The Infrared Atmospheric Sounding Interferometer (IASI) is a nadir-viewing Fourier transform spectrometer designed to measure the thermal infrared emission of the Earth–atmosphere system between 645 and 2760 cm-1. Measurements are taken from the polar sun-synchronous orbiting meteorological Metop series of satellites, every 50 km along the track of the satellite at nadir and over a swath of 2200 km across the track. With more than 14 orbits a day and a field of view of four simultaneous footprints of 12 km at nadir, IASI provides global coverage of the Earth twice a day at about 09:30 and 21:30 mean local solar time.

The Metop program consists of a series of three identical satellites successively launched to ensure homogenous measurements of atmospheric parameters covering more than 15 years. Metop-A and -B were successively launched in October 2006 and September 2012, respectively. The third and last satellite was launched in November 2018 on board Metop-C. In addition to its exceptional spatio-temporal coverage, IASI also provides good spectral resolution and low radiometric noise, which allows the measurement of a series of gas-phase species and aerosols globally (e.g. Clerbaux et al., 2009; Hilton et al., 2012; Clarisse et al., 2019).

In this study, we use the O3 profiles retrieved by the Fast Optimal Retrievals on Layers for IASI (FORLI-O3; version 20151001) near-real time processing chain set up at ULB (see Hurtmans et al., 2012, for a description of the retrieval parameters and the FORLI performances). The FORLI algorithm relies on a fast radiative transfer and a retrieval methodology based on the optimal estimation method (Rodgers, 2000), which requires a priori information (a priori profile and associated variance–covariance matrix). The FORLI-O3 a priori information consists of one single profile and one covariance matrix built from the global Logan–Labow–McPeters climatology (McPeters et al., 2007). The profiles are retrieved on a uniform 1 km vertical grid on 41 layers from surface to 40 km with an extra layer from 40 km to the top of the atmosphere considered at 60 km. Previous characterization of the FORLI-O3 profiles (Wespes et al., 2016) have demonstrated a good vertical sensitivity of IASI to the O3 measurement, with up to four independent levels of information on the vertical profile in the troposphere and the stratosphere (MUSt; LSt; upper troposphere-lower stratosphere – UTLS – 300–150 hPa; middle-low troposphere – MLT – below 300 hPa). The two stratospheric layers that show distinctive patterns of O3 distributions over the IASI decade (Fig. 1a) are characterized by high sensitivity (degree of freedom for signal – DOFS > 0.85; Fig. 1b) and low total retrieval errors (<5 %; see Hurtmans et al., 2012, and Wespes et al., 2016). The decorrelation between the MUSt and the LSt is further evidenced in Fig. 1d, which shows low correlation coefficients (<0.4) between the mean absolute deseasonalized anomalies (as calculated in Wespes et al., 2017) in the two layers (Fig. 1c). Note that the highest correlation coefficients over the Antarctic (∼0.4) are due to the smaller vertical sensitivity of the IASI measurements over cold surfaces (Clerbaux et al., 2009). The latest validation exercises for the FORLI-O3 product have demonstrated a high degree of precision with excellent consistency between the measurements taken from the two IASI instruments on Metop-A and -B, as well as a good degree of accuracy with biases lower than 20 % in the stratospheric layers (Boynard et al., 2018; Keppens et al., 2018). Thanks to these good IASI-FORLI performances, large-scale dynamical modes of O3 variations and long-term O3 changes can be differentiated in the four retrieved layers (Wespes et al., 2016). The recent validations have, however, reported a drift in the MUSt FORLI-O3 time series from comparison with O3 sondes in the Northern Hemisphere (NH) (∼3.53±3.09 DU decade-1 on average over 2008–2016; Boynard et al., 2018) that was suggested to result from a pronounced discontinuity (“jump”) rather than from a progressive change. Further comparisons with CTM simulations from the Belgian Assimilation System for Chemical ObsErvations (BASCOE; Chabrillat et al., 2018; Errera et al., 2019) confirm this jump that occurred on 15 September 2010 over all latitudes (see Fig. S1 of the Supplement). The discontinuity is suspected to result from updates in level-2 temperature data from Eumetsat that are used as inputs into FORLI (see Hurtmans et al., 2012). Hence, the apparent drift reported by Boynard et al. (2018) likely results from the jump rather than from a progressive “instrumental” drift. This is verified by the absence of drift in the O3 time series after the jump (non-significant drift of -0.38±2.24 DU decade-1 on average over October 2010–May 2017; adapted from Boynard et al., 2018). This is in line with the excellent stability of the IASI Level-1 radiances over the full IASI period (Buffet et al., 2016). From the IASI-BASCOE comparisons, the amplitude of the jump has been estimated as lower than 2.0 DU in the 55∘ S–55∘ N latitude band and 4.0 DU in the 55–90∘ latitude band of each hemisphere. The estimated amplitude of the jump is found to be relatively small in comparison to that of the decadal trends derived in Sect. 4; hence, it cannot explain the trend observed in the IASI dataset. Therefore, the jump is not taken into account in the MLR. The jump values will be, however, considered in the discussion of the O3 trends (Sect. 4).

