вход по аккаунту



код для вставкиСкачать
Int. J. Climatol. 20: 489–501 (2000)
Hadley Centre for Climate Prediction and Research, Meteorological Office, Bracknell, Berks, UK
Recei6ed 10 No6ember 1998
Re6ised 10 July 1999
Accepted 25 July 1999
Statistical and dynamical downscaling predictions of changes in surface temperature and precipitation for 2080 – 2100,
relative to pre-industrial conditions, are compared at 976 European observing sites, for January and July. Two
dynamical downscaling methods are considered, involving the use of surface temperature or precipitation simulated
at the nearest grid point in a coupled ocean–atmosphere general circulation model (GCM) of resolution 300 km
and a 50 km regional climate model (RCM) nested inside the GCM. The statistical method (STAT) is based on
observed linear regression relationships between surface temperature or precipitation and a range of atmospheric
predictor variables. The three methods are equally plausible a priori, in the sense that they estimate present-day
natural variations with equal skill.
For temperature, differences between the RCM and GCM predictions are quite small. Larger differences occur
between STAT and the dynamical predictions. For precipitation, there is a wide spread between all three methods.
Differences between the RCM and GCM are increased by the meso-scale detail present in the RCM.
Uncertainties in the downscaling predictions are investigated by using the STAT method to estimate the grid point
changes simulated by the GCM, based on regression relationships trained using simulated rather than observed values
of the predictor and the predictand variables (i.e. STAT – SIM). In most areas the temperature changes predicted by
STAT – SIM and the GCM itself are similar, indicating that the statistical relationships trained from present climate
anomalies remain valid in the perturbed climate. However, STAT – SIM underestimates the surface warming in areas
where advective predictors are important predictors of natural variability but not of climate change. For precipitation,
STAT – SIM estimates the simulated changes with lower skill, especially in January when increases in simulated
precipitation related to a moister atmosphere are not captured. This occurs because moisture is rarely a strong enough
predictor of natural variability to be included in the specification equation.
The predictor/predictand relationships found in the GCM do not always match those found in observations. In
January, the link between surface and lower tropospheric temperature is too strong. This is also true in July, when
the links between precipitation and various atmospheric predictors are also too strong. These biases represent a likely
source of error in both dynamical and statistical downscaling predictions. For example, simulated reductions in
precipitation over southern Europe in summer may be too large. © British Crown Copyright 2000.
KEY WORDS: Europe;
climate change; downscaling; climate models; surface temperature; precipitation
The problem of ‘downscaling’ output from global general circulation models (GCMs, current grid size
typically 300 km) to obtain information for localized areas has received increasing attention in recent
years (see Wilby and Wigley, 1997), motivated by the requirement of policy-makers for detailed regional
scenarios of climate change (e.g. Department of the Environment, 1996). Many downscaling techniques
are based on statistical relationships linking observations of local variables to the observed atmospheric
circulation. These relationships are then applied to the circulation simulated by a GCM in order to
generate predictions of local climate (e.g. Karl et al., 1990; von Storch et al., 1993). The use of such
* Correspondence to: Hadley Centre for Climate Prediction and Research, Meteorological Office, London Road, Bracknell, Berks
RG12 2SY, UK.
© British Crown Copyright 2000
methods is motivated by an assumption that GCMs simulate the large-scale atmospheric circulation better
than they simulate surface climate elements such as surface temperature and precipitation (e.g. Palutikof
et al., 1997), because the latter are particularly sensitive to subgrid-scale processes (convection, cloud
formation, turbulent transports, etc) which can only be represented approximately in the models.
The alternative is dynamical downscaling, in which predictions of (say) site-specific temperature or
precipitation are derived from values of surface temperature or precipitation simulated at nearby GCM
grid points. Here the assumption is that predictions based explicitly on relevant physical processes will be
better than those in which the underlying physics is only represented implicitly, via empirical relationships.
In practice, even dynamical downscaling will require an empirical element because the grid box predictor
variables are effectively spatial means for an area of 300× 300 km2 (e.g. Osborn and Hulme, 1997),
whereas the predictands are local values influenced by spatial heterogeneities in the regional physiography
(orography, coastlines, soil and vegetation types, etc). This mismatch in scales can be addressed by nesting
a high resolution regional climate model (RCM) inside the GCM (e.g. Giorgi et al., 1993; McGregor and
Walsh, 1993; Jones et al., 1995). Unfortunately, the resolution of RCMs (typically 50 km) is not yet
high enough to capture all aspects of the regional forcing. In addition, most research groups only have
a few years experience in developing and using RCMs, which is not long enough to optimize the
performance of such complex tools.
Despite their potential limitations, the use of downscaling schemes is essential if plausible regional
scenarios are to be constructed for impact assessments (Watson et al., 1998). It is, therefore, important to
compare statistical and dynamical downscaling methods in order to understand the reasons for the
differences in their predictions and, if possible, to identify the optimum technique.
Kidson and Thompson (1998) and Murphy (1999) have recently performed such comparisons by
assessing downscaling estimates of variability observed in the recent historical record. In both cases,
relevant model integrations were forced with time series of analyses of the observed atmospheric
circulation in order to remove the influence of model circulation biases. Kidson and Thompson (1998)
compared statistical and RCM estimates of daily and monthly temperature and precipitation observed at
stations in New Zealand, and found that, on average, the methods gave similar levels of skill. Murphy
(1999) compared three methods of downscaling values of surface temperature and precipitation observed
at 976 European observing stations. The predictor variables were:
(i) temperature or precipitation simulated at the nearest GCM grid point to the target location;
(ii) temperature or precipitation simulated at the nearest grid point of an RCM nested inside the GCM;
(iii) values of various atmospheric variables, including regional temperature, moisture, properties of the
near-surface wind and vertical stability, and large-scale patterns of variability in mean sea level
pressure. Subsets of these were converted into predictions of local temperature or precipitation, using
linear regression.
