1. Introduction
Changes in external conditions, such as the ongoing increase in CO2 and other so-called greenhouse gases, may affect not only the mean state but also the variability of climate. Both types of changes are relevant for the practical impacts of climate change; for example, changes in extremes are effected by both of them (Katz and Brown 1992). However, model-simulated changes in climate variability have been studied less comprehensively than the changes in mean climate.
The response of general circulation model simulated time-mean climate to increasing greenhouse gases has been studied extensively. In particular, the changes in mean temperature and precipitation tend to be addressed in almost every paper documenting a new climate change simulation. The common features and differences between different model experiments have been summarized by the Intergovernmental Panel on Climate Change (IPCC; Mitchell et al. 1990; Kattenberg et al. 1996; Cubasch et al. 2001). There are also several intercomparisons elsewhere in the meteorological literature, including Grotch and MacCracken (1991), Kittel et al. (1998), Giorgi and Francisco (2000), and Räisänen (2001; hereafter R2001).
Simulated changes in daily and interannual climate variability have also attracted attention at least since the late 1980s (e.g., Rind et al. 1989). A comparatively robust finding from such studies is generally larger precipitation variability in a warmer world, at least on the daily timescale that has been studied most frequently. Part but apparently not all of this is explained by increased mean precipitation. For example, Zwiers and Kharin (1998) reported, for an equilibrium 2 × CO2 experiment, a 4% increase in the global time-mean precipitation, a 7% increase in the average standard deviation of local daily precipitation, and an 11% increase in the average 20-yr extreme daily precipitation. In addition, several investigators (e.g., Rind et al. 1989; Cao et al. 1992; Gregory and Mitchell 1995; Zwiers and Kharin 1998; Boer et al. 2000; Kharin and Zwiers 2000) have found reduced daily and/or interannual temperature variability in high- and midlatitude areas in winter, at least partly as a result of reduced snow and sea ice. On the other hand, increased temperature variability has been reported in some studies over midlatitude continents in summer (Gregory and Mitchell 1995; Zwiers and Kharin 1998; Kharin and Zwiers 2000) or in the Tropics (e.g., Meehl et al. 1994). However, although some commonly occurring features have emerged, the agreement between the various studies is not complete. This may reflect differences in the models (and to some extent, measures of variability) used, but also the relatively low signal-to-noise ratio of the simulated changes.
What has been essentially missing thus far is systematic intercomparison between the changes in variability in different models. The present paper describes such an intercomparison of CO2-induced changes in the interannual variability of monthly temperature and precipitation. Nineteen atmosphere–ocean general circulation models (AOGCMs) participating in the second phase of the Coupled Model Intercomparison Project (CMIP2; Meehl et al. 2000), all forced by the same gradual increase in CO2, are included in this comparison.
The present study is motivated by, and aims to shed light on, a number of questions. How realistic is the magnitude of interannual temperature and precipitation variability in the control simulations? What are the typical features of CO2-induced change in interannual variability in the CMIP2 experiments, and how well do the experiments agree with each other? How large and how different are the simulated changes between different experiments, in comparison with the noise associated with internal variability in the statistics? How does the magnitude of the variability changes compare with the changes in the time-mean climate, and what are the implications for the changes in extremes? Some issues related to the physical interpretation of the variability changes are also discussed, although this subject would actually deserve a more comprehensive treatment than is possible here.
The paper starts with a description of the datasets (section 2) and the analysis methods used (section 3). The interannual variability in the control simulations is compared with observational estimates in section 4. The simulated changes in variability are documented in section 5, including an assessment of the agreement between the different experiments and the importance of noise for the interexperiment differences. In section 6, the magnitude of these changes is compared with the changes in time-mean temperature and precipitation to give an idea of how important the two types of changes are for changes in extremes. Section 7 discusses a few issues related to the interpretation of the variability changes. Finally, a summary is given in section 8.
2. Datasets
Each CMIP2 experiment consists of a control run with constant (“present day”) atmospheric CO2 and of a greenhouse run with a standard gradual (1% yr−1 compound) increase in CO2. Following the classification of Gates (1992), CMIP2 is thus a “level 2” intercomparison. In the present study, 19 models are included (see Table 1 where models from the participating institutions are identified). For all of these, 80 yrs of data are available at monthly time resolution for both the control and the greenhouse runs. The simulated global mean warming with a doubling of CO2 varies from 1.1°C (MRI2) to 3.1°C (CCSR2) with a mean of 1.75°C, and the change in global precipitation from −0.2% (ECHAM4) to 5.6% (CCSR2) with a mean of 2.5% (Table 2). To avoid problems with control run climate drift (see R2001), the climate changes with a doubling of CO2 are calculated here by comparing the last 20-yr period (model years 61–80) in the greenhouse runs with the same period in the control runs.
To compare the model-simulated variability with the variability in the real world, three datasets are used. One of these, the Climatic Research Unit (CRU) analysis of temperature and precipitation (New et al. 2000), is only available over the continents excluding Antarctica. The other two, the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) reanalysis (Kalnay et al. 1996; Kistler et al. 2001) and the Climate Prediction Center Merged Analysis of Precipitation (CMAP; Xie and Arkin 1997) are global. The NCEP–NCAR reanalysis is here only used for studying temperature variability.
The CRU dataset has a high resolution (0.5° lat × 0.5° lon) and is based on an interpolation of station-observed monthly temperature and precipitation anomalies (New et al. 1999, 2000). Disregarding the effects of measurement errors such as undercatch of precipitation, it is expected to provide a good approximation of local climate variability in areas with high station density. However, the interpolation, in which the grid-cell temperature and precipitation anomalies were computed as weighted averages of the anomalies at the eight closest stations, may suppress variability in data-sparse areas where the anomalies at neighboring stations are poorly correlated. In areas with very poor data coverage (e.g., Greenland), this problem is exacerbated by the use of zero-anomaly pseudostations in the interpolation procedure (New et al. 2000).
The NCEP–NCAR reanalysis dataset has a resolution of 2.5° lat × 2.5° lon. Over ice-free sea areas, the NCEP–NCAR surface air temperatures are primarily controlled by the sea surface temperature analyses used as the lower boundary condition. The temperatures over land and sea ice are more sensitive to errors in the assimilation model, although to some extent they are constrained by observations where these are available. In most land areas, however, the temperature variability in the NCEP–NCAR dataset is in good agreement with CRU.
