Empirically Downscaled Multimodel Ensemble ... - AMS Journals

Download PDF
7 downloads 0 Views 1MB Size Report
Comment
Nov 1, 2002 - Thompson 1998; Zorita and von Storch 1997; von. Storch et al. ...... Charles, S. P., B. C. Bates, P. H. Whetton, and J. P. Hughes, 1999: Validation of .... Meehl, G. A., W. D. Collins, B. A. Boville, J. T. Kiehl, T. M. L.. Wigley, and ...
3008

JOURNAL OF CLIMATE

VOLUME 15

Empirically Downscaled Multimodel Ensemble Temperature and Precipitation Scenarios for Norway RASMUS E. BENESTAD The Norwegian Meteorological Institute, Oslo, Norway (Manuscript received 10 September 2001, in final form 22 April 2002) ABSTRACT A number of different global climate model scenarios are used in order to infer local climate scenarios for various locations in Norway. Results from empirically downscaled multimodel ensembles of temperature and precipitation for the period 2000–50 are presented, based on common EOFs of large-scale temperature and sea level pressure fields. Comparisons with actual records for the past show that the multimodel ensemble range tends to span the observations. All scenarios for temperature change indicate a future warming, but the sea level pressure–based scenarios for precipitation are characterized by a large scatter about zero change. The primary cause for the large spread in precipitation trend estimates is attributed to differences between the various global climate scenarios. It is also acknowledged that the sea level pressure–based empirical models may underestimate the trends as they do not take directly into account increases in the precipitation due to increased temperatures. Nevertheless, in some locations the majority of the ensemble members suggest wetter springtime conditions.

1. Introduction Measurements of atmospheric CO 2 concentrations since the late 1950s indicate a gradual accumulation (Houghton et al. 2001). This increase is believed to alter the natural energy distribution on the earth, trapping heat near the surface and cooling the upper atmosphere, thus leading to a global climate change (Arrhenius 1896; Houghton et al. 2001). Most observations do indeed suggest warmer global mean surface temperatures, melting of sea ice, accumulation of heat in the upper oceans, and rising sea levels (Houghton et al. 2001). A climate change has potentially severe consequences for society, and it is therefore important to be able to forecast future climatic trends. Although the nature of the climate system is chaotic (Lorenz 1967) and its exact trajectory cannot be deterministically predicted years ahead, it may nevertheless be possible to predict the long-term climatic trends, given a systematic change in the boundary conditions. Such predictions are based on coupled atmosphere–ocean general circulation models (AOGCMs), which describe the large-scale dynamics and thermodynamics of the climate systems. The term ‘‘climate scenarios’’ is used henceforth in order to emphasize the fact that these climate models can only forecast plausible climatic trends and that the internal variations are more arbitrary. The state-of-the-art AOGCMs are designed to deCorresponding author address: R.E. Benestad, The Norwegian Meteorological Institute, P.O. Box 43, 0313 Oslo, Norway. E-mail: [email protected]

q 2002 American Meteorological Society

scribe the large-scale features of the global climate. Evaluation of AOGCMs suggests that these give a realistic description of the present-day large-scale climatic features (Busuioc et al. 2001; Banks et al. 2000; Tett et al. 1999; McGuffie et al. 1999; Benestad 2001b; Machenhauer et al. 1998; Rummukainen et al. 1998; Hansen et al. 1998; Carill et al. 1997; Tett et al. 1996). However, the global climate models are not able to give a good reproduction of local climate conditions (Grotch and MacCracken 1991). This is no surprise when considering the crude spatial resolution that is typical for these models. Small-scale processes not properly resolved by the model may be important. Local geographical features, such as mountain ranges, valleys, fjords, distance from the coast, altitude, and lakes have a profound impact on the local climates, and most of such features are not accounted for by today’s global climate models. Figure 1 shows a comparison between actual annual mean temperatures recorded at three different locations in southern Norway (Oslo, Nesbyen, and Bergen), corresponding values interpolated from the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) reanalysis (Kalnay et al. 1996), and one AOGCM simulation (ECHAM4 GSDIO1, see Table 1). The geographical differences between the three locations are not captured by either the reanalysis or the AOGCM. 1 The climate model integration is a ‘‘GSDIO’’ scenario that includes the effects of greenhouse gases, tropospheric ozone, and direct plus indirect effects of industrial aerosols (sulfur).

1 NOVEMBER 2002

3009

BENESTAD

FIG. 1. Comparison between observed annual mean temperature (black) and interpolated NCEP–NCAR values (gray). The temperature recordings were made at NMI weather stations in Oslo, Nesbyen, and Bergen. Also shown in dark gray (curves at farthest right) are interpolated temperatures from global climate scenarios produced by one AOGCM (ECHAM4 GSDIO). Two of the AOGCM curves (representing Nesbyen and Oslo) virtually overlap.

It is well-known that climatic variations in many locations are associated with the state of the large-scale circulation, such as the El Nin˜o–Southern Oscillation (ENSO) and the North Atlantic Oscillation (NAO). The NAO has a pronounced effect on the local climate variability in northern Europe (Busuioc et al. 2001; Chen and Hellstro¨m 1999; Zveryaev 1999; Jones et al. 1997; Appenzeller et al. 1998; Heyen et al. 1996; HellandHansen and Nansen 1920). It is also believed that changes in the snow cover (Brown 2000; Watanabe and Nitta

1998), sea ice extent (Hartmann 1994), and the heat transport by the Gulf Stream and the North Atlantic Drift will have an impact on the northern European climate. If the changes in the large-scale climatic features are known, it is therefore possible to infer local climate changes, given the relation between the large scales and the local variability. Empirical downscaling (Benestad 1999; Heyen et al. 1996; Kilsby et al. 1998; Kidson and Thompson 1998; Zorita and von Storch 1997; von Storch et al. 1993) is a way of adding the local geographical information to the global predictions as well as a way of overcoming the climate models’ inability of describing local climate on the grid-scale level (Grotch and MacCracken 1991). The main objective of this paper is to present local multimodel climate scenarios for a number of locations in Norway. Benestad (2002) presented empirically downscaled temperature scenarios for a number of sites in northern Europe, inferred using a new method involving common EOFs (Flury 1988; Sengupta and Boyle 1993; Barnett 1999) and the results described here were obtained using the same methodology. This study elaborates on the findings of Benestad (2002) by examining both the past as well as the future. The results for the past will be evaluated against the actual observations in order to provide a crude model evaluation. Both scenarios for precipitation and temperature are presented, and the sensitivity of the downscaled estimates to climate model differences and empirical model setup is explored. 2. The method and the data The global climate model scenarios used in this study are listed in Table 1, and include models with different spatial resolutions and degrees of sophistication. Of these, the NCAR climate system model (CSM) is one of the most sophisticated and includes a model of bio-

