Journal of Geophysical Research: Atmospheres

Long-term trends in Antarctic winter hydroxyl temperatures



[1] Observations of the hydroxyl nightglow emission with a scanning spectrometer at Davis station, Antarctica (68°S, 78°E), have been maintained over each winter season since 1995. Rotational temperatures are derived from the P-branch lines of the OH(6–2) band near λ840 nm and are a layer-weighted proxy for kinetic temperatures near 87 km altitude. The current 16 year record allows tentative estimation of the atmospheric response in the mesopause region to solar cycle forcing and the underlying long-term linear temperature trend. Seven years of new data have been added since the last reported trend assessments using these data. A multivariate regression analysis on seasonally detrended winter mean hydroxyl temperatures yields a solar cycle coefficient of 4.8 ± 1.0 K/100 solar flux units (SFU) and a linear long-term cooling coefficient of −1.2 ± 0.9 K/decade. These coefficients are consistent within uncertainties for nightly, monthly, and annual mean trend evaluations. A distinct seasonal variation in trend coefficients is found in 30 day sliding window or monthly trend analyses. The largest solar activity response (∼7 K/100 SFU) is measured in March, May–June, and September, and there is little or no solar response in April and August. The long-term trend coefficient shows the largest cooling rate (4–5 K/decade) in August–September, through to warming (2–3 K/decade) for the March and May–June periods. Comparisons of trend results are made with other important hydroxyl measurement sites. Variability in the remaining residual temperatures is examined using lag correlation analyses for the influence of planetary waves, the quasi-biennial oscillation, the polar vortex intensity, and the southern annular mode.

1. Introduction

[2] Over long time scales, the primary influences on temperature in the mesosphere and lower thermosphere (MLT) are seasonal variations due to meridional circulation patterns, planetary and gravity wave activity, solar flux variability and then changes in atmospheric gas composition. With regard to the latter, temperatures in the MLT region are expected to cool over the long-term in response to global increases of CO2 and other greenhouse gases. The low collision frequency at high altitude means greenhouse gases preferentially radiate absorbed energy to space, rather than warm the atmosphere through the kinetics of collisions.

[3] Many early modeling studies predicted that for a given increase in CO2 concentration, the magnitude of the response in the upper atmosphere would be much larger than for the lower atmosphere, and in particular, a high impact on temperatures in the high-latitude mesosphere [Roble and Dickinson, 1989; Rind et al., 1990; Berger and Dameris, 1993; Portmann et al., 1995; Akmaev and Fomichev, 1998, 2000]. More and more sophisticated generations of these models, for example the Thermosphere Ionosphere Mesosphere Electrodynamic General Circulation Model (TIME-GCM) [Roble, 2000], the Canadian Middle Atmosphere Model (CMAM) [Fomichev et al., 2007], the Whole Atmosphere Community Climate Model (WACCM) [e.g., Richter et al., 2008] and the Hamburg Model of the Neutral and Ionized Atmosphere (HAMMONIA) [Schmidt et al., 2006] include interactive chemistries for the substantial effects of changes in other atmospheric species such as O3, CO, CH4, NOx and water vapor, higher order dynamical forcing from planetary waves, tides and gravity waves, changes in sea surface temperature (SST) and sea ice distribution, and episodic events such as sudden stratospheric warmings [Liu and Roble, 2002].

[4] Middle atmosphere results of these latter models show a doubled CO2 concentration has maximum impact on temperature around the high-latitude winter stratopause (∼50 km; 1 hPa) and comparatively little effect near the mesopause (∼90 km; 0.003 hPa). Both CMAM and HAMMONIA for example, predict cooling maxima of 10–12 K in the high-latitude winter stratopause and 7–10 K in the high-latitude summer stratopause, with cooling minima of 3–5 K in the high-latitude winter mesopause and insignificant or even slight warming in the high-latitude summer mesopause. The lack of response in the mesopause region is a combination of both radiative and dynamical effects [Schmidt et al., 2006]. Radiatively, the infrared cooling primarily due to CO2 emissions in the 15 μm infrared band is less effective through the mesosphere due to radiative transfer (energy exchange) with the warmer underlying layers. Additionally, solar heating, due to CO2 absorption in the near-IR bands (2 and 2.7 μm) is important at altitudes ∼75 km and counteracts the radiative cooling. Dynamically, doubled CO2 also induces changes in the residual summer-to-winter circulation in the mesosphere. A slower ascent of the upper mesospheric circulation in summer decreases adiabatic cooling (i.e., a net warming which counteracts the CO2 radiative cooling). The dynamical wind, wave and temperature interactions are complex but both Schmidt et al. [2006] and Fomichev et al. [2007] propose the following mechanism: A stronger equator to pole temperature gradient due to warming in the tropical troposphere and cooling in the extratropical tropopause drives stronger midlatitude tropospheric westerlies, which filter more eastward propagating gravity waves, in turn resulting in less eastward drag in the upper mesosphere and consequent modification of the mesospheric ascent rate. The contribution of this effect is clearly more difficult to quantify as the results are dependent on choice of different gravity wave drag parameterization schemes and their potential modification and feedback with climate change.

[5] This kind of dynamical scheme for the coupling between the lower and upper atmosphere was also simulated for a stratospheric warming event by Liu and Roble [2002]. In this case, a dominant planetary wave 1 component was found to have been generated by the warming event, which decelerated and reversed the mean wind in the high-latitude winter mesosphere, thereby allowing more eastward propagating gravity waves through to the MLT, with a consequence of changing the meridional circulation from poleward descent, to equatorward ascent. The observed impact of the 2002 Southern Hemisphere stratospheric warming event is discussed later in this paper.

[6] Measured trends are at present far from realizing the doubled CO2 scenario's of these models (projected to be ∼70 years at 1% annual increase [Intergovernmental Panel on Climate Change, 2007]). Data sets longer than a decade in the MLT region are few. A benchmark review of measurement and model calculations of trends in the MLT region up to 2002 was conducted by Beig et al. [2003], with an update by Beig [2006] and a further review of solar activity responses by Beig et al. [2008]. More recently, She et al. [2009] have also summarized temperature trends from OH, rocket and lidar measurements.

[7] Although significant seasonal and latitudinal differences in the solar cycle response and long-term trend estimates are evident in these reviews, on a yearly average basis, a positive solar cycle response of 2–4 K/100 solar flux units (SFU) and a near zero long-term trend is relatively consistent for the longest mesopause region observation sets.

