Bayesian Networks and Social Objectives: A Canoeing Case Study

Much of the value of a river lies in the sociocultural values people attach to it. Though such 'social objectives' should form a fundamental part of integrated catchment management, they have typically been neglected due to the qualitative nature of many of the variables that predict them. One such social objective in the Don catchment, UK, is the management goal of maximising the recreational quality of rivers for canoeing within the constraints imposed by other management aims such as reducing flooding. Recreational quality is impacted by the modification of river weirs, an important management intervention in the catchment for which there are multiple potential options. An integrated catchment management decision support tool must predict the impact of the different modification options, raising the question; how to deal with the complex, uncertain, and subjective variables that determine a canoeist’s judgement of river quality? To tackle this issue, we employed a Bayesian Network (BN), which uses probability to describe relationships between variables. As probability can incorporate expert estimates, this enabled us to harness the knowledge of canoeist stakeholders. In this paper we discuss the experiences of building a BN with the collaboration of canoeists to predict how weir modification affects river recreational quality, and comment on implications for the overall utility of the approach for modelling social objectives. We conclude BNs are indeed suitable for modelling social objectives; probabilities capturing the uncertainty and subjectivity of variables such as ‘weir danger’. However the approach also has clear limitations. Though the canoeists found most parts of BN construction engaging, this was not so when eliciting judgements of probability, for which the use of questionnaires made it an abstract process. A further issue was that the number of questions needed in the probability elicitation stage increased dramatically with BN complexity. Interpolating probabilities from a limited number of questions is a partial solution, but even so the high number of questions still necessary was a barrier to completing this step. Ultimately this will constrain the potential complexity of BNs that require expert knowledge to define probabilistic relationships.