An examination of the differences between surface and free-air temperature trend at high-elevation sites: Relationships with cloud cover, snow cover, and wind



[1] Contrasts in high-elevation surface and free-tropospheric temperatures between 1971 and 1996 are examined by comparing surface temperatures from a subset of 72 stations in the GHCN (Global Historical Climate Network) and CRU (Climatic Research Unit) homogeneity adjusted surface data sets with free-air temperatures interpolated to the same locations from NCEP/NCAR Reanalysis R1. The selected stations are above the mean elevation of the surrounding topography, often located on mountain summits. Surface temperatures, free-air temperatures, and their difference (ΔT) are found to be related to independent surface cloud observations from the NDP-026C archive, local wind speed, satellite records of snow cover (NSIDC), and reanalysis wind components. Significant correlations are observed at most stations, and correlation spatial patterns are consistent for different subperiods of the record (e.g., presatellite era versus satellite era). Stepwise regression models built to predict surface temperatures, free-air temperatures, and ΔT from the above meteorological parameters typically explain 20–40% of the temperature variability on an annual basis and more for individual seasons. The stationarity of relationships between temperature and snow/cloud/wind is examined by comparing the temporal trends in the original temperatures with predicted trends from the best fit regression model and trends in model residuals. This provides an assessment of how much of any ΔT trends can be accounted for by changes in meteorology. Significant daytime ΔT residual trends occur primarily in Turkey and eastern China, but significant nighttime ΔT residual trends are more geographically widespread. While daytime residual trends may be the result of surface radiative cooling by increasing anthropogenic aerosol, attribution of nighttime residual trends is uncertain.

1. Introduction

[2] Numerous recent studies have found contrasting trends in surface and free-air tropospheric temperatures over the past 2 to 3 decades, typically with more warming at the surface than in the free troposphere. Although most investigations have been concerned with temperature trends on the global scale [National Research Council, 2000; Christy et al., 2003; Lanzante et al., 2003; Jones and Moberg, 2003; Fu et al., 2004], differing trends in surface and tropospheric temperature have also been noted on a regional and local basis [Seidel and Free, 2003; Pepin and Losleben, 2002; Pepin and Seidel, 2005]. Interpretation of such differences is complex, and it is currently uncertain whether the contrasting trends indicate a real long-term shift in the state of the climate system or merely result from inadequacies in and incompatibilities between the various observing systems involved, which include surface instruments, radiosondes and satellites.

[3] One difficulty with assessing potential errors in surface and tropospheric temperature data is that the measurements are not made at the same elevation. The present study overcomes this limitation by comparing high-elevation screen-level surface temperature observations (Ts) with free-air tropospheric data interpolated to the same location and elevation (Ta). Surface temperatures are obtained from a selection of high-elevation stations within the Global Historical Climate Network (GHCN) and Climate Research Unit (CRU) data sets with homogenous records (mostly in the Northern Hemisphere), and interpolated free-air temperatures are obtained from the NCAR/NCEP reanalysis R1. Since surface properties and atmospheric processes can have different effects on Ts and Ta, it is possible that changes in meteorological conditions may be partly responsible for contrasting trends in surface and free-air temperature.

[4] Time series of Ts, Ta, and their difference (TsTa, referred to as ΔT) are examined for relationships to other meteorological parameters including cloud cover, zonal and meridional wind components, surface wind speed, and snow cover. Strong relationships with these parameters, especially as measured by independent data, would increase confidence that at least some of the difference in trends observed between the surface and free-air data sets is real, rather than solely an artefact of data error. Furthermore, it is of interest whether relationships between ΔT and meteorological parameters are consistent over long periods of time (i.e., stationary). If so, any secular trends in ΔT should be accompanied by attendant changes in cloud, wind, or snow.

[5] In section 2, expected relationships between ΔT and other meteorological parameters are reviewed through discussion of previous literature. Section 3 introduces the data sets and preparation methods. Individual relationships are discussed in section 4, leading to regression models to predict surface, free-air temperatures, and ΔT. Section 5 examines temporal trends in temperatures (including ΔT) and discusses the extent to which cloud, wind, and snow changes are associated with the observed temperature changes.

2. Controls of ΔT

[6] Numerous field studies have examined instantaneous and climatological mean differences between surface (screen level) temperatures, usually measured on mountain summits, and temperatures in the free air nearby at equivalent elevations measured from radiosondes [Samson, 1965; McCutchan, 1983; Richner and Phillips, 1984; Tabony, 1985]. Barry [1992] provides a summary of such studies. The main temporal controls of the difference (ΔT) are cloudiness, advection (winds), snow cover, and topographical characteristics.

[7] Diurnal and seasonal structures in the differences are frequently reported, with mountain top sites in extratropical latitudes tending to be warmer than the free air (ΔT > 0) in the midday period and in summer, but colder (ΔT < 0) at night and in winter [McCutchan, 1983; Richner and Phillips, 1984]. This suggests that net local surface radiation flux is an important control. Cloud cover acts to suppress differences in surface and free air temperature by reducing the otherwise enhanced surface heating during day and enhanced surface cooling during night. Thus daytime positive values of ΔT (Ts > Ta) are reduced under increased cloudiness, while nighttime negative values are minimized. Low clouds are more influential than high clouds at night since low clouds are warmer and consequently emit more thermal radiation toward the surface. Moisture, in the form of specific humidity, controls latent heat transfer. In a dry atmosphere where evaporation is limited, free air temperatures aloft are likely to be relatively cold in comparison with the mountain surface [Molnar and Emanuel, 1999].

[8] Surface temperature often lags behind free-air temperature in cases of strong large-scale advection, particularly in sheltered locales such as valley bottoms. This means ΔT will be negative (Ts < Ta) for warm advection and positive for cold advection (Ts > Ta). Advection is closely related to the direction of the large-scale wind and is of particular importance in midlatitudes where air mass contrasts are great. To our knowledge no systematic investigation of this phenomenon has been published.

[9] Snow cover suppresses surface heating by acting as a heat sink, in that energy is transferred to the atmosphere in the melting process and the sensible heat flux is greatly reduced. This, combined with the high surface albedo, will make ΔT negative in most snow-covered areas, even by day.

[10] Absolute elevation is also important [Molnar and Emanuel, 1999], and mean annual values of ΔT become more positive at higher elevations, assuming radiative and convective equilibrium.

