基于区域相似度和特征降维的极化SAR影像分类

发布时间：2018-05-01 15:39

本文选题：极化SAR + 区域统计分割　；参考：《武汉大学》2013年博士论文

【摘要】：极化SAR(Polarimetric Synthetic Aperture Radar, PolSAR)影像分类是SAR影像处理分析的重要研究内容,经过几十年的发展,已提出了一系列的算法,在像素分类研究方面取得了长足的进步,然而像素分类无法克服斑点噪声影响,同时不能充分利用影像的空间信息,因而难以实现影像的高精度分类,而基于区域特征的影像分类能够有效克服上述不足,因此本文主要开展基于区域特征的极化SAR影像分类研究。区域特征分类包括影像分割和分割后的区域分类,文中首先总结了极化SAR影像在这两方面的研究进展,指出目前分割研究中的难点是如何在顾及计算效率前提下,降低斑点噪声影响,快速得到影像的准确分割结果；缺乏有效初始聚类方法,制约区域非监督分类的发展；而监督分类中的研究趋势则是极化散射特征的有效利用。针对上述难题,通过总结已有研究成果,引用借鉴图分割研究、聚类研究和数据降维研究方面的算法,文中发展了一套区域分割和分类算法,总体而言,在分割研究方面,文中采用区域统计分割,并改进初分割和区域合并处理；在非监督分类研究方面,提出基于区域相似度的初始聚类策略,在监督分类研究中,引入降维处理以提取特征,具体来说,本文的主要研究工作总结如下： (1)提出了一种应用区域统计特征的极化SAR影像分割方法。该方法采用分水岭算法初分割影像,针对传统差值梯度非恒虚警率(Constant False Alarm Rate, CFAR)和分水岭本身固有的过度过分割问题,引入了CFAR的均值比率梯度和测地学腐蚀重构,得到了一个良好的初分割结果；在区域合并方面,结合SAR数据的统计特征和假设检验,给出了一种能更好描述区域相似度的目标函数,确保得到更好的分割结果。 (2)将归一化割(Normal cut, NCut)应用到区域初始聚类研究。基于修正Wishart距离,构建分割后区域的相似度矩阵,利用Ncut算法处理相似度矩阵实现,得到影像的初始聚类结果。 (3)借鉴聚类趋势研究中相似度矩阵处理重排序策略,结合K均值算法,发展了一种区域初始聚类方法。利用聚类趋势研究中相似度矩阵排序方法,实现初步聚类,计算聚类中心与各区域相似度,利用K均值迭代得到稳定的初始聚类结果。 (4)引入近邻传播算法进行区域初始聚类研究。利用近邻传播算法处理相似度矩阵,通过调整参数,得到设定数目初始聚类结果。 (5)引入拉普拉斯映射和监督拉普拉斯映射,开展极化SAR影像特征降维研究。提取分割结果中各区域的极化散射特征,执行降维处理,达到去除冗余信息、从而有效利用极化散射特征的目的,确保得到高精度的监督分类结果。
[Abstract]:Polarimetric SAR(Polarimetric Synthetic Aperture Radar, PolSAR) image classification is an important research content in SAR image processing and analysis. After decades of development, a series of algorithms have been proposed, and great progress has been made in pixel classification. However, pixel classification can not overcome the influence of speckle noise and make full use of the spatial information of the image, so it is difficult to realize the high accuracy classification of the image. However, the image classification based on the regional features can effectively overcome the above shortcomings. Therefore, this paper mainly focuses on the classification of polarimetric SAR images based on regional features. Regional feature classification includes image segmentation and segmentation. In this paper, we first summarize the research progress of polarimetric SAR images in these two aspects, and point out that the difficulty of current segmentation research is how to take into account the computational efficiency. In order to reduce the influence of speckle noise, the accurate segmentation results can be obtained quickly. The lack of effective initial clustering method restricts the development of unsupervised classification. The research trend of supervised classification is the effective use of polarimetric scattering features. In view of the above problems, by summarizing the existing research results, referring to the algorithms of graph segmentation, clustering and data dimensionality reduction, this paper develops a set of region segmentation and classification algorithms. In the research of unsupervised classification, an initial clustering strategy based on regional similarity is proposed. In supervised classification research, dimensionality reduction is introduced to extract features. Specifically, the main research work of this paper is summarized as follows: In this paper, a new method of polarimetric SAR image segmentation based on regional statistical features is proposed. In this method, watershed algorithm is used to segment the original image. Aiming at the over segmentation problem inherent in the traditional differential gradient non-constant false alarm rate (False Alarm Rate, CFAR) and watershed itself, the mean ratio gradient of CFAR and geodesic corrosion reconstruction are introduced. A good initial segmentation result is obtained, and an objective function which can better describe the similarity of region is given by combining the statistical features and hypothesis tests of SAR data in the field of region merging to ensure a better segmentation result. (2) Normalized cut (NCut) is applied to the study of regional initial clustering. Based on the modified Wishart distance, the similarity matrix of the segmented region is constructed, and the initial clustering result of the image is obtained by using the Ncut algorithm to deal with the similarity matrix. Thirdly, a region initial clustering method is developed based on similarity matrix reordering strategy and K-means algorithm. The similarity matrix ranking method in clustering trend research is used to realize the preliminary clustering, to calculate the similarity between the clustering center and the regions, and to obtain the stable initial clustering results by using the K-means iteration. 4) introducing the nearest neighbor propagation algorithm to study the initial clustering of the region. The nearest neighbor propagation algorithm is used to deal with the similarity matrix and the initial clustering results are obtained by adjusting the parameters. 5) introducing Laplace mapping and supervised Laplace mapping to study the feature reduction of polarimetric SAR images. The polarimetric scattering features of each region in the segmentation results are extracted and dimensionally reduced to remove redundant information. Thus the polarimetric scattering features are utilized effectively and the supervised classification results with high accuracy are ensured.
【学位授予单位】：武汉大学
【学位级别】：博士
【学位授予年份】：2013
【分类号】：P23;P225.1

【相似文献】