The ozone layer protects life on Earth from harmful ultraviolet solar radiation and plays an important role in the radiation budget of the atmosphere. Its evolution is intimately coupled to climate change. Starting in the 1970s, the stratospheric ozone declined worldwide, with the largest decline of 4 to 8 % decade-1 seen in the upper stratosphere (WMO, 2014; Steinbrecht et al., 2017, and references therein). After international agreements on the reduction of ozone-depleting substances (Montreal Protocol and its Amendments), the first signatures of ozone recovery were observed, at least in certain latitude and altitude regions (e.g., Newchurch et al., 2003; Kyrölä et al., 2013; Bourassa et al., 2014; Eckert et al., 2014; Harris et al., 2015; Tummon et al., 2015; WMO, 2014). About half of the recent increase in upper stratospheric ozone is attributed to climate change (WMO, 2014). As ozone recovery will remain strongly influenced by climate change, continued monitoring of stratospheric ozone and its vertical structure is important.

The European Space Agency (ESA) Ozone Climate Change Initiative (Ozone_cci) aims to generate new high-quality satellite datasets that are essential to assess the fate of atmospheric ozone and better understand its link with anthropogenic activities (http://www.esa-ozone-cci.org). Participating in the Ozone_cci project are three instruments on board Envisat, GOMOS (Global Ozone Monitoring by Occultation of Stars), MIPAS (Michelson Interferometer for Passive Atmospheric Sounding) and SCIAMACHY (SCanning Imaging Spectrometer for Atmospheric CHartographY), as well as OSIRIS (Optical Spectrograph and InfraRed Imaging System) and SMR (Sub-Millimeter Radiometer), both aboard Odin, and ACE-FTS (Atmospheric Chemistry Experiment Fourier Transform Spectrometer) on SCISAT.

Satellite data provide good spatial coverage, but the temporal coverage of their data records is usually too short for trend analyses. For reliable estimates of ozone trends, long-term data records are needed in order to separate natural ozone variability (e.g., due to solar activity) and long-term trends. For the assessment of ozone trends, several merged ozone datasets with a good vertical resolution have been created (Harris et al., 2015; Tummon et al., 2015): GOZCARDS (Froidevaux et al., 2015), SWOOSH (Davis et al., 2016), SAGE II–OSIRIS (Bourassa et al., 2014) and SAGE II–GOMOS (Kyrölä et al., 2013). In construction of these datasets, different merging approaches are used. GOZCARDS and SWOOSH provide ozone mixing ratios on pressure levels, while SAGE II–OSIRIS and SAGE II–GOMOS datasets provide number density on a geometric altitude grid as used in the respective data retrievals.

This paper introduces a new dataset in which reliable satellite data providing ozone profiles on an altitude grid are merged into a climate data record for assessment of ozone trends. These are from the five Ozone_cci instruments, GOMOS, MIPAS, SCIAMACHY, OSIRIS and ACE-FTS, covering the period 2001–2016, which are merged with the data from SAGE II (Stratospheric Aerosol and Gases Experiment II, 1984–2005) and OMPS-LP (Ozone Monitor Profiling Suite-Limb Profiler, 2012–2016). Since the last ozone assessment (WMO, 2014), the Ozone_cci satellite data were processed with new retrieval versions. The stability of the individual instrument data records has been extensively studied; only sufficiently stable data are used for the merged dataset. The merging is performed on deseasonalized anomalies computed from each individual dataset. This method is often used for creating long-term data records (e.g., IPCC, 2013; WMO, 2014), as well as for trend analyses of ozone (e.g., Bourassa et al., 2014; Randel and Thompson, 2011; Sioris et al., 2014; Steinbrecht et al., 2017), stratospheric temperature (Randel et al., 2009; Seidel et al., 2011; Thompson et al., 2012) and water vapor (Jones et al., 2009). The main advantage of using deseasonalized anomalies is that biases due to different sampling patterns (including the difference in local time) and instrumental biases are automatically removed, if the sampling patterns do not change over time. While assessing the trends using deseasonalized anomalies, there is no need to fit the seasonal variations with harmonic functions (the seasonal variations do not necessarily allow a simple expansion into a few harmonics).

