由于在分类方法和空间分辨率等方面存在局限性，基于粗分辨率遥感数据的传统非监督分类结果在不同地物过渡带内往往误差较大。该文提出了基于地形限制特征的分类方法，在非监督分类的基础上，将非监督分类结果按照像元进行细分，并运用地形限制条件对细分后的像元进行二次判别分类。结果表明，分类精度明显提高，其中，农田和居民点分类精度的提高最为明显。这一方法使得完全同质的单元可以进行属性的变更，改善了像元空间分辨率差造成的误差；而地形限制特征的引入减少了传统非监督分类的不确定性，使模糊区域的分类有了较为明确的区分特征，提高了分类的精度。
Based on non_supervised classification methods, multi_temporal images, such as NOAA_AVHRR, SPOT_VEGETATION and MODIS series data, have been used to map regional, continental or global land cover patterns. Because of limitations of the classification method and spatial resolution of images, traditional non_supervised classification often results in many errors in the transition zones in which the spatial distribution of vegetation show segmental patterns. This problem can decrease the integral accuracy of the classification to a certain extent. In this paper, a new method, the topographical restrictive features, is presented to classify remotely sensed images using the Jinhe watershed in the Loess plateau as a case study. First, 25 relatively independent classes were obtained using the non_supervised classification method with multi_temporal NDVI data derived from the first two bands of MODIS data. Every cell with 500 m spatial resolution of the classification determined by the non_supervised method was divided into 25 homogenous cells with 100 m spatial resolution. Then, topographical features were defined. Information about aspect, slope, elevation, river net structure, and vegetation regional characteristics, derived from 1∶250 000 geographical spatial data, ETM+ image and yield data, were used to construct topographical restrictive features. Finally, the classification of every sub_cell was tested using the restrictive features and some cells were reclassified while maintaining their original classification. After the secondary classification, all cells were labeled by land cover type according to the land cover classification system, which was defined previously on the base of the IGBP land cover classification system and UMD system. Using this new method, the accuracy of the land cover classification increased from 71.88%, in traditional non_supervised classification method, to 84.09%. Inparticular, the accuracy of cropland and urban type classification improved. The method of cell sub_division avoids the shortcomings of traditional classification methods owing to the coarse resolution of the image processing, and makes it more highly probable that land cover types are homogenous within cells. The introduction of topographical restrictive features decreases uncertainty of traditional fuzzy classification and provides more precise distinctive features to classify the fuzzy zone, and thus improves the accuracy of classification.