PhysiOcean: A Spatio-Temporal Physics-Informed Operator Network for Multi-Point Coastal Dissolved Oxygen Forecasting

摘要

Forecasting dissolved oxygen (DO) in coastal waters is crucial for assessing environmental sustainability, particularly for evaluating habitat quality and the survival of aquatic organisms. Forecasting DO spatiotemporal evolution remains challenging due to nonlinear physical-biogeochemical couplings and sparse observations. However, existing data-driven methods either predict DO at isolated stations without spatial coherence, rely on rigid spatial decompositions that cannot generalize to unobserved locations, or ignore physical constraints altogether. To address these issues, we propose PhysiOcean, a physics-informed multi-head Deep Operator Network that maps time-series environmental forcings and spatial coordinates to DO concentrations in an end-to-end framework, which introduces three key innovations. Specifically, it employs a multi-scale causal temporal encoder to capture forcing effects from synoptic to seasonal timescales, along with a Fourier spatial decoder with learnable frequencies and EOF-guided pretraining to overcome spectral bias. Building on these, a multi-head fusion mechanism with nonlinear correction transcends the standard bilinear DeepONet formulation. We provide a PAC-style generalization analysis of how physics-informed constraints affect the model's Rademacher complexity. We validate PhysiOcean using a 32-year Chesapeake Bay monitoring dataset. PhysiOcean outperforms the best baseline, and maintains robust prediction at spatially unobserved locations. Our code is available at https://anonymous.4open.science/r/physiocean/.