湿地类型遥感影像分割最优尺度选择

Select optimal segmentation scale of wetlands using remote sensing data

影像分割是面向对象分析方法中的重要步骤和关键技术,分割尺度又是影像分割的核心问题,目前对于分割尺度的定量评价,国内外尚未建立统一的评价体系。以东洞庭湖为研究区,以高分一号卫(GF-1)卫星影像为研究对象,探索不同湿地类型的最优分割尺度。利用GF-1全色和多光谱融合后影像,采用归一化后方差均值与Moran’I指数构建全局评分指数,得到不同湿地类型的最优分割尺度,结果表明:最优分割尺度的大小与湿地分布的破碎程度直接关系;基于相同的有效影像特征以及波谱阈值,进行分层分类以及规则分类,分层分类的总体精度为89.62%,高于规则分类的总体精度为84.43%,说明分割尺度对分类精度有较大影响。

Image segmentation is the basis steps and key technology of object-oriented analysis methods. And the segmentation scale is the core of segmentation. However, the uniform evaluation system for quantitative evaluation of segmentation scale was not established yet. Dong Dongting Lake was the study area in this paper. Its purpose is to explore the optimal segmentation scale of different wetland types for GF-1 remote sensing images. After use the normalized average variance and normalized Moran‘I to build a global score index based the fusion image of panchromatic and multispectral GF-1 data, it can find the optimal segmentation scale of different wetland types. What’s more, the direct relationship can be found between the optimal segmentation scale and the size of its crushing distribution. Then, by compare the overall classiifcation accuracy of hierarchical classiifcation and rules classiifcations which were based the same effective image feature and spectrum threshold, the result show that the overall accuracy of hierarchical classiifcation that is 89.62% is higher than the overall accuracy of rules classiifcation which is 84.43%. It can be drawn that segmentation results have a greater impact on the classiifcation accuracy.