Ground-based FTIR O3 retrievals from the 3040 cm−1 spectral range at Xianghe, China

In this study, we present O3 retrievals from ground-based Fourier-transform infrared (FTIR) solar absorption measurements between June 2018 and December 2019 at Xianghe, China (39.75 ◦N, 116.96 ◦E). The FTIR spectrometer at Xianghe is operated with indium gallium arsenide (InGaAs) and indium antimonide (InSb) detectors, recording the spectra between 1800 and 11000 cm−1. As the harmonized FTIR O3 retrieval strategy (Vigouroux et al., 2015) within the Network for the Detection of Atmospheric Composition Change (NDACC) uses the 1000 cm−1 spectral range, we designed an alternative O3 retrieval 5 strategy in the 3040 cm−1 spectral range at Xianghe. The retrieved O3 profile is mainly sensitive to the vertical range between 5 and 40 km, and the degree of freedom for signal is 2.4±0.3 (1σ), indicating that there are two individual pieces of information in partial columns between the surface and 20 km and between 20 and 40 km. According to the optimal estimation method, the systematic and random uncertainties of the FTIR O3 total columns are about 13.6% and 1.4%, respectively. The random uncertainty is consistent with the observed daily 10 standard deviation of the FTIR retrievals. To validate the FTIR O3 total and partial columns, we apply the same O3 retrieval strategy at Maïdo, Reunion Island (21.08 ◦N, 55.38 ◦E). The FTIR O3 (3040 cm−1) measurements at Xianghe and Maïdo are then compared with the nearby ozonesondes at Beijing (39.81 ◦N, 116.47 ◦E) and at Gillot (20.89 ◦S, 55.53 ◦E), respectively, as well as with co-located TROPOspheric Monitoring Instrument (TROPOMI) satellite measurements at both sites. In addition at Maïdo, we compare 15 the FTIR O3 (3040 cm−1) retrievals with the standard NDACC FTIR O3 measurements using the 1000 cm−1 spectral range. It is found that the total columns retrieved from the FTIR O3 3040 cm−1 measurements are underestimated by 5.5 9.0 %, which is mainly due to the systematic uncertainty in the partial column between 20 and 40 km (about -10.4%). The systematic uncertainty in the partial column between surface and 20 km is relatively small (within 2.4%). By comparison with other 1 https://doi.org/10.5194/amt-2020-127 Preprint. Discussion started: 4 May 2020 c © Author(s) 2020. CC BY 4.0 License.

A ground-based FTIR spectrometer (Bruker IFS 125HR) has been installed in June 2018 at Xianghe (39.75 • N,116.96 • E; 50 m a.s.l.) to measure the atmospheric carbon dioxide, methane and carbon monoxide (Yang et al., 2019). The FTIR instrument at Xianghe is operated with indium gallium arsenide (InGaAs) and indium antimonide (InSb) detectors, recording the spectra with a spectral range from 1800 to 11000 cm −1 . Therefore, the NDACC standard O 3 retrieval strategy cannot be applied directly to the Xianghe spectra. Several other infrared microwindows, which have been applied to retrieve O 3 from the ground-based 5 FTIR spectra: Lindenmaier et al. (2010) summarized all the related FTIR O 3 studies, and it appears that the 3040 cm −1 range is often used within the ground-based FTIR community. Takele Kenea et al. (2013) used six micro-windows in the spectral range of 3039.37-3051.90 cm −1 for the O 3 retrieval at Addis Ababa, Ethiopia. García et al. (2014) tested O 3 retrievals in both 3040 and 4030 cm −1 ranges at Izaña, Spain, and they found that the precision of O 3 total column retrievals from the 3040 cm −1 range is 2%, which is much better than the 5% precision obtained in the 4030 cm −1 range. However, they found that 10 the total column of O 3 from the 3040 cm −1 range is about 7% smaller than that retrieved in the standard NDACC 1000 cm −1 range.
The aim of this paper is to study the FTIR O 3 retrieval in the 3040 cm −1 spectral range at Xianghe, and to evaluate the retrieval uncertainty. Section 2 presents the retrieval strategy and the characteristics of the FTIR O 3 retrieval at Xianghe. After that, we show the time series and seasonal variations of FTIR O 3 retrievals between June 2018 and December 2019. In section 3, 15 the same retrieval strategy is applied to Maïdo, Reunion Island (21.08 • N, 55.38 • E; 2155 m a.s.l.), which is a NDACC-IRWG affiliated instrument. At both sites, we compare the FTIR O 3 measurements with the nearby ozonesonde measurements and the co-located TROPOspheric Monitoring Instrument (TROPOMI) satellite measurements. In addition, the FTIR O 3 retrievals (3040 cm −1 ) are compared to standard NDACC FTIR O 3 retrievals (1000 cm −1 ) at Maïdo. Finally, the conclusions are drawn in Section 4. 20

