Climate Change & Tropospheric Temperature Trends

These are not robust results, and we must conclude that DEA have not demonstrated either a surface-troposphere disparity or an inconsistency between model predictions and observation.

The Fu et al. Method - A Viable Alternative to TLT?

Because MSU Channel 2 signal receives up to 15 percent of its signal (its raw digital counts) from the lower stratosphere (the 100-50 hPa layer), it very likely underestimates temperature trends in the lower to middle troposphere (the 850-300 hPa layer). Traditionally, this was accounted for by using MSU2 and the TLT as complementary lower troposphere products. We saw earlier that while TLT reduces the stratospheric Channel 2 “footprint”, it pays a price in sampling error and contaminating inputs from other sources such as Antarctic sea-ice and melt pools. Chiang Fu and his co-authors developed their method to avoid these problems. By using direct MSU4 temperature and trend data to correct MSU2 they avoid sampling errors associated with off-nadir MSU views and greatly minimize signal contamination from the surface. Even so, there has been much discussion as to whether their method is more reliable than data from synthetic channels like TLT. After the method was first published (Fu et al., 2004) discussion of its strengths and weaknesses occurred mainly in conference settings and some popular forums. The criticisms fell chiefly into two groups – concerns about the functional form of the corrected weighting function W_FT, and concerns about the reliability of using statistical methods to derive the T₂ and T₄ data used with it.

Roy Spencer of the UAH team expressed concerns about the first point. He argued that because W_FT goes negative above 100 hPa it will inevitably alias spurious warming into the troposphere trend. Spencer argued that the method might work, but only if trends are constant with altitude from the upper troposphere to the lower stratosphere (roughly 300 -50 hPa) – which they are not (Spencer, 2004). This would be a valid criticism if the method used W_FT strictly for the derivation of MSU2 brightness temperature with the layers above 100 hPa removed. This is not the case. What Fu and his colleagues actually did can be seen more clearly in Figures 59 to 61. Figure 59 shows W_FT compared with the weighting functions for MSU2 and MSU4, and Figure 60 shows the same information with MSU2 color banded according to the layers it detects. The region shown in light orange reflects the uncorrected free troposphere contribution to MSU2. The region shown in light blue reflects the tropopause and lower stratosphere, where 300 hPa can be considered the “lowest approach” altitude for the tropopause and 200 hPa a global mean. Figure 13 (right side) shows 1979-2001 upper-air trends as a function of altitude for several radiosonde products and single point trends for UAH Version D (Angell, 2003). Similar data is reproduced in Figures 31 and 33 as broad layer bar graph data for the longer 1958-1997 period using a different set of radiosonde products. It can be seen that the satellite era trends decrease with altitude. Within the uncertainty ranges shown, they go negative above altitudes of roughly 7 to 9 km with the global average being around 8 km (the 300-100 hPa layer). Comparing these trends with Figure 59 reveals that for the satellite era, the light blue layer has an overall negative trend and the orange layer a positive one. Because MSU2 sees the full weighting function of both, it will alias the cooling trends above 300 hPa into the warming trends below. Figure 61 shows Figure 59 shaded to reflect the layer coverage of the Fu et al. weighting function in comparison to its uncorrected MSU2 and MSU4 counterparts. The region shown in dark blue can be expressed in terms of MSU4 and is chosen so that its weighting will integrate to zero with altitude above 300 hPa. Below, the Fu et al. function will have the same weighting that MSU2 would have seen below 300 hPa if the stratosphere were not contributing to its signal (the combined light and dark orange regions). The characterization of this weighting function allows for these two regions to be separately expressed as multiples of T₂ and T₄ from which the actual free troposphere brightness temperature trend can be derived. Now it can be seen that Spencer misunderstood the Fu et al. method. W_FT goes negative above 90-100 hPa because it must do so to prevent a stratospheric cooling from being aliased into the free troposphere trend.

