As mentioned above, the most accurate description of the atmosphere is required to feed the RT models. In order to achieve this goal, the vertical profiles from a selection of atmospheric constituents in the different IR absorption bands measured by both instruments were extracted for each date from global analyses provided by the Monitoring Atmospheric Composition and Climate (MACC) project of the Copernicus programme

http://www.gmes-atmosphere.eu

The CO2 and CH4 three-dimensional global fields used for observation-to-profile interpolation rely on the global tracer transport model LMDZ , driven by the wind analyses from the European Centre for Medium-Range Weather Forecasts and using optimised surface fluxes following the configuration used by . For CO2 fields, the model has a horizontal resolution of 3.75∘ in longitude and 1.9∘ in latitude with 39 vertical layers, from close to the surface up to 8.6 hPa. For CH4 fields, the horizontal resolution is 3.75∘ in longitude and 2.5∘ in latitude with 19 levels, from the surface up to 295 hPa.

A bilinear interpolation weighted by the inverse of the square distance was used to match the corresponding profiles and the surface values associated with each one of the IASI observations. The position and time of these IASI observations were extracted from the Météo-France operational archive. In addition to the latitude, longitude, date and time values of the IASI observations, the following parameters were also extracted from this archive: IASI azimuth and zenith angles, solar azimuth and zenith angles, cloud cover from the Advanced Very High Resolution Radiometer (AVHRR) on board MetOp, land–sea mask values from the Météo-France global model ARPEGE and IASI measurements for a subset of 314 channels selected by and monitored operationally at Météo-France in 2013.

4A is an optimised line-by-line (LBL) radiative transfer model used as a reference RT model for the CNES/EUMETSAT IASI Level 1 Cal/Val activities and operational processing. Using the OpenMP technology, a parallel-spectrum processing capability was added to the model in order to reduce the required computing time per spectrum. Rather than using the onerous full LBL models like LBLRTM (Line-by-Line Radiative Transfer Model; ) or STRANSAC , less expensive but not-as-accurate optimised LBL models have been developed. For typical instruments like IASI and IASI-NG, precision is better than the instrumental noise. 4A allows a fast computation of the transmittance of a discrete atmosphere along the vertical at a very high spectral resolution as well as the Jacobians (with respect to temperature, mixing ratios and surface temperature and emissivity) for a user-defined observation level. The model relies on comprehensive atlases of monochromatic optical thickness for up to 50 atmospheric molecular species and 43 pressure levels. The atlases were created using the line-by-line and layer-by-layer model STRANSAC in its latest version at the release date with spectroscopy information from the GEISA (Gestion et Etudes des Informations Spectroscopiques Atmosphériques, French for “Management and Study of Spectroscopic Information”) 2011 spectral line data catalogue . The 4A model also includes up-to-date continua of N2, O2 and H2O. 4A uses surface emissivity values supplied by .

RTTOV is a fast RT model for passive visible, IR and microwave (MW) satellite-borne sensors and is used in various applications. Radiance data assimilation make use of RTTOV at Météo-France, the Met Office and ECMWF for example. Simulated satellite imagery is used by forecasters to compare the forecast outputs with the satellite radiances for nowcasting. Moreover, atmospheric retrieval can be obtained through the use of the 1D-Var provided by the EUMETSAT NWP Satellite Application Facility. Fast RT models can reproduce LBL radiances with a good accuracy and computational efficiency that fulfils NWP requirements of near-real-time monitoring and satellite radiance assimilation. This kind of models use computationally efficient parametrisations that allow them to simulate radiances at a fraction of the cost required by a LBL model.

RTTOV version 11.3 was used in this study. The latter takes as input vertical profiles of pressure, temperature, water vapour and optionally other trace gases, and scattering particle parameters along with associated surface parameters. The outputs are top-of-atmosphere (TOA) radiances, brightness temperatures (BTs) and the Jacobians for the selected instrument channels. To carry out the fast computation of optical depths, RTTOV uses linear regressions to compute optical depths on a fixed set of pressure levels. The coefficients for these linear regressions are pre-computed and stored in coefficient files which are specific to each instrument.