The CMAP precipitation analysis, which also has a resolution of 2.5° lat × 2.5° lon, was formed by combining rain gauge measurements with several types of satellite data and the NCEP–NCAR reanalysis. The different data sources were weighted according to their expected, geographically varying relative accuracy. In practice, gauge measurements have the highest weight where available. Uncertainties in this dataset are probably largest over polar regions, where there are few gauge measurements and most satellite-based techniques function poorly.
3. Measures of variability
After being calculated for each individual calendar month, SDT, SDP, and CVP were averaged arithmetically to obtain seasonal or annual mean values. All these statistics were first evaluated in each model's own coordinates, that is, at a resolution of about 300–600 km depending on the model. For comparison between different models, the measures of variability were then interpolated bilinearly to a common 2.5° lat × 3.75° lon grid. The estimates of observed temperature and precipitation variability were derived in the same manner. For the CRU dataset that has a much higher resolution than the models, an alternative variability estimate was also derived by averaging the analyzed temperature and precipitation over the 2.5° × 3.75° grid boxes before calculating their temporal variability. Compared with the values obtained directly in the 0.5° lat × 0.5° lon grid, this led on the average to a 15% decrease in SDP and CVP, but only to a 3% decrease in SDT. In the following, only the higher (0.5° lat × 0.5° lon grid) CRU estimate is used, with the motivation that this comes closer to the truly local climate variability, which is relevant in climate impact studies.
4. Variability in the control simulations
This section discusses the variability in the CMIP2 control runs and compares it with observational estimates of variability. The model-simulated variability is calculated, for consistency with the treatment of climate changes, for the last 20-yr period (years 61–80) in the control runs. The CRU temperature and precipitation variability is computed separately for the periods 1931–50, 1951–70, and 1971–90 and is averaged over these. The NCEP–NCAR temperature variability is averaged over the two 20-yr periods 1959–78 and 1979–98, whereas the CMAP precipitation variability is evaluated for the 1979–98 period. The exact choice of periods for the models and the various datasets was found inconsequential for the general conclusions.
Figure 1 displays the annual (mean of the variabilities in the 12 calendar months), global land, and sea area means of SDT, SDP, and CVP in the 19 models. The intermodel differences are substantial. For example, SDT over land varies by a factor of 1.5 from 1.33°C in CSIRO to 1.99°C in GFDL–R15, and SDT over sea by a factor of 2.2 from 0.49°C in CCCma to 1.09°C in GFDL–R15. Similar relative ranges are found for SDP and CVP. All the models simulate larger temperature variability over land than over the oceans, and the same is true for the relative (CVP) precipitation variability. However, the absolute precipitation variability (SDP) is, with few exceptions, larger over the oceans where the mean precipitation is also larger.
Each panel of Fig. 1 also includes observational estimates of variability, one for sea and two for land areas. The land estimates labeled CRU are actually complemented by the NCEP–NCAR or the CMAP data over the Antarctica (where no CRU data are available) that covers 9% of the global land area. The CRU and NCEP–NCAR land mean SDTs are very close to each other, and a fair agreement (mostly within 10%) also extends to the geographical details of SDT. Notable exceptions include parts of the equatorial Africa and South America (where NCEP–NCAR shows somewhat less variability than CRU) and Greenland (where the CRU data indicate spuriously low SDT and NCEP–NCAR is probably closer to the truth). The land mean SDP and CVP differ by about 10% between CRU and CMAP, the former being higher and the latter lower in CRU. The contrast in the CMAP minus CRU difference between these two quantities is partly due to the fact that the mean precipitation is generally larger (on the average 6%) in the CRU dataset. In addition, the CMAP data suggest much larger CVP than CRU over the Sahara. Due to the low mean precipitation, this area makes a negligible contribution to the global mean SDP, although its contribution to the global mean CVP is substantial.
The global 19-model means of variability are generally close to the observational estimates, most notably over the ocean. Considering the scatter between the different models and uncertainty in the observations, this seems somewhat fortuitous. However, in accord with the findings of Bell et al. (2000), most models overestimate temperature variability over land.
The close agreement in ocean mean SDT between the 19-model mean and NCEP–NCAR is in apparent conflict with Bell et al. (2000). They studied temperature variability in 16 AOGCMs and found 13 of them to simulate too little variability over the ocean. In addition to some differences in the set of models used, a different observational dataset, and slightly different analysis methods, Bell et al. focused on a different timescale. They analyzed the interannual variability of the annual mean temperature, rather than that of monthly temperatures. In fact, the CMIP2 19-model ocean mean SDT for the annual mean temperature is 10% below the NCEP–NCAR estimate (0.37° vs 0.41°C). This suggests that the models tend to underestimate the month-to-month persistence of temperature anomalies over the ocean.
The geographical distribution of the annually averaged variability is shown in Fig. 2. The first two columns depict the 19-model mean and an observational estimate. The latter is formed, over the continents excluding Antarctica, by averaging the CRU variability estimate with either NCEP–NCAR (SDT) or CMAP (SDP and CVP). This is not necessarily optimal since the expected accuracy of the two estimates may differ, but it yields results that are not generally far from either of them. In other areas, only the NCEP–NCAR or CMAP data are used. The last column highlights areas where most (at least 80%, i.e., 16 of 19) or all models are either below or above the observations.
For all of SDT, SDP, and CVP, the geographical patterns of the simulated variability agree generally well with the observational data. SDT yields, in the models, high values of up to 3°C over the Northern Hemisphere high-latitude continents and over the high-latitude Southern Ocean (Fig. 2a). The simulated variability decreases toward the equator and is relatively weak even over the ice-free midlatitude oceans. All these features conform well to the observational estimate in Fig. 1b, even though the latter suggests a stronger maximum near the Antarctic coastline. In addition, most or locally all models overestimate temperature variability over low-latitude continents (see Fig. 1c). During the summer, this tendency extends somewhat farther north to the midlatitudes (not shown). Over the oceans, there are both areas where the models tend to overestimate variability and areas where they tend to underestimate it. The latter include, in addition to the high-latitude Southern Ocean, the El Niño area in the eastern equatorial Pacific.
