Assessing the quality of humidity measurements from global operational radiosonde sensors



[1] The quality of humidity measurements from global operational radiosonde sensors in upper, middle, and lower troposphere for the period 2000–2011 were investigated using satellite observations from three microwave water vapor channels operating at 183.31±1, 183.31±3, and 183.31±7 GHz. The radiosonde data were partitioned based on sensor type into 19 classes. The satellite brightness temperatures (Tb) were simulated using radiosonde profiles and a radiative transfer model, then the radiosonde simulated Tb's were compared with the observed Tb's from the satellites. The surface affected Tb's were excluded from the comparison due to the lack of reliable surface emissivity data at the microwave frequencies. Daytime and nighttime data were examined separately to see the possible effect of daytime radiation bias on the sonde data. The error characteristics among different radiosondes vary significantly, which largely reflects the differences in sensor type. These differences are more evident in the mid-upper troposphere than in the lower troposphere, mainly because some of the sensors stop responding to tropospheric humidity somewhere in the upper or even in the middle troposphere. In the upper troposphere, most sensors have a dry bias but Russian sensors and a few other sensors including GZZ2, VZB2, and RS80H have a wet bias. In middle troposphere, Russian sensors still have a wet bias but all other sensors have a dry bias. All sensors, including Russian sensors, have a dry bias in lower troposphere. The systematic and random errors generally decrease from upper to lower troposphere. Sensors from China, India, Russia, and the U.S. have a large random error in upper troposphere, which indicates that these sensors are not suitable for upper tropospheric studies as they fail to respond to humidity changes in the upper and even middle troposphere. Overall, Vaisala sensors perform better than other sensors throughout the troposphere exhibiting the smallest systematic and random errors. Because of the large differences between different radiosonde humidity sensors, it is important for long-term trend studies to only use data measured using a single type of sensor at any given station. If multiple sensor types are used then it is necessary to consider the bias between sensor types and its possible dependence on humidity and temperature.

1 Introduction

[2] Atmospheric water vapor is the main natural greenhouse gas and significantly contributes to the Earth's radiation budget. Radiosonde humidity data are the main source of direct measurements of the vertical structure of the atmospheric humidity. Sonde data are assimilated into reanalysis [Dee et al., 2011] and are used as a priori information to retrieve atmospheric profiles from satellite data (e.g., 2011) as well as to initialize weather forecast models [Lorenc et al., 1996]. These data are also widely used for evaluating climatological features (e.g., Añel et al. [2008] and Ribera et al. [2008]). However, sonde data are prone to different errors especially in cold temperatures such as in the upper troposphere. The accuracy of the sonde data is significantly affected by the sensor design, calibration, data processing, and contamination. The uncertainty in the sonde measurements is especially very large in the upper parts of the troposphere owing to its dry and cold conditions [Elliott and Gaffen, 1991; Miloshevich et al., 2001; Nash et al., 2010]. A particular source of error is a dry bias caused by solar heating of the sensors [Vömel et al., 2007; Miloshevich et al., 2009] that affects data collected during daytime.

[3] Previous studies show very large spatio-temporal inhomogeneities in global sonde data [Elliott and Gaffen, 1991; Elliott et al., 1998; Ross and Elliot, 2001; Moradi et al., 2013a; Soden and Lanzante, 1996]. Other sensor intercomparison studies also show that even under controlled conditions, different sensor types show different accuracies throughout the troposphere, especially in mid-upper troposphere [Nash et al., 2010; Miloshevich et al., 2001]. In this study we evaluate the quality of humidity measurements from operational radiosonde sensors using satellite data observed at the microwave (MW) frequencies near the water vapor absorption line at 183 GHz.

[4] Section 2 introduces the satellite and radiosonde data sets that were used for the comparison. Section 3 explains the methodology and collocating criteria. Results and discussions are in section 4, and section 5 summarizes the comparison.

