面向对象的PolSAR影像分类研究

发布时间：2018-01-19 00:09

本文关键词： PolSAR 土地利用分类面向对象极化分解影像分割特征选择多分类器组合　出处：《吉林大学》2017年博士论文　论文类型：学位论文

【摘要】：近年来,随着雷达极化测量技术和合成孔径雷达(Synthetic Aperture Radar,SAR)成像技术的逐渐成熟,结合两者优势的极化SAR(Polarimetric SAR,PolSAR)应运而生,并逐渐取代了传统低分辨率、单极化体制的雷达,成为现代雷达系统发展的主流方向。有效的PolSAR影像解译技术是PolSAR成功应用的基础。PolSAR系统的极化数据要想体现其真正的价值,需要先通过影像解译技术的有效分析与解译。然而相比于比较成熟的PolSAR成像技术与系统设计,PolSAR影像解译技术的发展相对滞后,严重制约了PolSAR技术的应用,因此,开展PolSAR影像解译技术的研究对于提高PolSAR系统的应用水平具有重要意义。影像解译的一个重要内容就是土地利用分类。传统的基于像元的分类方法仅利用像元本身的特征进行分类,这类方法虽能较好地保留影像的细节信息,但对于PolSAR影像来说,其固有的相干斑噪声会使单个像元的散射测量值与真实值之间出现误差,造成分类结果中出现较多的小块区域和孤立像素,从而增加了解译的难度。面向对象分类方法通过分割获得同质性的区域对象,将对象内的像元看作一个整体能有效地降低相干斑噪声对影像信息的破坏程度。面向对象方法是遥感领域的新兴技术,在抑制PolSAR影像的相干斑噪声方面有重要作用。研究面向对象方法在PolSAR影像分类上的应用,对推动PolSAR影像分类的发展有着重要意义。针对目前PolSAR影像面向对象分类研究中存在的主要问题,考虑到现有的理论和问题的复杂性,本文结合极化测量理论、图像处理、计算机视觉、数据挖掘及模式识别等领域的最新技术,以吉林省长春市部分区域为研究区,全极化RADARSAT-2影像为数据源,按照“目标极化分解影像分割特征选择分类多分类器组合”的总体路线对相关问题进行了系统分析和深入研究。论文的主要工作和研究成果概括如下:1、多种经典的目标极化分解方法(Pauli分解、Krogager分解、Huynen分解、Barnes1分解、Barnes2分解、Cloude分解、H/A/α分解、Freeman2分解、Freeman3分解、Yamaguchi3分解、Yamaguchi4分解、Neumann分解、Touzi分解、Holm1分解、Holm2分解和Van Zyl分解)被同时用于分类。通过这些目标极化分解方法,本文共计提取了61个极化参数。将提取的所有极化参数进行合成,形成一幅多通道影像,以用于后续的特征提取和分类等研究。2、针对PolSAR影像的特点,提出了一种新的基于边缘信息和区域信息混合的PolSAR影像分割方法。为检验提出的分割算法的性能,选取了全极化RADARSAT-2影像的两个子区域进行分割实验,并与e Cognition软件中的FNEA算法的分割结果进行了对比。实验结果表明,本文提出的分割算法对全极化RADARSAT-2影像有很好的适用性,同时它还具有较好的抗噪性能,这对于PolSAR影像的分析解译是非常重要的。3、针对面向对象分类中存在的特征维数过高的问题,提出了一种结合ReliefF算法和遗传算法(Genetic Algorithm,GA)的混合特征选择方法,即首先利用Relief F算法作为特征预选器滤除相关性小的特征,然后以GA作为搜索算法,以分类精度作为评估函数在剩余特征中选出最优特征子集。该算法将Filter模式和Wrapper模式进行了有效地结合,充分利用了二者的优势。为验证提出的特征选择算法的有效性,与基于三种其他特征选择算法的土地利用分类结果进行比较。结果表明,本文提出的特征选择算法能够花费较少时间获得最高的分类精度,证明了该算法能够有效地用于面向对象土地利用分类。4、除了在PolSAR影像面向对象分类中常用的支持向量机(Support Vector Machine,SVM)算法和决策树(Decision Tree,DT)算法,本文还采用了随机森林(Random Forest,RF)算法、贝叶斯(Bayes)算法和k近邻(k-Nearest Neighbor,k NN)算法对研究区RADARSAT-2影像进行土地利用分类实验。实验结果显示,五种分类算法的分类精度从高到低的排序是:SVMRFDTk NNBayes;使用SVM算法进行土地利用分类的精度最高,总体精度和Kappa系数分别达到了92.60%和0.9022;而使用Bayes算法获得的精度最低,总体精度和Kappa系数分别仅为87.57%和0.8376。5、将多分类器组合方法引入到了PolSAR影像面向对象分类当中,首先利用Q统计量、相关系数和熵值三个指标,对SVM、DT、RF和k NN四个单分类器的所有组合形式的差异性进行度量,然后选取其中六个差异性较大的组合进行多分类器组合实验。实验结果表明:经过多分类器组合后,六个组合的总体精度和Kappa系数都得到了不同程度的提高,证明了多分类器组合方法在改善PolSAR影像面向对象分类效果上的有效性。
[Abstract]:In recent years, along with the radar polarization measurement technology and synthetic aperture radar (Synthetic Aperture, Radar, SAR) imaging technology is gradually mature, with both advantages of polarization SAR (Polarimetric SAR PolSAR) came into being, and gradually replaced the traditional low resolution, single polarization radar, has become the mainstream direction of the development of modern radar system of PolSAR images. Effective interpretation technique is polarization data base of.PolSAR system successfully used PolSAR to reflect its real value, need to go through effective analysis of image interpretation technology and interpretation. However, compared to the mature PolSAR imaging technology and system design, the development of PolSAR image interpretation technology is relatively backward, serious constraints the application of PolSAR technology, therefore, to carry out research on PolSAR image interpretation technology has an important significance for improving the application level of PolSAR system. A photointerpretation Is an important part of land use classification. The traditional pixel based classification method using only the feature pixel itself classified details, although this method can preserve the image, but for PolSAR image speckle noise makes the scattering measurement of single pixel error value and the true value between the inherent, small more areas and isolated pixels resulted from the classification results, thereby increasing the difficulty of interpretation. The object-oriented classification method by dividing the obtained region object homogeneity, the object within the pixel as a whole can effectively reduce speckle noise damage to the image information. The object-oriented method is a new technology in the field of remote sensing. There is an important role in suppressing speckle noise in PolSAR images. Research and application of object oriented method in PolSAR image classification, to promote PolSAR image Has an important significance to the development of the classification. The main problems existed in the research of the current PolSAR object oriented image classification, taking into account the complexity theory and the existing problems, combined with the theory of polarization measurement, image