Another challenge to the Fu et al. method was published in December of 2004 by the journal Nature. Simon Tett and Peter Thorne (hereafter, TT) of the UK Met Office used the Fu et al. method to derive new coefficients and free troposphere trends for the tropics (30 deg. S. to 30 deg. N Latitude) during the period 1978-2002 using the HadRT2.1s radiosonde analysis, the ERA-40 reanalysis (Uppala, 2003), and an ensemble of model runs (Tett and Thorne, 2004). These trends, which they denote as T_fjws in contrast with the T^tr_850-300 derived by other methods, were then compared to corrected MSU2 trends from UAH Version 5.0 (Christy et al., 2003), RSS Version 1.0, and surface trends. A comparison of their results is given in Figure 53. For non-satellite analyses, surface temperatures were derived from the products indicated. Satellite products were compared to surface trends from the HadCRUT2v dataset. ERA-40 reanalysis based surface trends were derived using zonal averages of 2-meter temperatures over land and SST’s over ocean regions. For their model comparisons TT used an ensemble of 6 runs of the atmosphere-only HadAM3 (Pope et al., 2000) and 4 runs of the coupled ocean-atmosphere HadCM3 model (Stott et al., 2000). Their HadAM3 and HadCM3 modeled results were forced with a suite of natural and anthropogenic inputs as described in the cited sources, and were identical with the exception of two corrections in HadAM3 – one for errors in ozone depletion and one for changes in sulfur cycle forcing (Tett and Thorne, 2004). Based on these results they concluded that,

• Fu et al. “trained” and tested their MSU2 and MSU4 coefficients (a2 and a4, respectively) using the same radiosonde dataset (Lanzante et al., 2003), obtaining false agreement and overfitting of the data. Their resulting corrections are overly small and result in overly warm free troposphere trends.
• For the Fu et al. methods to work, stratospheric trends must be relatively stable over the period analyzed, but in fact they are not. In particular, they claim that the lower stratospheric impact of the quasi-biennial oscillation (QBO) will be aliased into Fu et al. derived trends.
• With the exception of HadRT2.1s, free troposphere temperature trends as derived using the Fu et al. method applied to a suite of other upper-air products, show worse agreement with observation and larger confidence intervals than does the UAH Version 5.0 TLT product.
• Agreement between model run derived trends and those based on Fu et al. derived observations show good agreement only between the HadAM3 atmosphere-only run and RSS Version 1.0.

From a review of their methods and results, several comments can be made.

First, it is odd that TT base their comparison study on the tropics only. This is precisely the latitude band for which lapse rates are largest and trends are most variable for the period they studied. Extant radiosonde and reanalysis products are poorly characterized in this region as well. It is not clear why they did not extend their analysis to include a study of global and high latitude trends, and they offered no explanation for this. Such a study would have been particularly useful because the northern latitudes in particular are where their chosen radiosonde and reanalysis products are relatively well characterized and have good coverage. Furthermore, the high southern latitudes are where we expect the biggest differences between UAH and RSS products prior to correction by the Fu et al. method, and where we expect the largest contamination of the TLT record from sea-ice and summer melt pools signals. A test of the Fu et al. method in these regions would have been far more revealing than the region they chose. TT use ERA-40 for their reanalysis product, and this analysis has made great strides over the earlier ERA-15 product in dealing with issues like sea-ice and snow cover, particularly during the satellite era (Bromwich and Fogt, 2004). Comparisons with this product in these regions might have shed light on potential problems with the TLT record, but were not investigated.

