分割尺度对面向对象树种分类的影响及评价

Effect and Evaluation of Segmentation Scale on Object-Based Forest Species Classification

【目的】研究分割尺度对高空间分辨率遥感影像与星载全极化SAR数据协同面向对象树种分割与分类的影响,进而评价2种数据协同树种分类的适宜性。【方法】以QuickBird遥感影像和Radarsat-2数据为试验数据,在面向对象分类过程中采用3种分割方案(单独使用QuickBird遥感影像分割、单独使用Radarsat-2数据分割和QuickBird&Radarsat-2协同分割),每种分割方案采用10种分割尺度(25~250,步长为25),应用修正的欧式距离3(ED3_(modified))评价分割质量。对于3种分割方案采用各自特征及二者共同特征,分别应用支持向量机(SVM)分类器进行面向对象树种分类。【结果】在10种分割尺度上,QuickBird&Radarsat-2协同分割和单独使用QuickBird遥感影像分割的ED3_(modified)明显低于单独使用Radarsat-2数据分割获得的ED3_(modified)。QuickBird&Radarsat-2协同分割以分割尺度100进行分割的质量最好(ED3_(modified)=0.34)。3种分割-分类方案在小尺度上分类总精度(OA)较低,随着尺度增大,OA也再提高并在某个尺度达到最大值,之后OA随尺度增大而降低。QuickBird&Radarsat-2协同分割-分类在分割尺度100获得了最高分类精度(OA=85.55%;Kappa=0.86)。单独使用QuickBird遥感影像分割-分类在分割尺度150获得了最高分类精度(OA=81.11%;Kappa=0.82),单独使用Radarsat-2数据分割-分类在分割尺度125获得了最高分类精度(OA=66.67%;Kappa=0.68)。OA与ED3_(modified)高度相关(R^2=0.73)。【结论】在所有尺度(25~250)上,QuickBird&Radarsat-2协同使用的分割质量和分类精度高于单独使用其中一种数据源的分割质量和分类精度,相比单独使用Radarsat-2数据优势更加明显。分割尺度对面向对象树种分类结果有着重要影响。匹配良好的分割和参考对象能够得到更高精度的分类结果,同时,轻微的过度分割或分割不足不会明显影响分类结果。

【Objective】 The effects of different segmentation scales on the object-oriented tree species classification based on high spatial resolution remote sensing image and spaceborne polarimetric SAR data collaborated were studied,and the suitability of tree species classification based on the two kinds of data collaborated was also evaluated in this research.【Method 】 QuickBird remote sensing image and Radarsat-2 data are used as the experimental data.3 segmentation schemes,using QuickBird remote sensing image only,using Radarsat-2 data only,and using QuickBird and Radarsat-2 together,are applied in the object-oriented classification.In every segmentation scheme,10 segmentation scales（25-250,step 25） are adopted,and the modified Euclidean distance 3（ED3_（modified）） is used to to evaluate the segmentation quality.In the 3 segmentation schemes,the respective characteristics and the common characteristics are applied separately in support vector machine classifier to carry on object-oriented tree species classification.【Result】 On the 10 segmentation scales,the values of ED3_（modified） of segmentation with QuickBird and Radarsat-2 collaborated and QuickBird only are significantly lower than those with Radarsat-2 only.The best segmentation（ED3_（modified） = 0.34） is gotten at scale 100 with QuickBird and Radarsat-2 collaborated.On the 10 segmentation scales,the OA of 3 segmentationclassification schemes are low at the small scales.The OA improves as the scale becomes bigger,and reaches the maximum at a scale.Then the OA reduces with the scale increasing.The segmentation-classification using QuickBird andRadarsat-2 together gets the best accuracy at scale 100（OA = 85.55%; Kappa = 0.86）,and the scheme using QuickBird remote sensing image alone gets the best accuracy at 150（OA = 81.11%; Kappa = 0.82）,the scheme using Radarsat-2 data alone gets the best accuracy at 125（OA = 66.67%; Kappa = 0.68）,OA and ED3_（modified） are highly correlated（R^2=0.73）.【Conclusion】 At all scales（25-250）,the segmentation quality and accuracy of using QuickBird and Radardat-2 together are better than any other segmentation result and accuracies of using only one source of data,and has obvious advantages compared to only use Radarsat-2 data.Segmentation scale plays an important role in tree species classification.Good matching segmentation and reference objects can get higher classification accuracies.At the same time,the classification results are not obviously influenced by slightly over segmentation or insufficient segmentation.