Climate Change & Tropospheric Temperature Trends

What then, can be made of all this? How do the satellite and radiosonde trends relate to each other? Are they truly measuring the same quantities? If so, how do these datasets relate to the climatic variability that has been observed in the upper atmosphere in the last 25 years? For that matter, are the last 25 years representative of the century to come, and do they allow us to either detect or refute an anthropogenic impact on climate change? These are the questions that must be addressed with these records as they currently exist.

To address the length of the record, we must examine the last 25 years in light of the most significant climatic events that have impacted it. During this period, there have been at least two major volcanic eruptions that are known to have significantly reduced solar radiative heating of the troposphere causing cooling over 3 to 4 year periods (El Chicon in 1982 and Mt. Pinatubo in 1991). There were also 5 ENSO (El Nino Southern Oscillation) events and 4 La Nina events which resulted in warming and cooling, respectively. At least two of the ENSO events (1982-83 and 1997-89) and one La Nina event (1988-89) were particularly strong. Figures 20 and 21 show global temperature anomalies for the middle troposphere and lower stratosphere respectively, from 3 MSU/AMSU products and 2 radiosonde datasets (Seidel et al., 2003). The MSU/AMSU curves represent UAH Versions D and 5.0 (Christy et al., 2000; 2003) and RSS (Mears et al., 2003). The sonde curves are for HadRT2.1 and HadRT2.1s for the troposphere and stratosphere respectively (Parker et al., 1997), and LKS (Lanzante et al., 2003). In both figures, the bottom curve gives the average of all 5 datasets and the curves for each dataset above it give deviations from this average. Figure 22 shows 5 globally averaged and one tropically averaged time series from these datasets compared with the Quasi-Biennial Oscillation Index (QBI) and the Southern Oscillation Index (SOI), also from Seidel et al. (2003). In all 5 datasets the short-term variations in temperature anomalies (1-2 years) are larger than the resulting decadal trends and none of the trends shown are monotonic. The signatures of the El Chicon and Pinatubo eruptions are clearly visible as stratosphere warming and troposphere cooling in the years following each, as are the larger El Nino events since 1980. Also visible are the impacts of the QBI and SOI oscillations.

There is significant geographical variance as well, with fluctuations in the tropics being larger than the corresponding global trends particularly during the late 80’s and the late 90’s. With this much variance in the data over time scales representing significant portions of the record, the measurement of long-term trends becomes sensitive to how the trends are computed, and it is not clear how meaningful these trends are. This is particularly true of the radiosonde record, which is far more sensitive to variations in metadata that the satellite record and has suffered extensively from poorly kept station records. As was noted above, a shift of one level in a dataset due to corrections for metadata detected events can increase by as much as 50 percent the amount of time necessary to accurately derive a trend (Weatherhead et al., 1998).

The derived trends from MSU/AMSU also show much variation. Figure 8 shows the calculated tropospheric temperature trends as determined by the latest analyses of each investigating team. The values shown for RSS reflect their first version (Mears et al., 2003) and their most up-to-date figures as of this writing (Oct. 2004). UAH figures reflect Version 5.0 (Christy et al., 2003) and also their latest figures. Despite years of uncompromising dedication - in the case of the RSS and UAH teams, over 8 and 15 years of effort respectively - one is struck by the divergence in these trends. UAH observes little if any lower and middle tropospheric warming, but VG observe warming that equals, and in some places even exceeds that of the well documented surface warming. A closer look at these results reveals much about these differences impacted the measured trends.

Figure 23 shows the global time series of Channel 2 brightness temperature for the UAH and RSS teams, and Figure 24 shows their hot target temperature factors. The constant value of 0.03 used by Prabhakara’s team to characterize all sources of error (except diurnal drift and orbital decay) including IBE is shown for comparison. RSS target factors were derived using ocean-only data and are shown with and without the diurnal corrections derived this way. The UAH target factors were derived from land and ocean data. An examination of this data makes several things readily apparent.

