Developing models of an environmental process and using these models to generate predictions can be invaluable in understanding natural phenomena and in supporting decision making. However, it is important to be aware that inevitably, predictions made using environmental models are subject to uncertainty. This uncertainty arises from a variety of sources: Knowledge of environmental conditions, the internal state and the external forcings of a given system will never be perfect, and the mathematical representations of the system and its processes are always simplifications.
In order to gain confidence in an environmental model and its predictions, the associated uncertainty needs to be quantified. This is not trivial, since it requires an understanding of how the model structure, the model parameters, the input and the output differ from their real counterparts. Some of these errors are quantified more easily than others: It is impossible to know the true hydraulic parameters at every point of an aquifer, but areal precipitation and stream discharge can be measured reasonably accurately. Usually, the different uncertainty sources are therefore lumped together: For instance, some of the deviations of the predictions of a groundwater flow model from observations may be ascribed to measurement inaccuracy, the rest must be due to model structure uncertainty and parameter uncertainty.
A more detailed breakdown of the different uncertainty sources would make it possible to gain insight into model inadequacies and to guide efforts for improving a model. By quantifying the uncertainty sources and investigating how well they can be reduced, they can be prioritised and targeted more efficiently.
Such a breakdown is only possible if the modelled system is perfectly known. Since this is never the case for a real system, in this study a very complex and detailed model will be used as a "virtual reality" that will then in turn be modelled using simpler approaches. As the complex model and all the inputs and outputs are then perfectly known, this approach will allow for the assessment of the impact of different model simplifications, artificial measurement errors and prediction types on the prediction uncertainty.
A complex model of a well-instrumented river bend will be developed, which will then be used to generate synthetic observations from which simpler models are built. A detailed uncertainty analysis of the predictions made using these simple models will then be performed, and the influence of different simplifications and measurement errors on the prediction uncertainty will be investigated.
