基于多特征融合的高分辨率极化SAR图像分类方法研究

发布时间：2018-07-04 11:53

本文选题：高分辨率 + PolSAR图像分类　；参考：《电子科技大学》2017年硕士论文

【摘要】：作为极化SAR图像解译中的基本方法之一,极化SAR图像分类处理同样也是极化SAR成像技术迈向实际工程应用的重要步骤,因而日益成为研究人员探讨的热点问题。极化SAR传感器呈现出的其中一个关键技术方向是系统分辨率的显著提升。分辨率的提升带来更加丰富特征信息的同时也向极化SAR图像处理制造了新的麻烦。高分辨率极化SAR图像的空间结构信息更明显、地物起伏更剧烈、统计特性发生改变、数据量迅速增大、场景更为复杂等问题,使得原本中低分辨率前提下的各类极化SAR图像分类算法已不能适应高分辨率极化SAR图像中的实际应用需要。本文从现有的图像处理方式出发,针对上述情况,研究了适用于高分辨率极化SAR图像的分类算法。论文的具体研究内容主要包括如下四个方面:1.研究了包含综合后向散射信息的极化特征构建方法。高分辨率极化SAR图像地物起伏剧烈,传统的单一极化参量无法全面描述地物细节特性。经过极化代数运算或极化目标分解两类方法提取的极化特征,从不同角度解译了极化SAR数据。将它们收集起来共同组建高维极化特征空间,可以获取综合的回波散射信息。2.根据纹理结构信息更明显的特质,在高分辨率极化SAR图像分类中引入空间特征。为了有效捕获极化SAR图像的空间信息,选取了Gabor、MP、AP这三种特征。并在提取空间特征前计算SPAN基图像作为预处理手段以克服极化SAR图像固有相干斑噪声的影响。本文通过实验验证了空间特征作用于高分辨率极化SAR图像进行分类的有效性。3.将结合辨别分析与流形学习的LPDA算法作用于高分辨率极化SAR图像分类中,以解决待处理数据量增大的难题。该方法通过在一定准则下将特征向量的结构最优化,来保留类内流形结构并使类间可分性最大。实验证明,该降维算法能够在减少冗余的同时保持原有高维特征的分类正确性。4.面对高分辨率条件下极化SAR图像场景复杂性加大的情况,本文探讨了将极化特性与空间特性相结合为分类所用的算法。其中如何将两类特性结合的难题是采用信息融合技术解决的。两种融合策略被考虑,分别是基于特征级与决策级。第一种策略中,每个像素上两类特征向量首先通过矢量叠加或构造复合核函数结合在一起,再由一个SVM获得分类结果;第二种策略中,极化SAR图像首先由极化特征和空间特征分别进行分类,再将分类结果结合着给出最终分类图。实验结果证明了结合极化及空间信息为高分辨率极化SAR图像进行分类的优效性。
[Abstract]:As one of the basic methods in polarimetric SAR image interpretation, polarimetric SAR image classification and processing is also an important step of polarimetric SAR imaging technology towards practical engineering applications. One of the key technologies presented by polarimetric SAR sensors is the significant improvement of system resolution. The enhancement of resolution not only enriches the feature information, but also makes new trouble for polarimetric SAR image processing. The spatial structure information of high-resolution polarimetric SAR images is more obvious, the ground objects fluctuate more violently, the statistical characteristics change, the data volume increases rapidly, the scene becomes more complex and so on. All kinds of polarimetric SAR image classification algorithms based on low resolution and low resolution can no longer meet the needs of practical applications in high-resolution polarimetric SAR images. Based on the existing image processing methods, a classification algorithm for high resolution polarimetric SAR images is studied in this paper. The specific research contents of the thesis mainly include the following four aspects: 1. A method for constructing polarization features with integrated backscattering information is studied. High resolution polarimetric SAR images fluctuate dramatically, and the traditional single polarization parameter can not fully describe the detailed features of the ground objects. Polarimetric SAR data are interpreted from different angles by polarimetric algebraic operation or polarimetric target decomposition. When they are collected together to form a high dimensional polarization feature space, the integrated echo scattering information. 2. 2 can be obtained. Spatial features are introduced into high resolution polarimetric SAR image classification according to the more obvious characteristics of texture structure information. In order to capture the spatial information of polarimetric SAR image effectively, three features of Gabor-MPO AP are selected. In order to overcome the influence of the inherent speckle noise in polarimetric SAR images, the span based image is calculated as a preprocessing method before extracting the spatial features. In this paper, the effectiveness of spatial feature in high resolution polarimetric SAR image classification is verified by experiments. The LPDA algorithm combined with discrimination analysis and manifold learning is applied to the classification of high-resolution polarimetric SAR images to solve the problem of increasing the amount of data to be processed. By optimizing the structure of the eigenvector under certain criteria, the method preserves the structure of the inner manifold and maximizes the separability among classes. Experiments show that the algorithm can reduce redundancy and maintain the classification correctness of the original high-dimensional features. In view of the complexity of polarimetric SAR images under the condition of high resolution, this paper discusses the algorithm of combining polarization characteristics with spatial characteristics to classify SAR images. The problem of how to combine the two kinds of characteristics is solved by information fusion technology. Two fusion strategies are considered, one is based on feature level and the other is based on decision level. In the first strategy, two kinds of Eigenvectors on each pixel are first combined by vector superposition or by constructing a composite kernel function, and then a SVM is used to obtain the classification results; in the second strategy, Polarimetric SAR images are classified by polarimetric features and spatial features respectively, and then the classification results are combined to give the final classification map. The experimental results show that the high resolution polarimetric SAR images can be classified with polarimetric and spatial information.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN957.52

【相似文献】