Accuracy of NCEP/NCAR reanalyses and ECMWF analyses in the lower stratosphere over Antarctica in 2005



[1] This article compares the temperature, zonal, and meridional velocities issued by the 50-year National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP-NCAR) Reanalysis (NN50) and European Centre for Medium-range Weather Forecasts (ECMWF) operational analyses with independent observations collected during the Vorcore superpressure balloon campaign. The ECMWF analyses are found to be more accurate than the NN50 reanalyses. In particular, an overall warm bias in the polar Southern Hemisphere lower stratosphere is found in NN50 (+1.51 K), while a cold bias is found using ECMWF analyses (−0.42 K). The ECMWF temperature bias evolves from winter (−1.34 K in September) to summer (+0.31 K in January). A common feature of both analyses is the localization of larger discrepancies with respect to the observations for all variables over the Antarctic Peninsula, where orographic gravity waves were observed. More generally, two principal reasons are put forth for explaining the reported differences with the observations: unresolved small-scale pure and inertia gravity waves affecting both analyses and a large-scale misrepresentation of the vortex in NN50. Finally, simulated trajectories are computed starting from the positions of the balloon and advected using the ECMWF and the NN50 velocity fields. The simulated trajectories are compared to the real balloon trajectories. The spherical distance between the real and simulated positions exceeds 1000 ± 700 km on average in just 5 days using NN50 and after 10 days using ECMWF. The distances between the simulated and real balloons are found to increase faster in November and December, owing to the strong Rossby-wave activity in the stratosphere.

1. Introduction

[2] Meteorological model forecasts, analyses and reanalyses are often used as reference data sets to study atmospheric processes such as stratospheric ozone depletion or troposphere-stratosphere exchanges, among others. The analyzed temperature field is used in chemistry-transport models to assess several physical and chemical processes such as ozone destruction via the occurrence of polar stratospheric clouds. Furthermore, the wind velocities can be used to compute trajectories and back trajectories. For example, forward trajectories are computed in Match campaigns to assess the ozone loss of an air parcel [von der Gathen et al., 1995]. Conversely, backward trajectories are calculated to estimate the turbulent diffusion of an air parcel, and to retrace its history [Legras et al., 2003].

[3] These data sets integrate advanced data assimilation schemes to the model runs, incorporating satellite data and radiosonde profiles. However, there is a net hemispheric contrast in accuracy caused by the lack of monitoring stations available in the Southern Hemisphere (SH) in comparison to the Northern Hemisphere (NH) [Kistler et al., 2001]. In particular, poleward of 40°S, limited data are assimilated into the models [Parrondo et al., 2007] and few stratospheric observations are available [Gobiet et al., 2005] so that the reliability of analyses and reanalyses is best in the troposphere and in the lower stratosphere. Thus, it is very important to assess the accuracy of these analyses and reanalyses in the stratosphere.

[4] Studies such as those by Manney et al. [1996, 2003, 2005a, 2005b] have compared temperature analyses and reanalyses produced by different centers, in order to assess their accuracy and their reliability for polar studies. In particular, Manney et al. [2003] found that the occurrence of temperature minima and the volume of polar stratospheric clouds formed are very dependent on the analysis used. Strictly speaking, intercomparison studies only give relative information on the biases of the analyses. Yet, a consistent bias in 50-year National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP-NCAR) Reanalysis (hereinafter NN50) has been found by Manney et al. [2003, 2005a, 2005b] and suggests that NN50 reanalyses are not suitable for polar processing studies. Recently, Manney et al. [2005b] focused specifically on the 2002 Antarctic stratospheric warming, which has been intensively studied. The comparison between different analyses and reanalyses showed a strong warm bias in NN50 and an inconsistent vertical structure for European Centre for Medium-range Weather Forecasts (ECMWF) temperatures.

[5] Other studies have compared independent nonassimilated data such as balloon or satellite data with analyses. Knudsen et al. [1996] compared balloon data from the Polar Vortex Balloon Experiment (Povorbex), which took place in the winter NH from 1992 to 1995 with ECMWF analyses for temperature and velocities. They found a strong warm bias (+2.4 K) in the analyses. Gille et al. [1996] compared satellite measurements from the Cryogenic Limb Array Etalon Spectrometer (CLAES) with analyses from the National Meteorological Center (NMC) and U. K. Meteorological Office (UKMO) over all the globe in 1992 and 1993 and also found significant warm biases. Finally, Gobiet et al. [2007] compared temperatures issued from the CHAllenging Minisatellite Payload (CHAMP) radio occulation data from 2001 to present with a range of analyses and reanalyses, and found consistent warm biases of about +0.7 K over the SH high latitudes for the NN50 and an overall small warm bias (+0.5 K) for ECMWF analyses.

