The GUAN radiosonde network has been established by GCOS in order to make current and historical upper air data available for climate change detection and climate monitoring. GUAN provides global radiosonde observations, from homogeneously distributed upper air stations, that have a specific record length in addition to meeting the continuity requirement and data quality requirements as defined by GCOS (Daan, 2002). At present there are 171 GUAN stations worldwide. A station map and a station list can be found at http://www.wmo.int/pages/prog/gcos/index.php?name=ObservingSystemsandData. The GUAN data are distributed by the Global Telecommunication System and archived at the Deutscher Wetterdienst. The processing of GUAN data at the Deutscher Wetterdienst was consistently done, with one exception: in October 2008 the archiving system at the Deutscher Wetterdienst was changed and the GUAN processing software was adapted. However, the results of the comparison between ATOVS and GUAN data records do not exhibit any distinct feature in October 2008.

The quality of radiosonde observations is affected by a series of issues such as temporally and spatially varying radiosonde types and national practice (e.g. Soden and Lanzante, 1996; Christy and Norris, 2009; Moradi et al., 2010), as well as issues and differences in calibration procedures (e.g. Miloshevich et al., 2006; Vömel et al., 2006). Among the strongest impacts is the dry bias caused by solar radiation (Vömel et al., 2006), which leads to significant underestimations of humidity in the upper troposphere if not corrected. A series of correction algorithms have been developed by, for example, Miloshevich et al. (2004), Leiterer et al. (2005) and Miloshevich et al. (2009), which mainly focus on RS80 and RS92 radiosonde observations. Such corrections have not been applied to the utilised GUAN observations.

Examples of reprocessed radiosonde archives which include temperature and water vapour are the integrated global radiosonde archive (Durre et al., 2006) and its homogenised version (Dai et al., 2011). Dai et al. (2011) describe a few known discontinuities in humidity observations from radiosondes. These are as follows: the dew point depression was set to 30 ∘C under dry conditions at several stations, and temperature observations under cold conditions for “early radiosonde hygrometers” were unreliable and were reported as missing (Dai et al., 2011).

AIRS is an infrared cross-track-scanning instrument onboard the NASA Aqua satellite which also carries an AMSU-A radiometer. The NASA Aqua satellite has been in orbit since 2002. The level 2 AIRS data record which is used for comparison is the AIRX2RET product provided by the NASA Goddard Earth Science Data and Information Service Center (http://daac.gsfc.nasa.gov/); this product is based on AIRS and AMSU-A observations. The processing version is V5.0 for the data from 2002 to 30 September 2007 and V5.2 for data from 1 October 2007 onwards. AIRS L2 products come in swath-based 6 min length files, with 240 files covering 1 day. The products have a spatial resolution of 50 km × 50 km and the profiles are defined on 14 layers at 1000, 925, 850, 700, 600, 500, 400, 300, 250, 200, 150, 100, 70 and 50 hPa. The format is HDF4. The AIRS Standard Products consist of, among other things, cloud properties and profiles of temperature and water vapour. The products are the results of employing the combined AIRS-IR/AMSU-A microwave retrieval, which is described in detail in Susskind et al. (2003, 2006, 2011). The retrieval process also involves a cloud-clearing process which assumes that the radiative properties in each field of view are a function of cloud fraction only. A retrieval solution is rejected when the cloud fraction is larger than 90 %. A scattering (rain) index is not explicitly applied. Infrared and microwave surface emissivities are part of the solution. For the comparison to the ATOVS data record, the data field Qual_H2O was evaluated and “best” and “good” quality data were used in the comparison.

An evaluation of the AIRS version 5 TPW products' accuracy is given in Bedka et al. (2010), who compared the satellite products to ground-based observations at selected Atmospheric Radiation Measurement (ARM) sites. Using ground-based microwave radiometer observations at Barrow, Southern Great Plains–Lamont (SGP) and Nauru, the authors found an average relative error which is typically smaller than 5 % for all sites, except at SGP, where AIRS products are too moist when TPW is less than 10 kg m-2. At SGP, night-time observations exhibit a dry bias of approximately 10 % when TPW is greater than 10 kg m-2 (Bedka et al., 2010).

