The evolution of total ozone content (TOC) in Antarctica during austral spring is strongly linked to the important stratospheric ozone decline that was highlighted for the first time by Chubachi et al. (1985) and Farman et al. (1985). Nowadays the photochemical and microphysical processes leading to the massive and seasonal destruction of ozone in polar regions are well understood. The latest Ozone Assessment Reports (WMO, 2007, 2011, 2014) have confirmed the stabilization of ozone loss in Antarctica since 2000. The challenge now is to assess the impact of the observed reduction in the concentration of ozone-depleting substances (ODSs) (evaluated in the polar regions to ∼ 10 % in 2013 from the peak values in 2000; WMO, 2014) on the amplitude of the ozone destruction every year. During the last decade, several studies have been carried out to quantify a possible increase in total ozone column in the Antarctic polar vortex in spring directly linked to this decrease in the polar stratosphere. Most analyses use multilinear regression (MLR) models with different proxies to represent the interannual variability of ozone as a function of the 11-year solar cycle, the quasi-biennial oscillation (QBO), volcanic aerosols (Aer) or eddy heat flux (HF) (Salby et al., 2012; Kuttippurath et al., 2013; de Laat et al., 2015). These studies generally show a significant increase of TOC since 2000 for the September–November averaged period but they differ on the proxies used for the quantification of ozone interannual variability. De Laat et al. (2015) used a “big data” ensemble approach to calculate trends. Several scenarios were considered for the period over which the ozone record is calculated and for the different proxy records. They found that the significance of trends could vary from negligible to 100 % significant at 2σ levels depending on the scenario considered. They have also determined the optimal proxy records and ozone record scenarios to obtain the best regression. The limitation of MLR analysis is that only formal statistical error of trend is estimated and structural uncertainties linked to the single and arbitrary combination of proxies is not taken into account. De Laat et al. (2017) inferred trend values from daily ozone mass deficit (OMD) computed from a multi-sensor reanalysis (MSR) dataset without using any model but filtering the anomalous years with low polar stratospheric cloud (PSC) volume. The authors found positive and highly significant trend of OMD since 2000.

Solomon et al. (2016) evaluated trends in total ozone and ozone profiles records as well as healing characteristics by combining measurements (satellites and ozonesondes) and the Specified Dynamics version of the Whole Atmosphere Community Climate Model (SD-WACCM). They found a significant healing in September but not in October, during which ozone depletion is largest in the first 2 weeks. They also explain the difficulty of estimating the trend in October because the large variability of ozone linked to temperature variations and transport. The baroclinicity of the polar vortex in October and its displacement from the geographic pole can also contribute to the variability of the total ozone series averaged during the month of October.

The direct link observed between positive trends of total ozone within the polar vortex and the reduction of ODSs remains open to debate, given the natural variability of the Antarctic vortex and the possible contribution of greenhouse gases to the trends (Chipperfield et al., 2017).

The purpose of this paper is to provide an update of the ozone evolution inside the Antarctic vortex during the last decades, taking into account the vortex baroclinicity. The main aim is to determine the different contributions to ozone interannual variability and to estimate the post-2001 total ozone trend and related significance for different periods: September, which corresponds to the period of fastest development of catalytic photochemical ozone destruction, and mid-September to mid-October when the maximum ozone loss is reached.

This paper is organized as follows. Ozone datasets from satellites and MSR are presented in Sect. 2 and the description of the method used for total ozone column classification inside the vortex in Sect. 3. The influence of vortex baroclinicity on total ozone column inside the vortex is assessed in Sect. 4 by using a new classification method compared to standard ones based on a single isentropic level. Ozone trends before and after the turnaround year calculated using a multi-regression model for September and mid-September to mid-October are presented and discussed in Sect. 5. Results of trends using OMD records as a metric are also presented. The temporal evolution of the amount of very low total ozone values inside the vortex is evaluated in Sect. 6. Conclusions are finally presented in Sect. 7.

Total ozone global fields from satellite observations – Total Ozone Mapping Spectrometer (TOMS) and Ozone Monitoring Instrument (OMI) – and MSR are used in this study to cross-check trend estimation before and after a turnaround year over the 1980–2017 period.

In order to consider total ozone columns only within the polar vortex, the data classification is performed by evaluating the vortex's position at different isentropic levels from 1 May to 31 December, each year. Two classification methods are then applied in order to evaluate the impact of baroclinicity of the vortex on the averaged total ozone columns in both studied depletion periods. The first one is based on a single isentropic level, while the second one considers a range of isentropic levels.

