Comparing upper tropospheric humidity data from microwave satellite instruments and tropical radiosondes



[1] Atmospheric humidity plays an important role in the Earth's climate. Microwave satellite data provide valuable humidity observations in the upper troposphere with global coverage. In this study, we compare upper tropospheric humidity (UTH) retrieved from the Advanced Microwave Sounding Unit and the Microwave Humidity Sounder against radiosonde data measured at four of the central facilities of the Atmospheric Radiation Measurement program. The Atmospheric Radiative Transfer Simulator (ARTS) was used to simulate satellite brightness temperatures from the radiosonde profiles. Strong ice clouds were filtered out, as their influence on microwave measurements leads to incorrect UTH values. Day and night radiosonde profiles were analyzed separately to take into account the radiosonde radiation bias. The comparison between radiosonde and satellite is most meaningful for data in cloud-free, nighttime conditions and with a time difference of less than 2 hr. We found good agreement between the two data sets. The satellite data were slightly moister than the radiosonde data, with a mean difference of 1%–2.3% relative humidity (RH), depending on the radiosonde site. Monthly gridded data were also compared and showed a slightly larger mean difference of up to 3.3% RH, which can be explained by sampling issues.

1. Introduction

[2] Atmospheric water vapor is the main natural greenhouse gas and enhances the sensitivity of the climate to external forcings by about 70%, hence playing an important role in climate change prediction [Cess et al., 1990]. Global climate models that contain water vapor feedback with an approximately constant relative humidity in a warming climate predict that over the next century the Earth's surface will warm nearly twice as much as models that do not contain water vapor feedback [Minschwaner and Dessler, 2004]. The middle and upper troposphere contribute strongly to the water vapor feedback, especially in the tropics, owing to the low temperatures at these altitudes.

[3] The water vapor feedback mechanism in the upper troposphere appears to be in reasonable agreement both between different climate models [e.g., John and Soden, 2007] and between models and observations [e.g., Gettelman and Fu, 2008]. However, as shown, for example, by John and Soden [2007], the absolute values of upper tropospheric humidity (UTH) disagree significantly between different models. Although a number of studies have analyzed the distribution of water vapor in the upper troposphere [e.g., Gettelman et al., 2006; Gettelman and Fu, 2008; Soden et al., 2005; Buehler et al., 2008], large uncertainties remain in the available UTH data sets [de F. Forster and Collins, 2004].

[4] Two main sources of UTH information are radiosondes and satellite observations. The radiosonde data have been available since the 1940s from synoptic stations, but the data quality in the upper troposphere is questionable [Elliott and Gaffen, 1991]. In contrast to radiosonde measurements, satellite data have global coverage but these observations are available only since 1979 from infrared (IR) sounders (6.7 μm band) and since 1994 from microwave sounders (183.31 GHz). The IR channels are heavily influenced by clouds, while microwave measurements are less sensitive to clouds [Buehler et al., 2007].

[5] The first attempt to relate infrared radiances (Tb) to UTH was made by Soden and Bretherton [1993]. They presented a log linear relation (see equation (1)) between the brightness temperature of the 6.7 μm band from Geostationary Operational Environmental Satellite 7 (GOES-7) and Jacobian weighted relative humidity in the upper troposphere (approximately 500–200 hPa). They stated that the accuracy is within ±1K or ±10% relative humidity (RH). Buehler and John [2005] adapted this brightness temperature transformation method to Advanced Microwave Sounding Unit B (AMSU-B) data to retrieve Jacobian weighted UTH. A global UTH data set, derived from radiances of the AMSU-B instruments onboard the satellites National Oceanic and Atmospheric Administration (NOAA) 15, 16, and 17, hereafter referred to as N15, N16, and N17, respectively, is described in Buehler et al. [2008] (an updated version of the data set is available at In addition, UTH data are also available from several research satellite observations, for example, the Atmospheric Infrared Sounder [Gettelman et al., 2006]; the Microwave Limb Sounder (MLS) [Livesey et al., 2006]; the Michelson Interferometer for Passive Atmospheric Sounding [Ekström et al., 2008]; solar occultation limb sounders, for example, the Halogen Occultation Experiment [Harries et al., 1996]; and aircraft measurement by the Measurement of Ozone and Water Vapor by Airbus In-Service Aircraft project [Luo et al., 2007].

[6] AMSU data have previously been compared to European radiosonde data [e.g., Buehler and John, 2005; John and Buehler, 2005; Buehler et al., 2008], but the focus has never been in the tropics. UTH in the tropics is of large importance to the Earth's climate system. In light of this, in this study we use data from the Atmospheric Radiation Measurement (ARM) program [Stokes and Schwarts, 1994] measured at the Tropical Western Pacific (TWP) stations, located near the equator, and the subtropical Southern Great Plains (SGP) station. The measurements are made using high-quality modern sensors with good calibration and good vertical resolution. The aim of this study is to evaluate the capability of microwave satellite data to retrieve upper tropospheric humidity in the tropics. The midlatitude station, SGP, is used as a reference station to compare the quality of satellite UTH data in tropical and midlatitude regions. Another aim of this study is to investigate the ability of satellite data to quantify the radiosonde daytime solar radiation dry bias.

[7] The structure of the article is as follows: Section 2 provides information about the satellite and radiosonde data, focusing on the data properties relevant for the comparison. Section 3 presents the radiative transfer (RT) model, the UTH retrieval from satellite data, the cloud filter used, and the method used to collocate satellite and radiosonde data. Section 4 investigates and explains the results of the comparisons of the individual data and monthly mean data respectively, and section 5 presents the conclusions.

