摘要
为探索高分辨率遥感影像对城市复杂环境优势乔木树种分类的有效性,采用面向对象分类方法,基于WorldView-2影像对首都师范大学及周边地区(CNU)、北京师范大学及周边地区(BNU)两个研究区进行优势乔木树种(泡桐、法国梧桐、杨树、国槐、银杏)分类。首先对WorldView-2影像进行分割,获得树冠区域及其49个属性特征,包括31个光谱属性和18个纹理属性;随后利用随机森林RF与支持向量机SVM两种分类算法对树冠区域进行分类。CNU研究区SVM与RF总体分类精度分别为86.5%、75.8%,Kappa系数为0.801、0.648;BNU研究区SVM与RF总体分类精度分别为66.9%、65.3%,Kappa系数为0.541、0.520。实验表明WorldView-2影像能有效实现城市非阴影区域优势乔木树种分类,但异质性较高、树种分布分散的区域分类精度低于异质性较小、树种分布密集的区域;WorldView-2影像的4个新增波段尤其是红边波段的派生属性在分类过程中所占权重值较高。
Urban trees play important roles in urban ecosystems,such as air quality improvement,carbon sequestration,and urban flood control.This study aimed to evaluate the utility of WorldView-2image（acquired on September 14,2012）for urban tree species identification.Two study sites in Beijing,China including CNU and BNU were examined and five tree species including Royal Paulownia（Paulownia Sieb.）,London Plane（Platanus acerifolia）,Chinese white poplar（Populus tomentosa Carrière）,Chinese Scholar tree（Sophora Japonica）,and Gingko（Ginkgo biloba L.）were classified.Tree species distribution in CNU is spatially more variable than those in BNU.First,urban tree crowns in non-shadowed areas were extracted using object-based hierarchical threshold method.A total of 49 features including 31 spectral features and 18 textural features were extracted for both areas.Two machine learning classifiers including Support Vector Machine（SVM）and Random Forest（RF）were applied for dominant tree species classification within tree crown areas.The results showed that overall accuracy（OA）in CNU area was higher than BNU area regardless of classification algorithm（86.5% vs.66.9% using SVM and 75.8% vs.65.3% using RF）,indicating that spatial variability of tree species distribution had negative effects on classification accuracy;greater variability resulted in lower classification accuracy.At both study sites,SVM with optimal parameters produced higher OA than RF.Besides,the features derived from 4-additional bands,especially red-edge band,were more important than traditional bands during classification.This study proved that WorldView-2images were effective in dominant tree species classification at non-shadowed urban area.Future study will explore tree species classification in shadowed area by shadow restoration.
出处
《地理与地理信息科学》
CSCD
北大核心
2016年第1期84-89,F0003,共7页
Geography and Geo-Information Science
基金
国家自然科学基金资助项目(41401493
41130744)
教育部博士点基金项目(20131108120006)
北京市科技新星项目(Z151100000315092)
中国水利水电科学研究院科研专项基金(JZ0145B042014)
作者简介
李丹（1988-），女，硕士研究生，主要研究方向为遥感与GIS应用。
通讯作者E-mail：yke@cnu．edu．cn