The paper is organized as follows. Section 2 describes the ozone datasets from the individual instruments. Section 3 is dedicated to comparisons of the individual datasets and evaluation of the deseasonalized anomalies. In Sect. 4 we describe the merging method and associated uncertainties. Section 5 is dedicated to evaluation of ozone trends in the stratosphere using the merged SAGE–CCI–OMPS dataset. The information about data availability is provided in Sect. 6 and the conclusions are summarized in Sect. 7.

Ozone measurements available from different satellite instruments cover all seasons and various times of day and also have good latitudinal coverage. The information about individual datasets is collected in Table 1. All of the data used for creating the merged dataset have a sufficiently good vertical resolution of 1–3 km in the stratosphere and in the UTLS (upper troposphere and the lower stratosphere). For all instruments used here, ozone profiles are retrieved on the geometric altitude grid. The majority of the datasets – SAGE II, GOMOS, OSIRIS, SCIAMACHY and OMPS – provide number density ozone profiles; therefore this representation is adopted for the merged dataset. For ACE-FTS and MIPAS, the retrievals are in volume mixing ratio on an altitude grid. Conversion to number density profiles is performed using temperature profiles retrieved by these instruments, thus providing consistent (i.e., without using external information about temperature and pressure profiles) representation of number density ozone profiles. Since the publication of the WMO 2014 ozone assessment, new processing versions of the Ozone_cci datasets were introduced. Generally, the newly reprocessed ozone datasets have smaller biases and are more stable (further details are provided in the descriptions of individual datasets below).

The time series of the number of available ozone profiles per month from each instrument is shown in Fig. 1. Note that for some instruments, the selected time period is shorter than the full operation period (Table 1). The individual datasets have been compared with each other and with ground-based data and only time periods when the instruments were operating optimally are selected for merging (details are given later in Sect. 3).

The Ozone_cci datasets, which are used for merging, are included in the user-friendly HARMonized dataset of OZone profiles (HARMOZ) (Sofieva et al., 2013) and are available at the Ozone_cci web-page (http://www.esa-ozone-cci.org/?q=node/160). HARMOZ consists of the original retrieved ozone profiles from each instrument, which are screened for invalid data by the instrument experts and are presented on a vertical grid and in a common netCDF4 format, which simplifies the data usage. Below are more detailed descriptions of the individual datasets.

For creating monthly zonal mean data from the individual instruments, 10∘ latitude bands from 90∘ S to 90∘ N are used. For all sensors, the monthly zonal average is computed as the mean of ozone profiles xk(z): ρ(z)=1N∑xk(z), where N is the number of measurements (N>10). The uncertainty of the monthly mean σρ2 can be estimated as the standard error of the mean: σρ2=s2N, where s2=(xk-ρ)2 is the sample variance. Equation (2) is valid for random samples of uncorrelated data. As shown by Toohey and von Clarmann (2013), some deviations of the real standard error of the mean from that calculated using Eq. (2) can be observed for satellite observations. In our study, Eq. (2) is used as an approximate estimate of the standard error of the mean, since no estimates considering the impact of the correlations caused by the orbital sampling are currently available. In Eq. (2), we used a robust estimator for the sample variance: s=0.5⋅(P84-P16), where P84 and P16 are the 84th and 16th percentiles of the distribution, respectively. Monthly zonal mean, sample variability s and standard error of the mean from 15 to 50 km altitude in January 2008 are shown for the Ozone_cci instruments in Fig. 2. The ozone distributions shown in Fig. 2 are very similar for all datasets. Due to the large number of data available for averaging, the standard error of the mean is usually less than 1 % in the stratosphere.

