Weekly electrochemical cell (ECC) ozonesondes have been launched in tandem with radiosondes at Lauder since August 1986, measuring profiles of ozone, temperature, pressure, humidity, and wind speeds and directions from the surface up to about 35 km . The ECCs used for ozone sounding at Lauder are the Science Pump Corporation (SPC) series 4A/5A/6A (before 1996) and the Environmental Science (EN-SCI) Z series (after 1996), although there is some overlap period when both types were used. These ECC series were operated with a 1.0 % buffered potassium iodide (KI) cathode solution until July 1996 and a 0.5 % KI solution from August 1996 until present. These changes are relevant to the homogenisation process and are detailed in Table .

The homogenisation procedure, described in the Assessment of Standard Operating Procedures for Ozonesondes (ASOPOS 2.0) documentation and in the Ozonesonde Data Quality Assessment (O3S-DQA) activity , was applied to the Lauder ozonesonde time series, available at NDACC. These NDACC data, named “uncorrected data” hereafter, have been obtained by converting the raw currents measured with an ozonesonde to ozone partial pressures by subtracting a measured background current using a conversion efficiency of 1.0, the measured pump temperature, and the pump flow rate measured prior to launch in the lab and then correcting for the pump efficiency decrease with increasing altitudes. The O3S-DQA homogenisation, however, adds corrections to the pump temperature, the pump flow rate (due to the moistening effect), and the background current (avoiding too-high values) on top and uses a set of transfer functions applied to the conversion efficiency to remove biases due to changes in the instrument or operating procedures. For instance, as the 0.5 % KI solution has become the recommendation for the EN-SCI ECCs, a transfer function is applied to the profiles between 1994 and 1996, where the 1 % KI solution instead of the 0.5 % KI solution was used for the EN-SCI ECCs at Lauder. The re-processing of the Lauder data according to the O3S-DQA guidelines was carried out by the HEGIFTOM working group (Harmonization and Evaluation of Ground Based Instruments for Free Tropospheric Ozone Measurements, https://hegiftom.meteo.be, last access: 21 May 2024) within the TOAR-II (Tropospheric Ozone Assessment Report phase II, https://igacproject.org/activities/TOAR/TOAR-II, last access: 21 May 2024) initiative. Further details regarding the corrections are summarised in Appendix A.

In this study, we include a total of 1958 ozonesonde flights between August 1986 and June 2021, the data of which have been homogenised. Both homogenised and measured uncorrected datasets have been post-processed for linear trend calculations. The homogenised data are used in the MLR analysis. Linear piecewise regression was applied to interpolate the original ozone profiles from the surface to 30 km at a 1 km vertical resolution. We then exclude some extreme ozone values, identified as the values that are outside the 3-standard-deviation range, to create monthly means by averaging the data available for that month at each re-gridded vertical level.

We construct a multiple linear regression (MLR) model to identify the dominant factors that are associated with O3 variations and trends. The regression model approximates the annual mean ozone anomalies for each level, as well as for eight grouped layers where annual mean ozone anomalies are averaged (0–1.5, 1.5–3, 3–6, 6–9, 9–12, 12–15, 15–20, and 20–25 km). The auto-correlations that usually exist in the monthly varying data have been largely removed by averaging them into annually varying data. The purpose of this study is to analyse the interannual variations and trends in annual mean ozone. The regression models include nine terms representing the solar index (SI), which captures solar variability and is defined by the solar radio flux at 10.7 cm, the multivariate El Niño Southern Oscillation Index (MEI), the quasi-biennial oscillation at 30 and 10 hPa (QBO30 and QBO10, respectively), the tropopause height (HTTrop), the stratospheric temperature (TStrat), the surface relative humidity (RHsurf), the aerosol optical depth (AOD), and the equivalent effective stratospheric chlorine (EESC). The two QBO indices are orthogonalised with respect to each other. The EESC is defined as a relative measure of the potential for stratospheric ozone depletion that combines the contributions of chlorine and bromine from surface observations from ODSs and is calculated based on the ozone-depleting substances from the Coupled Model Project 6 (CMIP6) historical (until 2014) and Shared Socio-economic Pathway (SSP245) (for 2015–2021) scenarios . Surface relative humidity is measured by the radiosonde that has a humidity sensor. We define the tropopause based on the WMO lapse rate definition , which is calculated from the temperature measured by the radiosonde during each ozonesonde flight. The regressed O3 anomaly is expressed as 1Ozone(t)=a1⋅SI(t)+a2⋅SOI(t)+a3⋅QBO10(t)+a4⋅QBO30(t)+a5⋅EESC(t)+a6⋅AOD(t)+a7⋅HTTrop(t)+a8⋅TStrat(t)+a9⋅RHsurf(t)+ϵ(t). Here, Ozone(t) is the annual mean ozone anomaly; ϵ is the regression residual, minimised in the rms; and a1-9 are the regression coefficients for the corresponding regressors, all normalised to vanishing means and unit standard deviation (i.e. standardised). All forcings used in the regression are summarised in Table .