Recently, the AIRS version 6 products have been released. Improvements over version 5 are described at http://airs.jpl.nasa.gov/data/v6/; the first comparison results of version 6 and version 5 products to ECMWF can also be found on the webpage. The AIRS version 6 products may still be improved, and product validation and quality assurance are ongoing. Thus, its product validation state is “provisional”.

Three data records are used for the evaluation: ATOVS and GUAN data records for the period 1999–2011 and AIRS data record for the period 2003–2011. The data have not been collocated and GUAN data have only been used when at least two observations per day are available. ATOVS and AIRS data records exhibit similar annual cycles. However, a systematic difference between both data records is evident. This is discussed in Sect. 4.2.3. The annual cycle of the GUAN time series has larger amplitudes than the annual cycles of the satellite time series (not shown). The GUAN stations are located on islands and over land, with the majority of stations in the continental Northern Hemisphere. Schröder and Lockhoff (2013) show that the strength of the annual cycle is a function of region: strongest annual cycles are associated with the monsoon regions and the propagation of the ITCZ, largest regions of minimum strength are found over the oceans of the Southern Hemisphere, and land areas typically exhibit strong annual cycles. The former explains the presence of an annual cycle in the satellite data due to the imbalance in strength between the Northern and the Southern Hemisphere and the latter in combination with the asymmetric sampling between the Northern and Southern Hemisphere explains the annual cycle in the GUAN data. The annual cycle in TPW from GUAN has slightly larger amplitudes in 1999 and 2000 than from 2001 onwards (not shown). The larger amplitudes in the first 2 years are caused by stronger minima during boreal winters. When looking at the time series of deseasonalised anomalies (not shown), a breakpoint is found between June and July 2001. The strength of the breakpoint is computed as the difference between the average TPW using a period of 24 months prior to and after the break. Through use of this strength and the averaged standard deviation over the two periods as input to a two-sided t test, it is found that the strength of the change is associated with a coverage probability of 92 %. Thus, the breakpoint is not considered to be significant when applying a standard significance level of 0.05. In the following the strength and the coverage probability are computed and interpreted the same way.

The ATOVS TPW data record exhibits a breakpoint between January 2001 and February 2001 (not shown). The difference in TPW between the years 1999–2000 and 2002–2003 is 2.8 kg m-2 with a coverage probability of 99 %. Moreover, the annual cycle of TPW exhibits stronger minima during boreal winters during the first 2 years. This breakpoint is analysed in more detail in the following.

First, the breakpoint does not temporally coincide with the start of the use of SNOs in January 2001. Moreover, no breakpoint is visible between December 2010 and January 2011, when the use of SNOs ends.

Second, we assess the impact of Kriging on the homogeneity of the time series. We compared the PDF based on the CM SAF ATOVS data record products and ATOVS products, which have been arithmetically averaged on basis of daily values. Figure 4 shows the PDFs of TPW values for June 2002 separately for the CM SAF ATOVS data record products and for the arithmetically averaged monthly means. Obviously, the distribution of the CM SAF ATOVS data record products exhibits an increased number of TPW values at the high end of the distribution. This is reflected in the monthly mean TPW of 25.3 kg m-2 for the classically averaged data and of 26.2 kg m-2 for the CM SAF data record products, which gives an average difference of 0.9 kg m-2. Apart from sampling, gaps are caused by strong scattering events, e.g. in the presence of strong precipitation. During gap filling, Kriging uses valid observations in the vicinity of the gaps. The gaps neighbouring areas are typically humid, and thus Kriging fills these gaps with generally large values (see also Schröder et al., 2013). This explains the increased frequency of occurrence at the high end of the TPW distribution. The PDF does not change significantly when more than two satellites are used (not shown). The PDFs of the classically averaged monthly means for NOAA-15 and NOAA-16 alone are also shown in Fig. 4. The difference between the arithmetically averaged monthly means for NOAA-15 and NOAA-16 is 24.7 and 25.2 kg m-2, respectively, thus leading to a difference of 0.5 kg m-2.