Even so, their criticisms of the tropical record are flawed as well. TT rightly point out that the Fu et al. method is most reliable when stratospheric trends are relatively stationary by region and period. But then they point to the Quasi-Biennial Oscillation (QBO) as evidence that they are not and claim that Fu et al. are aliasing QBO trends into their MSU2 correction. Figure 22 shows the stratospheric QBO signal compared to monthly anomaly time series for 6 vertical layers averaged over several upper-air products. Included in this comparison are LKS, HadRT, RIHMI, Angell 63, Angell 54, and UAH Versions D and 5.0. All six time series shown are global, and the QBO signal was determined using 50-hPa altitude zonal wind patterns from radiosonde data at Singapore (Seidel et al., 2004). Because these time series’ draw upon a variety of products including both radiosonde and MSU, they are less subject to the idiosyncracies of any particular dataset, and as they are global in nature they present a better comparison to the Fu et al. methods than the tropical data used by TT. Three things are apparent. First, it can be seen that apart from a slight upturn prior to 1981 (at the beginning of the satellite era) and the upward punctuations of the El Chicon and Pinatubo eruptions, the MSU4 and 50-100 hPa time series’ are fairly monotonic and stable for the entire period TT examine, so this requirement is met. It may be argued that the two volcanic events destroy this continuity, but they are also reflected in the tropospheric MSU2 and 850-300 hPa records as proportionally large dips in those records shortly after the stratospheric spikes. Therefore, both layers will reflect this activity in comparable proportions with regard to trend comparisons. Second, a close examination of the stratospheric global MSU4 and 50-100 hPa layer records reveals that at best, the QBO impact on them is barely noticeable. The tropics where TT chose to do their analysis, is the one region where we expect the most significant QBO impact, but this region tells us the least about the applicability of the Fu et al. method to the global trends it was used for (Seidel et al., 2004). Finally, the QBO time series is highly periodic, and therefore self-canceling. Even if it did alias a significant signal into the tropospheric record, that signal would be largely removed by the trending process (Fu et al., 2004b). Furthermore, TT’s criticisms assume that the Fu et al. weighting function goes negative above 100 hPa and will therefore alias QBO effects into the free troposphere record that are not there currently. In fact, this is true only of the Fu et al. global weighting function. . The revised Fu and Johanson tropical weighting function does not go negative until around 75 hPa. Figure 54 shows this function compared to its MSU2 counterpart. It is evident that the MSU2 weighting receives more signal from this layer than the Fu et al. weighting, and for the latter the layer above 100 hPa will cancel out while the MSU2 contribution will not (Fu and Johanson, 2004; Fu et al., 2004b). This can even be seen in the global Fu et al. and MSU2 weightings shown in Figure 49.

TT’s reported disparities between T_fjws and T_850-300 trends in the tropics are also less revealing than they believe. In addition to the large lapse rates and temporal variability characterizing this region, the tropical tropopause often dips as low as 300 hPa. Tropopause trends are poorly characterized across all upper-air products and can significantly affect lower altitude trends if it is not excluded from the sampling (Seidel et al., 2004). For this region, Tfjws is representative of the entire troposphere from the surface up to 100 hPa rather than the 850-300 hPa layer alone. This can be seen clearly in TT’s own dataset. Their 1000-100 hPa layer trends agree quite well with their reported T_fjws trends (Fu et al., 2004b). Their ERA-40 derived trend forT_850-300 is 0.03 deg. K/decade. The ERA-40 vertical trend profile in this region is revealing. It is positive below 775 hPa, negative between 700 and 400 hPa, and strongly positive between 300 hPa and the tropopause – which itself may occur anywhere between 300 and 100 hPa in this region for the period TT analyze. Therefore, for this region the T^tr_850-300 trend may be much smaller than its T_fjws counterpart simply because of the vertical variability of this region (Fu et al., 2004b). But the global record will not reflect this.

TT also argue that model runs for the satellite era agree with T_fjws trends only for the atmosphere-only cases. They used HadAM3 (atmosphere-only) and HadCM3 (coupled ocean-atmosphere) forced with natural and anthropogenic inputs for their model comparisons and found that while their atmosphere-only and coupled model runs yielded similar results, the coupled runs (HadCM3) yielded a higher range of trends and overall, model results were consistent only with the corrected T_fjws trend of RSS Version 1.0. After applying the TLT and Fu et al. methods to these model runs, they concluded that TLT provides a better estimate of free troposphere trends than does T_fjws. For their coupled model evaluations, TT ran HadCM3 with the HadlSST ocean component. HadlSST is an ocean surface model based on observed SST and flux similar to the Ocean A component model that is part of the GISS SI200 suite (Tett, 2004). In fact, it was noted earlier that Ocean A is based largely on HadlSST. Here, TT are doing a comparison study of modeled and observed historical trends not unlike the one done by Douglass et al. (2004b) but with a different objective in mind. In regard to DEA’s model runs, we saw that historical SST based ocean models like HadlSST are a particularly good choice for historical trend studies, and TT take advantage of this. Even so, we also saw that these components have a tendency to overestimate upper-air trends compared to alternative deep ocean component models – possibly due to misrepresentations of high latitude sea-ice and SST. For at least two reasons, this will not likely be as big a problem for TT as it was for DEA. First, whereas DEA were testing the effectiveness of AOGCM’s for evaluating past, present, and future global change, TT seek only to evaluate two upper-air analysis methods as applied to a specific trend profile problem. As long as their model results are self-consistent and reasonably representative of actual upper-air trend profiles, they will be useful for this. Second, DEA sought a global evaluation of AOGCM’s in general, but TT conducted their study for the tropics only. Thus, at least some of the coverage and continuity problems that will have impacted DEA will not affect them, particularly the high latitude SST and sea-ice problems that are known to have impacted HadlSST and Ocean A (Hansen et al., 2002).

