Ozonesonde vertical profile measurements from 1980–2017 were downloaded from the World Ozone and Ultraviolet Radiation Data Center (WOUDC) (https://woudc.org/data/explore.php, last access: 29 September 2022), the National Oceanic and Atmospheric Administration (NOAA) (ftp://ftp.cmdl.noaa.gov/ozww/Ozonesonde/, last access: 29 September 2022), and the Harmonization and Evaluation of Ground-based Instruments for Free Tropospheric Ozone Measurements (HEGIFTOM) working group of the Tropospheric Ozone Assessment Report, Phase II (TOAR-II) (https://hegiftom.meteo.be/datasets/ozonesondes, last access: 30 September 2022). The global ozonesonde community is currently reprocessing and homogenizing data to account for changes in ozonesonde preparation and procedures, with the goal to reduce measurement biases associated with these changes (Tarasick et al., 2016; Van Malderen et al., 2016; Witte et al., 2018; Sterling et al., 2018; Ancellet et al., 2022). Where possible (12 of 25 sites), homogenized ozone profiles were used to ensure the most accurate ozone trends. Table 1 describes the ozonesonde profile information, dates, and whether the data are homogenized. While Payerne (Europe) has homogenized data, we use the original data since the site has only been homogenized since 2002. For data that are not homogenized, we ensure that they do not contain step changes (Figs. S1 and S2 in the Supplement). Updated tropical ozonesonde information is available from the Southern Hemisphere Additional OZonesondes (SHADOZ) (https://tropo.gsfc.nasa.gov/shadoz/, last access: 21 September 2022), but we did not include these data in this analysis because they did not meet our data requirements described below, typically due to not having enough profiles per month consistently throughout our time frame.

Most ozonesonde data were measured by electrochemical concentration cell (ECC) sensors, widely regarded as the most accurate sensor type (Tarasick et al., 2021). Four sites (Payerne, Uccle, Legionowo, and Lindenberg) in Europe switched from using Brewer–Mast (BM) sensors to ECC sensors partway through their data records, and data from both sensors were used since previous analyses showed good agreement between measurements (De Backer et al., 1998; Stübi et al., 2008). Only Hohenpeissenberg (Europe) used the BM sensor throughout the time period. Naha (Japan), Tsukuba (Japan), Sapporo (Japan), and Syowa (Antarctica) both used carbon iodine (CI) sensors prior to 2010 and ECC sensors after, and this switch could impact overall long-term trends (Tanimoto et al., 2015). Typical uncertainties for CI sensors range from 5 %–10 %, while they are 3 %–5 % for ECC sensors (Tanimoto et al., 2015). This could lead to substantial differences in calculated trends, and we discuss trends from these sites in the context of regional trends using sites with more reliable data (e.g., only one sensor type or homogenized data). We note that trends at these sites should be treated with caution. A recent study showed a drop in total column and stratospheric ozone measured by ECC instruments compared to satellite observations in the latter parts of their records for reasons still under investigation (Stauffer et al., 2020, 2022). We find that 5 of our 25 sites were impacted by these ozone measurement drops, although these drop-offs were typically limited to pressures above ∼50 hPa, so our results should not be affected. Out of an abundance of caution, at these impacted sites, we only used data from before the unexplained sharp drop-off in ozone concentrations, as data before these drops are still considered highly reliable (Stauffer et al., 2020, 2022), and this resulted in the removal of up to 1 year of data at each affected site.

Ozonesonde profiles were reduced to match the 47-layer GEOS-Chem reduced pressure levels by aggregating all observed ozone values between model-defined pressure edges. The following criteria for ozonesonde sites were used in this analysis from 1990–2017. Locations were selected based on data completion criteria adapted from Lu et al. (2019): (1) at least three observations per month, (2) at least two monthly observations per season, (3) at least eight monthly observations per year, and (4) at least 16 years of data. These data requirements were met by 25 ozonesonde locations throughout the globe for the 1990–2017 time period (Fig. 1). Nine of the selected sites have data extending back to the 1980s, and these trends are discussed where appropriate, although the main focus of this work is on trends after 1990.

Surface daytime baseline ozone data from 1990–2014 were obtained from the TOAR Surface Ozone Database (Schultz et al., 2017), which has been compiled and processed by the TOAR data team and made public via https://doi.org/10.1594/PANGAEA.876108 (last access: 5 May 2022). Each site in this database has at least 70 % of all hourly ozone measurements available for each year provided as monthly aggregates. Similar to the ozonesondes, sites used in this analysis were constrained by the following criteria: (1) at least two monthly observations per season, (2) at least eight monthly observations per year, and (3) at least 15 years of data throughout the time frame. TOAR site locations are shown in Fig. 2 below. All sites are in background locations, which is defined by individual data providers to the TOAR database with no formal unifying definition (Schultz et al., 2017). All sites were also classified as “rural,” which is defined as (1) NO2 column ≤8×1015 molec. cm-2 as measured by the Ozone Monitoring Instrument (OMI), (2) an averaged nighttime light intensity index of ≤25 within a 5 km radius of the site, and (3) a maximum population density of ≤3000 people km-2 within a 5 km radius of the site (Schultz et al., 2017). Of the 271 surface site locations meeting these requirements, 52 site locations are in the United States, and 173 are in Europe, biasing trend information to these areas (Fig. 2). However, there are 33 background sites in other regions spanning the globe that give insight to changes in surface ozone beyond the northern midlatitudes.

