Research On Mixed-pixel Unmixing Of Hyperspectral Imagery

Hyperspectral imaging spectrometer can provide nearly continuous curve of spectrum of ground objects with very high spectral resolution. It not only provides the possibility for classifying ground objects more nicely, but also gives high data guarantee on more precise mixed-pixel unmixing and sub-pixel classification. Some open problems and hot points in mixed-pixel unmixing methods of the hyperspectral image were studied in this dissertation. The major works of this dissertation are listed as follow:1,Based on the analysis on genesis of spectral mixing, the linear mixing model and stochastic mixing model were studied, which makes the preparation for follow-up study. According to related math model, a series of synthetic data were created for quantitative analysis for algorithms in latter chapter.2,The methods based on Neyman-Pearson Detection for Virtual dimensionality estimation were studied, and drawbacks of the methods to be used to estimate the number of endmembers were analyzed. Then, an Orthogonal Subspace Projection theory based algorithm, which is used to estimate the number of endmebers of hyperspectral image, was put forward. Finally, experiments on these two algorithms were carried out for comparison.3,Some typical endmember extraction algorithms were analyzed, and a Particle Swarm Optimization based endmember extraction algorithm was put forward. Through experiments, its ability to extract endmembers was proved. Additionally, based on the analysis of principle of Discrete Stochastic Mixture Model, a endmember extraction algorithm was put forward, and experiments proved its ability to solve the problem of endmember variation.4,Based on the researches on typical supervised linear spectral unmixing algorithms, by adding weight matrixes, the fully constrained least squares unmixing algorithm was improved and its improvement was analyzed experimentally. Theory of Nonnegative Matrix Factorization was analyzed, and its solving method was applied into unsupervised spectral unmixing. It is proved by experiments that this method is applicable to unmixing spectrum.5,Combined with the spectral mixing model based methods for sample training, a support vector machine based sub-pixel classification algorithm, which is applied to hyperspectral image, is put forward. Finally, by means of experiments, the validity of this method on linear and non-linear mixing data was proved.