Geoscience Reference
In-Depth Information
5.3 Case Study: Modelling Flow Processes at Reynolds Creek, Idaho
Reynolds Creek is a 234 km
2
rangeland catchment in the Owyhee Mountains of Idaho, managed by the
USDA North West Watershed Research Centre. It was the site of one of the very first attempts to evaluate
the predictions of a distributed process based hydrological model. Stephenson and Freeze (1974) used a
two-dimensional, finite difference, partially saturated Darcian subsurface flow model in an application
to a vertical slice through a complex hillslope within the Reynolds Creek catchment. Model predictions
were checked against field measurements made during a snowmelt season. After making initial estimates
of parameter values for the soil and rock layers in the slope, they carried out a trial and error calibration
of the model, adjusting the parameter values to try to improve the fit to the observations. At that time,
computer constraints severely limited the number of calibration runs that could be made. The results of
the best simulation are shown in Figure 5.5.
This study is interesting, even after 35 years, because it is one of the first studies to recognise that there
might be limitations to the application and validation of this type of model at particular sites. Stephenson
and Freeze conclude (1974, p. 293) that:
we recognise that our calibration is less than perfect but it is probably representative of
what can be attained when a fully deterministic mathematical model is applied to a field
site with a fairly complete, but as always imperfect, set of field measurements.
They also noted that validating such models was a particularly difficult problem since it presupposes
perfect knowledge of all the boundary conditions, parameter values and initial conditions required.
Imperfect knowledge always introduces a degree of flexibility (or uncertainty) to any attempt at validation
of the model.
More recently, the SHE model has been applied to the 40.4 ha Upper Sheep Creek subcatchment by
Bathurst and Cooley (1996). At this site, at an elevation of over 2000 m in the headwaters of Reynolds
Creek, average annual precipitation is of the order of 1016 mm, with more than 70% falling as snow.
Snow accumulation is highly variable, with a deeper pack building up in the lee of a ridge where the pack
may reach depths of more than 5 m.
Bathurst and Cooley simulated a single snowmelt period using both energy budget and degree-day
snowmelt models within the SHE model framework. A model discretisation based on 161 square grid
cells (50 m by 50 m) was used. All parameters for the model were specified on the basis of knowledge
of the catchment soils and vegetation. Initial snowpack characteristics were assigned on the basis of
snow course information and photographs; initial saturated zone thickness was assigned to reproduce the
initial flow at the start of the simulation. It is not clear from the paper how the initial unsaturated soil
moisture profiles were defined and only a 12-hour period was allowed for the model to “run in” before
the predictions were compared with observations.
The study aimed to test four hypotheses about the processes in the catchment based on how well the
model reproduces the stream discharge during the simulation period. These hypothesis varied in assump-
tions about the extent of frozen soil and depth of the effective impermeable layer. Unlike Stephenson and
Freeze (1974), Bathurst and Cooley made no attempt to validate the model predictions against internal
state measurements. The authors state (1996, p. 194) that
the traditional calibration approach of adjusting the parameter values (within each hypoth-
esis) to improve the agreement played a secondary role and was carried out under the
constraint that the [parameter] values must reflect the field measurements where these exist
or should otherwise lie within physically realistic limits.
The earliest run number reported in the paper is 69, the last 107. It is clear that this application, like that
of Stephenson and Freeze 20 years earlier, was limited by computer run times.
