This study uses six long-term merged limb satellite ozone profile datasets: GOZCARDS (Global OZone Chemistry And Related trace gas Data records for the Stratosphere), SAGE-CCI-OMPS (SAGE II, Climate Change Initiative and OMPS), SAGE-OSIRIS-OMPS (SAGE II, OSIRIS and OMPS), SAGEII-OSIRIS-SAGEIII (SAGE II, OSIRIS and SAGE III), SAGE-SCIA-OMPS (SAGE II, SCIAMACHY and OMPS), and SWOOSH (The Stratospheric Water and Ozone Satellite Homogenized). All datasets provide monthly zonal mean stratospheric ozone profiles and are widely used for long-term trend analyses e.g.. Most datasets begin in 1984/1985 with the launch of SAGE-II, while GOZCARDS extends back to 1979, providing the longest record. All datasets are extended until the end of 2024, enabling consistent post-2000 trend analyses. The temporal coverage of contributing instruments is shown in Fig. , and dataset characteristics including latitude coverage, vertical range, ozone units, while contributing instruments are summarized in Table . The datasets employ different merging strategies, including anomaly-based merging, debiased absolute-value merging, and absolute-value merging with retained seasonal structure. Despite these methodological differences, all datasets are internally consistent, quality-controlled, and suitable for assessing long-term stratospheric ozone trends. For consistency, all ozone profile datasets are converted to a common latitude–altitude grid and unified ozone units (Sect. ). A more detailed description of the datasets is given by and .

Six long-term merged total column ozone datasets: MSR2 (Multi-sensor re-analysis), GSG (GOME/SCIAMACHY/GOME2), GTO-ECV (GOME-type Total Ozone – Essential Climate Variable), SBUV (Solar Backscatter Ultraviolet) NASA (MOD (Merged Ozone Data Set)), SBUV NOAA (COH (Cohesive)), and WOUDC (World Ozone and Ultraviolet Radiation Data Centre) are analysed. All datasets extend through the end of 2024 and provide quality-controlled estimates of total ozone suitable for long-term trend analysis. The zonal mean total ozone records generally begin earlier than the profile datasets, with start dates from 1960 (WOUDC) and 1970 (SBUV-MOD). SBUV-COH starts in 1978, while GTO-ECV and GSG start in 1995. In the MLR, the latter two datasets are extended back in time to 1978, the starting year of the regression, using the SBUV-COH data . The temporal coverage, latitudinal range, and contributing instruments of each dataset are summarized in Table , while the contribution of individual instruments is illustrated in Fig. .

The datasets comprise a mix of satellite-only merged products and one dataset based upon ground-based observations (WOUDC). The MSR2 dataset provides globally gridded total ozone fields based on data assimilation using individual data records that are bias corrected using ground-based data from Brewer and Dobson instruments . The remaining satellite products apply homogenization and bias-corrections to match individual data records. WOUDC complements the satellite records by providing zonal mean data derived from ground-based total ozone measurements from Dobson, Brewer, and filter instruments. Zonal mean ground data are calculated combining station data and an ozone climatology . More details on the various merged datasets can be found in and .

In this study, all ozone profile datasets are expressed in units of number densities [molec cm⁻³] as a function of geometrical altitude [km]. This conversion is needed to directly compare the ozone trends in absolute units and will help in calculating the stratospheric column. The conversion was applied to GOZCARDS and SWOOSH (see Table ). Meteorological data (geopotential height, pressure, and temperature) from the ERA5 re-analysis data were used for the conversion of the datasets . Apart from the uncertainties in reanalysis temperatures, the conversion applied to monthly mean data, as provided in the merged data product, may lead to uncertainties compared to a merged data product based upon single ozone profiles converted before calculating the monthly means. The conversion of individual ozone profiles before averaging, gridding, and merging is considered to be more accurate, but is not possible if monthly mean merged data sets are used. We quantify this additional uncertainty using Aura MLS (Microwave Limb Sounder) ozone and temperature observations.