2 Satellite and Radiosonde Data

[5] The Advanced Microwave Sounding Unit-B (AMSU-B) and Microwave Humidity Sounder (MHS) are cross-track scanning five channel microwave radiometers. The AMSU-B channels operate at 89.0, 150.0, 183.31±1.00, 183.31±3.00, and 183.31±7.00 GHz [Saunders et al., 1995]. The MHS sensor is very similar to AMSU-B, but the second channel is moved to 157.0 GHz and the fifth channel has only one passband at 190.311 GHz [Bonsignori, 2007; Goodrum et al., 2007]. The instruments have a swath width of approximately 2300 km, with 90 scan positions. The earth incidence angle is 0.62° for the innermost scan positions and 58.5° for the outermost scan positions. The foot print size, defined with respect to the half power beam width, is approximately 16 km at nadir but increases toward the edge of the scan [Saunders et al., 1995; Goodrum et al., 2007]. In this study we used AMSU-B data from the instruments aboard NOAA-15 (2000–2002), NOAA-16 (2000–2005), and NOAA-17 (2002–2011), and MHS data from NOAA-18 (2005–2011), NOAA-19 (2009–2011) and MetOp-A (2006–2011). We limited NOAA-15 and NOAA-16 data to a few years after launch, since AMSU-B data from these satellites have suffered from asymmetry and calibration problems [Moradi et al., 2012].

[6] The Integrated Global Radiosonde Archive (IGRA) developed by the National Oceanic and Atmospheric Administration's (NOAA's) National Climatic Data Center (NCDC) consists of over 1500 globally distributed stations and most of them have two launches per day [Durre et al., 2006]. In this study, we evaluated the IGRA data for the period 2000–2011. We used the IGRA metadata and other sources such as Wang and Zhang [2008] to identify the radiosonde sensor type.

3 Methodology and Collocating Criteria

[7] There are generally two approaches to compare radiosonde profiles with the satellite data. The first approach is to simulate the satellite brightness temperatures (Tb) using radiosonde profiles and a radiative transfer (RT) model, then compare the observed and simulated Tb's. The second approach is to first retrieve the temperature and humidity profiles from the satellite observations using inversion techniques, then compare radiosonde and satellite retrieved profiles. In this study, we use the first approach and compare satellite observed and radiosonde simulated Tb's which avoid issues related to satellite retrieval algorithms and inversion techniques.

3.1 Radiative Transfer Simulations

[8] We used the atmospheric radiative transfer simulator [Eriksson et al., 2011], which is a line-by-line model to conduct the radiative transfer calculations. The RT model requires pressure, temperature, and water vapor mixing ratio profiles as input to simulate the satellite Tb's. Since water vapor mixing ratio is calculated using both temperature and humidity data, any error in the temperature data will be reflected in the calculated values for water vapor mixing ratio. However, previous studies show good accuracy for the sonde temperature data. For instance, Sun et al. [2010] compared the IGRA temperature profiles with the Constellation Observing System for Meteorology, Ionosphere, and Climate Radio Occultation (COSMIC-RO) data and concluded that the overall difference between the IGRA and COSMIC temperature profiles is 0.15 K throughout the troposphere and lower stratosphere.

[9] The AMSU-B/MHS water vapor sounding channels (183.31±1, 183.31±3, and 183.31±7 GHz) are sensitive to the amount of water vapor in different layers of the atmosphere. Therefore, these channels can be used to evaluate the quality of the radiosonde humidity data in different layers of the atmosphere. The sensitivity of MW sounding channels to different altitudes of the atmosphere is defined using Jacobians [Garand et al., 2001]. The water vapor Jacobians show the sensitivity of the measured radiances or Tb's to the perturbations of water vapor at different altitudes. Mathematically, Jacobians are the partial derivative of radiance with respect to the atmospheric parameter influencing the observed radiance. Perturbation method is normally used to obtain the Jacobians. In this method, the water vapor Jacobians (K) are defined as follows:

display math(1)

where, MR is water vapor mixing ratio, Q is equal to 0.05×MR, and the subscript i denotes to the layer of the atmosphere. This equation describes the change in Tb (in Kelvin) due to a small change in the water vapor concentration [Garand et al., 2001]. At each level, the water vapor concentration is perturbed by 10% (±5%) and the effect of this perturbation on the observed Tb is taken as the magnitude of the Jacobian at that level. The Jacobians of the AMSU-B water vapor channels are shown in Figure 1. Negative Jacobians mean that any increase in the water vapor concentration leads to a decrease in the observed radiance and vice versa. It has to be pointed out that the peak altitude of the water vapor Jacobians is a function of the precipitable water vapor (PWV). For instance, the 183±1 GHz peak altitude is about 6 km for a subarctic winter profile with a PWV of 5 kg.m2, but moves to an altitude of about 8 km for a subarctic summer profile with a PWV of 28 kg.m2(see Figure 1).

Figure 1.

The Jacobians for the AMSU-B channels operating at 183±1, 183±3, 183±7, 150.0, and 89.0 GHz. The plots are for (top) dry (PWV = 5 kg.m−2), (middle) semi-moist (PWV = 11 kg.m−2), and (bottom) moist (PWV = 28 kg.m−2) conditions.

