Geology Reference
In-Depth Information
USLE: model efficiencies were 0.73 for average
annual and 0.58 for annual values for both mod-
els. Klik and Zartl (2001) reported that WEPP rea-
sonably simulated soil losses for individual
storms as measured on plots in Austria, but only
after increasing standard inter-rill erodibility val-
ues by a factor of 9-10 and rill erodibility by a
factor of 1-3. With respect to annual values, the
performance of the calibrated WEPP and uncali-
brated RUSLE was similar. Stolpe (2005) found
that measured erosion rates in south-central Chile
were equally well predicted using the USLE and
WEPP (R
2
and no detectable progress was made in flood pre-
diction over two decades, despite increasing
model sophistication (Welles
et al
., 2007).
There are at least two basic reasons for this
rather uneasy situation: firstly, process-based mod-
els require much more input data, and with every
input value that is required there is necessarily
some uncertainty associated. As an example, WEPP
uses an inter-rill erodibility factor (
K
i
) that is calcu-
lated as the product of a base inter-rill erosion fac-
tor and seven subfactors (Alberts
et al
., 1995). Each
of these eight factors is calculated using an empiri-
cal relationship to account for canopy effects,
ground cover, roots, and so on. Assuming, as pro-
posed by Alberts
et al
. (1995), a baseline value of
5.3 × 10
6
(kg s m
−4
) for the base inter-rill erodibility
factor and assuming, for the sake of simplicity, that
each of the adjustment factors has a value of 1 and
that each of the factors has an uncertainty (error
standard deviation) of 10%, the resulting inter-rill
erodibility value will already have an uncertainty
of around 28%. An increase in the estimation
uncertainty of the subfactors to 20% would lead to
an uncertainty on the final value of around 60%.
Thus, as models become more sophisticated, the
uncertainty associated with estimating input val-
ues inevitably increases. Even if one would, for a
given point in the landscape, be able to estimate
K
i
with a very high degree of accuracy, we would face
a problem of uncertainty as the values for different
subfactors vary significantly over space. Basically,
further sophistication of a model will only result in
improved predictive capabilities if the increase in
prediction error resulting from an additional input
data is smaller than the reduction in prediction
error to a better model prediction (Van Rompaey &
Govers, 2002).
Secondly, even the most sophisticated model
will not be able to capture all the variability in
soil conditions that lead to variations in erosion
response. This can be shown by comparing ero-
sion measurements from replicate plots. Many
erosion study sites of the USDA contained repli-
cate plots: one might consider replicate plots as
being physical models of each other (i.e. 'identi-
cal twins'). It is therefore unreasonable to expect
that a simulation model can make a better
0.86), but predictions using the RUSLE
were weaker (R
2
=
0.50). Spaeth
et al
. (2003) com-
pared the performance of the USLE and RUSLE
using rainfall simulation data and found that both
models performed poorly: the greater flexibility of
the RUSLE did not offer any advantage over the
USLE. Results are broadly similar for catchment
models; again, spatially distributed, dynamic
models do not significantly outperform lumped
models when sediment yields are compared. Im
et al
. (2007) compared the HSPF model, which
uses detailed process descriptions but a rather
rough spatial discretization, with the spatially
distributed SWAT model with respect to the pre-
diction of runoff and sediment yield for a 12,000
ha watershed in Virginia, and found that both
models performed similarly with slightly better
results for HSPF. Parajuli
et al
. (2009) found the
process-based SWAT and the AnnAGNPS model
to perform similarly with respect to runoff and
sediment yield. SWAT did perform better, how-
ever, with respect to phosphorus export. Shen
et al
. (2009) reported that WEPP outperformed
SWAT with respect to runoff prediction in the
Zhangjiachong Watershed, but both models per-
formed similarly with respect to sediment yield,
with model efficiencies (after calibration) of 0.83
and 0.82 for WEPP and SWAT respectively.
Thus, the development of more sophisticated
models does not seem to lead to better predic-
tions. This situation is not at all unique with
respect to soil erosion; similar observations have
been made in hydrology, where the feasibility of
deterministic hydrological modelling has been
questioned for a long time (Grayson
et al
., 1992)
=
