The far-IR spectrum plays an important role in the earth’s radiation budget and remote sensing. The authors compare the near-global (80°S–80°N) outgoing clear-sky far-IR flux inferred from the collocated Atmospheric Infrared Sounder (AIRS) and Clouds and the Earth’s Radiant Energy System (CERES) observations in 2004 with the counterparts computed from reanalysis datasets subsampled along the same satellite trajectories. The three most recent reanalyses are examined: the ECMWF Interim Re-Analysis (ERA-Interim), NASA Modern-Era Retrospective Analysis for Research and Application (MERRA), and NOAA/NCEP Climate Forecast System Reanalysis (CFSR). Following a previous study by X. Huang et al., clear-sky spectral angular distribution models (ADMs) are developed for five of the CERES land surface scene types as well as for the extratropical oceans. The outgoing longwave radiation (OLR) directly estimated from the AIRS radiances using the authors’ algorithm agrees well with the OLR in the collocated CERES Single Satellite Footprint (SSF) dataset. The daytime difference is 0.96 ±2.02 W m−2, and the nighttime difference is 0.86 ±1.61 W m−2. To a large extent, the far-IR flux derived in this way agrees with those directly computed from three reanalyses. The near-global averaged differences between reanalyses and observations tend to be slightly positive (0.66%–1.15%) over 0–400 cm−1 and slightly negative (−0.89% to −0.44%) over 400–600 cm−1. For all three reanalyses, the spatial distributions of such differences show the largest discrepancies over the high-elevation areas during the daytime but not during the nighttime, suggesting discrepancies in the diurnal variation of such areas among different datasets. The composite differences with respect to temperature or precipitable water suggest large discrepancies for cold and humid scenes.
At the top of the atmosphere (TOA), the net incoming shortwave radiation is balanced by the outgoing longwave radiation (OLR) that escapes to space. Depending on the latitude, up to ~45% of OLR comes from the wavelength longer than 15 μm, normally referred to as the far-IR region (Harries et al. 2008). Water vapor pure rotational transitions dominate the gaseous absorption and emission in the far-IR region. Emission from the water vapor rotational band saturates at the upper and middle troposphere. Accordingly, this band dominates the clear-sky radiative cooling in the upper and middle troposphere (Clough and Iacono 1995; Mlynczak et al. 2005). The imaginary part of the refractive index of ice, which leads to absorption, has a minimum at 410 cm−1 (Warren and Brandt 2008). As a result, the far IR is also sensitive to the presence of ice clouds with the most prominent scattering effect around 410 cm−1. Turner et al. (2003) demonstrated the cloud phase detection by using both the far-IR and mid-IR measurements. Yang et al. (2003) conducted a feasibility study on using the far-IR measurements to determine the optical thickness of thin cirrus clouds. The optical properties of ice clouds for the far IR have been updated by Yang et al. (2005) and Baum et al. (2007). Using spectral radiances measured by Far-Infrared Spectroscopy of the Troposphere (FIRST) during the Radiative Heating in Underexplored Bands Campaigns II (RHUBC-II), Baugher et al. (2011, manuscript submitted to J. Geophys. Res.) investigated the effect of thin cirrus at the far IR. Rizzi and Maestri (2003) studied the impact of the cutoff wavelength of broadband radiometer on the estimated longwave cloud radiative forcing and pointed out the incorrect estimation of the contribution from the far IR. The absolute error of cloud forcing would be 2.7–5.3 W m−2 (relative error is 1%–6%) with the cutoff at 350 cm−1. The far IR plays an important role in the earth’s radiation budget and tropospheric adiabatic cooling and also has more potential in the remote sensing of our atmosphere.
Even though the importance of the far IR has been recognized for decades, no spectrally resolved measurements from space has ever been made for a large portion of the far-IR spectral region (<400 cm−1). This is mainly due to limitations of detector technology and other necessary instrumentation technology (Mlynczak et al. 2005) and is also due to the intrinsically low photon energy in this spectral region. During its 10 months of operation, the pioneering Infrared Interferometer Spectrometer-D (IRIS-D) instrument aboard the National Aeronautics and Space Administration (NASA) Nimbus IV spacecraft (Hanel et al. 1970, 1971) measured the infrared spectra between 400 and 1600 cm−1 with a spectral resolution of 2.8 cm−1 and a ground footprint ~100 km. About 700 000 calibrated spectra were obtained from IRIS-D and have been used in various studies, such as the detection of cirrus (Prabhakara et al. 1990, 1993), evaluation of climate models (Huang et al. 2002, 2006), and detection of spectral signatures of climate change (Harries et al. 2001). Another instrument aboard Nimbus IV, the satellite infrared spectrometer (SIRS; Wark et al. 1970), had three narrow bands in the spectral region less than 400 cm−1, each with a bandwidth of 5 cm−1. For the IR radiometers flying on National Oceanic and Atmospheric Administration (NOAA) polar-orbiting satellites and on geostationary satellites from different space agencies, no channels have covered the far IR while a channel at the mid-IR 6.7-μm water vapor band has existed in all such satellites. For recent space-borne hyperspectral observations such as Atmospheric Infrared Sounder (AIRS), Thermal Emission Spectrometer (TES), Infrared Atmospheric Sounding Interferometer (IASI), and Cross-Track Infrared Sounder (CrIS), the spectral coverage starts from ~667 cm−1; hence, they do not have observations of the far IR either.
