Journal of Geophysical Research: Atmospheres

Towards the use of cloud microphysical properties to simulate IASI spectra in an operational context

Authors


Corresponding author: F. Faijan, Centre de Météorologie Spatiale, av de Lorraine, BP 147, Lannion FR-22300, France. (francois.faijan@meteo.fr)

Abstract

[1] Capabilities and limitations of two fast radiative transfer models simulating cloudy spectra of the Infrared Atmospheric Sounding Interferometer (IASI) have been investigated in this paper. These models include a better modeling of the clouds than current operational fast forward models, such as scattering effects of the radiation in the cloud layer. An accurate simulation of the IASI spectra in the presence of ice clouds or for vertically extended clouds is a necessary step toward the use of these cloud-affected radiances in an operational context. Through a collocation of IASI observations with an independent data set from the Lindenberg ground-based station and the A-Train space-borne active instruments, this study first examines the accuracy of the fast radiative transfer models to simulate the cloudy spectra and provides some indications of confidence in the use of the cloud microphysical and optical properties. This first step shows the high sensitivity toward the cloudy inputs (crystal shapes, particle size, and cloud water content) which can reach several kelvins. It also provides a screening method to process cloudy radiances and to deal with the strong sensitivity of the fast radiative transfer model. The performance of the scattering fast radiative transfer models is then assessed in a global operational through simulations during a one week period over the globe, with the profiles of cloud variables from the ECMWF forecast used to compute the radiances. The applied screening method has reduced the standard deviation between simulated and observed spectra from 9 to 3 K.

1. Introduction

[2] The Infrared Atmospheric Sounding Interferometer (IASI) onboard the MetOp satellite is a spectrometer with 8461 spectral channels ranging from 3.7 to 15.5 μm. These potentially high informative observations are now assimilated, mainly in clear conditions, at many operational meteorological centers, and give a significant positive impact on forecast skill [Collard and McNally, 2009; Guidard et al., 2011]. However, more than 80% of the whole globe is covered by clouds with a large diversity of situations: a single layer of ice or water cloud, several separated layers of ice and liquid clouds in the same field of view, vertically extended clouds with a mixed phase (liquid and ice)… Recently, all NWP centers have begun to operationally handle cloudy data, starting with the assimilation of cloud-affected radiances with basic assumptions on the cloud only applicable to very restricted conditions such as overcast opaque single layers [e.g.,McNally, 2009; Pangaud et al., 2009; Pavelin et al., 2008]. Although it is a step in the right direction, IASI data are still under-exploited. In a numerical weather prediction (NWP) context, a more realistic simulation of cloudy radiances from NWP fields is still some way off but more sophisticated modeling that account for cloud scattering are now available in some fast radiative transfer models (FRTM). This is an essential step toward a better exploitation of cloud-affected IASI data in operational systems, mainly for semi-transparent ice clouds which are more complex to simulate with a “classical” FRTM due to the high spectral dependency of the cloud emissivity. Moreover, FRTMs also have an interest in the climate application where the ice clouds play an important role in the earth's global energy budget.

[3] The radiative transfer models (RTMs) usually include full line-by-line models or pseudo line-by-line models whereas FRTMs are based on a regression scheme of line-by-line models. In either case, the simulation of cloudy radiances is typically absent unless the model is coupled to a full scattering scheme. Nevertheless, many FRTMs, in operational NWP, use a simplified approach for the computation of cloudy radiances. In this approach, clouds are considered as an opaque or semi-transparent single layer at a defined pressure level with an effective emissivity Nε over a range of selected wavelengths (N the cloud cover in the instrument field of view and ε the cloud emissivity). The simulated cloudy radiances can be expressed, for each wave number, as:

display math

with Robs, Rclear and Rcloudrespectively the satellite observed radiances, the clear-sky radiances and the overcast cloudy radiances at the cloud top pressure CTP. This approach is used in this paper with a constant effective emissivity (Nε). Before computing the radiance spectra with equation (1), the cloud top pressure and emissivity are first extracted from the IASI measurements themselves using such methods as the CO2-Slicing method [Menzel et al., 1983; Wylie and Menzel, 1999], the minimum residual method [Eyre, 1989] or through a 1Dvar retrieval system [Pavelin et al., 2008]. These cloud retrieval schemes are mostly applied in the long-wave CO2band because of the constant cloud emissivity assumption and a very different value of emissivity in the short waves. At one step of the process, all the methodologies consist of a minimization of the residuals between the measured radiances and the simulated cloudy overcast radiances as function of the cloud pressure level, for a predefined set of channels. If the forward model used to compute the cloudy overcast radiances considers a simple cloud modeling of a single thin layer with a constant emissivity, retrieved cloud pressures can be biased, depending on the set of channels involved. As a consequence, the shape of the simulated spectrum can be quite different from the observed one, mainly for thin semi-transparent ice clouds whose observed spectrum slope can be quite large in the long-wave windows channels.

[4] Furthermore, for multilayered or vertically extended systems which represent more than half of the cloudy situations, the retrieved cloud pressure is an average value in the cloud layers which depends on the cover and thickness of the upper cloud layers and the residuals can also be large which is not the case with optically thin cirrus clouds.

[5] In assimilation systems, most of thin semi-transparent, multilayered and extended systems are discarded during the screening step, due to the inaccuracy of the radiative transfer calculations. This greatly reduces the amount of assimilated data, as shown inLavanant et al. [2011]who compared cloud products within the IASI footprints, from ten operational schemes over a 12 h global acquisition. The Lavanant et al. paper showed that the main meteorological cloud structures were correctly retrieved by all schemes, with a good agreement in the cloud height of the overcast and single layers. An accurate assimilation of these particular conditions can then be expected. However, for all the other clouds (semi-transparent, multilayer) the retrieved cloud heights and emissivities were very different due to the complex nature of the physics. For the sake of convenience, the cloud radiative transfer methodologies as described here (e.g.,equation (1)) will be referred to as “absorption approach” in the following sections. The ‘standard’ way to use the RTTOV [Saunders et al., 2010] model follows this approach, and is used in many NWP centers (e.g., RTTOV at UKMO, ECMWF, Météo-France…).

