Physical prior-guided SAM adaptation for underwater scene segmentation

Underwater image segmentation is fundamental to marine exploration and autonomous underwater vehicle navigation, yet its accuracy is severely compromised by wavelength-selective absorption and scattering inherent to aquatic environments. This paper presents Underwater-SAM, a physics-informed domain adaptation framework for the Segment Anything Model that integrates three synergistic modules. The Spectral Attenuation-Aware Attention module rectifies channel imbalance by modeling wavelength-dependent attenuation, while the Scattering-Aware Feature Enhancement module suppresses backscattering-induced haze in the feature space. Recognizing the inherent ambiguity of underwater object boundaries, the Boundary Uncertainty Estimation module further quantifies this uncertainty to produce probabilistic segmentation outputs. Comparative experiments and ablation studies on a large-scale underwater image benchmark demonstrate that the proposed method achieves optimal performance, substantiating the efficacy of physical prior-guided foundation model adaptation for underwater scenarios.