Quality assessment of O3 profiles measured by a state-of-the-art ground-based FTIR observing system

HAL is a multi-disciplinary open access archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. Quality assessment of O3 profiles measured by a state-of-the-art ground-based FTIR observing system M. Schneider, F. Hase, T. Blumenstock, A. Redondas, E. Cuevas


10
Ground-based measurements of highly-resolved infrared solar absorption spectra allow ongoing detection of the composition of the atmosphere in a cost-effective manner. They are essential for long-term monitoring and for validating satellite measurements and, thus, they are a vital component of the global atmospheric monitoring system. However, their application as reference measurement requires a precise documenta-15 tion of their quality. This is often done exclusively by theoretical studies. The errors are then calculated by a method suggested by C. D. Rodgers (Rodgers, 2000). These calculations give a good overview of the achievable data quality, however, they depend on the assumed error sources. Therefore, every assessment of data quality should be completed by a comparison to independent measurements of similar or better qual- 20 ity. Ozone is very suited for such an empirical quality assessment. It is an important atmospheric constituent and is monitored since many years by a great variety of measurement techniques. In this work we use ECC-sondes, launched weekly very close to the FTIR measurement site, for an empirical validation of the FTIR O 3 profiles. The FTIR O 3 profiles are obtained by a optimised retrieval approach (Schneider and Hase, edge of O 3 (mean profile and covariances) is taken from an ECC sonde climatology calculated from measurements between 1996 and 2006 as depicted in Fig. 1. It is important to mention that we use the same set of a-priori data for all retrievals. We do not vary our a-priori depending on season, a strategy often applied in other studies (e.g. Barret et al., 2002). This assures that all variability seen in our profile comes from the 10 measurement and can be easily interpreted. The O 3 amounts around the tropopause are highly variable. Under these conditions an inversion performed on a logarithmic scale is superior to a inversion performed on a linear scale (Hase et al., 2004;Schneider et al., 2006a;Deeter et al., 2007). Furthermore, the inversion on a logarithmic scale allows to constrain against isotopologue ratio profiles (Schneider et al., 2006b).

15
As a-priori for the typical ozone isotopologue ratio profiles and their covariances we use data reported by Johnson et al. (2000). The applied temperature a-priori profile is a combination of the data from the local ptu-sondes (up to 30 km) and data supplied by the automailer system of the Goddard Space Flight Center. The spectroscopic line parameters are taken from the HITRAN 2004 database (Rothman et al., 2005). For 20 H 2 O we apply the 2006 updates.

Theoretical quality assessment
When contemplating remotely-sensed vertical distribution profiles it is important to remember the inherent vertical resolution of these data. Figure 2 shows typical averaging kernels for the retrieved 48 O 3 profiles and demonstrates that the FTIR measure-25 ments contain information about the vertical distribution from the surface up to 40 km. The best vertical resolution is achieved between altitudes of 10 and 20 km, where the FWHM (full width half maximum) of the kernels is around 5 km. The trace (sum of diagonal entries) of the averaging kernel matrix is a measure of the degree of freedom in the measurement. It indicates the number of independent layers present in the retrieved profile. Summing up the diagonal entries of the averaging kernel matrix gives a good overview of the layers that are independently presented in the retrieved profile.

5
We identify as independent layer the altitude ranges where the sum of the corresponding diagonal entries reaches unity. The right panel of Fig. 2 gives an overview of these layers. It plots the altitude ranges for which the sum of the kernel matrix's diagonal elements reaches unity (x-axis) versus the altitude where this layer is centred (y-axis). The centre of the layer is the weighted mean (weighted by kernel matrix's diagonal 10 elements) of the altitudes contributing to the layer. It shows that up to 20 km the FTIR observing system is able to distinguish layers with a vertical extension of smaller than 8 km and in the middle stratosphere of around 10 km. The best resolution is achieved at the tropopause, where layers with a extension of 5 km can be distinguished. The uppermost layer that can be resolved extends from 26 km to the top of the atmosphere.

15
Our theoretical error estimation bases on the analytic method suggested by Rodgers (2000), which identifies three error classes: (a) smoothing error, (b) error due to uncertainties in input parameters (instrumental characteristics, spectroscopic data, ...), and (c) errors due to measurement noise: Interactive Discussion and a-priori state,p and p the estimated and real model parameters,ŷ and y the measured and simulated spectrum, and I the identity matrix. This procedure assures a very accurate error analysis. The assumed error sources are listed in Table 1. Figure 3 shows the standard deviation of the errors calculated for the 500 simulations according to Eq. (1). When considering vertically fine structured profiles, the smoothing 5 error is the leading error since the FTIR system only provides sufficient information about rough vertical structures. It reaches 25% in the tropopause region, where the actual profile is highly-structured. The most important parameter errors are remaining ILS distortions (uncertainties in modulation efficiency and phase error), uncertainties in the temperature profile, and errors in the applied pressure broadening parameter.

