Atmospheric Infrared Sounder (AIRS; NASA Aqua platform) observations over land are interpreted in terms of monthly mean surface emissivity spectra at a resolution of 0.05 μm and skin temperature. For each AIRS observation, an estimation of the atmospheric temperature and water vapor profiles is first obtained through a proximity recognition within the thermodynamic initial guess retrieval (TIGR) climatological library of about 2300 representative clear-sky atmospheric situations. With this a priori information, all terms of the radiative transfer equation are calculated by using the Automatized Atmospheric Absorption Atlas (4A) line-by-line radiative transfer model. Then, surface temperature is evaluated by using a single AIRS channel (centered at 12.183 μm) chosen for its almost constant emissivity with respect to soil type. Emissivity is then calculated for a set of 40 atmospheric windows (transmittance greater than 0.5) distributed over the AIRS spectrum. The overall infrared emissivity spectrum at 0.05-μm resolution is finally derived from a combination of high-spectral-resolution laboratory measurements of various materials carefully selected within the Moderate-Resolution Imaging Spectroradiometer/University of California, Santa Barbara (MODIS/UCSB) and Advanced Spaceborne Thermal Emission and Reflection Radiometer/Jet Propulsion Laboratory (ASTER/JPL) emissivity libraries. It is shown from simulations that the accuracy of the method developed in this paper, the multispectral method (MSM), varies from about 3% around 4 μm to considerably less than 1% in the 10–12-μm spectral window. Three years of AIRS observations (from April 2003 to March 2006) between 30°S and 30°N have been processed and interpreted in terms of monthly mean surface skin temperature and emissivity spectra from 3.7 to 14.0 μm at a spatial resolution of 1° × 1°. AIRS retrievals are compared with the MODIS (also flying aboard the NASA/Aqua platform) monthly mean L3 products and with the University of Wisconsin Cooperative Institute for Meteorological Satellite Studies baseline-fit method (UW/CIMSS BF) global infrared land surface emissivity database.
Surface emission depends on surface parameters, that is, emissivity and temperature. Emissivity of land surfaces substantially varies with vegetation, soil moisture, composition, and roughness (Nerry et al. 1988; Salisbury and D’Aria 1992). Because emissivity depends on wavelength, it is referred to as spectral emissivity; it also depends on the viewing angle.
Continental surface emissivity in the thermal infrared window is a key parameter for estimating the surface radiation budget. Spectrally integrated surface emissivity and the energy emitted from the surface are proportional. A 10% error (e.g., from 0.9 to 1.0) on the emissivity approximately corresponds to a 10% error in the energy emitted from the surface (a portion of which may be compensated by the reflected incoming radiation) (Prabhakara and Dalu 1976; Ogawa et al. 2003).
Analyses of the sensitivity of simulated energy balance to changes in soil emissivity (Zhou et al. 2003) revealed that, on average, over northern Africa and the Arabian Peninsula a decrease of the surface emissivity in the atmospheric window by 0.1 would increase ground and surface air temperature by about 1.1° and 0.8°C, respectively, and decrease surface net and upward longwave radiation fluxes by about 6.6 and 8.1 W m−2, respectively. Also, a constant emissivity is often used for land surfaces in energy balance studies and general circulation models, because of limited information on the spectral and spatial distributions and time variations of the land surface emissivity (Ogawa et al. 2003).
It has also been shown that accounting properly for the surface emissivity in the solution of the radiative transfer equation inverse problem substantially improves the meteorological profiles (temperature, moisture) and cloud (Plokhenko and Menzel 2000) characteristics retrieved from infrared vertical sounders. Also, over continental surfaces, knowledge of the infrared emissivity spectrum allows one to correct observed brightness temperatures from surface emissivity effect, making possible an accurate determination of semitransparent clouds and aerosols properties.
Therefore, from both observational and modeling point of views, an accurate knowledge of surface emissivity and its spectral, spatial, and temporal variations, especially in the atmospheric thermal infrared window, is necessary. The method developed here aims at determining simultaneously the surface infrared emissivity spectrum from 3.7 to 14 μm at 0.05-μm spectral resolution and the surface temperature by inverting analytically the radiative transfer equation. The method is applied to 3 yr of Aqua Atmospheric Infrared Sounder (AIRS) observations (April 2003–March 2006) of the tropical zone (30°S–30°N).