2. Measurements

2.1. Satellite Data

[8] The AMSU-B is a cross-track scanning five-channel (16–20; channels 1–15 are of AMSU-A) microwave radiometer. The AMSU-B channels operate at 89.0, 150.0, 183.31 ± 1.00, 183.31 ± 3.00, and 183.31 ± 7.00 GHz. The instrument has a swath width of approximately 2300 km with 90 scan positions. The satellite viewing angle is 0.55° from nadir for the innermost scan positions and 48.95° from nadir for the outermost scan positions. This corresponds to incidence angles of 0.62° and 58.5° from nadir, respectively. The AMSU-B footprint size, defined with respect to the half power beam width, is approximately 15 km at nadir but increases toward the edge of the scan [Saunders et al., 1995; Goodrum et al., 2007].

[9] The AMSU-B instruments are onboard the NOAA-15, -16, and -17 satellites. The NOAA-18 and -19 and Metop-A satellites, hereafter referred to as N18, N19, and MA, respectively, have the Microwave Humidity Sounder (MHS) onboard instead of AMSU-B. MHS is very similar to the AMSU-B, but the second channel has moved to 157.0 GHz and the fifth channel has only one passband at 190.311 GHz [Bonsignori, 2007; Goodrum et al., 2007]. In this study AMSU-B data from all available satellites, N15, N16, and N17, and MHS data from N18 and MA are used.

2.2. ARM Radiosonde Data

[10] The ARM program is a global change research program supported by the U.S. Department of Energy since 1989. The primary goal of the ARM program is to improve the understanding of the fundamental physics of the interaction between clouds and radiative feedback processes in the atmosphere. The ARM stations have been selected so models and parametrizations can be tested over a broad range of conditions to give confidence in their general applicability [Stokes and Schwarts, 1994]. A typical radiosonde provides vertical profiles of both the thermodynamic state of the atmosphere and the wind speed and direction. Radiosonde data of ARM include moisture, pressure, temperature, and horizontal wind.

[11] Radiosonde data from the Tropical Western Pacific (TWP-C1, -C2, and -C3) and Southern Great Plains (SGP-C1) stations are used in this study. These are located at Manus Island (Papua New Guinea), Nauru Island (the Republic of Nauru), Darwin (Australia), and Lamont (Oklahoma, USA), respectively. The station details are shown in Table 1 and Figure 1. The TWP stations, especially TWP-C1 and TWP-C2, are particularly close to the equator and are located in a tropical convective zone.

Figure 1.

Location of the stations superimposed on the mean UTH map retrieved from N17 AMSU data for the period 2002–2009. T1, T2, T3, and S1 stand for TWP-C1, TWP-C2, TWP-C3, and SGP-C1 stations, respectively.

Table 1. Geographical Attributes of the Studied Stationsa
StationAbbr.Latitude (°)Longitude (°)Altitude (m)Operating since YYMMUTH (%RH)σ (%RH)
  • a

    The last two columns show mean and standard deviation of UTH for the nighttime profiles of the period 2004–2008.

Southern Great Planes,Central Facility, Lamont, USASGP-C136.61−97.49315001034.2814.58
Tropical Western Pacific, Central Facility, Manus Island, PNGTWP-C1−2.06147.434970844.1712.39
Tropical Western Pacific, Central Facility, Nauru IslandTWP-C2−0.52166.927990337.3013.74
Tropical Western Pacific, Central Facility, Darwin, North AustraliaTWP-C3−12.42130.8830020425.7217.48

[12] Table 2 shows details of the radiosonde sensors, including the time period used in this study. Radiosonde data used in this study are mostly from Vaisala RS92 sensors. RS92 has two humidity sensors, where one is heated while the other measures. This eliminates condensation on the sensors when inside clouds or due to prolonged exposure to super-saturation. However, RS92 radiosondes suffer from a radiation dry bias when exposed to sunlight [Vömel et al., 2007]. This is believed to be caused by their lack of a radiation shield and by their large surface area due to the dual sensors.

Table 2. Radiosonde Sensors That Have Been Used to Measure Humidity at Different Stationsa
  • a

    Time periods are given in YYMM.

SGP C10010–01040105–05020502–0812
TWP C19708–02050206–05030504–0812
TWP C29903–02050206–05030504–0812
TWP C30204–0601 0601–0812

[13] In the lower troposphere the daytime dry bias was estimated to be 3%–4% for RS80 [Turner et al., 2003], 6%–8% for RS92 [Cady-Pereira et al., 2008], and 6%–8% for RS90/92 [Miloshevich et al. 2006] based on comparison to ground-based measurements of precipitable water vapor by microwave radiometers. Vömel et al. [2007] found the RS92 daytime bias to increase with altitude to up to 50% near 16 km. For the altitude region that channel 18 of AMSU-B is sensitive to (5–10 km), the dry bias ranges from 10% to 30%.

3. Methodology

[14] We use the profile-to-satellite approach, described for example in Soden and Lanzante [1996] and in Buehler et al. [2004], where satellite radiances are simulated from radiosonde profiles. The same transformation method is then applied to both measured and simulated radiances to convert the radiances to UTH. This approach is better than a direct comparison of radiosonde and satellite retrieved water vapor profiles, because it avoids the errors that are associated with the inverse problem for the satellite data. A complete error model for this kind of comparison was developed in Buehler et al. [2004] and is discussed in detail there. The error terms to consider are (1) radiometric noise of the AMSU measurement, (2) sampling error due to atmospheric inhomogeneity, (3) radiosonde measurement error in humidity and temperature, (4) RT model error (systematic), and (5) AMSU calibration error. From a statistical analysis of the error distribution, Buehler et al. [2004] estimated the combined error of terms 3, 4, and 5 to be less than 0.5K, which corresponds approximately to 3% RH in UTH.

