Nonstationarity, inconsistencies,heterogeneity, uniqueness of place, and surprises as epistemic uncertainties in hydrological modelling a framework for good practice
Hepple Lecture Theatre, School of Geography
Distinguished Professor Keith Beven, a former graduate of Geographical Sciences, University of Bristol, is receiving one of this years honorary degrees. Keith will be providing this seminar on some of his recent hydrology research and thinking.
Epistemic uncertainties are endemic in hydrological science and practice. This creates important issues for the application of statistical methods in hydrological analysis because treating epistemic uncertainties as if they were aleatory will in general lead to overconfidence in inference. The mathematical formalism of likelihood theory has been attractive to hydrologists but unthinking use might actually be misleading (even in some cases where the assumptions of a probabilistic error model might seem to be reasonably valid). Other methods might therefore be necessary but inevitably seem to invoke criticisms of subjectivity. Recent work based on trying to assess some limits of acceptability for models will be described while allowing for inconsistencies, nonstationarities and other effects of lack of knowledge. However, given the many different methods of trying to assess uncertainties then there is also a need to communicate the meaning of the outcomes to potential users (there are many examples in hydrology, hydraulics, climate change and other domains).
It is suggested that one way of being more explicit about the meaning of uncertainties is to associate each type of application with a condition tree of assumptions that need to be made in producing an estimate of uncertainty. The condition tree then provides a basis for discussion and communication of assumptions about uncertainties with users. Agreement of assumptions (albeit generally at some institutional level) will provide some buy-in on the part of users, and a basis for commissioning of future studies. Even in some relatively well-defined problems, such as mapping flood risk, such a condition tree can be rather extensive, but by making each step in the tree explicit then an audit trail is established for future reference. This can act to provide focus in the exercise of agreeing more realistic assumptions.