Both methods of classification described in the previous section were applied to satellite composite total ozone data series SAT and MSR-2 reanalysis at each grid point. For each year, daily mean total ozone amount inside the vortex was averaged over two periods: the whole month of September and the period of maximum ozone depletion between 15 September and 15 October. Figure 2 shows the evolution of total ozone average inside the vortex for the 15 September–15 October period between 1980 and 2017 for the MSR-2 data series computed with the standard classification method based on the single isentropic level (475 and 550 K) and with the second method using the 400–600 K range of isentropic levels. Error bars represent the 2σ standard error. Similar interannual total ozone variability is observed for the time series obtained by the different methods. The correlation coefficients between the range method and the standard one at 475 and 550 K are 0.98 and 0.99, respectively. Despite these good correlations, the data series are significantly different at the 2σ level. Larger ozone values are found with the standard method, especially for the 475 K level, which shows a mean difference with the TOC time series based on the range method of ∼ 15 % over the whole analysis period. Three years stand out in the comparison – 1995, 1999 and 2011 – during which the inside vortex region was systematically larger at 475 K compared to higher isentropic levels during the period. Similar results are observed for September (not shown). In this work, the second method is preferred since it takes into account the ozone loss at different isentropic levels, which strongly impacts the total column.

MSR-2 total ozone time series obtained in September and 15 September–15 October with the range classification method are displayed in Fig. 3. September presents ∼ 8 % larger ozone mean values than the 15 September–15 October period. Similar interannual variability is observed between the two periods as shown by the correlation coefficient of 0.98. The last 4 years present very similar ozone values of around 205 DU in September, while during the 15 September–15 October period they show larger variability.

MSR-2 total ozone data series inside the vortex are compared to SAT series as shown in Fig. 4, which displays the relative difference between MSR-2 and SAT for the 400–600 K range classification. Differences of about ±0.5 % are observed in the 1980s. Small differences are expected during this period since only TOMS data are used in both datasets until 1993. In the 1993–1995 period discrepancies between both curves are only due to the differences in the selection of MSR-2 data for the SAT record in order to have similar spatial coverage as the data from the other instruments incorporated in the SAT time series. These differences varying between -1 and 0.5 % represent an estimation of the impact of reduced spatial coverage in SAT dataset on the averaged total ozone level in September. The 15 September–15 October period presents negligible differences. The addition of GOME (1996–2005) in MSR-2 assimilation could explain the discrepancies with the SAT dataset that considers only TOMS-EP. From 2001, differences are larger and generally positive, reaching ∼ 5 % in September and ∼ 3 % in the 15 September–15 October period. These increased differences are especially visible during the period where data from instruments on board the ENVISAT platform (e.g. SCIAMACHY) are assimilated in the MSR-2 record. Overall, values in September present a mean bias of 1.3 % (dash blue line in Fig. 4) and during 15 September–15 October a smaller bias value of 0.5 % (dash red line in Fig. 4). Temporal evolution of the differences, e.g. negative trend in the 1980s and positive trend in the 2000s, can have an impact on the long-term ozone trends retrieved from both records. In general, differences between SAT and MSR-2 records are caused by MSR-2 starting to use multiple satellite total ozone columns records after 1996, procedures in MSR-2 to account for inter-instrument differences and data assimilation methodology that allows for filling gaps (van der A et al., 2015).

Despite the differences between those datasets, one purpose of this work was to analyse, in the same way, satellite data such as those included in the SAT record without any correction or adjustment and the MSR-2 record, which accounts for inter-instrument differences using ground-based total column data. Due to the larger differences observed between both datasets in September, especially in the 1995–2010 period, which may have an impact on trend analysis, it was decided to retrieve trends from the SAT dataset in the 15 September–15 October only.

In the next section, ozone data series based on the different classification methods are used to evaluate the impact of vortex baroclinicity on ozone trends inside the vortex for both studied periods.