The method follows four main steps: 1) An estimation of the atmospheric temperature and water vapor profiles is first obtained through a proximity recognition within the thermodynamic initial guess retrieval (TIGR) climatological library of about 2300 representative clear-sky atmospheric situations. With this a priori information, all terms of the radiative transfer equation are calculated by using the Automatized Atmospheric Absorption Atlas (4A) fast line-by-line radiative transfer model (Scott and Chedin 1981; http://ara.lmd.polytechnique.fr/). 2) Surface temperature is retrieved from AIRS data using a single AIRS channel (centered at 12.183 μm) chosen for its almost constant emissivity with respect to soil type. 3) Emissivity is then calculated for a set of 40 atmospheric windows (transmittance greater than 0.5). 4) The complete infrared emissivity spectrum at 0.05-μm resolution is then derived from a combination of the high-spectral-resolution laboratory spectra of selected materials [Moderate-Resolution Imaging Spectroradiometer/University of California, Santa Barbara (MODIS/UCSB) and Advanced Spaceborne Thermal Emission and Reflection Radiometer/Jet Propulsion Laboratory (ASTER/JPL) emissivity libraries] that are recognized as the closest to the set of 40 emissivity values obtained in step 3.
Eleven African regions have been analyzed, chosen for their good land-cover-type homogeneity and diversity (bare soil, sparsely vegetated, evergreen broad leaf forest, etc.). Comparisons are made with University of Wisconsin (UW) Cooperative Institute for Meteorological Satellite Studies/Space Science and Engineering Center (CIMSS/SSEC) baseline-fit method (BF) global infrared land surface emissivity database (Seemann et al. 2008) and MODIS Land Team products (Wan and Li 1997).
a. AIRS data
AIRS, on the National Aeronautics and Space Administration (NASA) polar platform Aqua, provides very-high-spectral-resolution measurements of radiation emitted by Earth’s surface and atmosphere in the spectral range 3.7–15.4 μm, at a spatial resolution of about 13 km (nadir). All channels belong to the three AIRS spectral bands (3.74–4.61, 6.20–8.22, and 8.80–15.4 μm). Each of its 2378 channels has a spectral resolution given by λ/Δλ = 1200, where λ is wavelength. A 324-channel subset has been distributed (Goldberg et al. 2003) and archived at Laboratoire de Météorologie Dynamique since April 2003 through an automatic process set up and run by the National Oceanic and Atmospheric Administration and sent from machines at Goddard Space Flight Center.
b. Infrared laboratory emissivity libraries: MODIS/UCSB and ASTER/JPL
MODIS/UCSB and ASTER/JPL emissivity libraries archive very-high-spectral-resolution thermal infrared laboratory measurements of the emissivity of different samples of typical Earth surfaces. These libraries are used to characterize the spectral features and range of variation of the emissivity of terrestrial materials. MODIS/UCSB and ASTER/JPL emissivity libraries have been downloaded from the Internet (http://www.icess.ucsb.edu/modis/EMIS/html/em.html and http://speclib.jpl.nasa.gov/, respectively).
In these two databases, the emissivity of a sample is usually determined by measuring its reflectance and converting it to emissivity by using Kirchhoff’s law, which holds for common terrestrial materials (Salisbury et al. 1994). We know that the mineral that is the most representative of continental surfaces is quartz, which also happens to display the strongest spectral features of any of the silicate minerals (Salisbury and D’Aria 1992). These so-called reststrahlen bands are observed between 8 and 10 μm and between 3 and 5 μm, where emissivity is shown to decrease with particle size (Hunt and Vincent 1968; Salisbury and Wald 1992; Moersch and Christensen 1995).
The signature in the 3–5-μm region strongly depends on the water and organic content of the soil. Green vegetation typically has a very high emissivity. Senescent (dry) vegetation has a more variable emissivity, especially in the 3–5-μm region, that depends on the type and structure of the cover type, soil water content, and so forth. Water, ice, and snow generally have a high emissivity, 0.94–0.99, in the thermal infrared region.
