into two categories.Subsequently the features of auroral images are visualized by t-SNE dimensionality reduction algorithm,analyzing the occurrence time of auroral images in both clusters,and further picking some samples randomly from the clustering results.All the above prove the validity of the clustering method proposed in this paper.