An assessment of the absolute accuracy of the Atmospheric Infrared Sounder v5 precipitable water vapor product at tropical, midlatitude, and arctic ground-truth sites: September 2002 through August 2008

Authors


Abstract

[1] The Atmospheric Infrared Sounder (AIRS) is the first of a series of satellite sensors that exploit high spectral resolution and broad spectral coverage of the midinfrared to improve the retrieval accuracy of passive infrared sounding. The AIRS atmospheric retrieval goals are to obtain 1 K accuracy for 1 km layers below 100 mb for temperature and 10% for 2 km layers for water vapor in clear and most cloud conditions. The AIRS total column precipitable water vapor (PWV) is obtained by integrating the vertical profile of water vapor mixing ratio derived from cloud-cleared radiances. The accuracy goal of the AIRS PWV product is 5%. This paper provides a validation of the AIRS PWV product at three distinct climate sites over the nearly full range of total water amounts observed on Earth (between 0.1 and 6.5 cm). Six years (September 2002 to August 2008) of AIRS v5 retrievals of PWV are evaluated against ground-based microwave radiometer (MWR) data at three Department of Energy Atmospheric Radiation Measurement program (ARM) sites. The accuracy of the MWR PWV retrieval is estimated to be between 1% and 3%. This study shows that the agreement between the MWR and AIRS retrievals of PWV is within 5% at all three ARM climate sites for most conditions. The notable exceptions are (1) very dry cases (PWV < 1 cm) over the Southern Great Plains (SGP) land site during both daytime and nighttime, where AIRS is too moist by 15%–30% and (2) nighttime observations over the SGP land site for PWV > 1 cm, where AIRS is too dry by about 10%. The moist bias for low water amounts (usually observed during the winter) over land could be a surface emissivity–related error since very little bias is seen at the ARM Arctic site for similar water amounts. The cause of the dry bias at nighttime over land for moderate water amounts is not determined by this study. However, a spatial map of the diurnal bias in monthly AIRS water amount suggests that this effect is related to meteorological conditions in the U. S. Great Plains, which in the summertime is characterized by a moist boundary layer. The diurnal error in AIRS PWV at the SGP site seen with respect to the MWR data are confirmed by PWV amounts derived from a coincident ground-based GPS receiver.

1. Introduction

[2] The hydrological cycle is of great interest to both climate scientists and weather forecasters because of water's central role in the radiation balance and convective stability of the Earth's atmosphere [Ramanathan, 1981; Chahine, 1992]. Precipitable water vapor (PWV) is defined as the amount of liquid water that would be produced if all of the water vapor in an atmospheric column were condensed. Because of the Clausius-Claperyon relation between temperature and saturation vapor pressure, global warming is predicted to increase the PWV over oceans; however, the response over land is less certain [Trenberth et al., 2005; Held and Soden, 2006; Soden and Held, 2006; Wang et al., 2008]. Accurate estimates of water vapor from satellites for both ocean and land areas will provide a valuable global constraint on climate model predictions of atmospheric water vapor [Gettelman et al., 2006; Pierce et al., 2006]. Satellite measurements of water vapor are also playing an increasingly important role in numerical weather prediction including forecasting severe weather and hurricane formation [Le Marshall et al., 2006; Randel and Park, 2006; Wu et al., 2006].

[3] Until recently, satellite profiles of temperature and water vapor have been available primarily from low spectral resolution infrared sensors [Smith, 1983; Susskind et al., 1984; Le Marshall, 1988; Ma et al., 1999; Seemann et al., 2003]. However, the utility of these retrievals is limited by a relatively low vertical resolution, which makes it difficult to separate contributions of atmospheric water vapor from variations in surface emission, especially over land. The NASA Atmospheric Infrared Sounder (AIRS) is the first of a series of satellite sensors, which combine accurate onboard radiometric calibration with higher spectral resolution coverage of the midinfrared to improve the retrieval accuracy of passive infrared sounding [Chahine et al., 2006]. AIRS was launched into a Sun-synchronous orbit in May 2002 on NASA's Earth Observing System (EOS) Aqua platform [Parkinson, 2003]. It provides the capability to retrieve atmospheric temperature and water vapor with a vertical resolution of a few kilometers throughout the troposphere. The Infrared Atmospheric Sounding Interferometer (IASI) has been making similar infrared spectrally resolved measurements from the European METOP-A platform since launch in 2007 with at least two additional platforms planned as part of the EUMETSAT Polar System (EPS) program [Klaes et al., 2007]. EPS is the European component of the joint European/U. S. polar satellite system (IJPS). The cross-track infrared sounder (CrIS) is the sensor that will continue the AIRS high spectral resolution infrared measurement record for the next several decades as part of the U. S. National Polar-Orbiting Operational Environmental Satellite System (NPOESS) [Cunningham et al., 2003]. Accurate measurements of the thermal infrared emission spectrum have the potential to provide altitude resolved climate trends of water vapor over both ocean and land with known uncertainties. A careful assessment of the accuracy of the AIRS retrievals of water vapor is necessary to develop confidence in using the derived products for climate studies.

[4] An assessment of the climate quality of products from the generation of EOS sensors is currently a priority for NASA as part of the transitional NPOESS Preparatory Program (NPP) [Schweiss et al., 2008]. The AIRS atmospheric retrieval goals are to obtain 1 K accuracy for 1 km layers below 100 mb for temperature and 10% for 2 km layer water vapor column density for a global ensemble of clear to partly cloudy ocean and land conditions [Fetzer et al., 2003]. Several studies have evaluated the accuracy of the AIRS version 4 (v4) temperature and water vapor retrievals [Tobin et al., 2006; Divakarla et al., 2006]. Tobin et al. [2006] created a “best estimate” validation data set from three dedicated rawinsonde launch campaigns conducted between September 2002 and September 2004. The rawinsondes were launched from three climatologically diverse ground-based observation sites operated by the Department of Energy Atmospheric Radiation Measurement (ARM) program at Aqua satellite overpass times with supporting ancillary information to provide absolute calibration [Ackerman and Stokes, 2003]. This study concluded that the AIRS water vapor retrieval performance was site specific, with 1 km layer RMS errors of 20% or less for the tropical ocean and from 25% to 35% below 400 mb at the U. S. Southern Great Plains (SGP) land site. Divakarla et al. [2006] came to similar conclusions through a comparison to the global operational radiosonde network at hundreds of ground stations. These studies showed that the AIRS v4 retrievals were close to achieving the desired accuracy goals for water vapor vertical profiles for nonfrozen ocean scenes but had larger errors over land areas.