Most RTTOV coefficient files are based on a fixed set of 54 levels, from 1050.0 to 0.005 hPa, but coefficients have also been generated on a set of 101 levels, from 1100.0 to 0.005 hPa, for some instruments, more explicitly the IR hyperspectral sounders AIRS, IASI, CRIS, IASI-NG and IRS. RTTOV can accept input profiles on an arbitrary set of pressure levels. A vertical internal interpolator maps the input profile onto the coefficient levels and, again, the computed optical depths back onto the input levels . In the latest version of the software, RTTOV 11.3, additional options have been added to improve consistency of the input vertical profile units .

The information contained in the IASI observation database has been used to prepare the most realistic possible atmospheric state according to the input of each RT model. Regarding the atmospheric chemistry, a total of 16 vertical atmospheric constituent profiles were ingested by 4A for the simulations: H2O; CO2; O3; N2O; CO; CH4; SO2; HNO3; OCS; CH3D; N2; and the CFCs 11, 12 and 14. For H2O, CO2, O3, CO, CH4, and SO2 species, the vertical profiles were taken from the above database. For the other species, a standard profile provided by 4A was used. All the profiles were linearly interpolated onto the 43 4A levels (43L). The reduction in the number of levels was carried out as no significant gain in accuracy was observed when using 60 levels, whereas the computing time is almost doubled in a single computer (evaluation against IASI observations not shown). RTTOV simulations were run over 60 levels with the latest coefficient files. The version 11 of RTTOV could only be supplied with six atmospheric constituents as inputs (compared with the 43 species possible with 4A), and only five vertical profiles of atmospheric constituents were provided to the model: H2O, CO2, O3, CO, and CH4. The sixth possible atmospheric species, albeit not used in our case, is N2O.

4A simulations have used surface emissivity values from internal tables over sea and sea ice. RTTOV uses the Infrared Surface Emissivity Model (ISEM; ) for sea surface emissivities. Land surface emissivities were taken for 4A from the University of Wisconsin (UW) IR atlas of emissivities for the year 2013. RTTOV simulations also used the UW atlas of emissivities but for the year 2007. Since RTTOV uses precomputed atlas values to speed up the calculations, it was not possible to use the 2013 atlases. As negligible differences are expected in the UW atlas emissivity values between the years 2007 and 2013, the RTTOV precalculated emissivity files have been used. Table  summarises the main differences between both RT runs.

The IASI viewing geometry was used for the simulations of both instruments despite the different scanning geometry of IASI-NG because scan geometry was not yet clearly defined at the time the dataset was built.

Although the AVHRR cloud cover values corresponding to each IASI pixel and cloud vertical profiles have been included in the database, we have chosen not to consider cloudy conditions because of computational cost of these kinds of simulations and also due to the large uncertainties in cloud radiative properties modelling.

Once the simulations were carried out, a random Gaussian noise was added to the simulations using the standard noise equivalent temperature difference (NEDT) at 280 K provided by the CNES (Fig. ). Two different noises were considered for IASI-NG depending on the prism material (KBr or ZnSe), which produced two IASI-NG simulation datasets.

A total of 5 242 047 IASI observations were compiled for the four selected dates. From this total, 3 463 432 observations correspond to sea pixels, 185 916 were from coastal regions and 1 591 529 were from over land. An overview of the distribution of the observation according to different latitude bands and AVHRR cloud mask values is given in Table . Five latitude belts have been considered: the North Pole, the mid-latitudes of the Northern Hemisphere, the tropics, the mid-latitudes of the Southern Hemisphere and the South Pole. For the AVHRR cloud mask, there is no value for some observations. The Northern Hemisphere presents more clear observations compared to the South. In total, clear cases represent 15.0 % of the total number of simulations. The no-flag, partly cloudy and cloudy simulations are respectively 1.6, 39.0 and 44.4 % of the 5 242 047 total simulations.

