How good are Bayesian belief networks for environmental management? A test with data from an agricultural river catchment

Abstract

Summary The ecological health of rivers worldwide continues to decline despite increasing effort and investment in river science and management. Bayesian belief networks (BBNs) are increasingly being used as a mechanism for decision‐making in river management because they provide a simple visual framework to explore different management scenarios for the multiple stressors that impact rivers. However, most applications of BBN modelling to resource management use expert knowledge and/or limited real data, and fail to accurately assess the ability of the model to make predictions. We developed a BBN to model ecological condition in a New Zealand river using field/GIS data (from multiple rivers), rather than expert opinion, and assessed its predictive ability on an independent data set. The developed BBN performed moderately better than a number of other modelling techniques (e.g. artificial neural networks, classification trees, random forest, logistic regression), although model construction was more time consuming. Thus, the predictive ability of BBNs is (in this case at least) on a par with other modelling methods but the approach is distinctly better for its ability to visually present the data linkages, issues and potential outcomes of management options in real time. The BBN suggested management of habitat quality, such as riparian planting, along with the current management focus on limiting nutrient leaching from agricultural land may be most effective in improving ecological condition. BBNs can be a powerful and accurate method of effectively portraying the multiple interacting drivers of environmental condition in an easily understood manner. However, most BBN applications fail to appropriately test the model fit prior to use. We believe this lack of testing may seriously undermine their long‐term effectiveness in resource management and recommend that BBNs should be used in conjunction with some measure of uncertainty about model predictions. We have demonstrated this for a BBN of ecological condition in a New Zealand river, shown that model fit is better than that for other modelling techniques, and that improving habitat would be equally effective to reducing nutrients to improve ecological condition.