欢迎访问《图学学报》 分享到：
图学学报

图学学报

• 视觉与图像 • 上一篇    下一篇

基于多分类器的 C5.0 决策树植被分类方法

  

  1. 1. 中北大学信息与通信工程学院，山西 太原 030051；2. 英国雷丁大学系统工程学院，Reading RG6 6AU, UK
  • 出版日期:2017-10-31 发布日期:2017-11-03
  • 基金资助:
    国家自然科学基金项目(61672472)；中北大学科学研究基金项目(XJJ2016024)；中北大学电子测试技术重点实验室开放基金项目
    (ZDSYSJ2015005)

Vegetation Classification Method Based on C5.0 Decision Tree of Multiple Classifiers

  1. 1. College of Information and Communication Engineering, North University of China, Taiyuan Shanxi 030051, China;
    2. School of Systems Engineering, University of Reading, Reading RG6 6AU, UK
  • Online:2017-10-31 Published:2017-11-03

PDF (PC)

174

摘要： 针对光谱角制图(SAM)和最大似然(MLC)分类器对 AVIRIS 高光谱遥感图像进行植
被分类精度均不高的问题，提出了一种基于多分类器的 C5.0 决策树植被分类方法。首先，利用
支持向量机(SVM)，进行核函数以及核函数参数选择，提取出 AVIRIS 高光谱图像中的植被信息。
其次，利用 C5.0 算法将光谱角制图和最大似然分类器组合，作为决策树的特征属性，学习样本
训练并生成分类规则；根据 C5.0 算法计算植被样本中对应分类器的信息增益率，选择信息增益
率最大的属性去分类样本；当叶样本的分类结果满足停止生长的阈值，输出样本分类的结果，
否则，回到开始，递归调用以上方法继续分类叶样本，直到所有子集仅包含一个植被类别的样
本完成决策。实验结果表明，与光谱角制图和最大似然分类器相比，本文提出的方法整体精度
分别提高了 6.04%、2.92%，不仅证实了多分类器组合的可行性和有效性，而且更加适用于
AVIRIS 高光谱图像中的植被调查。

关键词: 高光谱图像, 遥感, SVM, SAM, MLC, C5.0

Abstract: Aiming at the problem that the vegetation classification accuracy of spectral angle mapping
(SAM) and maximum likelihood classifier (MLC) for AVIRIS hyperspectral remote sensing images
both are low, a vegetation classification method based on C5.0 decision tree of multiple classifiers is
proposed. During the first stage, the kernel function and its parameters are selected by using support
vector machine (SVM) to extract the vegetation information of AVIRIS hyperspectral image. Then,
SAM and MLC combined by C5.0 algorithm, as characteristic property of the decision tree, are used
to learn samples training and generate classification rules. According to the C5.0 algorithms, the
information gain rate of the corresponding classifier in the vegetation samples is calculated, and the
attributes with the largest information gain rate are selected to classify the samples. While
classification results of leaf sample meeting the stopped growing threshold, outputs the result of
sample classification, otherwise, go back to the beginning, recursively tune the above method and
goes on classifying leaf sample, until all the subset contains only the sample of a vegetation type, and
complete decision. Experiments show that comparing with SAM and MLC, overall accuracy of the
第 5 期 刘 丹，等：基于多分类器的 C5.0 决策树植被分类方法 723
proposed method were improved by 6.04%, 2.92%, respectively, not only confirmed the feasibility
and effectiveness of the combination of multiple classifiers, and it is more suitable for vegetation
investigation in the AVIRIS hyperspectral image.

Key words: hyperspectral image, remote sensing, SVM, SAM, MLC, C5.0