[15] The profile-to-satellite approach has the desired consequence that errors in the UTH retrieval method cancel out. Such errors are therefore ignored in this study. We do not see this as a problem, since errors in the UTH retrieval method itself can be estimated accurately from simple model simulations and are well documented in Buehler and John [2005].

3.1. ARTS Model

[16] The Atmospheric Radiative Transfer Simulator (ARTS) [Buehler et al., 2005a] was used to simulate AMSU/MHS radiances from radiosonde profiles. The ARTS setup for AMSU/MHS simulation has been discussed by Buehler et al. [2004].

[17] Radiosonde data above 100 hPa are discarded to maintain a homogeneous vertical extent. This introduces a bias of approximately 0.033–0.090 K, depending on the channel, with a small random error of 0.006–0.057 K [Buehler et al., 2004]. Radiosonde profiles that do not contain data up to 100 hPa are not used in the study. This filters out about 5% of the SGP-C1 profiles, 8% of the TWP-C1 and -C2 profiles, and 4% of the TWP-C3 profiles.

[18] The vertical profiles of air pressure, temperature, and water vapor volume mixing ratio are used as input to the RT model ARTS. The ARM radiosonde data do not include ozone concentrations, therefore radiances are calculated without the impact of ozone. The impact of this omission on AMSU radiances is estimated to be largest for channel 18, with a maximum impact of 0.5 K [John and Buehler, 2004].

[19] We estimate that 0.5 K is also a reasonable approximation of the overall error of the RT model for this AMSU channel. The channel is close to the center of the 183.31 GHz water vapor line, so uncertain continuum absorption parameters play only a very small role. To cause significant radiance differences, quite large perturbations in the spectroscopic line parameters are necessary, which are unlikely given the well-established use of this line for atmospheric measurements. The same water vapor line was, for example, used by the MLS limb sounder satellite missions, where a large error in the line parameters would have been noticed (see Pumphrey and Buehler [2000]). The RT model ARTS itself has been validated against various other RT models [Melsheimer et al., 2005; Buehler et al., 2006], which leads to a rough estimate of the pure RT error (not including spectroscopic parameters) of less than 0.2 K.

3.2. Retrieving UTH from AMSU/MHS Data

[20] Following Buehler and John [2005], the log-linear relationship shown below is used to estimate UTH from the 183.31 ± 1.00 GHz radiances:

equation image

where Tb is the radiance expressed in brightness temperature and a and b are linear fit coefficients that are available separately for each viewing angle (see Table 1 in Buehler and John [2005]) so radiances do not need to be limb corrected [John et al., 2006].

3.3. Colocating Satellite and Radiosonde Data

[21] A radiosonde balloon typically drifts about 50 km horizontally while ascending from the ground to 100 hPa. Hence the average Tb of a target area, rather than that of an individual satellite ground pixel, is compared to the radiosonde simulated radiance [Buehler et al., 2004]. The target area is defined by a circle of radius 50 km centered at the launch site, which normally encompasses around 10–30 pixels. The pixels inside the target area have different viewing angles, and therefore the radiosonde brightness temperature is also simulated for the corresponding viewing angles. On average, a radiosonde takes about 20 min to reach 500 hPa and about 45–60 min to reach 200 hPa. Therefore the radiosonde time used to collocate with satellite data is defined as its launch time plus 45 min. The displacement of the air mass during this time is also considered. Radiosonde wind data between 700 and 300 hPa, the altitude range most important for the humidity channels, are used to calculate the average wind vector. The average wind vector is multiplied by the time difference between the satellite overpass and radiosonde time. If the calculated displacement is larger than 50 km, the data are excluded [Buehler et al., 2004]. In addition, the acceptable time difference between satellite overpasses and radiosonde times is limited to 2 hr. The number of clear-sky nighttime collocations for different stations and different years is given in Table 3.

Table 3. Number of Available Data Points per Year for Night Dataa
  • a

    The matches with more than 50 data points are used for the conclusions. The combinations with fewer than 50 data points, in total, are not shown.

SGP C1N15111917265940154
SGP C1N16131  3288143187
SGP C1N17 366090101604811406
TWP C1N17 367842133112311
TWP C1MA     14150113277
TWP C2N17  5442713198269
TWP C2MA     67748131
TWP C2N18    35213059
TWP C3N17   86597973225
TWP C3MA     1812499241

3.4. Cloud Filtering Method

[22] Microwave radiances are less sensitive to clouds than infrared radiances, but sufficiently optically thick ice clouds can affect microwave data. A cloud filter introduced by Buehler et al. [2007] was used to exclude cloud affected pixels from AMSU/MHS data. The cloud filter works as follows. First, channel 18 of AMSU/MHS is more sensitive to higher altitudes of the troposphere than channel 20. In clear-sky conditions, because of the natural lapse rate of air temperature, the brightness temperature of channel 18 (TB18) is lower than the brightness temperature of channel 20 (TB20). Ice clouds scatter outgoing radiation and reduce TB20 more strongly than TB18. Therefore, in the presence of an ice cloud ΔTB = TB20TB18, which is positive in clear-sky conditions, might become negative. Second, the cloud also reduces the value of TB18 directly, so a viewing-angle-dependent threshold, Tthr(θ), was utilized. In summary, the conditions for clear-sky data are ΔTB > 0 and TB18 > Tthr(θ). Data not fulfilling either of these conditions are considered cloudy. Values of Tthr for different viewing angles are given in Buehler et al. [2007]. The performance of this method for tropical regions will be discussed in section 4.1.

