Research On Feature Extraction And Classification Technology Of Hyperspectral Image

Promoted by the continuous development of imaging spectrometer technology,the acquisition technology of hyperspectral images is also developing day by day.The research on hyperspectral images with abundant spatial and spectral information at the same time has also received extensive attention.The feature extraction and classification technology of hyperspectral image is an important field of hyperspectral image research.However,hyperspectral images also have information redundancy.If they are directly used for classification,it will bring about multiple problems such as "synonyms spectrum" and "dimension disaster".This thesis studies the hyperspectral image feature extraction and classification technology from the perspectives of traditional spectral feature extraction,space-spectrum joint classification,and space-spectrum feature extraction and classification under the framework of deep learning.The main work and innovations of this thesis are as follows:(1)Aiming at the traditional feature extraction methods based on spectral features,firstly,the traditional feature extraction and dimensionality reduction methods of hyperspectral images are introduced.Then,the related methods of hyperspectral image feature extraction based on manifold learning are introduced and focused on.Based on the characteristics of the spectral image,the LPP algorithm in manifold learning is improved from the two aspects of introducing discriminant information and optimizing the weight matrix,forming an improved discriminant locality preserving projection algorithm(IDLPP).Finally,the relevant parameters are determined through experiments,which proves the effectiveness of the algorithm.(2)Aiming at the space-spectrum joint feature extraction and classification,the space-spectrum joint classification is realized from the three stages of pre-processing,classification process,and post-processing,and the SVM-SSCK-GF space-spectrum joint feature extraction framework is realized.First,based on the principle that pixels with similar spectrum and spatially adjacent pixels are more likely to belong to the same class of ground object,the original hyperspectral image is reconstructed from space to spectrum,and the spatial information is initially integrated into the hyperspectral image.Then according to the reconstruction image,the original SVM classifier is transformed to construct the space-spectrum composite kernel(SSCK),so that the space-spectrum information is used in the classification process of the classifier.Finally,the guided filtering is used to optimize the classification result map to further improve the classification accuracy.(3)Research on the realization of space-spectrum joint feature extraction and classification method based on deep learning framework.First,some basic theories of deep learning summary are presented and summarized,and some current commonly used networks in deep learning and techniques to improve network performance are introduced,And according to the characteristics of the 3D data of hyperspectral images,some of these techniques are integrated to form a 3D multi-scale residual block with attention mechanism for feature extraction,and the 3D-MRCNN network is formed with this structure as the main body of the network,and finally the network parameters are determined through experiments.Experiments have proved that under appropriate parameters,3D-MRCNN has a higher accuracy rate than other 3D-CNN structures,and requires fewer parameters to be trained.(4)Designed and developed a graphical software for hyperspectral image classification,realized the related algorithms proposed and improved in this thesis,and integrated them into the software together with some traditional classification methods.