[6] To increase the amount of cloudy data in operational systems, a different strategy to compute the spectra is necessary in the presence of ice clouds or vertically extended clouds, spectra for both types being often poorly simulated by the “absorption approaches.” To deal with these situations, different techniques that take into account the microphysical and optical properties of cloud particles have recently been explored in fast radiative transfer models, which can, for example, simulate the steep slopes in the radiance spectra in the long-wave window channels for clouds of small effective diameter (De). To simulate the IASI high-resolution spectral signature, the new methodologies have introduced single or multiple scattering parametrizations for different types of ice crystals or liquid water droplets. For a non-scattering medium, as it is supposed in “absorption models,” the absorption is the sole process of extinction of the radiation going through the medium whereas the new fast models consider that the extinction is the sum of the absorption and of the scattering. In such cloud modeling, effects resulting from the overlapping (in the vertical) of partially overcast cloudy layers also have an impact on the computation of cloudy radiances.

[7] Although, they have existed for decades, radiative transfer models, including multiple scattering, need quite demanding computational requirements to simulate cloudy high-resolution IR spectral radiances; this is the case for DIScrete Ordinates Radiative Transfer DISORT model [Stamnes et al., 1988]. The purpose of the parameterizations of fast models is to retain the largest degree of accuracy without compromising the computational efficiency of codes, in order to increase the amount of cloud affected radiances in the assimilation. For example, the scattering parameterizations can infer the dramatic change of the cloud emissivity for ice crystals in the long-wave CO210 micron band and simulate observed important slopes in the long-wave window channels as will be illustrated later in the article.

[8] This paper is organized as follows. Section 2gives an overview of two fast cloud-scattering radiative transfer models using different methodologies. Insection 3, the limitations and capabilities of these radiative transfer models are assessed using data sets collected during the Lindenberg ground-based campaign, followed by a discussion on their uncertainties. Insection 4, an investigation on a screening method to process cloudy radiances based on IASI/A-Train co-locations is described. Finally, insection 5, the use of the fast forward cloud models in conditions close to an operational context with the ECMWF (European Centre for Medium Range Weather Forecasts) cloud variables as an input is tested and some prospects are given.

2. Fast Cloud-Scattering Radiative Transfer Models

[9] In this study two fast cloudy radiative transfer models that simulate cloudy radiances by taking into account, unlike ‘the absorption models’, the scattering and absorption radiation by clouds for hyper-spectral IR sounder measurements have been evaluated. The absorption approach of RTTOV is currently used for simulating cloud-affected radiances in ECMWF, the UK Met-Office, Météo-France and other weather forecast centers. The scattering modeling in RTTOV, interfaced with NWP cloud variable profiles, is available since the RTTOV-9 version. This software is available on the NWP Satellite Application Facility (SAF) web site. The HISCRTM cloud software is used by different university teams to retrieve cloud microphysical properties from AIRS measurements [e.g.,Wei et al., 2004; Huang et al., 2004; Li et al., 2005]. Dr Jun Li supplied the software to us with the IASI coefficients. The main aspects in the two models, which are useful for the understanding of the following sections, are summarized hereafter.

2.1. RTTOV Cloud-Scattering Model

[10] The RTTOV model was originally developed at ECMWF [Eyre, 1991] to retrieve temperature and humidity profiles from TOVS (TIROS Operational Sounder) data. Subsequently the code went through several developments within the EUMETSAT NWP SAF, RTTOV-10.2 being the latest version (J. Hocking et al., RTTOV v10 Users guide. NWPSAF-MO-UD-023, 2011. Software available at:http://www.research.metoffice.gov.uk/research/interproj/nwpsaf/rtm/rtm_rttov10.html), used in this paper. The parameterization of multiple scattering of the radiation in RTTOV was first introduced in the RTTOV-9 version [Saunders et al., 2010] and a detailed description of the scattering scheme is given by Matricardi [2005]. The contribution of the thermal scattered radiation is obtained by replacing, in the radiative transfer equation, the absorption optical depth by an effective extinction optical depth τcld_ext defined as:

display math

where τcld_abs and τcld_scat represent respectively the cloud optical depths of absorption and scattering. b is the integrated fraction of the incident radiation energy scattered backward with the hypothesis that the diffuse radiance field is isotropic, based on the approach followed by Chou et al. [1999]. This assumption avoids explicit computations of multiple scattering and allows a computational efficiency comparable to the “absorption approach.” The extinction optical depth is computed at each atmospheric layer of RTTOV. It requires, in addition to the atmospheric temperature profiles, input profiles of cloud ice water content CIWC and/or cloud liquid water content CLWC. The CIWC and CLWC profiles are available from numerical weather forecast centers, such as ECMWF.

[11] For water clouds, a database of 5 cloudy optical properties was generated using the microphysical properties assembled in the Optical Properties of Aerosols and Cloud (OPAC) software package [Hess et al., 1998]. Given one of the 5 cloud types, the input CLWC profile is then converted by RTTOV into an effective extinction optical depth.

[12] For ice clouds, the size and shape of ice crystals can vary greatly from polar regions to midlatitudes and Tropics from bullet rosettes, hollow and solid columns, to plates and aggregates. However, the choice was made to represent the optical properties as randomly oriented hexagonal columns and of ice aggregates in RTTOV. Optical properties of hexagonal ice crystals were determined by using the Geometric Optics method [Takano and Liou, 1989] or the T-Matrix methodKahnert [2004] depending on the wavelength radiation and the size of the ice crystal. Then the radiative properties of clouds were built from the optical properties and 30 ice distributions representative of cirrus clouds. For aggregates, optical properties from the Baran and Francis [2004] database were used to build the radiative properties of clouds. These radiative properties were then used to determine regression coefficients function of CIWC, De and the crystal shape.

