2.1 Radiosonde Data

The high-resolution radiosonde data of this study are from the China L-band second-level observational basic data set (version 1.0) of the China National Meteorological Administration's National Information Centre. The humidity and temperature from the radiosonde instruments have been updated from the 59-type radiosonde device to an electronic radiosonde device, and the detector test element has been replaced by a wet-resistance approach. Large-scale modernization was mainly carried out from 2005–2007 to the end of 2010, 120 operational radiosonde stations in China were updated to the L-band radar-electronic sounding systems. The L-band sounding radar is a new generation of secondary wind-measuring radars, which are fully automated in China and are synchronized with GTSl digital electronic radiosondes to continuously and automatically measure pressure, temperature, humidity, wind speed, wind direction, and other meteorological elements. The sampling period is approximately 1.2 s (China Meteorological Administration, 2010). The accuracy of the radiosonde measurements shows that the L-band radiosonde has a significantly higher accuracy than the radiosonde Model 59. The GTS1 radiosonde temperature detection accuracy is comparable to the RS80 radiosonde, while the GTS relative humidity sensor has a significant dry bias compared to the RS80, especially under cloud conditions (Bian et al., 2011). In this case, we have manually eliminated these data during data processing (See Text S1 and Figure S2 in the supporting information, SI, for the details description and quality control of radiosonde data set).

In accordance with the rule for conventional upper-air meteorological observations, each station performs two conventional aerial surveys at 0800 and 2000 BJT per day, and additionally performs observations at 0200 and 1400 BJT depending on the special weather conditions or test requirements (J. Guo et al., 2016). Since 2011, more observations have been irregularly performed at 1400 BJT in summer during special weather conditions and depending on experimental needs. However, as only extended observations for the whole month were conducted in June 2013 and June 2014 according to statistics of the original database, we consider sounding observations only in these two months to validate the q and PWV products from the AIRS instrument.

To match the passing time of the Aqua satellite (around 1330 BJT), we use the sounding observations at 1400 BJT from 113 radiosonde sites (see the dots in Figure 1) in June 2013 and June 2014 to calculate the q and PWV and to evaluate the AIRS-based q and PWV products. For the convenience of discussion and according to their geographical location, elevation, and station density (Figure 1), the sounding data in the stations with good quality are simply divided into seven subregions (refer to Zhai and Eskridge, 1997): northeast China (12 sites), north China (17 sites), northwest China (24 sites), the Tibetan Plateau (13 sites), southwest China (14 sites), the Yangtze River Valley (17 sites), and south China (16 sites). In addition, detailed provincial administrative boundary and some special station names pointed in the below texts are identified in Figure S1 in the SI.

