New dynamic NNORSY ozone profile climatology

Abstract. Climatological ozone profile data are widely used as a-priori information for total ozone using DOAS type retrievals as well as for ozone profile retrieval using optimal estimation, for data assimilation or evaluation of 3-D chemistry-transport models and a lot of other applications in atmospheric sciences and remote sensing. For most applications it is important that the climatology represents not only long term mean values but also the links between ozone and dynamic input parameters. These dynamic input parameters should be easily accessible from auxiliary datasets or easily measureable, and obviously should have a high correlation with ozone. For ozone profile these parameters are mainly total ozone column and temperature profile data. This was the outcome of a user consultation carried out in the framework of developing a new, dynamic ozone profile climatology. The new ozone profile climatology is based on the Neural Network Ozone Retrieval System (NNORSY) widely used for ozone profile retrieval from UV and IR satellite sounder data. NNORSY allows implicit modelling of any non-linear correspondence between input parameters (predictors) and ozone profile target vector. This paper presents the approach, setup and validation of a new family of ozone profile climatologies with static as well as dynamic input parameters (total ozone and temperature profile). The neural network training relies on ozone profile measurement data of well known quality provided by ground based (ozonesondes) and satellite based (SAGE II, HALOE, and POAM-III) measurements over the years 1995–2007. In total, four different combinations (modes) for input parameters (date, geolocation, total ozone column and temperature profile) are available. The geophysical validation spans from pole to pole using independent ozonesonde, lidar and satellite data (ACE-FTS, AURA-MLS) for individual and time series comparisons as well as for analysing the vertical and meridian structure of different modes of the NNORSY ozone profile climatology. The NNORSY ozone profile climatology is available to the community as a comprehensive software library.


Introduction
Ozone is one of the most important trace gases in the atmosphere. Through its strong absorption of ultraviolet (UV) radiation it also influences the vertical profile of atmospheric temperature. The correlation between ozone and temperature profile has been widely studied by multiple linear regression analysis (Steinbrecht et al., 2003) with 5 model data and from long-term total ozone observations of TOMS (Total Ozone Monitoring Spectrometer) (Bhartia, 2003) and SBUV (Solar Backscatter Ultra-Violet) instruments (Heath and Park, 1975), as well as NCEP reanalysis temperature profile data (Kistler et al., 2001). Results show global patterns of total ozone column and lower stratospheric temperature correlated with the QBO (quasi-biennial oscillation), the so-10 lar cycle, and other parameters. Other atmospheric parameters like tropopause height are also correlated with ozone profile values and shape (Steinbrecht et al., 1998). Further studies of Steinbrecht et al. (2006) yield a quantification of major modes of interannual variation of total ozone column and lower stratospheric temperature from long-term observations and reanalysis as well as from coupled chemistry and climate 15