In both the models and in the light of the observations, SDP and CVP have very different geographical patterns (Figs. 2d,e,g,h). The former is largest in the Tropics, where the time-mean precipitation is high, the latter in arid subtropical regions with only sporadic precipitation. The observational data (CMAP) also show a pronounced maximum in CVP over the eastern equatorial Pacific, but this maximum is weaker and shifted to the west in the models that mostly underestimate the ENSO activity. In higher midlatitudes and polar regions, where precipitation falls frequently but in mostly relatively modest amounts, both SDP and CVP yield generally small values. The lowest simulated and observed CVP occurs over the Southern Ocean, apparently reflecting the continuous cyclone activity in this region. As discussed in the appendix, CVP is generally expected to increase with decreasing frequency of precipitation days.
Differences between the simulated and the observed precipitation variability do occur, but in several cases their interpretation is complicated by observational uncertainties. In most mid- and high-latitude areas, most models simulate larger SDP than is indicated by observations (Fig. 2f). This is not surprising, since the simulated mean precipitation also generally exceeds the observational estimate in these areas due to some combination of model errors and observational biases. In CVP, the difference between the models and the observational data is less systematic (Fig. 2i). The minima in CVP over the Southern Ocean and over the midlatitude North Atlantic and North Pacific are weaker in the models than according to the CMAP data, but the lack of gauge measurements in these areas suggests that this result should be taken with some caution. In addition to the eastern equatorial Pacific, the simulated SDP and CVP are generally lower than the CMAP estimate over Antarctica. Again, the cause of the difference is open to question. Despite smaller time-mean precipitation, the CMAP data indicate larger SDP over much of the Antarctic continent than over the Southern Ocean (Fig. 2e), and CVP jumps abruptly from low values over the Southern Ocean to surprisingly high values over Antarctica (Fig. 2h).
The amplitude of interannual climate variability varies with season. As an illustration of this, Fig. 3 compares the zonal mean seasonal cycles of SDT between the CMIP2 models and the observational data. The models qualitatively capture the observed difference between large extratropical temperature variability in winter and smaller variability in summer. Besides this, one may note that the relative overestimate in the simulated SDT is, in subtropical latitudes in both hemispheres, largest in the local summer. Although Fig. 3 also includes sea grid boxes, this primarily reflects model behavior over land.
Delineating why variability differs between different models, and between models and reality, is not a simple matter. Below, we only revisit two issues that have emerged in earlier investigations of model-simulated temperature variability: the possible impact of land surface hydrology, and apparent connections between the mean temperature and the variability.
Bell et al. (2000) suggested that the tendency of the models to overestimate temperature variability over land is related to the treatment of land surface hydrology. They found land-area temperature variability to be generally largest in models with simple bucket-type land surface schemes and smallest in models with more advanced physically based schemes (their Fig. 4), and associated the difference with more severe soil drying in the bucket-type schemes. Among the 19 CMIP2 simulations, the relationship between land-area temperature variability and the type of the land surface scheme is less clear. For example, of the four models that use a traditional single-layer bucket scheme, GFDL–R15 and GFDL–R30 do simulate very strong variability, but the others (BMRC and MRI1) are in the middle of the distribution. Nevertheless, the seasonal and geographical distributions of the 19-model mean SDT bias suggest that land surface processes may be involved. The relative overestimate in variability is largest where and when atmospheric circulation is relatively quiescent and high temperatures increase the chance of drying out of the soil, making temperature variability potentially sensitive to land surface processes: in low latitudes in general, and in northern midlatitudes in summer.
Duffy et al. (2000) found that models with high (low) global mean temperature tend to simulate weak (strong) temperature variability in polar regions (50°–90°N/S). We examined this relationship further by calculating, in each grid box, the correlation between the local mean temperature T and local SDT among the 19 CMIP2 models. The result is shown in Fig. 4 separately for December–January–February (DJF) and June–July–August (JJA). Correlation coefficients with an absolute value of 0.4 or larger, corresponding to approximate 90% significance, are shaded. The figure confirms a predominantly negative but geographically and seasonally varying correlation between T and SDT. A negative correlation from −0.4 to −0.8 is found, in both seasons, in most of the winter hemisphere. This likely reflects, in high-latitude sea areas and over the extratropical continents, enhanced variability over ice- and snow-covered surfaces, but the extension of the substantially negative values to much of the low-latitude oceans seems more difficult to explain. On the other hand, the correlation over continents at low latitudes and in the extratropical Northern Hemisphere in JJA is generally positive, albeit mostly relatively weak. This may be related to soil moisture. In the models in which evaporation becomes restricted by drying out of soil, this tends to increase temperature variability (Delworth and Manabe 1988, 1989; Tett et al. 1997) as well as the mean temperature.
5. Changes in variability at doubling of CO2
In this section, the simulated CO2-induced changes in SDT, SDP, and CVP are studied. The changes are evaluated for the 20-yr period (years 61–80) centered at the doubling of CO2 by subtracting from the greenhouse run values of SDT, SDP, and CVP during this period the corresponding quantity during the same period in the control run. To put the changes in variability in a perspective, the 19-model mean changes in annual mean temperature T and precipitation P are shown in Fig. 5. The warming is at a maximum over the Arctic Ocean but is otherwise generally stronger over land than sea areas. The 19-model mean precipitation increases in most of the world, but there are also wide areas of decrease in the subtropics and lower midlatitudes, including the southern United States and the Mediterranean region. The 19-model global mean temperature and precipitation changes are 1.75°C and 2.5%, respectively (see Table 2 for the model-to-model variation).
a. Geographical and seasonal distributions
The simulated changes in variability are illustrated in Fig. 6 with maps giving the changes in annual mean SDT, SDP, and CVP, and in Fig. 7 with Hovmöller diagrams showing the three-month zonal means of these changes. We first discuss the 19-model mean changes that are given both in absolute units (the left columns of Figs. 6 and 7) and in percent of the corresponding control run 19-model means (the middle columns of Figs. 6 and 7).
