Geology Reference
In-Depth Information
Ganspoel and Kinderveld catchments over a
three-year period. The data were used to evaluate
the WaTEM/SEDEM model (Van Oost
et al
.,
2000; Van Rompaey
et al
., 2001) in a Monte Carlo-
based analysis where two of the five model param-
eters were varied simultaneously to generate
5000 random parameter sets. This analysis there-
fore improved upon the univariate sensitivity
analyses described above, but might be best
described as a bivariate uncertainty analysis, as
all other parameters were held at 'reasonable' val-
ues. Exported sediment predictions were com-
pared against observations and evaluated using a
simple likelihood function for single events. The
approach demonstrated that uncertainty sur-
rounding predictions was wide when only one
likelihood measure was used to evaluate the
model performance. A further measure was then
introduced, evaluating the model predictions
against deposited sediment, and then the product
of the two likelihood measures was combined.
Not surprisingly, more models were rejected
using the combined likelihood approach, which
was reasonable as the model was being more
robustly tested. However, the authors argued that
this approach “… allowed us to narrow substan-
tially the uncertainty associated with the event
rainfall erosivity parameter”, as can be seen when
comparing Fig. 4.4A with Fig. 4.4B. Although this
may appear to be true, it should be noted that
many combinations of parameters still provided a
poor goodness-of-fit with the same value of rain-
fall erosivity (approximately 0.004) as giving the
best model predictions. In addition, the authors
failed to make the next step, which is to describe
the uncertainty around the single event obser-
vations in terms of uncertainty bounds or con-
fidence limits. In addition, we are only shown
an example of one event and the way in which
uncertainty associated with one parameter may
be constrained, which begs the question 'how
does uncertainty vary between different events?'
Clearly an understanding of how uncertainty in
erosion predictions varies between different
storms would be useful, as would the explicit rep-
resentation of uncertainty bounds around each
model parameter.
A further useful point was made by Van Oost
et al
. (2005) concerning the use of the spatial pat-
tern of erosion in a catchment as evaluation data.
The authors suggested that these 'soft' or 'quali-
tative' data might be useful in model evaluation,
although they did not make any use of this
approach themselves. It is suggested here that
evaluation of erosion models on more than just
plot or catchment outlet data is desirable. In the
same way as in Brazier
et al
. (2000), we tried to
evaluate erosion predictions for 'the correct
hydrological reasons', as it would be useful to see
sediment yields predicted in a way that reflects
the correct spatial patterns of erosion.
Unfortunately, despite the recognition of the
uncertainty associated with model predictions,
many models are still applied 'off-the-shelf' with-
out consideration of the implications of model
uncertainty. There are many examples of such an
approach to erosion modelling. All of the model
analyses reviewed here demonstrate that model
users
should
evaluate model uncertainty, yet it is
clear from the literature that
most
applications of
erosion models do not do this. Just one example
of this problem is illustrated by Mati
et al
. (2006),
who applied the EUROSEM model to the Embori
and Mukogodo catchments, Kenya. EUROSEM
has been shown to make reasonable (although
uncertain) predictions (see above), but here it is
applied to predict erosion without any considera-
tion of this uncertainty. While we must allow
that models are used for different purposes includ-
ing prediction, experimentation to learn about
process representation, to guide mitigation or soil
conservation, or perhaps to assess the impact of
land use change, it is clear that model developers
(as well as model users) must learn from previous
lessons and start to incorporate estimation of
uncertainty into their model applications. This
incorporation of previous research into our appli-
cation and development of contemporary erosion
models is especially important if we expect to
make any progress in developing 'better' erosion
models (see Chapter 5).
Vigiak
et al
. (2006) applied the MMF model
(Morgan
et al
., 1984) to a small catchment in the
West Usambara Mountains (Kwalei catchment,