[13] To estimate the input De profile, users can use one of the four possible parameterizations available in RTTOV. These parameterizations, based on measurement campaigns, are function of the CIWC and/or the temperature profiles provided by users: two of them [Boudala et al. 2002; Wyser, 1998] use both temperature and ice water content in the layer, Ou and Liou [1995] use the temperature only and McFarquhar et al. [2003] only uses the ice water content. RTTOV version 10.2 also allows an ice crystal effective diameter to be specified as user input besides the four parameterizations.

[14] Finally, the radiative transfer equation is solved by dividing the atmosphere into a number of homogeneous columns, using the maximum-random overlap assumption [Amorati and Rizzi, 2002]. Each column, which contributes to a fraction of the total radiance, is characterized by a different number of cloudy layers. Once the top of the atmosphere radiance has been computed for each homogeneous column, the cloudy radiance is computed as the sum of all the single column radiances weighted by the column fractional coverage. This allows the treatment of partially cloudy scenes, vertically extended clouds and mixed phase situations together.

[15] The RTTOV cloud-scattering model was mainly designed to be used in a NWP variational process and allows the use of NWP cloud profiles as control variables.

2.2. HISCRTM

[16] The fast High Spectral Cloudy Radiative Transfer Model HISCRTM was developed through joint efforts at the Texas A&M university and the University of Wisconsin-Madison to infer the optical thickness of semi-transparent ice clouds. A detailed description of the version used in this study is given inWei et al. [2004]. HISCRTM assumes that the clouds are located in a plane-parallel, single homogeneous and isothermal layer for any given IASI footprint. Because the surface reflectance in the IR region is small, only the first-order scattering between the surface and the cloud is considered and the upward radiation below the cloud is assumed to be isotropic (in the IR region). Water droplets are assumed to be spherical and the classical Lorenz–Mie theory is used to compute the cloud optical parameters. For ice crystals, observation campaigns indicate that in the upper portions of cirrus clouds, ice particles tend to be quite small and that the shape of these small ice crystals may be represented by droxtals. On the contrary, the middle layers of midlatitude synoptic-scale cirrus clouds are often composed of pristine particles such as hexagonal columns. And in the lower portions of cirrus clouds, ice particles tend to be larger and composed of irregular aggregates. Based on these observations, HISCRTM selects one of these 3 crystal shapes depending on the user defined particle size: droxtals for small particles (0–50 μm), hexagonal columns for moderately sized particles (50–300 μm), and aggregates for particles larger than 300 μm. The crystal optical properties were determined using either the Geometric Optics method or the FDTD method (depending on the crystal size and radiation wavelength). Cloudy microphysical (De) and optical (COT) properties are then obtained by combining the crystal optical properties previously defined with 30 particle size distributions [Fu et al., 1998]. The cloudy properties are finally used as input of DISORT to retrieve a database of reflection and transmission for i) satellite observing zenith angles ranging from 0° to 80°, ii) wavelengths from 3 to 20 μm, iii) optical thickness from 0.04 to 100, iv) effective particle sizes up to 200 μm. For a given IASI observation, the cloud transmission and reflections are then computed through interpolation in the pre-computed look-up tables to solve the cloudy radiative transfer equation with the cloud located at a user-defined level and for user-estimated cloud properties. The scheme assumes that multiscattering effects are negligible, of the order of a few percent at IR wave numbers.

[17] The input parameters of HISCRTM are the zenith satellite angle, the surface emissivity and surface temperature, the air temperature profile and the clear-sky transmittance at each level of the chosen clear-sky model and for the computation of the cloud absorption and scattering, the cloud cover, the visible cloud optical thickness COT at 0.55 μm, the ice/liquid phase, a mean De in the cloud and the CTP. No information is needed on the ice crystal shape as it is internally computed from the input effective size of ice crystals. For this study, the clear-sky transmittances are derived by RTTOV on the pressure or model levels of the input profile. The visible COT is derived from the CIWC or CLWC profiles and the COT profile is integrated along all the vertical levels to get a column integrated optical thickness, which is then located by the software in the layer at the input CTP.

[18] In this software, it is easy to use information from the observations themselves or from other instruments: e. g. CTP derived from the CO2-Slicing method or CTP, De and COT from a co-registered imager (with visible channels). The CTP information is important as, due to the nonlinearity in clouds, the model is very sensitive to a correct evaluation of the height of the cloud.

[19] The HISCRTM version used in this study reduces the cloud to a single pressure level and limits the comparison of the two models to situations with a unique phase, either liquid or solid. However, a new version is now available which allows multilayer cloud processing.

2.3. Computational Efficiency

[20] To give an order of magnitude of computational times, the CPU times were evaluated for the two models in cloudy conditions. They are respectively, in our work environment, of 300 ms for HISCRTM and 400 ms for RTTOV for a single layer cloud; this time value increases by 75 ms for each added cloudy layer in RTTOV. Nevertheless, this FRTM allows the computation to be performed in parallel on several processors with the OpenMP standard, which decreased significantly these computation times (reduced computational time by 4). These results are independent from the cloud phase.

3. Discussions on the FRTMs

3.1. Performances of FRTMs Through Validation Campaign

[21] IASI measurements are compared to the RTTOV and HISCRTM simulated spectra. We use CLWC and CIWC profiles measured during a ground-based field campaign in Lindenberg in order to compute the spectra. The limitations and capabilities of the two scattering models are assessed through this comparison using a set of typical situations.