[11] The topography at a particular locale can modify the relationships described above in complex ways. A mountain peak or summit site will be more strongly influenced by free-air advection than a valley site, so values of ΔT should be less temporally variable and closer to zero at these exposed locations. However, it is not clear whether the above relationships between ΔT, wind speed, cloud and snow become more or less influential at summit sites.

3. Data and Preparation Methods

[12] Although free-air temperatures and snow cover were available worldwide from global gridded data sets, availability of both simultaneous homogenous surface temperature records and high-resolution station cloud cover reports placed limits on possible locations for study. Since our research focuses on examining the factors responsible for past surface and free-air temperature differences rather than presenting a comprehensive global analysis of observed trends, data quality was of more importance than quantity (length of period, spatial coverage). Only 228 high-elevation (>500 m) GHCN/CRU surface temperature stations, mostly in the Northern Hemisphere, also had cloud cover data. Many of these stations were in areas of complex topography and did not have elevations greater than the average elevation of the surrounding terrain. Because it is important that the interpolated reanalysis free-air temperature represent true free-air conditions rather than an extrapolation below the Earth's surface, a subset of 72 stations with elevations well above the mean elevation of the four surrounding 2.5° × 2.5° reanalysis grid points was chosen for analysis (Table 1 and Figure 1). The plurality of these stations is exposed mountain summit sites, but there are a few sites in high-elevation basins or valleys. Sites are distributed over most of the major continents, but with a heavy concentration in Eurasia. The network does not sample the Southern Hemisphere well (only two sites), and South America is omitted. A variety of elevations (from 500 m to 3950 m) are sampled, with 7 sites located at greater than 3000 m. The mean station elevation is 1554 m. The 1971–1996 time period is analyzed because this is when reliable cloud and temperature data are both available. The trend analysis (section 5) is restricted to 1971–1990 because of data gaps in the 1990s at some stations.

Figure 1.

Geographical distribution of the 72 GHCN and CRU stations used in this study with symbols indicating site topography: summit or ridge (upward pointing triangle), slope (diamond), flat (square), basin (circle), incised valley (downward pointing triangle), and unknown (star).

Table 1. List of 72 Stationsa
WMO NumberNameLatitudeLongitudeElevation, mTopographical ClassEffective Sample Size
  • a

    Latitude and longitude are in decimal degrees. Topographical class is discussed in the main text: 1, summit or ridge; 2, slope; 3, flat; 4, basin; 5, incised valley; 6, not enough information.


3.1. Temperature

[13] Surface temperatures in the form of monthly mean maximum and minimum temperature anomalies were calculated for the 72 sites from the GHCN version 2 and CRU data sets [Peterson and Vose, 1997; Peterson et al., 1998; Jones, 1994; Jones et al., 1999; Jones and Moberg, 2003]. These data sets have been adjusted to improve temporal homogeneity and are frequently used in climate change studies.

[14] Free-air temperatures from the NCEP/NCAR reanalysis R1 [Kistler et al., 2001] were interpolated both vertically and horizontally to the exact location of the surface site using one or two of the four times daily grids (0, 6, 12 and 18 UTC). To ensure fair comparison with surface maxima and minima, grids corresponding to a time between 1230–1530 local solar time LST (maxima) and 0430–0730 LST (minima) at each longitude were used. Where necessary, temporal interpolation to a 3 hourly reading (i.e., midway between 6 hourly values) was obtained via simple averaging of the two relevant grids. Spatial interpolation was linear, with the vertical interpolation to the relevant height at the four nearest grid points done first. Interpolated temperatures are derived from aboveground pressure level data for the two levels immediately above and below the elevation of the surface site, not skin temperatures produced by the model [see Kalnay and Cai, 2003; Trenberth, 2004; Vose et al., 2004]. Because the surface sites are above the surrounding reanalysis grid points, local boundary layer effects, which would invalidate the linear assumption, are minimized. Since the reanalysis assimilated a mix of satellite and radiosonde observations but no surface observations, Ta and Ts come from independent sources. Despite the incorporation of a GCM in the reanalysis, tropospheric temperatures are highly dependent on the observations over the well-sampled Northern Hemisphere continents. Ease of interpolation to the station locations was the main reason free-air temperatures were obtained from the reanalysis rather than the original observations (radiosonde and satellite combination), and we must be cautious in our interpretation of the results (see final discussion). The known snow cover error in R1 does not impact the above surface temperatures used in this study (see Kanamitsu et al. [2002] for further discussion), but general concerns remain [Chelliah and Ropelewski, 2000].

[15] Monthly ΔT values were calculated by subtracting the mean free-air temperature from the mean surface temperature (TsTa) for each month (day and night separately). These were then converted into anomalies with respect to month. Daytime and nighttime values of ΔT are often referred to as ΔT max and ΔT min respectively.

3.2. Cloud Cover

[16] Observations of sky cover by all clouds (total cloud cover) and by only low clouds (low cloud cover) were taken from 3-hourly surface synoptic reports collected in the NDP-026C archive [Hahn and Warren, 1999]. Since many of the cloud cover/temperature relationships are strongly dependent on time of day, only those observations closest to the times of day of maximum and minimum temperature were used (1230–1530 LST and 0430–0730 LST, respectively). These observations were averaged to monthly mean values of total and low cloud cover (day and night separately) and then converted to anomalies with respect to month. Because surface observers have difficulty visually detecting clouds on dark nights, only those observations made under conditions of sufficient lunar illumination according the criterion of Hahn et al. [1995] were averaged. Although individual 3-hourly cloud cover values are not normally distributed, the monthly anomalies are close to normally distributed.

3.3. Snow Cover

[17] Snow cover for each month was assessed from the National Snow and Ice Data Center (NSIDC) Equal Area Scalable Earth (EASE) gridded weekly version 2 data set [Armstrong and Brodzik, 2002], which indicates for each week from 1971 to 1996 presence or absence of snow cover on a 25 km equal-area grid (Lambert's equal area azimuthal projection: The relevant grid pixel was identified for each station, and a snow index approximating the percentage of time with snow cover was calculated for each month by assigning 1 to snow-covered weeks and 0 to snow-free weeks and averaging the 4 or 5 weekly values. Individual weeks were assigned to the month with the most days in that week (e.g., week of 29 January to 4 February would be assigned to February). There are limitations to comparing a station point measurement (cloud, wind and temperature) with a pixel value (snow), especially at the edge of the global snow cover extent, and in areas of extreme relief (where there could be a systematic bias depending on the relative elevation of the station and the majority of the pixel area). However, these potential discrepancies are minor and have limited influence on snow index anomalies at any one location.

