高分二号影像树种分类最优分割尺度的非监督选择

Unsupervised Selection of Optimal Scale Parameters for Tree Species Classification Based on GF-2 Images

良好的影像分割结果能够提高树种分类的精度,而分割效果取决于最优分割尺度(optimal scale parameter,OSP)的选择。以往研究依赖人为设置的尺度序列,结果具有主观性。为避免此问题,以高分二号影像(GF-2)为数据源,提出一种基于有效尺度区间的非监督OSP选择方法,用于确定不同森林类型最佳分割结果出现的分割尺度。影像分割采用多分辨率分割(multi-resolution segmentation,MRS)算法,通过构建有效尺度区间估计函数(effective scale interval estimation functions,ESF),结合总体优度函数(overall goodness F-measure,OGF)得出不同森林类型在不同尺度区间下的OSP,最后依据监督分割精度分析结合谷歌地图目视判读确定最佳分割结果。结果表明,OGF在有效尺度区间Ⅲ获取的OSP得到了各森林类型的最佳分割结果,监督分割评价方法(F-measure)的最低和最高值分别为0.7311和0.8733。同时,在GF-2影像树种分类的分割任务中,OSP与树种和森林类型有关。研究结果为高分辨率遥感影像树种分类的对象提取提供技术支撑,同时为不同地物组成的复杂影像分割尺度参数选择提供方法。

Satisfactory image segmentation results can improve the accuracy of tree species classification,and the segmentation ef⁃fect depends on the selection of the optimal scale parameter(OSP).Previous studies have relied on manually set scale sequences,result⁃ing in subjectivity.To avoid this issue,this study used GF-2 images as the data source and proposed an unsupervised selection method of OSP based on effective scale intervals to determine at which scale the optimal segmentation results for different forest types occur.Multi-resolution segmentation(MRS)algorithm was used to segment images,and constructing effective scale interval estimation func⁃tions(ESF)and combining with the overall goodness F-measure(OGF),the OSPs of different forest types at different scale intervals were obtained.Finally,the optimal segmentation results were determined by supervised segmentation accuracy analysis combined with Google map visual interpretation.The results showed that the OSPs obtained from the effective scale intervalⅢachieved the best seg⁃mentation results for each forest type,with the lowest and highest F-measure of 0.7311 and 0.8733,respectively.Meanwhile,in the segmentation task of tree species classification based on GF-2 image,OSP was related to tree species and forest types.This paper pro⁃vided technical support for object extraction of object-based tree species classification based on high-resolution remote sensing images and also provided a methodological reference for the selection of scale parameters for images composed of different geographical objects.