基于空间抽样的区域地表覆盖遥感制图产品精度评估——以中国陕西省为例

地表覆盖数据是研究气候变化、生态环境、地理国情和人文经济等方面不可或缺的基础信息,因此其质量的优劣将直接影响相关决策的可靠性。本文针对我国研制的首套30 m分辨率全球地表覆盖数据产品,通过分析其海量、多维、非均质等空间特点,对传统的制图产品精度评估方法提出了改进：以地表类型进行分层抽样,样本量的计算采用优化模型并以地类所占面积比为权重逐层分配,样本的布设则考虑层内对象之间的空间相关性,在地表数据自然分布的基础上,通过分析空间相关性指数来提高样本的代表性和精度评估结果的可靠性。因此,本文提出针对区域地表覆盖遥感制图产品的空间分层抽样方法,其将精度评估分成抽样方案（样本的定量估计）和布设方案（样本的空间布设）2个部分,并以中国陕西省地表覆盖产品为例进行区域精度评估实验分析,从全区7大类地表类型数据中抽取1467个样本,经过样本检验与精度计算,该区域产品的总体精度为79.96%,Kappa系数为0.74。实验结果表明,本文提出的基于空间抽样的精度评估方法可行可靠,实验区域产品质量较好,并为后续针对全球范围的地表覆盖产品精度评估方法提供了参考。

Accuracy Assessment for Regional Land Cover Remote Sensing Mapping Product Based on Spatial Sampling: A Case Study of Shaanxi Province,China

Land cover refers to the complex of various materials of the earth's surface and its natural attributes and characteristics. Land cover data at global scale is fundamental and indispensable for global change studies, and its data quality has been the focus of a variety of international communities. In 2010, China launched the project of 30 m Global Land Cover remotely sensed products with 10 classifications for the base years of 2000 and 2010, which takes a four year period to complete, and the data quality issue will affect its subsequent applications. This paper mainly studies the accuracy assessment method for regional land cover data, as a preliminary of global validation. First, the spatial characteristics of land cover data are discussed, such as being massive, multi-dimensional, and heterogeneous. Then, a general approach based on spatial stratified sampling method for the accuracy assessment of regional land cover product is proposed to improve the traditional sampling method. Considering the spatial characteristics, this article presents the method of spatial stratified sampling: stratified sampling is conducted according to land cover type, the total sample size is calculated using probability and statistics optimal model, and the sample size of each layer is allocated according to the area ratio; while in the sample spatial allocation process, the representative of sample, which implies that the spatial correlation between samples should be low, needed to be considered. In each layer, each sample's spatial correlation is calculated by Moran's I index, and by setting a threshold value, the representative samples are chosen. The final sample set of each layer is randomly selected based on spatial analysis. The spatial sampling scheme is divided into three parts: sampling method, determination of sample size and the sample allocation. We have made an improvement in this paper, by designing a two-step scheme, which are quantitative estimation of sample size and spatial allocation. A case study of Shaanxi Province of China shows the method and process of accuracy assessment for regional land cover product. 1467 samples are selected by spatial stratified sampling in 7 strata. According to the confusion matrix, the overall accuracy is 79.96% and kappa index is 0.74. User's accuracies of all land cover types are more than 75%, while producer's accuracies fluctuate due to the low sample size caused by the low area ratio, which have little impact on the overall accuracy. Experimental results show that the proposed method is applicable to accuracy assessment of regional land cover product. Study on global land cover product will be performed as a main research direction in future.