Better use of prior information in the calibration of river system models

The Australian Water Resources Assessment (AWRA) modelling framework is jointly developed by the Bureau of Meteorology (BOM) and the CSIRO to support the production of the National Water Ressources Assessment and the annual National Water Accounts (BOM, 2009). In this context, AWRA models are run over the recent years in hindcast mode to provide estimates of certain terms of the water accounts. This paper is the first step towards a parameterisation strategy for the river system component of AWRA. This component will be used in the accounts to estimate ungauged lateral inflows to the river, transmission losses and storage in the river channel. The model considered in this paper is a simplified river system model coupling a rainfall-runoff, a routing model and a transmission loss model. Management components (e.g. diversions, storage in reservoirs) have not been included in the model based on the assumption that they could be estimated from measured data collected by the BOM. The calibration of this model against observed data across the Australian continent could be problematic with many areas remaining poorly gauged. The traditional approach based on the minimization of an objective function computed from local data(e.g. the Nash-Sutcliffe efficiency) may be challenged in data scarce environments where only a limited number of data values are available, leading to a loosely constrained optimisation and poorly identifiable parameter sets. To investigate this question, a simplified daily river system model combining a rainfall-runoff and a routing model was applied on 55 river reaches across Australia. A Bayesian inference scheme was used to derive the posterior distribution of the model parameters. Different priors were used to shift from noninformative to informative priors. The predictive uncertainty obtained with different priors was finally compared using deterministic (root mean squared error, root mean squared error on log transformed flow) and probabilistic scores (continuous rank probability score and reliability). The chief conclusion from the paper is the importance of the use of a-priori information in the calibration of hydrological models. Calibration data contain error which can introduce bias in the model calibration. An inference scheme constrained with prior knowledge can potentially mitigate this problem. In our study, it lead to parameters with better extrapolation capacity for half of the catchments considered. However, the model performance was degraded on the other half. Further work is required to determine the reasons for this counter-performance and better formulate the prior for these catchments.