[5] The AIRS total column precipitable water vapor is obtained by integrating the vertical profile of water vapor mixing ratio retrieved from cloud-cleared radiances. The accuracy goal of the AIRS PWV product is 5%. A comparison over ocean scenes of NASA AIRS v4 PWV to that derived from the NASA AMSR-E sensor on the same Aqua spacecraft was provided in the work of Fetzer et al. [2006]. The results indicate observational biases between the two sensors are generally less than 5% in PWV when identical scenes are observed. Larger systematic biases in sampling were noted for certain high latitude and subtropical areas where clouds may affect the AIRS retrieval methodology. Validation of PWV over land scenes was presented in Rama Varma Raja et al. [2008], who compared AIRS v4 retrievals to measurements from a network of more than 375 GPS receivers over the continental United States. That study concluded that the AIRS retrievals were in good agreement with the GPS PWV with a small dry bias in winter (<−0.5 mm) and a small wet bias in summer (about 1.2 mm). However, the study consisted of only 7 months of data (April–October 2004) and did not distinguish between daytime and nighttime observations, which can have different error characteristics.

[6] Preliminary validation of AIRS v4 PWV over land was also included in the work of Tobin et al. [2006] using a ground-based microwave radiometer at the ARM SGP site but only for the time periods of the special AIRS radiosonde launch campaigns. The preliminary results of PWV validation presented in the work of Tobin et al. [2006] are repeated in this paper for the newer AIRS v5 PWV product while at the same time extending the evaluation time period to a complete 6 year record of satellite overpasses. This paper also extends the Tobin et al. [2006] PWV validation by including validation at high-latitude arctic sites and providing a diurnal characterization of the results at each site. The resulting improved statistical precision allows us to evaluate the diurnal character of the AIRS PWV product error as a function of total column water vapor at three diverse climate sites.

[7] For ground-based PWV validation, a microwave radiometer provides an advantage over using rawinsondes in that it can provide a set of well-calibrated observations at coincident satellite overpass times. This reduces the error in the spatial and temporal sampling compared to radiosondes while providing improved absolute accuracy and precision. Turner et al. [2003] documented that the calibration of the Vaisala RS-80H rawinsondes used by the ARM program from 1994 until 2000 varied between calibration batches and with the age of the rawinsonde, which translated to increased uncertainties in the humidity measurements. The same study noted that the profiles from these rawinsondes contain a dry bias that is diurnally dependent (dry bias is larger during the daytime than at night). A similar diurnal bias was found for Vaisala RS-90/92 sondes [Vomel et al., 2007; Miloshevich et al., 2006]. Using the ARM SGP site microwave radiometer (MWR) to scale rawinsonde profiles as input to a line-by-line radiative transfer model and comparing the resulting radiances to highly accurate measurements from the Atmospheric Emitted Radiance Interferometer (AERI) has revealed that the calibration of the MWR, on which the retrieval of PWV critically depends, is very stable and shows no diurnal dependence [Cady-Pereira et al., 2008; Turner et al., 2003; Knuteson et al., 2004a, 2004b]. This supports the conclusion that, while rawinsondes sensors are useful for studies of vertical profiles of water vapor, the ARM MWR is a higher-accuracy reference that is ideally suited for satellite PWV validation [Liljegren et al., 1997; Revercomb et al., 2003; Turner et al., 2003].

[8] The purpose of this study is to establish the absolute accuracy of the retrievals of PWV from the AIRS sensor using ground-based ARM MWR observations. Coincident observations from ARM MWR and AIRS are presented that cover a 6 year time period from September 2002 to August 2008. These comparisons highlight both the seasonal and diurnal variability of PWV at three unique climate measurement sites, and the accuracy with which satellite algorithms are able to capture changes in water vapor. The remainder of this paper will describe the data and methods used in the analysis and present results at each of the three validation sites.

2. Data

2.1. AIRS

[9] The AIRS is a hyperspectral, scanning infrared sounder that measures emitted infrared radiation in 2378 spectral channels spanning the range from 3.7 to 15 μm [Aumann et al., 2003]. The spatial resolution at nadir is 13.5 km, and complete global coverage is attained daily using cross-track scanning. The AIRS operates on the NASA Aqua satellite in tandem with the Advanced Microwave Sounding Unit A (AMSU-A). Because of its high spectral resolution in the infrared and excellent absolute accuracy, the AIRS retrieval algorithms are able to measure vertical profiles of atmospheric temperature, moisture, and trace gases. The method for retrieving atmospheric moisture profiles is an iterative least squares solution that minimizes the dependence of the result on the first guess and simplifies assumptions about the properties of the clouds in the scene [Susskind et al., 2003, 2006, 2010]. The water vapor retrieval utilizes 66 spectral channels that are generally selected to cover a range of wavelengths of both on and off and both weak and strong water vapor absorption features throughout the water vapor continuum (between about 1300 and 1600 cm−1) and the atmospheric window regions. For the exact channels used, see Susskind et al. [2003]. Several studies have confirmed that both the AIRS radiances and the AIRS clear-sky forward model have an absolute accuracy of around 0.2 K for the spectral channels used in temperature and water vapor retrievals [Fetzer et al., 2003; Strow et al., 2006]. Infrared radiances are strongly affected by clouds and precipitation. To provide accurate quality controlled retrievals for fields of view containing clouds, AIRS infrared measurements are combined with AMSU-A microwave measurements. AMSU-A is capable of providing profiles from the surface up to 40 km, even in the presence of clouds. Clear column radiances are estimated using data from these two unique instruments, making the retrieval of temperature and moisture profiles possible in scenes that have up to 80% cloud fraction [Chahine, 1974, 1977; Susskind et al., 2003, 2006, 2010].

