Uncertainties in discerning changes
Several measuring systems are available to estimate the temperature variations and trends of the air above the surface, though all contain significant time-varying biases as outlined below.
The longest data sets of upper air temperature are derived from instruments carried aloft by balloons (radiosondes). Changes in balloon instrumentation and data processing over the years have been pervasive, however, resulting in discontinuities in these temperature records (Gaffen, 1994; Parker and Cox, 1995; Parker et al., 1997). Gaffen et al. (2000b) attempted to identify these biases by using statistical tests to determine "change-points" - sudden temperature shifts not likely to be of natural origin (e.g., instrument changes). However, they found that alternative methods for identifying change points yield different trend estimates and that the analysis was hampered by the lack of complete documentation of instrument and data processing changes for many stations. This study, however, only analysed change points in the time-series of individual stations in isolation. Another technique, used successfully with surface data, relies on differences produced from comparisons among several stations in close proximity. In addition, Santer et al. (1999) noted that temperature trends estimated from radiosonde data sets are sensitive to how temperature shifts are dealt with, which stations are utilised, and the method used for areal averaging.
Worldwide temperatures from the Microwave Sounding Unit (MSU) data (Christy et al., 2000) have been available from the beginning of 1979 for intercomparison studies. Parker et al. (1997) used the lower-stratospheric and lower-tropospheric MSU products to adjust monthly radiosonde reports for stations in Australia and New Zealand at times when instrumental or data-processing changes were documented. Some individual stratospheric corrections were as much as 3°C due to radiosonde instrument changes. The main disadvantage of the Parker et al. technique is that the raw MSU record has time-varying biases which must first be estimated and eliminated (Christy et al., 2000).
Gaffen et al. (2000b) compared trends for 1959 to 1995, calculated using linear regression, for twenty-two stations with nearly complete data records at levels between 850 and 30 hPa. Each of these stations is included in two data sets created since the SAR: (a) monthly mean temperatures reported by the weather balloon station operators (Parker et al., 1997; CLIMAT TEMP data) and (b) monthly mean temperatures calculated from archived daily weather balloon releases (Eskridge et al., 1995; CARDS data). Decadal trends at individual sites differed randomly between the two data sets by typically 0.1°C/decade, with the largest differences at highest altitudes. In a few cases the differences were larger and statistically significant at the 1% level. The discrepancies were sometimes traceable to time-of-observation differences of the data used to calculate the averages.
The analysis of trends requires long station data records with minimal missing data. The records for 180 stations in the combined Global Climate Observing System Upper Air Network (GUAN) and the Angell (1988, 2000) network do not generally meet this standard, as only 74 of the GUAN stations, for instance, have at least 85% of tropospheric monthly means available for 1958 to 1998. In the lower stratosphere (up to 30 hPa), only twenty-two stations meet this requirement (Gaffen et al., 2000b). These deficiencies present the dilemma of using either relatively small networks of stations with adequate data (the Southern Hemisphere, in particular, is poorly sampled) or larger networks with poorer quality data (adding uncertainty to the resulting trend estimates).
Characteristics, such as spatial coverage, of each data set derived from the weather balloon data are different. For example, Sterin (1999) used data from over 800 stations from the CARDS and telecommunicated data sets, with only gross spatial and temporal consistency checks. The data were objectively interpolated to all unobserved regions, introducing extra uncertainty. Parker et al. (1997) placed CLIMAT TEMP data into 5° latitude x 10° longitude grid boxes from about 400 sites, leaving unobserved boxes missing. Further data sets were created employing limited spatial interpolation and bias-adjustments, but uncertainties related to spatial under-sampling remain (Hurrell et al., 2000). Angell (1988) placed observations from 63 stations into seven broad latitudinal bands, calculated the simple average for each band and produced global, hemispheric and zonal mean anomalies.
Radiosondes measure temperatures at discrete levels, but satellite instruments observe the intensity of radiation from deep atmospheric layers. The advantage of satellites is the essentially uniform, global, coverage. The three temperature products that are commonly available from MSU are: the low to mid-troposphere (MSU 2LT, surface to about 8 km), mid-troposphere (MSU 2, surface to about 18 km, hence including some stratospheric emissions) and the lower stratosphere (MSU 4, 15 to 23 km, hence including some tropical tropospheric emissions) (Christy et al., 2000). No other data, such as from radiosondes, are used to construct these MSU data sets. It is important to note that the troposphere and stratosphere are two distinct layers of the atmosphere with substantially different temperature variations and trends. The altitude of the troposphere/stratosphere boundary varies with latitude, being about 16 to 17 km in the tropics but only 8 to 10 km at high latitudes.