[6] In comparison with temperature, observations of stratospheric winds are scarce and therefore do not constrain the models effectively [Polavarapu and Shepherd, 2005]. Keil et al. [2001] compared wind balloon data over both hemispheres in the high latitudes with analyses from the UKMO and the Data Assimilation Office (DAO) and found positive biases in the NH (up to 5 m s−1) and negative biases in the SH (reaching 7 m s−1). Manney et al. [2005b] intercompared the winds for ECMWF analyses, NN50 reanalyses and ECMWF 40-year reanalysis (ERA-40) and found very good agreement up to 10 hPa although NN50 reanalyses exhibit weaker westerlies.

[7] A number of studies have also focused on the computation of trajectories using the analyzed wind fields. For example, Keil et al. [2001] compared trajectories using the analyzed wind fields or persistence techniques, and Knudsen et al. [1996] computed isentropic trajectories at different levels around the balloon level and found large differences between real and simulated trajectories.

[8] The present study aims at assessing the accuracy of ECMWF analyses and NN50 reanalyses in the SH high-latitude stratosphere by comparing them to in situ measurements. As mentioned above, several studies have consistently reported large biases in NN50 reanalyses, and this study confirms these results with an independent data set. The following work draws on observations from the Vorcore superpressure balloon (SPB) campaign over Antarctica and the surrounding Southern Ocean [Hertzog et al., 2007], which took place during the winter (September 2005) to summer (February 2006) SH transition. Figure 1 illustrates the dense spatial sampling of observations, and justifies the validation possibilities for latitudes south of 40°S. Overall, the campaign resulted in over 150,000 observations which can be compared to the four-daily analyses and reanalyses produced by ECMWF and NN50. Indeed, these observations, namely the zonal and meridional velocities and the temperature, were not assimilated by these centers, and provide an independent data set.

Figure 1.

Trajectories of the SPBs flown during the Vorcore campaign.

[9] A similar study was carried out over the NH, using superpressure balloons SPBs launched from Kiruna, Sweden (69°N, 21°E) in January and February 2002. This study found a slightly warm bias (+0.8 K) in the NN50 reanalyses, and a small cold bias (−0.3 K) in the ECMWF analyses [Hertzog et al., 2004].

[10] Section 2 outlines the data sets used. Starting with a global view, section 3 presents the results obtained when comparing the data sets, then explores the temporal evolution as well as the latitudinal and longitudinal structure of the biases and standard deviations computed. Section 4 sets out to explain these biases and standard deviations distinguishing two dominant effects: small-scale and large-scale dynamics. Section 5 focuses on trajectories computed using the NN50 and ECMWF fields, launched every day at the current position of each balloon. The spherical distance between the real and the simulated positions is computed and its evolution in time is studied, with a particular emphasis on the state of the vortex.

2. Data Sets

2.1. Observations

[11] The Vorcore SPB campaign aimed at documenting the dynamics of the lower stratosphere during the 2005 SH polar spring. It sampled the different states of the stratospheric polar vortex from its steady state centered over the pole in September, through October and November when it was highly disturbed and displaced by planetary scale wave activity, until December when it broke down. Twenty-seven SPBs were launched into the stratospheric polar vortex from the station of McMurdo (77°S, 166°E), in Antarctica between 5 September and 28 October. They flew for an average of 2 months, with the longest flight lasting for 109 days. Depending on their size (8.5 m and 10 m diameter), they flew at different pressure levels: around 70 hPa or 50 hPa, which corresponds to the lower stratosphere (17–19 km). Owing to their constant volume, the SPBs have the particularity of drifting on isopycnic surfaces. They are very good tracers of horizontal motions, and therefore exhibit quasi-Lagrangian features.

[12] The SPB flights were terminated if the power inside the gondola reached a lower bound threshold, or if the SPBs flew north of 40°S. This occurred mainly after mid-December when the vortex had broken down, and the SPBs were no longer confined inside the vortex.

[13] The campaign and scientific payload are described in detail by Hertzog et al. [2007]. Briefly, the SBPs carried two temperature sensors, a Global Positioning System (GPS) receiver and a pressure sensor. The data were uploaded onto the ARGOS system for transmission to ground. Measurements were made every 15 min, enabling the computation of the 15-min averaged zonal and meridional velocities. On the basis of calibration studies and comparisons between the two temperature sensors, the temperature is accurate to 0.25 K. Moreover, Vorcore temperatures have been compared to temperatures derived from CHAMP radio occultation and the difference between data sets does not exceed 0.5 K [McDonald and Hertzog, 2008]. According to the manufacturer and careful laboratory calibration, the pressure sensor accuracy is 10 Pa with a 1 Pa precision. The GPS is accurate to 10 m in the horizontal and 20 m in the vertical. This horizontal precision yields a σh ∼ 0.02 m s−1 accuracy in the zonal and meridional velocities.

