顾及特征优化的极化SAR图像半监督分类  

A Semi-supervised Classification Method for Fully Polarimetric SAR Imagery Considering Feature Optimization

作　　者：程圆娥 袁春琦 吕志慧[1] 赫春晓[1] CHENG Yuan′e;YUAN Chunqi;LYU Zhihui;HE Chunxiao(Jiangsu Province Surveying&Mapping Research Institute,Nanjing 210013,China;North Information Control Group Co.,Ltd.,Nanjing 210013,China)

机构地区：[1]江苏省测绘研究所,江苏南京210013 [2]北方信息控制研究院集团有限公司,江苏南京211153

出　　处：《测绘与空间地理信息》2022年第8期49-53,共5页Geomatics & Spatial Information Technology

基　　金：江苏省科技厅项目——时空大数据整合挖掘研究(BM2018030)资助。

摘　　要：针对全极化SAR图像在监督分类中存在的人工标注样本费时费力以及大量未标记样本未有效利用等问题,提出一种顾及特征优化的半监督自训练分类方法。该方法以支持向量机作为半监督学习的基分类器。首先,引入混合编码遗传算法对多类极化特征进行有效选择和分类器参数优化。在此基础上,利用核模糊聚类算法从大量未标注样本中预选取出信息量较大的样本点作为候选点,并借助半监督自训练学习进一步标注候选样本,实现同时利用无标注和有标注样本进行分类。在AIRSAR和ALOS PALSAR影像上的实验表明,该方法能自适应地优选分类特征集,且在较少人工标注的样本下能获得较高的分类精度。Supervised classification methods usually require adequate labeled samples which are difficult and time-consuming to obtain,especially in synthetic aperture radar images.In order to make full use of the unlabeled samples,an improved self-trained semi-supervised classification method which considering feature optimization is proposed in this paper.The proposed semi-supervised classification method is based on support vector machines(SVM)approach.Firstly,a hybrid genetic algorithm is used to select the effective features and optimize the classifier parameters.Then,more reliable candidates are selected from the large pool of unlabeled samples by using Kernel Fuzzy Clustering algorithm.Finally,both the label and unlabeled samples are utilized for classifying via a self-trained semi-supervised learning model.Experiments on AIRSAR and ALOS PALSAR images show that the proposed method can adaptively preferably classify feature sets and achieve high classification accuracy with less manually annotated samples.

关 键 词：极化合成孔径雷达 图像分类 半监督学习 特征优化 

分 类 号：P209[天文地球—测绘科学与技术]