Stratospheric Processes and Their Role in Climate

1College of Ocean and Meteorology, Guangdong Ocean University, Zhanjiang, China 2George Mason University, Fairfax, VA, USA 3Cochin University of Science and Technology, Cochin, India 4Key Laboratory of Meteorological Disaster, Ministry of Education/Joint International Research Laboratory of Climate and Environment Change/Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters, Nanjing University of Information Science and Technology, Nanjing, China 5Chinese Academy of Meteorological Sciences, Beijing, China 6Hampton University, Hampton, VA, USA


Preface
SPARC has from its outset been concerned with "Stratospheric Indicators of Climate Change," "Stratospheric Processes," and "Troposphere/Stratosphere Modelling." The SPARC project GRIPS (GCM Reality Intercomparison Project) has focused on comparing troposphere/stratosphere general circulation models with one another, both in terms of their technical formulations and in their results. Of course, another aspect of GRIPS is to examine how well model results compare to observations. Direct model/data comparisons are not so straightforward, however. For instance, the stratosphere displays a great deal of interannual variability, so that model-data comparisons necessarily include statistics of both means and variances over comparable time periods. Additionally, stratospheric data are obtained from complicated inversions of radiances derived from satellite measurements, from direct but sparse balloon or rocket measurements, from time continuous but geographically sparse ground-based remote sensing instruments, and finally from analysis of stratospheric measurements either by statistical techniques or from data assimilation methods. The climatologies derived by these different methods do not agree in all respects. Finally, the entire concept of stratospheric trends means that stratospheric climatology is itself time varying.
The SPARC Reference Climatology Group was established to update and evaluate existing middle atmosphere climatologies for use in GRIPS, and in other SPARC activities. Rather than create a single new "super climatology," it was decided that a valuable contribution would be to (1) compile existing climatologies and make them easily available to the research community, and (2) carefully compare and evaluate the existing climatologies. The SPARC Data Center was established (in part) as a response to item (1), and this Report is a response to item (2). Here we present detailed intercomparisons of climatological wind and temperature data sets that are currently used in the research community, which are derived from a variety of meteorological analyses and satellite data sets. Special attention is focused on tropical winds and temperatures, where large differences exist among separate analyses. We also include comparisons between the global climatologies and historical rocketsonde temperature and wind measurements, and also with more recent lidar temperature data. These comparisons highlight differences and uncertainties in contemporary middle atmosphere data sets, and allow biases in particular analyses to be isolated. In addition, a brief atlas of zonal mean wind and temperature statistics is provided to highlight data availability and as a quick-look reference. This Report is intended as a companion to the climatological data sets held in archive at the SPARC Data Center (http://www.sparc.sunysb.edu).

-Introduction
Climatological data sets for the middle atmosphere are useful for empirical studies of climate and variability, and also necessary for constraining the behaviour of numerical models. The earliest comprehensive climatologies for the middle atmosphere were the 1964 and 1972 COSPAR reference atmospheres (CIRA), which were based largely on interpolation of single station balloon and rocket data. An updated version of CIRA in 1986 included early satellite observations of the stratosphere and mesosphere and has served as a community standard since that time. Around 1979, daily meteorological analyses with significant stratospheric coverage that included operational satellite temperature soundings began, and more recently (~1991) sophisticated model-based data assimilation schemes began to produce stratospheric analyses.
These analyses (supplemented in the 1990's by extensive retrospective reanalyses) have served as the basis for some more recent middle atmosphere climatologies.
Also, satellite observations from the Upper Atmosphere Research Satellite (UARS), launched in 1991 and continuing to operate in 2002, have provided additional climatological data sets for the middle atmosphere. Details of the circulation statistics derived from these various data sets will depend on several factors, including details of data inclusion and analysis techniques, and the respective time periods covered.
The objective of this work is to bring together several middle atmosphere climatological data sets, which are in current use in the research community, and make direct comparisons of some basic measured and derived quantities. These data sets are based on a wide variety of analysis techniques, including manual analyses, objective statistical analyses, and data assimilation systems. Our comparisons are used to identify biases in particular data sets, and also to highlight regions where there is relatively large uncertainty for particular diagnostics (i.e., where large differences are found among several data sets). Where possible, we provide some brief explanations as to why there are uncertainties and/or why the data sets might differ. However, more in-depth and detailed explanations are beyond the scope of this report.
The middle atmosphere climatologies considered here are primarily derived from global meteorological analyses and satellite data. Two independent data sets are also included for comparison, namely historical rocketsonde wind and temperature measurements (covering approximately 1965-1990), and lidar temperature data (covering the 1990's). Because the time and space sampling of the rocketsonde and lidar data are distinct from the global climatologies, direct comparisons require special attention, as discussed in Section 2C below.
An important aspect of numerical modelling for the middle atmosphere is to simulate not just the time mean structure, but also temporal (interannual) variability (e.g., Pawson et al., 2000). To that end we include here comparisons of both long-term means and interannual variance statistics. We also focus particular attention on the tropics, where there are relatively large differences among middle atmosphere climatological wind and temperature data sets [in particular, for variability related to the quasi-biennial oscillation (QBO) and semi-annual oscillation (SAO)]. We furthermore present some comparisons between the few available sources of mesospheric winds and temperatures. Finally, we provide a brief atlas of middle atmosphere circulation statistics as a quick-look reference, and to highlight data availability.
All of the monthly mean data presented and compared here are available to the research community via the SPARC Data Center (http://www.sparc.sunysb.edu). This atlas is intended to be a companion to those on-line data.

A. Description of fundamental data and types of global analyses
Two fundamental types of observations contribute to global (or hemispheric) stratospheric analyses. Radiosondes are balloon-borne instruments which provide vertical profiles of temperature, pressure and horizontal winds. These measurements cover the lowest 20-30 km of the atmosphere, with a nominal archive vertical resolution of ~ 2 km. The current radiosonde network provides approximately ~ 1100-1200 soundings per day, almost evenly split for measurements at 00UTC and 12UTC; the vast majority of stations are located over land masses of the Northern Hemisphere (NH). Almost all of these soundings (> 1000) reach at least 100 hPa, with ~ 800 reaching 30 hPa, and ~ 350 reaching up to 10 hPa. Satellite-derived temperature profiles provide the other major source of stratospheric data. Operational meteorological polar orbiting satellites provide near-global temperature profile retrievals twice daily up tõ 50 km altitude, but have the drawback of relatively low vertical resolution (> 10 km) in the stratosphere. A series of operational NOAA satellites has been in orbit since late 1978, containing a suite of instruments that are collectively called the TIROS Operational Vertical Sounder (TOVS) (Smith et al., 1979). An improved set of temperature and humidity sounders (called the Advanced TOVS, or ATOVS) is now replacing the older TOVS series, beginning with the NOAA-15 satellite launched in May 1998, NOAA-16 in September 2000, and NOAA-17 in June 2002. Temperature retrievals or radiances from TOVS and ATOVS data are a primary input to many of the global analyses here (including CPC, UKMO, UKTOVS, NCEP, ERA15 and ERA40 data sets).
Details of the various stratospheric analyses are described below, but the types of global or hemispheric analyses can be summarised as follows. The main objective is to depict the large-scale atmospheric behaviour by some sort of interpolation between sparse observations, or by combining different types of measurements (i.e., radiosondes and satellites). The simplest analyses provide global or hemispheric fields based on handdrawn analyses (FUB) or objective analysis gridding techniques (CPC and UKTOVS). More sophisticated analyses can be derived by the use of numerical forecast models to predict first-guess fields, and incorporate observations by optimal data assimilation (UKMO, NCEP, ERA15 and ERA40 data). A strength of the simple analyses is that they are most directly dependent on observations, but this can be a drawback in regions with less data availability. The data assimilation techniques have the advantage of incorporating knowledge of prior observations and dynamical balances, but have the disadvantage that details of the numerical model and assimilation method can influence the results. In fact most of the analyses discussed below are based on very similar radiosonde and satellite data, and so the differences revealed in our comparisons highlight the sensitivity of the final statistics to details of the data usage and analysis techniques.

B. Description of data sets
This section presents short descriptions of the climatological data sets included in the comparisons. These are intended to be brief, and more details for each analysis can be found in the listed references. Table 1 provides a summary list of some relevant details for each analysis. For brevity, each data set is referred to throughout this work by the following short acronyms. Since October 1991 stratospheric analyses have been produced daily using a stratosphere-troposphere version of the Met Office's data assimilation system (Swinbank and O'Neill, 1994). These analyses were formerly referred to as UK Meteorological Office (UKMO) stratospheric analyses, and we retain that name here. The stratosphere-troposphere data assimilation system is a development of the "Analysis Correction" data assimilation system (Lorenc et al., 1991), which was then in use for operational weather forecasting. In this system, observations are assimilated into a 42-level configuration of the Met Office Unified Model to produce a set of stratospheric analyses for 12 UTC every day. The analyses consist of temperatures, wind components and geopotential heights on a global grid of resolution 2.5° latitude by 3.75° longitude. The analyses are output on the UARS standard pressure levels, with six equally spaced levels per decade of pressure (100,68.1,46.4,31.6,21.5,14.7,10,…hPa). The analyses span the range 1000-0.3 hPa (approximately 0-55 km).
The stratospheric analyses were originally produced as correlative data for the NASA UARS (Upper Atmosphere Research Satellite) project, starting from October 1991. The analyses have been used in a number of research studies of stratospheric dynamics, and to help interpret UARS constituent measurements. Since October 1995 the separate UARS assimilation system was discontinued, but stratospheric analyses continue to be produced using a similar data assimilation system, which is run as part of the Met Office operational forecasting suite. Since November 2000 the Met Office stratospheric analyses have been produced using a new 3-D variational (3DVAR) data assimilation system (Lorenc et al., 2000). However, the majority of results presented here were derived from analyses produced before that change. A further note is that the UKMO temperature analyses at the uppermost levels (at and above 1 hPa) were adversely affected by an erroneous ozone climatology in the assimilation model after January 1998, but these do not influence the majority of the comparisons below (which focus on the time period 1992-1997).
The stratosphere-troposphere data assimilation uses essentially the same set of meteorological observations that are used for operational weather forecasting. In the stratosphere the most important observation types are TOVS (and more recently ATOVS) temperature profiles, together with radiosonde soundings of temperatures and winds (although most radiosondes only ascend as far as the lower stratosphere).

