A New Approach for Interferometric Phase Reconstruction over Decorrelated Regions with Deep Learning

Interferometric Synthetic Aperture Radar (InSAR), a powerful geodetic technique, can accurately provide ground deformation measurements with high spatial resolution and wide coverage. However, InSAR suffers from decorrelation due to heavy vegetation or large ground movements, especially when exploring areas with geohazard events. Ignoring or masking these incoherent areas can stabilize InSAR phase unwrapping but at the cost of losing valuable spatial information, such as the deformation near the epicenter of an earthquake. To retrieve information over the masked areas, a generative interferometric phase reconstruction (IPR) technique has been developed to handle various fringe patterns and dynamically changing decorrelated regions. In this paper, we proposed a new generative adversarial network (GAN) called PhaseNet, designed for IPR under different fringe rates and masked patterns. The model was trained using a simulated InSAR dataset under a pixel-space loss function. The structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) were used as reconstruction metrics to evaluate model performance. The model achieved an SSIM of 0.98 for the original simulated interferogram despite 15% of the pixels being masked. The largest disparity between the clear and reconstructed phases in the synthetic validation dataset remained within the range of ±1 radian in phase or sub-centimeter in deformation.