models.
Several available ozone profile climatologies are based on lookup tables. For construction of the SUNYA climatology of the State University of New York (Wang et al., 1995) total ozone column values based on TOMS from 1978 to 1992, stratospheric profiles based on SAGE II (Stratospheric Aerosol and Gas Experiment-II) from 1984 to 20 1992 and ozonesonde data were used. The stratospheric vertical distribution is based on the data of SAGE II which provides data up to 60 km at 1 km resolution from October 1984to November 1989. The resulting data set consists of monthly zonal means with 59 pressure levels from 0.28 hPa to 1000 hPa.
The UGAMP ozone climatology (Li and Shine, 1999)  averages on a 2.5 • × 2.5 • horizontal grid, with 47 levels ranging from the ground up to 0.001 hPa.
Monthly zonal mean ozone profiles for 17 zonal bands (80 • S-80 • N) are provided by the ozone profile climatology developed by Fortuin and Kelder (1998). Ozonesonde data from 30 different stations as well as data from SBUV, SBUV/2 and TOMS for the 5 period of 1980 to 1991 were used to prepare the climatology. It provides ozone partial column data on 19 vertical pressure levels (1000 hPa to 0.3 hPa) and includes information about ozone profile standard deviation. A further monthly zonal mean ozone profile climatology has been prepared by NASA Goddard Space Flight Center (GSFC) (McPeters et al., 2007) based on previous work of Logan (1999) and McPeter (1997) 10 for a tropospheric ozone climatology. It was compiled using ozone sonde data up to 24 km as well as SAGE II and UARS MLS data for the height range 29 km to 60 km and a weighted average in between. The GSFC is also authoring the TOMS V8 climatology (Barthia, 2003) which provides total ozone classified zonal mean ozone profile data for 10 • latitude bands. 15 Total ozone dependence is also a feature of the ozone profile climatology compiled by Lamsal et al. (2004). They provide separate sets for winter/spring and summer/fall seasons, for high and mid-latitude regions. For tropical and subtropical regions no seasonal dependent profile data are available. Primarily ozone profile data sets from ozonesondes (WOUDC, 2010) including SHADOZ measurements (Thompson et al.,20 2003), SAGE II (Cunnold et al., 1989(Cunnold et al., , 1991(Cunnold et al., , 2002a, POAM III (Lucke et al., 1999;Lumpe et al., 2002;Randel et al., 2003) as well as National Meteorological Center (NMC) and UKMO temperature profiles in the time range 1990 to 2000 contributed to this ozone profile climatology.
Two more recent ozone climatologies are described in Mze et al. (2010) and Ziemke 25 et al. (2011). The latter is a global climatology of tropospheric and stratospheric ozone columns derived from AURA OMI and MLS (Schoeberl et al., 2006;Levelt et al., 2006;Froidevaux et al., 2008) (15-50 km). Comparisons were made to eight SHADOZ stations in the altitude range from 15 km to 30 km; however the coarse spatial sampling of GOMOS and missing validaton for mid and high latitude restrict this climatology to tropical regions, while the climatology of Ziemke (2011) provides only 5 columnar values for the troposphere and stratosphere. Ozone profile climatology data are an important source of information in atmospheric sciences and especially within the retrieval community. For example, ann ozone profile climatology is needed as a-priori information for total ozone column retrieval using the TOMS (Bhartia, 2003), DOAS (Loyola et al., 2011), WFDOAS (Weber et al., 2005) 10 or GodFit (Lerot et al., 2010) algorithms. The NNORSY climatology presented in this paper was developed in the framework of the GDP5 project for the purpose of retrieving ozone total columns from GOME using the GodFit approach.
Other applications in atmospheric science using an ozone profile climatology are ozone profile retrieval from satellite data by means of optimal estimation to constrain 15 the solution according to Rogers' theory (Rogers et al., 1976;Rogers, 2000) applied e.g. by Mijling et al. (2010) or Lamsal et al. (2007), for background correction within retrieval schemes of other trace gases like CO 2 (Crevoisier et al., 2009), for data assimilation and validaton of 3-D chemistry and transport models, for several other applications or measurements of height resolved atmospheric parameters or for atmospheric cor-20 rection in space-based imagery of the surface. Furthermore, stratospheric ozone and water vapour profiles are correlated to meteorological regimes. Especially in the UTLS region different weather regimes exhibit unique profile shapes for water vapour and ozone (Follette-Cook et al., 2009;Hudson et al., 2009 in compiling column and/or profile data into an ozone climatology. Very few available ozone climatologies use total ozone column as a predictor for the ozone profile shape (e.g. Bhartia, 2003;Lamsal et al., 2003), but none of them use temperature profile data as a predictor. With both parameters the current state of the atmosphere for which a climatological ozone profile is desired can be specified. To model correlations between 5 dynamic input parameters (e.g. total ozone column and/or temperature profile) and a target vector (ozone profile in our case), data mining by means of machine learning techniques is an excellent approach. This was already proven with the NNORSY (Neural Network Ozone Retrieval System) scheme (Müller et al., 2002(Müller et al., , 2003 and compared, which was used to set up the new dynamic ozone profile climatology. A detailed 10 comparison study of NNORSY with other retrieval schemes can be found in Meijer et al. (2006). In Sect. 2 the main results of a user consultation will be shown, before in Sect. 3 the basic approach and the ozone profile measurement data sets used are briefly described. In Sect. 4, the methods employed, the NNORSY processing scheme as well 15 as the validation data will be specified. The pole-to-pole validation results are shown in Sect. 5 for the different climatology modes, by means of comparisons with single measurement stations, time series analysis as well as meridian ozone profile structure analysis, before we draw our conclusions in Sect. 6.