TABLE 1. Overview of the AOGCMs used in this study. CSIRO is the Commonwealth Scientific and Industrial Research Organization and GFDL is the Geophysical Fluid Dynamics Laboratory. The column marked by ‘‘ny’’ and ‘‘nx’’ give the number of data points there are along the meridional and zonal axes, and give an indication of the spatial resolution of the various models (the higher number, the better resolution). The abbreviation FA stands for ‘‘flux adjusted.’’ The last column (DT ) lists the predicted global mean temperature change (8C) between the periods 1991–2000 and 2041–50. For the NCAR–DOE scenario the periods were 1971–80 and 2026–35 because this integration ended in 2035. The GSDIO scenario includes the effects of greenhouse gases, tropospheric ozone, and direct plus indirect effects on industrial aerosols (sulfur). The GSA integrations include the effects of greenhouse gases and only the direct effects of aerosols. The b006 experiment simulates a 1% yr21 increase in the CO 2 level. Model

ny

nx

FA

Country

DT (8C)

CSIRO (GSA) CCCMA (GSA) ECHAM4 (GSDIO) ECHAM4 (GSA) ECHAM3 (GSA) HadCM2 (GSA) HadCM3 (GSA1ozone) NCAR–DOE (GSA) GFDL–19 (GSA) CCSR–NIES (GSA) NCAR–CSM (b006)

32 48 64 64 64 73 73 40 40 32 64

64 96 128 128 128 96 96 48 48 64 128

Yes Yes Yes Yes Yes Yes No No Yes Yes No

Australia (Gordon and O’Farrell, 1997) Canada (Flato et al. 2000) Germany (Roeckner et al. 1992; Oberhuber, 1993) German (Roeckner et al. 1996; Oberhuber, 1993) Germany United Kingdom (Johns et al. 1997) United Kingdom (Gordon et al. 2000) United States United States (Delworth et al. 1999) Japan (Emori et al. 1999) United States (Meehl et al. 2000)

1.20 1.38–149 0.89 1.13 1.13 0.99–1.07 1.26 2.67 1.62 1.38 0.80

3010

JOURNAL OF CLIMATE

geophysical and biogeochemical land–atmosphere interactions such as the effects of land surfaces on climate and atmospheric chemistry. The NCAR CSM and the Hadley Centre Coupled Ocean–Atmosphere General Circulation Model (HadCM3) models do not utilize flux correction that otherwise may artificially constrain the model solutions. The HadCM3 model also includes a scheme that describes stratospheric ozone. The ECHAM4 GSDIO integration listed in Table 1 includes the indirect effects of industrial aerosols as well as tropospheric ozone, whereas the ‘‘GSA’’ integrations only take greenhouse gases and the direct aerosol effects into account. The aerosols’ effect on the clouds, that is, the indirect effects, are not yet well known (Houghton et al. 2001). Although some models may be regarded as more realistic than others it is assumed here, to a firstorder approximation, that each scenario is equally probable. Benestad (2001a) showed that empirical downscaling involving common EOFs gives more reliable results than conventional methods projecting model results onto EOFs obtained from the observations. Therefore the same approach as described by Benestad (2001a) was used to downscale global climate scenarios from a number of different climate change experiments. The details of the downscaling analysis are described in the appendix. Benestad (2001a) and Benestad (2000b) have evaluated the downscaling models and documented their ability to describe most of the local warming trends. It has also been demonstrated that the relationship between large-scale structures and local climates may vary with the seasons (Busuioc et al. 2001; HanssenBauer and Førland 2001). For this reason, downscaling was applied to January, April, July, and October months, respectively, thus each season was analyzed separately. Then, the monthly mean anomalies from the AOGCM were spatially integrated (bilinear) onto the same grid as the observations. The interpolated AOGCM anomalies were subsequently merged with the observations, and an ordinary EOF analysis was applied to the combined data (the EOFs of the combined data are referred to as ‘‘common EOFs’’). Only the 20 leading modes were used as input for a stepwise screening process that selected a subset for these 20 modes for model calibration. The downscaling analysis employed statistical models calibrated with the part of the principal components (PCs) that represents the actual observations. The calibration was made on detrended series (Benestad 2001b) using a canonical correlation analysis (CCA; Benestad 1998). The models were developed using a stepwise screening process where each PC was included if they increased the Pearson correlation in a cross-validation analysis (Wilks 1995). The predictions, that is, downscaled scenarios, were made using the part of the PCs that represents the model simulations. The common EOFs ensure that the same observed large-scale spatial climatic patterns associated

VOLUME 15

FIG. 2. Map showing the locations of the predictor domains. The four different predictor domains used for these calculations include NATL (408–758N, 908W–408E), NEUR (508–708N, 308W–408E), NORD (508–758N, 208W–408E), and SCAN (558–758N, 08W–308E).

with local climate anomalies are used to infer local climate changes related to changes in the (large-scale) regional climate. As the downscaled results may be sensitive to the choice of predictor region (Benestad 2001a), a number of different spatial domains was used in the downscaling of the local climate. These predictor domains are shown in Fig. 2. Benestad (2000b) demonstrated that large predictor domains tend to underestimate the temperature trends, and hence more weight has been given to scenarios derived using the smaller domains by including more small domains than large domains in the multimodel ensemble downscaling (Table 2). The climate models are thought to give a more realistic description of the free tropospheric quantities and large-scale circulation than of surface parameters (Rummukainen 1997; Murphy 1999). Hence, the downscaling predictors should ideally be from the free troposphere. Benestad (2000b) found, however, a cooling trend in some cases when the predictors were based on just atmospheric circulation or tropospheric fields, in spite of a warming inferred from the gridbox values. The 2-m temperature [T(2m)] seemed to give the most credible results in general, although the free tropospheric fields sometimes may be a better choice. Hence, the downscaling analysis was carried out on various fields, but with the strongest emphasis on the T(2m) field. The predictor data used for model calibration were taken from the NCEP–NCAR reanalysis and the 1873–2000 reconstructions of sea level pressure (SLP) and T(2m) based on data from the University of East Anglia (Benestad 2000a). The downscaled temperature scenarios were derived from the climate change experiments using

1 NOVEMBER 2002

3011

BENESTAD

TABLE 2. Details of the different downscaled warming and precipitation trend estimates in Figs. 4–9. The first column sums up the number of different trend estimates listed in each row. Columns 5–8 indicate which regions were used for the predictors, where the NATL region is 408–758N, 908W–408E (27 3 8 grid points), NEUR 508–708N, 308W–408E (15 3 5 grid points), NORD 508–758N, 208W–408E (13 3 6 grid points), and SCAN 558–758N, 08W–308E (7 3 5 grid points). The domains are shown in Fig. 2. The domain WEUR 408–658N, 408W– 308E was only used for the downscaling of ECHAM4 GSDIO precipitation. N

Model (member)

