Research On Hyperspectral Image Classification Combing High-order Information And Sparse Subspace Clustering

With the development of hyperspectral imaging technology and the progress of computer science,the computer-based hyperspectral images interpretation method has subverted the traditional geological work method.Geologists can obtain and analyze the relevant data of the investigation area at home,which greatly improves the work efficiency.However,the rich spectral information in hyperspectral images can not only better explore the intrinsic structural features of ground objects,but also inevitably lead to the high dimension of hyperspectral data,which makes the common clustering technology unable to focus on the effective spectral features and affects the classification effect of hyperspectral images.The sparse subspace clustering algorithm,as one of the methods suitable for highdimensional data,has been successfully applied to the task of hyperspectral images classification,and has achieved some results.Its core is to design a representation model that can reveal the real subspace structure of high-dimensional data,so as to achieve accurate subspace clustering.However,there are still some problems in the specific application of hyperspectral image classification,such as high time complexity,excessive noise interference and limited classification accuracy.Based on the above analysis,aiming at the problems existing in sparse subspace clustering hyperspectral images model,the following research works are carried out in this paper:(1)This paper proposes a fast high-order sparse subspace clustering and cumulative Markov random field algorithm to explore the effectiveness of consistency constraints in highorder spaces for hyperspectral images classification.Firstly,the algorithm divides homogeneous and adjacent high-order regions to ensure the consistency of homogeneous region classification results.Secondly,a new regularization term is proposed to incorporate the spatial constraints between regions into the algorithm framework to further expand the scope of spatial influence.Finally,MRF is used to make up for the lack of detail caused by region segmentation.(2)This paper proposes a sparse subspace clustering algorithm with high-order spatial information.Firstly,high-order spatial information is obtained by dividing homogeneous and adjacent regions.Secondly,high-order data sparse representation and single sample sparse representation are integrated into a framework of sparse subspace clustering.Finally,the consistency of single sample sparse coefficients and high-order sparse coefficients are ensured by sparse coefficient mapping term,so as to improve the spatial consistency of high-order regions on the basis of preserving the details.(3)This paper proposes a target-aware locally guided sparse subspace clustering method.Firstly,nonnegative matrix factorization is introduced to obtain the category attributes from sparse reconstructed data,and then a local guidance matrix is constructed to guide the subsequent sparse representation process to meet the intra-class compactness and inter-class looseness.Finally,a new high-order space constraint term is used to ensure the consistency of homogeneous region classification results.This method realizes the combination of sparse selfrepresentation and category information,which is conducive to building a more accurate affinity matrix.The proposed algorithm model is tested on Indian pines data set and Pavia university data set respectively.The final hyperspectral images classification results show that,compared with the advanced algorithms in this field,the algorithm model proposed in this paper not only improves the classification accuracy significantly,but also makes significant progress in reducing noise interference and time complexity.