Figure 1 shows the emissivity mean value and associated standard deviation for each of the 324 AIRS channels calculated from soil, water, and vegetation samples of the MODIS/UCSB and ASTER/JPL emissivity libraries. As expected, quartz reststrahlen band features, centered at 4 and 9 μm, correspond to the lowest mean values (0.92 and 0.85, respectively) and display the highest standard deviation (0.2 and 0.12, respectively). This indicates the greatest sensitivity of emissivity at 4 and 9 μm to soil types as compared with, for example, the 11–15-μm band, which displays a much lower standard deviation of its emissivity (less than 0.025).
c. MODIS auxiliary dataset
Since the launching of the Aqua platform, the MODIS Land Team has been providing global maps of infrared land surface emissivity at six wavelengths, centered at 3.75, 3.96, 4.050, 8.55, 11.03, and 12.02 μm (corresponding to MODIS channels 20, 22, 23, 29, 31, and 32, respectively), and surface temperature at two different spatial resolutions (0.01° and 0.05°) and three different temporal resolutions (daily, weekly, and monthly), for both night (0130 LT) and day (1330 LT) observations (Wan and Li 1997). They also provide a clear/cloudy flag (Wan 2006) (see http://lpdaac.usgs.gov/modis/dataproducts.asp). As far as we know, MODIS Land is the only team distributing that kind of data.
For the purpose of this paper, only Aqua nighttime monthly averaged surface temperature and emissivity (MYD11C3 products) are used for comparison with AIRS-retrieved emissivities. Use is made of the MODIS cloud mask to determine clear-sky pixels for all applications and comparisons presented hereinafter. This choice ensures that potential discrepancies between AIRS-retrieved and MODIS-retrieved emissivities are not due to differences in the cloud-detection schemes.
d. UW/CIMSS/SSEC BF auxiliary dataset
The UW/CIMSS/SSEC team has developed the baseline-fit method to determine the emissivity at 10 wavelengths. Based on an analysis of MODIS/UCSB and ASTER/JPL libraries (section 2b), they extracted the most representative spectra (about 300) and identified 10 wavelengths (“hinge points”) situated within strategic regions of the emissivity spectrum. The MODIS emissivities at six wavelengths (section 2c) are used to estimate the emissivity at these 10 hinge points. This approach captures the most important features of the emissivity spectrum, that is, the signature of quartz reststrahlen bands.
The UW/CIMSS/SSEC BF dataset provides monthly averaged global infrared emissivity spectra from 3.7 to 14.3 μm at a 0.05° × 0.05° spatial resolution from January 2003 through November 2006 based on the analysis of MODIS data (section 2c) (Seemann et al. 2008) (see http://cimss.ssec.wisc.edu/iremis/). The fact that both MODIS and AIRS are aboard the Aqua platform makes the comparison between products derived from either MODIS or AIRS observations particularly coherent and easy.
3. The multispectral method (MSM)
At night, under clear-sky conditions and under the assumption of local thermodynamic equilibrium, the calculation of the monochromatic radiance I(λ, θ) emitted by the atmosphere and the surface at the wavelength λ in the direction θ is performed through the radiative transfer equation, which may be written as
The first and the second terms of Eq. (1) describe the upwelling radiance emitted by the surface of Earth and its atmosphere through the atmosphere. The third term corresponds to the downwelling radiance emitted by the atmosphere and reflected by the surface back to the satellite.
In Eq. (1), τ(λ, θ) is the (monochromatic) transmission function between the satellite and the current level, τs(λ, θ) is the transmission between the satellite and the surface, τ′(λ, θi) is the transmission between the surface and the current level under an incident angle θi, ɛs(λ, θ) represents the directional emissivity of the surface, and fr(λ, θi, θr) is the bidirectional reflectance distribution function (BRDF) of the surface (Nicodemus 1965). The BRDF gives the proportion of the incident radiation in the direction θi reflected in the direction θr. It depends on the wavelength and is determined by the structural and optical properties of the surface. Here Ts is the skin surface temperature and B(λ, T) is the usual Planck function; Ω+i indicates that the integration over the solid angle dωi is performed for the half-space above the surface.
