基于深度SVM和深度小波神经网络的极化SAR影像地物分类

发布时间：2018-05-04 10:28

本文选题：极化SAR + 深度SVM　；参考：《西安电子科技大学》2014年硕士论文

【摘要】：作为主动式雷达成像系统的合成孔径雷达具有全天时、全天候的特点,可以提供大尺度二维高分辨率地球表面反射率图像。目前,它已经获取了大量的SAR图像,这些图像被广泛地应用于战略目标探测、作物生长监测、森林监控等领域。极化合成孔径雷达是在合成孔径雷达的基础上发展起来的,它能够获取更为丰富有用的信息。与传统的单极化SAR相比,我们通常用极化矩阵形式表示极化合成孔径雷达测量的数据,而散射矩阵通常又与其对应的散射机理模型有关,因此其信息的挖掘和提取更为复杂。分类是对极化SAR数据进行解译和理解研究的一个重要方向,分类得到的结果图既可以作为中间结果继续用于进行边缘提取、目标检测和识别等操作,也可以作为最终结果直接输出给用户。因此,对极化SAR数据进行分类研究是很有必要的,而且具有一定的现实价值。时至今日,已经提出了不少关于极化SAR影像地物分类的算法。本文在传统算法的基础上,主要针对极化SAR图像分类器的设计进行了两大方面的研究,所取得的主要研究成果如下:第一部分工作为:深度SVM分类算法的研究。将最近提出的深度SVM的算法应用到极化SAR分类问题中,并对深度SVM的算法进行了一系列的改进。主要工作包括:(1)基于深度稀疏SVM的极化SAR影像地物分类。将深度SVM的模型和稀疏的最小二乘的LS-SVM算法相结合,提出了一种基于深度稀疏SVM的极化SAR影像地物分类方法,解决了原算法运行速度慢的问题。利用本章提出的算法进行仿真实验,得到的实验结果表明算法具有一定的可实施性,并且将算法与其他同种类型算法相比,体现了该算法的高效性和准确性。(2)基于协同深度SVM的极化SAR影像地物分类。将协同训练引入到深度SVM分类器的训练中,提出了一种基于协同深度SVM的极化SAR影像地物分类方法。通过协同训练的思想训练深度SVM分类器,使深度SVM在训练的过程中能同时利用有标记和无标记样本的信息,从而提高分类的准确率。(3)基于深度小波核SVM的极化SAR影像地物分类。将深度SVM分类器中的SVM的高斯核替换成小波核,提出了一种基于深度小波核SVM的分类器模型。针对提出的方法进行仿真实验,实验结果表明该方法能够更好的拟合数据,从而提高分类的准确率。第二部分工作为:神经网络算法的研究。通过对小波神经网络和深度学习的这两种算法进一步研究,提出了一种基于深度小波神经网络的极化SAR影像地物分类方法。深度小波神经网络采用深度学习的训练方式训练深度小波神经网络,避免了网络层数较多时出现梯度扩散的问题,并且能够提取出反映数据本质特性,刻画数据细节特征,突出不同地物类型之间差别的高维特征。实验表明,该方法能达到较好的结果。
[Abstract]:Synthetic Aperture Radar (SAR), as an active radar imaging system, has the characteristics of all-day, all-weather, and can provide large scale two-dimensional high-resolution reflectivity images of the earth's surface. At present, it has acquired a large number of SAR images, which are widely used in strategic target detection, crop growth monitoring, forest monitoring and other fields. Polarimetric synthetic Aperture Radar is developed on the basis of synthetic Aperture Radar, which can obtain more useful information. Compared with the traditional single-polarization SAR, we usually use polarization matrix to represent the data measured by polarimetric synthetic aperture radar, and the scattering matrix is usually related to the corresponding scattering mechanism model, so it is more complicated to mine and extract the information. Classification is an important direction in the interpretation and understanding of polarized SAR data. The result map can be used as intermediate results for edge detection, target detection and recognition. It can also be output directly to the user as the final result. Therefore, it is necessary and valuable to classify polarized SAR data. Up to now, many algorithms for ground object classification in polarimetric SAR images have been proposed. Based on the traditional algorithm, this paper mainly studies the design of polarimetric SAR image classifier. The main research results are as follows: the first part is the research of depth SVM classification algorithm. The recently proposed algorithm of depth SVM is applied to the problem of polarized SAR classification, and a series of improvements are made to the algorithm of depth SVM. The main work includes 1: 1) ground object classification of polarized SAR images based on deep sparse SVM. By combining the model of depth SVM with the sparse least square LS-SVM algorithm, a method of ground object classification for polarized SAR images based on deep sparse SVM is proposed, which solves the problem of slow running speed of the original algorithm. The simulation results show that the algorithm is feasible, and the algorithm is compared with other algorithms of the same type. The efficiency and accuracy of the algorithm are demonstrated. (2) based on the cooperative depth SVM, the ground object classification of polarimetric SAR images is presented. This paper introduces cooperative training into the training of depth SVM classifier, and proposes a method for ground object classification of polarimetric SAR images based on cooperative depth SVM. The depth SVM classifier is trained with the idea of cooperative training. The depth SVM can use the information of labeled and unmarked samples simultaneously in the process of training, so as to improve the accuracy of classification. (3) polarimetric SAR image ground object classification based on depth wavelet kernel SVM. The Gao Si kernel of the SVM in the depth SVM classifier is replaced by the wavelet kernel, and a classifier model based on the deep wavelet kernel SVM is proposed. The experimental results show that the proposed method can fit the data better and improve the accuracy of classification. The second part is the research of neural network algorithm. Based on the further study of wavelet neural network and depth learning, a method of ground object classification for polarimetric SAR images based on depth wavelet neural network is proposed. Depth wavelet neural network is trained by deep learning, which avoids the problem of gradient diffusion when there are more layers, and can extract the essential characteristics of data and depict the details of data. Highlight the high dimensional characteristics of the difference between different types of features. Experiments show that the method can achieve better results.
【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;TP183

【相似文献】