Atmospheric Measurement Techniques The MIPAS 2 D database of MIPAS / ENVISAT measurements retrieved with a multi-target 2-dimensional tomographic approach

We present a multi-year database of atmospheric fields of the upper troposphere, stratosphere and lower mesosphere retrieved from satellite measurements adopting a 2-dimensional tomographic approach. The full mission of the Michelson Interferometer for Passive Atmospheric Sounding (MIPAS) instrument, on board the European Space Agency ENVISAT satellite, is analyzed with the Geofit Multi-Target Retrieval (GMTR) system to obtain the MIPAS2D database with atmospheric fields of pressure, temperature and volume mixing ratio of MIPAS main targets H2O, O3, HNO3, CH4, N2O, and NO2. The database covers both the MIPAS nominal observation mode measured at Full Resolution (FR) from July 2002 to March 2004 and the nominal observation mode of the new configuration, measured at Optimized Resolution (OR) and introduced in 2005. Further to the main targets, minor species N2O5, ClONO2, COF2, CFC-11, and CFC-12 for the FR mission only have been included in MIPAS2D to enhance its applicability in studies of stratospheric chemistry. The database is continuously updated with the analysis of the ongoing measurements that are planned to last until the end of 2013 and extended to other targets. The GMTR algorithm is operated on a fixed vertical grid coincident with the tangent altitudes of the FR nominal mode, spanning the altitude range from 6 to 68 km. In the horizontal domain, FR measurements are retrieved on both the observational Correspondence to: B. M. Dinelli (bm.dinelli@isac.cnr.it) grid and an equispaced 5 latitudinal-degrees grid which is made possible by the 2-dimensional retrieval algorithm. The analysis of MIPAS OR observations is operated on the same altitude-latitude fixed retrieval grid used for the FR measurements. This choice provides a database with a homogeneous altitude and latitude grid, over the whole globe, covering to date about seven years of measurements. The equispaced latitude grid provides a new and convenient layout for the much needed synergetic studies of data from various instrumental and modeling sources. MIPAS2D is available to the scientific community through the two web sites http://www.mbf.fci.unibo.it/mipas2d.html , and http:// www.isac.cnr.it/ ∼rss/mipas2d.htm.


Introduction
Four decades of Earth observation satellites have brought scientists an incomparable number of measurements, leading our way to a better understanding of the Earth's system and climate. Although our picture of the Earth's atmosphere has substantially improved over the years, there still remain a number of open issues that severely affect 5 our knowledge of it. Among these issues are the behavior of the stratosphere and that of the Upper Troposphere-Lower Stratosphere (UTLS) region, and their influence on climate.
The global and multi-year coverage of satellite missions is giving us a detailed picture of physical and chemical quantities of the atmosphere, of their distribution, variability 10 and long term trends. In this contest limb sounding infrared emission observations have the excellent characteristics of allowing global coverage in both day-and night-time (they do not require the presence of sunlight) with high vertical resolution. On the other hand, the use of polar orbits permits to obtain full geographical coverage of the observations in a limited time interval. However, not only obtaining observations is vital 15 for understanding a changing atmosphere, but also the ability of properly extracting the relevant information from them. Although the short term variability and inhomogeneity of the atmosphere bury inside a signature of the processes governing the atmosphere, to date most of the online retrievals of remotely-sensed atmospheric measurements are still carried out with 1-dimensional algorithms assuming a horizontally homogeneous points of successive limb-scans is about 500 km in latitude (∼5 latitude degrees).
A minor part of MIPAS observations were also acquired using the so called special observation modes, designed to enhance the instrument performance for a number of scientific objectives ranging from the study of the UTLS region to volcanic eruptions. These modes exploit different spectral resolutions, vertical scanning range 5 and sampling steps, and can use sideways views.
Due to the deterioration of the interferometric slides, the FR mission was suspended in April 2004. The technical problem was overcome in January 2005 by operating MIPAS with an Optimized Resolution (OR) at 41% of the maximum spectral resolution used in the original configuration. Since the reduced spectral resolution implied 10 a reduced acquisition time of every limb view, all observation modes have been redefined increasing their spatial resolution. In the new OR nominal observation mode, each scan is made of 27 limb views with tangent altitudes ranging from 3 to 70 km at variable altitude steps. Their nominal values are: 70.0,66.0,62.0,58.0,54.0,50.0,46.0,43.0,40.0,37.0,34.0,31.0,29.0,27.0,25.0,23.0,21.0,19.5,18.0,16.5,15.0,15 13. 5, 12.0, 10.5, 9.0, 7.5, 6.0 km. While the vertical spacing between the tangent altitudes of the sweeps of the individual limb-scans is kept constant along the orbit, their offset is changed with latitude to follow the tropopause altitude (see Fig. 1 where the location of the tangent altitudes is reported as a function of the latitude for both FR and OR nominal modes). Although during the first two years of OR measurements 20 the instrument availability was poor, since January 2008 MIPAS is operated back at 100% of its duty cycle and is expected to perform well during the following years of the satellite life (until the end of 2013). The new duty cycle consists of 8 consecutive days of nominal mode measurements plus 2 days of observations in the so called "Upper Atmosphere" mode, where a wider altitude range at reduced vertical resolution