UKTOVS (Met Office Analyses of TOVS Data)
The Met Office also produced regular stratospheric analyses (not part of a model assimilation) from measurements made by NOAA polar orbiter satellites. The Met Office was formerly referred to as the UK Meteorological Office, and these analyses have been referred to in the research community as UKTOVS (which we retain here). Monthly means from these analyses are available for the period January 1979-April 1997. The analysis method is described by Bailey et al. (1993), but a brief description is given below. Scaife et al. (2000) present climatological data and interannual variations diagnosed from the UKTOVS data.
The UKTOVS fields are derived from an independent analysis of TOVS radiance measurements. The daily TOVS data were used to derive geopotential thickness values, covering the layers 100-20, 100-10, 100-5, 100-2 and 100-1 hPa. The thicknesses were then mapped onto a 5 degree resolution global grid, and added to the operational analysis of 100 hPa height (obtained from Met Office operational global analyses) to produce height fields up to 1 hPa. In turn, temperatures and horizontal balanced winds are derived from the height fields. Winds at the equator are interpolated from low latitudes, and resultant tropical variations (e.g., the QBO) are rather weak (as shown below).

CPC (Climate Prediction Center)
Operational daily analyses of stratospheric geopotential height and temperature fields have been produced by the Climate Prediction Center (CPC) of the US National Centers for Environmental Prediction (NCEP) since October 1978 (Gelman et al., 1986). NCEP was formerly called the National Meteorological Center (NMC), and these analyses have been referred to in the research community as "stratospheric NMC" analyses.
Here they are termed CPC analyses (to differentiate from NCEP reanalyses). The CPC stratospheric analyses are based on a successive-correction objective analysis (Finger et al., 1965) for pressure levels 70, 50, 30, 10, 5, 2, 1 and 0.4 hPa, incorporating TOVS and ATOVS satellite data and radiosonde measurements (in the lower stratosphere of the NH). This analysis system has been nearly constant over time (October 1978-April 2001. The TOVS temperature profiles are provided as layer mean temperatures between standard pressure levels; geopotential thicknesses are calculated from these temperatures, and added to a base level 100 hPa geopotential field taken from operational NCEP tropospheric analyses. Stratospheric temperatures at standard levels are derived by interpolation between the TOVS layer mean temperatures. The fields are produced each day for a nominal time of 1200 UTC, using 12 hours of TOVS data (0600-1800 UTC). The NCEP operationally analysed tropospheric fields over 1000-100 hPa are included, so that CPC analyses cover 1000-0.4 hPa. As a note, the CPC analyses were changed beginning in May 2001, with the data up to 10 hPa based on the NCEP operational analyses, and fields above 10 hPa based solely on ATOVS. However, comparisons here only include data prior to this change.
Satellite temperatures are the sole input to the CPC analyses over the Southern Hemisphere (SH) and tropics, and over the entire globe for levels above 10 hPa. For levels 70-10 hPa in the NH, radiosonde data (for 1200 UTC) are incorporated, using TOVS as a first guess field. TOVS temperatures have been provided by a series of operational NOAA satellites; these instruments do not yield identical radiance measurements for a variety of reasons, and derived temperatures may change substantially when a new instrument is introduced (Nash and Forrester, 1986). Finger et al. (1993) have compared the CPC temperatures in the upper stratosphere (pressure levels 5, 2, 1 and 0.4 hPa) with co-located rocketsonde and lidar data, and find systematic biases of order ± 3-6 K. Finger et al. (1993) provide a set of recommended corrections (dependent on time period and pressure level) to the CPC temperature analyses, which have been incorporated in the results here. However, in spite of these adjustments, the CPC analyses still probably retain artefacts of these satellite changes. One additional change of note is that before March 1984 NH fields were only archived over 20°N-pole; prior to this time CPC analyses over 0-20°N are interpolated values.
Horizontal winds are derived from the CPC geopotential data using the 'linear balance' technique (Randel, 1987). The calculations of longitudinally-varying wind fields become ill-behaved in the tropics, and eddy wind statistics are only reliable polewards of ~ 20°l atitude. Tropical zonal mean winds are estimated using the second derivative of geopotential height at the equator (Fleming and Chandra, 1989). Extensive climatologies of the stratosphere have been derived from these CPC data by Hamilton (1982), Geller et al. (1983) and Randel (1992).

NCEP Reanalyses
The NCEP/National Center for Atmospheric Research (NCAR) reanalysis project uses a global numerical weather analysis/forecast system to perform data assimilation using historical observations, spanning the time period from 1957 to the present (Kalnay et al., 1996). For brevity, the NCEP/NCAR reanalyses are referred to as NCEP here. The model used in the NCEP reanalysis has 28 vertical levels extending from the surface to ~40 km, and analyses of winds, temperatures and geopotential height are output on stratospheric pressure levels of 100, 70, 50, 30, 20 and 10 hPa.

ERA15 (ECMWF 15-year Reanalysis)
The European Centre for Medium Range Weather Forecasts (ECMWF) produced a global reanalysis for the period 1979-1993, based on data assimilation coupled with a numerical forecast model (Gibson et al., 1997). The forecast model used in that work spanned pressure levels 1000-10 hPa, with analyses output on stratospheric pressure levels of 100, 50, 30 and 10 hPa. An important detail is that the 10 hPa analysis level is at the top level of the model, and this has a detrimental impact on results at this level (as shown in the comparisons below).

ERA40 (ECMWF 40-year Reanalysis)
ECMWF is also producing an updated reanalysis, termed ERA40, covering the period 1957-2001. ERA40 will be a comprehensive set of global meteorological analyses, including the stratosphere up to 1 hPa, based on the use of variational data assimilation techniques. One important difference from the ERA15 reanalyses (in addition to the increased vertical domain) is that ERA40 will directly assimilate TOVS and ATOVS radiances, as opposed to retrieved temperature profiles. Production of the full ERA40 reanalyses is an ongoing activity at the time of writing of this atlas. However, we have obtained a subset of early production results for the time period 1992-1997, covering the parameters zonal mean temperature and zonal mean zonal winds (courtesy of Adrian Simmons of ECMWF), and these are included in our comparisons. The ERA40 analyses are available on 23 standard pressure levels spanning 1000-1 hPa, and also on each of the 60 levels of the assimilation model.

FUB (Free University of Berlin Analyses)
The meteorological analyses from the Free University of Berlin (FUB) are northern hemispheric, gridded products at four levels: 100, 50, 30, and 10 hPa. Monthly mean data at these levels are available since 1957 (geopotential height) and 1964 (temperature) on a 10-degree grid, with an increase in resolution to five degrees in the early 1970s as technological advances made the automated scanning of the hand-analysed charts a practical reality (the precise date depends on the pressure level). Daily analyses are produced only at the three upper levels (i.e., not at 100 hPa) and are provided only every second day in northern summer, when the flow evolves slowly. The analyses are performed by hand (subjective analysis) by experienced personnel using station observations of geopotential height, wind, and temperature; thermal wind balance is used as a constraint on the analyses, so that even though the wind field is not analysed, it plays an important part in the analysis procedure. Hydrostatic balance and the thermal wind are used as the analyses are built up from the 100-hPa tie-on level (for which data from FUB were used in the early years, but later operational products from the German Weather Service were substituted), to build accurate analyses from the station data at the stratospheric levels.
The FUB data have been used in a large number of studies of the climatology of the middle atmosphere, including trends and low-frequency variability (e.g., Labitzke and Naujokat, 1983;Pawson et al., 1993). Daily data have been extensively exploited to understand the occurrence of very cold regions which are associated with polar stratospheric cloud formation and ozone loss (Pawson and Naujokat, 1999). It should be stressed that these analyses do not include wind as a product; while their utility is restricted by this, they are a valuable record of the stratosphere between about 1957 and 2001, analysed in a consistent and uniform manner throughout this period. Full details of the FUB analysis, together with the entire data set, are available in compact disk (CD) format (Labitzke et al., 2002).

