Search:           


Climate Change & Tropospheric Temperature Trends

Part I: What do we know today and where is it taking us?

In later versions of their analysis products, UAH incorporated a merge procedure that models intersatellite discrepancies with a simple relationship that relates measured and actual atmospheric brightness temperature to non-linear instrument errors using pre-launch calibration coefficients, and to hot target temperature with a constant coefficient called the “hot target calibration factor” that is unique to each satellite. First, diurnal cycle errors were removed by comparing data from off-nadir views along the satellite’s track. Since at any given point these views are looking at different locations along lines of latitude, they are seeing different simultaneous times of day. Comparing the evolution of these separate views with time allows for the diurnal drift of each satellite to be evaluated and removed. After removing diurnal cycle errors, data from overlapping satellite records are smoothed to better facilitate comparison. Then these data points are incorporated in a set of some 30,000 equations describing the hot target factors and the offsets. These equations are solved simultaneously using a least squares method to calculate the hot target factors and a unified time series that minimizes the discrepancies between it and the individual records. Typically, a preliminary calculation is done that optimizes calculation of the target factors. These are then used in a more comprehensive analysis to derive the final global time series and decadal atmospheric temperature trends.

UAH released Version 5.0 (as opposed to “E”) in 2003 (Christy et al., 2003). In this version, data is analyzed for TIROS-N to NOAA-16 spanning from January 1979 to April 2002. Of these, NOAA-14 was the last to carry the original MSU package. With the launch of NOAA-15 in May of 1998, AMSU packages became the standard. Since these packages utilize more channels and views than their predecessors, the newer analysis was altered to incorporate the data into previous time series. The new lower troposphere channel MSUTLT was developed for this analysis from MSU Channel 2 and AMSU Channel 5, replacing the MSU2LT channel of earlier (see Figure 5). Likewise, two other channels called MSUTMT and MSUTLS were derived to replace Channels 2 and 4 respectively. With the newer data and methods incorporated, UAH Version 5.0 shows a lower troposphere (MSUTLT) temperature trend of +0.061 deg. K/decade and a middle troposphere (MSUTMT) trend of +0.013 deg. K/decade. Version 5.1 (Christy et al., 2004) updated Version 5.0 to include data from Nov. 1978 to late 2003 and an updated estimate of the NOAA-9 hot calibration target factor over that of earlier versions. UAH is updating their MSU2LT and MSUTLT products on an ongoing basis. As of October 2004, MSU2LT was showing a lower troposphere temperature trend of +0.077 deg. K/decade.

In recent years the evolution of UAH methods and analyses has been generally in the direction of increasing decadal tropospheric temperature trends. But their analyses still show less tropospheric warming than is expected from current AOGCM studies – results which continue to generate controversy.

The Prabhakara (PR) Team

In 1998 another analysis of MSU products was published by a team led by C. Prabhakara of the NASA/Goddard Space Flight Center (Prabhakara et al., 1998). Prabhakara and his colleagues analyzed MSU records from NOAA-6 through NOAA-14 spanning the period from 1980 to 1996 to generate their time series. Satellite orbital decay was taken into account by using only nadir view data (a decaying orbit will only affect the side looking off-nadir views). Diurnal, instrument calibration, and IBE effects were all accounted for in a simple and direct manner. First, PR limit the data from each satellite to only 3 years to minimize the impact of diurnal drifts. Then, all 3 error sources are accounted for by directly analyzing differences between the AM and PM data from each satellite overlap period and qualitatively estimating the simultaneous impact of all three. Then, the individual records were unified into a single time series using the usual least squares method. Analysis was limited strictly to MSU channel 2 (middle troposphere) so as to avoid the noise inherent in MSU2LT. They obtained a middle troposphere trend of 0.109 deg. K/decade – considerably different from that of UAH, and in better agreement with the predictions of AOGCM’s and theory (Prabhakara et al., 1998).