Included in these sites are eight high-elevation sites (>2800 m), which are discussed separately from the other surface sites and are marked with blue dots in Fig. 2. These sites include five mountaintop sites, which have been studied extensively to determine if ozone trends at these sites are dominated by FT air (Logan et al., 2012; Parrish et al., 2014; Cooper et al., 2020), generally using nighttime ozone values to avoid influence from local air masses. During the day, mountaintops often experience updrafts of polluted air from lower altitudes. While these sites have traditionally been used as another method for identifying lower FT ozone trends (Logan et al., 2012; Parrish et al., 2014), a recent analysis of three European mountaintop sites (Jungfraujoch, Sonnblick, and Zugspitze) found they were influenced by boundary layer air and were thus more representative of the lower troposphere (Cooper et al., 2020). Other mountaintop sites (Mauna Loa and Mt. Waliguan) have been found to be representative of FT air when the data are filtered appropriately to exclude air masses influenced by the boundary layer (Cooper et al., 2020; Lin et al., 2014; Xu et al., 2018). Here, we did not seek to reiterate trends since the 1990s reported in previous studies but rather used these high-elevation and mountaintop sites representative of regional or FT air to corroborate observed ozonesonde trends. Six of the sites (Centennial, Gothic, South Pole, Zugspitze, Jungfraujoch, and Sonnblick) were used as a point of comparison for lower tropospheric (>700 hPa) ozonesonde trends, and two sites (Mauna Loa and Mt. Waliguan) were used for FT ozonesonde trends (700–400 hPa). Trends from each site were reported using ozone measurements from various times during the 24 h diurnal cycle to capture regional or FT trends, and the times used are specified in Sect. 3.4.

Ozonesonde and TOAR surface data were analyzed using R statistical software (R Core Team, 2013). In this work, we reported trends in parts per billion (ppb) per decade and considered them significant if p<0.1. Trends were calculated using deseasonalized data and quantile regression due to the intermittent nature of the ozonesonde launches (Gaudel et al., 2020; Koenker and Bassett, 1978). Deseasonalization reduces the impact of autocorrelation. At each pressure level and site, we constructed a mean seasonal cycle for each site's time frame. This seasonal cycle was then used to deseasonalize individual observations on each pressure level. Quantile regression is an expansion of linear regression, which predicts trends for a distribution rather than using conditional means. An advantage of quantile regression for our dataset is that it does not require the aggregation of sparse data to monthly means. As most ozonesonde locations launch only three to four times each month, monthly mean values may not be statistically meaningful. Quantile regression is also robust for datasets containing outliers and intermittent missing values, making it appropriate for our ozonesonde dataset. Quantile regression has the added benefit of predicting trends for various percentiles of the distribution, allowing for the examination of extreme trends (e.g., 5th percentile). Linear trends were calculated using all profiles in the time frame.

To evaluate model ability to reproduce long-term ozone trends, we analyzed a variety of model configurations comprising different emissions inventories, chemical schemes, and resolutions. We used two simulations of GEOS-Chem v12.9.3 (GC) and a replay simulation of the National Aeronautics and Space Administration Goddard Earth Observing System (NASA GEOS) model coupled to the Global Model Initiative (GMI) chemical mechanism and meteorological information from MERRA-2 reanalysis data, hereafter referred to as MERRA2-GMI. We also used a shorter simulation from an earlier version of GEOS-Chem (v10-01) that spans 1980–2010 as a point of comparison. The details for each of these simulations are described below and in Table 2.

Regular ozonesonde launches at long-term sites during certain days of the week or month represent untargeted sampling that allows for a systematic characterization of the vertical distribution of the entire troposphere and above. However, concerns about the suitability of ozonesondes for long-term trend analyses have been raised previously (Saunois et al., 2012; Liu et al., 2016). The concern is that ozonesondes launched only a few times per month capture snapshots of ozone changes over time and may not fully capture trends. By contrast, ozone is measured continuously on an hourly basis at the surface sites, making it likely that these sites capture robust trends in long-term data, though they reflect only the trends in the atmospheric boundary layer. To assess the ability of our ozonesonde sites launching at least three times per month to accurately represent overall trends, we compared the lowest reliably available pressure level (800 hPa) to co-located surface TOAR sites within a 100 km radius. The 800 hPa pressure level is typically within the atmospheric boundary layer and should be mostly affected by similar processes as the surface sites.

In most seasons, we found that trends from the surface sites and the ozonesondes correlate significantly (r>0.5, p<0.1 for all), while wintertime often shows the worst agreement due to a lower boundary layer height. Summer is typically when trends match most closely, as the boundary layer is deepest then. At all five co-located sites during summer (Boulder (United States), Hohenpeissenberg (Europe), Payerne (Europe), Uccle (Europe), and Tateno (Japan)), trends between the surface and 800 hPa match in terms of direction, and the magnitude of trends differ by <30 % (Fig. S3). This suggests that ozonesondes launching at least three times per month are able to capture long-term seasonal trends. The absolute values of ozone can differ widely between measurement techniques, with surface sites being systematically lower due to the increased influence of dry deposition (Travis and Jacob, 2019). Previous work has typically used ozonesonde data that launch four times per month (Lu et al., 2019). However, along with our other data requirements, this restriction would limit the number of sites to just 15, eliminating nearly all Southern Hemisphere and polar sites and negatively impacting our global analysis. Here, we show that trends in low-level ozonesondes and TOAR sites largely match each other, and we conclude that we are able to use the ozonesonde sites launching at least three times per month to understand trends throughout the vertical column.

