【作者】 樊恒通；

【导师】 张友静；

【作者基本信息】 河海大学 ， 摄影测量与遥感， 2006， 硕士

【摘要】 城市绿化是城市生态系统的重要组成部分,采用高空间分辨率卫星数据及时、准确地获取城市植被信息(城市植被类型、分布及其结构),可为城市生态效益定量分析评价提供依据,满足城市绿化建设与管理部门的需求。 本文利用IKONOS影像,探讨了基于面向对象方法进行城市植被分类的最优分割尺度选择问题,构建了城市植被分类的类层次进行城市植被提取;为提高城市绿量估算的精度,在BP神经网络绿量估测模型的基础上,采用遗传算法对模型进行优化,并对参考因子进行改进,本文主要研究内容与结论如下: 1、影像预处理中,为识别城市植被中较小类型,提高植被分类精度,采用影像融合增强影像的解译能力。经分析比较,主成份变换3、4波段融合效果最好,可用于城市植被提取;高分辨率卫星影像中建筑阴影以及地形起伏影响,严重干扰了影像中地物的光谱信息,本文采取不同的方法对建筑物阴影和山体阴影分别进行校正:对城区建筑物阴影,采用面向对象的方法对影像分割后进行提取,采用朗伯模型进行校正;对紫金山山体阴影,结合DEM采用朗伯模型和经验统计模型分别校正,实验表明,朗伯模型存在过校正现象,经验统计模型校正能取得较为理想的效果。 2、利用面向对象的方法进行城市植被自动分类,提出了以实验法确定城市植被的最优分割尺度;基于商业软件构建了城市植被的类层次结构,应用对象的光谱、纹理以及上下文信息实现了城市植被的分类,分类总精度为85.5%,Kappa系数为0.826,取得了较好的分类效果。相比常用的基于像元的分类方法,面向对象的分类方法可以获得更高的分类精度。 3、利用遥感信息作为城市植被“绿量”的数据源,本文在神经网络估算模型的基础上,对估算参数进行了改变,选取植被指数以及环境因子、高程因子作为自变量;采用遗传算法对BP网络的权值、阈值进行优化,建立遗传优化的神经网络估算模型。实验结果表明,加入环境因子、高程因子,可以提高绿量的估算精度;采用遗传算法优化BP神经网络的权值、阈值,能够使网络收敛到全局最优解,提高了网络训练的稳定性。更多还原

【Abstract】 Urban virescence is an important part of the city ecosystem. High-resolution satellite image is adopted to extract the urban virescence information (The types, distributing and structure of urban vegetation) immediately and exactly, which is an important basis for evaluating quality of city ecosystems, and meeting the demand of deparment for urban planning.In this paper, the IKONOS image is used to extract urban vegetation through object-orient classification method. How to select optimal segmentation scale is discussed. in urban vegetation extraction.Then, the urban vegetation was extracted by an established class hierarchy. To increasing the accuracy of urban "Green Quantity" estimated, The optimized BP Neural Network model is used to estimating urban "Green Quantity" based on Genetic Algorithms The main contents and results are as follows:1. To enhance image interpretation ability, IKONOS PAN is merged with multi-spectral image in the image pro-processing. By anlysing four methods of image fusion, the PCA transformation is the best method for vegetation classification. Because of the building shadow and Terrain effect in high-resolution image, the spectral of geographic entity distorted seriously. This paper adopts different method to rectify mountain shadow and building shadow. On one hand, the building shadow was automated extraction by object-orient method and corrected by the Lambertian model;on the other hand, the mountain terrien normalization was corrected by Statistic-empirical method using DEM data. The experiment shows that Lambertian model was overcorrection seriously;but Statistic-empirical method can get better performance.2. In this paper, urban vegetation information was extracted by object-orient Method. A new method is developed for selecting optimal segmentation scale. Urban vegetation is classificated by a class hierarchy based on spectral, texture and context information which are established though commercial software. The gross accuracy is 85.5% and Kappa coefficient is 0.826. The experiment shows that object-orient method lead to improve urban vegetation classification accuracy compared to pixel-based classification method.3. Remote sensing is the source of geting the urban "Green Quantity" information. Based on BP Neural Network model, which is established, this paper changed two estimation parameters. It adopt environment factor and elevation factor which are correlate with vegetation, discard texture factor which is little correlate with vegetation. Then optimize the BP Neural Network model by Genetic Algorithms. The experiments show that the estimation of city "Green Quantity" increased because of adopting environment factor and elevation factor. The Neural Network model training was more stable because optimized by Genetic Algorithms.更多还原