Homogeneity criteria within IASI pixels for the preparation of an all-sky assimilation

IASI, homogeneous scenes, clouds, assimilation, ARPEGE model Abstract. This article focuses on a selection of satellite infra-red IASI observations and their simulation in the global Numerical Weather Prediction (NWP) system ARPEGE (Action de Recherche Petite Echelle Grande Echelle), using the sophisticated radiative transfer model RTTOV-CLD which takes into account the cloud multi-layers and the cloud scattering from atmospheric profiles and cloudy microphysical parameters (liquid water content, ice content and cloud fraction). The aim of this 5 work is to select homogeneous scenes by using information of the collocated Advanced Very High Resolution Radiometer (AVHRR) pixels inside each IASI field of view and to retain the most favourable cases for the assimilation of IASI infrared radiances. Two methods to select homogeneous scenes using homogeneity criteria already proposed en the literature were employed; criteria derived from Martinet et al. (2013) for cloudy sky selection in the French mesoscale model AROME (Applications of Research to Operations at MEsoscale), and the criteria from Eresmaa (2014) for clear sky selection in the global 10 model IFS (Integrated Forecasting System). An intercomparison between these methods reveals considerable differences, either in the method to compute the criteria or in the statistical results. From this comparison a revised method is proposed that is a compromise between the different tested methods, using the two infrared AVHRR channels to define the homogeneity criteria in the brightness temperature space. This revised method has a positive impact on the observation statistics minus the simulation statistics, while retaining 36% observations for the assimilation. It was then tested in the NWP system ARPEGE 15 and tested for the clear-sky assimilation. These criteria were added to the current data selection based on the Mc Nally and Watts (2003) cloud detection. It appears that the impact on analyses and forecast is rather neutral.

of multi-layer clouds was studied using diagnosed cloud schemes (Chevallier et al., 2004) and (Stengel et al., 2010). These studies also showed beneficial results. A step further was achieved with the study by Okamoto et al. (2014). They studied the assimilation of multi-layer cloud-affected infrared radiances using the all-sky assimilation approach already implemented for microwave imager at ECMWF. They particularly investigated the cloud effects on the differences between observations and simulations and thus proposed an appropriate quality check and dedicated observation errors.

5
In this study we are interested in IASI observations, where the radiances are considered with colocated clusters statistical properties of the Advanced Very High Resolution Radiometer (AVHRR) co-located with IASI on the METOP platform with a horizontal resolution of 1km at nadir (Cayla, 2001). Intuitively, collocated AVHRR data provide information on surface properties and the presence of clouds in the IASI Field Of View (FOV). They can therefore be used for cloud detection. The AVHRR cluster information associated with IASI has already proven to be useful for selection purposes in the context of 10 cloud data assimilation, with an explicit treatment of microphysical variables in the AROME model by Martinet et al. (2013). Eresmaa (2014) at ECMWF also used AVHRR cluster information for cloud detection and observation selection in clear sky. Martinet et al. (2013) selected cloudy scenes based on cloud homogeneity. This study was done in a 1D-Var framework using an advanced radiative transfer model (RTTOVCLD) including profiles for liquid water content, ice water content and cloud fraction to simulate cloud-affected radiances as background equivalents to AROME fields. The persistence of cloud information 15 brought by the analysis of cloud variables during a 3h forecast has then been evaluated successfully with an one-dimensional model AROME version (Martinet et al., 2014).
In this article, we try to determine whether or not collocated AVHRR and IASI information would facilitate the selection of homogeneous scenes which could be potentially used in an all sky assimilation approach. Section 2 describes the ARPEGE NWP system, the IASI instrument and the radiative transfer model RTTOV-CLD in cloudy sky conditions. In section 3, infor-20 mation about the AVHRR clusters is detailled, the strengths and weaknesses of the different methods to select homogeneous observations are discussed, the chosen method is presented together with a description of the selected observations. Section 4 depicts the impacts on analyses and forecasts of selected clear and cloudy IASI observations. Conclusion and perspectives are given in section 5.