FTIR O 3 retrievals at Xianghe
The FTIR site at Xianghe Observatory of Whole Atmosphere is operated by the Institute of Atmospheric Physics (IAP), the Chinese Academy of Sciences (CAS). The FTIR system includes a Bruker IFS 125HR instrument, an automatic weather station and a sun tracker system (Yang et al., 2019). The spectra suitable for O 3 retrievals are recorded with a maximum optical path difference of 180 cm, corresponding to a spectral resolution of 0.005 cm −1 . One specific optical bandpass filter (2000 -4000 25 cm −1 ) is inserted in front of the InSb detector in order to improve the signal-to-noise (SNR). The mean SNR of the spectra used in this study is about 1400.

Retrieval strategy
The SFIT4_v9.4.4 algorithm (Pougatchev et al., 1995) is applied to retrieve the O 3 profile using the optimal estimation method (OEM) (Rodgers, 2000) 30 where x r , x a and x t are retrieved, a priori and true state vectors (all retrieved parameters) and A is the averaging kernel, representing the sensitivity of the retrieved parameters to the true status. The SFIT4 algorithm minimizes the cost function where y and F (x) are the observed and fitted spectra, respectively, S is the measurement covariance matrix and S a is the 5 a priori covariance matrix. J (x) is the combination of the measurement information and the a priori information, with their weightings determined by S and S a . S is derived from the SNR of the spectra, with its diagonal values set to 1/SNR 2 and off-diagonal values to 0. S a is derived from the covariance matrix of the Whole Atmosphere Community Climate Model (WACCM) v6 O 3 monthly means between 1980 and 2020. The square root of the diagonal elements of S a are about 3% near the surface, 2% in the troposphere, 2.5% in the stratosphere and 1% above the stratosphere. re-analysis data. For the a priori profiles of O 3 and other interfering species, we use the mean of the WACCM model data between 1980 and 2020. Since the broadening effect of absorption lines is related to the pressure and temperature, we can obtain limited vertical information of O 3 by fitting the spectra.

Uncertainty estimation
According to Rodgers (2000), the error ( r = x r − x t ) of the retrieved O 3 profile is where b t and b are the true and used model parameters, e. g. solar zenith angle (SZA), spectroscopy, temperature; I is the unit matrix; G y is the contribution matrix; K b is the Jacobian matrix for the model parameters; m is the noise of the spectra. can be divided into three portions (Zhou et al., 2016), corresponding to smoothing (from the O 3 profile), interfering species and retrieval parameters in Table 2.