2.2 AIRS Data

The AIRS instrument aboard the Aqua satellite is the first high-spectral-resolution infrared sensor to provide the three-dimensional water vapor distribution of the atmosphere and can complement the limitations of other satellite observations (Aumann et al., 2003). The AIRS instrument is a cooled-grating spectrometer providing 2,378 infrared channels at wavelengths of 3.7–15.4 μm (3.74–4.61, 6.20–8.22, and 8.8–15.4 μm) and four visible channels at wavelengths of 0.4–0.94 μm, with a spectral resolving power of 1,200 (Aumann et al., 2003; Parkinson, 2003). As the Humidity Sounder for Brazil (HSB) failed after a short period in February 2003, data from the combined AIRS and AMSU instruments are used, namely, the AIRS/AMSU product (Lambrigtsen, 2003), which has been released and updated to version 6. However, power to the AMSU-A2 instrument on Aqua was lost on 24 September 2016, and attempts to restore it to normal operation were unsuccessful. Therefore, the version of AIRS+AMUS L2 and L3 products has not been updated after the AMSU-A2 instrument failure, but the AIRS-only version 6 products are almost as good as the combined AIRS+AMSU product, except under cloudy conditions (https://airs.jpl.nasa.gov/amsu_a2_anomaly). The AIRS-only L2 and L3 products are available for the full mission in addition to AIRS+AMSU products, and the detailed description of the AIRS product change is available online (https://disc.sci.gsfc.nasa.gov). Here, a comparison between AIRS L2, L3 products, and radiosondes observation (as the reference) are performed for the case study, which shows that AIRS L3 products have good consistence with radiosonde observations than AIRS L2 product (See Text S2 in the SI for the details description of the comparison results). In addition, the radiosondes usually have large drift distances at high levels, and their maximum distance can reach ~100 km (Figure 2), which is similar with the AIRS L3 resolution of 1° and is much greater than the AIRS L2 resolution of 50 km. Therefore, the AIRS (AIRS+AMSU and AIRS-only) L3 rather than L2 water vapor-retrieval products were used in our study, namely, the AIRS-only (AIRS + AMSU) q, and PWV products, which are version 6 and derived from L2 standard swath data with quality indicators of “best” or “good” (Olsen et al., 2013; Tian et al., 2013). The level 3 data with separate ascending (1330 BJT) and descending (0130 BJT) orbits are produced on a 1° × 1° latitude-longitude grid on daily, 8-day, and monthly resolution scales. The AIRS (AIRS+AMSU and AIRS-only) q retrievals at 1,000, 925, 850, 700, 600, 500, 400, 300, 250, and 200 hPa levels are used in the analysis. The data in June 2013 and June 2014 have been processed.

2.3 GPS Data

Here, the GPS PWV products from four International Global Navigation Satellite System Service (IGS) ground-based GPS sites (Beijing, Changchun, Shanghai, and Wuhan) were selected to compare with radiosonde PWV in June 2013 and June 2014. The twice-hourly (i.e., 0100, 0300, …, 2100, 2300 UTC) GPS PWV data set is available online (http://rda.ucar.edu/datasets/ds721.1/). The algorithms used to retrieve the GPS PWV are described in detail by Bevis et al. (1994), Wang et al. (2005), and Wang et al. (2007), and its product accuracy is roughly 1–2 mm. In order to minimize the impact of distance and elevation differences between GPS sites and sounding stations, we made the following matching criteria (Gui et al., 2017; Wang & Zhang, 2008): (1) the distance between the GPS and radiosonde stations is within 50 km; (2) the elevation difference between the GPS and radiosonde stations is within 100 m; (3) the PWV data derived from the GPS at 1300 and 1500 BJT are averaged to compare with the radiosonde data at 1400 BJT. On the basis of these conditions, the locations of the four GPS sites are shown in Figure 1, and Table 1 summarizes the locations, distances, and height differences of the selected GPS and radiosonde sites.

2.4 MODIS Data

Given the low spatial resolution of AIRS products, the subgrid cloud features cannot be well resolved. In contrast, a matching MODIS cloud product, to some extent, indicates whether a cloud in an instantaneous AIRS field exists or not and, thus, the testing of AIRS cloud products and the improved verification are mainly based on the cloud-product information in the MODIS atmospheric products as the testing standard (Li et al., 2005). Therefore, the L3 MODIS-Aqua CF and cloud-top-pressure data of the daily average (MYD08_D3) global grid over China in June 2013 and June 2014 were used. The spatial resolution of the MYD08_D3 daily data is 1° × 1°.

3.1 Estimation of q and PWV From Radiosonde Data

While sounding data are generally regarded as the atmospheric truth when evaluating satellite data, the temperature and dew point temperature from the soundings must still undergo quality control. Here, the Chinese L-band second-order observation basic data set (version 1.0) has undergone the quality control described in Zhai and Eskridge (1997). The water vapor mixing ratio q and PWV are calculated by (Wong et al., 2015, see also Alduchov & Eskridge, 1996 and Zhai & Eskridge, 1997)

(1)

where e = RHe_s(T) and e_s are the vapor pressure and the saturation vapor pressure (Pa), respectively, RH is the relative humidity (%), p is the pressure (Pa), and the vertically integrated water vapor (mm),

(2)

where p_i and p_i + 1 are the pressure levels (Pa) and g (= 9.806 m/s²) is the acceleration due to gravity, from which PWV may be calculated between the pressure levels p_i and p_i + 1.