[8] Temperature measurements from satellite instruments such as the Sounding of the Atmosphere using Broadband Emission Radiometry (SABER) instrument on the Thermosphere Ionosphere Mesosphere Energetics Dynamics (TIMED) spacecraft, launched in December 2001 [Mertens et al., 2002], the SCanning Imaging Absorption spectroMeter for Atmospheric CartograpHY (SCIAMACHY) experiment on Envisat, launched in March 2002 [von Savigny et al., 2004] and the Mesospheric Limb Sounder (MLS) on Aura, launched in July 2004 [Schwartz et al., 2008] are approaching the solar cycle length observation sets sufficient to provide long-term trend estimates and will in the near future provide an invaluable resource to help characterize seasonal and geographical differences in long-term trends.

[9] Observations of hydroxyl rotational emission lines, which result from the exothermic hydrogen-ozone reaction, have long provided a well established means of temperature measurement in the MLT region. However, imprecise knowledge of the source altitude and attribution of temperature variability to chemical processes or to changes in the layer height or shape has been a difficulty for these ground-based measurements. Early rocket experiments confirmed a layer origin near 87 km altitude with a mean thickness of ∼8 km [Baker and Stair, 1988] but in recent years, limb scanning satellite measurements such as SABER provide both temperature and OH volume emission rate (VER) as a function of tangent point altitude. French and Mulligan [2010] have recently compared OH temperatures with OH layer–weighted SABER profiles, computed with a number of VER and Gaussian layer weighting functions. Remarkably good agreement was found (within 2 K) for all layer weighting functions considered, indicating that the temperature is relatively insensitive to layer shape and peak altitude and the OH measurement remains a very good proxy temperature for the 87 km level.

[10] This work provides an update on the solar cycle and long-term trend analysis of the hydroxyl rotational temperature measurements taken through each winter at Davis station, Antarctica, which now extends for 16 consecutive years. It includes a further 7 years of measurements since the previous published trend assessment using these data [French et al., 2005]. It also includes a change in the instrument response calibration procedures (described in section 2) which produces a small correction to the derived rotational temperatures from previously reported work.

2. Instrumentation and Calibration

[11] Hydroxyl nightglow spectra used in this study are measured from Davis using a 1.26 m focal length f/9 scanning-grating spectrometer of Czerny-Turner design. It is a relatively simple and robust instrument which has the significant advantages of long-term stability and low operational cost for long-term studies.

[12] Scans of the OH(6–2) band P-branch lines, are made in the zenith with a 5.3° field of view and with an instrument resolution of ∼0.16 nm. Automated acquisition software starts and stops the observational routine, regardless of cloud conditions, when the sun is greater than 8° below the horizon. The observing season is limited at this latitude to between mid-February (∼day 48) and the end of October (∼day 300) each year. A thermoelectrically cooled GaAs photomultiplier, operated in pulse counting mode, records the emission. From 1997 the acquisition time per spectrum is about 7 min. Only the P-branch lines, background sample regions and the atomic oxygen line at λ844.6 nm (as an auroral activity indicator) are sampled and these are interpolated to a common time between consecutive spectra to allow for emission intensity changes during the course of the scan. During 1995–1996 continuous scans across the OH(6–2) region were made, which took of the order of 40 min to acquire. Further details of the instrument are provided by Greet et al. [1998] and French et al. [2000].

[13] The most critical element of long-term trend assessment for this type of instrument is the stability of the spectral response calibration. Here, it is maintained by reference to several Low Brightness Source (LBS) units, which are cross referenced annually to Australian National Measurement Institute (ANMI) standards. The instrument spectral response characterization over the 1995 to 2010 interval considered in this work, is comprised of over 2340 scans of the LBS units on the spectrometer at Davis and over 490 cross-reference scans of the LBS against ANMI standards. Significant changes to the instrument include replacement of the diffraction grating in 1995, replacement of the GaAs photomultiplier detector in 2001 and replacement of the acrylic roof hatch window in 2004. Care is taken to fully characterize each instrumental change by reference to the LBS units. The change in instrument spectral response correction is quantified by the P1(2)/P1(5) ratio (the most widely separated ratio used for a rotational temperature calculation and which contributes, on average, 59% of the weighted mean temperature). Over the 16 years, this correction ratio has not changed by more than 1.1%, corresponding to a maximum temperature change of 1.1 K (i.e., temperatures would change by a maximum 1.1 K if the same response correction was applied for all years). The correction uncertainty (standard deviation in repeated scans) is generally less that 0.3 K each year, with the exception of 1995 (1.8 K) due to calibration via a secondary calibration lamp and 2002 (1.2 K) due to detector cooling problems.

[14] Data used in this study have been corrected for an issue with the orientation of the LBS compared to previous published work [Burns et al., 2003; French and Burns, 2004; French et al., 2005]. This was due to measurements at Davis being taken with the LBS mounted in the vertical position but cross referenced with measurements taken at ANMI with the LBS in the horizontal orientation. A measurable difference in the spectral radiance of the LBS was detected in 2007 as a result of the altered shape of the tungsten filament in the different orientations of the LBS. Correcting the AMNI calibrations to match the Davis vertical measurements results in cooler OH rotational temperatures each by an average 0.9 K in absolute terms, but has a negligible affect on the trend assessment.

3. Temperature Data

[15] Sample temperatures are derived as a weighted average of temperatures from the three possible ratios from the P1(2), P1(4) and P1(5) emission lines. The weighting factor is the statistical counting error (based on the error in estimating each line intensity). P1(3) is contaminated by the unthermalized OH(5–1) P1(12) line so is not used [French et al., 2000]. P1(2) is corrected for the ∼2% temperature-dependent contribution by Q1(5). Background regions are selected to balance the small auroral contribution (mainly from the N21PG(3–2) and N2+ (4–2) Meinel bands) and solar Fraunhofer absorption (for spectra acquired during moonlit conditions) under each P-branch line. This is achieved with the use of theoretical N2 spectra programs provided by D. Gattinger (personal communication, 1998) and the solar spectrum available in the ModTran database [Berk et al., 1987]. The defined backgrounds were the mean of counts between 840.55 and 840.85 nm for P1(2), the mean of 843.8–843.9, 848.14–848.24 and 847.8–847.9 nm for P1(4) and 847.9–848.45 nm for P1(5). The success of these selections in accounting for auroral activity is demonstrated by plotting the temperature versus the atomic oxygen auroral indicator sampled each scan. No significant effect of aurora on the derived temperature is found. Error estimates for each P-branch line, taken as the square root of the total number of counts plus the standard error of the mean in the background are propagated through the rotational temperature equation to derive the statistical counting error.

