Understanding long-term trends in the mesosphere and lower thermosphere (MLT, ∼ 50–120 km) is becoming increasingly important as they can serve as a valuable indicator of climatic change. Trends in the MLT region exhibit significant variability, which can be influenced by solar activity, atmospheric dynamics, and anthropogenic factors (Laštovička, 2017, 2023, and references therein). Over the last few years, the upper atmosphere has been widely studied, particularly concerning long-term trends. While temperature remains the most extensively studied, other parameters (winds, water vapour, minor constituents) have been analysed as well (Laštovička, 2023; Cnossen et al., 2024, and reference therein).

The middle and upper atmosphere is experiencing a global mean negative temperature trend observed across datasets and models caused by the increasing concentration of CO2 (Laštovička, 2006; Qian et al., 2021; Cnossen, 2020). Increases in CH4 and H2O, along with decreasing stratospheric O₃, contribute to cooling in the stratosphere and mesosphere, but their influence becomes negligible in the upper thermosphere (Cnossen et al., 2024 and references therein). The mesospheric cooling trend has been observed in ground-based and satellite measurements with a magnitude of about 0.5–2 K per decade (Yuan et al., 2019; Bailey et al., 2021; Das, 2021; Li et al., 2021). Numerical simulations generally confirm the cooling trend except for the summer upper mesosphere, which shows near-zero or slightly positive temperature trends, likely reflecting dynamical effects (Qian et al., 2019; Solomon et al., 2019).

The long-term trends in dynamical processes, which are important in the MLT, are more complex. Observations and simulations reveal large spatial and temporal variability in wind trends, often differing between regions and seasons (Qian et al., 2019; Wilhelm et al., 2019). Research on atmospheric waves has been limited and concentrating mainly on tides (Laštovička, 2023). No significant long-term changes in diurnal tides have been observed, while semidiurnal components show mixed results that vary with altitude and latitude (Wilhelm et al., 2019). Overall, dynamical trends in the MLT are poorly constrained and regionally dependent, remaining a major source of uncertainty in understanding the MLT's long-term evolution (Laštovička, 2023 and references therein).

Mesospheric ozone plays an important role in atmospheric chemistry, dynamics, and energy balance, yet it remains one of the least-investigated parameters (Laštovička, 2023). Variations in its concentration can substantially influence the composition and behavior of the upper and middle atmosphere, thereby affecting the coupling between different atmospheric layers (Sinnhuber et al., 2012; Seppälä et al., 2009, 2025). In recent years, there has been a growing scientific interest in understanding solar-driven changes in ozone and temperature in the mesosphere and upper stratosphere (Salminen et al., 2019; Szeląg et al., 2022; Seppälä et al. 2024). These variations, although occurring alongside the rapid climate change driven by anthropogenic greenhouse gas emissions, have the potential to regulate regional climate on annual to decadal timescales (Baumgaertner et al., 2011; Langematz et al., 2005; Maliniemi et al., 2014; Rozanov et al., 2005, 2012; Seppälä et al., 2009, 2013). Understanding these complex interactions between human-induced and naturally driven climate variability is particularly important in the polar regions, which are experiencing some of the most significant changes.

While individual satellite instruments provide ozone measurements in the mesosphere, their operation is limited in time. To date, a comprehensive merged data set combining multiple sources of mesospheric ozone data is not available, and an assessment of global mesospheric ozone trends has not been conducted. Previous analyses (Bizuneh et al., 2022; Huang et al., 2014; Nath and Sridharan, 2014) based on 10–15 years of SABER (Sounding of the Atmosphere using Broadband Emission Radiometry) data were confined to narrow latitude ranges (5–15° N) or limited to lower and middle latitudes (48° S–48° N), thereby restricting insight into global long-term mesospheric ozone variability. Moreover, SABER ozone data are known to suffer from quality limitations (López-Puertas et al., 2023), underscoring the need for improved observational datasets and longer time series to establish reliable long-term trends.