Trends in ozonesonde data suggest tropospheric ozone has increased throughout the troposphere since the 1980s and 1990s. Of the 25 ozonesonde stations examined globally from 1990–2017, 14 show statistically significant increases from 800 to 400 hPa (Fig. 3). We caution that these results are derived from both homogenized and non-homogenized data depending on availability (see Table 1 for a list of homogenized sites). The impact of homogenization is shown in Fig. S4, with homogenization affecting trend magnitudes but rarely the sign of the trends compared to non-homogenized data. Across all pressure levels, these 14 sites average an increase of 1.8 ± 1.3 ppb per decade (3.5 % ± 2.6 % per decade), ranging from 0.1 to 5.3 ppb per decade (0.2 % per decade to 10.6 % per decade) since the 1990s. At the nine sites that have records from 1980, five show consistent increases averaging 1.3 ± 0.7 ppb per decade (2.6 % ± 1.4 % per decade) and ranging from 0.1 to 3.0 ppb per decade (0.1 % per decade to 5.9 % per decade; Fig. S5). Over half of all ozonesonde sites from 1990–2017 show increasing ozone in the free troposphere (700–400 hPa) at an average rate of 1.9 ± 1.3 ppb per decade (3.6 % ± 2.4 % per decade), but trends range widely, from 0.1 to 5.3 ppb per decade (0.1 % per decade to 9.9 % per decade). While relative trends (taken relative to the mean ozone concentration at each pressure level from 1990–2017) are remarkably constant through the troposphere at most sites, they tend to be larger closer to the surface (4.3 % per decade below 700 hPa on average, compared to 3.5 % per decade above 700 hPa), reflecting the importance of emissions changes on ozone trends (Fig. S6). Trends at sites that are not increasing show mostly insignificant decreasing trends, with few showing statistically significant decreases. The only records with strongly negative trends are the lower troposphere at Wallops Island (eastern United States, -2.2 ± 0.6 ppb per decade), the upper troposphere at Macquarie Island (Southern Ocean, -1.2 ± 0.4 ppb per decade), the lower troposphere at Broadmeadows (southeastern Australia, -1.2 ± 0.2 ppb per decade), and the extreme upper troposphere at Eureka (polar Canada, -2.7 ± 1.6 ppb per decade).

The strongest increasing trends from 1990–2017 occur in Japan, averaging 3.8 ± 0.8 ppb per decade (7.1 % ± 1.5 % per decade) across all pressure levels and ranging from 2.4 to 5.3 ppb per decade (4.4 % per decade to 9.9 % per decade). Caution should be taken to not over-interpret the Japanese trends, as a potential step change occurs at these sites around 2010 in the troposphere (Fig. S1). While this may be partially due to a change in sensor response, these step changes are not visible in the stratosphere (Fig. S1), suggesting that these trends mostly reflect the rapid increase in emissions over Asia in the past 4 decades. Similarly, all NH Polar sites except Eureka show increasing trends, averaging 1.6 ± 0.9 ppb per decade (3.1 % ± 1.7 % per decade; ranging from 0.4 to 3.3 ppb per decade). Over North America, the two Canadian sites (Edmonton and Goose Bay) show consistent increases throughout the tropospheric column, averaging 1.3 ± 0.9 ppb per decade (2.5 % ± 1.7 % per decade; ranging from 0.5 to 3.3 ppb per decade). Half of sites over Europe also show increasing trends, averaging 1.9 ± 1.1 ppb per decade (3.4 % ± 2.0 % per decade; ranging from 0.1 to 4.3 ppb per decade). Over the Southern Hemisphere, two of the six sites show smaller increasing trends from 1990–2017 compared to other regions, averaging 0.7 ± 0.6 ppb per decade (2.1 % ± 1.8 % per decade; ranging from 0.3 to 1.7 ppb per decade). Hilo (Hawaii) in the tropics shows insignificant trends below 600 hPa, averaging 0.6 ± 0.7 ppb per decade (1.3 % ± 1.5 % per decade), and insignificant increases above 600 hPa, averaging 1.1 ± 0.7 ppb per decade (2.2 % ± 1.4 % per decade).

Figure 4 depicts the shift in overall ozone distributions at all pressure levels between the first (1990–1994) and last (2013–2017) 5 years of the time series, with all sites grouped into five of the six regions (i.e., all except Hawaii). In each region, distributions from 800–400 hPa shift in a positive direction, with increases in medians averaging 2.5 ppb globally and ranging up to 3.5  ppb over Japan. Across all sites in these regions, the largest absolute and relative shifts occur in the lower troposphere (>700 hPa). Changes in medians average 2.2 ppb (5.1 %) in the lower troposphere and 1.3 ppb in the free troposphere (2.6 %).

The generally increasing ozone concentrations measured by ozonesondes are consistent with satellite and aircraft data. Satellite measurements from the Aura Ozone Monitoring Instrument/Microwave Limb Sounder (OMI/MLS) from 2005–2016 show widespread increases of ozone across the tropics and midlatitudes, ranging up to >3 DU per decade over Asia (Ziemke et al., 2019). This finding is corroborated by global chemistry climate models, which indicate that the tropospheric ozone burden has increased since 1990 (Myhre et al., 2017; Ziemke et al., 2019). In the simulations used in this work, we also find that the ozone burden has increased since 1980, which we discuss further in Sect. 4.3. Free tropospheric and tropospheric column ozone measured by IAGOS also suggests that ozone has increased across the Northern Hemisphere since the 1990s (Gaudel et al., 2020; Petzold et al., 2015; Cohen et al., 2018) at an average rate of 2 ppb per decade, which agrees with the average 2.0 ± 1.3 ppb per decade increase in Northern Hemisphere FT ozonesonde measurements. Although some variation is expected when comparing regions to individual sonde launch locations, our results show good agreement with previous analyses of FT ozone (700–300 hPa) since the 1990s using IAGOS flight data. Over Europe, Gaudel et al. (2020) found an increasing trend of 1.3 ± 0.2 ppb per decade, slightly lower than our result of 1.9 ± 1.1 ppb per decade but within uncertainty. Gaudel et al. (2020) report an increase of 1.3 ± 0.9 over the southeast US FT, which aligns with our findings in the upper troposphere at Wallops Island (Virginia, United States; 0.8 ± 0.3 ppb per decade). Over eastern North America, an increase of 1.7 ± 0.4 ppb per decade is in good agreement with ozonesonde measurements at Goose Bay (eastern Canada; 2.0 ± 0.7 ppb per decade). This remarkable agreement between ozonesondes and other measurement platforms lends further evidence that ozonesondes launching three times per month are able to capture long-term trends in tropospheric ozone.

There is much discussion about the number of profiles needed for statistical analyses of global ozone trends, and recent studies have suggested that 14 profiles per month are needed (Chang et al., 2020). However, this number of profiles is not possible under the current ozonesonde sampling landscape. Here, we show that careful selection and treatment of ozonesonde data can lend important insights to global ozone trends that are highly vertically resolved. We note that these trends may not be considered globally representative, but rather they offer an additional insight into ozone changes over the past few decades. That we find good agreement between ozonesonde trends and trends from other data sources suggests that ozonesonde information is an important part of the ozone monitoring landscape in determining global trends.

