Research On The Unmixing Algorithm Of Hyperspectral Image Based On Sparse Representation

Hyperspectral images contain dozens to hundreds of continuous and narrow spectral bands from the ultraviolet to the mid-infrared region,and contain abundant spatial information.Therefore,hyperspectral images have abundant spectral and spatial information.In recent years,hyperspectral data has been widely used in earth observation,environmental monitoring,food safety,precision farming,mineral mining,biometric applications,search and rescue operations and other fields.However,due to the complexity of material distribution on the earth surface,sensor height,imaging field angle and atmospheric transmission effect,mixed pixels are inevitable.The precise interpretation and target detection of image data are seriously restricted by mixed pixels.More and more researchers pay attention to the unmixing model because it can provide subpixel level ground surface information.In recent years,the hyperspectral unmixing algorithm based on sparse regression has attracted more and more attention as it effectively avoids the two bottleneck problems of hyperspectral unmixing,namely,the absence of pure pixels and the estimation of endmembers.So,the sparse unmixing algorithm has gradually become a hot research direction for researchers at home and abroad.Although researchers have proposed a large number of unmixing algorithms,due to the low spatial resolution of hyperspectral images and the large correlation between bands,as well as the influence of various imaging conditions,the design of the algorithm is more difficult.Therefore,the design of the unmixing algorithm and theoretical analysis still face many challenges.In this paper,based on the sparse representation theory,the research is carried out on the unmixing preprocessing,unmixing model and other contents,and the algorithm of corresponding model is designed.The main research work and innovative achievements of this paper are as follows:1.Hyperspectral images are often polluted by all kinds of mixed noises.These noises severely limit the accuracy of hyperspectral unmixing.To remove the noise,based on low-rank tensor decomposition,combined with non-local self-similar prior information,a tensor-based non-local low-rank denoising model is proposed,where non-local self-similarity uses mainly spatial correlation while low-rank tensor decomposition method uses mainly spectral correlation between bands.The traditional tensor-based methods are NP-hard and sensitive to sparse noise.In this paper,a new tensor singular value decomposition?T-SVD?and tensor nuclear norm?TNN?is used to effectively separate low-rank clean images from Gaussian noise and sparse noise?pulses,deadlines,stripes,speckle,etc.?,and The NP-hard task was also achieved well by the alternating direction multiplier method.Because the spectral and spatial information of the data are fully utilized,Gaussian noise and sparse noise in hyperspectral images are effectively eliminated.2.The key to the success of sparse unmixing is based on an overcomplete spectral library.However,the high mutual coherence between spectral library signals seriously affects the accuracy of sparse unmixing results.For the problem of high correlation of spectral library and insufficient sparse constraint,a new hyperspectral unmixing method based on CLSUn SAL dictionary reduction and weighted sparse regression was proposed.First,the algorithm uses a dictionary reduction strategy to identify a subset of the original spectral library.Second,this paper proposes a weighted sparse regression algorithm based on CLSUn SAL to further enhance the sparsity of endmembers in a given spectral library.Third,the reduced spectral library is applied to the weighted sparse regression algorithm.Experimental results show that compared with CLSUn SAL algorithm,this algorithm not only has better performance,but also has faster computing speed.3.The joint-sparsity model is superior to the single sparse unmixing method.However,the joint-sparsity model may lead to misidentification of pixels on the boundary.To solve this problem,researchers have developed a number of unmixing algorithms based on low-rank representation that take full advantage of the global structure of the data.In addition,the high mutual coherence of spectral libraries seriously affects the applicability of sparse unmixing.A new hyperspectral unmixing algorithm called non-convex sparse and low-rank constraints with dictionary pruning is proposed by adopting combined constraints imposing sparsity and low rankness.In particular,sparsity is imposed on the abundance matrix by using the mixed _2,pnorm,and use the weighted Schatten-p norm instead of the convex nuclear norm as the approximation of the rank.At the same time,five simulation experiments and one real hyperspectral data set experiment are used to verify the effectiveness of the proposed algorithm.