In the absence of direct and comprehensive observations of the far-IR radiances at the top of atmosphere, estimations can be derived via different approaches. A straightforward approach is to use a radiative transfer model with inputs consisting of temperature, water vapor, and cloud profiles (as well as profiles of other trace gases) to compute the spectral radiances or fluxes in the far IR. The most recent reanalysis datasets, such as the European Center for Medium Range Weather Forecasting (ECMWF) Interim Re-Analysis (ERA-Interim; Dee et al. 2011), Modern-Era Retrospective Analysis for Research and Application (MERRA; Rienecker et al. 2011) from NASA, and Climate Forecast System Reanalysis (CFSR; Saha et al. 2010) from NOAA/National Centers for Environmental Prediction (NCEP), exhibit considerable improvements from the previous generations of reanalysis products, especially on the assimilation of satellite observations and on the baseline adjustments for observing systems over multiple years (WCRP Observation and Assimilation Panel 2011). It would be worthwhile to survey how the three reanalyses, ERA-Interim, MERRA, and CFSR, differ in terms of the far IR flux and how such differences can be attributed back to the discrepancies in temperature and moisture fields among the reanalyses. Alternatively, given the close correlations between the far-IR channels and the channels in the water vapor v2 band (i.e., the 6.7-μm band), it is also possible to directly estimate the far-IR radiances or flux from observed spectral radiances over the H2O v2 band and other mid-IR bands. For example, Huang et al. (2008, 2010) established algorithms to directly infer the spectral flux with 10 cm−1 resolution in the far IR based on the AIRS spectrally resolved radiances in the middle IR, for both the clear-sky and the cloudy-sky observations. The algorithms were developed with maximum consistency with the Clouds and the Earth’s Radiant Energy System (CERES) Single Satellite Footprint (SSF) algorithms and were then validated against the collocated CERES observations over the tropical open oceans. It is meaningful to extend these algorithms to the AIRS observations over land, so a near-global coverage of such spectral flux dataset can be achieved and compared with the counterparts derived from three reanalysis datasets.
This study is motivated by previous studies and a scientific motivation to understand this important spectral region. Section 2 describes the collocated AIRS and CERES datasets used in this study, the reanalysis datasets, and the methods employed to compute the far-IR spectral flux from the reanalyses. Section 3 explains how we extend the algorithm in Huang et al. (2008) to different land scene types, validate the algorithms, and then use such algorithms to derive the clear-sky spectral flux over the far IR. Section 4 presents the comparisons between the far-IR fluxes computed from different reanalyses and the counterparts derived from section 3. Conclusions and further discussions are then given in section 5.
2. Datasets and data processing
a. Collocated AIRS and CERES observations
The AIRS and CERES data used here are similar to those used in Huang et al. (2008). AIRS is a grating spectrometer with 2378 channels over three bands but only two mid-IR bands, the 6.20–8.22-μm and 8.8–15.4-μm bands, are used here. AIRS collects data in a cross-track scanning mode with a ground footprint of 13.5 km in diameter at nadir view (Aumann et al. 2003). AIRS level 1B calibrated radiances (version 5) in the entire year of 2004 are used here with quality controls (Huang and Yung 2005). CERES SSF edition 2 is used, in which the CERES OLR is derived from the unfiltered longwave broadband radiance and corresponding Angular Distribution Model (ADM; Loeb et al. 2005). The ADM was constructed for different discrete intervals of different scene types. The scene types are classified in terms of their surface properties (ocean, desert, etc.). A CERES field of view (FOV) is classified as clear sky if the coincident MODIS pixel-level cloud coverage within the CERES FOV is ≤0.1%. Cloud coverage of each MODIS pixel is decided by a series of threshold tests applied to several MODIS channels. More details of CERES cloud-clearing algorithm can be found in Minnis et al. (2011). For clear-sky observations, a particular scene type is further divided into discrete intervals according to the precipitable water, surface temperature, and lapse rate (ΔTs, defined as the vertical temperature difference between the surface and 300 hPa above). Detailed information for classifying the scene types and discrete intervals are also included in the CERES SSF datasets. For a collocated AIRS and CERES clear-sky observation, we assume that they belong to the same discrete interval of the same scene type. Two criteria are adopted for collocating AIRS and CERES data. First, the time interval between AIRS and CERES observation is less than 8 s; second, the distance between the centers of their footprints is within 3 km. By doing so, we then are able to infer the spectra fluxes over the entire longwave spectrum from the collocated AIRS spectra by the spectral ADM approach (Huang et al. 2008). The details of constructing spectral ADM and validating the algorithms are described in section 3.