The dynamical methods (i) and (ii) were compared with the statistical method (iii) in terms of the ability
to predict interannual variations in monthly means. All three methods gave roughly equal levels of skill,
supporting Kidson and Thompson’s (1998) result. Average explained variances were higher in winter than
in summer and were higher for temperature than for precipitation. The dynamical methods were also
compared by evaluating errors in the predictions of daily distributions and the climatological mean and
variance of monthly averages. The RCM yielded improved skill relative to the GCM in estimates of both
climatological mean temperature and the daily and interannual variability of precipitation.
In this paper, Murphy’s (1999) study is extended to consider downscaling predictions of climate change.
For simplicity, only changes in multi-annual monthly means are considered. The intention is to determine
the spread of changes that can be generated from a single GCM simulation, using plausible downscaling
techniques. In order to explain differences between the statistical and the dynamical predictions a further
experiment is carried out in which the statistical method is used to predict the changes simulated by the
GCM. This provides an opportunity to assess the extent to which simple statistical regression equations,
calibrated from natural variability, can reproduce climate changes driven by changes in radiative forcing
© British Crown Copyright 2000
Int. J. Climatol. 20: 489 – 501 (2000)
in a complex climate model. It also allows the simulated predictor/predictand relationships to be
compared with corresponding observed relationships, thus providing an additional method of validating
the GCM itself.
The model integrations consist of a coupled ocean–atmosphere GCM simulation of climate change for
2080–2100, in which carbon dioxide (CO2) is increased to four times the pre-industrial concentration, and
a corresponding RCM integration driven by output saved from the GCM. The integrations simulate a
large change in climate in which the spatially-averaged surface warming over Europe ranges from 4 to
6°C, depending on season.
The models and experimental design are summarized in the following section. The downscaling
methodologies and the results are described in Sections 3 and 4. In Section 5 the statistical method is used
to predict the grid point changes simulated by the GCM, with concluding remarks in Section 6.
The GCM is the ‘HADCM2’ version of the Hadley Centre coupled model. The atmosphere and ocean
components are both grid point models of resolution 2.5°× 3.75° latitude/longitude. The ocean includes
an embedded sea-ice model. The ocean and atmosphere are coupled via surface exchanges of heat, water,
momentum and wind mixing energy.
A control integration of the GCM with CO2 held fixed was run for more than 1500 years with no
significant climate drift in the distributions of sea surface temperature or sea-ice. During the control
integration flux adjustments were added to the surface heat and water fluxes in order to maintain sea
surface temperatures and salinities close to the observed climatological values. The flux adjustments,
which varied with position and season but not from year to year, were also applied in the climate change
integration. See Johns et al. (1997) for a full description of the GCM and the control simulation.
The perturbed (climate change) integration of the GCM was initialized from a point 150 years into the
control run: CO2 was increased to represent the combined effect of changes in all greenhouse gases from
1860 to 1990, following which a further 1% per year compounded increase was applied from 1990 to 2100
(Figure 1). This creates a scenario of future greenhouse forcing similar to that of the IPCC IS92a
(Mitchell and Gregory, 1992). We consider changes for 2080–2100 relative to the control integration.
During this 20-year period, the CO2 concentration reaches four times the pre-industrial value (Figure 1).
Figure 1. Changes in CO2 relative to 1860 imposed during the perturbed integration of the GCM
© British Crown Copyright 2000
Int. J. Climatol. 20: 489 – 501 (2000)
Although this experimental design implies that the control integration should simulate pre-industrial
conditions, the CO2 concentration (323 ppm) and the sea surface temperatures used were actually
appropriate for the second half of the 20th Century (see Johns et al., 1997). Similarly, the observed data
used to calibrate required downscaling relationships were taken from the late 20th Century (see Section
3). It is assumed that these relationships are valid for the pre-industrial climate even though they are
obtained from 20th Century data.
The RCM is essentially identical to the GCM in the formulation of both the grid-scale dynamics and
the subgrid-scale physics (cloud and precipitation formation, radiative transfer, etc), differing only in
horizontal resolution (50 km cf. 300 km) and time step (5 min cf. 30 min). Its performance in
simulations for Europe and India has been documented in Jones et al. (1995, 1997), Bhaskaran et al.
(1996, 1998) and Noguer et al. (1998). The RCM is driven by a continuous time series of boundary
conditions archived from a previous integration of the GCM (i.e. ‘one-way nesting’). The required
boundary conditions are distributions of sea surface temperature and sea-ice and values of the atmospheric state variables (surface pressure plus temperature, moisture and wind components at each level)
at points bordering the RCM domain. The lateral boundary coupling is implemented by relaxing the state
variables in the RCM towards GCM values across a four-point buffer zone. See Jones et al. (1995) for
further details. In the present study, the RCM uses the European integration domain employed by
Noguer et al. (1998). This is designed to constrain the large-scale circulation to follow that of the driving
model as closely as possible without restricting the generation of meso-scale features. A 30-year time slice
from the GCM was used to drive the control integration of the RCM. The perturbed RCM integration
was driven by output for 2080 – 2100 from the perturbed GCM integration. Hereafter, the terms ‘control’
and ‘perturbed’ refer to the 30-year and 20-year RCM integrations, respectively, or to the corresponding
time slices from the GCM simulations.
