应用面向对象结合多时相哨兵-2A影像特征优选的毛竹林分布信息提取  

Mapping Moso Bamboo Forest Distribution in A Subtropical Region Using A Random Forest Classifier and Multi-temporal Sentinel-2A Data

作　　者：张旷典 郭孝玉[2] 康继 刘健[3] Zhang Kuangdian;Guo Xiaoyu;Kang Ji;Liu Jian(Fujian Agriculture and Forestry University,Fuzhou 350028,P.R.China;Sanming University;Fujian Key Laboratory of Resources and Environment Monitoring and Sustainable Management and Utilization,Sanming University)

机构地区：[1]福建农林大学,福州350028 [2]三明学院 [3]福建省资源环境监测与可持续经营利用重点实验室(三明学院)

出　　处：《东北林业大学学报》2023年第1期61-68,87,共9页Journal of Northeast Forestry University

基　　金：国家自然科学基金项目(41801279);福建省自然科学基金项目(2019J01820);国家级大学生创新训练项目(202011311011)。

摘　　要：提出一种面向对象及随机森林特征优选的分类方法,为毛竹林分布信息提取提供参考。依据多时相哨兵-2A(Sentinel-2A)卫星数据提取光谱特征、植被指数及红边植被指数特征、纹理特征共69个特征;设计8种特征组合方案,方案1~5为多时相方案,其中方案1——光谱特征波段;方案2——光谱特征波段+植被指数特征+红边植被指数特征;方案3——光谱特征波段+纹理特征;方案4——光谱特征波段+植被指数特征+红边植被指数特征+纹理特征;方案5——光谱特征波段+植被指数特征+红边植被指数特征+纹理特征+特征优选;方案6、7和8为3个单时相影像分类,将其分类结果与其他多时相方案进行对比。采用随机森林算法进行特征优选的毛竹林分布信息提取。结果表明:(1)多时相Sentinel-2A数据的短波红外波段特征、红边波段特征及红边植被指数特征在分类时重要性程度高,对毛竹林分布信息提取贡献度大;(2)使用随机森林面向对象的分类方法能够有效的减少“椒盐现象”;(3)所有特征参与并由随机森林算法特征优选的方案5对毛竹林的分布信息提取效果最佳,总体精度为85.94%,Kappa系数为0.7852,表明随机森林算法能够进行特征优选同时保持精度较高的毛竹林提取效果。因此,面向对象结合多时相Sentinel-2A影像,利用随机森林进行特征优选和分类的方法能够较为有效地提取毛竹林分布信息。A classification method of object-oriented and random forest feature optimization was proposed to provide reference for the extraction of distribution information of moso bamboo forest.By using multi temporal sentinel-2A satellite data,69 features including spectral features,vegetation index,red edge index and texture features are extracted;Five feature combination schemes are designed,scheme 1:spectral features,scheme 2:spectral features+vegetation index features+red edge index features,scheme 3:spectral features+texture features,scheme 4:spectral features+vegetation index features+red edge index features+texture features,and scheme 5:spectral features+vegetation index features+red edge index features+texture features;Scheme 6,scheme 7 and scheme 8 are three single temporal image classification,and their classification results are compared with other multi temporal schemes;Feature selection and distribution information extraction of moso bamboo forest were carried out by using object-oriented and random forest algorithm.The results are as follows:(1)The short wave infrared band features,red edge band features and red edge index features of multi temporal sentinel-2A data are of high importance in classification,and contribute greatly to the extraction of distribution information of moso bamboo forest;(2)The object-oriented classification method based on random forest can effectively reduce the“salt and pepper phenomenon”;(3)Scheme 5 with all features involved and optimized by random forest algorithm has the best extraction effect on the distribution information of moso bamboo forest,with the overall accuracy of 85.94%and kappa coefficient of 0.7852,which indicates that random forest algorithm can optimize the features and maintain the extraction effect of high accuracy for moso bamboo forest.Consequently,by using object-oriented and multi temporal sentinel-2A image,the method of feature selection and classification using random forest can extract the distribution information of moso bamboo forest effectively

关 键 词：面向对象 哨兵-2A卫星 随机森林算法 特征优选 毛竹林分布信息 

分 类 号：S757.2[农业科学—森林经理学]