中文摘要：

城市遥感图像分类是获取城市绿地空间分布状况的基础，然而混合像元的存在导致分类精度不高。因此，提出了混合像元分解模型结合神经网络法（MPD-NN法），利用其对北京市TM图像进行地物分类，并与较常用的决策树法分类结果比较，研究在图像现有空间分辨率的条件下提高城市分类精度。结果表明，MPD-NN法分类可较为有效地处理城市遥感图像存在的混合像元问题，可以提高城市分类精度，分类精度高于决策树法，其总体精度和Kappa精度均将近90％，水体、草地、农田、裸地、林地的分类精度在85％以上，高密度建筑区、低密度建筑区、水泥地的分类精度接近80％，也均高于决策树法，是遥感分类较精确识别城市下垫面地物的一种可行方法。

英文摘要：

Urban remote sensing classification is the foundation of extracting urban vegetation distribution. However,urban classification accuracy is very low because of mixed pixels. Therefore ,mixed pixel decomposing model with neural network（MPD-NN） is proposed to classify TM image in Beijing. Then the classification image from the method is compared with that from the decision tree method. Great effort is made to improve urban classification accuracy under TM spatial resolution. Results from this study show that MPD-NN could deal with urban mixed pixels very effectively and increase urban classification accuracy. With the total accuracy and Kappa accuracy nearly 90%, classification accuracy above 85 % for water,grassland, farmland, bare soil and woodland ,and classification accuracy nearly 70 for building and concrete, MPD-NN is more accurate than decision tree in classifying urban land categories and can be used to extract urban surface information with great accuracy.