[16] Correction factors account for the difference in Λ-doubling between the P-branch lines determined with knowledge of the instrument line shape from high-resolution scans of a frequency-stabilized laser. Langhoff et al. [1986] transition probabilities are used to derive rotational temperatures because they are closest to the experimentally determined ratios of French et al. [2000] for the OH(6–2) band. Choosing different transition probabilities from the published sets for the OH(6–2) band [e.g., Mies, 1974; Turnbull and Lowe, 1989] can change the derived temperatures by up to 12 K, but does not significantly affect the trend analysis reported here.

[17] Selection criteria limit extreme values of weighted standard deviation and counting error, slope and magnitude of the background and the rate of change of branch line intensities between consecutive scans. Further details of the rotational temperature analysis procedure are provided by Burns et al. [2003] and French and Burns [2004].

3.1. Trend Data Set

[18] The Davis rotational temperature data set currently comprises over 237,000 individual measurements which pass selection criteria, from which 3392 nightly averages are obtained. Nightly averages are calculated from all individual measurements within ±12 h of local midnight (∼18:50 UT) and are considered valid if there are at least 10 contributing samples (5 for 1995–1996). The time series of nightly averages are plotted in Figure 1a on top of the MSISE-90 model temperatures (a 4 h average around local midnight) for 87 km altitude and 69°S [Hedin, 1991].

Figure 1.

The Davis OH nightly average temperature record 1995–2010. A total of 3392 nightly averages. (a) The nightly averaged temperatures over the MSISE-90 model temperatures for 69°S. The error bar indicated is the mean standard error of the mean. Solid circles are the winter average temperatures. (b) The standard error of the mean for the nightly temperatures. (c) The nightly mean, and winter mean residual temperatures, after the smoothed climatology is removed, and (d) the F10.7 solar flux index. The gray line shows the contiguous daily F10.7 measurements, and the overplotted symbols indicate when there are coincident OH temperature measurements.

[19] MSISE-90 is shown here as a guide to illustrate the range of the full seasonal cycle that it is possible to sample at this latitude. A comparison study with MSISE-90 has previously been provided by Burns et al. [2002], who concluded that although there was reasonable agreement between model and data, the MSISE-90 model is limited in its representation of the seasonal cycle as only annual and semiannual terms are modeled, and a large (20 K) semidiurnal tide in the model at this latitude is not present in the data. Measured tidal amplitudes in the Davis OH temperatures are all < 2 K (therefore we do not expect the variation in observing hours over the season to significantly bias the nightly means). As pointed out by Azeem et al. [2007], the model also does not reproduce the observed solar cycle dependence.

[20] Standard error of the means (SEM) for the nightly averages are shown in Figure 1b. Higher SEMs are obtained in 1995 and 1996 due to the ∼6 times longer acquisition mode.

3.2. Mean Climatology

[21] Superposition of the time series by day of year is shown in Figure 2, which reveals the seasonal variation in nightly average temperatures. A fit of annual, semiannual and terannual terms to these data gives amplitudes of 42.7 K (maximum on day 175), 23.6 K (maximum on day 89) and 7.8 K (maximum on day 178), respectively; however, the seasonal fit is unconfined across the summer interval. More useful is the mean winter climatology, derived here as a 5 day running mean through the average of nightly temperatures for each day of year. This is similar to the climatological mean (or “seasonal composite”) derived by Azeem et al. [2007] for South Pole OH temperatures, except they use a 30 day running mean. Seasonal characteristics found in previous analyses [e.g., Burns et al., 2002] persist in the extended data set. These include the rapid temperature rise in autumn (February–March) at 1.2 K/d, a sharp transition to winter temperatures on day 85, a slow decline across the winter months (April–September) at −0.04 K/d and the spring decline (September–October) at −0.5 K/d. The mid-April (∼day 113) and mid-August (∼day 227) coolings are also characteristic, and appear to correspond to southward-to-northward reversals in the mean meridional flow, as measured by a collocated medium frequency radar. [e.g., see Dowdy et al., 2007, Figure 4].

Figure 2.

A superposed epoch of the 16 years of Davis OH nightly averaged rotational temperatures plotted as points. The solid green curve is a seasonal fit of annual, semiannual, and terannual terms. The solid orange line is a 5 day smoothed climatological mean with ±1 standard deviation indicated by the black lines. Vertical lines indicate days 106 and 259, the region over which the winter mean temperatures are calculated.

[22] Subtraction of the mean climatology from each year yields the residual temperatures plotted in Figure 1c. These residual or seasonally detrended temperatures are used for the trend analysis in section 4.

[23] Winter average temperatures are defined for each year as the average between days 106 and 259 (marked by vertical lines on Figure 2) to avoid the autumn and spring transition periods. These are plotted both as temperatures, and residuals as the solid points on Figures 1a and 1c.

4. Trend Analysis

[24] It is common practice for multivariate or multiple regression analysis to be used for trend analysis of temperature time series with various terms included to describe sources of temperature variability considered to be contained in the data [e.g., Azeem et al., 2007]. For MLT region observations, these sources include the seasonal variation (annual harmonics, largely a result of the mean meridional flow), planetary wave activity, a solar flux dependence, other shorter term periodic sources such as tides, gravity waves, the quasi-biennial oscillation (QBO), impulse events such as sudden stratospheric warmings, coronal mass ejections (CME), volcanic eruptions and long-term chemical effects such as ozone depletion. The remaining linear regression coefficient is then ascribed as the long-term trend.

[25] Where there is a measurable quantity or index characterizing the source of variability a regression coefficient can be determined which describes the dependence of the temperature measurement on that quantity. There are a number of measures of solar variability (F10.7, sunspot number, Lyman-α, MgII index), QBO (30 hPa, 50 hPa winds) and trace gas concentrations (CO2, O3, water vapor) for example which can be tested for their influence on the measured temperature. Periodic oscillations with known forcing mechanisms such at the seasonal variation, tides and some photochemistry can be modeled theoretically in the regression analysis. It is important to consider both magnitude and phase of the response to these sources as the variation may lead or lag its signature in the measured OH temperature.

