PUSIF-DUNet: Deep SAR Imaging Network With Joint Optimization on Learnable Probabilistic Undersampling and Unambiguous Reconstruction

Since the minimum antenna area constraint in the spaceborne single-channel synthetic aperture radar (SAR), it's difficult to achieve high-resolution and wide-swath (HRWS) imag ing simultaneously. A larger range swath requires a lower pulse repetition frequency (PRF) to increase the minimum interval between adjacent samples. However, low PRF results in aliasing of the azimuth spectrum, leading to ghost targets in the focused images. Compared to uniform undersampling, nonuniform sam pling can suppress the ambiguity by converting the structured spectrum aliasing caused by uniform undersampling to the nonstructured spectrum noise. However, existing methods face several challenges. First, ambiguity suppression techniques based on filtering exhibit limitations in effectively dealing with nonuni form sampling. Second, ambiguity suppression performance of predetermined nonuniform sampling deteriorates under low undersampling ratio. Third, current works regard nonuniform sampling pattern and sparse reconstruction as two independent problems, overlooking the impact of different SAR images and reconstruction methods on the optimal undersampling pattern. Forth, iterative-based sparse SAR imaging algorithms involve a large number of iterations, along with manual parameter setting that may result in suboptimal values, adversely impacting the imaging performance. To address these issues, we propose a novel end-to-end network for azimuth ambiguity suppression in single-channel SAR by jointly exploring an optimal undersam pling pattern and learning reconstruction network parameters, dubbed as PUSIF-DUNet. PUSIF-DUNet adopts the probabilistic undersampling (PU) scheme to learn an optimal Bernoulli un dersampling matrix instead of a deterministic pattern, that can reconstruct SAR images with effectively suppressed ambiguities and provide superior imaging performance at low undersam pling ratio. We adopt sparse SAR recovery algorithm to solve the nonuniform sampling imaging problem and unfold it into deep learning network. The sampling information fusion (SIF) technique incorporates the learned probabilistic undersampling pattern into each layer of deep unfolding network (DUN) to further exploit the information in sampling network and improve the performance. Extensive experiments in both simulated and real SAR scenarios demonstrate that our proposed network not only achieves state-of-the-art results in terms of both quantitative metrics and visual quality even at low undersampling ratios but also rapidly performs unambiguous SAR imaging from raw echo data.