摘要
由于单源数据的局限性,利用多源遥感数据进行对地观测联合分类是一种很有前景但又充满挑战的方法。但是多源数据之间存在成像机制的差距和信息的不平衡等问题,现有的方法在面对单源数据的特征提取和多源数据之间的特征融合时仍然存在一定的不足。提出一种基于多尺度异构特征提取的高光谱图像和合成孔径雷达图像分类方法,命名为多尺度异构跨模态注意力网络。具体来说,设计了多尺度异构特征提取模块提取高光谱图像和合成孔径雷达图像的空间-光谱联合特征,模块内部使用深度可分离卷积降低参数量,采用残差连接以增强不同层次特征的融合效果。此外,设计了跨模态融合注意力模块进一步挖掘高光谱图像的通道特征和合成孔径雷达图像的空间特征,通过计算跨模态特征在每个位置的映射关系,实现对互补特征的有效对齐和融合。在两个不同数据集上进行实验,相较于其他方法,多尺度异构跨模态注意力网络在Augsburg和Berlin数据集上的OA分别取得了至少2.91%和6.81%的提升,展现出了优秀的分类性能。
Due to the limitation of single-source data,the use of multi-source remote sensing data for joint classification of earth observations is a promising but challenging method.However,there are some problems such as the gap of imaging mechanism and the imbalance of information between multi-source data,so that the existing methods still have some shortcomings in feature extraction of single-source data and feature fusion of multi-source data.In this paper,we propose a hyperspectral image and synthetic aperture radar image classification method based on multi-scale heterogeneous feature extraction,named multi-scale heterogeneous cross-modal attention network.Specifically,a multi-scale heterogeneous feature extraction module is designed to extract the spatial-spectral joint features of hyperspectral image and synthetic aperture radar images.Depthwise separable convolution is used in the module to reduce the number of parameters,and residual connection is used to enhance the fusion effect of features at different levels.In addition,a cross-modal fusion attention module is designed to further explore the channel features of hyperspectral image and the spatial features of synthetic aperture radar images.To achieve effective alignment and fusion of complementary features,the mapping relationship of cross-modal features at each position is considered.Compared with other methods,MSHCNet achieves at least 2.91%and 6.81%improvement in OA on Augsburg and Berlin datasets,respectively,demonstrating excellent classification performance.
作者
刘怡
李渊
孙搏遥
郑宇航
叶珍
LIU Yi;LI Yuan;SUN Boyao;ZHENG Yuhang;YE Zhen(School of Electronics and Control Engineering,Chang’an University,Xi’an 710018,China)
出处
《空间电子技术》
2024年第6期57-65,共9页
Space Electronic Technology
基金
“慧眼行动”项目(编号:GYH2023033)。
关键词
多源遥感
高光谱图像
合成孔径雷达
分类
multi-source remote sensing
hyperspectral image
synthetic aperture radar
classification
作者简介
刘怡(2000—),甘肃兰州人,硕士研究生,主要研究方向为高光谱图像分析和多源遥感。E-mail:2022232009@chd.edu.cn;通讯作者:叶珍(1983—),陕西西安人,博士,副教授,主要研究方向为多源遥感图像处理、机器学习等。E-mail:yezhen525@126.com。