In November of 2000 PR published an updated analysis that refined their earlier work with explicit analyses of diurnal drift, instrument errors and IBE (Prabhakara et al., 2000). PR noted that for each satellite average brightness temperature was correlated with diurnal drift, indicating that the temperature changes that resulted from LECT drift impacted the hot target temperature. With this in mind they began by generating separate merged time series for the morning (AM) and Afternoon (PM) satellite data based on the corresponding overlapping records. Having separated the morning and afternoon records, they were able to compare trends with hot target temperature records. These were then used to derive an average hot target factor of 0.03 for all the satellites. With this factor, they derived a new global time series by merging independent satellite records. Their new middle troposphere trend was 0.13 deg. K/decade, also in good agreement with the predictions of AOGCM’s but significantly different from the UAH trend.

The Remote Sensing Systems (RSS) Team

In 2002 another analysis of MSU and AMSU products was published by a team of researchers from the Santa Rosa, CA based company Remote Sensing Systems (RSS). They produced time series for the middle troposphere and lower stratosphere but chose not to analyze the lower troposphere with synthetic channels like MSUTLT due to the noise inherent in these channels. Their time series, which covers tropospheric temperatures from 1979 to 2001, uses MSU Channel 2 and AMSU Channel 5 for the middle troposphere and MSU channel 4 and AMSU channel 9 for the stratosphere in the same manner as UAH analyses. It was the RSS team that first discovered the relationship between satellite orbital decay and measured brightness temperature (Wentz and Schabel, 1998). Their analysis accounts for errors from this decay, diurnal drift, satellite roll, interannual and intra-annual variations in radiometer gain, including IBE and non-linear calibration per the usual pre-launch coefficients (Mears et al, 2003) and spans the period from 1979 to 2001 (TIROS-N through NOAA-16). RSS used analysis methods similar in many respects to those of UAH, but there are a few important differences. Like UAH, they first remove diurnal errors independently of other errors. Then they remove instrument calibration errors, IBE, and satellite offsets with merge calculations that derive hot target factors and a unified time series using the usual least squares technique. Diurnal cycle errors were removed in a manner very different from that of UAH. RSS noted that the UAH method of comparing off-nadir views across the satellite flight path suffers from sampling error and other noise sources. They opted instead to model the diurnal cycle using a model of the cycle which was then directly compared with MSU observations at similar times. This method “leverages” MSU data with higher resolution and avoids the noise sources of the UAH method. Since the diurnal cycle is well understood, and as such, lends itself to modeling, they argued that this results in a better characterization of the diurnal cycle (Mears et al., 2003; 2003c). With the diurnal cycle accounted for, they evaluated overlapping satellite records using running 5-day averages (pentads) to smooth the data, and did merge calculations in a manner similar to that of the UAH team (Mears et al., 2003c), except that their hot target factor calculations are based on ocean data only. Ocean diurnal cycles differ from those of land masses in that they are much smaller and more stable due to the temperature moderating effects of water. This makes them much more predictable and easy to evaluate. Since the hot target factor is characteristic of the satellite itself, it is independent of the scene observed, RSS argued that an ocean only merge calculation would allow hot target factors to be determined more reliably prior to the final merge calculation for trends (Mears et. al., 2003; 2003b; 2003c).

The RSS analysis is noteworthy because their results, like those of PR, differ considerably from those of UAH despite a general similarity in analysis approach. Whereas UAH Version 5.0 indicates a statistically insignificant warming of 0.013 deg. C/decade for MSUTMT (middle troposphere), RSS observes a statistically significant warming of 0.097 deg. K/decade for the same layer. The difference is significant on two counts. First, given the similarities in their approach to analysis and the thoroughness and commitment brought to the datasets by both teams, the discrepancy is a testimony to the difficulties in characterizing MSU and AMSU datasets (there is a common misconception among many today that the satellite data is somehow “more reliable” than surface data – this is not true). Second, RSS’s results are not only higher than those of UAH, their middle troposphere results are also higher than UAH’s lower troposphere results (+0.061 deg. K/decade for Version 5.0 MSU2LT). As noted above, MSU2/AMSU5 receives a significant portion of its signal from the lower stratosphere, which is known to have cooled over the last 25 years due to ozone depletion (Bengtsson, 1999). RSS results will include this, indicating that if anything, their middle troposphere numbers are on the low side, making it likely that the discrepancy is even larger than the numbers imply. The same is true of the PR trend estimates which despite a simpler analysis approach are in good agreement with RSS trends but not with UAH. Like UAH, RSS is updating their analysis on an ongoing basis. As of October 2004, their MSU2/AMSU5 middle troposphere temperature trend was 0.131 deg. K/decade.