Downscaling predictions of changes in monthly precipitation and temperature for January and July are
given for point locations corresponding to 976 European observing stations (see Figure 2). As in Murphy
(1999), downscaling estimates X of station precipitation or temperature are derived from: (i) simulated
values of the predictand variable at the nearest GCM land point to the observing station (i.e. X=XGCM);
(ii) simulated values at the nearest RCM land point (X= XRCM); and (iii) statistical specification
equations linking the predictand variable to values of atmospheric predictor variables simulated by the
GCM over the domain of Figure 2 (X = XSTAT).
In the interests of brevity only changes in climatological monthly means are considered, although the
importance of changes in variability is also recognized (Katz and Brown, 1992).
Figure 2. Domain of regional model (excluding boundary buffer zone) and locations of observing stations. The locations of global
model (GCM) grid boxes over land are also shown
© British Crown Copyright 2000
Int. J. Climatol. 20: 489 – 501 (2000)
Plate 1. Spatial distributions of downscaling predictions of changes in observed diurnal mean surface air temperature TBAR (°C,
top six panels) and precipitation PRCP (mm day − 1, bottom six panels) for 2080 – 2100 relative to the pre-industrial period, for
January (JAN) and July (JUL). GCM—predictions made using the value of TBAR or PRCP simulated at the nearest GCM land
point to the observing station; RCM—predictions made using the value of TBAR or PRCP simulated at the nearest RCM land
point to the observing station; STAT—predictions derived from atmospheric predictor variables simulated by the GCM using
statistical regression. The distributions are plotted on a 0.5° × 0.5° grid, where the values at each grid point are inverse-distance
weighted means of the nearest five stations
© British Crown Copyright 2000
Int. J. Climatol. 20(5) (2000)
Plate 2. The left hand panels show changes in surface air temperature (°C) and precipitation (mm day − 1) simulated by the GCM
for 2080 – 2100 relative to the pre-industrial period, for January and July (repeated from Plate 1). The right hand panels show
statistical predictions of the simulated changes derived from atmospheric predictor variables using the STAT – SIM method
© British Crown Copyright 2000
Int. J. Climatol. 20(5) (2000)
3.1. Details of statistical method
The selection of predictors and calibration of regression relationships required for statistical downscaling is described in detail by Murphy (1999). The procedure is briefly summarized in this paper. The set
of predictor variables is listed in Table I. The CIRC predictor, which consists of a linear combination of
projections onto the first five empirical orthogonal functions (EOFs) of the observed mean sea level
pressure field, is included to represent the influence of large-scale flow regimes on either local temperature
or precipitation. The other predictors consist simply of the value of the relevant variable simulated by the
GCM at the nearest land point to the target station. The set of predictors is the same as that used in
Murphy (1999) with the addition of RH850, the inclusion of which provides a small improvement to the
skill of precipitation downscaling.
The predictors were linked to the predictands using linear specification equations with coefficients
determined by least squares regression, based on observations from 1983 to 1994. This period was used
for both calibration and verification of the relationships by employing a cross-validation procedure.
Various candidate predictor sets were tested (see Table II), with the number of included predictors varying
between one and four. The required values of the predictors were obtained from a GCM integration for
1983–1994 (GCM – ASSIM) during which values of surface pressure and atmospheric winds and temperatures were continuously assimilated from a time series of UK Meteorological Office operational analyses.
Because the atmospheric circulation in GCM – ASSIM was constrained to remain very close to the driving
analyses, values of the predictors obtained from this integration were as close as possible to the true
observations made at the relevant model grid point. Note, however, that the Q850, KIND and RH850
predictors are likely to be influenced by model biases because no moisture observations were assimilated
during GCM – ASSIM. This was because the available moisture analyses were not deemed to be
sufficiently reliable. In future, more reliable moisture information will be obtained by using values from
ECMWF re-analyses (Gibson et al., 1997). Use of re-analysis products will also allow the use of a longer
period to calibrate the specification equations, thus better sampling the low frequency variability present
in the observed record.
For a given station, predictand variable and month the best predictor set was identified as that which
achieved the highest correlation between estimated and observed monthly anomalies in the calibration
dataset. However, the risk of identifying a statistically unreliable relationship was reduced by correcting
the correlations to allow for the average drop in skill between calibration and independent verification
(details in Murphy, 1999). The frequency of selection amongst a range of candidate predictor sets is
shown in Table II for both January and (in parentheses) July. For temperature, the best individual
predictor is T850, which almost always appears in the chosen predictor set. In January, T850 is selected
on its own at 38% of stations and in combination with one or more advective predictors at the remaining
62% of stations. In July, advection is weaker and T850 is selected on its own at 69% of stations. For
Table I. Predictors used for statistical downscaling
Linear combination of projections onto large-scale flow patterns
Regional value of near-surface (10 m) wind speed
Regional value of near-surface westerly wind
Regional value of near-surface southerly wind
Regional value of near-surface vorticity
Regional value of temperature at 850 mb
Regional value of specific humidity at 850 mb
Regional value of relative humidity at 850 mb
Regional value of index of vertical stability, defined as
where T700 and T500 are temperature at 700 mb, 500 mb, respectively, and TD850 and TD700 are
dew-point at 850 mb, 700 mb, respectively
© British Crown Copyright 2000
Int. J. Climatol. 20: 489 – 501 (2000)
Table II. Frequency of selection (%) for statistical predictor sets in January (July)
Predictor set
Predictor set
precipitation, the selections are distributed more evenly between the alternative predictor sets, reflecting
the wide range of atmospheric properties related to condensation and precipitation formation in
convective or synoptic-scale ascent. The most important regional predictors are vorticity (VORT),
stability (KIND) and relative humidity (RH850). In January, the large-scale circulation (CIRC) is also
important: it is rarely selected on its own, but appears frequently in combination with one or more
regional predictors.
