When to Make Mountains out of Molehills: The Pros and Cons of Simple and
Complex Model Calibration Procedures
Abstract
Earth and environmental models are relied upon to investigate system
responses that cannot otherwise be examined. In simulating physical
processes, models have adjustable parameters which may, or may not, have
a physical meaning. Determining the values to assign to these model
parameters is an enduring challenge for earth and environmental
modellers. Selecting different error metrics by which the models results
are compared to observations will lead to different sets of calibrated
model parameters, and thus different model results. Furthermore, models
may exhibit ‘equifinal’ behaviour, where multiple combinations of model
parameters lead to equally acceptable model performance against
observations. These decisions in model calibration introduce uncertainty
that must be considered when model results are used to inform
environmental decision-making. This presentation focusses on the
uncertainties that derive from the calibration of a four parameter
lumped catchment hydrological model (GR4J). The GR models contain an
inbuilt automatic calibration algorithm that can satisfactorily
calibrate against four error metrics in only a few seconds. However, a
single, deterministic model result does not provide information on
parameter uncertainty. Furthermore, a modeller interested in extreme
events, such as droughts, may wish to calibrate against more low flows
specific error metrics. In a comprehensive assessment, the GR4J model
has been run with 500,000 Latin Hypercube Sampled parameter sets across
303 catchments in the United Kingdom. These parameter sets have been
assessed against six error metrics, including two drought specific
metrics. This presentation compares the two approaches, and demonstrates
that the inbuilt automatic calibration can outperform the Latin
Hypercube experiment approach in single metric assessed performance.
However, it is also shown that there are many merits of the more
comprehensive assessment, which allows for probabilistic model results,
multi-objective optimisation, and better tailoring to calibrate the
model for specific applications such as drought event characterisation.
Modellers and decision-makers may be constrained in their choice of
calibration method, so it is important that they recognise the strengths
and limitations of their chosen approach.