节点文献

基于Cloude-Pottier分解的全极化SAR数据非监督分类的算法和实验研究

The Unsupervised Classification Based on the Cloude-Pottier Decomposition for Fully Polarimetric SAR Data of Chinese Academy of Sciences

【作者】 曹芳

【导师】 洪文;

【作者基本信息】 中国科学院研究生院(电子学研究所) , 信号与信息处理, 2007, 博士

【摘要】 全极化SAR数据的地物分类是遥感领域中雷达极化的最重要的应用之一。全极化SAR数据的优势在于,利用极化目标分解方法可以识别地表覆盖物的散射机制,该散射机制对所有的全极化SAR数据均是稳定的,因此,它可以实现非监督的分类方法,即不需要地表覆盖物类型的先验知识数据库。本文开展了全极化SAR数据非监督分类的算法和实验研究。基于目前全极化SAR数据的非监督分类研究最流行的方法,本文采用将目标分解算法和统计模型的聚类方法结合起来实现全极化SAR数据的非监督分类,其研究内容主要是开展目标分解方法研究提取全极化SAR数据有关散射机制的信息,并结合该散射机制的信息和统计模式识别的方法开展全极化SAR数据非监督分类算法的研究。就目标分解算法而言,本文提出对目前最流行的Cloude-Pottier目标分解方法的改进,将回波功率参数引入到Cloude-Pottier分解中,并给出了Cloude-Pottier分解结果的直观表达方式。就非监督分类算法而言,本文首先提出了基于复Wishart分布和最大似然估计算法的Wishart SPAN/H/α分类算法,通过引入回波功率参数在一定程度上避免分类初始化的错误,提高分类器的性能;然后,考虑到Wishart SPAN/H/α分类算法初始化参数所表征的散射机制的信息不完全,本文进一步提出了采用SPAN/H/α/A四个特征参数进行初始化的Wishart SPAN/H/α/A分类算法作为Wishart SPAN/H/α分类算法的改进;由于Wishart SPAN/H/α/A分类算法初始化后的类数较多,本文又进一步提出了基于Wishart检验统计的区域合并算法来减少分类的类别数,获得有效的分类结果。Wishart SPAN/H/α/A分类算法使用预先给定的类别数进行区域合并,这一类别数并不一定是最优的。事实上,目前全极化SAR研究提出的分类算法均采用固定的类别数,关于全极化SAR数据类别数估计的研究还没有开展。本文通过引入模式识别中的交叉验证算法,提出了基于Wishart分布的交叉验证对数似然的概念,用该似然函数进行全极化SAR数据的类别数估计,根据全极化SAR数据的内部结构直接获得最优的类别数。基于该似然函数,本文又提出了一种新的全极化SAR数据非监督分类算法,带类别数估计的Wishart SPAN/H/α/A分类算法,该算法自动地根据全极化数据估计最优的类别数,并采用该最优类别数作为分类结果的类别数,消除分类类别数与数据内部结构不匹配的情况下可能导致的过拟合或欠拟合现象。另外,对于分类器定量评估方面,本文提出了采用交叉验证算法进行全极化SAR数据非监督分类器评估的方法,并尝试利用交叉验证算法实现全极化SAR数据非监督分类结果的评估。

【Abstract】 The land cover classification is one of the most important, applications inpolarimetry remote sensing. The main advantage of fully polarimetric SAR datais that it can use target decomposition algorithm to extract the informationof the scattering mechanisms, which are not data specific and can be used fortarget identification. Thus using fully polarimetric SAR data we can achieve anunsupervised classification without ground truth information.In this paper, an improvement for the Cloude-Pottier decomposition is givenfor analysis. We use the backscattering power information to improve the per-formance of the Cloude-Pottier decomposition and the transform algorithm isalso given to represent directly the decomposition results. Several unsupervisedclassification algorithms are also proposed to improve the classification perfor-mance step by step. Firstly, the Wishart SPAN/H/αclassification is proposed tointroduce the backscattering power information to the Wishart H/α/A classifi-cation. Then in order to include the scattering information within the parameterA, the Wishart SPAN/H/α/A classification is given to use the four parametersSPAN/H/α/A for initialization, and the Wishart test statistic is applied to re-duce the number of classes. The Wishart SPAN/H/α/A classification uses apredefined number of classes to perform the classification scheme. In fact, all theunsupervised classifications for fully polarimetric SAR data proposed nowadaysuse fixed number of clusters to classify the data. It is more reasonable to deter-mine the number of clusters directly from data analysis. According to the CrossValidation theorem in pattern recognition, We proposed the Cross-Validationlog-likelihood based on the complex Wishart distribution to estimate the opti-mal number of classes to reveal the inner structure of fully polarimetric SARdata. Using the Cross-Validation log-likelihood, a new unsupervised classifica-tion method, the Wishart SPAN/H/α/A classification with an adaptive numberof classes, is also given for interpretation. The number of classes is automati-cally optimized to avoid overfitting and underfitting for the inner structure of fully polarimetric SAR data. Moreover, since the Cross Validation algorithm is aquantitative estimation of the classification performance, it also has the potentialcapability to perform the validation of unsupervised classification.

  • 【分类号】TP75;TN958
  • 【被引频次】11
  • 【下载频次】920
  • 攻读期成果
节点文献中: 

本文链接的文献网络图示:

本文的引文网络