2.2. Analyses

2.2.1. ECMWF Analyses

[14] At the time of the experiment, the ECMWF spectral model operated a triangular truncation of 511 waves in the horizontal (T511), which corresponds to ∼40 km grid spacing in latitude and integrated a 4D variational data assimilation scheme [Rabier et al., 2000; Simmons et al., 2005]. This version (cyc29r2) included 60 model levels in the vertical, extending from the surface to 0.1 hPa (about 70 km) [Untch and Simmons, 1999]. The four-daily analyses were retrieved on a 0.5° × 0.5° grid over all the SH. The gravity wave drag parameterization combines a subgrid-scale orographic wave drag [Lott and Miller, 1997] and a Rayleigh friction scheme representing the deceleration induced by nonorographic waves. There are known problems to this method as it is linearly proportional to the strength of the winds and therefore acts strongly in regions where the winds are strong. For example, it results in a cold polar bias in the winter and strong westerlies in the midlatitudes [Hamilton, 1995; Hamilton et al., 1999].

2.2.2. NN50 Reanalyses

[15] The reanalysis system includes the NCEP global spectral model [Kanamitsu, 1989], with 28 sigma vertical levels and a horizontal triangular truncation of 62 waves, equivalent to a 210 km grid. The data assimilation scheme is a three-dimensional variational (3D-VAR) scheme cast in spectral space [Kalnay et al., 1996; Kistler et al., 2001]. The NN50 reanalyses are available four-daily with a 2.5° × 2.5° grid resolution and on 17 pressure levels ranging up to 10 hPa. An orographic gravity wave drag is implemented on the basis of work by Palmer et al. [1986]. The drag is produced by momentum deposition of small-scale waves, forced by the local orography, at critical levels or at the top of the model.

2.3. Interpolations

[16] To compare the data sets, the analyzed temperature, zonal velocity, and meridional velocity were interpolated on the balloon positions. A cubic spline, which extends over six temporal states and eight points in each direction in the horizontal, was used. The logarithm of pressure was used as the vertical coordinate, since the pressure in the balloon data set is more accurate than the GPS altitude. The vertical interpolation also used a cubic spline, extending over seven pressure levels for NN50 from 150 hPa to 10 hPa and over 11 model levels for ECMWF from about 132 hPa to 19 hPa. The levels used for this study correspond to pressure levels: 132, 113, 96, 80, 67, 55, 44, 36, 29, 23, 19 hPa in ECMWF, and 150, 100, 70, 50, 30, 20, 10 hPa in NN50.

[17] As an example, Figure 2 shows the zonal and meridional velocities, and the temperature interpolated with the ECMWF analyses and the NN50 reanalyses superposed onto the SPB observations during one flight (SPB7). This SPB flew from 17 September until 18 December. A good agreement is found between data sets, even though the time series exhibit large amplitude disturbances. The close-up of the observed time series illustrates the importance of short-scale disturbances in the observations, which are missed by the analyses partly because of the coarse temporal interpolation. The analyzed zonal and meridional velocities are very close to the observations while clear biases can be identified for the temperature. The NN50 reanalyses are warmer while the ECMWF analyses are slightly colder. This general result will be further detailed in the next section.

Figure 2.

SPB7 balloon data (black), ECMWF interpolated data (blue), and NN50 interpolated data (red), for zonal velocity, meridional velocity, and temperature.

3. Comparison Between Data Sets

3.1. General Results

[18] The differences between the interpolated analyses and reanalyses and the SPB data were computed for the temperature, zonal and meridional velocities. Figure 3 shows these results as histograms of the differences. The bins used to plot the histograms were respectively 0.25 K and 0.5 m s−1 for the temperature and velocity. NN50 reanalyses exhibit an overall warm bias (+1.51 K) while ECMWF analyses show a cold bias (−0.42 K), which may be caused by an insufficient parameterization of the gravity wave drag. Indeed, the Rayleigh friction used to mimic the gravity wave drag does not correctly represent the physics involved, such as the deposition of momentum flux at different levels, nor the anisotropy and intermittency of gravity wave sources. This issue will be further examined in section 4. Both analyses represent the zonal and meridional velocities very well. Nevertheless, the ECMWF distribution is slightly narrower than that for NN50 (as checked with the χ2 test), indicating a better accuracy of ECMWF winds. The large spacing between levels especially in the stratosphere and the relatively low top level of the NCEP model explains some of the deficiencies of NN50. Our results, compared to previous studies, highlight the fact that the observation system has been improved in the SH [Gille et al., 1996]. Similarly, Manney et al. [2003, 2005a] have found improvement in the representation of stratospheric Arctic winters between 1995/1996 and 2003/2004, in ECMWF analyses. Table 1 summarizes the main biases, standard deviations and higher-order moments. According to calibration tests in the laboratory, it is assumed that the observations are unbiased; see section 2.1. Hence, the mean of these differences represent solely the analyses bias. On the other hand, the standard deviations of the differences sum the systematic instrumental errors and the analyses uncertainties,

equation image

where σObs = 0.25 K, for temperature and σObs = 0.1 m s−1 for zonal and meridional velocities (σObs > σh ∼ 0.02 m s−1 to increase the confidence levels). Since the values of these systematic errors σObs are known, equation (1) can be inverted, so that the standard deviations presented in Table 1 only correspond to the analyses uncertainties.