Figure  presents the information available in the database associated with an IASI observation taken out in the South Pacific (179.0∘W, 24.55∘S) on 3 February 2013 at 21:13:46 UT. The simulated brightness temperatures presented in this figure come from the 4A RT model dataset. Figure a compares the results from full IASI (black) and IASI-NG (grey) spectrum simulations. IASI-NG presents a higher variability of the spectrum than IASI because of its higher spectral resolution. As a result, it is expected that new atmospheric constituents can be detected by IASI-NG and that those already detected, such as water vapour isotopologues, will be better characterised .

The differences between the IASI 4A simulated spectrum and the corresponding observation for the 314 available channels in the Météo-France archive range from -3 to 3 K, except for the region between 2280 and 2400 cm-1, where they increase to around 7 K (Fig. b). Those bands associated with the input vertical profiles present larger differences than the regions corresponding to surface properties. The region between 2280 and 2400 cm-1 is strongly affected by the so-called ghost effect . This effect was caused by the IASI compensation device but vanished when the device was turned off on 7 October 2015 . Finally the third horizontal panel presents the sea surface emissivity from the 4A model. In this case, the surface emissivity varies between 0.971 and 0.991.

The first row of Fig. d presents the values of temperature and cloud properties extracted from the MACC project. The vertical profiles of the different atmospheric constituents used for the simulation are shown in Fig. e. The additional information concerning this IASI observation can be found in Table : latitude, longitude, terrain elevation, date, time, surface pressure and temperature, ARPEGE land–sea mask, AVHRR cloud cover, instrument zenith and azimuthal angles, and solar zenith and azimuthal angles.

A first evaluation of the simulations was made using the IASI MetOp-A full-spectrum brightness temperatures from the orbit beginning 6 August 2013 at 20:35:59 considering only clear-sky observations over sea, including sea ice during night-time. Hence, 2877 out of the 91.800 IASI measurements carried out in each IASI orbit were retained.

The mean values (biases) present differences, which are displayed in Fig. . In the first band, the RTTOV bias is lower than the 4A one excluding the CO2 band, between 665 and 675 cm-1 and from 700 to 745 cm-1. It should be noted that RTTOV uses a set of predefined CO2 to generate the coefficient file, which dates from 2013 and varies between 300 and 500 ppmv (parts per million by volume). Since that date, the global CO2 concentration has increased by 2.5 % approximately, making this RTTOV profile training set possibly obsolete. 4A computes optical depths directly using the input vertical profiles, allowing for more accurate values for them than using a set of predefined lookup tables. The fact that RTTOV presents a lower bias than 4A in the rest of the band arises from not having considered the same sea surface model. As mentioned before, 4A uses constant emissivity values from , while in RTTOV the sea surface emissivity is computed by the ISEM model, which takes into account the skin temperature and the instrument viewing angle .

Whatever the spectral bands, differences could be due to the spectroscopic parameters used. The spectroscopy used by 4A came from the 2009 version of the GEISA database, and the spectroscopy used by RTTOV came from HITRAN 2012. In the second IASI band, RTTOV bias values are equal or lower to those of 4A. These differences arise most of the time from the different representation of the water vapour continuum at the stage of the simulations.

In band 3, the RTTOV bias and SDs are closer to zero at the beginning of the band but present larger values than 4A in the emission CO2 spectral window from around 2220 to 2380 cm-1. In the solar region, above 2400 cm-1, the 4A bias remains constant around 0.7–0.8 K, whereas RTTOV present lower bias values. The origin of these differences could come from (i) the different sea surface emissivity values used by the two models, (ii) the N2 continuum or (iii) line mixing effects around 4.3 µm.

