多特征优选的Sentinel-2A影像随机森林分类研究  

A Methodology of Random Forest Classification for Sentinel-2A Image Based on Multi-feature Optimization

作　　者：陈果[1,2] 李乐林[1,2,3] 陈浩 彭焕华[1,3] 赵茜 CHEN Guo;LI Lelin;CHEN Hao;PENG Huanhua;ZHAO Xi(Hunan Provincial Key Laboratory of Geo-Information Engineering in Surveying,Mapping and Remote Sensing,Hunan University of Science and Technology,Xiangtan 411201,China;National-Local Joint Engineering Laboratory of Geo-Spatial Information Technology,Hunan University of Science and Technology,Xiangtan 411201,China;Department of Earth Science and Spatial Information Engineering,Hunan University of Science and Technology,Xiangtan 411201,China)

机构地区：[1]湖南科技大学测绘遥感信息工程湖南省重点实验室,湖南湘潭411201 [2]湖南科技大学地球科学与空间信息工程学院,湖南湘潭411201 [3]湖南科技大学地理空间信息技术国家地方联合工程实验室,湖南湘潭411201

出　　处：《测绘与空间地理信息》2023年第3期19-23,共5页Geomatics & Spatial Information Technology

基　　金：国家科技基础资源调查专项(2019FY202502);湖南省教育厅科学研究优秀青年项目(17B093);国家级大学生创新训练项目--面向乡村振兴的融合多源遥感数据返贫监测研究(202210534038)资助。

摘　　要：为了探寻更高效、更准确的土地利用分类方法,本文选取遥感光谱、影像纹理特征、植被指数、生物物理指数及地形特征信息,构建多特征随机森林土地利用分类方案。以山西省晋中市祁县为例,设计7种方案对研究区域Sentinel-2A影像数据进行分类,采用混淆矩阵进行精度评估。研究结果显示,合理地选取特征变量可以提高随机森林分类的精度,结合不同特征组合随机森林分类精度和袋外数据误差进行特征重要性比较:土壤反射性指数>生物物理指数>纹理指数>地形特征;在相同分类条件下与其他机器学习分类方法(支持向量机、神经网络)相比,基于多特征随机森林优选的方法总体精度达到91.96%,Kappa系数为0.902,提取效率更快、精度更高。In order to explore more efficient and accurate land use classification methods,this paper selects remote sensing spectrum,image texture features,vegetation index,biophysical index and terrain feature information to construct a multi-feature random forest land use classification scheme.Taking Qixian county,Jinzhong city,Shanxi province as an example,seven schemes were designed to classify Sentinel-2A image data in the study area,and the confusion matrix was used to evaluate the accuracy.The research results show that the reasonable selection of feature variables can improve the accuracy of random forest classification,and combine the random forest classification results with different features and out-of-bag data errors to compare the feature importance:soil reflectivity index>biophysical index>texture index>topographic feature;Compared with other machine learning classification methods(support vector machine,neural network)under the same classification conditions,the overall accuracy of the method based on multi-feature random forest optimization reaches 91.96%,the Kappa coefficient is 0.902,and the extraction efficiency is faster,with higher accuracy.

关 键 词：土地利用分类 特征优选 随机森林 生物物理指数 Sentinel-2A 

分 类 号：P237[天文地球—摄影测量与遥感]