Figure 3.

From left to right, histograms of the differences between the interpolated analyses data and the SPB measurements for temperature, zonal velocity, and meridional velocity; for ECMWF (solid line) and NN50 (dashed line).

Table 1. Statistics of ECMWF/NN50 Analyses Minus SPB Observation
 Temperature (K)Zonal Velocity (m/s)Meridional Velocity (m/s)
  • a

    Not statistically significant.

Standard deviation1.241.752.433.412.383.13
Excess kurtosis0.811.482.151.320.751.57

[19] The biases for the zonal and meridional velocities are very small, still most of them are significant to 95%. The standard deviations are larger than 2 m s−1 in ECMWF analyses, and 3 m s−1 in NN50 reanalyses.

[20] A number of studies using satellite data and radiosondes have shown that the analyses biases depend on the altitude in the stratosphere. For instance, using radiosonde temperature profiles from Belgrano (78°S), Parrondo et al. [2007] report on an oscillatory vertical structure of the temperature field in the ECMWF analyses during the 2003 winter. This structure features negative biases at levels ranging from 40 hPa to 15 hPa and from 150 hPa to 65 hPa and positive biases at levels between 65 hPa and 40 hPa. Their results also show that the NN50 reanalyses exhibit a warm bias throughout ranging from 0.5°C to 3°C. Gobiet et al. [2005, 2007] also report on an oscillatory structure of the ECMWF temperature field in the SH winter. Their studies compared radio occultation data from the CHAMP mission with ECMWF analyses since 2001. The vertical structure described is consistent with the results from Parrondo et al. [2007]. We now investigate the dependence of the bias with altitude. Indeed, the Vorcore SPBs flew at two different altitude levels depending on their size: 17 km (70 hPa) and 19 km (50 hPa). The statistics were computed for the two levels and the results are summarized in Table 2. The biases and standard deviations are larger at higher altitudes for both analyses: for ECMWF, the bias is colder at higher altitudes, while for NN50, the bias is warmer. For NN50, our results are in good agreement with those from Parrondo et al. [2007]. However, they disagree for ECMWF as the bias evolves from a cold to warm bias at similar pressure levels. These differences, may be due to the different versions of the ECMWF analyses used: cyc26r1 and cyc26r3 for Parrondo et al. [2007] and cyc24r3 through cyc28r2 for Gobiet et al. [2005] while the version cyc29r2 was used in our study.

Table 2. Statistics of ECMWF/NN50 Analyses Minus Balloon Observation for 8.5-m and 10-m Diameter SPBs
 Temperature (K)Zonal Velocity (m/s)Meridional Velocity (m/s)
  • a

    Not statistically significant.

Bias: 8.5 m (≈17 km)−0.171.340.050.06−0.01a0.05
Bias: 10 m (≈19 km)−0.481.550.18−0.210.01a−0.15
Standard deviation: 8.5 m1.091.542.253.412.222.91
Standard deviation: 10 m1.271.792.53.412.453.21

3.2. Temporal Evolution

[21] The previous section reports on results over the whole period which extended between September 2005 and January 2006 and over the whole geographical domain covered by the SPBs. In this section, the temporal evolution of the zonal-mean biases and standard deviations is described. To this end, the statistics were computed for each month, in 5°-latitude bins. Figure 4 displays the number of observations for each month over each latitude interval and shows that there are over 200 observations for every bin but one, which supports the statistical significance of the results computed. The geographical distribution of the SPBs is also illustrated in Figure 4. In September, very few observations are recorded north of 55°S, as the SPBs are confined inside the vortex, which is centered over the pole. In October and November, the observations spread to the lower latitudes. At this time, the strong activity of planetary Rossby waves displaces and distorts the vortex. In December, this strong wave activity leads to the breakdown of the vortex, so that the SPBs are no longer confined and free to drift down to the midlatitudes.

Figure 4.

Number of 15-min observations per latitudinal bin from September to January, in logarithmic scale.

[22] Figure 5 shows the zonal-mean ECMWF (top left) and NN50 (top right) temperature biases, as a function of latitude. The corresponding standard deviations are shown in the bottom panels. Figure 6 shows the same for zonal velocity.