CIRA86 Climatology
The COSPAR International Reference Atmosphere, 1986 (CIRA86) of zonal mean temperature, geopotential height, and zonal wind has been described in detail in Corney (1985a, 1985b) and Fleming et al. (1988Fleming et al. ( , 1990. These reference climatologies extend from 0-120 km and are based on a variety of data sources, briefly summarised here. Temperatures for 1000-50 hPa are taken from the climatology of Oort (1983), which is based primarily on radiosonde data from the 1960s and early 1970s. Temperatures at 30 hPa over the NH are taken from FUB analyses, and for the SH are taken from the radiosonde climatology of Knittel (1974 (Hedin, 1983) used exclusively above 0.002 hPa (~ 90 km). All values were merged to obtain a smooth transition between the original data sets. For the geopotential height climatology temperatures were integrated upwards and downwards from the 30 hPa geopotential heights, which were taken from FUB analyses for the NH and Knittel (1974) for the SH.
The zonal wind climatology in the troposphere is taken from Oort (1983), with winds in the middle atmosphere above 100 hPa based on gradient winds derived from the geopotential height climatology. At high latitudes, the zonal wind is derived by assuming that the relative angular velocity remains constant poleward of 70° latitude. At the equator where the standard gradient wind calculation fails, the zonal wind (above 100 hPa) is based on the second derivative of geopotential height (Fleming and Chandra, 1989). The winds between the equator and 15°S (15°N) are computed by linear interpolation.

HALOE Temperatures
The Halogen Occultation Experiment (HALOE) instrument on UARS provides analyses of temperatures in the altitude range ~ 45-85 km (Russell et al., 1993;Hervig et al., 1996;Remsberg et al., 2002). HALOE uses a solar occultation measurement technique, providing 15 sunrise and 15 sunset measurements per day, with each daily sunrise or sunset group near the same latitude on a given day. The latitudinal sampling progresses in time, so that much of the latitude range ~ 60°N-S is sampled in one month; the measurements extend to polar regions during spring through late summer. The vertical resolution of these data are ~ 2 km, with sampling on UARS standard pressure levels (6 levels per decade of pressure). The results shown here are based on HALOE retrieval version 19.
The seasonal temperature analyses here use the combined sunrise plus sunset temperatures binned into monthly samples. The seasonal cycle is derived by a harmonic regression analysis of these monthly data over the period January 1992-December 1999, including annual and semi-annual harmonics at each height and latitude (spanning 60°N-S). This regression provides a useful method of interpolating the irregular temporal sampling of HALOE.

MLS Temperatures
Middle atmosphere temperatures have also been obtained from the Microwave Limb Sounder (MLS) instrument on UARS (Fishbein et al., 1996). The data here are from an independent retrieval described in Wu et al. (2002), covering the time period January 1992-December 1994. This retrieval is completely independent of other climatologies, using a single temperature profile (an annual mean) as the first guess and linearisation point. The valid altitude range is 20-90 km, with large uncertainties at the two ends; the temperature is reported on the UARS standard pressure grid (six levels per decade of pressure), but the actual retrieval was carried out at every other pressure surface. Compared to the MLS Version 5 (V5) retrieval, the data here have much better vertical resolution in the mesosphere, while it is about the same in the stratosphere. These data and further descriptions are available to the research community via an ftp site: mls.jpl.nasa.gov/pub/outgoing/dwu/temp.
The orbital characteristics of UARS allow MLS to obtain data from approximately 80°S-32°N or 32°S-80°N for alternating satellite yaw cycles (each approximately one month long). In order to handle these large data gaps in high latitudes, our analyses fit the seasonal cycle at each latitude and pressure level using harmonic regression analyses of monthly sampled data (including annual and semiannual harmonic terms in the analyses).

URAP Reference Atmosphere Winds
As part of the UARS Reference Atmosphere Project (URAP), Swinbank and Ortland (2002) compiled a wind data set using measurements from the UARS High Resolution Doppler Imager (HRDI; Hays et al., 1993), supplemented with data from the UKMO stratospheric analyses. The data set comprises zonal-mean wind data from the earth's surface to the lower thermosphere every month for a period of about 8 years starting from the launch of UARS, and the results here use statistics averaged over 1992-1998. The wind data set only includes the zonal (east-west) component of the wind and not the meridional (north-south) component. The wind data are stored on a pressurelatitude grid; the pressures are the UARS standard pressure levels, and the latitudes are equally spaced every 4° from 80°S to 80°N.
There were several periods when HRDI data were not available. The shorter data gaps in the stratospheric data were filled in using a time-smoothing procedure. In the mesosphere tidal variations in the wind field are much more important than in the stratosphere, so a different procedure was used that took into account the local time coverage of the observations (see Swinbank and Ortland, 2002 in preparation). This reduced the impact of tidal variations on the final wind data set as much as possible.
Wind measurements from HRDI span most of the stratosphere and also extend from the middle mesosphere to the lower thermosphere. In order to obtain as full as possible coverage of the atmosphere, the HRDI measurements were combined with UKMO stratospheric analyses. Together, the two data sets cover the troposphere and stratosphere, but there is a gap between the uppermost reliable level of the UKMO data and the lowermost reliable level of HRDI data in the mesosphere. So, the wind data in this region were supplemented with balanced winds calculated from the URAP temperatures (when available). When there were insufficient HRDI data in a particular month, climatological data derived from HRDI data were used instead. In order to provide complete coverage in the data set, interpolated and extrapolated winds were used when no other data were available. The data set includes a quality flag to indicate when the wind values are based on direct measurements and when they are largely derived from climatological or interpolated data.

C. Rocketsonde wind and temperature data
Measurements from small meteorological rockets provide an important source of wind and temperature information for the middle atmosphere, in the 25 to 85 km altitude region. A program of rocketsonde measurements began in the United States in the late 1950s. During the 1960s the program expanded to about a dozen stations making regular measurements once to three times per week (Schmidlin and Rocket, 1986).
Other locations made measurements for limited time periods and for special programs. The number of rocketsonde measurements peaked in the late 1970s, at about 1000 to 1500 yearly, including measurements from the former Soviet Union (USSR), Japan and several other countries. Most measurement locations were at middle latitudes of the Northern Hemisphere and tropical locations, but a few stations were located in polar and Southern Hemisphere mid latitude locations. The number of rocketsonde stations and frequency of observations decreased markedly in the 1980s, and by the 1990s fewer than a total of 100 rocketsonde measurements were made each year. Figure  The archived data collected from the 1950s to the 1990s from the rocketsonde network constitutes a valuable independent resource of in situ, fine vertical scale, temperature and wind information, for climatology and research. Indeed, when rocketsonde measurements were first made, they constituted the only source of middle atmosphere information above radiosonde levels. Now, it is especially important to be able to compare climatological summaries of the rocketsonde data at their very few measurement locations with climatologies constructed using remotely sensed global satellite information.
There are two types of small meteorological rocketsonde systems, thermistor and sphere. By far, most measurements have been made using an instrumented thermistor package and parachute, ejected from the rocket at apogee and tracked during descent by a ground radar. In situ measurements of thermistor temperature are transmitted to a ground station, so that a temperature versus altitude profile is obtained, with 1 km or finer vertical resolution, from apogee (up to 85 km for the USSR system and 70 km for the US system) to data cut-off (approx. 20 to 25 km). A pressure versus altitude profile is obtained with a support radiosonde observation, close in time and from a nearby location, that supplies the needed "tie on" pressure and height values. The horizontal wind versus altitude profile is obtained by ground radar tracking of the horizontal displacement of the descending parachute. Experimental studies for the US system have indicated temperature measurement precision (repeatability) of 1 to 3 K and wind precision of 3 m/s. Corrections have been applied to measured temperatures to account for solar short wave radiative heating, long wave radiative cooling, and frictional heating of the thermistor. Corrections are less than 1 K below 40 km, up to The number of sphere observations have been relatively small, from a few locations in middle northern latitudes and tropical locations. The sphere system consists of an inflatable mylar balloon, inflated at apogee, and tracked by high-precision ground radar as the inflated sphere descends. Atmospheric density from approximately 90 km down to 35 km is derived from the measured vertical profile of fall velocity of the sphere, assuming zero vertical atmospheric motion. Temperatures are derived from the density profile, using a "guess" temperature value at the top of the profile. The horizontal wind versus altitude profile is obtained by the ground radar tracking of the horizontal displacement of the descending inflated sphere. Studies have found average differences of 3 to 6 degrees K between sphere and thermistor measurements. However, the analyses here make no distinction between these two types of measurements.
The rocketsonde wind and temperature climatologies shown here are based on simple monthly averages, derived by binning all of the available observations during 1970-1989. Due to data availability (Figure A.1), we focus the comparisons on the tropics and extratropical NH. The extratropical bins are centred at 30°, 60°, and 80° latitude, including measurements within ± 10° of the central latitude. The tropical data are separated for measurements near 10°S (mostly from Ascension Is. at 8°S), and near 10°N (mostly from Kwajelein at 8°N). Based on this sampling, there are approximately 100-300 profile observations in each monthly bin, depending on latitude and altitude. Vertical sampling is made on the UARS pressure grid (six levels per decade of pressure). Figure A.2 shows an example of the data availability for rocketsonde zonal winds at 0.1 hPa for the latitude bin centred at 30°N. Two important considerations apply to the comparisons of rocketsonde data with global analyses. First, the time periods analysed here for the data are different (1992-1997 for the analyses, and 1970-1989 for the rocketsondes). This is most important for temperatures in the upper stratosphere, and mesosphere, which have experienced strong cooling (of order 2 K/decade near the stratopause, and possibly larger in the mesosphere) during the recent decades (WMO, 1999;Ramaswamy et al., 2001). A large part of the observed rocketsonde-analyses differences in these regions can be attributed to this cooling. Second, the monthly samples from analyses are based on zonal and monthly means of daily data, whereas the rocketsonde statistics are derived from infrequent samples at specific locations, taken over many years. Thus uncertainty levels for the rocketsonde means are significantly larger. Estimates of the standard error of monthly means from the rocketsonde (and lidar) data are calculated as sigmaclimatology = sigma / N , where sigma is the standard deviation of the individual soundings within each month and latitude bin (as in Figure A.2), and N is the corresponding number of measurements. These standard error estimates are included in the comparison figures below (although they are typically very small for the large number of rocketsonde measurements).

