The precision, accuracy, and potential sampling biases of temperature T and water vapor q vertical profiles obtained by satellite infrared sounding instruments are highly cloud-state dependent and poorly quantified. The authors describe progress toward a comprehensive T and q climatology derived from the Atmospheric Infrared Sounder (AIRS) suite that is a function of cloud state based on collocated CloudSat observations. The AIRS sampling rates, biases, and center root-mean-square differences (CRMSD) are determined through comparisons of pixel-scale collocated ECMWF model analysis data. The results show that AIRS provides a realistic representation of most meteorological regimes in most geographical regions, including those dominated by high thin cirrus and shallow boundary layer clouds. The mean AIRS observational biases relative to the ECMWF analysis between the surface and 200 hPa are within ±1 K in T and from −1 to +0.5 g kg−1 in q. Biases because of cloud-state-dependent sampling dominate the total biases in the AIRS data and are largest in the presence of deep convective (DC) and nimbostratus (Ns) clouds. Systematic cold and dry biases are found throughout the free troposphere for DC and Ns. Somewhat larger biases are found over land and in the midlatitudes than over the oceans and in the tropics, respectively. Tropical and oceanic regions generally have a smaller CRMSD than the midlatitudes and over land, suggesting agreement of T and q variability between AIRS and ECMWF in these regions. The magnitude of CRMSD is also strongly dependent on cloud type.
Temperature, water vapor, and clouds are central to processes strongly affecting the global circulation and energy budget. Accurate measurements of these variables are essential in monitoring climate change (Hansen et al. 2012), evaluating climate model simulations (Ma et al. 2012; Li et al. 2012; Jiang et al. 2012), and quantifying changes in the global circulation, energy and water cycles (Trenberth et al. 2005; Sherwood et al. 2010; Schneider et al. 2010; Wong et al. 2011).
The variety of approaches regarding cloud classification is helpful for sorting out complex processes relating temperature, water vapor, and clouds (Jakob and Tselioudis 2003; Jakob et al. 2005; Xu et al. 2005, 2007; Zelinka et al. 2013). Cloud classification has also been used to identify weaknesses in the representation of clouds and convection in general circulation models (GCMs) (Zhang et al. 2005; Williams and Tselioudis 2007; Chen and Del Genio 2009; Del Genio 2012). For example, Jakob and Tselioudis (2003) and Jakob et al. (2005) combine the cloud retrieval from the International Satellite Cloud Climatology Project (ISCCP; Rossow and Schiffer 1999) and atmospheric state profiles measured by the U.S. Department of Energy's (DOE) Atmospheric Radiation Measurement (ARM) Program (Ackerman and Stokes 2003) to relate cloud and atmospheric thermodynamic properties under different cloud regimes. They argue that exposing model shortcomings for different cloud regimes is an important step in improving GCMs by providing a crucial link between a given model's climate and cloud representation.
As previous work has shown, a multiyear, satellite-based climate data record of key parameters summarized by cloud types can serve as a powerful model constraint. However, satellite observations are subject to uncertainties that arise from instrument noise; temporal, spatial, and vertical resolution; the retrieval methodology; spectral coverage; and especially sampling characteristics (e.g., Schmit et al. 2009; Wang et al. 2007). It is crucial to separate variability observed in nature from these uncertainties. This is particularly important when applying satellite data to derive trends and long-term variability, since contributions of uncertainty can be as significant as those from the mean state retrieval (Keihm et al. 2009). Yue et al. (2011) showed that uncertainty in the lower tropospheric stability derived from the Atmospheric Infrared Sounder (AIRS) is dominated by extensive and uniform stratus. Guan et al. (2013) also found that satellite orbital sampling biases associated with the diurnal cycle and instrument swath width can impact the cloud water and upper-level humidity field climatology, especially when satellites are gridded at high spatial resolution.
The cloud-state-dependent sampling affects all satellite systems and is a strong function of clouds and precipitation. Clouds are not transparent in the infrared because of the strong absorption of condensed water (Liou 2002); therefore, cloud effects on infrared-based satellite instruments must be quantified before representative water vapor and temperature climatologies are reliably determined from infrared data. This concern has been raised in a number of climate processes studies using data from infrared satellite sensors. Wu et al. (1993) suggested that a dry bias is likely a result of cloud-induced sampling in the Television Infrared Observational satellite (TIROS) Operational Vertical Sounder (TOVS) based on broad spectral observations. Chaboureau et al. (1998) noted that TOVS produces higher total precipitable water vapor (TPWV) in the maritime stratocumulus cloud regions. Lanzante and Gahrs (2000) report a dry bias of 5% ~ 10% in upper-troposphere relative humidity from TOVS. Sohn et al. (2006) and John et al. (2011) both estimate the dry bias to be 20%–30%. Fetzer et al. (2006) show that the AIRS retrieved TPWV over the ocean is biased dry by 5%–10% within the convective areas in the deep tropics but is biased wet by about 15% in the subtropical stratocumulus decks relative to the Advanced Microwave Scanning Radiometer for Earth Observing System (EOS) (AMSR-E), carried on the National Aeronautics and Space Administration's (NASA) Aqua satellite with AIRS. While these differences are ascribed mainly to cloud-induced sampling effect on AIRS, this particular study did not take into account potential biases from AMSR-E. Even in the microwave, a frequency commonly assumed to be unaffected by cloud and used as a benchmark to access the infrared observational errors, the uncertainty in upper-tropospheric relative humidity is estimated to be up to 10% because of clouds, especially in areas with deep convection (Buehler et al. 2008). The numerous findings regarding biases in satellite-observed TPWV or upper-tropospheric water vapor have important ramifications on the long-term monitoring of water vapor vertical distributions observed by satellites. A regional study over East Asia (Kwon et al. 2012) compared temperature and water vapor retrievals from the Infrared Atmospheric Sounding Interferometer (IASI; Hilton et al. 2012) to radiosonde observations and sorted the differences according to the cloud top fraction and TPWV. However, TPWV itself is subject to biases related to cloud states, and cloud top fraction cannot solely represent the variety of cloud classes. To our knowledge, no study has examined in detail how water vapor vertical profile retrievals obtained from satellites are biased for a general set of cloud classes. Establishing the relationship between cloud classes and sampling error is a necessary step for robust applications of satellite data in climate research.
