Hyperspectral Image Classification Based On Tensor Decomposition And Low-Rank Matrix Recovery

Nowadays,hyperspectral images(HSIs)are widely used in geophysics exploration,agricultural remote sensing,ocean remote sensing and environmental monitoring due to its rich spectral information.The objective of HSI classification is to assign each spectral pixel to one of the classes based on the spectral characteristics.However,hyperspectral image classification is also facing some new problems,such as how to deal with the Hughes phenomenon of high dimensional data,how to use the spatial information of hyperspectral images and how to eliminate the difference of variance within different classes.In this thesis,the research status of HSI classification algorithm home and abroad is introduced,a group low-rank tensor decomposition model and a low-rank matrix with superpixel model are proposed for HSI classification.The main contents of this thesis are as follows:1.Classification based on Group Low-Rank Tensor Decomposition.Recent studies have shown within-class spectral variation seriously affects the performance of hyperspectral image classification.In this thesis,a novel group low-rank tensor decomposition(GLRTD)method is proposed to alleviate within-class spectral variation by fully exploiting the low-rank property of 3D HSI,which can significantly improve the classification performance.Specifically,the spectral dimension of the HSI is firstly reduced with principal component analysis(PCA)algorithm.Then,the dimension reduced image is segmented into a set of overlapping 3D tensor patches,which are then clustered into groups by Kmeans algorithm.By unfolding the similar tensors of each group into a set of matrices and stacking them,these similar tensor patches are constructed as a new tensor.Next,the intrinsic spectra tensor and its corresponding spectral variation tensor of each new tensor are estimated with a low-rank tensor decomposition(LRTD)algorithm.By aggregating all intrinsic spectra tensor in each group,we can obtain an integral intrinsic spectra tensor and separate its corresponding spectral variation tensor.Finally,the pixel-wise classification is performed only on the intrinsic spectra tensor,which can reflect the material-dependent properties of different objects.Experimental results on real HSI data sets demonstrate the superiority of the proposed GLRTD algorithm over several well-known classification approaches.2.Classification based on Low-Rank matrix recovery and superpixel(LR-SP).In the HSI,each superpixel can be regarded as a shape-adaptive region,which consists of a number of spatial neighboring pixels with very similar spectral characteristics.Therefore,it suggests that a superpixel in the HSI can be regarded as a low-rank matrix.First,the proposed LR-SP method adopts an over-segmentation algorithm to cluster the HSI into many superpixels.Low-rank matrix decomposition(LRMD)is utilized to extract the spatial features within and among superpixels.Experimental results on two widely used real HSIs indicate that the proposed LR-SP approach outperforms several well-known classification methods.