[10] As shown in Figure 1, in dry conditions, the Jacobians could become positive near the surface and touch the surface. In these cases, the measured radiance is affected by the surface emissivity. For instance, as shown in Figure 1, for a subarctic profile which is relatively dry, PWV=5 kg m−2, both Channels 4 (183±3 GHz) and 5 (183±7 GHz) are affected by the microwave emission from the surface, but for a subarctic summer profile with a PWV of 28 kg m−2, which is relatively humid, none of the water vapor channels is affected by the surface emissivity. Accurate surface emissivity data are required to simulate satellite radiances when they are affected by the surface. Few data sets have been developed for the ocean surface emissivity, yet there is no reliable database for the microwave land surface emissivity. Hence, we limit our study to the profiles where the satellite Tb is not affected by the land surface emissivity.

3.2 Collocating Criteria

[11] In this study, we use the terms upper, middle, and lower troposphere (UT, MT, and LT, respectively) to refer to the layers where the AMSU-B/MHS water vapor channels, 183.31±1.00, 183.31±3.00, and 183.31±7.00 GHZ, respectively, are most sensitive to. As shown in Figure 1, there is a slight overlap between these layers; however, this overlap does not introduce any error in our comparison since the simulated and observed Tb's both correspond to the same layer of the atmosphere.

[12] As mentioned before, because of the lack of accurate surface emissivity data for the microwave frequencies, it is important to remove profiles whose Tb's are affected by the surface emissivity. The Tb emitted by the surface and observed by the satellite is called the land surface contribution (LSC). Figure 2 shows the LSC as a function of PWV. The simulations were conducted using a subset of radiosonde data from the Atmospheric Radiation Measurement (ARM) program. We calculated LSC as the difference between the simulated Tb's when the surface emissivity varies from 0 to 1. As Figure 2 shows, the UT and MT channels are not sensitive to the surface emissivity if the PWV is greater than 5 and 10 kg.m−2, respectively. The LT channel is the most sensitive to the surface emissivity and becomes independent of surface emission only if PWV is greater than 30 kg.m−2. We used these PWV thresholds, i.e., 5, 10, and 30 kg.m−2, for the UT, MT, and LT channels, respectively, to screen out surface affected Tb's.

Figure 2.

Land surface contribution (LSC) for AMSU-B water vapor channels as a function of PWV. Note that the LSC is in logarithmic scale, therefore, we added 1 K to the LSC to avoid negative values introduced by numerical errors.

[13] Other collocating criteria including time difference, spatial distance, cloud screening method, and displacement filters are explained in Moradi et al. [2013a] and Moradi et al. [2010]. In this study, we allowed a time difference of up to 2 h between satellite overpass time and sonde launch time, and a spatial difference of 50 km between the geographical coordinates of the sonde launch site and the satellite spot.

3.3 Methodological Uncertainties

[14] Methodological uncertainties, such as uncertainties in the satellite data, RT calculations, and sampling errors, play an important role in this comparison; hence, these sources of uncertainty are explained in detail in this section.

[15] Sampling error is introduced by spatial and temporal differences between the sonde and satellite data. This error is estimated to be 3.3% RH per 3 h and 3.1% RH per 100 km in the troposphere [Sun et al., 2010]. We used a temporal threshold of 2 h and a spatial threshold of 50 km, hence, the sampling error is estimated to be about 0.3 K in Tb. The RT theory for the microwave frequencies around 183 GHz is well established and is simpler than the RT calculations for infrared and visible channels [Melsheimer et al.2005]. In addition, the channels we used are close to the center of the water vapor absorption line at 183.31 GHz, so that uncertainty of the continuum absorption parameters plays only a very small role in the RT calculations. Large perturbations in the spectroscopic line parameters are necessary to cause significant radiance differences, which are unlikely given the well established use of this line for atmospheric measurements. Altering water vapor spectroscopy data for the water vapor line at 183 GHz only introduces a small bias of less than 0.1 K in the RT calculations [Moradi et al., 2010; Payne et al., 2008]. Therefore, a rough estimate of the RT errors is about 0.2 K.

[16] The prelaunch specification for the uncertainty in the microwave water vapor channels, known as noise equivalent temperature (NEΔT), is about 0.5 K [Saunders et al., 1995]. There are several sources that could contribute to a post-launch uncertainty in the satellite data, including uncertainty in the calibration coefficients, geolocation, cross-polarization, and antenna pattern [Saunders et al., 1995; Moradi et al., 2013b]. However, previous studies, e.g., Moradi et al. [2010], show generally good consistency between different satellites for these particular channels. Besides, we only used a few years of data from NOAA-15 and NOAA-16, because the accuracy of data from these satellites decreased a few years after launch [Moradi et al., 2012]. In summary, our estimate for the total uncertainty is about 1.0 K.

