摘要
树种信息对林业资源监测和管理具有重要意义,及时准确地掌握树种及长势状况是防护林工程建设与效益评价的基础。为研究利用无人机高光谱数据进行防护林树种分类的效果,选取典型区域使用Matrice600型六旋翼无人机搭载Rikola高光谱成像仪获取高光谱影像,基于支持向量机递归特征消除算法(SVMRFE)选取原始波段最佳组合,再结合纹理特征、植被指数和数理统计特征,使用随机森林算法对所有特征进行重要性评估并与分类精度相结合进行特征优化,进而构建高光谱影像全波段、原始波段最佳组合、全部特征变量、基于随机森林(RF)特征优化后特征变量4种分类方案,分别采用最大似然法(MLC)、支持向量机(SVM)、随机森林对防护林优势树种进行分类。结果表明:所提出的基于交叉验证的SVMRFE算法选出的原始波段组合能更好地还原原始光谱特征;通过RF算法的特征重要性分析与分类精度相结合的方法可以有效选出重要特征,当使用全部特征的85%(包括17个光谱特征、3个纹理特征、5个植被指数和3个数理统计特征)进行分类时,总体精度最高为95.93%(Kappa系数为0.9475);所有特征中植被指数特征最重要,3种分类方法中RF算法分类总体精度(OA)最高。
Tree species information is of great significance to forestry resource monitoring and management,timely and accurate control of tree species and growth status is the basis for protection forest project construction and benefit evaluation.In order to study the effect of using UAV hyperspectral images to classify protection forest tree species,the advantages of UAV hyperspectral images high-resolution,multi-band,and short-period were used,taking the“Three North”protection forest which on the northern edge of the 150th Regiment of the Eighth Division of Xinjiang Production and Construction Corps as the research area,typical areas were selected and Matrice600 hexarotor drones equipped with Rikola hyperspectral imaging senstor was used to obtain hyperspectral images.Firstly,spectral features,textural features,vegetation indices(VIs)and characteristics of mathematical statics were extracted from the UAV hyperspectral image,the support vector machinerecusive feature elimination(SVMRFE)algorithm were used selection bands.Through the random forest algorithm to evaluate the importance of all features and combination with the overall classification accuracies was employed for feature reduction,and then four classification schemes of hyperspectral image full band,the best combination of original band,all feature variables,and feature variables based on random forest(RF)feature reduction were constructed.The classification results showed that the original band combination selected by the SVMRFE algorithm based on cross-validation proposed can better restore the original spectral features;when considering the four features(spectral features,textural features,hyperspectral Vis and mathematical statistics features)and after feature reduction,the three classifiers used,random forest(RF),maximum likehood classification(MLC)and support vector machine(SVM),the overall classification accuracies of RF was the highest,which were 95.93%,respectively.These results also suggested that vegetation indices were effective for discriminating tree species with similar spectral signatures.The overall results provided evidence for the effectiveness and potential of UAV hyperspectral data for protection forest tree species identification.
作者
赵庆展
江萍
王学文
张丽红
张建新
ZHAO Qingzhan;JIANG Ping;WANG Xuewen;ZHANG Lihong;ZHANG Jianxin(College of Information Science and Technology,Shihezi University,Shihezi 832000,China;Corps Space Information Technology Research Center,Shihezi 832000,China;School of Mechanical and Electrical Engineering,Shihezi University,Shihezi 832000,China;150 Regiment Agricultural Development Service Center,Shihezi 832000,China)
出处
《农业机械学报》
EI
CAS
CSCD
北大核心
2021年第11期190-199,共10页
Transactions of the Chinese Society for Agricultural Machinery
基金
新疆生产建设兵团科技计划项目(2017DB005)
兵团空间信息工程技术中心创建项目(2016BA001)
中央引导地方科技发展专项资金项目(201610011)。
关键词
树种分类
无人机
高光谱
特征挖掘
随机森林
tree species classification
unmanned aerial vehicle
hyperspectral
feature mining
random forests
作者简介
赵庆展(1972—),男,教授,主要从事农业信息化、空间信息系统集成与服务研究,E-mail:zqz_inf@shzu.edu.cn。
