Geology Reference
In-Depth Information
comparing the model predictions with field
observations and applying some measure of
goodness-of-fit. Where the comparisons are
based on numbers, the goodness-of-fit is often
assessed by statistical techniques, demonstrat-
ing that there is no significant difference
between the predicted and observed values, or
that there is a 1:1 relationship between them
expressed by a best-fit regression equation with
a slope close to 1.0 and which passes through
zero. The performance measure most favoured
at present by researchers is the Model Efficiency
Coefficient (
MEC
) (Nash & Sutcliffe, 1970),
defined by:
first check whether it has been validated by the
developer at the appropriate scale. However, the
user requires more than a generic validation. It is
important to show that the model operates well
in the conditions that apply in the study area. In
some cases, it may be necessary to modify the
model before it can be applied. Hessel (2002) had
to make changes to LISEM before it could be
used with the very high sediment concentra-
tions found in the runoff on the loess soils of
China. Fortunately, he was able to access suita-
ble research data both to develop and validate
the changes (see Chapter 12) but, in many parts
of the world, local data suitable for validation
are often non-existent or sparse. Although,
ideally, the validation should be based on the
principles and procedures described in Chapter 3,
in reality it is often restricted to showing that
the model outputs are close to the values
obtained from another study area for which field
data exist and which is geographically similar.
The user invariably has to begin by searching for
whatever data are available from whatever
sources. These may be from erosion plots at
research stations, but perhaps for periods of
three years or less, from consultations with
farmers and other land-users or from undergradu-
ate studies at the local university. Sometimes the
user may have to implement his or her own
research either by undertaking a period of field
observations during the season when erosion is
most likely, or by interpreting aerial photography
and mapping erosion features. In some cases, data
availability may be so sparse that a user attempt-
ing to demonstrate that the model outputs for,
say, a small catchment in Kalimantan, Indonesia,
are sensible may be able to do no more than show
that the model predictions for erosion plots under
similar climate, soils, slopes and land covers at
research stations in Java, Indonesia, or, even, in
Malaysia, are close to measured values. Similarly,
a user endeavouring to show that model predic-
tions of erosion on cleared land along pipeline
corridors in northern Argentina or west of Tbilisi
in Georgia are valid, may only be able to show
that model outputs are close to what is observed
on slopes devoid of vegetation in similar climatic
(
)
å
å
XX
-
2
obs
pred
MEC
=-
1
(2.9)
(
)
2
XX
-
obs
obs
where
X
obs
is the observed value,
X
pred
is the
value predicted by the model and
-
is the mean
of a set of observed values. The
MEC
is a meas-
ure of the variance in the predictions from a 1:1
prediction line with the measured values. Thus,
the closer
MEC
is to 1.0 in value, the better is
the model performance. Values are rarely > 0.7
(Nearing, 1998) and a value > 0.5 is considered
satisfactory (Quinton, 1997). Negative values
indicate that the model predictions are poor
and have a higher variation than the observed
values. Where validation is based on the model
predicting erosion in the right place in the land-
scape at the right time, comparisons are more
subjective, although they can be expressed in
terms of the number of observations being cor-
rectly predicted.
For model users, validation is a more complex
issue than for model developers. The latter need
to be able to demonstrate that their model is
rational and gives reasonable predictions when
compared with what is observed. Validation can
therefore be based on large datasets obtained from
research stations, although some caution is nec-
essary where, as indicated above, these data are
not at the same spatial or temporal scale as the
model. When selecting a model, the user should
