Climate Change & Tropospheric Temperature Trends

In September of 2003, Konstantin Vinnikov of the University of Maryland and Norman Grody of NOAA (hereafter referred to as VG) provided an independent analysis MSU/AMSU data from 1979 to 2002 using a methodology that that differs considerably from that of UAH and RSS. VG raised three concerns with previous analyses, UAH’s in particular.

• In situ corrections for diurnal cycle variations cannot be corrected for reliably due to an insufficient number of observations.
• Off-nadir brightness temperature views must be corrected to nadir before they can be used, and this introduces error – particularly when they are used in composite Channels such as UAH’s 2LT and TLT Channels.
• Data merging procedures based on comparisons of overlapping satellite records cannot account for drifts in hot target calibration factors and IBE because the effects of each cannot be reliably separated from diurnal cycle variations.

To address these concerns, VG analyzed the MSU/AMSU using a model they had developed elsewhere specifically for analyzing data from multiple satellites with seasonal and diurnal cycles and differing LECT’s (Vinnikov et al., 2002a, 2002b). In this model the long-term brightness temperature trend from each satellite is given as a set of Fourier components with terms representing the seasonal and diurnal cycles, the underlying climatic temperature trend, and a term for underlying constant bias intended to capture random noise and IBE. The satellite data is then fitted to this model and the Fourier components and bias constants are solved for using a least squares analysis. To account for orbital decay, they used only near nadir views. With this method they obtain a middle troposphere temperature trend of 0.22 – 0.26 deg. K/decade from MSU Channel 2 and AMSU Channel 5 (Vinnikov and Grody, 2003).

The Vinnikov and Grody method avoids the noise sources inherent in the analyses of earlier work and is flexible enough to account for a wide range of observed climatic effects. But because they have restricted their analysis to the low noise nadir views, their observations only cover a narrow earth scene directly under the satellite path. The regions between each orbital path that would have been observed by off-nadir views are missed. In addition, a single diurnal cycle is assumed and there is no discrimination between ocean and land. Because of these two omissions, the method relies heavily on zonal averaging of the data to account for unobserved regions and land/ocean variations. If VG have done their zonal averaging and statistics well, their numbers will be better than anyone else’s. If not, their calculated decadal trends could differ markedly from reality. These points are still being hotly debated.

The Radiosonde Record

Though the MSU/AMSU record provides the only truly global tropospheric temperature characterizations currently available, the issues discussed above present significant roadblocks to getting tropospheric temperature trends that are accurate enough to detect any potential anthropogenic fingerprint. Even after more than 15 years of investigations, many questions remain and the different teams, all equally committed, are obtaining widely varying results. In light of this, many researchers have sought independent verification of the MSU/AMSU record from alternate products. The one promising option for this is the radiosonde record.

For the last 50 – 60 years balloon borne instrument packages called radiosondes, whose sensory devices track temperature, humidity, and pressure have been launched at least once daily from many weather stations around the globe. The data from these devices provide a detailed temperature, pressure, and humidity history of the lower and middle troposphere at discrete altitudes that have been used for regional weather forecasts and meteorological studies. Once released, they ascend at known rates gathering data as they climb until they reach their maximum altitude. Data regarding wind patterns can also be obtained by monitoring radiosonde movements after release. Though not specifically designed for the purpose of assessing global change, in the absence of other in-situ data sources, they have been used for this purpose since then. As of 2000, there was a global network of some 900 upper air radiosonde launching stations located at points all around the world including polar regions. This network is predominately land based and weighted toward Northern Hemisphere locations. At least two thirds take twice daily measurements at 0000 and 1200 Coordinated Universal Time, or UTC (NRC, 2000). There are several datasets that have been compiled from this network spanning a period dating from the late 50’s to the present that have been used to evaluate global tropospheric temperature trends (Angell and Korshover, 1975; Angell, 1998; Parker et al., 1997; 2000; Brown et al., 2000; Gaffen, et al., 2000a; Lanzante et al., 2003). These datasets have been analyzed in different ways yielding different results depending on which data sources were used, sampling methods, and analytical comparisons were used (Angell, 1999; Parker et al., 1997; Santer et al., 2000; Lanzante et al., 2000; Hegerl and Wallace, 2002; Lanzante et al. 2003; Seidel et al., 2003).

