Research On Hyperspectral Image Mixture Pixel Decomposition And Compressed Sensing Reconstruction Algorithm

Hyperspectral images have abundant spatial,radiative and spectral information,In the course of data acquisition,transmission,and analysis,the increase in the amount of data is a new challenge for traditional data acquisition and compression.At the same time,due to the influence of the ground object itself,spatial resolution,atmospheric environment and other factors,the phenomenon of mixed pixel is commonly found in hyperspectral images.The mixed pixel decomposition can not only obtain the true information of pixels,but also help the image classification and object analysis.This topic focuses on the methods of decompose hyperspectral mixed pixel based on compressed sensing and reconstruction using the linear mixed pixel model of pixel elements.This paper studies the signal sparse representation,compressed sensing,decomposition of mixed pixels and reconstruction of hyperspectral images.The main research contents of this topic are as follows:(1)This paper introduces the characteristics and research status of sparse representation,compression sensing technology,and the decomposition of hyperspectral imagery,introduces sparse representation and compressed sensing framework to analyze the principle of compression reconfiguration,then through the analysis of the hyperspectral linear mixing model,explain the connection and difference between linear mixed model of pixel and compressed sensing theory,then based on the commonality between the two models,integrate the linear mixing model of pixel into the compressed sensing framework,and based on the known spectral library,realize the endmember extraction and abundance inversion;(2)Simulate hyperspectral mixed pixel phenomena in real scene by synthetic data,analyze the rationality and effectiveness of the proposed program,calculate abundance matrix based on spectral library using OMP algorithm and SOMP algorithm respectively,then use the two norms of the obtained abundance matrix to extract endmembers and estimate the corresponding abundances,the performance of the algorithm is evaluated using the parameters included spectral angular distance(SAD)of extracted endmembers and the root mean square error(RMSE)and standard deviation of the estimated abundance,at the same time,the results are compared with the results of endmember extraction and abundance estimation of the SMACC algorithm.The experimental results show that the endmembers extracted from the known endmembers under the framework of compressive sensing are truly existent,rather than the virtual endmembers,and the difference of two norm of abundance matrix is obvious,the effect of endmembers adaptive extraction is good,the abundance estimation accuracy is better than SMACC algorithm;(3)According to the characteristics of the linear mixed model of hyperspectral images,the spectrally compressed image still satisfies the linear mixed model,we discuss in the case of known endmember spectra and unknown endmember spectra respectively,abundance inversion of compressed data,using the inversion results and the endmember spectrum to reconstruct the hyperspectral images.Reconstruct the compressed data at different sampling rates,the results of the peak signal to noise ratio(PSNR)were compared with the standard OMP algorithm,experimental results show,In the synthesized data,the reconstruction effect of known endmembers is much better than the other two;In real data,the reconstruction effect of known endmembers and unknown endmembers are better than OMP algorithm;(4)According to the strong correlation of hyperspectral images in the spectral domain,the reconstruction accuracy is improved by distributed compressed sensing.Based on distributed compressive perception,using residuals error of spectral image,adaptive code stream allocation,according to the characteristics of guided filtering,distributed compressed sensing reconstruction based on guided filtering,and while considering the noise distribution of hyperspectral images,optimize the selection strategy for key and non-key bands.Experimental results show that compared to traditional distributed compressed sensing reconstruction algorithms,the proposed algorithm is improved in the reconstruction performance,at the same time,it has the effect of removing hyperspectral image noise,and it also can maintain vital plant information,effectively support plant hyperspectral data processing and analysis.This topic uses the relationship between the compressive sensing framework and the linear mixed model of the pixels,decomposition mixed pixels using compressed sensing,the results show that the proposed unmixing method is superior to the traditional unmixing method;simultaneously reconstruct hyperspectral image using linear mixed model,the results show that the method can extract endmembers and estimate abundance of hyperspectral images in the case of known and unknown endmembers to achieve reconstruction,the reconstruction effect is better than traditional methods;at the same time,based on distributed compressed sensing,guided filter assistance and adaptive grouping of AP algorithm improved,the reconstruction effect is better than the traditional method.