Satellite measurements sample a continuous ozone field at some locations and times. To characterize the nonuniformity of sampling, we computed inhomogeneity measures in latitude, Hlat, and in time, Htime (Sofieva et al., 2014a). Each inhomogeneity measure H is the linear combination of two classical inhomogeneity measures, asymmetry A and entropy E (for definition of these parameters, see Sofieva et al., 2014a): H=12A+(1-E). The inhomogeneity measure H ranges from 0 to 1 (the more homogeneous, the smaller H). For dense samplers (MIPAS, SCIAMACHY, OMPS), the inhomogeneity is close to zero for nearly all latitude bins. For other instruments, the inhomogeneity measure can be large for some latitude–time bins. The monthly zonal mean data with very high inhomogeneity are excluded from data merging (see below). The monthly zonal mean data from the individual instruments are available at the Ozone_cci web-page (www.esa-ozone-cci.org).

For each instrument, latitude band and altitude level, the deseasonalized anomalies are computed as follows: Δ(ti)=ρ(ti)-ρmρm, where ρ(ti) is the monthly mean value at a certain altitude and latitude band corresponding to time ti and ρm is the mean value for the corresponding month m, i.e. ρm=1Nm∑j=1Nmρj, Nm being the number of monthly mean values ρj in a given month m available from all years.

For the Ozone_cci instruments, the seasonal cycle is evaluated using the overlapping period 2005–2011. The seasonal cycle for SAGE II is computed using years 1985–2004 and for OMPS using the years 2012–2016. In computation of deseasonalized anomalies, we ignored data from those latitude–time bins with the mean inhomogeneity Htot=0.5(Hlat+Htime)>0.8 (for all considered satellite instruments, longitudinal inhomogeneity is negligible, Sofieva et al., 2014a).

Figure 3 shows examples of the seasonal cycles evaluated for the instruments considered. For instruments with coarse temporal and horizontal sampling, GOMOS and ACE-FTS, the seasonal cycle is evaluated less reliably. There are biases between instruments, which are in agreement with earlier validation and intercomparison results (Hubert et al., 2016; Rahpoe et al., 2015; Sofieva et al., 2017). Since the seasonal cycle for SAGE II and OMPS is evaluated using the periods different for that used for the Ozone_cci instruments, we added a corrected seasonal cycle (dashed lines in Fig. 3):

ρSAGE, corr(m)=ρSAGE(m)⋅ρSAGE, 2002–2004ρSAGE, 1985–2004⋅ρOSIRIS, 2005–2011ρOSIRIS, 2002–2004andρOMPS, corr(m)=ρOMPS(m)⋅ρOSIRIS, 2005–2011ρOSIRIS, 2012–2016, where ρinstrument, time_period is the average ozone for the indicated instrument and time period. (Note that this correction is not used in the merging algorithm, it is applied only for illustration in Fig. 3.) The small overall biases with respect to SAGE II data are observed for OSIRIS and GOMOS (except in tropics at 20 km). Very good agreement between these instruments is also reported in several previous studies (Adams et al., 2013, 2014; Hubert et al., 2016; Kyrölä et al., 2013).

The uncertainty of the seasonal cycle value σm for each month m is evaluated from uncertainties of individual monthly mean values σρ,j: σm2=1Nm2∑j=1Nmσρ,j2, which describes the propagation of the random uncertainties to the mean value. We would like to note that there is a significant interannual variability of ozone due to geophysical processes (which can be characterized through a multiple linear regression with different proxies). These geophysical variations between different monthly values for a particular month are not characterized by Eq. (5).

The amplitude and phase of the seasonal cycle is very similar for SAGE II, MIPAS, SCIAMACHY, OSIRIS and OMPS, as illustrated in Fig. 4, which shows the normalized seasonal cycle. The SCIAMACHY seasonal cycle is slightly different at 45 km at northern midlatitudes. This is caused by the altitude interpolation error, which is especially crucial at midlatitudes in the upper stratosphere due to changing from semiannual (above this altitude) to annual (below this altitude) variation regime. Due to a coarse altitude retrieval grid, SCIAMACHY data are more sensitive to the selected type of the interpolation than other instruments. When using a log-linear interpolation, very similar results as for MIPAS are obtained in the upper stratosphere. SAGE II seasonal cycle in the middle and upper stratosphere in the latitude band 40–60∘ S has significantly lower amplitude than those of other instruments, especially at 45 km, probably because of the sampling.