Finally, a specific feature of Kriging is discussed. Kriging requires two independent measurements such as those from different satellites or from the morning and afternoon overpasses of a single satellite. For the period January 1999 to February 2001, only NOAA-15 was available. Then, it may happen that the same location is not observed twice a day, e.g. due to the occurrence of strong precipitation events. When this happens, Kriging is not applied and the daily average is flagged as undefined. For the June 2000 case study, the number of valid observations is 12 % smaller in the Kriging product than in the arithmetically averaged product and the number of undefined values is 9 % larger in the Kriging product than in the arithmetically averaged product. Indeed, it is visible that the positions of minima in the number of valid observations and of undefined values coincide with the position of the Intertropical Convergence Zone (ITCZ) (not shown).

The methodology for the comparison of the ATOVS data record against the GUAN radiosonde data record is as follows. First, the GUAN data record is integrated to match the vertical layer and level definitions of the ATOVS data record water vapour products. For each day, only stations with at least two radiosonde launches per day are used and averaged to daily values. The ATOVS data record is spatially collocated to the position of the GUAN stations using a nearest-neighbour algorithm. The collocated daily averages form the basis for the comparison. We analyse the monthly bias and the bias-corrected root mean square error (RMSE) between ATOVS and GUAN data records. The number of valid collocations per month is greater than or equal to 450. The results shown in this section show bias and RMS based on all valid daily averages. Note that potential dependencies on climate regimes, TPW, and other regional dependencies are not resolved here. We expect occasionally larger bias and larger RMS on regional scale.

First of all, the difference in quality between the reprocessed and operational ATOVS products is discussed. Figure 5 presents the comparison results between TPW from the reprocessed and operational ATOVS data records and the GUAN data record. Figure 5 clearly shows that the TPW product from the reprocessed ATOVS data record exhibits a better quality and stability than the TPW from the operational ATOVS product. The bias of the operational ATOVS product compared to the GUAN data record shows a significant breakpoint between April and May 2009. At this time the following changes had been implemented in the operational ATOVS processing chain: migration of the processing chain, update of AAPP and IAPP, removal of NOAA-15 and NOAA-18 observations from the retrieval, and implementation of Metop-A and NOAA-19 observations in the retrieval. The obvious improvement for the reprocessed data record is that the breakpoint in the bias time series is largely reduced, and this also leads to an improved averaged bias. See Sect. 4.2.3 for further discussion. Moreover, the RMSE is slightly smaller for the reprocessed data record than for the operational product.

We now focus on the comparison between the reprocessed ATOVS data record and the GUAN data record that is shown in the left panel of Fig. 5. The TPW bias is typically smaller than 1 kg m-2 and the average bias is -0.16 kg m-2; furthermore, the bias is stable. Interestingly, no breakpoint is observed in the bias time series between 2000 and 2001. The breakpoints in the averaged, non-collocated TPW time series of GUAN and ATOVS have the same direction and a different strength and occur with a temporal difference of 5 months. This might translate into a breakpoint in the bias time series which is smaller than the ATOVS breakpoint itself and which is overlain by an anomaly between February and June 2001. This is not evident in the bias time series due to the collocation process: gap filling is mainly applied in the presence of strong precipitation. Associated areas are typically found in the ITCZ and storm track regions with poor coverage of GUAN stations. Due to the collocation procedure, data from these areas have a reduced impact on the breakpoint. In fact, the ATOVS time series of collocated TPW values exhibits a breakpoint which is reduced by 0.7 kg m-2. The GUAN data record and the ATOVS data record are sampled in a similar way; consequently, the collocated GUAN data also exhibit a breakpoint between January and February 2001, and the breakpoint is not evident in the bias time series.

