Coupled Bayesian Networks and recursive partitioning method for wetland ecological modelling

Bayesian Networks (BNs) are increasingly recognised as a useful tool for ecological modelling due to their ability to incorporate a broad range of data types and explicit representation of uncertainty through the use of probabilities. They allow considerable flexibility with respect to the detail and focus of the models allowing conceptually simple habitat condition models or more complex mechanistic representation of ecological response. However, BN outputs are sensitive to the model structure, particularly the selection and linking of variables and how the states are defined for each variable. In this paper, we use recursive partitioning to inform the configuration of the structure of BNs, used for ecological response modelling. The Narran Lakes Ecosystem is a nationally important floodplain wetland complex considered at risk. In 2005, a mesocosm experiment was conducted as part of the Narran Lake Ecosystem Project to investigate the seedbanks of ephemeral plant communities in response to a range of flood scenarios. From that data and existing literature, a conceptual model describing the major hydrological influences on the ephemeral herbfields was generated. BN ecological response models were then constructed using these data. The models were developed using information generated from recursive partitioning (regression tree and random forests) and BN learning approaches (Figure 1). The models have been incorporated into an environmental flow Decision Support System (DSS), IBIS DSS. The BN model is linked to a hydrological model of the Narran Lakes allowing the modelling of ecological response to flow series. Recursive partitioning analyses of biophysical and ecological data informed the development of BNs in two ways. Firstly, random forests analyses were used to identify important predictor variables: those variables that, statistically, best explain the ecological responses. Secondly, thresholds were identified using decision tree analysis to reduce subjectivity in the discretisation of variables in the BNs. Using the coupled BNs and recursive partitioning method improved the rigour and certainties associated with state discretisation, and allowed us to refine the choice of model variables, while maintaining the advantages associated with applying BNs within the DSS.(Figure presented).