b. Forward radiative transfer model: MODTRAN5
Moderate Transmission Code, version 5 (MODTRAN5), the latest version of MODTRAN, is based on newest spectroscopy compilation of HITRAN 2008. The most important improvement in MODTRAN5 is computing spectra at 0.1 cm−1 resolution. The comparison with line-by-line calculation shows its satisfactory performance and computing efficiency (Anderson et al. 2007). For developing the spectral ADM, the synthetic spectra computed by the MODTRAN5 are convoluted with the spectral response functions of individual AIRS channels (Strow et al. 2006). For computing the far-IR spectral flux from the reanalyses, the temperature and humidity profiles from each reanalysis are fed into MODTRAN5 and the TOA spectra fluxes are computed by a three-point Gaussian quadrature (Clough and Iacono 1995).
c. Reanalysis datasets
The latest reanalysis from ECMWF, the ERA-Interim, covers the period from 1989 to the present and is the third-generation reanalysis project (Dee et al. 2011; Simmons et al. 2007; Uppala et al. 2008) after the 15-yr ECMWF Re-Analysis (ERA-15; covering 1979–93; Gibson et al. 1999) and ERA-40 (covering 1957–2002; Uppala et al. 2005). Compared to the ERA-40, the ERA-Interim has a finer spatial resolution (1.5° latitude by 1.5° longitude) and more vertical levels (37 levels up to 1 hPa). The ERA-Interim employs a four-dimensional variational data assimilation (4DVAR) system and addresses various difficulties encountered with the ERA-40 project (Dee et al. 2011). It also has more effective and reliable assimilation of satellite data from different platforms over different periods. As a result, the ERA-Interim has achieved better qualities required for climate studies especially in the Southern Hemisphere than its precursors (Uppala et al. 2008; Dee et al. 2011).
NASA MERRA is an atmospheric reanalysis project covering the period from 1979 to the present. It is based on the Goddard Earth Observing System (GEOS-5) atmospheric general circulation model (AGCM) and the data assimilation system (DAS) (Rienecker et al. 2009, 2011). The detailed model description of GEOS-5 AGCM can be found in Rienecker et al. (2011). The GEOS-5 DAS employs a 3DVAR algorithm based on the gridpoint statistical interpolation scheme (GSI; Wu et al. 2002; Purser et al. 2003a,b). The MERRA data used in this analysis are on a horizontal grid of 0.5° latitude by ⅔° longitude with 42 vertical levels up to 0.1 hPa.
NOAA CFSR has been developed at the National Centers for Environmental Prediction (NCEP; Saha et al. 2010). Like the MERRA project, the CFSR also uses a 3DVAR technique. It covers 1979–2010 and the CFSR products are available on the 0.5° by 0.5° horizontal resolutions and 37 vertical levels up to 0.1 hPa.
In this study, data from 80°S to 80°N (near-global coverage or near-globe dataset) in 2004 from the three reanalyses are used. To facilitate the comparisons, the MERRA and CFSR reanalysis data are first regridded onto the same horizontal resolution of the ERA-Interim. Then the 6-hourly temperature and humidity outputs from these reanalyses are spatially subsampled and temporally interpolated onto the trajectories of collocated AIRS and CERES observations. Then, the subsampled profiles are fed into MODTRAN5. By doing so, we minimize the discrepancies due to temporal and spatial samplings among reanalyses and observations.
3. Inferring the far-IR flux from the collocated AIRS and CERES observations
a. Construction of the spectral ADMs
A key concept adopted for calculating flux from the radiance observation is the spectral anisotropic factor Rυ(θ), which is defined as
where Iυ(θ) is the upwelling radiance intensity at TOA for frequency υ and viewing zenith angle θ. The quantity Fυ is the corresponding upwelling flux. In the absence of scattering,
where subscript s denotes the surface and ɛvs denotes the surface spectral emissivity. Here, τ denotes the optical depth, B is the Planck function, z is altitude, and T(z) is air temperature at altitude z. The spectral ADM consists of a set of predetermined lookup tables of Rυ(θ) for each channel and each viewing zenith angle, so it can be used to invert the flux based on (1) for AIRS measured Iυ(θ).