After the removal of the seasonal cycle, the SAGE II deseasonalized anomalies are offset to the Ozone_cci mean anomalies in the years 2002–2005. The OMPS deseasonalized anomalies are offset to the mean Ozone_cci anomalies (which are based on OSIRIS and ACE-FTS measurements in this period) in the years 2012–2016. As a result, small offsets in SAGE II and OMPS anomalies due to the different reference time periods for evaluation of the seasonal cycle are removed, and the anomalies from all instruments are aligned.

Before merging, the deseasonalized anomalies of the individual instruments have been extensively intercompared with each other by computing and visualizing the time series of difference of individual anomalies from the median anomaly. This method turns out to be a sensitive method for detecting an unusual time series behavior of the individual data records. In particular, it was found that SCIAMACHY anomalies are larger in the beginning of the mission, for nearly all latitude bands and many altitude levels (see Fig. S3 in the Supplement, which shows the deviations of the SCIAMACHY deseasonalized anomalies from the median deseasonalized anomalies of SAGE II, GOMOS, MIPAS, SCIAMACHY, OSIRIS, ACE-FTS and OMPS). This might be attributed to possible pointing problems in the beginning of the mission; therefore we decided not to use the SCIAMACHY data before August 2003 in the merged dataset. Similarly, OMPS anomalies are lower in the first 3 months of the mission (see Fig. S6 in the Supplement); this might be related to relatively coarse sampling of OMPS in the first 3 months of the mission and possible problems with pointing. Therefore, OMPS data were included in the merged dataset starting from April 2012, when the instrument operated in its full capacity.

After the data selection, the anomalies from individual instruments are found to be in good agreement with each other. This is illustrated in Fig. 5, which shows the deviations of deseasonalized anomalies from each instrument relative to the median anomaly for latitudes 30–40∘ S. The deviations from the median anomalies are small, less than 5 % for the majority of data, and do not have statistically significant drifts with respect to the median anomaly (see also illustrations in the Supplement).

We computed the merged anomaly as the median of the individual instruments anomalies, for each altitude level z and for each latitude band θ and month t: Δmerged(θ,z,t)=median(Δk(θ,z,t)), where Δk indicates the individual instrument anomaly. Figure 6 illustrates the data merging: the upper panel shows the monthly zonal mean data, while the bottom panel shows individual anomalies and the merged (median) anomaly.

As observed in Fig. 6, the biases between the individual data records are removed by computing the deseasonalized anomalies. In the merging, we filtered out individual anomaly values (locally for each latitude band and altitude level), which differ from the median anomaly more than 10 % at latitudes 40∘ S–40∘ N and more than 20 % in other latitude bands. This filtering does not affect the absolute majority of cases; it removes only a few exceptional anomalies from GOMOS and ACE-FTS, which are due to lower sampling.

An additional illustration of the very good agreement between the individual deseasonalized anomalies and the merged anomaly is presented in Fig. 7, which shows the correlation coefficient between the merged anomalies and individual anomalies using the years 2001–2016. The correlation coefficient is above 0.9 for all instruments in most of the latitude bands and altitude levels.

Examples of merged deseasonalized anomalies for several latitude bands are shown in Fig. 8. In the upper stratosphere at midlatitudes a decrease from 1984 to 1995–1999 is observed, and then a gradual increase to the present. In the tropics, quasi-biennial oscillation (QBO) is observed.

