An error distribution-related function-trained two-dimensional InSAR phase unwrapping method via U-GauNet

Phase unwrapping (PU) has always been a critical and challenging step in interferometric synthetic aperture radar data processing. Due to the presence of noise, the interferogram usually presents phase inconsistencies, also called residues, which imply a nonunivocal solution. This work investigates the PU problem from a semantic segmentation perspective by exploiting convolutional neural network models and regards the PU as a two-stage process. In the first stage, the phase ambiguity gradient is directly estimated from the wrapped phase by the proposed U-GauNet which is based on the U-Net architecture, combines with the Global Attention Upsample module and removes the one-step decoder module. The U-GauNet is trained by the Cross-Entropy and a loss function which is related to the error distribution and proposed to punish the error zero gradient pixels that we find will make a significant influence to the PU result from aspects of the number and the percentage. In the second stage, the L1-norm is used to minimize the difference between the true ambiguity gradient and the results of the stage one to achieve the PU result. The proposed method is tested on the simulated data and true data of Three Gorges in China. Experimental results demonstrate the result of the proposed method is more credible than traditional 2-D PU methods.