【作者】 倪军； 韦立登； 张博栋； 尹嫱； 张帆； 项徳良；

【Author】 NI Jun;WEI Lideng;ZHANG Bodong;YIN Qiang;ZHANG Fan;XIANG Deliang;College of Information Science and Technology,BUCT;Beijing Institute of Radio Measurement;Artificial Intelligence Cross Research Center,BUCT;

【机构】 北京化工大学信息科学与技术学院； 北京无线电测量研究所； 北京化工大学人工智能交叉研究中心；

【摘要】 机载毫米波全极化合成孔径雷达(PolSAR)不受光照条件限制、地物散射信息丰富、分辨率高,逐渐成为重要的对地观测手段之一,而全极化SAR的图像分类算法是其图像解译的一个重要研究方向。以针对于Ka波段、分辨率为0.3米的未定标毫米波PolSAR影像基础,本文分别进行了无监督和监督分类算法研究。首先,无监督分类算法能够在无先验标签的情况下实现全极化SAR影像的分类和标注,本文以变分贝叶斯的Wishart混合模型(VBWMM)为基础,实现了毫米波全极化SAR的无监督分类过程。其次,监督分类算法能够在先验标签的基础上实现地物类别的高精度分类过程,本文以卷积神经网络(CNN)为基础,实现了全极化SAR影像的监督分类过程。同时,为了兼顾PolSAR影像的空间信息,本文提出了条件随机场(CRF)与超像素联合约束的空间分析算法,将无监督和监督分类的结果分别进行空间分析,以兼顾全极化SAR数据的空间统计特性,最终实现毫米波全极化SAR影像的无监督和监督分类任务。更多还原

【Abstract】 Airborne millimeter-wave fully polarimetric synthetic aperture radar(PolSAR) is not limited by light conditions,has abundant ground object scattering information and high resolution. It gradually becomes one of the important means of earth observation. The classification of PolSAR imagery is an important field in its image interpretation. Based on the Ka-band uncalibrated millimeter-wave PolSAR image with a resolution of 0.3m, unsupervised and supervised classification method are studied in this paper. Firstly, unsupervised classification method can realize the annotation of the PolSAR image without prior labels. Based on variational Bayesian Wishart mixture model(VBWMM), this paper realizes the unsupervised classification of the millimeter-wave PolSAR image. Secondly, the supervised classification method can realize the high-accuracy classification based on prior labels. In this paper, the convolutional neural network(CNN) is implemented to the mainly structure of supervised classification. Meanwhile, based on the unsupervised and supervised classification results, the conditional random field(CRF) restrained by superpixel segmentation is used for fusing spatial information of the PolSAR imagery, thus taking into account the whole polarimetric spatial statistical properties. Finally,the smooth unsupervised and supervised classification results will be obtained.更多还原