MLS, aboard NASA's Aura satellite, provides high vertical resolution ozone profiles spanning from the upper troposphere into the mesosphere . The long-term stability of Aura MLS has been assessed in comparisons to ground-based data e.g. and MLS is, therefore, suitable for assessing the conversion uncertainty.

The limb ozone profile merged datasets (Table ) are level 3 (L3) monthly mean zonal mean datasets having different ozone units, vertical grids, and spatial resolutions. For a direct comparison of the profile ozone trends, we convert the datasets to common ozone units and vertical grid and re-bin them to a common spatial grid. The advantage of using number density and geometric altitude is the calculation of partial ozone column amounts (e.g. stratospheric) by simple vertical integration of the ozone profile. The units and altitude grid conversion are only needed for SWOOSH and GOZCARDS.

The common latitude grid is defined by the lowest available resolution of the profile datasets, which ranges from 85° S to 85° N (center latitudes) in 10° steps. The merged ozone datasets were created from level-2 (L2) orbit data. The unit conversion of the monthly mean L3, as done in this work, instead of L2 data, may contribute to trend uncertainties. To evaluate the impact of this conversion, we compared two conversion approaches using the MLS/Aura L2 data (2004–2022). The first approach converts the MLS L2 individual profiles to the common ozone and altitude units and afterward creates the monthly mean L3 data. The second approach first determines the L3 data and then applies the unit conversion to the monthly mean L3 data using monthly mean zonal mean meteorological parameters: temperature, pressure, and altitude from ERA5. The linear trends of the difference timeseries between both approaches give an estimate of the trend uncertainty resulting from the unit conversion.

Figure shows the trend of the difference timeseries between both approaches as a function of latitude and month for selected altitude levels. Most of the trends are not significant (hatched areas). At 25 km, a negative trend of around 1 % per decade between April and August in the northern hemisphere (NH) and between January and July in the southern hemisphere (SH) is noted. The significance of these trends decreases with altitude, showing the most significant trends at 25 km. This is possibly because of stronger vertical ozone or temperature gradients near this altitude, which increase the uncertainties from the non-linear ozone unit conversion and vertical axis transformation. Especially in the SH, insignificant trends are observed at higher altitudes (40 km). In general, we conclude that the trend uncertainty due to unit and altitude grid conversions is small and insignificant, but in selected regions and months non-negligible.

Another contribution to uncertainties in long-term trends arises from relative drifts among the merged datasets. Here, relative drifts refer to systematic long-term differences between individual datasets and a common reference, in this case the multi-instrument mean (MIM). The spread of the drifts provides an estimate for possible contribution to trend uncertainties derived from multiple datasets.

To assess these relative drifts, a simple linear regression is applied to the time series of differences between the monthly mean merged datasets and the MIM. The percentage ozone difference timeseries is calculated as follows for each profile dataset: 1ΔO3(z,t)=100⋅O3(z,t)-MIM(z,t)MIM(z,t), where O3 is the ozone concentration of the merged dataset (in number density or DU). The linear trend (or relative drift) of the ΔO3 timeseries from all datasets is computed, and the standard deviation of the mean of these trends represents the 1σ drift uncertainty, which is independent of the common reference, here MIM.

Figure shows the 1σ drift uncertainty derived from the six merged profile datasets until the end of 1995 and since 2000. It ranges from 0.5 % per decade to 0.75 % per decade for the period 2000 to 2024 between 60° S and 60° N and above 20 km. The uncertainties are larger at higher latitudes and lower altitudes and are most likely due to higher uncertainties in ozone measurements. In the period between 1985 through 1995, the uncertainties show higher values from close to the tropopause up to 25 km for latitudes above 50° N / 50° S. Above this altitude and in the zonal band between these latitudes, the uncertainties are comparable in magnitude with those from the later period.