20
In 2004, in an early stage of the development of the new NNORSY Ozone profile climatology, a user consultation was organised in the framework of the ESA-funded CHEOPS project. A questionnaire was sent to more than 1200 scientists and experts in atmospheric sciences. Feedback was given by 86 scientists from 35 different countries. A main achievement of the consultation was a better knowledge about user re-25 quirements on the temporal and spatial resolutions as well as the height resolution and ozone profile units. Each responder has to select one of multiple choices per category. The main results are summarized in Table 1 Building on the feedback of the user consultation, the new NNORSY ozone profile climatology was developed, which features four different combinations (modes) of input parameters in order to be as flexible as possible for a wide range of potential applications (Table 2). Regarding spatial and temporal resolution there are in principle no restrictions because the new ozone profile climatology is provided as a software library 5 that takes arbitrary space-time coordinates as input. In practice, the temporal and spatial resolution of the NNORSY ozone profile climatology is mainly dependent on the resolution of the dynamic input parameters provided by the user.

Approach and data used
Currently available climatologies for ozone profile data are usually based on look-up 10 tables. User requirements show that for a new ozone profile climatology it is important to also use dynamic input parameters like total ozone and/or the temperature profile as predictors. These requirements (mainly temperature profile usage) lead to the NNORSY (Neural Network Ozone Retrieval SYstem) approach which was developed at our lab within the last decade for total ozone retrieval (Müller et al., 2002) as well 15 as for ozone profile retrieval from the GOME (Müller et al., 2003;Meijer et al., 2006) and SCIAMACHY instruments. Recently, NNORSY was further developed and implemented for GOME-2 onboard MetOp-A as well as for combined GOME-2/IASI ozone profile retrieval (Felder et al., 2011).
For the different climatology modes various neural networks have to be trained. In the 20 framework of several NNORSY projects for GOME, SCIAMACHY and GOME-2/IASI we built a large atmospheric profile data base with focus on ozone profiles. These data are collocated with analysis model data to add consistent temperature and pressure profile information. For the NNORSY climatology the following training data sources were used: Introduction -NASA's SHADOZ programme (Thompson et al., 2003) -Satellite solar occultation measurement data: -SPOT-3 POAM-II data version 6: (Lumpe et al., 1997) 5 -SPOT-4 POAM-III data version 4: (Lucke et al., 1999;Lumpe et al., 2002;Randel et al., 2003) -ERBS SAGE II data version 6.2: (Cunnold et al., 1989(Cunnold et al., , 1991(Cunnold et al., , 2002a -UARS HALOE data version 19: (Nazaryan et al., 2005;Eyring et al., 2006) The time range of training data spans 1995 to 2007 (Fig. 1). During ingestion of data 10 into the ZSW profile database rigorous quality checks are carried out, similar to the ones described by Hassler et al. (2008) with some extensions described below. Before the data are used for neural network training, statistics on the whole training dataset are calculated for each profile level and only data within the 99 % percentile on each level are used for training. For climatology applications of ozone profiles, there is no point in 15 trying to model very exceptional situations. Within the profile database each profile is collocated with temperature profiles from the GEOS-4 reanalysis (Bloom et al., 2005). All data are being interpolated to 61 height levels centred on 0.5 km up to 60.5 km with a sampling of 1 km. For additional quality control of ozone profiles, the temperature profile data often measured together with the ozone profile data and delivered within 20 most original data sources are cross-checked against the GEOS-4 temperature profile data. If the deviation between both temperature profiles is too large, the whole profile is flagged as questionable and not used for successive training or validation. Various comparisons showed that if the T-profile data of the measurement data source exhibits high deviations from T-profile analysis data then the ozone profile measurement is 25 usually of poor quality. Figure 2 shows the temporal and spatial distribution of the 782 Introduction