Even so, some coverage and continuity problems undoubtedly remain, and TT’s conclusions would have benefited from additional coupled model runs using at least one deep ocean component such as the Ocean B or Ocean E components of GISS SI2000 (Hansen et al., 2002; Sun and Hansen, 2003). While these are not without problems of their own, they have many strengths that SST components lack, and a suite of runs using both types of ocean component might have provided a clearer picture. In light of this, it is instructive to compare TT’s use of HadCM3 with that of Douglass et al. (2004b) and corresponding runs of GISS SI2000 using Oceans A, B, and E (Hansen et al., 2002; Sun and Hansen, 2003). DEA obtained SST forced HadCM3 data (Tett et al., 2002) directly from Tett and used it to generate vertical trend profiles for the tropics (30 deg. S to 30 deg. N Latitude). These vertical profiles are directly comparable to the data used by TT, the sole exception being that whereas TT report 1979-2001 trends by layer, DEA truncate their analysis to 1979-1996 so as to create the surface-troposphere trend disparity that their case depends on.

Figure 55 shows DEA’s 1979-1996 vertical trend profile for the same tropical region analyzed by TT (Douglass et al., 2004b). Included is a direct comparison of HadCM3 for the period 1975-1995 and GISS SI2000 forced with Ocean A for a similar period (1979-1998). Like TT, DEA use HadCM3 runs that were forced with natural and anthropogenic inputs, and did the same for their GISS SI2000 results. For this region and these periods, it is evident that both SST forced models give strikingly similar results indicating that they are largely comparable for this region and period. Extending the record to 2001 would not be likely to change this result significantly, as both models can be expected to capture the large 1997 ENSO event which dominates this portion of the record. Given the similarities between the two models, the impact of replacing the SST driven Ocean A component of GISS SI2000 that was used by DEA with a true deep ocean component like Ocean B. Figure 46B shows the difference. The left side figure shows the 1979-1998 vertical trend profile for the tropics and extra-tropics (40 deg. S to 40 deg. N Latitude - a region slightly wider than that used by TT) that is obtained using Ocean A SST forcing. The right side shows the comparable trend profile obtained from Ocean B forcing. It is clear that the deep ocean model yields a better fit with observation than the HadlSST based Ocean A component. Figures 57 and 58 show similar vertical trend profiles for the globe from runs using all three SI2000 ocean components, and transient global trends by vertical layer, respectively. These runs are not directly comparable to TT’s model studies, as theirs are tropical rather than global. But they are instructive nonetheless in that they demonstrate clear trend moderating effects that are generally closer to observation than the SST runs alone. They also appear to capture observed deep ocean heat content anomalies better, indicating that even though SST based models in theory, directly represent ocean heat content as surface fluxes, the dynamically modeled ocean components may offer better representations of these effects on regional and/or global scales.

It is reasonable to expect similar behavior from HadCM3. Had TT supplemented their HadlSST based runs with deep ocean components based runs we would expect their modeled tropical trends to be lower as well. This would likely have put them in a range where given the uncertainties in forcing and model component responses, they would be adequate representations of either UAH or RSS corrected tropospheric trends. With regard to the uncertainties inherent in an exercise like this one, note also that neither HadCM3 or GISS SI2000 with either forcing scenario reproduces the positive-negative-positive vertical trend variability that is observed in the tropics as we saw earlier (Fu et al., 2004b). Something approaching this behavior is noticeable to some degree in the high southern latitudes (Figure 46B, lower right), but is not captured in the tropics and extra-tropics. This alone should lead us to exercise caution when using model runs for comparisons with observation in localized latitude bands like the tropics that are highly variable and not well characterized in upper-air products. TT raise many important question regarding the use of the Fu et al. method and have shown that care must be taken when applying it. But their specific criticisms are lacking in many respects and fail to demonstrate that the method is not viable.