Under the Lambertian assumption, which is the case for viewing angles lower than 30° and large fields of view (Chédin et al. 2004), the BRDF may be simply written as
Therefore, the radiative transfer equation may be written as
The accurate calculation of the downwelling term may require large amounts of computational resources. To speed up the radiative transfer calculations, it is customary to bypass the time-consuming angular integration of the third term of Eq. (3) by using a monoangular-equivalent model. Previous studies have shown that this term may be accurately enough computed from (Turner 2004)
The MSM developed in this paper aims at determining infrared emissivity spectrum from 3.7 to 14 μm at a 0.05-μm spectral resolution, using observations from hyperspectral infrared sensors. This is done by analytically inverting the radiative transfer equation [Eq. (3)] for atmospheric window channels carefully selected to ensure an error between 1% and 4% on the estimated emissivity, depending on wavelength and soil type.
Estimating the emissivity for a given wavelength requires 1) knowing the thermodynamic state of the atmosphere to calculate its contribution to the radiative flux and 2) estimating the surface skin temperature.
4. Determination of the thermodynamic state of the atmosphere: Proximity recognition in TIGR
The thermodynamic state of the atmosphere observed is determined through a proximity recognition within the TIGR climatological dataset (Chedin et al. 1985; Chevallier et al. 1998; Scott et al. 1999). This library of atmospheres consists of 2311 clear-sky situations described by their temperature, water vapor, and ozone profiles (40 levels from 1013 to 0.05 hPa). The ozone profile is specified from the U.K. Universities’ Global Atmospheric Modelling Programme (UGAMP) ozone climatological dataset (Li and Shine 1999), taking into account the latitude, longitude, and time of each situation archived in TIGR.
Clear-sky transmittances, radiances, and Jacobian functions for all AIRS sounding channels are precomputed for each situation in TIGR by the 4A model. These results are also stored within the TIGR dataset.
Last, the situations in TIGR are stratified by a hierarchical ascending classification into five airmass types (tropical; temperate, known as “midlat1”; cold temperate and summer polar, known as “midlat2”; Northern Hemisphere very cold polar, known as “polar1”; and winter polar, known as “polar2”), depending on their virtual temperature profiles (Achard 1991; Chevallier et al. 1998).
To characterize how far two atmospheric situations in TIGR are, we have defined the following distance:
where XOBS and XTIGR stand for vectors composed of channel brightness temperatures (or differences of brightness temperatures between two channels), constructed from observed (XOBS) and simulated (XTIGR) data, respectively; σXTIGR, the standard deviation vector of XTIGR over TIGR dataset, is used to normalize the distance.
The 11 components of XOBS and XTIGR are detailed in Table 1. Note that components 1–6 are the tropospheric sounding channels, sensitive to temperature and, to a lesser extent, water vapor. The last five components are differences between AIRS channels to constrain water vapor content and temperature gradients. Figure 2 shows normalized weighting functions of all AIRS channels occurring in the calculation of D. As shown in Fig. 1, the emissivity mean value calculated from soil, water, and vegetation samples of the MODIS/UCSB and ASTER/JPL emissivity libraries of all the channels in D is between 0.975 and 0.984, and the associated standard deviation is always around 0.013 (about 1%). This implies that all the channels in D have a low sensitivity to soil type or composition. AIRS channels in D are also essentially sounding channels with surface transmission lower than 0.15. For these two reasons, for all radiative transfer calculations, the emissivity of the channels in D can therefore be considered as temporally and spatially constant. Based on the MODIS/UCSB and ASTER/JPL libraries, this value has been set at 0.98.
For each observation, we first calculate D for all atmospheric situations in TIGR. Then, we select, among all the possible situations, those characterized by a distance D between the minimum distance calculated over the whole TIGR database (Dmin) and Dmax = min (1.4Dmin, 1.5Dmoy), where Dmoy stands for the average distance between any situation of TIGR and its closest in TIGR. This allows rejecting observed situations that either are too marginal or are still contaminated by undetected clouds or by aerosols and characterized by unrealistic (too high) Dmin values. Temperature and water vapor atmospheric profiles are obtained by averaging those profiles for which distance D falls within the interval [Dmin, Dmax].
