Geology Reference
In-Depth Information
although as is discussed above, these papers
describe what is more appropriately termed sen-
sitivity analysis, not uncertainty analysis. Such
a call is better answered by a holistic framework
as demonstrated in Chapter 5, where a modified
version of the GLUE approach is used to evalua-
tion predictions from EUROSEM.
Hantush and Kalin (2005) took a similar
approach to that of Veihe and Quinton (2000) to
analyse uncertainty associated with the
KINEROS2 model (a development of KINEROS, a
kinematic runoff and erosion model; Woolhiser
et al
., 1990). The model is evaluated against
observed data from a 33.6 ha catchment, known
as W-2, which is located near Treynor, Iowa, US.
Monte Carlo simulations were used to generate
exceedance probability curves (for comparison,
these are akin to the cumulative distribution
function (cdf) curves of Veihe & Quinton, 2000,
or the cumulative distribution curves of Brazier
et al
., 2000) and uncertainty bounds based on
25th, 50th and 75th percentiles (see Fig. 4.3). The
authors demonstrate that the KINEROS2 model
provides more reliable predictions for larger
events and less reliable predictions of the dynam-
ics of smaller erosion events.
These model findings relate well to observa-
tions made by other authors. For example,
Nearing (2000) and Nearing
et al
. (1999) reported
greater variability in the response of replicate
plots to low-magnitude, high-frequency events,
than their larger, rarer counterparts. Thus, in a
sense, it is unrealistic to expect models such as
KINEROS2 to make more certain predictions
than the empirical data they are being evaluated
against. However, when the model is applied to
catchments for which calibration/evaluation data
do not exist, it is suggested that runoff and sedi-
ment yield predictions will only be “… within
order of magnitude of accuracy” (Hantush &
Kalin, 2005).
A further analysis of the uncertainty associated
with the erosion predictions of KINEROS2 is illus-
trated by Martinez-Carreras
et al
. (2007), in the
context of erosion occurring in a semi-arid badland
environment in the Ca l'Isard catchment (1.32 km
2
)
in the Eastern Pyrenees. First, calibration data
were collected in a small subcatchment to aid
parameterization of the model prior to application
at the catchment scale. Subsequently, the model
was run against four years of observed data using
the GLUE approach. Extending the argument of
Hantush and Kalin (2005), it might be thought that
as calibration data were available, we would expect
model performance to improve; however, this was
not generally the case. The model underpredicted
observed erosion rates and it was suggested that in
part this may be because the calibration data that
were used were not suitable. It is very important in
exercises of this type that the observed and pre-
dicted variables be
commensurate
(i.e. have the
same meaning). This is not always the case, and
the assumption of Hantush and Kalin (2005) and
Martinez-Carreras
et al
. (2007), as well as many
other authors, that model calibration improves
model prediction, may be true for specific sites or
model applications, but where the calibration is
affected by observational limitations there is no
guarantee that the effective parameter values
determined by calibration will be transferable to
sites where no calibration data are available (see
also sections 2.8.2 and 2.8.3).
As has been argued above, erosion models
may have many sources of error, all of which are
likely to contribute to the uncertainty that sur-
rounds model predictions. Jetten
et al
. (2003)
recognised that the current generation of erosion
models are very rarely (if ever) tested against
datasets describing not only the catchment
outlet hydrograph and sediment yield, but also
the
patterns
of erosion within the catchment.
In their overview of the Global Change and
Terrestrial Ecosystems Focus 3 programme, Jetten
et al
. (2003) acknowledged that erosion models
are only “moderately good” at predicting outlet
hydrographs and are “not very good for net soil
loss”. The authors called for better description
of the spatial distribution of soil erosion and bet-
ter use of these data to evaluate soil erosion
models (see also Section 3.5).
Incorporating spatial erosion data for model
evaluation was central to the work of Van Oost
et al
. (2005), who presented a dataset describing
the spatial variability in soil erosion from the
