A machine-learning-based model for water quality in coastal waters, taking dissolved oxygen and hypoxia in Chesapeake Bay as an example

Abstract

Hypoxia is a big concern in coastal waters as it affects ecosystem health, fishery yield, and marine water resources. Accurately modeling coastal hypoxia is still very challenging even with the most advanced numerical models. A data‐driven model for coastal water quality is proposed in this study and is applied to predict the temporal‐spatial variations of dissolved oxygen (DO) and hypoxic condition in Chesapeake Bay, the largest estuary in the United States with mean summer hypoxic zone extending about 150 km along its main axis. The proposed model has three major components including empirical orthogonal functions analysis, automatic selection of forcing transformation, and neural network training. It first uses empirical orthogonal functions to extract the principal components, then applies neural network to train models for the temporal variations of principal components, and finally reconstructs the three‐dimensional temporal‐spatial variations of the DO. Using the first 75% of the 32‐year (1985–2016) data set for training, the model shows good performance for the testing period (the remaining 25% data set). Selection of forcings for the first mode points to the dominant role of streamflow in controlling interannual variability of bay‐wide DO condition. Different from previous empirical models, the approach is able to simulate three‐dimensional variations of water quality variables and it does not use in situ measured water quality variables but only external forcings as model inputs. Even though the approach is used for the hypoxia problem in Chesapeake Bay, the methodology is readily applicable to other coastal systems that are systematically monitored.