A Numerical and Machine Learning Investigation of Water Quality in the Northern Gulf of Mexico, from Estuaries to Shelf
Doctor of Philosophy (PhD)
Oceanography & Coastal Sciences
In this study, numerical and machine learning (ML) models are developed and adapted to investigate the water quality of the Barataria Estuary and the Louisiana-Texas (LaTex) Shelf in northern Gulf of Mexico. By using a 3-D hydrodynamic ROMS model, the salinity variability of the Barataria Estuary is found highly influenced by the Mississippi River discharges and man-made river diversion systems. The Mississippi River freshwater can intrude into the estuary through the middle and the east passes. Sensitivity tests indicate that decreases in lower estuarine salinity due to the elevated Mississippi River discharges are more apparent than the salinity increases due to the reduced discharges and that during ~68% of the investigated time salinity in the lower estuary is lower than 5 due to the impacts of the proposed mid-Barataria sediment diversion system. Coupled with a 3-D hydrodynamic model, a modified NEMURO model is developed to explore the hypoxia dynamics in the LaTex Shelf. Results highlight the importance of the complexity of the lower-trophic community in bottom dissolved oxygen’s response to the changing nutrient loads. While sole nutrient reductions do not always guarantee a decrease in hypoxic area due to competition within the community, the hypoxia reduction goal of 5000 km2 is likely to be achieved with simultaneous reductions (~80%) on riverine nitrogen, phosphorous, and silicon. Built on the hindcast of the coupled physical-biogeochemical model, a ML hypoxic area forecast model for the LaTex Shelf is developed combining the prediction efforts by a zero-inflated generalized linear model and a generalized additive model and can predict up to 77 % of the total variability of the hindcast. Comparing against the Shelf-wide cruise observations, the prediction model provides a high R2 (0.9200 vs 0.2577–0.4061 by existing forecast models, same comparison hereinafter), a low root-mean-square error (2,005 km2 vs 4,710–9,614 km2), a low scatter index (15 % vs 36 %–95 %), low mean absolute percentage biases for overall (18 % vs 44 %–132 %), fair-weather summers (15 % vs 8 %–46 %), and windy summers (18 % vs 33 %–74 %) predictions.
Ou, Yanda, "A Numerical and Machine Learning Investigation of Water Quality in the Northern Gulf of Mexico, from Estuaries to Shelf" (2022). LSU Doctoral Dissertations. 5969.
Available for download on Friday, October 17, 2025