Run

T (2m) 4 2 3 3 4 6 2 6 2 4 8 2 2 48 total

ECHAM4 ECHAM4 ECHAM4 ECHAM4 NCAR CSM CCCma 1-3 CSIRO ECHAM3 1,2 GFDL HadCM3 HadCM2 1-4 NCAR–DOE CCSR–NIES

GSDIO GSA GSDIO GSDIO b006 GSA GSA GSA GSA GSA GSA GSA GSA

Precip. 5 3 3 9 3 3 3 3 12 3 3 50 total

ECHAM4 ECHAM4 NCAR CSM CCCma 1-3 CSIRO ECHAM3 GFDL HadCM3 HadCM2 1-4 NCAR–DOE CCSR–NIES

GSDIO GSA b006 GSA GSA GSA GSA GSA GSA GSA GSA

NATL

NEUR WEUR

T(2m) T(2m) F(700–500 hPa) T(850 hPa) T(2m) T(2m) T(2m) T(2m) T(2m) T(2m) T(2m) T(2m) T(2m)

u

u

u

u u u

u

u

SLP SLP SLP SLP SLP SLP SLP SLP SLP SLP SLP

u u u u u u u u u u u

uu

Predictor

u

NORD

SCAN

u u u u u u u u u u u u u

u u u u u u u u u u u u u

u u u u u u u u u u u

u u u u u u u u u u u

FIG. 3. Map showing the stations network for the (a) temperature and (b) precipitation measurements. The climate stations measuring the temperature include Oslo (marked with ‘‘O’’), Nesbyen (‘‘N’’), Ferder, Oksøy, Bergen, Ona, Røros, Domba˚ s (Kjøremsgrende), Værnes, Bodø, Tromsø, Karasjok (‘‘K’’), and Vardø. The precipitation measurements are made at Halden, Moss, Oslo, Røros, Verma, Hemne, Namdalseid (‘‘N’’), Dunderlandsdal (‘‘D’’), Sulitjelma (‘‘S’’), Barkestad (‘‘B’’), Tromsø, Geilo, Bja˚ en, Sviland, Bergen, and Lavik.

3012

JOURNAL OF CLIMATE

large-scale T(2m) field simulations as well as 500–700hPa thickness and 850-hPa temperature fields as predictors, whereas the precipitation trends were inferred using the large-scale SLP fields (Table 2). The simulated NAO indices were derived from interpolated SLP at 38.708N, 9.108W and 65.58N, 22.448W by taking their standardized difference. The predictand data used for model calibration was taken from the climate archive from the Norwegian Meteorological Institute (NMI, see online at http://met.no/ english/rpandpdpactivities/publications/1999.html). The locations of the various stations representing the downscaled local climates are shown in Fig. 3. The observed North Atlantic Oscillation index (NAOI; Hurrel 1995; Jones et al. 1997), which is the standardized pressure difference between the Azores and Iceland, was taken from the University of East Anglia Climate Research Unit’s Web site (http://www.cru.uea.ac.uk/cru/data/nao.htm). Control integrations often do not provide a good reference level if they are integrated with the best available forcings for the 1990s and the transient integrations are initialized by taking a time slice from the control as representing ‘‘1860’’ (Machenhauer et al. 1998; Benestad 1999). Therefore trend estimates were calculated by finding the best linear fit using least squares for at least 70 years (model date 1980 to the end of the integration) and the best-fit slope was used as an estimator of the warming rate. Thus, by using these linear trend estimates instead of difference between the downscaled results from the global warming experiments and control, the analysis is not severely affected by biases in the spinup histories and assumptions about past atmospheric greenhouse gas concentrations. Trend estimates, however, are sensitive to outliers at the beginning and end of the series, but longer records are less sensitive than short series. Thus, the sampling uncertainty may be reduced by using long intervals for trend estimation. 3. The results a. Reconstruction of the past The downscaled multimodel ensemble scenarios for a given location can be presented by ‘‘plume plots’’ showing the range of ensemble members as well as the temporal evolution of a chosen time series. Figure 4 shows a plume plot for the temperature in Oslo (location marked with an ‘‘O’’ in Fig. 3) for four different calendar months—January, April, July, and October—together with the actual temperature measurements (see Table 2 for details of the downscaling). For all four seasons, the observations tend to lie within the ensemble range, suggesting a realistic reproduction of the past.

VOLUME 15

The winter temperature is subject to the strongest shortterm variations in both the model results and the real world, and the variance described by these scenarios is comparable to the variance in the observations. The climate models tend to predict slightly warmer July and October conditions toward the end of the observational record than those inferred from the measurements. It is also evident from these results that single-model scenarios, such as the HadCM3 and the ECHAM4 GSDIO integrations, do not give a good description of the climatic evolution, but that a multimodel ensemble range approach may be more useful. Figure 5 shows the time series describing the Oslo monthly cumulative precipitation. The comparison between the model results and the observations suggests that the downscaling analysis captures slightly less of the rainfall variability than corresponding results for temperature. There are few signs of clear long-term trends either in the model results or past observations. Thus, the simulated ‘‘past’’ trends are in good agreement with the observed long-term trend. b. Scenarios The presentation of multimodel ensemble scenarios by box-and-whisker plots, where the outliers do not affect the box but are only shown as circles, gives a model consensus of the warming rate ranges where the extreme model results do not bias the location and range of the estimates. Figures 6 and 7 show box-and-whisker plots of multimodel ensemble temperature and precipitation change estimates, based on 48 and 50 derivations, respectively. Table 2 gives an overview of these derivations, showing the regions used for the predictors and which quantities and climate models have been used in the downscaling. The results for temperature (Fig. 6) show that there is a clear model consensus concerning a regional warming in the future for all locations, with strongest trends during winter. The most pronounced change is also inferred for locations with a continental-type climate, such as Nesbyen and Karasjok (marked in Fig. 3a as ‘‘N’’ and ‘‘K’’). These results are in good agreement with the corresponding results of Benestad (2002) derived using predictand data from the North Atlantic Climatological Dataset (NACD; Frich et al. 1996) as well as results from other evaluations of multimodel ensembles (Christensen et al. 2001; Ra¨isa¨nen 2001). In the spring, the multimodel ensemble predicts higher median values for northern Norway (Tromsø, Karasjok, and Vardø), but these estimates are also associated with larger spread. The climate models give a weaker spring→

FIG. 4. Plume plots showing the multimodel range in monthly mean temperature scenarios made for Oslo for the four calendar months: (a) Jan, (b) Apr, (c) Jul, and (d) Oct. Superimposed are the corresponding observed monthly mean temperature (heavy black). The past temperatures tend to lie within the ensemble range. The scenario curves have been adjusted so that they have the same mean value as the

1 NOVEMBER 2002

BENESTAD

FIG. 4. (Continued ) observations over time overlapping intervals with valid data. A 10-pt (yr) low-pass filter has been used to smooth the curves for clarity. The 48 different estimates were derived downscaling 17 different scenarios produced with the 10 different AOGCMs (listed in Table 1) using different predictor domains. The heavy yellow line shows the ensemble mean. The thin light and dark gray lines show the results derived using the HadCM3 and ECHAM4 GSDIO scenarios derived using the NORD and SCAN domains, respectively.

3013

3014

JOURNAL OF CLIMATE

FIG. 5. Same as for Fig. 4 but for precipitation in Oslo.

