基于多层网络模型的全极化SAR图像分类

发布时间：2018-01-13 08:46

本文关键词：基于多层网络模型的全极化SAR图像分类　出处：《武汉大学》2015年博士论文　论文类型：学位论文

【摘要】：全极化合成孔径雷达(Polarimetric Synthetic Aperture Radar, PolSAR)是一种先进的对地测量系统,能同时获取4个极化通道的SAR图像。因此,通过全极化SAR图像可以获得丰富的地物信息。图像分类是极化SAR系统的一个重要研究内容,在农林业规划、环境保护等领域都有着广泛的应用,开展全极化SAR图像分类研究对于提高SAR遥感的应用水平具有重要的理论意义及实用价值。近年来,在光学领域图像分类算法发展迅速,出现了很多新的模型或概念,如词袋模型,空间金字塔、稀疏编码,特征表达等,从事全极化SAR图像分类的研究人员纷纷借鉴光学图像分类中的优秀算法和概念,针对全极化SAR图像提出了很多新的的特征表达和特征编码方法,且取得了较好的成绩。2006年,Hinton等人在光学领域首次提出了深度学习的概念,开启了特征学习的研究,它能通过构建多层网络模型自动的从原始图像中学习出更本质的特征,从而有利于分类研究。此后,深度学习的相关研究如火如荼,在光学图像分类领域更是创造了诸多奇迹。本文引入深度学习的思想进而实现全极化SAR图像的分类,但极化SAR图像不同于光学图像,不能将光学图像中的深度学习模型直接用于全极化SAR图像分类,主要存在如下几个方面的问题：(1) SAR图像与光学图像在成像方式上有很大不同,光学图像是通过可见光传感器成像,可以获得地物的灰度信息,而SAR是通过微波传感器成像,然后以二进制复数形式记录地物的回波信息；另外,SAR图像所固有的相干斑噪声十分严重,信噪比极低,大部分信息都被淹没在相干斑噪声里,严重影响了全极化SAR图像解译及后续的应用,因此PolSAR数据需要处理后才能利用深度学习模型进行分类；(2) 全极化SAR能同时获取4个不同通道的SAR图像,在互易媒质的后向散射情况下,同一地物也对应着3幅单极化SAR图像,而原有的深度学习模型都是建立在单通道数据上的,不能充分的利用全极化SAR图像丰富的地物信息；(3) 深度学习往往需要大量数据对多层网络模型进行训练,而目标全极化SAR图像往往没有足够的数据来训练多层网络模型。为了充分利用深度学习的优势进行全极化SAR图像分类,就需要在深度学习和极化SAR图像之间构建一座桥梁,同时需要构建适合全极化SAR图像的多层网络模型用于特征学习和分类。本文从极化SAR基础理论出发,在描述极化SAR统计分布模型和极化分解等原理的基础上引入了深度学习,为了解决深度学习引入全极化SAR图像过程中遇到的一些问题,本文主要做了3个方面的工作：(1) 考虑到SAR图像的成像机理不同于光学图像,深度学习往往不能直接用于极化SAR图像分类,本文构建了一个基于统计分布的网络结构单元,以在深度学习和全极化SAR图像之间构建一座桥梁；(2) 为了充分地利用全极化SAR所包含的丰富地物信息,本文从特征融合和特征学习两个不同的角度构建了两个多层网络模型用于特征提取和分类。第1个多层网络融合了多种类型的特征,采用字典学习实现了空间金字塔特征表达,并构建了一个双层SVM实现了全极化SAR图像的分类。第2个多层网络是对多层反卷积网络进行了改进,构建了一个适合全极化SAR图像分类的多层反卷积网络,同时在反卷积网络中引入了一种新的软概率池化方法;(3) 考虑到目标极化SAR没有足够的数据用于多层网络的训练,本文引入了迁移学习的方法,采用相似的极化SAR数据对多层网络进行训练,将多层网络学习的特征作为中层表达,再用目标极化SAR图像对中层表达迁移学习,以能对目标极化SAR数据进行更准确的分类。论文在解决上述理论和技术问题的基础上,利用多个多层网络模型对全极化SAR图像进行了分类研究,在中国电子集团第三十八研究所获取的X波段单航迹海南省陵水县全极化SAR数据上进行了实验,实验结果表明多层网络模型在全极化SAR图像分类领域确实具有很大的潜力。
[Abstract]:Polarimetric synthetic aperture radar (Polarimetric Synthetic Aperture Radar, PolSAR) is a kind of advanced measurement system, SAR image is able to obtain the 4 polarization channels. Therefore, by fully polarimetric SAR images can provide abundant information. Image classification is an important research content of polarimetric SAR system, in agriculture and forestry planning, environmental protection and other fields have a wide range of applications, to carry out the classification of polarimetric SAR images has important theoretical significance and practical value to improve the application level of SAR remote sensing. In recent years, the development in the field of optical image classification algorithm rapidly, there are many models or new concepts, such as the bag of words model, space in Pyramid sparse encoding, feature, expression, researchers engaged in polarimetric SAR image classification algorithm and the concept of reference have excellent optical image classification, based on the full polarization SAR images are proposed. A lot of new feature expression and feature encoding method, and achieved good results in.2006, Hinton et al first proposed the concept of deep learning in the field of optics, the study on characteristics of learning, it can through the construction of automatic learning out of the essential characteristics of the original image from the secondary multi-layer model so as to facilitate the classification study. Since then, the related research of deep learning in optical image classification field like a raging fire, but also created many miracles. This paper introduces the idea of deep learning so as to realize the classification of polarimetric SAR images, but the polarization SAR image is different from optical image, can not be deep learning model in the optical image directly for polarimetric SAR image classification. Mainly has the following several aspects: (1) SAR and optical images are very different in the imaging mode, the optical image is visible through the sensor imaging, can In