AUTOMATIC SIMULATION OF SAR IMAGES: COMPARING A DEEP-LEARNING BASED METHOD TO A HYBRID METHOD

This study compares two approaches for simulating synthetic aperture radar (SAR) images. The first approach uses a conditional Generative Adversarial Network (cGAN) to learn statistical image distributions from optical images. In a second approach, we generate SAR images using a electromagnetic simulator taking into input material maps obtained by segmenting optical images. We propose two metrics to evaluate the quality of the simulation. We evaluate the methods on existing Sentinel-1 SAR images of France using the DREAM database. The results suggest that the physical simulator with automatically created material maps is better suited for generating realistic SAR images compared to the cGAN approach, even if a lot of work remains to be done on the complexity of the description of the scene.