基于图与改进Wishart距离的极化SAR分类研究

发布时间：2019-05-16 15:16

【摘要】：极化合成孔径雷达(Polarimetric Synthetic Aperture Radar,简称极化SAR)是一种先进的获取遥感信息的手段,通过测量地面每一个分辨单元在四种不同的极化组合下的散射特性,从而得到目标对应的极化信息,极化SAR比传统单极化SAR所记录的地物目标电磁散射特征信息更完整。极化SAR分类的意义重要在于它既能作为一个中间步骤,为极化SAR的解译提供帮助,协助极化SAR图像提取边缘信息、检测目标、识别目标,也可能是用户最终需求。传统的极化SAR分类均是以单个像素为分类单元,而极化SAR特有的相干斑噪声对以像素为分类单元的分类结果影响很大,因此本文构造包含点和含权值的边的全连图,对极化SAR数据先进行过分割,减小相干斑噪声对分类的影响,然后以分割后的区域为分类单元,结合数据的极化特征和结构信息实现对极化SAR数据的分类研究,文章主要包含了以下三方面的内容:1.提出一种基于图方法的极化SAR分割方法。该方法中先提取像素点的极化特征,结合极化SAR数据的Wishart距离构建图,然后基于图对极化SAR数据进行初始分割,最后对基于图方法初始分割后的区域进行一个分层合并,按照区域大小的等级设定不同的合并策略,得到一个相对均匀的分割结果。该算法引入了应用在自然图像上的分割算法图方法,结合极化SAR数据的特点改进权值的计算方法,合并过程中考虑了像素的空间信息,思路简明,便于理解。2.提出一种基于图方法过分割的极化SAR有监督分类方法。该方法利用了上面介绍的分割方法得到分割结果,以过分割后的区域为分类单元,利用Wishart距离计算每个区域与各个训练类别之间的距离,对每个区域进行类别划分。该分类方法为基于区域的有监督分类,减小了传统分类结果中出现杂点的情况,并且提高了分类结果的区域一致性,而且提高了极化SAR的分类精度。3.提出一种基于图方法过分割以及改进的Wishart距离的极化SAR二分树分类方法。该方法同上面提到的有监督分类方法一样,以图方法分割得到的区域为分类单元,计算每两个区域之间的不相似度构建二分树,最终得到的分类树的个数即类别个数。该方法中用到的不相似度计算方式为改进的Wishart距离,考虑到区域的尺寸大小。该方法减少了分类结果中出现局部收敛的情况,减小了一般无监督方法中出现杂点的现象,且提升了分类精度。
[Abstract]:Polarization synthetic Aperture Radar (Polarimetric Synthetic Aperture Radar,) is an advanced method to obtain remote sensing information. The scattering characteristics of each resolution unit on the ground under four different polarization combinations are measured. Thus, the polarization information corresponding to the target is obtained, and the electromagnetic scattering characteristic information of the ground object recorded by the polarization SAR is more complete than that recorded by the traditional unipolar SAR. The significance of polarization SAR classification is that it can be used as an intermediate step to help the interpretation of polarization SAR, assist polarization SAR images to extract edge information, detect targets, identify targets, and may also be the final requirements of users. The traditional polarization SAR classification takes a single pixel as the classification unit, and the speckle noise unique to the polarization SAR has a great influence on the classification results with pixels as the classification unit. Therefore, this paper constructs a fully connected graph containing points and edges with weights. Firstly, the polarimetric SAR data is segmented to reduce the influence of speckle noise on the classification, and then the polarimetric SAR data classification is realized by taking the segmented region as the classification unit and combining the polarization characteristics and structural information of the data. The article mainly contains the following three aspects: 1. A polarization SAR segmentation method based on graph method is proposed. In this method, the polarization features of pixels are extracted, and the Wishart distance of polarized SAR data is combined to construct the graph, and then the polarized SAR data is initially segmented based on the graph. Finally, a hierarchical merging of the regions after the initial segmentation based on the graph method is carried out. Different merging strategies are set according to the level of region size, and a relatively uniform segmentation result is obtained. The algorithm introduces the graph method of segmentation algorithm applied to natural images, and improves the calculation method of weights according to the characteristics of polarized SAR data. The spatial information of pixels is considered in the process of merging, and the train of thought is simple and easy to understand. 2. A polarization SAR supervised classification method based on graph method is proposed. In this method, the segmentation results are obtained by using the segmentation method described above. Taking the over-segmented region as the classification unit, the distance between each region and each training category is calculated by using Wishart distance, and each region is classified. The classification method is region-based supervised classification, which reduces the occurrence of miscellaneous points in the traditional classification results, improves the regional consistency of the classification results, and improves the classification accuracy of polarized SAR. 3. A polarization SAR binary tree classification method based on graph method oversegmentation and improved Wishart distance is proposed. This method is the same as the supervised classification method mentioned above, taking the region segmented by graph method as the classification unit, and calculating the dissimilarity between each two regions to construct the binary tree, and the number of classification trees is the number of categories. The dissimilarity calculation method used in this method is the improved Wishart distance, which takes into account the size of the region. This method reduces the local convergence in the classification results, reduces the phenomenon of miscellaneous points in the general unsupervised method, and improves the classification accuracy.
【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

【共引文献】