Using machine learning to model and predict water clarity in the Great Lakes

Abstract Over the last several decades, multiple environmental issues have led to dramatic changes in the water clarity of the Great Lakes. While many of the key factors are well-known and have direct anthropogenic origins, climatic variability and change can also impact water clarity at various temporal scales, but their influence is less often studied. Building upon a recent examination of the univariate relationships between synoptic-scale weather patterns and water clarity, this research utilizes nonlinear autoregressive models with exogenous input (NARX models) to explore the multivariate climate-to-water clarity relationship. Models trained on the observation period (1997–2016) are extrapolated back to 1979 to reconstruct a daily-scale historical water clarity dataset, and used in a reforecast mode to estimate real-time forecast skill. Of the 20 regions examined, models perform best in Lakes Michigan and Huron, especially in spring and summer. The NARX models perform better than a simple persistence model and a seasonal-trend model in nearly all regions, indicating that climate variability is a contributing factor to fluctuations in water clarity. Further, six of the 20 regions also show promise of useful forecasts to at least 1 week of lead-time, with three of those regions showing skill out to two months of lead time.