• The diurnally corrected and uncorrected RSS target factors differ only slightly from each other, the largest differences being that for TIROS-N. This demonstrates the relatively weak contribution of oceanic diurnal cycles to target factor and the RSS method of using ocean-only data for target factor calculations.
• Comparison of each team’s target factors shows that despite the differences in their merging methods, there is generally good agreement among them. The one obvious exception is NOAA-09, where the UAH team obtains a value that is much larger than anyone else’s. The RSS value for NOAA-09 agrees well not only with their target factors for other satellites. It also agrees with the other UAH team values, and these are in turn in good agreement with the corresponding RSS values. Prabhakara’s generalized constant target factor is generally similar. Given that all the JPL built MSU packages were nearly identical at launch time, this is not surprising. The only value that is a clear outlier is the UAH determined NOAA-09 factor.
• The UAH and RSS global time series’ are nearly identical at the beginning of the record, but begin to diverge from each other at a more or less constant rate during 1986, when NOAA-09 was in service and very near to the NOAA-09/NOAA-10 service life overlap. The nearly step function change in trend variation that happened here appears to account for most of the difference between the global decadal trends calculated by the UAH and RSS teams.

It is noteworthy that the divergence between the UAH and RSS time series’ begins during 1986, at or very near the NOAA-09/NOAA-10 overlap. This implies that the overall trend is very sensitive to how this overlap is handled. As part of their analysis, the RSS team extracted the residuals (remainders) left over for each dataset after they were combined to fit a final time series curve. The covariance matrix generated for the standard deviations of these residuals during merge calculations can be used to derive an estimate of the statistical uncertainties in each satellite’s dataset. Using a Monte-Carlo method, they superposed noise onto the after-the-fit residuals from their Ocean-Only merged time series and generated an ensemble of 30,000 sets of “noisy” deviations from the covariance matrix in 17 merging parameters and added them to the intersatellite differences from the fitted trend. These were then used to create a set of 30,000 “noisy” merging calculations that were used to derive the error estimates in the resultant trends. Performing the same analysis on the temperature differences between all pairs of co-orbiting satellites during overlaps in service, they were able to characterize the sensitivity of the overall trends to each service life overlap.

Figure 25 shows a graphical representation of the results. It can be seen that compared to the others, the NOAA-09 offset from the fitted trend, and its target factor, are both poorly constrained and largely dependent on each other, so that errors in one will contribute directly to errors in the other. It can also be seen that the NOAA_07/NOAA-09, NOAA_09/NOAA-09, and NOAA-09/NOAA-10 overlap overlaps have significantly more uncertainty than the others, and the overall decadal trend is particularly sensitive to how the NOAA-09/NOAA-10 overlap is handled. This can be demonstrated further by using the UAH target factors in the RSS ocean-only merging analysis. When the RSS team did this they obtained a 1979-2001 global MSU2 trend of 0.014 deg K/decade for the first case, which is much closer to the corresponding UAH value of -0.011 deg K/decade than to their measured trend of 0.099 deg K/decade. Using their target factors and substituting only the UAH NOAA-09 factor gives a trend of 0.022 deg K/decade, showing that the large majority of the RSS-UAH difference is accounted for by the NOAA-09 factor only (Mears et al., 2003c). Likewise, the RSS team also applied the UAH merging method and diurnal correction to their own pentad averaged datasets and obtained a NOAA-09 target factor of 0.075, which compares more favorably to the UAH value of 0.095 than to all other factors determined by both teams and Prabhakara’s team (Mears et al., 2003).

