PIPNet: A Deep Convolutional Neural Network for Multibaseline InSAR Phase Unwrapping Based on Pure Integer Programming

Multibaseline (MB) phase unwrapping (PU), as the core step in MB InSAR, breaks the limitation of phase continuity assumption. However, it still suffers from insufficient noise robustness and low unwrapping efficiency. This paper transforms the MB PU problem into pure integer programming (PIP) problem and innovates a deep convolutional neural network (DCNN) named PIPNet to solve PU problem. This PIPNet is built on the U-Net framework, incorporating Transformer modules, upsampling, and dense connections mechanisms to achieve powerful feature extraction capabilities. Interferometric phase, fringe frequency, and intercept data are input into multi-ports of PIPNet, with multi-source features fused in the encoding-decoding stages to collectively act on PU. Then, we innovatively designed a new joint loss function and PIP constraints to guide the iterative optimization of model parameters. The attention mechanism enhances the model ability to disentangle interferometric phase details. Finally, the trained model can achieve real-time and accurate calculation of phase ambiguity number. Theoretical analysis and two experimental results indicate that PIPNet-PU has much lower execution time than MCF, CA, and TSPA algorithms, and the second-best unwrapping accuracy for short- baseline interferogram and the best unwrapping ability for long- baseline interferogram. PIPNet-PU can achieve good unwrapping effects even in areas with abrupt topographic change and dense interferometric fringes.