D. Lidar temperature data
Lidars provide measurements of the vertical temperature profile in the middle atmosphere, and a number of specific sites have made lidar temperature measurements for a decade or longer. The Rayleigh lidar technique uses the backscattering of a pulsed laser beam to derive the vertical profile of atmospheric density, from which the temperature profile is deduced (Hauchecorne and Chanin, 1980;Keckhut et al., 1993). This technique provides an absolute temperature measurement over altitudes 30-75 km, which does not require adjustment or external calibration (derived temperatures above ~ 75 km can be influenced by first guess uncertainties and values below ~30 km by aerosols or details of the lidar measurements). The vertical resolution of the lidar data is approximately 3 km, and the profiles here are sampled on the UARS standard pressure grid.
For the climatological analyses here, we obtained a number of lidar temperature time series (for stations with relatively long records) from the Network for the Detection of Stratospheric Change (NDSC) web site: http://www.ndsc.ws/. The specific locations and available time records are listed in Table 2. The individual profiles are binned into monthly samples, focused on latitude bins centred at 20°N, 40°N and 80°N. We use all the lidar observations over 1990-1999, in order to most directly compare with the meteorological analyses over 1992-1997 (a slightly longer time record for the lidar data provides better monthly sampling). The total number of lidar observations and their latitudinal sampling is shown in Figure A.3. Our monthly and latitudinal sampling produces between ~ 20-80 measurements per bin for latitudes 20°N, and 80°N, and 300 per month for the bin centred at 40°N. The associated monthly means and standard deviations are calculated identically to those for the rocketsonde analyses.
As a note, one important source of variability for lidar and all data sets in the upper stratosphere-mesosphere is the diurnal and semi-diurnal tides, which have large amplitudes above the stratopause. Most lidar observations are taken at night and most rocketsonde observations are taken at a fixed time of day at each station. By contrast the zonal means of gridded global synoptic data sets average over all local times.

Data Intercomparisons
In this section we make direct comparisons among the different data sets for global fields of temperature, zonal winds, and zonal averaged eddy fluxes of heat and momentum. The first requirement for such comparisons is to choose a time period which maximises record length for overlap among most data sets. Here we choose the period January 1992-December 1997, which gives direct overlap of the UKMO, CPC, NCEP and ERA40 reanalysis, and FUB fields. The UKTOVS record is slightly shorter (to April 1997). The ERA15 reanalysis has a much shorter record during this 1992-1997 period (January 1992-December 1993). We include comparisons for these data by calculating differences only over this 1992-1993 record, rather than the full 6 years 1992-1997. We also include comparisons with the CIRA86 climatology, although it should be kept in mind that these data are derived from a very different time period (covering the 1960's-1970's). The FUB and NCEP re-analyses have data for the presatellite period (prior to 1979), and these are briefly compared separately in Section D below. Rocketsonde data span 1970-1989, while lidar temperatures cover 1990-1999.
A. Temperature

Zonal mean climatology
A cross section of January average zonal mean temperature is shown in Figure 1 based on UKMO analyses. The overall latitude-height structure is similar in all data sets, and comparisons are best made by considering differences with a single standard (UKMO in this case; note this is not an endorsement of the UKMO analyses as 'better' or more 'correct', but simply a choice of one data set as a reference). Included in Figure 1 are January temperature differences from the UKMO for each climatology, showing the overall character of the differences. The differences are typically ± 1-5 K, with systematic vertical or latitudinal patterns depending on each data source. Differences that are consistent across several data sets suggest a systematic bias in the UKMO reference, whereas differing biases across many data sets suggest a fundamental uncertainty in estimates of that quantity.

Figure 1 (continued)
The latitudinal structure of 100 hPa temperature for each data source during January, April, July and October is shown in Figure 2, and differences with the UKMO analyses are shown in Figure 3. The overall latitudinal structure in Figure 2 is similar between the data sets, but there is substantial spread in the tropics and also the polar regions. The UKTOVS data are a notable outlier (> 5 K warmer than all other data sets in the tropics), and are not included in the differences in Figure 3. Aside from UKTOVS, the 100 hPa tropical temperatures fall into two groups, biased warmer (CPC, NCEP and CIRA86) or colder (FUB, ERA15 and ERA40) than UKMO. As discussed in more detail below, the latter (cold) group is probably more realistic, and the former data sets (plus UKMO) have a true warm tropical bias of ~ 2-3 K at 100 hPa. The CIRA86 data exhibit warm biases by up to ~ 5 K in the winter polar regions, and it is likely that at least a part of this may reflect true cooling in the lower stratosphere between the 1960's and 1990's (e.g., Ramaswamy et al., 2001). CPC data show a warm bias over Antarctica in April, July and October, apparently distinct from the other data sources. Manney et al. (1996) compared CPC and UKMO lower stratospheric temperature analyses with polar radiosonde data during several winters, finding systematic warm biases (of order 1-3 K) for both CPC and UKMO data in the Arctic. In the Antarctic the CPC data showed similar 1-2 K warm biases, while UKMO biases were smaller (< 1 K).   Figure 4 shows difference statistics for the 50 hPa level, for January, March, July and October. Here the differences have been calculated with respect to the UKMO analyses at 46.4 hPa (the closest level), resulting in slight cold differences in the tropics for most data sets. There is reasonable overall agreement between the different data sets over a broad range of latitudes, except for warm biases in CIRA86 (warm by ~ 2-3 K) and cold biases in UKTOVS (cold by up to 5 K). There is also a wide range of biases over Antarctica during spring (October), with differences of ± 2-4 K, and no consensus between data sets. The climatology of 10 hPa zonal mean temperature in January and July, and differences with UKMO, are shown in Figure 5. There is a wider range of differences at this level (typically ± 2-4 K) than at 100 or 50 hPa, showing more uncertainty in the climatology.
In the tropics the CIRA86 and MLS data are on the warm side of the ensemble, and the ERA15 10 hPa temperatures are biased cold in the extratropics compared to all other data. The 10 hPa ERA40 data also show large cold biases over SH extratropics in July, and this is part of an oscillatory structure in the ERA40 temperature analyses which are especially large over Antarctica (seen in Figure 1). The cause of this feature is under investigation at ECMWF. Comparisons of temperatures at 1 hPa are shown in Figure 6. Here there are substantial differences of order ~ 5 K between the different data sets, with even larger differences over polar regions. MLS data are relatively warm and UKMO relatively cold compared to the other analyses. The UKTOVS show a slightly different latitudinal structure than the other data sets at high winter latitudes in both hemispheres. This level near the stratopause presents special problems in analyses, because it is not captured accurately in TOVS thick layer radiance measurements, plus it is near the top of the UKMO forecast/assimilation model (at 0.3 hPa).  The climatological vertical profiles of temperature in January and July are shown in Figure 8, for data at 80°S, the equator, and 80°N. These plots show the extreme range of global temperature variations, and illustrate regions where there are relatively larger differences between data sets. One obvious region of uncertainty is in the upper stratosphere and near the stratopause, and a further region showing substantial differences (~ 5 K) is in the Antarctic polar stratosphere in July, near the temperature minimum over ~ 25-30 km. A similar level of uncertainty is found in the Antarctic lower stratosphere in October (see Figures 3-4). The seasonal variations of polar temperatures in the lower stratosphere (100 and 50 hPa) are compared among the different climatologies in Figures 9-10. In the NH there is excellent agreement (within 1 K) between UKMO, CPC, FUB and NCEP reanalyses; the UKTOVS have slightly larger differences, while CIRA86 is much warmer in winter (due at least in part to the different time periods involved). In the SH the differences are somewhat larger; the CPC are warmer than UKMO by ~ 2-3 K throughout winter. Large differences are seen in CIRA86 data over the Antarctic during spring, and a large part of this is due to observed cooling between the respective time periods (associated with ozone depletion, e.g., Randel and Wu, 1999). The climatological minimum 50 hPa temperature over Antartica varies from 184-187 K between the different data sets for the more recent time period (1992)(1993)(1994)(1995)(1996)(1997), and these comparisons highlight the problem of accurate temperature analyses in the intensely cold Antarctic stratosphere.    Figure 9, but for statistics at 50 hPa. Figure 11 compares the temperature climatologies for January at 40°N, where optimal agreement might be expected in levels below ~ 25-30 km, due to maximum radiosonde coverage. Indeed, Figure 11 shows small differences over these altitudes (~ ± 2 K, aside from CIRA86) and the cluster of differences near -1 K suggest a small systematic bias in UKMO analyses. Above ~ 25 km the differences are larger (~ ± 3-5 K), showing the sensitivity to satellite data types and analyses. As at all other latitudes, relatively large differences are found near the stratopause, with the UKMO being systematically colder than most other analyses. Variability at the midlatitude stratopause is explored further in Figure 12, showing the seasonal cycle of 1 hPa temperature at 40°N and 40°S from each analysis. These comparisons demonstrate that some biases vary seasonally. In general, differences between analyses are somewhat larger during local winter. The UKTOVS have an accentuated annual cycle compared to the other analyses, with largest apparent cold biases during local winter. The MLS and CIRA86 data have a relatively warm stratopause throughout the year, while the UKMO, HALOE and ERA40 data are consistently on the cold side.

