Synthetic Aperture Radar Deep Statistical Imaging Through Diffusion Generative Model Conditional Inference

Abstract

Synthetic aperture radar (SAR) plays a crucial role in remote sensing because of its ability to operate in all weather conditions, both day and night. The traditional FFT-based SAR imaging algorithm suffers from severe speckle noise, which is almost inevitable owing to the coherent nature of the SAR system. Recently, the plug-and-play (PnP) SAR imaging method uses a plug-in denoiser as an image prior function to regularize the resulting image, thus suppressing speckle noise while maintaining the useful features of target objects. However, the existing plug-in denoisers used in statistical SAR imaging, either handcrafted or data-driven, are insufficient for complex remote sensing scenarios. More powerful image priors, such as the deep generative model for unconditional image generation, would be a better alternative regularizer for statistical SAR imaging. However, the most powerful diffusion generative model lacks an explicit latent space for conditional optimization to be adopted for SAR imaging from received signals. We propose a novel SAR imaging method based on conditional generation of a diffusion model. In detail, we embed the maximum a posteriori (MAP) formulation of SAR imaging from the received signal as a conditional guidance for diffusion generation, which overcomes the lack of latent space shortage. Compared with these statistical methods, our proposed methods exhibit exceedingly high performance both on simulated experiments and returned data imaging from RadarSat SAR data.