processing, computer vision, the latest technology in the field of data mining and pattern recognition, in some regions of Jilin province Changchun city the study area, fully polarimetric RADARSAT-2 images as the data source, in accordance with the "overall route selection target decomposition image segmentation feature classification and multi classifier combination" on issues related to the system analysis and research. The main work and research results are summarized as follows: 1, a variety of classic target polarization decomposition method (Pauli decomposition, Krogager decomposition, Huynen decomposition, Barnes1 decomposition, Barnes2 decomposition, Cloude decomposition, H/A/ alpha decomposition, Freeman2 decomposition, Freeman3 decomposition, Yamaguchi3 decomposition, Yamaguchi4 decomposition, Neuman N decomposition, Touzi decomposition, Holm1 decomposition, Holm2 decomposition and Van decomposition of Zyl) was also used for classification. Through these methods the target polarization decomposition, the total extraction of 61 polarization parameters. All polarization parameter extraction are combined to form a multi-channel imaging study, for the subsequent feature extraction and classification etc..2, according to the characteristics of PolSAR images, proposes a new segmentation method of PolSAR image edge information and region information based on mixed performance. To test the proposed segmentation algorithm, selects the two sub regional full polarimetric RADARSAT-2 image segmentation experiments, and results with the FNEA algorithm of E software in the Cognition segmentation compared. Experimental results show that the segmentation algorithm is proposed in this paper has good applicability of polarimetric RADARSAT-2 image, at the same time, it also has better anti noise performance, the analysis for PolSAR image interpretation It is very important to.3, according to the features of object oriented classification in dimension are high, we propose a combination of ReliefF algorithm and genetic algorithm (Genetic Algorithm GA) hybrid feature selection method, firstly the Relief F algorithm as a feature filter preselectors related characteristics of small, and then to GA as a search in the algorithm, the classification accuracy as the evaluation function to select the best subset of features in the remaining features. This algorithm combines Filter model and Wrapper model are combined effectively, make full use of the advantages of the two. The effectiveness of feature selection algorithm to verify the proposed election, and other three kinds of feature selection algorithm for land use classification results based on the comparison. The results show that the proposed feature selection algorithm can take less time to get the highest classification accuracy, proved that the algorithm can be effectively used for the object oriented land The classification of.4, in addition to the support vector machine commonly used in object classification in PolSAR images (Support Vector Machine SVM) algorithm and decision tree (Decision Tree DT) algorithm, this paper uses the random forest (Random Forest RF) algorithm, Bias (Bayes) algorithm and K nearest neighbor (k-Nearest Neighbor, K NN) algorithm of RADARSAT-2 image land use classification experiment. The experimental results show that the classification accuracy of the five classification algorithms from high to low ranking is: SVMRFDTk NNBayes; SVM algorithm is used to classify the land use of the highest accuracy, overall accuracy and Kappa coefficient respectively 92.60% and 0.9022; and using Bayes the accuracy of the algorithm is the lowest, the overall accuracy and Kappa coefficient are 87.57% and 0.8376.5, the method of the combination of multiple classifiers is introduced into a PolSAR object oriented image classification, the first use of Q statistic, correlation coefficient and The entropy of three indexes of SVM, DT, RF and K NN to measure four single classifiers all combinations of the differences, and then select one of the six different combinations of multiple classifiers combination experiment. The experimental results show that the combination of multiple classifiers, the combination of the six overall accuracy and Kappa coefficient improved, proved the effectiveness of multi classifier combination method in the improvement of PolSAR object oriented image classification results.

【学位授予单位】：吉林大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：P237

【相似文献】