Climate Change & Tropospheric Temperature Trends

Poor Vertical Resolution

Unlike radiosondes, which take discreet measurements at specific altitudes, all MSU products measure weighted averages of large portions of the sky. This is both a blessing and a curse. Bulk temperature measurements of atmospheric layers are more likely to be stable than highly localized vertical temperature profiles and meaningful for evaluation of long-term climate change. But at the same time, there are many large-scale phenomena important to tropospheric temperatures that occur on vertical and geographic scales that MSU products cannot resolve. For instance, tropospheric heating is impacted by various vertical vs. horizontal heat transport mechanisms that at times can isolate the upper atmosphere from the surface (NRC 2000; Spencer and Christy, 1992b; Gaffen et al., 2000b; Trenberth and Stepaniak, 2003; Trenberth and Stepaniak, 2004). MSU measurements are too coarse to resolve these effects, yet they can significantly impact local and globally averaged tropospheric temperature trends.

MSU channels also see portions of the sky that are expected to have different long-term behavior due to global warming. Channel 2 for instance, sees the mid and upper troposphere which are expected to warm under natural and anthropogenic forcings, but it also sees the lower stratosphere which is expected to cool due to ozone depletion and other effects (we will see later that this has recently been proven to be a crucial point). Some of the stratospheric “noise” problem can be avoided by using mixed channels such as MSU2LT and MSUTLT. Since these Channels use difference methods to weight observations toward the surface, they reduce the lower stratospheric signal “pollution”. But these difference methods rely on side looking views that because of the curvature of the earth and the increased length of optical viewing paths (which are long compared to temperature gradients over the regions being viewed). As such, they significantly amplify the sampling noise between contributing channels which more than doubles the error in MSU2LT vs. MSU2 (Spencer and Christy, 1992b; NRC 2000). Similar problems plague the later MSUTLT record. MSU2LT and MSUTLT receive a portion of their signal from the surface. The NRC (2000) estimated this contribution to be roughly 10% over ocean, 20% over land (NRC 2000; Spencer and Christy, 1992b). This might pollute their lower troposphere data with surface effects that are even more problematic than the stratospheric impacts on Channel 2, though there is some dispute about this. Spencer and Christy estimated the impact of this noise at approximately 0.03 deg K/decade (Spencer and Christy, 1992b) but recent estimates vary from those yielding errors as high as 0.1 deg K/decade, making the noise from MSU2LT and MSUTLT channels equal to or greater than that of most extant trend measurements from these products (Stendel et al., 2000), to those that show a negligible contribution (Litten et al., 2005). These factors could complicate how MSU data are to be weighted and widely varying answers can be obtained from a few equally reasonable assumptions (NRC 2000). One other factor has come to light recently as well. It now appears that the lower troposphere MSU2LT and TLT products might be polluted at high southern latitudes by Antarctic sea-ice. Because these channels receive 10 to 20 percent of their input from the surface and sea-ice has a very high albedo compared with the ocean, an increase in sea-ice will appear to them as spurious warming, a decrease will appear as cooling. The same affect will be present at the North Pole, but to a diminished degree owing to the smaller fraction of sea-ice present at higher latitudes. There is evidence to suggest that strong cooling trends at high southern latitudes in 2LT and TLT products may be related to annual and inter-annual sea-ice variations rather than lower troposphere temperatures (Swanson, 2003).

