Hydrodynamic and biochemical impacts on the development of hypoxia in the Louisiana-Texas shelf - Part 2: statistical modeling and hypoxia prediction

Abstract

This study presents a novel ensemble regression model for forecasts of the hypoxic area (HA) in the Louisiana-Texas (LaTex) shelf. The ensemble model combines a zero-inflated Poisson generalized linear model (GLM) and a quasi-Poisson generalized additive model (GAM) and considers predictors with hydrodynamic and biochemical features. Both models were trained and calibrated using the daily hindcast (2007-2020) by a three-dimensional coupled hydrodynamic-biogeochemical model embedded in the Regional Ocean Modeling System (ROMS). Compared to the ROMS hindcasts, the ensemble model yields a low root-mean-square error (RMSE) (3256 km2), a high R2 (0.7721), and low mean absolute percentage biases for overall (29%) and peak HA prediction (25%). When compared to the shelf-wide cruise observations from 2012 to 2020, our ensemble model provides a more accurate summer HA forecast than any existing forecast models with a high R2 (0.9200); a low RMSE (2005 km2); a low scatter index (15%); and low mean absolute percentage biases for overall (18%), fair-weather summer (15%), and windy-summer (18%) predictions. To test its robustness, the model is further applied to a global forecast model and produces HA prediction from 2012-2020 with the adjusted predictors from the HYbrid Coordinate Ocean Model (HYCOM). In addition, model sensitivity tests suggest an aggressive riverine nutrient reduction strategy (92%) is needed to achieve the HA reduction goal of 5000 km2.