[26] Some known sources of variability are less well characterized in terms of their impact at MLT altitudes. Planetary waves in particular strongly influence the OH temperatures with variations up to 40 K and periods of weeks to months [French and Burns, 2004] but vary greatly in character at different levels in the atmosphere due to wave–mean flow and wave-wave interactions. A planetary wave index or proxy, such as the Dynamical Activity Index (DAI) calculated from satellite ozone measurements (e.g., from the Total Ozone Mapping Spectrometer (TOMS) [Grytsai et al., 2005]) which is largely dominated by dynamical processes in the lower stratosphere may not reflect the variability at OH layer heights.

[27] This analysis begins with the simplest regression model containing only a solar activity index and long-term linear trend terms. Monthly and winter averages are used to reduce the influence of variations due to planetary wave and the seasonal variation is removed by subtracting the climatological mean as described above. This analysis provides the trend coefficients comparable with most other reported analyses of MLT region temperature trends.

4.1. Solar Cycle Correlation Analysis

[28] Temperature in the MLT region is known to respond to the solar activity cycle and estimates of the magnitude and phase of the relationship are sought in observational data sets as soon as they reach a sufficiently long sample time. Certainly, this natural variability must be taken into account when searching for longer-term trends in these data.

[29] The most recent review of observed and modeled solar cycle responses in the MLT region was conducted by Beig et al. [2008]. Generally, 2–4 K per 100 SFU is a consistent response in the observations and this agrees with numerical simulations (e.g., Marsh et al. [2007] with WACCM and Schmidt et al. [2006] with HAMMONIA). It is pointed out in this review that the thermal structure in the MLT may respond directly to variations in solar photon energy, or indirectly through modifying the ozone abundance in the lower atmosphere, leading to changes in the heating and cooling rates.

[30] The F10.7 cm solar flux index (1 solar flux unit = 10−22 Wm−2Hz−1) is shown in Figure 1d and it is apparent that a positive correlation exists. Figure 3 shows a scatterplot of the 3392 nightly average residual temperatures against the daily F10.7 value. The straight line is a best fit to the data and yields a solar response coefficient of 4.1 ± 0.25 K/100 SFU. This result remains consistent with previous analyses on shorter time series of these data [e.g., French et al., 2005]. The R2 (coefficient of determination) is 0.074 but the Pearson critical level for 95% significance is 0.034, so the correlation is significant. However, a large proportion of the temperature samples occur at low solar activity (60% < 100 SFU, 82% < 150 SFU), so to some extent the relationship is biased by the fewer points at high solar activity.

Figure 3.

A scatterplot of the 3392 nightly average residual temperatures against the daily F10.7 value. The straight line is a best fit to the data and yields a solar response coefficient of 4.1 ± 0.25 K/100 SFU. The R2 is 0.074 and the Pearson critical level for 95% significance is 0.034.

[31] As the day-to-day temperature variability is large (up to 40 K over 40–50 days due to planetary wave processes) compared to the derived solar cycle signal (∼5 K over a full solar cycle) we endeavor to reduce the effect of planetary wave variability on the correlation analysis by averaging on a monthly time scale. A lag correlation analysis of monthly averaged temperatures with monthly averaged F10.7, evaluated at 5 day lag intervals is shown in Figure 4. A broad correlation peak occurs where F10.7 leads the temperature by ∼160 days. The R2 maximum is 0.34 which is well above the 99% confidence level of 0.23. This peak is 8% higher than at than at zero lag (R2 = 0.26). Even so, only 34% (Rmax2 = 0.34) of the temperature variability in the monthly averages can be attributed to solar flux variation. The monthly averaged solar flux response at 160 days lag is 4.63 ± 0.60 K/100 SFU, compared to 4.2 ± 0.66 K/100 SFU at zero lag. We explore this apparent lag further using the multiple linear regression analysis in section 4.2. As alluded to earlier, the problem is the regression slope is dominated by the high solar activity points in one particular year.

Figure 4.

Lag correlation of monthly averaged residual temperatures with monthly averaged solar activity indices (F10.7, Lyman-α, E10.7, Qeuv (ergs), and S, the solar constant in W/m−2), evaluated at 5 day lag intervals.

[32] Other measures of solar cycle variability include for example the Solar EUV Monitor (SEM)/Solar Heliospheric Observatory (SOHO) EUV measurements [Judge et al., 1998], the NOAA MgII core to wing ratio, Lyman-α (121.6 nm) irradiance and the He 1083 nm infrared absorption index. Liu et al. [2006] compare these indices and find nonlinear relationships between them and the F10.7 index. It is possible that one of these indices may more accurately represent the process by which temperatures in the OH layer are influenced by solar variability. As Scheer et al. [2005] point out, there is no a priori reason to expect F10.7 to be the best possible index of geoeffective solar activity.

[33] The Solar Irradiance Platform (SIP) model [Tobiska et al., 2000] (formerly SOLAR2000; SIP model may be accessed at conveniently compiles many solar indices for long-term climate change studies and a selection of these were chosen to compare with F10.7 to examine if the correlation analysis changed or improved. Correlations for the Lyman-α irradiance and E10.7 (integrated EUV flux from 1 to 105 nm in SFU) indices, (together with their 81 day averages), the solar constant S(Wm−2) and the terrestrial extreme ultraviolet heating rate Qeuv (ergs cm−2 s−1) are also plotted on Figure 4 for comparison.

[34] In general, all indices show the same broad and offset correlation peak. The E10.7 index gives the highest correlation, but only slightly better than F10.7 and both show generally the same correlation features. The Lyman-α (121.6 nm) irradiance and EUV heating rate indices tend to show shorter offsets near 100 days. As Lyman-α is directly absorbed in the MLT region by molecular oxygen and water vapor, it is reasonable that this might be more directly correlated.

[35] The largest variation in flux between solar maximum and solar minimum occurs at the shortest (EUV) wavelengths. Major absorbers of energetic EUV radiation are molecular oxygen (O2) and ozone (O3) and these undergo different absorption processes at different levels in the atmosphere. The heating rate profile due to these two molecules has been characterized by Strobel [1978]. Photodissociation of O2 occurs in the MLT between about 80 and 130 km as it absorbs in the Schumann-Runge continuum (130–175 nm). The atomic oxygen produced is transported by eddy diffusion to lower altitudes where its three-body recombination occurs more readily and the release of energy (∼5.12 eV per reaction) is an important heat source below 90 km [Roble, 1995]. Through the mesosphere and upper stratosphere (40–95 km) O2 electronic and vibrational transitions occur with absorption in the Schumann-Runge bands (175–200 nm). Below about 50 km the well known photodissociation of O3 by absorption in the Hartley-Huggins bands (200–350 nm) dominates in the stratosphere, although the O2 Herzberg continuum (200–240 nm) absorption also contributes in this region. Significant energy is also deposited in the lower stratosphere through O3 absorption in the Chappuis bands in the visible and near infrared (450–750 nm). Further into the infrared, water vapor and CO2 become the dominant absorbers, but the variation in IR flux over the solar activity cycle is greatly reduced. Radiative, chemical, and physical transport processes redistribute the absorbed solar energy throughout the atmosphere on different time scales and the understanding of these processes continues to be undertaken in many coupled chemical-dynamical models [e.g., Roble and Dickinson, 1989; Rasch et al., 1995; Fomichev et al., 2002; Schmidt et al., 2006; Garcia et al., 2007].

