Hyperspectral Image Clustering Algorithm Based On Sparse Subspace Clustering

With the research and development of hyperspectral imaging technology,hyperspectral imaging has been noted extensively,and has gradually diversified in application fields.Combining the spatial dimension and spectral dimension of the images effectively is the contribution of hyperspectral images technology,and which realizes the “integration of image and spectral”.Hyperspectral images have rich spatial information and spectral information compared with RGB images and multispectral images,which can show more ground object information accurately.As one of the main research areas of hyperspectral image data analysis,there are many challenges for hyperspectral images clustering such as the high dimension and large amount of data.In view of the large amount of spectral information,complex spatial structure,and large amount of data in hyperspectral images,two methods based on spectral clustering and sparse subspace clustering are proposed in the thesis,fusing the spatial and spectral information of hyperspectral images.The main tasks of the thesis are as follows:Firstly,the relevant theories and data characteristics of hyperspectral images are expounded,and hyperspectral image data analysis technology,especially the development process and research status of cluster analysis technology are briefly described,and the research background and significance of the thesis are explained.Second,the related theory and implementation process of spectral clustering and sparse subspace clustering are summarized,and the Fréchet distance algorithm is briefly described,which provides a theoretical basis for the proposed algorithms.Several clustering algorithms and the clustering result evaluation index for hyperspectral images are introduced.Thirdly,in order to improve the clustering accuracy and applicability of spectral clustering,a spectral clustering algorithm based on Fréchet distance（called FSC）is proposed in the thesis.Spectral clustering algorithms are usually based on the assumption of manifold data,that is,two adjacent data points are assumed to have the same cluster label in a high-density region of a low-dimensional data manifold.However,for high-dimensional sparse data such as hyperspectral images,the nearest neighbors may actually be far apart,so the clustering accuracy is greatly reduced.To this end,in order to fully develop the spatial spectral information for hyperspectral images data,a similarity matrix of hyperspectral images through the Fréchet distance algorithm is reconstructed,then the reconstructed affinity matrix is applied to spectral clustering.Using the Fréchet distance to measure the similarity of data feature dimensions,which considers the high correlation of pixel spectral bands,and fully utilizes the spectral information of hyperspectral images,thereby expanding the applicability of spectral clustering algorithms.FSC is not only suitable for data with clear low-dimensional manifold structure,but also for high-dimensional or sparse data,such as hyperspectral images.Finally,to overcome the simplicity of sparse subspace clustering in the construction of sparse representation,and improve the utilization of spatial spectral information of hyperspectral images,in this thesis,a weighted sparse subspace clustering algorithm based on Fréchet distance（A Sparse Subspace Clustering Algorithhim Based on Fréchet Distance,called FSSC）is proposed.The using of spectral and spatial information of hyperspectral images of sparse subspace clustering is not efficient in the process of constructing similarity matrix based on single sparse representation.So in this thesis the Fréchet distance algorithm is introduced to measure the similarity between the spectral curves of pixels,and a spectrally weighted sparse subspace clustering model based on the sparse representation matrix is established,in order to solve a more accurate affinity matrix.A large number of experiments and results analysis on three classic hyperspectral image datasets（include Indian pines,Salinas and KSC）are conducted.The experimental results show that the two hyperspectral images clustering algorithms proposed,a spectral clustering algorithm based on Fréchet distance and sparse subspace clustering algorithm based on Fréchet distance can obtain the higher clustering accuracy than traditiona algorithms.