10
The m is derived from the SNR. The systematic uncertainties of both O 3 and interfering species a priori profiles are set to 10%, and their random uncertainties are derived from the WACCM data. According to the ATM2019 linelist, the systematic uncertainties of O 3 line intensity, air broadening and pressure broadening are 10-20%, 5-10% and 5-10%, respectively. In this study, we set 15%, 7.5% and 7.5% for the systematic uncertainties of the O 3 line intensity, air broadening and pressure broadening, respectively, and we assume that there are no random uncertainties. The uncertainties for temperature and H 2 O 15 are derived from the difference between NCEP reanalysis data and the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis data, where the mean difference is set as the systematic uncertainty and the standard deviation (STD) of the differences is set as the random uncertainty. The systematic uncertainty of the temperature profile is about 0.5 K for the whole altitude range, and its random uncertainty is about 2 K below 2 km and 1 K above. The random and systematic uncertainties for SZA are set to 0.5% and 0.1%, respectively. Table 2 shows the resulting total uncertainty on the retrieved O 3 20 total column and two partial columns. The systematic uncertainty is dominated by the uncertainty from the spectroscopy. The random uncertainty of the total column is 1.4%, which is coming mainly from the SZA and interfering species uncertainties.
SZA uncertainty. The random uncertainty of the upper partial column (20-40 km) is 2.2%, which comes mainly from the smoothing error and retrieval parameters uncertainties. To check the estimated random uncertainty, we calculated the mean of daily STD for all days with more than 4 measurements (see Table 2). Keep in mind that daily STD still includes the signal of the diurnal variation, therefore, it might be slightly larger the random uncertainty. In general, the STDs of the total column and the two partial columns are close to the estimated uncertainties, indicating that the random uncertainties have been estimated 5 correctly.  total column is 8.70 × 10 18 molecules/cm 2 , and the mean partial columns between the surface and 20 km, and between 20 and 40 km are 3.42 × 10 18 molecules/cm 2 and 5.05 × 10 18 molecules/cm 2 , respectively. The lower partial column (surface-20 km) has a minimum in August-September and a maximum in February-April, while the upper partial column (20-40 km) has a minimum in October-December and a maximum in May-July. The peak-to-peak amplitude of the seasonal variation in the partial column between surface and 20 km is 1.3 × 10 18 molecules/cm 2 , which is much larger than that in the partial column between 20 and 40 km of 0.4 × 10 18 molecules/cm 2 . Therefore, the seasonal variation of the total column is dominated by the lower partial column (surface-20 km). The FTIR O 3 retrieved lower partial column (surface-20 km) has a maximum sensitivity in the UTLS region (see Figure 2). The ozonesonde measurements between 2002 and 2010 at Beijing (Wang et al., 2012) showed that the high O 3 concentrations are in the UTLS in later winter and spring with a year-to-year variation, and the low O 3 5 concentrations in the UTLS in August-September. In the middle and upper stratosphere (20-40 km), the maximum observed in summer is mainly due to the higher photochemical production in this season (Perliski et al., 1989). On the purpose of validating the FTIR O 3 retrievals at Xianghe in the 3040 cm −1 spectral range, we first compare them with nearby ozonesonde and co-located TROPOMI measurements. Secondly, we apply the same retrieval strategy (3040 cm −1 ) to the FTIR observations at the Maïdo (Reunion Island) which is a NDACC affiliated site, and we compare them with the standard NDACC O 3 retrievals (1000 cm −1 ) at this site, as well as with nearby ozonesonde and co-located TROPOMI measurements. of the IAP ozonesonde measurements has been evaluated by comparison with other ECC ozonesonde measurements (Zhang et al., 2014b): the average difference in the ozone partial pressure between the IAP and ECC ozonesondes is 0.3 mPa from the surface to 2.5 km, close to zero from 2.5 to 9 km and generally less than 1 mPa for layers higher than 9 km. Note that we have applied the pressure pump efficiency corrections to the IAP ozonesonde (Zheng et al., 2018), resulting in higher ozone detecting performance relative to the results in Zhang et al. (2014b). The IAP ozonesonde measurements used in this study  (Thompson et al., 2003).

