Auroral Sequence Classification Based On Deep Feature Fusion And Attribute Constrained CNN Network

The shape of the aurora is closely related to the spatial physical processes such as the interplanetary magnetic field and the solar wind.By rationally classifying the aurora,it is helpful to understand the interaction between the sun and the earth and the changes in the Earth's own magnetosphere.At present,most of the research on aurora classification is based on single-frame images,while the research on aurora sequences is still relatively rare.Therefore,this thesis focuses on the direction of auroral sequence classification,and does two things: one is based on pseudothree-dimensional residual network and improved dense trajectory algorithm for auroral sequence classification;the other is based on attribute-constrained convolutional nerve The network combines the auroral sequence classification of long-and short-term memory networks.1.Aurora sequence classification method based on pseudo three-dimensional residual network and improved Dense Trajectory algorithm.It solves the problem that the traditional manual sequence features are used to characterize the aurora sequence but the recognition accuracy is not high.In the deep learning,the three-dimensional residual network with better effect on the sequence data is extracted to extract the depth feature of the auroral sequence,and the traditional method extracted by IDT algorithm is combined.The sequence features,the two are fused,and the aurora sequence is characterized from multiple aspects.In this thesis,the depth features extracted by the pseudo-three-dimensional convolution network are combined with the traditional sequence features extracted by the IDT algorithm.The support vector machine(SVM)is used to classify the aurora sequences,which significantly improves the classification effect.At the same time,the automatic segmentation of the unlabeled aurora long sequence is completed in combination with the sliding window strategy,and the auroral event detection experiment is carried out,which further proves the effectiveness of the method.2.Auroral sequence classification method based on attribute constraint convolutional neural network combined with long-term and short-term memory networks.In the past,the modeling method based on the classification of auroral sequences first extracted the auroral image manual features,and then used the model such as hidden Markov to model the aurora sequence for sequence classification.However,the manual features of these images are often not good enough,and modeling methods such as Hidden Markov cannot take into account context information in longer sequences.In order to solve these problems,this thesis further introduces the spot structure property by combining the auroral physical properties,and proposes a convolutional neural network based on attribute constraint to extract single-frame image features.The depth feature of this combined attribute is richer than the traditional feature information and has global information.With the local information of the spot,the long-short-term memory network that solves the longterm dependence problem is then introduced to model and classify the auroral sequence,which effectively improves the accuracy of the classification of auroral sequences.This part also combines the sliding window strategy to complete the automatic segmentation of the unlabeled aurora long sequence,and carries out the auroral event detection experiment,which further proves the effectiveness of the method.