comparison of climate change scenarios generated from

JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 104, NO. D6, PAGES 6603-6621, MARCH 27, 1999

Comparison of climate change scenarios generated from regional climate model experiments and statistical downscaling

L. O. Mearns, x I. Bogardi, e F. Giorgi, x'3 I. Matyasovszky, 4 and M. Palecki e,•

Abstract. We compare regional climate change scenarios (temperature and precipitation) over eastern Nebraska produced by a semiempirical statistical downscaling (SDS) technique and regional climate model (RegCM2) experiments, both using large scale information from the same coarse resolution general circulation model (GCM) control and 2 x COe simulations. The SDS method is based on the circulation pattern classification technique in combination with stochastic generation of daily time series of temperature and precipitation. It uses daily values of 700 mbar geopotential heights as the large-scale circulation variable. The regional climate model is driven by initial and lateral boundary conditions from the GCM. The RegCM2 exhibited greater spatial variability than the SDS method for change in both temperature and precipitation. The SDS method produced a seasonal cycle of temperature change with a much larger amplitude than that of the RegCM2 or the GCM. Daily variability of temperature mainly decreased for both downscaling methods and the GCM. Changes in mean daily precipitation varied between SDS and RegCM2. The RegCM2 simulated both increases and decreases in the probability of precipitation, while the SDS method produced only increases. We explore possible dynamical and physical reasons for the differences in the scenarios produced by the two methods and the GCM.

1. Introduction

One of the most significant factors that has hampered progress in the determination of the specific impacts of climate change on relevant Earth systems is the lack of detailed scenarios of regional climate change [$fnith and Tirpak, 1989]. To date, numerous climate change impact assessments have been performed using scenarios generated from the output of general circulation model (GCM) increased CO2 experiments, but these scenarios have lacked regional detail due to the coarse model spatial resolution (currently, between 3 ø and 6 ø of latitude, corresponding roughly to 300-600 km). These models perform relatively poorly at the regional scale [Grotch and McCracken, 1991], which limits the models'

• National Center for Atmospheric Research, Boulder, Col- orado.

2University of Nebraska, Lincoln. 3Also at Physics of Weather and Climate Group, Abdus

Salam International Centre for Theoretical Physics, Trieste, Italy.

4Eotvos Lorand University, Budapest. 5Now at Illinois State Water Survey, Champaign.

Copyright 1999 by the American Geophysical Union.

Paper number 1998JD200042. 0148-0227/99/1998JD200042509.00

accuracies at large scales as well [Mearns et al., 1990; Giorgi and Mearns, 1991; Risbey and Stone, 1996]. Two major techniques have been developed to make up for this deficiency: statistical downscaling of general circulation model results and nested regional climate modeling within GCMs [Giorgi and Mearns, 1991; McGregor, 1997; Wilby and Wigley, 1997]. Both methods rely on the coarser GCMs to provide large-scale information on the atmospheric circulation, but each generates regional details in substantially different ways. While these techniques have been explored for close to 10 years (longer in the case of statistical downscaling) and various claims of the relative merits of the two techniques have been made, to our knowledge there has not been any direct comparisons of these methods.

In this paper we provide a comparison of climate change scenarios due to a doubling of carbon dioxide concentration (2xCO2) produced with a nested regional model (the NCAR RegCM2 [Giorgi et al., 1993a, b]) and a semiempirical statistical downscaling (SDS) method [Matyasovszky et al., 1993, 1994]. Both methods make use of large-scale fields from equilibrium control run and 2 x CO2 experiments completed with the Commonwealth Scientific and Industrial Research Or-

ganization (CSIRO) GCM [Watterson et al., 1995]. The area of interest is eastern Nebraska, where scenarios from the two methods are compared at a number of locations.

66O3

6604 MEARNS ET AL.: REGIONAL AND STATISTICAL DOWNSCALING COMPARISON

Our purpose is (1) to determine how similar/dissimilar the resulting higher-resolution climate change scenarios are and how they compare with results from the GCM, providing the boundary conditions used for the nested model experiments and the synoptic scale circulation variables used in the SDS method; (2) to explore the causes for differences found, particularly in regard to different causal agents of the Earth atmosphere system and assumptions used in the two techniques; and (3) to determine what these differences mean in terms of the

strengths and weaknesses of these techniques. The paper is divided into the following sections: the

remainder of the introduction provides a general comparison of the two methods; section 2 describes the techniques used for this particular application, as well as the general characteristics of the study site. In section 3 we present the results of the experiments for temperature and precipitation; and in section 4 we present our discussion and conclusions.

Semiempirical statistical downscaling attempts to translate large-scale, coarse resolution GCM information into local, high-resolution statistics of surface variables of interest (usually temperature and precipitation) by using empirically derived relationships between observed large scale and local surface variables [Giorgi and Mearns, 1991, Wilby and Wigley, 1997]. A number of different types of empirical downscaling techniques have been developed for different surface variables at different temporal scales. Early efforts, such as that of Kim et al. [1984], established statistical relationships (using regression techniques) between the large area average surface temperature and the local surface temperature on a monthly timescale. Regression techniques evolved rapidly to include surface and upper air variables [e.g., Karl et al., 1990; Wigley et al., 1990), and consideration of daily time scales [e.g., Wilks, 1989].

An alternate statistical downscaling technique is the so-called weather pattern approach [(Lettenmaier, 1995; Wilby and Wigley, 1997], whereby large-scale weather patterns are classified into different circulation types. Then parameters required for the stochastic generation of, for example, daily precipitation, are conditioned on the different weather types. This approach has appeal because it is believed that more physical meaning is incorporated into the relationships, since weather patterns control aspects of daily precipitation patterns [Bardossy and Plate, 1991, 1992; Wilson et al., 1991, 1992; Matyasovszky et al., 1993, 1994;Letten- maier, 1995; Hughes et al., 1993; Katz and Parlange, 1993].

Numerous works on statistical downscaling have been published, and overviews of techniques can be found in the works of Giorgi and Mearns [1991],Lettenmaier [1995]; Hewitson and Crane [1996] and Wilby and Wig- ley [1997].

Regardless of the techniques employed, statistical downscaling has the following drawbacks when applied in a climate change context: (1) It must be assumed that the relationships between the large scale pattern

and the local variable remain the same under the perturbed climate as in the observations; (2) it can only be applied in regions where there are sufficient observed large-scale and local scale variables for developing the empirical relationships; (3) it cannot be employed where the relationship between the predictor and the predic- tand is weak; and (4) it may be valid only within the range of the input data.

Nested regional climate modeling explicitly considers the effects of mesoscale forcings by increasing the model resolution over the limited area of interest. A

high-resolution limited area model is driven by lateral boundary conditions provided by a coarse resolution GCM. The basic strategy underlying this approach is that the coarse resolution GCM simulates the response of the (large scale) general circulation to global forcings and the higher-resolution regional model describes the effects of local, sub-GCM grid scale forcings on the atmospheric circulation and distributions of climatic variables over the selected region. Since the original development of this technique for use under perturbed climate conditions [Giorgi, 1990; Giorgi and Mearns, 1991], numerous experiments with the NCAR RegCM have been performed over North America[Giorgi et al., 1994; 1998], Europe [Marinucci and Giorgi, 1992], and eastern Asia Hirakuchi and Giorgi, 1995]. Moreover, several climate modeling groups have developed nested regional modeling systems [Pielke et al., 1992; Jones et al., 1995, 1997; Lynch et al., 1995; Christensen et al., 1997, 1998; McGregor and Walsh, 1993; McGregor, 1997; Laprise et al., 1998; Machenhauer et al., 1998] and applied them to a variety of studies. For reviews of regional modeling experiments, see Giorgi and Mearns [1991], Giorgi [1995], Kattenberg et al. [1996], and Mc- Gregor [1997].

While the regional modeling method is obviously more physically based than any of the semiempirical downscaling approaches, it has some disadvantages as well. (1) The modeling is computationally expensive; (2) the control run simulations at the regional scale still suffer from inaccuracies; (3) the quality of the regional model control run is dependent on the quality of the of the GCM run providing the boundary conditions; and (4) nested models usually require much tuning of parameterizations when applied to new regions. Its ad- vantages include (1) the use of physically based models that can respond in physical ways to different external forcings; (2) it can, in principle, be applied anywhere on Earth (no data limitations); and (3) it can be used at a number of different resolutions.

Thus these two techniques have much in common but are methodologically quite different. In comparing these techniques we hope to shed some light on important questions such as follows: Do differences in the results obtained from the nested model and the SDS

model reflect on different strengths and weakness of the techniques, and is there any sense in which one can refer to one response being more "correct" than the other?