The uncertainty of individual deseasonalized normalized anomalies σΔi (for each month and each latitude–altitude bin) can be estimated as σΔi=1ρmσρ,i2+σm2, where σρ,i is the uncertainty of the monthly zonal mean value from Eq. (2), and σm is uncertainty of the seasonal cycle from Eq. (5). We estimated the uncertainties of the merged deseasonalized anomalies (which correspond to median value) as σΔ,merged=minσΔ,jmed,1N∑j=1NσΔ,j2+1N2∑j=1NΔj-Δmerged2, where σΔ,jmed is the uncertainty of the anomaly of the instrument corresponding to the median value. Equation (7) can be interpreted as follows (see also Fig. A1 in Appendix). If individual anomalies are significantly different, i.e., the corresponding error bars do not intersect (Fig. A1, left), the uncertainty of the merged anomaly is the uncertainty corresponding to the median value. In case several instruments report a similar anomaly (intersecting error bars), this provides more confidence of this anomaly value, and the resulting uncertainty of the merged anomaly is approximated as 1N∑j=1NσΔ,j2+1N2∑j=1NΔj-Δmerged2. For example, in the case of three coinciding anomalies with the same uncertainties σ, the uncertainty of the merged value will be σ/3. For the considered datasets, anomalies are usually very close to each other, so that several values are within the Δmerged±σΔ,merged interval, as illustrated by Fig. S8 of the Supplement.

The uncertainty of the merged dataset is illustrated in Fig. 9 for the same latitude bands as shown in Fig. 8. As expected, the uncertainties in the time period when only SAGE II data were available are larger than uncertainties for time periods when several instruments have contributed. The average uncertainty is usually less than 4 % before 2001 and below 1 % for the years 2002–2017. In the UTLS, uncertainties are larger than in the stratosphere and are in the range of 3–9 %. At midlatitudes, uncertainties are larger in winter than in summer due to larger ozone variability during winter; this is observed clearly in the period before 2001.

The merged deseasonalized anomalies can be used directly for analyses of trends in the vertical distribution of ozone. For this purpose, a multivariate regression has been applied to the merged SAGE–CCI–OMPS data: O3(t)=PWLT(t,t0)+q1QBO30(t)+q2QBO50(t)+sF10.7(t)+dENSO(t), where PWLT(t, t0) is a piecewise linear term (constant and a hockey-stick trend with the turnaround point in 1997), QBO30(t) and QBO50(t) are the equatorial winds at 30 and 50 hPa, respectively (http://www.cpc.ncep.noaa.gov/data/indices/), F10.7(t) is the monthly average solar 10.7 cm radio flux (ftp://ftp.geolab.nrcan.gc.ca/data/solar_flux/monthly_averages/), and ENSO(t) is the 2-month lagged ENSO proxy (http://www.esrl.noaa.gov/psd/enso/mei/table.html). As shown by Kyrölä et al. (2013) and Laine et al. (2014), the best estimate of the turnaround point is in 1997 for the majority of latitude bands and altitude levels. The sensitivity of regression results to the choice of turnaround point is discussed in Harris et al. (2015). Autocorrelations are removed using the Cochrane–Orcutt transformation (Cochrane and Orcutt, 1949).

Although uncertainties for the merged data are evaluated, they are not used in the regression analysis: different amounts of data available over time result in varying uncertainties over time (e.g., as shown in Fig. 9), which might improperly weight the time series. In our regression, all data points are considered with equal weights, and the uncertainty of the fitted parameters is estimated from the regression residuals.

The regression is performed for each latitude band and for each altitude level independently. The regression is similar to that used in Kyrölä et al. (2013). There are also other methods for evaluation of trends, e.g., using dynamical linear modeling (Laine et al., 2014). Since the main focus of our paper is generation of the merged SAGE–CCI–OMPS dataset, we selected a rather “standard” regression model (Eq. 9). A short discussion on the sensitivity of regression results to the choice of the regression model is presented below.