The same procedure was applied to the total ozone datasets, and results are shown for the period since 1996 (ODS peak) and 2000 in Fig. . We did not include the period before the ODS peak in this figure, because two of the total ozone datasets start around 1995 (GTO-ECV and GSG), leaving us with a limited selection of merged datasets and instruments before that year. Instead we show the relative drift uncertainty for two different start years (1996 and 2000), which were used in previous studies . Both relative drift uncertainties look very similar and show no significant differences. The drift uncertainties in the period 2000–2024 are slightly larger than in the period starting in 1996. When comparing the drift uncertainty for the total ozone datasets to the profile ozone datasets (Fig. ), we see that the total ozone dataset drifts are overall lower and range between 0.1 % per decade and 0.5 % per decade. The highest values can be seen at high latitudes and in the tropics, while the mid-latitudes show very low uncertainties. Higher drift uncertainties at higher latitudes are most probably due to the combination of higher uncertainties in ozone itself (higher solar zenith angles) and a larger natural variability in ozone. It is not clear why the drift uncertainties are enhanced in the tropics. Probably the low natural ozone variability masks the drifts to a lesser degree.

We apply a multiple linear regression (MLR) model to time series of ozone at each altitude-latitude grid point (for profile data) and each latitude (for total column data) including terms that account for various atmospheric processes, e.g. solar cyle, QBO, and El-Nino. The following equation shows the MLR equation with additional proxies (see Table ) as used in this study: 2y(z,t)=β1(z,t)⋅QBO1(t)+β2(z,t)⋅QBO2(t)+β3(z,t)⋅ENSO(t)+β4(z,t)⋅Solar(t)+β5(z,t)⋅SAODElChichon(t)+β6(z,t)⋅SAODPinatubo(t)+β7(z,t)⋅SAODHTHH(t)+β8(z,t)⋅AO(t)+β9(z,t)⋅AAO(t)+β10(z,t)⋅EHFS(t)+β11(z,t)⋅EHFN(t)+(β12(z,t)+β13(z,t)(t-t0))⋅Lpre(t)+β14(z,t)⋅LGap(t)+(β15(z,t)+β16(z,t)(t-t1))⋅Lpost(t)+ε(z,t) with y(z,t) being the monthly mean ozone as a function of altitude z and running month t and βk the regression coefficients.

Terms containing Lpre, LGap(t), and Lpost cover only consecutive periods with separate linear trend coefficients, β13 and β16, before and after the ODS peak in 1996, respectively. These three variables are defined as follows: 3Lpre(t)=1if t<t00if t≥t0, 4LGap(t)=0if t<t0 and t≥t11if t0≤t<t1, and 5Lpost(t)=0if t<t11if t≥t1. t0 is January 1996 and t1, December 1999. Note that the term with LGap(t) does not contain a linear trend coefficient during the gap period (1996–1999). This is needed to reduce the influence of the selection of a turnaround point between ozone decline and recovery because the timing of the trend reversal varies with latitude and altitude . By introducing a transition period rather than a single inflection year, the regression model allows for a more gradual change between the two trend segments.

To consider seasonality more explicitly, the regression model uses harmonic terms to represent the annual and semi-annual cycles of selected terms, as shown in Eq. (). 6βk(z,t)=βk0(z)+∑i=12βk1i(z)sin⁡2πit12+βk2i(z)cos⁡2πit12. Harmonic terms are only used for the QBO, ENSO, and fit coefficients in the trend terms (β12 to β16). This approach is robust for non-deseasonalized data and helps reduce uncertainty in the estimated trend, particularly in datasets with sparse temporal coverage.

Trend uncertainties are derived from the covariance matrix of the regression coefficients. Since ordinary least squares assumes uncorrelated residuals, it tends to underestimate uncertainties when applied to ozone time series. To account for autocorrelation, we correct for first-order autocorrelation (AR-1) in the regression residuals . This is achieved by calculating the lag-1 autocorrelation and applying a transformation to the regression model that reduces the serial correlation of the residuals . The transformed regression equation provides the adjusted fitting coefficients and their uncertainties.