The up to 10 K peak observed in the SD curves of both models at 2390 cm-1 from Fig.  is mainly caused by the so-called ghost effect . It is generally acknowledged that this effect emanates from a perturbation in the response of the instrument which is mainly caused by micro-vibrations of the interferometer separator blade, in turn induced by the instrument compensation device . As it has been found that IASI reconstructed radiances from a principal component compression do not exhibit this artefact , an additional comparison between the simulated spectra and reconstructed radiances was carried out (Figure not shown). Although a difference of 10 K in the SD was found for IASI channel 6942, the value decreases to 2.0 K with reconstructed radiances. This confirms that the peak of 10 K in the SD comes from the verification observation and not from the RT model.

In order to evaluate both model simulations over a longer period (the four selected days in 2013), simulations were compared against a subset of 314 channels of IASI observations stored in the Météo-France archive. Only 789 573 clear-sky observations over both land and sea from the 5 242 047 total observations, i.e. around 15 % of the recovered IASI observations, were cloud-free observations according to the AVHRR cloud cover. The average and SD of these differences for both 4A and RTTOV simulated datasets are shown in Fig. .

The SDs of 4A and RTTOV simulations against observation differences are similar all over the measured spectrum. There are only three slight differences: for surface channels in the range between 750 and 950 cm-1, RTTOV presents a slightly lower values compared to 4A: around 0.11 K on average, a reduction of about 3.2 % with respect to 4A. However, in the region between 1020 and 1225 cm-1, the SDs of RTTOV differences are on average 3.5 % higher than with 4A (around 0.1 K). The third area, where differences between RTTOV and 4A are noticed, is in the solar part of the spectrum, above 2400 cm-1. These larger SDs for RTTOV come from the fact that, contrary to 4A, the solar contribution is ignored.

Regarding the differences in biases, RTTOV exhibits values closer to zero than 4A, excluding the CO2 band between 700 and 800 cmcm-1, the surface-sensitive region between 1100 and 1200 cm-1 and the three channels above 2400 cm-1, probably due to the differences in the UW atlas surface emissivity values used. As was discussed previously, emissivity values corresponding to the year 2007 were used for RTTOV simulations instead those of the year 2013.

A comparison between IASI and IASI-NG radiance simulations is shown in Fig. , from 730 to 740 cm-1, belonging to the 15 µm CO2 absorption band. A single IASI observation from 6 August 2013 and acquired at 20:38:00 has been used in order to calculate the Earth IR spectrum with 4A at a resolution of 0.001 cm-1. The result is shown Fig. b. Figure a presents the CO2 absorption lines as extracted from the GEISA database . The position of these absorption lines corresponds with the sharp peaks observed in the 4A spectrum calculation of Fig. b.

The simulation of the IASI and IASI-NG spectrum by RTTOV and 4A models is shown in Fig. c and d respectively. The values of the corresponding IASI measurements are also drawn, with dots, in Fig. c. The simulations of the IASI and IASI-NG spectrum from both models are very close, almost undistinguishable in this spectral range. Additionally, the differences between the measurement and the IASI simulations are very small, showing the quality of the two models in this spectral window. IASI-NG signal presents a higher variability, giving values going from 231.5 to 266.5 K, compared with the the IASI range of 239.2 to 262.4 K. This higher variability will increase the low-concentration limit of trace gas detection of IASI-NG compared to IASI.

The background error covariance B matrix provided by the package assumes a priori information of a very high quality, i.e. an error on temperature retrievals of 0.5 K when IASI is supposed to be able to improve temperature retrieval errors below 1 K. To be consistent with the IASI specifications, the B matrix errors have been multiplied by 2. Background profiles were obtained from the true atmospheric state perturbed by a noise corresponding to the B matrix. For the observation error-covariance R matrix, a diagonal matrix formed by the Météo-France operationally specified IASI errors was used. As the IASI-NG noise is assumed to be close to half the IASI noise, the values of the associated observation error R matrix were divided by 4. This operational R matrix overestimates mostly the errors for water vapour channels as the inter-channel correlations in their errors were not taken into account.