Of the MSU/AMSU analysis product produced to date, the UAH and RSS products are the most mature and best characterized in terms of error correction and data inputs, and most discussion of lower atmosphere temperature trends revolves around them. As we will see, though these analyses have many similarities (the two teams have been sharing data and results with each other), the differences in their methods account for nearly all of their differing results.

Vinnikov & Grody (VG)

In September of 2003 another analysis was published which made use of a new and radically different approach to analyzing MSU/AMSU datasets and yielded a new trend estimate that differed from all other analyses to date. Konstantin Vinnikov of the University of Maryland and Norman Grody of the NOAA NESDIS group (VG) analyzed MSU/AMSU data from 1979 to 2002 using a method that is primarily statistical and model based rather than empirical to evaluate errors. In their analysis, MSU/AMSU data is fit parametrically to a mathematical model designed specifically to represent climatic data having diurnal and seasonal cycles (Vinnikov et. al., 2002a; 2002b; 2004). The model characterizes brightness temperature as seen from a satellite platform by superposing two trends – a daily trend that represents the diurnal cycle, and another that represents long-term variations. Each term is represented by a Fourier series which includes terms for the expected variation (physically determined) and a weather dependent anomaly (random variation) which includes all noise sources, including instrument biases. Only nadir and near nadir data are used so as to avoid potential errors associated with limb corrections and decaying satellite orbits. Data between satellite tracks that would ordinarily have been measured by off-nadir views are evaluated from the parametric model. For the merging of data from different satellites VG use pentad smoothing similar to RSS. Then they do merge calculations by deriving the Fourier coefficients in their parametrized function along with a least squares fit of the satellite observations to derive a final unified time series (Vinnikov and Grody, 2003). One of the issues confronting the merging of satellite data is the fact that for any two satellites with overlapping records, the contributions of different error sources might be quite different from other overlaps. Diurnal errors for instance, might be aliased into IBE or instrumental, making it is possible to conflate these errors with each other (though diurnal errors will be linear and instrument errors – notably IBE – will not be). There is therefore, no common correction for all satellites. The VG analytical method was developed to handle such problems. With this analysis, they obtained a middle troposphere trend of 0.22 – 0.26 deg. K/decade. This trend provides the best fit to date with AOGCM predictions, but is considerably different from the trends obtained by all other teams. The VG method avoids the problems inherent in characterizing independent error sources without aliasing them into each other, but is strongly dependent on the quality of their modeling. If this is good, their results will be more reliable than those of other teams. But if it is not, the resulting errors could be considerable. It must also be noted that their analysis differs from those of the other teams in two respects. They characterize diurnal effects with a single cycle and do not separate ocean and land cycles from each other. In addition, because their model is derived based from known climate patterns, it is not clear to what extent it can account for IBE and non-linear instrument errors, as these are satellite dependent and unrelated to atmospheric effects. It is likely that some of this error will be absorbed, but not all of it. This introduces a degree of uncertainty into their analyses, VG have argued that this is not enough to account for the low trend estimates of UAH.

