Research On Unmixing Models And Algorithms Based On Endmember And Abundance Constraints For Hyperspectral Images

Hyperspectral images are three-dimensional image data,composed of hundreds of continuous band information.It contains not only spatial information,but also a lot of spectral information.Therefore,hyperspectral images have been widely used in practical cases,such as drug process monitoring and military target detection.However,due to the low spatial resolution of sensors and the diversity of ground materials,mixed pixels containing multiple materials exist in hyperspectral images.These mixed pixels greatly affect the recognition and classification accuracy of ground materials.Therefore,hyperspectral unmixing,i.e.,decomposition of mixed pixels into constituent spectra,called endmembers,and their corresponding distribution fractions,called abundances,is an essential technology for further application and analysis of hyperspectral images.Hyperspectral unmixing aims to extract the component endmembers and the corresponding abundances in the mixed pixels.Therefore,the key to improve the unmixing accuracy is to make full use of the spectral and spatial information in the hyperspectral image,explore the prior features of the endmembers and abundances,and adopt appropriate feature constraints.Among the unmixing algorithms with the available spectral library,the collaborative sparse algorithms assume that adjacent pixels usually contain similar endmembers,and is exploited by constraining the number of non-zero rows of the abundance matrix.Collaborative sparsity simultaneously characterizes the sparsity of the abundance matrix and the spatial correlation between adjacent pixels,so it has attracted the attention of researchers.However,the traditional collaborative sparsity algorithms are hard to exploit the row sparsity accurately and easy to be disturbed by noise.In addition,in the blind unmixing with unknown endmember information,the traditional hyperspectral unmixing algorithms also have some problems such as insufficient utilization and inaccurate characterization of the prior information of endmembers and abundances.In view of the above problems in hyperspectral unmixing,this dissertation has carried out in-depth research,and proposed hyperspectral unmixing models and algorithms based on endmember and abundance constraints to accurately describe the row sparsity of the abundance matrix and improve the ability of endmember feature extraction and abundance information.The main contributions of this dissertation are as follows:(1)A hyperspectral unmixing method based on reweighted collaborative sparsity is proposed.Compared with the large number of endmembers in the spectral library,a hyperspectral image usually contains only a few endmember,which means that the corresponding abundance matrix contains a large number of zero rows(called collaborative sparsity).In addition,the l2,1 norm,used to characterize the collaborative sparse in the traditional collaborative sparsity algorithms,is a convex relaxation of the l2,0 norm,which is hard to guarantee the sparsity of solutions.Therefore,the proposed method introduces a reweighted l2,1 term to characterize the row sparsity of the abundance matrix.It enhances the characterization of the row sparsity of the abundance matrix by using the reweighted strategy.At the same time,the total variation(TV)regularization is introduced to describe the spatial correlation of adjacent pixels in hyperspectral images.A two-step iterative strategy under the framework of the alternating direction method of multipliers(ADMM)is used to solve the model.Its key idea is to calculate the current solution through the linear combination of the results of the previous two-step iteration,instead of only calculating the previous iteration,so as to avoid the loss of information in the iteration process.(2)A non-convex row-sparsity hyperspectral unmixing method is proposed.It is difficult to accurately characterize the row-sparsity of the abundance matrix,whether it is the convex relaxation or the approximate term of the l2,0 regularization.Therefore,this method introduces the non-convex non-convex l2,0 term into hyperspectral unmixing.Considering that the l2,0 norm is very sensitive to noise,the TV term is introduced to solve the sensitivity of the l2,0 norm in noisy scenes.By simultaneously utilizing the l2,0 term and the TV term,this method characterizes the row-sparsity and spatial correlation of the abundance matrix at the same time,so as to make full use of the rich spatial information in the hyperspectral image.The proposed model is solved in the framework of ADMM.The experimental results show that the proposed algorithm is effective in improving the unmixing peformance,and it can obtain clean background while ensuring the row-sparsity of the abundance matrix.(3)An endmember independence constraint hyperspectral unmixing method based on non-negative tensor factorization is proposed.Due to the high correlation of spectral curves in hyperspectral images,it is difficult to completely separate highly mixed spectral signals without characterizing the endmember features.In order to alleviate the influence of high correlation of spectral signals,this method utilize spectral information and spatial information in hyperspectral images from both endmember and abundance aspects.Specifically,for the endmember estimation,we introduce an endmember independence constraint to avoid obtaining similar endmember curves.For abundance estimation,we use the low-rankness in abundance maps to describe the spatial correlation of mixed pixels lying in homogeneous regions of the hyperspectral images.Finally,the method is solved under the augmented multiplicative update framework.(4)A hyperspectral unmixing method based on spatial feature extraction is proposed.In order to improve the precision of unmixing,many regularization terms describing abundance distribution are introduced into hyperspectral unmixing.However,the abundance distribution in the real scenes is complex and changeable,which is difficult to describe it with a single feature.In addition,due to the high correlation of spectral signals and various noises in hyperspectral images,the estimated abundance maps often contain some small values that do not exist in true images.Therefore,this method divides each abundance map into a feature layer and a sparse layer to protect the obtained abundance map from the above factors.The feature layer represents the main information of the abundance map.The sparse layer contains outliers dominated by the above factors.In particular,we designed a feature extraction regularization to describe the feature layer and used weighted ell1 norm to describe sparse layers.And it is solved in the framework of the augmented multiplicative update.In summary,aiming at the row sparse characteristics of the abundance matrix,endmember independence and abundance distribution characteristics in hyperspectral unmixing,this dissertation proposes hyperspectral sparse unmixing method based on reweighted collaborative sparsity,non-convex row sparsity hyperspectral unmixing method,endmember independence constraint hyperspectral unmixing method via non-negative tensor factorization and hyperspectral unmixing method based on spatial feature extraction.It is a useful exploration for accurately extracting endmembers and the corresponding abundances from highly mixed spectral signals.