A Wrap-Count-Based Phase Unwrapping Method for Large-Scale, Low-Coherence Interferograms Using Deep Learning

Unwrapping synthetic aperture radar interferograms with extensive low-coherence regions remains challenging, even when the deformation signal has a low gradient, such as that due to interseismic displacement. Here, we present a novel algorithm to unwrap low-gradient interferograms more efficiently and reliably using a semantic segmentation neural network. We first partition large-scale interferograms into overlapping patches and employ a trained Segformer network, making full use of spatial features, to identify decorrelated pixels while predicting the wrap count for coherent pixels. In a further step, we correct wrap counts or re-unwrap certain patches, based on a reliability metric. Finally, the patches are mosaicked to reconstruct the fully unwrapped interferogram by leveraging overlapping areas. The Segformer model is trained on over 20 000 simulated samples with varying decorrelation noise and more than 10 000 real-world samples from the COMET-LiCSAR portal. Synthetic experiments show that our approach significantly reduces the mean absolute error (MAE) compared to the classical minimum cost flow (MCF) method. Further validation on thousands of interferograms from the western Altyn Tagh Fault (ATF) and the western Haiyuan Fault (HF) confirms its superior performance in isolated regions separated by decorrelated noise and its ability to suppress the wide-ranging propagation of unwrapping errors, with the proportion of unwrapping errors reduced by 28%–83% for real interferograms. These results highlight the potential of our method to be used in large-scale automated processing of straining regions, such as the Alpine-Himalayan belt.