Determining Trends from MSU/AMSU Products

Figure 8 summarizes the tropospheric temperature trends derived from the most current MSU/AMSU analysis products. Despite the dedication of all teams, the spread of values shown is striking. For the preceding 25 years, UAH finds relatively little troposphere warming, but Vinnikov and Grody find warming consistent with the predictions of most current climate models and even greater than the observed surface warming. The results of the RSS and Prabhakara teams are in between.

It is evident that the uncertainties and differences in methodology of these analysis products have taken their toll and must be sorted out before a reliable picture of troposphere temperatures can be obtained. With the exception of Vinnikov and Grody, who used a radically different analytical method, each of the teams used basically the same approach, but with different methods of treating the individual sources of error in data collection. Of these methodological differences, 3 in particular, stand out: Satellite data merging methods, diurnal corrections, and Antarctic sea-ice contamination.

Diurnal Corrections

Removal of the diurnal sampling effect may be done in one of two ways - empirically using direct satellite measurements, or with physically based math models in conjunction with measurements. The UAH team adopted the former approach. They compared nadir and off-nadir Tb measurements to estimate the rate of drift. At any given equatorial crossing, the nadir view gives the true Tb at LECT, and off-nadir views give Tb at earlier and later times. View 1 represents a Tb that is 80 minutes earlier than View 11. By comparing these differences for millions of cross track scans throughout the lives of all satellites, UAH was able to estimate the temperature shifts that would result from any given satellite’s drift in LECT. These estimates were made separately for ascending and descending satellites. A constant bias due to satellite roll is removed by comparing the two. The final estimates of diurnal temperature shift were then applied to the MSU/AMSU record with appropriate adjustments for the differences between land and ocean diurnal cycles 4 (Christy et al. 2000; Christy et al. 2003). The advantage of this method is that it is based strictly on satellite observations with few extraneous assumptions. But because it is dependent on a comparison of nadir and off-nadir views, it is subject to sampling noise so that zonal averages must be used to calculate the diurnal cycle drift. The diurnal cycle cannot be removed separately for each measurement grid point (Mears et al., 2002; 2003c).

Given this inherent noisiness, the RSS team adopted the second approach. Diurnal cycles are well understood, particularly over oceanic areas where the cycles are moderated by the large thermal capacity of the world’s oceans, so it is possible to construct highly accurate models of them that can be calibrated by direct MSU observations. RSS calculated a climatology of local diurnal Tb anomalies using the NCAR Community Climate Model (CCM3) (Hack, 1995; Kiehl, 1996; 1998; 1998; Hurrell et al., 1998). This model, which is robust and has been tested for its ability to replicate broad features of global climate, produces surface and atmospheric temperature profiles on a 128 x 64 grid of the earth. Output from this model is then used in models of surface emissivity and radiative heat transfer to calculate observed brightness temperature at MSU/AMSU sensors (Wentz, 1998). Corrections for oceanic surface roughness and variations in emissivity with sea surface temperature (SST) were included, and an emissivity of 0.95 was assumed for land cycles. Using 5 year’s worth of hourly output from these models, a 5 year time series was constructed for each view angle at LECT. Results were tested by comparing the model predictions with direct MSU/AMSU measurements for ascending and descending cycles at the same location. To reduce noise, comparisons were made only for the 5 central MSU/AMSU fields of view, with the 4 off-nadir views corrected to nadir (Mears et al., 2003; 2003c). Figure 9 shows global MSU/AMSU brightness temperature differences between ascending and descending cycles for the month of June as measured by the center 5 fields of view (top) and as calculated by CCM3 (bottom). It can be readily seen that the agreement between the two is quite good. Since the calibration of this model is done on a global scale using MSU/AMSU near-nadir data for which noise levels are small compared to the magnitude of the diurnal cycle, this method avoids the uncertainties of the noisier UAH method, and the agreement between MSU/AMSU observation and the modeled cycle validates the calculated cycle as well (Mears et al., 2002; 2003c).

