Geoscience Reference
In-Depth Information
floating acoustic doppler velocity measurement devices, the costs of such measurements have become
smaller compared to more labour-intensive current metering. Even a small number of current metering
estimates of discharge might be feasible for some applications if it can be shown that the measurements
have value in reducing uncertainties for decision making.
This then raises the question of which measurements are the most cost-effective in the learning process,
and how many measurements might be necessary to achieve a desired reduction in uncertainty. This
remains an open question: there is almost no guidance in the hydrological modelling literature about
the value of measurements in model identification, except for some vague suggestions about how long a
discharge record is necessary to obtain an optimal model calibration (of the order of several years of data).
If these suggestions are correct then clearly it is unlikely that funding to collect several years of data in
an ungauged catchment would be made available, except perhaps in exceptional circumstances for high
capital expenditure projects, such as dams. However, Juston
et al.
(2009) demonstrate that, if selected
in an intelligent way, a small fraction of data points in a longer time series might contain almost all the
information of the entire data series. Rode and Suhr (2007) applied a water quality model to the River
Elbe and also found that a subset of the entire calibration data already provided good results. Rojas-Cerna
et al.
(2006) showed how a small number of measurements could be combined with prior estimates of
model parameters derived from a regionalisation process to improve the predictions of ungauged sites.
Binley and Beven (2003) also showed that a single set of geophysical measurements of a deep soil water
profile contained most of the information content of 18 months of weekly measurements in conditioning
a model of groundwater recharge. McIntyre and Wheater (2004) came to a similar conclusion for water
quality modelling, providing that the data were chosen from storm periods. This is an indication of the
potential value of limited observation data for conditioning the uncertainty in estimating the response of
an ungauged basin.
Seibert and Beven (2009) have addressed this problem in an application to 11 catchments draining to
Lake Malaren in Sweden. Models were calibrated using different selections of daily discharge val-
ues for one year, and then used to predict the catchment response for a 10-year period. For each
calibration, the best 100 models out of a sample of 10 000 runs were chosen based on the sum
of squared errors. They showed that a relatively small number of randomly chosen observations
(16-32 days) of data provided almost as much information as having a full year of data, but that a
poor (random) choice of days might decrease the evaluation performance. Even fewer measurements
might be needed when a hydrologically intelligent choice is made about when the measurements are
taken, in this case (see Figure 10.3) either the six highest peaks in a two-month period (MAX6) or peak
flow days followed by recession days (MAX1REC5, MAX2REC4). Interestingly, the ensemble mean
prediction over the 100 models, performed better in evaluation than the single best model of the set for all
the cases.
Others have also considered the use of limited streamflowmeasurements in constraining rainfall-runoff
models for ungauged catchments including Rojas-Cerna
et al.
(2006) and Winsemius
et al.
(2009). The
latter paper also suggests that “soft” information might be used where available (see also Seibert and
McDonnell, 2002). There remains a logistical problem of collecting even a small number of discharge
observations, especially in getting a team into the field to measure peak discharges (including the health
and safety issues of making measurements at the very highest peaks). One strategy to make this easier
might be to install a level sensor for a short period of time (e.g. during a single snowmelt season, in the
case of Sweden) and use the small number of measurements to develop an approximate rating curve for
a site. In doing so, however, it should be remembered that not only might the limited information in a
discharge record constrain the accuracy of model predictions but more effort might be required in the
estimation of the inputs to a model. Other, less obvious, sources of information for model calibration
have been suggested, including satellite estimates of river width as an index of discharge (see Sun
et al.
,
2010). This is a technique that might benefit from the enhanced accuracy that would be available if the
planned SWOT satellite is successfully deployed in the future. A number of studies have used remote