In independent verification, the average correlation between estimated and observed monthly temperature anomalies varies between 0.73 and 0.92 (Murphy, 1999, his figure 14a). For precipitation, skill is
lower but is still comparable with that obtained from dynamical downscaling (Murphy, 1999, his figure
14b). Average correlations range from 0.43 to 0.66, the lowest values occurring in summer.
3.2. Downscaling climate changes
Predictions of multi-annual monthly mean changes, denoted by DX, were obtained from each of the
three downscaling methods using the 30-year control and the 20-year perturbed simulations. For the
dynamical methods, DX consisted simply of a 20-year monthly mean grid point value of either surface
temperature or precipitation from the perturbed simulation minus the corresponding 30-year mean from
the control. For statistical downscaling, DX was obtained from the specification equation
DXSTAT =aDX1 +bDX2 +gDX3 +. . . ,
where DX1, 2, 3 etc. are changes in predictor variables (e.g. CIRC, T850, UWND) and a, b, g etc. are
regression coefficients determined from 1983–1994 observations, as described above. The required
average values DX1, 2, 3 etc. were obtained by differencing multi-annual means of the predictors from the
perturbed and the control integrations.
In the case of CIRC, predicted changes DXCIRC were obtained by projecting simulated anomalies in the
PMSL field onto EOFs of observed PMSL, i.e.
DXCIRC = % akPCk,
© British Crown Copyright 2000
Int. J. Climatol. 20: 489 – 501 (2000)
where PCk is the projection of EOFk onto the PMSL anomalies and the ak are regression coefficients
determined from the 1983 – 1994 observations. The multi-annual mean change DXCIRC was obtained by
projecting onto the multi-annual mean change in PMSL, i.e. DPMSL.
Plate 1 shows downscaled changes for both January and July. Accompanying statistics averaged over all
stations are provided in Table III. The changes predicted by the statistical method are referred to as
STAT, with GCM and RCM used to denote the dynamical predictions. The results are presented as
gridded 0.5°× 0.5° spatial distributions, where the value at each grid point is an inverse-distance weighted
mean of the values at the five nearest stations.
In January, all of the predictions show a substantial surface warming at most locations. The two
dynamical predictions generally agree to within less than 1°C (Table III), reflecting the use of a small
integration domain to constrain the large-scale circulation in the RCM to follow that of the GCM. The
differences between STAT and the dynamical predictions are larger, amounting to 2°C on average.
This is partly explained by the smaller spatially-averaged warming predicted by STAT (3.8°C, cf. 5.0°C
and 5.3°C for the GCM and the RCM, respectively). However, even if the spatial mean is removed the
root mean square (r.m.s.) difference between STAT and either GCM or RCM still amounts to 1.5°C.
This reflects differences in the patterns of warming (Plate 1). In particular, the STAT pattern shows a
much smaller warming over both western and eastern Europe. These are areas where advective predictors
receive relatively high weight in the specification equation, because the skill of T850 in estimating
historical interannual variations is not as high as in other parts of the domain (Murphy, 1999). The
divergence between STAT and the dynamical predictions occurs because changes in circulation do not
account for a large proportion of the simulated surface temperature increase. On the other hand, changes
in simulated T850 and surface temperature are much more closely linked (see Section 5); therefore, in
areas where T850 is the dominant predictor (e.g. over central and northern Europe), STAT, GCM and
RCM will all give similar predictions. These results are consistent with Jones et al. (1997), who found that
circulation anomalies generally contributed less than 1°C to the equilibrium response to doubled CO2 over
Europe, whereas the warming arising from diabatic processes in the absence of circulation changes
typically contributed 3 – 4°C. It is worth emphasizing, however, that these results apply to the timea6eraged changes: for example, they do not imply that the strong influence of circulation anomalies on the
interannual 6ariability of temperature will be reduced in the future.
In July, the warming predicted by the dynamical methods is again similar in most places, although
differences exceed 1°C over much of central Europe. This is because the RCM circulation in summer is
not so strongly constrained by the lateral boundary forcing supplied by the GCM (although the
differences between the models are smaller than in the previous experiment carried out by Jones et al.
(1997)). As in January, however, differences between RCM and GCM are usually smaller than differences
between STAT and either RCM or GCM. In general, STAT predicts a smaller warming than the
dynamical methods. The largest differences exceed 3°C and occur over northern Africa, the Iberian
peninsula and southeastern Europe (Plate 1). These are all areas where STAT explains a relatively small
portion (often less than 50%) of the predictand variance in the calibration dataset, mainly because of the
lower than average skill in the T850 predictor.
Table III. R.M.S. Differences Between Downscaled Changes
Surface temperature JAN (°C)
Surface temperature JUL (°C)
Precipitation JAN (mm day−1)
Precipitation JUL (mm day−1)
© British Crown Copyright 2000
Int. J. Climatol. 20: 489 – 501 (2000)
For precipitation, Table III shows r.m.s. differences of 0.8–1.0 mm day − 1 between the predicted
changes. These differences are generally larger than the mean magnitude of the change predicted by a
particular method, which lies in the range 0.6 – 0.7 mm day − 1 for both GCM and RCM and 0.4–0.5 mm
day − 1 for STAT. In contrast to the results for temperature, the RCM–GCM differences are comparable
with those of RCM – STAT and GCM – STAT. This is because the patterns of precipitation change
simulated by the RCM contain a significant component on scales too fine to be resolved on the GCM
grid, whereas this meso-scale signal is less important for temperature (Jones et al., 1997). When the
meso-scale component is removed from the RCM precipitation fields, the RCM–GCM differences reduce
to 0.4 mm day − 1 in both January and July. (The meso-scale signal is not well represented in Plate 1
because the plotted 6alues are a6erages o6er fi6e stations, which often represent an area much larger than an
RCM grid box).
