文献类型：期刊文献

中文题名：基于证据理论组合多分类规则实现大区域植被遥感分类研究

英文题名：Combining Multiple Classifiers Based on Evidence Theory for Large Scale Vegetation Types Classification by Remote Sensing Images

作者：胡博[1] 鞠洪波[1] 刘华[1] 郝泷[1] 刘海[1]

第一作者：胡博

机构：[1]中国林业科学研究院资源信息研究所

年份：2017

卷号：30

期号：2

起止页码：194-199

中文期刊名：林业科学研究

外文期刊名：Forest Research

收录：CSTPCD;;Scopus;北大核心:【北大核心2014】；CSCD:【CSCD2017_2018】；

基金：国家863计划课题(2012AA102001)

语种：中文

中文关键词：遥感;大区域;证据理论;植被分类

外文关键词：remote sensing; large area; evidence theory; vegetation classification

分类号：S771.8

摘要：[目的]利用遥感影像的时效性和宏观性特点,基于证据理论组合多分类规则的方法快速和高效地实现大区域植被遥感分类。[方法]首先,依据辨识框架的概念设计分类系统,并采用大区域样本快速采集方法提取训练样点;其次,将多个单分类规则得到的植被类型特征影像归一化处理为基本概率赋值作为表达对各类型信任程度的证据源数据,再将不同证据源的信任度信息依据证据理论组合;再次,将组合结果依据最大信任度原则确定植被类型;最后,在中国植被图与中国土地覆盖图的类型一致区域随机布点作为验证样本。[结果]各单分类器分类结果的总体精度范围为60%70%,两两规则组合分类结果的总体精度范围为70%80%,3个规则组合分类结果的总体精度达到80.84%。[结论]组合多分类规则的证据理论分类方法可以提高分类精度;参与组合的单分类器精度越高,相关证据源越多,组合分类结果精度越高。
Based on the evidence theory principle, the research will realize a combination of multiple classifiers quickly and efficiently for large scale vegetation types classification according to the temporal and the extensive features of remote sensing images. [Method]The classification system imitated the frame of discernment concept and extracted training samples with quick sampling obtaining method for large area vegetation. Taking the feature images of vegetation types obtained by different single classifiers as evidence sources, the feature images were normalized to the basic probability assignment for expressing the credibility and the basic probability assignments were combined based on the combination rules of evidence theory. The combination results were classified by cumulative belief value principle.[Result]The single classifier＇s accuracy range was 60%~70% while the pairwise combinatorial classifier＇s accuracy range was 70%~80%, but the combination of three classifiers accuracy was 80.84%. [Conclusion]The results showed that multi-classification based on evidence theory can improve the classification precision. The higher the single classifier＇s accuracy and the more the related evidence sources, the higher the classification results＇ accuracy would be.