In this paper, we introduce a new merged dataset of ozone profiles in the middle atmosphere, created using the data from several limb and occultation instruments: HALOE (Halogen Occultation Experiment), GOMOS (Global Ozone Monitoring by Occultation of Stars), MIPAS (Michelson Interferometer for Passive Atmospheric Sounding), ACE-FTS (Atmospheric Chemistry Experiment – Fourier Transform Spectrometer), MLS (Microwave Limb Sounder), and SOFIE (Solar Occultation For Ice Experiment). The paper is organized as follows. In Sect. 2, we provide short descriptions of the data used for the merged dataset. In Sect. 3, we describe the data preparation for the merging procedure, which is discussed in Sect. 4. Section 5 describes the new merged METEOR-O₃ dataset. Section 6 is dedicated to trends analyses in the upper atmosphere. The summary (Sect. 7) concludes the paper.

For the merged dataset of ozone mixing ratio profiles, we used profiles from the limb and occultation instruments that provide ozone data in the mesosphere and the lower thermosphere: HALOE, GOMOS, MIPAS, ACE-FTS, MLS, and SOFIE, as described below. For several instruments, the data are from the HARMonized dataset of Ozone profiles (HARMOZ) collection (Sofieva et al., 2013).

Since the data in the longest dataset with dense sampling, MLS, are retrieved on a pressure grid, pressure was selected as the vertical coordinate. HALOE ozone profiles are also retrieved on a pressure grid. MIPAS (both datasets), ACE-FTS and SOFIE ozone profiles are retrieved on a geometric altitude grid, but temperature and pressure profiles are also retrieved from the measurements, so the conversion to a pressure grid is straightforward. GOMOS provides ozone number density profiles on an altitude grid. Since the temperature and pressure profiles provided in the GOMOS files rely on a combination of ECMWF analyses and the MSIS90 model (Mass Spectrometer–Incoherent Scatter model; Hedin, 1983, 1991) and are therefore not very accurate, we used an altitude-pressure relationship derived from MIPAS_MA_UA measurements to convert the GOMOS profiles (registered in altitude) to a pressure grid. The GOMOS mixing ratios are computed consistently with the altitude-pressure-density conversion (see detailed description in the Supplement).

For creating monthly zonal-mean data from the individual datasets, 10° latitude bands from 80° S to 80° N are used. The specific challenge for combining ozone data in the mesosphere is that measurements are made at different local times, while diurnal variations are large in the MLT. Therefore, we first computed monthly zonal-mean ozone profiles for each illumination condition: daytime (solar zenith angle < 90°), nighttime (solar zenith angle > 108°), sunset or sunrise. For MLS and MIPAS make daytime and nighttime measurements, HALOE, ACE-FTS and SOFIE measure at sunset and sunrise.

For all datasets, the monthly zonal average is computed as the mean of ozone profiles xkt,z,θ,l, for each illumination condition l (daytime, nighttime, sunset, sunrise), month t, pressure level z and latitude band θ: 1ρt,z,θ,l=1Nt∑xkt,z,θ,l, where Nt is the number of measurements available in month t. We required Nt>10 in computation of monthly zonal-mean values. The uncertainty of the monthly mean σρ2t,z,θ,l can be estimated as the standard error of the mean: 2σρ2(t,z,θ,l)=s2(t,z,θ,l)Nt, where s2(t,z,θ,l) is the sample variance. In Eq. (2), we used a robust estimator for the sample variance, i.e., s=0.5(P84–P16), where P84 and P16 are the 84th and 16th percentiles of the distribution, respectively, similarly to the approach taken by Sofieva et al. (2017b).

To assess dataset consistency prior to merging, Figs. 2–3 demonstrate how well the various individual datasets reproduce the vertical structure and diurnal behaviour of ozone in the stratosphere and MLT. The monthly zonal-mean boreal wintertime (DJF) ozone distributions for MLS, MIPAS_NOM, MIPAS_MA_UA and GOMOS are presented in Fig. 2. The averaged values were calculated for the period 2005–2011 for these datasets. This period was chosen for intercomparison purposes, as it represents a time window with overlapping observations from MLS, MIPAS_NOM, MIPAS_MA_UA, and GOMOS. The top panel shows nighttime conditions, while the bottom panel represents daytime. Pressure altitudes are computed from pressure as z=16log⁡10(1013/P), where P is pressure in hPa. This pressure-derived altitude coordinate is used throughout the manuscript.