With regard to tests of the Fu et al. method, there is one more that needs to be considered. In the same December 2004 issue of Nature, alongside of TT, Nathan Gillett and Andrew Weaver of the University of Victoria, BC and Ben Santer (hereafter, GWS) published the results of their application of the Fu et al. method to AOGCM derived upper-air temperatures for the period 1958-1997 (Gillett et al., 2004). GWS used global upper-air temperatures from a four-run ensemble of the DOE PCM coupled ocean-atmosphere model forced with natural and anthropogenic inputs (Santer et al., 2003c; Washington et al., 2000) and used the Fu et al. method to derive values for the a₀, a₂, and a₄ coefficients. These were then applied to MSU2 and MSU4 brightness temperature trends that had been obtained by applying the respective weighting functions to PCM temperatures and using least squares methods to obtain the corresponding layer trends. The resulting T_FT (free troposphere) trends were then compared with the equivalent T850-300 and TLT trends that had been derived from PCM. Results are shown in Figure 56.

GWS found that the Fu et al. derived T_FT trends agree with the model “observed” T_850-300 trends to within +/- 0.016 deg. K/decade. Similar agreement was found for the northern and southern hemispheres and the tropics (Gillett et al., 2004). It is interesting to note that GWS’s T_FT trends also agree with their simulated T_LT trends for the same period and regions, indicating that the two do reflect similar upper-air layers. The significance of this test as compared to others is that the PCM modeled climatology is precisely known, and is therefore not subject to the observational uncertainties that plague the existing satellite, radiosonde, and reanalysis records (e.g. sampling noise, incomplete coverage and temporal record, differences in merge method, etc.). Whether or not it is accurate in its finest details compared to observations is beside the point. PCM does in fact reproduce the large scale behavior of the surface and troposphere and captures most of its more significant features. Therefore, it represents a valid “upper-air” environment against which the Fu et al. method itself can be tested. Because the objective of the GWS study was to test this method rather than reproduce observed climate variables, the robustness of the Fu et al. technique was demonstrated.

Thus, though a number of challenges have been made to the Fu et al. method, none of them withstands scrutiny. Given the relative stability of the stratospheric record and the independent test of the method’s robustness using modeled and multi-dataset applications, what criticisms remain regarding the statistical characterization of the method’s trend analysis are not likely to stand the test of time. As the quality of radiosonde, rawinsonde, and AMSU products grows, better characterizations of W_FT and T_FT will emerge that will allow for more complete investigations of the Fu et el. Weighting function and T_LT products. Until then debate regarding the two methods will likely continue, and both will be treated as complementary approaches to stratospheric trend removal.

The Road Forward

Despite the progress that has been made, many unanswered questions regarding upper-air trends remain and much work remains to be done. In their year 2000 report on upper-air datasets and global change the NRC identified four areas where changes needed to be made before significant progress could be made to expand upper-air datasets and reduce the existing uncertainties (NRC, 2000). Specifically, they called for,

• An international program to expand and update existing networks of upper-air temperature monitoring products to improve their quality and consistency. The existing products were not designed for monitoring global change and they lack the accuracy required to properly characterize long-term small trends and/or the needed level of global coverage. Improvements should include new temperature processing algorithms, better archiving of data, improved record keeping at upper-air stations, more consistent data retrieval methodologies, and expanded access to datasets by all research teams.
• A more comprehensive analysis of the uncertainties inherent in upper-air temperature retrievals and the data reduction and analysis methods applied to them. This should cover all the errors described in this paper, and an expanded search for further potential uncertainties. In regard to the satellite data, the uncertainties in the NOAA-09 record and inter-satellite merge methods are of particular concern. Lack of consistency in station records and data retrieval methods and are high priority problems for radiosonde products. All of this will need to be addressed.



Top


Page:   << Previous    1    2    3    4    5    6    7    8    9    10    11    12    13    14    15    16    17    18    19    20    21    22    23    24    25    26    27    28    29       Next >>