We have chosen the current operational channel selection of the ARPEGE model, consisting of 123 IASI channels out of the total 8461. For IASI-NG, the channel selection included the 123 channels corresponding to the same wave numbers as those of the IASI channel selection (see Table  for further details). This channel selection covers the whole atmosphere for temperature profile and only the troposphere for the humidity (Fig. ) even though the Jacobian peaks differ slightly between both instruments. These Jacobians have been obtained with RTTOV and correspond to Jacobians of both sounders averaged over the 1681 simulations. Humidity Jacobians depend on the water vapour profile, presenting higher values for higher humidity concentration (not shown). In the tropics, the humidity Jacobian peaks at higher altitudes in the atmosphere than in polar regions because of this dependency on the water vapour concentration.

A bias correction was applied to each latitude-band experiment. The values of this correction were obtained after a first run of the 1D-Var retrieval. The average difference between the first guess and the simulated brightness temperatures were removed from the observations to eliminate this bias for the second 1D-Var run.

To analyse the retrievals from the 1D-Var runs, the SD of the observation minus background and observation minus retrieval differences were computed for the brightness temperatures. For temperature and humidity vertical profiles, the SD computed were those of the truth–background and truth–retrieval differences. The analysis of the results consisted in comparing the ratio of SDs of the differences vs. the retrievals compared to the differences vs. the background: 1σreduction=σobs/truth-retσobs/truth-bck-1. Negative values of σreduction mean a reduction on the retrieval's SD – hence, a positive impact – whereas positive values of this index correspond to a negative impact or an increase in the retrievals SD.

Figure  presents the values of the σreduction index for the observations for the three latitude bands. High-stratospheric channels present an improvement in the error reduction of a 0.3 value, while the differences for tropospheric-peaking channels are better only by 0.1 on average. This difference is caused by the strong overestimations of the instrument noise for humidity channels because of the strong intercorrelation for these channels.

Similar plots to Fig.  are presented in Figs.  and but for the error reduction in temperature and humidity profiles respectively. We found a difference of up to 10 % in the reductions of temperature profile error for the tropical band around 400 hPa. Mid-latitude and polar latitude bands present lower error reduction differences, around 5 % in the troposphere, with a flatter shape on average. The different shape of the polar-latitude error reduction profile could have its origin in the lower number of profiles used (91 at polar latitudes, compared to 1068 at mid-latitudes). The error reductions in the stratosphere are lower than those observed in the troposphere, but IASI-NG still has higher error reductions than IASI. Apart from the first kilometres of the atmosphere, IASI-NG displays an error reduction twice as high as that given by IASI up to an atmospheric pressure of around 750 hPa. The worst performance of both instruments in the first atmospheric layers is related to the lack of sensitivity of IASI channels at these levels to the first atmospheric layers in the selection used in the Météo-France operational system combined with a possible lack of contrast with the surface. A new channel selection for IASI-NG shall be carried out including channels able to improve this lack of sensitivity and taking into account the IASI-NG bands 3 and 4 thanks to the IASI-NG noise reduction compared to IASI. The retrieval capability of IASI-NG at the low atmospheric levels shall also be studied with respect to the surface contrast.

The humidity error reduction profiles provided by both instruments present higher values than those found for the temperature error reduction profiles, where, in the best case, a 24 % error reduction was reached. The IASI-NG error reduction profile presents higher error reduction values, of around 5 %, from around 850 hPa to the tropopause compared to IASI. As was observed for temperature profiles, the error reductions in the low troposphere are smaller than those found for mid-troposphere or the upper atmospheric layers. These smaller error reductions are due to the lack of sensitivity of IASI channels to this atmospheric region combined with a weak background error in the humidity associated with these levels compared to those of the other regions, preventing observations from which to add more information in the 1D-Var.

A similar experiment to the IASI/IASI-NG experiment has been carried out to compare the performances of the two IASI-NG prism materials. Although a better error reduction was found for brightness temperature stratospheric channels, no impact was noticed in temperature or humidity profile retrievals. It should be noted that the channel selection used for this experiment was equivalent to the IASI one, just taking the channels corresponding to the same wave numbers. In order to investigate the differences between both materials, a new channel selection, adapted to IASI-NG characteristics, shall be performed.