[10] This study examines AIRS level 2 v5 moisture products, which are available beginning with September 2002 and which include a number of improvements to both moisture retrieval quality and error estimation over previous versions, especially over land [Susskind et al., 2010]. The specific data field examined in this study is totH2OStd. Additionally, totH2OStdErr and Qual_H2O are used for quality control of the PWV product. All variables are read from the swath name “AIRS L2_Standard_atmospheric&surface_product.” It should be noted that, during the time period included in this study, no correlation was observed between AIRS minus MWR PWV and viewing zenith angle.

2.2. MWR

[11] This study will focus on PWV comparisons over three ground-based observation sites operated by the U. S. Department of Energy Atmospheric Radiation Measurement program [Stokes and Schwartz, 1994]. The sites are a part of the ARM Climate Research Facility (ACRF), which provides routine quality assessment of the data record from facility instruments [Peppler et al., 2008]. The ACRF sites contain dual-frequency water vapor radiometers of the WVR-1100 series from the Radiometrics Corporation (see online at www.radiometrics.com), operating at 23.8 and 31.4 GHz and calibrated using a tipping curves method developed by the ARM program [Liljegren, 2000].

[12] The ARM Southern Great Plains (SGP) site is located in the central United States, near Lamont, OK (36.616°N, 97.50°W). The SGP site is located in a region that exhibits strong seasonal variability in water vapor amount making it ideal for testing PWV over a wide dynamic range. Two of the primary retrieval challenges in this area are the spatial and temporal variability of the land surface emissivity and the complexity of the storm systems that are characteristic of the region. The MWR at the SGP central facility site has been operating since 1993.

[13] The ARM Tropical Western Pacific (TWP) site is located on the small island of Nauru, which is a maritime location located approximately 1200 miles northeast of Papua New Guinea (0.521°S, 166.916°E). It is a useful site for validation because the surrounding ocean provides a well-known surface emissivity, relatively high levels of atmospheric water vapor, and little variation in moisture and temperature throughout the year. The TWP site at Nauru contains a MWR, which has been making measurements since 1998.

[14] The ARM North Slope of Alaska (NSA) site is located in Barrow, Alaska, adjacent to the Arctic Ocean (71.323°N, 156.616°W). During the winter months, the surrounding surface type is primarily sea ice and snow-covered land surface, but during the summer months, the surface type is either open ocean or tundra with many small melt ponds. Satellite retrievals are challenging at high latitudes due to the high solar zenith angle (for retrievals that use visible channels) and the extremely variable and potentially highly reflective surface type. The ARM site at Barrow provides valuable observations of the atmosphere under very cold and dry conditions, as well as in an area that would otherwise be data sparse. The MWR at Barrow has been observing atmospheric water vapor since 1997.

[15] All of the ARM MWR instruments are two-channel systems that measure downwelling microwave radiation at 23.8 and 31.4 GHz. The first frequency is on the side of a water vapor absorption line, and the signal is dominated by water vapor emission. The second frequency is an atmospheric window between absorption lines, and the signal is dominated by cloud liquid water. From these two frequencies, PWV and liquid water path (LWP) can be retrieved. The accuracy of the retrieved quantities depends on how well the radiometer is calibrated. The ARM MWRs use frequent tip curves [Han and Westwater, 2000] for calibration. An automated routine was developed that performs a tip curve whenever the sky is determined to be cloud-free and horizontally homogeneous, and these successful tip curves are used to maintain the instrument's calibration to better than 0.3 K, which equates to an uncertainty in PWV of less than 0.3 mm [Liljegren, 2000].

[16] The algorithm used to retrieve PWV from the each of the ARM MWR instruments is documented by Turner et al. [2007]. This algorithm is an advanced statistical retrieval that is tuned to agree with physical retrievals, where the latter are only performed at rawinsonde launch times. Additionally, systematic biases in the MWR observations that are a result of the calibration variability of the instrument are removed prior to the retrieval. This results in improved accuracy over the original statistical approach used by ARM. Turner et al. [2007] document that the PWV derived from the ARM MWR using this method shows excellent agreement with PWV derived from a scanning Raman lidar, calibrated to a well-characterized chilled mirror water vapor hygrometer.

3. Methodology

[17] This study considered time and space colocated AIRS and MWR retrievals of PWV over the ARM SGP, TWP and NSA sites. A file was generated of the satellite and ground site “matchups” at each site containing the AIRS PWV for the closest field of view passing the AIRS quality control tests and the ARM MWR PWV for a 10 min average centered each overpass time. The Sun-synchronous Aqua satellite has a nominal equator crossing time of 1330 (local solar time) in the ascending (daytime) mode. The mean AIRS overpass times for each of the ARM sites were 1936 UTC (ascending) and 0824 UTC (descending) for SGP, 0224 UTC (ascending) and 1421 UTC (descending) for TWP, and 2207 UTC (ascending) and 1344 UTC (descending) for NSA. The range of overpass times was just over 50 min for SGP, just over 40 min for TWP, and just over 2.5 h for NSA. The SGP and TWP sites each have distinct daytime and nighttime overpasses, which are associated with the ascending and descending orbits of Aqua, respectively. The overpasses of the high-latitude NSA site are more frequent, and thus, the time windows for the ascending and descending orbits are wider and each day may contain more than one ascending or descending overpass. Additionally, the high solar zenith angles observed at high latitudes make a day/night distinction more difficult. For the purposes of this study, an observation is classified as daytime if the solar zenith angle is less than 90°.