Finally, the present study only uses the daytime measurements (defined with a solar zenith angle to the sun <83∘) from the IASI-A (aboard Metop-A) instrument, which fully covers the first decade of the IASI mission. The daytime measurements are characterized by a higher vertical sensitivity (e.g. Clerbaux et al., 2009). Quality flags developed in previous IASI studies (e.g. Boynard et al., 2018) were applied a posteriori to exclude data with a poor spectral fit, with less reliability or with cloud contamination.

In an effort to unambiguously discriminate anthropogenic trends in O3 levels from the various modes of natural variability (illustrated globally in Fig. 1c as deseasonalized anomalies), we have applied to the 2.5∘ × 2.5∘ gridded daily MUSt and LSt O3 time series a MLR model that is similar to that previously developed for tropospheric O3 studies from IASI (see Wespes et al., 2017, 2018) but is here adapted to fit the stratospheric variations: 1O3(t)=Cst+xj=1⋅trend+∑n=1:2an⋅cos(nωt)+bn⋅sin(nωt)+∑j=2mxjXnorm,j(t)+ε(t), where t is the number of days, x1 is the trend coefficient in the data, ω=2π/365.25, an, bn, xj are the regression coefficients of the seasonal and non-seasonal variables, and ε(t) is the residual variation (assumed to be autoregressive with time lag of 1 d). Xnorm,j are the m chosen explanatory variables, commonly called “proxies”, which are normalized over the study period (2008–2017) with the following: 2Xnorm(t)=2X(t)-Xmedian/Xmax⁡-Xmin⁡. In addition to harmonic terms that represent the 1-year and 6-month variations, the MLR model includes the anthropogenic O3 response through a linear trend (LT) term and a set of proxies to parameterize the geophysical processes influencing the abundance of O3 in the stratosphere. The MLR uses an iterative stepwise backward elimination approach to retain, at the end of the iterations, the most relevant proxies (within a 95 % confidence level) explaining the O3 variations (e.g. Mäder et al., 2007). Table 1 lists the selected proxies, their sources and their temporal resolutions. The proxies describe the influence of the Quasi-Biennial Oscillation (QBO; visible from the deseasonalized anomaly maps in Fig. 1c with a typical band-like pattern around the Equator) at 10 and 30 hPa, of the North Atlantic and the Antarctic Oscillations (NAO and AAO), of the ENSO, of the volcanic aerosols (AERO) injected into the stratosphere, of the strength of the Brewer–Dobson circulation (BDC) with the Eliassen–Palm flux (EPF), of the polar O3 loss driven by the volume of polar stratospheric clouds (VPSC), of the tropopause height variation with the geopotential height (GEO), and of the mixing of tropospheric and stratospheric air masses with the potential vorticity (PV). The main proxies in terms of their influence on O3 over the period of the IASI mission are illustrated in Fig. 2. The construction of the EPF, VPSC and AERO proxies, which are specifically used in this study, is explained hereafter, while the description of the other proxies can be found in previous IASI studies (Wespes et al., 2016, 2017).

