【作者】 韦立登； 项徳良； 张博栋； 倪军； 尹嫱； 张帆；

【Author】 Lideng Wei;Deliang Xiang;Bodong Zhang;Jun Ni;Fan Zhang;Qiang Yin;Beijing Institute of Radio Measurement;Artificial Intelligence Cross Research Center, Beijing University of Chemical Technology;College of Information Science and Technology, Beijing University of Chemical Technology;

【机构】 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； 北京化工大学信息科学与技术学院；

【摘要】 机载毫米波全极化合成孔径雷达(PolSAR)具备不受光照限制、地物散射信息丰富、高分辨率等特点,随着其技术的不断发展和完善,已经成为重要的对地观测手段之一,而无监督分类是毫米波PolSAR数据的重要研究方向。针对于ka波段、分辨率为0.3米的未定标PolSAR影像,本文利用新的超像素分割算法与Wishart-H/A/α分类算法结合,实现毫米波全极化SAR的无监督分类过程。首先使用了线性压缩方法对数据进行压缩,使得未定标数据能够计算任务;然后,利用Wishart-H/A/α无监督分类算法得到每个像素的类别属性;同时,使用自适应极化超像素生成算法(Pol-ASLIC)实现超像素分割任务,以兼顾全极化SAR数据的空间统计特性;最后,利用超像素得到的空间信息与无监督像素标签信息融合实现最终的分类任务。本实验所用的PolSAR影像共包含10个地物类别,基于H/α和H/A/α提出的无监督分类方法分别实现了93.76%和93.09%的分类精度。更多还原

【Abstract】 Airborne millimeter-wave fully polarized synthetic aperture radar(PolSAR) is free from light restriction, has abundant scattering information and high resolution, etc. With the continuous development and improvement of its technology, it is an indispensable means of earth observation, and unsupervised classification is an important research direction of millimeterwave PolSAR data. For the Ka-band uncalibrated PolSAR images with a resolution of 0.3 m, this paper combines the new superpixel segmentation method with Wishart-H/A/α classification algorithm to realize the unsupervised classification. First, the linear compression method is used to compress the data so that the uncalibrated data can be calculated. Then, the Wishart-H/A/α method is implemented to obtain the category of each pixel. At the same time, the adaptive polarimetric superpixel generation algorithm(Pol-ASLIC) method is used to achieve superpixel segmentation so as to considering the spatial characteristics. Finally, the superpixel and unsupervised label are fused to obtain the final results. The PolSAR image contains 10 types, and the unsupervised classification method based on H/A/α and H/α achieves classification accuracy of 93.76% and 93.09%, respectively.更多还原