3.4. Horizontal Wind Components and Wind Speed

3.4.1. Regional

[18] Zonal (u) and meridional (v) wind components (ms−1) were extracted from the NCEP/NCAR reanalysis mean monthly files and interpolated vertically and horizontally to station locations in the same way as were free-air temperatures. The derived components are representative of the synoptic-scale gradient wind and not the local surface wind at the surface site, which will be influenced by topographical effects. Such winds are therefore used to represent large-scale air mass advection. The sign of the v component in the Southern Hemisphere was reversed such that positive/negative flow is poleward/equatorward (northerly/southerly flow).

3.4.2. Station Specific Wind Speed

[19] Wind speed recorded at the synoptic station was also available in the NDP-026C archive [Hahn and Warren, 1999]. Local surface wind speed can be substantially different from that derived from the regional wind components, especially at sheltered locations or in areas of complex topography.

3.5. Analyses

[20] All parameters were converted to anomalies. Correlations between monthly anomalies in surface, free-air temperatures, ΔT, and cloud, wind, and snow were examined for all sites. Stepwise regression models were derived for surface, free-air, and ΔT temperature anomalies for all 72 sites on the basis of cloud, wind, and snow anomalies. Although correlations and models were calculated for individual meteorological seasons (DJF, MAM, JJA, SON), for reasons of space the figures and tables mostly show results calculated over the whole year. Trends in actual temperature anomalies (1971–1990) were compared to those predicted by the best regression model based on cloud, wind, and snow anomalies at each site, and to trends in the residual differences between actual and predicted temperature anomalies. In this way the stationarity of locally relevant cloud, wind, and snow versus temperature relationships can be examined.

[21] Because trends were not removed from temperature and meteorological parameters before calculating model regression coefficients, the predicted trends are not entirely independent from the observed trends. This has little impact on our analyses, however, since correlations and coefficients based on detrended data are nearly the same as those for nondetrended data (not shown).

[22] As can be seen in Figure 1, the distribution and density of stations is highly nonuniform. Closely spaced stations possess less independent information than isolated stations, a factor that must be taken into account when calculating aggregate properties (mean, median, and percentage of significant values across all stations). The individual contribution of each station to the total spatial degrees of freedom was estimated by applying the eigenvector (EOF) method described by Bretherton et al. [1999] to concatenated Ts and Ta time series in a two-step process. The first step applied the EOF method to a combination of time series of isolated stations and regional averages (e.g., Europe, Turkey, China, etc.) to determine which single stations or groups of stations were independent from all of the others. The second step applied the EOF method to each group of stations to determine the spatial degrees of freedom of that group. Individual stations were assigned an effective sample size equal to the number of spatial degrees of freedom in the group divided by the number of stations in the group (last column of Table 1). Values ranged from 1.0 (Australia, South Africa, Baffin Island) to 0.07 (China), and the total effective sample size of all 72 stations was 13.

4. Results

4.1. Correlations

[23] With 5 time periods (4 seasons and all months together), 6 predictands, and 6 controlling variables (2 cloud, 3 wind and 1 snow) for 72 stations, the total number of correlations calculated was 12,960. Thus, of necessity, only a brief outline of the salient details is discussed. In the summary Tables 26 the arithmetic mean correlation can be rather meaningless because of the additive effect of strong negative and positive correlations, so the mean absolute correlation value (ignoring signs) is also given, along with the median value (more robust to outliers). All values are weighted by individual station effective sample size to account for spatial autocorrelation. In this and subsequent tables and figures max(imum) refers to daytime and min(imum) to nighttime. Although for some variables we have an a priori expectation that the correlation has a particular sign (e.g., cloud and snow), statistically significant correlations with the opposite sign can occur because there is a third parameter, such as advection, strongly controlling both such that expected relationship between them is overwhelmed.

Table 2. Correlations Between Total and Low Cloud Cover Anomalies and Temperature Anomalies (Ts, Ta and ΔT)a
 SurfaceFree AirDelta-T
  • a

    The majority of the table concerns total cloud, with low cloud correlations limited to the bottom two lines. All derived correlations in this table and Tables 37 (means, medians and percentages) take spatial autocorrelation into account by weighting individual station values by their relative effective sample size (see text). In this table and all subsequent tables.

  • b

    Sig. means p < 0.05 (significant).

  • c

    M.sig. means p < 0.10 (marginally significant).

Mean correlation−0.2810.079−0.207−0.065−0.2480.179
Median correlation−0.3070.070−0.238−0.040−0.2110.173
Mean (abs value)0.2840.1560.2240.1100.2610.218
Max correlation (+ve)0.0860.3870.3150.1870.1440.543
Min correlation (−ve)−0.740−0.435−0.564−0.432−0.714−0.299
Number sig.,b %79.556.775.238.981.865.8
Number m.sig.,c %
Sig. +ve, %
Sig. −ve, %79.512.973.532.880.65.2
Total −, %96.235.586.960.589.826.5
Mean cor. (low cloud)−0.2640.070−0.190−0.074−0.2430.180
Median (low cloud)−0.3310.063−0.222−0.071−0.2160.174
Table 3. Correlations Between u Wind Component Anomaly and Temperature Anomalies
 SurfaceFree AirDelta-T
  • a

    Sig. means p < 0.05 (significant).

  • b

    M.sig. means p < 0.10 (marginally significant).

Mean correlation0.0830.0540.0460.0590.1360.037
Median correlation0.0890.1030.0780.1430.1370.047
Mean (abs value)0.2130.2350.2120.2250.1840.144
Max correlation (+ve)0.5350.5310.5830.5390.5970.317
Min correlation (−ve)−0.422−0.499−0.475−0.482−0.414−0.558
Number sig.,a %70.672.962.675.960.340.0
Number m.sig.,b %4.01.712.
Sig. +ve, %47.446.139.851.352.824.9
Sig. −ve, %23.226.822.824.67.515.1
Total −, %
Table 4. Correlations Between v Wind Component Anomaly and Temperature Anomalies
 SurfaceFree AirDelta-T
  • a

    Sig. means p < 0.05 (significant).

  • b

    M.sig. means p < 0.10 (marginally significant).

