Hyperspectral Image Mixed Noise Analysis Using Low Rank And Sparse Based Method

In recent years,with the wealth of available spectral and spatial information,hyperspectral images(HSI)have the high spectral diagnosis ability to distinguish precise land-cover details even between the similar materials,providing the possibility of application in urban planning,agriculture and forestry,cadastral inventory,target detection,biometric and so on.Unfortunately,due to the limitation of imaging equipment and external environment,hyperspectral imaging sensors unavoidably introduce Gaussain noise,stripes,pixel missing,clouds,shadow and so on,into the acquired HSI data during the imaging process,which severely degrades the quality of the imagery and limits the precision of the subsequent processing.To alleviate the influence of noise,it is necessary to systematically develap HSI noise analytical method from the data pespective,and build up robustness HSI analysis framework from the task perspective.In this paper,we first reviewed the research progress about HSI noise analysis.Subsequently,on the basis of previous work,we summarized three problems to be solved,including mixed noise problem,color noise problem and model error noise problem.To deal with the noise problem,guilded by the most recent data mining theory and the characteristics of HSI,we proposed the low rank based HSI noise analytical method.From the data side,we reduced the mixed noise and color noise both theoretically and experimentally.From task side,we proposed a unified framework including noise analysis(mixed noise and model error noise)and endmember/abundance blind unmixing.The main contribution of this thesis are as follows:(1)To the mixed noise problem,we propose a concept of sparse noise,which include stripes,pixel missing,impulse noise,clouds and shadow.,and assume that the HSI is corrupted by Gaussian noise and sparse noise.First,on the basis of the low rank property of the local clean HSI and sparse property of sparse noise,we introduce a new HSI restoration method named low rank matrix recovery(LRMR).Second,by exploring the low rank property from the spectral perspective and piese-wize smoothness of spatial perspective,we propose a totalvariation-regularized low-rank matrix factorization HSI mixed noise removing method.The last,by utilizing the local low rank property of spectral information and global spectral-spatial smoothness sturcture prior,we propose a spatial-spectral total variation regularized local low-rank matrix recovery method.We explore the local and global,spectral and spatial priors of HSI,and integrate the low rank and sparse representation,spectral-spatial total variation into a unified framework,to realize the mixed mixed noise analysis and removing.(2)To the color noise problem,we propose a noise-adjusted iterative framework.Due to the low-dimensional property of clean HSI,we introduec a noise-adjusted iterative low-rank matrix approximation(NAILRMA)method for HSI denoising.Patchwise randomized singular value decomposition is first applied to each Casorati matrix.An iterative regularization technique is subsequently adopted,based on the patchwise LRMA,to further separate the signal and noise.As for the HSI,different bands have different noise levels.A noise-adjusted update mode is then proposed to update the input image of the iteration.Guided by the noise variance of each HSI band,an adaptive iteration factor selection is also proposed.The results on the simulate and real data experiments demonstrate the advantage of the proposed method in the case of color noise removing.(3)To the Gausian noise,sparse noise and model error noise in the HSI blind unmixng problem,we propose a noise robust analysis framework.To overcome the Gaussian noise influence in the bilind unmixing,we make the best use of the structure of the abundance maps,and propose a new blind HU method named total variation regularized reweighted sparse NMF(TV-RSNMF),which integrates the endmember estimation,abundance estimation,and abundance denoising.To alleviate the influence of mixed noise and model error noise,we separately model the sparse noise and Gaussian noise,and propose a robust sparse NMF model to unmix the hyperspectral data.In the experimental part,we studied the real HYDICE Urban dataset in depth and partitioned it into three subsets,i.e.,low-noise bands,noisy bands,and waterabsorption bands.We then compared the results of the proposed sparsity regularized RNMF methods with a low-noise image consisting of low-noise bands and a noisy image consisting of both low-noise and noisy bands,and concluded that the noisy bands can also provide appropriate and useful information for urban data unmixing.