3.1.1. EUMETSAT Lindenberg Validation Campaign

[22] An atmospheric sounding campaign within the EUMETSAT Polar System Program (PSP) in support of the calibration and validation of IASI was performed in Lindenberg, east of Germany, from June 01, 2007 to August 31, 2007 [Stiller et al., 2007]. A total of 290 radiosondes were launched at the overpass of the satellite. Furthermore, measurements with various ground-based remote-sensing systems supported the characterization of the state of the atmosphere. Radiosoundings, data from a microwave profiler, lidar ceilometer and Ka-band radar were combined to get an optimal estimate of the atmospheric state characterized by profiles of temperature, humidity, cloud liquid and ice water content and cloud top pressure. The cloud cover was determined by human observation and a whole sky imager. In situ measurements of surface meteorological parameters completed the data set.

[23] The RMS of departure (obs - simulated) for the 9 clear-sky situations using RTTOV for both the Lindenberg and ECMWF profiles has been evaluated (not shown). The RMS in the CO2window region is less than 0.7 K using the Lindenberg profiles and 1 K using the forecast profiles. For the water vapor channels, the RMS values are still small for the forecast profiles, of the order of 0.8 K, but reach up to 2 K with the Lindenberg inputs. From these results, it was decided to use as inputs, in this study, the atmospheric profiles and surface parameters from the ECMWF forecast instead of those of the radiosonde. Although this result is unexpected, it can be assumed that the NWP data are probably more representative of a 12-km field of view than radiosounding and local surface parameters. No bias corrections are applied on the IASI data in the following analysis of cloudy scenes, due to the lack of statistics in representative cloudy conditions and the difficulty to rely on clear-sky biases in all cloud type conditions.

[24] Although all cloudy situations were processed, only two cloudy situations are presented in details in this paper which are quite representative. The first case is an overcast low-level water cloud. The second is a typical semi-transparent ice cloud with a significant slope in the window channels, between 800 and 1000 cm−1. Through these 2 cases, the performance of FRTMs and the sensitivity to cloud parameters are discussed.

3.1.2. Case of a Low-Level Liquid Cloud

[25] Figure 1shows the CLWC Lindenberg profiles at the IASI overpass for June 2nd at 8h53. The cloud layer which is around 880 hPa and consistent with the MAIA cloud mask (L. Lavanant, MAIA-3 AVHRR cloud mask and classification. EUMETSAT contract, 2002. Documentation available at:http://www.meteorologie.eu.org/ici/maia/maia3.pdf) indicates the presence of an overcast opaque low-level cloud at 875 hPa. Additionally, the observation (Figure 2a) corresponds to the signature of an opaque low cloud with warm brightness temperatures (282 K) and a flat spectrum in the window channels. The CO2-Slicing CTP of 872 hPa and effective emissivity of 0.94 were used to simulate the spectrum with the RTTOV absorption approach and the departure to the observed spectrum is shown inFigure 2b. The HISCRTM and RTTOV (with the stratus continental parameterization) results are respectively presented in Figures 2c and 2d. Both cloud-scattering models simulate well, and in a similar way, the observed spectrum except in the 1100–1250 cm−1region where the simulated BTs are slightly too warm. The absorption assumption is accurate enough to simulate the observation in case of an opaque or nearly opaque cloud. Similar conclusions were found for other opaque liquid situations and also for high-level opaque ice clouds (not shown).

Figure 1.

Lindenbreg situation of June 02nd 2007, 08h53. IWC profile of a low-level overcast cloud.

Figure 2.

Lindenberg situation of June 2nd 2007, 08h53. Low-level liquid cloud. (a) Observed IASI spectrum. Residuals for (b) RTTOV with CTP and Nεfrom the CO2-Slicing method, (c) RTTOV with the stratus continental parameterization, and (d) HISCRTM with CTP from the CO2-Slicing method.

3.1.3. Case of a Semi-transparent Ice Cloud

[26] The situation of June 21st 2007 at 09h01 presents a good spatial and temporal co-location (less than 10 km and 5 min) between the IASI and the Lindenberg measurements. The situation, for which the CIWC profile is shown inFigure 3, corresponds to an overcast high-level thin cloud with a small visible COT of 2.04 (1.27 in IR). This is in agreement first with the strong 10 K slope in the IASI window region which is typical of COT values less than 5 and small De and second with the CO2-Slicing CTP at 240 hPa and the relatively small effective emissivity Nε(0.69). Moreover, the AVHRR cloud mask detected an overcast semi-transparent cloud and the observer reported a cirrostratus at the MetOp overpass.

Figure 3.

Lindenberg situation of June 21sh 2007, 09h01. IWC profile. The situation is overcast.

[27] For the current case, relatively large De values of 60 μm in HISCRTM and 57 μm (constant profile) in RTTOV were found manually to fit the IASI spectrum slope. Figure 4respectively shows the observed IASI spectrum (4a) and the simulated-observed residuals with the RTTOV absorption approach (4b), RTTOV (4c) and HISCRTM (4d). The residuals with the RTTOV absorption model are small but the method is unable to simulate the slope of the measured spectrum. The better fit to the slope using RTTOV or HISCRTM, indicates that the use of the cloud microphysical and optical properties could help to better describe the IASI spectrum for this type of clouds. To conclude with this case, having an accurate cloud ice profile in input, both models were able to correctly simulate the scattering effects of the ice particles.

Figure 4.

Lindenberg situation of June 21th 2007, 09h01. Semi-transparent ice cloud. (a) Observed IASI spectrum. Residuals for (b) RTTOV with CTP and Nεfrom the CO2-Slicing method, (c) RTTOV with the hexagonal ice crystal shape option and a mean De of 57 μm, and (d) HISCRTM with CTP from the CO2-Slicing method and a De of 60 μm.

