Geoscience Reference
In-Depth Information
provide useful information and evidence in complex environmental systems is one that still needs to
clarified (e.g. Beven, 2002b; Harris and Heathwaite, 2005; Harris, 2007).
It is difficult to be sure about predictions of water flows in catchments, and water management problems
involve more than just water flows. In satisfying the requirements of the EU Water Framework Directive,
chemical and ecological processes also need to be considered. These additional dimensions have their
own (even greater) uncertainties, as do the economic costs and impacts of management decisions. These
are indeed complex, multi-faceted problems but I believe that modelling still has an important role to play
in structuring the nature of a problem and in assessing the effects of a local change on the wider catchment
environment. This again is where the learning process within the models of everywhere framework will
serve to bring together scientists and stakeholders in solving management issues.
12.7 Models of Everywhere and Information
A learning process requires information. Post-normal science allows that the relevant information might
be purely qualitative, based on the practical experience of both scientists and stakeholders. However,
for the scientist, there is also an interest in what data might be of value in hypothesis testing and con-
straining the uncertainty in predictions. It was noted in Section 7.18 that there is little guidance in the
literature about the value of different types of data and normally, of course, we make use of whatever data
we can get hold of (while remembering to be wary of disinformation resulting from epistemic errors,
see Section 7.17).
As noted in the Preface, the last paragraph of the first edition of this topic read:
The future in rainfall-runoff modelling is therefore one of uncertainty: but this then implies
a further question as to how best to constrain that uncertainty. The obvious answer is by
conditioning on data, making special measurements where time, money and the importance
of a particular application allow. It is entirely appropriate that this introduction to available
rainfall-runoff modelling techniques should end with this focus on the value of field data.
Ten years on, we know a little bit more about the interaction between model uncertainty and measure-
ments. It has been the subject of a number of different workshops and sessions in major conferences. I
have suggested that one of the most productive frameworks for such an interaction might be in testing
quantitative hydrological models as hypotheses about how a hydrological system is functioning. Field
data, however, are not often collected in experimental catchments with this aim in mind, but rather with the
aim of extending our perceptual understanding about how catchments work. The quantitative modelling
often comes later, as an approximation.
The models of everywhere concept can potentially change this way of doing things, since the effect of
collecting additional data and information should be cumulative in the learning process. It is also possible
to be pro-active about gathering additional information in specific places. We would want that process
to be cost effective, which implies learning more about the value of different types of data collection.
However, the availability of a model of everywhere allows an experimental design to be tested and costed.
This is a form of what in statistical experimental design is sometimes called
pre-posterior prior analysis
.
It involves using the current model of the system as a prior estimator of the different observations that
might be collected. The benefit of actually collecting those values in constraining uncertainty in the model
predictions can then be assessed, allowing for the potential observation uncertainties. The cost of different
observation strategies can also be assessed, and the cost-benefit evaluated prior to actually making the
measurements. A similar process could also be used in choosing critical observations that might best
differentiate between models that differ in their predictions. Such a critical measurement might reveal