Short Temporal Record

MSU and/or AMSU products have been in service since 1979, making their time series about 26 to 27 years long as of this writing. This is too short for useful predictions of global change. Anthropogenic impacts on global climate (at least, as they relate to current concerns about global warming) are thought to have originated at the dawn of the Industrial age over 2 centuries ago, and can only be clearly separated from natural effects such as solar variability since the mid-20th century (NRC 2000; IPCC 2001). Against this time scale, the MSU record provides an incomplete time series at best. Many tropospheric and stratospheric phenomena which impact this record have occurred on time scales that make up a significant percentage of the MSU record as well. During the last 25 years there have been 4 ENSO events (El Nino Southern Oscillation) in 1983, 1987, 1991, and 1997, and at least 2 volcanic eruptions (El Chicon in 1982 and Mt. Pinatubo in 1991) that are known to have significantly impacted both stratospheric and tropospheric temperature time series. These impacts have ranged from one to five years so that collectively they have impacted a majority of the MSU record to at least some degree (Christy and McNider, 1994; Jones, 1994; Santer et al. 2001; Wigley and Santer, 2002). These effects must be separated from the background tropospheric record before meaningful predictions can be made. Thus, it is unlikely that the last 27 years will be representative of decades to come and care must be taken in basing predictions on them (NRC 2000).

Onboard Instrumental Biases

MSU sensory devices also suffer from a number of systematic equipment errors. The most important of these is drift in sensor gain (i.e. – the ratio of the perceived signal to the actual signal). MSU sensors detect upwelling atmospheric microwave emissions, amplify them for processing, and convert the radiosity measurements into temperatures based on known physical laws that relate the two. In the original MSU design, microwave emissions are detected by the two MSU antennas and juxtaposed with counts from the hot target microwave load at a known instrument temperature and deep space (2.73 deg. K) by the Dicke switch. Then they are passed through a 10-110 MHz low noise mixer, reamplified, and digitized by a 12 bit analog-to-digital converter to obtain the final data that is sent to ground stations. MSU sensors record temperature as “digital counts” which correlate to a microwave radiosity that in turn correlates to a temperature. The relationship between radiometer counts and temperature is non-linear, and the non-linearity must be calibrated for each MSU package prior to launch. If this calibration is not done accurately, non-linear errors in radiometer gain will result that can be very hard to correct for later. It now appears that this has happened for some of the MSU packages, NOAA-12 in particular (Mo, 1995). Uncertainties in non-linear calibration coefficient like these cannot be removed later with hot target calibrations.

In addition to direct equipment calibration non-linearities, another non-linear source of error has been identified that varies with the temperatures of the hot target and the physical radiometer components themselves. Though the hot target is in principle fixed at a known temperature, most experience at least some drift in temperature over their service lives. As the sun illuminates radiometer components from different angles during the satellite’s orbit, the sensor components themselves will also vary in temperature. These component temperature variations will affect the gain of the radiometer components and the temperature calibration in a non-linear manner. This has come to be known as Instrument Body Effect (IBE). Some of this IBE can be measured and corrected for in post launch analysis because differences in perceived temperature between pairs of satellites will identify those satellites with non-linear gain anomalies and allow for the combining of all overlapping satellite records in ways that reduce the resulting error – a process known as “merging”. A gain correction function that varies with time can then be applied to those satellites for which a non-linear signal response is found (Christy et al., 2000). In addition, over longer time series and many measurements much of the noise from these variations will average out and the correction functions will become more reliable. In this manner, a significant portion of non-linear noise can be accounted for. But this process cannot remove all errors, particularly residual calibration errors not correlated with radiometer component temperature, and some errors will persist (NRC 2000).

Orbital Decay and Attitude Drift

During its service life of 2 to 8 years, episodic solar wind events will cause the satellites to experience some orbital and attitude degradation. As the spacecraft orbit decays it will lose altitude, and MSU and AMSU sensors will see different portions of the sky from distances other than those expected (Wentz and Schabel, 1998). Slight variations in satellite attitude (roll orientation) will also alter MSU scan views. Nadir will be slightly off of nadir for instance, and other views will also be similarly degraded. This will have the effect of eroding the weighting functions for each scan and introducing spurious temperature variations (Mears et al., 2002). Fortunately, these two effects can be accounted for. The orbital mechanics for the POES spacecraft are well known and altitude degradation can be precisely determined. Loss of altitude can be combined with corrections to radiative transfer functions to remove orbital decay effects. Satellite attitude bias is systematic across all scan views so that when combined with known orbital decay, roll induced variations can be adjusted to the nadir view. These corrections have been applied to most recent datasets (Wentz and Schabel, 1998; Mears et al., 2002; Christy et al., 2003).

