Integrative environmental prediction using Bayesian networks

Abstract

The symptoms of coastal and estuarine eutrophication are the result of a number of interacting processes operating at multiple spatial and temporal scales. Thus, models developed to appropriately represent each of these processes are not easily combined into a single predictive model. We suggest that Bayesian networks provide a possible solution to this problem. The graphical structure explicitly represents cause-and-effect assumptions between system variables that may be obscured under other approaches. These assumptions allow the complex causal chain linking management actions to ecological consequences to be factored into an articulated sequence of conditional relationships. Each of these relationships can then be quantified independently using an approach suitable for the type and scale of information available. Probabilistic functions describing the relationships allow key known or expected mechanisms to be represented without the full complexity, or information needs, of highly reductionist models. To demonstrate the application of the approach, we develop a Bayesian network representing eutrophication in the Neuse River estuary, North Carolina from a collection of previously published analyses. Relationships among variables were quantified using a variety of methods, including: process-based models statistically fit to longterm monitoring data, Bayesian hierarchical modeling of cross-system data, multivariate regression modeling of mesocosm experiments, and probability judgments elicited from scientific experts. We use the fully quantified model to generate predictions of ecosystem response to alternative nutrient management strategies.