The 19-model annual mean SDT shows a pronounced decrease (locally down to −0.5°C or −20%) in high northern latitudes, particularly over the Arctic Ocean (Figs. 6a,b). The average change in almost all other areas is within ±0.1°C, excluding the somewhat larger decrease over the high-latitude Southern Ocean. In low latitudes in both hemispheres, the change is relatively small in absolute terms but mostly positive over the continents, in particular Africa, South America, Australia, and southern Asia. The decrease in SDT in high northern latitudes is largest in the northern autumn and winter and smallest in summer (Figs. 7a,b), following the seasonal cycle of the time-mean warming (not shown). Likewise, the decrease in SDT at 60°–75°N has a broad maximum in the southern winter. Slight increases in zonally averaged temperature variability dominate in low latitudes most of the year, and extend in the local summer to 70°N in the Northern and to almost 60°S in the Southern Hemisphere. Where and when zonal-mean SDT increases mainly reflect increasing variability over land; the changes over low- and midlatitude oceans are small and of varying sign.
The 19-model annual mean standard deviation of precipitation SDP increases in a large majority of the world (Figs. 6d,e). Most, but not all, of the subtropical and lower midlatitude areas where the time-mean precipitation decreases (Fig. 5b) also show a decrease in SDP. The absolute average increase peaks in the Tropics, especially the tropical Pacific Ocean, where SDP is largest even in the control simulations. However, the relative change is generally largest in high northern latitudes where the percent increase in the mean precipitation is a maximum. The seasonal variation of the SDP change is relatively modest in most latitude bands (Figs. 7d,e).
The 19-model annual mean CVP also increases in most areas (Figs. 6g,h), but the patterns of change differ substantially from the SDP change. The largest increases in CVP are seen in the subtropics and in lower midlatitudes, and they tend to coincide with areas of decreasing time-mean precipitation (cf., Fig. 5b). Slight decreases in annually averaged CVP occur sporadically in the Tropics and subtropics but happen more commonly in high latitudes, particularly in the northern polar region. As shown by Figs. 7g,h, this annual mean CVP decrease in high northern latitudes primarily results from a decrease in autumn and winter (when the relative increase in time-mean precipitation is largest, not shown); the change in summer is slightly positive. Overall, the changes in CVP are smaller in percent terms than those in SDP (cf., Figs. 6e and 7e with Figs. 6h and 7h). This suggests that increasing CO2 will generally have a smaller effect on relative than on absolute precipitation variability.
The 19-model means discussed above are averages of substantially varying changes in the individual CMIP2 experiments. As a measure of the relative agreement, the ratio between the mean changes and the (n − 1 form) interexperiment standard deviation is shown in the right columns of Figs. 6 and 7. Areas where the ratio exceeds 0.5 in absolute value are shaded, the motivation being that such a ratio approximates a lower limit of statistically significant agreement. Treating the 19 experiments as independent from each other, this ratio is equivalent to a t value of 0.5 × 191/2 = 2.18, which indicates a two-sided significance of slightly over 95%. For SDT, SDP, and CVP, respectively, areas where the local 19-model annual mean change is significant, in this sense, cover 17%, 46%, and 25% of the world. The same numbers for the seasonal zonal mean changes are higher (25%, 68%, and 49%) when averaged over the year and different latitude zones are weighted according to their area. In each case, the changes (generally increases) in SDP are more commonly significant with respect to the interexperiment variation than are the changes in SDT and CVP.
The increase in SDP is most robust in latitudes north of 45°N, where the mean-to-standard-deviation ratio for the seasonal zonal mean changes is generally between one and two. The ratio for the local, annual mean SDP changes also exceeds one in large parts of northern Eurasia, northernmost North America and the Arctic Ocean, and parts of the Southern Ocean. For SDT, the change with highest relative agreement is the decrease in variability in high northern latitudes during the winter half-year, and for CVP, the general increase in lower midlatitudes. Complete qualitative agreement on local annual variability changes, in the sense that the change would have the same sign in all 19 experiments, is rare. At most, such agreement occurs in about 1% of the world in the case of the SDP change (not shown). However, as indicated by the oval symbols in the third column of Fig. 7, the zonal mean SDP increase in mid-to-high northern latitudes fulfills this condition in most of the year. The same is sporadically true even for some of the other zonally averaged changes.
The differences between the individual models remain substantial even when the changes are averaged globally (Table 2). Globally and annually averaged SDT decreases in 16 of the 19 models, at most by 11% in CCSR2, and increases by 1%–2% in the remaining three. However, as there is in most models a cancellation between high-latitude decreases and low-latitude increases, the global mean SDT change is a slightly misleading quantity. The global annual means of SDP and CVP both increase in all but the CCSR1 experiment.
b. Analysis of variance
The differences in climate change between the CMIP2 experiments result from two factors: differences between the models themselves, and the noise associated with internal variability. To estimate the relative impact of these factors on the differences in time-mean climate change and to quantify the relative interexperiment agreement, a form of analysis of variance was used in R2001. Here, the same method is applied to the changes in interannual climate variability.
The results for local, annually averaged variability changes are shown in a zonally stratified form in Fig. 8. To illustrate the relative magnitudes of M2, D2, and N2, their zonal means are all normalized by the zonal mean of A2. The geographical variation in A2, which gives the 19-model mean squared amplitude of the changes, is huge. In the case of the SDT change, for example, its zonal means range from 0.004°C2 in the Southern Hemisphere midlatitudes to 0.27°C2 at the North Pole (not shown).
The common signal M2 is only a minor part of the squared amplitude of the SDT (excluding the high northern latitudes) and CVP changes (Figs. 8a,c). The agreement on the SDP changes is somewhat better, with the ratio M2/A2 exceeding 50% in most of the area north of 60°N (Fig. 8b). Most of the differences in the changes in interannual variability are attributed to the noise in the simulations, rather than directly to model differences—the top shaded area is in most latitudes larger than the middle unshaded area. This is especially true for the CVP changes, whose total squared amplitude consists mostly of noise in almost all latitude zones. The model-related differences only appear to dominate over the noise-related differences in the case of the STD change in high northern latitudes and near the Southern Ocean sea ice edge around 65°S.
