月球可见光和SAR影像特征匹配方法比较研究  

A comparative analysis of feature matching techniques for lunar optical and SAR imagery

作　　者：王晨旭 彭嫚 谢彬 邸凯昌 芶盛 WANG Chenxu;PENG Man;XIE Bin;DI Kaichang;GOU Sheng(Key Laboratory of Remote Sensing and Digital Earth,Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100101,China;University of Chinese Academy of Sciences,Beijing 100049,China;Key Laboratory of Earth and Planetary Physics,Institute of Geology and Geophysics,Chinese Academy of Sciences,Beijing 100029,China)

机构地区：[1]中国科学院空天信息创新研究院遥感与数字地球重点实验室,北京100101 [2]中国科学院大学,北京100049 [3]中国科学院地质与地球物理研究所地球与行星物理重点实验室,北京100029

出　　处：《遥感学报》2025年第2期389-401,共13页NATIONAL REMOTE SENSING BULLETIN

基　　金：国家重点研发计划(编号:2022YFF0503100);国家自然科学基金(编号:41771488)。

摘　　要：多模态影像匹配方法已在地球多源遥感影像中获得了广泛应用,但月球多模态影像的匹配尚缺少对比性研究。为了实现高分辨率月球光学图像与SAR图像的高精度匹配,本文选择月球中纬度、低纬度、南极、北极等多个实验区影像,使用SIFT (Scale-Invariant Feature Transform)、基于区域的CFOG (Channel Features of Orientated Gradients)、HOPC (Histogram of Orientated Phase Congruency)和基于结构特征的RIFT (Radiation-variation Insensitive Feature Transform)、 HAPCG (Histogram of Absolute Phase Consistency Gradients)、 HOWP(Histogram of the Orientation of the Weighted Phase descriptor)和深度学习SuperGlue、 LoFTR (Local Feature TRansformer)共8种特征匹配算法进行实验比较研究,通过正确匹配点数、均方根误差、重复率和覆盖度4种指标对匹配结果进行比较分析。结果表明,HAPCG算法使用了各向异性滤波并结合绝对相位方向梯度直方图构成特征向量,匹配效果最优。LoFTR算子使用了自注意层和互注意层机制,对纹理贫乏的月球影像效果次之。HOWP和SuperGlue匹配效果居中。CFOG、HOPC和RIFT效果最差。SIFT未能实现匹配。匹配点的分布和成像光照条件、影像重叠区域相关,中低纬度地区匹配效果优于南北极地区。对HAPCG匹配结果的Stokes第一参数进行了统计分析,雨海和高地实验区的匹配点的散射特性参数的平均值高于南极北极实验区,和地形特征相符。散点图显示出HAPCG匹配点对应的Stokes第一参数和光学影像灰度值存在相关性,证明了HAPCG对非线性辐射差异较大的月球光学影像和SAR影像匹配的稳健性。本研究为月球光学影像和SAR影像匹配方法的选择提供参考,有助于月球多源数据的应用。Multimodal image matching methods have been widely applied in the registration of multisource remote sensing images of the Earth,but comparative research on the application of multimodal registration of lunar images is lacking.To facilitate high-precision alignment between high-resolution lunar optical imagery and Synthetic Aperture Radar(SAR)imagery,this paper conducts an experimental comparison across various lunar regions,including mid latitude,low latitude,the Antarctic,and the Arctic,using a suite of eight algorithms:Scale-Invariant Feature Transform(SIFT),region-based Channel Features of Orientated Gradients(CFOG),Histogram of Orientated Phase Congruency(HOPC),the structural feature-based Radiation-variation Insensitive Feature Transform(RIFT),Histogram of Absolute Phase Consistency Gradients(HAPCG),and Histogram of the Orientation of the Weighted Phase descriptor(HOWP),along with the deep learning models SuperGlue and Local Feature Transformer(LoFTR).The performance of these algorithms is evaluated through four metrics:the number of correct matches,root mean square error,repeatability,and coverage.The findings reveal that the HAPCG algorithm,which integrates anisotropic filtering with a composite feature vector,outperforms the others in terms of matching quality.The LoFTR algorithm leverages self-attention and cross-attention mechanisms,and demonstrates robust performance,particularly for lunar imagery with sparse textures.The HOWP and SuperGlue algorithms exhibit midrange performance in terms of matching efficacy.By contrast,the CFOG,HOPC,and RIFT algorithms yield the least satisfactory results,and the SIFT algorithm fails to establish any matches.The distribution of matched points is influenced by factors such as imaging illumination conditions and the extent of the overlapping regions,and matches in mid-and low-latitude areas are more successful than those in polar regions.A statistical analysis of the Stokes parameter for the HAPCG matches indicates that the mean values of the scattering characteristic

关 键 词：月球 多模态匹配 SAR影像 光学影像 Stokes参数 

分 类 号：P2[天文地球—测绘科学与技术]