a. Comparisons with rocketsondes
As noted above, most of the extratropical rocketsonde data occur in latitude bins near 30°N, 60°N and 80°N, and we focus comparisons on these latitude regions. Figure 13 compares rocketsonde climatology profiles with analyses at 30°N for January and July statistics. The rocketsondes show good overall agreement in the stratosphere, and in locating the altitude of the stratopause. The rocketsonde temperatures in the mesosphere (~ 50-70 km) are warmer than the analyses, especially in January. Figure 13. Comparison of rocketsonde temperature statistics at 30°N with zonal mean analyses, showing statistics for January (left) and July (right). Line types denote the same data sources as in Figure 12.
The seasonal variation of temperatures near 30°N from rocketsondes and analyses are compared in Figure 14, for data at 10, 1 and 0.1 hPa. At 10 and 1 hPa the mean rocketsonde temperatures are slightly warmer than most analyses (except CIRA86 at both levels, UKTOVS at 10 hPa and MLS at 1 hPa). At 0.1 hPa the mean rocketsonde values are ~ 5-10 K warmer than MLS or HALOE data, and ~ 3-10 K warmer than CIRA86 (with maximum differences during NH winter). Similar comparisons for temperatures near 60°N are shown in Figure 15. Good agreement is seen at 10 hPa between rocketsondes and all analyses (except CIRA86, with a warm summer bias). At 1 hPa the rocketsondes are also in agreement with most analyses; the comparisons highlight cold biases for UKMO at the warm summer stratopause, and for the HALOE climatology in midwinter (possibly related to the harmonic analysis of sparse HALOE observations near 60°N). At 0.1 hPa, the rocketsondes are ~ 10-20 K warmer than CIRA86, MLS and HALOE, and only the MLS and rocketsonde data show a similar semi-annual seasonal cycle. Comparisons at 80°N (Figure 16) show a similar overall character, with analyses agreeing reasonably well with rocketsondes at 10 and 1 hPa (except for cold rocketsonde differences at 1 hPa during winter). Systematic warm rocketsonde differences are observed at 0.1 hPa, but these differences are smaller than those observed at 60°N. Figure 16. Seasonal cycle comparison of temperatures near 80°N between rocketsondes and zonal mean analyses. Details are the same as in Figure 14.

b. Comparisons with lidars
The lidar temperature measurements have the most data available in the latitude bands near 20°N and 40°N ( Figure A.3). Figure 17 shows a comparison of seasonal temperature variations near 20°N between lidar (from Mauna Loa) and zonal mean analyses at 10, 1 and 0.1 hPa. At 10 hPa the analyses form a relatively compact group, except for the warmer CIRA86 and UKTOVS data, and the lidar data are in good agreement with the larger group. At 1 hPa there is a wider spread (~ 5 K) between the analyses, with the lidar measurements generally toward the middle or lower range of analyses. At 0.1 hPa, there is overall reasonable agreement (to within ~ 5 K) between the lidar and temperatures from CIRA86, MLS and HALOE data.  (lidars from Table Mountain, OHP, Hohenpeissenberg and Toronto) are shown in Figure 18. The overall patterns are very similar to 20°N (Figure 17), with excellent agreement at 10 hPa (again with CIRA86 and UKTOVS as warm outliers). The lidars fall in the mid-range of analyses at 1 hPa, and exhibit reasonably good agreement at 0.1 hPa. We note that the overall good agreement between the lidars and satellite data at 0.1 hPa in Figures 17-18 is further evidence that the larger differences with rocketsondes in the mesosphere (Figures 13-16) are primarily due to the different time periods covered, given the knowledge of strong decadal-scale cooling near and above the stratopause (Ramaswamy et al., 2001).  Figure 19 compares the vertical profiles of temperature measured by lidar near 20°N and 40°N with zonal mean analyses, for NH winter (DJF) and summer (JJA) means, including profiles of the respective differences. Here we have simply averaged the monthly means, and use a sum of squares uncertainty for the lidar data. During DJF at 40°N all of the zonal mean analyses show cold biases with respect to the lidars over 40-45 km; the cause is unknown, but could possibly be due to spatial sampling of the lidars (centred over Europe and North America) and the time-mean longitudinal structure during NH winter. A similar bias is not evident during NH summer at 40°N, or for any season at 20°N. Aside from this bias, there are relatively small differences (of order ± 3 K) between the lidar and analysis data sets; relative outliers are the ERA40 data (persistent cold biases maximizing over ~ 5-2 hPa), and CIRA86 (warm biases over most of the stratosphere). Lidar data for the Arctic region are primarily available during winter, and a seasonal comparison is not possible. Figure 20 shows profile comparisons for 80°N for DJF, based on lidar data from Eureka, Ny Alesund and Thule. The lidars and analyses show good agreement up to the stratopause (with ERA40 still a cold outlier in the upper stratosphere), and the lidars are ~ 5-10 K colder than the MLS and CIRA86 data in the polar mesosphere.

Interannual variability in extratropics
Time series of monthly, zonal mean temperatures are compared here in order to quantify how well year-to-year variability is captured in the various data sets. The focus here is on the extratropical stratosphere during winter and spring (times of maximum variability). Interannual variability in the tropics is discussed separately below. Figure 21 shows comparisons for February temperatures in the Arctic polar region at 80°N for 1979-1999. Overall there is excellent agreement in detail at 100 hPa, and good correspondence between the different data sets at 10 and 1 hPa (aside from the approximately constant biases discussed above). Similar results are found for other months (not shown). Thus estimates of interannual variability in the Arctic are for practical purposes not dependent on data type. Figure 22 shows similar comparisons for the Antarctic (80°S) in October. Overall there is somewhat poorer agreement than for the analyses over the Arctic. At 100 hPa the CPC results show much less cooling during the 1990's than the other analyses. Note however that the 100 hPa CPC results are archived from operational NCEP tropospheric analyses, which are subject to continual analysis system improvements and changes; a similar plot at 50 hPa (where the stratospheric analyses are constant in time) does not highlight the CPC results as an outlier. Results at 10 hPa in Figure 22 show reasonable agreement, and the long records of CPC and UKTOVS data at 1 hPa have similar variability.

Longitudinal structure
Comparisons of the full 3-dimensional structures in the temperature climatologies show that differences are primarily a function of latitude, and accurately characterised by zonal means. However, zonal variations are evident in some difference fields, likely related to data availability and analyses, and a few examples are shown here. Figure 23 shows the climatological January NH 10 hPa temperature from UKMO, together with difference fields for CPC and FUB data. The differences with CPC data are positive and linked to the cold polar vortex (i.e., the coldest temperatures are analysed warmer in CPC results). Similar difference patterns tied to vortex structure are found in Southern Hemisphere spring statistics (not shown).
In contrast, differences with FUB temperatures in Figure 23 are primarily negative and not as obviously tied to the vortex, but rather they are largest over the North American and European-Asian land masses (where the majority of radiosonde data are available).

Tropical seasonal cycle in temperature
The tropics present special problems for analysis of stratospheric temperatures and winds. The tropical tropopause temperature minimum has a sharp vertical structure that is not well resolved by satellite measurements, and it is also problematic for assimilation/forecast models with vertical resolution of ~ 2 km. Temperature anomalies associated with the quasi-biennial oscillation (QBO) have relatively shallow vertical structures, which are also poorly resolved by nadir-viewing operational satellites. Thus it is not surprising to find a wide variance between climatological data sets in the tropics. Here we compare each data set for seasonal variation and interannual variability (the latter focusing on the QBO). The analyses here complement the tropical data comparisons shown in Pawson and Fiorino (1998a, b).
The seasonal variations of equatorial temperature at 100, 50 and 30 hPa derived from each climatological data set are shown in Figure 24. Included in these figures are estimates of monthly temperatures derived for the same 1992-1997 time period from radiosonde measurements at a group of eight near-equatorial stations (within 5° of the equator, including Belem, Bogota, Cayenne, Manaus, Nairobi, Seychelles, Singapore and Tarawa). The radiosonde statistics in Figure 24 include both the means and standard errors calculated from this eight-station group. The amplitude of the seasonal cycle in temperature is reasonably well captured in most data sets at 100 hPa, but there are clear biases among the data sets. In particular, the ERA15, ERA40 and FUB data are the coldest and agree best with radiosondes (except for FUB during January-March), whereas UKMO, CPC, NCEP and CIRA86 data each have a consistent warm bias of 2-3 K (and UKTOVS is almost 10 K too warm, and not shown in Figure 24). At 50 and 30 hPa the seasonal variations are smaller than at 100 hPa, and approximately captured in most analyses (aside from CIRA86 and FUB data at 30 hPa). At 50 hPa the CPC, NCEP and ERA40 data agree best with radiosondes, with FUB ~ 2 K warmer during NH winter. The UKMO analyses appear warm at 50 hPa, but the comparison is not exact since the UKMO pressure level is 46.4 hPa. Likewise, the ERA15 time period is different (1992-1993 versus 1992-1997), and the 1992-1993 period was anomalously warm in the tropics due to Pinatubo volcanic effects. The CIRA86 and UKTOVS (not shown) are warm outliers at 50 hPa. At 30 hPa most analyses are close to uncertainty estimates of the radiosondes, with FUB and CIRA86 being warm outliers in some months.