In the two bottom panels of Fig. 8, the same analysis is repeated for the changes in the time-mean temperature T and precipitation P. In stark contrast with SDT, the common signal of the time-mean temperature change is in most latitudes over 80% of the total squared amplitude (Fig. 8d). The ratio M2/A2 also tends to be higher for the change in the annual mean precipitation (Fig. 8e) than for the SDP change, although this difference is much smaller. Thus, the experiments agree better on the changes in time-mean climate than on the changes in variability. This is largely because the changes in time-mean climate are much more discernible from the noise in the simulations. In most latitudes, the noise is only calculated to explain about a third of the interexperiment differences in the time-mean precipitation change, and 10%–15% of the differences in temperature change. Its contribution to the total squared amplitude of the temperature change is very small.
This analysis was also made for the three-month zonal mean variability changes (not shown). The results showed a somewhat larger relative contribution from the common signal M2 than in the case of the local changes, and a somewhat smaller contamination by noise. However, noise appears to contribute substantially even to the zonal mean changes, explaining typically about a half of the interexperiment differences in the SDT and SDP changes and most of the differences in the CVP change.
6. Changes in variability versus changes in mean climate
Globally characteristic R ratios, obtained by replacing the numerator and denominator of (4) by their global root-mean-square (rms) amplitudes, are given in Table 3. They depend slightly on the details of the calculation. The ratios for seasonal changes tend to be slightly higher than those for the annual changes. Likewise, the ratios grow larger when evaluated for the changes in the individual models, rather than for the 19-model mean changes. This is, however, partly due to the limited sampling period: the fact that there is much more noise in the variability changes than in the changes in time-mean climate (Fig. 8) implies that lower ratios would be obtained if more than 20 yrs of data were used. In any event, the implications regarding the extremes of interannual variability are clear. The changes in temperature extremes will be primarily determined, in the light of these model results, by the changes in mean temperature. For the changes in precipitation extremes, the changes in the standard deviation and the mean precipitation are of more comparable importance although, in many areas, the former can be partly accounted for by assuming an unchanged coefficient of variation.
7. Possible causes of the variability changes
Changes in near-surface climate variability may be induced by a multitude of factors, including changes in surface state (such as snow and ice cover and soil moisture) and in the circulation and the thermodynamic properties (e.g., moisture content and cloudiness) of the atmosphere. An extensive study of these mechanisms is beyond the scope of the present work, but a few issues related to the interpretation of the CMIP2 simulated variability changes are nevertheless studied below. First, a partly successful attempt is made to explain the changes in SDT simply by the change in the mean temperature. Then, the general increase in precipitation variability is discussed.
a. Impact of time-mean warming on temperature variability
In section 4 it was found that, in high latitudes in winter, models with warm (cold) climate tend to simulate weak (strong) temperature variability (Fig. 4). Qualitatively in accord with this, the warming associated with doubled CO2 is generally accompanied in the CMIP2 models by reduced temperature variability in high latitudes. To study how well this connection between the simulated climate changes and the model-to-model differences in control climate holds in quantitative terms, a simple linear regression was used to predict the change in SDT from the change in T. The regression coefficients linking the two quantities were derived for each grid box and each calendar month separately, using the model-to-model differences in control run T and SDT. The changes in 19-model mean SDT, which the change in mean temperature should have produced according to this regression are shown in Fig. 10. The two panels of Fig. 10, depicting the annual mean and the seasonal zonal means of the regression-based changes, are directly comparable with the actual changes in Figs. 6a and 7a.
Several of the actually simulated features in SDT change are reproduced by the regression. In particular, Fig. 10a shows a decrease in SDT in high northern latitudes that has nearly the right magnitude and broadly the right geographical patterns. The minimum of −0.5°C near Spitsbergen, Norway, agrees well with Fig. 6a, as does the somewhat smaller decrease in SDT in Siberia. The regression also captures the decrease in SDT over the high-latitude Southern Ocean (although this extends too much over Antarctica) and the increase over Africa and South America. The global spatial correlation between the two fields is 0.71. The pronounced seasonal cycle of the change in the extratropical Northern Hemisphere is also reproduced, with the largest decrease in SDT in autumn and winter and little change or slight increase in summer (Fig. 10b). However, the regression-based changes also exhibit features that differ from the actually simulated changes, including a decrease over almost all sea areas.
These results suggest that the simulated SDT changes are at least in some areas a relatively direct consequence of the CO2-induced warming. In particular, the reduced variability in high northern latitudes and the high-latitude Southern Ocean likely reflects the decrease of sea ice and snow with increasing temperature. On the other hand, temperature variability is effected by other factors than the mean temperature as such, in both the control runs and regarding the CO2-induced changes. The partial disagreement between the simulated and the regression-based changes is thus not surprising. In addition to the changes in snow and ice, factors that may play a role include changes in the atmospheric circulation, the gradient of the time-mean temperature (Rind et al. 1989; Cao et al. 1992), and soil moisture (e.g., Zwiers and Kharin 1998).
b. Why does precipitation variability increase?
The CMIP2 simulations suggest a CO2-induced increase in the standard deviation of monthly precipitation in most of the world. The 19-model global mean SDP change is 4.2%, which is qualitatively similar but larger than the average 2.5% increase in mean precipitation (Table 2). In accord with this, there is also an average 2.4% increase in the coefficient of variation (CVP). Why does the monthly precipitation variability increase even in relative terms?
Equation (A5) in the appendix relates CVP to three characteristics of daily precipitation time series. An increase in CVP requires either an increase in the day-to-day persistence of precipitation, an increase in the coefficient of variation of precipitation among the wet days cvpw (which requires a change in the shape of the wet-day precipitation distribution), a decrease in wet-day frequency, or a combination of these. The monthly CMIP2 data do not allow us to discriminate between these alternatives. However, we see no obvious reason to wait for systematic changes in the day-to-day persistence of weather in a warmer climate, and therefore believe that the latter two factors are more important.