3.2 Matching and Comparison Methods

Intercomparison analyses conducted between satellite (grid scale) and ground-based measurements (point) remain challenging. Here, when using the sounding data to validate the AIRS product (For the convenience, AIRS-only and AMSU+AMSU products all abbreviated to AIRS products), the key issue is how the impacts of balloon drifts on the representativeness of sounding data match with the corresponding AIRS grid value. Therefore, we conducted several experiments to investigate how the drift varies at different heights induced by balloon transport in the troposphere, taking the maximum drift distances at 200, 500, and 850 hPa as examples. Figures 2a–2c shows the spatial distribution of the maximum horizontal balloon-drift distances at 200, 500, and 850 hPa, respectively. At 200 hPa, the maximum horizontal drift distances for most balloons is less than 70 km, while less than 25 and 10 km at 500 and 850 hPa, respectively. Figure 2d shows the corresponding frequency distribution of horizontal maximum drift distances, which are all less than 100 km, with distributions at 200, 500, and 850 hPa varying from 0–90, 2–25, and <10 km, respectively, and are less than the mean distance between adjacent two stations of 100–200 km, which can be considered as representative of sounding observations. In contrast, as the AIRS L3 data are also spatially averaged to a ≈110 km × 110 km grid, we suggest that the spatial representation of radiosonde water vapor and the corresponding water vapor at the AIRS L3 grid are similar. Therefore, for exact spatial-temporal matches, daily AIRS 1° × 1° products were bilinearly interpolated to match the locations (latitude and longitude) of sounding sites in the horizontal direction and to match the standard sounding profiles in the vertical direction at daytime (≈1330 BJT), following previous studies on satellite product validations (Chahine et al., 2006; Divakarla et al., 2006; Prasad & Singh, 2009).

To make the comparisons at the same vertical layers, based on the AIRS standard layer definition, we further calculate the value of q from radiosondes at each standard pressure layer. Then, the applicability of the horizontal direction is derived from the correlation coefficient (R), root-mean-square error (RMSE), and the mean bias (MB) of the AIRS q and sounding q values. The values of R, RMSE, RMS percent difference (Tobin et al., 2006), and MB of q between the AIRS retrievals and radiosonde observations are estimated at every station and for each standard (AIRS) vertical layer and then averaged vertically at every site. Comparing the AIRS q profile with the radiosonde q profile illustrates the difference in the vertical direction. The average values of RMSE, RMS, and MB for each station on each isobaric surface over the seven subregions are calculated to characterize the overall situation of the q deviation between the AIRS data and the observed data in each region. Moreover, to examine its profile, water vapor in the layers from the surface to 200 hPa, surface to 700 hPa, 700 to 400 hPa, and 400 to 200 hPa are calculated.

3.3 Statistical Metrics

For assessment of the performance of the AIRS product, we used the four statistical metrics R, RMSE, MB, and RMS defined by equations 3–6, respectively (Shi et al., 2018; Wong et al., 2015),

(3)

(4)

(5)

(6)

where n represents the number of samples, x_i indicates the value of x for sample i, that is, the radiosonde measurements (e.g., q, PWV), y_i indicates the value of y for sample i, that is, the satellite-retrieved and GPS products (e.g., AIRS q, AIRS PWV, and GPS PWV), denotes the average of x_i, and denotes the average of y_i. The correlation coefficient R is used to reflect the degree of the linear relationship between two variables, the root-mean-square error RMSE quantifies the differences between two variables, and the mean bias MB and the RMS percent difference respectively indicate the absolute and relative degree of overestimation/underestimation of each variable.

