Geology Reference
In-Depth Information
regions elsewhere in the world (see Section 14.7.4
for examples of using models hypothetically in the
absence of data, and Section 15.3 for specific issues
associated with studies of climatic change).
The availability of suitable data and the
demands of potential users have meant that vali-
dation has been largely limited to model outputs
such as the quantity of runoff and sediment leav-
ing an area over a unit of time. This is satisfac-
tory for relatively simple models where such data
are all that are predicted, but for more complex
process-based models it ignores the fact that the
same model outputs can be obtained in a variety
of ways. As already seen, with such models, pre-
dictions are made of the quantities of soil
detached by raindrop impact, detached by runoff,
transported by rainfall, transported by runoff,
and deposited at numerous points in the land-
scape. If the soil loss from an area is controlled by
the transport capacity of the runoff, the model
needs only to get the transport capacity of the
runoff from the lowest land unit in the landscape
correct, and to ensure that the availability of
detached particles on that unit either equals or
exceeds transport capacity, to get a reasonable
prediction. All the predictions of the other proc-
esses and what is happening elsewhere in the
landscape could be wrong, but if they do not
affect the output, this may never be known.
Unfortunately, detailed observations of the rates
of soil particle detachment, sediment transport
and deposition on different land units within
catchments are not generally available, and
data on their particle-size distribution are even
scarcer. It is therefore impossible to validate
complex process-based models fully. The impli-
cations of this for making judgments on model
outputs are explored in Chapter 7.
from high spatial and temporal variability which
means that a parameter cannot adequately be
expressed by a single value, errors involved in
estimating the values of those parameters that
cannot be easily measured, the use of guide val-
ues rather than measured values, and errors aris-
ing from predictions made by the various
equations used to run the model. The relation-
ship between these different sources of uncer-
tainty is far from clear, so it is not obvious
whether they are additive or multiplicative in
their effect, or whether some cancel others out so
that averaged overall, the errors have little influ-
ence. The procedures used to reduce the effects of
error (see Chapter 4) require large numbers of
datasets. For a generic evaluation of a particular
model, these sets can be generated artificially by
varying the values of input parameters randomly
within the limits of what is likely to occur in
field conditions. Multiple simulations can then
be run from which the number of simulations
needed to produce predictions within certain
error limits can be determined (Quinton, 1997;
Brazier
et al
., 2007).
The problems with the above approach are,
firstly, that the users need to know what level of
error is acceptable for their purpose. Given that
users are not always as familiar as model develop-
ers with the concepts of probability and uncer-
tainty, they may have little idea of how to
determine what is or is not acceptable. Some users
known to the author would not consider a model
worthwhile unless it was able to predict annual
soil loss in tonnes per hectare to two decimal
places, whereas in reality, some models will be
performing well if they can predict to the correct
order of magnitude. Secondly, a view of 'correct'
prediction is based on the assumption that the
measured or observed value is correct. In practice,
this disregards errors involved in the measure-
ment process. As already seen, most data used for
model validation come from measurements made
on erosion plots. As long ago as the late 1950s,
Hudson (1957) drew attention to the sources of
error associated with measurements from such
plots and, whilst some improvements in the design
of equipment have been made, the majority of
2.8.5 Uncertainty
The last decade has witnessed an increasing rec-
ognition that model predictions are subject to
uncertainty. One source of uncertainty is the dif-
ficulty of determining input parameter values.
Uncertainties arise from errors associated with
the techniques of measurement, errors arising
