Deep-Learning-Based Mask-Cut Method for InSAR Phase Unwrapping

Abstract

Two-dimensional phase unwrapping (2D-PU) is a key processing step for interferometric synthetic aperture radar (InSAR) and it plays an important role in InSAR data processing. For the phase unwrapping (PU) problem, many scholars began to consider using the deep learning (DL) technology in the field of artificial intelligence. By accumulating InSAR PU processing experience through DL, the learning-based PU method can surpass the traditional PU algorithm sometimes. Therefore, this article designs a mask-cuts (MCs) deployment network based on DL, which is named MCNet, and the PU method based on this network is also known as MCNet-PU. First, the residues images and its corresponding MCs images are obtained by using the traditional MC method as the training data and testing data. Second, the relationship between residues and MCs is learned through the training of the self-built MCNet. Then, the trained MCNet is used to obtain the MCs corresponding to the interferogram to be unwrapped. Finally, the unwrapped result is obtained by phase integration using the traditional flood fill method. Compared with the traditional MC method, MCNet does not need to use the quality map to guide the deployment of the MCs, nor does it need to refine the MCs, and it can make the deployment of the MCs more accurate. Experiments on simulated and real InSAR data show that the MCNet-PU method can improve the phase unwrapping success ratio (PUSR) by about 4%–15%, which shows the effectiveness of the method.