Across datasets, ozone distributions are similar, highlighting well-defined features. The nighttime ozone vertical distribution (top panels in Fig. 2) exhibits three distinct maxima. The primary ozone maximum occurs in the stratosphere, between approximately 30 and 35 km, where ozone mixing ratios reach about 10–11 ppmv (maximum in the tropics). It is produced through a photochemical equilibrium involving oxygen molecules, atomic oxygen, and solar ultraviolet radiation, and is modulated by natural variability and human activities (e.g. ozone-depleting substances, NO_x cycle; Brasseur and Solomon, 2005). The secondary maximum is found around 90–95 km, with the nighttime values as high as the stratospheric maxima (∼ 10 ppmv). The secondary ozone peak represents a short-lived photochemical equilibrium strongly influenced by temperature and by the concentrations of atomic hydrogen and atomic oxygen (Smith and Marsh, 2005). During polar winter, a tertiary ozone maximum of roughly 2–3 ppmv develops near 72 km, peaking close to the polar night terminator. This feature results from reduced concentrations of odd hydrogen during nighttime, decreasing odd-oxygen losses via HO_x catalytic cycles (e.g., Marsh et al., 2001; Sofieva et al., 2009).

In the middle and lower stratosphere, ozone exhibits relatively small diurnal variations because atomic oxygen concentrations are small compared to ozone, and ozone lifetimes exceed one day. As altitude increases into the MLT, diurnal variations become stronger. They are mainly controlled by the daytime photolysis of ozone and its reformation at night through the recombination of atomic and molecular oxygen. The efficiency of ozone production increases with altitude as atomic oxygen becomes more abundant (Brasseur and Solomon, 2005).

Figure 3 presents the monthly zonal-mean ozone distributions for boreal winter (DJF) as measured by the occultation instruments ACE-FTS, HALOE, and SOFIE. The DJF means were calculated for the periods 2005–2011 for ACE-FTS, 1991–2005 for HALOE, and 2007–2023 for SOFIE. Selecting 2005–2011 as the common period (as done for the limb measurements) is not possible here, because the time spans of the solar occultation instruments differ. Only ACE-FTS overlaps with this period, whereas HALOE and SOFIE operate largely outside it. Therefore, HALOE and SOFIE climatologies are shown for their full available periods, while ACE-FTS is presented for 2005–2011, consistent with Fig. 2. For the most part, similar features are observed as in Fig. 2. The ozone distributions show good consistency among all datasets, with spatial patterns clearly observed. However, a key difference is the nighttime ozone enhancement above 70 km. Solar occultation observations are limited to periods when the sun is rising or setting, rather than during complete darkness. The pronounced nighttime ozone enhancement in the MLT occurs during full darkness, outside the observation window of solar occultation instruments. The wintertime means from ACE-FTS and HALOE are in good agreement with those previously reported by Smith et al. (2013).

An example time series of nighttime ozone at northern mid-latitudes (30–60° N) from four instruments (MLS, MIPAS_NOM, MIPAS_MA_UA, and GOMOS), illustrating the vertical distribution of ozone over time, is shown in Supplement Fig. S4. Temporal variations in ozone are similarly represented across datasets, with key patterns over time clearly visible. Seasonal variations in ozone primarily manifest as an annual cycle in the stratosphere, with a maximum of about 10–11 ppmv in late spring/early summer. In the MLT region, a semi-annual cycle is observed, with ozone peaks of about 10–12 ppmv occurring around March–May and September–November.

The merging procedure in general is like that used by Sofieva et al. (2017b, 2023) but is adapted for the MLT region by considering the strong diurnal cycle of ozone. Below we present the details of the merging procedure. It consists of three main steps: (i) evaluation of deseasonalized anomalies, (ii) data pre-merging, and (iii) final merging of pre-merged datasets from individual datasets.

The deseasonalized anomalies can be directly used for ozone trend analyses. For other applications, we also created merged ozone mixing ratio profiles. For this purpose, we computed the merged seasonal cycle as the mean of seasonal cycles from MLS, MIPAS_NOM and MIPAS_MA_UA, for daytime and nighttime illumination conditions. These datasets were selected because of their dense sampling, and to ensure maximum spatial coverage. A similar approach is applied for the merged stratospheric ozone datasets (Sofieva et al., 2017b, 2023). This merged seasonal cycle is then applied to the merged deseasonalized anomalies. Thus, in addition to deseasonalized anomalies, nighttime and daytime ozone mixing ratio profiles and their uncertainty are provided in the merged dataset. Examples of the merged METEOR-O₃ mixing ratio profiles are shown in Fig. 10.