Diurnal Sampling Variations

As already mentioned, TIROS satellite orbits are sun synchronous with a daily orbital precession rate of 0.986” that preserves orientation and LECT’s throughout the year. However, for most TIROS platforms, this precession rate is not perfect, and slight variations in it introduce a drift in orbital path with respect to the sun and earth. This drift over extended periods causes the daily equatorial crossing times of each satellite to occur earlier or later in the day than expected with respect to local ephemeris time. Since temperature varies on a diurnal cycle from day to night, this introduces another spurious temperature variation into MSU products. Drifts of up to 0.5 hr/year have been observed. Diurnal sampling errors, though systematic in their effects, occur randomly and correlate imperfectly with observations, which can make them difficult to account for in datasets. One approach is to compare data from view angles to the left and right of nadir at predetermined daily times during ascending (northbound) and descending (southbound) orbital segments. Data from different view angles will correspond to different local times at different points in satellite orbit – from roughly an hour at the equator to several hours at the poles. Observations of these differences over many ascending and descending orbital cycles allow for a drift rate to be backed out of the data variations (Christy et al., 2000). The approach is straightforward, but the resulting calculation is quite involved and compounds many other uncertainties making the derivation of reliable numbers troublesome. Another approach is to use high resolution climate model simulations to recreate the diurnal brightness temperature variations expected along the satellite orbital path. The data from such models may then be used to correct each measurement to local noon and remove diurnal drifts. The modeled diurnal cycle temperature variations can in turn be validated against actual morning to evening temperature variations as observed by MSU platforms themselves (Mears et al., 2002). With proper validation of the climate modeling against observations, this methodology has the advantage of considerably reduced sampling noise compared to the cross scan methodology.

Intersatellite Data Drift

To date, NOAA has operated 11 TIROS-N and/or TIROS-ATN Class satellites since 1978. In theory, these spacecraft and their instrument packages should all behave identically. The MSU and AMSU packages in particular all contain the exact same components (per class) and are all calibrated prior to launch according to the same guidelines and protocols. In theory! In practice however, significant differences in measured brightness temperature have been observed between pairs of satellites whose service lives overlap - up to 0.4 deg. K (Hurrell and Trenberth, 1998; Christy et al. 1998; NRC, 2000). Much of this difference is related to IBE but not all of it. It is not clear where the remainder is coming from (NRC, 2000). Analyses to date have accounted for this drift by comparing data from pairs of satellites whose service lives overlap each other. Typically, during service life overlap, data from co-orbiting pairs of satellites are smoothed with running averages to remove noise and facilitate comparison. Then, after correcting for other noise sources, a single continuous time series can be created from the independent records by generating a single curve that minimizes the differences with the individual records – in other words, “merging” them. Typically, merge calculations of this sort are done with a least squares mathematical technique and the reliability of the merge is evaluated with a statistical analysis of the residuals (the leftover differences between the final time series and the individual satellite records).

For this to work, it is important to have a long enough overlap period between co-orbiting satellites for a statistically significant comparison of records. In most cases, overlaps are long enough for this to be done. But a few of the overlaps, particularly the NOAA-9/NOAA-10 overlap, are not. NOAA-10 was put into service in September of 1986. The NOAA-9 MSU unit failed unrecoverably in May of 1987 and produced no more data after that (the satellite was deactivated in February of 1998). As a result, only 90 days of overlapping data were generated between these two satellites - not adequate for a truly solid merge of the two. Accounting for this overlap has been the most problematic challenge facing the MSU/AMSU record. The merging procedure must also correct for the IBE and other equipment non-linearities as well. But for the reasons noted above, this can only be done to a certain extent.

