Research On Hyperspectral Image Classification Based On Support Vector Machine

The classification of ground objects is the basic content of hyperspectral image processing. In the real applications, the number of training samples is always limited. Statistical learning theory, the first theory that systematically studies the problem of machine learning with small size samples, put forward a kind of general machine learning method——Support Vector Machine(SVM). In this paper, SVM is applied in classification of hyperspectral image, the major work of it is as following four aspects:1. Systematically introduces some important theories and fundamental knowledge, such as empirical risk minimization principle, structure risk minimization principle, convex quadratic programming problem and kernel function reflection etc. and then discusses commonly used SVM applying techniques.2. Experimentally analyses the influence of the kernel function and the model parameters (including kernel parameter and error punishing factor) on SVM classification, and introduces two kinds of criterion for the performance of kernel parameter, and then educes two kinds of model parameters selecting algorithm——distance-based and angle-based kernel parameter searching algorithm. It is indicated that kernel parameter searching algorithms are better than traditional model parameters selecting algorithm (Grid searching algorithm, Genetic algorithm) through experimental analysis.3. Deeply analyses the characteristics of hyperspectral image data and the key problems of traditional classification methods, and systematically introduces and analyses existing multiclass SVMs methods. By using multicalss SVMs methods and traditional classification methods in hyperspectral image classification, this paper analyses and compares their performance, and deduces that multiclass SVMs methods can get better results than traditional methods.4. Theoretically introduces three kinds of reformative SVM algorithms: Least Squares SVM(Ls-SVM), Lagrangian SVM(L-SVM), Proximal SVM(P-SVM). Combining with the advantages of P-SVM, we proposes a support vector preselecting method (PSVM-SVP) .the results of experiments indicate that the application of PSVM-SVP can reduce the time for training and still keep the expansion ability of the C-SVM in the meantime.