综合多特征的极化SAR图像建筑物U-Net分类方法  

U-Net classification of buildings based on multi-feature synthesis in polarimetric SAR images

作　　者：李梅 沈麒凯 陈启浩[1] 刘修国[1] LI Mei;SHEN Qikai;CHEN Qihao;LIU Xiuguo(College of Geography and Information Engineering,China University of Geosciences,Wuhan 430070,China)

机构地区：[1]中国地质大学(武汉)地理与信息工程学院,武汉430070

出　　处：《测绘科学》2022年第9期146-153,162,共9页Science of Surveying and Mapping

基　　金：国家自然科学基金项目(41771467)

摘　　要：针对合成孔径雷达(SAR)图像不同类型建筑物的区分问题,该文提出了一种基于U-Net的极化SAR图像建筑物分类方法。该方法将极化SAR数据的Pauli分解参数、规范化圆极化相关系数和G~0统计纹理参数作为U-Net的输入,建立建筑物分类U-Net模型,同时考虑建筑物的高度和单体面积的情况下,将建筑物分为高层、中层、低层小面积、低层厂房类大面积建筑物4类。对武汉市城区GF-3极化SAR数据的各类建筑物分类精度均在80%以上,最高达94.2%。该方法与仅使用单类别特征的U-Net网络方法以及卷积神经网络方法相比,分类结果更完整、建筑边界更准确,也更适合于复杂中心城区的不同类型建筑物的分类。In order to distinguish different types of buildings in synthetic aperture radar(SAR)images,a classification method based on U-Net was proposed in this paper.This method combined the Pauli decomposition parameters,the normalized circular polarization correlation coefficient and the G~0 statistical texture parameters as the input of U-Net to establish building classification U-Net model.Then considering the height of the building and the single area of the building at the same time,the building was divided into four types:high-rise,middle-rise,low-rise small area,low-rise factory-like and large-area buildings.The classfiction accuracy of four types buildings based on GF-3 polarimetric SAR data in the urban area of Wuhan was over 80%,and the highest was 94.2%.Compared with the U-Net network method only using one class feature and the convolutional neural network method,this method had more complete classification results,more accurate building boundaries,and was more suitable for classification of different types of buildings in complex central urban areas.

关 键 词：极化SAR U-Net网络 分类 建筑物 深度学习 

分 类 号：P237[天文地球—摄影测量与遥感] TN957.52[天文地球—测绘科学与技术]