4.1 Horizontal Comparisons Between AIRS q and Radiosonde q

The correlation coefficients between the AIRS q and radiosonde q products are first calculated at each AIRS standard vertical layer for each site and then averaged vertically at every site. Figure 3 shows the geographical distributions of vertically averaged R values between the radiosonde and AIRS q products at 1400 BJT in June 2013 and June 2014. Among the 113 stations, 79 stations have R > 0.6, indicating that the AIRS satellite retrieval data and radiosonde observations have high correlations at these sites, whereas 33 stations have 0.45 < R < 0.6, indicating that both data sets are moderately correlated at these sites. In addition, only one site has a small R value (0.39), but the site still passes the 95% confidence test. The R distribution of the two kinds of q data sets in China shows a clear geographical difference as follows: the R value is relatively high with the maximum of 0.83 in northwest China and the Tibetan Plateau, followed by northeast China, north China and the Yangtze River Valley, with R values lowest in the southwest (including the Sichuan Basin, Yunnan) and south China, with an average value of approximately 0.57. Overall, most stations have high R values with an average value of 0.64, which indicates that the AIRS q data and the radiosonde q data are highly consistent.

As a correlation analysis verifies the consistency between the AIRS q and radiosonde q products without indicating any absolute differences, the metrics MB and RMSE can be used to characterize the deviation between satellite-retrieved data and radiosonde observations. Figure 4 shows the RMSE and MB values (AIRS—radiosondes) at each ground station, with the sizes and colors of the circles representing RMSE and MB values, respectively. The RMSE value shows a decreasing trend from southeast to northwest China. For seven regions, the largest RMSE value is mainly located in south China (≈1.51 g/kg), followed by the Yangtze River Valley (1.49 g/kg), and the southwest regions (1.21 g/kg), with smaller RMSE values appearing in north China (0.93 g/kg), northeast China (0.84 g/kg), northwest China (0.71 g/kg), and the Tibetan Plateau (0.63 g/kg). For each single station, the maximum (minimum) RMSE value ≈1.72 g/kg (0.37 g/kg) occurs in the Xisha (Beitashan) station (See Figure S1 in SI). The overall RMSE value over China is small. Sixty stations, which account for 53% of the total number of stations, have RMSE < 1 g/kg. The spatial differences of the MB values of the two data sets are evident over China, which are all negative in the southern of the Yangtze River, and nearly negative to the northeastern of the Yangtze River, the eastern coastal areas, and northeast China, with positive values in northwest China and the Tibetan Plateau (except for two sites). The largest negative value is −0.89 in Ganzhou station, Jiangxi Province (see Figure S1 in SI), whereas the largest positive value is 0.76 in the Jing River station, Shaanxi Province (see Figure S1 in SI). The MB value is negative in 52% of the sites, which indicates that the AIRS q values are lower than the radiosonde q data. However, that the AIRS q products overestimate the value of q when the water vapor content is low explains the difference in our results (see Qin et al., 2012). The absolute value of MB for 93 stations is within 0.5, which is 82% of the total number of stations and indicates that the deviation between the two data sets in China is small. The CF values indicate the fraction of a pixel covered by the cloud. Interestingly, Figures 3, 4, and 5a show that the areas with large differences of water vapor content correspond to the high-cloud-coverage zones. As the cloud fraction decreases from south to north, the errors of AIRS q data are relatively larger in the south due to the low retrieval accuracy of AIRS q, which is mainly influenced by high CF in these areas (refer to Susskind et al., 2003, 2006).

