基于多分类器集成的沅陵县天然林类型信息提取研究  

Research on the extraction of natural forest type information in Yuanling based on the integration of multiple classifiers

作　　者：张颖 肖越 鲁宏旺 龙江平 林辉[1,2,3] ZHANG Ying;XIAO Yue;LU Hongwang;LONG Jiangping;LIN Hui(Research Center of Forestry Remote Sensing&Information Engineering,Central South University of Forestry&Technology,Changsha 410004,Hunan,China;Key Laboratory of Forestry Remote Sensing Based Big Data and Ecological Security for Hunan Province,Changsha 410004,Hunan,China;Key Laboratory of State Forestry and Grassland Administration on Forest Resources Management and Monitoring in Southern Area,Changsha 410004,Hunan,China;Changsha Changchang Forestry Technical Consulting Co.,Ltd.,Changsha 410004,Hunan,China)

机构地区：[1]中南林业科技大学林业遥感信息工程研究中心,湖南长沙410004 [2]林业遥感大数据与生态安全湖南省重点实验室,湖南长沙410004 [3]南方森林资源经营与监测国家林业与草原局重点实验室,湖南长沙410004 [4]长沙市长长林业技术咨询有限责任公司,湖南长沙410004

出　　处：《湖南林业科技》2023年第3期28-36,共9页Hunan Forestry Science & Technology

基　　金：湖南省重点研发计划(2020NK2051)。

摘　　要：对天然林进行分类能够掌握天然林的林分组成与生长状况,是对天然林进行保护的前提,然而由于天然林林内林分组成复杂、龄组不一致等原因,使得如何使用遥感技术准确且有效地提取天然林信息成为了亟待解决的问题。针对于此,本研究以Sentinel-2影像为数据源,提取光谱特征、植被指数以及纹理特征等31个分类特征,采用随机森林算法进行了特征筛选。结合分层分类的思想,将最大似然算法(ML)、神经网络算法(ANN)、支持向量机算法(SVM)以及随机森林算法(RF)4种单分类器以加权投票的策略进行了集成(EL),以期提高天然林信息提取精度。结果表明:RF为4种单分类算法中的最佳算法,但EL的表现更为优秀,其总体精度达到了87.18%,相较于RF、SVM、ANN和ML分别提高了4.13、7.94、8.86、9.01个百分点;EL的Kappa系数达到了0.82,展现出了极佳的分类性能,表明EL能够有效提高天然林分类的精度。The classification of natural forests is a prerequisite for the protection of natural forests,as it enables the composition and growth of natural forests to be understood.However,the complex composition and inconsistent age groups of natural forests make the use of remote sensing technology to extract accurate and effective information about natural forests an urgent problem.For this purpose,31 classification features,including spectral features,vegetation indices and texture features,were extracted using Sentinel-2 images as the data source,and the random forest algorithm was used for feature selection.Combining the idea of hierarchical classification,four single classifiers,Maximum Likelihood(ML),Artificial Neural Net(ANN),Support Vector Machine(SVM),and Random Forests(RF),are integrated with a weighted voting strategy in order to improve the accuracy of natural forest information extraction.The experimental results showed that RF was the best algorithm among the four single classification algorithms,but EL performed better with an overall accuracy of 87.18%which was 4.13%,7.94%,8.86%and 9.01%better than RF,SVM,ANN and ML,respectively.The Kappa coefficient of EL reached 0.82,showing an excellent classification performance,indicating that EL could effectively improve the accuracy of natural forest classification.

关 键 词：遥感信息提取 天然林分类 集成学习 分层分类 Sentinel-2 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置]