It is important to note that we have not performed a seasonal analysis of ozonesonde data. Analyses of ozonesonde sites in the tropics point to the seasonal variability of ozone and show that trends are driven primarily by changes during certain months. For example, Thompson et al. (2021) did not find significant trends at Nairobi, Kenya, on an annual basis (consistent with our results) but found that FT ozone increased during February–April by 5 % per decade–10 % per decade, while it decreased during August–September. Other tropical sites show similar patterns – annual trends are insignificant, while seasonal trends are much larger. A seasonal analysis is beyond the scope of this paper, as the three to four launches per month may not give enough information for robust monthly or seasonal trend analysis. Future investigations of ozone trends should consider the impact of specific months or seasons, provided it can be done in a statistically meaningful way, to aid in identifying drivers of trends.

While surface ozone trends have been discussed in previous analyses (Gaudel et al., 2018; Cooper et al., 2020), we focus specifically on daytime ozone trends rather than on trends in monthly mean ozone that average all times of day (Parrish et al., 2014); we also consider a greater number of sites covering a larger geographical area than other studies attempting to characterize baseline ozone. Specifically, we include 271 sites, including additional sites in the poorly sampled Southern Hemisphere, while restricting site locations to rural background areas.

Despite regional decreases over the United States and Europe, surface ozone increases in most places globally since the 1990s, ozone distributions have generally shifted up across the time frame, and medians have largely increased (Fig. 4). At sites outside of the United States and Europe at low elevations, 73 % show increasing trends (24 of 33 sites). Including the US and European sites, we find that 42 % of global surface background sites (114 of 271) show ozone increases since the 1990s, with notable decreases at 48 of the 52 US sites and 100 of 186 European sites due to emissions regulations (Fig. 5). Surface ozone changes at individual sites globally range from -7.5 to +5.2 ppb per decade. Across all sites, increases average 1.0 ± 0.8 ppb per decade, and decreases average -1.4 ± 1.2 ppb per decade. For sites outside of the United States and Europe, increases average 1.4 ± 0.9 ppb per decade, and decreases average -1.3 ± 0.8 ppb per decade, with the largest increases occurring over Asia. Decreases in eastern China (LinAn) can be attributed to the prevalence of clean marine air masses impacting that site during fall that do not reflect the growing urban emissions in China (Xu et al., 2008). Our results are consistent with other global analyses of surface ozone data that have shown increases over varying time frames beginning in the 1990s at far fewer sites spanning a narrower slice of the globe (Cooper et al., 2020, 2014). Of the expanded 258 sites in the Northern Hemisphere analyzed here, we find increases at 103 sites (40 %), ranging from <0.1 to 5.2 ppb per decade. Focusing on Northern Hemisphere trends outside of the United States and Europe, we find increasing trends at 13 of the 20 sites (65 %), averaging 1.4 ± 0.9 ppb per decade (0.5 to 5.2 ppb per decade). At the 13 Southern Hemisphere sites analyzed here, we find increases at 11 sites (85 %) since 1990, ranging from 0.5 to 1.8 ppb per decade. Our results are consistent with findings from Cooper et al. (2020), who found that about half of Northern Hemisphere sites with significant trends (5 out of 10 sites) show increasing trends ranging from 0.7–1.7 ppb per decade, and 71 % of Southern Hemisphere sites (5 out of 7) show increasing trends (0.3 to 1.5 ppb per decade. Increases at surface sites are shown in Fig. 4, where the medians of all distributions except North America have shifted in a positive direction from the first 5 years of analysis (1990–1994) to the last 5 years (2010-2014), with changes in median ozone concentration across all regions averaging 2.0 ppb (6.0 %) at the surface, closely matching overall ozone increases at the 27 sites observed globally in Cooper et al. (2020) since 1995 (1.6 ppb).

High-elevation surface sites provide another line of evidence regarding regional baseline ozone (e.g., ozone that is not influenced by local emissions) trends, as they are regionally representative of the lower troposphere (Logan et al., 2012; Cooper et al., 2014; Parrish et al., 2012, 2014). Careful filtering of data at some of these high-elevation sites can also isolate the influence of lower FT air (Cooper et al., 2020; Lin et al., 2014). Analyses of high-elevation sites have focused primarily on Europe (Cooper et al., 2020; Logan et al., 2012), although a limited number of sites in North America, Japan, Hawaii, and China have also been studied (Parrish et al., 2012, 2014; Cooper et al., 2014; Lin et al., 2014; Xu et al., 2016). Here, we do not attempt to recalculate trends at these sites but rather examine previously reported trends and compare them to lower and free tropospheric ozonesonde trends. We show that the trends measured by ozonesondes match those of high-elevation and mountaintop surface trends in most locations, adding confidence to the trends we derive from ozonesondes launching at least three times per week.

Two mountaintop sites influenced by FT air are Mauna Loa (Hawaii) and Mt. Waliguan (China). At both of these sites, FT trends measured at the mountaintop sites show increasing FT (700–400 hPa) ozone trends. At Mt. Waliguan, FT trends can be isolated using nighttime ozone values, and measurements show an increase in FT ozone of 2.8 ± 1.6 ppb per decade from 1994–2013 (Xu et al., 2016) and 1.7 ± 0.5 ppb per decade from 1994–2016 (Cooper et al., 2020). This finding is attributed to both transport of increasing anthropogenic emissions and intensifying STE, which can explain 60 % of the springtime ozone increase (Xu et al., 2016, 2018). While we do not analyze any ozonesonde launch locations over China and therefore do not have a direct comparison to sonde information, it is important to recognize the pattern of increasing FT ozone at multiple sites throughout the globe. At Mauna Loa, the influence of FT air can be isolated under nighttime conditions with low relative humidity. Cooper et al. (2020) found that FT ozone at Mauna Loa has increased by 2.4 ± 1.0 ppb per decade since 1995. Annual trends from 1991–2010 were found to be 3.1 ± 0.7 ppb per decade (Oltmans et al., 2013), driven by increasing autumn trends (3.5 ± 1.4 ppb per decade; 1980–2012) (Lin et al., 2014). The trend reported in Cooper et al. (2020), which best matches our analysis time frame, is higher than the average FT trends we calculated over Hilo from ozonesonde measurements from 1990–2017 (0.9 ± 0.6 ppb per decade from 700–400 hPa) but falls within the range measured in the FT (range of -0.2 to 1.7 ppb per decade.