Intersatellite Data Merging Procedures

The MSU/AMSU record spans a period of nearly 25 years and uses data gathered from 9 satellites whose service lives overlap at various points during this period. Once the appropriate diurnal corrections have been made to the data of each, their records must be merged into one continuous time series. But each of these satellites will have an offset relative to the others due to the various linear and non-linear instrument errors already described. The most notable of these is the Instrument Body Effect (IBE) discussed previously. The exact cause of IBE is not known, but it is thought to result from a combination of uncertainties in the calibration of MSU and AMSU non-linear radiometer gain prior to flight, and a drift in radiometer gain that is proportional to the hot calibration target temperature (Christy et al., 2000; 2003) and has been observed in all TIROS-N class satellites. IBE is likely to be hardware related, it will therefore be unique to each satellite and independent of the flight path and the scene being viewed at any time. In addition to it there will be remaining residual errors not accounted for already in other corrections.

Because of these errors, no two MSU/AMSU records will be measuring temperature from the same reference point – all will be somewhat “out of sync” with each other. Before a consistent time series can be produced, the independent records for all satellites must be unified – that is, merged – to a single averaged record that minimizes the discrepancies between it and the individual records (e.g. the “residuals”). The standard procedure for doing this begins by smoothing each satellite’s record with “running averages” to remove sampling noise and short-term fluctuations, after which each record is expressed as a function of the actual measured brightness temperature, the product of the hot calibration target temperature and a constant target temperature factor, and a term for any remaining non-linearities. Then, for each period where any two satellites are co-orbiting and there are valid overlaps in their data, the differences between each set of simultaneous measurements can be expressed as a difference equation in terms of these functions. The full set of these for all overlapping records provides a large set of simultaneous equations in 17 unknowns (9 Hot Calibration Target Factors and 8 offsets for the full set of overlapping records). Since there is no independent reference temperature for calibration, one of the satellite’s observed temperature records (typically, NOAA-10) is arbitrarily set to zero as the datum. Solving these equations simultaneously using multiple regression techniques determines the target factors. Then, using these Target Factors and least squares methods to minimize the residual differences between the solution curve and the individual satellite records, a unified time series for the entire MSU/AMSU history can be constructed. The number of equations that will need to be solved depends on the type of averaging used. The effectiveness of this procedure is dependent on the lengths of the overlapping records – the longer these overlaps are, the more effectively differences can be characterized and removed. Figure 10b shows the service lives of the first 9 TIROS-N series satellites (up to NOAA-14) and their respective periods of service life overlap. It can be seen that of all overlapping periods of service, the NOAA-9/NOAA-10 and NOAA-7/NOAA-9 overlaps are unusually short – NOAA-9/NOAA-10 in particular being only about 90 days. The shortness of these overlaps has proven to be one of the thornier problems facing MSU/AMSU analyses. How they were handled by each team accounts for most of the differences between their results.

The UAH team began by applying 60 to 110 day smoothing to the individual records. Short overlaps were dealt with by setting a minimum acceptable overlap time for any dataset to be used in their analysis. This lead them to discard the TIROS-N/NOAA-6, NOAA-7/NOAA-9, NOAA-8/NOAA-9, NOAA-9/NOAA-10, and NOAA-10/NOAA-12 overlaps from their analysis. Figure 10a shows the selection of overlaps used in their analysis against the full record. The remaining overlaps were then used to for the derivation of hot target factors. Land and ocean data from each record were used to provide fully global coverage. Once the target factors were determined, the merging of records was accomplished using the selected overlaps to create a “backbone” against which all satellite records were compared to generate the final averaged time series (Christy et al., 2000; Christy et al., 2003). With this method, UAH avoided the noise and sampling problems presented by the shorter overlaps and had a more global basis for their derived hot target factors. According to the UAH team, their choice of overlaps to omit was guided by a decision to give top priority to minimizing the intersatellite trend differences – a goal which they appear to have achieved, reporting an RMS of all intersatellite annual trend differences of only 0.008 deg. C/yr (Christy et al., 2004). But a price was paid for this. While this is a desirable objective for intersatellite datasets, and one they have arguably achieved, the selection of overlaps used was based on a somewhat arbitrary choice of minimum required overlap, and the resulting decadal trend is fairly sensitive to which overlaps are omitted. This raises the possibility that the UAH merge minimized the trend itself rather than the uncertainty in the overall trend measurement (minimizing the actual offset residuals in a merge does not necessarily minimize the uncertainty in trend, especially where some record overlaps have been excluded from analysis). In addition, though the use of land and ocean data for calculating hot target factors gives them an observationally based global evaluation, diurnal temperature fluctuations over land are far greater both temporally and geographically, leading to increased noise in the hot target factor calculation.