The merged deseasonalized ozone anomalies are suitable for direct use in ozone trend assessments. For this analysis, we apply a multiple linear regression (MLR) to the METEOR-O₃ dataset: 8O3(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 two quasi-biennial oscillation (QBO) proxies (30 and 50 hPa equatorial winds; http://www.cpc.ncep.noaa.gov/data/indices/, last access: 12 May 2026), F10.7(t) is the monthly average solar 10.7 cm radio flux (https://www.spaceweather.gc.ca/forecast-prevision/solar-solaire/solarflux/sx-5-en.php, last access: 12 May 2026), and ENSO(t) is the 2-month lagged El Niño–Southern Oscillation (ENSO) proxy (https://www.esrl.noaa.gov/psd/enso/mei/data/meiv2.data, last access: 12 May 2026). The piecewise linear trend approach is motivated by the well-established ozone turnaround around in 1997 following the peak in ozone-depleting substances, which marks the recovery period (Harris et al., 2008). Similar formulations have been used in previous ozone trend studies (e.g., Bourassa et al., 2014; Kyrölä et al., 2013; Sofieva et al., 2017b). Uncertainties are derived from the residuals of the regression fits, and autocorrelation is corrected in both steps using the Cochrane-Orcutt transformation (Cochrane and Orcutt, 1949). Trends prior to 1997 are not shown because the available record (1991–1996) is too short for a robust estimation of ozone trends, particularly given the strong interannual variability in mesospheric ozone.

For comparison, the regression analysis was performed in two ways. In the first approach, the MLR was applied to the entire merged dataset covering the period 1991–2023. In this case, the bottom altitude range was limited to the coverage of HALOE observations, starting at approximately 37 km and running up to 93 km. In the second approach, the MLR was applied only to the period with the best spatio-temporal coverage (2004–2023), extending from about 22 km upward. Since trend estimates can be sensitive to the choice of regression period (Harris et al., 2015), this approach also serves as a robustness test with respect to the selected analysis window.

Figure 11 shows the trend results for the stratospheric ozone recovery period (1997–2023) and for the best spatio-temporal coverage period (2004–2023). The trends are very similar for both considered periods. Stratospheric ozone trends are in very good agreement with previous ozone assessments (Godin-Beekmann et al., 2022; Petropavlovskikh et al., 2019). The results indicate continued recovery of ozone in the upper stratosphere, with positive values of about 1 %–2 % per decade between 35 and 45 km (above the primary ozone maximum) that are particularly pronounced at middle and high latitudes in both hemispheres.

In contrast, MLT ozone exhibits negative trends, reaching about -1 % to -3 % per decade over 60–80 km. The strongest decreases, of approximately -8 % to -12 % per decade, occur between 80 and 90 km, below the secondary ozone maximum. Because the trends are expressed in percent, regions with lower background ozone can exhibit larger relative changes. Previous analysis based on 10 years (2002–2012) of SABER data confined to lower and middle latitudes (48° S–48° N) revealed a marginally positive not-significant ozone trend at altitudes of 60–80 km and a strong negative trend up to -10 % per decade near 85–100 km. Positive/negative ozone trends were accompanied by a cooling trend of about -3 K per decade (Huang et al., 2014). Regionally, Bizuneh et al. (2022), using 16 years of SABER data over low latitudes (5–15° N), have shown negative temperature and ozone trends (-0.85 K per decade and -0.12 ppmv per decade) in the lower mesosphere (60–80 km) but positive trends (1.25 K per decade and 0.27 ppmv per decade) in the upper mesosphere/lower thermosphere (85–100 km). Alongside natural drivers, the authors attributed these patterns to the influence of increasing greenhouse gas concentrations. Furthermore, rising H2O levels in the mesosphere and lower thermosphere can enhance HO_x production, which accelerates ozone loss and may further contribute to the observed negative O₃ trends.

To assess sensitivity to the trend formulation, additional regressions were performed using the Long-term Ozone Trends and Uncertainties in the Stratosphere (LOTUS) regression model (https://usask-arg.github.io/lotus-regression/, last access: 12 May 2026) with an independent linear term (ILT). The ILT includes two linear terms: a pre-1997 trend and a post-2000 trend. Separate intercepts allow the two periods to vary independently. In contrast, the PWLT predictors impose a common turnaround value in 1997. The resulting ozone trend patterns and magnitudes closely match those obtained with the original MLR (Supplement Fig. S6), demonstrating that, as with the trend period, the results are also not sensitive to the chosen trend representation.

