Combination Of Spatial Information And Spectral Information For Hyperspectral Imagery Classification

With the development of imaging spectroscopy, hyperspectral remote sensing technologyhas been developed and applied in many areas. And the hyperspectral imagery classificationtechnology is the basis of many applications. Hyperspectral imagery is rich of spectralinformation by combing the traditional image space dimension with spectral dimension.Hyperspectral imagery classification becomes one of the most important parts inhyperspectral data processing. The most of established hyperspectral imagery classificationmethods only consider the spectral feature information in the classification; and ignore theimportant role of spatial features information. To improve the accuracy of imageryclassification, this article is research on the algorithm which joins spatial features and spectralfeatures together for hyperspectral imagery classification. Main contributions of this thesis aregiven as follows:Firstly, it gives a whole view of hyperspectral imagery remote sensing technology,elaborates description of the background and significance of the research. Including theanalysis of hyperspectral imagery theory, its imagery data characteristics, typical existinghyperspectral imagery classification methods.Secondly, it introduces the hyperspectral imagery data feature extraction methods andclassifier models. In feature extraction section, the spectral feature extraction methods andspatial feature extraction methods are analyzed separately. By the simulation, the principalcomponent analysis, nonparametric weighted feature extraction, gray co-occurrence matrixand wavelet transform feature vectors show their advantages and disadvantages inhyperspectral imagery classifications. In classifier selection section the support vectormachine compared with the traditional supervised classification methods, and shows goodperformance during imagery classification. Then the least squares support vector machine ischosen in next simulation.Again, the research is based on integrating spatial information and spectral informationin hyperspectral imagery classification. The spatial texture features is extracted by Gabor filter,and the new features fusion the Gabor texture feature and spectral feature by kernel functions.The new combination of features is classified by least squares support vector classificationmodel, and the experiment result shows the classification accuracy is improved. To eliminate the "noise" pixels existing in the classification results, the neighborhood regularizations isadopted after classification.Finally, the algorithm based on multi-scale homogeneity discriminate to improvehyperspectral imagery classification result is proposed. Most of classification models focus onsingle-pixel spectral feature and spatial feature extraction, and ignore the pixel spatialcorrelation with its neighborhood, which makes the "noise" pixel existing in classificationresults. The imagery pre-classification result is divided into a series of non-overlapping areas,calculates the homogeneity degree of membership in the neighborhood of pixel, using thehomogeneity region determination rules reprocess classification results to eliminate "noise"pixels. Then the multi-scale zoning method compensates the single-scale disadvantages.