TS2O: A Token-Based Two-Stage Generative Architecture for SAR-to-Optical Translation

The widely used optical remote sensing imagery faces limitations due to illumination and weather conditions, whereas synthetic aperture radar (SAR) provides all-weather acquisition capabilities, yet it presents challenges in interpretation. This article introduces TS2O, a novel token-based two-stage framework for SAR-to-optical (S2O) translation that synergizes SAR’s acquisition robustness with optical imagery’s interpretability. Initially, token-based SAR-to-optical (TS2O) compresses both SAR and optical images into semantic tokens. Subsequently, our proposed frequency–attention fusion GAN, featuring frequency-aware enhancement and attention fusion modules, translates these SAR tokens to optical tokens, enhancing cross-modal feature alignment prior to image reconstruction. To further refine output quality, we introduce structure-guided loss, a composite loss function constraining brightness, contrast, and structural consistency. Experiments on the MSAW dataset show TS2O surpassing six GAN and two diffusion methods in image evaluation metrics and downstream building segmentation performance, demonstrating its capability for superior S2O translation with enhanced detail preservation.