Time series and seasonal variations
Detailed information about the ozonesonde measurements at Gillot can be found in Thompson et al. (2014); Witte et al. (2017), where the ozonesonde measurements are applied to understand the tropospheric ozone increases over the southern Africa re- 25 gion. Gillot is about 26 km away from Maïdo, and is considered representative for the ozone concentrations at Maïdo (Duflot et al., 2017). The ozonesonde measurements used in this study cover the period between April 2013 and July 2017.
We select FTIR measurements within a ± 3-hours window around each ozonesonde, and take the averaged FTIR retrieval and the ozonesonde measurement as one FTIR-sonde data pair. In total, we have 16 and 53 data pairs at Xianghe and Maïdo, respectively. As the vertical resolution of ozonesondes is much higher than that of the FTIR retrievals, the ozonesonde profiles 30 are smoothed with the FTIR averaging kernel to reduce the smoothing error in the comparison between both (Rodgers and Connor, 2003): where x F,a is the FTIR a priori profile, x s is the ozonesonde profile, x s is the smoothed ozonesonde profile, and A is the FTIR averaging kernel. To apply the smoothing correction, we have extended the ozonesonde profile to the top of atmosphere using the FTIR a priori profile.
The profiles of the FTIR retrievals and ozonesonde measurements, together with their relative differences at Xianghe and Maïdo are shown in Figure 4. In general, the relative difference profiles at these two sites are similar: within ±15% below 5 20 km and between -30 % and 10% between 20 km and 40 km. The total column observed by ozonesonde is 6.4 ± 6.0 (1σ) % and 9.0 ± 4.3 % larger than the FTIR (3040 cm −1 ) retrievals at Xianghe and Maïdo, respectively. To check the impact of the O 3 columns above the maximum height of the ozonesonde, we also compare the FTIR column between the surface to the maximum altitude of each co-located ozonesonde profile, where the ozonesonde measurements are 6.2 ± 6.1 % and 9.7 ± 7.0 % larger than the FTIR retrievals at Xianghe and Maïdo, respectively. As a result, the impact of extending the 10 ozonesonde profile to higher altitude with the FTIR a priori profile is relatively small compared to the large uncertainty. The comparisons between the total and partial columns (surface-20 km and 20-40 km) retrieved from the FTIR and the ozonesonde measurements are listed in Table 3.

TROPOMI satellite measurements
The Sentinel-5 Precursor (S5P) satellite, carrying the TROPOMI instrument, was successfully launched into a sun-synchronous

