Aurora Image Classification Based On Multi-Feature Latent Dirichlet Allocation

Aurora,also known as polar light,is a phenomenon of large-scale discharge that occurs at a high altitude in Arctic and Antarctica,which contains abundant physical meanings.Studying the mechanism of aurora is helpful to understand the Earth's magnetic field activity and the solar-terrestrial electromagnetic activity.With the continuous development of various observation methods,the number of acquired aurora images has increased dramatically.How to classify them according to morphology has become an important research direction.In the early days,aurora images were classified by visual inspection and manual marking and the gradual maturity of computer technology provides the possibility for rapid automatic classification.Several methods of aurora image classification based on computer vision have been put forward such as Fourier transformation,local Gabor features,morphological analysis,feature coding and topic model,but most of these methods are only aimed at the single feature of aurora,which is difficult to make a comprehensive and accurate description of aurora images.In order to solve the problems of poor universality,low accuracy and low efficiency in the traditional algorithm,an aurora image classification method based on multi-feature latent Dirichlet allocation(AI-MFLDA)is proposed in this thesis.The main contents and contributions are as follows:(1)This thesis summarizes the existed methods of aurora image classification,expounds the basic theory,development process and applications of the probabilistic topic model(PTM),introduces some extraction methods of image feature and several commonly used classifiers.(2)This thesis proposes an aurora image classification based on multi-feature latent Dirichlet allocation(AI-MFLDA).Multiple features,such as grayscale,structure and texture,are selected to describe the aurora images and transformed into one-dimensional(1-D)histogram before integration,which enhances the universality of the model.The multi-feature LDA is improved on the basis of the LDA model,using same Dirichlet prior distribution to generate the subject space for different features,which reduces the influence caused by the different feature distribution.In the process of classification,the generative differentiation strategy is utilized to improve the classification effect.(3)According to the classification system provided by China Polar Research Center,this thesis constructs an aurora image dataset and carries out a series of comparative experiments.The experimental results show that compared with traditional methods,the proposed AI-MFLDA can obtain higher classification accuracy(98.2%),while maintaining a lower feature dimension,and has a greater advantage in dealing with complex aurora morphology.Different feature combinations have a significant impact on the classification results,and the correlation parameter sensitivity analysis shows AI-MFLDA can maintain good stability when the number of topics is changed.When dealing with unknown samples,the aurora image classification method based on multi-feature LDA is still effective,which further reflects the universally applicability of the model.This thesis explores a more accurate and efficient method of aurora image classification,which can fully tap the potential information of aurora images.This achievement has some reference value for polar space atmospheric physics research.