MEARNS ET AL' REGIONAL AND STATISTICAL DOWNSCALING COMPARISON 6605

2. Methods

2.1. General Characteristics of the Study Region

We analyze the results at 12 locations (and corresponding RegCM2 grids) located in eastern Nebraska (Figure 1). One additional station (13 in Figure 1) is used for analysis of temperature. These locations fall within the areas of four different CSIRO grid boxes also indicated in the figure. The observed data consist of 32 years (1962-1993) of daily mean temperature and precipitation data from the Historical Daily Climate Data Set [Robinson et al., 1995]. The data come mostly from cooperative network and smaller first-order stations and have undergone extensive quality control. The area cov- ers about 2.5 ø in latitude and 3.5 ø in longitude.

The climate in this region is characterized by a steep east-to-west decreasing precipitation gradient and a lat- itudinal temperature gradient (Table 1). Temperature differences between the most southern versus the most

northern stations are about 2øC, both in summer and

in winter. East-to-west precipitation contrasts range from about 30 mm in the winter to about 57 mm in the

summer.

2.2. The CSlRO General Circulation Model

Simulations with the CSIRO GCM described by Waf- terson et al. [1995] were used to drive the RegCM2 and provide large-scale circulations for the downscaling model. The GCM was run at R21 horizontal spectral resolution (• 5 ø in physical space), with nine vertical levels and full physics representation. Clouds are di- agnosed as a function of relative humidity; precipitation occurs when supersaturation is attained in con- vectively stable conditions and via a "relaxed" moist convective adjustment in the presence of convective in- stability. Soil temperatures and water content are calculated through a force-restore method, and planetary boundary layer processes are represented via a local vertical diffusion scheme. The atmospheric model is coupled to a thermodynamic ice model [Parkinson and Washington, 1979] and to a mixed layer ocean model.

L I

* ?2

i '"

Figure 1. Station location map, also showing RegCM2 grid points (pluses) and CSIRO grid boxes (dashed line). Locations: (1) Auburn; (2) Hartington; (3) Kearney; (4) Pawnee City; (5) Albion; (6) Atkinson; (7) Central City; (8) Crete; (9) David City; (10) Fairbury; (11) Geneva; (12) West Point; and (13) Alma (used only for temperature analysis).

6606 MEARNS ET AL.' REGIONAL AND STATISTICAL DOWNSCALING COMPARISON

Table 1. Observed and RegCM2 Statistics for Selected Stations/Grids

Atkinson Hartington Auburn Kearney

Temperature ( o C) Winter

OBS -4.6 -5.3 -2.7 -3.9 RegCM2 -4.4 -4.5 -2.5 -2.0

Summer

OBS 22.0 22.9 23.9 22.6 RegCM2 22.4 22.3 24.1 22.9

this region was captured, as was the seasonal migra- tion of the midlatitude jet, associated storm track, and easterly motion of the Bermuda high. The most pro- nounced model deficiency was a misplacement of the center of maximum summer precipitation [Giorgi et al., 1998]. In 2 x CO2 conditions the CSIRO GCM produced a seasonally averaged warming over the Central Plains in the range of 4.5øC (summer) to 5.7øC (spring). Simulated precipitation change was mostly positive and greatest in the spring (~ 37%).

Precipitation (Total ram) Winter

OBS 46.4 48.8 78.3 51.0 RegCM2 58.5 62.2 76.3 55.3 % error 26.0 27.0 -2.0 8.0

Summer

OBS 257.3 270.7 314.9 268.2 RegCM2 307.9 328.7 354.0 303.1 % error 20.0 21.0 12.0 13.0

OBS, 30 years observed temperature or precipitation; RegCM2, temperature or precipitation from 5-year control run of the RegCM2.

Horizontal ocean flux correction is included. The global mean surface temperature increase under 2 x CO2 conditions is approximately 5øC. The global climatology of the model in both the control and the 2 x C02 simulations is described by Watterson et al. [1995].

The CSIRO GCM was selected for this work because

it provided a reasonably good simulation of present-day climate over North America compared to other GCMs. Giorgi et al. [1998] provide a description of the CSIRO model climatology over the Central Plains. The seasonal cycle of both temperature and precipitation over

2.3. Regional Climate Model (RegCM2)

2.3.1. Description. The NCAR RegCM2 is an augmented version of the NCAR-Pennsylvania State University mesoscale model MM4 [Giorgi et al., 1993a, b). It is a primitive equation, Crp vertical coordinate, grid point limited-area model with compressibility and hydrostatic balance. For application to climate studies, a number of physics parameterizations were incorporated in the model, namely a state-of-the-art surface physics package (the Biosphere-Atmosphere Trans- fer Scheme (BATS) [Dickinson et al., 1993], an explicit planetary boundary layer formulation [Holtslag et al., 1990], a detailed atmospheric radiative calculation package [Briegleb, 1992], a mass flux cumulus param- eterization scheme [Grell et al., 1994], and a simpli- fied explicit moisture scheme including an equation for cloud water[(Giorgi and Marinucci, 1996]. Note that BATS includes the effects of vegetation on the surface- atmosphere exchanges and explicitly calculates snow cover and soil water content [Dickinson et al., 1993].

The model domain (Figure 2) encompasses the western United States and the Central Plains at a horizontal

grid point spacing of 50 km. Fourteen vertical levels are used between the surface and the model top (80 mbar),

! ;

! ,,

; ,,

; [

,, 0

.. oooo ;

Figure 2. Regional climate model domain. Rectangular area indicates region oœ focus in the work of Georgi et al. [1998].

MEARNS ET AL.: REGIONAL AND STATISTICAL DOWNSCALING COMPARISON 6607

with a vertical resolution of approximately cr- 0.1 in the troposphere and five levels below about 1500 m. Two continuous simulations were carried out, a 5-year present day run and a 5-year 2 x CO2 run with the RegCM2 driven by time-dependent lateral meteorological fields provided by the output from the corresponding GCM simulations. The nesting technique employs a standard relaxation method for wind, temperature, water vapor and surface pressure applied over a lateral buffer zone of 10 grid points, and uses an altitude- dependent exponential weighting coefficient for the relaxation terms [Giorgi et al., 1993b]. The GCM forcing data were provided at 8-hour intervals, with lin- ear interpolation to each model time step (3 min). In the RegCM2 runs, initial soil water content was interpolated from the GCM-normalized soil model water content, and time-dependent sea surface temperatures (SSTs) were interpolated from the ocean component of the GCM. For comparison with the SDS method, we took RegCM2 grid points closest to each of the 12 observed stations.

2.3.2. Comparison of point observations with grid box areas. An important methodological issue in validation of climate models is how to appropriately compare climate variables from point observations with the grid box area values of the same variables from climate models. Mearns et al. [1995a, b] investigated the effect of aggregation of point observations when validat- ing a 60 km resolution run of the RegCM over the continental United States, including regions in the Great Plains. They found that the observed statistics (especially of higher-order moments of climate variables, such as frequency of precipitation) did change as the number of stations used to represent the grid area was varied, but at the 60 km scale the effect was quite small in most instances. The uncertainty due to possible mismatch of spatial scale was considerably less than the magnitude of model errors, most particularly for precipitation frequency.

More recently, Osborne [1997] and Osborne and Hulme [1997] have examined the relationship between station and grid box statistics of precipitation intensity and frequency and developed a technique whereby the statistical properties of grid box precipitation can be estimated from a small finite number of single-point stations. They applied the technique to the validation of relatively coarse resolution GCMs.

The point-area issue, however is of less significance when considering mean monthly or annual average values of temperature or precipitation, and is usually of less importance for temperature. We note this issue since we do compare the single observation station values with RegCM2 grid box values. The main effect would be seen in precipitation frequency and somewhat in daily variance of temperature. As described in Mearns et al. [1995a], the bias introduced by using only one observation station is relatively small. How- ever, we note that the probability of precipitation is

slightly underestimated and the daily intensity overestimated when using a single point to represent a 50 km grid box in the Great Plains. Daily variance of temperature is also slightly overestimated.

2.3.3. RegCM2 control run validation. De- tailed validation of the central Great Plains region of the RegCM2 domain is presented by Giorgi et al. [1998] and L. O. Mearns et al. (1998, manuscript in prepara- tion, hereinafter referred to as M98). Here we describe general characteristics of the control run for the region of focus in this article, eastern Nebraska.