Simulated CO2-induced changes in cvpw have probably never been reported explicitly. However, the analysis of Gregory and Mitchell (1995) of the U. K. Met Office High Resolution General Circulation Model (UKHI GCM) results over Europe indicates a change in the shape of the daily precipitation distribution that would produce this effect. Changes in precipitation frequency have been studied more often (e.g., Gordon et al. 1992; Hennessy et al. 1997; Zwiers and Kharin 1998), although with varying nonzero thresholds of what is counted as a wet day. While disagreeing in detail, all these studies share a general decrease in precipitation frequency in subtropics and lower midlatitudes (where the increase in CVP in the CMIP2 experiments is largest) and an increase in high latitudes (where CVP mostly decreases in CMIP2). One may also note that atmospheric moisture increases much more in model greenhouse experiments than the global precipitation. The average increase in moisture content in the CMIP2 experiments is 13% (approximately as expected from the Clausius–Clapeyron relationship), whereas the change in mean precipitation is constrained by the requirements of surface and atmospheric energy balance (e.g., Mitchell et al. 1987) to much lower values (Table 2). To the extent that instantaneous precipitation rates follow the absolute moisture content, this is suggestive (in a global mean sense) of fewer or shorter precipitation events in a warmer climate.
As a final note, the relationship (5) indicates that, at the limit of small perturbations, the relative change in the standard deviation of precipitation is the sum of the relative changes in the mean precipitation and the coefficient of variation. When the changes in SDP, P, and CVP are averaged over several models, grid boxes, and calendar months, the additivity deteriorates since the covariance between P and CVP comes into play. As illustrated in Fig. 11, which shows the relative 19-model zonal and annual mean changes in the three quantities, this approach nevertheless gives a useful first-order view of how the changes in P and CVP contribute to the average SDP change.
Close to the equator and poleward of 50°N and 50°S, the increase in SDP is mostly or totally due to the increase in the mean precipitation. In particular, in high northern latitudes the increase in P is counteracted by a slight decrease in CVP. In the subtropics and at lower midlatitudes, the reverse is true. The change in the mean precipitation is small or (at 10°–40°S) negative, and the increase in SDP is mainly or totally attributable to the increase in CVP.
8. Summary
Changes in external conditions may effect not only the mean state but also the variability of climate. Here, CO2-induced changes in the interannual variability of monthly temperature and precipitation have been studied using 19 AOGCM experiments participating in the CMIP2 intercomparison. The magnitude of interannual variability differs substantially among the CMIP2 control runs, but most aspects of the average geographical and seasonal distributions of variability appear realistic. However, the models tend to overestimate temperature variability over low-latitude land areas.
Gradual doubling of CO2 leads, in most models, to a decrease in temperature variability in the winter half-year in the extratropical Northern Hemisphere, especially over the Arctic Ocean where ice retreats. Similarly, variability tends to decrease over the high-latitude Southern Ocean during the austral winter. Over land in low latitudes and in northern midlatitudes in summer, a slight tendency toward increased temperature variability occurs.
Absolute (standard deviation) and relative (coefficient of variation) precipitation variability both generally increase in most areas. The standard deviation increases where the mean precipitation increases, but also in some of those areas where the mean precipitation decreases. The coefficient of variation increases especially where the mean precipitation decreases. Although other factors may contribute, comparison with earlier studies suggests that such increases in relative precipitation variability are at least partly related to a reduced frequency of precipitation days.
Changes in variability vary substantially among the 19 CMIP2 experiments. In contrast with the changes in time-mean temperature and precipitation, a majority of the interexperiment differences in variability change appear to be due to noise in the simulations, rather than a result of model differences. All in all, the simulated changes in interannual variability have a markedly lower signal-to-noise ratio than the changes in time mean climate. Taking the models as an analogue of reality, the same will probably be the case even with real future climate changes.
The changes in the interannual standard deviation of monthly temperature are, in general, a small fraction of the simulated time mean warming. The implication to the real future world with higher greenhouse gas concentrations is that, in most areas, the changes in the extremes of interannual temperature variability will be essentially determined by the average temperature change. However, the same is not necessarily true for the changes in interannual precipitation extremes. For the changes in temperature and precipitation extremes on the daily timescale, changes in variability are potentially more important than is the case with the extreme monthly values.
Acknowledgments
All CMIP2 modeling groups are acknowledged for conducting and making available the simulations requested by the CMIP Panel. CMIP is supported and the model data are distributed by the Program for Climate Model Diagnosis and Intercomparison (PCMDI) at the Lawrence Livermore National Laboratory (LLNL). This research has been funded by MISTRA and SMHI within the Swedish SWECLIM program.
REFERENCES
Barthelet, P., L. Terray, and S. Valcke, 1998: Transient CO2 experiment using the ARPEGE/OPAICE non flux corrected coupled model. Geophys. Res. Lett., 25 , 2277–2280.
Bell, J., P. Duffy, C. Covey, and L. Sloan, and the CMIP investigators,. 2000: Comparison of temperature variability in observations and sixteen climate model simulations. Geophys. Res. Lett., 27 , 261–264.
Boer, G. J., G. Flato, and D. Ramsden, 2000: A transient climate change simulation with greenhouse gas and aerosol forcing: Projected climate change to the twenty-first century. Climate Dyn., 16 , 427–450.
Boville, B. A., and P. R. Gent, 1998: The NCAR Climate System Model, Version One. J. Climate, 11 , 1115–1130.
Braconnot, P., O. Marti, and S. Joussaume, 1997: Adjustment and feedbacks in a global coupled ocean–atmosphere model. Climate Dyn., 13 , 507–519.
Cao, H. X., J. F. B. Mitchell, and J. R. Lavery, 1992: Simulated diurnal range and variability of surface temperature in a global climate model for present and doubled CO2 climates. J. Climate, 5 , 920–943.
Cubasch, U., and Coauthors. 2001: Projections of future climate change. Climate Change 2001: The Scientific Basis, J. T. Houghton et al., Eds., Cambridge University Press, 525–582.
Delworth, T. L., and S. Manabe, 1988: The influence of potential evaporation on the variabilities of simulated soil wetness and climate. J. Climate, 1 , 523–546.
Delworth, T. L., 1989: The influence of soil wetness on near-surface atmospheric variability. J. Climate, 2 , 1447–1462.
Duffy, P. B., J. Bell, C. Covey, and L. Sloan, and the CMIP investigators,. 2000: Effect of flux adjustments on temperature variability in climate models. Geophys. Res. Lett., 27 , 763–766.
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.