The other high-elevation sites have been found to be more representative of regional ozone trends in the lower troposphere than the FT. Two European mountaintop sites (Zugspitze, Sonnblick) show decreasing trends since 1995, -0.8 ± 0.6 and -1.0 ± 0.7 ppb per decade, respectively, while a third mountaintop site, Jungfraujoch, exhibits an insignificant trend of 0.2 ± 0.6 ppb per decade at night (Cooper et al., 2020). We find good agreement between closely located sonde and mountaintop trends. Both Zugspitze and Sonnblick are closely located to the Hohenpeissenberg ozonesonde location (within 100 km), which shows a decreasing trend of -0.8 ± 0.2 ppb per decade in the lower troposphere, within the range of trends reported for Zugspitze and Sonnblick. Jungfraujoch is near the Payerne ozonesonde location (within 100 km) and shows an insignificant decreasing trend of -0.2 ± 0.1 ppb per decade, which overlaps with the trend reported at Jungfraujoch. A consistent picture is difficult to put together for all of Europe considering the large variation in local trends, but overall our lower tropospheric ozone trends from sonde data encompass those found at mountaintop sites.

Over the United States from 1995–2017, Cooper et al. (2020) reported on two lower tropospheric high-elevation sites, Centennial (WY) and Gothic (CO). Trends are -1.5 and -1.9 ± 0.8 ppb per decade during the daytime, respectively. At the Boulder (CO) ozonesonde measurements, lower tropospheric trends average -0.5 ± 0.4 ppb per decade, agreeing with surface trends. Over the South Pole, only 24 h trends from 1995–2018 were reported by Cooper et al. (2020) due to the lack of a diurnal ozone cycle; these averaged 1.5 ± 0.6 ppb per decade. We find a consistent trend of 1.2 ± 0.1 ppb per decade in the lower troposphere at Syowa Station (coastal Antarctica). Differences in the increases at these two stations may occur as a function of station location and whether anthropogenic sources or meteorological variables are the main drivers of ozone trends at each station. At the South Pole, increases are associated with ozone-rich air from the upper troposphere and lower stratosphere, whereas Syowa, located at 69∘ S, is primarily impacted by marine air and air-mass transport from regions near South America (Kumar et al., 2021). It is also important to note that Syowa switched sensors from CI to ECC in 2010, which could impact trends.

Across all regions, we find that increases in 5th percentile ozone at the surface and aloft since the 1990s contribute to increases in median ozone. To estimate 5th percentile ozone trends, we first calculated the 5th percentile ozone in each month at each pressure level for all individual sites and then used the quantile regression method to calculate the trends of the 5th percentiles of measurements year-round. Figure 6 shows the trend of 5th percentile ozone across ozonesonde and surface sites grouped into the six regions. At most locations globally (178 of 271 surface sites and 13 of 25 sonde sites), 5th percentile ozone has increased in both ozonesonde and surface trends, averaging 1.8 ± 1.3 ppb per decade, with 59 % of those sites showing increases of greater than 1.0 ppb per decade and ranging up to 4.9 ppb per decade at surface sites and 5.6 ppb per decade at ozonesonde sites. Notably, while 5th percentile surface ozone has increased significantly in the United States and Europe, peak surface ozone values decreased in recent years (Fig. 4), reflecting reductions in regional anthropogenic emissions of ozone precursors (Yan et al., 2018a, b). In contrast, in the FT over Japan and the Southern Hemisphere, the entire ozone distribution has shifted higher. Over Japan, these increases have been attributed to transport from the Asian continent and reduced NOx emissions leading to decreased titration of ozone (Akimoto et al., 2015). Over the Southern Hemisphere, these increases occur in response to changing precursor emissions and large-scale dynamics, including an expansion of the Hadley cycle which may allow more stratospheric, ozone-rich air to enter the troposphere (Lu et al., 2019; Zeng et al., 2017; Cooper et al., 2020).

Increasing 5th percentile concentrations are consistent with other analyses that suggest baseline ozone has been increasing, especially in the Northern Hemisphere. Increases in 5th percentile ozone have been attributed to a number of factors: decreased titration from NOx as a result of emissions decreases on a local scale, especially over urban areas in the United States and Europe (Yan et al., 2018b; Simon et al., 2015; Lin et al., 2017; Gao et al., 2013; Clifton et al., 2014); increases in methane concentrations (Lin et al., 2017); changes to large-scale processes such as STE (Parrish et al., 2012); and transport of ozone from the tropics and subtropics (Zhang et al., 2016; Gaudel et al., 2020). While all of these factors likely play a role in increased 5th percentile ozone in the Northern Hemisphere, multiple previous analyses suggest that regional, baseline ozone increases observed in rural locations with little impact from local emissions are best explained by transport from the tropics (Zhang et al., 2016, 2021). The largest emissions of ozone precursors have shifted toward low-latitude nations, especially in Southeast, East, and South Asia, where increased convection and temperature lead to more efficient ozone production compared to the midlatitudes. This ozone is then transported poleward (Zhang et al., 2016). Tropospheric ozone increases in the middle troposphere (550 to 350 hPa) over midlatitudes can be largely explained in models through transport of ozone from low latitudes, with STE playing an important role in the upper troposphere (above 350 hPa) (Zhang et al., 2016; Gaudel et al., 2020). Only 15 % of the ozone increase over the western United States between 1980–2014 has been attributed to an increase in methane concentrations (Lin et al., 2017).