Figure 5.

(top) Mean temperature differences for (left) ECMWF and (right) NN50. (bottom) Standard deviation of the temperature differences for (left) ECMWF and (right) NN50. Both are plotted according to month, from September to January.

Figure 6.

(top) Mean zonal velocity differences for (left) ECMWF and (right) NN50. (bottom) Standard deviation of the zonal velocity differences for (left) ECMWF and (right) NN50. Both are plotted according to month, from September to January.

[23] One interesting feature in Figure 5 is the adjustment of the ECMWF analyses from a cold bias in September (−1.34 K) to a slightly warm bias in January (+0.31 K) independent of latitude.

[24] These results agree with comparisons between ECMWF analyses, NN50 reanalyses and radio occultation data from the CHAMP mission between 2001 and 2004 from June to September, as reported by Nedoluha et al. [2007]. In particular, the absence of latitudinal structure in the ECMWF analysis is consistent with their results.

[25] For NN50, the biases are warm for all months, with a relatively smaller bias between 60°S and 70°S. The corresponding standard deviations are larger than in ECMWF analyses. However in both analyses, the biases and standard deviations are smaller in January. Indeed, after the breakdown of the vortex, the easterly summer circulation sets in so that the stratospheric variability and the velocities are smaller.

[26] The winter cold bias and the summer warm bias are both consistent with a deficit in gravity wave drag in the ECMWF model. However, the consistently warm bias in NN50 reanalyses cannot be explained through a lacking gravity wave scheme, see section 2.2.

[27] Throughout the campaign, the ECMWF zonal velocity biases are very small. Furthermore, the ECMWF biases and standard deviations do not depend on latitude. On the other hand, the latitudinal structure of the NN50 reanalyses exhibits more erratic variations for both means and standard deviations with biases ranging from −2 m s−1 to +5 m s−1. For example in November, the structure in latitude means that the winds in NN50 are too westerly equatorward of 70°S and too easterly poleward of 70°S. Similar features are evident for the rest of the period studied. The meridional velocity statistics are not shown but essentially exhibit the same features as the zonal velocity: small biases and standard deviations for ECMWF, and a strong heterogeneity for both the bias and the standard deviation for NN50.

3.3. Longitudinal Structure

[28] In the previous section, it was shown that the NN50 temperature biases and standard deviations were randomly distributed in latitude, in particular in November. Further insight into the geographical distribution is given here.

[29] The biases and standard deviations are averaged over intervals of 5° in latitude and 20° in longitude. A single box was used for latitudes poleward of 85°S. We used a minimum of 50 observations per box to compute the biases and standard deviations.

[30] Figure 7 shows maps of the temperature biases and standard deviations in November. At that time, the temperature and velocity fields are very disturbed, because of the influence of strong planetary Rossby waves as shown by Hertzog et al. [2007, Figure 11]. In Figures 5 and 6, this corresponds to larger biases and/or standard deviations in November for NN50. Figure 7 highlights the coexistence of strong positive and negative biases with a disperse geographical distribution in NN50 reanalyses. A likely reason for this behavior is a misrepresentation of the large-scale structures: the stratospheric vortex and in particular its edge region where gradients are strongest, and therefore large departures can be found. This will be further discussed in section 4.1. It can also be noted that large biases are found at the most equatorward latitudes. As these latitudes correspond to the storm tracks, a possible explanation is that tropospheric disturbances may be misrepresented by NN50 reanalyses.

Figure 7.

Temperature statistics in November, in boxes of 20° in longitude and 5° in latitude. The two left panels show mean differences for ECMWF and NN50, respectively, and the two right panels show standard deviations for ECMWF and NN50, respectively. Boxes including less than 50 observations are dashed.

[31] For ECMWF, Figure 7 confirms the overall homogeneous structure of the biases and standard deviations. This supports the hypothesis that large-scale structures are well resolved by the ECMWF analyses.

[32] The standard deviation distribution is qualitatively similar to those of the biases (homogeneous for ECMWF, heterogeneous for NN50). Nevertheless, for both analyses, the standard deviations are larger over the Antarctic Peninsula where large amplitude gravity waves were observed during the balloon campaign, [Hertzog et al., 2007]. These events may cause strong temperature disturbances, but are generally too localized to be properly resolved in the models.

4. Discussion

[33] The previous section reviewed the biases and higher-order moments found when comparing the observational data set with the ECMWF analyses and NN50 reanalyses. These errors can be distinguished as being caused by two shortcomings in the models. The first one is due to unresolved small-scale phenomena. This includes disturbances caused by gravity waves, as demonstrated by Vincent et al. [2007], and inertial oscillations. For instance in Figure 1, the cycloidal aspect of the SPB trajectories, which can be observed South of Australia, is typical of such oscillations. The second effect is due to large-scale phenomena, and includes errors due to a misrepresentation of the vortex both horizontally and vertically and the extent and position of the vortex edge.

