A Novel Two-Stage Learning-Based Phase Unwrapping Algorithm via Multimodel Fusion

Phase unwrapping (PhU) is one of the key steps in interferometric synthetic aperture radar (InSAR) data processing, and it is a considerable challenge for PhU in regions with high-noise and large-gradient changes. Deep learning phase unwrapping (DLPU) can better solve this problem. However, a single DLPU algorithm still finds it difficult to obtain robust PhU results in regions with large-gradient changes. In addition, the performance of the same training model varies greatly for different data. To solve this problem, this paper combines a deep neural network model with the traditional PhU model and proposes a novel two-stage learning-based phase unwrapping (TLPU) algorithm via multimodel fusion. The major advantages of TLPU are as follows: 1) A high-resolution U-Net (HRU-Net) model trained on a dataset constructed according to InSAR interferometric geometry is utilized for the PhU for the first time, which effectively improves the performance of the DLPU. 2) TLPU utilizes the traditional PhU method to optimize the results of DLPU, addressing the issue of weak generalization ability of a single DLPU, while improving accuracy in areas with large-gradient changes. Experimental analysis was carried out using LT-1 data, and the results show that the proposed TLPU algorithm can achieve superior excellent results in large-gradient change regions compared with the commonly used PhU method, with root mean square errors of only 1.63 rad in experiment 1 and 1.96 rad in experiment 2.