A multivariate Chain-Bernoulli-based prediction model for cyanobacteria algal blooms at multiple stations in South Korea

Predicting the occurrence of algal blooms is of great importance in managing water quality. Moreover, the demand for predictive models, which are essential tools for understanding the drivers of algal blooms, is increasing with global warming. However, modeling cyanobacteria dynamics is a challenging task. We developed a multivariate Chain-Bernoulli-based prediction model to effectively forecast the monthly sequences of algal blooms considering hydro-environmental predictors (water temperature, total phosphorus, total nitrogen, and water velocity) at a network of stations. The proposed model effectively predicts the risk of harmful algal blooms, according to performance measures based on categorical metrics of a contingency table. More specifically, the model performance assessed by the LOO cross-validation and the skill score for the POD and CSI during the calibration period was over 0.8; FAR and MR were less than 0.15. We also explore the relationship between hydro-environmental predictors and algal blooms (based on cyanobacteria cell count) to understand the dynamics of algal blooms and the relative contribution of each potential predictor. A support vector machine is applied to delineate a plane separating the presence and absence of algal bloom occurrences determined by stochastic simulations using different combinations of predictors. The multivariate Chain-Bernoulli-based prediction model proposed here offers effective, scenario-based, and strategic options and remedies (e.g., controlling the governing environmental predictors) to relieve or reduce increases in cyanobacteria concentration and enable the development of water quality management and planning in river systems.