Thus, the overall trend differences between UAH and RSS can be attributed to how each team smoothed their data and conducted their dataset merge and target factor derivations, particularly for the NOAA-09 target factor. The inclusion or omission of NOAA-09/NOAA-10 overlap and/or the other NOAA-09 overlaps, allows for a great deal of variability in outcome and makes it possible to minimize either the trend or the uncertainty depending on what is chosen or ignored. This introduces a degree of arbitrariness into the UAH analysis that renders their results questionable. By including all of the datasets with appropriate weighting for short and/or noisy overlaps, as the RSS team did, this arbitrariness can be avoided while still accounting for the related uncertainties. It is significant that by neglecting several overlaps, including NOAA-09/NOAA-10, UAH obtained a NOAA-09 target factor that is anomalously high not only compared to all other target factors found by other teams, but also to their own values for all other target factors. In light of this, it now appears that the UAH team selected their overlaps and merging methods to minimize the global trend in tropospheric temperatures rather than the overall uncertainty. Their stated objective was to minimize the intersatellite trend differences, and this they largely succeeded in doing (Christy et al., 2004). But by selectively omitting some satellite overlaps, they have obtained a NOAA-09 target factor that is inconsistent with their own results for other satellites and the target factors derived by other teams. It must be noted that the hot target calibration factor is a property of the hot target itself and not the scene viewed or the data gathered. As such, a large factor is indicative of an issue with the hot target itself. The hot targets carried by MSU and AMSU packages on POES satellites are little more than simple blackbody emitters regulated by simple electronic devices and monitored with two PRT thermocouples. There will always be at least some variability in the manufacturing of such devices, as there would in the manufacture of any product. But a variance of 300 percent for one of these devices as compared with all other such devices in service is unconvincing. Add to this the noisiness associated with the nadir/off-nadir differencing derivation of their composite TLT numbers for the lower troposphere, and the associated pollution of these numbers with data from the surface, and it is not surprising that many researchers today question the UAH lower and middle troposphere temperature trends. In their favor, UAH trends to compare favorably with some radiosonde datasets, a fact that continues to give credibility to their derived trends. But the issues with radiosonde analyses limit their viability as an independent check and cannot be used to conclusively validate UAH in light of the other problems. There is also the issue of Antarctic sea-ice and melt-pool area impacts on UAH 2LT and TLT products. Because these factors will not affect the products of other teams who have based their results on Channel 2 alone, it is not surprising that UAH lower troposphere products show smaller trends for this layer than would be expected from Channel 2 and surface temperatures alone (Swanson, 2003), particularly as there is independent reason to believe that many regions of the Antarctic have shown surface and lower troposphere warming in recent decades (Vaughan et al., 2001; Thompson and Solomon, 2002).

It remains to consider Vinnikov and Grody (2003). As noted, their analysis avoids much of the noise associated with other methods, but leaves us having to account for IBE, discrepancies between land and oceanic diurnal cycles, and the way their data is zonally averaged between satellite tracks. Given that they use pentad smoothed datasets, as did the RSS team, and only low noise nadir views, it is reasonable to estimate the impact each of these effects has on their analysis by comparing them with the corresponding RSS estimates. Figure 26 shows RSS Version 1.0 and UAH Version 5.0 decadal trends for land and ocean data, with and without diurnal corrections. It can be seen that the RSS diurnal correction contributes about 0.007 deg K/decade to the oceanic trend and 0.064 deg K/decade to the land trend. It is reasonable to conclude that depending on how the VG diurnal cycle is zonally averaged, their numbers would be corrected by up to half of this amount, or around 0.03 deg K/decade. Interestingly, VG observed a trend in their diurnal maxima and minima that contradicts what has been observed (Vinnikov and Grody, 2003). Their analysis gives a diurnal cycle of the warming trend that yields a daytime maximum and a nighttime minimum, indicating that the diurnal cycle is increasing in amplitude, which is not observed to be the case (Karl et al., 1993). When VG correct for this in their analysis, they get a downward shift of 0.04 deg K/decade, not unlike this 0.03 deg K/decade interpolation, yielding their lower bound of 0.22 deg K/decade. Strictly speaking, the potential error they addressed in warming trend diurnal cycle is a correction of temporal phase, not geographical location. If VG’s numbers were corrected for both diurnal effects, it would be reasonable to expect them to drop to something like 0.22 – 0.03, or 0.19 deg K/decade. For IBE, VG rolled all error estimates into a single linear correction which could be expected to absorb part, but not all of the IBE accounted for by the target factors in the UAH and RSS analyses. Since IBE is determined by the product of hot target factor and target temperature variation, we can estimate the impact of it by considering the observed variations in target temperature for selected MSU packages. Figure 27 shows MSU2 monthly hot target temperatures for a few MSU packages. Inspection of these variations reveals that while the amplitude of variation is in some cases quite large (e.g. NOAA-12), the normalized average fluctuations are on the order of 2 to 7 deg K/decade with the average somewhere in between. The UAH and RSS target factors shown in Figure 24 average to around 0.018, neglecting the UAH NOAA-09 factor. From these values we can make a very rough estimate of the impact of IBE corrections to the VG trend as being on the order of -0.036 to -0.126 deg. K/decade at the very largest. With their Version D release, UAH produced a more thorough estimate of IBE impacts by estimating the dependence of the MSU calibrated antenna digital count errors with variations in Hot Target temperature (Christy et al., 2000). They obtained an estimated impact of 0.026 deg. K impact per deg. K temperature variation in the Hot Target. Based on these estimates, it appears that if Vinnikov and Grody’s zonal averages of observed temperatures between satellite paths are reasonable, correcting their analysis to include IBE and both land and ocean diurnal cycles will likely cause their numbers to converge with those of the RSS or Prabhakara teams rather than the UAH team.

