Synthetic Aperture Radar (SAR) enables robust imaging under all weather and lighting conditions, but the scarcity of labeled SAR data limits the use of modern vision models. Novel-view synthesis offers a promising way to augment training data, yet existing methods struggle with speckle noise and radiometric variability inherent to SAR imagery.
We introduce a SAR-specific self-supervised representation learning framework based on co-domain augmentations that operate directly on pixel magnitudes. By combining multiplicative Rayleigh speckle and random monotonic intensity remapping, our method learns features that are invariant to speckle realizations while preserving structural and geometric cues. These learned representations are then used to supervise a latent-diffusion novel-view generator adapted from zero-1-to-3 through a projected feature-matching loss, replacing fragile pixel-space comparisons with noise-robust feature-space supervision.
Experiments on MSTAR and MSTAR-OOD demonstrate substantial improvements in identity preservation, pose consistency, and perceptual quality for both seen and unseen targets. Although evaluated on object-centric SAR for automatic target recognition, the proposed framework is content-agnostic and naturally extends to scene-level SAR novel-view synthesis.