The ARPEGE model and its 4D-Var system
The ARPEGE model is the global NWP model at Météo-France, used operationally since the early 1990s (Courtier et al., 1991). This system is fully integrated within the ARPEGE-IFS software that was conceived, developed and maintained in collaboration with ECMWF.
This model is a spectral global model with a stretched grid having a horizontal resolution around 7.5 km over France and 30 37 km over the antipodes. It has 105 vertical levels according a following-terrain pressure hybrid coordinate, with the first level at 10 m above the surface and an upper level at around 70 km. Clouds and precipitation are described by using three different scheme in the ARPEGE model. The stratiform clouds in terms of cloud profile and precipitation are explicitly modeled from the microphysical condensation scheme by Lopez (2002). The large-scale effects of deep convection are parametrized from a mass-flux scheme derived from Bougeault (1985) and the shallow convection ones with the Bechtold et al. (2001) one. In these last two cases, the cloud fraction and the liquid water, ice and precipitation profiles are diagnosed.
ARPEGE has four analyses per day at 00, 06, 12 and 18 UTC. Since June 20th, 2000 the operational data assimilation system of the ARPEGE model is a 4D-Var. This implementation, as detailed in Janiskova et al. (1999) and Rabier et al. (2000), is used 5 to provide an analysis which corresponds to the best atmospheric state knowing observations, an a-priori state, dynamical and physical constraints. The background error statistics are derived from a climatological matrix and an 25-member assimilation ensemble which runs for every analysis times. The control variables considered are temperature, specific humidity, vorticity, divervence and the logarithm of the surface pressure.
At each analysis around 7 million observations are assimilated. They include conventional observations (from radiosound-10 ing, aircraft, ground stations, ships, buoys, etc.) and satellite data. These latter include radiances in the infrared and microwave spectra such as AIRS (Atmospheric InfraRed Sounder), IASI, CrIS (Cross-track Infrared Sounder), SEVIRI (Spinning Enhanced Visible and InfraRed Imager), AMSU-A (Advanced Microwave Sounding Unit-A), MHS (Microwave Humidity Sounder), ATMS (Advanced Technology Microwave Sounder) and atmospheric motion vectors. Scatterometers provide information on ocean surface wind. Zenithal total delay signals and from radio-occultation measurements from the Global Naviga- 15 tion Satellite System (GNSS) are also assimilated.
With the advent of hyperspectral sounders such as AIRS and IASI, a variational bias correction (VarBC) method (Auligné et al., 2007) has been operationally implemented at Météo France and notably in the ARPEGE model. The VarBC scheme aims to minimize systematic innovations in radiances while preserving the differences between the background and other observations in the analysis system.

20
The observation operator translates the atmospheric variables quantities into the measured quantities for comparison with the actual measurements. For satellites radiances, it includes a radiative transfer model. It has then some limitations. Indeed, several atmospheric conditions are difficult to model and impose to exclude them in the assimilation process. The infrared radiances as IASI observations affected by clouds must be treated more carefully than in clear-sky conditions as their modelisation is more difficult. 25 The assimilation of clear radiances at Météo France is based on the McNally and Watts cloud detection scheme (McNally and Watts (2003)) which was designed to detect and isolate cloudy radiances from the clear-sky spectrum for a particular pixel. The method consists in finding the altitude at which the cloud affects the radiances and in filtering out the contaminated channels. The observed spectrum is compared to a clear sky simulated spectrum from the model guess. Channels are ordered according to their altitude sensitivity. This ranked partition is computed separately for each band sensitivity (CO2, water vapour, 30 ozone. . . ). In addition, a cloud characterization is made using cloud parameters (a cloud top pressure (PTOP) and an effective cloud fraction (Ne)) deduced from a CO 2 -slicing algorithm ( Pangaud et al. (2009)). These two parameters are used to model the radiative impact of cloud as a single layer cloud, with an emissivity set to 1 using a clear sky radiative transfer model.

Main features of the IASI instrument
IASI is a key element of the Metop series payload of European polar orbiting meteorological satellites (Cayla, 2001). It was designed by CNES (Centre National d'Etudes Spatiales) in cooperation with EUMETSAT. The first flight model was launched in 2006 on board the first European meteorological satellite Metop-A in polar orbit. The second instrument, mounted on the Metop-B satellite, was launched in September 2012. The third instrument will be mounted on the Metop-C satellite, which is 5 scheduled to be launched during the Autumn 2018. The horizontal resolution of the instrument is 12 km at the nadir. IASI is dedicated to operational meteorological soundings with a high level of accuracy (specifications on temperature accuracy: 1 K for 1 km and 10% for humidity (Chalon et al., 2001)). Its measurements are also useful for atmospheric chemistry to estimate and monitor different trace gases such as ozone, methane or carbon monoxide on a global scale (Hilton et al., 2012).

