应用多时序特征的哨兵系列影像对南方丘陵区树种识别  

Tree Species Identification in Southern Hilly Areas Using Sentinel Series Images with Multi-Temporal Features

作　　者：王洁[1] 李恒凯[1] 龙北平 张建莹 Wang Jie;Li Hengkai;Long Beiping;Zhang Jianying(Jiangxi University of Science and Technology,Ganzhou 341000,P.R.China;Jiangxi Provincial Coal Geology Bureau of Surveying and Mapping Brigade)

机构地区：[1]江西理工大学,赣州341000 [2]江西省煤田地质局测绘大队

出　　处：《东北林业大学学报》2024年第3期60-68,共9页Journal of Northeast Forestry University

基　　金：国家自然科学基金项目(42161057)。

摘　　要：树种分类是森林资源调查和监测的重要工作,杉木和油茶作为袁州区主要经济树种,准确获取树种空间分布信息,对产量估算和资源管理具有重要意义。以江西省宜春市袁州区为研究区,试验融合时序哨兵-1(Sentinel-1)、哨兵-2(Sentinel-2)等数据,结合中国南方丘陵区树种特点,提取植被指数、红边植被指数、地形特征和纹理特征等构建特征变量组合,分别利用分离阈值法(SEaTH)和特征权重算法(ReliefF)进行特征重要性排序和特征优选,分析各特征对树种分类的影响。结果表明:(1)在使用光谱特征和植被-水体指数的基础上加入不同特征后,树种分类精度均有提升,其中纹理特征的加入更有利于树种分类。(2)结合随机森林算法和特征权重算法(ReliefF)对树种分类的精度最高,总体精度为85.33%,Kappa系数为0.81,优于相同特征组下的支持向量机算法和分类回归树算法。Tree species classification is an important task in forest resource investigation and monitoring.In the Yuanzhou region,Chinese fir and camellia oil tree stand out as the primary economic tree species.Accurately obtaining spatial distribution information of tree species is of great significance for yield estimation and resource management.Taking the study area of Yuanzhou District,Yichun City,Jiangxi Province,the experiment fused the data of time-series Sentinel-1(Sentinel-1) and Sentinel-2(Sentinel-2),and combined with the characteristics of the tree species in the southern hilly areas of China,and extracted the vegetation indices,the red-edge vegetation indexes,the topographic features and the textural features to construct the combinations of the characteristic variables.The feature importance ranking and feature selection were performed using the SEaTH method and the ReliefF algorithm,respectively.The impact of each feature on tree species classification was analyzed.The results showed that:(1) The addition of different features improved the accuracy of tree species classification on the basis of spectral features and vegetation-water index,with texture features being more beneficial for tree species classification.(2) The combination of Random Forest algorithm and ReliefF algorithm achieved the highest accuracy in tree species classification,with an overall accuracy of 85.33% and a Kappa coefficient of 0.81,outperforming Support Vector Machine algorithm and Classification and Regression Tree algorithm with the same feature set.

关 键 词：树种分类 哨兵-1 哨兵-2 特征优选 随机森林 中国南方丘陵 

分 类 号：S771[农业科学—森林工程]