On an annual average, mean temperatures are within 1 ø of observed and the north to south temperature increase is well represented. For example, the annual mean observed temperature at Fairbury, a southern Ne- braska site, is 10.5øC and the simulated temperature 10.8 •, whereas for Hartington in the north, the corresponding observed and modeled values are 9.6 ø and 8.5 •, respectively. The precipitation gradient, increasing precipitation from west to east, is well captured by the model. For example, the western location of Atkin- son, Nebraska, receives on average 608 mm annual total, .while Hartington farther east, receives about 660 mm total. The annual average totals for the corresponding model grid boxes are 606 and 686 mm, respectively. Winter and summer contrasts for several stations and

the corresponding RegCM2 grid boxes (Table 1) indi- cate very good spatial representation of the observed climate by RegCM2, but seasonal values of precipitation are overestimated by 25% for the northern stations in winter and about 20% in the summer. However, as has been noted elsewhere [Mearns et al., 1995b, M98], the relatively good reproduction of mean precipitation is based on compensating errors in the frequency and intensity fields of precipitation. For the model, frequency is usually double that of the observations and intensity underestimated by greater than 50•0.

Naturally, the validation data set for the empirically downscaled method is more accurate than the control

run of the RegCM2, since the validation is on the observed relationship not on the control run of the CSIRO (see below).

2.3.4. Semiempirical statistical downscaling (SDS). The semiempirical downscaling approach is based on a stochastic linkage between large-scale atmospheric circulation and local climatic variables. Ob- served atmospheric circulation and observed local climatic variables are used to estimate the parameters of the stochastic linkage model and to validate the model by comparing the stochastic properties of simulated climatic variables with an independent observed data set. Then, large-scale circulation output of the GCM is used with the calibrated stochastic linkage model to obtain simulated time series of the climatic variables under 1

x CO2 (control) and 2 x CO2 conditions. Two main points motivate the use of this method: (1) there is a clear stochastic linkage between large-scale circulation and several local climatic variables, and (2) large-


scale circulation is one of the more accurate outputs of GCMs. On the other hand, the methodology assumes that the stochastic linkage stays invariant under chang- ing climate.

Large-scale circulation can be characterized by the spatial distribution of several variables. Lower or mid- dle tropospheric geopotential heights are applied for this purpose. In the present SDS model the 700 mbar geopotential height is used.

The SDS method used here is based on that of Bar-

dossy and Plate [1991, 1992], which was modified in the following ways: An objective classification technique is used to classify circulation pattern (CP) types; a Markov chain is used to simulate the sequence of CP types; and a general transformation is used rather than a power transformation to render the distribution

of daily climate variables more normal [Bogardi et aL,

2.3.4.1. Analysis and modeling of atmospheric circulation.' Three types of daily circulation pattern (CP) data are used.

(1) Historical data are represented by the National Center for Environmental Prediction (NCEP) grid point analyses of the height of 700 mbar pressure fields available from the National Center for Atmospheric Re- search (NCAR). The analysis is based on daily values (1200 UT) at 40 points on a diamond grid covering the sector 25ø-60øN, 80ø-125•W for the period January 1962 to December 1993.

(2) A 5-year-long data series for the same pressure level has been obtained from the CSIRO GCM output corresponding to the 1 x CO2 (control)run.

0.2

0.15

0.1

0.05

CP1 CP2 CP3 CP4 CP5 CP6 CP7 CP8 CP9

r----] Historical • lxCO2 • 2xCO2 [ 0.2

o.15

0.1

0.05


Historical • lxC02 • 2xC02 Figure 3. Relative frequency of circulation pattern (CP) types for observed, CSIRO control (1 x CO2) and 2 x CO2' (a) winter, (b) summer.

MEARNS ET AL.' REGIONAL AND STATISTICAL DOWNSCALING COMPARISON 6609

(3) An analogous series has been obtained from the 2 x C02 scenario.

To model the occurrence of daily CP types, first the daily 40 point field was classified. In the present paper the periods from April to September (summer) and from October to March (winter) have been examined separately. Pressure height values at each grid point are standardized to exclude the annual cycle. Classification over shorter time periods would be preferable, but the available data set is not sufficiently large to estimate the stochastic properties (transition probabilities) of twice nine classes of CP types (separate for winter and summer). On the other hand, using two seasons for CP classification has resulted in a very good representation of the present climate [Bogardi et al., 1993].

The classification is performed by principal component analysis (PCA) coupled with k-means cluster analysis because a conjunctive use of these techniques usually provides the most separable system of clusters with the most concentrated clusters [Gong and Richman, 1992; Bartholy, 1992].

The relative frequencies of CP types are shown in Figure 3 for the two 6-month periods. Figure 4 illustrates several representative circulation pattern types based on their spatial patterns of 700 mbar heights. Figure 4a represents a wet CP type in winter, which is characterized by a large-amplitude trough over the western central United States. It results in southwest-

erly flow over the central and eastern United States. This pattern is wet for central areas because it is in a location relative to the trough favored for baroclinic in- stability. Figure 4b represents a dry winter type. It is a pattern with a weak ridge over western-central U.S.

and with strong west-northwesterly near zonal flow over the northern part of the continent. The first type for summer (Figure 4c) is similar to the first winter one but obviously with smaller gradients. The pattern is quite wet. The second summer type (Figure 4d) corresponds to a warm and dry pattern, but on occasion, wet and relatively cold air originating from the Pacific can yield considerable amounts of local precipitation. Other illus- trations of summer and winter CP types are presented and ex151ained in detail by Bogardi et al. [1993] and Matyasovszky et al. [1993].

The differences between relative frequencies of CP types (Figure 3) corresponding to the historical 1 x CO2 (control) and 2 x CO• cases are not significant. Al- though we used the slightly different relative frequencies and transition probabilities in the 1 x CO• and 2 x CO• scenario, this effect alone cannot represent the actual synoptic driving force of climate change. CP characteristics within a given CP type may change, which means that knowledge of the CP type alone may not be sufficient to characterize the CP types under climate change conditions. Therefore an additional parameter, namely the spatially averaged height of the 700 mbar pressure level has been introduced, which can be considered as an indicator of the air pressure (low or high) and the temperature (cold or warm). The anticipated warming due to the increasing concentration of CO• is accompa- nied with the expansion of the atmosphere, so pressure levels are located at greater heights.

Figure 5 shows the spatially averaged geopotential height of the 700 mbar pressure field for each CP type, for the observed, 1 x CO2 (control), and 2 x CO• CP data sets. Three important observations can be made.

2880'

ß .

' i 2800 '"' b ,'"" " ß -.. ..

,,, ',. ,, .. ,,.

,.-' 2880 .,' .. ß

ß

,,' ... ,' .... 29-20 .," ..

,,' ,. 2960 ' ' :

....." ':

.. . ......

Figure 4. Representative atmospheric CP types: (a) CP3 in the w•ter half of the year, (b) CP5 in the winter half of the year, (c) CP4 in the summer half of the year, (d) CP7 in the summer half of the year. Contours are 700 mbar geopotential heights in meters.

6610 MEARNS ET AL' REGIONAL AND STATISTICAL DOWNSCALING COMPARISON

Figure 4. (continued)

(1) The historical and the 1 x CO2 pressure heights are similar, which indicates that the GCM reproduces quite accurately this large-scale property; (2) the geopotential height significantly increases in the 2 x CO2 case; (3) the increase is larger in the winter season. On the basis of these findings, the stochastic downscaling is ex- pected to provide 1 x C02 (control) climate close to the observed climate, i.e., with a small bias, and to produce larger changes in the winter than in other seasons. Our results confirm these expectations.

The occurrence and the persistence of the daily CP types is described by a Markov chain model fully characterized by the matrix of transition probabilities from one CP type to another.

2.3.4.2. Relating local climate to CP types and average geopotential height: To reproduce the space-time statistical structure of local climate variables, a suitable model must be chosen. Autoregres- sive processes represent a well-developed and commonly used tool to model time series. They have been developed principally for Gaussian processes, but climatic factors do not usually follow a Gaussian distribution. Therefore it is desirable to construct a transformation

establishing a relationship between the distribution of a local climatic factor and a normal distribution. Let the

vector Z(t) = (Z(t, ul), Z(t, u2),..., Z(t, UK)) represent a daily climatic variable at locations ux, u2,..., UK and time t and let W(t) be a K-dimensional normal random vector at time t. We suppose for simplicity that each component of the vector W(t) has unit variance. The time dependency of W(t) is described using mul- tivariate first order autoregressive (AR(1)) processes. The transformation of the random vector Z(t) into the normal vector W(t) depends on the climatic variable

under consideration. In eastern Nebraska, daily mean temperature cannot be described by a simple normal distribution [Matyasovszky et al., 1994]. Because the likelihood of having temperatures higher or lower than the most probable value differs, a distribution is used that is similar to the Gaussian except for its symme- try. The binormal distribution developed by Toth and $zentimrey [1990] satisfies this requirement. Binormal distributions are fitted to the temperature data series conditioned on CP types.