10
IASI is a passive IR remote-sensing instrument using an accurately calibrated Fourier transform spectrometer to cover the spectral range from 3.62 µm (2760 cm −1 ) up to 15.5 µm (645 cm −1 ) with 8461 channels. Its spectral resolution is 0.5 cm −1 with a spectral sampling of 0.25 cm −1 . The IASI spectrum can be divided into three major bands: from 645-1210 cm −1 : CO 2 , window and ozone channels mainly sensitive to temperature, called long-wave (LW) channels;
Only a subset of 314 channels (300 channels selected by Collard (2007) and 14 additional channels for monitoring purposes) used in operations at Météo-France, is considered in this study 2.1.2 Towards the assimilation of cloudy infrared IASI radiances 20 Assimilation of cloudy radiances is a crucial challenge for NWP centres as the cloudy observations discard represent an underexploitation of hyperspectral sounders and an error source in sensitive meteorological areas (McNally, 2002;Fourrié and Rabier, 2004). As mentioned in the introduction, studies about all-sky infrared assimilation have started. The radiative transfer model RTTOV-CLD for cloudy sky, included in RTTOV version 11 (Saunders et al., 2013), offers a realistic modeling of the cloud scattering. This model also allows to better describe the cloud emissivity as well as cloud scattering, using the 25 microphysical cloud profiles (water content, cloud ice content and cloud cover).
To simulate the radiances observed in cloudy conditions using RTTOV-CLD, we use two main types of clouds: firstly liquid water cloud, which corresponds to the Stratus Continental and Stratus Maritime ; secondly the ice water cloud of the Cirrus type, using Baran parameterisation (Vidot et al., 2015) to define the optical properties.
To illustrate the benefit brought by RTTOV-CLD, Figure ( in clear sky (figure 1.b) to simulate IASI observations, despite the presence of some cases with almost similar values, many cloud structures are not well simulated due to the lack of cloud information in the radiative tranfer simulation. On the other hand, when IASI observations are simulated using RTTOV-CLD (figure 1.c), a good agreement is obtained and similar cloud structures are found, for example, over the North Atlantic (30N-70N, 40W-0W) and above (30S-70S, 60W-0W) the Southern Atlantic Ocean. In this case, clouds in mid-latitudes are better simulated than in the Tropics. This may be explained by the fact 5 that clouds are better simulated in the ARPEGE model for mid-latitudes than in the Tropics.

Selection method of homogeneous observations
The assimilation of cloudy radiances in NWP models remains an innovative challenge. In the context of the preparation of all-sky assimilation, we plan to assimilate clear or cloudy observations that are completely covered the IASI FOV in a homogeneous way, discarding the cases of fractional cloud observations. These scenes are supposed to be better characterized 10 and simulated than fractional cloudy scenes in NWP models. Indeed, by selecting homogeneous cloudy scenes in both model and observation spaces, we improve the agreement between observations and background simulations. This selection of cases seen as homogeneous by both IASI and the model avoids misplacement errors. In this section, limited to cases over sea to avoid problems related to the land surface properties, we describe several methods for analysing the homogeneity of the scene in the observation and model space. However these methods were applied over all surface in the assimilation experiments of 15 section 4.

AVHRR clusters
In order to select homogeneous pixels, the AVHRR imager information collocated within IASI on the MetOp platform is used.
The spatial resolution of AVHRR observations is around 1 km at nadir and measures the radiation emitted in six broad-band channels: one visible channel, two near-infrared channels, a shortwave infrared channel and two long-wave infrared channels 20 (10.5 µm and 11.5 µm). Two IASI Level 1c products provided by EUMETSAT were used: the AVHRR clusters (Cayla, 2001) and the percentage of cloudy AVHRR pixels in the IASI FOV (product GEUMAvhrr1BCldFrac: Pequignot and Lonjou (2009)).
The AVHRR pixels are clustered into homogeneous classes in the radiance space, (visible and infrared channels) using the Kmean classification algorithm. For each AVHRR cluster and each AVHRR channel, the mean radiance, the standard deviation and the class coverage in the IASI FOV are given.

