地球信息科学学报 ›› 2013, Vol. 15 ›› Issue (5): 744-751.doi: 10.3724/SP.J.1047.2013.00744

• 遥感科学与应用技术 • 上一篇    下一篇

一种面向对象的像元级遥感图像分类方法

李小江1,2, 孟庆岩1, 王春梅1, 刘苗1, 郑利娟1, 王珂1   

  1. 1. 中国科学院遥感与数字地球研究所 遥感科学国家重点实验室, 北京 100101;
    2. 中国科学院大学, 北京 100049
  • 收稿日期:2013-03-07 修回日期:2013-05-13 出版日期:2013-09-29 发布日期:2013-09-29
  • 通讯作者: 孟庆岩(1971-),男,博士,研究员,主要从事遥感应用研究。E-mail:mqy@irsa.ac.cn E-mail:mqy@irsa.ac.cn
  • 作者简介:李小江(1988-),男,硕士生,主要从事环境遥感研究。E-mail:lixiaojiang.good@163.com
  • 基金资助:

    欧盟第七框架项目(FP7-PEOPLE-2009-IRSES-IGIT);广东省省院产学研合作资金项目(2012B091100219)、中国科学院对外重点合作项目(GJHZ1003)。

A Hybrid Model of Object-Oriented and Pixel Based Classification of Remotely Sensed Data

LI Xiaojiang1,2, MENG Qingyan1, WANG Chunmei1, LIU Miao1, ZHENG Lijuan1, WANG Ke1   

  1. 1. State Key Laboratory of Remote Sensing Science, Beijing 100101, China;
    2. Graduate University of Chinese Academy of Sciences, Beijing 100049, China
  • Received:2013-03-07 Revised:2013-05-13 Online:2013-09-29 Published:2013-09-29

摘要:

本文提出一种面向对象的像元级分类方法(混合模型),并将其与单纯的以像元和面向对象的两种方法同时应用于分辨率分别为30m和0.5m的环境星CCD数据和航空影像进行对比分析。分类结果中不同地物类别之间光谱可分性的大小,很大程度上可反映分类结果的可靠性。若地物类型之间的光谱差异大,说明分类方法能将光谱差异大的地物很好地划分,显示出较可靠的分类结果;相反,如果分类结果中地物类型光谱差异小,则反映分类方法不够可靠。鉴此,本文通过计算分类结果中不同类别所对应的原始遥感影像像元之间的J-M(Jeffries-Matusita Distance)距离来度量分类结果中地物之间的光谱可分性,并用J-M距离比较分析了3种图像分类方法对2种不同分辨率影像的分类结果中各个类别之间的光谱可分性的变化。分析结果表明,混合模型不但能够得到较连续的分类结果,同时能够保持分类结果中类别之间的可分性。本文对分类结果进行了精度验证,结果发现混合模型的分类精度较其他2种方法要高。2种不同分辨率的遥感影像分析结果得到相同的结论,表明该模型适用于中分辨率和高分辨率影像。

关键词: 多种分辨率, 面向对象, 混合模型, J-M(Jeffries-Matusita)距离, 地物光谱可分性

Abstract:

Three classification methods: maximum likelihood at pixel level, maximum likelihood at object level and a hybrid model of pixel based and object-based classification method were used to classify these two sorts of images, i.e., 30m resolution HJ-1-A CCD imagery and 0.5 m resolution aerial images. Spectral separability in training samples between different classes directly affect the classification accuracy; on the contrary, spectral separability between different classes in the classification result can be used to evaluate the accuracy of classification. Based on this assumption, classification results of these three methods were evaluated by spectral separability between different classes. Result reveals that compared with the maximum likelihood at object level, classification at pixel level can make full use of the spectral information in remotely sensed image. Analysis also shows that hybrid model can improve the classification accuracy. Multi-resolution image analysis proves that the hybrid model is suitable for applications at different scales.

Key words: object-oriented, hybrid model, J-M(Jeffries-Matusita) Distance, multi-resolution, spectral separability