基于规则的面向对象分类与监督分类对比研究——以WorldView-2影像为例

Comparative Research on Rule-based Object-oriented Classification and Supervised Classification——Taking WorldView-2 Image as an Example

针对传统监督分类方法在处理高分辨率影像时会产生“同物异谱,同谱异物”的现象,本文以杭州市WorldView-2遥感影像数据作为基础数据源,设计了一种基于规则的面向对象分类方法(ROOC)以提取研究区域地物,提高地物分类精度。首先,针对不同地物的光谱特征差异构建分类规则函数;其次,依据地物图斑在不同分割层下的大小及形状不一致的特征,引入特征波段并参与多尺度分割;最后,将光谱特征分类规则函数与多尺度分割结果相结合构建ROOC方法,通过目视评价和定量分析的方法与MLC、SVM和NNC 3种监督分类结果进行对比。结果表明,ROOC方法的总体分类精度为81.5%,较MLC、SVM、NNC分别提高了10.0%、10.5%、8.0%;总体Kappa系数为0.7685,较MLC、SVM、NNC分别提高了0.1278、0.1295、0.1057。因此,ROOC方法通过将地物光谱特征差异与多尺度分割相结合,能够更准确地识别WorldView-2影像中光谱特征相近的地物,能有效减少地物混分现象,提高分类精度。

In view of the phenomenon of“the same object with different spectrum while the same spectrum extracted from the different objects”when traditional supervised classification methods in dealing with highresolution images,a rule-based object-oriented classification method(ROOC)is designed to extract the objects in the study area and improve the classification accuracy of the objects based on WorldView-2 image data source of Hangzhou.Firstly,the classification rule function is constructed according to the difference in spectral characteristics of different features;secondly,based on the inconsistent features of the size and shape under different segmentation layers of the map spots,multi-scale segmentation is performed on the image and feature bands are added to participate in the segmentation;finally,the spectral feature classification rule function is combined with the multi-scale segmentation results to construct the ROOC method,and it is compared with the MLC,SVM and NNC these three supervised classification methods through visual evaluation and quantitative analysis.The results show that the overall accuracy of ROOC method is 81.5%,whose classification accuracy increases by 10.0%,10.5%and 8.0%,respectively;the overall Kappa coefficient of ROOC method is 0.7685,which increases by 0.1278,0.1295 and 0.1057,respectively.Therefore,by combining the difference of ground spectral features with multi-scale segmentation,ROOC method can more accurately identify the ground features with similar spectral features in WorldView-2 images,which effectively reduces the phenomenon of mixed features and improves the classification accuracy.