4. Results and Discussion

[23] Figure 2 shows scatterplots of collocated N17, AMSU-B channel 18 Tb data versus TWP-C1 simulated brightness temperatures for this channel. This figure includes separate plots for nighttime all skies (Figure 2a) and nighttime cloud-free (Figure 2b) and daytime cloud free (Figure 2c) data. Each data point corresponds to the average brightness temperature in the target area as discussed in section 3.3. Vertical lines represent the standard deviations of AMSU-B brightness temperatures inside the target area. These represent the inhomogeneity of the atmosphere observed by the satellite. This quantity is used to weight each collocation for calculating statistics, such as bias and linear fit parameters, where a high standard deviation corresponds to a small weight (see Appendix A for details).

Figure 2.

Scatterplots of N17 AMSU Tb versus radiosonde Tb for the entire TWP-C1 data set. Vertical lines show the standard deviation of each data point given by the standard deviation of pixels inside the target area. The dashed and solid lines are diagonal and regression lines, respectively.

4.1. Cloud Filter in the Tropics

[24] The performance of the cloud filter in the tropics was examined using nighttime data only. AMSU/MHS TB18 observed under cloudy skies are lower than those recorded in cloud-free conditions and are much lower than the corresponding ARTS-simulated brightness temperatures. In Figure 2a, cloudy data points are located at the bottom left section of the plot. The cloud filtered data are shown in Figure 2b. As the figure demonstrates, the cloud filter indeed removes most outliers and greatly improves the agreement between satellite and radiosonde. The cloud filter also performed equally well in comparisons made for all other radiosonde-satellite pairs (not shown).

4.2. Solar Radiation Dry Bias

[25] Partitioning the radiosonde data into day and night profiles allows us to investigate the daytime dry bias that particularly the RS92 sondes are expected to have due to solar radiation. We define nighttime and daytime as follows: Nighttime refers to the period between 1 hr after sunset and 1 hr before sunrise, and daytime to between 1 hr after sunrise and 1 hr before sunset. For nighttime collocations, AMSU/MHS data may be within ±1 hr of sunrise or sunset as long as the radiosonde time is at night. This is reasonable as the satellite measurements are not influenced by solar radiation. Averaged collocation time for each satellite-station pair is shown in Figure 3 (symbols denote the average collocation time and vertical bars its standard deviation).

Figure 3.

Average collocation time for different satellite-station pairs. Different stations are coded using different markers. Filled and unfilled markers stand for daytime and nighttime, respectively.

[26] Figures 2b and 2c show nighttime and daytime comparisons, respectively. For the TWP-C1/N17 pair, the collocations take place around 10:00 LT or 22:00 LT. The relative bias of the radiosonde data, in humidity space, compared to satellite measurement is 8.3% at night and 24.9% during the day. This corresponds to about 16.6% dry bias caused by solar heating.

[27] Radiation dry bias and the difference between day and night biases for all satellite-station pairs are given in the last columns of Table 5. The radiation dry bias is 5%–20%, which is consistent with the dry bias reported in Vömel et al. [2007]. The SGP-C1/N15, TWP-C3/N17, and TWP-C3/MA combinations show a smaller radiation dry bias, about 5%–8%. This may be explained as the day-night differences are not substantial for these combinations because collocations occur close to sunrise-sunset, as can be inferred from Figure 3. This is in agreement with a smaller dry bias (maximum of 4%) estimated by John and Buehler [2005] for radiosonde data for 2001–2003 from European stations using N15 (morning/evening collocations) and a larger dry bias using N16 (noon/midnight collocations) that was on average 9.6% for RS80 and 11.5% for RS90 sensors. The results are also consistent with the findings of Miloshevich et al. [2009] that RS92 day time biases are from 5% to 15% in the 500–200 hPa layer.

4.3. One-to-One Comparison

[28] For the analysis presented in this section we use only cloud-free nighttime data. The number of collocations per year for each station-satellite pair is reported in Table 3. We use mainly the RS92 data (compare Table 2) for our analysis, but RS90 data for the N17/SGP-C1 combination are also quoted in Tables 4 and 5 for comparison. Combinations of 50 data points or more are used for the analysis. We describe the comparison with the help of statistical parameters, including bias, standard deviation of the bias (σbias), slope, root-mean-square deviation (RMSD), and correlation coefficient. They are shown in Table 4 (radiance space) and in Table 5 (humidity space) for all available station-satellite combinations. Further details on calculating these quantities are given in Appendix A and in Buehler et al. [2004].

Table 4. Statistical Parameters for AMSU/MHS Tb Minus Radiosonde Tba
  • a

    The bias, standard deviation of the bias (σbias), and RMSD are in Kelvin. The columns D240 and D270 show the distance between the regression line and the diagonal line, in Kelvin, at 240 and 270K sonde brightness temperature. N is the number of data points used to calculate the statistics and r is the correlation coefficient. The parameter tscore is the t score of a paired t test. The tscore limit is around ±1.97. For SGP-C1/N17 the first line shows data from RS92 and the second line data from RS90 (all other data are RS92). The last rows show statistics calculated using RS92 data from all the stations but for different satellites.