Mean correlation0.1850.3010.2650.327−0.086−0.013
Median correlation0.2030.2910.2790.315−0.128−0.030
Mean (abs value)0.2220.3060.2650.3270.1890.150
Max correlation (+ve)0.5360.6230.5500.6300.4380.457
Min correlation (−ve)−0.366−0.321−0.0130.025−0.599−0.555
Number sig.,a %69.980.784.597.862.843.5
Number m.sig.,b %
Sig. +ve, %
Sig. −ve, %3.90.50050.223.8
Total −, %
Table 5. Correlations Between Station Wind Speed Anomaly and Temperature Anomalies
 SurfaceFree AirDelta-T
  • a

    Sig. means p < 0.05 (significant).

  • b

    M.sig. means p < 0.10 (marginally significant).

Mean correlation0.0440.063−0.0100.0060.1080.078
Median correlation0.0200.048−0.0430.0120.1160.101
Mean (abs value)0.1000.1140.1080.0830.1450.138
Max correlation (+ve)0.4130.3720.3420.2020.3970.419
Min correlation (−ve)−0.266−0.231−0.312−0.239−0.587−0.339
Number sig.,a %25.830.440.336.157.842.8
Number m.sig.,b %
Sig. +ve, %17.026.519.720.153.235.4
Sig. −ve, %8.83.920.616.04.67.4
Total −, %
Table 6. Correlations Between Snow Cover Anomaly and Temperature Anomalies
 SurfaceFree AirDelta-T
  • a

    Sig. means p < 0.05 (significant).

  • b

    M.sig. means p < 0.10 (marginally significant).

Mean correlation−0.168−0.151−0.152−0.149−0.105−0.043
Median correlation−0.165−0.137−0.131−0.126−0.132−0.058
Mean (abs value)0.1710.1550.1560.1540.1290.064
Max correlation (+ve)0.0830.0780.0880.0950.2170.143
Min correlation (−ve)−0.547−0.541−0.447−0.455−0.524−0.335
Number sig.,a %65.051.473.161.253.419.2
Number m.sig.,b %4.423.72.616.92.78.9
Sig. +ve, %00002.41.9
Sig. −ve, %65.051.473.
Total −, %95.795.293.893.169.174.3

4.1.1. Temperatures and Cloud Cover

[24] Table 2 summarizes annual correlations between surface, free-air and ΔT temperatures, and total cloud cover anomalies. Most sites show significant correlations (p < 0.05). Mean and median correlations are usually quite similar. As expected, cloud cover is associated with lower surface maxima and higher surface minima at most sites. The nighttime cloud warming influence is absent in the “free-air” temperatures where the mean correlation is weakly negative. This is confirmed by the strong relationships between ΔT and cloud cover anomalies at most stations, negative by day (more cloud causes surface cooling relative to free air) and positive by night (more cloud leads to surface warming relative to free air). Nighttime correlations between surface and ΔT temperatures and cloud cover anomalies are also strong using low clouds only, since low clouds emit more radiation downward (bottom line of Table 2).

[25] Stations near to one another usually show similar correlations, but much variability occurs at larger scales, suggestive of a regional climate effect (Figure 2). Anomalous stations occasionally occur, but many of these can be explained with some knowledge of local climate. For example, stations with weak correlations between surface maximum temperature and total cloud cover occur mostly in northern and western China (e.g., Hoboksar, Baytik-Shan) where incursions of cold air from the Siberian high in winter bring bitterly cold but cloud-free conditions even in the daytime (Figure 2a). On the other hand, the strongest negative relationships (below −0.5) center on 30°N and 105°E (Sichuan and Shaanxi provinces) [see Yu et al., 2004]. The daytime cloud/temperature correlation weakens with increasing latitude (metacorrelation r = 0.451), consistent with a more prominent cloud cooling effect at lower latitudes due to higher mean sun angle and the counter-influence of daytime cooling under clear skies in winter at higher latitudes (polar air masses). Correlations also become slightly stronger at higher elevations, presumably because of stronger radiative forcing and dependence of surface heating on direct radiation.

Figure 2.

Maps of positive (upward pointing triangle) and negative (downward pointing triangle) correlations between anomalies in temperature and total cloud cover. (a) Surface maximum temperatures, (b) surface minimum temperatures, (c) free-air maximum temperatures, (d) free-air minimum temperatures, (e) ΔT maximum, and (f) ΔT minimum. Triangle size is proportional to correlation magnitude.

[26] In the free air both nighttime and daytime cloud/temperature correlations are stronger at the higher-elevation sites in Europe and Southern China (Figures 2c and 2d). Both the day and night ΔT/cloud correlations (Figures 2e and 2f) are also stronger at the higher-elevation sites in southern China, emphasizing the increased importance of cloud cover at high elevations in preventing rapid relative surface cooling through efficient radiation transfer.

[27] The fact that significant relationships between total cloud cover anomalies, surface, free-air and ΔT temperatures occur at most sites increases our confidence in the quality of surface and free-air temperature data sets, especially since the cloud data are independent of both.

4.1.2. Temperatures and Regional Wind Components

[28] Tables 3 and 4 provide summaries of the station correlations between temperature anomalies and reanalysis u and v wind component anomalies, respectively.

[29] The relationships with the u wind component are inconsistent and often weak, especially at night, since the characteristics of westerly advection depend on the upstream surface (i.e., factors such as continentality and relief). The relationships between u, surface, free-air temperatures and ΔT are mapped in Figure 3. Increased westerly advection causes surface and free-air warming at more than half the stations, but there are also stations with strong negative correlations (approximately 25% of sites (Table 3)). Over most of Europe an increase in u is associated with increase in surface maxima, while in Asia positive/negative relationships are concentrated in the south/north of the continent. A broadly similar pattern is seen for surface minima, and for free air temperatures. The big positive correlations between u and surface/free-air temperatures in southern China occur east of the Tibetan Plateau, suggestive of a downslope warming effect. Stronger zonal flow (high u) is associated with a sharpened meridional temperature gradient in China, both at the surface and in the free air. Relationships with ΔT are less consistent.

Figure 3.

(a–f) As in Figure 2 but for correlations between anomalies in temperature and u wind component.

[30] The v wind component represents meridional advection and exhibits consistent correlations with temperature anomalies. Surface and free-air temperatures are nearly always significantly positively correlated with the v wind component (positive defined as poleward flow) during both day and night (Table 4). Although we may expect that relationships between wind components and free-air temperatures will be stronger than those with surface temperatures because of their common source in the reanalysis, free-air temperatures additionally represent smoothed large-scale anomalies that are less influenced by local radiative effects (aspect, exposure, patterns of cloud formation). The relationship between v and ΔT is usually negative during the day (over 60% of stations). Under cold advection in polar air masses (negative v), ΔT is usually high (warm surface in comparison with free air); implying subdued and lagged cooling of the surface. Surface heating similarly lags under warm advection (positive v and low ΔT).