3.2. Impact of Uncertainties of Cloudy Inputs

[28] We have shown, through 2 cloudy situations of the Lindenberg campaign, the capabilities of RTTOV and HISCRTM to simulate the IASI cloudy spectra. Although the obtained biases for these two simulations are acceptable in a data assimilation process, some questions are worth asking. What are the impact and significance of the effective diameter (De) in these two models? What is the impact of the crystal shape? What is the impact of cloudy uncertainties on the computed spectrum? And more generally, how acceptable are the uncertainties are for an operational use?

3.2.1. Effective Diameter De

[29] The definition and the meaning of the effective diameters (De) of cloud particles composed of nonspherical ice crystals is ambiguous since it does not represent a physical or measurable property of the ice crystal. Moreover, no universal definition is used in FRTMs.

[30] For semi-transparent cloud situations described insection 3.1.3, we evaluated the different parametrizations of the effective diameter of RTTOV (Boudal, McFarquhar, Wyzer and Ou & Liou) for randomly oriented hexagonal columns and aggregate crystals, the 2 available shapes in RTTOV. We also assessed the impact of the effective diameter in HISCRTM fo 4 effective diameters (15, 30, 60 and 90 microns).

[31] 1. HISCRTM: The difference brightness temperatures in HISCRTM, for two extreme values, are of the order of several kelvins (6 Kelvins between 15 and 90 microns at 1200 cm−1). The De value changes the slope in the CO2 10 μm band. Note the inflection point at 906 cm−1 where De is very weakly correlated to the brightness temperature (Figure 5c). If it seems easy to determine the De value in the model HISCRTM, via a simple algorithm presented in Wei et al. [2004], this work is less obvious for RTTOV because a change of De has impacted significantly not only the slope, but also the spectral region of the window channels.

Figure 5.

Impact of the 8 possible choices in RTTOV on the simulated IASI spectra. Boudala is in black, McFaquhar in red, Ou & Liou in green, Wyser in blue. (a) Randomly oriented hexagonal columns. (b) Ice aggregates. (c) The sensitivity of De in HISCRTM (De = 15, 30, 60, 90 μm).

[32] 2. RTTOV: Figure 6 shows the ice crystal effective diameter De as function of the CIWC and different cloud top temperatures for the 4 possible parameterizations in RTTOV. This figure illustrates the large variability of the spectra due to the De parameterizations. The differences can reach 20 K in the simulated brightness temperature depending on the user's choice (Figure 5a) between the parameterizations of Boudala and Wyzer. There are numerous facts explaining this huge variability, including the De definition used to build the parameterization [McFarquhar and Heymsfield, 1998; McFarquhar et al., 2007].

Figure 6.

Impact of the 4 possible parameterizations in RTTOV on the effective diameter. Boudala is in red with triangles, McFaquhar in gray with diamonds, Ou & Liou in blue with circles, Wyser in black with stars.

3.2.2. Crystal Shape

[33] Figure 5 also illustrates the sensitivity to the crystal shape. In HISCRTM, because the shape is correlated with the size of the crystal (see section 2.2), it is difficult to measure the impact of the crystal shape on the brightness temperature.

[34] In RTTOV, the uncertainty of the crystal shapes (randomly oriented hexagonal columns or aggregate) can lead to uncertainty on the brightness temperature of several kelvins (5 K based on the parameterization of McFarquhar around 1200 cm−1). This uncertainty decreases as the optical thickness of the cloud increases [Wendish et al., 2007]. During our study, we use the randomly oriented hexagonal crystals because the simulated spectra shapes with the aggregate crystals were never observed with the IASI observations during our study.

3.2.3. Cloud Water Content

[35] We have evaluated the impact of cloud water content uncertainties on the brightness temperature. For this, we have statistically computed the error covariance matrix of the cloud ice/liquid water content from ECMWF forecast data by determining the departure between the analysis and the 12 h forecast. This has been done for mid latitudes in the northern hemisphere during the Lindenberg campaign period as a first approach. The statistic errors of CIWC and CLWC are represented in Figure 7. Applying an error on the CIWC profile of Figure 3of a size compatible with these statistics, the brightness temperature difference between this result and the result obtained with the unmodified CIWC profile reached 9 K in the long wave window channel region (not shown). This study gives an order of magnitude of the potential impact on the simulated brightness temperature uncertainties with the a-priori representative errors of CIWC and CLWC in an operational context. Due to in situ measurements, cloud profile errors from Lindenberg should be lower resulting in a lower uncertainty on the brightness temperature.

Figure 7.

Covariance error of CIWC (full line) and CLWC (dash line) obtained from a statistic of ECMWF forecast in the middle North latitude during the Lindenberg campaign.

[36] The impact of the cloud water content uncertainty should decrease with the increase of cloud water content (or COT) due to the saturation of FRTMs. Figure 8 illustrates this phenomenon by showing the difference in brightness temperatures (simulated clear minus cloudy at 900.5 cm−1) function of IR COT and VIS COT at a given pressure level and for the 2 FRTMs. The black dotted line represents the opaque cloud at this pressure level. So, COT uncertainties do not have the same impact on the computed brightness temperature with a small COT value and large COT value.

Figure 8.

Clear minus cloudy BT at 900.5 cm-1, function of the IR COT for RTTOV in red and the Visible COT for HISCRTM in blue.

[37] Finally, unlike HISCRTM, RTTOV directly uses the cloud water content profile which is a major advantage in a NWP context. However, this vertical distribution is also an uncertainty source. To illustrate it, we just reversed the cloud water content vertical distribution of the cloud, the column integrated IR COT and all other parameters remaining unchanged. The impact on the departure of the simulated minus observed spectra is larger than 3 K for the long-wave window channels (Figure 9).

Figure 9.

Impact of cloud vertical distribution in RTTOV. (a) The two cloud profiles, located on the same layers with the same COT but two different shapes. (b) The resulting departure with the observation.