We use the NIWA-UKCA model simulations from the Chemistry-Climate Model Initiative project (CCMI-1: ) to assess the impact of the major anthropogenic forcings, including greenhouse gases (GHGs) and ozone-depleting substances (ODSs), on ozone changes at Lauder. The Lauder ozonesonde measurements cover both the ozone depletion and the recovery periods of 1987–2020. Global chemistry–climate models (CCMs), including NIWA-UKCA , generally have coarse spatial resolutions. Therefore, it is often not ideal to use the simulations from the CCMs for reproducing the observed trends at a specific location. Instead, the CCMs can be used to attribute the trends to various forcings on a wider spatial and temporal scale. Here, we calculate the ozone trends using the NIWA-UKCA simulations to gauge the impacts of GHGs and ODSs on ozone changes at Lauder over the observational period in a limited area (averaged over 40–50° S and 160–180° E). We also show the simulation results on a global scale in context. The CCMI-1 simulations from NIWA-UKCA used here consist of all the forcings (including time-varying GHGs, ODSs, and ozone and aerosol precursors) of the coupled atmosphere–ocean reference experiment “RefC2”, covering the simulation period of 1960–2100 (we keep the same experiment naming convention as defined by ) and its corresponding fixed single-forcing sensitivity experiments (sen-C2-fODS, sen-C2-fGHGs, sen-C2-fCH4, and sen-C2-fN2O) in which ODSs, the combined GHGs, methane (CH4), and N2O are individually fixed at their 1960s levels. The impact of each single forcing on ozone is derived from the differences in ozone between the RefC2 ensemble mean (five members) and the ensemble averages of the corresponding fixed single-forcing simulations (one to three ensemble numbers for each experiment). We directly assess the impacts of changes in ODSs, combined GHGs, methane, and N2O on ozone trends based on available simulations. However, no simulation was performed to directly assess the impact of CO2 within CCMI-1; instead, it will be assessed by subtracting the impacts of methane and N2O from the impact of the combined GHGs . We again separately examine the changes in modelled ozone trends over the periods of 1987–1999 and 2000–2020. The detailed description of the model and experiments can be found in and in and the references therein.

The homogenised and the uncorrected datasets are directly compared without any temporal interpolation but are both interpolated to a 1 km vertical grid using piecewise linear regression. Figure  shows the percentage difference between vertical ozone profiles from the two datasets for all flights. The correction procedure and the impact of each correction are described in more detail in Appendix A. Overall, corrections lead to mostly increased ozone values in the homogenised time series, reaching 6 % to over 10 % before 1995 due to the pump temperature correction (Fig. c in the Appendix). The pump flow rate correction results in a uniformly positive effect of less than 2 % in general (Fig. d). There are scattered increases in ozone in the homogenised time series compared to the uncorrected data, especially between 2012 and 2015, when a modified background current correction is applied (Fig. b). The effect of the changes in the KI solution concentration on the conversion efficiency (Fig. a) is mainly negative between 1994 and 1996 but positive in the beginning of the time series (1986), when a smaller cathode sensing solution amount was used (2.5 mL instead of 3 mL).

The differences between the homogenised and the uncorrected monthly mean ozone time series are calculated, excluding outliers defined for ozone outside the 3-standard-deviation interval of all data points for that level (Fig. ). This step removes fewer than 1 % of data points from the monthly mean ozone calculations. Most of these outliers are around 10 km, where the ozone is subjected to large dynamical variations.