The greatest strengths of the radiosonde record are the length of time over which reliable measurements have been made (which is considerably longer than the MSU/AMSU record), the maturity of the temperature sensing technology used, and the fine resolution with which sondes measure vertical temperature profiles and lapse rates ⁴. Even so, the sonde record faces a number of challenges. How these are dealt with bears directly on the usefulness of this data as an independent check of the MSU/AMSU record. Sonde record limitations can be broadly described as follows.

1. Instrumental Variation:    Over the years a number of makes and model of radiosonde have been put into service around the world. Instrument packages on these platforms have varied according to the manufacturer and model of sonde involved, and have evolved considerably in their sophistication and reliability since the post-WW II period. Newer devices utilize temperature thermocapacitor - wire resistor – thermisters based designs, or bimetallic sensor designs (NRC 2000). Variations in the calibration, reliability and general accuracy of these devices introduce errors which must be corrected for. In addition, a “technology growth” correction is necessary to account for the fact that advances in sensor technology since the birth of the sonde network have introduced variations in accuracy and life of sonde based temperature sensors in the last 50 years. An overview of the evolution of sonde based instrumentation and the potential sources of error associated with it was published by the World Meteorological Organization in 1996 (WMO 1996), and it is as relevant today as it was then.
2. Sampling Variation:    The 900 odd upper air radiosonde stations around the world are at discrete, locations that tend to favor the Northern Hemisphere. The large majority are at continental land locations. There has never been a truly global coverage pattern of the sort that reliable estimates of global change require. This has led to the need to either extrapolate sonde data to global scales with the aid of models, or limit their use to partial comparisons with other datasets. Furthermore, radiosonde stations have come and gone over the years due to budget cuts, policy changes, and regional conflicts - particularly in regions of political unrest like sub-Saharan Africa or the former Soviet Union after the fall of communism. Thus, the number and location of many sonde stations has not been constant since their records began and temporal corrections and/or weightings must be introduced into these datasets to account for this “geographic noise”. Variations in land surface characteristics have also affected data from some stations. Heavy development or land use changes could lead to increased desertification or forestation for some stations over longer time periods. These can in turn impact the local surface emissivity of regions monitored by these stations. Growing urbanization will do the same. Some stations have varied the times at which daily sonde reading are taken throughout their history. These variations will create spurious temperature variations in a manner not unlike that of the diurnal cycle variations experienced by MSU products when their LECT’s vary over their service lives (NRC, 2000). There are a variety of ways to account for such effects, but each has weaknesses as well as strengths.
3. Vertical Data Homogeneity:    Though sonde datasets have far better vertical resolution than satellite products, they very widely in the extent of their vertical coverage. At their best, radiosondes can achieve altitudes of up to 50 km, but older or less well designed balloons commonly burst at much lower altitudes creating inhomogeneities in the altitude range covered by some datasets. Differing sonde instrument packages have also varied over the years in the number of readings taken per ascent and the specific altitude levels of each. Measurements have been taken and transmitted with anywhere from 10 m to 50 m resolution depending on sonde rise and/or data sampling rates. Older archived datasets for many global stations may contain only 20 data levels per sounding - far less than most modern records (NRC, 2000). Some datasets are based on all available data, with variations in these factors (Angell, 1988). Others, such as CLIMAT TEMP reports are based on monthly normalized data from a global subset of upper air stations and from preset pressure determined altitudes (Parker et al. 1997). Missing data and temporal variations in how they were gathered can introduce significant errors in these monthly estimates (NRC, 2000). In addition, sonde datasets have been built over the years under widely varying surface and boundary layer conditions reflecting differing locations and local weather conditions. Variations in the daily launch times mentioned previously can also affect vertical profile readings, as these can very throughout any given day. These factors were never considered an issue before given that, like the original MSU packages, sonde records were produced to support local meteorological studies and/or weather forecasts. But for global change studies they cannot be ignored, particularly when they are being used to independently validate other data sources with differing issues.
4. Onboard Instrumental Biases:    Like MSU and AMSU packages, sonde based temperature sensors are also subject to instrumental biases from various sources including incident solar and infrared radiation (which is not related to ambient atmospheric temperature), drifts in calibration, and noise driven variations in measurement accuracy. Device designs have evolved in their reliability over the years as well, and these factors are not necessarily constant over the lifetime of measurements from any given station (NRC, 2000). Variations in how instrumental factors have been accounted for over the years in analyses have impacted results, particularly for the upper troposphere (Gaffen, 1994). Toward the end of the 20th century for instance, there was a growing shift toward the use of sonde instrument packages built by the Vaisala company (Parker and Cox, 1995). These packages have measurement characteristics which differ from those of earlier types and the differences can be aliased into the record of the sonde stations that implemented them.
5. Data Reduction and Analysis Variations:    Over the years, dataset problems like those discussed above have been dealt with in varying ways. Analytical techniques that account for onboard instrument biases have evolved as understanding of them grew. There have also been variations in the way monthly mean data have been evaluated, leading many stations to use differing methods for reporting monthly means at varying locations worldwide and over the years records have been kept. This has been a problem for the CLIMAT TEMP monthly reports most commonly used for MSU/AMSU trend comparisons. Analysis of years worth of monthly averaged data from globally distributed stations using consistent rules has shown that significant errors can result from these variations (Gaffen et al. 2000b). Another issue plagues the sonde record that is of particular relevance to MSU/AMSU comparisons. MSU and AMSU packages measure the weighted average temperature of broad layers of the troposphere and stratosphere, but sondes generate specific temperature vs. altitude profiles. These profiles vary widely with location and date and bear little resemblance to MSU/AMSU weighting functions. As such, sonde data cannot be directly compared to MSU/AMSU datasets until they have been processed so that they better represent what these sensors would have observed had they been co-located. There are differing ways this can be done. The simplest approach is to average the sonde data over the same atmospheric layer seen by the corresponding MSU/AMSU channel. This neglects the altitude determined weighting function these channels see (Figure 7). More commonly, sonde datasets are weighted in a manner that reflects the MSU/AMSU weighting functions, yielding more meaningful comparisons. This however does not account for the impact of atmospheric moisture, which MSU and AMSU sensors are relatively insensitive to but sondes are not. This can introduce further error, though typically errors of this sort are typically small in comparison to the decadal trends being measured (NRC, 2000).

