Image Classification And Dynamic Process Analysis For Dayside Aurora On All-sky Image

Aurora is created by atmospheric atoms and molecules colliding with electronsand protons from outer space when they precipitate into the atmosphere. Theseenergetic particles are composed of those particles discharged in magneto dynamicsprocess of the solar wind-magnetosphere interaction, and some particles from the solarwind. Among geophysical phenomena processing features of polar region, aurora is theonly one can be seen by naked eye. The aurora phenomena provide a convenientprojection of effects from complex and energetic plasma processes of the outermagnetosphere. Much has been learned about the ionosphere and magnetosphere fromauroral events. The spatial structure and temporal evolution of auroral luminosity areascribed to cumulative effects of the solar wind–magnetosphere interaction and thephysics of the magnetosphere–ionosphere interaction. Therefore, the study on auroralappearance and evolvement is helpful to study influence of the sun to earth, and issignificant to acquire information about space weather. Thus this study has greatsignificance for science research and practical application.Certain physical processes in the magnetosphere and ionosphere are responsiblefor the auroral appearance. In earlier studies, there have been several types of auroraidentified, which have turned out to be correlated with specific magnetosphericregimes and dynamic activities. Variations in the solar wind parameters seem to havea strong influence on auroral appearance. The valuable two-dimensionalmorphological information can be acquired by high spatial and temporal resolutions,which enables scientists to successively observe auroral behavior. The auroral studybased on image is attracting more attentions. However, most of current studies are―case study‖which is performed manually, so that annually increasing huge numberof auroral images haven’t been utilized. Provided with so abundant valuable data,tools which automatically analyze aurora are urgently needed, which has beensubjects of many research efforts.The study of applying image processing and pattern recognition techniques inautomatic auroral image analysis is far from complete. Several problems need furtherinvestigation. In this thesis, several good techniques in image processing and patternrecognition are introduced and improved in studying aurora phenomena. Manyrepresentation methods properly characterizing auroral statistic and dynamicinformation are proposed. Automatic analysis on huge number of auroral images is achieved. The experimental results provide morphological proofs for classificationschemes available, and also offer morphological interpretation of auroral types. Thus,the feasibility and effectiveness of techniques of image processing and patternrecognition in studying space science are verified.In sum, the author’s major contributions are outlined as follows:1. By systematically analyzing dayside aurora characters in all-sky imager (ASI)observations and having attempted various methods, a spatial texture basedrepresentation method including features of intensity, shape and texture, was utilizedto characterize ASI images. The combination of the local binary pattern (LBP)operator and a delicately designed block partition scheme achieved capabilities ofrepresenting both global shapes and local textures. The representation method wasused in automatic recognition of four primary categories of discrete dayside aurorausing observations between years2003–2009at the Yellow River Station,Ny-lesund, Svalbard. The supervised classification results on labeled data in2003were in accordance with the results labeled by scientists considering both spectral andmorphological information.2. The occurrence distributions of the four categories were obtained byautomatic classification technologies. Regardless of being discriminated by somephysical properties or morphology, certain auroral types may occur periodically andrepeatedly in the auroral oval, which has been verified in numerous space sciencestudies. The obtained occurrence distributions of four dayside auroral categoriesbased on morphology confirm that there are dominant morphological characteristicsin different regions of the dayside oval. The peak positions of occurrence of the fourauroral types are in accordance with synoptic distribution based on the intensities ofphoto-emissions on three wavelengths (427.8,557.7and630.0nm) along themagnetic meridian during daytime (0300–1500UT/0600–1800MLT). Theexperimental results support that auroral morphology identifiable. The results providemorphological evidence for available classification schemes and offers morphologicalinterpretation of auroral types.3. By now, there is no generally accepted set of auroral types for auroralclassification studies. It’s uncertain that the auroral classification schemes availablereflect the real natural structure of the auroral data. The unsupervised clusteringmethod is utilized in classifying ASI images, which don’t depend on any predefinedclassification. In unsupervised way, dominant morphological characteristicsthroughout dayside oval are found, which thereby test and complement the available classification schemes. Being independent of any prior knowledge or temporalinformation, auroral data exhibit clustering tendency. The regularity in occurrencedistributions of clustering labels verifies the feasible of studying auroral dynamicprocesses depending on morphology. These achievements properly verifyeffectiveness of pattern recognition method in studying aurora phenomena.4. Considering the fluid nature of auroral motions and non-uniformity of motionson different scales, we introduce a fluid flow algorithm and modified theregularization methods to estimate motion fields under dynamic auroral situations.The modified motion estimation method doesn’t depend on brightness constancyassumption and use the matching regularizer for auroral motions on different scale.The modified method can effectively, accurately and robustly capture the auroralappearance structures and estimate auroral motion.5. Based on extracted motion fields, the auroral event can be characterized andautomatic event detection is achieved. Because auroral activities are extremelycomplex, changeful and predictable, we use spatiotemporally statistics of motionvectors to represent auroral motions. By using the representation, two sequences withdifferent lengths can be compared. Automatic multi-event detection is achieved,which support statistical analyzing auroral dynamic process based on numeroussuccessive observation.6. A novel method of modeling dynamic texture (DT) is proposed to characterizeauroral motions. The method synergizes spatial and temporal aspects of DTs, anddoesn’t limit the length of sequence. Its validity is firstly verified in a common DTdataset. The new method can effectively and robustly represent auroral motions,which support segmentation and clustering of auroral events. Thus, by using theproposed method of automatically analyzing auroral event, great amount of auroralactivities can be quickly browsed, inspected and eliminated.