4 Results and Discussion

[17] We evaluated relative humidity data from 19 different sonde sensor types compared to the satellite observations, taking the satellite Tb's as the reference. The sensors and their acronyms used throughout this section are given in Table 1. The first letter of the acronyms indicates country of the manufacture: C for China, I for India, R for Russia, U for United States, and V for Vaisala (Finland). As shown in Table 1, the radiosonde sensors can be classified into three groups based on the radiosonde sensor type. The oldest radiosondes use gold-beater's skin, which is a moisture-sensitive membrane made out of calf intestines. This type of sensor is used in some Russian and Chinese sensors. The second category belongs to the carbon hygristor that is used in most radiosondes made in the U.S. and the Chinese GTS-1 radiosonde. Theses sensors are made of a carbon particle coated glass or plastic membrane, whose resistance varies with relative humidity. The last group belongs to the Vaisala radiosondes that are equipped with a polymer sensor. Some of the Russian made sensors also use a Vaisala sensor module. The total number of collocated data points for daytime and nighttime data are given in Tables 2, 3, and 4 for the UT, MT, and LT, respectively. As mentioned before, the only collocating criterion that differs for different channels is the PWV threshold, which is higher for the LT channel than for the other channels. Therefore, we have less collocations for the LT channel. The percentage of data removed by the PWV filter in the LT and MT are given in Table 5. The UT channel is not included in Table 5, because this channel is not sensitive to the surface, and the PWV filter only removes a few profiles for this specific channel. However, in most cases, the PWV filter removes more than 50% of the data for the LT channel. Therefore, the results presented for the LT are only valid for very moist conditions.

Table 1. List of Radiosonde Sensors Studied in This Paper
AcronymNameSensor Type
C-GTS1SHANGHAI GTS1 1680 MHZCarbon hygristor
C-GZZ2GZZ-2 403 MHZGoldbeater's skin
R-ATEXATTEX MTP-5HGoldbeater's skin
R-MARSMARSGoldbeater's skin
R-MRZ3MRZ-3AGoldbeater's skin
R-RF95AVK-RF95-ARMAGoldbeater's skin
U-SPCNSIPPICAN 1649-540Carbon hygristor
U-VZB2VIZ/SIPPICAN B2Carbon hygristor
X-UNCLNo metadataUnknown
Table 2. Statistics for the Sonde Versus Satellite Comparison in the Upper-Tropospherea
  1. a

    R stands for correlation coefficient, b for the slope of the fitted line, N number of the collocated data points, n stands for the nighttime and d for the daytime data. tt shows the results of the statistical t test and the null hypothesis is rejected if the t test value is equal to 1.

Table 3. Same as Table 2, but for the Middle-Troposphere
Table 4. Same as Table 2, but for the Lower-Troposphere
I-MARK    0.870.7168
Table 5. Percent of the Data Rejected by the PWV Filtera
  1. a

    The UT channel (183±1 GHz) is not sensitive to this filter, so just the values for the MT (183±3 GHz) and LT (183±7 GHz) channels are included. The subscripts n and d stand for the daytime and nighttime data, respectively.

I-MARK00 7
R-ATEX1 72 

4.1 Overall Comparison

[18] Figure 3 shows the nighttime and daytime biases for different sensors relative to the satellite data. The differences between nighttime and daytime biases known as daytime radiation bias [Luers, 1997; Vömel et al., 2007; Nash et al., 2010] are also explicitly given in Figure 3. The data measured shortly before sunrise or after sunset may also be affected by solar radiation, therefore, we excluded nighttime data measured 30 min before sunrise or after sunset. It is estimated that each Kelvin bias is equal to 10% error in terms of relative humidity. This value was derived by adding 5% and 10% to the relative humidity profiles from the ARM Program, then conducting radiative transfer calculations using both original ARM data and data with altered humidity profiles. The results showed that adding 5% and 10% to the relative humidity profiles introduces about 0.5 K and 1.0 K bias, respectively, in terms of Tb. These values are almost the same for all the water vapor channels. Near the water vapor absorption line at 183 GHz, the emitted radiance is absorbed by water vapor; therefore, the observed or calculated Tb's are inversely related to the concentration of atmospheric water vapor. Therefore, negative biases (sonde minus satellite Tb) show a wet bias for the sonde measurements and vice versa. Since the total uncertainty is expected to be about 1 K, all sensors with a bias less than 1 K are interpreted to have a small bias. It is not possible to quantify these small biases using current reference data.