Cloud classification methods from satellite observations have matured over the years. Xu et al. (2005) identify four cloud regimes using clustering analysis by combining cloud products from the Visible and Infrared Scanner (VIRS) and broadband radiation measurements from the Clouds and the Earth's Radiant Energy System (CERES) on the Tropical Rainfall Measuring Mission (TRMM) satellite. The most widely used cloud classification method is the cloud top pressure and cloud optical thickness categorization of the ISCCP cloud products (Rossow and Schiffer 1991, 1999). Similar methods have been applied to the Moderate Resolution Imaging Spectroradiometer (MODIS) data (e.g., Li et al. 2007; Marchand et al. 2010). In addition, radiance spectral characteristics have been used to identify different cloud types over various surface conditions based on the cloud and surface spectral radiative properties from MODIS (Li et al. 2003), the Advanced Very High Resolution Radiometer (AVHRR; Pavolonis et al. 2005), and the Visible Infrared Imager Radiometer Suite (Hutchison et al. 2005).
The NASA A-Train satellite constellation includes several active and passive sensors that are coordinated to allow for nearly simultaneous, collocated observations of three-dimensional atmospheric thermodynamics, clouds, aerosols, radiation, and other parameters important in the understanding of climate change and weather prediction (L'Ecuyer and Jiang 2010). With a record including over 10 years of nearly continuous observations, the A-Train dataset is an invaluable data source for climate studies. The A-Train temperature and water vapor profiles are retrieved from the infrared and microwave radiances measured by the AIRS instrument and the Advanced Microwave Sounding Unit (AMSU) on board the Aqua satellite, which are subject to significant cloud impacts (Tobin et al. 2006; Susskind et al. 2006; Fetzer et al. 2006). Detailed cloud classification data are readily available from the CloudSat satellite based on a combination of radar reflectivity profiles and atmospheric conditions (Wang and Sassen 2001). Sassen and Wang (2008) showed that the CloudSat cloud class climatology largely agrees with that from ISCCP.
In this study, we use the CloudSat cloud classification to quantify the cloud-state-dependent sampling effects in AIRS thermodynamic profiles that are retrieved from the combination of infrared and microwave radiances from the AIRS instrument and AMSU. The European Centre for Medium-Range Weather Forecasts (ECMWF) model analysis is also collocated with the AIRS/CloudSat data and categorized by the CloudSat cloud classes. It serves as a benchmark for this study, although the ECMWF model analysis is itself subject to various uncertainties, especially over areas with sparse observations. AIRS radiances have been assimilated in the ECMWF model analyses and show consistent improvements in the quality of the analyses and forecast skill (McNally et al. 2006). However, only a small fraction of clear AIRS spectra is used, and no tropospheric channels are used over land. It is not apparent how the assimilation of clear AIRS radiances will affect the representation of atmospheric vertical structure, especially in the lower troposphere and in cloudy conditions (McNally et al. 2006; Reale et al. 2008). Therefore, the ECMWF and AIRS retrievals are treated independent of each other. The pixel-scale collocation between ECMWF and the satellite field of view (FOV) per pixel is important because it removes biases that arise from satellite orbital sampling in the presence of a diurnal cycle and the inherent “smoothing out” of pixel-scale variability in a gridded dataset. Moreover, quality flags in the AIRS data are used to subsample the collocated ECMWF model analysis to provide exactly the same samples from AIRS and ECMWF.
Therefore, three datasets that are categorized by CloudSat cloud classes are included in this study: 1) AIRS retrievals, 2) collocated ECMWF data subsampled to the AIRS quality control (EC-S), and 3) collocated ECMWF data not subject to AIRS quality control (EC). The AIRS and EC-S datasets have the same sample sizes taken from the same times and locations; thus, any difference between them suggests either an observational bias in the AIRS retrieval, a model bias in the ECMWF data, or some combination of both. Differences in the EC-S and EC datasets are the result of subsampling by the AIRS quality control and are due to sampling effects that are a function of cloud state within the AIRS FOV.
This article is organized as follows. Section 2 describes the data sources and the collocation method and an approach to summarize the collocated AIRS FOVs with much finer resolution CloudSat cloud classification data. Section 3 shows the global and regional characteristics of the vertical profiles of AIRS temperature and water vapor within various cloud classes. Section 4 quantifies the observational and cloud-state-dependent sampling biases in the AIRS data by comparing with EC and EC-S versions of the ECMWF analyses. The centered root-mean-square differences (CRMSD) of AIRS and ECMWF are also presented. A summary and set of conclusions are presented in section 5.