4.2 Vertical Comparisons Between AIRS q and Radiosonde q Products

Vertical profiles of water vapor mixing ratio derived from the AIRS and radiosonde observations directly reflect any vertical differences. The distribution of the average vertical profile of the AIRS q and radiosonde q at 1400 BJT in June 2013 and 2014 over seven subregions are shown in Figure 6. Due to the topography and altitude, the northwest and southwest China do not give consistent results below 1,000 hPa. The Tibetan Plateau has high altitudes (≈700 hPa) and no data below 700 hPa. Figure 6 shows that both the AIRS and radiosonde-derived q in the different regions of China decrease with increasing height. From 500 to 200 hPa, the vertical distributions of the two data sets in the seven regions are the same, and less than 4 g/kg, which indicates a high degree of coincidence. Below 500 hPa, especially at low levels, the differences between the two data sets are larger. South China, the Tibetan Plateau, and southwest China have relatively large deviations in the lower layers, which are mainly related to the elevation of the plateau, the topography of southwest China, and the abundant water vapor in south China (Zhai & Eskridge, 1997). The interpolation techniques within the subregions at different sites can also cause certain deviations. From 500 to 200 hPa, we find that the AIRS q values over the Tibetan Plateau and northwest China are higher than the radiosonde q values, whereas that of the other five regions are lower than radiosonde q values. Different areas have varying vertical profiles of q, with q values in the lower troposphere in south (northwest) China the largest (smallest). The corresponding q values for the AIRS and radiosonde observations at 1,000 hPa are 18.6 g/kg (8.53 g/kg) in south China and 17.8 g/kg (10.04 g/kg) in northwest China, respectively. For the Tibetan Plateau, the two kinds of q data at 700 hPa correspond to values of 6.02 and 4.84 g/kg. Below the height of 850 hPa, the AIRS q values are generally larger than the radiosonde q values across almost all regions, except for over northwest China. The satellite-retrieved data may be affected by the near-surface complexity below 850 hPa altitude (Zhao et al., 2012). The vertical profiles of AIRS and radiosonde q values show good consistency with the changes in height, with all reasonably reflective of the variation characteristics of q with decreasing height over the seven subregions.

Figure 7 illustrates the magnitudes of RMSE, RMS, and MB of the AIRS and radiosonde q data sets for the seven subregions. The RMSE values have a decreasing trend with increasing height (Figure 7a). Above 400 hPa, the RMSE values of the seven areas are within 1 g/kg. Apart from the fact that the upper atmosphere is not affected by near-surface influences, the value of q in the upper atmosphere is considerably smaller than that in the lower layers. The vertical distribution of RMSE values of q in south China and the Yangtze River Valley are similar, where the RMSE values corresponding to each layer are largest relative to the other subregions (3.67 and 3.4 g/kg), appearing at the height of 850 hPa. The RMSE values at each altitude in southwest China are also larger than those in other subregions, with a maximum value of 3.03 g/kg at 700 hPa. The vertical distributions of RMSE values in north, northeast, and northwest China are similar, with the RMSE reaching maximum values (2.54, 2.38, and 1.7 g/kg) at 925, 1,000, and 850 hPa, respectively. Figure 7b shows that, at 1000–250 hPa, RMS percent differences range from 20 to 50%, and all increasing with height. Above 250 hPa, the RMS values in each region are relatively large, with the maximum value (more than 80%) found in the Northwest China. Maybe it is related to radiosonde data scarcity and poor quality of radiosonde humidity data in the upper troposphere. Figure 6c shows that, above 700 hPa, the mean biases of the seven regions differ significantly. Above 400 hPa, all MB values decrease to nearly zero with increasing height. For 1,000–700 hPa, the maximum absolute MB values are found in the lower troposphere over south China, the Yangtze River Valley, north China, and northeast China, with the maximum MB value of −2.56 g/kg over south China. According to previous studies (Divakarla et al., 2006; Susskind et al., 2003, 2006; Tobin et al., 2006), the retrieval errors of q profiles are mainly induced by the cloud height, which can be directly represented by the cloud top pressure related to the magnitude of the cloud top height (Figure 5b). By comparison of Figures 5b and 7, we find that the larger vertical MB values are mainly located at the existing cloud heights or below for most regions, where the cloud top pressures are >500 hPa (Figure 5b). Therefore, we deduce that the AIRS q errors at the lower to middle layers are mainly caused by vertically distributed clouds.

In general, the above results suggest that the cloud fraction and cloud height are the main causes of the horizontal and vertical differences of AIRS q and radiosonde q values in the summertime (June) over China. That the accuracy of the AIRS q product is mainly affected by clouds is because clouds have a significant effect on the observed infrared radiances used in the temperature and moisture profile retrievals, with smaller effects on channel noise (Aumann et al., 2003; Rosenkranz, 2001). However, we suspect that the quality-controlled and revised sounding observations can correct inversion errors caused by the clouds to a certain extent, which is worth exploring in future work. Note that, in the upper troposphere, the GTS1 humidity data are (possibly) of poor quality and cannot pass quality control that may be related to low-temperature environment and humidity slants dry, which may be contributing to the large differences of horizontal and vertical above 300 hPa (e.g., RMS).

