Geology Reference
In-Depth Information
site, Mississippi, as part of the USDA/ARS USLE
database (Zhang, personal communication,
1999). The GLUE approach was used to explore
parameter and model structural uncertainty
against 12 years of runoff and erosion data at the
two sites by running the model in continuous
simulation mode and varying the 16 most sensi-
tive model parameters simultaneously. No
a
priori
assumptions were made about parameter
distributions, so parameter values were ran-
domly sampled from uniform distributions
with initially wide minima and maxima.
Approximately 5.7 million model runs were car-
ried out in order to explore, as fully as possible,
the 16-dimensional model parameter space. All
other parameters within the model were held
constant. Each model prediction was evaluated
against observed data using a combined-likeli-
hood function that assessed the model perform-
ance in terms of its ability to predict both runoff
and erosion, or to make good erosion predictions
for the correct hydrological reasons. Behavioural
parameter sets or model structures that satisfied
predetermined criteria for goodness of fit were
then retained to describe the uncertainty sur-
rounding model predictions. It should be noted
here that the predetermination of what is an
acceptable model prediction is clearly subjec-
tive, although every effort was made in this work
to ensure that the criteria were neither too strict,
resulting in the rejection of all models as non-
behavioural, nor too relaxed, resulting in the
acceptance of all models as behavioural, which
is clearly not the case.
Results show, for both the UK and US sites,
that the model consistently overpredicted small
events and underpredicted large events, leading
to the overestimation of observed data during
years when little erosion occurred and the under-
estimation of observed data during years when
erosion was significant. Figure 4.7 illustrates the
modelled erosion predictions for both sites
(shown here as minimum, maximum and median
predictions to describe uncertainty bounds).
Predictions were highly uncertain. Although
most years of observations were captured within
the model uncertainty bounds, certain years,
notably the dry year of 1963 on the Holly Springs
plots, were overpredicted by a factor of seven, by
even the minimum uncertainty bounds.
Brazier
et al
. (2000) also showed that model
predictions that are produced through the ran-
dom generation of parameter values are better
than those produced by either estimation or
optimization techniques presented by previous
authors (Zhang
et al
., 1996), and a common
approach when parameterizing erosion models.
This is an important finding, as it underlines
how poorly the model structure is able to
describe the system that is being modelled. It
also questions the approach of 'estimating'
parameter values in general, without consider-
ing the uncertainty that these estimations
implicitly incorporate into model predictions. It
is clear that a better approach would be to quan-
tify parameter uncertainty, so as to be explicit
about how well the effective parameter values of
erosion models can be estimated and thus what
levels of uncertainty we might expect, given our
current understanding and representation of the
processes of erosion in models.
Finally, in common with all of the other
approaches discussed herein, this work shows
that model predictions are far from precise, and
that even with the 'new generation' of erosion
prediction technology that WEPP was heralded to
be, much work is still to be done to provide us
with reliable or certain predictions of erosion.
Qualification of model predictions, through
uncertainty analysis, must not be seen as a nega-
tive process. Far from it, all of the above
approaches allow us varying degrees of insight
into the performance of erosion models, which
can direct model developers towards areas of
process-understanding that need to be improved
and better represented within soil erosion mod-
els. Analysis of model uncertainty is also a vital
tool in the toolkit of model users. Decision-
makers or land managers who apply erosion
models to foresee the effects of management deci-
sions must do so with some consideration of
uncertainty. Ideally this would be done within a
framework such as GLUE (see Chapter 5 for a fur-
ther example), or at the very least with some
