Research On Sparse Unmixing And Sub-pixel Mapping Based On Endmember Dictionary Of Imaging Spectral Images

With the gradual maturity of remote sensing technology,imaging spectral images became widely used in modern society.Due to the limitations of imaging spectrometers,the spatial resolution of most imaging spectral images is low,which leads to a large number of mixed pixels in remote sensing images,brings a lot of inconvenience to the subsequent data processing.Therefore,the spectral unmixing technique of imaging spectral images has important theoretical research significance and practical application value.In this paper,starting from the characteristics of imaging spectral images and the principle of imaging,the study of sparse unmixing techniques for imaging spectral images based on endmember dictionary is carried out,and an endredundant dictionary training algorithm for hyperspectral images based on feature categories is proposed.And a spectral unmixing algorithm based on K-SVD and endmember dictionary has received good results.It is further applied to the sub-pixel mapping field,and a subpixel mapping algorithm based on spatial correlation constraints is designed and implemented,which effectively improves the positioning accuracy of the traditional algorithm.Specifically include:An endmember redundant dictionary training algorithm based on K-SVD and standard spectral library is designed and implemented.The method is based on the standard spectral library.Firstly,spectral features are used to classify the spectral library data.Then a single dictionary is trained for the spectral curve of each type of feature,and the most representative spectral curve of the feature is selected to constructed the endmember redundancy dictionary.Using the method of spectral angle matching and spectral information divergence,the endmembers in the dictionary are further selected and optimized so that the atoms in the endmember dictionary can reduce the correlation while ensuring sufficiency,to eliminate the subsequent spectral unmixing process.The negative effects of irrelevant items lay the foundation for the efficient unmixing of the mixed spectral pixels in the subsequent imaging spectrum.A mixed pixel unmixing algorithm based on endmember dictionary is designed and implemented.This method first established a linear spectral mixing model,and then used the KNN method to detect the mixed pixels in the image.Final,used the spectral redundancy dictionary trained as a terminal element set,the sparse decomposition method is used to unmixing the mixed pixels.Experiments with simulated data and real data showed that the method has good unmixing effect,high robustness,and improved the problem of low precision of the traditional unmixing method under low signal-tonoise ratio.A subpixel mapping algorithm based on sparse unmixing of endmember dictionary is designed and implemented.This method constructs two cost functions based on spatial correlation.Then based on the previous research,the pixel component exchange algorithm is used to achieve the sub-pixel location of the imaging spectral image.Experimental results showed that the proposed algorithm has good effect of sub-pixels mapping at the boundary of features,and is close to the distribution of actual features.It validates the effectiveness of the sub-pixel mapping algorithm and in turn verifies the accuracy of the unmixing algorithm.