In comparing with the spectral ADM in Huang et al. (2008), the largest difference here is the different surface spectral emissivity for different land surface types. Otherwise, the procedure for constructing the spectral ADM is identical to the one described in Huang et al. (2008). Specifically, the training datasets consist of ERA-Interim 6-hourly temperature and humidity profiles of January, April, July, and October 2005. Land surface type is determined from the 1-km resolution land coverage dataset from the U.S. Geological Survey (USGS; Loveland et al. 2000), which has 18 different surface types following the definition of the International Geosphere Biosphere Programme (IGBP) surface classification with one additional type of tundra. The dominant surface type within an ERA-Interim grid box (1.5° × 1.5°) is assumed to be the surface type of that grid box. Spectral emissivity of the 18 different surface types are obtained from the Advanced Spaceborne Thermal Emission Reflection Radiometer (ASTER) Spectral Library version 2.0 (Wilber et al. 1999; Baldridge et al. 2009). The library covers 2–16 μm with a spectral resolution of 4 cm−1. Figure 1 shows the spectral emissivity of such 18 surface types. The CERES SSF algorithm uses seven different surface types, and their relations with the 18 IGBP surface types can be found in Table 2 of Loeb et al. (2005) as well as in the caption of Fig. 1.
To avoid the complexity involved with spectral emissivity of snow and ice surfaces at different aging stages (Haggerty and Curry 2001), we chose to skip these surface types because their geographical distributions are mostly limited within polar and subpolar regions. Observations over other surface types in the high latitudes are still used in our analysis. Following the CERES SSF algorithm and Huang et al. (2008), the spectral ADM is built for each applicable discrete interval of each surface type. The discrete interval is defined based on the precipitable water pw, surface skin temperature Ts, and the temperature difference between the surface and 300 hPa above the surface (lapse rate). For each discrete interval of a given surface type, 800 profiles are randomly chosen from the training datasets (or all qualified profiles if the total number of qualified profiles for a discrete interval is less than 800). These profiles are then fed into MODTRAN5 to derive the Rυ(θ) at each AIRS channel. These profiles are also used to derive the principal components for each discrete interval so the flux over the spectral region not covered by the AIRS instrument can be estimated by a multivariate linear regression scheme (details can be found in Huang et al. 2008).
For AIRS channels with large opacity, the TOA radiances are not sensitive to different surface spectral emissivity due to the emissions from the surface being absorbed and reradiated within the atmosphere. The different surface type matters only for channels with sensitivity to the surface emission. Figure 2 shows examples of Rυ(θ) for different surface spectral emissivity at two AIRS window channels (933.041 and 1079.384 cm−1) for four discrete intervals, among which the only difference is the lapse rate while pw and Ts are identical. In general, Rυ(θ) increases as the surface spectral emissivity increases up to a threshold (e.g., ~0.98 for 933.041 cm−1 and ~0.95 for 1079.384 cm−1). Beyond the threshold for each, Rυ(θ) levels off and starts oscillating.
In practice, we find that, for scene types with large diurnal variations in temperature, such as savannas, dark desert, and bright desert, the performance would be improved if we construct the ADM separately for the AIRS daytime and nighttime observations (the new approach) instead of one ADM for all observations (the old approach). This is largely due to the fact that the Aqua satellite has an equatorial crossing time at 0130 and 1330 LT, close to both the peak and the trough of the diurnal cycle of surface temperature. The difference between these two approaches can be seen in Table 1, which shows the monthly-mean differences between OLR derived from AIRS radiances (OLRAIRS) and the collocated CERES OLR (OLRCERES). The new approach reduces the difference between OLRAIRS and OLRCERES for all scene types as well as for both daytime and nighttime observations with the exception of the nighttime dark desert scene (which changes from 0.04 to 0.4 W m−2, with both being small enough compared to the differences for other scene types). For observations over the oceans, we also find an improvement when the spectral ADMs are constructed for extratropical and tropical oceans separately (Table 1). As a result, in total we construct nine sets of spectral ADMs for forest, grassland, daytime savannas, nighttime savannas, daytime dark desert, nighttime dark desert, daytime bright desert, nighttime bright desert, and extratropical ocean. The spectral ADMs for the tropical ocean scenes are directly taken from those created by Huang et al. (2008). These 10 sets of spectral ADMs are what we used to derive the near-global spectral fluxes at a 10 cm−1 interval and the far-IR portions of such spectral fluxes will be used in the following sections.