Tables Figures
Back Close

Full Screen / Esc
Printer-friendly Version

Interactive Discussion
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | ozone profile data used, while in Fig. 3 latitudinal histograms of ozone profile data from the different sources are displayed. It shows for instance that for WOUDC more data are available in the Northern Hemisphere while POAM data are only available for higher latitudes.
For setting up a climatology, the seasonal dependence is also important. In Fig. 4 5 this dependence of the ozone profile measurements used for training is shown by data source. Only for POAM II the histogram shows significantly more data in the second half of the year. The influence on the overall distribution is quite small because the absolute amount of data from the other limb sounders like SAGE, HALOE and POAM III is much higher.

10
A common feature of limb sounder data is the poor quality of profile retrievals around and below the tropopause region. During ingest of limb sounder ozone profile data into the training database, all ozone profile data at and below the tropopause were flagged and not used for successive neural network training. The tropopause was calculated from GEOS-4 reanalysis data using a mixture of WMO lapse rate definition and 2.5 15 PVU surfaces.

Neural network training
The neural networks for different climatology modes were trained separately using the same number of training patterns. The problem to overcome for training climatology modes with total ozone column input parameter was that no independent measure-20 ments of the total ozone column for each ground and satellite ozone profile measurement (training pattern) were available. Total ozone column measurements from satellites like GOME, GOME-2, TOMS or OMI were ruled out because their retrieval algorithms already use different ozone profile climatologies and/or model data, and we would like to avoid any possible feedback loop to already existing ozone profile clima- 25 tologies. Therefore we first trained the neural network for climatology mode TLLT (Table 2) which already gives a very good estimate of the ozone profile. For this training we did not split up the ozone profile measurement database for training, test and evaluation. This neural network training for mode TLLT was used to supplement the missing parts of ozone profiles for a given data source (e.g. above burst point for ozonesondes and 5 below tropopause for limb sounder data). After supplementing missing ozone profile levels with first neural network climatology for mode TLLT, the total ozone column was calculated by integrating the resulting ozone profiles. This total ozone column value was later on used for training of modes TLLO and TLLTO where it is used as input parameter. For training of the neural networks, the ozone profile database was divided 10 into training and test dataset. For a first validation during development mainly ACE-FTS, and AURA-MLS ozone profile data were used which were not part of the training or test data. With this setup, all four NNORSY climatology modes were trained and tested. Figure 5 shows the training results for climatology mode TLLO as an example. For full geophysical validation, additional sonde and lidar measurements from NDACC 15 stations are used which will be described in the next section.