When reporting mean trends from multiple datasets, we use two different formulas for the uncertainty, explained in the LOTUS report and in more detail in . For consistency with the previous analyses , we present both uncertainty estimates in our paper. First we use the LOTUS uncertainty estimate , which is defined as the maximum uncertainty of two terms: 7σmean2=max⁡1N2∑i,jCijσiσj,1neff∑(xi-x¯)2N-1. The first term represents the variance of the mean of correlated trend estimates, accounting for the individual regression uncertainties and the correlations of model-data residuals. The LOTUS report uses constant correlation coefficients Cij for the pre-1997 period and for the post-2000 period. In this paper we use latitude- and altitude-dependent correlation coefficients, which do not change the results substantially . The second term is based on the sample variance of the individual trend estimates; the effective number of independent datasets, neff is defined as follows: 8neff=N2∑i,j=1NCij .

The second approach, the J-distribution, is a more conservative estimate and characterizes the total ensemble variance of trend estimates , which is defined as follows: 9σmean2=1N∑σi2+∑(xi-x¯)2N-1. The first term accounts for uncertainties in the trends of the individual datasets and the second term represents the contribution from taking the statistical mean of the individual trends. The joint distribution is considered a conservative estimate and produces larger uncertainties than other methods . Both the LOTUS estimate (Eq. ) and the J-distribution (Eq. ) consist of two components: the first arises from uncertainties in the MLR fit, while the second characterizes the spread of trend estimates across datasets. However, the two approaches differ in the interpretation of these terms. The J-distribution assumes that the dominant contribution to the MLR uncertainty, estimated from fit residuals, reflects unresolved variability not captured by the trend model. As a consequence, trend estimates from different datasets are effectively treated as fully correlated with respect to this component, and averaging does not reduce this contribution. In the case of Cij=1, the first terms in Eqs. () and () become identical. In contrast, the LOTUS estimate accounts for partial correlations between trend uncertainties, with correlation coefficients derived from fit residuals and ranging between approximately 0.6 and 0.95 for the datasets considered here. The second term in the J-distribution represents the full ensemble variance, whereas the LOTUS estimate corresponds to the standard error of the mean for correlated values. Since neff≈1.4 for the considered datasets, this term is only slightly reduced in the LOTUS formulation compared to the J-distribution. The J-distribution combines both contributions additively, while the LOTUS estimate uses the maximum of the two terms. For consistency with previous analyses , both the LOTUS and J-distribution uncertainty estimates are used in this study. A detailed comparison of the two approaches is provided in .

A list of the proxies used here is given in Table . The proxy timeseries have been normalized to the range from -0.5 to 0.5, except for the stratospheric aerosol optical depth (SAOD) (0 to 1 range). They have not been detrended, which means the dynamics-related trends are captured by the proxies rather than by the linear trend terms. This way the linear trends approximate the long-term ODS-related ozone change . In contrast to the common proxy setup of the LOTUS model , we add the Arctic and Antarctic Oscillation (AO/AAO), the Brewer-Dobson Circulation (BDC) as expressed by the eddy heat fluxes (EHFN, EHFS), and three separate SAOD proxies . Both EHFN and EHFS are proxies derived from ERA5 reanalysis ín the respective hemisphere (45–75°). The Brewer-Dobson circulation in boreal and austral winter has an ozone response in the tropics opposite to that in the extratropics e.g.. There is also SH ozone response to the Brewer-Dobson circulation in the NH . For this reason, EHFN and EHFS are included in the MLR for all latitudes. Similar arguments apply to AO/AAO proxies.

