Empowering Hybrid-Level Contrastive Learning for Hyperspectral Underwater Target Detection in Nearshore Environment

UAV-borne hyperspectral remote sensing has emerged as a promising technique for underwater target detection (UTD). Existing hyperspectral UTD (HUTD) methods in deep-sea environments typically rely on bathymetric models to characterize spectral attenuation caused by the water column, often coupled with restoration- or prediction-based pipelines for target detection. However, in nearshore environments, high turbidity and complex seabed topography introduce nonlinear and heterogeneous attenuation effects that cannot be accurately modeled by conventional bathymetric models. Moreover, the highly dynamic nature of nearshore waters induces significant spectral variability in target signatures, rendering restoration- and prediction-based pipelines ineffective due to their limited capacity to model such variability. To address these challenges, we propose the Hyperspectral Underwater Contrastive Learning Network (HUCLNet), a data-driven framework for HUTD that eliminates dependence on bathymetric models. Rather than modeling spectral attenuation, HUCLNet establishes a semantically meaningful latent space with enhanced target-background separability through a hybrid-level contrastive learning framework. A reliability-guided clustering strategy is introduced to refine the contrastive learning inputs and improve representation robustness. Furthermore, a self-paced learning mechanism flexibly integrates the clustering and contrastive modules to stabilize training and accelerate convergence. Extensive experiments demonstrate that HUCLNet outperforms state-of-the-art methods across various evaluation metrics.