Exploiting simulation model results in parameterising a Bayesian network - A case study of dissolved organic carbon in catchment runoff

This work is part of CLIME project (Climate and Lake Impacts in Europe), which assesses climate change effects on lake dynamics. In CLIME, a decision support system (CLIME-DSS) is based on a causal Bayesian network that summarises the most important relationships between climate variables and lake characteristics. A Bayesian network is a probabilistic graphical model, where nodes represent random variables and arcs between the nodes represent conditional dependencies. In a Bayesian network, relationship between the dependent variable and its explanatory variables is described for discrete variables as a conditional probability table (CPT). The aim of this study is to demonstrate how expert knowledge provided by researchers, and results of an environmental simulation model, are exploited in constructing a Bayesian network. A case study addresses the impact of climate change on concentrations of dissolved organic carbon (DOC) in catchment runoff. The environmental simulation model is a DOC model (Jennings and Naden, 2004), which is coupled with the hydrological routine of the Generalized Watershed Loading Function (GWLF) model. The output of the model is the daily stream water DOC concentration and the daily load of DOC entering a lake. The DOC/GWLF model has been calibrated and validated against historical data from three catchments in Europe. A Bayesian network for describing interrelations between the climate and DOC concentrations is constructed on the basis of expert opinions and the structure of the DOC/GWLF simulation model. Those variables that are present both in the DOC model, and in the network structure based on the expert opinions, are included in the final structure of the Bayesian network. One Bayesian network is constructed for each of the three study sites. The GWLF/DOC model was run under a variety of climatic conditions using one set of calibrated parameter values at a time. The meteorological input variables were compiled from results of Regional Climate Models (RCM). Subsequently, the RCM and GWLF/DOC model results were analysed to compute conditional frequency tables for the links between each dependent node and its explanatory variables in the Bayesian network. The procedure of Kokkonen et al. (2005) was utilised to estimate link strength values from the conditional frequency information. The link strength values were optimised against the conditional frequencies determined from the model simulations. Finally, all values in the CPTs were generated using the optimised link strength values. In order to apply the Bayesian networks within the study region, RCM results are utilised for creating distributions of explanatory variables for all computation grid cells in Europe. The distributions are constructed for different scenarios characterising current and future climatic conditions. These spatial data on the climatic variables together with the Bayesian networks allow the CLIME-DSS users to study the predicted climate change effects across Europe. Three Bayesian networks were applied to predict how decomposition and summertime DOC concentrations change in the future in Lough Leane in Ireland. The application revealed that variability of the predicted annual decompositions and summer DOC concentrations was very different between the three Bayesian networks. The predicted direction of change in DOC concentrations, however, from the control scenario to the future climate scenario was same for all three Bayesian network parameterisations. The model application demonstrates how Bayesian networks can be used as diagnostics for assessing the conformity of model regionalisation.