Selection criteria for homogeneous observations
This study intends to focus on those IASI pixels that contain only one cluster, which corresponds to a homogeneous scene.
However only 2% of daytime IASI observations over sea contain only one class. The aggregation is built with all available AVHRR channels (visible, NIR, IR), several classes can be produced with the K-mean classification even with relatively small standard deviations for the IR channels. An IASI FOV with several classes, each one having a small standard deviation and a 30 mean radiance close to the ones of the other classes, can thus be more homogeneous than a FOV with a single class.  and RTTOV-CLD (c) for surface channel (1271, 962.5 cm −1 ) for 30 January 2017 daytime over sea from ARPEGE 6-hour forecast fields.
For this reason, the number of AVHRR clusters within each IASI pixel has not been used as a homogeneity criterion, but these characteristics have been used to calculate the overall AVHRR cluster statistics, aggregating the information provided by all clusters in the IASI FOV. We tested four methods for selecting homogeneous scenes by calculating homogeneity criteria in the observation space as well as in the model space, using the AVHRR channels. The first two ones are described in the literature and we propose two other ones which are detailed below.
3.2.1 Homogeneity criteria derived from Martinet and al., (2013) These homogeneity criteria are based on a single AVHRR infrared channel 11.5 µm, which is used to compute three homo-5 geneity tests, the first two tests are calculated in the observation space and the third one in the model space:

Intercluster homogeneity
The intercluster homogeneity is based on σ inter defined as: Where L j is the mean radiances of cluster j at channel 11.5 µm , L mean represents the radiance weighted average. The 10 weighting is determined by C j is the cluster fraction of each class inside the IASI pixel. N is the number of classes in the IASI pixel.
A small calculed standard deviation σ inter means that all classes observe a similar cloudy scene in the infrared channel. If this standard deviation is too high, each class observes a different scene (clear or cloudy) and the IASI pixel is very heterogeneous.
Intracluster homogeneity 15 In order to finalize the homogeneity criterion in the observation space, it is also necessary to check if each class itself is sufficiently homogeneous, using the following formula: Where the σ j are the standard deviations of each cluster j calculated for the infrared channel 11.5 µm. The IASI observation is considered homogeneous if it verifies the following criteria:

Background departure check
Finally, in order to obtain a similar criterion in the model space, each AROME grid point within IASI FOV was used to simulate the equivalent AVHRR channel 11.5 µm with RTTOV-CLD. Homogeneous IASI observations are preserved if the ratio of the standard deviation of the AVHRR simulations and the simulated mean radiance of the AVHRR is less than 8%.
Adaptation of the method 5 In our study, which focuses on the ARPEGE global model, we chose to use the simulated brightness temperature from the guess profiles coming from a 6-hour forecast and interpolated using 12 points surrounding the observation position. The homogeneous cases are retained as long as the difference between AVHRR observations and simulations is less than 7 K. This method will be noted M2013. (2014) 10

Homogeneity criteria derived from Eresmaa
The study of Eresmaa (2014) aimed to propose an imager assisted cloud detection for the global ECMWF NWP system and was based on the hypothesis that each AVHRR cluster are made of fully clear or fully cloudy pixels.
Therefore, his selection criteria is only intended to diagnose and retain observations when they were completely clear, using the last two infrared channels of AVHRR (10.5 µm and 11.5 µm). This detection is based on three checks called the homogeneity check, the intercluster consistency check and the background departure check. If a IASI pixel do not satisfy one 15 of these checks, it is not free of cloud and is rejected.
The standard deviation of the brightness temperatures of the two infrared channels from all pixels present in the FOV is used for the first check. If both standard deviations are over the pre-determined threshold values (0.75 and 0.80 K, respectively), it means that a cloud is potentially observed and the IASI observation is rejected.
The intercluster consistency check relies on the comparison between the properties of the different clusters within the IASI 20 FOV. The distance of each cluster to the background in both infrared AVHRR channels as well as the distance between each pair of clusters. A cloud is detected if there is a pair of clusters covering more than 3% of the IASI FOV and for which the intercluster distance exceeds the minimum value of the distances between these clusters and the background.
The distance between 2 clusters j and k is computed as the squared-summed intercluster departure: where R j i is the mean brightness temperature of cluster j for channel i. In addition the distance of the cluster j to the background is computed with: Where R BG i is the background brightness temperature for AVHRR channel i. The observation is rejected due to the diagnostics of the presence of a cloud if the following inequality is true and the coverage of clusters j and k is over 3%: The last check on the background departure is computed as a fractional-weighted mean of the squared-summed background departures: where N is the number of clusters in the IASI FOV and f j is the fractional coverage of cluster j. The presence of cloud is diagnosed if D mean exceeds the threshold value of 1K².

