论文总字数：29808字

摘 要

首先融合实验区多源遥感数据，把nDSM作为融合后影像的一个波段处理。高度信息也应被赋予权重参与分割，借助ESP工具确定两个不同尺度的对象分割体系。利用特征视图和训练区样本统计并确定各类地物的分类特征和特征阈值，过程中需要反复迭代以求最优参数组合，其中高度信息、绿波比、长宽比是重要的分离指标。使用基于知识规则的模糊分类方法，先试验性地独立计算出各单项地类的提取结果，所有地类规则都清晰之后再统一进行分类，最后结合阈值分配类方法优化分类结果。

对实验结果进行基于隶属度分类的分类稳定性和最优分类结果精度评价，在地表真实图像上随机生成1000个样本点，计算混淆矩阵；同时使用ENVI软件平台对影像进行基于像元的决策树分类，尽量与面向对象分类实验中使用同一组分类规则，对比两项实验精度结果。研究证明基于面向对象的分类方法能取得较好结果，能有效提高高分影像用地分类的精度和效率。

关键词：高空间分辨率遥感影像，面向对象，用地分类，模糊分类，eCognition

A STUDY ON OBJECT-ORIENTED

LAND USE CLASSIFICATION

OF HIGH SPATIAL RESOULTION IMAGE

Abstract: In this thesis, land use classification of high spatial resolution remote sensing image data based on object- oriented method was investigated with eCognition, using the spectral imagery combined with LIDAR intensity data in Australia.

Different data sources were integrated and nDSM was pre-processed into a band of the integrated imagery. Two different scales of segmentation system of image were established based on ESP tools when height information should also be weighted during the segmentation. All kinds of feature used in classification and relative characteristic thresholds were estimated and determined by feature views and training sample statistics. The process required repeated iteration in order to find the optimal parameter combination, which height information, green ratio, ratio of length to width were important separation indexes. Tentatively extract the operated results of each individual class depended on knowledge rules before the overall fuzzy classification was carried out. Finally, the algorithm of assigning class by characteristic threshold was conducted to optimize the classification.

Accuracy evaluation included classification stability and best classification result based on membership function, and error matrix based on 1000 sample points chosen from ground truth image to calculate overall accuracy. In contrast, a decision tree in ENVI which was based on similar rules of previous object-oriented experience was structured for classification in the unit of pixel. This study shows that based on the object-oriented classification method can achieve better results, effectively improves the efficiency and accuracy for classification of high spatial resolution image.

KEY WORDS: high spatial resolution remote sensing image, object-oriented, land use classification, fuzzy classification, eCognition

目 录

摘要………………………………………………………….……………….……………………Ⅰ

Abstract…………………………………………………………….…….………………………Ⅱ

第一章 绪 论 1

1.1 引言 1

1.2 面向对象分类方法研究现状 1

1.3 研究目标、研究内容和技术路线 3

第二章 实验准备和实验数据 5

第三章 方法与试验 7

3.1 多波段影像与nDSM的融合 7

3.2 影像多尺度分割 8

3.3 对象分类 13

第四章 精度评价 25

4.1 分类稳定性 25

4.2 最优分类结果 26

4.3 基于真实地表样本点的混淆矩阵 26

4.4 基于像元的分类精度对比 29

结论 ……………………………………………………………………………………………….34

致谢………………………………………………………………………………………………..36

参考文献（Reference） 37

绪 论

1.1 引言

卫星遥感是人类观察、分析、描述所居住的地球环境的有效手段^[1]。随着传感器硬件技术和成像方式的发展，商用的WorldView2影像分辨率可达0.5m，WorldView3更可达到0.31m，高分影像放大了同类地物内部各像元之间的光谱差异^[2]，基于像元光谱信息的统计模式分类已经无法满足信息提取的精度和效率要求，成为制约高分辨率卫星数据应用的主要瓶颈^[2]。如何能高效、精确、智能化地从遥感影像中解译出能满足各项应用的专题信息，是迫切要研究解决的问题。

传统分类方法只能以像元的光谱值作为分类依据，有光谱变异度高、同物异谱、异物同谱等弊端，而且对高分影像来说，使用以单个像元为单位的常规分类方法，如监督分类或非监督分类^[34]，容易仅仅关注单个像元光谱信息而忽略与邻近像元组成的真实地表对象的几何结构特征，限制信息提取的精度^[3]，造成空间数据大量冗余及分类不精。

剩余内容已隐藏，请支付后下载全文，论文总字数：29808字