Geoscience Reference
In-Depth Information
degree-day snowmelt calculations could also produce acceptable discharge predictions (Figure 5.6b) but
only after calibration of the degree-day coefficient to a value that was high relative to those reported in
the literature.
5.4 Case Study: Blind Validation Test of the SHE Model on the Slapton
Wood Catchment
Ewen and Parkin (1996) have outlined a methodology for the
blind validation
of a hydrological model that
involves the specification of tests and criteria of success before the model simulations are compared with
observed discharges or other observations. This is a form of the “proxy basin” validation test of Klemes
(1986) (see also Section 5.7). In an application of this methodology to the 1.4 km
2
Rimbaud catchment
in the Maures Massif near Toulon in southern France, Parkin
et al.
(1996) tested the SHETRAN version
of SHE using a grid of 100 m and only prior estimates of the parameter values based on information
about the soils and vegetation. The Rimbaud catchment is one of a number of nested subcatchments
in the Real Collobrier basin, managed by CEMAGREF. Uncertainty in these estimates was allowed by
specifying a range for each parameter. Some general information about the runoff responses was used to
set the criteria for success in the model evaluation. The evaluation was “blind” in that the modellers did
not have access to the observed discharge record from the catchment before making the model runs. At
that time, computer run times of the SHE model were still a significant issue and only a limited number
of runs to explore the parameter space and estimate the range of model predictions were possible.
The model was evaluated on four criteria set before the start of the blind test. It was required that
the predictions bounds bracket 90% of the observed discharges, 90% of the peak discharges, 11 out of
13 monthly runoff volumes and the total runoff volume. In fact the model was totally successful in only
the last of these criteria. Only 78% of the discharge hydrograph, 47% of the peak flows (the model tended
to generate runoff by an infiltration excess mechanism, overestimating the peaks with recessions that
were too steep) and 10 of the 13 monthly flows were within the prediction bounds. No attempt was made
to evaluate any internal predictions in relation to measurements. The studies of Refsgaard and Knudsen
(1996) and Feyen
et al.
(2000) also demonstrated somewhat limited success in trying to validate the SHE
model in this way. Somewhat greater success has been claimed by Lange
et al.
(1999) in predicting peak
flows in a semi-arid environment using a simpler distributed model based on Hortonian infiltration excess
concepts, but with surface runoff generation based directly on field measured plot infiltration experiments.
A more recent study of this type is the application of the SHETRAN version of SHE to the 0.94 km
2
Slapton Wood catchment in Devon, UK by Bathurst
et al.
(2004), following an intensive period of
hydrological observations that included internal state measurements of water tables and soil moisture
profiles. The catchment was represented by a grid of 376 squares of 50 m. Blind predictions were made
of 10 features of the phreatic surface, soil water potential and surface runoff responses based on prior
estimates of the parameter values. Output uncertainty bounds were determined as a function of uncertainty
in the model parameter values. Again the number of runs made to estimate the uncertainty bounds was
limited by available computer time.
Bathurst
et al.
(2004) note how the resulting uncertainty bounds were subjectively modified in magni-
tude and timing (before the modellers had sight of the evaluation data). This, it was argued, was to take
account of the additional uncertainty arising from instrumentation and data processing errors (there is
also the commensurability issue of what the measurements at points in the catchment mean relative to
what the model is predicting on a 50 m grid). They also note how they had some prior information about
the nature of the response of the catchment (initial condition information, the double peaked hydrographs
in this catchment, a marshy area at the head of the stream) from previous publications about the catchment
even though they were not allowed to see the observations with which the model would be compared.
