Research On Hyperspectral Image Unmixing Algorithm

Compared with multispectral remote sensing,hyperspectral images provide more information on the number of bands and the range of wavebands.These rich spectral information can be used to detect and recognize the ground objects.However,a problem that can not be ignored in the use of hyperspectral data is the existence of mixed pixels.The unmixing method is the most effective tool for sloving mixed pixel problems.The paper is based on the NLSMA?Non-Local Spectral Mixture Analysis?method.For the shortcomings of using Kd-tree method to find non-locally similar blocks that occupy too much memory and have a long computation time,we propose a joint sparse unmixing method based on low-rank decomposition.The low-rank decomposition model decomposes the matrix into a low rank matrix,a sparse matrix and an error matrix whose constraints are the?_?-norm of sparse matrix and error matrix is less than the threshold.The optimization algorithm of the model is less and the running time is too long,so we appropriately relax the constraint conditions and replace the?_?-norm of the constraint items with?₂-norm.Experiments show that the method can achieve the same effect and greatly improve the speed of operation.After low-rank decomposition of the image,many similar blocks composed of many similar pixels are obtained.It is assumed that the spectral signals of the pixels in these similar blocks consist of the same endmembers,but with different abundance coefficients.The NLSMA method uses the joint sparse method to perform a full-limit nonnegative solution for each similar block.It uses the T-MSBL?Transform-Multiple Sparse Bayesian Learning?method to solve the multiple measurement vectors?MMV?problem to obtain the types of ground objects that may be included in similar blocks,and then according to FCLS?Fully Constrained Least Squares?method finds the corresponding abundance vector for each pixel in a similar block.Observing the water abundance map obtained by this method,we can see that there are many non-water bodies whose local abundance coefficient is not zero.Therefore,considering the NDWI indicator to propose the water body firstly,it will provide a framework to propose a certain specific ground object individually.Compared with other MMV algorithms,the T-MSBL method considered the temporal correlation between the observation vectors and obtains better results.However,this method consumes a lot of time.Therefore,the MSBL?Multiple Sparse Bayesian Learning?method is considered.The experiment shows that the method has faster calculation speed and the accuracy is similar to the T-MSBL method.Using the method proposed in the paper,the experiments were performed on the published hyperspectral dataset.The experimental results show that the average accuracy of the 9categories of classification is 93.35%.According to the V-I-S model,the 9 types of ground objects are merged into 5 categories,and the overall classification accuracy reaches at 96.88%.Compared with NLSMA method,the proposed method could greatly improve the speed of operation when the accuracy is similar to the NLSMA.