Classification Of Aurora Static Images And Analysis Of Aurora Dynamic Process

Aurora is an important window for scientists to study the atmospheric of physics.Based on the research of aurora morphology and its dynamic process, large amounts ofinformation about the magnetosphere and solar-terrestrial space electromagneticactivities can be obtained. As the digital all-sky imager (ASI) system were put into useat Arctic Yellow River Station, millions of aurora images are collected and storedannually, which provide a lot of very important data for aurora research. For thesemassive aurora images, how to analysis them and their occuring mechanism quickly andeffectively is a hot topic today, which has great significance of research and applicationvalue to analyse aurora morphology and evolution mechanism. The thesis carried ontwo parts: classification for static aurora images and analysis for aurora dynamicprocess.On the study of calssification for the static aurora images, reducing the highdimensions of feature vectors is involved unavoidably. Therefore, this thesis putsforword an algorithm based on discriminative locality alignment with gaussian processlatent variable model (DLA-GPLVM) for feature dimension reduction of the auroraimages. The aurora data set with BOW features and two public small sampl databaseswith high dimension are contained for the experiments. The results on three databasesdemonstrate that the novel menthod DLA-GPLVM is very effective for reducing thedimension on the data set with high dimension and small sample. This novel metnodcan deal with the dimension reduction for small sample with high-dimensional.However, it is not only time consuming but also poor effective on dimensionalityreduction of high-dimensinal in the field of big data.According to the characteristics of the large aurora data, biological inspiredfeatures (BIFs) with sparse representation using manifold learning are proposed, whichcan improve the accuracy of aurora images calssification further. BIFs can be usedrepresent the visual characteristics of the cerebral cortex by simulating the human visualattention model. Then, discriminative locality alignment (DLA) is used to datareduction for the high-dimensional BIFs. Support vector machine (SVM) and nearestneihbor (NN) are used to classify the static aurora images respectively. Theexperimental results proved that BIFs-DLA has higher classification accuracy, lowercomputing complexity and stronger robustness than methods available. It is not enough to study the static characteristics for auroras with varied andcomplex motion process. Therefore, two methods are proposed to study the dynamicprocess of aurora. Arc aurora sequences are analysed based on lattice boltzmann method(LBM) and poleward moving auroral forms (PMAFs) are discussed based on the gravitycenter of one image respectively. Considering the aurora as non-rigid, the motion offluid particles is used to simulate the partilces movement of the aurora dynamic process,which can obtain the macroscopic parameters of microscopic particles. Throughstatistical macroscopic parameters, the dynamic process of arc aurora sequences can beanalysed further. Analysing the pseudo-color image or fitting the slope of the center ofgravity in aurora sequence, the PMAFs can be identified quickly. Experiments haveshown that our methods are more effective for analysing the dynamic process of aurorasand provide ideas for the research of occurance aurora mechanism further.