Geoscience Reference
In-Depth Information
a hydrological survey, then appropriate process representations can be used for different types of source
areas (e.g. Uhlenbrook and Leibendgut, 2002; Scherrer and Naef, 2003).
The disadvantage of this approach is the fundamental problem of calibration of the parameters. As
with any distributed model there may be hundreds or thousands of parameter values that must be defined
before a simulation can be run. Databases of parameter values are available for such models to provide
a working first estimate, but there is no guarantee that they will be the effective values required for
particular places and at a particular discretisation scale to give good predictions (Beven, 2000). Indeed,
there is no guarantee that the conceptual nature of the process representations will be able to give a
good simulation of the surface and subsurface runoff generation in a particular application. Calibration
of parameter values will help, but clearly such models pose great difficulties for calibration given the
very large number of parameters of different types that could be changed. It is possible to reduce the
dimensionality of the calibration problem, for example, by fixing the relative magnitudes of the spatial
pattern of the initial estimates and adjusting the whole field of values by a calibration multiplier. There
is no firm assurance, however, that this will provide the right runoff predictions in the right places and
for the right reasons, even if the predictions of the hydrograph are improved by calibration.
Others have claimed generality for their process representations, suggesting even that they might be
more generally representative than those of the continuum distributed models of Chapter 5. Vinogradov
et al.
(2010) make this claim on the basis that the model has produced good results over a wide range of
scales using parameter values estimated only on the basis of soil and vegetation characteristics. They are
correct when they assert that any model that claims generality should be widely tested in this way, but it is
difficult to accept a model as general that ignores the effects of topographic convergence and divergence
on surface and subsurface flows and uses point scale parameter values to predict runoff production over
each HRU as if it was homogeneous (though in this respect their model is similar to many others). But
this takes us back to the perceptual model of the individual hydrologist as to what is important: on large
catchments in the steppes of Russia, topographic convergence and divergence might seem much less
important than in small humid temperate catchments in the uplands of the UK.
Does all this matter if good predictions of discharge can be achieved over a wide range of catchments,
especially if good predictions can be achieved while avoiding the problems intrinsic to catchment-
specific model calibration with such models? This class of models is, for the most part, empirical in
nature (even though the term “physically based” is now seen quite often as a descriptor for SWAT,
including in the recent review by Gassman
et al.
, 2010). There is then an
instrumentalist
argument that
empirical success in prediction is an adequate justification. But, what then should be considered as a
hydrological success in this respect, given that we know that there are limitations of input data, model
structures and the observations with which a model might be compared (e.g. Harmel
et al.
2006). This
is a question that we return to several times in the chapters that follow in considering model calibration
and evaluation, predicting future change in catchment runoff, considering what might be expected of the
next generation of rainfall-runoff models, predicting the response of ungauged catchments, and models
of everywhere.
6.8 Key Points from Chapter 6
•
In considering the variability of hydrological responses within a catchment area, it may be difficult,
and may not be necessary, to simulate those responses in a fully distributed way if a simpler way can
be found of representing the distribution of responses in the catchment. This implies finding a way to
define whether different points in a catchment act in hydrologically similar ways.
•
The Probability Distributed Moisture (PDM) model does this in terms of a purely statistical repre-
sentation of conceptual storage elements. It is an attempt to represent the variability in the catchment
but does not allow any mapping of the predictions back into the catchment space for comparison with