2. Data and method
a. Data sources
We use the standard level 2 (L2) products of “A Multi-Sensor Water Vapor Climate Data Record Using Cloud Classification” program under NASA's Making Earth System Data Records for Use in Research Environment (MEASURES) project (Fetzer 2012). Under this program, observations and retrievals from multiple instruments of the A-Train constellation have been collocated, including AIRS, CloudSat, the Cloud–Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO), and Microwave Limb Sounder (MLS). This study is limited to the collocated observations of AIRS and CloudSat. The collocation method is described in section 2b. The matched product consists of AIRS level 1B radiances and L2 retrieved variables and CloudSat profiles collocated within the AIRS instrument FOV (13.5 km, nominal, at nadir) for radiance and the AIRS/AMSU FOV (45 km, nominal, at nadir) for retrievals. An index product is created simultaneously that consists of the time, footprint index, and granule number of AIRS and AMSU FOV, and the index array of collocated CloudSat profiles. Vertical profiles of cloud, temperature T, and water vapor mixing ratio q are taken from the MEASURES L2 dataset.
The collocated ECMWF model analysis data are obtained from the CloudSat ECMWF auxiliary (ECMWF-AUX) data product, which consists of the T, q, and surface temperature fields from the ECMWF analysis that has been interpolated to CloudSat cloud profiling radar bins. Based on the index data product, the collocated ECMWF model analysis data are collected within each AIRS/AMSU FOV.
1) Atmospheric Infrared Sounder
AIRS is a suite of microwave and infrared instruments on the Aqua satellite launched on 4 May 2002 into a sun-synchronous orbit with an equator crossing time of 1330 LT (0130 LT) on the ascending (descending ) node. It includes the AIRS instrument, which is a hyperspectral infrared grating spectrometer with 2378 channels between 3.7 and 15.4 μm, and the AMSU-A, with 15 microwave channels between 23 and 90 GHz (Aumann et al. 2003). The nadir spatial resolution for the AIRS instrument is 13.5 km, and there are 3 × 3 AIRS footprints coregistered to a single AMSU footprint (Lambrigtsen and Lee 2003) with an FOV of 45 km. The AMSU FOV is equivalent to the AIRS/AMSU FOV, and the horizontal resolution of retrieved quantities. The combination of infrared and microwave radiances facilitates the retrieval of vertically resolved T and q profiles in cloudy conditions up to an infrared effective cloud fraction (ECF) from approximately 0.7 (Susskind et al. 2006) to 0.9 (Yue et al. 2011), depending on the cloud regime. The yield of the highest-quality retrievals decreases rapidly at higher ECF (Fetzer et al. 2006; Yue et al. 2011). The AIRS/AMSU combined retrieval product is referred to as AIRS data in the rest of the manuscript for simplicity, and the AIRS/AMSU FOV is simply referred to as AIRS FOV. The fraction of the vertical profiles from successful retrievals per total number of observed AIRS scenes, defined as yield in this study, is strongly dependent on the cloud class within the AIRS FOV and varies with surface type and latitude, but also altitude because of cloud effects.
The AIRS sounding suite observes up to 324 000 vertical profiles of T and q per day. The matched AIRS–CloudSat observations used in this study are derived from version 5 (V5) of the AIRS L2 support product (Olsen 2007). The L2 support product has a horizontal resolution of approximately 45 km at nadir and is reported at 100 vertical levels. The AIRS system was designed to retrieve T with 1 K of root-mean-square uncertainty in 1-km layers in the troposphere and q with 15% absolute uncertainty for every 2-km layer. The AIRS T and q profiles have been validated against various in situ observations and satellite measurements, which shows AIRS has met these prelaunch specifications (Fetzer et al. 2006, and references therein).
2) Cloud classification: CloudSat cloud scenario
We use the vertical profiles of the cloud classification obtained from CloudSat to group the T and q vertical profiles from AIRS and ECMWF. The CloudSat instrument is a 94-GHz cloud profiling radar with a horizontal resolution of approximately 1.4 km × 2.5 km and sampling every 1 km (Stephens et al. 2008). Cloud types are separated into eight categories and are reported in the CloudSat level 2B cloud classification (2B-CLDCLASS) product if clouds are detected with a signal confidence level ≥20 on a scale of 0–40 (Sassen and Wang 2008). The maximum reflectivity and its horizontal and vertical coherence, the presence of precipitation, and ECMWF T profiles as well as surface topography height are used to classify clouds into stratus (St), stratocumulus (Sc), cumulus (Cu), altocumulus (Ac), altostratus (As), nimbostratus (Ns), deep convection (DC), or high cirrus and cirrostratus (Ci). There are very few occurrences of St, thus the “Sc” category as shown here represents all detectable shallow boundary layer clouds (Yue et al. 2011). “Clear” is determined if the CloudSat cloud fraction along the collocated CloudSat ground track is zero, and the AIRS ECF is less than 0.01 within the collocated AIRS FOV. This combined approach for determining clear sky greatly reduces mismatches that are caused by clouds located off of the CloudSat track but within the AIRS FOV (Kahn et al. 2008). As a result, a total of eight classes, including clear, are used in this study.
