文献类型：期刊文献

中文题名：C5.0决策树Hyperion影像森林类型精细分类方法

英文题名：Precise classification of forest types use Hyperion image based on the C5.0 decision tree algorithm

作者：王怀警[1] 谭炳香[1] 房秀凤[1] 李世明[1] 李太兴[2]

第一作者：王怀警

机构：[1]中国林业科学研究院资源信息研究所遥感室;[2]吉林省白河林业局

年份：2018

卷号：35

期号：4

起止页码：724-734

中文期刊名：浙江农林大学学报

外文期刊名：Journal of Zhejiang A & F University

收录：CSTPCD;;北大核心:【北大核心2017】；CSCD:【CSCD2017_2018】；

基金：国防科工委高分辨率对地观测系统重大专项(30-Y20A37-9003-15/17-3);浙江省省院合作林业科技项目(2017SY04);国家自然科学基金资助项目(31370635)

语种：中文

中文关键词：森林经理学;Hyperion;C5.0决策树;分层分类;森林类型分类;高光谱

外文关键词：forest management;Hyperion;C5.0 decision tree;hierarchical classification;classification of forest types;hyperspectral

分类号：S757.4

摘要：以吉林省白河林业局为中心研究区,利用星载高光谱Hyperion数据并结合其他辅助数据,综合利用影像光谱特征、纹理特征、地形特征、典型地类和主要森林类型外业调查样本数据,探究针对C5.0决策树算法的高光谱影像土地覆盖类型多层次信息提取与森林类型识别的有效方法。在分析典型地物光谱特征的基础上,优选8种纹理特征,引入主成分分量及与主要森林类型空间分布相关的敏感地形因子,采用分层分类的策略,根据光谱特征将地类划分层次,在层次间建立基于C5.0决策树算法的决策树模型,对研究区的地类进行细分。为便于对比,以相同的策略采用支持向量机(SVM)分类器进行分类。最后,结合野外采集样本并参考高分辨率影像,采用分层随机抽样的独立检验样本对森林类型精细识别结果进行精度验证。结果表明:C5.0决策树算法可综合利用高光谱影像的光谱、纹理及其他辅助数据,自动寻找出区分各类别的最佳特征变量及分割阈值,运算速度快,占用内存较小且无需人为参与,其分类精度达到优势树种级别,总体分类精度达81.9%,Kappa系数0.709 8。
To explore the potential for using the C5.0 decision tree algorithm with hyperspectral data in precise classification of forest types, Baihe Forestry Bureau in Jilin Province was used as the center of the study area.Hyperion hyperspectral data, other auxiliary data, comprehensive utilization of an image＇s spectral feature, texture and terrain features, typical land types, and field investigation data of main forest types were used to support the study. Using the hierarchical classification strategy and according to characteristics of the spectrum,forest types were divided into different levels based on analysis of the typical spectral characteristics, eight optimum texture features, a principal component analysis, and sensitive terrain factors related to spatial distribution of the main forest types. Then the decision tree model based on the C5.0 decision tree algorithm was established for different levels. The robust Support Vector Machines（SVM） classifier was selected for comparison using the same strategy. Finally, verification samples were generated from the classification results using stratified random sampling method, combined with high-resolution remote sensing images, to evaluate the classification accuracy. Results showed that the C5.0 decision tree algorithm could comprehensively utilize the spectrum, texture, and other auxiliary data of hyperspectral images to automatically determine the best feature variables and segmentation thresholds for distinguishing each category. The algorithm also had the advantage offast computing speed, minimum occupied memory, and no human involvement. Classification accuracy reached the level of the dominant tree species, along with an overall classification accuracy of 81.9% and Kappa coefficient of 0.709 8. This method can be used for the precise mapping of forest types of hyperspectral remote sensing images, and can be used as reference for forest type classification of GF-5 hyperspectral images.