A Study Of Hyperspectral Unmixing Based On Low-Rank And Sparse Representation

Hyperspectral unmixing is a hot research topic in the current remote sensing field.It is to extract deeper and more accurate information in a single pixel.The main content is to obtain the actual material existence and the proportion of the corresponding material by decomposing the mixed pixels.Through more accurate analysis and calculation of individual pixel components,research on mapping and identification topics can be greatly promoted.Because the remote sensing technology itself has a low spatial resolution,the mixed pixels correspond to an actual area block on the earth's surface.According to the characteristics of surface material distribution,the materials distributed in a certain area block are generally the same,then the corresponding material proportion also have certain structural features.The traditional method for hyperspectral unmixing generally performs the unmixing work on the properties of the whole hyperspectral image to do the constraint.Based on the feature that the material distributed in the regional block is basically the same,this paper extracts and utilizes higher-level local spatial information by extracting the hyperspectral region similarity block.The specific work is as follows:1.In the local area of hyperspectral image,the distribution of the end-member is basically the same,the use of superpixel division?Simple Linear Iterative Cluster,SLIC?algorithm to split the whole hyperspectral image into many regions with high similarity.In the acquired regional block,the distribution of end-member is the same,then on the basis of NMF,it can use characteristic of the abundance and then get more precise and effective solutions of mixed results.2.Nuclear norm refers to the sum of the singular value of the matrix.Here it gives each singular value a weight according to the importance of the singular value,and the important weight is relatively large,whereas the other is small.And weighted nuclear denotes weighted sum of matrix's singular value,can be more flexible and effective on expression of low rank information,have relatively good anti-noise properties at the same time.Based on the superpixel segmentation,the spatial feature information of each small block is utilized.The weighted kernel norm algorithm is used to obtain more effective unmixing accuracy,and it has a certain inhibitory effect on external noise interference.3.Structured sparseness(l_2,1 norm)is different from sparseness of l₂ and l₁ norm,needs to be more row sparseness.It is better in solving the efficiency of the solution.However,for the solution of the whole hyperspectral image,the obtained abundance image is a sparse representation in essence,l_2,1 norm take advantage of the sparse features,but it does not take full advantage of structuring.The method is based on the superpixel segmentation,and the characteristic performance of abundance in the region similar to the region is obvious.Because the existence of the block terminal element is certain and the same.The structural features are used effectively to obtain more effective solution mixing accuracy.4.Information entropy measures the amount of the information.The greater the uncertainty of variables,the greater the amount of information.Entropy is used to extract the information of singular value in the matrix,and the low-rank nature and sparse property of the matrix can be expressed concisely.According to the spatial feature information in the obtained local block,information entropy function is used to constrain the singular value of the matrix,and more accurately represent the spatial information in the current block.The effective and sufficient use of local information can obtain more accurate solutions.