The EPF proxy consists of the normalized upward component of the EP flux crossing 100 hPa and spatially averaged over the 45–75∘ latitude band for each hemisphere. The fluxes are calculated from the NCEP/NCAR 2.5∘ × 2.5∘ gridded daily reanalysis (Kalnay et al., 1996) over the IASI decade. The VPSC proxy is based on the potential volume of PSCs given by the volume of air below the formation temperature of nitric acid trihydrate (NAT) over 60–90∘ north and south and calculated from the ERA-Interim reanalysis and from the MLS climatology of nitric acid (Ingo Wohltmann, personal communication, 2018; Wohltmann et al., 2007; and references therein). The PSC volume is multiplied by the EESC to account for the changes in the amount of inorganic stratospheric chlorine that activates the polar ozone loss. The O3 build-up and the polar O3 loss are highly correlated with wintertime accumulated EP flux and PSC volume, respectively (Fusco and Salby, 1999; Randel et al., 2002; Fioletov and Shepherd, 2003; Rex et al., 2004). These cumulative EP flux and PSC effects on O3 levels are taken into account by integrating the EPF and VPSC proxies over time with a specific exponential decay time according to the formalism of Brunner et al. (2006; see Eq. 4). We set the relaxation timescale to 3 months everywhere, except during the wintertime build-up phase of O3 in the extratropics (from October to March in the NH and from April to September in the Southern Hemisphere – SH) when it is set to 12 months. For EPF, it accounts for the slower relaxation time of extratropical O3 in winter due to its longer photochemical lifetime. For VPSC, the 12-month relaxation time accounts for a stronger effect of stratospheric chorine on spring O3 levels: the maximum of the accumulated VPSC (Fig. 2) coincides with the maximum extent of the O3 hole that develops during springtime and that lasts until November. Note that correlations between VPSC and EPF are possible since the same method is used to build these cumulative proxies. VPSC and EPF are also dynamically anti-correlated to some extent since a strong BDC is connected with warm polar stratospheric temperatures and, hence, reduced PSC volume (e.g. Wohltmann et al., 2007).

The AERO proxy is derived from the aerosol optical depth (AOD) of sulfuric acid only. That proxy consists of latitudinally averaged (22.5–90∘ N: AERO-N; 22.5–90∘ S: AERO-S; and 22.5∘ S–22.5∘ N: AERO-Eq) extinction coefficients at 12 µm calculated from merged aerosol datasets (SAGE, SAM, CALIPSO, OSIRIS, 2-D model simulation and Photometer; Thomason et al., 2018) and vertically integrated over the two IASI stratospheric O3 columns (AERO-MUSt and AERO-LSt). Figure 2 shows the AERO proxies (AERO-N, AERO-S and AERO-Eq) corresponding to the AOD over the whole stratosphere (150–2 hPa), while Fig. 3 represents the latitudinal distribution of the volcanic sulfuric acid extinction coefficients integrated over the whole stratosphere (a) and, separately, over the MUSt (b) and the LSt (c) from 2005 to 2017. The AOD distributions indicate the need for considering one specific AERO proxy for each latitudinal band (AERO-N, AERO-S and AERO-Eq) and for each vertical layer (AERO-MUSt and AERO-LSt). Note that, as an alternative proxy to AERO, the surface area density of ambient aerosol, which represents the aerosol surface available for chemical reactions, has been tested, giving similar results.

Note also that, similarly to what has already been found for tropospheric O3 from IASI (Wespes et al., 2016), several time lags for ENSO (1-, 3- and 5-month lags; namely, ENSO-lag1, ENSO-lag3 and ENSO-lag5) are also included in the MLR model to account for a possible delay in the O3 response to ENSO at high latitudes.

