The Aurora Image And Aurora Sequence Classification Based On Convolution Neural Network

Aurora is a natural light phenomenon,which often appers in high latitude region and is caused by the collision between the high-energy charged particle flow from the earth's magnetosphere or outer space and atmospheric molecules or atoms.The study of the auroral morphology and its evolution can obtain a lot of information about the solar wind–magnetosphere interaction and geomagnetic physics,which is helpful to understand the influence of the sun to the earth and the change rule of the space weather.This paper studies the aurora sequence and the static image.It mainly focuses on two aspects of aurora image research:One is the classification of the aurora static images based on multichannel fusion and convolutional neural network;Another is the classification of the aurora sequence based on dynamic pool image network.A classification method based on multi-channel fusion feature and convolution neural network is proposed.The traditional feature is to specify some characteristics of the extracted image which is more or less loss of some useful information of the original aurora image,such as texture feature,global features and local features,resulting in a poor classification effect.Using multi-channel fusion technology to fuse the original image information and the designated effective traditional feature information into a fusion image.Using pre-training convolution neural network to automatic extract effective feature information of fused image.The multi-channel characteristics and deep learning are combined to obtain the characteristics of high efficient representation of aurora images.A method based on dynamic image network is proposed for aurora sequence classification.Auroras are constantly changing over time,and the aurora image sequence is the carrier of auroral dynamic information,which contains many rich information of aurora.A dynamic image network is used to describe the sequence of aurora images.First of all,by using the dynamic ranking pooling function to characterize the aurora image sequences,the aurora dynamic images are obtained.Then,the characteristics of dynamic image feature training are extracted with the convolution neural network.Finally,the features are classified.The results on the auroral observation data from the Chinese Yellow River Station demonstrate the validity and reliability of the proposed two methods.