Research On Clustering Of Ultraviolet Aurora Oval Images

Aurora is affected by large-scale dynamics such as geomagnetic substorm driven by solar wind,its morphology and morphological transformation are closely related to the coupling effect of solar wind-magnetosphere-ionosphere.Generally,ultraviolet(UV)imager can observe main information of global auroral activities in a relatively complete way.UV auroral oval image clustering can realize categorization of auroral oval and its morphology,which is of great significance for statistical analysis and classification standard establishment of auroral oval and its morphology,as well as for the study of coupling between solar wind and geomagnetic field and its relating dynamic process.In this thesis,a feature-based image clustering concept was chosen to conduct research of categorization of aurora oval and its morphology with Polar UV auroral oval image data.In addition to some conventional processing procedures such as denoising,contrast enhancement and data screening,Graph Cut algorithm was introduced to extract auroral oval morphology from images and geomagnetic coordinate localization of auroral oval was added,which provided a strong support to effectively categorize auroral oval and its morphology in purely image-based manner.With tuned unsupervised deep representation learning model(Mo Co),abstract features of large-scale UV auroral oval images were extracted,advantages of the model over traditional algorithms were then verified through comparison experiments and evaluation.Based on feature sample set that well represents images,experiments of six feature clustering algorithms were carried out,while best algorithm(GMM)was confirmed with the help of clustering quality evaluation metrics.In the end,UV auroral oval image clustering results were obtained by assigning feature labels to their corresponding images.Furthermore,this thesis designed a method for physical rationality assessment of image clustering results,and built a space environment parameter dataset,which verified that strength of influence of important parameters,such as auroral electrojet index,on auroral oval morphology and their correlation degree with clustering results remained relatively consistent,i.e.,clustering results were in accord with scientific perception.With combination of unsupervised deep representation learning model and image feature clustering algorithm,a UV auroral oval image clustering model,Mo Co-GMM,which can effectively mine and utilize abstract image features representing auroral oval morphology,was established,realized objective categorization of auroral oval and its morphology.Achievements of this work provided experimental support for further research on physical process related to aurora,and laid a foundation for the study of UV auroral oval images from Solar wind – Magnetosphere-Ionosphere Link Explorer(SMILE)to be launched in 2024.