基于分层分级的遥感图像植被分类方法

Vegetation classification of remote sensing image based on hierarchical clustering analysis

以陕西省榆阳区2014年和2002年8月份的OLI/ETM+图像作为基础数据源,比较分层分级方法与3种直接监督分类方法识别7种主要植被类型的精度.对去除非植被信息后的解译底图依次进行小波滤波、L-V特征空间分析,生成包含目标植被类型和与其光谱特征近似地物的一级图层;在一级图层上,依据地物间光谱特征和形态特征差异执行面向对象分割、支持向量机(SVM)监督分类、数学形态学开闭运算,生成只包含目标植被类型的二级图层;精度评价合格后从解译底图上去除该植被类型的覆盖区域;按上述方法依次处理灌溉耕地、旱耕地、有林地及果园、灌木林地、高盖度草地及低盖度草地5个专题图层,复合各专题图层的提取结果形成一期遥感分类图;与SVM,BP神经网络(BPANN)、最大似然法(MLC)监督分类结果比较分类精度.分层分级方法有效降低椒盐效应,减少混分漏判现象:使运行性陆地摄像仪(OLI)图像总体分类精度分别提高了6.85%,9.13%,18.21%,Kappa系数分别提高了8.03%,10.65%,21.27%;使ETM+图像总体分类精度分别提高了8.10%,10.43%,17.78%,Kappa系数分别提高了9.26%,12.01%,21.15%.

This paper aims to seek out the most suitable method for vegetation classification of OLI/ETM＋image in 2014 and Aug.2002 in Yuyang district,by the comparison study of supervised classification based on SVM/BPANN/MLC and hierarchical clustering analysis.Firstly,non-vegetation information was removed from original OLI/ETM＋image,and then it was used as interpretation base map for further processing.Secondly,wavelet filtering,threshold segmentation in L-V feature space were carried out in interpretation base map to obtain primary layer covering target vegetation types and terrestrial objects with similar spectral characteristics.In primary layer,object oriented segmentation,supervised classification based on SVM,opening-closing operation in mathematical morphology were carried out to get secondary layer of precise target vegetation types.Thirdly,after accuracy evaluation in secondary layer,coverage area of target vegetation type was removed from interpretation base map for next vegetation type.By repeating the preceding three steps,five kinds of secondary layer were obtained as thefollowing order：irrigable land,dry land,woodland and orchard,shrubbery,high coverage grassland and low coverage grassland.Next,classification results of the five secondary layers were recombined together to acquire the vegetative classification results of OLI/ETM＋image in Yuyang district.Finally,classification accuracy of supervised classification based on SVM/BPANN/MLC and the suggested method was compared.Research results show that the method of hierarchical clustering analysis proposed in this paper is more effective in removing salt and pepper noise in classified images.In OLI image,growth rates of 6.85%,9.13%,18.21%for overall classification accuracy and 8.03%,10.65%,21.27% for Kappa coefficient were achieved compared with supervised classification based on SVM/BPANN/MLC.In ETM＋image,growth of 8.10%,10.43%,17.78% for overall classification accuracy and 9.26%,12.01%,21.15%for Kappa coefficient were achieved compared with supervised classification based on SVM/BPANN/MLC.