Since the SAR, several issues have emerged regarding MSU temperatures. Mo (1995) reported that for one of the longest-lived satellites (NOAA-12, 1991 to 1998) the non-linear calibration coefficients were erroneous, affecting MSU 2 and MSU 2LT. Wentz and Schabel (1998) discovered that satellite orbit decay introduces gradual, spurious cooling in MSU 2LT. Christy et al. (1998, 2000) found that instrument responses often differ between the laboratory assessments and on-orbit performance, requiring further corrections. Additional adjustments were also made by recalculating and removing spurious temperature trends due to diurnal effects induced by the east-west drift of the spacecraft (Christy et al., 2000). The magnitude of the spurious trends (1979 to 1998) removed from version D compared to version C were: orbit decay, -0.11; instrument response, +0.04 and diurnal drift, +0.03°C/decade.
Figure 2.12: (a) Seasonal anomalies of global average temperature (°C), 1958 to 2000, relative to 1979 to 1990 for the lower troposphere, as observed from satellites (MSU 2LT) and balloons (UKMO 2LT), and for the surface (adapted from Jones et al., 2001). Also shown (bottom graph) are the differences between the surface temperature anomalies and the averages of the satellite and balloon-based observations of the lower-tropo-spheric temperature anomalies. (b) As (a) but for the temperature of the lower stratosphere, as observed from satellites (MSU 4 and SSU 15X) and balloons (UKMO 4). The times of the major explosive eruptions of the Agung, El Chichon and Mt. Pinatubo volcanoes are marked. Also shown (bottom graph) are the differences between the MSU 4 and UKMO 4 based temperature anomalies.
Version D of the MSU data is used in Figure 2.12. The SAR presented version B that for the low to mid-troposphere indicated a global trend about 0.05°C/decade more negative than version D (for 1979 to 1995). Quite separately, Prabhakara et al. (1998) generated a version of MSU 2 without corrections for satellite drift or instrument body effects, in many ways similar to MSU 2 version A of Spencer and Christy (1992a, 1992b).
The Stratospheric Sounding Unit (SSU) detects the intensity of thermal emissions and measures deep layer temperatures at altitudes above 20 km (Nash and Forrester, 1986). As with the MSU products, adjustments are required for radiometer biases, diurnal sampling and orbital drift (Chanin and Ramaswamy, 1999).
Rocketsondes and lidar
Data sets generated from rocketsondes have been updated (Golitsyn et al., 1996; Lysenko et al., 1997), providing temperature information to as high as 75 km. Important difficulties arise with these data due to different types of instrumentation, tidal cycles (amplitude 2°C) and to assumed corrections for aerodynamic heating. The last set of adjustments has the most significant impact on trends. The approximately 11-year solar cycle forces a temperature perturbation of >1°C in the mid- to upper stratosphere (30 to 50 km). Keckhut et al. (1999) and Dunkerton et al. (1998) created a quality-controlled data set of these measurements, which is used in Chanin and Ramaswamy (1999). The very limited number of launch sites leads to some uncertainty in deduced temperatures, and most launches were terminated in the mid-1990s.
Rayleigh lidar measurements began in 1979 at the Haute Provence Observatory in southern France and have spread to locations around the world. Lidar techniques generate the vertical profile of temperature from 30 to 90 km, providing absolute temperatures within 2.5°C accuracy. Chanin and Ramaswamy (1999) have combined MSU 4, SSU, radiosonde, lidar and rocketsonde data to estimate 5-km thick layer temperature variations for altitudes of 15 to 50 km, generally limited to the Northern Hemisphere mid-latitudes.
The principle of reanalysis is to use observations in the data assimilation scheme of a fixed global weather forecasting model to create a dynamically consistent set of historical atmospheric analyses (Kalnay et al., 1996). Within the assimilation scheme, potentially errant data are amended or excluded using comparisons with neighbours and/or calculated conditions. However, small, time-dependent biases in the observations, of magnitudes important for climate change, are virtually impossible to detect in the model, even in areas of adequate in situ data. Furthermore, in areas with few in situ data the reanalyses are often affected by inadequate model physics or satellite data for which time-varying biases have not been removed.
Though interannual variability is reproduced well, known discontinuities in reanalysed data sets indicate that further research is required to reduce time-dependent errors to a level suitable for climate change studies (Basist and Chelliah, 1997; Hurrell and Trenberth, 1998; Santer et al., 1999, 2000; Stendel et al., 2000). It is anticipated that future assessments of climate change will utilise reanalysis products to which substantial improvements will have been made. Data from the NCEP reanalysis are included below for comparison purposes, but longer-term stratospheric trends from NCEP are especially suspect due to a large shift in temperature when satellite data were incorporated for the first time in 1978 (Santer et al., 1999).
Other reports in this collection