Following Huang et al. (2008), the validation here is done in two steps. The first step is termed as the “theoretical validation.” For each discrete interval of a given surface type, we randomly choose 200 profiles from 6-hourly ERA-Interim data in 2004 (with no overlaps between the training datasets which come from ERA-Interim data in 2005), feed them to MODTRAN5 to compute the synthetic AIRS spectra as well as synthetic spectral fluxes, and finally compare the spectral fluxes estimated from the synthetic AIRS spectra (“predicted spectral fluxes”) with the directly calculated synthetic spectral fluxes (“directly computed spectral fluxes”). Figure 3 summarizes differences between the predicted and directly computed spectral fluxes for all discrete intervals recorded in the CERES SSF data. Generally the differences are within ±0.02 W m−2 per 10 cm−1 with only a few exceptions: the largest differences [~(0.03–0.06) W m−2 (10 cm−1)−1] are seen between 200 and 600 cm−1 for two discrete intervals that have a small lapse rate or even an inversion boundary layer and either a warm surface underneath a dry atmosphere or a cold surface underneath a humid atmosphere. Only a limited number of cases from our training datasets are found for such intervals. The largest negative difference found is about −0.03 W m−2 (10 cm−1)−1 around the ozone band for the discrete interval with an inversion layer and a warm surface underneath a dry atmosphere. Similar to the case in Huang et al. (2008), the difference is not sensitive to the choice of daytime and nighttime profiles, nor is it to the viewing zenith angles.
The second step in validation is to compare the OLR derived from the AIRS spectra with this algorithm (OLRAIRS) with the collocated CERES OLR (OLRCERES). We apply our algorithm to all applicable collocated AIRS and CERES clear-sky observations in 2004 between 80°S and 80°N. The OLRAIRS is simply the summation of the spectral flux derived from each AIRS spectrum. The differences between OLRAIRS and OLRCERES are then grouped and studied according to surface type. Table 2 shows the mean differences and corresponding standard deviations of OLRAIRS – OLRCERES difference for each surface type. The daytime-mean differences are between −0.71 and 1.67 W m−2 with standard deviations less than 3 W m−2. Statistics of the nighttime differences are similar to that of the daytime one. The histograms of OLRAIRS – OLRCERES differences based on all collocated observations are shown in Fig. 4. Similar to the statistics shown in Fig. 7 of Huang et al. (2008), the mean difference is less than 1 W m−2 and the standard deviation is less than 2 W m−2. Among ~2.48 million collocated clear-sky observations, 99.96% have OLRAIRS – OLRCERES differences within ±10 W m−2.
The geographical maps of OLRAIRS – OLRCERES for daytime and nighttime are shown in Figs. 5a,b, respectively. In general, the annual-mean differences averaged onto 2.5° × 2° grids are within ±2 W m−2. For the daytime, the largest differences (~±6 W m−2) exist in Tibet Plateau and some areas in the tropical East Africa. The nighttime differences over the land tend to be smaller than the daytime counterparts. The differences over the extratropical oceans tend to be modestly positive (~2 W m−2) for both daytime and nighttime observations.
The above comparisons and validations show that the performances of the algorithms and spectral ADMs developed for land surface types are comparable to those in Huang et al. (2008) for the tropical ocean surface. This gives us confidence to further examine the far-IR portion of the spectral fluxes derived by this approach in the following sections.
4. Comparisons with far-IR flux computed from reanalysis datasets
In this section, we compare the spectral flux derived in section 3 with their counterparts computed from the three reanalysis datasets described in section 2. As mentioned before, the 6-hourly reanalysis outputs are first subsampled to the same time and location as the observations used in section 3; then such subsampled profiles are fed into MODTRAN5 to compute spectral fluxes. Figures 6a,c show the temperature and humidity profiles averaged over all subsampled data from the ERA-Interim. Deviations of the MERRA and CFSR data from the ERA-Interim are shown in Figs. 6b,d. The most noticeable discrepancies are the large deviations of the CFSR temperature profile in the lower stratosphere from the ERA-Interim and MERRA.
