Hyperspectral Aquatic Understanding Foundation Model: A physics-aware and cross-platform learning strategy

Hyperspectral foundation models play a critical role in Earth observation, but existing ones are mainly developed for terrestrial applications and remain insufficient for aquatic scenarios. This limitation arises from two key challenges. First, existing aquatic hyperspectral datasets are usually built for isolated tasks, leading to substantial heterogeneity in preprocessing and annotations, which hinders their integration for large-scale pretraining. Second, aquatic hyperspectral imaging is fundamentally different from land observation, making it difficult to directly transfer terrestrial-oriented foundation model designs and learning strategies to aquatic scenarios. To address these challenges, we construct ATR-HyperAquatic-600K, a large-scale aquatic hyperspectral dataset containing about 600,000 image patches from satellite and airborne platforms, multiple sensors, diverse water environments, and broad geographic coverage. Based on ATR-HyperAquatic-600K, we further propose the Hyperspectral Aquatic Understanding Foundation Model (HyperAUDI), the first hyperspectral aquatic understanding foundation model for unified aquatic hyperspectral interpretation. HyperAUDI develops a physics-aware satellite pretraining strategy, which incorporates aquatic imaging mechanisms into representation learning to enable aquatic-oriented spectral–spatial representation learning. To further bridge satellite and airborne observations, HyperAUDI formulates cross-platform learning as a coarse-to-fine continual pretraining process, where satellite-derived historical knowledge guides fine-grained airborne representation learning. Extensive experiments on multiple aquatic downstream tasks, including benthic classification, bathymetry estimation, oil spill detection, and underwater target detection, demonstrate that HyperAUDI achieves state-of-the-art performance and learns transferable representations across heterogeneous aquatic scenes.