4.3 Intercomparison Between AIRS PWV and Radiosonde PWV Products

The PWV is the total atmospheric water vapor contained in a vertical column per cross-sectional unit (Ichoku et al., 2002; King et al., 1992; Shi et al., 2018). To perform a reliable comparison between the radiosonde and AIRS PWV products, radiosonde PWV data are first assessed using ground GPS data, which typically share similar techniques and can also successfully retrieve the PWV. The comparison of ground-based GPS PWV with radiosonde PWV is presented in Figure 8 at 1400 BJT for Beijing, Changchun, Shanghai, and Wuhan. The GPS PWV and radiosonde PWV at the four stations show excellent agreement, which is consistent with previous results (Gui et al., 2017), with R values of 0.95, 0.9, 0.98, and 0.98 for the Beijing, Changchun, Shanghai, and Wuhan sites, respectively. Moreover, the mean MB and RMSE values of the four stations are 1.77 and 3.97 mm, respectively. With the exception of the MB value of the Changchun station (−0.83 mm), the MB values of all stations indicate that the GPS PWV is slightly higher than the radiosonde PWV because of the difference in the length of the data used, the available PWV data at each station, and the different detection techniques adopted (Jiang et al., 2016).

The spatial distributions of the AIRS-derived mean PWV and radiosonde PWV (dots) are shown in Figure 9. The radiosonde PWV in June shows a gradual decline in magnitude from the southeast coast to the northwest inland, reaching maximum values of 57 mm in south China. The second largest value appears in the Yangtze River Valley and the Sichuan Basin, where the PWV is approximately 45 mm. The PWV in Xinjiang and Inner Mongolia is below 25 mm, and the PWV in the Tibetan Plateau is the lowest (below 10 mm). The variation in PWV is mainly associated with temperature and the source of water vapor, which is consistent with previous research by Zhai and Eskridge (1997). The limited amount of radiosonde stations cannot represent the spatial distribution of PWV over all of China, so the spatial variation of PWV derived from the AIRS data set is given. A large difference of the averaged radiosonde PWV at 1400 BJT is well captured by the AIRS PWV. The PWV obtained by the two detection techniques give a consistent spatial distribution, but the AIRS PWV tends to underestimate the PWV when considering the radiosondes as a reference, especially in south China.

To further quantify the differences between the two types of data, the scatterplots of AIRS PWV and radiosonde PWV in China and its seven subregions have been plotted in Figure 10, illustrating excellent consistence, with the R, RMSE, and MB values of 0.91, 7.68, and −2.79, respectively. For the seven subregions, except for south China, the correlations for north China, northwest China, northeast China, the Tibetan Plateau, the Yangtze River Valley, northeast China and southwest China are all significant, with R values of 0.89, 0.82, 0.8, 0.76, 0.74, and 0.68, respectively. The corresponding MB values are −0.06, −0.05, −2.04, −4.49, −1.99, and −5.7, and RMSE values of 4.63, 4.16, 4.23, 10.43, 5.16, and 9.92, respectively. The AIRS PWV tends to be systematically lower than the radiosonde PWV, and the difference between the two kinds of PWV also increases when the PWV is large. However, generally, the scatterplots of radiosonde PWV and AIRS PWV demonstrate a close correspondence, which gives confidence that AIRS PWV product can be used as an alternative when the radiosondes are unavailable.