We find that models comprising different resolutions, time-varying emissions, assimilated meteorological inputs, and chemical schemes tend to underestimate observed long-term ozone trends throughout the troposphere, and the direction of trends at some individual sites is not captured (Fig. 7). Across all 25 sites evaluated, the average 800–400 hPa observed ozone trend by ozonesondes is 0.8 ± 1.7 ppb per decade from 1990–2017, reflecting the wide spread in observed trends, and the three simulations underestimate this trend mostly in the northern extratropics. Globally, MERRA2-GMI captures ∼75 % of the trend at 0.6 ± 0.7 ppb per decade, but both GC simulations drastically underestimate it and do not differ significantly between the different resolutions (0.15 ± 0.7 ppb per decade for the 4∘×5∘ version; 0.1 ± 0.9 ppb per decade for the 2∘×2.5∘). This result represents <20 % of the overall average observational trend for both GC simulations from 1990–2017. Notably, MERRA2-GMI typically performs better in the upper free troposphere than the GC simulations in the northern midlatitudes, matching 44 % of the observed trend from 600 to 450 hPa, while the GC simulations only capture 24 %. An important note is that a notable step change occurred in MERRA2-GMI ozone after 1998, associated with an observing system upgrade incorporated into MERRA-2 (Stauffer et al., 2019). This step change impacts pressure levels mostly above our analysis range, and model ozone at pressures <450 hPa may be affected. MERRA2-GMI trends at pressures <450 hPa should thus be interpreted with caution. While the error bars do overlap between models and observations in all simulations, most of this error is due to regional variability, and trends between models and measurements within regions often do not overlap.

Figure 8 shows shifts in ozone distributions from 1990–2017 between 800 and 400 hPa. Most overall shifts in distribution from 800–400 hPa are captured by models in a qualitative sense, but shifts tend to be underestimated, most strongly by the GC simulations. Both GC simulations capture the observed increases in all regions except the NH Polar region and Europe, where the models both show decreasing trends in contrast to observations (Fig. 4; also shown in Fig. 7). The median ozone increases are underestimated by an average of 3 ppb in both simulations. In contrast, MERRA2-GMI captures the observed increases everywhere but underestimates these increases over North America by 0.9 ppb. MERRA2-GMI also overestimates the median increase over Europe, Japan, and the NH Polar region by 1.6, 2.1, and 1.7 ppb, respectively, yet it captures within 0.5 ppb the overall median increases over the Southern Hemisphere.

Increases in 5th percentile ozone across North America are marginally captured by all models, but only MERRA2-GMI captures that trend over the NH Polar region and Europe. Shifts of the entire distribution that are observed over the Southern Hemisphere and Japan are captured by all models, although these shifts are typically underestimated (SH: 2 ppb observations, range of 1.1–2.8 ppb from models; Japan: 6.1 ppb observations, range of 1.8–4 ppb from models), with MERRA2-GMI replicating the shifts most reliably.

It is unlikely that the differences in trends between GEOS-Chem and MERRA2-GMI are primarily due to differences in the underlying emissions inventories. MERRA2-GMI used the MACCity inventory, and GEOS-Chem used the CEDS inventory. Typically, CEDS estimates higher magnitudes of NOx emissions and larger trends than MACCity (Fig. S7). However, we find that GEOS-Chem (using CEDS) produces smaller ozone trends than MERRA2-GMI, which suggests that the trend differences between models are more likely to be due to factors other than the emissions inventories, such as model resolution.

Average trends in daytime ozone at surface locations overlap between models and observations (Fig. 9), although individual sites are typically not captured well. The average observed increasing surface ozone trend is 1.0 ± 0.8 ppb per decade, and all simulations overlap (GC 4×5: 0.6 ± 0.8 ppb per decade, GC 2×2.5: 0.6 ± 0.7 ppb per decade, MERRA2-GMI: 1.4 ± 1.0 ppb per decade). The direction of trends at the surface is generally captured by MERRA2-GMI, with the model capturing increasing trends at 67 % of the surface sites also exhibiting increasing trends. Both GC simulations perform more poorly, with GC 2×2.5 capturing increasing trends at 37 % of sites and GC 4×5 capturing increasing trends at just 19 % of sites. In both GC simulations, the trends predicted by the models at many locations, especially over North America and Europe, are opposite in sign to trends in the observations. At high-elevation sites, which are more representative of regional air, the models do a better job of predicting the observed direction but do not capture the magnitude of trends. At these sites, MERRA2-GMI captures the sign of the trends at five of eight sites but underestimates these trends by 0.3 ppb per decade on average. Both GC simulations capture the sign of the trends at six of the eight high-elevation sites, but the 4∘×5∘ simulations overestimate the trends by 0.8 ppb per decade on average, and the 2∘×2.5∘ simulation overestimates the trends by 0.5 ppb per decade on average. The directions of regional changes are captured well by the models, but resolutions may be too coarse to get the surface trends at individual locations, especially in the GC simulations.

Figure 10 shows shifts in surface regional ozone distribution medians in the models. All models qualitatively capture median shifts in Europe, North America, and the Southern Hemisphere, but models tend to underestimate these shifts. GC underestimates distribution shifts by 1.8 ppb on average, and MERRA2-GMI underestimates by 1.9 ppb on average. MERRA2-GMI reproduces the median shift in the Southern Hemisphere well (1.5 ppb in observations and 1.8 ppb in model). All simulations capture a median shift opposite in sign to the observations in Japan and the NH Polar region. The discrepancy between models and observations in both regions can be traced to the models' failure to capture the increased frequency of high-concentration ozone values during the 2010–2014 period. However, the models do capture the increase in frequency in low-concentration ozone values in Japan.