SGP-C1 RS90N170.900.97−0.26−0.100.66−
Table 5. Statistical Parameters for AMSU/MHS UTH Minus Radiosonde UTHa
  • a

    Bias, standard deviation of the bias (σbias), and RMSD are in %RH. Subscript d stands for daytime data; all other data are nighttime. N shows the number of data points and Δ B = % Bias − % Biasd. The columns D0% and D60% show the distance between the regression line and the diagonal line, in %RH, at 0%RH and 60%RH sonde UTH values. N is the number of data points used to calculate the statistics and r is the correlation coefficient. For SGP-C1/N17 the first line shows data from RS92 and the second line data from RS90 (all other data are RS92). The last rows show statistics calculated using RS92 data from all the stations but for different satellites.

SGP-C1 RS90N170.910.960.604.452.68−2.780.173.081903.4418.5149814.06

[29] All station-satellite combinations show a negative bias in Table 4 and a positive bias in Table 5. In other words, the satellite UTH is higher than the radiosonde UTH. This is consistent with earlier studies [e.g., Buehler et al., 2008; John and Buehler, 2005; Buehler and John 2005]. The bias is between 0.6 and −1.6 K in radiance space and between 1.1% and 2.4% RH in humidity space (5.1%–15.4% relative difference). For the SGP-C1/N17 combination, RS90 bias (4.5%) is less than RS92 bias (7.5%) in humidity space. This is not consistent with Miloshevich et al. [2006]. However, they reported that the accuracy of the RS92 data in the middle or upper troposphere, for temperatures between −20°C and −50°C, is about 10% RH (see Table 3 of Miloshevich et al. [2006]. Therefore, the difference between RS90 and RS92 bias is still less than the accuracy of the radiosonde UTH. TWP-C3 shows a larger bias (−1.6 and −1.5K in radiance space) compared to the other stations. This is due to the atmospheric conditions, as the average UTH at this station is very low (25% RH) compared to the average UTH at the other stations. Table 4 contains two additional bias parameters, D240 and D270, to make this last point clearer. These are the differences of the linear fit from the diagonal at 240 K and 270 K in the radiance space scatterplot. In contrast to the traditional bias (defined as the weighted mean of the difference) these bias parameters do not depend on the different mean conditions at the different stations. They show that TWP-C3 is indeed consistent with the other stations. Similar bias parameters D0% and D60% were added to the humidity space data in Table 5.

[30] In humidity space, the slopes are close to unity and range between 0.94 to 1.05. In dry conditions, satellite UTH is higher than radiosonde UTH for all cases. This is consistent with the results from Buehler et al. [2004] and John and Buehler [2005]. During moist conditions, satellite UTH is sometimes higher than radiosonde UTH and sometimes lower. This can be seen in Table 5, where satellite-radiosonde combinations with a slope larger than 1 indicate moister satellite measurements than radiosondes. The correlation coefficients are very significant in all the cases and range from 0.97 to 0.99 in radiance space. Another very useful statistical parameter that reflects all other parameters is the weighted RMSD, defined in Appendix A. This parameter is presented in both radiance space (Table 4) and humidity space (Table 5). Radiance space RMSD ranges from 1.3 K for TWP-C2/N17 to 2.1 K for TWP-C3/N17. Humidity space RMSD ranges from 2.0% RH for TWP-C3/N17 and TWP-C3/MA to 3.6% RH for TWP-C1/N17 and TWP-C1/MA.

[31] We also performed a statistical t test to evaluate the difference between satellite and radiosonde UTH. We used a paired t test because both satellite and radiosonde UTH are observed under similar conditions. The null hypothesis was that the averages of sonde and satellite Tb/UTH are statistically the same. The results are shown in Table 4. The critical t value at 0.05 confidence level is around 1.97, so in all cases tscore is larger than the critical t value and the null hypothesis fails. Thus, although the bias is small, it is statistically significant.

[32] One known source of bias is that our radiative transfer simulation neglects absorption by ozone (and also the stratosphere, since the radiosonde profiles were cut at 100 hPa). According to Figure 3 of John and Buehler [2004], this will introduce a radiance dependent bias with approximately D240 = −0.15 K and D270 = −0.37 K. This means that up to approximately one quarter to one third of the observed bias can be explained by this effect. Another issue for radiative transfer calculations is the error in water vapor spectroscopy data. For example, we modified the ARTS temperature and pressure exponents using new values proposed by Payne et al. [2008] and did the RT simulations again. The difference between two versions was less than 0.1 K. As has been discussed by Payne et al. [2008], this difference is significant in cold and dry regions like polar climates but it is negligible in humid conditions.

[33] The remaining error is likely to be due to either the satellite measurements or the radiosondes. As to the satellite measurements, there may well be a calibration bias in the channel 18 AMSU-B/MHS data. Without better reference measurements this is difficult to estimate quantitatively. However, at least studies so far (this work, as well as Buehler et al. [2004], John and Buehler [2005], and Buehler et al. [2008]) show generally good consistency between different satellites for this particular channel (with somewhat larger errors for N15 than for the other satellites, as shown in Buehler et al. [2005b], due to a radio interference problem of the AMSU-B sensor on that satellite).

[34] Overall, the available evidence points toward the radiosondes being the dominant error source, since they have several known issues. Soden et al. [2004] and Miloshevich et al. [2006] showed that radiosondes underestimate UTH for very dry conditions and introduced correction procedures to rectify deficiencies in the radiosonde humidity.

