Aarhus Universitets segl

No. 469: Modelling chlorophyll-a concentrations in Danish coastal waters using a Bayesian modelling approach

Shetty N, Christensen JPA, Damgaard C & Timmermann K. 2021. Modelling chlorophyll-a concentrations in Danish coastal waters using a Bayesian modelling approach. Documentation report. Aarhus University, DCE – Danish Centre for Environment and Energy, 62 pp. Scientific Report No. 469.


 Bayesian hierarchical models were developed to predict chlorophyll-a concentration in Danish water-bodies, as a function of land-based nitrogen and phosphorus loadings, physicochemical and climatic predictors. The objective of the model development was to support the Danish implementation of the Water Framework Directive (WFD) by providing tools applicable for estimating chlorophyll-a reference conditions and maximum allowable nutrient input (MAI) to the Danish coastal waters covered by the water framework directive.

We developed single station Bayesian models for 46 Danish water quality monitoring stations representing 43 water-bodies, as well as an overall group model that included data from 42 monitoring stations representing 39 water bodies.  In the group model, nutrient loadings, water temperature, salinity, and water column stability were the best predictors of chlorophyll-a concentration in the 39 water bodies represented by the model. For the single station model, we found that nutrient loading was the most abundant predictor for chlorophyll-a concentration, followed by temperature and salinity. As expected, we found a positive slope coefficient between nutrient loadings and chlorophyll-a concentration in all but two stations.

Model evaluation plots and performance statistics revealed that most of the models could capture the levels and year-to-year variation in chlorophyll-a concentrations reasonable well indicating that the models can be used to produce reliable predictions of chlorophyll-a concentrations in Danish coastal waters. Not surprisingly, model performance for single station models was, in general, better than for the group model, but the risk of “over parameterization” is also higher for single station models.