Finally, autocorrelation in the noise residual ε(t) (see Eq. 1 in Wespes et al., 2016) is accounted for in the MLR analysis with time lag of one day to yield the correct estimated standard errors for the regression coefficients. They are estimated from the covariance matrix of the regression coefficients and corrected at the end of the iterative process by the autocorrelation of the noise residual. The regression coefficients are considered significant if they fall in the 95 % confidence level (defined by 2σ level).

In the seasonal formulation of the MLR model, the main proxies (xjXnorm,j; with xj, the regression coefficient, and Xnorm,j, the normalized proxy) are split into four seasonal functions (xsprXnorm,spr+xsumXnorm,sum+xfallXnorm,fall+xwintXnorm,wint) that are independently and simultaneously adjusted for each grid cell (Wespes et al., 2017). Hence, the seasonal MLR adjusts four coefficients (instead of one in the annual MLR) to account for the seasonal O3 response to changes in the proxy. If that method avoids over-constraining the adjustment by the year-round proxies and, hence, reduces the systematic errors, the smaller daily data points covered by the seasonal proxies translate to a lower significance of these proxies. This is particularly true for EPF and VPSC, which compensate each other by construction. As a consequence, the annual MLR is performed first in this study and then complemented with the seasonal one when it is found helpful for further interpreting the observations.

Figure 4 shows the latitudinal distributions of the O3 columns in the two stratospheric layers over the IASI decade (first panels in Fig. 4a and b), as well as those simulated by the annual MLR regression model (second panels) along with the regression residuals (third panels). The root mean square error (RMSE) of the regression residual and the contribution of the MLR model to the IASI O3 variations (calculated as σO3Fitted_model(t)σO3(t), where σ is the standard deviation relative to the regression model and to the IASI time series; bottom panels) are also represented (bottom panels). The results indicate that the model reproduces ∼ 25 %–85 % and ∼ 35 %–95 % of the daily O3 variations captured by IASI in the MUSt and the LSt, respectively, with the best representation in the tropics and the worst around the SH polar vortex, and that the residual errors are generally lower than 10 % everywhere for the two layers, except for the spring O3 hole region in the LSt. The RMSE relative to the IASI O3 time series are lower than 15 and 20 DU at global scale in the MUSt and the LSt, respectively, except around the SH polar vortex in the LSt (∼30 DU). On a seasonal basis (figure not shown), the results are only slightly improved: the model explains ∼ 35 %–90 % and ∼ 45 %–95 % of the annual variations and the RMSEs are lower than ∼12 and ∼23 DU everywhere, in the MUSt and the LSt, respectively. These results verify that the MLR models (annual and seasonal) reproduce well the time evolution of O3 over the IASI decade in the two stratospheric layers and, hence, that they can be used to identify and quantify the main O3 drivers in these two layers (see Sect. 3).

The MLR model has also been tested on nighttime FORLI-O3 measurements only and simultaneously with daytime measurements, but this resulted in a lower-quality fit, especially in the MUSt over the polar regions. This is due to the smaller vertical sensitivity of IASI during nighttime measurements, especially over cold surfaces, which causes larger correlations between stratospheric and tropospheric layers (e.g. 40 %–60 % at high northern latitudes versus ∼ 10 %–20 % for daytime measurements based on deseasonalized anomalies) and, hence, which mixes counteracting processes from these two layers. For this reason, only the results for the MLR performed on daytime measurements are presented in this paper.

The distributions of the linear trend estimated by the annual regression are represented in Fig. 8a for the MUSt and the LSt (left and right panels, respectively). In agreement with the early signs of O3 recovery reported for the extratropical middle–upper stratosphere above ∼ 25–10 hPa (> 25–30 km; Pawson et al., 2014; Harris et al., 2015; Steinbrecht et al., 2017; Sofieva et al., 2017; Ball et al., 2018), the MUSt shows significant positive trends larger than 1 DU yr-1 poleward of ∼35∘ N–S (except over Antarctica). The corresponding decadal trends (>10 DU decade-1) are much larger than the discontinuity of ∼ 2–4 DU encountered in the MUSt record on 15 September 2010 and discussed in Sect. 2.1. The tropical MUSt also shows positive trends but they are weaker (<0.8 DU yr-1) or not significant. The largest increase is observed in polar O3 with amplitudes reaching ∼2.0 DU yr-1. The mid-latitudes also show significant O3 enhancement, which can be attributed to air mass mixing after the disruption of the polar vortex (Knudsen and Grooss, 2000; Fioletov and Shepherd, 2005; Dhomse et al., 2006; Nair et al., 2015).

