A Deep Learning Framework for Angular Resolution Enhancement of Multichannel Forward-Looking SAR Under Space-Variant and Perturbed Kernels

Abstract

With multiple channels resolving left/right ambiguity, multichannel synthetic aperture radar (SAR) has the ability to perform forward-looking imaging. However, its angular resolution, especially in regions close to the flight path, is limited by the aperture size. In this article, an enhanced imaging scheme under space-variant and perturbed kernels for multichannel forward-looking SAR is proposed. In the scheme, the ideal space-variant convolution matrix is obtained first with the provided system parameters. Then, the linear mapping equation between the preliminary imaging result and the scene is established. The alternating direction method of multipliers (ADMM) is adopted to solve the optimization problem, and the procedure is unrolled into network form. Meanwhile, to overcome the influence of nonideal factors such as residual errors and noise, the influence of a perturbed kernel function is considered by introducing a perturbation factor and synchronously estimating it during iterations, and the plug-and-play (PnP) module with denoising is added to the network. At last, the experimental results are illustrated to verify the effectiveness of the proposed scheme.