[18] In addition to the precipitable water vapor totH2OStd, the AIRS L2 product contains an estimate of the error in the retrieved total column water vapor. The data used in this study were limited according to this reported error estimate to ensure not only that all retrievals fell within a consistent range of estimated uncertainties but also that enough data were included to achieve adequate statistical significance of the results. The field totH2OStdErr is shown in Figure 1 for each of the three sites. Also shown are the threshold values used for each site: 19% for SGP, 16% for TWP, and 35% for NSA. These thresholds were chosen based on inspection of Figure 1 to remove outliers detected by the AIRS error estimation procedure while retaining the bulk of the measurements. Use of this error estimate variable in a limits test is quite effective by itself but was complemented in this study by the use of AIRS quality control flags in order to filter out bad retrievals to the maximum extent possible. The points shown in Figure 1 have not been filtered using the AIRS quality flags, which would have removed many of the outliers that are above the indicated cutoff used in this study. This study has not addressed the adequacy of those quality flags, since the data used in this analysis restricted to a subset of the entire AIRS data set. The AIRS and MWR quality control criteria used in this study are unique to each site and are listed in Table 1.

Figure 1.

Percentage error in the AIRS PWV as reported in the level 2 data file for three ARM sites (top, SGP; middle, TWP; bottom, NSA). Shown as a solid line on each plot is the threshold value used in this study (19% for SGP, 16% for TWP, and 35% for NSA).

Table 1. Quality Control Criteria Used in the Analysis of the AIRS and MWR Data
SGPTWPNSA
Distance between AIRS observation and site ≤ 50 kmDistance between AIRS observation and site ≤ 50 kmDistance between AIRS observation and site ≤ 50 km
Qual_H2O ≠ 2Qual_H2O ≠ 2Qual_H2O ≠ 2
Qual_Surf ≠ 2Qual_Surf ≠ 2Qual_Surf ≠ 2
PGood ≥ PSurfStdPGood ≥ PsurfStdPGood ≥ PSurfStd
AIRS cloud fraction ≤ 0.8AIRS cloud fraction ≤ 0.8AIRS cloud fraction ≤ 0.8
AIRS viewing zenith angle ≤ 45AIRS viewing zenith angle ≤ 45AIRS viewing zenith angle ≤ 45
totH2OStdErr ≤ 19%totH2OStdErr ≤ 16%totH2OStdErr ≤ 35%
Uncertainty in MWR 10 min mean ≤ 0.75%Uncertainty in MWR 10 min mean ≤ 0.3%Uncertainty in MWR 10 min mean ≤ 1.3%
PWV (AIRS or MWR) between 0 and 10 cmPWV (AIRS or MWR) between 0 and 10 cmPWV (AIRS or MWR) between 0 and 2 cm

[19] Since cloud clearing techniques are implemented in the AIRS retrieval, we included observations only where the AIRS reported cloud fraction was 80% or less. We noted no correlation between AIRS and MWR differences and cloud fraction for cases that passed the other quality control criteria.

[20] For this study, the 20 s MWR data were averaged over 10 min intervals centered on the AIRS overpass time in order to reduce the random noise in the PWV retrievals. The uncertainty in the 10 min average was calculated as the standard deviation of the individual 20 s PWV values divided by the square root of the number of points within the time interval. The uncertainties in the 10 min mean MWR value computed for each overpass at each of the three sites are shown in Figure 2. Thresholds of 0.75%, 0.3%, and 1.3% were selected for SGP, TWP, and NSA, respectively. Matchups with the MWR were excluded if the uncertainty in the MWR mean exceeded the threshold value. MWR observations with abnormally high uncertainties could include cases with very few data points in the 10 min average or cases where the water vapor was changing rapidly around the satellite overpass time. MWR observations in which optically thick water clouds were likely to be present were also excluded from the 10 min average. The criteria used to screen out these cases was that retained “clear-sky” observations have MWR-retrieved liquid water path (LWP) values less than 200 g/m2 and brightness temperatures less than 100 K at both 23 and 31 GHz.

Figure 2.

Percentage uncertainty in the 10 min mean MWR PWV for the three ARM sites (top, SGP; middle, TWP; bottom, NSA). Shown as a solid line on each plot is the threshold value used in this study (0.75 for SGP, 0.3 for TWP, and 1.3 for NSA).

[21] Matchups were limited to cases where the distance between the AIRS observation and the MWR observation was 50 km or less. For each satellite overpass, the closest AIRS retrieval to the ARM MWR location was selected for which all of the quality control criteria were met. This has the advantage of providing a spatially unbiased sampling in the neighborhood of the ground site while maintaining the 50 km AIRS spatial resolution. Figure 3 shows an example profile from AIRS (in red) plotted with two Vaisala RS92 rawinsonde profiles. The sondes were launched approximately 3 h before and after the AIRS overpass time and show that the temperature and moisture profiles were fairly constant on this particular evening. In this case, the AIRS, MWR, and sonde PWV values are all within about 0.3 cm of each other. It is apparent that the rawinsonde measures water vapor at a much finer vertical resolution than AIRS is able to resolve using passive infrared remote sensing. This fact should be considered when evaluating the AIRS PWV product since a significant percentage of the total water column is contained in the surface boundary layer near the sources of water vapor flux.

Figure 3.

An example of AIRS retrieved temperature and water vapor profile is compared with two Vaisala RS92 rawinsonde profiles at the ARM SGP site. Sondes were launched approximately 3 h before and after the AIRS overpass time of 0831 UTC on 31 July 2005.