3.2.4. Toward an Operational Use

[38] We listed all cloudy input and their associated uncertainties on the computed brightness temperature for the 2 FRTMs in Table 1. The obtained uncertainties give an order of magnitude for a thin cloud (section 3.1.3) which is larger than an acceptable bias in data assimilation. In order to use these FRTMs in an operational context, we thus have to implement a strong and robust screening method in order to be confident enough with the cloudy inputs for the cloudy data to be assimilated, as explained in the next section. We also have to select the De value for each FRTM. Because we did not have information about it, we arbitrarily decided to rely on each IASI observation to choose this De. Regarding HISCRTM, De was adapted to describe the slope between 780 and 960 cm−1, in RTTOV we chose among the 4 parameterizations.

Table 1. Impact of Input Cloud Uncertainties on the Brightness Temperature
Cloudy Input UncertaintyΔTB RTTOVΔTB HISCRTM
Crystal shape5 K-
De20 K6 K
Cloud water content9 K9 K
Vertical distribution of cloud water content3 K-

[39] Finally, we consider, in our study, correct cloudy radiances when the cloud water profiles are accurate enough (see screening method section 4.2) with the De value which minimizes the residuals (obs – simulated).

4. Coregistered MetOp /A-Train Data Set

[40] This study with a set of Lindenberg profiles provided an insight in FRTMs performances and their sensitivity toward cloud profiles. The main objectives of this part are, first, to define a screening method based on IASI observations and input cloud variables only, in order to select the situations having input cloud information consistent with observations and second to statistically validate the two FRTMs using the cloud profiles inferred from the A-Train active sensors.

4.1. A-Train Dardar Cloud Profiles

[41] In this section, we used a data set [August et al., 2011], made in the context of the ConcordIasi campaign [Rabier et al., 2010], of collocated observations from the IASI and AVHRR instruments on MetOp, and from the Cloud Profiling Radar (CPR) and Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) on the A-Train satellites constellation. The spatial and temporal departures between the instruments are less than 14 km and 10 min. The CIWC and De profiles from the synergetic Dardar products between the CPR, CALIOP and the MODIS IR radiometer [Delanoë and Hogan, 2008, 2010] were added to the data set. The Dardar CIWC profiles are retrieved using a variational estimation of three physical parameters: extinction to backscatter ratio, visible extinction coefficient and the ice normalized number concentration parameter. The CALIOP products (cloud flag, COT, cloud reflectivity), the CPR products (cloud phase, COT, De and CIWC) and the Dardar profiles (CIWC, De) were obtained via the ICARE thematic center icare.univ-lille1.fr that provides the research community with various products related to the atmosphere. De defined in Dardar and CPR are different by their definition, Dardar uses theFoot [1988] parameterization and gives a De profile whereas CPR uses a Modis product and gives a mean value for the whole cloud. Hereafter, we refer to the Dardar De value.

[42] Of a total of 20896 co-locations from September 2010 to January 2011, 7931 (38%) of them were declared cloudy and overcast by the AVHRR cloud mask and Dardar. This large number of co-locations allows statistics between independent measurements, but given the orbit properties of MetOp and the A-train satellites the spatial and temporal coincidences are only found in high-latitude regions for latitudes poleward of 60. In this study, cloud inputs are derived from the A-Train, and other parameters (temperature, humidity, ozone profiles and surface parameters) are extracted from the French ARPEGE forecast. During the Concordiasi campaign, (F. Rabier, personal communication, 2012) a good accuracy of this optimized version of ARPEGE was found in comparison with observational measurements on the Antarctic continent.

4.2. Elaboration of a Screening Method

[43] Bearing in mind the high sensitivity of the scattering models to small errors in the input cloud information, this section proposes to define a method for discarding situations with too large inconsistencies between the cloud inputs and the observed IASI radiances. Indeed variational systems assume linear or near-linear processes which require that the a-priori is not too far away from the solution. In this particular study, many reasons may explain the inconsistencies. First, MetOp and A-Train instruments, which have different footprint sizes, do not exactly observe the same column of atmosphere and possibly not the same cloud. Furthermore, the frequency domains of the IASI, CPR and CALIOP instruments are different, leading to different interactions of the wavelength to the ice particles and very different capabilities to penetrate in the cloud. As the consequence the 3 sensors do not “see” the same cloud thickness and have different information. This may affect the representativeness of Dardar's CIWC in the IR domain.

[44] The definition of the screening method was driven by HISCRTM which makes use of less a-priori cloud information. The screening is built as follows. First, situations with a CTP from the CO2-slicing or AVHRR, found outside the CIWC profile, are removed. Then, the situations are classified as opaque or semitransparent depending on their VIS COT. As seen inFigure 8, the signal is completely absorbed in the cloud layer for integrated VIS COT values greater than 5 (also indicated in Winkler et al. [2003]) and the cloud is classified as opaque.

[45] 1. For the opaque clouds, a CTP is determined from the CIWC profile, which corresponds to the pressure level for which the integrated COT of the upper cloudy layers is equal to 5. The situations with a difference with the AVHRR CTP larger than 100 hPa (Figure 10, left) are discarded.

Figure 10.

Scatterplots of HISCRTM residuals. (left) Opaque clouds and (right) semi-transparent clouds. Straight lines symbolize the filters.

[46] 2. For the semi-transparent clouds, the filter is based on the CO2-slicing effective emissivity Nε. As all collocated situations are overcast, Nε is an estimation of the cloud emissivity. Figure 10 (right) presents the cloud emissivity as a function of the Dardar cloud water path. The color scale represents the difference of brightness temperature (DBT) between the HISCRTM simulated and observed spectra at 906 cm−1. This channel was chosen because of its small sensitivity to the De parameter (see Figure 5c). To select the situations with small HISCRTM DBTs, we applied a linear band-pass filter symbolized by the straight lines in the figure. Note that IWP + LWP are in a logarithmic scale to be consistent with the radiative transfer equation which links emissivity, absorption, reflectance or transmittance with the IWP and LWP as an exponential (through the COT calculation).

