利用SPOT4-VEGETATION 的遥感数据产品生成的NDVI 与NDWI 植被指数时间序列图像集，通过 ISODATA 非监督分类方法，编制中国西北地区土地覆盖图。以TM 影像人工解译结果作为真实值，通过对西 北五省共计47 个均匀分布且异质性强的25km× 25km 样本区内的土地覆盖类型及其面积进行统计分析，修正 了SPOT4-VEGETATION 的土地覆盖分类系统，建立了各省验证结果的样本统计直方图并计算其回归系数。结 果表明SPOT4-VEGETATION 中国西北地区土地覆盖图在总体上具有较高的准确性。影响遥感数据自动分类 精度，造成土地覆盖误判的原因主要来源于两个方面：即异物同谱和混合像元问题。对于前者通过叠加各种 辅助数据如DEM 等可以降低误判的机率；对于后者运用混合像元分解的各种算法可以提高分类精度。

This paper introduces the methods of compiling the land-cover map of Northwest China（with an area of approximately 3，100，000 km2）using SPOT4-VEGETATION data sets and the validation techniques using high spatial resolution TM images. Based on the spectral reflectance，NDVI（Normalized difference Vegetation Index）and NDWI（Normalized Difference Water Index）time series data sets from SPOT4-VEGETATION，the land-cover map of Northwest China is compiled by applying the ISODATA unsuperviesd classification method. In order to evaluate this land-cover map’s accuracy， 47 sampling units are selected from the whole mapping region. Each sample has a 25km × 25km unit and evenness distribution as well as high heterogeneity. With the assumption of the TM interpretation in the 47 sampling units as the true land cover condition，we calculated every land cover type and its area percentage existing in every sampling unit in the SPOT4-VEGETAION map and the TM interpreted map，respectively. According to the statistical results，the landcover classification system of the SPOT4-VEGETATION is modified，the sampling statistical histogram of the validation results in every province is established and the regression coefficient value is also calculated respectively. The validation results show that the land cover mapping of Northwest China using SPOT4-VEGETATION data sets and the ISODATE method can get an improving accuracy attributing to the high quality of SPOT4-VEGETATION data products and the effective method of spectral index combination of NDVI and NDWI . The reasons that reduce or impact the landcover classification accuracy mainly derive from two aspects：one is attributed to different landcover classes with same spectral characteristics， and the other is due to the mixed pixel problem. TO the former，adding some auxiliary information data such as DEM can reduce the possibility of misjudgment；to the later，the methods of mixed pixel decomposition and sub-pixel mapping may be good ways to increase the land cover classification and mapping accuracy.