Forecasting coastal hypoxia using a blend of mechanistic and artificial intelligence models

Daily fluctuations in coastal hypoxia significantly impact marine ecosystems, requiring forecasts that balance efficiency and accuracy. Statistical models are computationally efficient but often fall short in prediction performance, while mechanistic models are accurate yet resource-intensive. Here, we present a lightweight artificial intelligence (AI) model for daily hypoxia forecasting on the Louisiana-Texas shelf, trained and validated using a 14-year mechanistic ROMS hindcast. The AI model integrates observed riverine nutrient loads and 2-day hydrodynamic forecasts and achieves strong predictive performance: median (± 1 s.d.) accuracy of 0.85 ± 0.07 and F1 score of 0.72 ± 0.18 against the hindcast test set, and 0.67 ± 0.10 accuracy with 0.62 ± 0.14 F1 score against shelf-wide cruise observations. The model remains robust when applied to independent hydrodynamic forecasts (accuracy = 0.71 ± 0.09; F1 score = 0.64 ± 0.17). Beyond forecasting, the AI model enables rapid scenario testing for coastal management. Nutrient reduction assessments suggest that reductions exceeding 90% may be required to meet Gulf Hypoxia Task Force goals. Ablation experiments identify water column stratification as the dominant predictor of daily hypoxia events. This study demonstrates the potential of AI to enhance real-time water quality forecasting, support management decision-making, and inform adaptive cruise planning in dynamic coastal systems.