VOLUME 15

1 NOVEMBER 2002

BENESTAD

time warming than in the other seasons, except for in the northernmost locations. In July, slightly warmer scenarios are found in the mountainous regions near Røros and Domba˚s than elsewhere, but this difference may not be significant. For October, the multimodel ensemble predicts the smallest temperature increases in the north. The spread in the estimated warming presented in this paper may be attributed to both different climate scenarios (‘‘intermodel scatter’’) and different predictor domains (‘‘intramodel scatter’’). Figure 8 provides a further analysis of the spread between the various estimates for Oslo, showing how the warming estimates relate to the various global climate scenarios. The January intramodel spread is most pronounced with the NCAR–Department of Energy (DOE) and the University of Tokyo Center for Climate System Research–National Institute of Environmental Studies (CCSR–NIES) January scenarios (Fig. 8a). The highest of these estimates were derived with the smallest predictor domain, a result that may suggest that these two models are not able to give a good reproduction of the large-scale spatial temperature structures. With the exception of the NCAR–DOE and the CCSR–NIES scenarios, the contribution to the spread is of similar magnitude from both intermodel and intramodel differences. The HadCM2 ensemble members exhibit a broad range for the downscaled estimates, with a range similar to the whole multimodel ensemble excluding the extreme NCAR–DOE and CCSR–NIES models. Thus, the intermodel spread reflects both the differences in model formulation and initial conditions. The same analyses for April and July (Figs. 8b,c) indicate that the scatter is about the same for intermodel and intramodel estimates. The intermodel scatter tends to dominate over the spread due to different downscaling choices in October (Fig. 8d). The strongest warming is predicted in January by the NCAR–DOE model (Fig. 8a: DT 5 6.48C between 2000 and 2050) closely followed by the CCSR–NIES. The weakest January warming was estimated from member 3 of the HadCM2 ensemble (DT 2000 ... 2050 5 0.18C). The downscaling of the ECHAM4 GSDIO scenario based on the 700–500-hPa thickness fields and 850-hPa temperatures gave, in general, similar results as those based on the T(2m) fields, except for the April scenarios [The SCAN (558–758N, 08–308E) and NORD (508–758N, 208–408E) regions of 700–500-hPa thickness gave 0.208 and 20.558C, respectively, whereas the calculations based on T(850hPa) yielded 0.358, 0.408 and 20.058C with the domains SCAN, NEUR (508–708N, 308W– 408E), and NATL (408–758N, 908W–408E), respectively.] There was no systematic preference as to which predictor domain yields the highest estimates for Oslo in January and July, but the highest April values were usually derived with the smallest predictor domain (SCAN) and in October the NORD domain gave the highest values. The scenarios for precipitation change between 2000

3015

and 2050 (Fig. 7) show larger scatter than the temperature scenarios, and in no location is there any indication of a clear trend since the boxes tend to cross zero. In January, the scenarios for southern Norway nevertheless indicate a slight wet bias (Halden: 36 of the 48 estimates suggest wetter future climate; Moss: 36; Oslo: 30; Geilo: 36), whereas the 34 of the 48 scenarios point to less precipitation near Røros. The locations where the results give the clearest indication of a robust signal are Namdalseid in April as well as Barkestad and Sulitjelma in July, where most of the scenarios describe wetter conditions (Fig. 3b marked as ‘‘N,’’ ‘‘B,’’ and ‘‘S’’). However, even for these locations there are some estimates suggesting drier future spring seasons [Namdalseid: 8 negative precipitation estimates distributed amongst the HadCM2 ensemble, NCAR–DOE, and CCSR–NIES; Barkestad: 11 negative estimates including the Canadian Centre for Climate Modelling and Analysis model (CCCma), HadCM2, NCAR–DOE, and CCSR–NIES; Sulitjelma: 12 negative estimates from CCCma, HadCM2, NCAR–DOE, and CCSR–NIES]. The multimodel estimates also show a wet bias for Hemne, Dunderlandsdal, and Barkestad (Fig. 3b marked as ‘‘S,’’ ‘‘D,’’ and ‘‘B’’), as 35 estimates of a total of 50 were positive for these locations [henceforth the notation ‘‘(35/50)’’ will be used]. For July, the majority of the estimates indicate more rain in Sulitjelma (37/50) and Barkestad (39/50). The global scenarios that predicted drier summers at Sulitjelma include members 1, 2, and 3 of the CCCma ensemble, HadCM2 member 3, NCAR– DOE, and CCSR–NIES. Member 4 of the HadCM2 ensemble also suggest a negative rainfall trend at Barkestad. The minority of the predictions with different signs to the ‘‘consensus’’ include both ‘‘reasonable’’ as well as ‘‘extreme’’ models, and the lack of agreement points to high uncertainties about the precipitation trends. The intramodel scatter in the downscaled precipitation change between 2000 and 2050 is small, in general, compared to the intermodel spread. There are nevertheless some cases during April (Fig. 9b) where the intramodel spread is of similar magnitude as the intermodel scatter (CCCma members 2 and 3 and ECHAM3 member 1). The largest intramodel scatter is found for ECHAM3 in January, NCAR–DOE in April and July, and ECHAM3 in October. The greatest 2000–50 increase in precipitation is predicted by ECHAM4 GSA in (January at Hemne: 51.4 mm month 21 ), and the most pronounced decrease is produced by member 3 of the HadCM2 ensemble (January at Lavik: 256.75 mm month 21 ). 4. Discussion In order to use climate models for the projection of future climates, they must demonstrate that they can reproduce the past. Individual climate model integrations may not give a good reproduction of the climatic

3016

JOURNAL OF CLIMATE

VOLUME 15

FIG. 6. Box-and-whisker plot representation of 48 warming trend estimates based on 17 different AOGCM scenarios (10 different AOGCMs) using the (a) Jan, (b) Apr, (c) Jul, and (d) Oct T(2m) fields in addition to the ECHAM4 GSDIO T 850hPa , and F 500hPa 2 F 700hPa thickness (see details in Table 2). The change was estimated by computing the linear trend estimates from model date 1980 to the end of the transient run and multiplying by 50 yr. The whiskers extend to the most extreme data point that is no more than 1.5 times the interquartile range from the box (default). The outliers are given as circles. The boxes represent the interquartile range and the horizontal line gives the median. The details of the individual downscaled scenarios can be found in Benestad (2000b).

evolution over only a few decades if the chaotic natural variations (noise) are much more pronounced than the slow trends caused by changes in the climate forcings. Plume plots, however, are more suitable for comparing past records with corresponding simulated values. En-

sembles of many scenarios may capture the underlying trends caused by changing boundary conditions if the natural short-term climate variations are independent and cancel each other. Using scenarios from different models may provide a crude sample of the uncertainties

1 NOVEMBER 2002

BENESTAD

3017

FIG. 6. (Continued )

associated with different model biases. Of course, shortcomings that are common traits among the models will not be resolved by such a multimodel approach. There are some caveats attached to the multimodel ensemble approach adopted here. Allen et al. (2000) argue that the use of intermodel spread as a measure for uncertainty may be problematic because the models do not necessarily span the full range of known climate system behavior, and this is especially true for small ensembles. On the other hand, the criterion that the estimates for the past must span the observations for pro-

jections for the future to be credible may also contain serious flaws. It is, for instance, not difficult to include unrealistically extreme models to the ensemble so that the observations always lie within the maximum and minimum. The usefulness of this approach, however, depends on the choice of representative climate models that increase with diminishing intermodel range. There is no guarantee that even if the multimodel ensemble spans the observations in the past, the future climatic state will lie within the ensemble range. There are also some important limitations to empirical

