训练样本数量选择和总体分类精度的关系研究  

Relationship between number of training samples selection and accuracy of overall classification

作　　者：王春来 张森原 崔璐 葛玉停 张金禄 张淼泓 

机构地区：[1]河南黄河勘测信息工程院,河南郑州450045 [2]黄河水文勘察测绘局,河南郑州450000

出　　处：《河南城建学院学报》2015年第3期51-55,60,共6页Journal of Henan University of Urban Construction

摘　　要：基于特征基元分类和基于像元分类的研究对象不同,训练样本数量的选择也有很大的区别。采用高分辨率遥感影像数据,利用SVM分类器,通过实验得出：基于特征基元分类训练样本选取的数量在6~8倍于波段数目时分类精度就能达到较高的水平,并进入一个稳定阶段;而基于像元分类训练样本选取数量在24~30倍于波段数时分类精度才达到较高的水平。The research objects between the classification based on the feature per- parcel and pixel are dif- ferent; there is a big difference in the choice of the number of training samples. In this paper, using high - resolution remote sensing image data and the classifier of SVM, the experiment showed that the accuracy can reach a high level based on feature per - parcel classification when the number of samples reach to 6 - 8 times of the band numbers and enter a stable phase ; but the accuracy can reach a high level based on pixel when the number of samples reach to 24 -30 times of the band numbers.

关 键 词：监督分类 特征基元 训练样本 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]