Bayesian Networks (BNs) are increasingly being used as decision support tools to aid the management of the complex and uncertain domains of natural systems. They are particularly useful for addressing problems of natural resource management by complex data analysis and incorporation of expert knowledge. BNs are useful for clearly articulating both the assumptions and evidence behind the understanding of a problem, and approaches for managing a problem. For example they can effectively articulate the cause effect
relationships between human interventions and ecosystem functioning, which is a major difficulty faced by planners and environment managers. The flexible architecture and graphical representation make BNs attractive tools for integrated modelling. The robust statistical basis of BNs provides a mathematically coherent framework for model development, and explicitly represents the uncertainties in model predictions. However, there are also a number of challenges in their use. Examples include i) the need to express conditional probabilities in discrete form for analytical solution, which adds another layer of uncertainty; ii) belief updating in very large Bayesian networks; iii) difficulties associated with knowledge elicitation such as the range of questions to be answered by experts, especially for large networks; iv) the inability to incorporate feedback loops and v) inconsistency associated with incomplete training data. In this paper we discuss some of the key research problems associated with the use of BNs as decision-support tools for environmental management. We provide some real-life examples from a current project (Macro Ecological Model) dealing with the development of a BN-based decision support tool for Integrated Catchment Management to illustrate these challenges. We also discuss the pros and cons of some existing solutions. For example, belief updating in very large BNs cannot be effectively addressed by exact methods (NP hard problem), therefore approximate inference schemes may often be the only computationally feasible alternative. We will also discuss the discretisation problem for continuous variables, solutions to the problem of missing data, and the implementation of a knowledge elicitation framework.