As stated earlier, within each CP type, daily temperature and precipitation also depend on the actual spatial average height of the 700 mb height field. The basic idea is to include the spatially averaged pressure height into the analysis. The annual cycle of the pressure height is considered as an analogy of the difference between present and 2 x CO• climates. The relationship between the probability distribution of temperature or precipitation and the spatial average height is described using historical data. A regression approach is used. Further details on the use of average geopotential height are included in section 4, on using the model to generate the climate change scenario.

The main difficulty of modeling precipitation is its space-time intermittence. The precipitation occurrence at a given location must be conditioned on precipitation at other locations; then the precipitation amount (if any) must be conditioned on occurrence and amount at other locations. This approach requires the estimation of many parameters. As before, the conditional (on CP type) probability distribution of precipitation is far from normal, so a general transformation is needed to establish a relationship between precipitation and a normal distribution. A Fourier series expansion of this


3150

E 31oo

• 3050 ._

o

o

o 3000

295O

_ - ,•

i i i i i i i i


Historical • lxC02 • 2xC02 I 3150

3100

3050

3000

2950 CP1 CP2 CP3 CP4 CP5 CP6 CP7 CP8 CP9

Figure 5. Spatially averaged geopotential height of the 700 mbar pressure level for each CP type, for the observed, control (1 x CO2), and 2 x CO2 CP data sets' (a) winter, (b) summer.

transformation is used and then the Fourier coefficients

are calculated as a function of the geopotential height as described by Matyasovszky et aL [1994].

2.3.4.3. Model calibration and validation with

simulation: The daily temperature and rainfall data sets are divided into two parts: the first half of the data (1962-1977) is used to estimate the statistical parameters of the double conditional distributions of daily temperature and precipitation and the rest of the data (1978-1993) is used to validate the model. The following simulation procedure is used: (1) generate a possible daily CP type according to the Markov process estimated from the calibration data set; (2) select pa-

rameters of the autoregressive process according to the generated CP type and average geopotential height of the CP type and use this process to generate a daily temperature and rainfall amount; (3) repeat the previous two steps to generate a time series of daily temperature and precipitation; and (4) compare the stochastic properties (e.g., mean, standard deviation) of the generated time series and the observed calibration time series.

As an example, Figure 6 shows the observed and simulated histograms of daily mean temperature for Hart- ington. From spring to fall the simulated histograms satisfactorily fit the observed histograms (July, Fig-

6612 MEARNS ET AL' REGIONAL AND STATISTICAL DOWNSCALING COMPARISON

0.12

0.08

0.06

0.04

0.02

a o , ,

-30 -20 -10 i i

o lO

0,12

0.1

0.08

0.06

0.04

0.02

b o i i

-30 -20 -10 0 10

0'12 t 0.1

0.08

0.06 f 0.04 f 0'01

1 õ 20 I I

25 30

o C

35

0.12

0,1

0.08

0.06

0.04

0,02

d o i i

15 20 i i

25 30

o

c

Figure 6. Simulated and observed frequency distributions of temperature for Hartington: (a) January observed, (b) January simulated, (c) July observed, (d) July simulated.

ures 6c, 6d). However, in winter (January, Figures 6a, 6b) the simulation behaves poorly in that the asym- merry of observed distribution is not well reproduced. As was mentioned earlier the probability distribution of temperature under the climate of eastern Nebraska is not normal in winter. Therefore the distribution of

temperature conditioned on CP types was described by a binormal distribution, but the unconditional distribution (superposition of conditional ones) becomes very close to normal.

The $DS method used in this paper has been applied numerous times in Nebraska to observed surface variables and 500 or 700 mbar heights from the NCEP data, and numerous validations have been presented [e.g., Bogardi et al., 1993;Matyasovszky et al., 1994]. Thus we do not present a complete validation here. We note, however, that the model performs quite well, and

daily temperature and precipitation characteristics for the 12 or 13 stations are very well reproduced. We note that the SDS model validates much better than

the regional climate model (described above), but this is necessarily the case. The SDS model is calibrated on observed data, and then validated with observed data. Although the 700 mbar heights from the CSIRO GCM were more accurate than GCMs used with this method

in the past[(Matyasovszky et al., 1994], when we used the control run 700 mb heights directly, some months were well reproduced, but others were not (the case of October is discussed in section 3).

In the case of the RegCM2, no observed data are directly used in the model control or doubled CO2 simulations, and the regional model is driven by another (global) model that also has biases. Hence, given the state of art of regional and general circulation models,


it is inevitable that the semiempirical statistical downscaling method will present a better validation than the regional climate model.

2.3.4.4. Use with GCM circulation pattern data: This final procedure is based on the calibrated model and consists of a simulation procedure similar to the one used for the model validation. Specifically, the first two steps of the procedure change: instead of the historical CP type and geopotential information, the corresponding 1 x CO2 and 2 x CO2 results of the GCM are used.

We performed this in two different ways. In the first approach, the 1 x CO• case is not used directly. The 2 x CO• case is modified by adjusting GCM- generated geopotential heights under 2 x CO• by the difference between observed geopotential height and GCM-generated geopotential height under 1 x CO•. The means of incorporating the effect of changes in average geopotential height without using the 1 x CO2 case directly involves making a correction that takes into account the difference between the observed and

the control run average geopotential heights. Proba- bility distributions are estimated corresponding to the 1 x CO2 and 2 x CO• monthly mean heights of the 700 pressure field by selecting two subsets corresponding to geopotential heights coming from a neighborhood of monthly mean heights of the 700 mb geopotential fields under GCM-generated 1 x CO• and 2 x CO• climates. Then the probability distributions calculated from the whole historical data set is adjusted according to the difference between the two above mentioned distributions (the A CO• case). Specific forms of this adjustment for temperature and precipitation can be found in the work of Matyasovszky et al. [1994]. Re- suits of this technique are shown in this paper.

The second approach directly uses the 1 x CO2 and 2 x CO• results from the GCM. As is seen (Figure 5) the CSIRO GCM has, on average, a small bias in repro- ducing 700 mb geopotential heights for the present climate, and therefore there is little difference between the two approaches concerning estimated climate changes. The monthly mean temperatures estimated by SDS in the 1 x CO• case are smaller than the observed means, especially in the fall. There are two reasons for this. The geopotential heights obtained from the CSIRO 1 x CO• run are somewhat systematically smaller than the observed heights (Figure 5). Secondly, the basis for defining the CP types is a half year, but the frequency of types within a half year is not constant. In October, for instance, when the SDS mean temperatures under 1 x CO2 differ substantially from the observed means, the bias in the frequency of CP types from the CSIRO is considerably larger compared to in December, say, when the bias is much smaller.

3. Comparison of Climate Change Results

To compare the climate change results from the CSIRO, RegCM2, and the SDS method, we applied a

number of statistical tests to the results of the experiments. However, the results of these tests should be viewed as providing only one, among other, pieces of information about the results. Since the data being compared have been created in such different ways and because of the very different sample sizes obtained, the results of the statistical tests must be interpreted with caution.

The results of the climate model experiments have definitive sample sizes. Only 5 years of control and doubled CO• runs of the RegCM2 were created. And only 5 years of the CSIRO climate model runs were used to provide boundary conditions. In the case of the SDS results, parameters for generating time series of climate change are estimated from a combination of the 16 years of the observed data sets of surface observations, and 5 years of daily 700 mbar heights from the one and 2 x CO• experiments of the CSIRO. Then daily time series of present and 2 x CO• temperature and precipitation are simulated using these estimated parameters. Any length of time series could be generated, and in that regard, the potential sample size is in- finite. To obtain good statistics of the time series, fairly long series should be generated. In the present context, originally 160 years were generated. This means that a very large difference in "sample size" exists between the SDS time series and time series from the climate model runs. This condition should be taken into consideration

in interpreting results of the statistical tests. We do not emphasize the test results.

