Physics-guided underwater image enhancement via multimodal spatial distance-decay constraint

Underwater images are affected by light absorption, scattering, and non-uniform illumination, which often lead to color distortion, uneven brightness, and structural blurring, while the scarcity of paired training data further limits model performance. To address these issues, a physics-guided underwater image enhancement method based on multimodal spatial distance–decayed unsupervised contrastive learning is proposed. To alleviate the lack of training data, an unsupervised contrastive learning framework based on unpaired images is constructed, enabling clear-domain data to participate in model optimization directly. To handle uneven illumination and low contrast, a patch-aware frequency block is introduced, which enhances edge details and local contrast through coordinated low- and high-frequency branches in the spatial domain, while adaptively correcting global brightness in the frequency domain. In addition, a spatial distance–decayed contrastive constraint guided by depth and RGB modalities is designed to enhance local structural continuity while maintaining discriminative capability. Finally, a physics-guided constraint based on the underwater imaging model is incorporated to ensure the physical plausibility of the enhanced results. Experimental results on five underwater datasets collected under diverse scenarios demonstrate that the proposed method achieves significant improvements in color restoration, brightness correction, and structural detail recovery.