This result is consistent with Plate 1, which shows that the broad-scale patterns of precipitation change
in both RCM and GCM are similar to each other. On the other hand, the STAT patterns lack some of
the features apparent in the model distributions. In January, for example, both dynamical methods predict
spatially-averaged increases exceeding 0.4 mm day − 1, whereas STAT predicts a mean change close to
zero. In July, RCM and GCM predict reductions exceeding 0.5 mm day − 1 over much of southwestern
Europe, whereas STAT predicts much smaller reductions or, in some places, small increases. The reasons
for the differences between the statistical and the dynamical predictions are discussed in Section 5.
In Section 4, the application of three plausible downscaling strategies led to a distribution of climate
change scenarios containing a substantial spread, particularly for precipitation. Failure of the alternative
methods to agree on a common scenario creates a major source of uncertainty for regional impact
assessments; hence, it is important to understand the reasons for differences between the downscaled
patterns of change. In this paper, this is attempted via an idealized experiment in which the STAT method
is used to predict grid point changes in surface temperature and precipitation simulated by the GCM. This
‘perfect model’ approach (hereafter STAT – SIM) is a test of whether statistical specification equations,
calibrated under present climate conditions, remain valid when applied in perturbed future climates: such
a test cannot be carried out in reality until observations of future climate change become available.
Furthermore, comparisons can be made between the simulated and observed predictor/predictand
relationships, which provides important information about the physical realism of the GCM.
STAT – SIM predictions were obtained for January and July using the specification equation (1).
Predictions were made for the GCM land point nearest to each observing site in Figure 2. The procedure
was identical to that used for the STAT predictions of Plate 1, except that the required regression
coefficients (a, b, g etc. for Equation (1) and a1, . . . , a5 needed to evaluate the CIRC predictor using
Equation (2)) were obtained from the 30-year control integration of the GCM by linking monthly values
of simulated grid point precipitation or temperature to simulated values of the predictor variables. In the
case of the CIRC predictor, the EOF patterns needed for the calculation of principal components of
PMSL in Equation (2) were also determined from the GCM control simulation rather than from
5.1. Comparison of specified and simulated changes
The results (Plate 2) show that STAT – SIM generally predicts the simulated temperature changes quite
well. The r.m.s. difference amounts to 1.6°C in January (32% of the mean warming) and 0.8°C in July
(17% of the mean warming). Whilst the errors are not trivial, a method which could predict the observed
change to within 20 – 30% would be regarded as extremely useful for impact assessments. Because T850 is
the dominant predictor (as when the specification equations are calibrated from observations; Table II),
the predictions given by STAT – SIM and by T850 alone (not shown) are very similar in most places. A
© British Crown Copyright 2000
Int. J. Climatol. 20: 489 – 501 (2000)
notable exception to this occurs over western and southwestern Europe in January, where T850 has a
relatively low weight in the specification equations. Here the prediction based on T850 alone is more
accurate than the STAT – SIM prediction because the latter is strongly dependent on advective predictors,
which play a large role in determining interannual variability but a much smaller role in determining the
mean climate change.
In January, the STAT – SIM precipitation predictions show an r.m.s. difference from the simulated
change of 0.66 mm day − 1. This exceeds the mean magnitude of the simulated change; hence, by this
criterion, the STAT – SIM predictions are less skilful than a prediction of zero change everywhere. One
problem is the failure of STAT – SIM to reproduce the spatial average increase of 0.47 mm day − 1
simulated by the GCM. This change is part of a general increase in wintertime precipitation poleward of
45°N, which is a common feature of GCM simulations of CO2-driven global warming (e.g. Intergovernmental Panel on Climate Change (IPCC), 1990, chapter 5) arising from increased poleward moisture
transport in a warmer, moister atmosphere (e.g. Manabe and Wetherald, 1975). Here the influence of
increased moisture can be estimated using the relationship between Q850 and precipitation calibrated
from the control integration. The mean predicted increase associated with Q850 turns out to be 0.78 mm
day − 1, which is more than sufficient to explain the spatial mean change in the GCM (see also Crane and
Hewitson, 1998). Unfortunately, the STAT – SIM prediction is insensitive to increases in specific humidity
because Q850 is only included in the selected predictor set on 1.3% of occasions. This is because other
predictors are better at explaining the interannual anomalies of precipitation simulated in the control run.
It should be noted, however, that the Q850/precipitation relationships do not yield particularly good
predictions of the spatial distribution of the simulated changes; thus, the distribution errors apparent in
the STAT – SIM predictions of Plate 2 (e.g. the absence of a band of increases \0.5 mm day − 1 near to
the northern and the northeastern boundaries of the domain) are not necessarily attributable to the
absence of specific humidity in the chosen predictor sets.