[22] In addition to AIRS PWV comparisons with the MWR instruments at each site, this study also considers comparisons with GPS measurements at the SGP site only. Validation of AIRS using GPS measurements is useful because there is a network of GPS instruments in place that provides good spatial coverage over CONUS [Ware et al., 2000; Wolfe and Gutman, 2000]. The methodology for retrieving PWV from GPS measurements using this network of receivers is based on the principles described by Bevis et al. [1992] and is described fully by Ware et al. [2000]. Rama Varma Raja et al. [2008] have compared AIRS v4 retrieval products with GPS retrievals of PWV over CONUS for limited time periods. This paper continues those earlier GPS studies by analyzing the comparison of GPS water vapor at the ARM SGP site to the AIRS v5 retrieval products for a 6 year period.

4. Results

[23] This study considers colocated AIRS and MWR retrievals of PWV over the ARM sites for a 6 year period between September 2002 and August 2008. In order to facilitate the interpretation of results, the analysis is presented first for each site separately and then for all sites combined. Note that mean percentage errors presented in the following sections were calculated by dividing each AIRS-MWR PWV difference by the MWR value and averaging over all of the observations within each month or PWV bin. Larger percentage errors at small PWV amounts are to be expected because the denominator in these cases is small.

4.1. SGP

[24] During the time period examined, a total of 976 AIRS overpasses of the SGP site met the criteria shown in Table 1. Of these, 524 samples were daytime observations and 452 were nighttime observations. Table 2 lists the bulk comparison statistics between AIRS and MWR PWV for matchups at each of the three ARM sites. The AIRS data at the SGP site show a very slight positive bias during the daytime and a negative bias at nighttime. The linear fit for the daytime (nighttime) data has a slope of 0.99 (0.79) and an intercept of 0.07 cm (0.36 cm). This indicates excellent agreement between the two instruments during the daytime but shows a significant negative bias at nighttime. This bias is largest at SGP for PWV values greater than about 2 cm, which are present during the warmer months.

Table 2. Overall Statistics for the AIRS Minus MWR Matchups at ARM Validation Sitesa
 SGP AllSGP DaySGP NightTWP AllTWP DayTWP NightNSA AllNSA DayNSA Night
  • a

    Results are for the time period 1 September 2002 through 31 August 2008. PWV units are in centimeters.

Mean MWR PWV2.212.162.264.864.894.840.680.850.44
Linear fit slope0.911.000.800.920.930.910.960.950.95
Uncertainty in the slope0.00840.0110.010.0120.0180.0150.0060.0080.009
Linear fit intercept0.190.0760.360.390.310.470.0170.0320.013
AIRS-MWR bias−0.00600.070−0.094−0.0043−0.0370.036−0.0089−0.0077−0.011
AIRS-MWR StDev0.340.310.360.220.230.200.170.190.12
AIRS-MWR RMS0.340.310.370.220.240.200.170.190.12
Correlation r0.960.970.960.960.950.970.940.920.93

[25] Figure 4 shows the daytime and nighttime monthly mean values in PWV for the matchups of AIRS and MWR at the SGP site over the 6 year comparison period. Also shown is the mean diurnal difference, monthly mean day minus monthly mean night, for each sensor. The error bars shown on this plot indicate the uncertainty in each monthly mean (with a 95% confidence interval or k = 2), calculated by dividing twice the monthly standard deviation by the square root of the number of matchups for each month. The error bars shown for the diurnal differences are the root sum square of the daytime and the nighttime monthly mean uncertainties. Observations from both instruments generally indicate a positive diurnal difference (day > night) during the warmer months (May–September) and a slightly negative diurnal difference (day < night) during the colder months (October–April). Compared to the MWR, AIRS overestimates the diurnal difference during the warmer months by as much as nearly 0.6 cm (in August) relative to the MWR. However, since the number of daytime cases is not equal to the number of nighttime cases, some of this apparent diurnal signal may actually be due to temporal sampling errors in estimating the monthly mean values of PWV.

Figure 4.

Monthly mean daytime, nighttime, and diurnal difference (mean day-mean night) in PWV over SGP from MWR and AIRS. Error bars represent the uncertainty in each monthly mean at the 95% confidence level.

[26] In order to further investigate the apparent diurnal bias in AIRS PWV at the SGP site, we restrict the analysis to a subset of matchups where the daytime observations are within 13 h of the nighttime observations. Figure 5 shows the diurnal difference for AIRS and MWR, calculated by averaging these “same day” differences. The error bars shown are the calculated uncertainty (95% confidence range) in the mean of the differences taking into account the number of samples. The mean differences are similar to those shown in Figure 4; however, the uncertainties are reduced due to the method of differencing of matched day/night observations rather than the comparison of 6 monthly mean values. Figure 5 shows that the error in the AIRS measurement of the diurnal cycle in PWV is statistically significant during the summer months at the SGP site. Note that the MWR diurnal signal is statistically equal to zero for the months of June, July, and August whereas these 3 months show the largest diurnal signal in the AIRS PWV product (0.4–0.6 cm or about 10%). Inspection of Figure 4 shows that this AIRS summertime diurnal bias is primarily due to an AIRS underestimate of the PWV during the nighttime overpasses of the SGP site.

Figure 5.

Monthly mean diurnal difference, as calculated using individual daytime/nighttime observations within a 13 h period over SGP. Error bars are the uncertainty in the mean difference (at the 95% confidence level). The bottom plot shows the number of pairs of observations included in each monthly mean.

[27] An illustration of the regional extent of the error in the AIRS diurnal cycle of PWV is provided in Figure 6 which shows the AIRS v5 level 3 PWV product as a diurnal difference (daytime minus nighttime estimates) over North America for the month of July 2003. The diurnal difference in PWV observed by AIRS at the SGP site is typical of the diurnal difference observed by AIRS throughout the central and southwestern United States. However, the magnitude of the AIRS summertime diurnal PWV signal is 3–5 times larger than the 1–1.8 mm previously reported by ground-based sensors in the central United States [Dai et al., 2002]. Figure 6 combined with the analysis at the SGP site relative to the ARM MWR implies that a nighttime dry bias in the AIRS PWV retrievals exists for the unique conditions of the summertime U. S. Great Plains and U. S. Southwest, but this error is not evident in the Eastern U. S., Pacific Northwest, Alaska, or Canada. Inspection of this monthly diurnal difference for other months of the year show that this diurnal effect disappears during the fall and winter months but reappears each summer in the U. S. Great Plains beginning about May in the south and progressing north each month. The cause of this AIRS nighttime dry bias has not been identified, but the authors note that differences in land cover/land use, surface emissivity, and unique boundary layer meteorology all could be contributing factors.

