土地覆被遥感产品真实性检验方法对比  

Comparison of Accuracy Assessment Methods of Remote Sensing based Land Cover Products

作　　者：王冰泉 冉有华 Wang Bingquan;Ran Youhua(Northwest Institution of Eco-Environment and Resources,Chinese Academy of Sciences,Lanzhou 730000,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院西北生态环境资源研究院,甘肃兰州730000 [2]中国科学院大学,北京100049

出　　处：《遥感技术与应用》2022年第1期196-204,共9页Remote Sensing Technology and Application

基　　金：国家自然科学基金项目(42071421);高分辨率对地观测系统国家重大专项(21-Y20B01-9001-19/22)资助。

摘　　要：真实性检验是保障土地覆被遥感产品生产质量,支持土地覆被遥感产品应用的重要环节。土地覆被遥感产品真实性检验实践总与理论存在一些差异,其主要体现在用于检验的参考数据样本数量和空间分布两个方面,而这种差异对检验结果的影响还不十分清楚。以中国陆地范围为实验区域,GlobeLand30作为全样本参考数据,ESA CCI-LC作为检验对象,通过设计实验,评估了样本数量、抽样方式和样本单元的不同对检验结果的影响。结果表明:(1)样本量对检验结果有重要影响。当样本量减半(约600个样本点),点样本真实性检验的总体精度与在理论样本量下的总体精度接近,但是,在样本量较小的情况下,不同类型的精度对样本量的敏感程度不同,面积权重小的类别对样本量更敏感,尤其是在采用分层随机抽样的情况下;(2)在理论样本量基础上,无论是点样本真实性检验还是群样本真实性检验,抽样方式对总体精度影响不大,但是对各个类型的精度有重要影响,尤其是面积权重小的类别;(3)群样本检验稳定性比点样本检验差。在理论样本量基础上,样本单元的增大将增加检验结果的不确定性,随着样本单元的增加,虽然检验结果波动性变小,但是检验结果相较于点样本检验结果仍存在一定偏差。因此,对于以获取总体精度为目标的真实性检验,对于点样本,样本量可以减少到理论样本量的一半,总体精度的绝对误差在1%以内,标准差约为2%,而对于群样本,则需要增加到理论样本量的5倍。对于以获取各类型精度为目标的真实性检验,建议增加样本量为理论样本量的2倍以上,以确保面积权重小的类型被分配足够的样本点。Accuracy assessment is important for the application of land cover products.In practice,there are many differences for sample size and spatial distribution of the reference data used for accuracy assessment,and the impact of this difference on accuracy assessment results is not very clear.This paper validated the ESA CCI-LC using GlobeLand30 as the reference data in China.We test the impact of sample size,sampling model,and sample unit on the overall accuracy and various types of accuracy.The results are as following.Firstly,the sample size makes a difference to the accuracy assessment results.The overall accuracy based on point sample is close to that of the theoretical sample size while the sample size in half(about 600 sample points),however,the sensitive degree of different classes of precision on the sample size is different,classes with low area weight are more sensitive to sample size,especially stratified random sampling used in the case of small sample size.Secondly,based on the theoretical sample size,the sampling model has little effect on the overall accuracy whether it is the accuracy assessment of point sample or cluster sample,but it has an impact on the accuracy of each class,especially the class with a small area weight.Finally,the stability of accuracy assessment of simple random sampling based on cluster sample is worse than the accuracy assessment based on point sample.The increase of the sample unit will increase the uncertainty of the accuracy assessment result on the basis of the theoretical sample size,with the increase of the sample unit,although the volatility of the accuracy assessment result becomes smaller,the result is still large compared to the point sample.Therefore,in order to reduce the uncertainty of the accuracy assessment results of land cover remote sensing products,the sample size can be increased for point sample,and the number of sample units can be increased or the size of sample units can be appropriately reduced for cluster sample.

关 键 词：抽样方式 样本量 真实性检验 ESA CCI-LC GlobeLand30 

分 类 号：P237[天文地球—摄影测量与遥感] K903[天文地球—测绘科学与技术]