생성형 변환 모델을 통한 SAR 데이터의 의미론적 해석 및 시각적 이해 확장 = Extending Semantic Interpretation and Visual Understanding of SAR Data via Generative Translation Models

다국어 초록 (Multilingual Abstract)

Synthetic aperture radar (SAR) offers critical all-weather observation capabilities, yet its interpretation remains challenging due to inherent speckle noise and non-intuitive scattering characteristics. Consequently, directly applying vision-language models (VLMs) trained on natural images to the SAR domain is limited by significant modality gaps and the scarcity of high-quality SAR-text datasets. To overcome these challenges, this study proposes a two-stage framework that leverages SAR-to-optical translation to bridge the domain gap. First, we introduce a conditional Brownian Bridge Diffusion Model integrated with a SAR feature guidance module. This approach transforms SAR images into optical-like representations while preserving structural fidelity, thereby addressing the geometric distortions and hallucinations common in generative adversarial network (GAN)-based methods. Second, the translated images are analyzed by a domain-adapted VLM, utilizing the GeoRSCLIP visual encoder and a LoRA-tuned LLaVA model to generate precise semantic captions. Experimental results using Sentinel-1 and Sentinel-2 datasets demonstrate that the proposed translation model outperforms existing GAN models in terms of PSNR and SSIM. Furthermore, the framework achieves significant improvements in captioning metrics, including BLEU, ROUGE-L, and BERT-Score, compared to direct SAR interpretation. This study validates that high-fidelity modality translation can effectively extend the reasoning capabilities of pre-trained VLMs to the SAR domain without requiring extensive SAR-specific annotations.