Figure 6.

AIRS diurnal PWV difference (monthly mean day - monthly mean night) in mm, over North America for July 2003. The location of the ARM SGP site is noted with an arrow.

[28] In order to further quantify the diurnal nature of the AIRS PWV product performance, we compute monthly statistics for matchups between AIRS and MWR for the 6 year period separately for daytime and nighttime overpasses. Figure 7 shows the AIRS minus MWR PWV matchups averaged over each month as a percentage difference. Positive values indicate that AIRS is moister than the MWR, negative values that AIRS is too dry. The largest percentage differences in both the daytime and nighttime data (about +15% for both) are for winter months. Throughout the year, the daytime AIRS retrieval has a near zero or slight moist bias. The nighttime bias is positive during the winter months but negative during the summer months. During the warmest and wettest months of the year, the daytime bias is between about 3% and 6%, and the nighttime bias is between −4% and −10%.

Figure 7.

Monthly mean percentage difference between AIRS and MWR PWV over SGP. Positive values indicate that the AIRS PWV exceeds the MWR PWV. Error bars indicate the uncertainty in the AIRS-MWR difference at the 95% confidence level.

[29] Because of the relatively large number of cases obtained in this 6 year time period, we are also able to characterize the AIRS v5 PWV product as a function of total water vapor column amount. Figure 8 shows the percentage difference between AIRS and the MWR in PWV bins between 0 and 5.5 cm. Note that the daytime percentage difference with respect to MWR is within the AIRS science team recommended ±5% range for total water vapor amounts greater than 1 cm at the SGP site but exceeds 5% for values below 1 cm. The nighttime percentage difference exceeds +5% for PWV amounts less than 1 cm but shows a significant negative bias for PWV amounts greater than 2 cm. This is consistent with the previous monthly analysis because the winter months at the SGP site are near 1 cm PWV or less while the summer months are typically 2 cm of PWV or higher.

Figure 8.

Percentage difference between AIRS and MWR (blue) and AIRS and GPS (red) for (top) daytime and (bottom) nighttime over SGP calculated for 0.5 cm bins. The error bars shown are the uncertainty in the bias for each bin at the 95% confidence level. Dashed lines are the AIRS science team suggested product accuracy of ±5%.

[30] Also shown in Figure 8 are AIRS comparisons to PWV derived from a GPS receiver at the SGP site. These GPS measurements are coincident in time and space with the AIRS and MWR matchups so they can be compared directly. Since the same AIRS data are used in both comparisons, the difference between AIRS and MWR and AIRS and GPS can be interpreted as the uncertainty in the validation data. The accuracy of the ARM MWR PWV is better known than that of the GPS PWV, but the GPS validation at the SGP site does confirm the AIRS and MWR results. This reinforces the conclusion that the differences seen between the AIRS and MWR PWV are due to errors in the AIRS retrievals and not in the ground-based validation data.

4.2. TWP

[31] The ARM TWP site on the small equatorial island of Nauru provides a unique validation location for satellite retrievals of water vapor because the surrounding ocean surface provides a well-known surface emissivity background and the variability of water vapor (and temperature) is small, which reduces time and space sampling errors. During the time period examined, a total of 556 AIRS overpasses of the TWP site met the criteria given in Table 1. Of these, 308 were daytime observations and 248 were nighttime observations. Table 2 lists the bulk comparison statistics between AIRS and MWR at the TWP site. The AIRS PWV values show a near-zero bias relative to the ARM MWR PWV during both the daytime and the nighttime at this site. The linear fits of both the daytime and nighttime observations are similar and exhibit a significant offset. However, due to the climatology of the Nauru site, many of the MWR observations are clustered in a narrow PWV range (between 4 and 6 cm), making a linear fit less useful for determining the accuracy of the satellite retrieval.

[32] Given the maritime tropical location of the TWP observation site, the range of PWV values is small throughout the year. Both the daytime and nighttime monthly mean PWV values were between about 4.4 and 5.25 cm for all months. The monthly mean diurnal differences for AIRS and MWR were examined by averaging individually matched daytime and nighttime observations that occurred within the same 13 h period. The results confirmed that the diurnal difference in PWV in the tropics is very small throughout the year (between ±0.35) and that the AIRS and MWR reported diurnal differences agreed to within the calculated uncertainties for all months. Figure 9 shows the monthly mean percentage difference between AIRS and MWR PWV for daytime and nighttime overpasses. The AIRS retrieval exhibits a slight dry bias in the daytime PWV for some months of the year at the level of about 2% but with a statistical uncertainty of nearly the same magnitude so firm conclusions cannot be drawn with the limited statistics. At night, AIRS agrees with the MWR within the statistical uncertainty for all months except November, which shows a small moist bias. Since the number of overpasses of an equatorial site is relatively low, a refined analysis of diurnal effects in PWV at the TWP site will require additional years of observation. The AIRS minus MWR PWV average over all months at the TWP site is better than 3%, which indicates excellent agreement.

Figure 9.

Monthly mean percentage difference between AIRS and MWR PWV over TWP. Positive values indicate that the AIRS PWV exceeds the MWR PWV. Error bars indicate the uncertainty in the AIRS-MWR difference at the 95% confidence level.

