Hyperspectral Imagery Classification Aiming at Protecting Classes of Interest
Hyperspectral Imagery Classification Aiming at Protecting Classes of Interest
复制标题
DOI：
10.1109/csie.2009.990
复制
发表时间：
2009-03
期刊：
影响因子：
--
通讯作者：
Wang Liguo;Deng Luqun;Lei Ming
中科院分区：
文献类型：
--
作者：
Wang Liguo;Deng Luqun;Lei Ming
分类是高光谱图像处理的一项重要技术。在传统的高光谱图像分类方法中，所有类别都是平等对待的。但是，其中一些方法应给予更多的关注，因此，特别强调兴趣类的分析效果是有意义的。在这种情况下，提出了两种处理方法，以保护感兴趣的类在最小二乘支持向量机的分类过程中：删除训练样本和改变对角元素。前一种方法通过在SVM训练过程中删除不感兴趣类的样本，保留感兴趣类，从而大大提高了分类精度。后一种方法通过对惩罚矩阵的对角元素赋予不同的权值，使感兴趣类的样本得到更多的关注，从而提高了相应的分类精度。实验结果表明，本文提出的方法能够有效地提高感兴趣类的分类效果。
Classification is an important technique of hyperspetral imagery processing. In traditional classification methods of hyperspectral imagery, all classes are treated equally. Some of them, however, should be given more regard, and so, it is significant to emphasize particularly on the analysis effect of classes of interest. In this case, two kinds of processing methods are proposed to protect classes of interest in process of least square SVM based classification: deleting training samples and changing diagonal elements. In former method, by deleting samples of uninterested classes in process of SVM training, interested classes are left and their classification accuracies are improved greatly. In latter method, by attaching different weights to diagonal elements of punishment matrix, samples of interested classes are given more regard and so the corresponding classification accuracies are improved. Elaborate experiments show that the proposed methods can improve the classification effect of classes of interest.