Comparisons of Radiosonde Products

Given the issues with MSU/AMSU analysis products, several teams have produced comparison studies of the two. Some have sought independent validation of MSU results (Christy et al., 2000; 2003). Others have sought to refine the reliability of radiosonde analyses using MSU products as an independent check, or to clarify the correlations and interdependencies between sonde and MSU products (Angell, 2000; 2003; Seidel et al., 2003). These studies have yielded differing results. Some found no differences in trend between the surface and tropospheric temperature records since 1958 (Angell, 2000). Others indicate that the troposphere has warmed relative to the surface before 1978 and then cooled with respect to it thereafter (Gaffen et al., 2000a). For many of these results, variance in the data is similar in magnitude to the trends being measured, particularly in regions such as the tropics and high southern latitudes that are particularly important for comparisons with different MSU/AMSU analysis products. This variance in results has impacted attempts to validate MSU/AMSU analysis products using sonde analyses.

The UAH team in particular has compared their results with selected networks of radiosonde sites as part of their Version D and 5.0 releases, and more recently, a lower troposphere intercomparison with their Version 5.1 release (Christy and Norris, 2004). In Version D (Christy et al., 2000) UAH derived a time series for the 97 radiosonde stations they used covering North America, Bermuda, Iceland, and the Western Pacific. These time series were then compared to Channel 2LT data from UAH’s 2.5 degree monthly maps at the same locations. Care was taken to construct the selected sonde network prior to evaluation of the corresponding MSU data points to insure that the two were independent. Direct radiosonde records, which give a vertical temperature profile with specific readings being taken at regular height intervals as the device ascends (once for every 5 meter rise is common), were converted into their corresponding weighted averages as they would have been read by the MSU with a constant surface emissivity being assumed (the MSU would directly detect averaged fluctuations of actual surface emissivity as it influenced the 2LT data). Good agreement was found between the two. Both datasets agreed on lower tropospheric trend to within 0.005 deg. K/decade with an annual correlation of 0.97. UAH notes however that the decadal trend observed for these sites was +0.16 deg. K/decade – considerably higher than the Version D global trend of -0.01 deg. K/decade demonstrating that the sites chosen, though noteworthy for their reliability as sonde sites go, were not representative of the whole earth. Seeking sonde data with more global coverage, they also compared their Version C and D datasets with Angell 63 and HadRT2.0 analyses (Angell, 1988; Parker, et. al., 1997) after checking the data for internal consistency and binning it into a global grid to facilitate comparison with their own globally gridded results. Figure 28 shows how the 4 datasets compared. It can be seen that the overall agreement is fairly good, though there are noticeable differences in the 1984-85 and 1990-95 timeframes and lower correlations with MSU (0.91 for HadRT2.0 and 0.93 for Angell). It should also be noted that the Angell 63 network predates Angell 54 and has not been corrected for the 9 stations known to have anomalous records (Angell, 2003), and the HadRT2.0 dataset has not been corrected for anomalous record changes either, as were later HadRT products such as HadRT2.1.