Figure 7a shows the near-global averaged clear-sky spectral greenhouse parameter gΔυ computed from the collocated AIRS and CERES observations. The clear-sky spectral greenhouse parameter is defined as (Ackerman et al. 1992)
where the integral at the right side is the radiant energy emitted from the surface over the frequency interval Δυ. The quantity FΔυ is the flux at the TOA over the same frequency interval, and gΔυ indicates the fraction of radiant energy emitted from surface that has been trapped within the climate system. As expected, the gΔυ from AIRS shows large spectral greenhouse effects over the CO2 band, the O3 band, the H2O v2 band, and the H2O rotational band. Figure 7b shows the differences between gΔυ derived from the reanalysis and the gΔυ in Fig. 7a. The gΔυ from the ERA-Interim and MERRA generally tracks with each other (Fig. 7b), with the largest difference in the H2O v2 band. They both agree with the gΔυ of AIRS within ±2% and without any prominent spectral features. On the other hand, differences between CFSR and AIRS have more significant spectral features. The CFSR has the smallest gΔυ in the CO2 band, which is consistent with the large positive deviation of its lower-stratospheric temperature profile from the others (Fig. 6b). Higher stratospheric temperature in the CFSR correlates with more emission to space over the CO2 band; therefore, a smaller gΔυ can be expected. The CFSR has the largest gΔυ over the ozone band. For each reanalysis, gΔυ is computed based on its own ozone profiles. Note the situation of the ozone band is much more complicated than the CO2 band because the TOA flux of this band is sensitive to the stratospheric temperature, ozone absorption and emission, water vapor continuum in the lower troposphere, and the surface emission. Since the focus of this study is the far IR, hereafter we focus our comparisons only on three bands, 0–200 cm−1, 200–400 cm−1, and 400–600 cm−1. For the U.S. Standard Atmosphere, 1976 profile, the peaks of weighting functions for such three bands are around 10, 6, and 3 km, respectively.
Figure 8 shows the monthly-mean time series of near-global flux differences between those computed from the reanalyses and those derived from the collocated AIRS and CERES observations. For comparison, the flux difference over H2O v2 band (1400–1800 cm−1) is also plotted. For all comparisons, the month-to-month variation is small and the differences are between −0.8 and 0.5 W m−2. For all three reanalyses, the near-global mean fluxes over 0–200 cm−1 and over 200–400 cm−1 are larger than the observed ones, while the near-global mean fluxes over 400–600 cm−1 are smaller than the observed ones. For 0–200 cm−1, three reanalyses have similar difference from the observations. For 200–400 cm−1, the difference is largest for the MERRA data. Table 3 summarizes the differences in Fig. 8 together with the differences in OLR. While the OLR differences between reanalyses and observations are ~1% or even less, the percentage difference in a particular band can be considerably larger than 1%. For each reanalysis, the signs of differences in different bands can be same or opposite to the sign of the OLR difference. Hence, band-by-band differences here compensate each other when only OLR difference is examined. For all three reanalyses, the H2O v2 band has the largest percentage difference while the far-IR 400–600 cm−1 band tends to have the smallest percentage differences.
Figure 9 shows the spatial distributions of annual-mean differences between reanalyses and observations (averaged onto 2.5° × 2° grids). Since the differences over 0–200 cm−1 and 200–400 cm−1 have the same sign for all three reanalyses (Fig. 8), we combined the two bands together here for brevity. For the 0–400 cm−1 band, the flux differences between the reanalysis and the observation are positive (~2 W m−2 or less) over most parts of the land and over the tropical oceans. Over the extratropical oceans, the differences tend to be negative for all three reanalyses but the amplitude is small (~1 W m−2 or less). Such a contrast in signs between tropical and extratropical oceans cannot be seen in the flux difference map for the 400–600 cm−1 band (Fig. 9). Note the spectral ADMs for the tropical oceans were trained with the ERA-40 data over 4 months in 2001 and 2002 (Huang et al. 2008), while those for the extratropical oceans were trained with the ERA-Interim reanalysis data in 2005. Since 0–400 cm−1 is more sensitive to upper–middle-tropospheric humidity while 400–600 cm−1 is more sensitive to the lower-tropospheric humidity, such contrasts between the tropical and extratropical oceans are likely due to the changes of humidity assimilations in the ERA-Interim compared to the ERA-40, which affect the upper to middle troposphere more so than the lower troposphere.
