BAYESIAN NETWORK MODELS FOR ENVIRONMENTAL FLOW DECISION MAKING IN THE DALY RIVER, NORTHERN TERRITORY, AUSTRALIA TERENCE U. CHAN, a * BARRY T. HART, a MARK J. KENNARD, b,c BRADLEY J. PUSEY, b,c WILL SHENTON, a MICHAEL M. DOUGLAS, c,d ERIC VALENTINE c,d and SANDEEP PATEL c,d a Water Studies Centre, Monash University, Clayton, Australia b Australian Rivers Institute, Griffith University, Brisbane, Australia c Tropical Rivers and Coastal Knowledge (TRaCK), Commonwealth Environmental Research Facility, Australia d School for Environmental Research, Charles Darwin University, Darwin, Australia ABSTRACT This paper reports the development and application of two Bayesian Network models to assist decision making on the environmental flows required to maintain the ecological health of the Daly River (Northern Territory, Australia). Currently, the Daly River is unregulated, with only a small volume of water extracted annually for agriculture. However, there is considerable pressure for further agricultural development in the catchment, particularly with demand for extra water extraction during the dry season (May–November). The abundances of two fish species—barramundi (Lates calcarifer) and sooty grunter (Hephaestus fuliginosus)—were chosen as the ecological endpoints for the models, which linked dry season flows to key aspects of the biology of each species. Where available, data were used to define flow–fish habitat relationships, but most of the relationships were defined by expert opinion because of a lack of quantified ecological knowledge. Recent field data on fish abundances were used to validate the models and gave prediction errors of 20–30%. The barramundi model indicated that the adult sub-population was key to overall fish abundance, with this sub-population particularly impacted by the timing of abstraction (early vs. late dry season). The sooty grunter model indicated that the juvenile sub-population dominated the overall abundance and that this was primarily due to the amount of hydraulically suitable riffle habitat. If current extraction entitlements were fully utilized, the models showed there would be significant impacts on the populations of these two fish species, with the probability of unacceptable abundances increasing to 43% from 25% for sooty grunter and from 36% for barramundi under natural conditions. Copyright # 2010 John Wiley & Sons, Ltd. key words: environmental flows; barramundi; sooty grunter; Bayesian network models; decision support; Daly River Received 16 July 2010; Accepted 29 July 2010 INTRODUCTION Numerous methods are now available for assessing the environmental flow regime required to sustain or restore the ecological integrity for a particular river system (Tharme, 2003). These environmental flow methods generally attempt to achieve a flow regime similar to that which would have occurred naturally (Poff et al., 1997; Hillman et al., 2003). Reviews of the range of available methods by Arthington and Pusey (2003) and Tharme (2003) show clearly that these methods rely heavily on expert opinion, often lack the required hydrological data (simulated daily natural flows), and generally lack transparency on how various flow components are related to ecological outcomes (Hart and Pollino, 2009). The work reported here, and in a companion paper, Shenton et al. (2010), builds on the review by Hart and Pollino (2009) that identified the potential for Bayesian Network (BN) models to assist decision makers in assessing ecological risks of different water management scenarios and establishing optimum environmental flow regimes. BN models are increasingly being used in natural resources management (Ellison, 1996; Batchelor and Cain, 1999; Varis and Kuikka, 1999; Nyberg et al., 2006; Castelletti and Soncini-Sessa, 2007; McCann et al., 2007; Ticehurst et al., 2007; Wang et al., 2009) and more recently, specifically for determining environmental flow allocations (Hart and Pollino, 2009; Stewart-Koster et al., 2010). BN models have a number of properties that make them particularly useful for ecological and environmental management applications. In particular, they show cause–effect relation- ships directly through a simple causal graphical structure, but are also easily constructed, extended and modified; they have a natural way to handle missing data; they explicitly incorporate uncertainty in relationships; they are an accessible and intuitive modelling approach; they can show good predictive accuracy even with small sample sizes; they allow the conditional probabilities between variables to be constructed using either observed data, other models, or expert knowledge; they can easily be updated as new data *Correspondence to: Terence U. Chan, Water Studies Centre, Monash University, Clayton, Australia. E-mail: terry.chan@sci.monash.edu.au Copyright # 2010 John Wiley & Sons, Ltd. RIVER RESEARCH AND APPLICATIONS (wileyonlinelibrary.com) DOI: 10.1002/rra.1456 River Res. Applic. 28: – (2012) Published online in Wiley Online Library 2 2 September 2010 83 301 4