An assessment of the predictability of column minimum dissolved oxygen concentrations in Chesapeake Bay using a machine learning model


Subseasonal to seasonal forecasts have the potential to be a useful tool for managing estuarine fisheries and water quality, and with increasing skill at forecasting conditions at these time scales in the atmosphere and open ocean, skillful forecasts of estuarine salinity, temperature, and biogeochemistry may be possible. In this study, we use a machine learning model to assess the predictability of column minimum dissolved oxygen in Chesapeake Bay at a monthly time scale. Compared to previous models for dissolved oxygen and hypoxia, our model has the advantages of resolving spatial variability and fitting more flexible relationships between dissolved oxygen and the predictor variables. Using a concise set of predictors with established relationships with dissolved oxygen, we find that dissolved oxygen in a given month can be skillfully predicted with knowledge of stratification and mean temperature during the same month. Furthermore, the predictions generated by the model are consistent with expectations from prior knowledge and basic physics. The model reveals that accurate knowledge or skillful forecasts of the vertical density gradient is the key to successful prediction of dissolved oxygen, and prediction skill disappears if stratification is only known at the beginning of the forecast. The lost skill cannot be recovered by replacing stratification as a predictor with variables that have a lagged correlation with stratification (such as river discharge); however, skill is obtainable in many cases if stratification can be forecast with an error of less than about 1 kg m−3. Thus, future research on hypoxia forecasting should focus on understanding and forecasting variations in stratification over subseasonal time scales (between about two weeks and two months).

Estuarine, Coastal and Shelf Science