As explored earlier, observations suggest that increases in surface ozone are at least partially attributable to an increase in low-percentile ozone over North America, NH Polar, and Europe (Fig. 4). At the surface, increases in low quantile ozone values are captured by both GC simulations over North America, the NH Polar region, and Europe. Both GC simulations also capture the decreasing high tails in North America and Europe. In contrast, MERRA2-GMI does not capture the observed increases of low quantile ozone at the surface in North America or Europe, and it does not capture the decreasing high tails in Europe. While all models capture the increasing high tail in Southern Hemisphere observations, the increase in frequency of low-concentration ozone values is reproduced only by GC 2×2.5.

While models tend to underestimate ozone increases globally, we find that the model ozone burdens in GC and MERRA2-GMI show global increases throughout the time frame (Table 3), suggesting that the models capture at least some portion of the global ozone increase from 1980–2017. However, each of our simulations shows a smaller ozone burden than previous analyses and model intercomparisons (Table 4). MERRA2-GMI gives an overall ozone burden that is ∼10 % lower than other estimates on average. In GC simulations, the magnitude of the ozone burden is considerably lower (by ∼14 %–18 %) than in a previous version (GEOS-Chem v10-01) and other model intercomparisons. Table 4 also summarizes chemical production, chemical loss, and dry-deposition terms, and these are all lower in GC than in most other models. The only term in the ozone budget to increase between model versions is STE, which increases from the earlier version by 161 Tg yr-1 on average and places it in the range of other models.

Systemically low model ozone burdens, especially in the northern midlatitude free troposphere, are a known issue in recent versions of GEOS-Chem (Mao et al., 2021; Murray et al., 2021). We find that the underprediction of free tropospheric ozone persists across the last 4 decades of simulations, particularly in winter–springtime midlatitudes to high latitudes. While surface ozone tends to be overestimated by GC (as well as MERRA2-GMI), FT ozone in GC is underestimated by ∼10 ppb (Fig. S8). These underestimates may be caused by recent model developments such as improved halogen chemistry (Wang et al., 2021) or NOy reactive uptake by clouds (Holmes et al., 2019) that have increased sinks of ozone or NOx. Neglect of lightning-produced oxidants may also be responsible for the ozone underestimates (Mao et al., 2021). Shah et al. (2022) found that including particulate nitrate photolysis in a recent version of GEOS-Chem increases ozone concentrations by up to 5 ppb in the northern extratropics FT, which is not yet included in the model but will help to resolve this discrepancy in future analyses. By comparison, MERRA2-GMI and the earlier version of GEOS-Chem, both without the above model updates, nearly ubiquitously show ozone values that are much higher and closer to observations, and values are within 5 % of observations at northern midlatitudes in both simulations, although MERRA2-GMI tends to overestimate FT ozone at midlatitudes and high latitudes (Fig. S8) (Hu et al., 2017). However, it is important to note that the earlier version of GEOS-Chem does not perform better than the more recent version in capturing long-term trends (Fig. S9), as it yields less than 10 % of observed trends from 1990–2010. Such widespread model underestimation of tropospheric ozone across a long period highlights the need for better understanding of the processes that promote ozone production, such as VOC chemistry, biomass burning emissions, or the chemical evolution of smoke plumes (Bourgeois et al., 2021; von Schneidemesser et al., 2016). Improvements are especially important in the FT, where long-term transport of ozone is critical to understanding tropospheric ozone trends.

Previous analyses have identified significant challenges facing models in reproducing observed tropospheric ozone trends in recent decades (Parrish et al., 2014; Young et al., 2018). In the Northern Hemisphere midlatitudes, chemistry–climate models were only able to reproduce ∼50 % of the observed ozone trend (Parrish et al., 2014), consistent with our current analysis using chemical transport models. Another analysis using GEOS-Chem from 1995–2017 found that the model underestimated global ozone trends compared to aircraft measurements and that aircraft emissions are a potential source of trend underestimation in the model (Wang et al., 2022). Tarasick et al. (2019) also pointed out the role of data representativeness: uncertainty in estimated observational trends stems largely from data representativeness rather than the accuracy of historical data, pointing to the importance of increasing ozone monitoring station number and frequency, especially when the evaluation of model skill necessarily relies on comparison to sparse datasets. The models examined in this work capture the general tendency of increasing ozone from 1980–2017, and the multi-model average increase in global tropospheric ozone burden is 10 % or 28 Tg (Table 3). However, they often underestimate tropospheric ozone trends at globally distributed sites (60 % of trend captured with MERRA2-GMI, <15 % for GC). Our findings that models are not able to reproduce recent ozone trends contrast with an analysis of GEOS-Chem and GISS-E2.1 that found the model accurately reproduced preindustrial ozone concentrations (Yeung et al., 2019). Notably, the GEOS-Chem simulations in that analysis were performed by running the standard model without anthropogenic combustion and fertilizer sources. This result implies that a large issue in reproducing recent decadal trends may come from uncertainties in anthropogenic emissions, including neglected precursor emissions (Granier et al., 2011; Hassler et al., 2016) and underestimated aircraft emissions (Wang et al., 2022). Although the Yeung et al. (2019) results imply that natural sources are well represented in models, natural sources of NOx and VOCs such as lightning, biogenic emissions, and soils are subject to many uncertainties: this includes land surface properties, the impact of land use change on biogenic VOC emissions and ozone dry deposition (Tai et al., 2013; Fu and Tai, 2015), meteorological variables, and the sensitivity of ozone chemistry to emissions (Young et al., 2018; Banerjee et al., 2014; Hudman et al., 2012).