Effectively addressing these uncertainties requires that we examine the operational history of each reporting station as well as its data. But for a variety of reasons weather stations around the world have varied widely in the quality and completeness of their records. This has been particularly problematic for MSU/AMSU comparison studies because many of the most important regions for comparison (e.g. the tropics and northern Africa) lie in regions of chronic political instability and weather stations in these regions have had difficulty keeping their funding base and maintaining consistent locations and operational records. Other areas that are even more important for MSU/AMSU comparisons (e.g. the Southern Pacific Ocean) are under-sampled due to a scarcity of stations. Where datasets and operational records are incomplete, it is necessary to interpolate data and/or metadata from the station’s past records or from the nearest stations with continuous simultaneous records.

There are two basic dataset products that have been used in radiosonde analyses relevant to MSU/AMSU comparisons. The first is based on individual soundings containing all gathered data at each level of ascent for all stations reporting data in this manner (Angell, 1988). The National Climatic Data Center (NCDC) has consolidated numerous such datasets from around the world, along with all relevant metadata about each station’s operational history into a single database called the Comprehensive Aerological Reference Data Set, or CARDS database (Eskridge et al., 1995). The second is based on monthly averaged data taken from a globally distributed subset of weather stations at 9 mandatory pressure levels between 850 and 30 HPa. These are known as CLIMAT TEMP reports (Parker et al., 1997). Of all weather stations worldwide, roughly 45 percent provide both CARDS and CLIMAT TEMP datasets (NRC, 2000). Both have suffered from discontinuities in data and metadata and must be adjusted accordingly before they can be used. Typically there are 3 steps involved. First, anomalous discontinuities in the record must be identified and separated from real climatic effects. Second, the size of these discontinuities must be determined. Last, appropriate corrections must be made that will remove their impact on the trend. In doing all this there are 2 pitfalls to be avoided – natural variations in regional climate might be confused with changes in a station’s operational practices, and changing operating practices might be mistaken for real climate variations. Both will introduce spurious warming or cooling into the long term record from that station that will impact a global time series. To date, several radiosonde analysis products have been developed. They differ from each other in their use of either CARDS or CLIMAT TEMP data and their methods for identifying anomalous events and correcting for them. All have strengths and weaknesses and none stands out as uniquely more reliable than the others. The ones that have been of most interest for MSU/AMSU comparisons are the following.