10
Even though a systematic error source, the pressure broadening coefficient produces random errors sinceĜ andK p of Eq. (1) depend on the actual atmospheric state. Figure 4 depicts the systematic errors of the retrieved profiles (mean value of the errors calculated from the 500 simulations). We find that the uncertainty in the pressure broadening coefficient is the most important systematic error source.

15
If we only consider rough vertical structures (layers with extension of 5-10 km the smoothing error becomes less important. The right panel of Fig. 2 gives an idea about a reasonable choice for the extension of these layers: e.g. a first layer representing the troposphere (surface -10 km, subsequently called the TR layer), a second layer the tropopause region (12.5-17 km, subsequently called the TP layer), and a third layer the 20 middle stratosphere (20-30.5 km, subsequently called the MS layer). Tables 2 and 3 present the error estimations for these layers. We additionally include the estimations for the layer ranging from the surface to 30.5 km which corresponds to the altitudes covered by nearly all ECC sondes. The partial column amount errors of these layers are investigated in great detail. While for the VMR profile errors (Figs. 3 and 4) we 25 restricted the discussion to an estimation of the mean and standard deviation, for we separate these partial column amount errors in random error, systematic sensitivity error, and systematic bias error component. Figure 5  O 3 amount of the MS layer taking the smoothing error as example. The slope of the linear regression line gives the systematic sensitivity error (in our example −7.2%), the offset at the climatological value the systematic bias error (−0.2 DU or −0.1% if referred to the climatological amount of the MS layer), and the scattering around the regression line the random error (2.0 DU or 1.3%). This error treatment is described in Schneider 5 and Hase (2008), which should be consulted for more details.
For the layers as depicted in the right panel of Fig. 2 the sensitivity error does not significantly exceed −10% (i,e. the FTIR system has a sensitivity of around 90%). This is a very satisfactory value, and demonstrates that the high quality measurements together with an advanced retrieval strategy allow an adequate monitoring of these rough 10 atmospheric structures. An artificial increase of this sensitivity by applying seasonally dependent a-priori data is not necessary and would only difficult the interpretation of the FTIR data. Table 3 as well as Fig. 4 show that the systematic errors are dominated by the smoothing error and uncertainties in the spectroscopic line parameters.

15
For this assessment we compare the FTIR O 3 profiles to regularly performed ECCsonde measurements. The ozone sonde program started in November 1992 applying ECC-sondes (type: Scientific Pump 6A). The sondes are launched weekly from Santa Cruz de Tenerife (35 km northeast of the Observatory) and since October 2006 from Güimar (15 km south of the Observatory). In March 2001 Izaña's ECC-sonde to-20 gether with the Brewer, DOAS, and FTIR activities have been accepted by the NDACC (Network for Detection of Atmospheric Composition Change (http://www.ndacc.org/), formerly called NDSC: Network for Detection of Stratospheric Change (Kurylo, 1991(Kurylo, , 2000).
The ECC-sondes generally burst between 30 and 34 km. To use as many sondes as 25 possible and to homogenize the study we use only ECC data measured up to 30.5 km. This altitude is reached by around 90% of all sondes. These criteria provide 53 coin- and 2007 allows to conclude that both techniques have a precision of better than 0.5%. However, Schneider et al. (2008) also observes a systematic difference of 4-5%, which is consistent to laboratory studies of Picquet-Varrault et al. (2005). Therefore, we mainly attribute it to inconsistencies between the spectroscopic data in the UV (Brewer), on the one hand, and the infrared (FTIR), on the other hand.

10
Here we make an analogous brief study for the ECC amounts. As aforementioned we only apply ECC data for altitudes below 30.5 km. However the residual O 3 partial column above this altitude is still around 20% of the total O 3 amount. From the HALOE climatology (Grooß and Russell III, 2005) we deduce a 1σ value for the O 3 variability above 30.5 km of typically 10-15%. Assuming a vertical correlation length 15 of 2.5 km (which corresponds to the length derived from the ECC data around 30 km), we estimate a 1σ variability for the O 3 residual of 4 DU. The left panel of Fig. 6 shows the correlation between Brewer total O 3 amounts and ECC partial O 3 amounts below 30.5 km for all 80 Brewer/ECC coincidences during 2005 and 2006. We calculate a difference of 60.6 ± 6.8 DU. Approximately 4 DU of the scatter between the Brewer 20 and ECC data is caused by the ignorance of the ECC O 3 residual. Since the Brewer total column amounts are very precise (around 1.5 DU), there is a remaining scatter of around 6.8 2 −4 2 −1.5 2 ≈5.3 DU (or around 2.0% if referred to the typical amount), which can be attributed to errors in the ECC data or to the observation of different airmasses by the Brewer, on the one hand, and by the ECC sonde, on the other hand. Introduction of the NDACC ozonesonde PI meeting in Potsdam, July 1998, available at the NDACC web site: www.ndacc.org). It is argued that this approach accounts for part of the real variabilities present in the O 3 residual. The right panel of Fig. 6 depicts the Brewer total O 3 versus the ECC total O 3 calculated for the so-estimated O 3 residual. Naturally the systematic difference is smaller if compared to the left panel. However, we observe a 10 slightly poorer correlation, indicating that, at least at the subtropical site of Izaña, such a simplified estimation of the O 3 residual introduces more noise than real information.

