QuickBird卫星图像信息识别

Information Identification on QuickBird Image

信息识别是目前高分辨率遥感应用中的最大障碍。以株洲市Qu ickB ird图像为研究对象,将研究区分为道路、水、林地、农用地、裸露地和居民点6种地类,分别进行目视判读、计算机监督分类和非监督分类,其精度分别为98.2%、72.64%和60.71%。同时,还对研究区内的Qu ickB ird、ETM+和TM图像进行计算机监督和非监督分类对比,结果表明无论是监督分类还是非监督分类,Qu ickB ird图像的分类精度均低于ETM+和TM图像,这说明空间分辨率的提高对传统的计算机分类结果没有改善,传统的基于像元的分类技术在应用于Qu ickB ird图像时表现出严重的缺陷。因此,本文回避了像元灰度统计法,采用先将图像分割,将以像元为基础的Qu ickB ird图像转化为以对象为基础的图像,这样将研究区共分割出10 000多个对象,建立对象的面积、周长、长度、宽度、长/宽、矩形度和圆形度计算模型;根据研究区各地类特征确定特征因子阈值,模拟目视判读过程,重新对研究区进行分类,结果6种地类的综合分类精度达到91.6%,这说明基于对象的多特征分类对于Qu ickB ird图像识别有明显的改善作用。

The information identification is the most difficult in the application of high-resolution remote sensing images. This paper focused on the method of QuickBird image＇s information extraction. Six kinds of land cover types： road, water, forest, agriculture, nuke and residence in the study area, Zhuzhou, Hunan, were identified by visual interpretation, supervised classification and non-supervised classification respectively, and the accuracy is 98.2% , 72.64% and 60.71% correspondingly. At the same time, pre-processed images of QuickBird, ETM ＋ and TM were identified by supervised classification and non-supervised classification respectively and the accuracy of QuickBird image classification is lower than that of ETM＋ and TM. This showed that resolution＇ s improvement couldn＇t raise the accuracy of classification by the method of traditional classification. This paper avoided the traditional classification based on pixel-to-pixel and applied new method of classification based on object＇ s gray character and shape characters. QuickBird image based on pixels was changed into new image based on objects first by segmentation, then models of measuring objects＇ area, perimeter, length, width, length/width, rectangle and roundness were built. Six kinds of land types were classified again by computer stimulating visual interpretation in the study area, the composite accuracy of classification is up to 91.6% , this showed that the method based on objects is a very effective way to improve the accuracy of classification of QuickBird image.