ESA MIPAS Level 1b data
The first of the two main processing phases of the MIPAS mission ground segment produces ESA Level 1b data. Level 1b processing decodes the downlinked source packets of MIPAS interferograms transforming them into geolocated and calibrated atmospheric radiance spectra (Level 1b spectra). The spectra are collected into Level 5 1b files each containing the calibrated spectra acquired during a complete orbit. Details of the Level 1b processing phases can be found in Kleinert et al. (2007) and references therein.
Level 1b spectra are processed by ESA to obtain altitude profiles of atmospheric targets (Level 2 data). The same spectra are processed by our team using the GMTR 10 to obtain the MIPAS2D database. In this section we recall some of the main features related to the quality of MIPAS2D products and Level 2 products in general.
Level 1b data for the FR mission were released by ESA and updated in successive improved processing. The current database of MIPAS Level 1b data has been produced with different versions of the ESA Instrument Processing Facility (IPF): 15 version 4.61 and 4.62 for the FR mission, and version 4.65 (up to 14 July 2006), and version 4.67 (afterwards) for the OR mission. Figure 2 reports the noise level for a 10 cm −1 wide spectral region centered at 690 cm −1 for the full MIPAS mission. The figure shows that, as expected from radiometric considerations, the change in MIPAS spectral resolution (occurred in 20 January 2005) corresponds to a downward jump in the noise level. Figure 2 also shows a discontinuity in the first part of the OR mission where the noise level first increases exponentially and then suddenly drops. This discontinuity was caused by the first ice decontamination process operated after the restart of MIPAS mission. The strong change in the noise level is due to the removal of the heavy ice layer accumulated 25 on the detector windows during the long inactivity of the MIPAS instrument. The implications of this discontinuities in the results of the analysis of the full mission are discussed in Sects. 4.3.2 and 4.4.
Level 1b spectra are reported on a frequency grid of 0.025 cm −1 for the FR mission and of 0.0625 cm −1 for the OR mission. The interferometric spectral resolution of the measurements, that is given by 1/2 L where L is the maximum optical path difference of the interferometer (20.016 cm and 8.2 cm, respectively), is slightly different from these values. These implies that to be reported on the above mentioned frequency 5 grids, the spectra have to be resampled. An error in the spectra-resampling procedure which is operated with an interpolating function that is not fully consistent with the FT spectral resolution affects the IPF versions 4.65 and 4.67 applied to the measurements of the MIPAS OR mission. In our retrieval analyses we cope with this problem by emulating in the forward model the spectral distortions introduced by the resampling 10 error. A new corrected ESA Level 1b processor is almost ready. Tests performed running a GMTR analysis on the same orbit, processed with IPF 4.67 and with an error corrected prototype software, showed negligible differences always laying within the retrieval error bars, validating the approach used for the OR measurements retrievals. The selection process of the Level 1b data used to generate the MIPAS2D database 15 is discussed in detail in Sect. 4.2.

ESA MIPAS Level 2 data
Level 2 Ridolfi et al., 2000). ORM implements the Global-fit (Carlotti, 1988) algorithm by using several approximations to meet the very demanding runtime 2646 requirements of the near real-time processing. For each target, the retrieval is performed on a set of small spectral regions called Micro Windows (MWs), selected in order to have the best retrieval accuracy of the final Level 2 products (Dudhia et al., 2002). All main targets are retrieved in cascade, using the results of previously retrieved targets in the subsequent steps of the retrieval chain, and following 5 a sequence designed to minimize the error propagation. An off-line retrieval is also performed by ESA with the same Level 2 processor on more consolidated Level 1b data using the best available correlative data, therefore improving the quality of ESA Level 2 off-line products. ESA Level 2 products, derived with the described retrieval strategy, are based onto 10 three main assumptions that can lead to shortages in the retrieved data: 1. ORM assumes a horizontally homogeneous atmosphere. This assumption can significantly affect the error budget of the retrieved quantities, since the horizontal homogeneity assumption can fail in presence of significant horizontal structures along the instrument LOS. 15 2. Within the analysis of each limb-scan, ORM performs a sequential retrieval of the targets. The sequential retrieval leads to a propagation of the retrieval error on successively retrieved targets by effect of the reciprocal contamination in the analyzed spectral features.
3. MIPAS limb-scans are individually analysed by ORM, using the geolocation of 20 the tangent points to define the retrieval grid. As a consequence, the retrieved profiles are calculated on a sparse horizontal and vertical grid, and the horizontal resolution is implicitly defined by the separation between adjacent scans (see von Clarmann et al., 2009b).
The approximations described at points 1. and 2. affect significantly the error 25 budget of the retrieved quantities. The impact of these systematic errors is limited with a careful selection of the spectral points of each MW to be included in the analyses.

