GSA Student working group: Bayesian statistics

Cyanobacterial blooms alter lake ecosystems through production of surface scums, toxins, and taste and odor issues. Near-term forecasts of blooms would help lake managers preemptively manage water quality and provide advance warning to the public of potential recreational water closures. Importantly, forecasts should include uncertainty estimates to provide a robust assessment of bloom forecast confidence. Our project, a collaboration among members of the Ecological Forecasting Initiative (EFI), the Global Lake Ecological Observatory Network (GLEON), and the Lake Sunapee Protective Association (LSPA), aims to identify the dominant sources of uncertainty to near-term forecasts of cyanobacterial densities in oligotrophic Lake Sunapee, NH, USA, which experiences blooms of the cyanobacterium Gloeotrichia echinulata. We partitioned uncertainty among model-process error, initial conditions error, observational error, and parameter uncertainty for several Bayesian state-space models predicting cyanobacterial density. Candidate models were calibrated and validated using subsets of a 14-year, ongoing time series of weekly G. echinulata densities collected from May to October. We evaluated whether candidate forecast models including wind direction and speed, thermal stratification stability, and water temperature covariates improved one-week-ahead predictions of G. echinulata compared to a random walk model. Preliminary uncertainty partitioning of the predictive intervals for our models revealed that initial conditions and process errors were the primary drivers of total prediction uncertainty. Our work suggests that more intensive G. echinulata observations and exploration of further candidate models could improve near-term bloom predictions and advance understanding of how best to forecast cyanobacterial blooms.