Geology Reference
In-Depth Information
not only important to reflect upon misapplica-
tions of models, but also on (mis)conceptions that
may direct which models we will develop in the
future and how we will test them. Here, we dis-
cuss a number of such misconceptions that we
consider to be important. We explore to what
extent commonly accepted conceptions about
erosion models are indeed tenable, considering
the available empirical evidence and what the
possible implications of these conceptions are for
future model development and applications.
and dynamic, process-based models has been
compared. In most cases these publications focus
on predictions of soil erosion rates only, i.e. the
amount of soil leaving a pre-defined area subject
to erosion, divided by the surface area considered
and the amount of time over which measure-
ments took place. Most often, data are collected
on areas subject to net erosion over the whole
surface, so the plots rarely contain zones which
are characterized by net deposition.
The outcome of these comparison exercises is
somewhat surprising. Overall, average annual and
even annual values of soil erosion rates as meas-
ured on erosion plots are not better predicted by
using a dynamic, process-based model rather than
a statistical approach. As already mentioned,
Tiwari
et al
. (2000) found that, using the data used
to develop the USLE, the overall performance of
the WEPP model was somewhat worse than that
of the USLE. The RUSLE performed similarly to
the WEPP model for average annual values, and
similarly to the USLE for annual values. Other
studies show similar results. Bathurst and Lukey
(1998) used the SHETRAN model to simulate
sediment fluxes in experimental catchments
strongly affected by soil erosion in the French
Alps, and found that SHETRAN performance was
very mixed, with good agreement for some events
while predictions were out by more than an order
of magnitude for other events; overall the correla-
tion between predicted and measured data was
not significant. Using the same data and a simple,
empirical model based on rainfall characteristics
only as an input, Brochot and Meunier (1995)
arrived at good predictions over a wide range of
events (R
2
7.3.1
Misconception: 'Better' models lead
to more accurate predictions
(of soil erosion rates)
The continuous development of soil erosion mod-
els has been driven, to a large extent, by the desire
to improve predictions. Although not always
explicitly stated, it was assumed that statistical
'average' models were inadequate to describe the
complexity of soil erosion processes and did not
allow sufficiently for the impact of temporal and
spatial variability as well as the role of different
subfactors in model input variables. While evi-
dence of the importance of temporal variability
and the role of soil and vegetation subfactors
gradually accumulated in the literature from 1960
onwards, the seminal papers published by Foster
and Meyer (1972, 1975) in conjunction with the
advent of cheap computer power caused the
development and testing of process-based models
to progress relatively quickly from 1980 onwards.
There is no doubt that erosion models have
been 'improving' over time: rather than being
purely empirical and simulating average condi-
tions only, many models are now capable of
accounting for temporal variability and often
contain some form of process description.
Research on erosion processes is still ongoing and
may lead to a further improvement of the process
descriptions in the future.
One might therefore expect that the predictive
capability of such newer models would exceed
that of old models such as the USLE. Over recent
years, several publications have appeared where
the performance of statistical, average models
0.72-0.78). Risse
et al
. (1993) exten-
sively tested the USLE and arrived at R
2
values of
0.58 for annual and 0.75 for average annual val-
ues, respectively, while a similar study by Zhang
et al
. (1996) using the WEPP model obtained val-
ues of 0.50 and 0.65 for annual and annual average
erosion rates before optimization, and values of
0.54 and 0.68 after optimization; these results are
thus in line with those of Tiwari
et al
. (2000).
Rapp
et al
. (2001) did not find any improvement
in soil loss predictions from 206 natural runoff
plots when the RUSLE was used instead of the
=