This algorithm was applied to the atmospheric situations of the TIGR database. For each situation in TIGR, we look for the situations that fall within the circle of distance (its distance D falls within the interval [Dmin, Dmax]). Of course, the atmosphere considered is excluded as corresponding, by definition, to D = 0. The error induced by the proximity recognition scheme on the brightness temperature is then estimated by computing the difference between the average brightness temperature of the situations within the circle of distance and the initial one. For all AIRS window channels with a surface transmission greater than 0.5, this error is between 0.5 and 1.5 K. This result is consistent with accuracies that will be discussed later in section 6a.
5. Estimation of the surface temperature
a. The single semitransparent window channel method
For each AIRS observation, we now know the thermodynamic state of the corresponding atmospheric situation. By using the 4A model, we are thus able to evaluate the atmospheric contribution to the radiative flux of all terms of Eq. (3). For each AIRS observation, the remaining unknowns in the radiative transfer equation are then emissivity and surface temperature [see Eq. (5)].
As already seen, Fig. 1 shows that channels may be found that have the double property of being atmospheric windows and having an almost constant emissivity whatever the surface is. This is the case, for example, of AIRS channel 528 (133 in the subset of 324) at 12.183 μm, with a mean emissivity of 0.977 and a standard deviation of 0.012. As seen in section 4, its emissivity can be considered as spatially and temporally constant (0.98). Therefore, at 12.183 μm, the surface temperature remains the only unknown of the radiative transfer equation. The surface temperature is then evaluated by
with λ0 = 12.183 μm and ɛs(λ0) = 0.98.
b. Comparison with MODIS surface temperature
Figure 3 displays, for nighttime observations of the zone 30°S–30°N and at a 1° × 1° spatial resolution, the comparison between monthly averaged surface temperature estimated, as explained in section 5a, from AIRS channel 528 (Figs. 3a,d) and the surface temperature provided by MODIS (Figs. 3b,e) for June 2003 (left column) and December 2003 (right column). Also shown in this figure are histograms of the difference between these two surface temperatures (Figs. 3c,f), calculated for all grid points of the 1° × 1° maps. The mean and the standard deviation of these histograms are given in Table 2. Considering the difficulty of calculating the skin temperature for continental surfaces, the mean (0.25 and −0.28 K for June and December, respectively) and the standard deviations (1.76 and 2.06 K, respectively) are consistent with recently reported errors (Wan 2003; Wan et al. 2004; Bosilovich 2006; Zhang et al. 2007).
6. Determination of a complete infrared surface emissivity spectrum
a. Channel selection: The error amplification factor approach
To be selected for retrieving surface emissivity, a channel must obviously be sensitive to the surface, which implies a satellite-to-surface transmittance high enough (here, higher than 0.5), and not too sensitive to the variability of the atmospheric situation, or, more precisely, to small errors in its description (section 4). This latter sensitivity may be quantified theoretically through the error amplification factor (EAF).
The error amplification factor of a variable X (EAFX) on the emissivity ɛ is defined as
EAFX can be interpreted as follows: a 1% error on the estimation of parameter X induces an EAFX % error on the emissivity.
By deriving Eq. (5) with respect to surface temperature Ts and the brightness temperature Tb, we have
Except for EAFTs which is a surface-dependent variable, errors in the estimation of atmospheric thermodynamic profiles (section 4), transmission function, instrument noise, or imperfection of the cloud mask all have an impact on the brightness temperature and are therefore represented by EAFTb. This limits the number of variables that necessitate a systematic uncertainty analysis to two: Tb and Ts.
For typical tropical atmospheric and surface conditions, EAFTs and EAFTb have opposite signs and are in a range of 3–25 essentially depending on surface transmission and wavelength. In general, for channels with a transmission at the surface greater than 0.5, the highest values (about 6–8) are reached in the 4-μm region and the lowest values (about 3–5) are reached between 8 and 12 μm. Note that EAFs are systematically greater around 4 μm than between 8 and 12 μm. This implies that the accuracy of infrared emissivity determined from satellite observations is theoretically lower at 4 μm than in the 8–12-μm band independent of the inversion method. Now, considering each channel with a transmission at the surface greater than 0.5, we evaluate its maximum theoretical uncertainty on the retrieved emissivity [err(ɛs)] by calculating the EAFs defined by Eqs. (9) and (10). According to errors currently reported on Ts and instrument noise, the reference uncertainty for surface temperature (∂Ts/Ts) and brightness temperature (∂TB/TB) are fixed at 0.6% (≈2 K) and 0.5% (≤1.5 K), respectively. A channel is selected if the following condition is fulfilled:
From the subset of 324 channels, 40 AIRS channels have thus been selected. They are listed in Table 3. The uncertainty on the emissivity retrieved by the MSM through the simultaneous use of the 40 selected channels is expected to be much lower than the single channel err(ɛs) of Eq. (11). It is discussed below in section 6c. The 40 AIRS channels selected are distributed among the spectral bands displaying the strongest emissivity variability with respect to soil and vegetation types (see Fig. 1) and thus allow accurate reconstruction of the full infrared emissivity spectrum.