Figure 23.
Top panel shows climatological January average 10 hPa temperatures from UKMO analyses. Middle panel shows the differences with CPC analyses (CPC-UKMO), and lower panel differences with FUB (FUB-UKMO). Contour interval in the lower panels is 1 K. Figure 24. Comparison of the seasonal variation in equatorial temperature from available analyses at 30 hPa (top), 50 hPa (middle) and 100 hPa (bottom). Because of relatively large biases, UKTOVS data are not included here. The circles show a climatology derived from radiosonde measurements at 8 near-equatorial stations (over 5°N-5°S), and the error bars denote the plus/minus two standard errors within the station group. Figure 25 shows similar statistics for temperature variations at 10, 1, 0.1 and 0.01 hPa, where at each level the dominant variation is a semi-annual oscillation (SAO). There is approximate agreement in the amplitude and phase of the SAO at 10 hPa across many data sets, but biases on the order of 5 K between the different analyses. Similar mean biases are seen at 1 hPa, along with larger differences in SAO amplitude and phase (quantified below). The 0.1 hPa temperatures show quite good agreement between CIRA86, MLS and HALOE, whereas at 0.01 hPa (~ 80 km) there are substantial differences in detail between these three data sets.

a. Tropical rocketsondes
Rocketsonde comparisons in the tropics are available near 10°N and 10°S; the temperature comparisons are similar at both latitudes, and we focus here on results near 10°N. Figure 26 shows the 10°N rocketsonde climatology and analysis temperatures at 10, 1 and 0.1 hPa. The SAO amplitude and phase evident in rocketsonde data at 10 and 1 hPa is reasonably consistent with the various analyses, although note that observations near 10°N do not capture the full SAO amplitude at the equator (as shown below in Figure 28). At 0.1 hPa the rocketsondes show a less coherent SAO than that inferred from the CIRA86, MLS and HALOE data, and furthermore the rocketsondes are approximately 10 K warmer.

Figure 26.
Comparison of the seasonal cycle of temperatures near 10°N between rocketsondes and zonal mean analyses, at 0.1 hPa (top), 1 hPa (middle), and 10 hPa (bottom).

b. Semi-annual oscillation (SAO)
Because the SAO dominates the seasonal variation in equatorial temperatures, it is useful to quantify the SAO amplitude and phase structures derived from the different data sets analysed here. A more comprehensive review and climatology of the SAO (extending to 100 km) is provided in Garcia et al. (1997). The results here are based on simple harmonic analyses of the different data sets for the time periods available.
The vertical structure of the temperature SAO amplitude and phase are shown in Figure 27, including results from each data set. As well-known from previous analyses (e.g., Hirota, 1980), the temperature SAO has a double peaked structure in altitude, with maxima below the stratopause (~ 45 km) and mesopause (~ 70 km), and these maxima are approximately 180 degrees out of phase. The maximum near 45 km has an amplitude of ~ 4 K in MLS, HALOE, CIRA86 and ERA40 data sets, and substantially weaker amplitude in CPC, UKMO and UKTOVS data. For the maximum near 70 km the CIRA86, MLS and HALOE show a range of amplitudes of ~ 4-7 K. For further comparison of the upper level peak, the dots in Figure 27 show results derived from Solar Mesosphere Explorer (SME) temperature data for 1982-1986 (taken from Garcia and Clancy, 1990). These SME results show similar amplitude and phases as the other data sets, but don't exhibit an absolute peak near 70 km. Figure 27. Comparison of the amplitude (left) and phase (right) of the semi-annual oscillation (SAO) in equatorial temperature derived from each temperature data set. Phase refers to month of the first maximum during the calendar year. The dots show the mesospheric results derived from SME satellite data, taken from Garcia and Clancy (1990). Figure 28 compares the amplitude and phase structure of the temperature SAO as a function of latitude at 2 hPa and 0.046 hPa (near the amplitude maxima seen in Figure 27). The different data sets at 2 hPa show a clear separation in terms of SAO amplitude, with the ERA40, MLS, HALOE and CIRA86 data having amplitudes near 4 K, approximately twice as large as the CPC, UKMO and UKTOVS results. The rocketsonde results (shown as dots near 8°N and 8°S) show amplitudes that agree better with values from the former (larger amplitude) group. Phases at 2 hPa are in good agreement between all data sets (maximum near April 1). Comparisons at 0.046 hPa show less agreement in amplitude between the few data sets. The HALOE and MLS data both show an equatorial maximum (5-7 K), which is not evident in CIRA86. Rocketsondes have weaker SAO amplitudes than the other data sets at this high altitude.

Tropical temperature interannual variability
Interannual anomalies in 100 and 70 hPa equatorial temperature for the period 1985-1999 are shown in Figure 29, with anomalies calculated as differences from the 1992-1997 mean seasonal cycles shown above. The size of the variations at 100 hPa are small (~ ± 1-2 K), and there are substantial differences among the different data sets regarding the details of these small changes. At 70 hPa (where FUB and UKTOVS are not available) the interannual temperature changes are ~ ± 2-3 K, larger than those at 100 hPa, and there is overall good agreement among the different data sets (with the latter half of the 1990's being relatively cold). Interannual temperature changes at 50, 30 and 10 hPa are shown in Figure 30. Variability at these levels is dominated by the QBO, and while a QBO is evident in each data set, the amplitude varies considerably. For comparison, we include in the 50 and 30 hPa plots in Figure 30 temperature anomalies derived from radiosonde measurements at Singapore (1°N). These comparisons suggest the QBO temperature anomalies are most accurately captured in FUB analyses, and underestimated in each of the other data sets to some degree. Negative temperature anomalies (easterly shear) are particularly poorly sampled. The interannual anomalies at 10 hPa show a clear QBO signal in the FUB analyses after 1991; the QBO is evident but much weaker in the other data sets. Prior to 1990 there are large differences among anomalies derived from the five available data sets (note the FUB 10 hPa temperatures are only available during winter months prior to 1992).

Figure 30.
Time series of interannual anomalies in zonal mean temperature at the equator from various analyses, together with results from radiosonde measurements at Singapore (1°N). Statistics are shown for 10 hPa (top), 30 hPa (middle) and 50 hPa (bottom).
The strength of the QBO temperature variation in the different data sets is quantified in Figure 31, where the equivalent "QBO amplitude" is plotted as a function of latitude (at 30 hPa) and height. The QBO amplitude is defined as 2 times the rms deviation of deseasonalised anomalies during 1992-1997; this is the equivalent amplitude of a harmonic oscillation, following Dunkerton and Delisi (1985). For comparison, Figure 31 also includes the equivalent result derived from Singapore radiosonde measurements at 50 and 30 hPa. The FUB analyses exhibit the largest QBO signal in temperature, peaking above 4 K at 30 hPa. The next best results are derived from the assimilated data sets (ERA40, ERA15, UKMO and NCEP, in that order), although each underestimates the amplitude to some degree, especially the UKMO and NCEP data. The CPC and UKTOVS data sets provide relatively poor estimates of the temperature QBO. (b) Shows the vertical structure of the QBO temperature amplitude at the equator. For comparison, plus signs show results derived from Singapore radiosonde data.
Interannual temperature anomalies at 1 hPa are shown in Figure 32, derived from UKMO, CPC and UKTOVS data. The only records extending prior to 1990 are the CPC and UKTOVS analyses, and these show differences of 2-3 K for the pre-1990 period (and hence substantial differences in decadal trends). A separate data set included in Figure 32 shows brightness temperature (radiance) anomalies for 1979-1998 derived from the series of Stratospheric Sounding Unit (SSU) measurements (which are part of the TOVS data that go into the CPC and UKTOVS analyses). The SSU data set in Figure 32 used overlap periods between the different satellite instruments to make adjustments, in an effort to produce a homogeneous long-term data set (see Ramaswamy et al., 2001). The SSU channel 27 data in Figure 32 are representative of a thick layer of the upper stratosphere, spanning approximately 34-52 km, with a peak near 44 km (between 1-2 hPa in pressure), and thus comparisons with the 1 hPa analyses are not exact. Nonetheless, the SSU time series suggests that interannual variations in the CPC temperatures at 1 hPa may have uncorrected biases which influence estimates of long term variability (such as decadal trends or 11-year solar cycle changes); somewhat better overall agreement is seen between SSU and UKTOVS results. The UKMO analyses in Figure 30 show cold anomalies after 1997, which are related to an erroneous ozone climatology in the assimilation model during this time (error introduced on 28 January 1998).

