OSCAR: Optical-aware Semantic Control for Aleatoric Refinement in Sar-to-Optical Translation

TL;DR

OSCAR tackles the ill-posed SAR-to-Optical translation by introducing an Optical-aware SAR Encoder learned through cross-modal distillation from a satellite-trained optical teacher, enabling optical-like semantic priors to guide SAR representations. A Semantically-Grounded ControlNet then injects both global class-aware prompts and dense hierarchical visual prompts into a latent diffusion process, stabilizing structure and texture under speckle noise. An Uncertainty-Aware Objective dynamically modulates reconstruction losses with pixel-wise confidence, reducing artifacts in ambiguous regions. Across BigEarthNet-v2 and SEN12MS, OSCAR achieves state-of-the-art perceptual and semantic fidelity, validated by ablations showing the critical roles of cross-modal alignment and multi-level semantic guidance for robust SAR-to-Optical synthesis.

Abstract

Synthetic Aperture Radar (SAR) provides robust all-weather imaging capabilities; however, translating SAR observations into photo-realistic optical images remains a fundamentally ill-posed problem. Current approaches are often hindered by the inherent speckle noise and geometric distortions of SAR data, which frequently result in semantic misinterpretation, ambiguous texture synthesis, and structural hallucinations. To address these limitations, a novel SAR-to-Optical (S2O) translation framework is proposed, integrating three core technical contributions: (i) Cross-Modal Semantic Alignment, which establishes an Optical-Aware SAR Encoder by distilling robust semantic priors from an Optical Teacher into a SAR Student (ii) Semantically-Grounded Generative Guidance, realized by a Semantically-Grounded ControlNet that integrates class-aware text prompts for global context with hierarchical visual prompts for local spatial guidance; and (iii) an Uncertainty-Aware Objective, which explicitly models aleatoric uncertainty to dynamically modulate the reconstruction focus, effectively mitigating artifacts caused by speckle-induced ambiguity. Extensive experiments demonstrate that the proposed method achieves superior perceptual quality and semantic consistency compared to state-of-the-art approaches.

Figures (8)

• Figure 1: Qualitative comparison of S2O translation results on three datasets. We compare our proposed method against state-of-the-art baselines (StegoGAN, ControlNet, cBBDM). Below each example, the corresponding Class-aware Prompt generated by our Optical-aware SAR Encoder is displayed. As shown, our method demonstrates superior structural fidelity and perceptual quality compared to other state-of-the-art models, effectively preserving semantic details while suppressing artifacts.
• Figure 2: Schematic of the proposed Cross-Modal Semantic Alignment. A Cross-Modal distillation strategy is employed where the trainable SAR Student mimics the frozen Optical Teacher.
• Figure 3: Overview of the proposed OSCAR framework. The architecture utilizes the Student SAR model (pre-trained in Fig. \ref{['fig:distill']}) as the Optical-aware SAR Encoder. Key components include the injection of hierarchical semantic information extracted by the Student SAR model, class-aware text prompts for global semantic context, and the estimation of a pixel-wise confidence map to optimize the uncertainty-aware objective.
• Figure 4: Qualitative comparison on BigEarthNet-v2 and SEN12MS. While baseline methods often suffer from geometric distortion or texture blurring, OSCAR generates structurally coherent and photorealistic results.
• Figure 5: Qualitative comparison of ablation variants on BigEarthNet-v2.