4.2. Long-Term Trend Multivariate Regression Analysis

[36] Beyond simple correlations with solar activity indices an initial multivariate fit model of the form TOH = L.Year + S.F10.7 + C seeks to detect the long-term linear trend coefficient (L) coupled with the solar cycle coefficient (S) in the time series of monthly averaged, or winter averaged OH residual temperatures (TOH). As stated above, residual temperatures are derived by removing the climatological mean from each year, in effect, removing the seasonal variation components from the regression model.

[37] Winter average temperatures are taken as the average of all nightly average temperatures between days 106 (15 April) and 259 (15 September) although the results are not significantly different if the full observing window (nominally days 049–296) is selected.

[38] Figure 5a shows the multivariate regression fit to the winter average temperatures between 1995 and 2010, calculated with no F10.7 lag. Figures 5b and 5c show the decoupled solar cycle and long-term trend components. Clearly, some years depart significantly from the regression fit, notably 2007 and 2002 are much warmer, while 2001 and 2010 much cooler than this simple model would predict. The long-term trend and solar cycle coefficients are compared in Table 1 for the full year, winter, monthly and nightly means computed with this regression model. The coefficients are highly consistent, within 1σ errors, between each of the analysis sets. The R2 value increases with the averaging window, consistent with the averaging out of planetary wave variability.

Figure 5.

A multivariate regression fit to (a) the residual winter average (D106–259) temperatures from 1995 to 2010 with (b) extracted solar and (c) long-term trend components. Straight lines are best linear fit (defined by equations). Error bars are standard error of the mean.

Table 1. Long-Term Trend and Solar Cycle Coefficients and 95% Confidence Limits for the TOH = L.Year + S.F10.7 + C Regression Model Derived From the Yearly Mean, Winter Mean, Monthly Mean, and Nightly Mean Residual Temperaturesa
 SamplesLong-Term Trend L (K/decade)Solar Cycle S (K/100 SFU)
  • a

    Boldfacing denotes trend coefficients that are the key parameters derived from the analysis.

Full year means R2 = 0.72161.03 ± 0.77 0.64 > L > −2.96 [95%]4.35 ± 0.87 6.24 > S > 2.46 [95%]
Winter means R2 = 0.70161.18 ± 0.87 0.71 > L > −3.06 [95%]4.79 ± 1.02 6.99 > S > 2.60 [95%]
Monthly means R2 = 0.271281.00 ± 0.62 0.23 > L > −2.24 [95%]3.89 ± 0.69 5.25 > S > 2.53 [95%]
Nightly means R2 = 0.0833921.07 ± 0.25 −0.57 > L > −1.57 [95%]3.77 ± 0.69 4.28 > S > 3.27 [95%]

[39] This analysis shows a statistically significant solar cycle response of around 4 K/100 SFU and a very small long-term cooling trend of the order of 1 K per decade in the Davis OH data series. The 95% confidence limits do not rule out a zero trend in the 16 year data set although it is excluded at the 1σ uncertainty level.

4.2.1. Trend Result Comparisons

[40] These results are directly comparable with other reported observations of solar cycle and long-term change in OH temperatures. Most of the long-term trend results are reported in the review by Beig et al. [2003] (see their Table 5) and the update by Beig [2006] with an overview of solar activity responses by Beig et al. [2008] (see their Table 3a). More recent updates of these entries are also provided by Dyrland and Sigernes [2007], Scheer et al. [2005] and Offermann et al. [2010]. Table 2 summarizes the most recently published solar cycle and long-term trend coefficients derived from hydroxyl measurements for selected sites. Further comparisons, including with trends derived from other techniques can be made utilizing the Beig et al. reviews.

Table 2. Solar Cycle and Long-Term Trend Estimates From Hydroxyl Layer Measurements at Selected Sites, Ordered by Latitude, Compared With the Davis Winter Mean Resultsa
SiteData SpanLong-Term Trend L (K/decade)Solar Cycle S (K/100 SFU)Reference
  • a

    Davis winter mean results are from Table 1.

South Pole (90°S)1994–2004Not significant4 ± 1Azeem et al. [2007]
Davis Station (69°S, 78°E)1995–2010−1.2 ± 0.94.8 ± 1.0This work (2011)
El Leoncito (32°S, 69°W)1986–2002−10Not significantReisin and Scheer [2002]Scheer et al. [2005]
Cachoeira Paulista (23°S, 45°W)1987–2000−10.8 ± 1.5∼5Clemesha et al. [2005]
Wuppertal (51°N, 7°E)1988–2008−2.3 ± 0.63.5 ± 0.21Offermann et al. [2010]
Zvenigorod (56°N, 37°E)2000–2005−5 ± 1.74.5 ± 0.5Semenov et al. [2002]Pertsev and Perminov [2008]
Stockholm (57°N, 12°E)1991–1998+5 ± 21.6 ± 0.8Espy and Stegman [2002]
Yakutia (63°N, 129°E)1982–2008−5.3 ± 1.2N/AAmmosov and Gavrilyeva [2010]
Svalbard (78°N, 15°E)1980–2005+2 ± 10 ± 0.5Sigernes et al. [2003]Dyrland and Sigernes [2007]

[41] Of particular interest to this work are the other long time series Antarctic results from South Pole station. Azeem et al. [2007] derive a solar cycle coefficient of 4 ± 1 K/100 SFU from 11 years (1994–2004) of Michelson Interferometer observations of the OH(3–1) band using a very similar multiple regression analysis of 30 day mean temperatures. While their long-term trend coefficient (+1 ± 2 K/decade) is not statistically significant the solar cycle result is in complete agreement with the present Davis result. In addition, as Azeem et al. [2007] point out, it is about one third of the earlier estimate by Hernandez [2003] of 13.2 ± 3.5 K/100 SFU derived from 13 years (1991–2003) of Fabry-Perot spectrometer OH observations, also at South Pole.