The ozone hole is generally defined as the region with total ozone columns lower than 220 DU. This standard value was used in different studies to evaluate the ozone depletion from the OHA (e.g. Newman et al., 2006; Solomon et al., 2016) or the OMD (e.g. de Laat and van Weele, 2011; de Laat et al., 2017) metrics. In order to evaluate how the ozone hole is influenced by very low ozone values, the surface relative to the vortex area occupied by ozone values lower than different threshold levels is computed for each day and averaged over different periods (September, 15 September–15 October and October). The top panel of Fig. 12 shows the evolution of these average relative areas for five different thresholds: 220, 200, 175, 150 and 125 DU for the 15 September–15 October period. MSR-2 datasets are used for this analysis and vortex areas are estimated using the 400–600 K range classification, the results of previous sections having shown that the range classification better constrains the OHA. Results show increasing areas during the 1980s, a stabilization in the 1990s and a larger interannual variability since 2001. In contrast to the 220 DU threshold case, the evolution of relative areas corresponding to lower thresholds shows a delayed increase from the beginning of the 1980s to the early 1990s, reaching a maximum in all cases in 2000. After 2000, a larger interannual variability is generally observed and from 2006 a steady decrease is seen for thresholds lower than 200 DU. In all cases, several anomalous years are observed with important reduction of ozone depletion: 1988, 1991, 2002, 2004, 2010 and 2012. Note that these years correspond to a high contribution of HF and GRAD proxies to the regressed ozone values (Fig. 9, second panel). If we exclude these anomalous years, the 220 DU relative area remains fairly stable at about 90 % of the total vortex in average since 1990. In the most recent years, relative areas for 125 and 150 DU thresholds decrease to less than 10 and 30 %, respectively, from their peak value of 21 and 57 % reached in 2000. If such trend persists, the frequency of very low ozone values (e.g. below 125 DU) is expected to become negligible in the coming decade.

In addition, OMD were computed for the same thresholds (bottom panel of Fig. 12). The evolutions of OMD present similar behaviour as the relative area, but in this case OMD at 220 DU threshold shows a visible decrease since 2000. Nowadays OMD at a threshold lower than 150 DU presents very small values lower than 0.2 Mt.

Solomon et al. (2016) have highlighted for the first time a delay in the formation of the ozone hole after 2000. This shift can be explained by the slower ozone loss rates after sun appearance over the pole due to ODS decrease in the polar stratosphere. In this work, such a time shift was investigated by computing the first day when ozone levels below certain thresholds occur inside the vortex (using the 400–600 K classification range), from 1 September to 15 October (Fig. 13). The same thresholds values as for Fig. 12 were used. In order to avoid influence of spurious values, the number of 1∘ × 1∘ grid cells with total ozone columns below the various thresholds in the first day (or start day) has to be larger than 10. For each curve, day values equal to 244/245 correspond to years when ozone levels below the corresponding threshold have appeared at or before the beginning of September. For the 220 DU threshold, the dark blue curve shows that this is the case since 1983. For the 200 DU threshold, lower ozone values appear before the beginning of September after the mid-1980s. For the other thresholds, we observe a decrease, with some variability, of the start day during the 1980s, the two lowest thresholds in the 1990s and an increase after 2000–2005. This increase is most visible on the 125 DU threshold curve and to some extent also on the 150 DU threshold curve. In 2016, ozone levels below 150 DU have appeared in the beginning of September, similar to ozone holes at the end of the 1990s, but levels below 125 DU still appeared later. No values for a particular year in the threshold curves indicate that total ozone levels were above that threshold during the whole period considered. This is the case for the two lower thresholds before 1985 and for the 125 DU threshold in 2002, 2004 and 2017.

MSR-2 and SAT (TOMS and OMI with gaps in 1993–1995 filled by MSR-2) datasets have been used to evaluate total ozone trends within the southern polar vortex over the 1980–2017 period. A multi-regression model is applied to ozone values averaged over the month of September and the 15 September to 15 October period in order to compute long-term trends before and after the ODS peak in the polar stratosphere that occurred around 2001 (Newman et al., 2007). The 15 September–15 October time range corresponds to the period of maximum ozone depletion. It was not commonly used in previous works. Proxies and time windows for averaging them are selected following de Laat et al. (2015) work.

For the classification of total ozone measurements inside the vortex, the classical Nash et al. (1996) method is used. In order to evaluate the impact of vortex baroclinicity on trend analysis, classifications using a single isentropic levels (475 and 550 K) and a range of levels (400–600 K) are tested. Systematic differences are found between the various datasets. However, the interannual variability is similar, with correlation coefficients ranging from 0.98 to 0.99 in both studied periods. While larger trend values are generally found with the 475 K classification, the differences with trends related to the 400–600 K range classification are not significant at 2σ level.

The use of combined piece-wise linear and parabolic functions for the trend proxies (PWT) in the 1980–2000 and 2001–2017 periods provides a good representation of the total ozone long-term behaviour inside the vortex (after removal of interannual variability), especially for the 15 September–15 October period, probably in relation to the effect of ozone loss saturation. The classical PWLT used in previous studies seems to overestimate the trends during the recovery period.