The linear ozone trends before 1997 and after 1997 are shown in Fig. 10. Shaded areas show regions where trends are statistically different from zero at the 95 % level. In the period 1984–1997, statistically significant negative trends from -4 to -8 %dec-1 are observed in the upper stratosphere. In the period 1997–2016, the ozone trends in the upper stratosphere are ∼+2 %dec-1, and they are statistically different from zero at the 95 % level in the extratropics. The ozone trends in large latitude bands, analogous to those studied in WMO (2014), Harris et al. (2015) and Steinbrecht et al. (2017), are shown in Fig. 11 (the anomalies in large latitude bands are created from 10∘ latitude anomalies). The estimated ozone trends are consistent with previous results (Bourassa et al., 2014; Harris et al., 2015; Kyrölä et al., 2013; Tummon et al., 2015; WMO, 2014) and indicate that the ozone recovery has started. A slightly different ozone trend analysis using the SAGE–CCI–OMPS merged dataset by Steinbrecht et al. (2017) has resulted in nearly identical ozone trends (Fig. 3 in Steinbrecht et al., 2017), which indicates only a weak dependence of ozone trend estimates on the regression method (and collection of proxies).

To study the sensitivity of ozone trend results to filtering of suspicious data, we created a version of the merged dataset in the same way as above, but keeping the early periods of SCIAMACHY and OMPS operations, and performed the same analysis. Keeping all data results in very minor changes in ozone trends after 1997, as illustrated by Fig. S9 in the Supplement, the changes are mostly less than 0.3 %dec-1. This is not unexpected, because the median estimate used in data merging is insensitive to outliers.

Although conversion of mixing ratio to number density for MIPAS and ACE-FTS is performed using the retrieved temperature profiles, this might introduce a minor inconsistency. To evaluate a potential effect on ozone trends, we created a version of the merged dataset, but without MIPAS and ACE-FTS data. This merged dataset contains only data that were retrieved in number density on a geometric altitude grid. Minor changes (<0.3 %dec-1 at latitudes 40∘ S–40∘ N, up to ∼1 %dec-1 at 40–60∘ S and 40–60∘ N) in ozone trends after 1997 are observed (Fig. S10 in the Supplement). The effect is rather small because MIPAS operated in 2005–2012, when other datasets are also available, and the coverage by ACE-FTS is limited. Furthermore, anomalies from individual datasets are in good agreement with each other, and therefore significant changes in evaluated ozone trends are not expected.

The satellite data quality degrades in the UTLS. The merging principle seems to be optimal also for the UTLS, as it automatically removes biases, which can be significant in this region. The trends in the UTLS estimated using the merged SAGE–CCI–OMPS data follow the expected trend (declining ozone in the tropics just above the tropopause due to intensification of the Brewer–Dobson circulation). This is also in agreement with the dedicated studies on ozone trends in the tropical UTLS (e.g., Sioris et al., 2014).

We have presented the merged dataset of ozone profiles from several satellite instruments: SAGE II on ERBS, GOMOS, SCIAMACHY and MIPAS on Envisat, OSIRIS on Odin, ACE-FTS on SCISAT, and OMPS on Suomi-NPP. The merged dataset has been created with the aim of analyzing ozone trends in the stratosphere. For the merged dataset, we used the most recent retrieval versions of the satellite datasets. The datasets from the individual instruments have been extensively validated and intercompared; only datasets which are in good agreement, and do not exhibit drifts with respect to collocated ground-based observations and with respect to each other, are used for merging.

The long-term SAGE–CCI–OMPS dataset is created by computation and merging of deseasonalized anomalies from individual instruments and associated uncertainties of the merged data are estimated. The merged SAGE–CCI–OMPS dataset consists of deseasonalized ozone anomalies in 10∘ latitude bands from 90∘ S to 90∘ N. The data are provided on an altitude grid from 10 to 50 km covering the period from October 1984 to July 2016.

Ozone trends in the stratosphere are evaluated by applying a multiple linear regression to the merged SAGE–CCI–OMPS dataset. Negative ozone trends in the upper stratosphere before 1997 and positive trends after 1997 are observed. The upper stratospheric trends in the extratropics are statistically significant and indicate onset of ozone recovery.