We chose these additional regressors to better capture dynamic variability in ozone, such as the high ozone in the NH in 2024 . The Arctic Oscillation (AO), Antarctic Oscillation (AAO), and Brewer–Dobson circulation (BDC) proxies are winter/spring accumulated and weighted by the monthly ozone-variability to account for their seasonally integrated influence on stratospheric ozone . AO, AAO, EHFS, and EHFN regulate the meridional ozone transport into high latitudes from autumn through spring and accumulate ozone over winter/spring at high latitudes. They also regulate stratospheric temperatures in the polar region and, therefore, control the size of the polar vortices and associated polar chemical ozone loss . The low/high bias in spring ozone persists into summer, even though ozone variability decreases from spring to summer. By accumulating these proxies over the dynamically active season and weighting them by the monthly ozone variability, the regression model better captures the integrated ozone response. Extreme values in the timeseries are better captured by the regression model with these additional proxies . Figure shows the regression coefficients of the proxy terms obtained from the MLR fit to SWOOSH data. The latitude-altitude distribution of the regression coefficients is as expected: the dominant lower stratospheric influence from the HTHH volcanic eruption (HTHH proxy) reducing ozone in the SH e.g. and higher absolute values of EHFS and EHFN increasing ozone at high middle latitudes, as well as the strong influence of AO in the NH.

In this study, the aerosol proxy is separated into three distinct time periods corresponding to the El Chichón period (1979–1989), the Pinatubo period (1990–1999), and the post-2000 period (2000–2024), which includes a series of medium volcanic eruptions as well as the HTHH eruption . Volcanic eruptions have different impacts on stratospheric circulation, chemistry, and ozone distribution, depending on magnitude, season, and latitude of the event . This variable response is illustrated in panels i–k of Fig. 6. The Australian wildfire (end of 2019) and HTHH had a stronger initial impact in the SH than in the NH because the volcanic aerosols (and stratospheric water vapour) were transported mostly into the SH resulting in an ozone loss (band of negative values in panel i). The Pinatubo eruption had a stronger impact in the NH (boreal winter) as illustrated by the band of negative coefficients from the tropics to the polar region (panel j, ). The dynamic response to the El Chichón eruption showed the opposite effect. Due to changes in the SH circulation the dynamical ozone increase exceeded the chemical ozone loss caused by the eruption (see panel k, ). However, nearly all El Chichon coefficients are not statistically significant at 95 % confidence level. Because of very different ozone responses the major eruptions of El Chichón and Mount Pinatubo are treated separately in the regression. With the 2022 eruption of Hunga Tonga–Hunga Ha’apai (HTHH), a third strong volcanic event occurred that clearly affected stratospheric dynamics and composition . Multiple SAOD proxies are also used because the eruptions of El Chichón and, in particular, Pinatubo produce substantially stronger and longer-lasting perturbations to ozone than later eruptions and thus dominate the regression if a single aerosol proxy is used. Treating these periods separately allows the regression model to better represent the variable volcanic forcing .

The aerosol data are taken from the GLOSSAC climatologies . The solar proxy is represented by the Mg II index time series . The Quasi-Biennial Oscillation (QBO) is described by the first two principal components from a PCA analysis of mean ERA5 stratospheric winds between -6 and 6° latitudes from 70 to 10 hPa. ENSO is described by the ENSO 3,4 index time series. In the past, the El Niño–Southern Oscillation (ENSO) index (MEI V2, ) was used, but we suspect a possible drift starting in the early 2020s, making the index during the recent strong El Niño in 2023/24 too weak (Fig. ). Additional seasonal components are added to the solar, QBO, and ENSO proxies as mentioned above.

We first evaluate the total ozone datasets (Table ) for trends after 2000. The trends as a function of latitude (here in 5° steps), shown in Fig. , are mostly positive for all latitudes from 2000 to 2024. The blue lines show the results from our full setup, i.e., with dynamical proxies and three AOD, and the orange lines show the results for the so-called LOTUS setup, i.e., no dynamical proxies and single AOD. Most of the total ozone datasets agree very well. In the NH, total ozone recovers by +1 to +1.5 DU per decade (+0.25 % per decade to +0.5 % per decade). In the SH, the trends polewards of 40° increase on average to +3 to +4 DU per decade (+1.3 % per decade to +1.5 % per decade) and become significant when using the full proxy setup. With the LOTUS settings, SH trends are lower remaining at about +1.5 DU per decade and insignificant above 45° S (see also Fig.  for the trends in % per decade). The difference between the two proxy setups is the strongest in the SH above 35° while good agreement is seen elsewhere. The hemispheric asymmetry in the ozone trends with the full setup is in line with observations of the hydrogen chloride (HCl) asymmetry. In the SH, HCl has declined faster due to a faster circulation, fostering in turn a faster increase in ozone compared to the NH . Results shown in Fig.  are very similar to those from , who used a very similar proxy setup.