By contrast, the RSS team used a more unified approach. Their datasets were smoothed using running 5-day averages (pentads) rather than the longer period smoothing of the UAH team, allowing them to include all the data equally weighted. Shorter overlaps (e.g. – NOAA-9/NOAA-10) are de-emphasized because of they are a smaller portion of the overall time series, but they are not ignored altogether. Target factors and intersatellite offsets are simultaneously calculated for all satellite record overlaps from the same set of equations, and the unified time series and trend are produced from this (Mears and Wentz, 2003b). The use of all satellite datasets is an important advantage of this method. No dataset is arbitrarily ignored, making their analysis insensitive to any arbitrariness in what data is used to calculate a backbone for their time series. Yet because they use of pentad smoothing of their data rather than longer period smoothing, they can still account for the problems of shorter overlaps without introducing any arbitrariness in their time series associated with ignoring them. A comparison of overlaps used in the analyses of UAH and RSS is shown in Figure 10. Another important difference is that the RSS team calculated their target factors using ocean only data. Due to its large heat capacity, water retains temperature very well making oceanic diurnal temperature fluctuation more consistent and far less subject to geographical noise (it is no coincidence that extreme seasonal temperature differences are more characteristic of central continent regions such as the American Midwest, and coastal regions like the Pacific Northwest are more seasonally moderate). Also, unlike land regions, oceanic brightness temperature fluctuations are typically small on diurnal and seasonal scales. This relative stability allows for target factors to be calculated with more precision. This is important because the diurnal fluctuation is part of the data sampled when evaluating target factors, and this introduces noise in the target factor calculation that is not accounted for directly by the diurnal corrections themselves. Since hot target factors are ultimately a property of the satellite hardware itself rather than the scene being viewed, once determined, they may be used for all analysis. Thus the RSS team has been able to determine hot target factors with more precision and make use of a more complete and less arbitrary dataset when evaluating their decadal trends.

Figure 11 gives the 1979 to 2001 decadal trend in global Channel 2 (AMSU Channel 5) brightness temperature as determined by UAH Ver. 5.0 (Christy et al., 2003) and RSS Ver. 1.0 (Mears et. al., 2003). Figure 11a gives the RSS regional profile, Figure 11b shows the UAH trend, and the differences between the two is in Figure 11c. It can be seen from Figure 11c that despite the differences between UAH and RSS global trends, the two teams are in good agreement with each other over most of the northern hemisphere above 30 deg. North latitude. They differ markedly though for the tropics and high southern hemisphere (particularly the southern Pacific Ocean), where the UAH team sees much more cooling than RSS. Other notable differences occur in northern Africa where RSS observes more warming, and the Himalayas where UAH observes more warming. Thus, the differences between the global trends observed by these teams has a strong regional component weighted heavily toward the tropics and southern hemisphere. Furthermore, the most significant regional differences between the two analyses are larger than global trends derived.

The Method of Vinnikov and Grody




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 >>
Climate Change
General Science
Troposphere Temperatures
Negative Climate Feedbacks
The Hockey Stick
Polar Ice-Caps & Sea-Level Rise
Solar Climate Forcing
Resources & Advocacy
Christianity & the Environment
Global Warming Skeptics
The Web of Life
Managing Our Impact
Caring for our Communities
The Far-Right
Ted Williams Archive