基于面向对象SVM和谱聚类的极化SAR分类

发布时间：2018-04-30 05:42

本文选题：极化SAR + 支持向量积　；参考：《西安电子科技大学》2014年硕士论文

【摘要】：极化合成孔径雷达(Pol-SAR)是一个全天候,多通道,多参数的雷达成像系统,它可以获得一定波长和视角下目标的极化散射信息。和合成孔径雷达(SAR)比较,极化SAR拥有更丰富的极化内容,再加上极化SAR数据具有高维性,并且数据相对较复杂,如何结合现有的技术对极化SAR数据进行高效、准确的分类已成为极化SAR领域的研究热点。为了克服传统的分类方法时间复杂度过高的问题,本文提出了基于支持向量的聚类方法。由于极化SAR存在着较大的相干斑噪声,对后续的分类产生了极大的影响,所以本文根据相干斑噪声模型,提出了利用面向对象的方法对极化SAR进行分类,其主要工作如下:(1).本文提出了基于面向对象和SVM的极化SAR分类方法。传统的SVM分类精度高,速度快,但是其分类极化SAR时易受相干斑噪声影响、分类杂点较多,本文将基于像素的极化SAR分类和基于区域的极化SAR分类方法进行了有效的结合,首先将极化SAR数据的相干矩阵T利用SVM进行分类得到初始分类,然后将极化SAR数据的Pauli特征利用面向对象的方法进行过分割,最后在过分割的图像上对SVM的初始分类结果利用投票的方式进行二次分类,从而得到最终结果。由于该方法有效的利用了极化SAR数据的散射以及空间信息,所以具有不受相干斑噪声影响,边缘保持好,准确率高等优点。(2).本文提出了基于面向对象和谱聚类的极化SAR分类方法。传统的面向对象的方法可以对图像进行过分割,但是过分割后如何高效的融合是一个问题。谱聚类可以对极化数据进行很好的分类,但是当数据量大时,时间复杂度高,容易造成内存溢出。本文提出的方法首先利用面向对象的方法将极化SAR数据进行过分割,从而起到降维的目的,然后将过分割后的图像的每一个单元块当做一个对象,然后对这些对象进行谱聚类,最后将图像以对象为基本单元进行分类。由于该方法利用面向对象的方法对极化SAR进行了降维,所以时间复杂度大大降低,又因为是以对象为单元进行聚类,所以很好的克服了噪声的影响。(3).本文提出了基于支持向量积和谱聚类的极化SAR分类方法。谱聚类是以谱图理论为基础的,和传统的聚类方法比,它有着很多优点,比如在任意的样本空间都可以聚类并且能收敛于全局最优解、对不规则数据不敏感、准确率高等。但极化SAR数据量通常很大,直接求解其相似度矩阵不可行。所以本文提出了先降维后聚类的方法。首先选择少量样本,对其利用快速SVM进行训练,从而得到其支持向量,然后利用谱聚类对支持向量进行聚类,并计算出相应的类心,最后计算剩余样本到各个类心的距离并进行分类。和SVM类比,该方法可以有效的提高分类准确率,并解决了谱聚类中数据量过大,内存容易溢出和计算复杂度过高的问题。
[Abstract]:Polarimetric synthetic Aperture Radar (Pol-SAR) is an all-weather, multi-channel, multi-parameter radar imaging system, which can obtain polarimetric scattering information of targets at a certain wavelength and angle of view. Compared with synthetic Aperture Radar (SAR), polarimetric SAR has much richer polarization content, and polarization SAR data have high dimension and relatively complex data. Accurate classification has become a research hotspot in polarized SAR field. In order to overcome the problem of high time complexity of traditional classification methods, a support vector based clustering method is proposed in this paper. Due to the existence of large speckle noise in polarized SAR, it has a great influence on the subsequent classification. In this paper, based on the speckle noise model, an Object-Oriented method is proposed to classify polarized SAR. The main work is as follows: 1. In this paper, a polarimetric SAR classification method based on object oriented and SVM is proposed. The traditional SVM classification method has high accuracy and high speed, but it is easy to be affected by speckle noise when it is classified by polarized SAR, and there are many clutter. In this paper, the polarimetric SAR classification based on pixel and the polarimetric SAR classification method based on region are effectively combined. First, the coherent matrix T of polarized SAR data is classified by SVM to get the initial classification, then the Pauli feature of polarized SAR data is over-partitioned by object-oriented method. Finally, the initial classification results of SVM are classified by voting method on the over-segmented image, and the final results are obtained. Due to the effective use of scattering and spatial information from polarized SAR data, this method has the advantages of not affected by speckle noise, good edge maintenance and high accuracy. In this paper, a polarimetric SAR classification method based on object-oriented and spectral clustering is proposed. The traditional object-oriented method can over-segment the image, but how to fuse efficiently after over-segmentation is a problem. Spectral clustering can classify polarimetric data well, but when the data is large, the time complexity is high, which can easily cause memory overflow. The method proposed in this paper firstly uses the object-oriented method to over-partition the polarimetric SAR data so as to reduce the dimension. Then, each unit block of the over-segmented image is treated as an object, and then these objects are clustered by spectrum. Finally, the image is classified with object as the basic unit. Because the method reduces the dimension of polarized SAR by using object-oriented method, the time complexity is greatly reduced, and because the object is taken as the unit to cluster, it can overcome the influence of noise. In this paper, a polarimetric SAR classification method based on support vector product and spectral clustering is proposed. Spectral clustering is based on spectral graph theory. Compared with traditional clustering method, it has many advantages, such as clustering in arbitrary sample space, convergence to global optimal solution, insensitivity to irregular data, high accuracy and so on. However, the amount of polarimetric SAR data is usually very large, so it is not feasible to solve the similarity matrix directly. So this paper puts forward the method of reducing dimension first and then clustering. Firstly, a small number of samples are selected, and the support vectors are obtained by using fast SVM, then the support vectors are clustered by spectral clustering, and the corresponding cluster centers are calculated. Finally, the distance between the remaining samples and each class center is calculated and classified. Compared with SVM, this method can effectively improve the classification accuracy, and solve the problems of too large amount of data in spectral clustering, easy memory overflow and high computational complexity.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN957.52

【相似文献】