OS-SBDA: a dual-track Schrödinger Bridge adversarial framework for unpaired optical-to-SAR image translation

Unpaired optical-to-synthetic aperture radar (Opt2SAR) image translation addresses the scarcity of strictly aligned multimodal datasets. However, the fundamentally distinct physical mechanisms of optical reflectance and microwave backscattering hinder accurate cross-modal mapping, making this domain gap difficult to bridge for existing generative models. We propose the Optical-to-SAR Schrödinger Bridge Diffusion Adversarial (OS-SBDA), a dual-track Schrödinger Bridge adversarial framework for unpaired optical-to-SAR image translation, which formulates the Opt2SAR task as a distribution matching problem. Rather than relying on unconstrained stochastic processes, our approach approximates a discrete energy-constrained Schrödinger Bridge trajectory connecting the optical and SAR distributions. To ensure SAR statistical consistency and structural fidelity along this trajectory, the framework integrates a dual-track feature alignment mechanism: a patch-wise contrastive learning module preserves the spatial topology of the optical source, while a global Gram matrix penalty accurately injects the high-frequency speckle statistics inherent to SAR backscattering. Extensive evaluations on the QXS-SAROPT dataset demonstrate the superiority of OS-SBDA over established fully supervised and unsupervised benchmarks. Notably, the energy-constrained bridge converges in only 5 function evaluations, as measured by the Number of Function Evaluations (NFE), compared to 50 NFE for continuous-time diffusion baselines, accelerating inference by a factor of 27 while minimizing distribution approximation errors. The architecture yields exceptional perceptual quality, achieving a Fréchet Inception Distance of 84.40 and high correlation in spatial textures, effectively suppressing the non-physical oversmoothing artefacts typical of standard generative models. OS-SBDA provides an efficient, statistically-aligned paradigm for robust cross-modal remote sensing data synthesis.