CloudSat underestimates the frequency of some clouds like Ci and Sc/St because of limited sensitivity to small hydrometeors in Ci and ground clutter in the lowest 3–4 range bins where Sc and St are common (Tanelli et al. 2008). Multiple studies have shown that the cloud information along the CloudSat track is representative of cloud conditions within the collocated AIRS FOV (e.g., Kahn et al. 2008; Yue et al. 2011).
3) ECMWF model analyses
The ECMWF model analysis is taken from the CloudSat ECMWF-AUX dataset, an auxiliary product containing the ECMWF model analysis data sampled over the CloudSat ground track (CloudSat AN-ECMWF ancillary data, http://www.cloudsat.cira.colostate.edu/ICD/AN-ECMWF/AN-ECMWF_doc_v7.pdf). The global model analyses are generated using a four-dimensional variational data assimilation (4D-VAR) method by assimilating the observations every 6 h day−1 with a short-range forecast from the ECMWF (Courtier et al. 1998). The forecast fields are saved every 3 h at a spatial resolution of approximately 40 km at the equator. A combination of forecast and analysis results is used to produce the AN-ECMWF dataset. For the spatial interpolation, the geolocation data of each CloudSat ray are used as the reference point, and then the orbital sampling is carried out with a bilinear interpolation on the four bounding ECMWF grid points and the CloudSat reference point. This procedure is performed for the two forecast times bounding the time of the CloudSat ray, and then a linear temporal interpolation is performed on the values obtained for each forecast time. The final result is the ECMWF data field interpolated to both the location and the time of each CloudSat ray. A detailed discussion can be found in the CloudSat data documentation on ECMWF-AUX (http://www.cloudsat.cira.colostate.edu/ICD/ECMWF-AUX/ECMWF-AUX_PDICD_3.0.pdf) and AN-ECMWF.
The ECMWF data within any given AIRS FOV are assembled using the index product, and the mean state is calculated. In this study, the T and q vertical profiles in the ECMWF model analysis are used as a benchmark to compare with and assess any biases in the AIRS retrievals. Although absolute comparisons with radiosondes are the most capable method of establishing observational biases (e.g., Tobin et al. 2006; Divakarla et al. 2006), a cloud-type comparison with ECMWF allows for a global analysis with very large sample sizes and is the most promising approach for quantifying cloud state sampling biases.
b. Matching method
The Aqua and CloudSat spacecraft time separation is typically a few minutes. A nearest neighbor approach is used to collocate AIRS and CloudSat profiles based on their latitude and longitude pairs (Kahn et al. 2008; Fetzer 2012; Yue et al. 2011). As a result, approximately 45–50 CloudSat profiles spaced every 1 km are coincident with each AIRS FOV (with ~45 km nominal resolution). Occasionally, there are fewer cloud profile coincidences because the CloudSat ground track varies with AIRS scan angle and with latitude. Collocated data are presented in this study at the AIRS spatial resolution. Over each AIRS FOV, vertical profiles of cloud scenarios are collected along the CloudSat track. However, given that infrared instruments are most sensitive to the upper portion of clouds, the cloud scenario from only the topmost cloud layer in the CloudSat data is considered and used to characterize the AIRS observations and ECMWF model analysis. We define the cloud type within each AIRS FOV only when more than 95% of the collocated cloudy profiles have the same cloud scenario; otherwise, a “mixture” category is defined. To simplify the problem, the mixture type is not considered further in this study. For AIRS FOVs classified for a single cloud type, the statistical mean of ECMWF is taken by averaging over the collocated CloudSat FOVs. Table 1 lists the frequencies of occurrence for the eight cloud classes (including clear) and classification mixtures over the globe.
c. Quality control of AIRS data and sampling of ECMWF
The AIRS V5 data include a case-by-case, level-by-level type of quality control (QC) using two characteristic pressures: PBest and PGood. They are defined as the pressure level p down to which the profile is considered of best and good quality, respectively. To determine PBest and PGood for any given FOV, first an error estimate profile is calculated for T, then based on the preselected thresholds for each level, PBest and PGood are defined as the largest pressure level at which the error estimate in the next three pressure levels are smaller than the thresholds at each corresponding pressure level. These thresholds vary between land and ocean and are larger for PGood than for PBest to allow a looser quality control and better spatial coverage (Susskind et al. 2011) when using PGood.
The error estimate profiles depend on many factors including the surface type, terrain complexity, and small-scale variability of surface temperature and emissivity. The largest uncertainty source is related to clouds within the FOV, which implies that the values of PBest and PGood are mainly affected by cloud. Using collocated CloudSat and AIRS data, Yue et al. (2011) find that the percentage of successful AIRS retrievals to the surface is strongly correlated with cloud coverage within the AIRS FOV. Although Yue et al. (2011) focuses on maritime boundary layer clouds, these conclusions can be extended more generally, as discussed in section 3. In our study, we include AIRS T and q vertical profiles for levels that satisfy p ≤ PBest or p ≤ PGood, which include both partial profiles as well as full profiles to the surface.
Since ECMWF-AUX analyses are already matched to CloudSat and AIRS, the orbital sampling bias has been removed from the comparison. The AIRS quality control is applied to the collocated ECMWF model analyses to generate the EC-S data, which have the same sampling as AIRS to ensure a profile-to-profile comparison between model and satellite data that helps quantify cloud-induced sampling biases.