The PWV can explain the overall horizontal distribution of water vapor, but give no vertical information. To further study the vertical distribution of water vapor content over China, and compare the AIRS layer data with radiosonde data at different altitudes, we calculate the water vapor content in the following four layers: surface to 200 hPa, surface to 700 hPa, 700 to 400 hPa, and 400 to 200 hPa, representing the whole, lower, middle, and upper troposphere, respectively (Zhai & Eskridge, 1997). We focus on the radiosonde data, whereas the AIRS data are used for comparison. Figure 11 describes the spatial distributions of the mean of layered water vapor (LWV) for the (a) whole troposphere, (b) lower troposphere, (c) middle troposphere, and (d) upper troposphere from AIRS and radiosonde (dots) data at 1400 BJT in June 2013 and June 2014. The radiosonde LWV illustrated in Figure 11 shows that approximately two thirds (66%) of the water vapor content is stored in the lower troposphere. The water vapor rapidly decreases with increasing altitude, and the upper troposphere has less than 5% of the total water vapor content in the whole troposphere. For the Tibetan Plateau, the layer from the surface (average at 649.5 hPa) to 400 hPa is thinner than that in the other regions (Zhang et al., 2013), and this layer contains approximately 84% of the total water vapor content. Moreover, the maximum and minimum zones at the lower, middle, and upper layers do not exactly match vertically. The differences of water vapor content between the AIRS and radiosonde data are mainly observed in the lower troposphere, where the largest differences are at Guangdong, Guangxi, and Fujian Province (see Figure S1 in SI), followed by northeast China, the Yangtze River Valley and the Sichuan Basin; the differences in the other regions are relatively small. In the middle troposphere, the areas with clear differences are mainly distributed in the Yunnan-Guizhou Plateau and the Sichuan Basin. In the upper troposphere, the spatial distribution of the water vapor content between the AIRS data and radiosonde data is relatively consistent, and the difference is negligible, which is related to the fact that the upper troposphere is not affected by the near-surface layer and considerably lower levels of high LWV.

4.4 Impact of Cloud Fraction on the AIRS PWV

As cloud contamination often leads to uncertainty or artifacts in the satellite inversions (Foster et al., 2006), to examine the effect of the CF on the correlation between the AIRS PWV and radiosonde PWV, the seven subregions are further analyzed by obtaining the CF through MODIS CF data, matched with the corresponding AIRS PWV timestamp and sorted by the CF value in ascending order. The samples that correspond to the smallest (largest) one third of the cloud value have been classified as low–(high–) cloud cover conditions, and the others as partly cloudy conditions. Table 2 presents the statistical information of three bins for the CF containing the value range and mean CF for each bin in each area. The mean CF values of the low-cloud-cover conditions in southwest China, the Yangtze River Valley and south China exceed 0.4. The cloudiness in these areas is relatively high in summer, and the range of values of overcast conditions is above 0.9, which indicates that cloud cover increases in summer over China.

The effect of the CF on the correlation between the AIRS PWV and radiosonde PWV is investigated by dividing the paired data into three equal-sample bins. Figure 12 shows that the value of R tends to decrease when all paired data are taken under highly cloudy conditions in China and the seven subregions, in comparison with that under low–cloud cover conditions. With the increase of CF, the R values are reduced by 13.75% (from 0.8 to 0.69), 9.89% (from 0.9 to 0.82), 19.54% (from 0.87 to 0.7), 22.62% (from 0.84 to 0.65), 30.67% (from 0.75 to 0.51), 42.7% (from 0.89 to 0.51), and 87.95% (from 0.83 to 0.12) for northeast China, north China, northwest China, the Tibetan Plateau, southwest China, the Yangtze River Valley, and south China, respectively, which indicates that the CF exerts a significant impact on the correlation between the AIRS PWV and radiosonde PWV. Moreover, RMSE values increase with the increase of CF, MB values vary from positive to negative (except those for the Tibetan Plateau, southwest China and south China), whereas the absolute values increase from relatively small to larger magnitudes. The CF considerably affects the accuracy of the AIRS PWV-retrieved products, with differences under overcast conditions significantly higher than those under low–cloud cover conditions. As the statistics for south China show that the accuracy of AIRS PWV is low because of the high cloudiness in summer, caution must be taken when using AIRS data in this region during the summer.