MMPhU-Net: A Novel Multi-Model Fusion Phase Unwrapping Network for Large-Gradient Subsidence Deformation
Yandong Gao, Jiaqi Yao, Nanshan Zheng, Shijin Li, Hefang Bian, Yu Tian
Abstract
The problem of phase unwrapping (PhU) in the large-gradient deformation areas is the bottleneck problem of interferometric synthetic aperture radar (InSAR) data processing. However, the extraction of large-gradient deformation areas is one of the key issues in coal mining deformation monitoring. Here, we propose a novel multimodel fusion PhU Network, abbreviated as MMPhU-Net, and apply it to the extraction of large-gradient deformation areas. The major advantages of MMPhU-Net are as follows: First, MMPhU-Net combines the advantages of different basic network models, which can improve the model convergence speed and phase gradient estimation accuracy. MMPhU-Net can improve the lack of recognition effect of a single basic model. Second, different from existing deep learning PhU methods, MMPhU-Net directly estimates the gradient ambiguity numbers, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i> , so its phase gradient estimation completely breaks through the (-π, π) limitation. Therefore, MMPhU-Net can obtain ideal PhU results in large-gradient deformation areas. In addition, optimization algorithm models are used to optimize the estimation results of the multimodel fusion network. Subsequently, the obtained <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i> and a novel two-step filtering method are combined to obtain the final PhU results. Through the verifications of simulated data sets and realistic GaoFen-3 SAR data sets, the proposed MMPhU-Net method can achieve superior excellent results than the commonly used PhU method.