一种利用多尺度直方图的遥感影像分类方法  

An Image Classification Method Based on Multi-scale Histogram

作　　者：李亮[1] 王蕾[1] 李胜[1] 应国伟[1] 

机构地区：[1]四川省第三测绘工程院,成都610500

出　　处：《遥感信息》2017年第3期149-154,共6页Remote Sensing Information

基　　金：测绘地理信息公益性行业科研专项经费(201512026);四川测绘地理信息局科技支撑(J2014ZC12;J2014ZC16)

摘　　要：针对单一直方图对像斑特征表达不充分的问题,提出了一种利用多尺度直方图的遥感影像分类方法。首先,通过影像分割获取像斑,选取训练样本像斑;其次,提取像斑在不同灰度级下的直方图,形成像斑的多尺度直方图特征;再次,利用G统计量度量各尺度下的直方图距离,加权组合直方图距离构建像斑在单波段上的特征距离;然后,计算各波段不同灰度级下的信息熵,自适应确定各波段对应的权重,加权组合单波段特征距离构建像斑的特征距离;最后,依据像斑特征距离最小的原则,获取影像分类结果。在QuickBird遥感影像上的实验表明,与单一尺度直方图分类法相比,该方法的分类精度较优。In order to solve the problem that single histogram is not sufficient to describe the feature of the image object,a classification method based on multi-scale histogram for remote sensing images was proposed in the paper.Firstly,training sample objects were chosen after image segmentation.Secondly,histograms of objects at different gray levels were extracted,and Gstatistic was used to measure the distance between histograms at different levels.And then,feature distance on the single band was built by a weighted combination of histogram distance at different gray levels.Next,an adaptive method for determining the weight was used according to the entropy of the band at different gray levels.Finally,image classification results were obtained according to the minimum feature distance of the objects which was built by a weighted combination of feature distance on the single band.The experimental result on the QuickBird image shows that the proposed method can provide higher accuracy,compared to the classification method by the single-scale histogram.

关 键 词：直方图 影像分类 像斑 G统计量 最小距离 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]