[42] Other hydroxyl layer trend estimates in the Southern Hemisphere include those by Clemesha et al. [2005] who derive 6.0 ± 1.3 K/solar cycle (∼5 K/100 SFU) and −10.8 ± 1.5 K/decade from 14 years of OH(9–4) and OH(6–2) band photometer observations from Cachoeira Paulista (23°S, 45°W) and Scheer et al. [2005], who find no evidence yet of a solar cycle effect (−0.14 ± 0.33 K/100 SFU) in a relatively short time series of 5 years (1998–2002) of OH(6–2) band observations from El Leoncito (32°S, 69°W), but a large long-term trend of −10 K/decade if additional earlier campaign data from 1986 to 87 and 1992 (a 15 year span) were included [Reisin and Scheer, 2002].

[43] Interhemispheric high-latitude comparisons can be made with the important long-term measurement sites at Zvenigorod (56°N, 37°E), Wuppertal (51°N, 7°E), Svalbard (78°N, 15°E), Yakutia (63°N, 129°E) and Stockholm (57°N, 12°E). Table 2 lists the range of coefficients for these sites. In general, a solar cycle response near 4 K/100 SFU is most consistent, although the higher northern latitude sites show a smaller, or no response. Long-term trend estimates are much more variable; 10 K/decade cooling to 5 K/decade warming, and the Davis estimate certainly falls within these limits. Further aspects of these trends are explored in section 4.2.2.

4.2.2. Seasonal Variability of Trends

[44] Increasingly, trend surveys are reporting solar cycle and long-term trends separated into winter and summer seasons, or even trends for each month. This is entirely reasonable, as the temperature response in the MLT region is likely to be subject to the seasonal dependencies in UV radiation, wind and wave activity, and in the concentration of chemical species (particularly O and O3).

[45] To explore this seasonality in the Davis data series further, the multivariate analysis was computed in a 30 day sliding window across each of the 16 years, and extended to include the F10.7 lag at each lag step, thereby producing a correlation map shown in the bottom left of Figure 6. Each grid point on the map contains the regression analysis of the 16 points derived from the 30 day residual temperature averages centered on the day of year on the ordinate axis, with the 30 day F10.7 average, centered on the day of year and offset by the lag on the abscissa. Month labels on the ordinate axis mark the center of each month. Clockwise from Figure 6 (bottom left) are the Pearson correlation coefficient (R), the R2 coefficient of determination (significant at the 95% level > 0.514), the solar cycle and the long-term trend coefficients returned from the regression fit. A strong seasonal dependence is revealed. Correlation maxima occur in March, May–June and September and minima in April and August coincident with the mid-April and mid-August temperature dips identified in the climatology (section 3.2). Correlations are mostly positive, except in mid-August, and biased on the positive lag side (maximizing at the 160 day offset we found previously). However, the broad spread indicates that the correlation is relatively insensitive to F10.7 lag. Note that the sloped line structure in the correlation maps is due to the 1:1 matching of the temperature window to the offset F10.7 window.

Figure 6.

Coefficients of a multivariate lag correlation analysis computed in 30 day sliding windows across each of the 16 years of residual temperatures, with F10.7 lag on the abscissa. (bottom left) Pearson correlation coefficient (R), (top left) the R2 coefficient of determination (significant at the 95% level > 0.514), and the (top right) solar cycle and (bottom right) long-term trend coefficients. Month labels on the ordinate axis mark the center of each month.

[46] Generally, the solar cycle response mirrors the R2 coefficient, with the largest solar response (∼7 K/100 SFU) in March, May–June and September and little or no solar response in April and August. The long-term trend coefficient shows the largest cooling rate (4–5 K/decade) in August–September, through to warming (2–3 K/decade) for the March and May–June periods.

[47] Several iterations of this analysis were performed to test the resilience of the F10.7 lag maximum at 160 days. We considered whether the correlation was dominated by the short-term variations or the long-term cycle by separating the F10.7 data into a smoothed (1 year) running mean variation, and the residual (short-term) variations after the long-term running mean was subtracted. We then also sequentially removed each individual year from the series to see if the variations in any particular year were responsible for the 160 day offset in the correlation map. It was apparent from these analyses (not shown here) that indeed, the short-term variations in one year, 2002, were largely responsible for the 160 day offset correlation maximum in mid-June. Although we do not wish to exclude 2002 from the trend analysis it was a particularly unusual year. It is the warmest in the 16 year record, with the highest F10.7 activity and contained the unprecedented Southern Hemisphere stratospheric warming [Baldwin et al., 2003; French et al., 2005]. The coincidental high correlation of temperature with short period solar activity variations in 2002 has a dominant influence on the slope of the temperature versus F10.7 relationship over one solar cycle of measurements and in this respect the 160 day lag result is misleading. More data (at least through another solar maximum) are required to resolve this matter and until this can be achieved we view the F10.7 lag results with caution and proceed with evaluating the trend coefficients at zero lag.

[48] Figure 7 shows the 30 day sliding window solar cycle (Figure 7a) and long-term (Figure 7b) trend evaluations at zero lag (i.e., this is the vertical transect through Figure 6 (right) at zero lag). One sigma error bars and the 95% confidence limits are also shown. Again it is noted that overall 95% confidence limits do not yet rule out a zero long-term trend in these data. The (true calendar) monthly evaluations are also plotted and labeled.

Figure 7.

The 30 day sliding window (5 day step) evaluations of (a) solar cycle and (b) long-term trend coefficients at zero F10.7 lag (i.e., the vertical transect through Figure 6 (right-hand panels) at zero lag). One-sigma error bars and the 95% confidence limits (upper and lower traces) are as marked. The (true calendar) monthly evaluations are also plotted and labeled. Overlaid on Figure 7b are the equivalent monthly trend results by Offermann et al. [2010] from Wuppertal (labeled WUP) and Espy and Stegman [2002] from Stockholm (labeled STO). These are approximate values scaled off their respective seasonal trend plots and are offset by 6 months to match the Southern Hemisphere season.

[49] Overlaid on Figure 7b are the recent results by Offermann et al. [2010] from Wuppertal (labeled WUP) and Espy and Stegman [2002] from Stockholm (labeled STO). These are approximate values scaled off their respective seasonal trend plots and are offset by 6 months to match the Southern Hemisphere season. There is a surprising similarity in the seasonal trend characteristics with the Wuppertal measurements, not only in trend magnitude but in particular the April (Autumn) and August (Spring) cooling maxima and little apparent cooling over the winter months. To some extent the Stockholm results also show some similarity, given that these trends were derived from only 7 years of measurements with no fitting of a solar cycle component. In contrast, the seasonal trend results by Semenov et al. [2002], obtained by combining hydroxyl observations from six northern midlatitude stations show a more-or-less sinusoidal pattern with long-term cooling of ∼10 K/decade over the winter months and little or no cooling over the summer months.

