SAR Data Generation Using Denoising Diffusion Probabilistic Models

The lack of measured data is a common issue plaguing the synthetic aperture radar (SAR) automatic target recognition (ATR) community. For deep learning approaches such as convolutional neural networks (CNNs), a large amount of data is needed to achieve desirable recognition rates. Several solutions to synthetically generating SAR data have been proposed, including using radar signature simulators and generative AI (GenAI). Radar signature simulators are expensive, proprietary, and rely on the ability to correctly model target shape, pose, radar cross section (RCS), and background clutter. Generative adversarial networks (GANs) have been applied to the problem of synthetically generating SAR data but suffer from convergence issues and often have worse generative performance compared to denoising diffusion probabilistic models (DDPMs). In this paper, we propose a conditional DDPM for synthetically generating SAR data. We show the benefit of using conditional DDPMs to generate synthetic SAR data by both GenAI and ATR metrics. Moreover, we show that diffusion-generated data improves ATR performance and permits realistic dataset variability.