Figure 32.
Interannual anomalies in zonal mean temperature at 1 hPa from available data sets (top), together with similar results derived from SSU satellite measurements (bottom). The SSU data represent temperatures over a ~15 km thick layer centered near 1-2 hPa.

Temperatures for pre-1979 data
The FUB and NCEP reanalysis data sets have stratospheric temperatures (up to 10 hPa) for the pre-1979 period, i.e., prior to the availability of satellite data. The inclusion of satellite temperatures in the NCEP reanalysis has a significant impact on regions with relatively little radiosonde data (Mo et al., 1995). As an example, Figure 33a shows time series of 100 hPa equatorial temperature from the NCEP reanalysis, showing a clear discontinuity of approximately 3 K prior to and after 1979, coincident with the inclusion of satellite temperatures. The spatial structure of this temperature discontinuity in NCEP reanalyses is shown in Figure 33b, comparing 15-year averages before and after 1979. The largest differences are observed in the tropics and SH middle and high latitudes (regions of few radiosondes); altitudes near the tropopause are warmer with the inclusion of satellite data, while the region of 30-50 hPa is cooler. Corresponding stratospheric zonal wind discontinuities of ± 2-4 m/s are found in the tropics and SH midlatitude and polar regions (not shown here). Extreme caution should be used in analysing long-term variability of NCEP data in these regions.  1960-1974 and 1985-1999. Note the jump in temperatures near 1979, associated with the introduction of satellite data into the reanalyses. b) Bottom panel shows a cross section of temperature differences between 15-year means before (1960-1974) and after (1985-1999) the satellite discontinuity.
Differences between NCEP and FUB temperature analyses over the NH are shown in Figure 34, comparing 5-year periods before (1974)(1975)(1976)(1977)(1978) and after (1993)(1994)(1995)(1996)(1997) the satellite changes in NCEP data (presumably the FUB data are more homogeneous). Significant differences between the two time periods are found in the tropics at 100 hPa, where the NCEP reanalyses are colder than FUB for 1974-1978, but warmer for 1993-1997. The artificial warming in NCEP data between these periods is consistent with the results in Figure 33. Figure 34 also shows systematic differences in 30 hPa temperature over the entire hemisphere (with NCEP warmer for the satellite period), but the differences are relatively small (~ 1 K). The comparisons in Figure 34 suggest temperatures in other regions are less sensitive to the satellite discontinuity. Figure 34. Differences between FUB and NCEP annual average zonal mean temperatures at 30, 50, and 100 hPa, comparing the pre-satellite (1974-1978) and post-satellite (1993-1997) time periods. Error bars show the ± 2-sigma variability values associated 5-year means (just included on the 1993-1997 results for clarity).

Zonal mean climatology
The January zonal mean zonal wind climatology (for 1992-1997) from UKMO data is shown in Figure 35, together with differences from the other climatological data sets. In general there are relatively large differences in the tropics, related to the uncertainties in deriving balanced winds in low latitudes from temperature (height) data. Detailed comparison of tropical winds are discussed separately below. The difference patterns in NH extratropics show a consistent pattern of positive values near 60°N for each contemporaneous data set (CPC, UKTOVS, NCEP and ERA40), indicating a slightly weaker polar night jet in UKMO data. More complicated global patterns are seen in the CIRA86 differences. Figure 36 shows UKMO climatology and differences with CPC and ERA40 winds for July and October. In July there are relatively small differences (outside of the tropics), and the strength of the intense SH polar jet is similar in each analysis. During October the UKMO SH polar jet is somewhat stronger than the CPC and ERA40 jets. Note these October wind differences are consistent (via thermal wind) with the colder Antarctic polar vortex analyzed in UKMO data (see Figures 3-4). Figure 37 compares January and July zonal wind climatologies from each data set at 100, 10 and 1 hPa. Outside of the tropics there is reasonable agreement between most analyses, with the CIRA86 climatology showing some biases in detailed structure. The westerly jets are strongest in UKTOVS data.

a. Comparison to rocketsondes
Rocketsondes provide direct measurements of zonal winds, and are unique for comparing to winds derived from analyses (given the caveat of differing time periods). Seasonal climatologies of zonal winds derived from rocketsondes are compared with the various analyses at 30°N and 60°N in Figure 38, for pressure levels 10, 1 and 0.1 hPa. At 30°N there is quite good agreement in the winds at all levels; note especially the strong subtropical mesospheric jet (at 0.1 hPa) in rocketsondes, CIRA86 and URAP data. At 60°N there are some systematic differences with rocketsondes during local winter (November-March) at 10 hPa, with the rocketsonde winds substantially weaker. Given the overall good agreement at other times and locations in Figure 38, these differences may be related to real (decadal-scale) time changes. Note these wind changes are consistent (via thermal wind balance) with observed cooling of the polar lower stratosphere (discussed above).

Tropical seasonal cycle
Tropical stratospheric winds present particular problems, because there are few direct wind measurements on a daily basis above the lower stratosphere. Also, due to the smallness of the Coriolis parameter, determination of balanced wind in the tropics requires a more accurate estimate of horizontal temperature gradients than at higher latitudes. Thus special attention is required in assessing the quality of tropical winds.
The seasonal variations of equatorial zonal winds at 100 and 50 hPa are shown in Figure 40. An annual cycle is evident at 50 hPa (maximum easterlies during July-September), and a semi-annual variation at 100 hPa, and aside from the CIRA86 results there is overall agreement to within a few m/s among the different data sets. Similar statistics are compared for the 30, 10, 5 and 1 hPa levels in Figure 41. There is approximate agreement among analyses at 30 and 10 hPa, aside from the CIRA86 (strong easterly biases). There is a substantial spread among the data sets at 5 hPa, with ERA40 exhibiting relatively large differences compared to other levels (note the corresponding differences seen in Figures 35-36). There is a pronounced SAO in zonal wind at 1 hPa in Figure 41, which shows similar magnitudes and phases in each data set (discussed more below).

a. Tropical rocketsondes
Rocketsondes are particularly valuable for ground-truth measurements of tropical winds in the middle atmosphere, given the uncertainties in balance wind estimates discussed above. Extensive rocketsonde data are available near 10°N and 10°S, and because there is a substantially different seasonal cycle at these latitudes, we include comparisons for both 10°N and 10°S in Figure 42. The overall impression from Figure 42 is that there is remarkably good agreement between the rocketsonde climatologies and analyses at 10 and 1 hPa, with appropriate seasonal variations and cross-equatorial differences mirrored in all data sets (except for some notable UKTOVS and CIRA86 biases at 10 hPa and 10°S). The rocketsondes at 10°S, 0.1 hPa also show approximate agreement with the CIRA86 and URAP data sets; there are fewer rocketsondes available at 10°N, 0.1 hPa.

b. Semi-annual oscillation
The vertical and latitudinal amplitude and phase structure of the zonal wind SAO is shown in Figure 43, comparing each data set along with rocketsonde results at 8°N and 8°S. The vertical structure shows an amplitude maximum near the stratopause (~ 50 km), with reasonable agreement between various data sets. A second amplitude maximum near the mesopause (~ 80 km) is suggested in URAP winds. The latitudinal structure at 1 hPa shows maximum SAO amplitude near 10-20°S for most data sets, which is distinct from the equatorially-centred SAO in temperature ( Figure 28). The rocketsonde results near 8°N and 8°S suggest a latitudinal asymmetry consistent with analyses (i.e., larger zonal wind SAO in the SH subtropics). The rocketsonde SAO amplitudes are approximately 25% larger than most analyses, while the phases are in good agreement.  Figure 44 shows interannual anomalies in equatorial zonal wind at 50, 30 and 10 hPa during 1985-1999 derived from the various analyses. The QBO dominates variability in these time series, and included in Figure 44 are anomalies derived from Singapore radiosonde data, which are a standard reference for the QBO (e.g., Naujokat, 1986). The QBO signal is evident in each analysis, but the amplitude varies strongly between different data (and with altitude). In general the assimilated data sets (UKMO, NCEP, ERA15 and ERA40) have the largest amplitudes, and most closely approach the Singapore data, whereas the balance winds derived from CPC and UKTOVS are much too weak. The strength of the QBO in the different data sets is quantified in Figure 45, where the equivalent QBO amplitude (defined earlier) is plotted as a function of latitude (at 30 hPa) and height. For comparison, Figure 45 also includes the 30 hPa QBO amplitude derived from tropical radiosonde climatologies in Dunkerton and Delisi (1985), plus the equivalent result from Singapore radiosonde measurements over pressure levels 70-10 hPa (using data as in Figure 44). Overall the ERA40 data exhibit the largest QBO amplitude (in good agreement with the radiosonde climatology and Singapore data), with the ERA15, UKMO and NCEP re-analyses somewhat weaker, and CPC and UKTOVS (balance winds) as severe underestimates. The ERA40, ERA15, UKMO and NCEP data show approximately similar amplitudes between 70 and 30 hPh, whereas above 30 hPa there are much larger differences, and only the ERA40 approaches the Singapore amplitudes over 20-10 hPa. Above 10 hPa there is a factor of two difference between the ERA40 and UKMO results, and here the UKMO amplitude is almost certainly too weak. Figure 46 shows wind anomalies at 1 hPa from UKMO, CPC and UKTOVS data. As with temperatures (Figure 32), there is some approximate agreement after 1990, but larger differences between CPC and UKTOVS for the earlier period. The UKTOVS time series suggests a QBO signal at 1 hPa in the early record, which is seen in all three data sets after 1990 (and which is evident at 1 hPa in rocketsonde data, e.g., Gray et al., 2001).  (b) Shows the vertical structure of QBO amplitude at the equator. For comparison, results of Dunkerton and Delisi (1985) are shown (dots), together with estimates from Singapore radiosondes (plus signs).