To assess differences between the LOTUS model and the full regression model as well as the impact from using the third HTHH SAOD proxy, we look at selected modelled and observed total ozone timeseries. The MLR was applied to the September 60–90° S (ozone hole season) and the 35–60° S annual mean of the six datasets, the results are shown in Fig. . A slightly different MLR setup is used here for a more direct comparison to previous analysis by : no gap terms, and in the polar region the pre-ODS period ends in 1999 and post-ODS peak period starts in 2000 (in the extrapolar region 1995 and 1996, respectively) . These results can be directly compared to results from . For the southern mid-latitudes, report an annual trend of 0.7 % per decade for the period 1996–2020 applied to the median timeseries from five total ozone datasets. Extending the analysis to 2024, we obtain trends of 0.6 % per decade (panel c) and 0.4 % per decade with the LOTUS model (panel a). The inclusion of additional dynamical proxies improves the representation of short-term variability, particularly for the local ozone minima in 2020–2022.

A higher difference between the MLR setups can be seen in the southern polar region, where ozone variability and chemical ozone depletion are strongly influenced by the dynamical processes. For September mean ozone between 60 and 90° S, report a trend of 12.0 % per decade for 2000–2020. In contrast, we derive trends of 3.0 % per decade (LOTUS, panel b) and 8.2 % per decade (full regression), panel (d). Despite the low SH polar ozone in the early 2020s, the full regression trend, interpreted as the smooth ozone recovery related to ODS changes, remains stable within uncertainties after adding four additional years. The impact of dynamical proxies is most pronounced in the polar region (panel d). It is imperative to note, that the LOTUS setup (panel b) was developed for extra-polar latitudes and thus is not suited for polar regions, which is why it only serves as an illustration here.

To demonstrate the impact of the third SAOD (“HTHH”) proxy, we look at the September trends in the SH polar region using different SAOD setups (panel d to f). Polar ozone recovery was first identified in the early phase of the ozone hole season (September) e.g.. The first setup (panel e) uses a single AOD proxy for the complete time period like in the LOTUS setup, the second (panel f) uses two SAOD proxies, one until 1989 (“El Chichon”) and the other from 1990 to 2024 (“Pinatubo”) , and the last (panel d) uses three SAOD proxies by adding the period 2000–2024 (“HTHH”) and limiting the second SAOD to 1990–1999. The difference between the first (e) and the second (f) setup is small, with trends after 2000 being around +3.5 % per decade in both cases. The addition of the third (d) SAOD (“HTHH”) more than doubles the post-2000 trend to 8.2 % per decade, making it very similar to +12.0 % per decade until 2020 from . These results show the strong sensitivity of the regression and the influence of the dynamical proxies in the southern polar region. Volcanic eruptions lead to perturbations in ozone that usually last for a few years . In the early 2020s, a series of high-stratospheric aerosol events (Australian wildfires in 2019/2020, the 2021 Le Soufrière, and 2022 HTHH volcanic eruption) was responsible for the above-average size ozone holes between 2020 and 2023 . The HTHH-SAOD proxy allows a better separation of these perturbations from the slow (background) decreases in ODS responsible for ozone recovery. The effect from the SAOD-HTHH proxy is however negligible for extrapolar total ozone, but the MLR pattern of the regression coefficient of the HTHH proxy shown in panel (i) of Fig. seems physically reasonable. In summary, the use of the dynamical and SAOD-HTHH proxies in the MLR as well as the comparisons with the earlier results for the period ending in 2020 prove that ozone recovery is ongoing despite the ozone perturbations in the early 2020s in response to the series of volcanic and wildfire events.