Adaptation of the method
Since this method assumes that each cluster is made of pixels that are either all clear or cloudy, its homogeneity tests have been 10 adapted to the selection of clear and cloudy pixels, with criteria that would fit our purpose, with the first test in the observation space and the second one in the model space. Selection thresholds were modified and all simulations from background made with RTTOV-CLD.
-Inter-cluster homogeneity. This test uses the standard deviation of the infrared brightness temperature, calculated on all clusters occupying the IASI field of view. The standard deviation is calculated in the same way as Eresmaa (2014) 15 but the IASI pixel is considered homogeneous if the two standard deviations (one for each channel) are below their predetermined threshold values of 0.75 K and 0.8 K respectively.
-Background departure check. In this test, we used the D mean proposed by Eresmaa (2014) but the IASI pixel is considered as homogeneous if D mean is less than 49 K².
This method is referenced as E2014 in the following.

20
The threshold values of the homogeneity criteria derived from Martinet et al. (2013) and Eresmaa (2014), are based on the analysis of statistics, applied to all IASI FOVs of the different situations (day/night at sea). Threshold values are specified in such a way that the standard deviation between the observations and simulations is not too large while keeping a fair amount of the observations.

25
This third method, (called Obs_HOM thereafter) proposes a homogeneity check in the brightness temperature space calculated only on the observation space, using both infrared AVHRR channels (10.5 µm and 11.5 µm). This inter-cluster homogeneity criterion is based on the relative standard deviation of AVHRR clusters inside the IASI pixel. This test is satisfied when all classes observe a very similar scene in the AVHRR infrared channels. To evaluate the interclass homogeneity, the standard deviation of the mean brightness temperature of clusters which occupy the IASI FOV has been calculated using the following formula: Where: R i,j is the mean brightness temperature of cluster j on channel i, R i,mean represents the weighted average on channel i, N is the number of classes in the IASI pixel, and C j is the cloud fraction. to the relative cluster standard deviation of the mean radiances (on the y-axis) for intercluster homogeneity for (a) the AVHRR IR Channel (10.5 µm) and (b) the AVHRR IR Channel (11.5 µm) Figure 2 provides a calibration to determine the thresholds to be used to define homogeneous scenes. These thresholds should lead to a sufficient size of the selected dataset and avoid selecting the fractional cloud as much as possible. Therefore we decided to select an observation if the relationship between intercluster homogeneity and mean radiance for both AVHRR 10 IR channels (10.5 µm and 11.5 µm) are less than 0.8%.

Compromise for the homogeneous scene selection
Based on the previous methods, we propose a fourth one which represents a compromise between them. Two AVHRR infrared channels (10.5 µm and 11.5 µm) are used, and we define two homogeneity criteria in the observed and simulated brightness temperature spaces. 15 The first criterion for homogeneity is the interclass homogeneity check which was used in the third method, calculated in the observation space (presented in section 3.2.3). Similarly, in model space, we used D mean (presented in section 3.2.2). Only observations that fulfilled the two following criteria were selected: -Ratio between intercluster homogeneity and mean radiance for two AVHRR IR channels (10.5 µm and 11.5 µm) < 0.8%.
-Sum of the average distances between each cluster and the background < 49 K².
This method is named COMPR in the following. All the four methods are sumerized in Table 1. To ensure that the monitoring is focused on overcast and clear scenes, the percentage of cloudy AVHRR pixels in the IASI field was used to assess the choice of homogeneity criteria.

10
Our dataset is made of 50% of the observations entirely covered by clouds and 12% of clear observations according to the AVHRR cloud cover. The bias and standard deviation of observations minus simulations (O-G), are shown in Figure 3.(a) for the 314 IASI channels. As expected, the best statistics are obtained for channels less affected by clouds (e. g. CO2 and water vapour high peaking channels).
The mean standard deviations are larger for window channels sensitive to the surface, therefore to the presence of clouds: 15 11.7 K with a bias of -0.6 K for window channels between 770-980 cm −1 and 11.0K with a bias of -0.7 K for window channels between 1080-1150 cm −1 . Channels between 650-770 cm −1 show an averaged standard deviation of 2.5 K and a bias of 0.06 K.
The M2013 selection method ( figure 3. (b)), reduces the standard deviation of 3.7 K with a bias of -0.16 K and to a standard deviation of 3.6 K with a bias of -0.37 K for window channels between (770-980 cm −1 ) and (1080-1150 cm −1 ), respectively.