4.1. Small-Scale Dynamics

[34] In order to assess the importance of these small-scale effects, the entirety of the balloon data and the interpolated analyses data were filtered with a low-pass filter with a 15-hour cut-off period which corresponds to the longest inertial period in the balloon data set. Figure 8 shows the zonal velocity recorded by SPB7 over a few days, and the corresponding filtered data. This filtering removes all oscillations caused by pure and inertia-gravity waves, so that the solid line in Figure 8 exhibits oscillations only due to large-scale dynamics.

Figure 8.

Raw (dashed line) and filtered (solid line) zonal velocity for SPB7.

[35] Figure 9 shows the distribution of the differences between the filtered data sets for temperature, zonal velocity and meridional velocity. The small-scale disturbances mainly correspond to the wave oscillations so that their effect should vanish when averaged. Hence, the biases in the filtered time series are not significantly changed, for both the overall bias and the monthly biases (not shown). On the other hand, the standard deviations are greatly reduced. This is particularly true for the zonal and meridional velocities, as the quasi-inertial waves have a strong signature on the velocity fields, and are ubiquitous in the balloon observations [Hertzog et al., 2002]. The distribution of the velocity differences are much more peaked and narrow, in particular for ECMWF, when compared to the histograms in Figure 3. For temperature, the effect is less striking.

Figure 9.

From left to right, histograms of the differences between the filtered interpolated analyses data and the filtered SPB measurements, for temperature, zonal velocity, and meridional velocity; using ECMWF (solid line) and NN50 (dashed line).

[36] The difference between standard deviations with or without filtering gives an estimate of the contribution of inertia-gravity waves to the overall analyses standard deviations,

equation image

where σLARGE SCALE is estimated from the filtered time series, σANALYSES is the overall standard deviation computed in section 3.1 and σGW+INERTIAL corresponds to the standard deviation induced by inertia-gravity waves apparent in the balloon observations. Table 3 shows the standard deviations associated with inertia-gravity waves for temperature and velocities, for both analyses. The standard deviations associated with small-scale motions have very similar amplitudes for both analyses, suggesting that they both misrepresent the small scales. Furthermore, the standard deviations are much larger for velocities than for temperature. This suggests that quasi-inertial gravity waves make up a substantial part of these standard deviations, as these waves mainly induce disturbances in the horizontal velocities.

Table 3. Standard Deviations due to Small-Scale and Large-Scale Effects and Total
 Temperature (K)Zonal Velocity (m/s)Meridional Velocity (m/s)
σLARGE SCALE1.061.551.042.580.982.22

4.2. Large-Scale Dynamics

[37] Since the large-scale structure of the stratosphere is dominated by the vortex at high latitudes in winter, we study in this section the analyses bias and standard deviation distributions with respect to equivalent latitudes [Butchart and Remsberg, 1986], that is the position of the SPB with respect to the vortex. Namely, for each observation, the differences between the balloon equivalent latitude and that of the vortex edge (which is defined as the location of the maximum PV gradient) were calculated. The ECMWF fields were used to compute these two parameters, since the previous sections show a better agreement of the ECMWF analyses with the observations. The computation of equivalent latitude was stopped on 14 December, after the vortex has broken down into several pieces. The difference between the balloon observations and the analyzed fields were computed over 5° latitude bins, for each month, using once again a threshold of 50 observations per bin. Figure 10 shows the biases and standard deviations for temperature.

Figure 10.

(top) Mean temperature differences for (left) ECMWF and (right) NN50 versus equivalent latitude. (bottom) Standard deviation temperature differences for (left) ECMWF and (right) NN50 versus equivalent latitude. All are plotted according to month, from September to December.

[38] Inside the vortex, for ECMWF, the distribution is similar to the one plotted for zonal means (Figure 5), showing an adjustment from colder temperatures in September to warmer ones in December, inside the polar vortex. The mean and standard deviations are almost independent of the position of the SPB with respect to the vortex, which demonstrates that the ECMWF analyses accurately represent the vortex structure and its variability.

[39] On the other hand, the results for NN50 show an increase of the biases and standard deviations at the edge of the vortex, in particular in October and November, supporting a misrepresentation of the edge of the vortex by NN50 reanalyses.

[40] Furthermore, Table 3 shows the standard deviations due to large-scale dynamics. The standard deviations in the NN50 reanalyses are larger for all variables in particular for the wind fields, confirming that the NN50 reanalyses demonstrate shortcomings concerning the representation of large-scale structures in the stratosphere. These results are consistent with those from previous studies [Manney et al., 2005b] and confirm the deficiencies of NN50 reanalyses.