Figure  shows ozone profile trends as a function of latitude and altitude for the period 2000 to 2024 using the same setup as for total ozone (Eq. ). Generally, there is good agreement between the datasets. Above 30 km the trends are below +0.1 DU km⁻¹ per decade as ozone concentration decreases with altitude. Below 30 km, the ozone trends increase up to 0.5 DU km-1 per decade and contribute substantially to the total ozone trends. Negative trends with high uncertainties are mostly visible below 20 km, close to the tropopause. In the UTLS region, uncertainties in limb ozone measurements are higher and likely responsible for higher trend uncertainties .

Ozone profile trends in units of % per decade are shown in Fig.  of the Supplement. The trends observed in our study (Fig. ) mostly agree with the results from , but there are some differences. The positive trends in the upper stratosphere are consistent between both studies (about +1 % per decade to +3 % per decade). A belt of negative trends just above the tropopause appears to be more persistent from 60° S to 60° N in Fig. 1 of and negative trends in the UTLS appear to be less negative in our case. This is also seen when comparing to a MLR setup without the AO, AAO, EHFS, and EHFN and with a single SAOD proxy (LOTUS setup), as shown in Figs.  and .

In the lower to middle stratosphere (above 20 km altitude), a hemispheric asymmetry in trends is noticeable, with higher trends in the SH, a pattern similar to what is observed for total ozone in Fig. . This asymmetry is also seen when looking at trends in three broad latitude bands. Figure shows the ozone profile trends for the zonal bands, 35–60° S (SH), 20° S–20° N (tropics) and 35–60° N (NH). Below 30 km altitude, trends are smaller in the NH compared to the SH. There is overall good agreement in the broad band trends between the datasets. All datasets agree on a positive trend above 30 km in all three latitude bands, maximizing at between 1 % per decade to 3 % per decade near 43 km altitude (Fig. ). The SH displays larger variations in trends between datasets in this altitude region. The tropics show negative trends below 25 km for all datasets. In the extratropics trends below 20 km can be both positive and negative depending on the dataset. Below about 20 km altitude, the uncertainty in the mean trend increases.

The broad zonal band trends (Fig. ) show the band trends, similar to . The corresponding representation in % per decade (Fig. ) can be compared with those shown in Fig. 4 of , which were derived using the LOTUS setup and cover the same time period up to 2024. Upper stratospheric trends agree well between both studies. Differences are visible in the lower stratosphere. As mentioned before, the lowermost stratospheric mean trends are negative in the extratropics in , while in our study they are positive in the SH and on average closer to zero in the NH (see also Figs.  and ). show consistently positive trends in the upper stratosphere, while finding mixed and negative trends in the middle and lower stratosphere from ground-based measurements

The larger variations in SH trend profiles for altitudes above 25 km, seen in panel a) of Fig.  (see also Figs. –) is due to the smaller positive trends seen in GOZCARDS compared to all other merged datasets. Applying the MLR to deseasonalized ozone data (rather than absolute values), GOZCARDS trends in the SH are getting closer to the trends of the other data, as shown in .

The differences between the LOTUS uncertainty estimate (shaded area) and the joint-distribution approach (dashed lines) are largest at lower altitudes, where the trends from the individual datasets show higher differences. This is particularly evident in the mid-latitudes around 17 km, most likely driven by both higher trend differences and differences in the underlying dataset uncertainties. While the LOTUS uncertainty estimate generally envelops the individual trends, with some exceptions for datasets exhibiting pronounced trends at specific altitudes, the joint-distribution approach spans a wider range, as it characterizes the total ensemble variance.

We calculate the stratospheric columns by integrating the ozone profiles upward from the ERA5 climatological thermal (lapse rate) tropopause. Another method to calculate the stratospheric ozone column trends is to integrate the ozone layer trends shown for example in Fig. . A comparison between the two methods (Fig. ) show that the trends from both methods are generally very similar, however, the uncertainty of the vertically integrated layer trends are smaller than for the trend derived from stratospheric column trends. The tropopause height is determined at the center of each latitude band using interpolation. The MLR is then applied to the stratospheric columns. The difference between total column and stratospheric column trends can then be attributed to the tropospheric contribution to total ozone trends.

