10.7 Intercomparison of Methods
Few formal comparative studies of different regionalisation techniques have
been carried out. To date, published work has mostly focused on the comparison
between RCMs and statistical downscaling techniques. Early applications of RCMs
for climate change simulations (Giorgi and Mearns, 1991; Giorgi et al., 1994)
compared the models against observations or against the driving GCMs, but not
against statistical/empirical techniques.
Kidson and Thompson (1998) compared the RAMS (Regional Atmospheric Modelling
System) dynamical model and a statistical regressionbased technique. Both approaches
were applied to downscale reanalysis data (ECMWF) over New Zealand to a grid
resolution of 50 km. The statistical downscaling used a screening regression
technique to predict local minimum and maximum temperature and daily precipitation,
at both monthly and daily timescales. The regression technique limits each
regression equation to five predictors (selected from Empirical Orthogonal Functions
(EOFs) of atmospheric fields). Both monthly and daily results indicated little
difference in skill between the two techniques, and Kidson and Thompson (1998)
suggested that, subject to the assumption of statistical relationships remaining
viable under a future climate, the computational requirements do not favour
the use of the dynamical model. They also noted, however, that the dynamical
model performed better with the convective components of precipitation.
Bates et al. (1998) compared a southwestern Australia simulation using the
DARLAM (CSIRO Division of Atmospheric Research Limited Area Model) model with
a downscaled DARLAM simulation where the downscaling model had been fitted
independently to observational data. The downscaling reproduced observed precipitation
probabilities and wet and dry spell frequencies while the DARLAM simulation
underestimated the frequency of dry spells and over estimated the probability
of precipitation and the frequency of wet spells. In a climate change followon
experiment, again using both methods, Charles et al. (1999b) found a small decrease
in probability of precipitation under future climate conditions.
Murphy (1999) evaluated the UK Meteorological Office Unified Model (UM) RCM
over Europe against a statistical downscaling model based on regression. Monthly
mean surface temperature and precipitation anomalies were downscaled using
predictor sets chosen from a range of candidate variables similar to those used
by Kidson and Thompson (1998) (EOFs of atmospheric fields). The results showed
similar levels of skill for the dynamical and statistical methods, in line with
the Kidson and Thompson (1998) study. The statistical method was nominally better
for summertime estimates of temperature, while the dynamical model gave better
estimates of wintertime precipitation. Again, the conclusion was drawn that
the sophistication of the dynamical model shows little advantage over statistical
techniques, at least for present day climates.
Murphy (2000) continued the comparative study by deriving climate change projections
for 2080 to 2100 from a simulation with the HadCM2 AOGCM. The dynamical and
statistical downscaling techniques were the same regional and statistical models
as used by Murphy (1999). The statistical and dynamical techniques produced
significantly different predictions of climate change, despite exhibiting similar
skill when validated against present day observations. The study identifies
two main sources of divergence between the dynamical and statistical techniques:
firstly, differences between the strength of the observed and simulated predictor/predictand
relationships, and secondly, omission from the regression equations of variables
which represent climate change feedbacks, but are weak predictors of natural
variability. In particular, the exclusion of specific humidity led to differences
between the dynamical and statistical simulations of precipitation change. This
point would seem to confirm the humidity issue raised in Section
10.6.3 (Hewitson and Crane 1996, Crane and Hewitson, 1998, Charles et al.,
1999b; Hewitson 1999).
Mearns et al. (1999) compared RCM simulations and statistical downscaling using
a regional model and a semiempirical technique based on stochastic procedures
conditioned on weather types which were classified from circulation fields (700hPa
geopotential heights). While Mearns et al. suggest that the semiempirical approach
incorporates more physical meaning into the relationships than a pure statistical
approach does, this approach does impose the assumption that the circulation
patterns are robust into a future climate in addition to the normal assumption
that the crossscale relationships are stationary in time. For both techniques,
the driving fields were from the CSIRO AOGCM (Watterson et al., 1995). The variables
of interest were maximum and minimum daily temperature and precipitation over
centralnorthern USA (Nebraska). As with the preceding studies, the validation
under present climate conditions indicated similar skill levels for the dynamical
and statistical approaches, with some advantage by the statistical technique.
In line with the Murphy (2000) study, larger differences were also noted by
Mearns et al. (1999) when climate change projections were produced. Notably
for temperature, the statistical technique produced an amplified seasonal cycle
compared to both the RCM and CSIRO data, although similar changes in daily temperature
variances were found in both the RCM and the statistical technique (with the
statistical approach producing mostly decreases). The spatial patterns of change
showed greater variability in the RCM compared with the statistical technique.
Mearns et al. (1999) suggested that some of the differences found in the results
were due to the climate change simulation exceeding the range of data used to
develop the statistical model, while the decreases in variance were likely to
be a true reflection of changes in the circulation controls. The precipitation
results from Mearns et al. (1999) are different from earlier studies with the
same RCM (e.g., Giorgi et al., 1998) that produced few statistically significant
changes.
Extending the comparison beyond simple methodological performance, Wilby et
al. (2000) compared hydrological responses using data from dynamically and statistically
downscaled climate model output for the Animas River basin in Colorado, USA.
While not a climate change projection, the use of output from an RCM and a statistical
downscaling approach to drive a distributed hydrological model exemplify the
objective of the downscaling. The results indicate that both the statistical
and dynamical methods had greater skill (in terms of modelling hydrology) than
the coarse resolution reanalysis output used to drive the downscaling. The statistical
method had the advantage of requiring very few parameters, an attribute making
the procedure attractive for many hydrological applications. The dynmical model
output, once elevationcorrected, provided better water balance estimates than
raw or elevationcorrected reanalysis output.
Overall, the above comparative studies indicate that for present climate both
techniques have similar skill. Since statistical models are based on observed
relationships between predictands and predictors, this result may represent
a further validation of the performance of RCMs. Under future climate conditions
more differences are found between the techniques, and the question arises as
to which is “more correct”. While the dynamical model should clearly
provide a better physical basis for change, it is still unclear whether different
regional models generate similar downscaled changes. With regard to statistical/empirical
techniques, it would seem that careful attention must be given to the choice
of predictors, and that methodologies which internally select predictors based
on explanatory power under present climates may exclude predictors important
for determining change under future climate modes.
