MCJ-UNet:一种双/多通道联合InSAR相位解缠网络

MCJ-UNet:A Dual/Multi-channel-joint Phase Unwrapping Network for Interferometric SAR

干涉合成孔径雷达(InSAR)可实现地表高程的高效获取,在地形测绘中应用广泛。双/多通道InSAR技术可借助不同通道(基线、频点)的高程模糊度差异,解决相位欠采样问题,完成高程陡变区域的干涉相位解缠,实现InSAR技术在测绘困难区域的有效应用。该文即面向高效高精度相位解缠需求,利用深度学习这一有力工具,结合不同通道的相位特征及相互约束关系,提出了一种双/多通道联合干涉相位解缠网络:Multi-Channel-Joint-UNet(MCJ-UNet)。该网络的构建以双通道(双频、双基线)InSAR为基本观测构型,并可实现向多通道构型的扩展,其构建的核心思路主要包括3点:首先,将干涉相位解缠中的模糊数估计问题转化为语义分割问题,并采用UNet网络完成分割处理;其次,引入挤压激励模块(SE)动态调整信息权重,以增强网络不同通道对其所需信息的感知能力;最后,利用多通道联合约束下的相位残差优化损失函数,实现网络调谐。此外,为避免语义分割结果的边缘细节误差对解缠效果的影响,该文还提出了一种基于多通道联合约束的解缠误差自修正方法,以保证解缠质量。模拟地形仿真数据、真实地形仿真数据以及TerraSAR-X实测数据验证了所提方法的有效性。

Interferometric Synthetic Aperture Radar(InSAR)enables the efficient retrieval of surface elevation and has extensive applications in terrain mapping.Dual/multi-channel InSAR techniques utilize the differences in the elevation ambiguity of different InSAR channels(i.e.,baselines and frequencies)to perform Phase Unwrapping(PU).This enables the effective application of InSAR in regions with abrupt terrain changes.In response to the growing demand for efficient and precise PU,this study leverages deep learning and proposes a dual/multi-channel joint PU network,i.e.,Multi-Channel-Joint-UNet(MCJ-UNet),which effectively combines multi-channel phase characteristics and their mutual constraint relationships.The proposed network is constructed based on the dual-channel(i.e.,dual-frequency and dual-baseline)InSAR observation configuration.It can also be extended to multi-channel InSAR.The core concept of the proposed method can be summarized as follows.First,the method transforms the elevation ambiguity estimation problem in PU into semantic segmentation,and the UNet network is employed to accomplish the segmentation processing.Second,the squeeze-and-excitation module is introduced to dynamically adjust the information weights,enhancing the network’s perception of the required information across different channels.Third,a phase residual optimization loss function is employed in the context of multi-channel joint constraints to achieve network tuning.In addition,to mitigate the effect of edge detail errors in semantic segmentation results on PU performance,a selfcorrecting approach for PU errors based on multi-channel joint constraints is proposed.The proposed MCJUNet is verified by computer simulations based on simulated and real terrains and experiments based on real TerraSAR-X data.