Angell 54

For over 3 decades many tropospheric radiosonde studies used an analysis product prepared by Jim Angell, currently of the NOAA Air Resources Laboratory in Silver Springs, MD. This product was based on data taken from a selection of 63 global radiosonde stations taken at mandatory layers in the troposphere (850-300 hPa), tropopause (300-100 hPa), and lower stratosphere (100-50 hPa) layers. Layer mean temperatures were ascribed to each using the hydrostatic equation an temperatures measured at each bounding pressure surface. Trends for each layer were then derived for various global regions by averaging the annual temperature anomalies using data from all stations within each (Angell and Korshover, 1983; Angell, 1988). The resulting analysis has been referred to as Angell 63. While the coverage of this network has been considered to be good and the record relatively complete, the consistency and quality of the data were not explored in much depth and the resulting trends were taken at face value. Recently, other investigators have examined individual Angell 63 station datasets more closely and determined that changes in instrument packages and operating procedures have contaminated the data from several of them (Gaffen, 1994; Gaffen et al., 2000a; Lanzante et al., 2003). To quantify the impact of this, Angell reanalyzed the 1958-2000 datasets for the 63 station network focusing on the 300-100 hPa layer. This layer, which reflects the tropopause, was selected because other layers had less complete datasets, and the full time period was chosen so as to make characterization of anomalies more apparent. A least squares regression analysis was performed for the network and the confidence intervals of each stations record was then compared to the network trend to determine which ones deviated most from the regression. Anomalous stations were defined as those whose 2 standard error of regression values were larger than 0.2 deg. K/decade. Angell identified 9 of the original 63 stations, located mainly in the tropics, as providing anomalous data. These 9 were removed and the remaining 54 station network was reanalyzed. This analysis product has come to be known as Angell 54 (Angell, 2003).

Figure 12 shows the original Angell 63 station network with the 9 anomalous stations highlighted in red. Note that all 9 of the anomalous stations removed are located in the tropics. The remaining 54 stations show reasonable global coverage emphasizing land based locations, though the density of the network is quite low given the regional variability of tropospheric temperature trends and lapse rates. Coverage of ocean regions is somewhat sparce, particularly in the Southern Pacific. Figure 13 shows the upper air trends obtained from this network for the 850-300 hPa, 300-100 hPa, and 100-50 hPa layers in 4 regions – the northern and southern hemispheres, the tropics, and the globe as compared with several other radiosonde, radiosonde-satellite re-analysis, and surface products, and UAH Version D (Angell, 2003). Here, the tropics are defined as the region from 30 deg. N to 30 deg. S Latitude, whereas the IPCC WG1 Year 2001 Report defines the tropics as between 20 deg. N to 20 deg. S Latitude (IPCC, 2001). Figure 14 shows just the Angell 54 upper air trends for the same altitudes and regions for 1958-2000 (Left), 1979-2000 (Center), and the change from the former to the latter (Right). As in Figure 13, the horizontal bars indicate 2-sigma confidence intervals. The most noticeable feature is the stronger cooling of the low stratosphere (100-50 hPa) for the 1979-2000 period vs. 1958-2000, the difference amounting to nearly 0.4 deg. K/decade. Also noticeable is the cooling in all regions of the tropopause (300-100 hPa) in the later period relative to the whole record, and possible even a shift from warming to cooling. Based on the mean trends, the lower to middle troposphere (850-300 hPa) appears to have shifted from warming over the longer record to less warming or possibly cooling everywhere except in the northern hemisphere. It also appears that lapse rates have increased during the latter period indicating that the surface appears to have warmed with respect to the lower troposphere. Figure 13 also shows that Angell 54 is in good agreement with other sonde and surface-air products, and UAH Version D. But the confidence intervals shown, as determined by twice the standard error of measurement and the observed discrepancies between other products, indicate spreads that accommodate both UAH and RSS analyses. It can also be seen that a significant range of variation is possible for all regions, particularly the tropics.