[31] Figure 4 displays spatial maps of the correlations between the v wind component and temperature anomalies. Surface maximum temperatures (Figure 4a) are strongly influenced by v where outbreaks of polar air masses are strong. The few stations with apparently anomalous relationships (negative) include Huajialing, Xifengzhen and Emei in the Sichuan Basin in China. This particular area has a distinct microclimate [Arakawa, 1969; Zhang and Lin, 1992], being protected from incursions of polar air from the north by mountain ranges. Additional downslope warming under northerly winds creates the unusual situation of northerly air masses being warmer than southerly ones at these stations. Other anomalous stations such as Gaziantep in Turkey (37.1°N, 37.4°E, 855 m) and Fraserburg in South Africa (31.9°S, 21.5°E, 1300 m) are prone to downslope winds in equatorward airstreams. The anomalous behavior of Emei (29.5°N, 103.3°E, 3049 m) remains in the map for surface minimum temperatures (Figure 4b), but disappears in the free-air temperature maps (Figures 4c and 4d). The reanalysis behavior is thus decoupled from the surface at this location.

Figure 4.

(a–f) As in Figure 2 but for correlations between anomalies in temperature and v wind component.

[32] The correlations between ΔT and v exhibit much more variability (Figures 4e and 4f). In most areas poleward advection (high positive v) leads to a heat deficit on the mountain surface (western Europe, southern Turkey, most of northern China) but a minority of stations show the opposite relationship (e.g., Arxan, Xi-Ujimqin-Qi).

4.1.3. Temperatures and Local Wind Speed

[33] The relationships between local wind speed and temperatures are variable (Table 5) and usually only significant at about half the stations (maps not shown). Although correlations with surface and free-air temperatures are often weak, in many cases relationships with ΔT are much stronger. The surface often becomes warmer in comparison with the free air under conditions of strong ventilation. Thus stronger winds make an already negative ΔT less negative at night. Some of the explanation for the weaker correlations may be because the local wind speed observation is highly variable with time and possibly unrepresentative of a long time period.

4.1.4. Temperatures and Snow Cover

[34] Table 6 and Figure 5 present correlations for the 64 out of 72 sites that experienced snow cover. At most locations the correlations between snow cover anomalies and temperature anomalies are significantly negative, albeit relatively weak. For example, the weighted mean correlation for surface temperature during daytime is −0.168. Correlations tend to be stronger than this in areas where snow cover is sporadic and hence interannual variability is high (e.g., central North America, south central Asia, eastern China around 40°N). Contrastingly, sites in high northern latitudes experiencing extensive snow cover every winter exhibit weakly negative correlations. Correlations with surface temperature are slightly weaker during nighttime than daytime. During the day, correlations with free-air temperatures are slightly weaker than those with surface temperatures. This means that there is a weak negative relationship between daytime ΔT and snow cover at most locations, again strongest in the zone of greatest snow cover variability (Figures 5e and 5f).

Figure 5.

(a–f) As in Figure 2 but for correlations between anomalies in temperature and snow cover.

4.1.5. General Comments

[35] In conclusion, relationships between temperatures, cloud cover, local wind speed, and snow cover tend to be strongest at the surface, while those with the regional wind components tend to be strongest in the free air. This produces significant relationships between ΔT and cloud, wind, and snow parameters at most stations, the predominant signs of which agree with those expected from consideration of past studies (section 2). However, it is easy to oversimplify our findings. Although significant correlations exist at many stations, the sign of the relationship varies according to location, meaning that the mean absolute correlation value is often higher than the arithmetic mean correlation. Thus any prediction of ΔT must be locally based, and is dependent on local and regional climate.

4.2. Effect of Seasonal Breakdown

[36] Most correlations, especially those with cloud cover (being related to the radiation flux), are seasonally dependent, and subdividing by season often improves mean values. Table 7 shows mean seasonal correlation values for select predictors (total cloud cover, u and v wind components). Correlations for individual seasons are on average higher than ones for all months, but this is not always the case. Cloud correlations are best in summer because of greater insolation, while wind correlations are better in winter because of greater air mass contrasts. The majority of climates represented are strongly seasonal, particularly in Eurasia where the contrast between northerly and southerly monsoon regimes is distinct.

Table 7. Effect of Seasonal Division on Mean Correlation Strength Between Temperature and Total Cloud/Wind Component Anomalies
SurfaceFree AirDelta-T
Total cloud
U wind
V wind

4.3. Effect of Topography

[37] Each of the 72 sites was classified according to topography using a subjective definition derived from investigation of topographical charts. This is a local definition based on the 5–10 km scale. Because we only selected sites higher than the mean of the surrounding reanalysis grid points, mountain summit sites are frequent. The six classes (numbers of sites in parentheses) included 1: summit or ridge (23), 2: slope (7), 3: flat (15), 4: basin (2), 5: incised valley (16), 6: not enough information to classify because of confusion over exact location or poor mapping (9). Analysis of variance was applied to assess whether topographic class influences correlation magnitudes. Significant differences between classes often occurred, particularly for the mountain summit class. Figure 6 displays box plots for correlations between daytime ΔT and total cloud (Figure 6a), u wind component (no significant differences between classes) (Figure 6b), v wind component (Figure 6c), and snow cover (Figure 6d). The mountain summit class shows stronger cloud cover and v wind component correlations than every other known category, but weaker snow cover correlations than other classes. The flat, basin and valley sites show the opposite. Snow cover not surprisingly has more control over temperatures at basin and valley sites where sky-view factors are lower and surface effects will be more important.

Figure 6.

Boxplots showing for each topographic class the distribution of correlations between anomalies in ΔT maximum and (a) total cloud cover, (b) u wind component (no significant differences), (c) v wind component, and (d) snow cover. Boxes span the distance between first and third quartiles (the interquartile range, IQR), and the middle line marks the median. Whiskers extend out to the greatest value less than 1.5 × IQR beyond the first and third quartiles, and points mark outliers.

