FPUNet: A Fourier-Enhanced U-Net for Robust 2-D Phase Unwrapping of Noisy InSAR Interferograms

Two-dimensional phase unwrapping (PU) of interferometric synthetic aperture radar (InSAR) data remains difficult when steep deformation gradients and multi-source disturbances violate the Itoh condition. This study proposes FPUNet, a Fourier-enhanced encoder–decoder for joint denoising and 2-D PU, in which frequency-domain global context modeling is combined with complementary multi-scale spatial aggregation and attention-based feature refinement. Specifically, the bottleneck cascades a Fourier Mixed Residual Block (FMRB), atrous spatial pyramid pooling (ASPP), and a convolutional block attention module (CBAM) to suppress noise while preserving deformation-related fringe structures. FPUNet is trained end-to-end on realistically simulated Sentinel-1 interferograms generated from Shuttle Radar Topography Mission (SRTM) digital elevation models using a physics-informed composite loss that enforces data fidelity, gradient consistency, spectral regularization, and selective rewrapping consistency. On a synthetic benchmark of 1800 test interferograms, FPUNet achieves an RMSE of 0.79 rad, improving over a plain U-Net (1.61 rad) and producing fewer large fringe-number errors than least-squares, SNAPHU, PUNet, and DLPU. Experiments on real Sentinel-1 data over the Datong mining area and the 2022 Menyuan and Luding earthquakes further indicate improved phase closure and rewrapping consistency, particularly in high-gradient coseismic fringes, supporting FPUNet as a robust PU module for InSAR deformation monitoring.