Another possible source of uncertainty in reproducing ozone trends is model representation of STE, which plays an important role in driving interannual variability and helps explain ozone changes that are not attributable to emissions changes alone (Liu et al., 2017, 2020; Ordóñez et al., 2007). Previous studies have suggested that STE has increased over the last few decades (Neu et al., 2014; Griffiths et al., 2020) and is projected to increase over the next century due to increasing greenhouse gas emissions that strengthen Brewer–Dobson circulation, enhancing mean advective transport (Butchart et al., 2006; Hegglin and Shepherd, 2009; Abalos et al., 2019). This increase in STE has been found to contribute to increases in tropospheric ozone in regions including North America, China, and the Southern Hemisphere (Liu et al., 2020; Xu et al., 2018; Lu et al., 2019). Recent analyses using an earlier version of GEOS-Chem suggest that STE in models may not be sufficient at high northern latitudes (Hu et al., 2017; Jaeglé et al., 2017). An issue with CTM simulations is that they require the aggregation of meteorological fields from their native resolution both spatially and temporally, which can cause losses in transport, especially vertical transport (Yu et al., 2018). Of the models we evaluate, MERRA2-GMI most accurately captures trends from 800–400 hPa and at the surface, perhaps due to its finer resolution that allows the meteorological products to be used at native resolution (∼50 km). The coarse resolution of the GC simulations means that remote sites can exist in the same grid cell as urban areas, limiting accurate representation of ozone in areas with sharp gradients (Lin et al., 2017).

Sensitivity simulations can provide further evidence behind model issues in reproducing ozone trends. Using the GC 4×5 simulation, we perform two sensitivity tests to examine the impact of emissions and meteorology on ozone trends from 1980–2017: (1) constant anthropogenic emissions (“Meteorology”) and (2) constant meteorology (“Emissions”). In the “Meteorology” simulation, all changes in ozone concentrations result from changes in meteorology, as anthropogenic emissions are cycled annually at 1980 values. Note that, in the “Meteorology” simulation, only anthropogenic emissions from the CEDS inventory are cycled (e.g., NOx, SO2, CO, NH3, NMVOCs, black carbon, and organic carbon). Conversely, in the “Emissions” simulation, ozone changes stem from changes in emissions, with the meteorology cycled annually at 1980 values. Additionally, we examine stratospheric influence on tropospheric ozone using MERRA2-GMI's STO3 tracer (described in detail in Sect. 2.2.2) (Liu et al., 2020). Comparison of STO3 trends and tropospheric ozone trends within the MERRA2-GMI model can reveal the extent to which model trends at a given location are driven by stratospheric ozone. This can be a substantial effect, and a previous analysis by Griffiths et al. (2020) found that an increase in STE drove a small increase in tropospheric ozone burden from 1990–2010.

Figure 11 shows that, in GEOS-Chem, ozone trends at different altitudes are driven by different processes. At higher altitudes (i.e., 400 hPa), dynamics are an important driver of base GC trends in Europe and the NH Polar region. Here, the “Meteorology” simulation accounts for the majority of trends in the base simulation at 400 hPa, while the “Emissions” simulation shows opposite trends to the base. This result suggests that changing meteorological fields and dynamics such as intra-hemispheric transport and vertical transport from the stratosphere drive the ozone changes in the base simulation over these regions. At 600 and 800 hPa, meteorological fields still play a role in driving base simulation ozone trends, but emissions play a larger role closer to the surface. Non-anthropogenic emissions (e.g., soil NOx, lightning, or biogenic VOCs) are not held constant in the “Meteorology” run, and some of the ozone trend contribution at lower altitudes in this simulation may also be attributed to these natural emissions.

Figure 12, which investigates the role of STO3 in explaining ozone trends in MERRA2-GMI, also shows the importance of transport for understanding ozone trends. At 400 hPa over Europe, North America, and the NH Polar region, ozone trends are largely attributable to the stratospheric ozone influence. This aligns with the GEOS-Chem sensitivities that suggest meteorological inputs drive model trends at 400 hPa. Stratospheric influence is also prevalent at lower pressure levels for Europe and North America, consistent with previous analyses of ozone trends over these regions (Liu et al., 2020; Ordóñez et al., 2007). At the surface, the influence of STE is negligible in all regions. Importantly, MERRA2-GMI captures trends at 400 hPa remarkably well in Europe, North America, and the NH Polar region, which can be attributed to the ability of MERRA2-GMI to capture STE, likely due to its high resolution (Knowland et al., 2017). Model ability to capture vertical transport is important in reproducing ozone trends. GEOS-Chem and MERRA2-GMI show similar stratospheric trends (Fig. S10) but different trends at 400 hPa (Figs. 11 and 12), suggesting that transport from the stratosphere is most important for capturing trends at 400 hPa. Increases in MERRA2-GMI STO3 in the troposphere may stem from both changes in STE dynamics and recovery of the ozone hole. MERRA2-GMI has been shown to capture a decrease in lower stratospheric ozone in the northern extratropics from 1998–2016, when ozone-depleting substances were no longer increasing (Wargan et al., 2017). This decreasing trend was attributed to changes in lower stratospheric ozone circulation that may be due to climate change, but evidence for this is unclear. This decrease is offset by an increase in upper stratospheric ozone due to ozone layer recovery. The extent to which either dynamics or ozone recovery impacts increasing STO3 in MERRA2-GMI is currently difficult to quantify.

In the other regions examined (Hawaii, Japan, and the Southern Hemisphere), the “Emissions” simulation is able to explain more of the simulated ozone trend than the “Meteorology” simulation. MERRA2-GMI agrees with GEOS-Chem in Japan and the Southern Hemisphere in that transport of ozone, either horizontally or from the stratosphere, does not explain ozone trends well at most pressure levels. This is in contrast with a recent analysis from Lu et al. (2019), which attributes observed Southern Hemisphere ozone changes primarily to changes in large-scale dynamics, although their focus was austral autumn. The large uncertainty bars in Figs. 11 and 12, which represent the standard deviation of trends across sites, show that the magnitudes of ozone trends and the primary drivers of these trends can vary across individual sites in a region. Future work must therefore focus on optimizing both emissions estimates and transport parameterizations in models to best capture observed ozone trends. Our model evaluations also reveal that the recent version of the GEOS-Chem model underpredicts free tropospheric ozone over the past 4 decades, particularly in the winter–springtime northern extratropics. Such widespread model underestimation of tropospheric ozone highlights the need for better understanding of processes that promote model ozone production.