Figure 3.

Nighttime and daytime biases in terms of Tb in Kelvin in (top) upper, (center) middle, and (bottom) lower troposphere. The daytime radiation bias (daytime minus nighttime bias) is also shown. The dashed lines show the total uncertainty introduced by different sources.

[19] Overall, as shown in Figure 3, all Russian and a few other sensors, including C-GZZ2, U-VZB2, and V-R80H, show a wet bias in the UT; all other sensors show a dry bias. Russian sensors have the largest biases in UT ranging from −4 to −6 K. These large biases are likely due to the long response time of the humidity sensors, i.e., the sensor is still responding to the lower and more moist altitudes when ascending through the troposphere [Soden and Lanzante, 1996]. The bias becomes larger in UT because of the rapid exchange of water molecules between the sensor and the surrounding environment. The dry bias of the Vaisala sensors is introduced by a few factors including the following: contamination of the humidity sensor by chemical substances, long-term instability of the sensor polymer, and time-lag error [Miloshevich et al., 2001; Wang et al.2002].

[20] The standard deviation of the differences between sonde and satellite measurements (STD), which is a measure for the random errors, is shown in Figure 4. All Chinese, Indian, and U.S. sensors have a STD greater than 4.5 K in UT. In fact, all of these sensors fail to respond to humidity changes in UT. Therefore, none of these sensors is suitable for UT studies. As expected in most cases the differences between the daytime and nighttime STDs are negligible, since the daytime radiation bias is a systematic error, which does not affect the STDs.

Figure 4.

Standard deviations of the differences between the sonde and satellite Tb's for nighttime (blue) and daytime (red) data in (top) upper, (center) middle, and (bottom) lower troposphere.

[21] The differences between daytime and nighttime biases show the effect of solar radiation on the daytime measurements and are shown in Figure 3. In most cases this bias is a dry bias known as daytime radiation dry bias. The RF95 and VZB2 sensors show a negligible radiation wet bias in UT, which is smaller than the methodological uncertainty. The daytime radiation bias depends on the sensor design and the observation time. Near sunrise or sunset, the daytime radiation bias may be small. Therefore, the observation times for collocated sonde data in relative local time (rLT), see Appendix A, are shown in Figure 5. As shown in Figure 5, most daytime sonde measurements for the Chinese and Indian sensors are from the early mornings and late evenings. On the other hand, most of the Vaisala daytime data are collected around local noon, which intensifies the radiation dry bias for those sensors. Bian et al. [2010] reported a daytime radiation dry bias of about 9% (about 1 K) below 500 hPa and 35%–50% (about 3.5–5 K) above 400 hPa for C-GTS1. We found smaller daytime radiation dry bias, which is due to the fact that we have used operational data that are measured in the early mornings and late evenings; while most of the daytime data for Bian et al. [2010] were measured around 14:00 local time.

Figure 5.

Number of collocated observations in relative local time in (top) upper, (center) middle, and (bottom) lower troposphere.

[22] In the MT, Russian sensors still have a wet bias, but all other sensors have a dry bias. The wet bias of the Russian sensors in MT is about 3 K less than their bias in UT. As shown in Figure 3, all sensors have a dry bias in LT. The Russian sensors have the lowest biases in LT. The STDs and daytime radiation biases decrease toward LT. It is worthwhile to mention that the stronger diurnal variation and vertical gradient of humidity in LT may introduce a larger sampling error in this layer. However, since we used a temporal threshold of 2 h, the error should be negligible even in the lower troposphere. As mentioned before, in LT, only the very moist profiles were used in the comparison due to the high value for the PWV threshold. The LT biases might be different if we could process all the data including all dry profiles. As mentioned before, the current method is not able to utilize all the data including all dry profiles.