[50] At least with the Wuppertal comparison, the agreement between time-shifted Northern Hemisphere and Southern Hemisphere results suggest that the trends are modulated by seasonal effects rather than global change effects. The spring (SH August, NH February) cooling is a particularly strong feature and we see from examination of the Davis temperature series that the mid-August temperature dip is deepest during the solar maximum years (2000–2003). Changes in mean wind and planetary wave activity with the solar cycle have been previously reported [e.g., Jacobi et al., 1998; Schmidt and Brasseur, 2006] and it is possible that the particular combination of wind and wave events that cause the mid-August minimum are strengthened at solar maximum. We can only infer so much from one solar cycle of measurements however, and this feature is the subject of future analysis and modeling work.

4.3. Shorter-Term Variations

[51] In this section, efforts are made to characterize other known or anticipated sources of variability in the OH temperature series by extension of the multivariate fit model. A lag correlation regression analysis for each variable is performed in the same manner as section 4.2.2 except now using the residual temperatures after climatological mean, zero-lag solar cycle and long-term trend terms we quantified in section 4.2 have been removed.

4.3.1. Planetary-Scale Wave Variations

[52] Temperature oscillations of up to 40 K over periods of 10–60 days are observed in the residual time series. Planetary wave (PW) activity is likely to be a dominant source of this variability both directly and through modulation of the gravity wave flux. Figure 8a presents a Lomb-Scargle periodogram of the amplitudes of waves with periods between 5 and 50 days in the residual temperatures (plotted in Figure 8b), and shows that the large amplitude, long period oscillations vary greatly in character from year to year. In some years they propagate throughout the winter observing period, while in others there are only short bursts of activity at different times in the season.

Figure 8.

(a) A Lomb-Scargle periodogram showing the amplitude of wave periods between 5 and 50 days in the residual OH temperature data series. The x axis is the center of a 100 day sliding window, tick marks denote 1 January of each year. (b) Monthly average (connected points) and nightly average hydroxyl temperatures (with seasonal climatology, solar cycle and long-term trends removed). (c) Zonal Fourier decomposition of the 10 hPa geopotential height for 67.5°S from NCEP Reanalysis-2 data, showing the strength of PW 1–3 and their combined SUM as labeled. (d) A zonal wind anomaly polar vortex intensity (m/s positive eastward) calculated relative to the mean zonal wind climatology from NCEP Reanalysis-2 10 hPa level data. (e) The standardized 30 hPa Quasi Biennial Oscillation (m/s negative eastward) and (f) the Southern Annular Mode Index (hPa) calculated from zonal mean sea level pressure from 12 stations between 45°S and 65°S [Marshall, 2003].

[53] Various indices such as the Dynamical Activity Index (DAI) based on Total Ozone Mapping Spectrometer (TOMS) data [Grytsai et al., 2005] or the Polar Vortex Intensity (PVI; see section 4.3.2) capture characteristics of planetary wave activity. Following the DAI technique of Fourier decomposition of zonal mean ozone measurements into PW-1 and PW-2 components, here, a measure of PW activity appropriate to Davis is derived by zonal Fourier decomposition of the 10hPa geopotential height at 67.5°S from NCEP Reanalysis-2 data [Kanamitsu et al., 2002]. Variations in the strength of PW 1–9 components are derived, and Figure 8c plots the first three (PW 1–3) to show their relative strengths, together with their combined sum. Although the 10 hPa level is still in the upper stratosphere (∼32 km altitude) it is the top level available in the NCEP reanalysis data set.

[54] It is well known that propagation of planetary and gravity waves from the lower atmosphere into the MLT region is strongly dependent on filtering effect of the zonal mean wind. In an eastward flow planetary waves propagate while in a westward flow the waves are absorbed [Charney and Drazin, 1961]. Eastward zonal flow dominates the winter observing period at Davis and westward flow in the summer, with reversals occurring in early March (∼day 70) and early October (∼day 285). This is clearly apparent in the 10 hPa PW amplitudes where wave activity is at a minimum over the summer months and builds during the winter months, maximizing in September–October. This seasonal distribution of PW activity is also seen for the Northern Hemisphere by Offermann et al. [2009], who have analyzed SABER data to derived wave climatologies from zonal mean temperature standard deviations.

[55] Lag correlation analyses of temperature and wave index were performed using the nightly mean residual temperatures (Figure 8b) and the summed planetary wave index and each of the harmonic modes independently. Correlation is not significant over the entire time series, but monthly sliding window samples give significant correlations (R2 up to 0.8) for the combined sum and the PW-1 and PW-2 modes individually with maximum positive correlations typically within a few days lag. It is clear that there are times when planetary waves in the upper stratosphere, defined by this index, propagate through to the OH layer. The highest correlations occur in late winter (September–October) when the zonal eastward winds permit PW propagation and their amplitudes are largest. A notable 14 day PW-1 event occurred in August–Oct, 2002 for example and has previously been examined by French et al. [2005].

[56] However, it is also clear that there are times when there is no correlation, indicating that the waves either do not penetrate to the MLT region, or are substantially modified in character by wave-wave or wave–mean flow interactions. As such, these PW indices are not consistent enough to be used to reduce variability in the OH temperatures and improve the trend analysis. Further development of the coupling and interactions of waves from the lower atmosphere using satellite observations, assimilated data sets and models may yet provide a representation of the observed PW activity and this is a direction for future study.

4.3.2. Polar Vortex Intensity

[57] Another measure of wave driving in the stratosphere, which has a strong impact on temperature, is calculated as the deviation of the average polar night jet strength from climatology. In an analysis of the unusual conditions in the Southern Hemisphere winter of 2002, Newman and Nash [2005] use the zonal mean zonal wind anomaly relative to zonal mean climatology, derived from reanalysis data as a measure of the polar vortex intensity (PVI). Figure 8d plots the 10 hPa PVI (in m/s westward). A notable feature of the trace is the strong disturbance to the westerly winds in late 2002 associated with the unusual stratospheric warming and early breakup of the ozone hole that year.