3.1. Temperature

In presenting the results for temperature we discuss four representative stations for all months of the year. We consider the mean monthly change in temperature and the change in daily variance of temperature. Change in variability is important from an impact point of view, since it can have substantial effects on the frequency of extreme events [Katz and Brown, 1992]. The four representative stations (indicated in Figure 1) are Hartington, Atkinson, Kearney, and Auburn. These stations were chosen to provide distinct contrasts across eastern Nebraska. Hartington versus Auburn provides a north-south contrast, along the eastern border of the region; Atkinson and Hartington provide an east-west contrast to the north, and Kearney versus Auburn provides something of an east-west contrast toward the south. Also, these stations fall in different CSIRO grid boxes, so all four are represented (Figure 1). Each plot shows the results of the CSIRO model, RegCM2, and the SDS method. The top panel of each plot shows the seasonal cycle of change in mean temperature, and the bottom panel shows the change in the standard deviation of daily mean temperature.

We performed statistical tests on the changes in the mean and in the innovation variance of daily temperature. The innovation variance is the variance remaining after the autocorrelation structure of the time series has been removed. Most of the decreases in innovation vari-


ance were significant (at the 0.05 level), and the change in innovation variance and raw variance were very similar. Details of these tests as applied to other regional model experiments may be found in the work of Mearns et al. [1995a], and development of the tests may be found in Katz [1988].

3.1.1. Major findings. The findings are as follows.

3.1.1.1. Seasonal cycle of mean temperature change: The most striking contrast among the CSIRO, RegCM2, and SDS is the very large seasonal cycle of temperature change in the SDS results (Figure ?), which is not found in the physical model results. Values are generally higher than those of the two climate models in

winter but much smaller in summer (less than 1ø), and this is the case at all locations. The very small changes were actually found only for June, July, and August, rather than for the full half year defined as the summer period in the SDS method. The climate model results generally resemble each other, except in March at the NE CSIRO grid.

3.1.1.2. Daily temperature variability: The change in variance of daily temperature is of the same order of magnitude for the three simulations (Figure 7), but generally only variance decreases are found in the SDS case. The largest decreases in both SDS and RegCM2 occur during the winter months. For the RegCM2 often the smaller decreases in spring fail to

10

9

½- 8

• 7

o 5 o

4

10

9

½' 8

• 7

0 5 o

4

2

1

0

0 0 o

Monfh

10

9

½- 8

• 7 F 6

o 5 o

4

2

1

0

0 0 o

Auburn

-- ,% ---- SDS I -• .-_./•• ---REGCM / •- -"/ .......... •g::'"x --- CSIRO I z• - z.--' .... ,'•,. ,,,, /,.C,-..'.,:

, , ..... , , , j . ••_-•-' ./2=_: _ z._ __ _ •__•.•_•_.?•__:..•____.•.• ,,,•'"' •"" .... " ! -''--' - 1 I I %• I t/ I I I I I I I 'l

1 2 3 •4 5 6 7 8 9 10 11 12

Monfh

Kearney

10 id, , , , , , .... . 9

6 ß REGCM / i r- I/ ..... %•------"'•x .•-- csmo • ,, ' ø ¸ 5 •];.. % X:x -•-•--._-,,'-•'•.//'%_:

<• 4 3

2

1

Monfh Monfh

Figure ?. Mean temperature (degrees Centigrade) contrasts (2 x CO2-1 x CO2) and change in standard deviation of daily temperature for four stations: (a) Hartington; (b) Auburn; (c) Atkin- son; (d) Kearney. The top panels show mean temperature change for SDS, RegCM2, and CSIRO; and the bottom panels show change in the standard deviation of daily temperature.


reach significance (at the 0.05 level). The RegCM2 and sometimes the CSIRO exhibit variance increases in the

warmer months (although the changes usually lack statistical significance) while SDS never produces increases in variance at that time. However, all three are similar in that decreases in variance dominate during the winter and early spring at most locations.

3.1.1.3. Spatial patterns of temperature change: Temperature changes are generally highest in winter and spring in the northeastern climate model grids (Figure 7a) but this spatial pattern breaks down in summer. Figure 8 displays the spatial pattern of temperature change in winter (December, January, and February) for the two climate models and the SDS re- suits. In the RegCM2 (and four CSIRO grids) there is a clear pattern of smallest temperature changes in the southwest and largest in the northeast. The range of change across the RegCM2 grids is 1.6øC. For the SDS results, larger changes are found in the northeast, but the other patterns seen in the RegCM2 and CSIRO are not evident. For example, on the western edge of the area, the north-south contrast is nonexistent in the SDS results. The spatial range of temperature for SDS is half that of the RegCM2, 0.8 ø. The mean for the whole field is 6.6 ø (spatial standard deviation of 0.24 ø) for SDS and 4.7 ø (spatial standard deviation of 0.48 ø) for the RegCM2. Results for the near stations/grids, Auburn and Pawnee, show that the temperature results are very similar both for the RegCM2 and the SDS. Such a comparison is obviously not possible with the GCM. Overall, we see much greater spatial variabil-

4.$ .•i ß 5.4 .... •! • 4..5 ;;

4.4 • 6.7 ß 5.2

Figure 8. Changes in mean temperature (degrees Centigrade) in winter for RegCM2, SDS, and CSIRO for all stations/grids. RegCM2 values are dark, below station location point; SDS values are gray, above the station location point; values for the CSIRO grids are larger dark numbers in corner of each CSIRO grid.

ity in the RegCM2 mean temperature changes than in the results of SDS.

3.1.2. Discussion of results. The large seasonal cycle of temperature change in the SDS technique re- suits mainly from the very small changes in summer temperature. Under 2 x CO2,700 mbar heights generated by CSIRO are frequently larger than the maximum of observed heights. In order to handle such a case and avoid extrapolation the CSIRO-generated heights were substituted by the maximum of observed heights. Observed temperature and precipitation values corresponding to this height were then used in the stochastic model. Therefore the changes in summer obtained by SDS are estimates at the lower limits of changes. The results for SDS in summer, which are consistent with previous applications of this statistical downscaling method [Matyasovszky et al., 1994] are therefore less robust than those for the two climate model results, in that they are artificially affected by the limitation of the 700 mb height to the maximum of the observed value. This implies that the large seasonal cycle of temperature change in the SDS is likely unrealistic.

This result illustrates well one of the potential dif- ficulties of statistical downscaling methods: if the cli-

ß

mate change simulation exceeds the range of the data used to develop the statistical model, then the validity of the model itself may be in question, unless suitable modifications to the statistical model are carried out.

The similarities in the responses of the three methods in change in daily variance of temperature likely reflect the large scale control over this characteristic of temperature, especially in winter. As explored by Mearns et al. [1995a, M98], Gregory and Mitchell [1995], and Cao et al. [1992], many climate models show decreases in daily variance of temperature, particularly in the winter months in midlatitudes. The regional and CSIRO models respond similarly in daily temperature variance throughout the domain (M98). Dynamical explanations for this are found in the above references.

The tendency to see localized increases in variability in the summer in both climate models, has also been reported in other nested model experiments over the continental United States [Mearns et al., 1995a]. In these earlier experiments, the increased variability was traced to increases in the local meridional temperature gradient as well as to possible local effects associated with surface processes. The absence of this response in SDS indicates that a circulation-type criterion based exclusively on 700 mbar height may not be adequate in capturing local effects.

The much smaller spatial variability for SDS indicates that the model does not spatially separate well the spatial aspects of climate change. In the physical models, surface forcing can significantly contribute to the spatial variability of climate change. An example is given by the snow-albedo feedback during the winter. As snow is depleted under surface warming, snow cover decreases. Therefore the surface albedo decreases and

in turn enhances the surface warming. This effect is ev-

6616 MEAR,NS ET AL' REGIONAL AND STATISTICAL DOWNSCALING COMPARISON

ident in both the CSIRO and the RegCM2 simulations. It is unlikely that the spatial details of this surface effect is communicated to the 700 mbar heights, so the SDS method most likely fails to capture it.

3.2. Precipitation For precipitation we compare three quantities: mean

daily precipitation, the probability of precipitation, and the median of daily intensity of precipitation. We analyze mean daily precipitation spatially for all stations for the four cardinal months, and we analyze probability of precipitation and daily intensity of precipitation for half of the individual stations. We perform statistical tests on all three quantities. The tests used are described by Mearns et aZ. [1995b] and are taken from Katz [1983]. The tests on probability of precipitation and the median of intensity were performed on the 5

years of RegCM2 and CSIRO climate experiments and 120-year time series for the SDS method. In this section, statistical significance refers to significance at the 0.05 level.

3.2.1. Mean daily precipitation. Figure 9 shows the spatial distribution of changes in mean daily precipitation for the four cardinal months for SDS and the two climate models. While we have not indicated

statistical significance of these changes on the maps, we note that in general, about half the changes for SDS are significantly (at the 0.05 level) different, whereas very few are for the RegCM2.