[23] Our results are consistent with Bian et al. [2010], who reported that the GTS1 sensor fails to respond to humidity changes in the mid-upper troposphere. The results are also consistent with Miloshevich et al. [2006] who reported a large uncertainty for the U-VMK2 measurements in the UT, because the sensor stops responding to humidity changes at some point and continues to measure the same conditions as it ascends. Overall, our findings are consistent with Sun et al. [2010] but exhibit notable differences for some sensors. Relative to MetOp-A MHS data, Sun et al. [2010] reported a dry bias for most radiosonde types but a moist bias for the Russian sensors. We also found a dry bias for most sensor types, and a wet bias for the Russian sensors. However, the absolute biases are different. The main reason for this discrepancy is that Sun et al. [2010] only used one year of the MetOp-A data, which limits the number of collocations, but our statistics are based on a long period of data from several satellites. Our results are not directly comparable with Nash et al. [2010], since they compared the profiles, but we have an average bias value for each profile and in each layer. Generally, the absolute biases reported in Nash et al. [2010] are smaller than those reported here, which may be explained by the fact that we have used operational data, whereas Nash et al. [2010] used data measured under controlled conditions during a campaign. Note that the GTS-1 sondes evaluated in this study and Bian et al. [2010] are manufactured by Shanghai Changwang Company and use a carbon hygristor sensor, while the GTS-1 sondes evaluated in Nash et al. [2010] are manufactured by Huayun (GTS(U)1-1) and Daqiao (GTS1-1) Companies and use a polymer sensor.

[24] We performed a t test analysis to examine whether the differences between sonde and satellite Tb's are statistically significant. The results are shown in Tables 2, 3, and 4 for UT, MT, and LT, respectively. The null hypothesis was that the bias is not different from zero at the significance level of 0.05. In most cases, the null hypothesis was rejected, which means that the biases are significantly (at 5% level) different from zero.

[25] We also examined the effect of replacing upper tropospheric water vapor profiles with a constant value from the 300 hPa level. The main purpose was to test whether the sonde measurements provide valuable information in the UT or the sensors have just become unresponsive to water vapor. For this test we used data from the ARM stations, including Tropical Western Pacific (TWP), Southern Great Plains (SGP), and North Slope of Alaska (NSA) and replaced the relative humidity values for the layer expanding from 300 hPa to the stratosphere with the value measured at 300 hPa. Then we conducted the RT calculations using both the original and the modified data sets. The simulations show that the effect of replacing the UT data depends on the climatic conditions of the station as well as the satellite channel. Since the UT channel is the most sensitive to the water vapor signal above 300 hPa, the effect of a “nonresponsive” sensor is expected to be strongest in this channel and should decrease strongly toward the LT channel. The calculated biases for “nonresponsive” profiles over the original data in UT are 1.9 K, 3.1 K, and 8.0 K at TWP, SGP, and NSA, respectively. In MT the bias decreases to 2.2 K, 1.7 K, and 5.8 K at the three sites and in LT this bias drops to 0.1 K, 0.2 K, 0.7 K, respectively. The bias found at SGP, which is a mid-latitude site, is comparable to the biases we found for the sensors made of goldbeater's skin and carbon hygristor, which implies that these sensors have biases comparable to a “nonresponsive” sensor.

[26] Figure 6 shows the simulated sonde Tb's versus the satellite Tb's for a few representative sensors. The slopes of the fitted lines for all the sensors are given in Tables 2, 3, and 4 for the UT, MT, and LT, respectively. The following facts should be taken into account to correctly interpret the scatterplots: (1) higher Tb's show a drier atmosphere and vice versa, (2) the sonde sensor has a dry bias if the sonde simulated Tb is larger than the satellite observed Tb and vice versa, and (3) the radiosonde sensor has a dry bias in dry conditions and a moist bias in moist conditions if the slope of the fitted line is greater than one and vice versa, if the regression line crosses the diagonal line near the midpoint of satellite Tb's. Obviously, the bias does not depend on the atmospheric conditions if the slope is close to unity. The C-GTS1 and C-GZZ2 sensors overestimate UT humidity in moist conditions and underestimate it in dry conditions. The C-GTS1 sensor underestimates MT humidity in dry conditions. The C-GZZ2 and R-MARS (both in LT) and I-IMDR (throughout the troposphere) overestimate tropospheric humidity in dry conditions and underestimate it in moist conditions. The U-VMK2 and V-RS92 sensors underestimate LT humidity in moist conditions.

Figure 6.

Sonde simulated Tb's versus satellite observed Tb's for chosen radiosonde types and the three water vapor channels. The columns from left to right show the UT, MT, and LT channels, respectively, and the sonde type is printed on the plots.