For Version 5.0 (Christy et al., 2003), UAH generated 3 composite simulated radiosonde temperatures designated as RLT, RMT, and RSL (to facilitate comparison with their TLT, TMT, and TLS Channels, respectively) and compared them with MSU results on a range of scales. Analysis was done using CARDS data for a single station and a small network of 28 stations chosen for their internal consistency of method and equipment, and an additional comparison was made using a more global network with fewer controls. For their single station comparison, Minqin, China (38.6 deg. N, 103.1 deg. E, 1367 m elevation) was used. Minqin was chosen for its unique consistency in equipment, methods, and continuity of record. The 28 stations listed in their appendix, were chosen to optimize the same levels of consistency to the greatest degree possible. Excellent agreement was found for Minqin with a monthly correlation between the two datasets of 0.90 for the lower troposphere and 0.95 for the middle troposphere, monthly standard deviation differences of less than 0.53 deg. K, and only 0.01 deg. K/decade difference in overall trend. Good agreement was also found for their 28 station network, with the corresponding figures being 0.94 and 0.93 for monthly correlation, 0.15 and 0.12 deg. K for monthly standard deviation variance, and an overall trend difference of -0.05 deg. K/decade.

For their larger global analyses, UAH compared TLT data with RLT data generated from 4 other sonde analysis products – HadRT2.0, Angell 63 (as with Version D), RIHMI, and a temperature reanalysis product from the National Centers for Environmental prediction (NCEP) generated from in situ sonde and satellite observations (other than MSU) ⁸. Comparison was also made with results from a general circulation climate model (Kalnay et al., 1996). In each case, temperature vs. altitude data for a vertical column of 850 to 300 hPa altitudes (corresponding to the lower troposphere) was assigned a weighting similar to that observed by MSU devices to generate an RLT simulated MSU measurement to facilitate comparison with TLT data. Good agreement was found for all products, with trend differences of less than 0.02 deg. K/decade for all products. Figure 29 shows time series of annual lower troposphere temperature anomalies from UAH’s Version 5.0 TLT data and the corresponding products from NCEP and HadRT. NCEP is a reanalysis product that uses data from in situ sources, satellite data (other than MSU, but using NOAA sounding profiles similar to those used in MSU products), and results from a global circulation model (Kalnay, 1996). HadRT data for this comparison was taken from HadRT 2.1 which corrects for discrete temperature changes using comparisons with MSU Ver. D data, and thus is not entirely independent of MSU results. Likewise, NCEP data is dependent to some extent on radiosonde products and is not entirely independent from them either.