The averaged RMSE is 3.25 kg m-2 and the RMSE is stable from 2001 onwards, with values around 3 kg m-2. The RMSE exhibits maximum values in the first 2 years of the time series. This is expected since, for the years 1999 and 2000, only one satellite is available for the processing, while for the rest of the processing at least two satellites are used. This behaviour was also observed in a comparison between SSM/I-based and ERA-Interim TPW products: the RMSE decreased with the transition from a single-satellite product to a multiple-satellite product (Schröder et al., 2013). Finally, in contrast to the bias, the RMSE exhibits an annual cycle with maxima during boreal summers. When analysing the standard deviation of TPW from GUAN data record (not shown) a pronounced annual cycle is visible with a sharp increase in standard deviation between May and July (maximum value: 4.3 kg m-2). The maximum in July is followed by a slow decrease until February (minimum value: 2.3 kg m-2). Besides a potential dependency of the uncertainty of the radiosonde observations on TPW, we argue that the dominating factor for the annual cycle in RMSE is that, with increasing TPW values during boreal summers, the natural variability in water vapour also increases and that the increase in natural variability enhances the representativity uncertainty between the point and areal observations.

Figure 6 presents the comparison results between the ATOVS and the GUAN LPW products, again in terms of bias and RMSE. The LPW bias for layer 5 (surface–850 hPa) is around -0.7 kg m-2, while the LPW biases for layers 3 and 4 (850–700 and 700–500 hPa) are between 0 and 0.6 kg m-2. As for the TPW, a slight increase in bias is present in early 2009, and an explanation for this increase is given in Sect. 4.2.3. The LPW biases for layers 1 and 2 are relatively small, with values around 0.002 and 0.003 kg m-2, respectively. The LPW bias for layer 2 exhibits an unexplained anomaly of approximately 0.2 kg m-2 in 1999. Maximum RMSE values are slightly larger than 2.5 kg m-2 and are found for layer 5. LPW RMSE values for layers 2 to 4 exhibit a decrease after the first 2 years, as was found for TPW. Furthermore, the RMSE typically exhibits an annual cycle similar to the TPW RMSE.

In view of the results shown in Fig. 3 we briefly want to characterise the relative bias of the ATOVS specific humidity product (not shown). The relative bias increases with height and ranges from 4 % at 1000 hPa to approximately 90 % at 200 hPa, with this latter value being 3 or more times larger than the values of the other levels and with ATOVS being more humid than the radiosondes. This may again partly be explained by the dry bias in radiosondes. However, the relative bias is of similar order to the maximum values given in Fig. 3. This may point to a wet bias in the ATOVS product in the upper troposphere. However, a verification is hard to accomplish due to the lack of fully independent and high-quality reference data.

When relative values are considered (not shown), the bias is the smallest (largest) for LPW5 (LPW1) with average values of -6 and 36 %, respectively. The moist bias in the upper troposphere can be expected due to the observed dry bias in uncorrected radiosonde observations (e.g. Soden and Lanzante, 1996). The relative RMSE systematically increases with height. The lowest layer has an average relative RMSE of 30 %, whereas for the highest layer this is 107 %. Our results exhibit similar magnitudes to the results presented in Reale et al. (2012), who compared humidity profiles from various satellite products, among them the NOAA ATOVS product (Reale et al., 2008), with the Global Telecommunication System radiosonde data and found typical values of 25–50 % below 400 hPa and of 100 % or more in the upper troposphere.

The GUAN radiosonde data and ATOVS are assimilated in the ERA-Interim reanalysis. Consequently, the bias and RMS of the comparison between the ATOVS data record and the GUAN radiosonde data might be underestimated due to this dependency. Although it is outside of the focus of this work, we briefly note again that various uncertainties contribute to the observed differences. Here, we compare point measurements with areal observations. Thus, the representativity uncertainty impacts the observed differences. To our knowledge the representativeness uncertainty is not known at each GUAN station. However, for assimilation purposes, high-resolution models have been used to assess such uncertainties. An analysis example is given in Waller et al. (2014) for specific humidity – they found a strong dependence of the representativity uncertainty on height and weather state. Furthermore, the comparison of the ATOVS and the GUAN data is based on daily averages and the differences in sampling between the radiosonde and the satellite observations contribute to the observed differences. To give an example of the diurnal sampling uncertainty, we use the work of Dai et al. (2002). Using high-temporal-resolution Global Positioning System data from stations over North America, they found that the uncertainty in seasonally averaged TPW is within ±3 % or ±0.5 kg m-2 when the sampling is changed from 30 min to twice daily at 00:00 and 12:00 UTC. Finally, we refer the reader to the work of Pougatchev et al. (2009) and Sun et al. (2010), who compared temperature and relative humidity profiles from radiosonde data to IASI and to the Constellation Observing System for Meteorology, Ionosphere, and Climate observations in order to analyse the uncertainty arising from temporal and spatial mismatches.

