You are here:
Computational Approaches to Predict Indices of Cyanobacteria Toxicity
Citation:
Kreakie, B., F. Nojavan, Jeff Hollister, AND Bryan Milstead. Computational Approaches to Predict Indices of Cyanobacteria Toxicity. National Water Quality Monitoring Council (NWQMC), Tampa, FL, May 02 - 06, 2016.
Impact/Purpose:
This is an abstract for an invited talk at the National Water Quality Monitoring Council (NWQMC) meeting in Tampa, Florida. This talk will be in section titled: Freshwater Harmful Algal Blooms: Advances in Analytical Tools, Assessments, and Prediction.
Description:
As nutrient inputs increase, productivity increases and lakes transition from low trophic state (e.g. oligotrophic) to higher trophic states (e.g. eutrophic). These broad trophic state classifications are good predictors of ecosystem health and the potential for ecosystem services (e.g. recreation, aesthetics, and fisheries). Additionally, some ecosystem disservices, such as cyanobacteria blooms, are also associated with increased nutrient inputs. Thus, trophic state can be used as a proxy for cyanobacteria bloom risk. To explore this idea, we construct two random forest models of trophic state (as determined by chlorophyll a concentration). First we define an “All Variable” model that estimates trophic state with both in situ and universally available data, and then we reduce this to a “GIS Only” model that uses only the universally available data. The “All Variables” model had an RMSE of 0.09 and R2 of 0.8; whereas, the “GIS Only” model was 0.22 and 0.48 for RMSE and R2, respectively. Examining the “GIS Only” model (i.e., the model that has broadest applicability) we see that in spite of lower overall accuracy, it still has better than even odds (i.e., prediction probability is > 50%) of being correct in more than 1091 of the 1138 lakes included in this model. The “GIS Only” model has tremendous potential for exploring spatial trends at the national level since the datasets required to parameterize the model are ubiquitous. To further extend this analysis, we used the results of a random forest modeling as a means of variable selection from which we developed a Bayesian multilevel model of microcystin (a common cyanobacteria toxin with human and animal health implications) concentrations. The Bayesian multilevel model allows us to ascertain the key drivers at the continental scale while still incorporating local influences at the ecoregion scale. This reduces uncertainty and improves risk assessment at the eco-regional or state scale. Our preliminary results show associations between microcystin and turbidity, total nutrients, and N:P ratios. Upon release of a comparable 2012 NLA dataset, we will apply Bayesian updating in an adaptive management framework. The results will help develop management strategies to reduce microcystin impacts and improve lake quality.