Geoscience Reference
In-Depth Information
Sub-ET1
Sub-ET2
16
16
14
12
10
8
6
4
2
14
12
10
8
6
4
2
2
46810
E
PAN
(mm day
-1
)
12
14
16
246810
12
14
16
E
PAN
(mm day
-1
)
(a)
(b)
Sub-ET3
ET = Sub-ET1+Sub-ET2+Sub-ET3
16
14
12
10
8
6
4
2
16
14
12
10
8
6
4
2
R
2
= 0.77
246810
E
PAN
(mm day
-1
)
12
14
16
246810
12
14
16
(c)
(d)
E
PAN
(mm day
-1
)
FIGURE 8.6
Individual (a-c) and combined (d) test period outputs produced from Sub-ETs in preferred
E
PA N
solution for estimating pan evaporation at NLWS.
http://www.ec.gc.ca/rhc-wsc/). The meteorological data were obtained from Canada's National
Climate Data and Information Archive and are for Greenwood Airfield (44°58′40″N, 64°55′33″W),
situated almost 10 km east from Wilmot gauging station (http://climate.weatheroffice.gc.ca/).
Eureqa (Cornell University, 2013) was used to perform GP modelling. The software package
supports a number of basic data cleansing operations, including smoothing, normalisation and out-
lier removal - but in this study, they were not applied.
Eureqa
automatically ranks each solution in
terms of its complexity; this is calculated from the number of nodes present in the solution, that is,
size of solution. Collectively, these features mean the potential effects of data pre-processing and
model complexity operations can be easily evaluated. The mathematical functions selected for the
Annapolis experiment are shown in Table 8.6 and other settings in Table 8.7. Goodness-of-fit met-
rics for the developed models were calculated using HydroTest, a standardised, open access website
that performs the required numerical calculations (Dawson
et al., 2007).
Eureqa
modelling of daily discharge for the Annapolis River produced a range of solutions
of varying complexity. The principal rainfall-runoff drivers are found to be lagged discharge
and lagged precipitation, as expected. The complexity-accuracy trade-off is nevertheless found
to exhibit threshold behaviour, in which snow cover is eventually included at higher levels of
complexity to account for multifaceted cold season processes. Figure 8.8 illustrates that fitness
was generally poorer for models that did not include a snow-related factor, which first appeared
in model size 25.