Figure 30 gives a summary of 95 percent confidence intervals for monthly anomalies between their TLT (lower troposphere) and TMT (middle troposphere) data, and their Minqin, 28-station network, HadRT, and NCEP analyses. Despite problems with instrument and method consistency for the larger scale analyses and gaps in data (Christy et al., 2003) there is generally good agreement between all datasets. Since the publication of UAH Ver. D, the Angel 63 product has been updated to Angell 54 as described above. Removal of the anomalous 9 stations resulted in a significant warming of the 1958-2000 record compared to the original up to an altitude of 100 hPa, particularly in the tropics. The global changes in trend are smaller, but also show an increase. Most of the change however occurs before 1979 and the beginning of the MSU record. There is general consensus from this data that globally, the troposphere and the surface have warmed at roughly the same rate since 1958, but that prior to 1979 the troposphere warmed more than the surface, and since 1979 has cooled with respect to it. Removal of the anomalous sonde stations improves the agreement between most sonde datasets and the UAH Version D MSU record (Angell, 2003). Christy and Norris’ Year 2004 intercomparison study also found good agreement between the MSUTLT record and their 89 station radiosonde record, with the global trend differences generally falling within +/- 0.04 K/decade of each other depending on the subsample chosen and whether the data had been adjusted for discontinuities or not (Christy and Norris, 2004). They also found consistently larger differences between their Version 5.1 MSUTMT product and the corresponding product from RSS, though it is not at all clear that RSS-UAH discrepancies can shed any light on the MSUTLT record (for which RSS has no comparable product).

Independent radiosonde analysis products have yielded varying results depending on the methods and networks used. Globally, Brown et. al. (2000) find that since 1958 the surface to lower troposphere has warmed by about 0.20 deg. K/decade. Lanzante et al. (2003) find a warming of around 0.15 deg. K/decade between 1958 and 1997 for the same layers using datasets adjusted via the LKS methodology. Gaffen et al. (2000a) find a warming of 0.08 deg. K/decade. With the update of Angell 63 to Angell 54, Angell (2003) finds 850-300 hPa tropospheric warming equal to that of the surface and a tropical warming of 0.13 deg. K/decade for 1958 to 2000. Figures 31 and 32 respectively show 1958-1997 temperature trends for 3 atmospheric layers from 5 radiosonde analysis products, and similar 1979-2001 trends for 3 MSU products and one radiosonde product. Each figure also shows a typical confidence interval that generally represents each of these datasets (Seidel et al., 2003). The HadRT trends represented here are from HadRT2.1s. For the 850 – 300 hPa layer, Angell 54 yields 0.10 deg. K/decade while HadRT and LKS yield 0.08 and 0.125 deg. K/decade respectively. RIHMI yields about 0.45 deg. K/decade, and with respect to the typical 2s confidence interval shown, is a clear outlier with respect to the other sonde products.

Over this longer period, these results are more or less consistent with the predictions of many AOGCM’s and traditional theories of global atmospheric vertical energy and temperature transport. But the shorter periods of 1958-1979 and 1979-2000 (the non-satellite and satellite era’s respectively) reveal a complex evolution of regionally and global lapse rates. For 1979-2000, Gaffen et al. (2000a) and Brown et al. (2000) find that in the tropics, the lower troposphere warmed with respect to the surface between 1958 and 1978 (just prior to the beginning of the MSU record) and then cooled with respect to it afterward. Angell (2003) finds that between 1958 and 1979 the global 850-300 hPa layer warmed with respect to the surface by nearly 0.13 deg. K/decade, and then cooled with respect to it between 1979 and 2000. Brown et al. (2000) and Hegerl and Wallace (2002) observed similar trends. For the shorter period of 1979-1997, LKS shows a global trend of 0.07 deg. K/decade for the 850-300 hPa layer, and -0.05 deg. K/decade for the 1015-10 hPa layer weighted to simulate MSU2 (lower and middle troposphere), and HadRT2.0 yields 0.09 deg. K/decade and -0.11 deg. K/decade respectively. For weightings simulating MSU2LT, the corresponding figures from these products are 0.075 deg. K/decade for LKS and -0.01 deg. K/decade for HadRT2.0. Extending the satellite era record from 1997 to 2001 yields good regional and global agreement between UAH and HadRT2.0 for MSU2LT, but considerable disagreement for MSU2, where the HadRT2.0 data appears to diverge from both UAH and RSS both globally and in the northern hemisphere and tropics (Seidel et al., 2003). The LKS dataset does not extend past 1997.

