A deep learning-based approach using dual-label samples for InSAR phase unwrapping

Phase unwrapping (PU) is a critical procedure in the field of interferometry synthetic aperture radar (InSAR). Improving the accuracy of phase unwrapping in areas with phase discontinuities is a critical task that has been delegated to deep learning methodologies. However, addressing the uncertainties associated with annotated sample labels remains a significant challenge. In this work, we present a dual-label sample training strategy designed to mitigate the impact of these uncertainties on the estimate of wrap-count gradients. Additionally, by analyzing the spatial distribution characteristics of labels, we propose a semantic segmentation model (DSANet) that excels in extracting features of specific shapes. Through rigorous testing with simulated and real-world datasets, coupled with comparative analyses against statof-the-art phase unwrapping (PU) algorithms, our findings substantiate the efficacy of our presented dual-label sample training strategy and underscore the robust unwrapping capabilities of our proposed methodology.