随着遥感数据获取能力的不断增强，自动化程度已经成为大尺度遥感土地覆被分类面临的关键问题。然而，现有训练样本的人工选取方法成为制约土地覆被分类自动化的瓶颈。本文以河南、贵州两省为研究区，提出一种基于多源数据的土地覆被样本自动提取方法，以构建适用于大尺度的土地覆被自动分类。首先，以2010年1∶10万土地利用数据CHINALC和30 m分辨率全球土地覆被数据GlobleLand30为样本数据源；然后，利用空间一致性分析及异质性分析确定样本初选区域；最后，通过样本提纯去除无效样本。结果表明：（1）应用多源数据的土地覆被样本自动提取方法获得的分类产品总体分类精度高于人工样本提取方法制作的全球土地覆被产品MCD12Q1。（2）与单源样本自动提取方法相比，应用多源数据的土地覆被样本自动提取方法，可获得更好的分类稳定性。综上，多源数据的土地覆被样本自动提取方法可在保证精度的同时，提升土地覆被分类的自动化程度。

The capability of remotely sensed data acquisition is constantly improved. Thus, enhancing the automation degree for land cover classification at a large scale by remote sensing has become a key problem. However, present manual methods of selecting samples are becoming the bottleneck of automatic land cover classification. Many global and national land cover datasets based on remote sensing have been produced in the past two decades for different international or national initiatives. However, the rich knowledge implied in these products has not been fully exploited.
The overall objective of this study is to set up an automatic land cover classification approach at a large scale by remote sensing through an automatic method of collecting land cover samples based on multisource datasets. The practical goals are to improve automation degree of land cover classification and enhance the accuracy of land cover classification.
Henan and Guizhou provinces were selected as the study areas based on their types of land covers. First, the national land use database of China at a scale of 1∶100000 (CHINALC) and global land cover data (GlobleLand30) at a resolution of 30 m were selected as the data sources for the sample collection. Second, the initial sample areas were collected based on the spatial consistency analysis and heterogeneity analysis. Third, invalid samples were removed from the initial samples through the technology of sample purification. Finally, the Jeffries–Matusita distance was used to measure the classification feature separability of the samples between the different land cover types to prove the feasibility of the proposed method. The accuracy of the land cover product by the proposed method of sample collection was assessed and compared with the globe land cover product MCD12Q1.
Experimental results show that the following: (a) The overall accuracy of the classification product through the proposed automatic method of sample collection based on multisource datasets was higher than that of the global land cover product MCD12Q1, which was classified based on the manual method of sample selection. The overall accuracy of the land cover classification product based on the proposed method was 78% in Henan province and 57% in Guizhou province, whereas that of MCD12Q1 was 74% and 55%, respectively. The Kappa coefficients of the former were 0.54 and 0.25, respectively, whereas those of the latter were 0.42 and 0.15, respectively. (b) Compared with the method of sample collection based on a single source land cover dataset, the proposed automatic method of sample collection based on multisource datasets had better classification stability and higher classification feature separability. The standard deviation of the Kappa coefficient and accuracy of products by 8-time experiments were less than 0.004. The classified results were also more stable.
Unlike the method of sample collection based on a single source land cover dataset, the proposed automatic method of land cover sample collection based on multisource datasets not only improves the automation degree of land cover classification, but also enhances the accuracy of land cover classification.