SDS almost always produces mean precipitation increases, at all locations, except for two stations in July. The RegCM2 produces coherent subregions of both increases and decreases in each of the four car-

dinal months. However, very few of the changes are

J Jan ua ry ! •:!:•'•;, i April

.•6 : -.01 :

2.54 .s$ .... .. 2.11 .31

, o.4• s

•.::.:•::

Ocfober • .

1.11 ß ?:':'L .•9 !.17 ..............

1.1t

.45 1.•7 ß 1.25 .

- .02 .•9 1.08 .6 4 •.44 ::: .21

1.96 .

.:

Figure 9. Change (2 x CO2-1 x CO2) in mean daily precipitation (mm/d) for the four months indicated, for the 12 locations/grids: RegCM2, SDS, and CSIRO. RegCM2 values are dark, below station location point; SDS values are gray, above the station location point; values for the CSIRO grids are larger dark numbers in corner of each CSIRO grid.


statistically significant. The CSIRO generally produces increases at all four grids, although significant increases only occur in April. In January, RegCM2 predicts small (insignificant) mean precipitation decreases except in the southeast corner of Nebraska. SDS produces increases throughout with (significant) larger magnitudes than RegCM2.

In April, RegCM2 predicts mean daily precipitation decreases in the central and northeast part of the domain (but none significantly so), while SDS calculates only relatively large increases. That same pattern is seen in July. In October, RegCM2 produces mean precipitation decreases in the southeast part of eastern Ne- braska, while SDS calculates only increases. Across all months and locations, in 40% of the cases, SDS and RegCM2 disagree regarding the direction of change of mean daily precipitation.

The CSIRO model, for the four grids indicated in Fig- ure 1, exhibits precipitation increases in April and Octo- ber, as well as increases in July, except for the SE grid. Largest (and most significant) changes occur in April. In January the northern two grids show (insignificant) increases and the southern two (insignificant) decreases.

Thus in April and October the RegCM2 produces mean precipitation decreases that are reflected neither in the driving GCM nor in the SDS results. This is not unusual. Nested model simulations often produce changes in precipitation which are opposite in sign from those produced by the driving GCM [Giorgi et al., 1994; 1998]. This tends to occur primarily in areas located downwind of major mountain systems, such as the Rocky Mountains, and is due to the different topo-

graphical fields in the GCM and nested model, which significantly influence aspects of climatology such as the precipitation shadowing effect or circulations around to- pography. We should point out that differences in the simulated sign of precipitation change between GCM and statistical downscaling simulations also often occur [e.g.,yon $torch et al., 1993, Wilby et al., 1998].

The results for SDS confirm earlier results obtained

for Nebraska using the Canadian Climate Centre (CCC) GCM and the Max Planck Institute (MPI) GCM [Matya- sovszky et al., 1994]. In those cases the actual magnitude of change was smaller, as was the geopotential height increase under 2 x CO2.

3.2.2. Probability of precipitation. Table 2 presents statistics on changes in probability of precipitation for five locations. For SDS, only increases in probability are found, but the changes are very small. RegCM2 produces changes in both directions, and vary- ing by season. Only decreases occur in January and October, while April and July exhibit mainly increases. RegCM2 changes, in general, are larger than those of SDS, but fewer differences are 'significant due to the small sample size of the RegCM2 runs. For the CSIRO there is general agreement with RegCM2 in direction of change of probability of precipitation with large significant increases occurring for all four CSIRO grids in April.

The results for the SDS method are again in general agreement with the earlier results for the CCC and MPI GCMs. In the work of Matyasovszky et al. [1994], most of the changes in probability of precipitation were also quite small, and most were insignificant. However,

Table 2. Probability of Precipitation

Atkinson Hartington Kearney Auburn Pawnee

January SDS

RegCM2

April SDS

RegCM2

July SDS

RegCM2

October SDS

RegCM2

2 x CO2 0.16' 0.15' 0.17' 0.19' 0.15 1 x CO2 0.12 0.11 0.14 0.16 0.13 2 x CO• 0.54 0.57 0.37* 0.50 0.46 1 x CO• 0.55 0.54 0.50 0.52 0.54

2 x CO• 0.25 0.25 0.28* 0.29 0.26 1 x CO• 0.24 0.24 0.22 0.30 0.25 2 x CO2 0.77* 0.73 0.67 0.67 0.68 1 x CO• 0.63 0.65 0.63 0.70 0.68

2 x CO• 0.29 0.29 0.26 0.28 0.31 1 x CO• 0.27 0.29 0.25 0.28 0.31 2 x CO• 0.75* 0.78* 0.74 0.72 0.70 1 x CO2 0.64 0.63 0.68 0.63 0.61

2 x CO• 0.24* 0.23* 0.22* 0.27* 0.20 1 x CO•. 0.17 0.17 0.18 0.19 0.18 2 x CO• 0.55 0.44 0.44 0.35* 0.32* 1 x CO• 0.55 0.43 0.54 0.52 0.51

SDS, semiempirical statistical downscaling results. RegCM2, regional climate model results. 2 x CO2 = 2 x CO• results. lx CO• = control or 1 x CO• results.

*Difference 2 x CO•. - 1 x CO2 is significantly different from 0.0 (at the 0.05 level).


with those GCMs, there were some small decreases in probability.

3.2.3. Changes in the median of intensity. In the colder months (January and October), SDS uni- formly produces mean intensity increases, as does RegCM2 except for two locations in January (Table 3). In April, SDS again produces intensity increases, but RegCM2 results are mixed. In July both methods produce mixtures of increases and decreases although there is little correlation between the two methods for individ-

ual stations. In general then, as with the probability of precipitation, the RegCM2 produces greater variability (across space and time) of changes in intensity compared to the SDS method. Both methods exhibit the greatest variability of change in July when local convective processes are dominant in precipitation forma- tion. Note that the RegCM and CSIRO results for both the precipitation intensity and the probability of occurrence should be interpreted very cautiously in view of the uncertainties found in the model simulation of these

quantities (see section 2.3 and Tables 2 and 3). In the case of the CCC GCM [Matyasovszky et al.,

1994], largely only increases in mean intensity were found using the SDS method, but both increases and decreases were found for downscaling from the MPI GCM. Since changes in probability with SDS are usually very small, changes in mean daily precipitation are mainly driven by changes in mean intensity.

4. Discussion and Conclusion

We have demonstrated that substantial differences

in the regional details of climate change are produced by two different means of downscaling from the same

large-scale GCM experiments. The meaning of these contrasts are not easy to discern. One broad-based conclusion is obvious; that is, some of these differences must pertain to the fact that only aspects of the perturbed climate communicated to the 700 mbar heights are taken into account by the SDS method, while all aspects of the climate system (i.e., those aspects success- fully modeled) at all vertical levels of the atmosphere, including the surface come into consideration for the regional model and the GCM. Indeed, one of the underlying assumptions of semiempirical downscaling is that the local predicted variable is primarily a function of synoptic forcing [Hewitson and Crane, 1996].

The aspects of the driving GCM utilized by each technique is different. Moisture, temperature, and wind fields throughout the vertical levels from the GCM serve as boundary conditions for the regional climate model, while SDS considers only the 700 mb heights. It might be interesting to take the 700 mbar heights from the RegCM2 and recalculate the statistically downscaled climate change to see if this bears greater resemblance to the regional climate model results than those downscaled from the GCM.

As noted earlier, finding large differences in response between physical models and semiempirical statistical models has been demonstrated in other contexts. [Wilby et al. [1998], compared climate change results downscaled from the Hadley Centre's coupled GCM simulations using various SDS methods and found that changes in daily precipitation produced by the different methods were generally smaller than those from the general circulation model; contrasting directions of change for precipitation were also found. Cubasch et al. [1996] compared high resolution time slice experi-

Table 3. Median of Intensity (mm/d)

At kinson Hart ington K earney A u b urn Pawnee

January SDS

RegCM2

April SDS

RegCM2

July SDS

RegCM2

October SDS

RegCM2

2 x CO2 5.5* 5.2* 4.2* 4.6 6.4* 1 x CO2 2.6 4.4 3.7 4.2 4.6 2 x CO2 0.5 0.4 0.5 0.4 0.4 1 x C02 0.6 0.6 0.5 0.3 0.5

2 x C02 5.3 6.4* 5.5 7.4* 8.6* 1 x C02 5.0 8.1 4.7 5.5 7.3 2 x C02 0.9 1.0 1.2 1.5' 1.7' 1 x C02 0.9 0.9 1.4 0.8 1.0

2 x CO2 12.8' 8.9* 18.5' 9.4 7.8* 1 x CO2 7.5 6.4 10.2 9.0 6.3 2 x CO• 2.3 2.4 3.5 2.9 2.4 1 x CO• 2.3 2.1 2.5 3.3 3.0

2 x CO• 10.9' 5.2 4.9 6.4 7.9 1 x CO• 7.9 5.2 5.1 6.0 7.6 2 x C02 1.0 1.6 1.3 1.4 2.6 1 x C02 0.8 1.4 1.3 1.5 1.2

*Ratio of the medians (2 x CO2/1 x CO2) is significantly different from 1.0 (at the 0.05 level).


ments derived from various coarse resolution GCM ex-

periments with an SDS approach and found that for the Iberian Peninsula, different directions of change in precipitation were calculated. The cause of these differences was not clear. In neither of these studies was

it possible to clearly discern why the various methods produced different results for climate change conditions.