As in the MUSt, the LSt is characterized in the southern polar latitudes by significantly positive and large trends (between ∼1.0 and 2.5 DU yr-1). In the mid-latitudes, the lower stratospheric trends are significantly negative, i.e. opposite to those obtained in the MUSt. This highlights the independence between the two O3 layers sounded by IASI in the stratosphere. Poleward of 25∘ N the negative LSt trends range between ∼-0.5 and -1.7 DU yr-1. Negative trends in lower stratospheric O3 have already been reported in extra-polar regions from other space-based measurements (Kyrölä et al., 2013; Gebhardt et al., 2014; Sioris et al., 2014; Harris et al., 2015; Nair et al., 2015; Vigouroux et al., 2015; Wespes et al., 2016; Steinbrecht et al., 2017; Ball et al., 2018) and may be due to changes in stratospheric dynamics at the decadal timescale (Galytska et al., 2019). These previous studies, which were characterized by large uncertainties or resulted from composite-data merging techniques, are confirmed here using a single dataset. The negative trends which are observed at lower stratospheric middle latitudes are difficult to explain with chemistry-climate models (Ball et al., 2018). It is also worth noting that the significant MUSt and LSt O3 trends are of the same order as those previously estimated from IASI over a shorter period (from 2008 to 2013) and latitudinal averages (see Wespes et al., 2016). This suggests that the trends are not very sensitive to the natural variability in the IASI time series, hence supporting the significance of the O3 trends presented here.

The sensitivity of IASI O3 to the estimated trend from MLR is further verified in Fig. 8b, which represents the global distributions of relative differences in the RMSE of the regression residuals obtained with and without a linear trend term included in the MLR model ((RMSE_w/o_LT - RMSE_with_LT)/RMSE_with_LT × 100; in percentage). An increase of ∼ 1.0 %–4.0 % and ∼ 0.5 %–2.0 % in the RMSE is indeed observed for both the MUSt and the LSt, respectively, in regions of significant trend contribution (Fig. 8a), when the trend is excluded. This demonstrates the significance of the trend in improving the performance of the regression. Another statistical method that can be used for evaluating the possibility to infer, from the IASI time period, the significant positive or negative trends in the MUSt and the LSt, respectively, consists of determining the expected year when these specified trends would be detectable from the available measurements (with a probability of 90 %) by taking into account the variance (σε2) and the autocorrelation (Φ) of the noise residual according to the formalism of Tiao et al. (1990) and Weatherhead et al. (1998). The 95 % confidence interval for that expected trend detection year can also be determined. Such a method has already been used for evaluating the trends derived from IASI in the troposphere (Wespes et al., 2018). It represents a more drastic and conservative method than the standard MLR. The results are displayed in Fig. 8c for an assumed specified trend of |1.5| DU yr-1, which corresponds to a medium amplitude of trends derived here above from the MLR over the mid-polar regions (Fig. 8a). In the MUSt, we find that ∼2–3 additional years of IASI measurements would be required to unequivocally detect a trend of |1.5| DU yr-1 (with probability 0.90) over high latitudes (detectable from ∼ 2020–2022 ± 6–12 months), whereas it should already be detectable over the middle and lower latitudes (from ∼ 2015 ± 3–6 months). In the LSt, an additional ∼7 years (± 1–2 years) of IASI measurements would be required to categorically identify the probable decline derived from the MLR in northern mid-latitudes, and even more to measure the enhancement in the southern polar latitudes. The longer required measurement period at high latitudes is due to the larger noise residuals in the regression fits (i.e. largest σε) at these latitudes (see Fig. 4a and b). Note that a larger specified trend amplitude would obviously require a shorter period of IASI measurement. We find that only ∼2 additional years would be required to detect a specified trend of |2.5| DU yr-1 which characterizes the LSt at high latitudes (data not shown).