b. From the discrete set of 40 emissivity values to the full spectrum (3.7–14.0 μm) at 0.05-μm resolution
The method detailed above allows calculating emissivity for a subset of 40 AIRS channels. This obviously represents an important step forward in comparison with low-spectral-resolution sounders. However, accurate simulation of the upwelling thermal infrared energy flux emitted by the surface requires, besides the surface temperature, knowledge of the whole infrared emissivity spectrum.
To compute the full high spectral resolution of the surface emissivity, the MSM couples the 40 AIRS channel emissivities previously determined and a subset of high-resolution infrared laboratory measurements of the emissivity of different samples of typical Earth surfaces from MODIS/UCSB and ASTER/JPL libraries.
First, a subset of 165 emissivity spectra of various soils and vegetation types representative of Earth’s ecosystems was selected from MODIS/UCSB and ASTER/JPL libraries. These 165 high-spectral-resolution emissivity spectra are then interpolated linearly at the 0.05-μm resolution considered here, giving a total of 207 wavelengths between 3.7 and 14.0 μm. The resulting interpolated spectra dataset is hereinafter referred to as the 165-MOD-AST database.
For a given set i of 40 AIRS channel emissivities retrieved from observations as explained in section 6a, an “emissivity distance” d(i, j) to the jth emissivity spectrum of the 165-MOD-AST database is computed as
where ɛ(k, i) are the emissivities retrieved for the 40 selected AIRS channels and ɛdatabase(k, j) are the corresponding emissivities for the sample j in the 165-MOD-AST database.
Then, we select, among all the possible 165 spectra, those characterized by a distance d(i, j) between the minimum distance calculated over the whole 165-MOD-AST emissivity database (dmin) and dmax = 1.4dmin. A first-guess emissivity spectrum is finally obtained by averaging those spectra the distance d(i, j) of which falls within the interval [dmin, dmax].
For some surface types (typically Saharian bare soils) or minerals, the number of samples available in the 165-MOD-AST database is limited. Moreover, laboratory measurements of individual materials cannot reproduce the complexity spanned by the satellite that combines different ecosystems, soils, and so on. To overcome this difficulty, a fitting procedure has been developed that aims at minimizing the distance d(i, j) by refining the first-guess spectrum while maintaining its general shape. This is done by dividing the whole spectral interval into six different spectral bands (B1 = 3.7–5 μm, B2 = 5–8 μm, B3 = 8–8.6 μm, B4 = 8.6–9.5 μm, B5 = 9.5–10 μm, and B6 = 10–14 μm) and by computing, in each of these bands, the mean difference δk (k = 1, 6) between the 40 AIRS-retrieved emissivities and their corresponding emissivities for the first-guess spectrum obtained as explained above. The δks are then used to translate the first-guess spectrum into a more accurate one. Of course, between each spectral band the continuity of δ is insured. Note that band B2 is devoid of retrieved emissivities but plays a role in the continuity solution. In every case, the highest corrections are observed in B3 (6%–7%) and B4 (4%–5%). However, the mean correction over the spectrum is always less than 2%, which is consistent with accuracies discussed in section 6a.