2647
However, under not uncommon circumstances, the residual errors are significant (see for instance Arnone et al., 2009a). The aforementioned independent analyses of MIPAS observations performed at Oxford (Dudhia et al., 2007) and IMK (von Clarmann et al., 2003;Milz et al., 2005) are based on similar assumptions; in particular they retrieve all targets on the horizontal 5 grid defined by the geolocation of MIPAS scans assuming the atmosphere horizontally homogeneous. Therefore the obtained products are affected by problems similar to the ESA Level 2 analysis. Recently von Clarmann et al. (2009a) have introduced in the retrievals of the OR observations a horizontal temperature gradient (applied only in the atmospheric region close to the tangent points) to cope with the horizontal 10 inhomogeneities of the sampled atmosphere that were making their analysis unstable.

The 2-dimensional geofit multi-target retrieval (GMTR) algorithm
As a response to the shortages highlighted in the previous section, Carlotti et al. (2006) developed the GMTR 2-D algorithm. The GMTR retrieval system is an open source package, specifically designed for MIPAS measurements, developed under an ESA 15 contract and delivered to ESA for inclusion in the BEAT tools repository (http://envisat. esa.int/services/beat). An upgraded version of GMTR that, among the novel features, is able to handle the Level 1b data of MIPAS OR configuration, was used to generate the MIPAS2D database. The most important features of the GMTR algorithm are briefly recalled below. A detailed description of the analysis system can be found in Carlotti 20 et al. (2006).
The GMTR is based on the Geo-fit approach (Carlotti et al., 2001) upgraded with the Multi-Target Retrieval (MTR) functionality (Dinelli et al., 2004). The Geo-fit performs a tomographic retrieval on observations collected along a whole orbit operating a 2-D discretization of the atmosphere which enables to model the horizontal atmospheric 25 structures. The MTR approach fits simultaneously different targets thus eliminating the systematic error components due to the propagation of the uncertainties on pressure, Interactive Discussion temperature and on the amount of molecules that generate interfering spectral features. The GMTR version used to generate the MIPAS2D database also includes the possibility of performing the retrievals using the Optimal Estimation (OE) or maximum a posteriori approach (Rodgers, 2000). This new functionality was introduced to preserve the possibility to adopt a regular retrieval grid also in case of lack of 5 observations (due to cloud filtering) and of the variable altitude offset of the OR limb scans along the orbit, and has been used for the database generation.

MIPAS2D retrieval grids
In the Geo-fit approach each limb observation contributes to determining the unknown quantity at a number of different locations among those spanned by its line of sight.

10
MIPAS nominal mode measurement strategy, with LOS laying approximately in the orbit plane, enables to gather information about a given location of the atmosphere from all the lines of sight that cross that location regardless of the scan they belong to. The exploitation of this characteristic is implemented in GMTR by defining a 2-D retrieval grid described by a vertical coordinate (spanning the altitude range of the 15 observations) and an angular coordinate (following the satellite position along the orbit) called Orbital Coordinate (OC) and performing the radiative transfer through a 2-D discretised atmosphere. Since GMTR simultaneously fits observations from a complete orbit that originates and terminates at the North Pole, the OC spans from 0 • at the North Pole, through 180 • at the South Pole up to 360 • back to the North Pole. 20 The fact that the observations of a whole orbit are simultaneously analysed makes it possible to define a retrieval grid that is totally independent from the measurements geolocation. This is an advantage with respect to the 1-dimensional retrievals that can determine atmospheric profiles on custom made vertical grids, but are tied to the position of the limb-scans for the geolocation of the retrieved profiles. Interactive Discussion scan has its own tangent altitudes because of the different refractive indices of the inhomogeneous atmosphere sampled by the lines of sight. Moreover, the motion of the satellite during the acquisition time of each sweep is only partially compensated by the measurement strategy and the tangent points of each scan are not vertically aligned. If the 2-D retrieval grid were defined at the tangent points, the spread of 5 the tangent altitudes would lead to a challenging interpolation process for both the construction of the atmospheric fields on the 2-D discretization of the atmosphere and the representation of the results. The 2-D interpolation processes can be easily performed if one of the 2-D grid dimensions (vertical or horizontal) is fixed over the whole orbit. For this reason we have adopted a vertical retrieval grid located at fixed 10 altitudes along the whole orbit, corresponding to the vertical grid of the nominal mode of the FR mission, i.e. from 6 to 68 km with a 3 km step from 6 to 42 km and then at 47, 52, 60 and 68 km. The same vertical grid has been used for both FR and OR missions in order to obtain a homogeneous database. As a consequence of this choice, the horizontal retrieval grid can be defined 15 at arbitrary positions and it is not forced to coincide with the geolocation of the measurements. A natural choice for the horizontal grid would be to use the OC of the tangent points of the measurements. However, since the tangent points of individual limb-views of the same scan are not vertically aligned, in order to ease the interpretation of the retrieval results we set the retrieval grid points vertically aligned at 20 selected values of the OC. A natural choice for the OC values is the average orbital coordinates of the tangent points of each limb-scan where the maximum information is expected. By convention, we have labelled "NOM" the dataset of MIPAS2D products obtained with this horizontal grid. The NOM grid follows the horizontal sampling of MIPAS measurements. Figure 3 shows the geolocation (latitude versus longitude) 25 of all retrieved profiles for the NOM grid of the FR mission in the equatorial region.
Although the latitude of the NOM profiles tends to cluster around preferred locations, a large fraction of profiles does not, therefore requiring horizontal interpolation even just to obtain a latitude-consistent database of the available measurements. Moreover Introduction Tables  Figures   Back  Close Full Screen / Esc

