文献类型：期刊文献

中文题名：面向对象CART决策树方法的湿地遥感分类

英文题名：Remote Sensing Classification of Wetlands based on Object-oriented and CART Decision Tree Method

作者：姚博[1] 张怀清[1] 刘洋[1] 刘华[1] 凌成星[1]

第一作者：姚博

机构：[1]中国林业科学研究院资源信息研究所

年份：2019

卷号：32

期号：5

起止页码：91-98

中文期刊名：林业科学研究

外文期刊名：Forest Research

收录：CSTPCD;;Scopus;北大核心:【北大核心2017】；CSCD:【CSCD2019_2020】；

基金：国家林业科学数据平台项目(2005DKA32200-04);国家自然科学基金项目(31370712);国家重大专项(21-Y30B05-9001-13/15-2)

语种：中文

中文关键词：CART决策树;湿地信息;湿地类型;北京地区;Landsat8;OLI

外文关键词：CART Decision Tree;wetland information;wetland types;Beijing;Landsat8 OLI

分类号：TP751

摘要：[目的]以北京地区为例,对大区域多类型湿地信息提取方法进行研究。[方法]采用面向对象多尺度分割算法及光谱差异分割算法分割Landsat8 OLI遥感影像,辅助Google Earth高清影像及2015年人工解译结果,使用分层抽样法随机产生训练样本与验证样本;综合运用光谱、形状、纹理特征及拓扑关系,构建CART决策树模型提取研究区湿地信息;与最大似然法、面向对象最邻近方法的分类结果进行对比。[结果]利用面向对象CART决策树方法,分类结果的总精度为88.05%,Kappa系数为0.844,相较于面向对象最邻近方法,总体精度、Kappa系数相差不大,但针对部分湿地类型,如河流、沼泽湿地,精度提高了10%~20%;比使用最大似然分类法的总精度高近30%,Kappa系数提高0.355。[结论]对于湿地分布广泛、类型及数量较多的地区,面向对象CART决策树方法分类结果较好,是一种快速、有效的分类方法。
[Objective] Beijing was taken as the research field to discuss the extraction methods for classification of multi-type wetlands in large areas.[Method] Object-oriented multi-scale segmentation algorithm and spectral difference segmentation were used to segment the Landsat8 OLI image, and stratified sampling method was used to generate random training samples and validation samples by Google Earth high-definition image and manual interpretation results for 2015. Subsequently, CART Decision Tree was constructed to extract wetland information by combining spectral, shape, texture features and topological relation. The results were compared with maximum likelihood method and object-oriented and nearest neighbor method.[Result] Using object-oriented and CART Decision Tree, the total accuracy of the results was 88.05%, and the Kappa coefficient was 0.844. Compared with the object-oriented and nearest neighbor, the overall accuracy and Kappa coefficient showed less difference, but for some wetland types, such as rivers and swamp, the accuracy increased by 10% to 20%;the total accuracy was nearly 30% higher than that of the maximum likelihood classification, and the Kappa coefficient increased by 0.355.[Conclusion] Object-oriented and CART Decision Tree is a fast and effective method for wetland classification in the areas with wide distribution, multi-types and quantities of wetland.