It is not possible from these results to conclude that one method provides more "correct" responses to external forcing of climate compared to the other. We can say that for temperature, there are clear limitations for the SDS method when changes in 700 mbar heights (and probably summer temperatures) extend beyond the range in the observed data. While the regional climate model simulations provide a more complete response to possible factors affecting the distribution of local climate, the biases in the model control run, especially for the probability and intensity of precipitation, preclude attaching greater significance to these model results under doubled CO2 conditions.

Clearly more comparative investigation of these methods is called for. It would be desirable, for example, to produce longer-nested regional model simulations using longer runs from a GCM to increase the sample size for the climate model results and improve the estimation of parameters for the SDS method. Also, comparison over a larger region would probably facilitate finding more complete explanations of results.

Numerous SDS techniques and regional climate models are today available. We do not even as yet know if different regional models respond similarly to control and perturbed climate boundary conditions provided from the same GCM. This will be an important piece of research for determining how differently nested regional models respond to the same lateral forcing, which in turn will contribute to understanding how and why semiempirical downscaling methods differ from nested region modeling approaches in providing high- resolution climate change scenarios. Given the strong need expressed for higher-resolution climate change scenarios, the need to understand why different downscaling methods perform differently has become very important; especially since there is mounting evidence that higher resolution scenarios can lead to different estima- tions of the effect of climate change on resource systems such as agriculture [Mearns et al., this issue]. We en- courage more research programs that will lead to rig- orous intercomparisons of SDS methods and nested regional climate modeling.

Acknowledgments. Seventy percent ($50,000) of this research was funded by the U.S. Department of Energy's (DOE) National Institute for Global Environmental Change (NIGEC) through the NIGEC Great Plains Regional Center at the University of Nebraska-Lincoln. (DOE Cooperative Agreement No. DE-FC03-90ER61010.) Financial support does not constitute an endorsement by DOE of the views expressed in this article/report. We thank Larry McDaniel for providing computing and graphics support for this re-

search. We also thank the two anonymous reviewers for useful comments and suggestions. The National Center for Atmospheric Research is sponsored by the National Science Foundation.

References

Bardossy, A., and E. J. Plate, Modeling daily rainfall using a semi-Markov representation of circulation pattern occurrence, J. Hydrol., 122, 33-47, 1991.

Bardossy, and E. J. Plate, Space-time model for daily rainfall using atmospheric circulation patterns, Water Re- sour. Res., 28, 1247-1259, 1992.

Bartholy, J., Methodological study on clustering local and regional data series of precipitation, in Proceedings of the 5th International Meeting on Stochastic Climatology, J123-J124, Am. Meteorol. Soc., Boston, Mass., 1992.

Bogardi, I., I. Matyasovszky, A. Bardossy, and L. Duckstein, Application of a space-time stochastic model for daily precipitation using atmospheric circulation patterns, J. Geo- phys. Res., 98, 16,653-16,667, 1993.

Briegleb, B. P., Delta-Eddington approximation for solar radiation in the NCAR Community Climate Model, J. Geophys. Res., 97, 7603-7612, 1992.

Cao, H. X., J. F. B. Mitchell, and J. R. Lavery, Simulated diurnal range and variability of surface temperature in a global climate model for present and doubled CO•. climates, J. Clim., 5, 920-943, 1992.

Christensen, J. H., B. Machenhauer, R. G. Jones, C. Schar, P.M. Ruti, M. Castro, and G. Visconti, Validation of present day regional climate simulations over Europe: LAM simulations with observed boundary conditions, Clim. Dyn., I3, 489-506, 1997.

Christensen, O. B., J. H. Christensen, B. Machenauer and M. Bozet, Very-high resolution regional climate simulations over Scandinavia, Present climate, J. Clim., in press, 1998.

Cubasch, U., H. yon Storch, J. Waszkewitz, and E. Zorita, Estimates of climate change in Southern Europe derived from dynamical climate model output, Clim. Res., 7,(2), 129-149, 1996.

Dickinson, R. E., A. Henderson-Sellers, and P. J. Kennedy, Biosphere-Atmosphere Transfer Scheme (BATS) version IE as coupled to the NCAR Community Climate Model, NCAR Tech. Note, Natl. Cent. for Atmos. Res., Boulder, Colo., 1993.

Giorgi, F., On the simulation of regional climate using a limited area model nested in a general circulation model, J. Clim., 3, 941-963, 1990.

Giorgi, F., Perspectives for regional earth system modeling, Global Planet. Change, I0, 23-42, 1995.

Giorgi, F., and M. R. Marinucci, A study of the sensitivity of simulated precipitation to model resolution and its implications for climate studies. Mon. Weather Rev., 12,•, 148-166, 1996.

Giorgi, F., and L. O. Mearns, Approaches to regional climate change simulation: A review, Rev. Geophys., 29, 191-216, 1991.

Giorgi, F., R. Marinucci, and G. Visconti, 2XCO2 climate change scenario over Europe generated using a limited area model nested in a general circulation model, 2, Cli- mate change scenario, J. Geophys. Res., 97, 10,011-10,028, 1992.

Giorgi, F., M. R. Maxinucci, and G. T. Bates, Development of a second generation regional climate model (RegCM2), I, Boundary layer and radiative transfer processes, Mon. Weather Rev., 121, 2794-2813, 1993a.

Giorgi, F., M. R. Maxinucci, G. DeCanio, and G. T. Bates, Development of a second generation regional climate model (REGCM2), II, Cumulus cloud and assimilation of lat-


eral boundary conditions, Mon. Weather Rev., 121, 2814- 2832, 1993b.

Giorgi, F., C. Shields Brodeur, and G. T. Bates, Regional climate change scenarios over the United States produced with a nested regional climate model, J. Clim., 7, 375- 399, 1994.

Giorgi, F., L. Mearns, C. Shields, and L. McDaniel, Regional nested model simulations of present day and 2 x CO2 climate over the Central Great Plains of the United States, Clim. Change, J0, 457-493, 1998.

Gong, X., and M. B. Richman, An examination on methodological issues in clustering North American precipitation, in Proceedings of the 5th International Meeting on Stochastic Climatology, J103-J108, Am. Meteorol. Soc., Boston, Mass., 1992.

Gregory, J. M., and J. F. B. Mitchell, Simulation of daily variability of surface temperature and precipitation over Europe in the current and 2 x CO2 climates using the UKMO climate model, Q. J. R. Meteorok Soc., 121, 1451- 1476, 1995.

Grell, G. A., J. Dudhia, and D. R. Stauffer, A description of the fifth generation Penn State•NCAR Mesoscale Model (MM5), NCAR Tech. Note, NCAR/TN-398-•-STR, Natl. Cent. for Atmos. Res., Boulder, Colo., 1994.

Grotch, S. L., and M. C. MacCracken, The use of general circulation models to predict regional climate change, J. Clim., J, 286-303, 1991.

Hewitson, B.C., and R. G. Crane, Climate downscaling: Techniques and application, Clim. Res., 7(2), 85-95, 1996.

Hirakuchi, H., and F. Giorgi, Multiyear present-day and 2 x CO• simulations of monsoon-dominated climate over eastern Asia and Japan with a regional climate model nested in a general circulation model, J. Geophys. Res., 100, 21,1•05-21,125, 1995.

Holtslag, A. A., E. I. F. deBruijn, and H. L. Pan, A high resolution air transformation model for short-range weather forecasting, Mon. Weather Rev., 118, 1561-1575, 1990.

Hughes, J.P., D. P. Lettenmaier, and P. Guttorp, A stochastic approach for assessing the effects of changes in synoptic circulation patterns on gauge precipitation, Water Resour. Res., 29, 3303-3315, 1993.

