Research On Classification Of Hyperspectral Images Based On Support Vector Machine

With the development of digital signal processing technology, computer technology and communication technology, remote sensing imagery processing takes increasingly important effects in the fields of military and civil applications. Compared with multispectral remote sensing, the hyperspectral remote sensing has a higher spectral resolution. The research of hyperspectral images (HSI) classification is one of the main contents of the hyperspectral remote sensing application. Hard classification which aims at simplifying the complex phenomenon is one of the significant processing methods in the interest information extraction of hyperspectral images. On the other hand, mixed pixels are widely existent in hyperspectral images for its low spatial resolution. The analysis and processing of mixed pixels are of more importance and significance. As a main technique of mixed pixel processing, spectral unmixing is to work out the mixing proportion of each class included in mixed pixel. It is a more accurate soft classification technique. At present, there are lots of hard classification methods, but their classification performances are not very perfect, or some methods themselves are to be improved. Traditional spectral unmixing methods are inefficient for the participation of unrelated classes and for the deficiency of spectral unmixing model. In this case, the techniques of hard classification and soft classification (spectral unmixing) are researched mainly based on support vector machine (SVM) theory in the paper.Firstly, a band selection method based on stepwise deletion of similar feature and a gauss low pass filtering method are proposed. The band selection method is constructed on the computation of feature similarity. The method is unsupervised and of low computational complexity. The proposed gauss low pass filtering method aims to weaken or eliminate the high frequency component of hyperspectral images but not change its low frequency component. The filter is designed to smooth hyperspectral, to reduce within-class distance and to enlarge between-class distance. In this case, the filtering method is helpful of following classification processing of hyperspectral images.Secondly, SVM theory is researched systemically, including theory basis, classification principle, the generalization form linear classification to nonlinear classification, the extent from bi-class problem to multi-class problem, and the main implementation techniques. The classification performance of SVM is also tested in this paper. The studies of theory basis and classification principle benefit to understand the unique dominance of SVM, the studies of optimization algorithm and development version are helpful of improving its application efficiency, and the study of multi-class extension makes it possible for the technique to process classification problem with large number of classes conveniently. These studies provide necessary theory basis for the progress of the dissertation.Thirdly, the classification performance of SVM is improved; including the usage of fuzzy method before classification, distance based fuzzy SVM, and secondary training in multi-class problem. In order to overcome the blindness of sample selection, fuzzy method select training samples based on fuzzy xxx, followed by SVM based classification. The weighted SVM makes use of distance measure to fuzzy weigh the punishment terms in least square SVM, and so overcomes the bad influence brought by outliers in process of training. In the idea of secondary training, all parameters in two kinds of SVM are regulated effectively. Through the secondary parameter regulation for some weak classifiers, new general classifier is formed by sub-classifiers with optimal parameters.Finally, the feasibility and method of applying SVM to spectral unmixing (a kind of soft classification) and the selection method of class sub-set are researched. On the one hand, the principle of applying linear SVM to spectral unmixing is introduced, and then the application is intended to nonlinear SVM. The unmixing effect of the application is also tested. On the other hand, the selection of correlative class sub-set is conducted based on spatial correlation and on class-of-interest is researched, and the more accuracy spectral unmixing is implemented on the class sub-set.Experiments show that SVM has good performance both in hard classification and in soft classification, and proper preprocessing measure, improving measure of SVM, and selection measure of class sub-set are all helpful of getting better analysis effect.