Printer-friendly Version
Interactive Discussion the distance between these preferred latitudes vary between about 4.6 • and 5.3 • along the orbit, so that all NOM measurements require a horizontal interpolation if they have to be reported on a constant latitude grid (as used by models). To minimize the need of interpolation in averaging processes and to ease the intercomparison with data from other instruments or atmospheric models, the best horizontal grid is a fixed layout of 5 equispaced OCs for all the analyzed orbits. We label "GRD" the dataset corresponding to this fixed horizontal grid. The GRD retrieval grid originates at the North Pole and extends along the full orbit with profiles at OCs (and thus latitudes) equispaced by 5 • .
Both horizontal grids have been applied to the FR measurements. The OR nominal mode samples the atmosphere with finer vertical and horizontal 10 steps with respect to the FR nominal mode. Moreover the tangent points of the measurements span an altitude range that is latitude dependent (see Fig. 1). The spread of the tangent altitudes in both vertical and horizontal domains makes less meaningful the GMTR analysis on a NOM grid; furthermore the exploitation of the finer vertical and horizontal grid would require the use of a regularization scheme, making 15 the results of the analysis not directly comparable with the FR results. For this reason for the first release of the database the analysis of the OR mission has been performed only on the GRD grid.

The MIPAS2D database
In this Section we summarize the retrieval strategy, the selection of the observations 20 used for the analyses, the sensitivity of the results to the adopted parameters and the filtering applied to obtain the released MIPAS2D version 1.0.

Retrieval strategy
For the generation of MIPAS2D, GMTR was configured with the aim of both a homogeneous database along the two parts of the mission and to facilitate the use 25 2651 of MIPAS2D data with atmospheric models and other observational datasets. All the main targets are retrieved for both the FR and OR missions on the GRD retrieval grid. As discussed in Dinelli et al. (2004), because of their correlations, pressure, temperature, water vapor and ozone are jointly retrieved by exploiting the MTR functionality of the code. All the other species are then retrieved sequentially 5 assuming the previously retrieved atmospheric fields. The adopted sequence: HNO 3 , CH 4 , N 2 O, and NO 2 was designed to minimize the systematic error deriving from the non-correct modeling of the previously retrieved species. For the FR mission the additional NOM database was produced, also including the minor species N 2 O 5 , ClONO 2 , COF 2 , CFC-11, and CFC-12 that are retrieved individually in this order after the main targets. The 2-D fields of all targets were retrieved simultaneously with the 2-D field of the absorption coefficients (one for each MW) of the atmospheric continuum. The retrievals are performed using the OE technique (Rodgers, 2000) in order to avoid singularities in the inversion procedure caused by the presence of atmospheric regions not sampled by the observations included in the retrieval because 15 of e.g., measurement strategy, corrupted sweeps or presence of clouds in the LOS.
With OE the a priori information and the associated a priori errors used for the retrieved parameters are an important choice because they can affect the quality of the obtained products. The choice of the a priori information (used also as initial guess) is discussed in Sect. 4.2.2. The a priori errors used in the analyses have been defined as a fraction 20 of the initial guess of the parameter value plus a constant term introduced to prevent strong constraints where the a priori fields assume very small values. Table 1 shows the values used for all the considered targets. As shown in this table, for all the molecules but NO 2 the percent part of the a priori error was set to 80%. For NO 2 , since a high variability is expected due to either its day-night variation or to the high 25 altitude variability of its VMR, we have chosen to use a large value for the constant part and a small percent part in the whole altitude range. The a priori error of pressure was tuned to 30% in order to account for MIPAS pointing errors.
For technical reasons the GMTR analysis system does not allow for the retrieved 2652 parameters to assume negative values: if during the retrieval iterations the correction to be applied to one parameter leads to a negative value, the parameter is arbitrarily set to 1.0E-14. The spectra affected by the presence of clouds in the LOS are filtered out using the method proposed by Spang et al. (2004) adopting a cloud index (CI) threshold equal to 5 4.0. This conservative value prevents the use of spectra even partially contaminated by clouds. The sensitivity to the adopted retrieval configuration is described in Sect. 4.3.

Level 1b data selection
Since ENVISAT is on a nearly-polar orbit, the atmospheric regions where the 10 geographical distance between the air masses observed by consecutive orbits is minimum are at the Earth's Poles. Therefore one of the two Poles is the best choice to originate the analysed orbit. For the MIPAS2D database, as discussed in Sect. 3, the North Pole identifies the starting and ending point of the analysed orbits. However the first limb-scan reported in ESA Level 1b files is always at latitudes far from 15

90
• N; therefore in order to obtain a complete orbit to be analysed with GMTR the measurements contained in the Level 1b files of two consecutive orbits are merged in a single analysis (see Carlotti et al., 2006). Whenever consecutive orbits are not available, the retrieval is performed on the measurements acquired after reaching the North Pole of a single Level 1b file without closure of the geographical loop.