In July, the r.m.s. difference between the simulated precipitation changes and the STAT – SIM estimates
is 0.47 mm day − 1, i.e. 70% of the mean magnitude of the simulated change. Thus, typical errors in
STAT – SIM are significantly larger than in the corresponding temperature estimates. Nevertheless, the
large-scale pattern of the simulated precipitation changes is quite well captured, particularly in the area
with reductions of \ 0.5 mm day − 1 covering southwest Europe. Such reductions have been found in
several recent greenhouse-warming simulations (Kattenberg et al., 1996) and are generally largest in areas
where soil moisture is reduced as a result of enhanced evaporation, reductions in cloud cover and other
factors (Wetherald and Manabe, 1995). Associated with these changes are large reductions in RH850,
which is the key predictor responsible for the success of the STAT – SIM predictions in this region. If
predictor sets including RH850 are excluded when selecting the specification equations, the predicted
reductions over southwest Europe become much smaller.
5.2. Comparison between simulated and obser6ed predictor/predictand relationships
Several recent studies have assessed climate models in terms of their ability to reproduce observed links
between large-scale circulation anomalies and regional precipitation over selected parts of Europe (von
Storch et al., 1993; Noguer, 1994; Busuioc et al., 1999; Wilby and Wigley, 1999) have compared simulated
and observed correlations between precipitation in parts of the US and various atmospheric predictors.
All of these studies report errors in the predictor/predictand relationships in the models. Such errors imply
deficiencies in the simulation of relevant physical relationships, with adverse consequences for the
reliability of both dynamical and statistical downscaling predictions.
In this paper, the observed and the simulated correlations with each of the predictor variables are
compared for both surface temperature (Table IV) and precipitation (Table V). The implications of the
comparisons for downscaling estimates of climate change are also assessed by comparing the STAT
predictions of Plates 1 and 2. The STAT – SIM distributions of Plate 2 are predictions of simulated surface
temperature and precipitation for 2080 – 2100. Those in Plate 1, hereafter STAT – REAL, are predictions
of actual (as yet unobserved) temperature and precipitation for 2080–2100. Hence, differences between
© British Crown Copyright 2000
Int. J. Climatol. 20: 489 – 501 (2000)
Table IV. Average variance of surface temperature explained by individual predictors
STAT – SIM and STAT – REAL arise exclusively from differences between the predictor/predictand
relationships found in the GCM and observations.
For surface temperature, errors in the simulated correlations with advective predictors are associated
with systematic biases in the climatological mean circulation. In January, for example, the GCM fails to
reproduce the observed extension over northern Europe of the strong southwesterly flow on the southern
flank of the Icelandic low (not shown). This results in weaker than observed correlations between surface
temperature and CIRC (Table IV). In summer, the mean westerly flow over western and central Europe
is weaker than observed, leading to weaker than observed correlations with UWND and WSPD.
However, for the STAT estimates of climate change the most important issue is the variance explained by
T850, because this is usually the dominant predictor in the specification equation. In general, the
simulated correlation between surface temperature and T850 is too high (Table IV), especially in July. For
this reason, the STAT – SIM estimates of surface warming are consistently higher than those of
STAT – REAL (Plate 2 cf. Plate 1). The spatial average difference is 1.1°C in January and 1.4°C in July.
These differences imply errors in the surface warming simulated by the GCM. However, it does not
necessarily follow that the dynamical downscaling estimates of Plate 1 are less reliable than those of
STAT – REAL. This is because the deficiencies in the simulated surface temperature–T850 relationship
could introduce biases into the predicted T850 changes as well as the surface changes. Nevertheless, the
present results demonstrate that such deficiencies represent a key source of uncertainty in both the
dynamical and the statistical downscaling results; hence, their removal is an obvious priority for the future
development of the model.
For precipitation as predictand, the simulated and observed correlations with the predictor variables
correspond reasonably well in January when averaged over all points (Table V). The large-scale patterns
of change in the two STAT predictions also correspond fairly well, even though there are regional
differences introduced by differences in regional values of the predictor/predictand correlations (not
shown). In July, the explained variances are much smaller in the observations than in the model for all
the leading predictors (Table V), leading to larger differences between STAT – SIM and STAT – REAL.
The differences are particularly marked over southern Europe, where STAT – REAL predicts much
smaller reductions in precipitation than STAT – SIM. As discussed earlier, the large reductions predicted
by STAT – SIM arise from reductions in RH850. However, the observed link between RH850 and
precipitation is much weaker than the simulated link, so the reductions in RH850 do not trigger such large
precipitation reductions in STAT – REAL.
One possible explanation for the stronger predictor/predictand links in the GCM is that the GCM
predictand variables are areal means whereas the observational predictands are point values. This was
tested by reconstituting the observed predictands as averages over all stations within a 150 km radius of
the target location. The average number of stations from which the new predictands are constructed is
4.95: this is sufficient to provide a reasonable indication of the effect of area-averaging, even though a
greater number of stations would be required to form a true grid box mean (Osborn and Hulme, 1997).
Comparison of the last two columns of Table V shows that the explained variances do increase when
© British Crown Copyright 2000
Int. J. Climatol. 20: 489 – 501 (2000)
Table V. Average variance of precipitation explained by individual predictors
Observed relationships,
area-mean predictands
pseudo-area mean predictands are used; however, the increases are not large enough to explain the
differences from the corresponding simulated relationships. This suggests that simulated precipitation
could be too sensitive to variations in many of the large-scale and regional atmospheric properties
represented in Table V, in which case the changes predicted by the GCM (shown in Plate 1 and repeated
in Plate 2) could be too large. A similar suggestion has been made by Wilby and Wigley (1999), who
found that precipitation in the present GCM depends too strongly on moisture at locations in the US.
