Research On The Classification And Cutting Method Of All-sky Aurora Images Based On Machine Learning

Aurora is a colorful and luminous phenomenon that appears in the high latitudes of the Earth's north and south poles.It is the most concentrated expression of the physical and chemical processes of the polar regions(especially the magnetosphere-ionosphere interaction).Through long-term observation and research on aurora,it is helpful to analyze the mechanism of aurora and understand the rapid and complex evolution of aurora,which is of great significance for studying the changes of the Earth's magnetic field and the electromagnetic activity of the Earth's space.millions of all-sky aurora images taken by optical imagers in China per year,but only the way of observing massive aurora images with the human eye is particularly burdensome.Therefore,how to efficiently and accurately classify and segment these massive aurora images has always been an important topic in the field of aurora research.In this thesis,an auroral image automatic classification method and a weakly supervised auroral image automatic segmentation strategy are proposed for the characteristics of auroral data.In the study of automatic classification method of Aurora image,this thesis combines Spatial Transformer Networks(STN)with Convolutional Neural Network(CNN),In order to adaptively select the aurora of interest for the network model during training and feature learning,the classification model is supervised and optimized by the Large-margin softmax(L-softmax)loss function,so that encourages intra-class compactness and inter-class separability between learned aurora features.Altogether 8000 all-sky auroral images captured at the Arctic Yellow River Station during years 2003-2007 are classification in this thesis.The experimental results are characterized by feature visualization and accuracy comparison.The proposed method has stronger representation ability and higher classification accuracy.In the study of the weakly supervised auroral image automatic segmentation strategy,this thesis uses the full convolutional neural network as the basic network architecture.The traditional machine learning Seed Region Growth(SRG)algorithm generates data labels,which greatly solves the pressure of deep learning and segmentation network manual labeling data.Specifically,we first train an initial segmentation model,called Modell,with the simple single-arc auroral images(only one distinct auroral arc)and their segmentation labels obtained by the seeded region growing method.Then,an enhanced segmentation model,called Model2,is learned with the hot-spot auroral images and their segmentation labels based on the Model1 as well as the complicated multi-arc images(two or more arcs coexist)and their segmentation labels obtained by the seeded region growing method.Finally,the segmentation results are further optimized by the fully connected conditional random field model.Altogether 2715 auroral images captured at the Arctic Yellow River Station during years 2003-2007 are segmented in this thesis.The experimental results are compared quantitatively and qualitatively with the state-of-the-art results and manual labels,of which the intersection-over-union value between the ground truth and the segmented results is as much as 60%,which proves the effectiveness of the proposed method.