Validation
The validation of the new climatology relies on independent ground-based and satellite ozone profile measurements. The ground-based data comprise ozonesonde and lidar observations associated with the Network for the Detection of Atmospheric Com-  et al., 2003) and are interpolated to the standard NNORSY height grid with 1 km sampling. In parallel to the validation of the NNORSY climatology, the ozone profile climatology used for the TOMS Version 8 ozone column retrieval algorithm (Bhartia, 2003) was compared to the same correlative dataset. 5 Validations started with the analysis of comparisons at all individual ground stations, for dynamic climatology mode TLLO using total ozone as input parameter. The example given in Fig. 6 also includes ozone profile data of the TOMS V8 climatology (Bhartia, 2003), since it is based on the same predictors. Figure 6 shows that for the selected stations from four different regions, the overall agreement of NNORSY ozone 10 profiles with sonde data is good. However, in ozone hole conditions the NNORSY as well as the TOMS V8 climatology do not represent the full extent of the high ozone depletion around 20 km altitude, which is expected because the corresponding ozonesonde measurements show that the ozone concentration is almost zero. In the tropical stratosphere the TOMS V8 climatology profile at ozone peak altitude is 15 too low while NNORSY is in good agreement with the sonde measurements. A similar pattern can be observed for the measurement at Uccle station where TOMS V8 climatology is lower than NNORSY, but the smoothed ozonesonde measurement falls around the mean at ozone peak levels, which also shows the smoothing effect for both climatologies very well. 20 Figure 7 shows the vertical structure of the statistical agreement of NNORSY (mode TLLO) and TOMS V8 climatology at the same stations already used in Fig. 6. At Nairobi NNORSY underestimates ozone in the height range of 10-20 km and overestimates it ∼10 % below 10 km. The largest underestimation up to 30 % is at 18 km. At this height level there are very steep gradients of the ozone profile and for example only a slight 25 shift of 0.5 km in the height assignment can cause such a bias. Below this height the ozone concentration is very low and therefore the absolute deviation is small. around 10 km altitude while TOMS V8 shows an overestimation in the height range of 6-9 km. For Uccle station the agreement of NNORSY is very good while TOMS V8 exhibits an overestimation of up to 30 % around 10 km altitude.

Time series
Comparisons over time were performed against satellite measurement data (ACE-FTS These differences can partly be explained by the diurnal cycle of ozone concentration and the other part seems to be stem for some kind of scan pattern bias in the AURA-MLS data itself which can also be seen in direct comparisons of MLS data with HALOE V19 data. Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | measurements from Mouna Loa (tropics), for NNORSY climatology mode TLLO. The agreement for Mouna Loa is very good in the stratosphere but show some negative bias in tropopause while for Ny-Alesund and Uccle the main differences can be seen in the UTLS region and for Neumayer station during ozone hole conditions in the altitude range between 10 km-20 km, where the absolute ozone concentration can be very low 5 inducing very steep gradients in the ozone profile (Fig. 6). Therefore the smoothing error can be a major contribution to the overall difference shown. A negative bias can be recognized in the inter-tropical UTLS region around 18 km (Fig. 10), and around 10 km in the polar regions (Fig. 12). For the high and mid-latitude regions the agreement between measurement and NNORSY climatology is within 10 % 10 while in the inter-tropical zone standard deviation is higher (Fig. 7).

Meridional structure and total ozone sensitivity
In order to compare the four different modes of the NNORSY climatology qualitatively and quantitatively, Fig. 13 shows the meridional structure for all climatology modes. It can be seen that modes with (TLLO, TLLTO) and without (TLL, TLLT) total ozone 15 as predictor are more similar and that the agreement with total ozone used as predictor is better. The NNORSY mode TLL does not use any dynamic input parameter and therefore this climatology generates ozone profiles only dependent on date and geographical location. Figures 14 and 15 show a sensitivity study with respect to the total ozone input 20 parameter dependent on latitude for the Southern and Northern Hemisphere, respectively. The interval for varying the total ozone input parameter is dependent on latitude and the plots will give an indication for the valid range of total ozone input for the NNORSY climatology.