[35] A less significant source of error is the time lag of the humidity sensor at very cold temperatures [Miloshevich et al., 2009; Vömel et al., 2007; Wang et al., 2002]. The relative humidity sensor response time increases exponentially with decreasing temperature and the consequence error can be as large as ±20% RH at −40°C and ±40% RH at −70°C if the ambient RH changes rapidly [Miloshevich et al., 2001]. However, this error is expected to be most significant at very high altitudes [Miloshevich et al., 2009], where the AMSU-B channel 18 is not sensitive.

4.4. Radiosonde Homogeneity and General Comparison Features

[36] Overall, the ARM radiosonde data used in this study are considerably more homogeneous than the all-European radiosonde data that were used in John and Buehler [2005]. In that study, the bias at 245 K was found to vary from approximately −1.5 K to +1 K between the different stations for the same satellite. The definition of the bias used there was similar to our D240, except for the small difference in reference temperature, which can be neglected here.

[37] For the ARM stations and nighttime RS92 data, the D240 for the N17 satellite ranges only from −0.61 K to −0.26 K. We attribute this very good result mainly to the fact that the same sensor was used for all data. For the station SGP-C1 we also have RS90 data from early years (see Table 2). Indeed, for those data the D240 with the N17 satellite is +0.66 K, distinctly different from all other N17 comparisons.

[38] The ARM data set is homogeneous enough to combine all data for the N17 satellite into a single scatterplot, as shown in Figure 4. Combined statistical parameters for all stations for N17 and other satellites are also included in Tables 4 and 5.

Figure 4.

Scatterplots of N17 AMSU data versus radiosonde RS92 data from all stations in (a) radiance space and (b) humidity space. Different symbols mark different stations. The dashed and solid lines are diagonal and regression lines, respectively.

4.5. Monthly Gridded UTH

[39] We also compared the monthly averages of the radiosonde UTH to the matching cell of the monthly mean gridded AMSU/MHS UTH data. The monthly data set was originally developed by Buehler et al. [2008] and has been updated regularly. The grid size of this database is 2.5 × 2.5°. We limited the comparison to the months with at least n = 25 radiosonde profiles. This threshold was derived from the desired accuracy of the monthly mean, σm = σ/equation image. The standard deviation of the data, σ, was estimated as 5% RH from Table 6, and we wanted an accuracy of σm = 1% RH, so n = (5/1)2 = 25.

Table 6. Statistical Parameters for Monthly Gridded AMSU/MHS UTH Versus Monthly Radiosonde UTHa
StationSatBiasBias%Slopeσsondeσsatequation imageequation imagetsNNPBiasdBiasd%SlopedNdΔB%BiasaBiasa%SlopeaNaNPa
  • a

    The columns are as follows: Sat is satellites, σ is standard deviation, equation image is long-term average of collocated sonde UTH, equation image is long-term mean of collocated satellite UTH, ts is t score, N is number of collocated months, NP is total number of profiles, Δ B% is the difference between day- and nighttime biases and subscripts d and a refer to day only and day and night together, respectively. The slope, standard deviation, and bias are in %RH. The first 12 columns represent parameters from nighttime data. The accepted limit of ts ranges from ±2.2 (for N = 10) to ±1.99 (for N = 80).

SGP C1N15−0.74−1.660.684.483.5134.3933.652.898036711.575.560.92877.210.120.740.809610060
SGP C1N160.211.040.744.483.8034.3934.61−0.878036713.4912.090.938711.051.505.040.819610060
SGP C1N17−0.54−1.330.824.173.8234.0833.532.326536714.6116.360.847217.691.765.950.817710060
SGP C1N181.524.830.903.924.0432.8834.40−4.253136715.2018.771.143913.933.2010.710.984210060
SGP C1MA0.461.680.844.174.0532.6733.13−0.841536714.9618.300.712116.622.287.800.862410060
TWP C1N15−3.14−6.440.654.823.7846.5743.438.234923994.6812.030.976318.470.822.140.811025386
TWP C1N16−1.43−2.810.734.864.2146.5345.103.794823994.8012.460.826115.271.944.690.80935386
TWP C1N17−0.19−0.240.844.564.1945.9645.780.644223996.2716.360.865216.593.297.850.92775386
TWP C1N181.493.500.814.554.3045.5247.01−3.573223996.8618.251.163314.754.2510.221.00425386
TWP C1MA1.383.360.865.125.7144.2445.62−1.561823997.3619.791.141916.434.5811.191.05245386
TWP C2N15−0.020.670.786.605.5736.0336.010.065126473.9813.001.017712.331.875.780.951125692
TWP C2N161.394.450.806.435.5536.8338.22−3.744426474.7115.740.897211.293.139.430.91955692
TWP C2N171.364.200.875.805.5235.8537.21−3.393426476.5121.790.985917.604.2312.610.99775692
TWP C2N183.269.691.025.896.3234.2937.55−7.742226476.4323.341.113213.655.0115.921.06425692
TWP C2MA1.334.410.907.287.0933.7735.10−1.491026477.1827.031.211722.624.5514.711.07245692

[40] Furthermore, we separated the data once more into day and night to investigate the impact of the radiation dry bias on monthly comparisons. We used the same day and night thresholds as for the one-to-one comparison. Monthly means were calculated by averaging individual profiles. We collocated nighttime-daytime monthly mean UTH to corresponding gridded data from descending-ascending orbits of N15, N16, and N18 and vice verse for N17 and MA. The statistics of this comparison, including bias (in %RH), relative bias (bias%) and slope (in %RH) are reported in Table 6. The nighttime bias is less than 3.3% RH or 10% in relative difference. In relative difference, the TWP-C1/N17 combination shows the lowest bias, just −0.2%. For nighttime data, the slope is nearly always less than one and ranges from 0.65 (TWP-C1/N15) to 1.02 (TWP-C2/N18).

