| As one major technique for people to observe information of the surface of the earth, spectral remote sensing has become more and more important for our lives and our social. With the tendencies of "3-High" (high spatial resolution, high spectral resolution and high frequency resolution) and "3-Multi"(rnulti-sensor, multi-flat and multi-angle), spectral remote sensing is providing more and more data. Unfortunately, our ability for abstracting information from spectral remote sensing images still largely lags behind technical developments. Therefore, it is considerable significant to improve the ability by studying new theories and methods.Classification is one of the methods for people to analysis information and to obtain knowledge. Traditional classification methods are based on empirical risk minimization inductive principle, which will lead to optimal accuracy only when infinite training samples provided. With limited training samples, the accuracy of traditional classification methods is always unsatisfactory. The problem is even worse in the classification of hyper-spectral remote sensing image because of the "Hughes Phenomenon". Statistical Learning Theory (SLT), the first theory that systematically studies the problem of machine learning with small size sample, presents a new inductive principle, structural risk minimization (SRM) principle, which tell us how to selection suitable classification model according to sample amount so as to obtain high generalization ability. Support vector machine (SVM) is a new general machine learning method based on SRM. In this paper, SVMs are applied in classification of spectral remote sensing imagery. The major works and contributions of this paper are as follows.? Three major kinds of supervised classification methods are systematically summarized and analyzed with theory results from statistical learning field from following aspects: the relationship among training accuracy, estimation accuracy of learning model and general ability of classifier; the interaction among complexity of learning model, dimension of feature vector and sample amount; the way in which these factors effect training accuracy and testing accuracy; model selection method.? SVMs are compared with some typical classification methods by several classification experiments with multi/hyperspectral remote sensing images. These typical classification methodsinclude not only classical methods-Minimum Distance (MinDis), Mahalanobis Distance(MahDis) and Maximum Likelihoods (MaxLike) -, but also most modern methods-Spectral Angle Mapping (SAM), LOOC and Radio Based Function Neural Network (RBFNN). Following conclusions can be drawn from experimental results. First, generally speaking, SVMs are more power than MinDis, MahDis, MaxLike, SAM and LOOC. Second, SVMs can avoid theHughes Phenomenon; Third, SVMs are generally faster than RBFNN in both training stage and classification stage. And the classification accuracies of SVMs are less sensitive to parameter selection than RBFNN. Thus SVMs are more convenient than RBFNN.Through systematically analysises of existing multi-class SVMs (M-SVMs) methods, it is shown that hierarchy multi-class SVMs (H-SVMs) can be relatively effective. Futher analysis shown that existing methods that measure separability between different classes are not suitable for kernel feature space. A new method is presented for separability measure in feature space based on the characters of RBF kernel function and SVMs. Based on the new separability measure, two kinds of H-SVMs, Binary Tree SVMs (BT SVMs) and Single Layer Clustering SVMs (SLC SVMs), are presented. They are both implements of following ideal: the higher a pair of two sub-classes is in the hierarchy, the easier to separate them. In this way, we can not only achieve higher classification accuracy by alleviate error accumulation from top to bottom, but also rise classification speed by reduce support vectors in classifier. Experimental results justify the rationality of the new separabili...
|
PDF Full Text Request