In summary, the three sets of data (AIRS, EC-S, and EC) can now be compared and are stratified into eight classes (seven cloud types and one clear category) as described previously. The difference between AIRS and EC-S indicates the AIRS observational biases or ECMWF model biases, and we make no attempt here to determine which one may dominate. (This topic warrants further investigation, and our group is presently establishing benchmarks with radiosondes.) The difference between EC and EC-S is purely a result of subsampling from the AIRS quality control and is taken here to represent the cloud-state-dependent sampling biases inherent in AIRS retrievals. This approach also has the potential to quantify biases in other infrared sounders including the Cross-Track Infrared Sounder (CrIS; Bloom 2001) and IASI (Hilton et al. 2012).
3. Mean profiles of atmospheric thermodynamic properties and yield by cloud classes
a. Global vertical profiles
Figure 1 shows the global mean vertical profiles of AIRS retrieval yield for each cloud class over ocean (left) and over land (right). The PBest and PGood quality-controlled data are shown by solid and dashed lines, respectively. The profiles of the mean yield (over all conditions regardless of cloud types) are indicated by the bold black curves in each panel.
Generally, the yield in the free troposphere over land is larger than over ocean for the seven cloud types. This is partially because the threshold values to determine PBest and PGood are larger over land to ensure ample spatial sampling of AIRS retrievals over land. However, because of the uncertainty in the surface emissivity and complexity of the orography, the yield for clear land is smaller than for clear ocean. In addition, T and q in the lower troposphere are difficult to determine accurately because of the lack of thermal gradient between the surface and atmosphere, the impacts of low clouds, and increased small-scale horizontal temperature variability (Kahn and Teixeira 2009). Therefore, the yield decreases near the surface over both land and ocean, with a more prominent effect over land. Furthermore, the spatial scales of a particular cloud type may vary differently over land and ocean with the diurnal cycle, the underlying surface, and the large-scale dynamics that organize clouds at smaller scales (Hahn and Warren 2003; Wood and Field 2011).
The AIRS yield is typically between 60% and 70% globally throughout most of the troposphere, while it decreases to 50% near the surface over ocean and at 800 hPa over land. As a result, a T and q climatology of vertical structure below 800 hPa over land derived from AIRS data should be treated with caution because of reduced sampling.
DC and Ns have the lowest yields since they have the deepest vertical extent and are associated with heavy precipitation and large values of cloud liquid water path, which blocks the infrared radiation emitted from the surface, as well as within and below the cloud. Midlevel clouds such as Ac and As have the second lowest yields. Because of a larger threshold on horizontal extent (see CloudSat project level 2 cloud scenario classification product process description and interface control document; http://cloudsat.cira.colostate.edu/ICD/2B-CLDCLASS/2B-CLDCLASS_PDICD_4.0.pdf), the yield for As cloud is lower than that for Ac. Since the CloudSat defined Ci category includes cirrus, cirrocumulus, and cirrostratus, the low sensitivity of the radar signal to small ice particles suggest the Ci category is mostly cirrus with moderate to high optical depth. The yield under Ci is 67% over ocean and 75% over land in the free troposphere, lower than shallow convective clouds and clear sky.
As shown by previous studies (e.g., Sassen and Wang 2008; Yue et al. 2011), CloudSat Sc is a combination of various types of low clouds in the planetary boundary layer, which include stratocumulus of sufficient depth and extensive coverage, trade wind cumulus, or low clouds associated with baroclinic wave activity in the extratropics. The relationships between AIRS yield and the various types of boundary layer clouds identified by CloudSat are discussed in great detail by Yue et al. (2011). Cu clouds usually have small horizontal coverage but somewhat greater vertical extent than Sc, and the cloud clearing method of AIRS is usually successful. For all cloud types, the difference between PBest and PGood is only noticeable at pressures greater than 700 hPa over ocean and 500 hPa over land, with PGood yields significantly larger than PBest.
Figures 2 and 3 show the global mean ocean (Fig. 2) and land (Fig. 3) profiles of T and q for each cloud type from AIRS retrieval (top panels), EC-S (middle panels), and EC (bottom panels). We can clearly see that the dependence of T and q vertical structures on cloud types is qualitatively similar in all three datasets. In general, T changes over land from the smallest to the largest in sequence from Ns, As, Clear, Sc, Ac, Cu, Ci, to DC. There are some variations near the surface for AIRS and EC-S. Over ocean, the same dependence of T variations on cloud type is found in EC. However, both AIRS and EC-S present the warmest T under Ci and clear sky and intermediate T for DC. There are also strong differences in T and q by latitude for a given cloud type (not shown).
