Research On The Unmixed Technology Of Hyper Spectral Image On Compressive Sensing

With constant improvement in spatial resolution of satellite sensors, the ability to eliminate the spectral secondary reflection effect is continuously enhanced, which brings new research opportunities for hyperspectral image classification, but mixed pixel decomposition of hyperspectral image has become a new research difficulty in the field of remote sensing due to issues such as the diversity of complex measured ground feature, multiple scattering of ground object spectra and real-time change, etc. This paper has carried out research on demixing preprocessing, feature extraction of endmembers as well as subspace partition of endmembers etc. by making full use of sparsity of abundance coefficient vector in spectral library with compressive sensing theory as a research tool aimed at such issue such as low decomposition accuracy existed in hyperspectral demixing process, redundant end members and slow processing speed etc. The main research work includes the following three aspects:â‘ In allusion to uncertain observation vector caused by random measurement matrix in the process of sparsity demixing of hyperspectral image, this paper establishes integer programming model in search of width of measurement matrix as well as sparsity, introducing chaos ant colony algorithm to determine the parameters of the measurement matrix. Based on pseudo-random sequence generated by the chaotic Tent Map system, it has constructed improved chaotic Toeplitz matrix according to the cycle /block diagonal splitting structure. Using the correlation between the measurement matrix and the projection matrix, this paper introduces block SVD decomposition strategy in stochastic gradient descent algorithm so as to get optimum improved measurement matrix of chaotic Toeplitz. Experimental tests show that in the process of extracting features of endmembers, the optimized measurement matrix not only has effectively reduced the number of endmembers maintaining a high processing speed, but also has ensured a more accurate feature extraction.â‘¡As for issues of declined quality of the hyperspectral image caused by real-time additive noises of clouds and atmosphere, etc. in the collection process, quantization noise in the quantization process and linear fuzzy because of relative motion between hyperspectral imager and the ground objects, as well as the extraction accuracy of endmembers from sparsity demixing generation, the paper proposes a solution of combining improved splitting Bregman iteration and improved minimum approximate conjugate gradient. The improved minimum approximate conjugate gradient algorithm is used to detect peaks and mutated portion of spectral signals, and parallel optimization acceleration strategy is adopted to improve convergence rate of splitting Bregman iteration policy. The paper has introduced iterative weighted strategy to transform dictionary learning, and to suppress structure noisy of original signals. Experimental results show that, for different noises, this strategy can not only extract and retain the real-time signal peaks and mutated portion splendidly, and has the ability of robustness to reconstruct the original testing signal, which has helped the subsequent realization of improving decomposition precision from the sparsity demixing of multiple endmembers in the mixed region.â‘¢In terms of issue how to divide the high-dimensional data into subspace of endmember to improve the decomposition accuracy and processing time of the mixed elements, the paper has proposed a real-time low-rank decomposition strategy of combining improved alternating direction method and robust principal component analysis(RPCA) to achieve low rank decomposition by predicting, learning and training the online dictionaries, and it has introduced block adaptive SVD decomposition strategy to improve rate of low-rank decomposition by using the redundancy compensation to reduce error in low rank decomposition. The policy is applied to the sparse subspace clustering(SSC) of hyperspectral signal to achieve the purpose of improving precision of demixing decomposition by dividing subspace of endmembers. Experimental studies have shown that this strategy is an effective classification method in real-time feature dimension reduction, and achieves classification of effective bands in feature dimension reduction of real endmembers.Finally, in the main research of this paper, beneficial attempts have been carried out in the main results such as measurement matrix construction and optimization obtained from compressive sensing theory, robust reconstruction with noise background of hyperspectral image and low-rank decomposition of high-dimensional signal, etc. to issues in hyperspectral sparsity demixing such as feature extraction of endmembers, the pretreatment of demixing, subspace partition of endmembers and so on, and it focuses on improving the feature extraction accuracy and extraction speed of endmembers in sparsity demixing so as to achieve multiple-endmembers sparsity demixing of hyperspectral image.