Abstract: The accuracy of terrain elevation or surface deformation retrieval relies heavily on the quality of InSAR phase unwrapping. Conventional phase unwrapping techniques, rooted in non-machine learning models (such as path-following or minimum norm), face challenges in producing accurate unwrapping outcomes within areas of low coherence or high phase gradients (dense interference fringes). Deep neural network models offer distinct advantages in nonlinear representation and feature expression, widely employed in digital image processing research, wherein InSAR phase unwrapping parallels image regression.This paper presents an InSAR phase unwrapping approach utilizing the R2AU-net. Initially, pairs of wrapped and unwrapped phases are simulated through mathematical fractal methods, circumventing inherent errors and artifacts introduced by integrating external DEMs into the phase. This approach maintains terrain feature diversity and complexity while providing the requisite dataset for model training. Subsequently, the R2AU-net phase unwrapping model, built upon the foundational U-net model, incorporates attention mechanisms to augment the model's convolutional feature selection capacity, thereby improving unwrapping performance in regions of low coherence or dense striping. The utilization of recurrent residual convolutional structures addresses the vanishing gradient issue, enhancing the model's feature representation capability.Ultimately, experimental analyses are conducted using both simulated and real data. The results demonstrate that the proposed R2AU-net phase unwrapping model effectively retains terrain elevation or real surface deformation information, thereby bolstering the reliability of unwrapping outcomes. In terms of performance, it surpasses established methods such as the Goldstein branch-cut method, SNAPHU method, as well as CNN and U-Net phase unwrapping models.

Key words: InSAR, phase unwrapping, deep neural network, U-Net, surface deformation