In hyperspectral image classification, labeled samples are difficult to obtain. Semisupervised classification methods can make use of the information contained in the large number of unlabeled samples to improve classification accuracy. Transductive support vector machine (TSVM) is an extension of support vector machine ( SVM) in semisupervised learning. In this paper we use Concave-Convex Procedure ( CCCP) to optimize the nonconvex objective function of TSVM. The noconvex function is decomposed into the combination of convex part and concave part. So the problem is changed into a convex optimization problem. In hyperspectral image, each band's ability to distinguish materials is not in the same range. In order to make better use of bands' classification abilities,the spectrally weighted vector ia introduced. The evaluated approaches of weighted values for two classes and multi classes are introduced. As a result, different bands are assigned different weighted values. Then the kernel function is modified by the spectrally weighted vector. Experiments prove the proposed method's superiorities. Thus the proposed method can be applicable to large-scale hyperspectral image classification.%高光谱图像分类中有标签样本获取较为困难,而半监督分类可以利用到大量未标签样本所含信息,来提高分类准确率.直推式支持向量机是标准支持向量机在半监督学习问题上的一种扩展.本文采用凹凸过程规划将直推式支持向量机的非凸目标函数分解为凸函数和凹函数的组合,将非凸问题转化为凸优化问题求解.并且针对高光谱图像不同波段鉴别地物类别的能力的差异,为了充分利用各个波段的分类能力,引入了光谱权值的概念,探讨了两类分类和多类分类的权值估计策略.对不同的波段赋予不同的权值,从而改进了直推式支持向量机的核函数.实验表明了本文提出算法的优越性,适用于较大规模的高光谱图像分类.
展开▼