20
A selection of the ESA Level 1b files has been necessary because of the several Level 1 reprocessings performed by ESA and leading to the release of non-unique Level 1b files for each orbit over time. The selection has been performed adopting criteria based on the geographical coverage of the files and on quality indicators. Whenever multiple Level 1b files exist for the same orbit we retained the one processed 25 with the latest version of the IPF spanning the wider geographical region. If a further selection is needed, the most recent released file is retained unless evident anomalies 2653 are found during an automated and manual check. An ad-hoc list of Level 1b data was then generated with a unique file for each orbit. Figure 4 shows in black the histogram of the number of Level 1b orbits per month analysed to produce the MIPAS2D database. The figure also reports in gray the total number of available orbits per month, including the orbits measured in the special observation modes, that have 5 not been included in the current analysis.

Auxiliary data
As mentioned in Sect. 4.1 we exploit the OE method for the retrieval of our Level 2 products and the a priori information is provided by the same atmospheric fields used as initial guess for the Gauss-Newton iterative procedure. In principle, if the 10 size of the a priori errors has been correctly evaluated, the result of the retrieval is independent from the initial guess used for the target quantity provided that the analysed observations contain sufficient information (see Sect. 4.3). However, an initial guess close to the real fields reduces the number of Gauss-Newton iterations needed to reach a stable result, thus limiting the computing time. Therefore, whenever 15 available, ESA Level 2 data were used as initial guess in the analysis of the FR mission. As for the OR mission, where ESA Level 2 data are not yet available, we have used the results obtained for the FR mission for a good-quality orbit of the corresponding month of 2003. Only for a very limited number of cases, where this strategy was not possible, the initial guess profiles were taken from the climatological IG2 database (Remedios et 20 al., 2007). The same database has been used to complement the initial guess profiles outside the vertical retrieval range and to define the atmospheric state relative to all the non-target molecules of the retrieval procedure.
For each target (or group of targets in the case of MTR) the analysis has been performed using dedicated MWs (different for the FR and OR mission) selected with 25 the algorithm MWMAKE, described in Dudhia et al. (2002)  as well as those used for the retrieval of the minor species not included in the ESA operational analysis, have been generated for a 2-D analysis that exploits the MTR functionality. For HNO 3 , CH 4 , N 2 O and NO 2 we have chosen to use the MWs of the ESA Level 2 ground processor. The possibility of using dedicated MWs, selected for a 2-D analysis, is under investigation for future releases of the database.

5
Auxiliary data for the GMTR analysis system are also the Occupation Matrices (OM) that define the altitude range where each MW is used in the retrievals and the altitude dependent "spectral masks" used to exclude specific spectral points from the inversion for a reduction of the size of the systematic errors affecting the retrievals. All these auxiliary data have been generated with the MWMAKE program. Further auxiliary 10 input data were taken from the GMTR package (see Carlotti et al., 2006).

Data quality and parameter sensitivity
The strategies adopted for the selection of the retrieved parameters and for the setup of the analysis environment were optimized with a series of dedicated test retrievals. In particular, an extended test was performed over all MIPAS measurements acquired 15 from 1-20 October 2003 in order to assess the sensitivity of the results to the initial guess profiles and to the size of the errors assigned to the a priori information.
We report the results of these activities in the following subsections where we describe the size of the systematic errors affecting the Level 2 products included in the database (Sect. 4.3.1), the quality assessment of these products (Sect. 4.3.2), the 20 behavior of the retrieval procedure when single orbits or orbit segments are analysed (Sect. 4.3.3), the sensitivity to pointing calibrations (Sect. 4.3.4) and the difference between the NOM and the GRD datasets (Sect. 4.3.5).

Systematic errors
The algorithm used to select the MWs (Dudhia et al., 2002) enables an a priori 25 assessment of the systematic errors affecting the final products of the analysis. These 2655 errors are assumed to be common to all the analyses that use the same set of MWs and they are characteristic of each target and change for the FR and the OR Level 2 products. The systematic errors whose effects have been quantified are: errors due to the assumptions used in the forward model internal to the retrieval system (i.e. local thermal equilibrium, shape of the profiles outside the retrieval 5 range, spectroscopic errors) instrumental errors (i.e. instrument line shape, radiometric and frequency calibration) influence of the interfering species onto the retrieved targets.
Details on how these errors have been evaluated can be found in Dudhia et al. (2002) 10 and in the web site http://www.atm.ox.ac.uk/group/mipas/err/.
As an example Fig. 5 reports, as a function of the retrieval altitudes, the total systematic errors affecting the retrieved values of temperature (red lines) and ozone VMR (blue lines) for the FR and OR missions. In the altitude range from 20 to 50 km, the total systematic errors affecting ozone are below 10% (as low as 4%), and below 15 1 K (as low as 0.5 K) for temperature. The total systematic errors for both FR and OR MWs are reported in the database in specific files, one for each target. These files also report the major systematic error components contributing to the total evaluated systematic error. 20 In order to asses the contribution of MIPAS observations to the atmospheric fields reported in MIPAS2D we have adopted a quantifier (called individual information gain) defined for the retrieved values of each target quantity as in Dinelli et al. (2009). The individual information gain is defined as:

