Reducing SAR-EO Domain Gap via Semantic Alignment for SAR Segmentation

Minseok Seo, Doyi Kim, Sohee Son

Abstract: Synthetic Aperture Radar (SAR) imagery is widely utilized in remote sensing due to its all-weather and day-night imaging capabilities, which offer advantages over Electro-Optical (EO) imagery. However, SAR image interpretation remains challenging, leading to significant interest in automated analysis using deep learning. A primary obstacle is the limited availability of labeled SAR datasets, due to their complex scattering and speckle noise. Typically, aligned EO data with abundant labels is often used for training; however, segmentation performance on SAR images remains suboptimal due to inherent differences in radiometric characteristics. This paper addresses the SAR-EO domain gap by leveraging semantic consistency between SAR and EO images. Despite visual differences, both modalities share highly consistent semantic label spaces. Exploiting this, we propose a semantic-level domain adaptation approach that transfers knowledge from EO to SAR images. Our model requires only SAR data at inference time and achieves state-of-the-art performance in SAR semantic segmentation, effectively bridging the domain gap between SAR and EO.

External IDs:dblp:journals/access/SeoKS25