4.4. Effect of Time Period

[38] There is concern that the introduction of satellite data in 1979 had a nonstationary influence on reanalysis temperatures [Sturaro, 2003]. Thus correlations of monthly temperature anomalies with each parameter were calculated for all sites for the separate periods 1971–1978 (presatellite) and 1979–1996 (satellite). Differences in correlation values between the two periods were small and usually insignificant (nearly always less than 0.1 (figures not shown)). Spatial patterns of correlations were compared for the two periods, and also for the subperiods 1979–1987 and 1988–1996 (both during the satellite era). Interestingly, the spatial correlation patterns for ΔT were more different between the two subperiods within the satellite era than between 1971–1978 and 1979–1996 (Table 8). Thus any reanalysis inhomogeneity has had less impact on correlations than internal changes within the satellite era. Overall persistence in correlation patterns is high, suggesting that most relationships are stationary (temporally consistent). The consequences of this are discussed further in section 5. Relationships with snow were the least temporally consistent, possibly because of limited snow extent in the later period.

Table 8. Metacorrelations Between Spatial Patterns of Meteorological Parameter/Temperature Correlation Magnitudes For Different Time Periods
 SurfaceFree AirDelta-T
1971–1978 (Presatellite) Versus 1979–1996 (Satellite)
Total cloud0.7950.7880.7150.6320.8500.807
Low cloud0.7690.7050.7180.6220.7680.722
U wind0.8210.8570.8380.8370.5860.783
V wind0.7420.8350.6380.7390.7340.696
Wind speed0.6240.4370.5360.0930.7450.561
1979–1987 Versus 1988–1996 (Within Satellite Era)
Total cloud0.6840.6130.7640.6290.7420.569
Low cloud0.6990.5910.8120.6990.6290.599
U wind0.7400.7330.8840.8440.5560.723
V wind0.6730.7510.8800.8890.5730.518
Wind speed0.4320.2520.6990.5350.5450.471

4.5. Stepwise Regression Models

[39] Forward stepwise regression was employed for surface, free-air temperatures, and ΔT (day and night) at each of the 72 stations with six predictor variables: total and low cloud cover, u and v wind components, local wind speed, and snow cover. A significance of p = 0.1 was required for entry into the model (432 models for each season or annual). Table 9 summarizes model performance for the annual case. The first two columns show the mean overall r2 for the best annual model for all stations and for stations above 2000 m (adjusted for spatial autocorrelation). The other columns give the number of times each predictor was considered significant in the final model, listing the number of times the coefficient for that predictor was positive/negative. Coefficient consistency factors show whether the sign of a predictor coefficient is consistent. If this factor is high (bold >90%) then there is confidence that the effect is real (rather than influenced by multicollinearity between predictors).

Table 9. Summary of Forward Stepwise Regression Models to Predict Surface Temperature, Free-Air Temperature and ΔT Anomalies From Six Parameters (Total Cloud, Low Cloud, u and v Wind Components, Local Wind Speed and Snow Cover Anomalies)a
All SitesMean r2Number of Times the Predictor Appears in Models With +/− Coefficient
CloudWindWind SpeedSnow
All SitesElev > 2000 mTotalLow OnlyV WindU Wind
  • a

    Variables have to have p value of less than 0.1 to enter the model. Numbers in the second row are coefficient consistency factors, calculated by (n/t) where n = frequency of most common coefficient sign and t = total number of times the predictor appears. Values above 90% are in bold. Mean r2 values take spatial autocorrelation into account.

Max temp  97.987.296.371.759.197.7
Min temp  51.784.098.251.956.3100.0
Free air0.2830.2644/478/3465/032/165/262/45
Max temp  92.281.0100.066.783.995.7
Free air0.2510.1618/3115/2068/035/188/130/45
Min temp  79.557.1100.066.061.9100.0
ΔT (day)0.2350.3020/374/3611/2529/827/82/29
ΔT (night)0.1670.28028/629/411/2715/2216/93/21

[40] Total cloud cover has a consistent cooling effect by day (for surface, free air and ΔT) and the v wind component and snow cover are consistent for surface and free-air temperatures (but slightly less so for ΔT). Other predictors are often less consistent, especially the u wind component and local wind speed. Although there were strong individual correlations at some stations (>0.7) the mean percentage of temperature variation explained among all stations varies from 16 to 30%. For surface maxima, approximately a third of stations have r2 above 0.4 and the mean r2 is 0.30. The mean r2 is slightly lower for ΔT than for surface or free-air temperatures but higher at sites above 2000 m above sea level.

[41] The spatial variation in annual model r2 is shown in Figures 7a–7f for the six predictands. Surface maxima are most successfully modeled in North America, Europe and southern China. Surface minima are also modeled well in the same areas. The worst areas for free-air models are in northern and western China centered on Xinjiang province during the day (e.g., Balguntay, Yiwu) and also Sichuan province at night (e.g., Daocheng, Emei Shan). In contrast with the surface models, free-air model performance deteriorates with elevation, especially by night (metacorrelation r = −0.368, p = 0.002). ΔT models show different patterns with the worst models being much more widely scattered. The best daytime models are in southern China and Europe (r2 > 0.5). Nighttime ΔTr2 values improve with elevation (metacorrelation r = 0.460, p = 0.0001) and show a different distribution. The best predictions are in Northern and Central China, while Turkey and southeastern China show poorer performance (e.g., Tianmu Shan, Nanyue).

Figure 7.

Maps of variance explained (r2) by stepwise regression model. (a) Surface maximum temperatures, (b) surface minimum temperatures, (c) free-air maximum temperatures, (d) free-air minimum temperatures, (e) ΔT maximum, and (f) ΔT minimum.

[42] Models developed for individual seasons tended to increase r2 (Figure 8), particularly in winter when the wind variables were useful predictors, and in summer when total cloud cover was a useful predictor. On average there is an increase of approximately 0.1 in mean model r2 (above the annual models), and the number of station models with r2 > 0.5 increases dramatically. Winter improvements are especially successful, when variable snow cover and wind components (advection) are both important. Nevertheless in the majority of cases at least half of the temperature variability remains unexplained by meteorology.

Figure 8.

Average fractional variance explained (r2) by models for all months and individual 3-month seasons. Values are corrected for spatial autocorrelation.

5. Observed Trends in Temperatures and Other Parameters

[43] Temporal trends in Ts, Ta and ΔT and their controlling parameters were measured by least squares linear regression. Significance is assessed using the method of Santer et al. [2000], taking temporal autocorrelation into account. 3 stations had too many data gaps to calculate reliable trends and were excluded from further analysis. The original trends themselves are significantly different from those using a wider range of 1084 stations over a longer period [Pepin and Seidel, 2005]. However, the purpose here is not to be globally comprehensive but to examine the relationships between Ts, Ta and ΔT trends and simultaneous changes in cloud, wind and snow, where reliable data are available.