Individual information gain
where S rj is the diagonal element of the variance-covariance matrix (VCM) of the retrieval and S aj is the diagonal element of the VCM of the a priori information for parameter j . q j provides a measure, in terms of bits of information, of the information gain provided by the analysed measurements with respect to the a priori knowledge of each retrieved value (Rodgers, 2000). The value of q j depends on the size of the 5 a priori errors used in the retrieval, and therefore it is not an absolute quantifier. In regions where the information in the measurements is low, due to either cloud coverage or lack of atmospheric sampling, the retrieval error comes close to the a priori error and the individual information gain value approaches zero. Any other value greater than zero indicates that some information about the retrieved quantity was extracted from 10 the measurements. Our baseline is to consider q j values below 0.5 (that indicates that the a priori error has been reduced of a factor of √ 2) as an indication that the retrieved value comes mainly from the a priori knowledge of the target quantity.
Although retrieved values with low information gain are included in the database, these represent only a minor fraction as shown in Fig. 6 that reports the frequency at 15 which a certain value of the individual information gain is obtained in the database. The figure shows that the individual information gain distributions peak at different values for the FR and OR data. This is due to the fact that the radiometric noise error is different for the two missions (see Sect. 2.1), while the error used for the a-priori is the same in the two analyses. The data with low values of q j can be filtered out 20 a-posteriori by the users or properly handled through weighted averages (since they always correspond to poorly determined parameters) (see Sect. 4.4). It is however useful to keep in the database also values with low information gain because they may still be used in specific studies. As discussed below, negative values are artificially assigned to the individual information gain in order to flag data characterized by specific 25 caveats.
Because of the large size of the files, the Averaging Kernel (AK) matrices relative to each analysed orbit are not included in the database. Only an example of the 2-D AK for both NOM and GRD grids is included in the database for both FR and OR missions; 2657 AKs for selected parts of the database can be supplied under request.

Single orbits and orbit segments
We have seen in Sect. 4 that the best performance of the 2-D approach of GMTR is achieved when it is operated on a North Pole-to-North Pole complete orbit. Nevertheless, non-complete orbits, having a significant number of scans, are analysed 5 and the results are included in the database because of their not-negligible fraction over the total number of MIPAS observations. Only orbit segments that cover a time period shorter than 20 min of continuous observations have been rejected from the analysis. In the case of NOM data, the horizontal grid used to retrieve the orbit segments is restricted to the geolocation range spanned by the available MIPAS scans. In the GRD case, the same horizontal grid, covering the complete orbit, has been used also for orbit segments; in this case, retrieved values laying outside the MIPAS measurement range are flagged with a negative value of the individual information gain (see also Sect. 4.4).
The quality of the results obtained with incomplete orbits was tested by analysing 15 orbit segments artificially created from a full orbit. The comparison of the results of these analyses with those obtained from the corresponding full orbit showed no significant deviations, apart from the atmospheric fields retrieved at the edges of the orbit segment. In MIPAS2D the data corresponding to these regions are properly characterized by low values of the corresponding individual information gain (see also 20 the discussion on filtering in Sect. 4.4).

Sensitivity to pointing calibrations
Since GMTR does not retrieve pointing information, the impact on MIPAS2D results of the calibration routinely performed on the MIPAS instrument elevation angle needs to be investigated. During most of the FR mission, MIPAS pointing errors were affected 25 by a significant seasonal cycle induced by a mispointing of the ENVISAT satellite.

2658
This problem was corrected on December 2003 (orbit 9321) significantly improving the pointing stability. As a result, the vertical grid used as input to ESA Level 2 analyses is affected by seasonal and abrupt changes in coincidence with pointing calibrations. The erroneous tangent altitude values reported in the Level 1b files and their abrupt changes can also affect the results of the GMTR analysis, although the 2-D approach 5 reduces the impact of the pointing error thanks to the compensation effect operated by the 2-D ray tracing of the multiple observation geometries that contribute to determine each profile. Nevertheless, some residual inconsistency remains in terms of pressure step-like changes in time series as shown in Fig. 7 that reports the evolution of FR daily mean pressure within the 20 • S/20 • N equatorial latitude band. Figure 7 also 10 highlights the impact of the correction adopted in December 2003 when a more stable pointing was achieved significantly improving the accuracy of the tangent altitudes hence reducing the pressure oscillations. If MIPAS2D data are plotted versus altitude they appear to be partially affected by the pointing uncertainties.

Comparison of NOM and GRD results
The consistency between MIPAS2D products obtained in the two configurations of the horizontal grid (NOM and GRD) was tested using daily averages within the latitude bands selected for the development of the IG2 database by Remedios et al. (2007). For the majority of the data present in the two databases we have obtained a satisfactory 20 agreement in terms of both average values and standard deviations. The averages were only marginally different in the most unfavourable cases of largest atmospheric variability (e.g. at the breaking of the southern polar vortex or in the disturbed winter northern polar region). However, even in these cases, the differences are far below the weighted ESD, thus pointing towards a robust consistency. As an example of these 25 comparisons, Fig. 8 shows mean profiles (obtained averaging the retrieved values weighted by their retrieval error) under typical conditions (temperatures in Novemberleft panel) and under two worse-case scenarios (ozone at the breaking of the southern 2659 polar vortex -central panel, and NO 2 in the disturbed winter North Pole -right panel). In each case, the mean profile (bold line) and the standard deviation (dashed line) are shown (black for NOM and red for GRD). Also shown are horizontal bars representing the weighted ESDs affecting each retrieved level of the profiles, altough most are not visible because of their small value.

