Geoscience Reference
In-Depth Information
also be important during flood forecasting, though in a forecasting situation it can be at least partially
compensated for by the use of data assimilation (see Chapter 8).
7.18 The Value of Data in Model Conditioning
As in the quotation from Brian Ebel and Keith Loague at the head of this chapter, there have been many
exhortations for improved interaction between the field hydrologist and the modeller in providing data of
different types for improved model conditioning (Seibert and McDonnell, 2002; Weiler and McDonnell,
2004; Sidle, 2006; Fenicia
et al.
, 2007b, 2008a). However, a search of the literature will reveal very little
in the way of advice or guidance as to the value of different types of data in conditioning rainfall-runoff
models. In fact, there is also very little advice or guidance on the quality of the rainfall, evapotranspiration
and discharge data that are used to drive rainfall-runoff models. Even data sets that have been used in many
modelling studies might still have deficiencies (e.g. the Leaf River catchment example discussed in Beven,
2009b). This is, perhaps, not surprising, given the epistemic nature of the errors associated with such
data (they are differentiated as epistemic exactly because we do not know too much about their nature).
There were some early studies about how much data was required to optimise a hydrological model
(e.g. Gupta and Sorooshian, 1985; Jakeman and Hornberger, 1993) but there has been little in the context
of constraining prediction uncertainty (but see Seibert and Beven (2009) for a case of a daily conceptual
model within a GLUE framework). This is, however, actually a rather important problem because it is
relevant to the issue of planning a measurement campaign for understanding the response of an ungauged
catchment (where providing information for a decision might justify the expense).
There is also the possibility of using different types of data within a multiple criterion evaluation. A
number of authors have suggested that environmental tracer data might help constrain the representation of
runoff processes (e.g. Seibert and McDonnell, 2002; Iorgulescu
et al.
, 2005, 2007; Vache and McDonnell,
2006; Fenicia
et al.
, 2008b). In the short term such data can distinguish (with some uncertainty) between
water in the hydrograph that is event water and that which is pre-event water displaced from storage. In
the longer term, tracer data can give some indication of the residence times of water in the catchment
system (eg. McGuire and McDonnell, 2006; Botter
et al.
, 2009, 2010). Back in Chapter 1, however, the
point was made that distinguishing between event and pre-event water was not the same as distinguishing
surface and surbsurface processes; that while an ideal tracer follows the water velocities, the hydrograph
is controlled by the wave celerities (see also the kinematic wave analysis in Section 5.5.3). In terms
of constraining models, therefore, the information provided by tracer observations might be relevant to
different types of parameters (or effective storage volumes) than that provided by discharge observations
and hydrographs (see also Chapter 11). This has not always been recognised in studies that have used
tracer information in model conditioning. There are other types of data that might also need additional
model components to be useful in conditioning. Stream temperatures, for example, have been used to
infer discharge increments in river reaches but require an intepretative model, with its own uncertain
parameters to be useful. Most remote sensing imaging is also of this type.
This is an area that still demands much more research. This is probably best posed in the context of
funding for data collection. If some expense can be justified to try to reduce the uncertainty associated
with rainfall-runoff model predictions, how should that money be spent?
7.19 Key Points from Chapter 7
•
Limitations of both model structures and the data available on parameter values, initial conditions and
boundary conditions generally make it difficult to apply a rainfall-runoff model (of whatever type)
without some form of calibration or conditioning of model parameter sets.