[47] As a summary, a screening method independent from FRTMs, has been defined to process opaque and semi-transparent clouds. It is based on observations (CTP, Nε) and the input cloud profile (IWP and LWP) only.

4.3. Results

[48] As mentioned in section 3.2.4, the De parameter has been chosen to minimize the residuals between simulated and observed IASI spectra. Regarding HISCRTM, De was adapted to describe the slope between 780 and 960 cm−1, in RTTOV we chose among the 4 parameterizations and the De Dardar profile.

[49] Using the screening process described previously, 70% of the co-locations were classified as semi-transparent. Then, out of the 2317 opaque clouds, 1562 (67%) were selected. For semi-transparent clouds, only 536 (10%) among the 5614 situations were selected, due to the deliberate strictness of the filter.Figure 11presents the spectral statistics of departures separately for semi-transparent and opaque clouds and the histograms of departure for channel 906 cm−1for both categories. The histograms are close to a Gaussian shape with zero bias, except for RTTOV for opaque clouds having a 2 K bias. This bias may be explained by the fact that the Dardar CIWC profile is less accurate when the lidar and radar are not used simultaneously (J. Delanoë, personal communication, 2012). This bias does not appear in the HISCRTM processing, because the cloud top pressure is probably accurate and the integrated value of cloud water content is larger than the saturation value. Nevertheless, the simulations are not very different from the observed radiances. In the case of semi-transparent clouds, the bias fluctuates around zero whatever the FRTM, with a standard deviation under 2 K, except for RTTOV which is between 750 and 900 cm−1. These results are fully acceptable to a further variational process and are mainly due to the severity of the semi-transparent cloud filter.

Figure 11.

A-Train data set: results for HISCRTM (green) and RTTOV (blue), for (a, c, d) semi-transparent and (b, e, f) opaque clouds. Figures 11c–11f are the histograms of departures for a channel nearly insensitive to De, and Figures 11a–11b are the statistics of departures for all channels. Full lines represent the standard deviations and dashed lines the biases.

[50] Figure 12shows the scatterplots of observed and RTTOV simulated radiances of semi-transparent clouds for the 5 parameterizations. The figure highlights the high sensitivity to De parameters. For example, the statistics differences between Boudal and McFarquhar parameterization show De differences of 10 μm and a bias larger than 2 K for the semi-transparent clouds (not shown in the figures).Table 2indicates the percentage of each parameterization used in this study. 75% of the time, the Dardar parameterization has been chosen for opaque clouds. Note that this parameterization is not an available choice in RTTOV. For semi-transparent cases, no parameterization dominates the choice.

Figure 12.

Scatterplots of simulated and observed BT with the four available RTTOV and Dardar parameterizations. The value of De is the mean De on the no saturating layers (COT < 5).

Table 2. Choice of Parameterization in RTTOV for the A-Train Data Set, Which Best Described the Observed IASI Spectrum (in Percent)
 OpaqueSemi-transparent
Ou & Liou4.93%15.30%
Wyser11.14%16.60%
Boudala2.82%20.71%
McFarquar6.02%19.22%
Dardar75.10%28.17%

5. Simulations of IASI Spectra Using ECMWF Cloud Profiles

[51] In the previous sections, it was shown that the use of FRTMs including cloud optical properties is useful for a better treatment of ice phase or extended cloud single layers. However, these studies focused on cloud profiles extracted from observational data. In this section, we have extended the study to global conditions during a 6-day period in February 2011, using ECMWF cloud variables as input of the FRTMs. The spatial and temporal departures between IASI and NWP forecasts are less than 0.5 degrees in latitude and longitude and less than 1h30. It is proposed to investigate the current capabilities in an operational context of both cloud-scattering models with an a-priori knowledge of the atmospheric cloud state vector, with a view to future exploitation within 1DVar and NWP assimilation.

[52] In this study, partly covered situations (AVHRR cloud cover in IASI fov <95%) and multilayer situations (those with 2 layers separated by more than 200 hPa) were first discarded. Then, the screening method described in section 4.2 was used to select the situations having NWP cloud profiles consistent with the observations. The parameterization among the four that minimizes the residuals between simulated and IASI observed spectrum was chosen in RTTOV and the De value which described the slope of the observation was selected in HISCRTM (section 3.2.4).

[53] Figure 13presents the statistics and the histograms of calculated-observed radiances for semi-transparent and opaque clouds separately. For opaque clouds, the biases fluctuates around zero whatever the FRTMs used, with smaller standard deviations for HISCRTM, probably due to a preliminary rejection of complex multilayer situations and to the use of the observed spectrum itself to define the cloud level. Note that the RTTOV bias observed in the A-Train data set (Figure 10) has been almost completely removed. For semi-transparent clouds, biases are of the same order of magnitude than with the A-Train data set (Figure 10), but the standard deviations are increased by about 1 K for both HISCRTM and RTTOV. Many factors may explain this relatively slight degradation. First, in the high-latitude A-Train data set, clouds are mostly in an ice phase whereas in this study, the three phase conditions (liquid, mixed and ice) are often observed and processed in the same situation. For these cases, the HISCRTM input phase is set to ice, which is not appropriate. Second, although inconsistent NWP cloud profiles with IASI observations were discarded, the diversity in the vertical distribution of clouds for this global data set is much more important than in the high-latitude A-Train data set.Table 3 summarizes the statistics in bias and standard deviation at 906 cm−1for the 5 latitude bands (NH polar, NH midlatitude, Tropics, SH midlatitude and SH polar). The number of co-registrations and the percentage of processed situations are also given. For both RTTOV and HISCRTM, the standard deviations increase when moving from polar to tropical latitudes, whatever the cloud type. In the Tropics, clouds are often more vertically extended than in other regions, increasing the possible errors in the cloud vertical distribution and leading to larger errors in the RTTOV calculated radiances (seeFigure 9). And for HISCRTM, the assumption of a single layer located at a defined pressure level is poor. In the Tropics, the screening efficiency should be improved in a later study, mainly for multilayer clouds.