Filtering and consolidated datasets
The whole MIPAS mission was analysed with GMTR using an automated procedure. Before including the final results into the database, the quality of the results was checked, in order to remove corrupted and low quality data. At the time of writing, MIPAS measurements up to December 2008 have been analysed and only the 10 acceptable results have been included into the database. In total 7188 FR orbits and 8057 OR orbits have been analysed. As discussed in Sect. 2 part of the FR and of the OR missions were measured adopting special observation modes that are not analysed in this first release of MIPAS2D. Furthermore a fraction of the nominal mode measurements was discarded because of either failures in reaching convergence or 15 poor quality of the retrievals often due to poor Level 1b data. As a result of the data quality filtering, the FR NOM dataset includes 5830 orbits (all FR measurements), while the GRD dataset includes 5830 FR orbits and 6843 OR orbits. Considering that MIPAS2D can be used for a variety of applications, the filtering stage was tuned at a level that enables to exclude products that appear to be unphysical 20 but not those that provide unexpected values of the retrieved quantities. The filtering criteria are based on the quality of the spectral fitting, on threshold values of the information gain (see Sect. 4.3.2) and on the presence of Level 1b corrupted data. The first stage of filtering was on the value of the final χ 2 , defined as the normalized L 2 noise weighted norm of the residuals (observed minus simulated spectral radiances) 25 at the final step of the iterative retrieval procedure. This quantifier provides a measure of how well the results of the forward model simulations are able to reproduce the measurements of the full orbit taking into account the measurement errors. A χ 2 2660 equal to 1 indicates that the spectral residuals are within the measurement noise level. The filtering was operated adopting a threshold of 7 on the χ 2 of the MTR retrieval of pressure, temperature, water vapour and ozone. No filtering was applied to the other targets based on the χ 2 value. Although a high χ 2 does not imply that the retrieved quantities are not realistically describing the analysed spectrum, high values 5 were found to be associated with orbits having whole parts affected by poor Level 1b data. Since all the orbits leading to a MTR χ 2 above 7 were discarded, all other targets are consequently not available for these orbits. On the contrary, targets with a day-night variability such as NO 2 can lead to high χ 2 values in coincidence of low χ 2 values of the MTR: these products were all kept into the database. Most rejected 10 orbits have failed to reach an acceptable convergence during the retrieval. The final χ 2 for the MTR retrieval of pressure, temperature, water vapor and ozone for the full MIPAS mission (GRD dataset) is shown in Fig. 9. The figure shows the long term self-consistency of the FR and OR datasets, with no indication of a trend or increased spread along the extension of the MIPAS mission. Figure 9 also highlights the effects of 15 the different noise levels of the FR and OR missions; actually the χ 2 value is negatively correlated with the absolute value of the noise (see Fig. 2), indicating that the retrievals are affected by systematic errors of similar size in both FR and OR measurements. A part from the automatic removal in the retrieval iterations of negative values, the MIPAS2D results are intentionally not filtered against anomalous values or strong 20 deviations from climatological expectations in order to avoid the introduction of biases and the loss of information if anomalous atmospheric events are encountered. This choice however implies that the user should take some caution in using the database. Unless seeking specific data under very unfavourable situations, information gain below 0.5 (see Sect. 4.3.2) is a proxy for poor retrieval information, and the user should 25 carefully evaluate the impact of the use of these data. As an example, the distribution of information gain for temperature results shown in Fig. 6 shows that the suggested thresholds on the information gain affect only a minor fraction of the datasets.
Further filtering of the data was performed on the basis of the consistency of the 2661 values of the information gain. Because of numerical instabilities of the matrix inversion algorithm, the retrieval error for some parameter assumes unrealistically very small values, consequently their information gain assumes unacceptable high values. These data were removed flagging them with negative values of the information gain. In the case of the GRD dataset, negative information gain is also used as a quality flag to 5 identify retrieved data that are located outside the range of MIPAS observations (see Sect. 4.3.3). We stress again that the results flagged with negative information gain should be discarded by the end users since they are kept into the database only for the sake of uniformity of the released products. As a consequence of the chosen strategy to originate and close the orbits at the 10 North Pole the profiles in the artic region derive from measurements that are not consecutive in time. The time lag between the data determined before and after the North Pole is about one hour; for this reason the users should be careful in the usage of the TIME information reported in the files of the database. If required, for selected cases it is possible to repeat the analysis originating and closing the orbits at the South 15 Pole.

