结合深度学习的多通道InSAR高程反演方法

Multi-channel InSAR Elevation Reconstruction Method Combining Deep Learning

针对干涉合成孔径雷达探测技术中的高程反演问题,提出一种结合深度学习的多通道InSAR高程反演方法。该方法通过构建一种多通道特征连接与融合网络(MCFCF Net)来建立多通道干涉图和高程图的直接映射关系。首先,网络采用融合了SE注意力模块的残差单元,在保护干涉图信息完整性的同时提升了对重要通道的关注力;其次,用密集连接的方式实现多通道特征复用,加强了特征图的传播,提高了网络对多幅干涉图信息的融合能力;最后,MCFCF Net为对称结构,有利于获取多通道干涉图间的关联信息并准确还原高程。不同类型地形的多通道InSAR高程反演实验表明了该方法有效性和稳健性。

This paper presents a multi-channel InSAR elevation reconstruction method based on deep learning to solve the problem of elevation inversion of interferometric synthetic aperture radar technology,where a multi-channel feature connected and fusion net,named as MCFCF Net,is built to establish direct mapping relationship between multi-channel interferograms and elevation map.Firstly,the residual unit integrated with the SE attention module is used to improve the attention to important channels while protecting the integrity of phase information from interferograms.Then,the propagation of feature maps is enhanced by means of dense connection,which improves the network’s ability to fuse the information of multiple interferograms.Finally,the MCFCF Net is a symmetric structure,which is conducive to obtaining the correlation information between multi-channel interferograms and accurately reconstructing the elevation maps.Elevation reconstruction experiments for different types of terrain demonstrate the effectiveness and robustness of the proposed method.