Towards Fully Synthetic Training: Exploring Data Augmentations for Synthetic-to-Measured SAR in Automatic Target Recognition

The scarcity of measured Synthetic Aperture Radar (SAR) images, driven by high costs and time constraints, poses a significant challenge to developing robust Automatic Target Recognition (ATR) systems. Artificially generated SAR images, unlike real measured data captured by radar systems, provide a potential alternative. However, discrepancies between these synthetic SAR images and the real-world measured SAR images hinder effective generalization. This paper introduces data augmentation techniques to train effective SAR-ATR models exclusively on synthetic SAR data, achieving high accuracy on real measured SAR data. We employ image augmentation techniques that mimic essential radar signal components, including target variations with color jitter, system characteristics with Gaussian noise, and environmental effects using random erasing. We assessed these techniques individually and in combination across multiple deep learning architectures. Our experiments showed that applying Gaussian noise and random erasing with the ConvNeXt-Base model achieved a promising accuracy of 92.01% on measured SAR data, despite training exclusively on synthetic SAR data. Qualitative analyses, including confusion matrices, t-SNE visualizations, and class activation maps, further show improved class separation and feature learning, especially for challenging targets. These findings highlight the potential of effective augmentation strategies to bridge the synthetic-to-measured SAR gap, offering a scalable solution for SAR-ATR systems where measured data are limited or unavailable. The code is available at https://github.com/adu-baffour/Synthetic-to-Measured-SAR/.