FTIR versus ECC-sonde
When validating remotely-sensed vertical distribution profiles it is important to remember the inherent vertical resolution of these data. There are two possibilities to ade-15 quately validate remotely-sensed profiles: (a) degrade the vertical resolution of the vertically highly-resolved data towards the vertically poorly-resolved data. By this means we exclude the smoothing error from the comparison. In our case the ECC in-situ measurements are vertically highly resolved. (b) Another possibility is to compare only the rough structures that are supposed to be resolvable by the remote sensing measure-20 ments. We estimated these structures in Sect. 3 and depict them in the right panel of Fig. 2. In the following we compare FTIR and ECC profiles applying both method (a) and method (b). The smoothing (or degradation) of the vertically highly-resolved ECC profile x ECC is done by convolving it with the FTIR averaging kernelsÂ: The result is an ECC profile (x ECC ) with the same smoothing error as the FTIR pro-5 file. Consequently the difference between FTIR and smoothed ECC profile eliminates the smoothing error component, which is the leading error component. Equation (2) requires ECC profile data beyond 30.5 km. However, this data is not available and we extend the ECC profile with the zonally averaged HALOE climatological profile used as a-priori in the FTIR retrieval. Consequently the smoothed ECC profile is a combination of two experiments: the ECC and HALOE experiments. Furthermore, applying a climatological profile above 30.5 km introduces additional random errors in the smoothed ECC profile close to 30.5 km. Figure 7 depicts the mean and standard deviation for the difference between FTIR and smoothed ECC profile. These calculations are comparable to the error estimations presented in Figs. 3 and 4. The grey shaded area indicates 15 the total random error of the FTIR profiles (excluding the smoothing error, thick black line in Fig. 3). The light grey shaded area indicated the sum of the FTIR and ECC random errors. As ECC random error we assumed 6% suggested in Smit and Sträter (2004). From the surface up to 26 km we found systematic differences between the FTIR and smoothed ECC data of −9 to +9%. These differences become significant 20 around 12 and 18 km and are probably due to incorrect line parameterisations (error in the pressure broadening coefficient (compare to Fig. 3)). Interactive Discussion a mean 5% (between the surface and 30.5 km). It is highest between 10 and 15 km where it reaches 10%. It is in a satisfactory agreement to the expected errors of around 7% (light grey shaded area of Fig. 7). The slightly higher standard deviation may be due to the observation of different airmasses by FTIR and ECC or due to a weak overestimation of the theoretical vertical resolution of the FTIR. It is important to mention 5 that Fig. 7 provides no comprehensive documentation of the quality of the FTIR profile. This is only possible together with the averaging kernels: Schneider et al. (2005a) also reports an agreement within 5-10% to ECC sondes but for FTIR profiles inverted by applying spectra of poorer quality and non-optimised retrieval strategies. This leads to broader averaging kernels and provides a stronger smoothing of the ECC profile. The 10 agreement is similar, but the compared vertical structures are much rougher.