Figure  shows as an example the ozone timeseries in the broad zonal band 35–60° S for stratospheric ozone and total ozone columns of all datasets. The stratospheric ozone column was bias-corrected by subtracting the mean during the 1998–2008 period from each dataset (anomaly) and adding back the multi-instrument mean to obtain bias-corrected absolute ozone values for each dataset. The difference between the total and stratospheric columns of about 20 to 30 DU yields the zonal mean tropospheric column. There is a gap in the stratospheric column timeseries after the eruption of the Pinatubo volcano in 1992, because of filtering of columns that did not reach the tropopause. This is mainly due to retrieval artifacts in the presence of layers of high aerosol extinction obstructing the limb view in the lowermost stratosphere. The total ozone datasets seem less affected. Due to the weaker eruption of HTHH in 2022 in terms of stratospheric aerosol load, this impact on stratospheric column calculation is also reduced except for SAGE-SCIA-OMPS. For this dataset, observations from February 2022 to May 2023 were excluded.

Figure shows the results of the stratospheric column trends as a function of latitude which can be compared to the total ozone trends (Fig.  ). The tropics and the NH show very similar results with low ozone trends in the tropics (+0.0 to +1 DU per decade) and slightly higher ozone trends above 40° N (up to 2 DU per decade) for both proxy setups. The differences which we saw in the SH total ozone can be seen here as well. Trends using the full setup are larger (up to +3.5 DU per decade) than using the LOTUS setup (+2 DU per decade). Both proxy setups yield similar trends in the SH for stratospheric and total columns, but the stratospheric column trends show a much higher spread between the trends of the different datasets compared to total ozone.

In Fig. 3b of combined latitude-dependent stratospheric column trends are shown and can be compared to our results. Above 40° latitude in each hemisphere stratospheric column trends using the LOTUS setup are slightly higher (around 2 DU per decade) in our analyses compared to (around 1 DU per decade). This difference likely arises from the comparatively strong trends in the OSIRIS dataset, which exert a larger influence on the multi-dataset mean in our six-dataset ensemble than in the eight-dataset mean used by . Below 40° the stratospheric column trends agree.

Figure  shows trends for all datasets from 2000 to 2024 in the three broad zonal bands. The bars represent the stratospheric ozone column (light blue) and total ozone column trends (orange) of the individual merged datasets. The last two bars show the mean stratospheric ozone column (dark blue) and the mean total ozone column (brown).

The agreement between stratospheric and total column trends is remarkably close across all latitude bands (panels a to c). This suggests that the contribution of tropospheric ozone to the total column trends is small and well below the trend uncertainties. Significant positive stratospheric and total column trends are visible in the SH (about +2 to +4 DU per decade). On average, smaller trends (about +1 DU per decade) are seen in the tropics and the NH in agreement with Figs.  and . The two merged ozone profile datasets that include OSIRIS (SAGEII-OSIRIS-OMPS and SAGEII-OSIRIS-SAGEIII) show larger trends in comparison to the other ozone profile datasets in the extratropics. Both profile datasets that contain Aura/MLS at the end of their data record (GOZCARDS and SWOOSH) show, not unexpectedly, nearly the same trends.

The bottom row of Fig.  (panels d to f) shows our MLR results using the LOTUS setup. The mean profile trends are about +1 DU per decade in the SH and are smaller than with the full proxy setup except for the two merged datasets including OSIRIS and are mostly positive in the other two bands, similar to the full proxy trends.

The ground-based dataset WOUDC total ozone trends are consistent with the trends of the satellite-based datasets. Although WOUDC total ozone values are occasionally slightly higher than the satellite-based datasets (Fig. ). The LOTUS and the J-distribution uncertainty estimates are showing similar results for total ozone, while the J-distribution uncertainty is mostly larger than the LOTUS uncertainty for the mean stratospheric column trends .