The effect on total O3 of the counteracting trends in the northern mid-latitudes and of the constructive trends in the southern polar latitudes in the two stratospheric layers sounded by IASI is now investigated.

Figure 9 represents the global distributions of the contribution of the MUSt and the LSt into the total O3 columns (Fig. 9a; in percentage), of the adjusted trends for the total O3 (Fig. 9b in DU yr-1) and of the estimated year for a |1.5| DU yr-1 trend detection with a probability of 90 % (Fig. 9c). While no significant change or slightly positive trends in total O3 after the inflection point in 1997 have been reported on an annual basis (e.g. Weber et al., 2018), Fig. 9b shows clear significant changes: a negative trend at northern mid-latitudes and high latitudes (up to ∼2.0 DU yr-1 north of 30∘ N) and positive trend over the southern polar region (up to ∼3.0 DU yr-1 south of 45∘ S). Although counteracting trends between lower and upper stratospheric O3 have been pointed out in the recent study of Ball et al. (2018) to explain the non-significant recovery in total O3, we find from IASI a dominance of the LSt decline that translates to negative trends over some regions of the NH mid-latitudes and high latitudes in TOC (Fig. 9b). This is explained by the contributions of 45 %–55 % from the LSt to the total column, vs. ∼ 30 %–40 % from the MUSt (Fig. 9a) in the mid-latitude and polar regions over the whole year. In addition, the increase in total O3 at high southern latitudes is dominated by the LSt, although both layers positively contribute around Antarctica, compared to the trend distributions in Fig. 8. Note that most previous ozone trends studies, including Ball et al. (2018), excluded the polar regions due to limited latitude coverage of some instruments merged in the data composites.

While the annual MLR shows a significant dominance of LSt trends over MUSt trends in the northern mid-latitudes and significant constructive trends in the southern latitudes, total O3 trends are not ascribed with complete confidence according to the formalism of Tiao et al. (1990) and Weatherhead et al. (1998) discussed in Sect. 4.1. The detectability of a specified trend of |1.5| DU yr-1 (Fig. 9c), which corresponds to the medium trend derived from MLR in middle and high latitudes of both hemispheres (Fig. 9b), would need several years of additional measurements to be unequivocal from IASI on an annual basis (from ∼ 2022–2024 over the mid-latitudes and from ∼2035 over the polar regions). A higher trend amplitude of ∼|2.5| DU yr-1 derived from the MLR would be observable from ∼ 2020–2025 (figure not shown).

The use of the annual MLR could translate to large systematic uncertainties on trends (implying large σε), which induces a longer measurement period required to yield significant trends. These uncertainties could be reduced on a seasonal basis, by attributing different weights to the seasons, which would help in the categorical detection of a specified trend. This is investigated in the subsection below by focusing on the winter and the spring periods.