[41] Standard deviations of satellite gridded data and monthly averaged radiosonde data are reported in Table 6 and are quite similar. The standard deviation ranges from 3.9% (SGP-C1/N18) to 7.3% RH (TWP-C2/MA) for radiosonde data and from 3.5% (SGP-C1/N15) to 7.1% RH (TWP-C2/MA) for satellite data. The long-term averages of collocated satellite and radiosonde UTH are also included. The long-term averages of SGP-C1, TWP-C1, and TWP-C2 for nighttime radiosonde data are approximately 32.7%–34.4%, 44.2%–46.6%, and 33.8%–36.8% RH.

[42] Long-term averaged UTH values derived from combinations of recent satellites are a few percentages of RH less than those from older ones (see Table 6). For instance, at TWP-C2, the mean UTH has decreased from 36.0 for N15 to 33.8 for MetOp-A. To investigate these variations, the annual averages of radiosonde and satellite UTH values are shown in Figure 5. The annual averages are calculated from all the available nighttime clear-sky data, regardless of the time difference between sonde and satellite. Annual mean UTH shows relatively large variations at all stations. The variation range is 8% RH at SGP-C1, 5% RH at TWP-C1, 15% RH at TWP-C2, and 10% RH at TWP-C3. Satellite and sonde UTH are consistent at all the stations, except for MetOp-A at SGP-C1, which show a large drift from the others.

Figure 5.

Interannual nnual variation of nighttime sonde and satellite UTH at different stations (station name given in the caption of subplots). Annual averages of both sonde and satellite UTH were calculated using nighttime data, independent of the time difference between sonde and satellite. Vertical error bars show the ratio of the annual standard deviation of UTH to the square root of the number of data points in that year. (c) Outgoing longwave radiation (OLR) is plotted together with UTH to show the anticorrelation between OLR and UTH. The legend shown in subplot Figure 5d is valid for all the subplots.

[43] We do not have a long enough UTH data record to investigate this drift. But it is well known that UTH is the dominant factor controlling outgoing longwave radiation (OLR) [Chung et al., 2009]. Therefore, we employed NOAA OLR data [Liebmann and Smith, 1996] to evaluate the drift. Figure 5c shows OLR data together with UTH data for TWP-C2, which has the highest UTH variability. The negative correlation between OLR and UTH is evident in the plot. The long-term record of annual mean OLR is shown in Figure 6. It is evident from this figure that the apparent trend between 2002 and 2008 is just a part of the regular interannual variation.

Figure 6.

Interannual variation of outgoing longwave radiation (OLR) in Wm−2 at the station TWP-C2.

[44] TWP-C2 and SGP-C1 are located at the border of dry and humid regions (see Figure 1), therefore the interannual variation of the position of air masses shifts these stations from a humid region to a dry one and vice versa. This causes a large variation in UTH at TWP-C2 and SGP-C1 but a small variation at TWP-C1 and TWP-C3, as those stations are located in more stable humid and dry regions. Annual UTH maps (not shown) confirm this interpretation.

[45] We performed a t test, see Table 6, to statistically compare the averages of monthly gridded data and monthly averaged radiosonde data. We argue that the satellite-radiosonde data are paired because both are influenced by the same weather conditions. The null hypothesis is that monthly averages of the gridded and radiosonde data are the same. For many station-satellite combinations, the differences are not statistically significant for the monthly gridded data, although they are significant in the one-to-one comparison. The reason for this apparent paradox is that the dominant error in the monthly comparison is the poor temporal sampling, particularly of the radiosonde data. The sampling problem leads to large random variations that mask the small systematic differences. The monthly comparison does not indicate a significant difference for the following satellite-station combinations: SGP-C1/N16, SGP-C1/MA, TWP-C1/N17, TWP-C1/MA, TWP-C2/N15, and TWP-C2/MA. We did not have enough data points for TWP-C3 station, thus this station is not included in Table 6.

[46] Relative biases for day- and nighttime data are also reported in Table 6. The difference between night- and daytime bias is between 7% and 23%. MetOp-A combinations show the largest differences because the radiosonde data of these combinations are from recent years which are recorded using RS92 sensors. For daytime data the relative bias ranges from 6% for SGP-C1/N15 to 27% for TWP-C2/MA combination. To date, radiosonde-satellite data comparisons performed by other researchers have not separated day-night biases, thus we calculated statistical parameters for combined day-night data. These parameters, including bias and relative bias, are also reported in Table 6. In this case, the bias ranges from 0.1% RH for SGP-C1/N15 to 5.0% RH for TWP-C2/N18. This corresponds to a relative bias of 0.7% and 15.9%, respectively. The slope is larger than 0.8 (SGP-C1/N15) and smaller than 1.1 (TWP-C1/MA, TWP-C2/N18, and TWP-C2/MA).

[47] Buehler et al. [2008] initially validated this monthly gridded UTH at four stations located in different latitudes from subtropical, 15.93°N, to high latitudes, 60.14°N. They reported that for N16 the bias ranges from −4.33 to +4.90% RH, the standard deviation is less than 3% RH, and the correlation coefficients is above 0.85. For the current comparison, the AMSU N16 bias is less than those reported by Buehler et al. [2008] and ranges between 1.5% and 3.1% RH.

