高光谱图像分类的全面加权方法研究

RESEARCH ON ALL-AROUND WEIGHTING METHODS OF HYPERSPECTRAL IMAGERY CLASSIFICATION

像元分类是高光谱数据分析的最基本、最重要内容之一,而基于支持向量机(SVM)的分类方法以其高效性得以广泛使用.原始的SVM分类模型中并没有体现出样本、特征、类别对于分类或分析的不同重要性,从而影响了处理效果.为此,将各样本偏离其类中心的距离映射为样本加权系数;将类内散度矩阵应用于特征加权方法;将SVM方程系统中的单位矩阵对角元素加以调整来完成类别加权.不同加权方法既可以单独使用也可以联合使用.实验表明,所提出的加权方法有助于进一步提高高光谱图像的分类效果.

Pixel classification is one of the most basic and important contents of hyperspectal imagery （HSI） analysis, and SVM based method is very popular in HSI classification for its high efficiency. The importance of samples, features, and classes, however, is not reflected in original SVM based classification model, and the classification effect is deteriorated consequently, in this study, the distance of each sample deviating from its class-center was mapped into the sample as weighting coefficient. And within-class scatter matrix was introduced into the feature weighting measure, and the diagonal elements in SVM equation system were adjusted for the purpose of class weighting. The weighted methods can be used solely or jointly. Experiments show that the proposed weighting methods are helpful to improve the effect of HSI classification.