In Fig. 4, an example of emissivity spectrum calculated by MSM is presented (red curve). We also show on the same plot emissivities calculated for the 40 AIRS channels (blue crosses), the first-guess emissivity spectrum (purple curve), and the six spectral bands, B1, B2, B3, B4, B5 and B6, defined above (vertical bars).
c. Expected accuracy on emissivity spectra retrieved by MSM
Theoretical simulations have been used to evaluate the accuracy of the MSM approach. A simulation consists in randomly selecting an atmospheric situation from the TIGR database (section 4) and an emissivity spectrum from the 165-MOD-AST emissivity database. For each atmospheric situation, a surface temperature is generated as the sum of the temperature of the atmosphere at the lowest level and a random number with zero mean and a standard deviation of 4 K. This insures that 99% of the skin temperatures chosen are between Tair_ground − 12 K and Tair_ground + 12 K. Corresponding AIRS channel brightness temperatures are then calculated using the 4A line-by-line model, and noise equivalent temperature (NEΔT) is added to the input brightness temperatures to account for the AIRS instrument noise. Last, MSM is applied. Five thousand such simulations were carried out.
Mean and standard deviation of the difference between the calculated emissivity spectrum and the true one for the 5000 cases processed provides an estimate of MSM expected accuracy. They are shown in Fig. 5 for nadir-viewing conditions and a surface pressure equal to 1013 hPa. Similar statistics were obtained for the retrieved surface temperature. The standard deviation, 0.75 K, is lower than recently reported errors (Wan 2003; Wan et al. 2004; Bosilovich 2006; Zhang et al. 2007). These results are representative of the accuracy that may be expected from the MSM. From Fig. 5 one can see that the emissivity uncertainty varies from 0.007 to 0.03 depending on wavelength. The highest uncertainties are observed in the atmospheric windows, 0.03 around 4 μm and 0.02 in the 8–10-μm band, where both emissivity variability with soil type (Fig. 1) and EAFs (section 6a) have an important impact on the retrieved emissivity spectrum. These accuracies are consistent with the maximum theoretical error on emissivity determined in section 6a. They are also consistent with the fact that in the 4-μm region EAFs are about 2 times those in the 8–10-μm region (section 6a).
7. Application of the MSM to AIRS observations
Three years of nighttime AIRS observations (from April 2003 to March 2006) between 30°S and 30°N have been processed and interpreted in terms of monthly mean surface skin temperature and emissivity spectra from 3.7 to 14.0 μm at a spectral resolution of 0.05 μm and a spatial resolution of 1° × 1°. To estimate the quality of the multispectral method, retrievals were compared with the MODIS/Aqua monthly mean L3 products and with the UW/CIMSS/SSEC BF global infrared land surface emissivity database.
a. Radiation model bias removal
Because the algorithms detailed in the preceding sections result from the consideration of radiative transfer model simulations, their application to real data implies that possible brightness temperature systematic biases between simulations and observations have been eliminated. This may be done by comparing simulations and observations for a set of collocated satellite and radiosonde data. Here, collocations are from the 40-yr European Centre for Medium-Range Weather Forecasts reanalysis radiosonde archive. For each channel, the systematic bias is obtained by averaging the difference between simulations from the 4A model and observations over the whole time period considered. Brightness temperature systematic biases are computed over sea. Daytime observations are not considered because of possible contamination by solar radiation of the short-wavelength channels. Only nighttime observations made by AIRS at around 0130 LT are considered.
b. Emissivity spectral variations: Comparison with MODIS and UW/CIMSS/SSEC BF
Emissivity spectral variations have been studied for 11 areas selected for their good land-cover-type homogeneity. These regions (see Fig. 6) are situated in Africa, which displays a latitudinal distribution of vegetation that allows one to select a variety of large enough and homogeneous scenes.
The MSM has been applied to the 11 regions selected for June 2003. Then, mean infrared emissivity spectra and corresponding standard deviations over each given geographical region have been calculated. Resulting spatially averaged infrared emissivity spectra (from 3.7 to 14.0 μm at a 0.05-μm resolution) are displayed in Fig. 7 (in red). On this figure are also plotted the spatially averaged emissivity spectra extracted from the UW/CIMSS/SSEC BF dataset for same month of June 2003. The six MODIS-channel spatially averaged emissivities are also shown (in black). For each selected zone, emissivity monthly spatial distribution variability (mean ± 2 standard deviations) is represented in purple for the MSM, in light blue for UW/CIMSS/SSEC BF data, and in black for MODIS data.
