改进Fisher判别分类的遥感影像变化检测  

Change detection of remote sensing images based on modified fisher classification

作　　者：陈科[1] 张保明[1] 谢明霞[1,2] 

机构地区：[1]信息工程大学测绘学院 [2]75719部队

出　　处：《测绘科学》2010年第4期160-162,96,共4页Science of Surveying and Mapping

摘　　要：Fisher判别分类的好坏关键在于训练样本集选取的精度和在降维过程中样本特征信息的损失程度,基于此问题,本文根据不同时相同一地区的遥感影像的差值影像中各像素本身的灰度值及其邻域平均灰度值特征获得其一维和二维直方图,针对差值影像无噪和带噪两种情况,根据直方图信息选取Fisher判别分析所需的训练样本,同时为了尽可能降低判别分析过程中有用信息的损失,将所得到的原训练样本集进行非线性变换,使其映射到高维空间中,利用映射后的训练样本求得Fisher判别规则。实验结果表明:与基于原训练样本的Fisher判别分类和基于寻找更多的样本特征的Fisher判别分类方法生成结果相比,在差值影像无噪和带噪情况下,本文提出的方法具有更好的变化检测精度和抗噪性。The keys of effectiveness evaluation of fisher classification are the precision of choosing training samples and the degree of samples information decrease. The paper firstly used multi-temporal remote sensing images of the same region to gain the one-dimen- sional histogram and two-dimensional histogram and to choose training samples aiming at noise and non-noise image according to those histograms. Secondly, it mapped the training samples to higher dimensional space through nonlinear transform in order to minimize the loss of useful information during the process of fisher classification. Finally, it gained the rule of classification by using the mapped training samples to classify the unlabelled pattern in remote sensing images. Through experiments, it was proved that the change detec- tion approach of remote sensing images based on decision rules of the paper had higher precision and capability of anti-noise comparing with fisher classification based on initial training samples in lower dimension and that based on seeking much more sample features.

关 键 词：Fisher判别分类 变化检测 非线性变换 直方图 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]