Global stratospheric trends have been less variable and are characterized by better agreement between radiosonde and MSU products. While stratospheric trends are not directly related to tropospheric temperatures and considerations of anthropogenic greenhouse warming, they are of indirect importance because of their impact on MSU2 data. Strong stratospheric cooling trends (due to ozone depletion) will introduce some spurious cooling into MSU2 trends, which are otherwise generally considered to be synonymous with the lower and middle troposphere. For the period of 1958-1997, Angell finds cooling trends of -0.48 deg. K/decade and -0.41 deg. K/decade for the globe and the tropics, respectively. HadRT shows comparable cooling of -0.39 deg. K/decade globally and -0.37 deg. K/decade, and LKS finds cooling of -0.38 and -0.36 deg. K/decade respectively for the same regions (Seidel et al., 2003). RIHMI shows far smaller cooling trends than the other sonde products with global and tropical trends of -0.20 deg. K/decade and -0.21 deg. K/decade respectively. Once again, RIHMI is an outlier compared to the other sonde products. Similar figures are obtained for the satellite era, indicating that during the last half century stratospheric lapse rates have been less variable than their tropospheric counterparts. Regional data reveal some interesting observations. Angell 54 finds much more long-term cooling for the southern hemisphere than other products (-0.79 deg. K/decade vs. -0.42 deg. K/decade for HadRT2.0, the next closest result). For 1979-1997, all analysis products are more regionally self-consistent, but LKS shows much more cooling than the other products. In all cases, the radiosonde products appear to show more cooling than MSU. This might be due to the fact that MSU4 is fairly strong at 100 hPa which is quite near the tropopause, a region characterized by high lapse rates and rapid transitions from warming to cooling, particularly in the tropics. In addition, some sonde datasets such as Angell 54 show more cooling than others (e.g. LKS) due to differences in southern hemisphere coverage that will cause some datasets to selectively emphasize the south pole ozone hole more than others (Angell, 2003). This may also explain part of the increased cooling observed between sonde datasets and MSU. It is also known that the shift toward Vaisala radiosondes, and the further evolution of these packages during the 1990’s (e.g. the evolution of designs from the RS11 to RS90 Series of devices) has also introduced a spurious apparent cooling, particularly in the stratospheric record. Differences in how this evolution was corrected for may result in differences in apparent cooling trends. For the 300-100 hPa layer, which globally straddles the tropopause, there is considerable disagreement on the 1958-1997 trend between all products at statistically significant levels, with some products disagreeing even on the sign of this trend (Seidel et al., 2003).

Though there are statistically significant differences between various radiosonde and satellite products, confidence intervals are quite large, particularly for radiosonde simulated MSU2 and MSU2LT data in the southern hemisphere and tropics. In many cases the trends show 2-s confidence intervals that are larger than the trends being measured. Stratospheric confidence intervals tend to be much larger than their tropospheric counterparts, reflecting the many problems associated with incomplete datasets due to failed equipment at higher altitudes, particularly the bursting of inferior balloons at random stratospheric altitudes before datasets could be completed. In general, there is much more confidence in data from the northern hemisphere that the southern hemisphere and tropics, reflecting the better quality of data and coverage from that region Seidel et al., 2003). But all current radiosonde datasets agree that globally, over the longer term (1958 to 2000) the surface and 850-300 hPa layers have warmed at comparable rates, but since 1979 the surface has warmed relative to the 850-300 hPa layer with the estimates ranging from 0.04 to 0.14 deg. K/decade for the various datasets (Angell, 2003). There appears to have been a step function change in lapse rate around 1976-77 that is not yet fully understood. Thus, the long-term trend in tropospheric temperatures are consistent with the predictions of the best extant AOGCM’s, but over shorter intervals these temperatures have evolved in more complex ways. It is not at all clear that trend estimates for the last 25 years can be either generalized to model predictions or extrapolated into the future.