5.1. Surface, Free-Air and ΔT Trends: 1971–1990

[44] Maps of Ts, Ta and ΔT trend magnitudes are shown in Figure 9, and aggregate properties with weighting by station effective sample size are listed in Table 10. Many trends are insignificant (Table 10), particularly those for daytime Ts. Quoted errors for the mean trends in Table 10 are based on a 95% confidence interval. During the day, surface trends are not systematically greater than or less than free-air trends, and negative ΔT trends are just as frequent as positive ΔT trends. The mean trend (all stations) for daytime ΔT is not significantly different from zero. Contrastingly, surface warming at night (Figure 9b) exceeds free-air warming (Figure 9d) at most stations, so the trend in ΔT is often positive (mean of +0.21 ±0.2 °C/decade). Comparison of our Ts and Ta trends with the Northern Hemisphere figures reported by Fu et al. [2004] shows strong similarities on a diurnal mean basis, but further discussion of these figures is not advised since our study is not global, and examines a different time period.

Figure 9.

Maps of original surface, free-air and ΔT trends for the 69 stations. (a) Surface maximum temperatures, (b) surface minimum temperatures, (c) free-air maximum temperatures, (d) free-air minimum temperatures, (e) ΔT maximum, and (f) ΔT minimum.

Table 10. Surface Temperature, Free-Air Temperature and ΔT Trends for 69 Global Stations Using Monthly Anomalies (1971–1990)a
 SurfaceFree AirDelta-T
  • a

    Trends are presented in degC/decade, and mean/median values take spatial autocorrelation into account. Errors for the mean trend are based on a 95% confidence interval.

  • b

    Significant means p < 0.05.

Mean trend0.25 ± 0.160.43 ± 0.250.29 ± 0.120.23 ± 0.12−0.05 ± 0.120.21 ± 0.20
Max trend0.951.270.840.890.520.98
Min trend−0.44−1.03−0.44−0.34−0.68−1.16
U. quartile0.420.750.490.430.160.50
L. quartile0.−0.26−0.04
Number +ve, %
Number −ve, %
Number sig.,b %3.641.327.121.235.753.4
Sig. +ve, %3.041.326.020.614.640.8
Sig. −ve, %0.601.10.621.112.6

5.2. Cloud, Wind and Snow Trends: 1971–1990

[45] Table 11 summarizes trends in nontemperature variables. Cloud cover trends are significant at slightly less than half of sites but trends of both signs are common. There are significant decreases in wind components, local wind speeds, and snow cover at around a quarter of sites.

Table 11. Summary of Trends in Cloud, Wind and Snow Cover Anomalies, 1971–1990, 69 Global Stations (per Decade)a
VariableDaytime Clouds, % cover/decadeNighttime Clouds, % cover/decade 
  • a

    Means and medians take spatial autocorrelation into account.

  • b

    Significant means p < 0.05.

  • c

    Eight stations have no (or extremely minimal) snow cover recorded.

Mean trend1.841.872.012.33 
Median trend−0.21−0.560.390.28 
Number +ve, %40.241.854.552.7 
Number −ve, %59.858.245.547.3 
Number sig.,b %39.839.133.137.8 
Sig. +ve, %17.529.516.926.8 
Sig. −ve, %22.39.616.211.0 
VariableReanalysis Winds, (m/s)/decadeSurface Winds, (m/s)/decadeSnow Anomaly
V WindU WindDay Wind SpeedNight Wind Speed
Mean trend−0.090.02−0.17−0.12−0.020
Median trend−0.050.04−0.12−0.04−0.008
Number +ve, %45.555.534.436.628.1
Number −ve, %54.544.565.663.471.9
Number sig.,b %36.438.326.228.336.7
Sig. +ve, %10.424.
Sig. −ve, %

5.3. Relationships Between Temperature Trends and Trends in Cloud, Wind and Snow

[46] To examine whether the temperature trends in Figure 9 are accounted for by attendant changes in cloud, wind and snow, the original trends (their magnitude and significance) were compared with trends predicted by the stepwise regression model (predicted trends) and trends in the residual differences between actual and predicted anomalies (residual trends). The predicted trends represent what would result if temperature trends were determined solely by meteorological factors. Note that observational uncertainty in the data will reduce the magnitude of calculated regression coefficients, thus leading to potential underestimation of the predicted trends. The residual trend is the component of temperature change unconnected to changes in meteorology, and NOT the difference between original and predicted trends. Because seasonal models were more successful than annual models, fits and residuals were derived for each season separately using the most appropriate model for that season, and then combined into one long time series. The geographical distribution of predicted and residual trends is shown in Figures 10 and 11 respectively (see Figure 9 for the original trends).

Figure 10.

As in Figure 9 but for predicted trends.

Figure 11.

As in Figure 9 but for residual trends. Type A stations are cyan, types B and D are red, type C is blue, types E and G are green, and types F and H are black (see Table 12 and text for definitions).

[47] Eight classes of response can be identified, depending on which combination of original [O], predicted [P] and residual [R] trends is significant (Table 12). Only the first four types of response have actual significant observed trends and the other response types (E–H) are of limited interest for the purposes of this study.

Table 12. Frequencies of Type A–H Stationsa
TypeSignificanceSurfaceFree AirDelta-T
  • a

    The second column shows the combination of trends that is significant. O, original; P, predicted; R, residual. For a detailed definition of categories see the main text.

Total 696969696969

[48] Type A stations have significant original, predicted and residual trends. This means that the observed trend is only partially due to meteorological changes. Type B stations have significant predicted trends but not significant residual trends, meaning that the observed trend can be accounted for by the changes in meteorological factors over the same period. Type C stations have no significant predicted trend and a residual trend remaining. This means that variability in meteorological factors does not account for the observed temperature trend. At type D stations, observed trends are significant, but both predicted and residual trends are insignificant. This is rare.