Dobson and Zenith Scattered Light-Differential Optical Absorption Spectroscopy (ZSL-DOAS) measurements. It is found that
the mean bias between the TROPOMI and ground-based measurements is +0.1% and the STD of the relative differences is about 2.0%, which is within the mission requirements (Garane et al., 2019).
We select TROPOMI satellite OFFL data within a ± 6 hours temporal window and within a ± 1.0 • latitude and ± 3.0 • longitude box of each FTIR O 3 measurement at Xianghe and Maïdo. As the FTIR measurements at Xianghe start in June 5 2018, in this study, we compare the FTIR measurements with TROPOMI OFFL data between June 2018 and December 2019 at both sites. As mentioned in Section 2.1, the FTIR a priori profile is derived from the WACCM model, while the a priori profile of the TROPOMI retrieval is from a column-classified ozone profile climatology (Heue et al., 2018). In order to reduce the influence of different a priori profiles, we substitute the satellite a priori profile for the ground-based FTIR a priori profile when comparing both datasets where x r is the adapted FTIR profile by using satellite a priori profile as the a priori profile; x r is the original FTIR retrieved profile; x s,a and x F,a are the satellite and FTIR a priori profiles. TROPOMI provides the column averaging kernel (A s ) together with the total column, therefore, we applied the smoothing correction to the adapted FTIR profile: 15 where T C s,a is the TROPOMI a priori total column and T C r is the FTIR retrieved total column after a priori profile substitution and taking TROPOMI vertical sensitivity into account. Figure 5 shows the time series of the co-located FTIR and TROPOMI O 3 total columns, together with their differences and correlations at Maïdo and Xianghe. Similar to ozonesonde measurements, the TROPOMI measurements are 5.5 ± 2.0 % and 6.1 ± 1.3 % larger than the FTIR (3040 cm −1 ) total columns at Xianghe and Maïdo, respectively. In addition, there is no clear time dependence in the relative differences between FTIR and TROPOMI total columns.
There is a good correlation between the FTIR and TROPOMI measurements at Xianghe (R=0.99) and Maïdo (R=0.96). The seasonal and synoptic variations (phase and amplitude) of total columns of O 3 from the FTIR and TROPOMI measurements are very close to each other at both sites. As an example, FTIR and TROPOMI measurements show that there is a large 5 enhancement of O 3 total column on 31 January 2019 at Xianghe (see Figure 6). Keep in mind that we should focus on the total column and two partial columns of FTIR measurements instead of the FTIR retrieved O 3 profile due to its limited vertical information. According to the FTIR measurements, both partial columns increase on that day, but the large increase of the total column mainly results from the enhancement of the lower partial column from the surface to 20 km altitude. There is 12 https://doi.org/10.5194/amt-2020-127 Preprint. Discussion started: 4 May 2020 c Author(s) 2020. CC BY 4.0 License. one ozonesonde profile available on 31 January 2019, which confirms that the O 3 mole fraction is much larger compared to the FTIR a priori profile above 10 km, especially in the UTLS region. The smoothed ozonesonde profile is close to the FTIR retrieved profile below 23 km, which is consistent with our results in Table 3.

FTIR (1000 cm −1 ) retrievals
Maïdo is an NDACC station, where FTIR measurements using a MCT detector are carried out (Baray et al., 2013;Zhou et al., 5 2018). The harmonized O 3 standard retrieval strategy using 1000-1005 cm −1 has been performed at Maïdo, so that we can compare the FTIR (3040 cm −1 ) with the FTIR (1000 cm −1 ) retrievals for total column as well as for two partial columns. The FTIR O 3 (1000 cm −1 ) retrieval has a DOFS of about 4 to 5, because in this spectral range it benefits from more O 3 lines with different intensities. The systematic and random uncertainties of the total column from FTIR O 3 (1000 cm −1 ) retrievals are about 3.0% and 1.0%, respectively. The systematic and random uncertainties of the surface to 20 km partial column retrievals 10 are about 3.2% and 2.5%, respectively, and of the 20 to 40 km partial column retrievals are about 3.4% and 1.5%, respectively.
Both precision and accuracy are better using O 3 (1000 cm −1 ) than O 3 (3040 cm −1 ), which explains why the MCT spectral region is preferred at NDACC stations where these measurements are available. The systematic uncertainty is also dominated by the spectroscopy (HITRAN2008; Rothman et al. (2009)), where we set 3% for the uncertainty of line intensity (NDACC-IRWG recommendation, based on total column comparisons with Dobson and Brewer measurements, e.g. in Vigouroux et al. (2008)).