Examples of results
The MIPAS2D database is intended to promote the use of MIPAS products and stimulate scientific collaboration.  Figure 10 shows time series relative to six targets obtained with both FR and OR MIPAS2D results. The maps in Fig. 10 report average values computed over 2662 a 5 days period (this averaging criterion will be used for all the coming figures but Fig. 17) at the pressure level 31.5 hPa (about 24 km) and for latitudinal bins of 5  Fig. 10 shows the evolution of the polar vortices over the years, with colder temperatures for the Southern Polar winter than for the Northern Polar winter. The ozone map shows a clear depletion at the South Pole in September, correlated with the breaking of the vortex. The HNO 3 map shows a build up at the 15 South Pole in April, May and June, then a depletion corresponding to the winter period with a progressive increase up to the breaking of the polar vortex. We also see the seasonal alternation of high and low peak ozone values at mid latitudes. The water vapor map shows the dehydration occurring at the South Pole during spring and at the tropics around January each year. CH 4 and N 2 O, that can be used as trace species for 20 the dynamic of the Stratosphere, show similar latitudinal behavior. Both gases show a seasonal latitudinal displacement of their peak values within the tropical dynamical barriers, as the meridional circulation changes with season. This behaviour is seen also in HNO 3 but with an opposite sign of the horizontal gradient.
In the GMTR analyses most of the initial part of the reconstructed orbits, and 25 therefore OC below 180 • , lies in the day side. The exact range of the OC lying in the dayside depends on the solar zenith angle and therefore varies with season. Figure 11 shows the time series relative to ozone values at 68 km on the orbital coordinate of the GRD retrieval grid. At this altitude the photolysis of ozone into atomic oxygen leads 2663 Printer-friendly Version Interactive Discussion to order of magnitude differences between daytime and nightime concentrations of ozone. This leads to the abrupt changes at the day-night terminators and the depicted seasonal evolution (Brasseur and Solomon, 2005). Figure 12 shows the NO 2 time series at 47 km on the orbital coordinates of the GRD retrieval grid. NO 2 concentrations are close to zero during the day time while they 5 suddenly increase at dusk, with larger changes at the South Pole than at the North Pole. Events of abrupt NO 2 enhanced concentrations are visible at the winter poles (e.g. July 2003 at the South Pole) due to downward transport of mesospheric NO 2 or during solar proton events (e.g. at both poles in October-November 2003).  Figure 16 shows the results of the GMTR analysis on the NOM grid (FR mission only) and reports the time series for the VMR of minor species N 2 O 5 , ClONO 2 , COF 2 , CFC-15 11 and CFC-12. All these maps show the 5-days average of the data in the antarctic latitudinal band (65 • S-90 • S). For these species the vertical coverage of the analysis has to be reduced due to the very low information content of the measurements outside the plotted range. The figure highlights the seasonal evolution of the species and gives an example of the potentiality of the results contained in the MIPAS2D database. In 20 fact these additional species are useful to obtain an improved picture of the ozone chemistry including ozone catalysts and their reservoirs. A further example of the potentiality of the MIPAS2D database in terms of applicability to stratospheric dynamics is shown in Fig. 17 where the ozone VMR at 21 km is reported for the measurements acquired from 13 to 16 September 2002 (left 25 hand panel) and from 23 to 26 of the same month (right hand panel) when the Southern polar vortex splitted into two lobes for the first time ever recorded (Simmons et al., 2003). The split up of the polar vortex around the 24 September is clearly traced by the split up of the ozone hole.

2664
MIPAS2D has been exploited to study the influence of transient luminous events on the atmosphere (Arnone et al., 2009b) and to analyze trends of ozone and several molecular species related to ozone chemistry: HNO 3 , N 2 O, NO 2 , N 2 O 5 , ClONO 2 , COF 2 , CFC-11 and CFC-12 by Papandrea et al. (2009). Starting from January 2008 MIPAS duty cycle is back to 100% and the mission has been extended to the end 5 of 2013. Therefore the analysis strategy adopted in that paper is well suited also for a refined analysis and a correct monitoring of possible ozone recovery, as required by the Montreal protocol and successive amendments as well as shorter term studies.

Conclusions and future development of the database
We have analysed all the measurements acquired by MIPAS in the nominal observation 10 mode from the beginning of the mission up to December 2008 by applying for the first time a tomographic approach to such an extended sets of data. The obtained MIPAS2D database contains 2-D fields of pressure, temperature and VMR of H 2 O, O 3 , HNO 3 , CH 4 , N 2 O and NO 2 that have been obtained on a fixed altitude-latitude grid for the whole mission. The analyses have been perfomed with the GMTR 15 system that enables the tomographic retrieval of MIPAS data taking properly into account the horizontal inhomogeneities of the atmosphere. The resulting database has been stored into the web sites http://www.mbf.fci.unibo.it/mipas2d.html and http: //www.isac.cnr.it/ ∼ rss/mipas2d.htm and is available to end users upon request. The database is extended with the 2-D distribution on the FR horizontal observation grid of 20 the targets mentioned above complemented with the VMRs of N 2 O 5 , ClONO 2 , COF 2 , CFC-11, and CFC-12. To date these minor species have been retrieved for the FR mission only. Other target species, whose emission spectrum falls into the frequency range measured by MIPAS, can be retrieved on demand. The obtained results are complemented with detailed auxiliary files. The database is continuously updated with the analysis of new available Level 1b files of the OR mission, and is beeing expanded to other MIPAS special observation modes. 2672 Fig. 4. Distribution of the number of Level 1b orbits per month processed in the MIPAS2D (black), and of the total number of available Level 1b orbits (gray). Note that the total number of Level 1b orbits include also the special observation modes.