Jones, R. G., J. M. Murphy, and M. Noguer, Simulation of climate change over Europe using a nested regional climate model, I, Assessment of control climate, including sensitivity to location of lateral boundary conditions, Q. J. R. Meteorok Soc., 121, 1413-1449, 1995.

Jones, R. G., J. M. Murphy, M. Noguer, and M. Keen, Sim- ulation of climate change over Europe using a nested regional climate model, I: comparison of driving and regional model responses to a doubling of carbon dioxide, Q. J. R. Meteorok Soc., 123, 265-292, 1997.

Karl, T. R., W. C. Wang, M. E. Schlesinger, R. W. Knight, and D. Portman, A method of relating general circulation model simulated climate to the observed local climate, I, Seasonal statistics, J. Clim., 3, 1053-1079, 1990.

Kattenberg, A., et al., Climate models projections of fu- ture climate, in Climate Change 1995: The Science of Climate Change, Contribution of Working Group I to the Second Assessment Report of the Intergovernmental Panel on Climate Change, chap. 6, pp. 285-357, edited by J.T. Houghton, L.G. Meira Filho, B. A. Callander, N. Harris, A. Kattenberg, and K. Maskell, Cambridge Univ. Press, New York, 1996.

Katz, R. W., Statistical procedures for making inferences about precipitation changes simulated by an atmospheric general circulation model, J. Atmos. Sci., JO, 2193-2201, 1983.

Katz, R. W., Statistical procedures for making inferences about climate variability, J. Clim., 1, 1057-1064, 1988.

Katz, R. W., and B. G. Brown, Extreme events in a chang- ing climate: Variability is more important than averages, Clim. Change, œ1, 289-302, 1992.

Katz, R. W., and M. B. Parlange, Effects of an index of atmospheric circulation on stochastic properties of precipitation, Water Resour. Res., 29, 2335-2344, 1993.

Kim, J. W., J. T. Chang, N. L. Baker, D. S. Wilks, and W. L. Gates, The statistical problem of climate inversion: Determination of the relationship between local and large- scale climate, Mon. Weather Rev., 112, 2069-2077, 1984.

Laprise, R., et al., Climate and climate change in western Canada as simulated by the Canadian Regional Climate Model, Atmosphere-Ocean, 36(2), 119-167, 1998.

Lettenmaier, D., Stochastic modeling of precipitation with applications to climate model downscaling, in Analysis of Climate Variability: Applications of Statistical Tech- niques, edited by H. yon Storch, and A. Navara, chap. 11, Springer-Verlag, New York, 1995.

Lynch, A. H., W. L. Chapman, J. E. Walsh, and G. Weller, Development of a regional climate model of the Western Arctic, J. Clim., $, 1555-1570, 1995.

Machenhauer, B., M. Windelband, M. Botzet, J. H. Chris- tensen, M. Deque, R. G. Jones, P.M. Ruti, and G. Vis- conti, Validation and analysis of regional present-day climate and climate change simulations over Europe, sub- mitted to Q. J. R. Meteorok Soc., in press, 1998.

Marinucci, M. R., and F. Giorgi, A 2x CO• climate change scenario over Europe generated using a limited area model nested in a general circulation model, I, Present-day climate simulation, J. Geophys. Res., 97, 9989-10009, 1992.

Matyasovszky, I., I. Bogardi, A. Bardossy, and L. Duckstein, Estimation of local precipitation statistics reflecting climate change, Water Resour. Res., 29, 3955-3968, 1993.

Matyasovszky, I., I. Bogardi, and L. Duckstein, Comparison of two GCMs to downscale local temperature, precipitation and under climate change, Water Resour. Res., 30, 3437-3448, 1994.

McGregor, J. J., Regional climate modelling, Meteorok At- mos. Phys., 63, 105-117, 1997.

McGregor, J. J, and K. J. Walsh, Nested simulations of per- petual January climate over the Australian region, J. Geo- phys. Res., 98, 23,283-23,290, 1993.

Mearns, L. O., S. H. Schneider, S. L. Thompson, and L. R. McDaniel, Analysis of climate variability in general circulation models: Comparison with observations and changes in variability in 2 x CO2 experiments, J. Geophys. Res., 95, 20,469-20,490, 1990.

Mearns L. O., F. Giorgi, L. McDaniel, and C. Shields Brodeur, Analysis of daily variability and diurnal range of temperature in a nested regional climate model: Compari- son'with observations and doubled CO2 results, Clim. Dyn., 11, 193-209, 1995a.

Mearns, L. O., F. Giorgi, C. Shields Brodeur, and L. Mc- Daniel, Analysis of the variability of daily precipitation in a nested modeling experiment: comparison with observations and 2 x CO2 results, Global Planet. Change, 10, 55-78, 1995b.

Mearns, L. O., T. Mavromatis, E. Tsvetsinskaya, C. Hays, and W. Easterling, Comparative responses of EPIC and CERES crop models to high and low spatial resolution climate change scenarios, J. Geophys. Res., this issue.

Osborn, T. J., Areal and point precipitation intensity changes: Implications for the application of climate models, Geophys. Res. Left., 2•, 2829-2832, 1997.

Osborn, T. J. and M. Hulme, Development of a relationship between stations and grid-box rain day frequencies for climate model evaluation, J. Clim., 10, 1885-1908, 1997.

Parkinson, C. L., and W. M. Washington, A large-scale nu-


merical model of sea ice. J. Geophys. Res., 8•, 311-377, 1979.

Pielke, R. A., et al., A comprehensive meteorological modeling system- RAMS, Meteor. Atmos. Phys., J9, 69-91, 1992.

Risbey, J., and P. Stone, A case study of the adequacy of GCM simulations for input to regional climate change, J. Clim., 9, 1441-1446, 1996.

Robinson, D. A., D. J. Leathers, M. A. Palecki and K. F. Dewey, Some observations on climate variability as seen in daily temperature structure, Atmos. Res., 37, 119-131, 1995.

Smith, J., and D. Tirpak (Eds.), Potential Effects of Global Climate Change on the United States, Environ. Prot. Agency, Washington, D.C., 1989.

Toth, Z., and T. Szentimrey, The binormal distribution: A distribution for representing asymmetrical but normal- like weather elements, J. Clim., 3, 128-136, 1990.

von Storch, H., E. Zorita, and U. Cubasch, Downscaling of global climate change estimates to regional scales: An application to Iberian rainfall in wintertime, J. Clim., 6, 1161-1171, 1993.

Watterson, I. G., M. R. Dix, H. B. Gordon, and J. L. Mc- Gregor, The CSIRO nine-level atmospheric general circulation model and its equilibrium present and doubled CO2 climate, Aust. Meteorok Mag., ZZ, 111-125, 1995.

Wigley, T. M. L., P. D. Jones, K. R. Briffa, and G. Smith, Obtaining sub-grid-scale information from coarse resolution general circulation model output, J. Geophys. Res., 95, 1943-1953, 1990.

Wilby, R. L., and T. M. L. Wigley, Downscaling general circulation model output: A review of methods and limitations, Prog. in Phys. Geogr., œ1(4), 530-548, 1997.

Wilby, R. L., T. M. L. Wigley, D. Conway, P. D. Jones, B. C. Hewitson, J. Main, and D. S. Wilks, Statistical downscaling of general circulation model output: A comparison of methods, Water Resour. Res., 3,i, 2995-3008, 1998.

Wilks, D. S., Conditioning stochastic daily precipitation models on total monthly precipitation. Water Resour. Res., 25, 1429-1439, 1989.

Wilson, L. L., D. P. Lettenmaier, and E. F. Wood, Sim- ulation of daily precipitation in the Pacific Northwest using a weather classification scheme, in Land Surface- Atmosphere Interactions .for Climate Modeling: Obser- vations, Models, and Analysis, edited by E. F. Wood, Kluwer Acad., Norwell, Mass., 1991.

Wilson, L. L., D. P. Lettenmaier, and E. Skyllingstad, A hierarchical stochastic model of large-scale atmospheric circulation patterns and multiple station daily precipitation, J. Geophys. Res., 97, 2791-2809, 1992.

I. Bogardi and M. Palecki, Department of Civil Engineer- ing, University of Nebraska, Lincoln, NE 68588-0531. (e- mail [email protected])

F. Giorgi and L. O. Mearns, National Center for Atmo- spheric Research, PO Box 3000, Boulder, CO 80307. (e-mail [email protected]; [email protected])

I. Matyasovszky, Department of Civil Engineering, Eotvos Lorand University, Budapest, Hungary. (e-mail matya@lu dens. elte.hu)

(Received February 19, 1998; revised August 5, 1998; accepted September 22, 1998.)

comparison of climate change scenarios generated from

Documents