In order for it to be possible to compare the ATOVS data record with the AIRS data record, the AIRS profiles are vertically integrated according to the ATOVS layer definitions (see Sect. 2). Then, the swath-based products are gridded onto the ATOVS spatial grid, and finally all data are averaged to obtain monthly means, which form the basis for the comparison. The number of valid collocations per month is typically larger than 60 000.

Figure 7 presents the comparison results of AIRS and ATOVS TPW products. It can be seen that the TPW bias changes from approximately 1 kg m-2 in 2003 to approximately 2 kg m-2 in 2011. A breakpoint is present between April and May 2009 which temporally coincides with the removal of NOAA-15 data from ATOVS processing. The strength of the breakpoint is 0.5 kg m-2 and exhibits a coverage probability of 97 %. The RMSE is relatively stable, with values around 2.4 kg m-2.

The LPW bias is shown in Fig. 8 and exhibits similar features as the bias for TPW, except for the LPW bias for layer 5, which exhibits an annual cycle. The breakpoint observed in the comparison of TPW in early 2009 is also evident for LPW for layers 3 to 5. Relative to the TPW bias and the LPW bias for layers 3 to 5, the RMSE is stable over time. The LPW bias for layer 1 and the LPW RMSE for layer 3 exhibit a distinct feature between late 2005 and early 2009. This coincides with the use of NOAA-18 observations in the retrieval while MHS onboard this particular satellite experienced a series of technical issues (see http://www.ospo.noaa.gov/Operations/POES/NOAA18/mhs.html).

The RMSE between the ATOVS and the AIRS data records does not exhibit a pronounced annual cycle and is typically smaller than the RMSE between the ATOVS and the GUAN data records likely because the number of valid collocations is larger and equally distributed over the Northern and Southern Hemisphere and because the comparison of point measurements with areal observations likely exhibits larger representativity uncertainties than the comparison of two areal observations. However, the biases for TPW and LPW are larger between the ATOVS and the AIRS data records than between the ATOVS and the GUAN data records. The relatively large bias between the ATOVS and the AIRS data records is discussed and analysed in more detail. Schneider et al. (2012) compared the TPW of a SSM/I+Medium Resolution Imaging Spectrometer (MERIS) product to TPW products from GUAN, AIRS (V5) and ATOVS (operational CM SAF ATOVS data) for the period 2004–2008. For the comparison to AIRS, the AIRS cloud mask has been applied because TPW from MERIS is only available under clear-sky conditions. They found biases (RMSEs) of 0.5 and -1.1 kg m-2 (2.3 and 3.3 kg m-2) relative to AIRS and ATOVS, respectively. Due to the clear-sky bias (e.g. Sohn and Bennartz, 2008; Mieruch et al., 2010), these values cannot be directly compared to the results of this work. Nevertheless, this study shows that the ATOVS product is more humid than the AIRS product, and the comparison of the ATOVS data to the AIRS data exhibits similar RMSE values to our analysis. Schneider et al. (2012) also observed a greater bias in their comparison to the AIRS data record relative to their comparison to GUAN data record.

Bedka et al. (2010) compared the AIRS V5 data record to ARM observations at Nauru, Barrow and SGP. The average RMSE values are between 2.0 and 3.4 kg m-2 and envelope the average RMSE of 2.4 kg m-2 observed here. The bias between AIRS and ARM data is on average smaller than 0.1 kg m-2. However, their comparison results to all sites exhibit a day–night contrast which is most pronounced for SGP, with day–night differences of 1.6 kg m-2. They conclude that the relative difference at SGP for values larger than 10 kg m-2 is 10 % during boreal summers and decreases during boreal winters. They argue that surface emissivity, land use and cover, and unique boundary layer conditions may contribute to this difference.

