结合特征选择和大尺度谱聚类的极化SAR图像非监督分类

发布时间：2018-02-17 07:37

本文关键词： 极化合成孔径雷达特征选择非监督分类谱聚类　出处：《信号处理》2016年06期 　论文类型：期刊论文

【摘要】：非监督分类是极化SAR图像解译的重要手段,但其分类结果易受到高维特征的影响。针对此问题,本文提出一种结合特征选择和大尺度谱聚类的极化SAR图像非监督分类方法。该方法首先深入分析并提取了极化SAR图像分类中常用的特征参数,包括基于测量数据及其简单线性变换的特征和极化目标分解的特征。然后通过聚类森林特征选择算法进行特征降维处理,去除冗余信息。最后利用过分割产生代表点并构建原始数据与代表点间的二分图,通过大尺度谱聚类算法完成图像的非监督分类。实验结果表明,该方法能够选取有效的特征组合,并得到较为满意的分类效果。
[Abstract]:Unsupervised classification is an important means of polarimetric SAR image interpretation, but its classification results are easily affected by high-dimensional features. In this paper, an unsupervised classification method for polarimetric SAR images based on feature selection and large-scale spectral clustering is proposed. Firstly, the commonly used feature parameters in polarimetric SAR image classification are analyzed and extracted. It includes the features based on the measurement data and its simple linear transformation, and the feature of polarization target decomposition. Then, the feature reduction dimension is processed by clustering forest feature selection algorithm. The redundant information is removed. Finally, the representative points are generated by over-segmentation and the binary graph between the original data and the representative points is constructed, and the unsupervised classification of the images is accomplished by large-scale spectral clustering algorithm. The experimental results show that, The method can select effective feature combinations and obtain satisfactory classification results.
【作者单位】：武汉大学电子信息学院;上海卫星工程研究所;内蒙古工业大学雷达技术研究所;
【基金】：国家自然科学基金(61271401,41501414) 复旦大学电磁波信息科学教育部重点实验室开放基金项目(EMW201504)
【分类号】：TN957.52

【相似文献】