3018

JOURNAL OF CLIMATE

VOLUME 15

FIG. 7. Box-and-whisker plot (as in Fig. 6) representation of 50 precipitation change estimates for the period 2000–50, based on 16 different AOGCM scenarios (10 different models) using the SLP fields.

downscaling analysis in general. It is important to keep in mind that empirical downscaling assumes that the relationship between the large-scale anomalies and small scales will be the same in the future as in the past. This assumption can be tested to some extent by examining the relationship between the local and largescale anomalies in the observational record (HanssenBauer and Førland 2000) as well as in the global climate scenarios (‘‘local’’ is then represented by a grid box)

(Benestad 2001a). The results presented here hinge on the assumption of stationarity. Substantial intramodal scatter may point to an inability of the climate models to reproduce the essential features in the observed spatial structures in the predictor fields. Large predictor domains place greater demands on the model skill than smaller domains, and it is interesting to note that the largest domain (NATL), in conjunction with both the HadCM3 and NCAR CSM

1 NOVEMBER 2002

BENESTAD

3019

FIG. 7. (Continued )

models, produces lower temperature estimates than the smaller predictor regions, except for in April. On the other hand, the predictor domains should not be too small as the climate models’ skillful scale may not be smaller than about 50 grid points (Grotch and MacCracken 1991). Thus, the SCAN domain, which gave the highest April values for the temperature change, may be smaller than the models’ skillful scale and may not give reliable scenarios (7 3 5 5 35 grid points). The issue about domain size also raises the question about

which region is optimal for the downscaling of climate scenarios. Here, the solution to this dilemma has been to use various regions to get a range of estimates. The empirically downscaled results obtained in this study may be compared with studies based on dynamical downscaling in order to improve our knowledge of climate change. The nested dynamical models are more physically consistent and are thought to be less prone to nonstationary relationships between the local and large scales. However, even the nested models involve

3020

JOURNAL OF CLIMATE

VOLUME 15

FIG. 8. Scatterplots showing the 2000–50 warming estimates (y axis in 8C) vs model scenario (x axis) for (a) Jan, (b) Apr, (c) Jul, and (d) Oct. There are several estimates in each model category corresponding to different predictor domains.

empirical relationships, such as for subgrid parameterization, and hence, also assume stationary relationships. One advantage of empirical downscaling is that low computational demands tend to allow longer time series analysis, and hence, simulated past long-term trends can be compared with observed values. Furthermore, shorter time series are more prone to sampling fluctuations. It is also less costly to produce a large number of different scenarios based on multimodel ensembles by empirical downscaling than by dynamical downscaling. For example, the climate change synthesis by Christensen et al. (2001) analyzed the results from only three different climate scenarios (ECHAM4 GSDIO, ECHAM4 GHG2, and HadCM2 GHG), the longest of which was 20 yr long (the shortest scenario was 8 yr long). Sampling fluctuations associated with natural decadal and interdecadal variability may be reduced by the use of an ensemble composite (8 1 20 1 10 1 10 5 48 yr), but model-dependent uncertainties are poorly resolved with small ensembles. The results obtained in this study are in line with Christensen et al. (2001) in the sense that 2 Greenhouse gas (GHG) integrations include greenhouse gases only and no sulfur emission.

the strongest warming takes place during winter in four different dynamical downscaling experiments. Both Ra¨isa¨nen (2001) and Christensen et al. (2001) found a general increase in precipitation for the Nordic region and the latter study reported a tendency toward generally wetter autumns (15% increase in the rainfall rate). Christensen et al. (2001) also found the largest intermodel spread during spring due to large discrepancy between the NMI RegClim model (Bjørge et al. 2000) and the others. The contrast between the lack of general trends reported here and the increase in precipitation found by Christensen et al. (2001) may partly be due to sampling errors and different ensemble sizes. On the other hand, the disagreement between the present results and the precipitation change reported by Ra¨isa¨nen (2001), which are based on a large multimodel ensemble but only 20-yr periods, point to shortcomings in the empirical downscaling analysis. Whereas the temperature estimates are believed to describe most of the local variations (Benestad 2001a), it is important to keep in mind that the SLP field can only describe the part of precipitation caused by large-scale circulation and may not account for increases associated with increased humidity (Wilby et al. 1998; Charles et al. 1999). Hanssen-

1 NOVEMBER 2002

BENESTAD

3021

FIG. 8. (Continued )

Bauer and Førland (2000) reported that although SLPbased empirical models can reproduce most of the longterm trends and decadal variability from 1940 to the present, they fail to describe the winter precipitation in southeastern Norway between 1900 and 1940. Thus, the change in precipitation presented here only describes the part of the precipitation associated with changes in the atmospheric circulation, and may underestimate the precipitation change caused by changes in the airmass characteristics (e.g., increased rainfall associated with higher temperature). A remedy for this shortcoming would be to include humidity (temperature) among the predictors (Hanssen-Bauer et al. 2001; Murphy 2000). The fact that most of the estimates gave increased precipitation for a few locations brings up the question as to whether these are significant or purely coincidental. This issue is related to ‘‘the problem of multiplicity’’ (Wilks 1995) as one would expect to find some locations where most of the estimates have the same sign if many different sites are examined and the estimates for each are independent. When considering numerous independent estimates such as trends at different locations, then the probability of yielding a few cases with such results by chance increases, and these estimates do therefore not necessarily suggest a common trait among the models.

A significant amount of the rainfall over Scandinavia can be associated with the NAO, therefore the trends in the precipitation will be affected by long-term changes in the NAO. Figure 10 shows the intermodel spread in the NAO index as a function of time, also suggesting that there is little consensus within this multimodel ensemble as to how the NAO is affected by global warming. The lack of model consensus for how SLP will react to global warming does not seem to be limited to the models examined in this study, as Ra¨isa¨nen (2001) also found substantial differences in the projected SLP fields among the transient integrations from the second Coupled Model Intercomparison Project (CMIP2). Ra¨isa¨nen (2001) also found little agreement in precipitation within the CMIP2 ensemble, except for over limited regions with relatively high agreement (one over Scandinavia). However, these were not tested for statistical significance regarding the problem of multiplicity, and it is therefore unclear as to whether these regions of ‘‘model consensus’’ were due to chance rather than real model agreement. 5. Conclusions A comparison between empirically downscaled multimodel ensemble reconstructions of the past climate

3022

JOURNAL OF CLIMATE

VOLUME 15

FIG. 9. Scatterplots showing the 2000–50 precipitation change estimates (y axis in mm month 21 ) vs model scenario (x axis) for (a) Jan, (b) Apr, (c) Jul, and (d) Oct. There are several estimates in each model category corresponding to various domains.

since 1860 and monthly mean temperature measurements from Oslo suggests that the simulations give a realistic description of the past climate conditions. Although the models tend to produce slightly high July and October temperatures toward the end of the twentieth century, it is believed that such multimodel ensemble results are capable of giving realistic local climate scenarios.