Comparison of partial column amounts
A straight forward comparison of partial columns has the advantage that the results are easy to interpret. Figure 8 depicts  of around −11% for the TP layer (see Table 3). This estimated FTIR sensitivities are confirmed by the comparison to the ECC amounts. Figure 8 shows  Interactive Discussion bias in the FTIR data is mainly produced by errors in the spectroscopic parameters. We estimate that it is more important for the TR and TP layer than for the MS layer (see Table 3). This is quantitatively confirmed by the intercomparison. If related to the climatological amounts there is a systematic bias between FTIR and ECC amounts of −0.75 DU 21.61 DU =−3.5% and 0.54 DU 12.61 DU =+4.0% for the TR and TP layer, respectively, and a 5 smaller bias of 2.62 DU 161.01 DU =+1.6% for the MS layer. For the layer between the surface and 30.5 km it is 4.00 DU 218.62 DU =+1.8%. This is a relatively low value remembering that between the total column amounts of FTIR and Brewer a bias of up to 5% was observed (Schneider et al., 2008). FTIR and ECC amounts are in good agreement. The observed differences between both experiments are close to the differences expected 10 due to the FTIR errors. This indicates that the ECC sonde provides very precise O 3 data and/or that the FTIR error estimation is rather conservative.
In Sect. 4.1 we compared the Brewer total O 3 amounts to the ECC O 3 amounts between the surface and 30.5 km. This comparison reveals a remaining scatter of around 2%, which can be contributed to errors in the ECC measurements or to the 15 observation of different airmasses by the two experiments. Brewer and FTIR analyse the same airmass (they have the same observation geometry) and the disagreement between the FTIR and the ECC data is very similar to the disagreement between the Brewer and the ECC data. This is demonstrated in Fig. 9, which depicts the differences between the ECC partial column amounts (PC ECC = 30.5 km 2.37 km x ECC d z) and the Brewer 20 total column amounts (TC Brewer ) versus the differences between the ECC and FTIR partial column amounts (PC FTIR = 30.5 km 2.37 km x FTIR d z). We can use the Brewer data to account for the observation of different airmasses or for the errors in the ECC data. Therefore we normalise the ECC data to the Brewer total column amounts: Here 60.6 DU is the mean value of the residual O 3 amount. Figure 10 compares the normalised ECC partial column amounts to the FTIR partial column amounts. This figure gives a good insight in the real performance of the FTIR system. The normalisation according to Eq. (3) produces changes of about 5% if referred to the typical O 3 amounts (see Fig. 9). In the stratosphere the natural variability is of a similar magnitude (5-10%), and the normalisation to the Brewer data significantly improves the correlation between FTIR and ECC partial columns (compare panel (c) of Fig. 10 and 8). On 5 the other hand, in the troposphere or tropopause the natural variability is very large (between 30 and 70%) and an alteration of the ECC data by only 5% has no significant effect In 4.1 we analysed the ECC quality by the Brewer total O 3 measurements. The variability of the residual O 3 above the sonde's burst altitude is 4 DU (or 1.4%). The 10 ignorance of this variability together with the error of the Brewer data of around 0.5% limits the validity of this test to 1.4 2 +0.5 2 =1.5%. This uncertainty range becomes even larger if the balloon bursts already below 30.5 km. As demonstrated in this paper a state-of-the-art FTIR system provides partial column amounts of good quality. At supersites like Izaña the quality of the ECC sondes can be checked by the FTIR partial 15 column amounts. This enables to perform good quality checks even for sondes whose balloon's burst at lower altitudes. For sondes reaching 30.5 km the quality check can be performed with a precision of 0.8% (0.8% is the estimated random error for the FTIR partial column amounts below 30.5 km; see Table 2). 20 We made an extensive theoretical error estimation for O 3 profiles measured by a stateof-the-art ground-based FTIR observing system. The FTIR system provides high quality data with a vertical resolution of 4.5-7 km for O 3 amounts below 20 km and of around 10 km in the middle stratosphere. The altitude range with the best vertical resolution coincides with the tropopause region. The application of a unique a-priori facilitates the 25 interpretation of annual cycles. We do not recommend the usage of seasonally varying a-priori data since the measurement alone contains sufficient information. At the Izaña Observatory ECC sonde, Brewer, and FTIR measurements are performed continuously at high quality. It offers world-wide unique conditions for intercomparing these different O 3 monitoring techniques, and to document their quality. Such quality documentations are essential if the measurements are to be applied for the validation of satellites. The comparison between ECC and FTIR amounts agrees with 5 our theoretical estimations about the FTIR precision. Furthermore, it indicates that the ECC sonde provides data with a precision of better than 5% from the surface up to the middle stratosphere. This demonstrated good performance of the ECC measurements is an in-field confirmation of extensive laboratory studies (Smit and Sträter, 2004). We found no significant systematic difference between the ECC and FTIR O 3 Schneider, M. and Hase, F.: Technical note: Recipe for monitoring of total ozone with a precision of 1 DU applying mid-infrared solar absorption spectra, Atmos. Chem. Phys., 8, 63-71, 2008, http://www.atmos-chem-phys.net/8/63/2008/. 4978, 4980, 4983, 4994 Schneider, M., Redondas, A., Hase, F., Guirado, C., Blumenstock, T., andCuevas, E.: Compar-5 ison of ground-based Brewer and FTIR total O 3 monitoring techniques, Atmos. Chem. Phys. Discuss., 8, 285-325, 2008, http://www.atmos-chem-phys-discuss.net/8/285/2008