20
For brightness temperature channels in the range between (650-770 cm −1 ), we obtain a standard deviation of 0.8 K with a bias of 0.14 K. The E2014 selection method (figure 3. (c)) improves the bias to 0.11 K and -0.16 K with a standard deviation of 2.0 K for the window channels between (770-980 cm −1 ) and (1080-1150 cm −1 ), respectively, while the standard deviation of  the brightness temperature channels between (650-770 cm −1 ) is reduced to 0.6 K with a 0.13 K bias. As expected, the impact is larger for surface sensitive (and thus cloud sensitive) channels.
With the Obs_HOM method ( figure 3. (d)), small statistics improvement is obtained: the standard deviation is slightly decreased to 10.5 K and the bias to -0.2 K for window channels between 770-980 cm −1 . For window channels between 1080-1150 cm −1 , the bias is reduced to -0.5 K and the standard deviation to 10.0 K, while the brightness temperatures between 5 (650-770 cm −1 ) present a bias of 0.2 K and a standard deviation of 2.2 K.
The COMPR method reduces the bias to -0.09 K and a standard deviation to 2.2 K for window channel between 770-980 cm −1 and a bias to -0.29 K and a standard deviation to 2.1 K for the second range of window channels between 1080-1150 cm −1 (Fig. 3.(e)). A lower bias and standard deviation result (0.1 K and 0.6 K, respectively) is found for the channels between 650-770cm −1 as scenes are constrained be clear or cloudy both in the observations and in the models.

10
To complete the comparison, the Probability Density Function (PDF) of the O-G differences was studied (Fig.4). Three channels were assessed: the window channel 1271 (962.5 cm −1 , whose weighting function peaks at around 1000 hPa), the mid-tropospheric water vapour channel 2701 (1320 cm −1 , weighting function maximum at around 400 hPa) and the lowtropospheric water vapour channel 5403 (1955 cm −1 , weighting function peaking at around 900 hPa), The distribution asymmetry is relatively small for mid and low tropospheric water vapour channels. The impact of clouds 15 is evident on the window channel, with differences ranging from -90 to 64 K. After the homogeneity criteria is been applied, narrower Gaussian distributions are observed for all channels with a significant improvement for the window channel. Using the M2013 criteria, differences in O-G for the window channel are significantly reduced, from -18 K to 20 K, and from -7 to 9 K using the E2014 criteria (Figure 4.g, h, i).
With Obs_HOM criteria (Figure 4.(j, k, l)), the O-G distribution is not much improved for all channels except for the low-20 tropospheric water vapour channel where the range is reduced from 60 K to 40 K. When the homogeneity criterion in the model space is added using the COMPR selection, the O-G distributions become symmetrical (and Gaussian) and centrered around zero for the three previously selected channels (Figure 4.(m, n, o)), which indicates the data are correctly diagnosed as homogeneous. Table 2 summerizes statistics about the different datasets. The bias and standard deviation obtained by the M2013 method 25 have some reasonable statistics before the assimilation (-0.6 K for the bias and 3.75 K for the standard deviation, for the window channels). The E2014 selection method seems relevant for selecting homogeneous scenes in terms of bias and standard deviation (0.11 K and 2.0 K respectively, for the window channels). However, the number of selected observations presents a disadvantage for this selection method, since we keep only 22% of the observations of which 10% are totally clear, 6% are totally covered by clouds and 6% are heterogeneous. These observations are distributed throughout the globe, but we keep 30 more observations on high latitudes.
The Obs_HOM method allows to keep 67% of observations, which 12% are totally clear and 32% are totally covered by clouds, but this method does not give acceptable statistics (bias of -0.2 K and standard deviation of 10.5K). When the test on observations minus simulations of the infrared channels AVHRR are added by the fourth method, results are improved. For window channels the bias is reduced to -0.09 K and the standard deviation to 2.1 K compared to -0.6 K and 11.7 K for all  observations, which presents a good score compared to the M2013 and Obs_HOM methods. In addition 36% of the observations is retained, compared to the whole dataset, with 10% of clear observations and 11% of cloudy observations of the total amount, which is a better result compared to E2014, which removes many more observations, and shows that the proposed methodology is effective.
The cloud cover distribution corresponding to the amount of observations that is kept (36%) is made of 28% of clear observa-5 tions and 29% of the observations totally covered by clouds. In addition, 14% of the observations have a cloud cover of less than 10% and 4% of the observations have a cloud cover exceeding 90%. The observations kept are distributed in different parts of the globe (Figure 5.a) although we have been able to retain different cloud types, including high clouds even in the tropics for few cases only ( Figure 5.b). This may be explained by the weakness of the model clouds in these areas.
The main objective of the study is to select homogeneous IASI observations in clear and cloudy sky which are well simu-10 lated with RTTOV-CLD and could be used in data assimilation. Comparison of different methods of selecting homogeneous scenes showed that the M2013 method improves the first guess departure statistics (bias of -0.16 K and standard deviation of 3.17 K) but it keeps more heterogeneous observations (25%) according AVHRR cloud cover than the E2014 method, which significantly improves the statistics (bias of 0.11 K and standard deviation of 2 K) and favours more clear observations but keeps only 22% of the observations. The Obs_HOM method, which focuses only on homogeneity in the observation space, 15 does not strongly improve the statistics but it filters 33% of heterogeneous observations. However the addition of the criterion on the simulated observations in the COMPR method improves the scores on IASI simulations (bias of 0.09 K and standard deviation of 2 K), while retaining 36% of the observations.
After the selection criteria were implemented in the assimilation system of Météo France, their impact was tested through a 4D-Var assimilation experiments in the ARPEGE global model. The impact of the homogeneity criteria for data selection on all observation simulations, on analyses and forecasts is evaluated.