4.2 Distribution of Sonde Humidity Data

[27] As mentioned before, the comparison of sonde versus satellite data indicates that most radiosonde sensors fail to respond to the relative humidity changes in UT and sometimes even in MT. However, the comparison does not identify the altitude where sonde stops responding to relative humidity. We use the density plots of sonde data that show the frequency of occurrences of relative humidity values as a function of altitude to describe this issue. Figure 7 shows the density plots for nighttime sonde data. There was no obvious difference between daytime and nighttime distributions; therefore, only the nighttime distributions are shown in Figure 7. The distributions are evaluated separately for different regions to account for the regional variation of relative humidity. The distributions are separated based on the first digit of the WMO code that reflects the geographical region. This digit is 0 and 1 for Europe, 2 and 3 for Russia, 4 for Middle East, India and South Asia, 5 for China, 6 for Africa, 7 for North America, 8 for South America and Antarctica, and 9 for Oceania. The distributions of the RS92 humidity data from the ARM stations are also shown in Figure 7 to be used as a reference. As discussed before, the satellite observations are only sensitive to the altitudes below 12 km, so the sonde biases above that altitude cannot be evaluated using microwave satellite data.

Figure 7.

Vertical distribution of the relative humidity data from different sensor types. The data are partitioned based on the sensor type as well as geographical region (first digit of the WMO code). The plots in the bottom row are for data from the ARM stations. The red dashed line indicates the altitude up to which the microwave measurements are sensitive to water vapor.

[28] All Vaisala sensors, including those from the ARM program, show a similar distribution. They indicate extremely frequent very dry measurements above 12 km. We cannot tell if this dry region is a natural signal or a dry bias in the Vaisala data. Another interesting feature is that a small number of the RS92 soundings show relatively high values in the UT. These data most likely suffer from sensor icing. This means that even the Vaisala RS92 sensor, equipped with the sensor heating system, is not 100% free of sensor icing. The Russian sensors show a substantially more moist mid-upper troposphere than Vaisala sensors used in the same region. However, in LT, both Russian and Vaisala sensors show a similar distribution. The Russian Vaisala sensor sonde (R-R95V) uses the Vaisala RS80 sensor module and should therefore show a similar RH distribution as the RS80 sensor. However, the comparison between these sensors in region 2 shows they seem to be very different. Thus, the implementation or the processing unit of a sensor may also contribute to the final errors in the measurements.

[29] Most of the RH data from both Indian sensors (IMDR and MARK) range systematically between 5% RH to 20% RH in the mid-upper troposphere. Many of the soundings from the Chinese sensors indicate extremely dry measurements above 4 km. This is an indication that these sensors fail at those altitudes and report a constant low value. Obviously, the GZZ2 data are more moist than the GTS1 data throughout the middle to upper troposphere. Similarly, many measurements from the U.S. sensors fail through the troposphere that are indicated by frequent very dry measurements above 2 km. All U.S. sensors show a moist region above 12 km, but as mentioned before, the current satellite observations are not able to capture that region. Three different sensor types are used in Antarctica: R-MRZ3, V-R80A, and V-R92S. The Russian and Vaisala sensors show different patterns over Antarctica. The Russian sensor shows a moist mid-upper troposphere like the pattern we found over Russia, but the Vaisala sensors show a pattern similar to what we found for the Vaisala sensors in other regions. The distributions of the Vaisala RH data in Antarctica are also consistent with those from the ARM RS92 data measured at the NSA station.

[30] Overall, the humidity sensors made of goldbeater's skin and also the carbon hygristor (see Table 1) are only sensitive to water vapor in the LT and just barely in the MT. These sensors fail to respond to humidity changes in the upper troposphere and even middle troposphere. This can be identified from the density plots of the RH data as well as the large standard deviation of the differences between sonde and satellite Tb's. Only the Vaisala sensors made out of polymer respond to the UT humidity changes.

4.3 Temporal and Spatial Inhomogeneities

[31] Studies based on long-term radiosonde humidity data should take into account that sensor changes have a crucial effect on evaluating the temporal and spatial variation of tropospheric humidity, especially in the UT. For instance, Figure 8 shows the temporal variation of the differences between sonde and satellite Tb's at Anchorage (WMO no. 70273). That station has used a V-R80A sensor until the end of 2008 then changed to a U-VMK2 sensor. In UT, the VMK2 bias is much larger than the R80A bias, but in the mid-lower troposphere the difference between the two sensors is small. The figure clearly shows that the discontinuity introduced by the sensor change would lead to a wrong interpretation of the climate trends. Therefore, it is highly recommended to use data from a single type of sensor for trend analyses of tropospheric humidity. Since differences between different sensor types are larger in UT than in the mid-lower troposphere, using different sonde sensors to study UT humidity should be avoided.

Figure 8.