[49] Note that raw frequencies listed in Table 12 do not represent a global average because of the highly uneven density of stations. Comparison of Figures 9 and 11 indicates that at many sites in North America, Europe and Central Asia (70–100°E), daytime ΔT residual trends are either insignificant (types B, H) or much smaller than original ΔT trends. This suggests that differences between trends in daytime surface and free-air temperatures for these high-elevation regions of the globe can largely be explained by changes in meteorological conditions. In parts of Turkey and eastern China, the negative residual daytime ΔT trends could indicate that the strength of the relationship between temperature and cloud cover has changed over time [Sun et al., 2000]. Large increases in anthropogenic aerosols in these two regions may have radiatively cooled or retarded the warming of the surface relative to the atmosphere [e.g., Tayanç et al., 1997; Luo et al., 2001; Kaiser and Qian, 2002; Che et al., 2005], a factor not taken into account in our model. In northern China there are several type A stations where the meteorology only partly accounts for (often positive) ΔT change. Although aerosol loading in China is small north of 40°N [Luo et al., 2001; Kaiser and Qian, 2002], the cause of our positive residual daytime ΔT trends in this area is unknown. This may have arisen because of predicted trends being underestimated because of observation uncertainty.

[50] Large nighttime ΔT residual trends are present in many regions of the globe. Positive trends predominate, resulting from greater warming of the surface than the free air. Further investigation into the causes of this contrast is required; speculative reasons include land use change and urbanization. Most of these stations are again sited in China. Type A and C stations are more common than type B, demonstrating that at the majority of locations variance due to meteorology cannot (solely) account for the observed nighttime trends. Such significant nighttime ΔT residual trends could indicate that (1) important parameters are missing from the models that could have accounted for the trends, (2) the residual trends are real and occur over and above meteorological forcing due to other forcing factors (e.g., increasing carbon dioxide concentration or aerosol loading or land use change), or (3) the original trends in ΔT must be spurious because they cannot be explained by meteorology, and therefore result from systematic changes in instrumental errors over time.

[51] It is our feeling that, although r2 is modest (0.3–0.4) in most models, the most influential controlling meteorological factors (surface controls, advection and radiation balance) have been included, and so inclusion of further variables is unlikely to improve r2 dramatically. This leaves the latter two interpretations.

6. Summary and Discussion

[52] We have examined the relationships between surface temperatures, free-air temperatures, their difference (ΔT), and meteorological parameters (cloud, wind and snow cover) for 72 exposed high-elevation stations with reliable observations. Significant, but subtly different, relationships exist at the majority of locations, suggesting that meteorological processes are partially responsible for causing differences between observed surface temperatures (GHCN and CRU) and free-air temperatures (reanalysis R1 interpolated to similar elevations). This implies that the surface and free-air observations are measuring physically different quantities and so should be expected to differ (both in instantaneous values and in trends). The cloud and snow data and some of the wind data (local wind values) are independent of both the surface and free-air temperature data sets, adding more weight to this interpretation.

[53] The signs of the relationships between ΔT and meteorological parameters at the majority of sites are as expected from a consideration of the physical processes involved (outlined in section 2). ΔT becomes more negative under cloudy conditions during the day, but more positive at night (e.g., cloud minimizes the difference in temperature between the surface and the free atmosphere). In most locations, warm advection (positive v) makes ΔT more negative because of the lag effect. Relationships with u are more variable, dependent on regional climatic controls. Relationships with snow cover are usually negative since the snow acts as a heat sink at the surface in comparison to the free atmosphere.

[54] Typically a third to a half of the variance in ΔT on a seasonal basis can be explained by variation in these meteorological parameters. The remaining half or more of the variance could be controlled by unknown parameters or result from instrumental error or sampling problems (for example, time differences between reanalysis and surface observations, which, although fairly constant at any one location, could have a variable influence on ΔT).

[55] Trends in surface temperature, free-air temperature, and ΔT were shown at many sites. ΔT trends are often significant but inconsistent by day (meaning that the surface is not warming at the same rate as the free air) but positive during the night (meaning enhanced surface warming relative to the free atmosphere). Once observed cloud, wind and snow changes are taken into account, daytime residual temperature trends are small for most stations outside of Turkey and eastern China. Negative daytime residual trends in Turkey and eastern China may result from increased radiative surface cooling by anthropogenic aerosols. Significant nighttime residual trends are more widespread, and only part of the variance in ΔT can be explained through meteorology at these locations since the contrast between measured free-air and surface temperatures is not supported by attendant changes in cloud, wind and snow.

[56] To ensure that predicted trends were independent from trends in the data used to calculate the model coefficients, a similar methodology was applied after detrending all anomalies (temperatures and other parameters) before running the stepwise regression. The detrending had minimal effect on model r2 or on results in nearly all cases. This confirms that the temperature variance due to meteorology is more or less independent of that due to trends.

[57] A few caveats must be mentioned. The original trends derived using the 69 stations are not necessarily representative of global conditions or longer time periods. Many are in eastern Asia. Indeed they are dissimilar to trends calculated for a larger number of high-elevation stations (1084) over a longer time period (1948–2002) [Pepin and Seidel, 2005] where positive trends in ΔT were predominant. Thus the findings concerning coherence between ΔT trends and controlling parameters should only be extrapolated spatially with extreme care. It would be profitable to expand this type of analysis to longer time periods and more comprehensive data sets, but this requires compatible simultaneous homogenous data for many parameters.

[58] The preceding results indicate that, for many regions, differences in daytime maximum temperature trends measured at high-elevation stations and in the free atmosphere primarily result from changes in cloud cover, wind, and snow. Where this is not the case, there have been large increases in anthropogenic aerosols over the past several decades. Attribution of differences in nighttime minimum temperature trends is uncertain and may indicate the occurrence of unaccounted meteorological changes or presence of inhomogeneities in the data. Since the reanalysis is a mixture of free-air data (satellite and radiosonde), it is not possible to expand these findings here to discuss differences between individual free-air and surface data sets such as LKS [Lanzante et al., 2003] and/or MSU2 [Mears et al., 2003; Vinnikov and Grody, 2003; Fu et al., 2004] versus CRU [Jones and Moberg, 2003]. A similar analysis relating such inter-data-set differences to other atmospheric variables would be a useful, but difficult, exercise mainly because of problems in ensuring temporal and spatial consistency in comparisons. Nevertheless the use of other variables (cloud, wind, surface properties, atmospheric moisture etc) to help examine the validity of inter-data-set temperature differences deserves further consideration.


[59] The majority of this work was undertaken while N. Pepin was a visiting scientist at the NOAA Air Resources Laboratory in Silver Spring, MD. Funding was provided by the National Research Council of the United States of America. J. Norris was supported by an NSF CAREER award, ATM02-38527. The helpful guidance of Dian Seidel and the Climate Variability and Trends Group is appreciated.