Experimental design 5
To evaluate the impact of our homogeneity criteria on the assimilation process over sea and land, during daytime and night, four experiments were performed over one month from 06/12/2017 to 17/01/2018. 314 IASI channels were used in the simulation, and 129 channels (Tables A1 and A2 in Appendix) were used for assimilation as operationally.  In these experiments, no cloudy observations detected with the CO2-slicing method was assimilated.

Impact on observation
The number of assimilated channels for each observation is different depending the areas, e.g. in the tropics, there are less clear channels (between 30 and 40 channels) in the REF (Figure 6.a), EXP.A (Figure 6.b) and EXP.B (Figure 6.c), which is explained by the presence of high clouds in this region. In EXP.A (Figure 6.b) with the Obs_HOM criteria, some observations 20 were filtered in the tropical area, and even more in EXP.B (Figure 6.c) where the criterion used is even more stringent, more observations are filtered in areas corresponding to high clouds.

Impact on background and analyses
The analysis departure data discussed below are made by comparing the analysis between the REF and the two experiments (EXP.A and EXP.B) to evaluate the impact of the criteria for selecting homogeneous IASI observations (refer to Table 3). . This implementation removes some IASI observations from the assimilation and this reduction has an impact on the analysis. In Figure 8.a, a negative temperature difference is located in the Atlantic Ocean near to the South-West African coast. A similar behaviour on temperature is noticed from the impact of COMPR (EXP.B) as shown in Figure 8.c. Weaker and patchy impact is reported on specific humidity that is mainly located in the tropics. EXP.A and EXP.B seem to remove some 20 temperature and humidity analysis increments from the operational experiment (REF) just at some isolated locations.
In order to assess the impact of the new selection of IASI observations on the analyses and forecast, first guess departures (FG departures) corresponding to the difference between the observations and the simulation from the 6-h forecast and the analysis departures (AN departures) are computed. As biases and standard deviations of FG and AN departures were very weak for IASI, CRIS and AMSU-A instruments, and humidity measurements performed by radiosondes, relative differences 25 have been performed between experiments to highlight detailed comparisons. Concerning the CrIS observations ( Fig. 9.b), the differences results are mainly not significant excepted for the EXP.B where the standard deviation increases at around the 850 water vapour channels for both FG and AN departures, and a significant improvement for the AN departures at the channels 160.