[58] The same lag correlation as for the planetary wave indices was performed for the PVI and again, while a significant correlation was not found for the entire time series, 30 day sliding window samples do yield significant correlations (R2 up to 0.7) around zero lag in windows across late winter. Where significant correlation is obtained the coefficient is negative, i.e., stronger easterly winds are associated with greater planetary wave penetration and higher residual temperatures. It is most likely that the PVI gives some insight into the ‘transparency’ of the atmosphere to the propagation of waves through to the MLT region where they influence the temperatures measured here, but in the present analysis the index is again not consistent enough to be used to reduce residual temperature variability.

4.3.3. Quasi-Biennial Oscillation

[59] The QBO has a well defined signature in the equatorial stratospheric zonal winds and although remote from the polar MLT region, is worth testing for an influence. Figure 8e plots the 30 hPa QBO index, derived as the monthly averaged zonal wind anomaly (negative eastward) compared to the mean zonal flow. In this case, monthly averaged OH residual temperatures were used in the lag correlation analysis to search for a QBO response. Batista et al. [1994] have previously sought to identify a QBO signature in OH temperatures by dividing residual temperatures into the east and west phase of the QBO. No apparent difference was found in 9 years of measurements at the low latitude station Cachoeira Paulista, Brazil (23°S, 45°W). For the Davis data, we correlate the monthly average residual temperatures with the monthly averaged QBO index and a peak in correlation coefficient is found where the QBO leads the OH temperature by about 3 months, but the correlation is only at the 5% level (R2 = 0.05). It is difficult to resolve the QBO signal as the monthly averaged residual temperatures are still dominated by planetary wave variability. However, it is expected that QBO should lead the OH temperature as this is consistent with the downward phase progression of the QBO response in the atmosphere. If the OH temperatures are divided into the east and west phase of the QBO, with a 90 day lag, the mean temperature is 0.7 K (standard error of the mean 0.4 K) higher in the QBO west phase (+ve) than the east phase (−ve). Gao et al. [2010] have recently examined the global distribution of QBO (as well as annual and semiannual oscillations) in OH intensities using the SABER data set. They found QBO amplitude peaks at the equator, and at 35°N and 35°S, with a fall off at higher latitudes. Furthermore, a latitudinal dependence of the QBO period was found; the oscillation becoming longer at higher southern latitudes. Thus, as well as being masked by the planetary wave variability, the QBO signal at high southern latitudes is expected to be small in amplitude and extend over a longer period than the tropical 30 hPa index compared to here. Given these factors, the present analysis cannot identify a significant QBO response in the Davis OH temperatures.

4.3.4. Southern Annular Mode

[60] Another dominant mode of variability in the mid to high southern latitude atmospheric circulation pattern is known as the Southern Annular Mode (SAM). The monthly SAM index, plotted in Figure 8f is calculated from zonal mean sea level pressures (MSLP) from 12 stations and is the difference between the normalized monthly MSLP at 45°S and 65°S (data available at [Marshall, 2003]). The index essentially characterizes fluctuations in the strength of the SH circumpolar flow due to the pressure gradient. In the positive phase of the SAM, pressures over Antarctica are relatively low compared to those in the midlatitudes and enhanced Southern Ocean westerlies occur, intensifying the circumpolar vortex [Marshall, 2003]. A general increase in the SAM index beginning in the 1960s has been reported by Marshall [2003] and linked to stratospheric ozone depletion and Antarctic surface temperature trends [e.g., Thompson and Solomon, 2002].

[61] Similarities can be seen between the SAM and PVI indices in Figure 8 and this is reasonable as they both measure aspects of zonal wind strength, albeit at different atmospheric levels. Results of the correlation analysis on monthly averaged data were consequently similar to the PVI index; while reasonably good correlations exist for windowed intervals, overall the correlation is weak. Again, the strength of the SAM index gives a measure of the ‘transparency’ of the stratosphere to vertical wave propagation, but in order to determine if waves propagate to the MLT region the index needs to be calculated through the lower atmosphere.

5. Summary and Conclusions

[62] In this paper, 16 years of hydroxyl (6–2) band rotational temperature measurements from Davis station, Antarctica have been analyzed for a solar cycle dependence and underlying long-term trend. Seven years of new data have been added since the last published trend analysis and a small correction has been made due to a calibration lamp orientation effect.

[63] Correlation analyses with various solar flux indices shows general agreement between them with a broad correlation maximum biased toward positive lag (solar index leads the temperature response). Further investigation showed that this bias was largely the result of short period variations in 2002. This year is unusual in that it is the warmest in the 16 year record, with the highest F10.7 activity and contained the unprecedented Southern Hemisphere stratospheric warming. We view the lag result with caution and require more data to resolve this matter.

[64] Multivariate regression analysis on seasonally detrended mean winter OH temperatures yields a solar cycle dependence of 4.8 ± 1.0 K/100 SFU and a long-term cooling trend of 1.2 ± 0.9 K/decade (R2 coefficient 0.70). These coefficients remain consistent around 4 K/100 SFU and −1 K/decade if the regression is performed with annual, monthly or nightly means. In general, the solar activity response is consistent with other observers and the long-term trend falls within the larger range of reported hydroxyl layer trends.

[65] A distinct seasonal variation in trend coefficients is found in the 30 day sliding window or monthly trend analyses. The largest solar activity response (∼7 K/100 SFU) is measured in March, May–June and September and little or no solar response in April and August. The long-term trend coefficient shows the largest cooling rate (4–5 K/decade) in August–September, through to warming (2–3 K/decade) for the March and May–June periods. A surprising similarity in the seasonal trend characteristics is noted when compared with the Wuppertal measurements if a 6 month offset is applied to match the Southern Hemisphere season. The April (SH Autumn) and August (SH Spring) cooling maxima and little apparent cooling over the winter months is consistent suggesting that the trends are modulated by seasonal effects rather than global change effects.

[66] Attempts have been made to characterize other sources of variability in the residual temperatures using indexes for planetary wave strength, zonal wind anomaly (PVI), the quasi-biennial oscillation and southern annular mode. No statistically significant correlations were found with the PVI, QBO or SAM indices over the entire data set; however, there was clear evidence of planetary wave modes identified in the 10 hPA NCEP reanalysis data penetrating to OH layer heights at different times in the series. Further work will focus on these events to identify the conditions for wave propagation.


[67] The authors thank the dedicated work of the Davis optical physicists over many years in the collection of airglow data and calibration of instruments. Helpful discussions with G. B. Burns are gratefully acknowledged. This work is supported by the Australian Antarctic Science Advisory Council.