HadRT

Another sonde analysis product has been developed by David Parker and Margaret Gordon of the Hadley Centre, U.K. MET Office. This one uses monthly gridded CLIMAT TEMP data from a globally distributed network of 400 weather stations, considerably larger than Angell 54. The methodology is best described in Parker et al. (1997). Several HadRT analysis products are currently available. HadRT2.0 uses data from 1958 to the present on a 5-degree latitude by 10 degree longitude grid. No bias corrections are applied to station data for this release. Versions 2.1 (troposphere and stratosphere) and 2.1s (stratosphere only) identify anomalous changes in station records using comparisons with MSU data from UAH Version C (Christy et al., 1998) for Version 2.1, and Version D (Christy et al., 2000) for Version 2.1s (stratosphere MSU corrections only). Later versions add additional corrections to data from 1958 to 2000, the most recent as of this writing being HadRT2.3 (troposphere) and HadRT2.3s (stratosphere) which use Laplacian methods applied to the second derivative of the corresponding NCEP reanalysis ⁸ temperature fields to fill in gaps in the data (Reynolds, 1988). Of the available products, HadRT2.1(s) is the one that to date has been most frequently used for MSU comparisons.

The HadRT method uses sonde temperature data taken at 9 standard altitudes between 850 and 30 hPa. These subjected to certain quality control measures (Parker & Cox, 1995), and given a weighting derived from the MSU Channel 2 weighting function so as to simulate a “bulk” layer temperature that would have been measured by MSU devices looking directly at the same vertical column of the atmosphere. Then for each global grid location, this weighted data is compared on a case by case basis with collocated monthly MSU anomalies and known changes in instrumentation and analysis methods taken from Gaffen (1996). Differences between the sonde and MSU anomalies are evaluated before (Δ1) and after (Δ2) the most recent known changes in operating procedure and/or equipment, excluding data prior to known previous changes. If Δ2 - Δ1 was found to differ significantly from zero at the 95 percent confidence level, a seasonally invariant correction equal to Δ2 - Δ1 was applied to the earlier data for the layer being evaluated as far back as the next previous identified instrument change. If no earlier changes are identified the correction is applied back to January 1979 when the MSU record began. Δ2 is evaluated for the entire record following the most recent instrument change and incorporates bias corrections already evaluated in previous steps. In this manner, time series for each global grid are evaluated one step at a time going back as far as 1979. Each adjustment is then apportioned to the individual mandatory altitude layers according to average bias adjustments estimated at each layer in the tropics and extra-tropics for different classes of instrument and/or analysis method changes (Parker et al., 1997). This method benefits from the leveraging of the MSU data which is consistent and truly global in a way that raw sonde data is not. But CLIMAT TEMP data were gathered at observation times (0000 UTC or 1200 UTC) that may differ from location to location, and often the time series for a given gridded location may combine both. This can result in inconsistencies and complicate comparisons with other similar methods. It is also presumed that natural events such as volcanic eruptions will affect sonde and MSU products in a similar manner. The MSU flight history has seen at least 2 major volcanic eruptions which are known to have impacted global temperature profiles in both the troposphere and lower stratosphere. MSU devices rely on radiative measurements from space while sondes utilize local direct temperature measurements using thermocaps (dielectric temperature sensors) which may be impacted quite differently. It is not at all clear that both products have responded to these events in a similar manner. Since corrections are based on the MSU record, the two are not entirely independent. To some degree this compromises the usefulness of this product for evaluation of MSU analyses, although the corrections were applied only to a relatively small portion of the data, so the impact is not severe.

UAH

The UAH team has developed their own method of preparing sonde analysis products. Their method compares monthly anomalies of sonde simulated MSU data (RaobT_b) to actual MSU data in a manner similar to that of the HadRT method (Christy et al., 2000). There are a few key differences though. Whereas HadRT uses monthly gridded CLIMAT TEMP data, the UAH method uses raw CARDS data taken directly from specific stations and wherever possible, considers data gathered at 0000 UTC separately from that taken at 1200 UTC. In addition, this method evaluates record discontinuities based only on bulk temperatures derived from collocated MSU and RaobT_b data, as opposed to HadRT which the applies corrections on a layer by layer basis. First, a time series is formed by differencing RaobT_b and MSU data. Then, a 30 month running average time series is formed from this one and used to identify anomalous discontinuities in record. It has been shown that using this method, differences of about 0.3 K in the tropics to 0.6 K at higher latitudes are significant (Free et al., 2002).