Gates, W. L., 1992: AMIP: The Atmospheric Model Intercomparison Project. Bull. Amer. Meteor. Soc., 73 , 1962–1970.
Giorgi, F., and R. Francisco, 2000: Evaluating uncertainties in the prediction of regional climate change. Geophys. Res. Lett., 27 , 1295–1298.
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., P. H. Whetton, A. B. Pittock, A. M. Fowler, and M. R. Haylock, 1992: Simulated changes in daily rainfall intensity due to the enhanced greenhouse effect: Implications for extreme rainfall events. Climate Dyn., 8 , 83–102.
Gregory, J. M., and J. F. B. Mitchell, 1995: Simulation of daily variability of surface temperature and precipitation over Europe in the current and 2 × CO2 climates using the UKMO climate model. Quart. J. Roy. Meteor. Soc., 121 , 1451–1476.
Grotch, S. L., and M. C. MacCracken, 1991: The use of general circulation models to predict regional climate change. J. Climate, 4 , 286–303.
Hennessy, K. J., J. M. Gregory, and J. F. B. Mitchell, 1997: Changes in daily precipitation under enhanced greenhouse conditions. Climate Dyn., 13 , 667–680.
Hirst, A., S. P. O'Farrell, and H. B. Gordon, 2000: Comparison of a coupled ocean–atmosphere model with and without oceanic eddy-induced advection. Part I: Ocean spinup and control integrations. J. Climate, 13 , 139–163.
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, spinup and validation. Climate Dyn., 13 , 103–134.
Kalnay, E., and Coauthors. 1996: The NCEP/NCAR 40-Year Reanalysis Project. Bull. Amer. Meteor. Soc., 77 , 437–471.
Kattenberg, A., and Coauthors. 1996: Climate models—Projections of future climate. Climate Change 1995: The Science of Climate Change, J. T. Houghton et al., Eds., Cambridge University Press, 285–357.
Katz, R., and B. Brown, 1992: Extreme events in a changing climate: Variability is more important than averages. Climatic Change, 21 , 303–315.
Kharin, V. V., and F. W. Zwiers, 2000: Changes in the extremes in an ensemble of transient climate simulations with a coupled atmosphere–ocean GCM. J. Climate, 13 , 3760–3788.
Kistler, R., and Coauthors. 2001: The NCEP/NCAR 50-Year Reanalysis: Monthly means CD-ROM and documentation. Bull. Amer. Meteor. Soc., 82 , 247–268.
Kittel, T. G. F., F. Giorgi, and G. A. Meehl, 1998: Intercomparison of regional biases and doubled CO2-sensitivity of coupled atmosphere–ocean general circulation experiments. Climate Dyn., 14 , 1–15.
Knutson, T. R., T. L. Delworth, K. W. Dixon, and R. J. Stouffer, 1999: Model assessment of regional surface temperature trends (1949–1997). J. Geophys Res., 104 , 30981–30996.
Manabe, S., R. J. Stouffer, M. J. Spelman, and K. Bryan, 1991: Transient responses of a coupled ocean–atmosphere model to gradual changes of atmospheric CO2. Part I: Annual mean response. J. Climate, 4 , 785–818.
Meehl, G. A., M. Wheeler, and W. M. Washington, 1994: Low-frequency variability and CO2 transient climate change. Part 3: Intermonthly and interannual variability. Climate Dyn., 10 , 277–303.
Meehl, G. A., G. J. Boer, C. Covey, M. Latif, and R. J. Stouffer, 2000: The Coupled Model Intercomparison Project (CMIP). Bull. Amer. Meteor. Soc., 81 , 313–318.
Mitchell, J. F. B., C. A. Wilson, and W. M. Cunnington, 1987: On CO2 climate sensitivity and model dependence of results. Quart. J. Roy. Meteor. Soc., 113 , 293–322.
Mitchell, J. F. B., S. Manabe, V. Meleshko, and T. Tokioka, 1990: Equilibrium climate change and its implications for the future. Climate Change: The IPCC Assessment, J. T. Houghton, G. J. Jenkins, and J. J. Ephraums, Eds., Cambridge University Press, 131–164.
New, M., M. Hulme, and P. Jones, 1999: Representing twentieth-century space–time climate variability. Part I: Development of a 1961–90 mean monthly terrestrial climatology. J. Climate, 12 , 829–856.
New, M., 2000: Representing twentieth-century space–time climate variability. Part II: Development of 1901–96 monthly grids of terrestrial surface climate. J. Climate, 13 , 2217–2238.
Nozawa, T., S. Emori, T. Takemura, T. Nakajima, A. Numaguti, A. Abe-Ouchi, and M. Kimoto, 2000: Coupled ocean–atmosphere model experiments of future climate change based on IPCC SRES scenarios. Preprints, 11th Symp. on Global Change Studies, Long Beach, CA, Amer. Meteor. Soc., 352–355.
Power, S. B., R. A. Colman, B. J. McAvaney, R. R. Dahni, A. M. Moore, and N. R. Smith, 1993: The BMRC coupled atmosphere/ocean/sea-ice model. BMRC Res. Rep. 37, Bureau of Meteorology Research Centre, Melbourne, Australia, 58 pp.
Räisänen, J., 2001: CO2-induced climate change in CMIP2 experiments: Quantification of agreement and role of internal variability. J. Climate, 14 , 2088–2104.
Rind, D., R. Goldberg, and R. Ruedy, 1989: Change in climate variability in the 21st century. Climatic Change, 14 , 5–37.
Roeckner, E., L. Bengtsson, J. Feichter, J. Lelieveld, and H. Rodhe, 1999: Transient climate change simulations with a coupled atmosphere–ocean GCM including the tropospheric sulfur cycle. J. Climate, 12 , 3004–3032.
Russell, G. L., and D. Rind, 1999: Response to CO2 transient increase in the GISS coupled model: Regional coolings in a warmer climate. J. Climate, 12 , 531–539.
Tett, S. F. B., T. C. Johns, and J. F. B. Mitchell, 1997: Global and regional variability in a coupled AOGCM. Climate Dyn., 13 , 303–323.
Tokioka, T., A. Noda, A. Kitoh, Y. Nikaidou, S. Nakagawa, T. Motoi, S. Yukimoto, and K. Takata, 1995: A transient CO2 experiment with the MRI CGCM: Quick report. J. Meteor. Soc. Japan, 73 , 817–826.