[41] To further illustrate this point, we now consider the vertical structure of the vortex as represented by the ECMWF analyses and NN50 reanalyses near the drifting levels of the SPBs. Figure 11 shows the modified potential vorticity (MPV) [Lait, 1994] on isentropic surfaces situated in the lower stratosphere from 15 November to 15 December. The top and bottom panels of Figure 11 are directly issued from ECMWF analyses at isentropes 430 K and 475 K while the middle panels correspond to an intermediate isentrope (450 K) in the NN50 reanalyses. Given that typical values of the vertical gradient of potential temperature in the lower stratosphere (i.e., ∼20 K/km), the distance between the successive isentropic levels shown is on the order 1 km.

Figure 11.

Vertical structure of the vortex (depicted by the MPV fields) as a function of time: columns correspond to the dates 15 November, 1 December, and 15 December. (top and bottom) Two isentropic levels from ECMWF (475 K and 430 K), and (middle) an isentropic level in NN50 (450 K). The color scale is the same for all graphs.

[42] On 15 November the ECMWF defines a sharp vortex edge on both the bottom and top levels, with a large gradient separating the air inside from the air outside the vortex. On the other hand, the NN50 reanalyses shows a more blurred and therefore thicker edge at the intermediate level. Secondly, on 1 December the vortex is breaking down, and the fact that the dissolution is initiated from higher altitudes is evident from the comparison of the two ECMWF maps at 475 K and 430 K. Again, the MPV gradients at the vortex edge are more blurred in the NN50 reanalyses than in both levels in ECMWF analyses. Manney et al. [2005b] found similar results at higher altitudes.

[43] These results agree with the equivalent latitude distributions, and plead for some deficiencies in the representation of the vortex in the NN50 reanalyses. This different behavior between analyses is also very likely due to the different resolutions (0.5° × 0.5° for ECMWF and 2.5° × 2.5° for NN50), and thus the ability of ECMWF analyses to resolve the large gradients associated with the vortex edge.

5. Trajectories

5.1. Method

[44] Trajectories and back trajectories are powerful tools to study the transport properties of the troposphere and the stratosphere. For example, many studies have focused on the intrusions of stratospheric air into the troposphere and vice versa [Cristofanelli et al., 2006]. Nevertheless, the accuracy of analyses and reanalyses limits the validity of such studies. The errors in trajectory calculations have been reviewed by Stohl [1998]. The assessment of the total error of trajectories combines a range of errors and is difficult to estimate. Thus, these trajectories last a maximum of 10 days for the Match technique [Morris et al., 2005], and 9 days for Legras et al. [2003] trusting that the cumulative errors due to the biases from the analysis will be limited.

[45] We tested this by computing forward trajectories lasting up to 60 days. These simulated trajectories were launched every day at midnight from the current position of the SPBs. The time integration used a fourth-order Runge Kutta method, with a time step of 10 min. This time step was reduced near the pole or when the wind velocity became large. The particularity of the SPBs was that they drifted on isopycnic surfaces, so we computed an average value of the flight density and used this reference density for interpolation. If the simulated trajectory reached latitudes equatorward of 40°S, the trajectory was automatically stopped, as for the SPBs launched during the experimental campaign. We computed the spherical distance between SPB and simulated trajectory, and studied its evolution with time, using both NN50 reanalyses and ECMWF analyses.

5.2. Results and Discussion

[46] Figure 12 presents the mean spherical distance between the SPB position and the simulated position as a function of the simulated flight duration for all simulated flights. The envelopes correspond to the spherical distance of the 25% (and 50%) closest points to the mean spherical distance. The 50% envelope is seemingly asymmetric (this is also true for the 25% envelope albeit less obvious): the distance between the top branch of the 50% envelope and the mean is larger than the distance between the bottom branch and the mean. Indeed the distribution of spherical distance is skewed toward larger distances. For both analyses, we observe a fast growing regime during which the distance increases very rapidly for up to 15 days, and more slowly afterward. After 5 days, the spherical distance between real and simulated positions is larger than 400 km for both analyses. The maximum distances are larger using the NN50 reanalyses than using the ECMWF analyses. The distance reaches 1000 ± 800 km after 5 days and 3000 ± 1400 km after 30 days using NN50 fields. Using ECMWF analyses, the distance reaches 1000 ± 1000 km after 10 days, and 3000 ± 1700 km after about 50 days. So, this result agrees with the previous sections, which concluded that the ECMWF analyses, and especially the wind field, give a better representation of the stratosphere than the NN50 reanalyses. For longer simulations, the distance saturates because of the finite size of the vortex in which the SPBs are trapped, during the winter circulation. In November and December, this upper bound distance increases as the vortex is stretched by the activity of Rossby waves.

Figure 12.