For the 400–600 cm−1 band, the differences between reanalysis and observation are generally negative. Over land, the differences are generally smaller during the nighttime than the daytime. Moreover, the largest differences (−6 to −4 W m−2) happen during daytime in the Tibet Plateau and Andes Mountains, where the surface elevation is high enough and surface emission can reach the top of atmosphere (the peak of weighting function for 400–600 cm−1 band intercepts the surface of such high-elevation areas). For all three reanalyses, the differences during the nighttime over the Tibet Plateau and Andes Mountains are not as prominently noticeable as those during the daytime, suggesting a disagreement between reanalyses and observations on the diurnal cycle of such high-elevation areas with large complexities in topography. We speculate that, during the daytime, the complicated topography in such high-elevation areas leads to considerably inhomogeneous distribution of surface temperature even within one model grid (e.g., permanent shaded area versus sunshine area). As a result, it is difficult to correctly represent such large heterogeneity of surface temperature during the daytime in the data assimilation. This cloud lead to such large discrepancies between reanalyses and observation in the daytime over 400–600 cm−1 only (not in the nighttime, nor over the 0–400 cm−1 band which is not sensitive to surface of such high-elevation area). The spatial maps of the ERA-Interim and MERRA differences over 400–600 cm−1 are relatively more uniform than that of the CFSR. The tropical mean difference between the CFSR reanalysis and observation over 400–600 cm−1 is about 2 W m−2 smaller than the extratropical counterparts. Similar contrasts between the tropics and extratropics have been noted before for the surface temperatures in the CFSR (Wang et al. 2010).
Another way to disclose the differences among different datasets is to composite the flux differences with respect to surface temperature Ts and precipitable water pw, which are shown in Fig. 10 for Ts composites and Fig. 11 for pw composites. The standard deviations for the ERA-Interim composites are shown as shaded values (mean ±1σ) in both figures. The standard deviations of the composites from the other two reanalyses are similar to those from the ERA-Interim (not shown). In general, the composite differences from the three reanalyses are similar to each other. Especially over 0–200 cm−1 and 200–400 cm−1, all three composite differences normally increase with Ts as shown in Figs. 10a,b. They slightly increase with pw up to 5 cm and then decrease with pw until ultimately they become negative instead of positive (Figs. 11a,b). When two bands are combined (i.e., 0–400 cm−1), differences tend to be negative around −0.9 W m−2 when pw > 6 cm and to be positive around 0.6 W m−2 when the temperature is above 290 K. This suggests that the negative composite difference for the instance of pw > 6 cm is mostly due to the cold humid scenes instead of the tropical warm humid scenes. Note that the CFSR Ts is never higher than 320 K; therefore, no composite of CFSR difference is available for those bins above 320 K (Fig. 10). For the band of 400–600 cm−1, three composite differences with respect to temperature are similar to each other (Fig. 10c), except for very cold surface temperature (225 K ≤ Ts ≤ 235 K), where the composite MERRA difference is much more negative than the other two and outside the 1σ range from the composite ERA-Interim difference. The composite differences with respect to pw (Fig. 11c) show similar behaviors for the MERRA and ERA-Interim composites while the CFSR composite difference tends to be more negative when pw > 6 cm (but still within the 1σ range from the composite means of the other two reanalyses).
The spatial maps of the differences and the composites of differences both show that the three reanalyses generally agree with each other. Agreements between the ERA-Interim and MERRA are usually better than those between ERA-Interim and CFSR. All three reanalyses show a noticeably negative difference from the observations over 400–600 cm−1 in the Tibet Plateau and Andes Mountains over the daytime but not over the nighttime, pointing to discrepancies in diurnal cycle of surface temperature in such high-elevation region. For the 0–400 cm−1 region that we have had no direct spectral observations from yet, the composites indicate that the largest discrepancies between the AIRS-derived fluxes and the fluxes computed from reanalyses are over those cold but humid scenes. Cold scenes have stronger emission in the 0–400 cm−1 region than warm scenes, and humid scenes have stronger water vapor absorption in the same spectral region than dry scenes. These probably explain the reasons why largest discrepancies (in absolute flux) are seen in such scenes.
5. Conclusions and discussion
We here extend the work of Huang et al. (2008) to AIRS observations over the land and over the extratropical oceans. The clear-sky OLR derived by this approach agrees well with the collocated CERES clear-sky OLR, with the daytime difference being 0.96 ±2.02 W m−2 and the nighttime difference being 0.86 ±1.61 W m−2 for the entire year of 2004 (Fig. 4). Since there have been no direct observations of far-IR spectrum from the space, we further compare the far-IR fluxes derived from the collocated AIRS and CERES observations with the counterparts computed from three most recent reanalyses. Note the radiances from a subset of the AIRS channels have been assimilated into all reanalyses analyzed here. The comparison shows that, in terms of the near-global monthly-mean averages, fluxes over the band 0–400 cm−1 based on reanalyses are all slightly larger than those derived from the AIRS observations by ~(0.31–0.54) W m−2 (equivalently 0.66%–1.15% in percentage). Over the band 400–600 cm−1, the differences between reanalyses and observations are negative, ranging from −0.6 to −0.3 W m−2 (−0.89% to −0.44% in percentage). Given that the two bands are sensitive to emissions from different parts of the troposphere, such differences likely suggest that the reanalysis–observation discrepancies (temperature, humidity, or both) are in opposite signs for the upper–middle and lower troposphere.
