Assessment of Environmental Variables that Affect Harmful Algal Blooms and Hypoxia in Lake Erie Using Multi-media Modeling and Machine Learning

Predicting water quality in lakes is important because healthy lakes provide diverse ecosystem services and environmental benefits that positively influence our quality of life and the strength of our economy. A combination of urban areas, industries, and agricultural activities have undoubtedly contributed to an increased loading of nutrient pollution into Lake Erie, particularly phosphorus and nitrogen. Today, harmful algal blooms (HABs) in Lake Erie are a frequently occurring seasonal issue. The overgrowth of algae likely triggers hypoxia in the lake as a result of the biological oxygen demand required for breakdown processes by microbes. In this study, we use a suite of physical modeling systems with in-situ measurements of chlorophyll- (chlor-) and dissolved oxygen (DO) in Lake Erie to serve as proxies of HABs and hypoxia. Observations are provided by the Lake Erie Committee Forage Task Group (LEC FTG) and the Great Lakes National Program Office (GLNPO) for the period 2002-2012. Modeling systems involved are the: 1) Weather Research and Forecasting Model (WRF); 2) Variable Infiltration Capacity Model (VIC); 3) Community Multiscale Air Quality Model (CMAQ); and 4) Environmental Policy Integrated Climate Model (EPIC). Meteorological weather variables from WRF, hydrological variables from VIC, nitrogen deposition from CMAQ, and agricultural management practice variables from EPIC for the 11 year period are used to fit a random forest machine learning model to predict concentrations of each response variable. Via random forest regression, the predictive model for chlor- is able to explain 57% of the variance in chlor-. Separate predictive models are developed for DO across various depths in the lake. The model is able to explain above 72% of the variance in lake depths that experience DO concentrations less than 5 mg/L. The importance of explanatory variables is evaluated, and the contribution of each covariate in the model is analyzed with Accumulated Local Effect (ALE) plots to better understand the occurrence of HABs and hypoxia.