A Novel Model for Interferometric Phase Reconstruction Based on Multi-Stage Conditional GANs

Abstract. Reconstructing the interferometric phase in decorrelated regions is a significant challenge in interferometric synthetic aperture radar (InSAR) techniques, as decorrelation disrupts the continuity of phase fringes and obscures critical information. This paper presents a novel two-stage generative adversarial network (GAN) framework to address this issue. In the first stage, GAN is designed to reconnect fragmented phase fringes. In the second stage, GAN focuses on reconstructing the phase in masked regions guided by the reconnected fringes achieved from the first stage. The proposed model was trained on a simulated topographic phase with the SRTAM product. The proposed model achieves a structural similarity index (SSIM) of 0.9 and a peak signal-to-noise ratio (PSNR) of 30.4. Then, we conducted a quantitative evaluation with a real interferogram from the Greater Bay Area (GBA). The experiment results demonstrated the generalization capabilities of the proposal model, with an average correlation of 0.8 between the predicted and actual phases. The proposed approach can effectively preserve phase continuity, reconstruct masked areas, and mitigate the impact of decorrelation. It shows potential for use in topographic retrieval and ground deformation monitoring in InSAR applications.