A Research On Dimensionality Reduction And Classification Of Hyperspectral Image Based On Support Vector Machine

Statistical pattern-classification of hyperspectral imagery (HSI) is a difficult endeavor, due to the high-dimensional feature spaces often tend to result in ill-conditioned problem. Obstacles, such as the houghes phenomenon, arise as the data dimensionality increases, thus fostering the development of subspace learning methods, which are able to deal with high-dimensional data sets and limited training samples. The goal of subspace learning is to obtain a low-dimensional representation of high-dimensional data samples while preserving most of the ’intrinsic information’ contained in the original data. As a discriminative classifier, support vector machines (SVMs) rely on discriminant functions and posterior class distributions, based on which many state-of-the-art classification methods are built. SVMs, whose ability in dealing with large input spaces and producing sparse solutions has been largely demonstrated, has shown good performance in hyperspectral image classification. This thesis detailedly introduced a series of subspace learning method, combined with support vector machine classifier, then systematically evaluate the classification performance of different dimension reduction algorithm with two different kinds of hyperspectral data sets. In particular, we present a multi-kernel learning framework based on multiple characteristics of hyperspectral image, then estimate quantitatively with its state-of-the-art competitors. The main contributions of this thesis are as follows:(1) This thesis particularly introduced the conventional linear dimension reduction(DR) methods such as PCA & LDA, and some representative nonlinear manifold learning algorithms such as LLE, ISOMAP, LE, LPP & LFDA and semi-supervised subspace learning strategy, furthermore, systematically reveals the basic principle of subspace learning algorithms. Subsequently, we discussed the intrinsic dimension of hyperspectral image data and introduce the concept of virtual dimension.(2) A research on hyperspectral image classification based on support vector machine (SVM). First briefly introduces the basic theory of support vector machine classifier, which is the statistical learning theory. Then, this thesis puts forward a variety of different support vector machine models such as linear, nonlinear and semi-supervised methods. Finally, we discussed different evaluation criterion of hyperspectral image classification, and combined with a variety of subspace learning method based on support vector machine (SVM) to carry out classification results evaluation, and compared the classification performance of different dimension reduction algorithms under the same conditions.(3) A locality-preserving multi-kernel learning framework are proposed to deal with hyperspectral image classification tasks with multiple features as input. Based on the existing research, we incorporate multiple features such as original spectral features, local characteristics and spatial properties into a multi-kernel learning framework, and explore the possible optimal combination model. Subsequently, we compare the proposed framework with its state-of-the-art competitors, the results show that our methods has obvious advantages on the classification results.