The q profiles from all three datasets show that the driest air is in association with Ns and the moistest with DC. Despite the similarity in the relative ordering of the different cloud types, there are significant quantitative differences among the three datasets. Generally speaking, AIRS and EC-S are relatively drier in the free troposphere and moister in the boundary layer than EC. This is related to the AIRS sampling characteristics in clouds. The retrieval within a given FOV is weighted toward the clear portion of the atmosphere. This cloud-induced effect decreases as the cloud cover in the FOV decreases, consistent with cloudier areas being moister (e.g., Fetzer et al. 2006). As a result, the difference between EC-S and EC are smaller as the cloud type changes from large-scale cloud types with higher liquid water path (DC and Ns) to more broken and vertically shallow cloud (Ac, Cu, and Sc). Another factor at play is that AIRS preferentially retrieves in clear or partly cloudy regions; therefore, FOVs filled by large patches of clouds are missed from the retrieval. However, this effect can lead to either a dry or moist bias. One common feature of AIRS and EC-S is that the near-surface portion of profiles has larger variability than EC. This is caused by small sample sizes near the surface after flagging for the AIRS quality control, especially the PBest flag.
Figure 4 shows the difference between the profile of each cloud type and the mean profile over the global ocean or land for all three datasets. The magnitude of the spread among the lines indicates the meteorological variability sampled by these data. Since AIRS and EC-S have the same exact spatial and (nearly) temporal samples, their difference is from the observational or model state, or some combination thereof. EC represents the full range of meteorological conditions along the satellite ground track. Therefore, the difference between EC and EC-S is entirely due to the application of the AIRS quality control flag. As shown by Figs. 2 and 3, the difference between AIRS and EC-S is very small except in Ns over land, where the sample size is smallest. The profiles for DC and Ns mark the warmest and coldest boundaries of the spread in all three datasets. However, the range is larger in AIRS and EC-S than those for EC, which is because of large differences between AIRS/EC-S and EC in Ns and DC, while results for other cloud types are similar from all three datasets. For example, at 500 hPa, AIRS and EC-S have a T (q) range 5 K (1.1 g kg−1) wider than that in EC.
The close agreement between AIRS and EC-S shows that the AIRS retrieval fairly represents regimes when the AIRS retrieval is successful. Thus, a climatology based on quality-controlled retrievals, even with low yield, can be representative of the true state of the atmosphere, which is in agreement with previous studies (Fetzer et al. 2006; Tian et al. 2006). However, this is not the case for DC and Ns. The global mean differences over both land and ocean indicate that AIRS forms a biased climatology that is more representative of the clear areas surrounding clouds that exist in regions with large horizontal T and q gradients.
b. Regional results
Six oceanic regions are listed in Table 2. Figure 5 shows the AIRS yield profiles for these regions. The relative ordering of the regions' yields by cloud type are nearly identical for all regions (with a few subtle changes) to Fig. 1. The cold tongue has the highest mean yield, while the midlatitude regions have much reduced yields near the surface because of the extensive cloud cover often associated with midlatitude storm tracks and high-latitude ocean characterized by frequent cold air outbreaks (Fetzer et al. 2006).
In general, the tropical regions are in better agreement among AIRS, EC-S, and EC. The largest differences are in Ns and DC in the extratropical oceanic regions. The entire vertical profile in DC is drier in the AIRS and EC-S compared to EC.
4. Observational biases and cloud-state-dependent sampling biases
In this section, the biases between AIRS and EC are quantified.
a. Global results
Figure 6 shows the difference of the T and q vertical profiles among AIRS, EC-S, and EC. As mentioned earlier, we define the difference between AIRS and EC-S as an observational bias for simplicity (middle column in Fig. 6), although this bias may also be due to ECMWF biases that are more or less apparent within particular cloud types (Tian et al. 2006).
1) Observational biases
Small observational biases characterize AIRS and EC-S comparisons within Ns and DC cloud types, where much drier profiles are obtained from AIRS (Fig. 6). The observational bias is characterized by a vertical structure that is dry at pressures less than 800 hPa and wet near the surface in AIRS, with an S-shaped feature below 600 hPa. Note that in discussions on profiles, below (above) means levels below (above) given pressures in height. The magnitude of this bias depends on the cloud type and ranges from −1 to +0.5 g kg−1 below 600 hPa, but can be larger in DC and Ns especially over land. The largest q observational biases are found for DC, which can exceed −3.5 g kg−1 at 800 hPa over land. Since water vapor concentrations in the atmosphere decrease rapidly with height (scale height ~ 2 km), the relative differences (in percent) are presented in Fig. 7. A dry observational bias is found for pressures less than 800 hPa for all cloud types, and the magnitude generally increases with height. Nearer the surface, the average profiles (black bold lines) have a wet bias less than 5%.
For T, the observational biases are small and generally within ±1 K throughout the troposphere, which is within the AIRS prelaunch specifications. The exceptions are for Ns and DC, especially over land, where the cold bias near the surface is as large as −3 K. The biases for Ns and DC are also more sensitive to PBest and PGood than the other cloud types. However, the profiles flagged by PBest show a better quantitative agreement near the surface with those from ECMWF data indicating a higher accuracy with this stricter quality control.
2) Cloud-state-dependent sampling biases
The biases induced by the cloud-state-dependent sampling of AIRS are obtained by differing EC and EC-S (Figs. 6, 7, right columns). These differences originate from the application of AIRS quality flags that largely depend on cloud conditions within the AIRS FOV. Therefore, these biases are strongly correlated with cloud type. For oceanic q in precipitating clouds such as DC, Ns, and As, dry biases reach maxima of −2 g kg−1 between 800 and 700 hPa and a slightly wet bias of +0.5 g kg−1 appears near the surface. Wet biases occur throughout the atmospheric column with a magnitude less than 0.5 g kg−1 and maximize near the surface for other cloud types.