Mean spherical distance (bold line) for (left) ECMWF and (right) NN50 as a function of time since simulated launch. The dotted envelope includes 25% of the closest data points to the mean. The solid envelope includes 50% of the closest data points to the mean.

[47] Furthermore, the sensitivity of the Match technique to the type of meteorological analysis used was studied by Hoppel et al. [2005]. They compared the ozone loss rates produced using UKMO, NN50 and ECMWF fields and found that the ozone loss rates computed using ECMWF fields were closest to those observed, in agreement with our results.

[48] Similar results were obtained when simulated trajectories were computed from the positions of SPBs which flew in January and February 2002 in the NH [Hertzog et al., 2004]. Two growth regimes were observed, and the spherical distances found were smaller for ECMWF (270 ± 230 km after 5 days) and comparable for NN50 (1700 ± 1400 km after 5 days). This highlights the better representation of the NH in the ECMWF analyses.

[49] Secondly, the trajectories were separated according to the month during which they were launched. Figure 13 exhibits the two regimes described above for each month, with a switch from one to the other occurring roughly around 15 days as in Figure 12. Two-day averages were performed for smoothing purposes. In agreement with results from section 3.2, the distances are smaller and grow at a slower rate in January. For both analyses, the distances are larger in November and December which are the months when the vortex is highly distorted and displaced owing to the strong Rossby wave activity. The associated large atmospheric variability, as described by Hertzog et al. [2007, Figure 11], is likely to be responsible for the deterioration of the statistics during these months.

Figure 13.

Mean spherical distance (averaged over 2 days) for (left) ECMWF and (right) NN50, by month of launch as a function of time since simulated launch.

6. Summary

[50] Observations from the Vorcore superpressure balloon campaign were used to evaluate the accuracy of ECMWF analyses and NN50 reanalyses in the lower stratosphere during the 2005 SH spring. ECMWF analyses were found to agree closely with the observations with virtually no bias on the zonal and meridional velocities and a small cold bias (−0.42 K) on the temperature. The velocities from the NN50 reanalyses are also very close to the balloon observations although they exhibit larger dispersion. Overall, the NN50 reanalyses displayed a strong warm bias (+1.51 K).

[51] For ECMWF, the evolution in time of the temperature biases showed an adjustment from September (−1.34 K) to January (+0.31 K), independent of latitude. On the contrary, in the NN50 reanalyses, a randomly distributed geographical structure of the bias is observed throughout the campaign with strong positive and negative biases. For both analyses, an increased variability is found over orographic regions such as the Antarctic Peninsula where large-amplitude gravity waves have been observed.

[52] It is inferred that two phenomena produce the main discrepancies: the first is caused by the insufficient resolution of pure and inertia-gravity waves in the analyses while the second one is caused by misrepresentations of large-scale structures in the NN50 reanalyses. The importance of small-scale dynamics in the time series is showed by filtering the balloon and interpolated data sets with a low-pass filter and reevaluating the biases and standard deviations. Both analyses seem to be affected in a similar way by these shortcomings, although this may be partly due to the coarse temporal interpolation. Indeed, the analyses and reanalyses are output every 6 hours only.

[53] To evaluate the misrepresentation of the vortex by the NN50 reanalyses, the temperature biases were estimated as a function of the position of the SPB with respect to the edge of the vortex using the equivalent latitude. Maps of modified potential vorticity also imply a blurred representation of the vortex, also possibly caused by the resolution of the NN50 reanalyses (2.5° × 2.5°) while the ECMWF analyses are output on a 0.5° × 0.5° grid.

[54] Secondly, we compared the trajectories of the SPBs with isopycnic trajectories computed using ECMWF and NN50 velocity fields. The spherical distance between the real and the simulated positions was computed as a function of the flight time. The ECMWF wind fields produce more accurate trajectories than the NN50, with spherical distance reaching 1000 ± 700 km for NN50 and 400 ± 400 km for ECMWF after 5 days. The distances are largest in November and December, corresponding to the largest part of the Rossby wave activity.

[55] This work confirms the fact that NN50 reanalyses are not suitable for polar studies, as previously demonstrated by Manney et al. [2003, 2005a, 2005b]. This study complements Hertzog et al. [2004] in providing a large number of in situ observations to assess the accuracy of analyses and reanalyses in the stratosphere. A new SPB campaign is planned, releasing 12-m-diameter superpressure balloons from McMurdo, Antarctica. Therefore, a similar study will be undertaken to assess any improvements in the model or the assimilating system of these meteorological centers.


[56] The authors would like to thank two anonymous reviewers for insightful comments. The authors would like to acknowledge the French Space Agency (CNES) as well as the NSF and the French Polar Institute (IPEV) for their longstanding support of the Vorcore campaign. The Laboratoire de Météorologie Dynamique is a member of the Institut Pierre Simon Laplace.