Temporal discontinuity of the differences between the sonde and satellite Tb's introduced by the sonde sensor change at Anchorage in upper, middle, and lower troposphere, from top to bottom, respectively. N16 to N19 show the data from NOAA-16 to NOAA-19 satellites and MPA shows the data from MetOp-A satellite. The vertical line shows the date when the radiosonde sensor type was changed.

[32] The current radiosonde network is not appropriate to study the spatial variation of tropospheric humidity especially in the UT. Figure 9 shows the spatial variation of nighttime biases in the upper, middle and lower troposphere. Obviously large differences are observed in the UT resulting from the low-quality of UT data. Most European and Australian stations are using Vaisala sensors, which have a better quality than other sensors. The spatial distribution of the biases, especially in the LT, may also be affected by the diurnal variation of relative humidity. However, since we used a small time difference for the collocations, this effect should be small. Figure 9 also shows that different sensor types operating in the same region still have different biases, e.g., different sensor types used in the U.S. network show different biases. Furthermore, similar sensor types operating in different regions have the same bias, e.g., Vaisala sensors used in the U.S. and Scandinavian countries have similar biases. The spatial inhomogeneity in sonde data, especially in the UT, has been discussed in more detail in the earlier studies [Elliott and Gaffen, 1991; Elliott et al., 1998; Ross and Elliot, 2001; Moradi et al., 2013a; Soden and Lanzante, 1996].

Figure 9.

Spatial inhomogeneity in the radiosonde biases relative to the satellite data in the upper, middle, and lower troposphere, from top to bottom, respectively. NOAA15 to NOAA19 show the data from the NOAA-15 to NOAA-19 satellites and METOPA shows the data from the MetOp-A satellite.

5 Summary and Conclusions

[33] The quality of tropospheric humidity data from different operational radiosonde sensors was evaluated for a 12 year period (2000–2011) using satellite data from Advanced Microwave Sounding Unit (AMSU-B) aboard NOAA-15, NOAA-16, NOAA-17, and Microwave Humidity Sounder (MHS) aboard NOAA-18, NOAA-19, and MetOp-A. The main purpose of this study was to identify sensors that are more or less reliable than the other sensors in different layers of the troposphere, i.e., lower, middle, and upper troposphere. Similar to earlier studies, this study reveals a spatial inhomogeneity in the data since different sensors tend to be used in specific regions. Since sonde data are widely used in different fields including forecast models, reanalysis, and satellite retrieval validations, the results of this study can be widely used in different fields of atmospheric and hydrologic sciences.

[34] The results show that most sonde sensors have a systematic dry bias throughout the troposphere. Only Russian sensors have a large systematic wet bias in the mid-upper troposphere. Some sensors including those from China, India, and the United States show a large random error in the mid-upper troposphere. These sensors fail to respond to humidity changes in the upper and sometime even in the middle troposphere. It should be pointed out that some of these sensors are not manufactured for measuring upper tropospheric humidity. For example, the manufacturer specifications for the Sippican MarkIIa (U-VMK2) RH sensors is 5%–100% RH at temperatures above −50 °C, so that the measurements may not be reliable in the cold upper tropospheric conditions [Miloshevich et al.2006]. The Russian sensors have a large wet bias in the upper troposphere decreasing toward the mid-lower troposphere, so that their lower tropospheric bias is comparable with the biases for the Vaisala sensors. Russian sensors also have a relatively small standard deviation of less than 2 K in the lower troposphere, which is comparable to the standard deviation of the Vaisala sensors. Overall, Vaisala sensors perform better than other sensors in the mid-upper troposphere; however, both Vaisala and Russian sensors can be used for studies related to the lower tropospheric water vapor, such as precipitable water vapor.

Appendix A: Relative Local Time

[35] Relative Local Time (rLT) is calculated using local standard time (LST) as follows:

display math(A1)

[36] In this time scale, the sun always rises at 06:00 A.M. and sets at 06:00 P.M. The entire day is divided into four quarters, i.e., midnight to sunrise, sunrise to noon, noon to sunset, and sunset to midnight. For each quarter, min is the start time of that quarter, max is the end of that quarter, and n is the quarter number. For instance, for the quarter between sunrise and noon, min is equal to sunrise in local time and max is equal to 12:00. Midnight is counted as 00:00 for the first quarter and 24:00 for the last quarter.


[37] We would like to thank three anonymous reviewers for their valuable comments.

[38] This research was partially supported by National Oceanic and Atmospheric Administration (NOAA) via grant to the Cooperative Institute for Climate and Satellites (CICS).