C. Zonally averaged heat and momentum fluxes
The zonally averaged fluxes of heat v' T' ( ) and momentum u'v' ( ) are fundamentally important diagnostics of atmospheric wave behaviour and large-scale transport. Their calculation is based on co-variances of eddy winds and temperatures (in longitude), and these fluxes provide sensitive diagnostics of planetary wave behavior and coupling with the mean flow (in both observations and models). The eddy heat and momentum fluxes are primary quantities involved in calculation of the Eliassen-Palm (EP) flux and its divergence (Andrews et al., 1987). Of primary importance is the poleward eddy heat flux v' T' ( ) in the extratropical lower stratosphere, which is proportional to the vertical wave activity flux (EP flux) from the troposphere into the stratosphere (e.g., Andrews et al., 1987). The fluxes considered here are calculated from daily data and then monthly averaged (i.e., they contain both stationary and transient components). Because daily data are involved in these calculations, we focus on comparisons among UKMO, CPC and NCEP re-analyses. The time period covered is 1992-1997 (a few ERA15 results are also shown for reference, but these are for the period 1988-1993 and are thus not directly comparable). Statistics are compared for NH winter-spring and SH spring seasons, when stratospheric planetary waves and fluxes have maximum amplitude.
January climatological heat fluxes at 100, 50 and 10 hPa are compared in Figure 47. ( ) in the NH varies bỹ 10-20 % between the analyses, with NCEP on the stronger side and CPC slightly weaker. These uncertainties are consistent with Newman and Nash (2000), who include comparisons with other data sets (for shorter periods). This 10-20 % difference is thus the current level of uncertainty in this derived quantity over NH midlatitudes. Similar statistics for the SH in Figures 47-48 show the CPC data as an outlier with substantially smaller fluxes than the other analyses. Interannual variability of heat fluxes at 50 hPa is shown in Figure 49 for February and March in the NH and October in the SH (including results from ERA15 data). Reasonable agreement between analyses is seen in the NH (differences of ~ 15 %), and the (weak) biased CPC estimates are evident in the SH (although the year-to-year variations are captured to some degree in CPC data).   ( ) are shown in Figure 50.
Reasonable agreement is found between the analyses at 100 hPa in both hemispheres (outside of the tropics). At higher levels there are larger differences. The NCEP reanalyses have particularly small u'v' ( ) in the NH at 50 hPa, and the CPC analyses appear systematically small in the SH, similar to the heat fluxes in Figure 46.

Summary of largest biases
This study has focused on comparing climatological data sets for the middle atmosphere, which are currently used in the research community. Overall the climatologies developed from analyses (and lidar measurements) for the 1990's agree well in most aspects, although each data set can exhibit 'outlier' behaviour for certain statistics.
The following is a list of the largest apparent biases in each climatological data set, as derived from the foregoing intercomparisons. These are identified when individual data sets are 'outliers' from the group, for these particular features. • warm biases of ~ 5-10 K over much of the stratosphere (20-50 km) • relatively large easterly biases in tropical winds (derived from balanced winds) • SAO's in mesospheric wind and temperature not well resolved

MLS
• warm biases (~3-7 K) over much of the stratosphere (20-50 km) Comparisons of the recent climatologies with the historical data sets (CIRA86 and rocketsondes) show reasonable agreement, but the effect of decadal-scale cooling throughout the middle atmosphere is evident, particularly in the polar lower stratosphere (Figures 9-10) and in the upper stratosphere and mesosphere (e.g., Figures 13-16). Decadal changes may also influence zonal mean winds at high latitudes ( Figure 38). Despite these differences, the overall quality of the CIRA86 global climatologies is remarkable, given that they were derived from several combined data sets, covering different altitudes and time periods. While the direct wind measurements afforded by UARS and data assimilation techniques provide improved winds (especially in the tropics), the balance winds calculations of CIRA86 captured the overall global climatology reasonably well.

Outstanding uncertainties and problem areas
The comparisons here also allow us to highlight aspects of middle atmosphere climatologies that are relatively uncertain. These are identified for statistics that show relatively large variability among each of the different data sets, suggesting a fundamental level of uncertainty or high sensitivity to the details of data analysis. These include: 1) The tropical tropopause region is biased warm (compared to radiosonde data) in many analyses. Relatively smaller biases are found in the ERA15, ERA40 and FUB analyses, which are more strongly tied to radiosonde measurements. The warm biases in this region of sharp temperature gradients probably result from a combination of low vertical resolution in the analyses, plus the less than optimal use by most analyses of thick-layer satellite temperature measurements.
2) The temperature and 'sharpness' of the global stratopause shows large variability among different data sets. This is probably due to the relatively low vertical resolution of TOVS satellite measurements, and also the fact that the stratopause is near the upper boundary for several analyses (UKMO, CPC, UKTOVS).
3) Temperature variability in the tropics (associated with the QBO) is underestimated in all analyses (except FUB), compared to radiosonde measurements. The underestimates are particularly large for analyses that rely primarily on low resolution TOVS satellite data (CPC and UKTOVS). The temperature SAO near the stratopause is also underestimated in these data sets (as well as UKMO).
4) QBO variations in zonal wind are underestimated to some degree in most analyses, as compared to Singapore radiosonde data. The best results are derived from assimilated data sets (ERA40, ERA15, UKMO and NCEP, in that order), and only ERA40 has realistic wind amplitudes above 30 hPa. The use of balance winds in the tropics (derived from geopotential data alone) is problematic for the QBO, and produces large underestimates of variability in CPC and UKTOVS data sets.

An Atlas of Middle Atmosphere Temperatures and Zonal Winds
This section presents a brief atlas of monthly mean temperatures and zonal winds (over 0-85 km), together with estimates of interannual variability in these quantities over 0-50 km. The zonal mean temperature climatology is derived using UKMO analyses over 1000-1.5 hPa, combined with the HALOE temperature climatology over pressures 1.5-0.0046 hPa (~ 85 km). The monthly HALOE climatology is best sampled for the latitude range 50°N-S, and mesospheric temperatures poleward of 50°N and 50°S are derived from the MLS climatology, with offsets at each pressure level used to match the HALOE data at 50°N and 50°S (i.e., the polar latitudinal gradients from MLS are used). The monthly zonal wind climatology (over ~ 0-85 km) is derived from the URAP wind analyses, based primarily on UKMO and UARS HRDI data (see Section 2.9). In order to provide smooth monthly estimates, we use a harmonic analysis of the available time series over 1992-1997, including annual and semi-annual harmonic components.
Estimates of the interannual variability of the zonal mean temperatures and zonal winds are derived using UKMO analyses for the time period 1992-2000. The UKMO analyses have a shorter time record than the CPC or UKTOVS data sets, but provide improved estimates of tropical zonal winds. Interannual standard deviations are calculated from the standard formula where x represents the ensemble (climatological) mean, x i is the monthly mean for each year, and N is the number of available years (N = 9 for 1992-2000).
As a note, the UKMO temperature analyses had some significant errors introduced after January 1998 at the uppermost levels (at and above 1 hPa), due to an ozone climatology problem in the assimilation model (as seen in Figure 31). In order to avoid large effects on the interannual variability estimates, temperature variability at and above 1 hPa use statistics derived from the shorter record 1992-1997.
Climatological means and standard deviations are shown below, in the forms of: (1) latitude-height cross sections for each month (Figure 51), (2) latitude-time sections at a few selected pressure levels (Figure 52), and (3) height-time sections at a few latitudes (Figure 53). In the latitude-height and height-time sections we include heavy dashed lines indicating the location of the tropopause (defined by the lapse rate criterion and taken from the NCEP reanalyses), and the stratopause (defined by the local maximum in temperature near 50 km). zonal wind (top right) for each month, derived from the data sets discussed in Section 5. Contour interval is 5 K for temperature (values below 210 K shaded), and 5 m/s for zonal wind (with zero contours omitted). The heavy dashed lines denote the tropopause and stratopause. Lower panels show the corresponding interannual variability of monthly means for temperature (left) and zonal wind (right) derived from UKMO analyses over 1992-2000. Contour interval is 1 K for temperature (values above 3 K shaded), and 2 m/s for zonal wind (values above 8 m/s shaded).

Figure 52.
Latitude-month sections of zonal mean temperature (top left) and zonal wind (top right), together with estimates of interannual variability (lower panels), at a number of selected pressure levels. Contour intervals are 5 K and 5 m/s for the means (upper panels), and 1 K and 2 m/s for the lower panels. Interannual variability plots are only available for data at and below 1 hPa. Shading is included for mean temperatures below 210 K, and for wind (temperature) variability greater than 8 m/s (3 K), respectively.