The reports on early signs of total O3 recovery (Salby et al., 2011; Kuttippurath et al., 2013; Shepherd et al., 2014; Solomon et al., 2016; Kuttippurath and Nair, 2017; Weber et al., 2018) have all focused on the Antarctic region during spring–summer, when the ozone hole area is at its maximum extent, i.e. the LSt O3 levels at minimum values. Kuttippurath et al. (2018) have, in particular, reported a significant reduction in Antarctic O3 loss saturation occurrences during spring. Here we investigate the respective contributions of the LSt and the MUSt to the TOC recovery over the southern latitudes during spring and also during winter when the minima in O3 levels occur in the MUSt (down to ∼60 DU in polar regions), in comparison with the northern latitudes. Figures 10 and 11, respectively, show the SH and the NH distributions of the estimated trends from seasonal MLR (left panels) and of the corresponding year required for a significant detection of |3.0| DU increase per year (right panels) during their respective winter (JJA and DJF; Figs. 10a and 11a) and spring (SON and MAM; Figs. 10b and 11b) for the total, MUSt and LSt O3 (top, middle and bottom panels, respectively). Figure 10a and b clearly show significant positive trends over Antarctica and the southernmost latitudes of the Atlantic and Indian oceans, with amplitudes ranging between ∼ 1 and 5 DU yr-1 over latitudes south of ∼ 35–40∘ S in total, MUSt and LSt O3 (∼3.6±2.7, ∼3.0±1.3, ∼3.6±3.1 and ∼3.7±1.7, ∼1.3±0.7, ∼3.7±1.6 DU yr-1: spatial averages over JJA and SON for the three O3 columns, respectively). These trends over 10 years are much larger than the amplitude of the discontinuity in the MUSt time series (Sect. 2.1) and than the trends estimated in Sect. 4.1 (see Fig. 8 for the MUSt and the LSt) and 4.2 (see Fig. 9 for TOC) over the whole year. In MUSt, significant positive trends are observed during each season over the mid-latitudes and polar latitudes of both hemispheres (Figs. 10 and 11 for the winter and spring periods; the other seasons are not shown here) but more particularly in winter and in spring, where the increase reaches a maximum of ∼4 DU yr-1. In the LSt, the distributions are more complex: the trends are significantly negative in the mid-latitudes of both hemispheres, especially in winter and in spring of the NH, while in spring of the SH, some mid-latitude regions also show near-zero or even positive trends. The southern polar region shows high significant positive trends in winter–spring (see Fig. 10). For the total O3 at middle to high latitudes, given the mostly counteracting trends detected in the LSt and in the MUSt and the dominance of the LSt over the MUSt (∼ 45 %–55 % from the LSt vs. ∼ 30 %–40 % from the MUSt into total O3 over the whole year), these latitudes are governed by negative trends, especially in spring of the NH. High significant increases are detected over polar regions in winter–spring of both hemispheres but more particularly in the SH where the LSt and MUSt trends are both of positive sign.

The substantial winter–spring positive trends observed in MUSt, LSt and total O3 levels at high latitudes of the SH (and of the NH for the MUSt) are furthermore demonstrated to be detectable from the available IASI measurement period (see Fig. 10, right panels: an assumed increase of |3.0| DU yr-1 is detectable from 2016±6 months and from 2018±1 year in the MUSt and the LSt, respectively). The positive trend of ∼4 DU yr-1 measured in polar total O3 in winter–spring would be observable from ∼ 2018–2020 ± 1–2 years and the decline of ∼-3 DU yr-1 in winter–spring of the NH in the LSt would be detectable from ∼ 2018–2020 ±9 months (not shown here). Note that the higher negative trends found above the Pacific at the highest latitudes (see Fig. 10) correspond to the regions with longest required measurement period for significant trend detection and, hence, point to poor regression residuals. About ∼50 % and ∼35 % of the springtime MUSt and LSt O3 variations, respectively, are due to anthropogenic factors (estimated by VPSC × EESC proxy and linear trends in MLR models). This suggests that O3 changes, especially in the LSt, are mainly governed by dynamics, which contributes to a later projected trend-detection year in comparison with the MUSt (Figs. 10 and 11) and which may hinder the O3 recovery process.

Overall, the large positive trends estimated concurrently in the LSt, MUSt and total O3 over the Antarctic region in winter–spring likely reflect the healing of the ozone layer with a decrease in polar ozone depletion (Solomon et al., 2016) and, hence, demonstrate the efficiency of the Montreal Protocol. To the best of our knowledge, these results represent the first detection of a significant recovery in the stratospheric and the total O3 columns over the Antarctic from one single satellite dataset.