5. Summary and Conclusions

[48] Although several studies have compared AMSU/MHS derived UTH with radiosonde data, none have done so for the tropics. We compared the brightness temperature and upper tropospheric humidity, UTH, retrieved from AMSU/MHS radiances, with the ARM radiosonde data of the tropical stations TWP-C1, -C2, and -C3 and the subtropical station SGP-C1.

[49] We analyzed the statistics separately for day and night. Several researchers have shown that Vaisala radiosondes, especially RS90 and RS92, have a large dry bias for daytime soundings (up to 50% in the upper troposphere) that is caused by solar heating of the RH sensor [Vömel et al., 2007; Miloshevich et al. 2006; Wang et al., 2002]. Our own study confirmed this strong radiation dry bias for daytime radiosonde data, particularly RS92.

[50] Microwave radiances are largely insensitive to clouds. However, thick ice clouds are known to have some influence on measurements. We therefore employed a cloud filter which is based on AMSU/MHS humidity channels. The performance of this filter was qualitatively evaluated and confirmed in the tropics.

[51] The main comparison was done for cloud-free nighttime data with a time difference of less than 2 hr between radiosonde and satellite. Additionally, a displacement filter discarded data when the product of wind speed and time difference was larger than 50 km.

[52] We found good agreement between measured and simulated UTH with RMSD less than 3.6% RH and mean difference (bias) less than 2.4% RH. The bias was small but statistically significant. We attribute it to several sources. Part of the bias (up to about one third) can be explained by the omission of ozone absorption (and the entire stratosphere) from the radiative transfer calculations and also the uncertainties in the water vapor absorption model and water vapor spectroscopy data. Apart from that, the most important factor is likely to be the limited accuracy of the radiosondes, but uncertainties in the absolute calibration of the satellite data may also play a role.

[53] It should be noted that this level of agreement would have been considered spectacularly good only a few years ago, when the typical level of agreement in such studies was several tens of percentages rather than a few percentages [Soden et al., 2004]. There are three main factors that contribute to the good agreement. The first factor is the improved quality of the radiosonde data, particularly in this study where all data were from the same sensor type. The second factor is the methodology to use a radiative transfer model to map the radiosonde data to radiance space and then treat it in the same way as the satellite data. The third factor is careful collocation with appropriate filters to exclude cases where the comparison is not valid.

[54] Besides the one-to-one comparison, we also compared monthly gridded satellite UTH against ARM radiosonde data. In that case the absolute value of the relative bias of the nighttime comparison ranged from 0.2% to 9.7%. Some station-satellite combinations showed negative and some others positive bias.

[55] While the results of the one-to-one comparison and the gridded monthly mean comparison are consistent, they are of a quite different practical value. According to a t test, the difference between satellite and radiosonde UTH is statistically significant for all the one-to-one combinations, but only for some of the gridded monthly mean combinations. The reason for this is that the level of agreement between the two data sets is so good that the gridded monthly mean data comparison is dominated by sampling errors.

[56] This finding has repercussions for future satellite validation studies. As data quality improves, collocation requirements become more and more stringent. This trend is likely to continue, so we foresee that a growing fraction of the total effort in future studies will be spent on carefully collocating the data or in planing measurement campaigns to deliver coincident data.

Appendix A:: Statistics

[57] The bias is defined as the weighted mean of the differences between satellite Tb/UTH (Tsat/UTHsat) and sonde Tb/UTH (Tsnd/UTHsnd):

equation image

where ΔB = TsatTsnd or ΔB = UTHsat − UTHsnd and w = 12, σ is the standard deviation of the pixels located inside the target area, see section 3.3. Similarly, relative bias (Bias%) is defined as the weighted mean of the relative differences (ΔB% = (TsatTsnd)/Tsnd or ΔB% = (UTHsat − UTHsnd)/UTHsnd). Bias% is calculated using equation A1 but ΔB is replaced with Δ B%.

[58] The standard deviation of the bias (σbias) is calculated as follows:

equation image

[59] The distance between the regression line and the diagonal line at the reference points, D240, D270, D0%, and D60%, is calculated as follows:

equation image

where a and b are offset and slope of the fitted line and are calculated separately in radiance and humidity space. xref is the reference point at which the distance is calculated and is equal to 240, 270, 0, and 60 for D240, D270, D0%, and D60%, respectively.

[60] The RMSD is defined as follows:

equation image


[61] Radiosonde data were obtained from the Atmospheric Radiation Measurement Program sponsored by the U.S. Department of Energy, Office of Science, Office of Biological and Environmental Research, Climate and Environmental Sciences Division. Thanks to the ARTS radiative transfer community, many of whom have indirectly contributed by implementing features to the ARTS model. Thanks to Patrick Eriksson, Larry Miloshevich, and David Parker for their comments. Special thanks to Oliver Lemke for his assistance. Thanks to Lisa Neclos from the Comprehensive Large Array-data Stewardship System of the U.S. National Oceanic and Atmospheric Administration for providing the AMSU data. Special thanks to three anonymous reviewers for their valuable comments. Viju John was supported by the U.K. Joint DECC and DEFRA Integrated Climate Programme-DECC/Defra (GA01101). Isaac Moradi is financed by the Swedish Space Board, grant 48/09. This work contributes to COST Action ES604, Water Vapour in the Climate System (WaVaCS).