Geology Reference
In-Depth Information
model (Knisel, 1980), parts of which underlie the
WEPP model (see later discussion), and its com-
panion the Groundwater Loading Effects of
Agricultural Management Systems (GLEAMS)
(Leonard
et al
., 1987), were also products of the
1980s revolution in 'physically-based' model
development, and were subjected to sensitivity
analyses (Silburn & Loch, 1989) which focused on
the erosion-sedimentation part of the model.
Despite early efforts to ensure that CREAMS/
GLEAMS were physically-based and therefore did
not need to be calibrated, the sensitivity analyses
show similar limitations to the empirically-based
USLE, perhaps not surprisingly as parts of the
USLE were incorporated into the erosion compo-
nent of the model. The model was shown to be
sensitive to a variety of parameters including the
USLE erodibility and vegetation cover parame-
ters, peak runoff rate, storm erosivity, slope
length and sediment size distribution (Silburn &
Loch, 1989), demonstrating the legacy of former
model development in the new generation of ero-
sion prediction tools. In addition, Silburn and
Loch (1989) found parameter interaction to play a
role in controlling model output, although this
was not adequately assessed via the univariate
sensitivity analysis.
The Système Hydrologique Européen (SHE)
soil erosion model was also subject to a sensitiv-
ity analysis (Wicks
et al
., 1992), although the
goals of this work were more explicitly to under-
stand whether the model could be applied with-
out calibration. Results demonstrated that the
model could predict overland flow response to
simulated rainfall on saturated soil, but was less
able to predict flow generated on dry soil, and
indeed required detailed calibration to predict
erosion or sediment yield as the erosion submodel
was very sensitive to correct representation of
the hydrology (Wicks
et al
., 1992).
Veihe and Quinton (2000) and Veihe
et al
.
(2000) analysed the EUROSEM model using two
techniques - a Monte Carlo-based approach
which is discussed in the following sections, and
a simple sensitivity analysis which varied para-
meters by
the soil has a significant impact on runoff genera-
tion and subsequent soil erosion predicted by the
model. However, Veihe and Quinton (2000) rec-
ognised that such simple sensitivity analyses do
not inform the user about the distribution of out-
put from the model, and as such they advised that
more complex, multiple parameter variation is
carried out using Monte Carlo techniques.
In part to address the need to understand the
influence of multiple model parameters, recently
a more complex form of sensitivity analysis has
been developed by Wei
et al
. (2007) using the
Rangeland Hydrology and Erosion Model (RHEM)
based on the WEPP model (Nearing
et al
., 1989).
In this exercise, the authors selected 14 of the
model parameters relating to the hydrology and
erosion subcomponents of the model and varied
these parameters by 5%, one at a time, to calcu-
late a local sensitivity index at each of 10,000
randomly selected points in the 14-dimensional
parameter space (as defined by pre-determined
parameter ranges (Wei
et al
., 2007). The 'local-
ized' sensitivity indices were then summarized
to compare the relative importance of each
parameter to model output and, via classification
of the local sensitivities of each parameter, to
illustrate the distributions of parameter sensitiv-
ity within the model. Wei
et al
. (2007) reported
that results were in agreement with those of
Tiscareno-Lopez
et al
. (1995), who performed a
similar (although less complex) univariate sensi-
tivity analysis on the WEPP model.
Two lessons can be learnt from this work:
(1)
As with all univariate sensitivity analyses,
the neglect of the interaction of parameters in
combination with each other severely under-
mines the conclusions of the research. By only
varying one parameter at a time, it is assumed
implicitly
that there is no interaction between
parameters, when this is clearly not the case. Wei
et al
. (2007) attempted to deal with this problem
via regression analysis of parameters to explore
what the relationships between parameters may
be. However, this is only performed between
pairs of parameters, ignoring the other 12 at any
one time, so does not constitute a full analysis of
parameter interaction.
10%. Results of the latter approach
indicated that the presence of rock fragments in
±