This paper compared downscaling predictions of greenhouse gas-driven climate change for 2080–2100,
relative to pre-industrial conditions. Changes in climatological mean surface temperature and precipitation at 976 European observing sites were predicted by two dynamical downscaling methods and one
statistical method. The dynamical methods involved the use of temperature or precipitation simulated at
the nearest grid point to the target location in a global coupled ocean–atmosphere GCM and a high
resolution regional atmospheric model (RCM) nested inside the GCM. For statistical downscaling
(STAT), predictor sets were chosen for each site from a range of atmospheric properties. The specification
equations linking the predictors to the predictand were calibrated from observations using linear
regression. The three downscaling methods can be regarded as equally plausible candidates for use in
studies of climate change impacts, having been found to show similar skill when applied under present
climate conditions (Murphy, 1999).
The main results are:
(i) For temperature, STAT is the outlier of the three methods, predicting changes which differ from the
dynamical methods by typically 40 – 50%.
(ii) For precipitation, there is a wide spread between all three methods. Differences between the
dynamical predictions are increased by the fine-scale detail present in the RCM distributions.
The results illustrate how a wide range of local scenarios can be constructed from a single GCM
simulation of climate change. In order to narrow the uncertainty it is necessary to assign estimates of
reliability to the different predictions, although this is difficult when the methods perform with equal skill
under present climate conditions. In this paper, an attempt was made to understand the differences
between the statistical and the dynamical predictions using an idealized experiment in which the STAT
method is used to predict the changes in surface temperature and precipitation simulated by the GCM.
This was carried out by calibrating the specification equations using simulated rather than observed
values of the predictors and predictands. This experiment yielded the following results:
© British Crown Copyright 2000
Int. J. Climatol. 20: 489 – 501 (2000)
(i) In most areas the specification equations estimate simulated future temperature changes reasonably
well, with a typical error of 20 – 30%.
(ii) There are local exceptions to (i), notably in regions where advective predictors are dominant in
explaining natural variability but contribute less strongly to climate change.
(iii) For precipitation, the specification equations are generally less successful in estimating the simulated
change. This is particularly applicable in January when STAT fails to capture a spatial mean increase
in simulated precipitation arising from increases in moisture in the warmer atmosphere. Unfortunately, moisture is rarely selected in the specification equation because it does not explain enough of
the variance of precipitation in the control simulation.
(iv) The link between surface and 850 mb temperature (T850) in the GCM is stronger than the
corresponding link found in observations. This is likely to lead to errors in both the dynamical and
the statistical estimates of surface warming. Also, the strength of the link in the model implies that
the assumption underlying statistical downscaling—namely that GCMs simulate values of the
atmospheric predictor variables better than they simulate values of the surface predictand variables—may not be justified in the case of T850.
(v) For precipitation, the simulated predictor/predictand correlations are too high in July, suggesting
that simulated precipitation is too sensitive to changes in the large-scale circulation and regional
vorticity, stability and relative humidity (see also Wilby and Wigley, 1999). For example, the
simulated reductions in precipitation over southern Europe in summer, which are linked to large
reductions in relative humidity, may be too large.
In order to reduce the uncertainty associated with downscaling predictions, the models (GCMs and
RCMs) must be improved by reducing biases in the simulation of present climate and improving the
representation of climate change feedbacks, particularly those involving cloud (e.g. Senior, 1999). In a
downscaling context, progress should be reflected in improved correspondence with observed predictor/
predictand relationships and also in a reduced spread between regional predictions from different GCMs
(see Kattenberg et al., 1996). Various possibilities exist for improving statistical downscaling methods,
including increasing the range of predictor variables, using longer calibration series and investigating
alternative functional forms for the specification equations. For example, recent studies suggest that the
present method could be improved by including mid-tropospheric divergence and moisture as predictors
(Cavazos, 1999; Wilby and Wigley, 1999). Ultimately, increased confidence in estimates of regional
climate change will only be established by convergence between the dynamical and the statistical
predictions or by the emergence of clear evidence supporting the use of a single preferred method.
The author wishes to thank the Climate Prediction Center, Washington DC, and the National Center for
Atmospheric Research, Boulder, CO, for making available the observed station data. This work was
supported by the Department of the Environment, Transport and the Regions under Contract PECD
Bhaskaran, B., Jones, R.G., Murphy, J.M. and Noguer, M. 1996. ‘Simulations of the Indian summer monsoon using a nested
regional climate model: domain size experiments’, Clim. Dyn., 12, 573 – 587.
Bhaskaran, B., Jones, R.G. and Murphy, J.M. 1998. ‘Intraseasonal oscillation in the Indian summer monsoon simulated by global
and nested regional climate models’, Mon. Wea. Re6., 126, 3124 – 3134.
Busuioc, A., von Storch, H. and Schnur, R. 1999. ‘Verification of GCM-generated regional seasonal precipitation for current climate
and of statistical downscaling estimates under changing climate conditions’, J. Clim., 12, 258 – 269.
Cavazos, T. 1999. ‘Large-scale circulation anomalies conducive to extreme precipitation events and derivation of daily rainfall in
northeastern Mexico and southeastern Texas’, J. Clim., 12, 1506 – 1523.
Crane, R.G. and Hewitson, B.C. 1998. ‘Doubled CO2 precipitation changes for the Susquehanna Basin: down-scaling from the
Genesis general circulation model’, Int. J. Climatol., 18, 65 – 76.
Department of the Environment. 1996. Re6iew of the Potential Effects of Climate Change in the United Kingdom, HMSO, London.