Positive trends in total O3 over Antarctica were already determined earlier by Solomon et al. (2016) and by Weber et al. (2018) during September over earlier periods (∼2.5±1.5 DU yr-1 over 2000–2014 and 8.2±6.2 % per decade over 2000–2016, respectively). The larger trends derived from the IASI records (see Fig. 10b; ∼3.7±1.7 DU yr-1 or ∼14.4±5.8 % per decade on average in TOC during SON) suggest that the O3 response could be speeding up due to the accelerating decline of O3-depleting substances (ODSs) resulting from the Montreal Protocol. This has been investigated here by estimating the change in trend in MUSt, LSt and total O3 over the IASI mission. Knowing that the length of the measurement period is an important criterion for reducing systematic errors in the trend coefficient determination (i.e. the specific length of natural mode cycles should be covered to avoid any possible compensation effect between the covariates), the ozone response to each natural driver (including VPSC) taken from their adjustment over the whole IASI period (2008–2017; Sect. 3, Fig. 5) is kept fixed. The linear trend term only is adjusted over variable measurement periods that all end in December 2017, by using a single linear iteratively reweighted least squares regression applied on gridded daily IASI time series, after all the sources of natural variability fitted over the full IASI period are removed (typical examples of linear trend adjustment can be found in Fig. S2 of the Supplement). The discontinuity found in the MUSt IASI O3 records on September 2010 (see Sect. 2.1) is not taken into account in the regression; hence, it might over-represent the trends estimated over periods that start before the jump (i.e. 2008–2017, 2009–2017, 2010–2017). The zonally averaged results are displayed in Fig. 12 for the statistically significant total, MUSt and LSt O3 trends and their associated uncertainty (accounting for the autocorrelation in the noise residuals; within the 95 % confidence level) estimated from an annual regression. Note that the results are only shown for periods starting before 2015, given that considering shorter periods induces larger standard errors associated with the trends. In the LSt, a clear speeding up in the southern polar O3 recovery is observed, with amplitudes ranging from ∼1.5±0.4 DU yr-1 over 2008–2017 to ∼5.5±2.5 DU yr-1 over 2015–2017 on zonal averages. Similarly, a speeding up of the O3 decline at northern mid-latitudes is found with values ranging between ∼-0.7±0.2 DU yr-1 over 2008–2017 and ∼-2.8±1.2 DU yr-1 over 2015–2017. In the MUSt, a weaker increase is observed over the year around ∼60∘ latitude of the SH (from ∼0.8±0.2 DU yr-1 over 2008–2017 to ∼2.5±1.3 DU yr-1 over 2015–2017). Given the positive acceleration in both LSt and MUSt O3 in the SH, this is where the total O3 record is characterized by the largest significant recovery (from ∼1.7±0.7 DU yr-1 over 2008–2017 to ∼8.0±3.5 DU yr-1 over 2015–2017). Surprisingly, the speeding up in the O3 decline in the NH is more pronounced in the total O3 (from ∼-1.0±0.4 DU yr-1 over 2008–2017 to ∼-5.0±2.5 DU yr-1 over 2015–2017) compared to the LSt, despite the opposite trend in MUSt O3. This could reflect the O3 decline observed in the northern latitudes in the troposphere (∼-0.5 DU yr-1 over 2008–2016; see Wespes et al., 2018), which is included in the total column.

Overall, the larger annual significant trend amplitudes derived over the last few years of total, MUSt and LSt O3 measurements, compared with those derived from the whole studied period (Sect. 4.1 and 4.2) and from earlier studies, translate to trends that remain detectable over the increasing uncertainty associated with the shorter and shorter time segments (see Fig. S3 of the Supplement), especially in both LSt and total O3 in the SH. This demonstrates that we progress towards a significant emergence and speeding up of the O3 recovery process in the stratosphere over the whole year. Nevertheless, we calculated that additional years of IASI measurements would help in confirming the changes in O3 recovery and decline over the IASI period (e.g. ∼4 additional years are required to verify the trends calculated over the 2015–2017 segment in the highest latitudes in the LSt). In addition, a longer measurement period would be useful to derive trends over successive segments of the same length that are long enough to reduce the uncertainty, in order to make the trend and its associated uncertainty more comparable across the fit.