order to obtain the gray information of objects, while the SAR is through the microwave imaging sensor, and then to echo information of plural recording binary objects; in addition, coherence inherent speckle noise in SAR images is very serious, the signal-to-noise ratio is very low, most of the information that is submerged in the speckle noise, seriously affecting the polarization SAR image solution translation and subsequent application, so PolSAR data needs to be processed before the use of deep learning model classification; (2) fully polarimetric SAR can obtain 4 different channels of SAR image at the same time, the reciprocal medium backscatter case, the same object is corresponding with 3 single polarization SAR images, and the original depth learning models are based on single channel data, feature information of fully polarimetric SAR image can't take advantage of the rich; (3) deep learning often need training on multilayer network model for large amounts of data, and the goal of all Polarimetric SAR images often do not have enough data to train the multilayer network model. In order to make full use of deep learning the advantages of fully polarimetric SAR image classification, we need to build a bridge between deep learning and polarization SAR image, also need to build features for learning and classification of multilayer network model for fully polarimetric SAR images. This paper from the polarization SAR the basic theory, the deep learning is introduced based on SAR statistical distribution model to describe the polarization and polarization decomposition principle, in order to solve some problems encountered in deep learning into full polarimetric SAR image process, this paper has 3 aspects: (1) considering the imaging mechanism of SAR image of Yu Guangxue deep learning image, often can not be directly used for classification of polarimetric SAR images, this paper constructs a network structure based on the statistical distribution of the unit, in the depth of learning and Build a bridge between fully polarimetric SAR image; (2) in order to enrich the feature information of full use of polarimetric SAR contains, the feature fusion and feature learning from two different angles constructed two multilayer network model for feature extraction and classification. The first characteristics of various types of multi network fusion, using dictionary learning can express the space characteristic of Pyramid, and the construction of a double SVM to achieve the classification of fully polarimetric SAR image. Second multilayer network is of multilayer deconvolutional networks is improved, construct a suitable polarimetric SAR image deconvolution multilayer network classification, and introduces a new soft probability pool in the method of deconvolution in the network; (3) taking into account the target SAR does not have enough data for training multilayer network, this paper introduces a method of transfer learning, using polarimetric SAR data of similar Multilayer neural network training, will feature multilayer network learning as the middle expression, then the target polarimetric SAR image transfer learning of middle expression, to carry out a more accurate classification of polarimetric SAR data. Based on the theory and technology to solve the above problems, the use of multiple multilayer network model classification research on full polarization SAR images acquired in the study of thirty-eighth China Electronics Group X band single track in Lingshui County of Hainan province fully polarimetric SAR data for the experiment results show that it has great potential in the field of multilayer network model of fully polarimetric SAR image.

【学位授予单位】：武汉大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：P237

【相似文献】