So how do radiosonde and MSU analysis products compare with each other? Are they truly complementary, and does one vindicate the other as so many people hope? Perhaps, but with so many factors influencing both, and given the complexities of how the upper atmosphere has evolved in the last 50 years, the devil is in the details. It must be remembered that while the various tropospheric analyses from MSU differ in their resulting trends, these datasets are truly global in their coverage while their radiosonde counterparts are not. Regional temperature trends can vary widely, despite the superficial appearance of global agreement. Similar global trends often mask large variations and/or consistencies in regional trends. Sonde analyses will be useful in MSU comparison studies only if they have adequate coverage and confidence levels in regions where MSU trends from different teams disagree. Figure 11A shows 1979-2001 MSU global decadal trends as determined by the RSS team (a) the UAH team (b), and the difference between the two (c) from Mears et al. (2003). Figure 11B shows a similar comparison for 1979-2002 trends (Mears et al., 2003b). It can be seen that the two analyses are in relatively good agreement with each other over most of the northern hemisphere. The most significant differences between them are in the tropics, northern Africa, parts of Siberia, and in the high southern latitudes (Mears et al., 2003). It is these regions that discriminate between RSS and UAH global trends. Comparing Figures 11A(c) and 11B(c) with the regions covered by the networks used in various radiosonde analysis products reveals that the bulk of the sonde coverage is elsewhere. For instance, the Minqin sonde site (38.6 deg. N, 103.1 deg. E, 1367 m elevation) used in UAH Version 5.0 for their single station comparison, is at a location where there is very little difference between the UAH and RSS trends (Christy et al., 2003; Mears et al., 2003). The Angell 54 station network has only 6 stations representing both the south temperate and south polar zones where many of the largest differences are. Several of the southern hemisphere sites are in a mid-latitude band where UAH and RSS again largely agree (e.g. Hedland in Australia, Amsterdam in the Indian Ocean, and Pascua in the Western Pacific). The LKS network has better coverage than Angell 54, as does HadRT and RIHMI, but all are still relatively sparse in regions that are of greatest interest for MSU/AMSU comparisons, particularly the South Pacific. Furthermore, the LKS dataset only covers the period up to 1997 and thus covers only ¾ of the extant MSU/AMSU record, and in particular, it omits much of the significant warming that characterizes the end of the 20th century and the beginning of the 21st. The impact of these differences can be seen in regional trends and confidence levels.

For 1978-1997, UAH Versions D and 5.0 show excellent agreement with LKS both regionally and globally for MSU2. But there is much less agreement for MSU2LT where LKS shows higher trends both globally and in all regions. The opposite is true for HadRT, which shows relatively good agreement with both UAH products for MSU2LT, but more global and regional cooling for MSU2. RSS consistently shows more warming UAH and LKS, but confidence intervals largely overlap except for the southern hemisphere, which may well over-represent cooling due to the sparse network sampling problems already discussed. HadRT2.0 consistently shows more cooling than UAH and RSS for what is likely the same reasons, though agreement is better with UAH (a fact that might possibly be related to the two not being fully independent). It can also be seen that while LKS shows excellent agreement with UAH for MSU2, it shows considerably more stratospheric cooling than other sonde and MSU products, which suggests that it may be under-representing MSU2. If so, correcting for this may well bring LKS into better MSU2 agreement with RSS, which shows less MSU4 cooling than UAH, HadRT, and LKS (Seidel et al., 2003).

Extending the comparisons to 2001 changes the picture yet again. For the longer period, LKS data cannot be used, as it only extends to 1997. For this period, HadRT again shows good agreement with both UAH products both regionally and globally, but confidence intervals are high with respect to the observed trends, particularly in the tropics. For MSU2, there is less agreement, though again UAH is closer to HadRT than RSS. But confidence intervals are large enough to include all 3 MSU products, and HadRT differs considerably from MSU even in the northern hemisphere where UAH and RSS data are best characterized. Note also that once again, HadRT shows considerably more stratospheric cooling (MSU4) than UAH and RSS, raising the likelihood that it may be under-representing MSU2 (Seidel et al., 2003).