基于亚分数混淆矩阵的中国典型区大尺度土地覆盖数据集评价  

Assessment of Large-scale Land Cover Datasets in Typical Areas of China based on Sub-fractional Error Matrix

作　　者：蒋璐媛 肖鹏峰[1] 冯学智[1] 李云[1] 朱榴骏[1] 

机构地区：[1]南京大学地理与海洋科学学院地理信息科学系,江苏南京210023

出　　处：《遥感技术与应用》2015年第2期353-363,共11页Remote Sensing Technology and Application

基　　金：国家973计划项目(2011CB952001)资助

摘　　要：在中国东北、华北、华中、华南、西北、青藏、内蒙古7个自然地区分别选择典型区A、B、C、D、E、F、G,以Landsat TM/ETM+影像分类结果为参考数据,采用亚分数混淆矩阵对5种大尺度土地覆盖数据集的精度进行定量评价,为数据集的使用提供科学依据。亚分数混淆矩阵可避免参考数据与待评价数据尺度转换时引入的误差,能反映不同优势类比重情况下数据集的总体精度和分类方法误差。结果表明:GLC2000在全部典型区的总体精度最高,为65.64%;UMD总体精度最低,为43.06%。GLC2000在主要土地覆盖类型为林地和耕地以及草地区域具有较高的分类精度;UMD在各区域的分类精度均最低或较低。5种土地覆盖数据集对于城镇、其他的分类精度在各典型区均较低;对于草地和水体的分类精度则是在西北干旱区和青藏高原区的典型区较高。This paper evaluated the accuracy of five large-scale land cover datasets based on sub-fractional error matrix,by taking Landsat TM/ETM+image classification results of seven typical areas in China as the reference data,which provides the scientific basis for the use of datasets.The sub-fractional error matrix can avoid errors caused by the scale difference between reference data and datasets,and evaluate the accuracy on sub-pixel scale and reflect the classification accuracy and classification method error with different dominant fraction.The results show that:the overall accuracy of GLC2000 is the highest in all typical areas,is at 65.64%;and UMD is the lowest in all typical areas.GLC2000 has a higher classification accuracy in the areas covered by forest,cropland and grass;the classification accuracy of UMD is the lowest or the lower one in each typical area.The five land cover datasets have a lower classification accuracy in urban and other;while with a higher classification accuracy of grass and water in each typical area of arid region of northwest China and Tibet Plateau region.

关 键 词：亚分数混淆矩阵 大尺度 土地覆盖数据集 精度评价 

分 类 号：P237[天文地球—摄影测量与遥感]