Empirical downscaling analysis of simulations from different global climate scenarios suggests further warming in Norway for the future, but there is no clear consensus for the precipitation, perhaps with the exception of a few locations in mid-Norway during the spring season. The lack of clear trends in the precipitation may be due to the fact that the effect of increased humidity is not taken into account because most of the

1 NOVEMBER 2002

BENESTAD

3023

FIG. 9. (Continued )

driving models point toward wetter future climate over Scandinavia. Another factor playing a role is the lack of agreement among the models as to what will happen with the SLP fields as the world gets warmer. Empirical downscaling based solely on SLP may underestimate the long-term precipitation changes. Here precipitation scenarios are closely associated with changes in the large-scale circulations, whereas results obtained from dynamical downscaling also include the

effects of increased humidity and a warmer climate. Christensen et al. (2001) and Ra¨isa¨nen (2001) found an increase in the total precipitation whereas a hint of a positive signal was only found for a few places during spring in this study. The difference between these results may be reduced by including predictors describing changes in the airmass qualities caused by global warming. Uncertainties associated with the predictor domain

3024

JOURNAL OF CLIMATE

VOLUME 15

FIG. 10. Estimates of the NAO index using the SLP fields from the multimodel ensemble. The model range is represented by the shaded area and the ensemble mean is shown as dark gray.

used in the downscaling were compared with the spread in climate change estimates among the various climate models. In most cases the intermodel spread is greater than the sensitivity to the predictor domain, but there are also some cases for which temperature scenarios using different predictor domains produce larger spread than using different models. Large scatter in the downscaled estimates derived from different predictor domains may point to an inability of the climate models to reproduce the observed large-scale structures in the predictor fields. Acknowledgments. I am grateful to E. J. Førland, I. Hanssen-Bauer, and two anonymous reviewers for valuable advice regarding interpretation of the precipitation scenarios and preparing this work. O. Vignes assisted in retrieving the ECHAM4 GSDIO data. This work was done under the Norwegian Regional Climate Development Program under the Global Warming (RegClim) program, and was supported by the Norwegian Research Council (Contract NRC-120656/720) and the Norwegian Meteorological Institute. Part of the analysis was carried out using the R data processing and analysis language (Ellner 2001; Gentleman and Ihaka 2000), which is freely available online (http://www.R-project.org/).

APPENDIX Description of Downscaling Method A matrix, Xobs , can be used to describe the temporal evolution of observed geographical temperature distributions. The matrix dimension is M 3 N and the matrix holds the information on the spatial structure over M different locations. At each location the temperature is observed over a time period, and the time series can be represented by a column for each time x t 5 [x1 , x 2 , . . . x M ] t , and there are N observations made over a time interval: X 5 [x1 , x 2 . . . x N ]. Likewise, the gridded temperature anomalies from the climate model scenarios can be represented as XAOGCM . The respective mean values have been subtracted before combining the data. The climate model results were interpolated onto the same spatial grid as the observations in order to carry out a common EOF analysis, and the gridded observations were detrended in order to avoid inflated correlation coefficients and related undesirable effects in the calibration of downscaling models (Benestad 2001b); however, the trends were not removed in the climate model data as this would remove the climate change signal. A composite dataset was constructed by

1 NOVEMBER 2002

3025

BENESTAD

merging the detrended observations with the model data along the time axis, X 5 [Xobs , XAOGCM ], by appending the rows of the climate model matrix to the rows of the observations. The combined dataset was subject to singular value decomposition (SVD; Strang 1995; Press et al. 1989) in order to compute the right and left inverses of the data matrix: X 5 ULVT .

(A1)

The left inverse is a rectangular matrix (U) holding the right eigenvectors along its columns, each describing an ‘‘eigenpattern’’ (also referred to as ‘‘modes’’). Here U is an M 3 K matrix where K represents the total number of different modes, and is equivalent to empirical orthogonal function (EOF) analysis (Lorenz 1963; North et al. 1982; Wilks 1995; von Storch and Zwiers 1999) when a geographical weighting is applied to the data so that the grid boxes representing smaller surface area carry proportionally less weight [W ij 5 Ïcos(F j ), where cos(F j ) is the latitude associated with the grid box]. A spatial weighting was applied to the data prior to the SVD and an inverse scaling function was applied to the spatial patterns U obtained from the SVD. The eigenpatterns described by the columns in U hold no information about the time evolution of the data. The right inverse (V) is also referred to as the ‘‘principal components’’ in the geophysical literature, and this rectangular matrix describes the time evolution of the various eigenpatterns. The matrix V has the dimensions N 3 K, and the columns of V hold a set of weights used to determine the importance of each mode at each time when these are recombined to describe the spatial anomalies. The principal components can be divided into two T sections: VT 5 [VTobs, VAOGCM ]. The matrix L is a diagonal matrix holding the eigenvalues in descending order along its diagonal. The higher-order modes were dropped in order to simplify regressional analysis and filter away noise. Here, only the leading 20 modes were kept for a stepwise screening (U M,K → U M,20 , L K,K → L 20,20 , V N,K → V N,20 ). The data were subsampled, as recommended by North et al. (1982), to reduce any effects caused by autocorrelation. Subsampling the data also reduces effects introduced by discontinuities between the observed and simulated data if the respective parts have nonzero autocorrelation. The estimate for V was then computed by projecting the entire dataset onto the eigenpatterns computed using the subsampled data: VT 5 L 21UTX. The part of the principal components corresponding to the observed temperature reconstruction (VTobs ) was used for model calibration based on canonical correlation analysis (CCA; Preisendorfer 1988; Wilks 1995; von Storch and Zwiers 1999), and since the principal components were used, the analysis is similar to the method described in Barnett and Preisendorfer (1987). The empirical models were seasonally stratified, with one respective model for each calendar month. A cross-

validation (Wilks 1995, 194–198) analysis was employed to determine which modes should be included. The patterns that improved the cross-validation Pearson correlation were selected for the model calibration. The period 1891–1990 was used for model development. The CCA model, which consists of a linear transform (a matrix C) from the predictor V (principal components of gridded data) to the predictand Y (station temperature records) can be described as ˆ 5 CVTobs . Y

(A2)

The matrix C in Eq. (A2) is the statistical model that can be used for prediction. The CCA between VTobs and Y yields canonical variates (U and V), canonical correlation maps (G and H), and a correlation matrix (M) that forms the basis for the statistical model. Here Y can therefore be predicted from VTobs according to ˆ 5 GM(HTH) 21HTVTobs. Y

(A3)

The CCA-based downscaling model is the matrix C 5 [GM(HTH) 21HT ]. This model can be used with the principal components representing the AOGCM results, as these describe exactly the same spatial structures as VTobs. Hence, downscaled scenarios can be derived according to T ˆ scenario 5 CVAOGCM Y .