From the comparisons of near-global averages, near-global spectral greenhouse parameters, and the annual-mean spatial map and composites with respect to surface temperature and total column of water vapor, it can be seen that the results from the three reanalyses are generally tracking with each other, especially between the ERA-Interim and MERRA. The differences over the 0–400 cm−1 band between reanalyses and the observations tend to increase with surface temperature but not with column water vapor when the column water vapor is larger than 4 cm. Both reanalyses and the spectral fluxes inferred from the AIRS data have their own limitations and uncertainties. For reanalyses, the uncertainties mainly arise from the quality of assimilated temperature and humidity profiles. For spectral fluxes inferred from the AIRS data, the uncertainties mainly come from the uncertainty of far-IR and mid-IR spectroscopy (including water vapor continuum model) used in the algorithm. The uncertainty in IR gaseous spectroscopy is usually no more than a few percent. State-of-the-art retrieval of humidity profiles from the mid-IR radiances has ~(16%–25%) uncertainty (Chahine et al. 2006). If we assume comparable uncertainty in the assimilation of IR radiances in the mid-IR H2O bands, it would suggest that the uncertainty associated with the assimilated humidity profiles dominates over the uncertainty with spectroscopy in our comparison results.
Figures 10 and 11 indicate that the composite differences of the water vapor v2 band do not resemble those of the far-IR rotational band. Figures 7 and 8 also show the contrast between the far-IR and mid-IR water vapor bands, even though the information content of two is highly correlated. Note that these differences are shown for clear-sky cases only. Once cloud (especially ice cloud) is taken into account, it would be even more challengeable to infer the far-IR flux based on mid-IR measurements because of the spectral dependences of many cloud optical properties. Such spectral dependences are eventually determined by the cloud microphysical properties, which is hard to uniquely inferred from the mid-IR passive remote sensing alone. Therefore, to further and better understand the far-IR portion of the spectrum, well-calibrated global far-IR measurements are still indispensable. Such measurements are not replaceable by other measurements, especially in the presence of clouds. Such measurements will complete our understandings to the longwave radiation budget and the role of far IR in climate, as well as help us to compare and understand results from such state-of-the-art data assimilation efforts especially the applicability of such reanalyses in the climate-oriented radiation budget studies. Preliminary studies and technological development have been directed toward space-borne far-IR spectrally resolved measurements, such as Radiation Explorer in the Far-Infrared (REFIR; Esposito et al. 2007), the Far-Infrared Spectroscopy of the Troposphere (FIRST; Mlynczak et al. 2006), and the Tropospheric Airborne Fourier Transform Spectrometer (TAFTS; Cox et al. 2007). The Climate Absolute Radiance and Reflectivity Observatory (CLARREO) mission, which primarily aims at rigorously observing climate change over the decadal time scale and using such rigorous and thorough observations as the most critical metric to determine the accuracy of climate change projections, can also provide spectral measurements of the far IR with high accuracy and calibration traceable to the SI standards for the kelvin and the watt. A collateral benefit of CLARREO would be the far-IR observations from space that can advance our understandings to various issues related to the far IR.
This work shows the feasibility of extending the algorithm established in Huang et al. (2008) to the global scale for the clear-sky observations. Similarly, the algorithm in Huang et al. (2010) can be extended to global cloudy observations as well, which is the theme of our ongoing follow-up study. The number of discrete intervals in CERES cloudy ADM is order of magnitude more than that of CERES clear-sky ADM. How to efficiently develop and validate spectral ADM for such large number of discrete intervals for cloudy scenes and different surface types is the central topic in our follow-up study.
We thank two anonymous reviewers for their suggestions that improve the clarity of this paper. The AIRS data were obtained from NASA GSFC DAAC, and the CERES data were obtained from NASA Langley DAAC. The ECMWF ERA-Interim data (http://data-portal.ecmef.int/data/d/), NASA MERRA reanalysis data (http://disc.sci.gsfc.nasa.gov/daac-bin/FTPSubset.pl), and NOAA/CFSR data (http://dss.ucar.edu/datasets/ds093.1/) are obtained online. The first author is thankful to Ms. He and Dr. Song for their help in data processing. This work was supported by NASA Grants NNX11AH55G and NNX11AE68G awarded to the University of Michigan.