中文摘要：

人类活动对生态环境具有显著影响，大尺度土地利用／覆盖变化（Land Use／Cover Change，LUCC）作为人类活动最直接的表征，能够很好地反映这一过程，因此进行精确而迅速的大尺度土地利用／覆盖分类与提取方法研究尤为关键。全球覆盖产品GlobCover（2005／2006）数据已经具有良好的空间精度和数据准确度，但仍然存在一些分类误差。为提高地表覆被分类精度，本文以G10bcover（2005／2006）的巴西数据为例，以2005年LandsatTM／ETM影像为主要信息源，结合相应地学知识与辅助数据，利用人机交互逐栅格修改方法得到2005年土地利用数据产品。结果表明：通过对GlobCover数据和本次成果数据进行精度评价与对比分析，GlobCover数据巴西地区的总体精度为67．17％，Kappa系数为0.58，改进后产品总体精度为93．39％，Kappa系数为0．91。此外，改进后数据显示巴西常绿阔叶林面积最大，面积比例达45．67％；农地／自然植被镶嵌面积次之，比例为19．19％；封闭灌丛面积最小，比例为12．34％。农地／自然植被镶嵌和灌丛与草地2种地类的修改比例最大，其中混合像元地类比例减少3．54％，灌丛与草地比例增加3．81％。综上，改进方法可以有效地提高土地利用／覆盖分类的效率和精度，为后续大尺度LUCC产品的制作和以LUCC产品为基础的相关研究提供参考。

英文摘要：

Human activities have significant impacts on ecosystems. As the most direct characterization of human activity, large-scale land use/cover change is used to analyze the impacts of human activities on ecosystems. Therefore, scientists have paid much attention on the classification and extraction methods of land use/cover products. It was suggested that GlobCover （2005/2006） product was precise enough for the scientific study. How- ever, the product has some limitations. In order to improve the quality of this product, this study developed new method for mapping and monitoring national land cover information in Brazil. The new Brazilian land use/cover data in 2005 were developed by using human-computer interactive discrimination at per-cell level based on Glob- Cover （2005/2006） data and the combination of geographic knowledge and the major data source of Landsat TM/ ETM images. The results indicated that data accuracy and cost-efficiency were both improved by the developed method. The classification accuracy was improved from 67.17% in the GlobCover to 93.39% in our new dataset. Kappa coefficient was also improved from 0.58 to 0.91. Evergreen broadleaf forest area in Brazil was the highest among all the land cover types, with an area ratio of 45.67%. Farmland/natural vegetation mosaic area followed with an area ratio of 19.19%. The third largest land cover type was closed shrub with an area ratio of 12.34%. Modification ratio of agricultural land/natural vegetation mosaic and shrub and grassland was the largest. Among them, the proportion of mixed pixels of land class decreased 3.54%, while shrub and grassland increased 3.81%. As a result, the new developed method was proved to be more efficient and accurate. It can be used for largescale land use/cover classification and analysis in further study.