As quantitative understanding of flow-ecology relationships improves, statistical and other models are being used to assess the environmental outcomes of changed flow regimes. However, whether these models can be used to predict outcomes under novel flow conditions has not been well tested. This is important because managers are expected to consider hypothetical flow futures and plan water resource use accordingly.
We used data-driven statistical models of responses to flow variability in the Goulburn River, south-east Australia, to make predictions of i) outcomes from different scenarios for delivering inter-valley transfers (IVTs) for irrigation over summer, and ii) the hypothetical restoration of natural, unregulated flows to the river. For case ii) we also used expert-based mechanistic predictive models, previously developed for an environmental flow assessment, to predict the effect of natural flows.
Against expectations, predictions for bank vegetation and bank condition for different IVT scenarios had slightly better outcomes than for the ‘base case’ where only environmental flows were delivered. Similarly, the statistical models showed no strong effects of natural flows. Conversely, the expert-based models found larger effects; fish outcomes were favoured by natural flows, while bank stability and littoral vegetation declined.
The results highlight the importance of considering the original purpose of a model before applying it elsewhere. While model-based predictions are potentially useful for informing flow-management decisions, we need to be realistic regarding the suitability of existing models for this purpose. Non-mechanistic statistical models, like the ones used here, may not be useful for assessing hypothetical conditions – i.e., models designed to assess past outcomes may not be useful for predicting future states. Conversely, the expert-based models were explicitly developed to make predictions regarding different environmental flow scenarios, and so proved better for extrapolation. Either way, new models may often be required to make robust predictions.