Based on these considerations, differing dataset merging methodologies have been applied to the MSU/AMSU record, resulting in trend differences that at one time or another have spanned a 0.76 deg. K/decade spread (Mears et al., 2002, Christy et al., 1998; 2000; 2003; Vinnikov & Grody, 2004). In some cases the short overlaps have been omitted entirely from long-term datasets (Christy et al., 2000; 2003). In others, it has been used and evaluated against the remaining long-term trends (Mears et al., 2002). Of these, the methodologies with the lowest error characteristics and the best fit with overall long-term trends have been used and corrections have been made as the sources of error were discovered (Christy et al., 1998; Christy et al., 2003; Mears et al., 2002). But even after nearly 15 years of error removal, there are still considerable differences in the results arrived at by the latest analyses. To date, it appears that the poor overlap of the NOAA-9 record with that of conjoining satellites is the largest contributor to MSU/AMSU dataset uncertainties, accounting for at much a 65 percent of the total spread between currently existing analyses based on MSU/AMSU products (Karl et al., 2002).

Antarctic Sea-Ice and Melt Pool Albedo Impacts

MSU and AMSU devices measure atmospheric temperatures as digital counts of radiation under the presumption that the primary source of this radiation is upwelling emissions from the atmosphere that follow the weighting functions of each channel (see Figure 7). They are not able by themselves to determine where the radiation counts they are measuring originated from. Thus, the accuracy of their measurements of troposphere temperatures and trends depend on the presumption that all digital counts are the result of atmospheric radiation and the reliability of the channel weighting functions. The high southern latitudes (above 60 deg. S) contain significant portions of sea-ice of Antarctic origin. Likewise, high northern latitudes exhibit significant sea-ice fraction of their own, though these are of considerably smaller extent than their Antarctic counterparts. During the austral summer melt season this sea-ice shrinks to an area of roughly 3 million km², and then increases to 18 million km2 during the winter and spring (Cavalieri et al., 1997). Because of its high albedo, sea-ice area increases will increase the surface generated digital counts to MSU and AMSU sensors, and will appear to them as a spurious warming. Likewise, decreases in this sea-ice and the corresponding increases in melt pond area will appear as a cooling. Sea-ice area data has been gathered since the late 70’s by Scanning Multichannel Microwave Radiometer (SMMR) and Special Sensor Microwave/Imager (SMM/I) devices flying aboard NASA Nimbus-7 and Defense Dept. Meteorological satellites. These devices are passive microwave detectors similar to the MSU and AMSU devices aboard NOAA POES satellites, and measure sea-ice as brightness temperature readings. Data from these platforms shows that between 1979 and 2002 Arctic sea-ice shrunk by about 360,000 km2 and overall Antarctic sea-ice increased slightly, though West-Antarctic sea-ice decreased (Cavalieri et al., 1997). However, radiosonde data from the Antarctic region also indicates that during the satellite era there may have been lower troposphere warming as well as high troposphere cooling due to ozone depletion (Thompson and Solomon, 2002). In addition, the breakup of the Larsen A and B ice shelves (in January of 1995 and March of 2002, respectively) also might have contributed to relative sea-ice and relative ocean/melt pool areas. The Antarctic peninsula has warmed considerably in recent decades, as shown by various direct and proxy measurements reveal (Vaughan et al., 2001). The resulting warmer temperatures have likely increased melt pool area on the ice that remains during austral summer. These melt pools are indistinguishable to MSU and AMSO devices from the surrounding ocean, and thus must be considered as part of the surface radiation balance from higher latitudes. Recent comparisons of daily anomalies at southern latitudes higher than 60 deg. Has revealed a strong correlation between Antarctic ice summer melt periods and 2LT and TLT cooling trends (Swanson, 2003). Due to its higher altitude weighting, these fluctuations will exhibit only a minor impact on the Channel 2 trend and thus will have little impact on the MSU/AMSU products of the RSS, Prabhakara, and VG teams.