For T, the cloud types that are optically thick or contain precipitation, such as DC, Ns, and As, lead to large cold biases as a result of the sampling, which approach −6 K in the middle troposphere. For other cloud types, warm biases tend to occur near the surface and are as high as 4 K for Sc.
The bias in clear sky is likely caused to a large degree by cloud contamination. In the subtropical regions with persistent Sc, the quality flags remove these uniform Sc FOVs and lead to wet biases in the free troposphere. The preferential sampling over clear and less cloudy conditions also causes warm biases near the surface. The smallest cloud-state sampling biases occur within Ci and Cu. The largest near surface bias is found in Sc.
3) Total biases of climatology based on AIRS retrievals
The maximum (statistical) biases of AIRS retrievals are a combination of the cloud-state sampling biases, as represented by the difference between EC-S and EC (right columns in Figs. 6, 7), and the observational biases, represented by differences between AIRS and EC-S (middle columns of Figs. 6, 7). The total biases (left columns) are very similar to the cloud-state sampling biases, indicating that the latter dominate the uncertainty in the AIRS retrieval. For most cloud types, the bias ranges from −0.7 to +1.5 g kg−1 for q and from −1 to +4.9 K for T over the ocean. The wet biases near the surface can be as large as +1.8 g kg−1 over the ocean for Sc, and the largest dry biases between 700 and 800 hPa are found for Dc and Ns, which can be as significant as −3 and −1.5 g kg−1, respectively. Strong cold biases as large as −5 K for DC and Ns and −2 K for As cloud are found over the ocean in most of the troposphere. Warm biases occur near the ocean surface ranging from +1.5 K for Ci and +4.9 K for Sc.
The comparison over land is complicated by the low yield near the surface. As a result, the difference profiles are not as smooth as those over the ocean, and caution is required when interpreting the near-surface differences. The largest bias over land is seen for DC and Ns, with significant contributions from both observational and cloud related sampling biases. However, the frequency of occurrence of these cloud types is low, so the global mean observational biases are still within ±1 K for T and −0.5–0 g kg−1 for q. The global mean total biases over land are again largely dependent on cloud-state-dependent sampling, ranging from −3 to +1 K for T and from −1 to +0.5 g kg−1 for q above 800 hPa.
b. Regional differences of biases
The results of biases are similar for the tropical ocean, deep tropics, and warm pool regions; thus, we only show the warm pool region in Fig. 8 for brevity. To complement the results with a contrasting region, Fig. 9 shows results for the cold tongue, as its T and q characteristics are quite different from other tropical oceanic regions. The northern hemispheric (NH) midlatitude oceanic region is shown in Fig. 10 for comparison to the two tropical regions. The southern hemispheric (SH) midlatitude results (not shown) are similar to the NH.
One significant difference between tropical and midlatitude regions is that the cloud-state-dependent sampling biases are much smaller in tropics than in the midlatitude and global regions, especially for T. Large cloud sampling biases in q occur for DC and Ns in tropical regions and are as large as −1.5 g kg−1 at 800 hPa. In the midlatitudes, the dry biases associated with these two cloud types occur below 300 hPa and are as high as −2.5 g kg−1 near 800 hPa. The observational biases contribute more to the total biases in both T and q in tropical regions than in the midlatitudes. The magnitude of the bias in tropical T is small and within ±1 K on average. The maximum biases occur at the surface for DC. The small T biases are primarily due to the fact that the variations and gradients of T in the tropics are generally small compared to the midlatitudes.
One point of interest is the difference between the cold tongue and warm pool regions. Over the cold tongue, warm biases in T are found for DC in both the observational and the cloud-sampling biases below 700 hPa while cold biases occur over other tropical regions and in the global mean. For q, AIRS is wetter at pressures greater than 800 hPa in DC, while dry biases are seen throughout the entire troposphere in the other geographical regions.
c. Centered root-mean-square difference
Biases indicate the mean differences of AIRS and ECMWF; however, they do not quantify the similarity of the variability in the T and q fields from the two datasets. Therefore, we calculate the CRMSD of AIRS with respect to ECMWF data, which is added quadratically to the bias to determine the full root-mean-square difference (Taylor 2001). Figures 11 and 12 show the CRMSD in the AIRS T and q, respectively, for the warm pool, cold tongue, and NH midlatitude ocean regions, as well as for global ocean and land. To calculate CRMSD, both AIRS and ECMWF must have the exact same sample dataset. Therefore, the CRMSD correspond to the comparison between AIRS and EC-S, where the related “bias” is the observational bias.
The CRMSD have a clear dependence on cloud type. For T, they generally increase from clear sky, through Ci, Cu, Sc, Ac, and As to DC and Ns, except that for the cold tongue region between 300 and 600 hPa, the CRMSDs for DC and Ns are smaller than other cloud types. Tropical and oceanic regions have smaller CRMSD than the midlatitudes and land regions, similar to the observational bias. The largest CRMSD is found near the surface ranging from 1 to 4 K depending on the region, surface type, and cloud type. Similar to the bias profiles, a stricter quality control using PBest gives smaller CRMSD compared to using PGood.
