DMP-PUNet: A novel network for two-dimensional InSAR phase unwrapping under severe noise and complex fringes conditions

Abstract

In the processing of Interferometric synthetic aperture radar (InSAR) data, two-dimensional (2-D) phase unwrapping (PU) is crucial for ensuring the quality of InSAR data inversion. Traditional methods, based on the assumption of phase continuity, often struggle with abrupt terrain changes and the influence of severe noise, leading to poor performance or failure. To address these challenges, this paper presents a dilated multi-path phase unwrapping network (DMP-PUNet) for 2-D PU under conditions of severe noise and complex fringes. To train this model, we developed a multi-effect interferometric phase simulation (ME-IPS) strategy that aims to simulate interferometric phases that closely resemble real-world conditions by comprehensively considering various factors, including terrain and digital elevation model (DEM) errors, atmospheric turbulence, vegetation effects, baseline geometry, multiple scattering, and noise. This simulation, combined with quasi-real interferometric phase data obtained from DEM inversion algorithms, forms the comprehensive training dataset. Finally, experiments on simulated data, quasi-real data, the InSAR-DLPU dataset, and InSAR data demonstrate that DMP-PUNet outperforms existing methods. For simulated data, DMP-PUNet achieved an overall average mean absolute error (MAE) in residuals of 0.221 rad, improving accuracy by 54.75 % with an average processing time of 0.81 s. For quasi-real data, the average MAE was 0.320 rad, a 119.06 % increase in accuracy, with an average processing time of 0.82 s. For the InSAR-DLPU dataset, overall, the MAE of DMP-PUNet was 20.34 % to 64.96 % lower than that of the best-performing baseline method (DLPU), with an average processing time of 1.90 s. For InSAR data, DMP-PUNet performed stably, with lower noise levels, smooth phase transitions, and deformation spatial patterns and profile shapes that conform to the laws of mining subsidence, averaging a processing time of 1.71 s, outperforming existing methods.