(A4)

REFERENCES Allen, M. R., P. A. Stott, J. F. B. Mitchell, R. Schnur, and T. L. Delworth, 2000: Quantifying the uncertainty in forecasts of anthropogenic climate changes. Nature, 407, 617–620. Appenzeller, C., T. F. Stocker, and M. Anklin, 1998: North Atlantic oscillation dynamics recorded in Greenland ice cores. Science, 282, 446–449. Arrhenius, S. A., 1896: On the influence of carbonic acid in the air upon the temperature of the ground. Philos. Mag. J. Sci., 236– 276. Banks, H. T., A. Wood, J. M. Gregory, T. C. Johns, and G. S. Jones, 2000: Are observed decadal changes in the intermediate water masses a signature of anthropogenic climate change? Geophys. Res. Lett., 27, 2961–2964. Barnett, T. P., 1999: Comparison of near-surface air temperature variability in 11 Coupled Global Climate Models. J. Climate, 12, 511–518. ——, and R. W. Preisendorfer, 1987: Origins and levels of monthly and seasonal forecast skill for United States surface air temperatures determined by canonical correlation analysis. Mon. Wea. Rev., 115, 1825–1850. Benestad, R. E., 1998: CCA applied to statistical downscaling for prediction of monthly mean land surface temperatures: Model documentation. KLIMA Rep. 28/98, 96 pp. [Available from DNMI, P.O. Box 43, Blindern, 0313 Oslo, Norway.] ——, 1999: Pilot studies of enhanced greenhouse gas scenarios for Norwegian temperature and precipitation from empirical downscaling. KLIMA Rep. 16/99, 86 pp. [Available from DNMI, P.O. Box 43, Blindern, 0313 Oslo, Norway.] ——, 2000a: Analysis of gridded sea level pressure and 2-meter temperature for 1873–1998 based on UEA and NCEP re-analysis II. KLIMA Rep. 03/00, 45 pp. [Available from DNMI, P.O. Box 43, Blindern, 0313 Oslo, Norway.] ——, 2000b: Future climate scenarios for Norway based on linear empirical downscaling and inferred directly from AOGCM re-

3026

JOURNAL OF CLIMATE

sults. KLIMA Rep. 23/00, 127 pp. [Available from DNMI, P.O. Box 43, Blindern, 0313 Oslo, Norway.] ——, 2001a: A comparison between two empirical downscaling strategies. Int. J. Climatol., 21, 1645–1668. [ DOI 10.1002/joc.703] ——, 2001b: The cause of warming over Norway in the ECHAM4/ OPYC3 GHG integration. Int. J. Climatol., 21, 371–387. ——, 2002: Empirically downscaled temperature scenarios for northern Europe based on a multi-model ensemble. Climate Res., 21, 105–125. Bjørge, D., J. E. Haugen, and T. E. Nordeng, 2000: Future climate in Norway. Research Rep. 103, 41 pp. [Available from DNMI, P.O. Box 43, Blindern, 0313 Oslo, Norway.] Brown, R. D., 2000: Northern Hemisphere snow cover variability and change, 1915–97. J. Climate, 13, 2339–2355. Busuioc, A., D. Chen, and C. Hellstro¨m, 2001: Performance of statistical downscaling models in GCM validation and regional climate change estimates: Application for Swedish precipitation. Int. J. Climatol., 21, 557–578. Carill, A. F., C. G. Mene´ndez, and M. N. Nun˜ez, 1997: Climate change scenarios over the South American region: An intercomparison of coupled general atmosphere–ocean circulation models. Int. J. Climatol., 17, 1613–1633. Charles, S. P., B. C. Bates, P. H. Whetton, and J. P. Hughes, 1999: Validation of downscaling models for changed climate conditions: Case study of southwestern Australia. Climate Res., 12, 1–14. Chen, D., and C. Hellstro¨m, 1999: The influence of the North Atlantic Oscillation on the regional temperature variability in Sweden: Spatial and temporal variations. Tellus, 51A, 505–516. Christensen, J. H., J. Risnen, T. Iversen, D. Bjørge, O. B. Christensen, and M. Rummukainen, 2001: A synthesis of regional climate change simulations—A Scandinavian perspective. Geophys. Res. Lett., 28, 1003–1006. Delworth, T. L., and Coauthors, 1999: Coupled climate modelling at GFDL: Recent accomplishment and future plans. CLIVAR Exchanges, 4, 15–20. Ellner, S. P., 2001: Review of R, version 1.1.1. Bull. Ecol. Soc. Amer., 82, 127–128. Emori, S., T. Nozawa, A. Abe-Ouchi, A. Numaguti, M. Kimoto, and T. Nakajima, 1999: Coupled ocean–atmosphere model experiments of future climate change with an explicit representation of sulfate aerosol scattering. J. Meteor. Soc. Japan, 77, 1299– 1307. Flato, G. M., G. J. Boer, W. G. Lee, N. A. McFarlane, D. Ramsden, M. C. Reader, and A. J. Weaver, 2000: The Canadian Centre for Climate Modelling and Analysis global coupled model and its climate. Climate Dyn., 16, 451–467. Flury, B., 1988: Common Principal Components and Related Multivariate Models. Wiley Series in Probability and Mathematical Statistics, Wiley and Sons, 258 pp. Frich, P., and Coauthors, 1996: North Atlantic Climatological Dataset (NACD Version 1)—Final Report. Danish Meteorological Institute Scientific Rep. 1, 47 pp. Gentleman, R., and R. Ihaka, 2000: Lexical scope and statistical computing. J. Comput. Graphical Stat., 9, 491–508. Gordon, C., C. Cooper, C. A. Senior, H. Banks, J. M. Gregory, T. C. Johns, J. F. B. Mitchell, and R. A. Wood, 2000: The simulation of SST, sea ice extents and ocean heat transports in a version of the Hadley Centre coupled model without flux adjustments. Climate Dyn., 16, 147–168. Gordon, H. B., and S. P. O’Farrell, 1997: Transient climate change in the CSIRO coupled model with dynamic sea ice. Mon. Wea. Rev., 125, 875–907. Grotch, S., and M. MacCracken, 1991: The use of general circulation models to predict regional climate change. J. Climate, 4, 286– 303. Hansen, J. E., M. Sato, R. Ruedy, A. Lacis, and J. Glascoe, 1998: Global climate data and models: A reconciliation. Science, 281, 930–932. Hanssen-Bauer, I., and E. Førland, 2000: Temperature and precipi-

