基于多核支持向量机的高分辨率遥感影像建筑物提取研究

发布时间：2018-06-28 16:57

本文选题：遥感影像 + 改进分水岭分割　；参考：《江西理工大学》2015年硕士论文

【摘要】：遥感技术快速发展,使得遥感影像分辨率不断提高,数据量急剧增加,因此能够更加精确地提取和分析影像中所包含的丰富地物信息。然而,目前高分辨率遥感影像信息自动提取准确度不高,建筑物作为其中一类极其重要的人工地物目标,将其各种信息较好的提取出来对于推动高分辨率遥感影像在目标识别分类以及城市土地规划管理等方面具有十分重要的意义。在影像信息提取中分类是一个关键问题,以核函数为基础的支持向量机(SVM)分类是解决影像目标分类问题的一种有效方法。针对核函数的选择不同和核参数设置的不同使得基于单核SVM模型的分类性能差异较大而不能准确提取建筑物的问题,本文提出构建多核学习支持向量机(MKSVM)分类模型用于提取影像建筑物。该MKSVM分类模型是针对不同特征对建筑物提取所贡献作用的不同,通过权重的方式将不同的基核函数线性相加构建的一种分类器,模型包含其中所有基本核函数的特性,具有很好的学习能力、推广性能和灵活性。本研究在Visual studio 2013平台下基于C++语言编写的ORFEO TOOLBOX(OTB)分布式影像处理算法开源库,研制面向对象的高分辨率遥感影像建筑物信息提取系统,实现采用两次分割分类的方式分步提取建筑物。首先通过区域合并的改进分水岭方法分割影像,并基于光谱特征采用K近邻监督分类提取只包含建筑物、道路和水泥广场等在内的不透水层,在此过程中同时单独提取影像中的蓝色厂房;然后对不透水层进行均值漂移分割,根据提取的建筑物光谱、纹理和形状等特征,将样本输入到多项式(POLY)与径向基(RBF)核函数构建的MKSVM分类数学模型训练得到分类器;最后根据训练得到的MKSVM分类器提取建筑物。通过与单核SVM分类提取建筑物结果进行精度比较,相对RBF核函数,两个区域的MKSVM分类的用户精度提高了1.5%左右,相对POLY核函数,两个区域的MKSVM分类的用户精度提高了3%到5%;两个区域的Kappa系数也分别提高到了0.8579和0.8415。实验结果表明,本文研制的影像建筑物信息提取系统运行可靠,通过与单核SVM提取建筑物相比,基于MKSVM的面向对象分类方法提取建筑物准确率更高,具有更好的分类能力。
[Abstract]:With the rapid development of remote sensing technology, the resolution of remote sensing image is improved and the amount of data is increased rapidly. Therefore, it is possible to extract and analyze the rich information of ground objects in the image more accurately. However, at present, the accuracy of automatic extraction of high-resolution remote sensing image information is not high, and buildings are one of the most important artificial objects. It is very important to extract all kinds of information to promote the high-resolution remote sensing image in target recognition and classification, urban land planning management and so on. Classification is a key problem in image information extraction. Support vector machine (SVM) based on kernel function is an effective method to solve the problem of image target classification. Because of the difference of kernel function selection and kernel parameter setting, the classification performance of SVM model based on single core is very different and the building can not be extracted accurately. In this paper, a multi-kernel learning support vector machine (MKSVM) classification model is proposed to extract image buildings. The MKSVM classification model is a kind of classifier which is constructed by linearly adding different basis kernel functions according to the contribution of different features to the building extraction. The model includes the characteristics of all the basic kernel functions. Good learning ability, promotion performance and flexibility. In this paper, the object oriented building information extraction system of high resolution remote sensing image is developed based on the ORFEO toolkit (OTB) distributed image processing algorithm open source library based on Visual studio 2013. The method of twice segmentation and classification is used to extract buildings step by step. Firstly, the improved watershed method is used to segment the image, and based on the spectral features, K-nearest neighbor supervised classification is used to extract the impervious layer, which includes only buildings, roads and cement squares, etc. In this process, the blue workshop in the image is extracted separately, and then the impermeable layer is segmented by the mean shift, according to the features of the building spectrum, texture and shape, etc. The samples are input into the MKSVM classification mathematical model which is constructed by polynomial (Poly) and radial basis function (RBF) kernel function. Finally, the building is extracted according to the training MKSVM classifier. Compared with the results of single kernel SVM classification, the user accuracy of MKSVM classification in two regions is improved by about 1.5%, and the relative POLY kernel function is compared with RBF kernel function. The user accuracy of MKSVM classification in two regions is improved by 3% to 5%, and the Kappa coefficient of the two regions is increased to 0.8579 and 0.8415 respectively. The experimental results show that the information extraction system of image building developed in this paper is reliable. Compared with single kernel SVM, the object oriented classification method based on MKSVM has higher accuracy and better classification ability.
【学位授予单位】：江西理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：P237

【参考文献】