5
The a priori profiles of the FTIR O 3 (1000 cm −1 ) retrievals are the same as those of the FTIR O 3 (3040 cm −1 ) retrievals (see Section 2.1). To take the low vertical resolution of the FTIR (3040 cm −1 ) retrieval into account, the FTIR (1000 cm −1 ) retrieved profile is smoothed with the FTIR (3040 cm −1 ) averaging kernel where x a is the FTIR a priori profile; x 1000 is the FTIR (1000 cm −1 ) retrieved profile, x 1000 is the FTIR (1000 cm −1 ) 10 retrieved profile after smoothing with the FTIR (3040 cm −1 ) averaging kernel (A).
The time series of the hourly means retrieved O 3 total column and partial columns (surface-20 km and 20-40 km) in the 3040 cm −1 and 1000 cm −1 spectral ranges, together with their differences and correlations are shown in Figure 7. Both O 3 datasets show the same seasonal variations in the total column and the two partial columns. The mean and STD of their relative differences are also listed in  (20-40 km) retrieved in the 3040 cm −1 spectral range is 10.8 ± 1.8 % smaller than the one retrieved in the 1000 cm −1 spectral range. García et al. (2014) found that there is an underestimation of 7% in the FTIR O 3 (3040 cm −1 ) total column compared to FTIR O 3 (1000 cm −1 ) retrievals at Izaña based on the HITRAN2012 spectroscopy (Rothman et al., 2013), which is generally 20 in good agreement with our result (8.4 ± 1.1 %) at Maïdo. In this study, we also looked at comparisons between the two partial columns. The biases observed between FTIR O 3 (3040 cm −1 ) and FTIR O 3 (1000 cm −1 ) on one hand and between FTIR O 3 (3040 cm −1 ) and ozonesondes on the other hand are similar (see Table 3), pointing to an underestimation of the FTIR retrieved total and partial columns products in the 3040 cm −1 spectral range; the bias is coming mainly from the 20-40 km partial column bias.  cm −1 ). The resulting averaging kernel shows that the retrieved O 3 profile is mainly sensitive to the vertical range between 5 and 40 km, and the DOFS is 2.4±0.3 (1σ), indicating that we can retrieve two independent partial columns, one from the surface to 20 km and a second one from 20 to 40 km altitude. Based on the optimal estimation method, we have estimated 10 the systematic and random uncertainties of the retrieved FTIR O 3 total columns to be about 13.6% and 1.4%, respectively, in which the random error is generally in good agreement with the observed daily STD of the FTIR retrievals.
The FTIR retrieval systematic uncertainty is then verified by comparing the FTIR O 3 retrievals in the 3040 cm −1 spectral range with nearby ozonesonde and co-located TROPOMI measurements at Xianghe and Maïdo, and with NDACC standard FTIR O 3 (1000 cm −1 ) retrievals at Maïdo. There is a systematic underestimation by 5.5-9.0% in the FTIR O 3 (3040 cm −1 ) 15 total column retrievals, which is within the estimated systematic uncertainty and mainly due to the spectroscopic uncertainties.
According to ozonesonde measurements and standard NDACC FTIR O 3 retrievals, the underestimation of the FTIR (3040 cm −1 ) O 3 total column mainly results from the underestimation by 10.1-10.8% in the upper partial column (20-40 km). The systematic uncertainty is relatively small in the lower partial column (surface-20 km), which is within 2.4%.
At Xianghe, the FTIR retrieved O 3 partial columns between surface and 20 km show a maximum in February-April 20 and a minimum in August-September, with a peak-to-peak amplitude of 1.3 × 10 18 molecules/cm 2 , while the 20-40 km partial columns show a maximum in May-July and a minimum in October-December, with a peak-to-peak amplitude of 0.4 × 10 18 molecules/cm 2 . As the amplitude of the seasonal variation in the lower partial column (surface-20 km) is much larger than the one in the upper partial column (20-40 km), the seasonal variation of the total column is dominated by the lower partial column. for useful discussion about the ozonesonde measurements, and Ball William (ETH zürich) for useful discussions. We also acknowledge the NDACC-IRWG network for providing the retrieval code and data, and ESA for providing the TROPOMI products. The work done by MZ and BIRA colleagues has been supported through the Copernicus Atmospheric Monitoring Service contracts (CAMS-84 and CAMS-27). Feng, Z., De Marco, A., Anav, A., Gualtieri, M., Sicard, P., Tian, H., Fornasier, F., Tao, F., Guo, A., and Paoletti, E.: Economic losses due to ozone impacts on human health, forest productivity and crop yield across China, Environ. Int., 131, 104 966,