AIR-PolSAR-Seg-2.0:大规模复杂场景极化SAR地物分类数据集  

AIR-PolSAR-Seg-2.0:Polarimetric SAR Ground Terrain Classification Dataset for Large-scale Complex Scenes

作　　者：王智睿 赵良瑾 汪越雷 曾璇 康健 杨健[6] 孙显 WANG Zhirui;ZHAO Liangjin;WANG Yuelei;ZENG Xuan;KANG Jian;YANG Jian;SUN Xian(Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100094,China;University of Chinese Academy of Sciences,Beijing 100049,China;School of Electronic,Electrical and Communication Engineering,University of Chinese Academy of Sciences,Beijing 100049,China;Key Laboratory of Target Cognition and Application Technology(TCAT),Beijing 100190,China;School of Electronic and Information Engineering,Soochow University,Suzhou 215006,China;Department of Electronic Engineering,Tsinghua University,Beijing 100084,China)

机构地区：[1]中国科学院空天信息创新研究院,北京100094 [2]中国科学院大学,北京100049 [3]中国科学院大学电子电气与通信工程学院,北京100049 [4]目标认知与应用技术国家级重点实验室,北京100190 [5]苏州大学电子信息学院,苏州215006 [6]清华大学电子工程系,北京100084

出　　处：《雷达学报(中英文)》2025年第2期353-365,共13页Journal of Radars

基　　金：国家自然科学基金(62331027)。

摘　　要：极化合成孔径雷达(PolSAR)地物分类是SAR图像智能解译领域的研究热点之一。为了进一步促进该领域研究的发展,该文组织并发布了一个面向大规模复杂场景的极化SAR地物分类数据集AIR-PolSAR-Seg-2.0。该数据集由三景不同区域的高分三号卫星L1A级复数SAR影像构成,空间分辨率8 m,包含HH,HV,VH和VV共4种极化方式,涵盖水体、植被、裸地、建筑、道路、山脉等6类典型的地物类别,具有场景复杂规模大、强弱散射多样、边界分布不规则、类别尺度多样、样本分布不均衡的特点。为方便试验验证,该文将三景完整的SAR影像裁剪成24,672张512像素×512像素的切片,并使用一系列通用的深度学习方法进行了实验验证。实验结果显示,基于双通道自注意力方法的DANet性能表现最佳,在幅度数据和幅相融合数据的平均交并比分别达到了85.96%和87.03%。该数据集与实验指标基准有助于其他学者进一步展开极化SAR地物分类相关研究。The ground terrain classification using Polarimetric Synthetic Aperture Radar(PolSAR)is one of the research hotspots in the field of intelligent interpretation of SAR images.To further promote the development of research in this field,this paper organizes and releases a polarimetric SAR ground terrain classification dataset named AIR-PolSAR-Seg-2.0 for large-scale complex scenes.This dataset is composed of three L1A-level complex SAR images of the Gaofen-3 satellite from different regions,with a spatial resolution of 8 meters.It includes four polarization modes:HH,HV,VH,VV,and covers six typical ground terrain categories such as water bodies,vegetation,bare land,buildings,roads,and mountains.It has the characteristics of large-scale complex scenes,diverse strong and weak scattering,irregular boundary distribution,diverse category scales,and unbalanced sample distribution.To facilitate experimental verification,this paper cuts the three complete SAR images into 24,672 slices of 512×512 pixels,and conducts experimental verification using a series of common deep learning methods.The experimental results show that the DANet based on the dual-channel self-attention method performs the best,with the average intersection over union ratio reaching 85.96% for amplitude data and 87.03% for amplitude-phase fusion data.This dataset and the experimental index benchmark are helpful for other scholars to further carry out research related to polarimetric SAR ground terrain classification.

关 键 词：极化合成孔径雷达 公开数据集 复数图像 地物分类 深度学习 

分 类 号：TN957[电子电信—信号与信息处理]