5
Results obtained for AMSU-A ( Fig. 9.c) are mainly satisfactory with FG departure standard deviation differences reduced by around 0.05 % for channels 6, 9, 10, 11 and 12 with the EXP.A and EXP.B. However, a significant degradation of 0.2 % is observed for channel 8. The AN departure results follow more or less the same behaviour. Finally, no significant standard deviation difference is observed concerning the TEMP-q observations (Figure 9.d).
Results shown in this part report non-negligible impact of the homogeneous criteria implemented into EXP.A and EXP.B 10 on the analyses and the short range forecasts. Indeed, as seen in the previous section (section 5.2), selected IASI observations are removed over the more cloudy locations and then impact the humidity and temperature analyses (as seen in Figure 8).
Statistical results in Figure 9 report a non-negligible decrease of the dispersion within the FG departure and AN departure for IASI observation and AMSU-A for several channels but some negative impact have to be noted for other wavenumbers. More attenuated and mainly non-significant impacts can be recorded for CrIS and TEMP-q observations. Thus, the analyses and 15 short range forecasts have been slightly changed compared to REF.

Impact on forecast scores
The forecast from EXP.A and EXP.B 00UTC for the period 7 December 2017 to 17 January 2018 were compared to REF ones and evaluated against radiosondes and operational analyses from ECMWF. Rootmean square forecast errors at the 12-h forecast ranges with respect to the ECMWF analyses were computed for temperature, relative humidity and wind. Similar computations 20 were made against radiosondes . No major difference can be found between the three experiments. Very small improvements of the 12-hour forecast with respect to the ECMWF analyses were found in the Southern hemisphere for temperature and wind at around 700 hPa (Figs. 10). This reduction of 2% for temperature and 0.5% for the wind is significant accordind a Bootstrap test with a 99.5% confidence level. Other improvements are found at 200 hPa for temperature (1.5%) and at 500 hPa for wind (0.5%). Regarding the evaluation against radiosondes, very small, but not significant, improvements for the wind were found 25 in the troposphere in the Southern hemisphere and in the Tropics.

Conclusion and perspectives
A new method using of collocated AVHRR cluster information to improve the selection of homogeneous IASI observation scenes within the numerical weather prediction ARPEGE model has been developed at Météo-France for data assimilation purposes and has been presented in this study. 30 The first step consisted in adapting the IASI observation operator based on the RTTOV radiative transfer model by using the RTTOV-CLD module with cloudy microphysical parameters (liquid water content (ql), ice content (qi) and cloud fraction) partures, it was found that these two methods were not satisfactory in an operational context (in assimilation) due to a large IASI observation reduction. Then, two new sets of criteria were defined and implemented within the ARPEGE model: -The first criterion looks for the consistency between different clusters occupying the same IASI FOV by examining this homogeneity relative to the weighted average brightness temperature of the AVHRR clusters; it is only based on observations.
10 -In addition, the second criterion assesses the coherence of each cluster compared to the background; it is in fact a good compromise between the previous criterion and the two historical ones with accurate statistics and a sufficient number of observations that passed the check.
Therefore, assimilation experiments were conducted to assess the impact of these new selecting homogeneous IASI observation features in the current clear sky assimilation. This revised check was added to the McNally and Watts (2003) cloud detection. 15 The results obtained in this case show that the scenes categorization has been facilitated and cloudy observations can be better filtered out compared to what is done in the operational ARPEGE version. 3% of all observations are rejected with the compromise method and only 1% for the method based only on homogeneity in observation space which is more convenient for the assimilation. The impacts on the first guess and analysis departures (showing more Gaussian shape) are generally low but with a beneficial reduction on the standard deviation of first guess departures mainly on the IASI and AMSU-A observations.

20
Regarding the forecasts scores, neutral impact is reported when these selection criteria are taken into account on top of the McNally and Watts (2003) algorithm.
However, this step has been necessary to prepare the future which will consist of the assimilation all sky within the ARPEGE model. These methods of observation selection allow to separate the clear-sky and cloudy scenes and manage each route in an 25 independent way. Then, it could be available to directly assimilate the cloudy radiances into the 4D-Var ARPEGE by adapting the observation errors for more cloudy situations. However, hydrometeors used in the RTTOV-CLD are not available into the background error covariance matrix and then cloudy and convective situations are badly represented and will penalise the cloudy direct assimilation. In order to bypass this problem a the second solution under study is to retrieve information within cloudy observation by a Bayesian inversion method, in a first step, and assimilate these retrieved products in terms of 30 temperature and/or humidity profiles into the 4D-Var in a second step. This method called 1D-Bayesian + 4D-Var was already studied for microwaves radiances (Guerbette et al., 2016;Duruisseau et al., 2018)