Consequently, the trend in ozone can be affected by the homogenisation. Here, we calculate simple linear trends in the two datasets for comparison. Figure  displays linear trends from the surface to 30 km for both homogenised and uncorrected datasets over the pre-1999 and post-2000 periods, respectively. Also displayed is the observed FTIR ozone trend for the post-2000 period for comparison. All trend calculations use annual mean ozone anomalies to minimise the auto-correlation in data as opposed to using the monthly mean data. This shows that there are systematic differences in ozone trends in the homogenised data compared to in the uncorrected data over the 1987–1999 period (Fig. a). During this period, the vertically resolved trends in the uncorrected dataset are slightly negative throughout most of the domain above 10 km and positive below 10 km, although these trends are generally insignificant at the 95 % confidence level. In contrast, the trends in the homogenised data are negative throughout the domain (∼ -2 % per decade to -13 % per decade), with most of the trends below 5 km and above ∼ 24 km being statistically significant at the 95 % confidence level. This result is more consistent with the impact of increasing ozone-depleting substances (ODSs) over this period.

In the post-2000 period, the calculated trends are very similar between homogenised and uncorrected ozone profiles (Fig. b). Both show significant positive trends of up to ∼ +2 % per decade in the free troposphere and a significant negative trend of ∼ -2 % per decade to -6 % per decade above ∼ 16 km in the stratosphere, which peaks around 18 km. We note that the lower-stratospheric ozone negative trend of ∼ -3 % per decade to -6 % per decade between 15 and 20 km looks visibly smaller in magnitude than the trend presented in , where it exceeds -6 % per decade in the same region, but both of these are within the uncertainty ranges displayed by . Distinct negative trends of ∼ 2 % per decade to 4 % per decade also exist in the upper troposphere and the lower stratosphere between 8 and 16 km, albeit with large statistical uncertainty, highlighting the large dynamical variability typical for this region. We find that the vertically resolved ozone trends calculated by excluding the outliers are very similar to the trends calculated by including all data points. The only difference is that, by excluding the outliers, we have reduced the trend uncertainties around the 10 km region (not shown). The vertically resolved trends in observed ozone for the post-2000 period are in excellent agreement with the trends derived from the FTIR ozone data (Fig. b). Note that the updated FTIR ozone data presented here, obtained using an updated retrieval strategy , are markedly different from the FTIR data shown by . The negative trends in the lower stratosphere in both the sonde and the FTIR ozone data shown here are noticeably larger in magnitude than the trends in the satellite data shown by , which have typically insignificant trends of less than -2 % per decade.

The regression model (Eq. ) was constructed using annual mean ozone anomalies of the homogenised ozonesonde data (hereafter, we refer to the homogenised ozone dataset as the “observed” ozone at Lauder). The regression is performed for each level from the surface to 30 km at a 1 km resolution, and the linear trend of the predicted ozone at each level was then calculated. Figure  shows that the vertically resolved ozone trends from the MLR-predicted ozone are quite similar to the simple linear trends in the observed ozonesonde data, but the uncertainties in the MLR-predicted ozone trends are generally smaller than those in the observed trend. The trends in the MLR-predicted ozone are also systematically smaller in magnitude in the stratosphere above ∼ 18 km for both periods, but the difference is slightly larger for the post-2000 period, though both are within the uncertainty ranges of the observed trends. This indicates that the regressors used in the MLR model do not fully capture the observed trend there. Despite the slight differences in magnitude, trends calculated here and by point to the fact that significant negative trends in the lower stratosphere exist in Lauder ozonesonde data, and these negative trends are underestimated by the satellite products.

We then group the vertical ozone profile into eight layers from the surface to 25 km in order to identify the drivers of ozone variability and trends for each vertical layer. The same MLR is performed individually for each layer. The observed and regressed ozone anomalies, together with the leading contributions from individual regressors, are shown in Fig. . The ozone variance explained by the regression is given by the multiple regression coefficient of determination R2 (Fig.  and Table ). The standardised individual regression coefficients for each regressor can be used to measure their contributions to the total variance explained at that level (Table ).