Voss, R., R. Sausen, and U. Cubasch, 1998: Peridiocally synchronously coupled integrations with the atmosphere–ocean general circulation model ECHAM3/LSG. Climate Dyn., 14 , 249–266.
Washington, W. M., and Coauthors. 2000: Parallel climate model (PCM) control and transient simulations. Climate Dyn., 16 , 755–774.
Xie, P., and P. A. Arkin, 1997: Global precipitation: A 17-year monthly analysis based on gauge observations, satellite estimates, and numerical model outputs. Bull. Amer. Meteor. Soc., 78 , 2539–2558.
Yukimoto, S., M. Endoh, Y. Kitamura, A. Kitoh, T. Motoi, and A. Noda, 2000: ENSO-like interdecadal variability in the Pacific Ocean as simulated in a coupled GCM. J. Geophys. Res., 105 ((D10),) 13945–13963.
Zhang, X., G. Shi, H. Liu, and Y. Yu, Eds.,. 2000: IAP Global Ocean–Atmosphere–Land System Model. Science Press, 249 pp.
Zwiers, F. W., and V. V. Kharin, 1998: Changes in the extremes of the climate simulated by the CCC GCM2 under CO2 doubling. J. Climate, 11 , 2200–2222.
APPENDIX
Relationship between Daily and Interannual Precipitation Variability
Annually and geographically averaged variability in the 19 control simulations over land (horizontal axis) and over sea (vertical axis): (a) std dev of temperature, (b) std dev of precipitation, (c) coefficient of variation of precipitation. The models are lettered as indicated in the legend, and the closed circle gives the 19-model mean. The horizontal and the two vertical lines in each panel represent observational estimates of variability. In (b) the letters A, H, L, and M are clustered near (0.85, 1.45) mm d−1, and the letters Q, S, and R near (0.85, 1.25) mm d−1. See text for further explanation
Citation: Journal of Climate 15, 17; 10.1175/1520-0442(2002)015<2395:CICIIT>2.0.CO;2
Comparison of annually averaged variability in the control simulations with an observational estimate: (left column) 19-model annual means of the (a) std dev of monthly temperature, (d) std dev of precipitation, and (g) coefficient of variation of precipitation. (b), (e),(h) The variability inferred from observational datasets (see text). (c),(f),(i) Percent fraction of models in which the variability exceeds observations
Citation: Journal of Climate 15, 17; 10.1175/1520-0442(2002)015<2395:CICIIT>2.0.CO;2
Running 3-month zonal means of temperature variability SDT (in °C): (a) average of the 19 CMIP2 control runs, (b) observational data, and (c) the ratio between the 19-model mean and the observational estimate
Citation: Journal of Climate 15, 17; 10.1175/1520-0442(2002)015<2395:CICIIT>2.0.CO;2
Model-to-model correlation between mean temperature and temperature variability in (a) DJF and (b) JJA. Values are multiplied by 10; contours at ±4, ±6, and ±8 × 10−1. Dark (light) shading is used for correlations above 4 × 10−1 (below −4 × 10−1)
Citation: Journal of Climate 15, 17; 10.1175/1520-0442(2002)015<2395:CICIIT>2.0.CO;2
Changes in 19-model (a) annual mean temperature (in °C) and (b) precipitation (% of 19-model control run mean) at doubling of CO2 (model years 61–80)
Citation: Journal of Climate 15, 17; 10.1175/1520-0442(2002)015<2395:CICIIT>2.0.CO;2
(a),(d),(g) Annual 19-model mean changes in SDT, SDP, and CVP at doubling of CO2 (years 61–80). (b),(e),(h) The same in percent of the control run 19-model mean. (c),(f),(i) The ratio between the mean change and the model-to-model standard deviation. A slight smoothing is used to enhance legibility
Citation: Journal of Climate 15, 17; 10.1175/1520-0442(2002)015<2395:CICIIT>2.0.CO;2
As in Fig. 6, but for running 3-month zonal averages of the SDT, SDP, and CVP changes. The oval symbols in the right column indicate that the change has the same sign in all 19 models
Citation: Journal of Climate 15, 17; 10.1175/1520-0442(2002)015<2395:CICIIT>2.0.CO;2
Components of the partition [see Eq. (3)] of local annual climate changes. The dark shaded areas represent the common signal M2, the lighter shaded areas are the noise-related variance N2, and the unshaded areas are the model-related variance D2, all normalized so that their sum corresponds to 100 units. (a)–(c) Changes in interannual variability (ΔSDT, ΔSDP, and ΔCVP), and (d), (e) changes in time-mean temperature (ΔT) and precipitation (ΔP)
Citation: Journal of Climate 15, 17; 10.1175/1520-0442(2002)015<2395:CICIIT>2.0.CO;2
The ratio between (a) the 19-model annual mean changes in SDT and T, (b) the change in SDP and the absolute value of the change in P, and (c) the CVP contribution to the change in SDP and the absolute value of the change in P. The shading values in (a) differ from those in (b) and (c)
Citation: Journal of Climate 15, 17; 10.1175/1520-0442(2002)015<2395:CICIIT>2.0.CO;2
The changes in 19-model mean SDT predicted from the changes in time-mean temperature with the regression model of section 7a: (a) local annual mean changes, (b) 3-month mean zonal averages. The unit is 0.1°C and positive values are shaded
Citation: Journal of Climate 15, 17; 10.1175/1520-0442(2002)015<2395:CICIIT>2.0.CO;2
Percent changes in 19-model zonal and annual mean P (dashed line), CVP (dotted line), and SDP (solid line)
Citation: Journal of Climate 15, 17; 10.1175/1520-0442(2002)015<2395:CICIIT>2.0.CO;2
The 19 CMIP2 models
Changes in globally and annually averaged time-mean climate and interannual climate variability in the 19 experiments with a doubling of CO2 (model years 61–80). The highest value is shown in bold and the lowest in italics
The ratio between the global rms amplitudes of the changes in variability and the changes in mean climate. The first row shows the ratios for the 19-model mean changes, and the second row shows the ratios for the changes in the individual models. Ann = annual mean