In addition to the relative (% per decade) analysis, we have also derived trends based on absolute (ppmv) ozone values (Fig. 12). These trends are predominantly negative between 60 and 90 km, in qualitative agreement with regional studies (Bizuneh et al., 2022). In the 60–80 km range, trends vary from about -0.01 to -0.02 ppmv per decade during nighttime. At altitudes above 80 km, negative trends of about -0.15 to -0.2 ppmv per decade are observed, with the peak decreases occurring close to the secondary ozone maximum during nighttime. At northern mid-latitudes above 90 km, the trends weaken and in some regions transition to slightly positive values, reaching up to 0.05 ppmv per decade during nighttime.

To investigate the seasonal dependence of ozone trends, we applied a two-step multiple regression technique like that described by Szeląg et al. (2020). In the first step, natural cycles (solar, QBO, and ENSO) were estimated and removed from the data using the traditional MLR formulation given above (Eq. 8). The regression was performed using data from each (three-month) season separately. In the second step, the residual time series (after removal of natural variability) was used to estimate linear trends for the recovery period (2000–2023) using a simple linear regression. This two-step approach provides a robust estimation of seasonal trends by ensuring sufficient data for detecting natural cycles and by fitting linear trends only within periods where ozone changes are approximately linear.

Variations of ozone trends over the period 2000–2023 for each latitude and altitude are shown for each season separately in Fig. 13. In the upper stratosphere, ozone trends are positive throughout all seasons and across most latitudes, consistent with previous studies (Szeląg et al., 2020). The strongest positive trends, up to about 2 %–4 % per decade, are observed at middle and high latitudes.

In the lowermost mesosphere (50–60 km), ozone trends are generally close to zero, with most not statistically significant, suggesting little long-term change in this region. Between approximately 60 and 80 km, negative trends (about -2 % to -4 % per decade) dominate, indicating a persistent decrease in mesospheric ozone. The largest negative values, reaching up to -15 to -20 % per decade, occur between 80 and 90 km, particularly in the Northern Hemisphere during boreal summer (JJA) and in the Southern Hemisphere during austral summer (DJF) where the ozone absolute values are very low. During the equinoctial seasons, negative ozone trends peak at mid-latitudes and in the tropics. In some seasons and latitude bands, however, trends become positive, for example during local winter between 75–80 km and 60–80° latitudes.

We have developed a new merged dataset of middle atmospheric ozone (∼ 22 to 100 km), METEOR-O3, constructed from multiple limb emission and occultation satellite instruments (HALOE, MLS, ACE-FTS, MIPAS, GOMOS, and SOFIE). The merging procedure, adapted from previous methods for stratospheric datasets, accounts for the strong diurnal variability of ozone in the mesosphere. The dataset provides monthly zonal-mean ozone anomalies from 1991 to 2023, with 10° latitude resolution and vertical coverage from approximately 22 up to 100 km.

The merged dataset shows excellent internal consistency among datasets and provides robust coverage of the middle and upper atmosphere, with typical uncertainties below 2 % and slightly higher uncertainties (5 %–10 %) near the mesopause. The deseasonalized anomalies derived from METEOR-O3 are well suited for direct use in trend analyses.

Ozone trend evaluation was carried out on pressure levels using two complementary regression methods. The one-step MLR analysis was applied to the full merged dataset (1991–2023) and to the period of best spatio-temporal coverage (2004–2023). For the full-period analysis, only post-recovery (post-1997) trends were presented. The results show positive trends of about 1 %–2 % per decade in the upper stratosphere (35–45 km), consistent with the ongoing ozone recovery observed in earlier studies (Godin-Beekmann et al., 2022; Petropavlovskikh et al., 2019). In the mesosphere, ozone trends are negative, with values of approximately -1 % to -3 % per decade in the 60–80 km region and up to -8 % to -12 % per decade between 80 and 90 km. While previous SABER-based studies were temporally and regionally limited (e.g., Bizuneh et al., 2022; Huang et al., 2014), the present analysis extends to a global scale, providing for the first time a comprehensive assessment of long-term mesospheric/lower thermospheric ozone variability.

The METEOR-O3 dataset enables global evaluation of long-term changes and seasonal variations. It offers a valuable resource for model validation and for improving understanding of upper atmospheric processes.