MCR-PFNet: A novel InSAR interferometric phase filtering method for complex noise and large gradient deformations

Abstract

Phase filtering is crucial for ensuring accurate phase unwrapping in the processing of Interferometric Synthetic Aperture Radar (InSAR) data. Traditional filtering methods often struggle to effectively suppress noise and accurately preserve phase edge information when processing InSAR interferometric phases with complex noise and large gradient deformations. In response to this challenge, we have proposed the MCR-PFNet model based on residual attention convolution for InSAR phase filtering. MCR-PFNet integrates residual blocks, the convolutional block attention module (CBAM), and a multi-head self-attention module, enabling the simultaneous extraction of both local and global phase features, while filtering out noise and preserving phase details. To further enhance the generalization capability of MCR-PFNet, additive Gaussian noise and local phase jumps were introduced into the training dataset, and the MCR-PFNet model was trained with a custom-designed periodic phase loss function. The filtering performance was evaluated on simulated and real datasets. The results demonstrate that MCR-PFNet excels in scenarios with complex noise and large gradient deformations. In the simulated wrapped phase experiments, MCR-PFNet outperformed other methods in three metrics: PSNR (55.820–64.017 dB), SSIM (0.965–0.990), and MSE (0.026–0.170 rad²). In two sets of real data experiments, MCR-PFNet significantly improved the quality of the interferograms. In the real InSAR data of coal mining subsidence, the filtered NOR decreased to 517, PRR reached 86.987 %, and Metric Q reached 95.278 %. In the real InSAR data of earthquakes, NOR was reduced by 4.016 %–79.997 % compared to other methods, while PRR and Metric Q increased by 0.116 %–25.899 % and 0.111 %–13.803 %, respectively.