Spectral signatures of quartz reststrahlen bands dominate systematically the emissivity spectra of desert (Figs. 7a–c) and sparsely vegetated regions (Fig. 7d–k). This result is explained by the fact that quartz dominates the mineralogical composition of Earth soils. In these same quartz reststrahlen bands, emissivity increases with the proportion of vegetation. For example, the dense equatorial forest of zone 10 (Fig. 7j) masks the ground, and its signature cannot be seen from space. This explains why the retrieved emissivity remains close to 1 throughout the spectrum.
In general terms, the MSM provides results consistent with the UW/CIMSS/SSEC BF and MODIS products. This is, for example, the case for the two MODIS channels centered at 11.03 and 12.02 μm. Around 4 μm, differences of about 2%–5% are observed for zones 5 (Fig. 7e), 7 (Fig. 7g), and 11 (Fig. 7k). These differences are consistent with the retrieval accuracy expected in this particular spectral region (see section 6c). In the most important infrared quartz reststrahlen band (8.0–10 μm), large differences of about 10% are only observed for zone 3 (Fig. 7c) and, to a lesser extent, zone 5 (Fig. 7e). These zones are both situated in the Sahara desert. An explanation could be that Sahara desert samples are not well represented in MODIS/UCSB and ASTER/JPL emissivity libraries. Note also that in the 8–10-μm band only a single MODIS channel centered at 8.55 μm is available.
Figure 8 shows emissivity spectra retrieved by MSM for zone 5, characterized by a savanna vegetation cover type, for January 2004 (in red), April 2004 (in green), July 2004 (in blue), and October 2004 (in black). In the quartz reststrahlen band at 8–10 μm, emissivity is maximum in July (wet season) and minimum in January (dry season). In this band, the amplitude of emissivity seasonal variation is at a maximum at 8.3 μm and is on the order of 0.15. These results are consistent with those discussed in Chédin et al. (2004). For other spectral regions, emissivity seasonal variations are much smaller or are not important.
Results from the MSM approach have been archived for the period from April 2003 to March 2006 at the space–time resolution of 1° × 1° and 1 month. A monthly climatological dataset of infrared emissivity spectra (3.7–14 μm at 0.05-μm spectral resolution and 1° × 1° spatial resolution) for the tropical zone (30°S–30°N) will be soon made available on the Internet (http://ara.lmd.polytechnique.fr).
We have interpreted three years of Advanced Infrared Sounder data in terms of full infrared spectra of continental surface emissivity at a spectral resolution of 0.05 μm. From simulations, we have shown that the expected accuracy of the retrieved emissivity spectra, which depends of the wavelength and on the soil type, is on the order of 3% around 4 μm and is much less than 1% in the 10–12-μm spectral region.
The multispectral method developed in this paper relies on the physical processing of the radiative transfer equation and makes use of accurate a priori information on the state of the atmospheric situation observed as well as on very-high-spectral-resolution laboratory measurements of numerous earth surface samples carefully selected within the MODIS/UCSB and ASTER/JPL emissivity libraries.
The spectral variations of the emissivity have been studied for 11 zones selected for their good land-cover-type homogeneity. Signatures of quartz reststrahlen bands always dominate the emissivity spectrum for arid and semiarid regions. This result is explained by the fact that the quartz dominates the mineral composition of terrestrial materials. In these reststrahlen bands, emissivity increases with the proportion of vegetation. These results confirm those obtained from High-Resolution Infrared Radiation Sounder (HIRS) observations. In general, MSM results are consistent with MODIS and UW/CIMSS/SSEC BF data. However, by fully exploiting AIRS high spectral resolution, this method allows a complete coverage of the infrared emissivity at a high spectral resolution.
Infrared emissivity spectra estimation by the MSM is fast, simple, robust, and adaptable to all hyperspectral infrared sensors [here AIRS and soon the Infrared Atmospheric Sounder Interferometer (IASI) launched in October 2006 aboard the Meteorological Operation (MetOp) platform]. It can easily be implemented within the assimilation scheme of numerical models.
We are happy to thank R. Armante for help and fruitful discussions. Warm thanks are also given to the three anonymous referees for their particularly constructive and helpful comments and criticism.
Corresponding author address: E. Péquignot, Laboratoire de Météorologie Dynamique, Ecole Polytechnique, 91128 Palaiseau CEDEX, France. Email: firstname.lastname@example.org