Geology Reference
In-Depth Information
compared with the increased amount of informa-
tion needed to run the model, something that is
almost never accounted for in model comparison
studies.
The above suggests that the problem may be
(partly) solved by complementing output-based
model evaluation with other information. Jetten
et al
. (2003) have already discussed this issue and
proposed that more attention should be given to
what may be called internal model validation,
i.e. the degree to which the model is indeed capa-
ble of representing the spatial and temporal vari-
ations in soil (and water) redistribution within a
catchment (see also Chapter 3). While data on
internal sediment redistribution are usually not
available on an event basis, tracer studies and soil
truncation studies (Govers
et al
., 1996; Quine
et al
., 1997; Polyakov
et al
., 2009) can provide
information on sediment redistribution patterns
over the long term. Other spatial data may also be
useful, such as the spatial distribution of ephem-
eral gullies (Nachtergaele
et al
., 2001) or the spa-
tial distribution of deposition zones after an
important event.
However, we should realize that there may be
fundamental limits with respect to what may
be achieved using spatial data; the spatial and tem-
poral resolution of the spatial data required that
we may collect will in general be at least an order
of magnitude smaller than that of the input data
for a spatially distributed, dynamic model, as well
as of the simulated model response. Furthermore,
these data will be characterized by important
uncertainty. For instance, correctly estimating
deposition volumes on a field is hampered by the
variation in field microtopography. The uncer-
tainty on such estimates may therefore well
exceed values typical for reservoir sedimentation
(i.e. 30-35%; Verstraeten & Poesen, 2002). Thus,
spatially distributed data on sediment distribution
may help to identify major shortcomings in model
results, but will not allow solution of the equifi-
nality problem entirely: the use of spatial data
may help to constrain the range of 'possible'
parameter values and to compare those values
with physically meaningful values. By doing so we
may be able to investigate to what extent a pro-
posed model is indeed capable of describing the
observed response. If the observed response can-
not be described with reasonable parameter val-
ues, then this might be an indication that the
model is structurally flawed, and is not capable of
describing the processes controlling the field
response.
7.3.3 Misconception: The major advantage
of process-based models is that they can be
applied without a priori calibration
The development and testing of process-based
models has often been promoted through the idea
that statistical models can be applied only to con-
ditions for which they have been developed, while
a process-based model may be applied to condi-
tions different from those for which the model
was developed and tested. The implicit idea
behind this is that process-based models describe
the basic processes leading to runoff and sedi-
ment detachment, transport and deposition.
Hence, such a model should, in principle, be
applicable without extensive
a priori
model cali-
bration and it should be possible to use such
models for
ex ante
evaluations.
This assumption is wrong for several reasons.
One of the most important is that, while process-
based models may use deterministic, well-tested
process descriptions, the inputs that are necessary
to run them (e.g. erodibility, soil hydrological
characteristics) are usually estimated through
statistical procedures. For instance, the effective
hydraulic conductivity (
K
erange
(mm h
−1
) ) for range-
land in WEPP is estimated using the following equa-
tion when the surface has >45% rills (Alberts
et al
.,
1995):
K
=-
14.29
-
3.40 ln(
ROOT
10)
+
37.83
s and
erange
+
208.86
orgmat
+
298.64
RR
-
27.39
RESI
+
64.14
BASI
(7.1)
where
ROOT
10 is the root biomass in the top
0.1 m of the soil (kg m
−2
),
sand
is the fraction of
sand in the top 0.1 m of the soil,
RR
is the random
roughness (m),
RESI
is the fraction of litter surface