The regression model matches the observed anomalies well, particularly in the stratosphere, the upper troposphere, and near the surface (Fig.  and Table ). With R2 values ranging from 0.28 to 0.61 in the troposphere and 0.57 to 0.71 in the stratosphere, the stratospheric ozone variations and trends are better captured by the MLR model than tropospheric features. Indeed, interannual variations in ozone anomalies in the upper troposphere and the lower stratosphere (9–15 km) are especially well explained by variations in tropopause height (Fig.  and Table ). The downward trend in the stratospheric ozone between 9 and 20 km is clearly driven by the significant positive linear trend in tropopause height (Fig. c). The QBO at 30 hPa, together with tropopause height, also explains a large part of the ozone variability for the 15–20 km layer (Fig.  and Table ). Above 20 km, the QBO at both 30 and 10 hPa, the AOD, and the stratospheric temperature (TStrat) have a significantly negative trend (Fig. b). We note that the correlation between AOD and the Lauder stratospheric ozone is positive, e.g. after the Mt Pinatubo eruption in 1991, despite the potential ozone depletion in the years following a volcanic eruption (Fig. h). This lack of ozone depletion was attributed to the perturbation of the stratospheric dynamics by the Mt Pinatubo eruption that obscured the chemical effect in the southern extra-tropics . We also note that the prolonged decline in ozone above 15 km at the end of the time series (around 2020) cannot be explained by the MLR model (Fig. g and h); this might be the result of the Australian bush fires in January 2020, which depleted stratospheric ozone . The lack of this process in MLR might also explain the weaker negative trend in the regressed ozone compared to that in the observed ozone above 18 km (Fig. b).

It is well established that increases in CO2 influence temperature, humidity, and circulation, which in turn affect ozone chemistry and transport . Warming in the troposphere and cooling in the stratosphere due to the increase in CO2 have driven the increase in tropopause height over the last several decades based on radiosonde observations, reanalysis data, and modelling . The tropopause height derived from the Lauder sonde data shows a significant positive trend of 117 ± 82 m per decade (at 95 % confidence) over the observational period, calculated as the simple liner trend in the annual mean anomaly (Fig. c), which is larger than the trend of ∼ 50–60 m per decade in the Northern Hemisphere (20–80° N) over 2001–2020 based on radiosonde data .

However, in the middle and upper troposphere (6–9 km), the regression function explains less of the ozone variation compared to at levels above and below (Fig. d). Here, although the solar influence is the strongest in relative terms, influences from all other regressors, except the QBO at 10 hPa, contribute non-negligibly to explaining the ozone variations at this level.

In the lower and free troposphere (below 6 km), the sharp decreases in ozone during the early period of the record and the large negative anomalies in 1997–1998 are well reproduced by the regression, along with the subsequent increases in ozone there. This trend transition in tropospheric ozone coincides with the evolution of EESCs, which increased from the late 1980s before declining after 1997; this indicates that the impact of stratospheric ozone changes due to changes in ODSs could impact tropospheric ozone through transport. Indeed, the interannual variation in the free tropospheric ozone is shown to be influenced by the QBO at 30 hPa (Fig. b and c). In the troposphere, increases in ODSs are expected to drive an increase in ozone after the late 1990s, whilst the response to the CO2 increase is more complex. For instance, the associated increasing humidity would lead to more chemical destruction of ozone in the troposphere, and the increase in temperature may result in more ozone production through NOx–CH4 (and volatile organic compound) chemistry e.g.. Here, relative humidity and surface ozone are anti-correlated (Fig. a). Relative humidity has a large negative impact on surface ozone (Table ). Moreover, we have not considered changes in ozone precursor concentrations and other meteorological parameters in the regression that could substantially impact tropospheric ozone. Therefore, more explanatory variables can be included in an MLR model that is specifically focused on tropospheric ozone; this is the subject of a future study.

We examine the modelled vertically resolved ozone trends in the vicinity of Lauder (40–50° S and 160–180° E) from the NIWA-UKCA model over the ozone depletion (1987–1999) and recovery (2000–2020) periods separately, along with the effects of changes in ozone trends due to individual forcings, using the same approach as in . Meanwhile, in order to help understand Lauder ozone changes in a global context, we also show the modelled zonal mean ozone trends covering all latitude bands and the changes that are attributable to ODSs and GHGs (Appendix B). The modelled and attributable trends are linear trends in diagnosed annual mean ozone anomalies calculated from model simulations.