The CRMSDs of q are calculated as the percent differences with respect to the mean q field of EC-S at each pressure level. Its magnitude increases with height from about 10% at the surface to nearly 80% near 200 hPa. The CRMSDs of q also have a dependence on cloud types. However, unlike T, the relative ordering on cloud type varies strongly with region and altitude. Although clear sky has the smallest values of observational bias for pressures greater than 500 hPa (Fig. 7), clear sky CRMSDs have an intermediate magnitude among all conditions. Near the surface, DC has relatively small CRMSD; except in tropical regions, Ns generally has the largest CRMSD.
5. Summary and conclusions
Vertical profiles of temperature T and water vapor q from the Atmospheric Infrared Sounder (AIRS) satellite instrument and the European Centre for Medium-Range Weather Forecasts model analysis are stratified by a CloudSat cloud classification methodology that is collocated in space and time. Global and regional mean T and q profiles for seven cloud types and clear sky are obtained from AIRS and compared to those from the ECMWF. An observational bias and a cloud sampling bias are quantified across a variety of cloud types.
We retain AIRS T and q profiles for levels that satisfy p ≤ PBest or p ≤ PGood, which include both partial and full profiles to the surface. These quality indicators are largely dependent on the cloud conditions within the AIRS FOV. They are also applied to the collocated ECMWF analyses that are subsampled to the AIRS quality control (EC-S), so that the same samples in space and time are compared to quantify the observational bias. The differences between EC-S and all of the collocated ECMWF model data (EC) are solely a result of the quality control of AIRS and give a quantifiable estimate of the cloud-state-dependent sampling biases that are inherent in the AIRS retrieval.
The AIRS retrieval yields are greater than 90% for clear skies and 50%–80% for shallow convective or stratiform clouds, but less than 10% for optically thick and/or precipitating clouds such as DC and Ns. Near the surface, the retrieval yield typically decreases because of the presence of low-level cloud over ocean, while the decreases are more significant over land because of the uncertainties in surface emissivity and complex topography. Regional yield profiles have a similar dependence on cloud types (i.e., the relative ordering of the yield) to the global retrieval yield profiles. However, the tropical regions have higher yields than the midlatitudes for all cloud types. This is likely because of the horizontal scale of a given cloud type over land or ocean, the variability and gradients in T and q, and sampling as a function of the diurnal cycle and meteorological regime. The relative importance of these effects warrants further investigation. The global AIRS yield is between 60% and 70% throughout the troposphere and decreases to 50% below 900 hPa over ocean. The yields are higher in the free troposphere over land compared to over ocean, which is partially because of the different cloud characteristics over land and ocean and the selection of larger thresholds in AIRS quality control over land. The AIRS yields over land decrease to less than 50% for pressures greater than 800 hPa, and any climatology near the land surface may contain significant sampling biases.
The observational biases are very small over the ocean, within ±1 K in T and from −1 to +0.5 g kg−1 in q. Larger T observational biases (−2 K at the surface for ocean and −3 K at 800 hPa for land) are found for DC and Ns clouds. The largest q observational biases are found for DC and can exceed −3.5 g kg−1 at 800 hPa over land. An S-shaped feature is observed with AIRS, revealing slightly drier air in the free troposphere and slightly wetter air near the surface. Larger observational biases in DC and Ns strongly suggest that AIRS cannot faithfully represent the thermodynamic properties of these cloud types. The average observational biases overall remain small since the frequency of occurrence of DC and Ns are lower than the other cloud types. Also, biases can be of opposite sign for different cloud type (summarized below), so biases cancel when averaged across regimes. Therefore, the global mean of AIRS remains very similar to the ECMWF analysis.
The cloud-state-dependent sampling introduces significant biases and is the dominant form of uncertainty. Biases associated with DC and Ns can be as large as −6 K for T in the middle troposphere and −2 g kg−1 for q between 800 and 700 hPa depending on the cloud type. Cloud-induced biases tend to be negative (cold, dry bias) in the free troposphere and positive (warm, wet bias) near the surface for optically thick and precipitating clouds. Warm biases in T are observed for either shallow or broken clouds. This behavior is caused by preferential AIRS sampling along the edges of cloud systems that tend to be drier in the free troposphere and moister in the boundary layer. The wet bias in clear conditions could be caused by contamination of clouds not detected by CloudSat or AIRS, while poor sampling in the extensive stratocumulus cloud induces a wet bias in the free troposphere (e.g., Fetzer et al. 2006).
The center root-mean-square differences (CRMSD) between AIRS and ECMWF are strongly dependent on cloud type. In general, tropical and oceanic regions have smaller CRMSD than the midlatitudes and land regions. The magnitude of CRMSD for T increases from clear sky, Ci, Cu, Sc, Ac, and As to DC and Ns, indicating deteriorating agreement between the two datasets as well as larger variability in T. For the cold tongue region, DC and Ns have smaller T CRMSD between 300 and 600 hPa. For q, the CRMSDs show different cloud type dependence at different altitudes. Relatively smaller CRMSD of q is found for DC below 500 hPa and somewhat larger CRMSD is found for Ns.
The research described in this paper was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration. This project was supported by NASA's Making Earth Science Data Records for Use in Research Environments (MEASURES) program. We acknowledge the support of the AIRS Project at JPL. The authors thank Baijun Tian, Bjorn Lambrigsten, Ed Olsen, and Joel Susskind for useful feedback during the preparation of this manuscript. We also thank the anonymous reviewers for their helpful suggestions and comments.