VOLUME 15

tation variations in Norway 1900–1994 and their links to atmospheric circulation. Int. J. Climatol., 20, 1693–1708. ——, and ——, 2001: Verification and analysis of a climate simulation of temperature and pressure fields over Norway and Svalbard. Climate Res., 16, 225–235. ——, O. E. Tveito, and E. J. Førland, 2001: Precipitation scenarios for Norway: Empirical downscaling from the ECHAM4/OPYC3 GSDIO integration. KLIMA Rep. 10/01, 39 pp. [Available from DNMI, P.O. Box 43 Blindern, N-0313 Oslo, Norway.] Hartmann, D. L., 1994: Global Physical Climatology. International Geophysics Series, Vol. 56, Academic Press, 411 pp. Helland-Hansen, B., and F. Nansen, 1920: Temperature variations in the North Atlantic Ocean and in the atmosphere. Smithsonian Institution Miscellaneous Collections, Vol. 70, No. 4, 408 pp. Heyen, H., E. Zorita, and H. von Storch, 1996: Statistical downscaling of monthly mean North Atlantic air-pressure to sea level anomalies in the Baltic Sea. Tellus, 48A, 312–323. Houghton, J. T., Y. Ding, D. J. Griggs, M. Noguer, P. J. van der Linden, and D. Xiaosu, Eds.,2001: Climate Change 2001: The Scientific Basis. Cambridge University Press, 944 pp. Hurrel, J., 1995: Decadal trends in the North Atlantic ocsillation and relationships to regional temperature and precipitation. Science, 269, 676–679. Johns, T. C., R. E. Carnell, J. F. Crossley, J. M. Gregory, J. F. B. Mitchell, C. A. Senior, S. F. B. Tett, and R. A. Wood, 1997: The second Hadley Centre coupled ocean–atmosphere GCM: Model description, spin-up and validation. Climate Dyn., 13, 103–134. Jones, P. D., T. Jonsson, and D. Wheeler, 1997: Extension to the North Atlantic oscillation using early instrumental pressure observations from Gibraltar and south-west Iceland. Int. J. Climatol., 17, 1433–1450. Kalnay, E., and Coauthors, 1996: The NCEP/NCAR 40-Year Reanalysis Project. Bull. Amer. Meteor. Soc., 77, 437–471. Kidson, J. W., and C. S. Thompson, 1998: A comparison of statistical and model-based downscaling techniques for estimating local climate variations. J. Climate, 11, 735–753. Kilsby, C. G., P. P. P. Cowpertwait, P. E. O’Connel, and P. D. Jones, 1998: Predicting rainfall statistics in England and Wales using atmospheric circulation variables. Int. J. Climatol., 18, 523–539. Lorenz, E., 1963: Deterministic nonperiodic flow. J. Atmos. Sci., 20, 130–141. ——, 1967: The nature and theory of the general circulation of the atmosphere. WMO Publ. 218, 59–96. Machenhauer, B., M. Windelband, M. Botzet, J. H. Christensen, M. De´que´, R. G. Jones, P. M. Ruti, and G. Visconti, 1998: Validation and analysis of regional present-day climate and climate change simulations over Europe. Max-Planck-Institut fu¨r Meteorologie Tech. Rep. 275, 80 pp. McGuffie, K., A. Henderson-Sellers, N. Holbrook, Z. Kothavala, O. Balachova, and J. Hoekstra, 1999: Assessing simulations of daily temperature and precipitation variability with global climate models for present and enhanced greenhouse climates. Int. J. Climatol., 19, 1–26. Meehl, G. A., W. D. Collins, B. A. Boville, J. T. Kiehl, T. M. L. Wigley, and J. M. Arblaster, 2000: Response of the NCAR Climate System Model to increased CO 2 and the role of physical processes. J. Climate, 13, 1879–1898. Murphy, J., 1999: An evaluation of statistical and dynamical techniques for downscaling local climate. J. Climate, 12, 2256–2284. ——, 2000: Predictions of climate change over Europe using statistical and dynamical downscaling techniques. Int. J. Climatol., 20, 489–501. North, G. R., T. L. Bell, and R. F. Cahalan, 1982: Sampling errors in the estimation of empirical orthogonal functions. Mon. Wea. Rev., 110, 699–706. Oberhuber, J. M., 1993: Simulation of the Atlantic circulation with a coupled sea ice–mixed layer–isopycnal general circulation model. Part I: Model description. J. Phys. Oceanogr., 23, 808– 829.

1 NOVEMBER 2002

BENESTAD

Preisendorfer, R. W., 1988: Principal Component Analysis in Meteorology and Oceanology. Elsevier Science, 426 pp. Press, W. H., B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling, 1989: Numerical Recipes in Pascal. Cambridge University Press, 759 pp. Ra¨isa¨nen, J., 2001: CO 2-induced climate change in CMIP2 experiments: Quantification of agreement and role of internal variability. J. Climate, 14, 2088–2104. Roeckner, E., and Coauthors, 1992: Simulation of present-day climate with the ECHAM model: Impact of model physics and resolution. Max-Planck-Institut fu¨r Meteorologie Tech. Rep. 93, 171 pp. ——, and Coauthors, 1996: The atmospheric general circulation model ECHAM4: Model description and simulation of present-day climate. Max-Planck-Institut fu¨r Meteorologie Tech. Rep. 218, 99 pp. Rummukainen, M., 1997: Methods for statistical downscaling of GCM simulations. SWECLIM Rep. 80, SMHI, 29 pp. ——, J. Ra¨sa¨nen, and P. Graham, 1998: Regional climate simulations for the Nordic region—First results from SWECLIM. SWECLIM Rep., SMHI, 22 pp. Sengupta, S. K., and J. S. Boyle, 1993: Statistical intercomparison of global climate models: A common principal component approach PCMDI Rep. 13, Lawrence Livermore National Laboratory, Livermore, CA, 44 pp. [Available online at http://wwwpcmdi.llnl.gov/pcmdi/pubs/pdf/13.pdf.]

3027

Strang, G., 1995: Linear Algebra and Its Application. Harcourt Brace, 505 pp. Tett, S. F. B., J. F. B. Mitchell, D. E. Parker, and M. R. Allen, 1996: Human influence on the atmospheric vertical temperature structure: Detection and observations. Science, 274, 1170–1173. ——, P. A. Stott, M. R. Allen, W. J. Ingram, and J. F. B. Mitchell, 1999: Causes of twentieth-century temperature change near the Earth’s surface. Nature, 399, 569–572. von Storch, H., and F. W. Zwiers, 1999: Statistical Analysis in Climate Research. Cambridge University Press, 484 pp. ——, E. Zorita, and U. Cubasch, 1993: Downscaling of global climate change estimates to regional scales: An application to Iberian rainfall in wintertime. J. Climate, 6, 1161–1171. Watanabe, M., and T. Nitta, 1998: Relative impacts of snow and sea surface temperature anomalies on an extreme phase in winter atmospheric circulation. J. Climate, 11, 2837–2857. Wilby, R. L., T. M. L. Wigley, D. Conway, P. D. Jones, B. C. Hewitson, J. Main, and D. S. Wilks, 1998: Statistical downscaling of general circulation model output: A comparison of methods. Water Resour. Res., 34, 2995–3008. Wilks, D. S., 1995: Statistical Methods in the Atmospheric Sciences. Academic Press, 467 pp. Zorita, E., and H. von Storch, 1997: A survey of statistical downscaling results. Tech. Rep. 97/E/20, GKSS, 42 pp. Zveryaev, I., 1999: Decadal and longer changes of the winter sea level pressure fields and related synoptic activity over the North Atlantic. Int. J. Climatol., 19, 1177–1185.