4.3. NSA

[33] Satellite infrared retrievals of water vapor in the Arctic can be challenging, but for that reason accurate validation is even more important. The ARM NSA site near the northernmost point of the United States at Barrow, Alaska, is in a particularly good location for validation in winter because the ocean and tundra are frozen; however, in the summer, the surface is a mixture of open ocean, floating ice pack, and warm tundra. The site is also often obscured by mixed phase clouds or fog so the test of MWR liquid water content is particularly important for the NSA site. Because of the high solar zenith angle in the arctic, differentiation between day/night can be ambiguous. For this study, we have defined observations with solar zenith angles less than 90° as “day” and greater than or equal to 90° as “night.” This means that there are many more daytime measurements in the warmer summer months and more nighttime measurements during the Arctic winter. For the purposes of this study, the comparison was restricted to MWR PWV amounts less than or equal to 2 cm. This is because there were a large number of outliers in the MWR data with PWV amounts much greater than 2 cm that were attributed to MWR measurement artifacts, possibly due to ice buildup on the sensor for brief periods. Note that the cutoff of 2 cm is within two standard deviations of the 6 year monthly mean MWR PWV for all months of the year and thus limiting the comparison to PWV amounts below 2 cm is an effective way to eliminate the MWR outliers. The two-channel MWR at the NSA site used for this validation study is known to have poor signal to noise for extremely dry conditions; however, the relatively large number of samples obtained in the Arctic due to multiple satellite overpasses each day partially compensates for this degraded performance [Turner et al., 2007].

[34] During the time period examined, a total of 3219 AIRS overpasses of the NSA site met the criteria given in Table 1. Of these, 1875 were daytime observations and 1344 were nighttime observations. The bulk comparison statistics between AIRS and MWR over the NSA site are given in Table 2. The AIRS data show a near-zero bias during both the daytime and the nighttime, and the bulk statistics suggest excellent agreement between the two instruments in a mean sense. The linear fit slope and offset are included for reference although the range of PWV values is limited.

[35] The monthly mean daytime and nighttime PWV over NSA ranges from between 1.2 and 1.4 cm in the warmer months of the year (June–August) to extremely low values (between 0.25 and 0.3 cm) in the coldest months (December–March). Both the AIRS and the MWR capture this seasonal variation in PWV well in an average sense. It is not as useful to define a diurnal signal in the Arctic because the daytime and nighttime cases vary so greatly with season so we do not include a figure for the NSA similar to Figure 5. However, we can compute daytime and nighttime statistics separately by month. Figure 10 shows the differences between AIRS and MWR in percent. Note the wide variation in the number of daytime (or nighttime) cases by month of the year. AIRS exhibits a dry bias at NSA during the daytime from September through May, though the wintertime months carry a large uncertainty due to the very small number of daytime observations. During the summer (June, July, and August), AIRS has a slight moist bias of around 2%–3% during the daytime. AIRS exhibits a similar tendency at night, with a dry bias between October and April, and a moist bias during August.

Figure 10.

Monthly mean percentage difference between AIRS and MWR PWV over NSA. Positive values indicate that the AIRS PWV exceeds the MWR PWV. Error bars indicate the uncertainty in the AIRS-MWR difference at the 95% confidence level.

4.4. Comparisons at All Sites

[36] Each of the three ARM sites used in this study is unique with respect to weather and climate, annual range of water vapor amount, and ocean/land surface characteristics. Considering the three sites together gives an idea of how AIRS performs under a variety of meteorological conditions but with the understanding that these three ARM sites do not represent all geographic locations for which the AIRS data will be used. Figure 11 shows a scatterplot of the MWR versus AIRS PWV for all three sites together for daytime (left) and nighttime (right). The linear fit for the daytime (nighttime) data has a slope of 0.99 (0.97) and an intercept of 0.014 (0.012). This suggests that the AIRS retrieval performs reliably across the entire range of PWV values included in this study (near 0 cm to just over 6 cm). Figure 12 (top) shows a scatterplot of the AIRS to MWR PWV ratio versus the MWR PWV for SGP (976 points, shown in red), TWP (556 points, shown in blue), and NSA (3219 points, shown in green). The largest ratios (just under 2) are for very dry cases (<1 cm), which is consistent with the results presented in the work of Tobin et al. [2006]. The ratios at SGP and TWP are almost all between about 0.75 and 1.25, with the AIRS retrieval performing particularly well over the tropical ocean. The middle plot shows the same set of data; however, the y axis shows the percentage difference between the AIRS and MWR PWV. Figure 12 (bottom) shows a comparison of 5421 coincident MWR and Vaisala RS-92 rawinsonde PWV observations over NSA (1596 points), SGP (2348 points), and TWP (1596 points) for the same 6 year time period. The rawinsondes were quality controlled visually, and sondes with bad RH profiles, broken temperature sensors, or profiles that truncated prematurely were removed from the analysis. These sondes are time and space coincident with the MWR but not with the AIRS overpasses. The percentage difference between the MWR and the ARM rawinsonde observations is within about ±10% for almost the entire range of PWV values. This shows that the scatter in the percentage differences of AIRS minus MWR shown in Figure 12 (middle) are mainly due to errors in individual AIRS retrievals.

Figure 11.

Scatter plot of MWR PWV versus AIRS PWV for three sites. Daytime is shown to the left, and nighttime is on the right. Green is NSA (Barrow, AK), red is SGP (Lamont, OK), and blue is TWP (Nauru Island).

Figure 12.

All plots show all PWV data from both day and night, colored according to site (red, SGP; blue, TWP; green, NSA). (top) Ratio of spatially and temporally colocated AIRS to MWR observations of PWV. (middle) Percentage difference between spatially and temporally colocated AIRS and MWR observations of PWV. (bottom) Percentage difference between spatially and temporally colocated MWR and rawinsode observations of PWV.