A new proxy (GRAD) representing the vortex stability over both studied periods is included in the multilinear regression. This proxy improves the representation of total ozone interannual variability by the regressed values especially over the last decade. It results in a ∼ 0.05 larger value for the R2 determination coefficient, lower fitted residuals and smaller trend uncertainties for the different classification methods and datasets. In general, the best agreement between observations and regressed values is found for the 15 September–15 October period. While the HF combined with GRAD proxies reproduce the interannual variability of ozone quite well, other proxies such as Aer, QBO, SF and AAO present smaller explanatory power and contribute less to reduce trend uncertainties.

In the period of increasing ODSs (1980–2000), the MLR analysis shows negative and significant trends for both studied periods, similar to values found in previous studies (e.g. Kuttippurath et al., 2013, and de Laat et al., 2015). The 15 September–15 October period presents slightly larger negative trends in absolute value than the month of September.

In the 2001–2017 period, positive trends are obtained for all scenarios. The largest trends and highest significance are found for the September period, with a trend value of 1.84 ± 1 DU for the MSR-2 total ozone record using the 400–600 K range classification method. For the 15 September–15 October period, a lower trend of 1.42 ± 0.92 DU is obtained using the same record. Better fit and smaller residuals are obtained for that period. Differences with trend results from the other SAT dataset evaluated in the study are not statistically significant.

The ratio between trends before and after 2001 varies according to the studied period. Only September trends present a ratio of ∼ 3 as expected for an ozone response to ODS evolution. However, as for other trend studies based on MLR fit to observations, it is not possible from this analysis alone to fully attribute the retrieved trends to ODS evolution. For such a study, a combination of model and observations is needed. Potential feedbacks between chemistry, radiation and dynamics will play a role in ozone recovery. A recent study indicates an increase in temperatures within the vortex core from MERRA reanalyses during the period 2000–2014 in austral spring and summer (Solomon et al., 2017). Such a temperature increase that could be linked to ozone increase could play a role in the decrease of occurrence of low ozone values within the vortex and subsequent ozone increase.

The evolution of OMD was also analysed using MSR-2 data. MLR analysis on this metric confirms the findings obtained for total ozone columns, e.g. a general improvement of the fits with the GRAD proxy and the main explanatory power provided by the GRAD, HF and PWT proxies. The 2001–2017 OMD trends are larger in absolute value for September (-0.86 ± 0.36 Mt yr-1) than for 15 September–15 October (-0.65 ± 0.33 Mt yr-1). They are significant at 2σ level in both cases. These results are in general good agreement with those obtained in de Laat et al. (2017). Similar reductions of 53 and 35 % of OMD are computed for September and 15 September–15 October, respectively. This is consistent with the 30–40 % change in ODSs relative to their level in 1980, when total ozone values below the 220 DU threshold started to appear systematically (WMO, 2011).

The structural uncertainties of the MLR analysis linked to the selection of proxies were not fully analysed in this work, as in de Laat et al. (2015). The main sensitivity tests concerned the baroclinicity of the vortex and the impact of its stability during the studied periods. Trend differences in the various scenarios analysed provide some quantification of related uncertainties and are lower than the statistical trend uncertainties. Further, the large determination coefficients obtained for both periods analysed give confidence in the retrieved trends. The heat flux proxy that provides the largest explanatory power in the various fits is a well-known driver of vortex temperature conditions that are the primary causes of polar ozone depletion in periods of high ODS levels. The influence of the GRAD proxy in recent years highlights the importance of the vortex stability for the containment of the ozone hole during the period of maximum depletion.

Polar ozone recovery was also evaluated by examining the temporal evolution of relative areas occupied by ozone levels below various thresholds within the vortex. Very small total ozone columns (< 150 DU) did not occur inside the vortex before the late 1980s and early 1990s. For the 125, 150 and 175 DU thresholds, relative areas display a steady decrease since the beginning of the 21st century, while for the 200 and 220 DU thresholds, the relative area's evolution is quite stable. All relative area curves are marked by increased variability since 2000. Relative areas related to the lowest thresholds show a more rapid decrease, which further points towards polar ozone recovery. OMD records based on the same thresholds show a similar behaviour.

In summary, this work present clear symptoms of polar ozone recovery. Recovery is found for the month of September and for the first time for the period of maximum ozone depletion, e.g. from 15 September to 15 October. For both studied periods, recovery is deduced from the significant positive trends in total ozone, significant negative trends of ozone mass deficit and from the steady decrease of the occurrence of low ozone values within the polar vortex. As ODSs continue to decrease in the next years, it is likely that ozone recovery in the polar vortex in spring will become more evident.