Geology Reference
In-Depth Information
-
||
simulations of the second peak. However, the tim-
ing of the main peak is well simulated and so is
the initial rise of the hydrograph. The 'spikiness'
of the modelled discharge response is again appar-
ent. In terms of the
SS
simulations, results are not
different to event 1 and, again, no parameter set
survived the test against the sediment data.
Figure 5.6 shows the relationship between
model performance for
Q
and
SS
and parameter
values for both events, with only those simula-
tions which pass the
-
||
0.4 with the same parameter sets. Lowering
the performance threshold would result in over-
lapping parameter distributions, but at the expense
of simulation accuracy of individual events.
=
5.5 Discussion
The events simulated in this study were charac-
teristic of those found in temperate areas in
response to frontal rainfall systems, which typi-
cally generate long-duration but low-intensity
rainstorms. Simulating such events required
parameter sets that minimized both infiltration
and erosion, since runoff had to be generated, but
erosion rates were low, probably due to the cohe-
sive nature of the soil. Most of the parameter val-
ues identified are sensible for the conditions
described above. However, as the better-perform-
ing parameter regions for the two simulated
events did not always overlap, we cannot identify
consistent parameter sets for the conditions at
Den Brook even if variability in initial conditions
is taken into account. Therefore we would not
recommend 'optimizing' model parameters as a
general approach, as we also demonstrated that
many parameter values gave similarly good simu-
lations. Some of the most sensitive parameters
had non-overlapping behavioural distributions
for the two events (when considering only dis-
charge), particularly
RFR
,
FMIN
,
G
and
XL
. This
indicates that both events could not be described
with the same sets of parameters using our behav-
ioural threshold, unless the level of model per-
formance for individual events was compromised.
This may be due to errors in the model structure,
the input data (which in this study is not
accounted for) or real variations in hydraulic and
soil surface microtopographical characteristics
between events. As EUROSEM is an event-based
model, it relies on the user to recognize changes
in parameters in order to take count of surface
changes. However, the changes in the optimal
values of
FMIN
and
RFR
are counter-intuitive:
both increase between event 1 and 2 when it
would be expected that the soil surface would
become smoother and less permeable due to
0.4
Q
performance cri-
terion displayed. It is clear from Fig. 5.6 that many
more discharge simulations performed better for
event 1 than for event 2. However, it is also clear
that no sediment simulations performed better
than the
MAE
threshold of 150 mg l
−1
. Although
rejected as non-behavioural, by showing the
model realizations with respect to suspended sed-
iment error we are able to see which parameters
the model is sensitive to. Few of the parameters
have much influence over the model's perform-
ance for the simulations we ran. Key parameters
for discharge are: saturated hydraulic conductiv-
ity (
FMIN
), field length (
XL
; a surrogate for catch-
ment area), capillary drive (
G
) and surface
roughness (
RFR
). For sediment we can add cohe-
sion (
COH
) and erodibility (
EROD
) to the list.
The dotty plots show different responses for the
two events. The better-performing values of satu-
rated hydraulic conductivity, surface roughness
and erodibility are higher for event 2 than for
event 1, while those of slope length and capillary
drive are lower for event 2 than for event 1. Given
that the true area of the field lies somewhere
within the sampled range of slope length times
slope width, which might not coincide with the
area of highest performance, it is clear that the
slope length parameter can compensate for model
deficiencies here. Indeed, for a true area corre-
sponding to a slope length towards the low end of
the sampled range, model performance would be
worse than the chosen
-
||
=
0.4 threshold. It is also
clear that updating the posterior parameter distri-
bution of event 1 with that of event 2 results
in rejection of all sampled parameter sets due to
the non-overlapping distributions in Fig. 5.6, i.e.
the two events cannot be simulated better than
=
