提高城市TM图像分类精度的两种方法比较  

Comparison of Two Methods to Enhance Classification Accuracy in Urban TM Image

作　　者：王修信[1,2] 秦丽梅[1] 罗涟玲[1] 农京辉[1] 朱启疆[2] 

机构地区：[1]广西师范大学计算机科学与信息工程学院,广西桂林541004 [2]北京师范大学遥感科学国家重点实验室,环境遥感与数字城市北京市重点实验室,地理学与遥感科学学院,北京100875

出　　处：《广西师范大学学报（自然科学版）》2009年第4期19-22,共4页Journal of Guangxi Normal University:Natural Science Edition

基　　金：国家自然科学基金资助项目(40571109);北京市自然科学基金重点基金资助项目(4051003);北京师范大学与北京市教委共建资助项目(JD100270540);广西师范大学博士科研启动基金资助项目(2007)

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。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.

关 键 词：遥感图像 城市分类 混合像元分解模型 神经网络法 决策树法 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]