Figure 13.

Same as Figure 11 but for the global data set over a 6 day period.

Table 3. Global Data Seta
 RTTOVHISCRTM
Semi-transparentOpaqueSemi-transparentOpaque
  • a

    Statistics of calculated-observed brightness temperatures at 906 cm−1 depending on the latitude. The numbers correspond to: number of situations/percent of selected situations/bias/standard deviation.

North polar6299/3.84/0.36/1.305395/17.17/−0.29/1.876299/3.84/−0.17/1.445395/17.17/0.14/1.09
North mid7766/2.60/0.34/2.259838/8.30/0.09/2.957766/2.60/0.11/2.429838/8.30/−0.02/1.44
Tropical4279/2.13/0.75/4.814246/5.28/−1.83/4.494279/2.13/0.49/3.424246/5.28/−0.88/2.66
South mid2939/5.13/0.84/3.4210765/17.99/−0.49/3.062939/5.13/−0.45/2.5210765/17.99/−0.37/1.61
South polar6237/3.40/0.04/1.388302/3.84/13.55/2.246237/3.84/−0.29/1.548302/3.40/0.02/1.12

[54] Figure 13 shows the maps of brightness temperature departures at 906 cm−1 for the selected and processed situations over the 6 days experiment, separately for the two FRTMs and the two cloud types. The maps on Figure 14 show the high geographical heterogeneity of the situations selected by the screening. Tropical latitudes are poorly described. A study conducted at ECMWF [Forbes and Tompkins, 2011] indicated an underestimation of ECMWF CIWC in the Tropics, which explains the poor simulations of some situations in this region and also reveals the good performance of the screening which removed most of these data. The maps also confirm the better simulation of IASI radiances with HISCRTM for opaque clouds and with RTTOV for semi-transparent clouds. Note that for RTTOV, a small subset gives large departures, above 10 K, mainly located in tropical areas for semi-transparent clouds (red points on the map) and for south hemisphere opaque situations (blue points). These situations, which are outside the histograms ofFigure 13, have a non-negligible impact on the statistics and it is obviously recommended not to assimilate them. To conclude withFigure 13, no real difference is observed between sea and land situations and no separate statistics are presented here.

Figure 14.

Maps of simulated-observed brightness temperatures at 906 cm-1 over a 6 day period. (a, c) HISCRTM and (b, d) RTTOV for semi-transparent (Figures 14a and 14b) and opaque (Figures 14c and 14d) clouds.

6. Conclusions and Outlook

[55] The two HISCRTM and RTTOV cloud forward radiative transfer models including cloud microphysical properties were assessed in this paper. In the first part of the article, through examples of collocated IASI observations and in situ cloud profiles from the Lindenberg ground-based campaign we investigated the behavior of both models. Then, extending the study to collocated A-Train and IASI/AVHRR measurements, we have shown the positive contribution of these scattering models to simulate cloudy spectra for ice crystals when cloud profiles are accurate enough. Thanks to the Dardar cloud products, a robust screening method was defined. This method is based on the IASI/AVHRR observations and the cloud profiles only, and is independent from the FRTMs. Standard deviations of departure of about 2 K for both RTTOV and HISCRTM were obtained in the CO2window channels after discarding inconsistent situations. Note that, to expand on a global experiment, it would be interesting to use Dardar profiles with AIRS hyper-spectral sounder.

[56] In the last part of the article, an operational context was simulated with the use of cloud variables profiles from a ECMWF forecast to compute the radiances during a one-week period over the globe. In this part, the screening method obtained previously was applied, and standard deviations of residuals lower than 3 K were obtained, whatever the type of clouds and region (slightly larger in the Tropics). These results are of the same order of magnitude than those obtained with the A-Train data set which shows the robustness of the screening method.

[57] This study also revealed the large impact of the ‘unknown’ De on the RTTOV simulated spectrum and the large impact of uncertainties of the cloud water content. The choice of the De profile or parameterization makes the use of the package difficult for users and in the current studies, simulations were made with the parameterization that minimized residuals between the simulated and observed spectra. This is a first step toward a more physical method for the De choice.

[58] This study showed the poor description of the cloud variables in the tropical regions with the rejection of a large amount of situations by the screening. A further study with collocated AIRS and active sensors of the A-Train together with the last HISCRTM version, which allows two cloud layers and two different cloud phases, should help to understand the problems of multi layers clouds in these regions.

[59] As mentioned in section 2, RTTOV is designed to be used in a NWP assimilation with cloud profiles as control variables of the variational process and to provide cloud information on each pressure or model level. This is of course not possible with HISCRTM which makes use of a small number of cloud column information. HISCRTM is more appropriate in a stand-alone inversion with possibly good retrievals of the integrated COT and mean cloud effective radius on the column, also useful information in the NWP context.

[60] This study is the first step toward the operational retrieval of the microphysical properties of clouds from IASI measurements. We will then continue that work, after selecting informative channels for the retrieval of the clouds variables, by the implementation of a first scheme using HISCRTM to initialize and constrain a 1dVar run using RTTOV with the De profile or a De mean value on the column as control variable. The Dardar profiles of the A-Train data set will be used for the validation.

Acknowledgments

[61] This work has been financially supported by Météo-France and CNES. The authors would like to gratefully thank Jun Li for providing the HISCRTM package with the IASI coefficients. We would like to thanks Icare center which provide us Dardar profiles. We thank our colleagues at CMS, Pascal Brunel for his help in implementing RTTOV and Jérôme Vidot for his useful discussions regarding the validation aspects. Finally, we would like to acknowledge Tony McNally and Thierry Phulpin for their comments which permitted considerable improvements in the definition of this study.