不同样本集划分策略对农作物遥感分类精度的影响  

Effects of Different Sample Set Partition Strategies on Crop Remote Sensing Classification Accuracy

作　　者：刘洋 李强子 杜鑫 王红岩 张源 张喜旺 沈云祺[1,2] 张思宸 余仕奇 LIU Yang;LI Qiangzi;DU Xin;WANG Hongyan;ZHANG Yuan;ZHANG Xiwang;SHEN Yunqi;ZHANG Sichen;YU Shiqi(Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100101,China;University of Chinese Academy of Sciences,Beijing 100049,China;Henan University,Zhengzhou 475001,China)

机构地区：[1]中国科学院空天信息创新研究院,北京100101 [2]中国科学院大学,北京100049 [3]河南大学,河南郑州475001

出　　处：《河南农业科学》2024年第6期144-153,共10页Journal of Henan Agricultural Sciences

基　　金：国家重点研发计划项目(2021YFD1500103);中国科学院战略性先导科技专项(XDA28070504);国家自然基金面上项目(42071403);高分辨率对地观测系统国家科技重大专项(71-Y50G10-9001-22/23)。

摘　　要：农作物分布的提取精度对于后续的农田参数反演和作物单产估算等均具有深刻的影响,而农作物分类识别过程中,训练样本的准确性和数量对其最终分类结果的影响是至关重要的。针对样本较少且分布不均匀的问题,通过实地标识和目视解译2种方式构建农作物分类样本数据集,设计5种样本数据集构建方案:方案(1)全部采用实测样点(70%训练、30%验证);方案(2)全部采用目视解译样点(70%训练、30%验证);方案(3)实测样点与目视样点分别选取相同比例的训练样本与验证样本,再结合构建训练样本集与验证样本集(70%训练、30%验证);方案(4)目视解译样点作为训练样本,实测样点作为验证样本;方案(5)目视样点与实测样点选取相同数量的样本进行结合构建样本集(70%训练、30%验证)。研究不同方案下的农作物遥感分类精度。结果表明,除方案(4)外,(1)(2)(3)(5)4种样本数据集划分方案的整体精度均在95%以上,分类结果较好,表明采用目视解译补充样本点,可以有效解决样本点较少和分布不均匀的情况。方案(3)作为研究区作物识别提取的最佳分类方案,总体精度达到97.6%,Kappa系数达到0.970,且玉米、水稻、大豆单类精度均超过97%,表明目视解译样点与实测样点分别选取训练样本与验证样本再结合构建训练样本集与验证样本集,不仅可以提升分类结果的精度,而且可以提高分类结果的真实性、准确性。The extraction accuracy of crop distribution has a profound impact on the subsequent inversion of farmland parameters and estimation of crop yield per unit area.In the process of crop classification and recognition,the accuracy and number of training samples are crucial to the final classification results.Aiming at the problem of small number of samples and uneven distribution,the crop classification sample data set was constructed by two ways of field identification and visual interpretation,and five sample data set construction schemes were designed:①all the measured sample points(70%training,30%verification)were used in the scheme;②all visual interpretation sample points were used(70%training,30%verification);③the same proportion of training samples and verification samples were selected from the measured sample points and visual sample points respectively,and then the training sample set and verification sample set were constructed combinedly(70%training,30%verification);④the visual interpretation sample points were used as training samples,and the measured sample points were used as verification samples;⑤the same number of samples were selected from the visual sample points and the measured sample points to construct a sample set(70%training,30%verification).The accuracy of crop remote sensing classification was studied under different schemes.The results showed that except④,the overall accuracy of①②③⑤four sample data set partition schemes was more than 95%,and the classification results were good.Using visual interpretation to supplement sample points could effectively solve the problem of fewer sample points and uneven distribution.As the best classification scheme for crop recognition and extraction in the study area,scheme③had an overall accuracy of 97.6%and a Kappa coefficient of 0.970,and the accuracy of corn,rice and soybean was all more than 97%,indicating that the combination of training samples and validation samples selected from visual interpretation samples and measured sam

关 键 词：遥感 农作物分类 目视解译 样本集划分 

分 类 号：S127[农业科学—农业基础科学]