Comparing Deep Learning Approaches for SAR Imaging: Electromagnetic and Segmentation-Driven Simulation versus Image-to-Image Style Transfer

Abstract This article explores two innovative approaches to simulating synthetic aperture radar (SAR) images from their optical modality equivalents using deep learning methods; one approach also incorporates a physical simulator. The goal is to generate realistic images of large scenes, which could be used for training supervised recognition algorithms or pilot training. We explore, on one hand, the use of conditional generative adversarial networks (cGAN) for supervised transfer from the optical to the radar modality. On the other hand, we use EMPRISE, a physical simulator developed by ONERA, which utilizes a description of materials present in the scene, and we consider achieving this description through supervised semantic segmentation of the optical image. The evaluation of the realism of the obtained SAR images focuses on the fidelity of textures, as well as the presence of bright scatterers. We propose to use the Bhattacharyya distance between the second and third log-cumulant diagrams of the real and simulated images to assess the fidelity of textures. We consider three complementary criteria for evaluating bright point similarity, in particular the chamfer distance between the point clouds detected in both real and simulated images. In our dataset, the cGAN model fails to generate bright points effectively, whereas EMPRISE generates a point cloud with an approximate chamfer distance of 15 pixels. Moreover, EMPRISE tends to overgenerate bright points in the simulated images but demonstrates superior performance in urban areas. Concerning textures, the Bhattacharyya distance for cGAN is approximately 1.72, indicating degraded performance over forested regions, whereas for EMPRISE, this distance is about 0.10, showing the best results in the same forest areas. The results demonstrate that while the cGAN approach produces images with realistic content, it falls short in accurately simulating micro-textures and bright scatterers. Despite being constrained by segmentation into a limited number of classes and exhibiting modest performance, the EMPRISE simulation proves to be more effective than cGAN at reproducing real SAR features.