MSU/AMSU Analysis Products

It has already been noted that the MSU and AMSU records were originally intended for regional weather forecasting and climate studies. For applications such as these, the errors described above are usually small in comparison with the data being generated and are of little consequence. But when studying long-term global climate change, small errors can add up quickly, and a great deal of data massaging is necessary to make them useful. One can get an idea of the degree of massaging involved from the fact that after nearly 15 years, the most mature of these analyses has recently undergone it 5th major revision (Christy et al. 2003) and one update of it (Christy et al., 2004). Yet its conclusions differ considerably from those of other analyses based on different data reduction methods.

As of this writing, there are 4 noteworthy analyses of tropospheric temperature trends based on MSU and AMSU products. Each makes use of a different approach to dealing with the uncertainties in MSU and AMSU records and each has its strengths and its weaknesses. All account for solar wind induced orbital decay and known calibration related instrument biases. The chief differences between them lie in how they deal with diurnal sampling errors and IBE, and how they merge independent MSU/AMSU records to create a 20 to 25 year continuous time series.

The University of Alabama, Huntsville (UAH) Team

The earliest use of MSU products for climate change studies dates to the early 90’s when a team led by John. R. Christy of the University of Alabama, Huntsville and Roy W. Spencer of the NASA Marshall Space Flight Center published the first revision of their MSU based lower troposphere studies. Christy and Spencer’s team (hereafter referred to as UAH for University of Alabama, Huntsville) assembled a full time series of MSU data from 1978 to 1990 for the lower troposphere. It was the UAH team that derived the MSU2LT and MSUTLT synthetic channels to emphasize the lower troposphere (peak weighting at roughly 3.5 km) by differencing Channel 2 nadir and off-nadir views. They also developed another synthetic channel known as MSU2RT in their earlier analyses, though this product is seldom used any more, having been discarded in favor of the better characterizes 2LT and TLT products.

In their first analysis, UAH combined data from TIROS-N through NOAA-12 using a simple merging procedure to remove then known instrument biases and formed a consistent lower troposphere temperature time series that was designated as UAH Version A (Spencer & Christy, 1990; Spencer & Christy, 1992a; 1992b). By 1994, enough new data had been gathered to justify a reanalysis. It had been discovered that NOAA-11 had experienced eastward drifts in its diurnal cycles over its service life that increased its LECT’s and introduced a spurious warming into the dataset. It was at this time that UAH developed the MSU2LT channel to replace MSU2RT. Using this new channel, and incorporating updates to corrections for missing data and satellite overlaps, Version B was released (Christy et al., 1995).

At the time, Versions A and B spurred much debate because they indicated a decadal cooling trend for MSU2R and MSU2LT. Version B showed a cooling of -.058 deg. K/decade, which was 0.03 deg. K/decade cooler than that of Version A. These trends were inconsistent with what climatologists expected for the lower troposphere based on the predictions of the best AOGCM’s of the day. Shortly after the release of Version B, it was discovered that NOAA-12 was suffering from unusually large instrument biases, and that NOAA-7 had experienced a considerable drift in it LECT compared to that of NOAA-11. Correcting for these errors lead to Version C (Christy et al., 1998), which produced a trend that was 0.03 deg. K/decade more positive than that of Version B. After the release of Version C an additional source of instrument gain error was discovered that impacted MSU2 and MSU2LT. A new set of NES-DIS non-linear calibration corrections became available for NOAA-12 which allowed for more accurate estimates of its non-linear instrument biases. Then, it was discovered that solar wind drag had caused the orbits of all NOAA POES satellites to decay, resulting in losses of altitude of up to 15 km. These altitude changes introduced spurious cooling into the MSU time series (Wentz and Schabel, 1998). Incorporation of these effects lead to the release of UAH Version D, which yielded a trend of +0.06 deg. K/decade for MSU2LT and 0.04 deg. K/decade for MSU2 (Christy et al., 2000).