Unsupervised Learning for Forward-Looking MIMO SAR Reconstruction based on U-Net

The angular resolution presents a significant challenge in forward-looking multiple-input multiple-output synthetic aperture radar (FL-MIMO-SAR) applications, stemming from factors like antenna size, slow vehicle speeds, and physical constraints. In this article, we propose an unsupervised learning method based on U-Net to enhance angular resolution and address the issue of lacking ground truth data, which is common in practical scenarios. Our approach introduces a model-based image reconstruction framework, incorporating prior knowledge of the ground truth into the reconstruction process to enhance image quality. Through simulations, we demonstrate that U-Net outperforms convolutional neural networks (CNNs), achieving higher convergence speeds and producing images with superior angular resolution.