极化SAR图像特征提取与分类方法研究

发布时间：2018-03-12 19:08

本文选题：极化SAR　切入点：特征提取　出处：《电子科技大学》2016年硕士论文　论文类型：学位论文

【摘要】：极化SAR凭借其全天时、全天候的工作特性,已走在了遥感信息获取技术的前列。随着获取的信息越来越多,如何快速而准确的解译这些信息,已成为目前研究的热点问题。而极化SAR图像分类作为图像解译的重要步骤,引起了学者越来越多的重视。此外,极化SAR图像分类在军事情报、土地利用、森林监测等多个领域的作用也不可小觑。因此开展极化SAR图像分类研究,对拓宽极化SAR系统的应用能力具有重大的现实意义。显然,分类特征的提取和分类算法的设计是实现分类的重要前提。信息爆炸时代衍生出了海量的分类特征,如何从这些特征中挑选最有限、最高效、最本质的特征组合,是优化特征空间、提高分类器性能的关键。此外,为了将极化SAR更好的运用到实际研究中,也为了从更全面的角度分析问题,文中对比分析了单、双、全极化SAR在分类性能上的差异及原因。本文以Radarsat-2和ALOS-PALSAR全极化数据为例,重点研究了极化SAR图像特征提取技术、极化SAR图像分类方法、极化SAR图像分类特征选择以及单、双、全极化SAR分类性能对比等内容。主要工作和成果如下:(1)极化SAR图像特征提取。基于极化SAR的特性,提取了极化比、雷达植被指数等极化SAR特有的特征参数;构建协方差矩阵和相干矩阵,提取其元素;依托Cloude分解和Freeman分解,提取目标分解参数;基于灰度共生矩阵,提取纹理参数。至此,共提取了33维分类特征,并重点分析了不同类型的特征对分类结果的贡献度,结果表明纹理参数贡献度最高,极化参数最低。(2)极化SAR图像分类算法实现。本文分别采用支持向量机和随机森林算法,对实验区做监督分类,验证了算法在实验区内的有效性,并对比分析了二者的分类效果。另外,采用网格搜索法寻得了支持向量机的最优参数。(3)极化SAR图像特征选择。采用遗传算法,定义适合的适应度函数,经过多次寻优迭代,最终将原始33维特征空间简化为10维。采用支持向量机和随机森林算法对特征选择前后的分类效果进行评价,结果显示二者的整体精度和Kappa系数均稍有提高。体现了特征选择的必要性和优越性。(4)单、双、全极化SAR分类性能对比。从定量和定性角度评价分类效果,可得:在分类性能方面,全极化SAR最优,双极化次之,单极化最差;分析不同双极化组合的分类差异,结果表明HH-VV极化组合可作为全极化SAR的一种合理替代,并从H/α空间相似性的角度给出原因;同样,对不同单极化通道的分类效果进行评价,并阐述原因。
[Abstract]:Polarimetric SAR has been in the forefront of remote sensing information acquisition technology with its all-day, all-weather working characteristics. As more and more information is obtained, how to interpret this information quickly and accurately, Polarimetric SAR image classification, as an important step in image interpretation, has attracted more and more attention of scholars. In addition, polarimetric SAR image classification in military information, land use, The role of forest monitoring and other fields can not be underestimated. Therefore, it is of great practical significance to develop the classification of polarized SAR images to broaden the application capability of polarimetric SAR system. The extraction of classification features and the design of classification algorithms are the important prerequisites for classification. In the age of information explosion, there are a lot of classification features derived from them, how to select the most limited, efficient and essential feature combinations from these features. It is the key to optimize the feature space and improve the performance of classifier. In addition, in order to better apply polarized SAR to practical research, and to analyze the problem from a more comprehensive perspective, this paper compares and analyzes single and double, In this paper, we take Radarsat-2 and ALOS-PALSAR full polarization data as examples, we focus on the study of polarimetric SAR image feature extraction technology, polarimetric SAR image classification method, polarization SAR image classification feature selection and single, double, single, double, single, double, single, double, single, double, and single, double, and single, double, and single, double, and single, double, The main work and results are as follows: 1) feature extraction of polarimetric SAR images. Based on the characteristics of polarimetric SAR, the characteristic parameters of polarimetric SAR, such as polarimetric ratio and radar vegetation index, are extracted. The covariance matrix and coherent matrix are constructed to extract the elements; the target decomposition parameters are extracted by Cloude decomposition and Freeman decomposition; the texture parameters are extracted based on gray level co-occurrence matrix. So far, 33 dimensional classification features are extracted. The contribution of different types of features to the classification results is analyzed. The results show that the texture parameters are the highest, and the polarization parameters are the lowest. The polarimetric SAR image classification algorithm is implemented. Support vector machine and stochastic forest algorithm are used in this paper, respectively. The effectiveness of the algorithm in the experimental area is verified by the supervised classification of the experimental area, and the classification effect of the two methods is compared and analyzed. The optimal parameter of support vector machine (SVM) is found by grid search method. The feature selection of polarimetric SAR image is obtained. The fitness function is defined by genetic algorithm, and after several optimization iterations, Finally, the original 33 dimensional feature space is simplified to 10 dimension. Support vector machine and stochastic forest algorithm are used to evaluate the classification effect before and after feature selection. The results show that the overall accuracy and Kappa coefficient of both are improved slightly, which reflects the necessity and superiority of feature selection. The classification performance of single, double and fully polarized SAR is compared. The classification effect is evaluated from quantitative and qualitative aspects. In terms of classification performance, the fully polarized SAR is the best, the double polarization is the second, and the single polarization is the worst. The results show that the HH-VV polarization combination can be used as a reasonable substitute for the fully polarized SAR. The reasons are given from the angle of spatial similarity of H / 伪, and the classification effect of different single-polarization channels is evaluated, and the reasons are explained.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TN957.52

【参考文献】