Hybrid Numerical and AI/ML Models for Operational Hypoxia Forecasting on the Louisiana-Texas Shelf

As climate change continues to escalate threats to coastal ecosystems, accurate and timely hypoxia forecasts are critical for sustaining the livelihoods and well-being of coastal communities. This presentation introduces a hybrid forecasting framework that integrates advanced numerical models with cutting-edge artificial intelligence (AI) and machine learning (ML) techniques to predict both the Hypoxia Area (HA) and the spatial distribution of hypoxia on the Louisiana-Texas (LaTex) shelf. This forecasting system has been operationalized at hypoxia.lsu.edu, part of the comprehensive Gulf of Mexico Coastal Hazards Forecast System developed at LSU.

Our approach includes an ensemble regression model that combines a zero-inflated Poisson generalized linear model (GLM) with a quasi-Poisson generalized additive model (GAM). Trained on a 14-year hindcast from the Regional Ocean Modeling System (ROMS), this model accurately predicts the overall extent of hypoxia, achieving high accuracy with low root-mean-square error (RMSE) and high R2 values. Sensitivity analyses underscore the necessity of aggressive nutrient reduction strategies to meet HA reduction targets, highlighting the model's practical application in environmental management.

Complementing this, we developed a multi-classifier AI model that integrates U-Net and DeepLabv3+ architectures for daily hypoxia distribution forecasts. This AI model, trained and validated on the same numerical hindcast, outperforms existing statistical models for seasonal forecasts, achieving an F1 score, precision, and recall of 0.75, with an overall accuracy of 0.84 against the hindcast test set and a mean accuracy of 0.65 against shelf-wide cruise measurements. The model’s robustness has been confirmed through independent hydrodynamic forecasts, particularly NGOFS2-FVCOM products, which provide the most accurate hypoxia predictions among tested datasets.

Ablation studies reveal that water column stratification is the most critical factor influencing daily hypoxia forecasts, with sedimentary biogeochemical processes playing a secondary but significant role. This insight emphasizes the importance of incorporating both physical and biogeochemical predictors to capture the submesoscale dynamics that drive bottom dissolved oxygen variability.

By operationalizing this hybrid numerical and AI/ML forecasting system through hypoxia.lsu.edu within the Gulf of Mexico Coastal Hazards Forecast System, we provide a powerful tool for predicting both the extent and distribution of hypoxia. This system is poised to enhance the resilience of coastal ecosystems and communities by offering precise, actionable insights to support informed decision-making in the face of climate change.