Conclusions
This paper reports on the development and verification of a new generation ozone profile climatology based on the NNORSY approach. In order to study the geophysical soundness of the climatology, output ozone profiles have been compared to correlative measurements acquired by a network of ground-based ozonesonde and lidar 5 stations deployed from pole to pole. The quantitative and qualitative agreements between the climatology and ground-based observations resemble those reported for the first NNORSY ozone profile retrievals for ERS-2/GOME (Müller et al., 2003). A monthly mean agreement often better than ±10 % is obtained in the stratosphere and the interpercentile 68 % is close to 10 %. For ozone hole conditions the NNORSY climatology 10 has a positive bias in situations where ozone concentration in the 20 km altitude range is very low. The better comparison results obtained when smoothing the ozonesonde and lidar data at 3 km resolution confirm this value as a good estimate of the NNORSY vertical resolution. In general the different climatology modes have similar agreement but the dynamic 15 modes of the climatology -including the ozone total column and/or the temperature profile as input parameters -yield better results than the static mode TLL, mainly by reducing the standard deviation of the mean difference. Under ozone hole conditions the vertical structure of early ozone depleted profiles (September) is not reproduced well by the modes TLL and TLLT. Mode TLL strongly overestimates the ozone con-20 centration in ozone depleted situations, while for mode TLLT the overestimation is less and at lower altitudes (Fig. 13). Especially for Antarctica modes including total ozone give much better results for ozone depleted profiles with slight overestimation at ozone depleted altitudes. For the inter-tropical tropopause modes TLL and TLLO underestimate the ozone 25 concentrations between 10 km and 20 km, while modes TLLT and TLLTO exhibit slight overestimation. This can be an indication that including the temperature profile as a predictor for the ozone profile helps the neural networks to implicitly estimate tropopause height and thus to perform a coarse airmass type classification for different climatological regimes. If we try to compare the NNORSY ozone profile climatology to the TOMS V8 climatology, in general the relative differences between TOMS V8 and ground-based ozonesonde data are similar to those between NNORSY climatology and ground-5 based measurements. In the inter-tropical troposphere the TOMS V8 climatology compares better to correlative ozonesonde data than NNORSY. Positive biases are observed for TOMS V8 near the tropopause that are not present in the NNORSY climatology. This may be explained by the difference in vertical resolution between TOMS V8 using Umkehr layers and NNORSY which is estimated to ∼3 km (Müller et al., 2003).

10
The NNORSY ozone profile climatology is available as software library as well as standalone software. The software is written in Fortran and will compile with GNU Fortran Compilers which are available on most computer system used in the atmospheric science community. This opens a broad range of applications for the NNORSY ozone profile climatology ranging from data assimilation, trace gas retrieval from satellite data 15 as well as climate modelling tasks. Especially for total ozone column and ozone profile retrieval schemes using optimal estimation the new NNORSY ozone profile retrieval has the potential to speed up retrieval, and potentially improve retrieval results.    Station Ny-Alesund for arctic region, Station Uccle for northern mid-latitude region, 4 Nairobi for tropics and Neumayer for Antarctica. The red line shows NNORY 5 climatology ozone profile, the blue line the co-located ground based raw ozone profile and 6 the green line the same ground-based profile smoothed to estimate the NNORSY vertical 7 resolution (3 km FWHM). The yellow line shows the corresponding TOMS V8 ozone 8 profile. The total ozone column value in Dobson Units are results by integrating ozone 9 profiles for NNORSY and TOMS V8 climatology while GDP denotes the total ozone retrieval of GDP Version 4.1 from ERS-2/GOME data. Fig. 6. Typical comparison of single profiles of NNORSY climatology mode TLLO and TOMS V8 Climatology with measurement stations in different climatological regions. Station Ny-Alesund for arctic region, Station Uccle for northern mid-latitude region, Nairobi for tropics and Neumayer for Antarctica. The red line shows NNORY climatology ozone profile, the blue line the co-located ground based raw ozone profile and the green line the same ground-based profile smoothed to estimate the NNORSY vertical resolution (3 km FWHM). The yellow line shows the corresponding TOMS V8 ozone profile. The total ozone column value in Dobson Units are results by integrating ozone profiles for NNORSY and TOMS V8 climatology while GDP denotes the total ozone retrieval of GDP Version 4.1 from ERS-2/GOME data. 5,2012 New dynamic NNORSY ozone profile climatology   FTS and AURA-MLS data. A one month running mean has been applied to the relative 5 difference. 6 without comparisons for ACE-FTS data. The ACE orbit does not allow solar occultation 6 measurements within the tropical region. 7 Nairobi (equator) and Neumayer (Antarctica). "O3S" denotes ozone sonde, "LID" 7 ground-based lidar measurements. A one month running mean has been applied for the 8 relative difference time series. For comparison of single measurements for these stations 9 see