PS 13-114 - Ecological forecasting of cyanobacterial densities in low-nutrient lakes: Testing the role of incorporating multiple sites and spatial processes to reduce uncertainty and improve predictions

Background/Question/Methods

Cyanobacterial densities in lakes are a growing concern due to their ability to produce toxins and surface scums, but their abundance is often highly variable in space and time. Near-term ecological forecasts that include partitioned uncertainty estimates are advantageous for water resource managers who benefit from advanced warning of potential blooms to mitigate the negative impacts to water quality, fisheries, and recreation on lakes. Using a long-term dataset of Gloeotrichia echinulata densities at four locations in a low-nutrient lake, we fit Bayesian state-space models to test the effect of incorporating multiple sites and spatial processes on cyanobacterial density predictions and uncertainty estimates. Bayesian state-space models were fit with weekly estimates of G. echinulata data from 2009-2014 and results were compared to data from 2015-2016 at one-week and four-week forecast horizons. Prior research has shown that water temperature and wind are important predictors of cyanobacterial density at a single site, but the best models often still missed rare, large bloom events. By sub-dividing the cyanobacteria data into different growth categories, we tested whether the environmental covariates that are significant drivers of low-density growth differ from those for bloom events. We also identified the dominant sources of uncertainty for each model.

Results/Conclusions

Across the four sites, G. echinulata densities varied seasonally and spatially with larger bloom events primarily occurring at two sites. Water temperature and growing degree days were important predictors of cyanobacterial densities at all four sites in 2015 and 2016. In addition, a minimum water temperature threshold may be important for initiating the start of the growth phase while growing degree days may be more important for predicting bloom events. We found that the dominant sources of uncertainty were the model process error and initial conditions at the one-week forecast horizon, while errors in driver data were a larger proportion of the variance than initial conditions at the four-week horizon. By including data from all sites and spatial processes such as wind speed and direction, we will test whether a lake-wide model will improve predictions for both low-density growth and rare, large bloom events. Partitioning the forecast uncertainty will be useful for determining whether incorporating data from multiple sites and differentiating the growth patterns reduces the model process error. This could help water quality managers decide whether limited resources would be better devoted to increasing spatial or temporal frequency of sampling to improve predictions of cyanobacterial densities in low-nutrient lakes.