© British Crown Copyright 2000
Int. J. Climatol. 20: 489 – 501 (2000)
Gibson, R.K., Kållberg, P., Uppala, S., Hernandez, A., Nomura, A. and Serrano, E. 1997. ECMWF Re-Analysis Project Report
Series, 1, ERA Description, ECMWF.
Giorgi, F., Bates, G.T. and Nieman, S.J. 1993. ‘The multiyear surface climatology of a Regional Atmospheric Model over the
western United States’, J. Clim., 6, 75–95.
Intergovernmental Panel on Climate Change (IPCC). 1990. Scientific Assessment of Climate Change, Cambridge University Press,
Johns, T.C., Carnell, R.E., Crossley, J.F., Gregory, J.M., Mitchell, J.F.B., Senior, C.A., Tett, S.F.B. and Wood, R.A. 1997. ‘The
second Hadley Centre coupled ocean–atmosphere GCM: model description, spinup and validation’, Clim. Dyn., 13, 103 – 134.
Jones, R.G., Murphy, J.M. and Noguer, M. 1995. ‘Simulation of climate change over Europe using a nested regional climate model.
I: assessment of control climate, including sensitivity to location of lateral boundaries’, Q.J.R. Meteorol. Soc., 121, 1413 – 1449.
Jones, R.G., Murphy, J.M., Noguer, M. and Keen, A.B. 1997. Simulation of climate change over Europe using a nested regional
climate model. II: comparison of driving and regional model responses to a doubling of carbon dioxide’, Q.J.R. Meteorol. Soc.,
123, 265 – 292.
Karl, T.R., Wang, W.-C., Schlesinger, M.E., Knight, R.W. and Portman, D. 1990. A method of relating general circulation model
simulated climate to the observed local climate. Part I: seasonal statistics’, J. Clim., 3, 1053 – 1079.
Kattenberg, A., Giorgi, F., Grassl, H., Meehl, G.A., Mitchell, J.F.B., Stouffer, R.J., Tokioka, T., Weaver, A.J. and Wigley, T.M.L.
1996. Climate models—projections of the future’ in Houghton, J.T., Meiro Filho, L.G., Callendar, B.A., Harris, N., Kattenberg,
A. and Maskell, K. (eds), Climate Change 1995. The Science of Climate Change. The Second Assessment of the Intergo6ernmental
Panel on Climate Change, Cambridge University Press, Cambridge, pp. 285 – 357.
Katz, R.W. and Brown, B.G. 1992. ‘Extreme events in a changing climate: variability is more important than averages’, Clim.
Change, 21, 289 – 302.
Kidson, J.W. and Thompson, C.S. 1998. ‘A comparison of statistical and model-based downscaling techniques for estimating local
climate variations’, J. Clim., 11, 735–753.
Manabe, S. and Wetherald, R.T. 1975. ‘The effects of doubling the CO2 concentration on the climate of a general circulation model’,
J. Atmos. Sci., 32, 3–15.
McGregor, J.L. and Walsh, K.J. 1993. ‘Nested simulations of perpetual January climate over the Australian region’, J. Geophys.
Res., 99 (D10), 20889–20905.
Mitchell, J.F.B. and Gregory, J.M. 1992. ‘Climatic consequences of emissions and a comparison of IS92a and SA90’, in Houghton
J.T., Callendar, B.A. and Varney, S.K. (eds), Climate Change 1992. The Supplementary Report to the IPCC Scientific Assessment,
Cambridge University Press, Cambridge, pp. 171–182.
Murphy, J.M. 1999. ‘An evaluation of statistical and dynamical techniques for downscaling local climate’, J. Clim., 12, 2256 – 2284.
Noguer, M. 1994. ‘Using statistical techniques to deduce local climate distributions. An application for model validation’, Met.
Apps., 1, 277 – 287.
Noguer, M., Jones, R.G. and Murphy, J.M. 1998. ‘Effect of systematic errors in the lateral boundary forcing on the climatology of
a nested regional climate model over Europe’, Clim. Dyn., 14, 691 – 712.
Osborn, T.J. and Hulme, M. 1997. ‘Development of a relationship between station and grid-box rainday frequencies for climate
model evaluation’, J. Clim., 10, 1885–1908.
Palutikof, J.P., Winkler, J.A., Goodess, C.M. and Andresen, J.A. 1997. ‘The simulation of daily temperature time series from GCM
output. Part I: comparison of model data with observations’, J. Clim., 10, 2497 – 2513.
Senior, C.A. 1999. ‘Comparison of mechanisms of cloud-climate feedbacks in GCMs’, J. Clim., 12, 1480 – 1489.
von Storch, H., Zorita, E. and Cubasch, U. 1993. ‘Downscaling of global climate estimates to regional scales: an application to
Iberian rainfall in wintertime’, J. Clim., 6, 1161–1171.
Watson, R.D., Zinyowera, M.C. and Moss, R.H. (eds). 1998. The Regional Impacts of Climate Change, Cambridge University Press,
Wetherald, R.T. and Manabe, S. 1995. ‘The mechanisms of summer dryness induced by greenhouse warming’, J. Clim., 8,
3096 – 3108.
Wilby, R.L. and Wigley, T.M.L. 1997. ‘Downscaling General Circulation Model output: a review of methods and limitations’, Prog.
Phys. Geogr., 21, 530–548.
Wilby, R.L. and Wigley, T.M.L. 1999. ‘Precipitation predictors for downscaling: observed and general circulation model
relationships’, Int. J. Climatol., 19 (in press).
© British Crown Copyright 2000
Int. J. Climatol. 20: 489 – 501 (2000)
Без категории
Размер файла
419 Кб
Пожаловаться на содержимое документа