Unwrap-Net: A deep neural network-based InSAR phase unwrapping method assisted by airborne LiDAR data

Abstract

In Interferometric Synthetic Aperture Radar (InSAR) data processing, accurately unwrapping the phase is crucial for measuring elevation or deformation. DCNN models such as PhaseNet and PGNet have improved the efficiency and accuracy of phase unwrapping, but they still face challenges such as incomplete multiscale feature learning, high feature redundancy, and reliance on unrealistic datasets. These limitations compromise their effectiveness in areas with low coherence and high gradient deformation. This study proposed Unwrap-Net, a novel network model featuring an encoder-decoder structure and enhanced multiscale feature learning via ASPP (Atrous Spatial Pyramid Pooling). Unwrap-Net minimizes feature redundancy and boosts learning efficiency using SERB (Residual Convolutional with SE-block). For dataset construction, airborne LiDAR terrain data combined with land cover data from optical images, using the SGS (Sequential Gaussian Simulation) method, are used to synthesize phase data and simulate decorrelation noise. This approach creates a dataset that closely approximates real-world conditions. Additionally, the introduction of a new high-fidelity optimization loss function significantly enhances the model’s resistance to noise. Experimental results show that compared to the SNAPHU and PhaseNet models, the SSIM of the Unwrap-Net model improves by over 13%, and the RMSE is reduced by more than 34% in simulated data experiments. In real data experiments, SSIM improves by over 6%, and RMSE is reduced by more than 49%. This indicates that the unwrapping results of the Unwrap-Net model are more reliable and have stronger generalization capabilities. The related experimental code and dataset will be made available at https://github.com/yangwangyangzi48/UNWRAPNETV1.git.