S3U-SAR: Physics-driven AI deciphers aircraft from radar images

Synthetic Aperture Radar (SAR) is crucial for all-day, all-weather target interpretation, but existing methods rely on unordered local scattering centers that often miss physically existing components with weak signals. To overcome this, a team led by Yifei Yin introduces a new paradigm: Semantic Scattering Structure Understanding. Their framework, S3U-SAR, defines semantic scattering keypoints that associate local electromagnetic responses with specific aircraft parts (e.g., wings, fuselage). Visibility-aware attributes preserve components that are physically present but weakly observable. These keypoints are organized into a stable semantic scattering structure, guided by multi-dimensional physical priors including scattering heterogeneity, rigid-body topology, and speckle uncertainty. A confidence-gated joint supervision strategy alleviates optimization conflicts during training.

The team also created KP-SAR-Aircraft-1.0, the first fine-grained benchmark for this task. Extensive experiments show S3U-SAR achieves state-of-the-art performance compared to baselines, with superior robustness and transferability across different aircraft categories and datasets. This work moves beyond texture-based or local-center methods, providing a physically grounded, component-aware interpretation of radar imagery—a significant step for defense, surveillance, and remote sensing applications.