基于边缘信息增强的面向对象高分影像信息提取研究

发布时间：2018-04-26 18:01

本文选题：高分影像 + eCognition　；参考：《安徽理工大学》2016年硕士论文

【摘要】：随着信息技术的发展,地理信息获取也变的越来越重要。地理国情普查作为我国一项重要的地理信息统计工作,要求必须掌握我国自然地理要素的基本情况。遥感技术是目前获取地理信息的一种最重要技术手段之一,在测绘、城市规划、国土资源、地质矿产勘查、农林业、军事等行业和领域中都有广泛的应用。随着遥感平台的不断更新,高分辨率可用卫星、WordlView、QuickBird、IKONOS、 GeoEye以及国产的高分系列卫星最高空间分辨率可达到亚米级,高分影像的出现使得遥感影像能够展现出更加丰富的地理信息。较常用的遥感影像分类方法主要包括监督分类、非监督分类以及决策树分类等。它们依据地物的光谱特征进行分类,比较适合中低分辨率的遥感影像。但是对于高分影像,常用的基于像元的分类方法不仅不能有效地利用其丰富的空间结构,还会由于其地物光谱异质性小而引起较多“椒盐”噪声,最终导致分类结果精度降低。因此,对于高分辨率遥感影像传统的分类方法已经越来越难于满足其信息提取的要求。为了更好的提取高分影像信息,研究人员提出了能够充分利用高分影像纹理、空间等信息的面向对象影像信息提取技术。目前面向对象的信息提取方法主要包括：基于最大面积和目标函数的最优分割尺度选择；基于Gram-Schmidt等不同融合方法的面向对象的遥感影像分类；基于总体精度或Kappa系数的面向对象分割分类评价指标等。本文在eCognition软件的基础上以淮南地区的WorldView-III影像为例,对其进行了预处理,多尺度分割,边缘信息增强以及边缘信息增强后的影像信息提取,最后对于不同分类方法做出了精度评价,具体研究如下：(1)以淮南地区高分辨率遥感影像为例,将原始影像进行Pansharping融合,并对其分别做监督分类、非监督分类,生成不同方法下遥感影像分类图。(2)重点研究了基于eCognition软件的面向对象分类方法。首先借助ESP(Estimation of Scale Parameter)尺度评价工具,通过不同实验,结合目视判别效果确定了最优分割参数：其中尺度因子为110,形状因子为0.5,紧凑度因子为0.5。影像分割完成后对其进行最邻近分类,多次循环选取最优分类样本及样本特征(NDⅥ、Shap-index、Length/Width、mean-Brightness、mean-max.difff等),最终生成基于面向对象的影像分类图。(3)对高分影像进行基于Canny算法的边缘检测,将检测后的影像层参与地物多尺度分割,并设置其分割权重,得到基于边缘检测的多尺度分割影像图。对高分影像做基于Getis-Ord的局部空间自相关性统计分析,在此基础上再将其进行基于Canny算法的边缘检测,将检测后的影像层参与地物多尺度分割,并设置其分割权重,得到基于信息增强组合变换的多尺度分割影像图。(4)对不同影像信息提取方法分类精度进行对比分析,发现基于像元的分类方法中最大似然分类器在本次实验中精度最高,其中总体分类精度为84.9382%,Kappa系数为0.8152,而面向对象分类的总体分类精度达到了92.3076%,Kappa系数达到了0.9018,基于Getis-Ord和Canny边缘信息增强后的面向对象分类在本次实验中精度最高,其总体分类精度为94.6667%,Kappa系数为0.9354。
[Abstract]:With the development of information technology, the acquisition of geographic information is becoming more and more important. As an important geographical information statistics work in China, the geographical situation census requires that the basic situation of the natural geographical elements of our country must be mastered. Remote sensing technology is one of the most important technical means for obtaining geographic information at present, in surveying and mapping, in urban planning, Land and resources, geological and mineral exploration, agroforestry, military and other industries and other fields are widely used. With the continuous updating of remote sensing platform, high resolution can be used by satellite, WordlView, QuickBird, IKONOS, GeoEye, and homemade high grade series satellites can reach the highest spatial resolution, and the appearance of high image makes remote sensing image More abundant geographic information is displayed. The commonly used remote sensing image classification methods mainly include supervised classification, unsupervised classification and decision tree classification. They are classified according to the spectral characteristics of the ground objects, and are suitable for middle and low resolution remote sensing images. But for high resolution images, the commonly used classification methods based on pixels are not only used. It is not effective to make use of its rich spatial structure, and also cause more "salt and pepper" noise because of its small spectral heterogeneity, which eventually leads to the reduction of classification accuracy. Therefore, the traditional classification method for high resolution remote sensing images has become more and more difficult to meet the requirements of its information extraction. Information, researchers have proposed an object oriented image information extraction technology that can make full use of high image texture, space and other information. At present, the object oriented information extraction methods mainly include the optimal segmentation scale selection based on the maximum area and the target function, and the object oriented teleconnection based on different fusion methods such as Gram-Schmidt. Sense image classification; based on the overall accuracy or Kappa coefficient of object oriented segmentation and classification evaluation index. Based on the eCognition software, this paper takes the WorldView-III image in Huainan as an example to carry out preprocessing, multi-scale segmentation, edge information enhancement and image information extraction after the enhancement of edge information. The precision evaluation of the same classification method is made as follows: (1) taking the high resolution remote sensing image of Huainan as an example, the original image is fused by Pansharping, and the classification is supervised and classified, and the different methods of remote sensing image classification are generated. (2) focus on the object-oriented classification based on the eCognition software. Method. First, with the aid of ESP (Estimation of Scale Parameter) scale evaluation tool, the optimal segmentation parameters are determined through different experiments and visual discrimination effect: the scale factor is 110, the shape factor is 0.5, and the compact factor is the nearest neighbor classification after the completion of the 0.5. image segmentation, and the optimal classification sample is selected repeatedly. Sample features (ND VI, Shap-index, Length/Width, mean-Brightness, mean-max.difff, etc.), and finally generate object based image classification. (3) edge detection based on Canny algorithm for high score images, and the detected image layer is involved in multi-scale segmentation, and the segmentation weight is set up, and multiscale based on edge detection is obtained. On the basis of the Getis-Ord based local spatial autocorrelation statistics analysis, the high score image is analyzed based on the local spatial autocorrelation, and then the edge detection based on the Canny algorithm is carried out. The detected image layer is involved in multi scale segmentation, and the segmentation weight is set up, and the multiscale segmentation image based on the information enhancement combination transformation is obtained. (4) The classification accuracy of different image information extraction methods is compared and analyzed. It is found that the maximum likelihood classifier based on the pixel based classification method has the highest accuracy in this experiment, of which the overall classification precision is 84.9382%, the Kappa coefficient is 0.8152, while the overall classification precision of the object oriented classification is 92.3076%, the Kappa coefficient reaches 0.9018, based on this method, the maximum likelihood classifier is based on the object classification method. The object oriented classification with enhanced edge information of Getis-Ord and Canny has the highest accuracy in this experiment. The overall classification accuracy is 94.6667%, and the Kappa coefficient is 0.9354.

【学位授予单位】：安徽理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：P237

【相似文献】