Aurora Image Retrieval And Fine-grained Classification Of Arc-shaped Aurora Images Based On Convolutional Neural Networks

Aurora is the luminous phenomenon that the charged particles from the sun impact the earth's atmosphere.Optical observation is an important means to study the auroras.It is of great research value to retrieve the interested auroral images from these massive auroral data for exploring the information of magnetospheric changes and the electromagnetic activities between the sun and the earth.In the aurora,the arc aurora appears most frequently.It is very important to subdivide the arc Aurora into single arc and multi arc.The main reason for the research of Aurora image from the initial manual processing to the present computer automatic recognition is that the artificial representation image depends on the experience of researchers,has strong subjectivity,and is difficult to fully characterize the aurora image.Convolutional Neural Networks(CNN)is a kind of deep learning algorithm for machine independent hierarchical learning of Aurora image features.Compared with manual processing,it simplifies the process of feature extraction and can more fully represent the aurora image.In this paper,the aurora image is effectively represented based on CNN,and the aurora image retrieval and fine-grained classification of arc aurora image are realized.The main results of this paper are as follows:Aiming at all sky Aurora image,this paper proposes an aurora image retrieval system based on VGG16 feature fusion by fusing the local and global features of the aurora image.In this paper,the retrieval experiment of the whole sky Aurora image data set marked by the Arctic Yellow River Station from 2003 to 2004 is carried out.The mean Average Precision(mAP)is 68.9%.By comparing with three commonly used depth models,Alexnet,Inception-v4 and VGG16,it is proved that the method proposed in this paper can improve the effect of Aurora image retrieval.In addition,in order to intuitively display the retrieval results of large-scale Aurora data,part of the 2003-2004 all sky Aurora image and the 2004-2009 unmarked all sky Aurora image data set are used as the query image and retrieval database respectively.The retrieval results show that the method in this paper improves the visual matching degree of the aurora image.For all sky arc aurora image,this paper combines convolution attention module with VGG16 depth model,and proposes a fine-grained classification method for arc aurora image.In this paper,we use the all sky arc aurora images from 2003 to 2004.The accuracy of single arc and multi arc is 100%and 94.5%respectively.It is proved that the model proposed in this paper can pay attention to the effective area of the auroral arc and learn from it.In addition,the occurrence time distribution of two kinds of auroral arcs is studied.