Abstract: [Purposes]To address the low unwrapping accuracy in regions with complex noise and phase discontinuities, as well as the difficulty of traditional phase unwrapping algorithms in balancing precision, robustness and computational efficiency,the study proposes a deep learning method called quality-guided residual U-Net (QG-ResUNet).[Methods]This approach innovatively incorporates pixel-level phase quality maps as prior information through an attention mechanism into a residual-connected U-Net architecture,guiding the network to focus on high-reliability phase regions while intelligently handling low-quality areas.[Findings]Using normalized wrapped phase and quality maps as dual-channel inputs,the network end-to-end predicts detwisted phase.On a simulated dataset featuring diverse complex terrains and noise levels,QG-ResUNet significantly outperforms traditional methods like SNAPHU,baseline U-Net,and PhaseNet in RMSE,MAE,and SSIM metrics,reducing RMSE to 0.51 rad and improving SSIM to 0.95.Ablation studies demonstrate that the quality map-guided mechanism reduces RMSE by approximately 34.6%,validating its core role.[Conclusions]Testing on real Sentinel-1 data further shows that this method effectively handles complex noise and discontinuities,generating high-quality disentanglement results.

Key words: phase unwrapping, InSAR, deep learning, residual U-Net