[37] In order to quantify the accuracy of the AIRS PWV product, it is useful to examine biases separately for each site as a function of daytime and nighttime and PWV amount. Table 3 lists the AIRS minus MWR PWV bias and standard deviation for all sites, calculated for 0.5 cm PWV bins between 0 and 6.5 cm. Statistics were calculated for bins with ≥4 observations. Figure 13 shows the same bias values, converted to percent, and plotted against the median MWR PWV within each bin. Daytime (top) and nighttime (bottom) overpasses are shown separately. Results from the SGP, TWP, and NSA sites are plotted in red, blue, and green, respectively. Error bars are the uncertainty in the mean of AIRS minus MWR PWV differences at the 95% confidence level for a given bin. The larger error bars indicate either a very small number of data points or a high standard deviation within a given bin or both. At the NSA and TWP sites, both the daytime and nighttime bias errors are close to the AIRS science team suggested range of ±5% for all PWV bins. At the SGP site, the bias error falls within 5% for daytime observations greater than about 1 cm. Both the daytime and nighttime AIRS retrievals at low PWV (<1 cm) show a moist bias of 15%–30%, while the nighttime AIRS retrievals at PWV values above about 2 cm show a significant dry bias of about 10%.

Figure 13.

Percentage difference between AIRS and MWR PWV for 0.5 cm bins, (middle) for daytime data only and (bottom) nighttime data only. Data from SGP, TWP, and NSA are plotted in red, blue, and green, respectively. Error bars are the uncertainty in the mean percentage difference for a given bin (at the 95% confidence level).

Table 3. AIRS Minus MWR PWV Bias and Standard Deviation Within 0.5 cm PWV Bins for Daytime and Nighttime matchups at Each ARM Validation Sitea
PWV Range (cm)NSA Day (cm)NSA Night (cm)SGP Day (cm)SGP Night (cm)TWP Day (cm)TWP Night (cm)
  • a

    Statistics are for the time period 1 September 2002 through 31 August 2008.

0.00.50.0045 ± 0.0980.0014 ± 0.0830.082 ± 0.0750.25 ± 0.16  
0.51.0−0.0026 ± 0.19−0.047 ± 0.160.079 ± 0.100.17 ± 0.20  
1.01.50.019 ±0.230.0049 ± 0.210.049 ± 0.130.15 ± 0.18  
1.52.0−0.12 ±0.25−0.039 ± 0.210.023 ± 0.130.071 ± 0.29  
2.02.5  0.067 ± 0.33−0.11 ± 0.27  
2.53.0  0.057 ± 0.33−0.24 ± 0.26  
3.03.5  0.19 ± 0.45−0.31 ± 0.34−0.13 ±0.220.13 ± 0.16
3.54.0  0.11 ± 0.44−0.41 ± 0.29−0.012 ± 0.190.11 ± 0.18
4.04.5  0.015 ± 0.43−0.42 ± 0.34−0.0033 ± 0.220.085 ± 0.17
4.55.0  −0.0083 ± 0.54−0.38 ± 0.400.026 ± 0.210.062 ± 0.20
5.05.5   −0.82 ± 0.26−0.021 ± 0.220.0096 ± 0.17
5.56.0    −0.16 ± 0.25−0.0086 ± 0.20
6.06.5    −0.18 ± 0.30−0.18 ± 0.17

5. Conclusions

[38] Satellite retrievals of PWV from AIRS over three climatologically diverse sites were compared with ground-based MWR data. Each location exhibited unique retrieval challenges, indicating that geographically specific considerations, such as a good characterization of the surface emissivity, may be important for accurate satellite retrievals. The following conclusions may be drawn from this comparison.

[39] 1. Southern Great Plains land site: The AIRS daytime PWV is determined to be within the absolute accuracy goal of 5% for PWV amounts greater than 1 cm, but AIRS exhibits a moist bias in excess of 15% for low PWV amounts (<1 cm) for both daytime and nighttime overpasses. Additionally, the AIRS retrievals have a significant nighttime dry bias of about 10% with respect to the MWR at PWV amounts greater than about 2 cm. The nighttime dry bias is largest during the summer months of the year when water amounts are highest. Preliminary investigation suggests that the nighttime dry bias exists on a significant spatial scale over the U. S. Great Plains and desert Southwest but not in the Eastern United States or Canada. The moist bias for low water amounts (observed during the winter) at the SGP site could be a surface emissivity related error since no bias is seen at the ARM Arctic site for similar water amounts.

[40] 2. Tropical Western Pacific ocean site: The AIRS retrievals over the tropical ocean are within the AIRS science team suggested error of 5% for the entire range of PWV values observed in this study: between 3 and 6.5 cm.

[41] 3. North Slope of Alaska Arctic site: The AIRS retrievals of PWV from Barrow show a near-zero bias for PWV values less than 2 cm and a fractional error of the bias that is within the 5% accuracy goal of the AIRS science team. However, it is important to note that the RMS error in this is still large, and individual samples can have errors of 50% or higher for very dry cases.

[42] In summary, the AIRS PWV Version 5 product is very useful for climate process studies over both ocean and land, which require knowledge of the total water vapor column with an accuracy of about 5% in monthly mean values; however, there are two main caveats. The product accuracy is degraded for (1) low water amounts over mixed bare soil and vegetation and (2) nighttime only measurements at moderate water amounts in the U. S. Southern Great Plains. The use of the AIRS version 5 level 3 PWV product for evaluation of monthly mean diurnal changes in water vapor has been shown to contain bias errors greater than the natural signal observed with ground-based sensors. In particular, the nighttime version 5 level 3 PWV product should used with caution for climate studies due to a possible underestimate in the total water vapor column over certain land types. This issue deserves further investigation. The authors anticipate that future versions of the AIRS PWV product will provide improvements to the retrievals over land in general and the PWV product in particular. The ability of the AIRS PWV product to detect significant climate trends in water vapor over ocean or land will be the subject of future work as it requires the characterization of possible time dependent bias errors not addressed in this paper.

Ancillary