基于深度支持向量机的极化SAR图像分类

发布时间：2018-07-24 08:07

【摘要】：极化合成孔径雷达(Polarimetric Synthetic Aperture Radar,PolSAR)由于能够提供更多丰富的目标散射信息,近年来成为遥感领域的新热点。其地物分类问题作为极化SAR图像解译的重要研究内容和关键技术,在民用和军事领域具有重大的理论意义和应用价值。支持向量机(Support Vector Machine,SVM)是统计学中一种有效的监督分类方法,目前已广泛应用于多个领域,该方法在机器学习领域受到了重视和长足的发展。本文采用SVM算法对极化SAR地物分类问题进行研究,所取得的主要研究成果为:1.本文着重研究了支持向量机算法中的最小二乘支持向量机(LSSVM)算法,考虑到在利用LSSVM模型解决分类问题时,采用传统的求解算法有如下弊端:计算复杂度高、求出的解不稀疏,模型受样本噪声的影响较大的问题,本文将模糊支持向量机与LSSVM的稀疏求解算法相结合,提出了模糊稀疏LSSVM算法,考虑到模糊隶属度对模糊LSSVM的重要性,本文采取了两种基于样本到所属类别类中心距的计算方法来度量模糊隶属度,即基于欧式距离的度量方法和基于核距离的度量方法。并将提出的算法对极化SAR数据进行分类,从分类结果和对比实验上看,提出的算法性能更强。2.对LSSVM的核函数进行了研究。由于在求解非线性分类问题时,首要的任务是选择符合条件的核函数,将样本映射到高维空间,从而实现高维空间线性可分,因而核函数的选取是关键。常用的核函数为径向基核函数,但其在拟合较为复杂的函数时,得到的效果并不十分满意,本文首先分析了支持向量机核函数的条件,然后将稀疏LSSVM分类器的核函数改为了符合核函数条件的Morlet小波核函数,提出了一种小波核稀疏LSSVM算法,经过对极化SAR数据分类的对比实验证明,Morlet小波核稀疏LSSVM模型对极化SAR数据的分类精度更高。3.对提出的小波核稀疏LSSVM算法进行扩展。将该算法与深度SVM的网络架构相结合,提出了一种深度小波核稀疏LSSVM模型,模型主要设计了2个隐层单元,将低层的支持向量对应的激活值作为高层的训练样本,依此训练出深度小波核稀疏LSSVM分类器,考虑到算法的时间复杂度,将第二层的求解采用Lasso算法,依此构建出深度Lasso模型,将深度Lasso算法和深度小波核稀疏LSSVM算法分别采用了UCI数据和极化SAR数据进行验证。
[Abstract]:Polarimetric synthetic Aperture Radar (Polarimetric Synthetic Aperture) has become a new hotspot in the field of remote sensing due to its ability to provide more abundant target scattering information. As an important research content and key technology of polarimetric SAR image interpretation, the classification of ground objects has great theoretical significance and application value in civil and military fields. Support Vector Machine (SVM) is an effective supervised classification method in statistics, which has been widely used in many fields. In this paper, SVM algorithm is used to study the classification of polarimetric SAR ground objects. The main research results are as follows: 1. In this paper, the least squares support vector machine (LSSVM) algorithm in support vector machine (SVM) algorithm is studied. Considering the disadvantages of the traditional algorithm when using LSSVM model to solve the classification problem, the computational complexity is high, and the solution is not sparse. The model is greatly affected by the sample noise. In this paper, the fuzzy sparse LSSVM algorithm is proposed by combining the fuzzy support vector machine (FSVM) with the sparse solution algorithm of LSSVM, considering the importance of fuzzy membership to fuzzy LSSVM. In this paper, we adopt two methods to measure fuzzy membership based on the distance between the sample and the center of the class, that is, the Euclidean distance based metric method and the kernel distance based measurement method. The proposed algorithm is used to classify the polarimetric SAR data. The proposed algorithm has better performance from the classification results and comparative experiments. The kernel function of LSSVM is studied. When solving nonlinear classification problem, the most important task is to select the kernel function that meets the conditions, and map the sample to the high-dimensional space, thus realizing the linear separability of the high-dimensional space, so the selection of kernel function is the key. The commonly used kernel function is radial basis kernel function, but the result is not very satisfactory when fitting the more complex function. In this paper, the condition of kernel function of support vector machine is analyzed. Then, the kernel function of the sparse LSSVM classifier is changed to the Morlet wavelet kernel function which accords with the kernel function condition, and a wavelet kernel sparse LSSVM algorithm is proposed. The contrast experiment of polarized SAR data classification shows that the sparse LSSVM model based on Morlet wavelet kernel has higher classification accuracy for polarimetric SAR data. The proposed wavelet kernel sparse LSSVM algorithm is extended. Combining the algorithm with the network architecture of deep SVM, a sparse LSSVM model with deep wavelet kernel is proposed. Two hidden layer units are designed in the model, and the activation values corresponding to the support vectors in the lower layer are used as the training samples at the higher level. Based on this, the sparse LSSVM classifier with depth wavelet kernel is trained. Considering the time complexity of the algorithm, the Lasso algorithm is used to solve the second layer, and then the depth Lasso model is constructed. The deep Lasso algorithm and the deep wavelet kernel sparse LSSVM algorithm are validated by UCI data and polarized SAR data, respectively.
【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】