Finally, Fig. 9 shows a map of the TPW bias between ATOVS and the AIRS data records. Obviously the bias is dominated by regions of strong precipitation and frequent cloud occurrences such as in the ITCZ and storm track regions. Relatively large values are observed at mountainous areas such as the Alps, but the maximum differences are observed over tropical land surfaces with relative differences of about 15 %. Of course, differences in retrieval setup and associated uncertainties contribute to this bias. Of particular relevance in view of the spatial distribution of the bias are differences in cloud detection, in cloud clearing (not applied for the ATOVS data record) and in the handling of scattering events (screened in the ATOVS data record). In the ATOVS data record, AMSU-B observations are used which also allow a retrieval under cloudy conditions, while in the AIRS data record, cloud clearing needs to be applied to the AIRS data in order to retrieve TPW. In general clear-sky observations exhibit a systematic underestimation of TPW relative to almost all-sky observations (e.g. Sohn and Bennartz, 2008). Thus, the different instrumentation might contribute to the observed differences.

As outlined earlier, the gap-filling process of the Kriging contributes to the observed difference between the ATOVS and the AIRS TPW products with the TPW from ATOVS being larger in precipitating areas than the TPW from AIRS. Schröder et al. (2013) compared the CM SAF SSM/I TPW product to the SSM/I TPW product from the University of Hamburg and from the Max Planck Institute for Meteorology (Andersson et al., 2010). The only difference in the generation of both products is again that the CM SAF product is based on post-processing using Kriging. The spatial distribution of their results is very similar to spatial distribution in Fig. 9. Thus, Kriging is a significant contributor to the observed bias between the ATOVS and the AIRS data records.

As precipitation over tropical land surfaces exhibits a pronounced diurnal cycle with maxima in the late afternoon and evening (e.g. Yang and Slingo, 2001) the differences between TPW from Metop-A (Equator-crossing time of ascending node: ∼ 21:30 local time), NOAA-16 (∼ 19:00 local time) and NOAA-19 (∼ 13:30 local time) have been analysed for the year 2011 (not shown). While the differences between the TPW from NOAA-19 and from Metop-A are relatively small, the differences between the TPW from NOAA-16 and from the other two satellites exhibit pronounced minima over tropical land surfaces. We found minimum differences of approximately -3 kg m-2, or -2 %, with smaller TPW values from NOAA-16 than from NOAA-19 and from Metop. These minima over tropical land surfaces nearly vanish when NOAA-16 and Metop-A differences are computed for the year 2008. Then, the NOAA-16 Equator-crossing time is approximately 16:00 local time. It seems that the diurnal sampling in combination with the diurnal cycle of deep convection over tropical land surfaces has an impact on the ATOVS product. However, channel 4 observations from NOAA-16 have not been used as input to the retrieval. Information from channel 4 is valuable to separate information on near-surface properties from the temperature and water vapour signal of the lower troposphere. When the differences in the TPW products from the CM SAF ATOVS data record and from the AIRS data record are compared for the months of July and August for the period between 2008 and 2011 (not shown), the general coincidence of large biases and precipitating areas is still present. However, land surfaces in the northern extratropics exhibit an increase in bias as well, which might also be associated with convective precipitation. Because convective precipitation has a short-term impact on surface emissivity, differences in handling surface emissivities contribute to the overall difference as well.

Because the TPW bias between the ATOVS and AIRS data records exhibits a breakpoint in early 2009, we had a closer look at the TPW time series from the ATOVS data record and at the TPW bias time series between the ATOVS and the GUAN data records. Between April and May 2009, the ATOVS TPW anomaly time series exhibits a breakpoint with a strength of 0.75 kg m-2 with coverage probability of 98 % and the bias between TPW from ATOVS and from GUAN, a breakpoint of strength 0.37 kg m-2 with coverage probability of 97 %. Obviously the removal of NOAA-15 observations from the retrieval in May 2009 introduces a breakpoint into the ATOVS time series. After the removal of NOAA-15 observations from the retrieval, from May 2009 onwards, we still use two satellites for the ATOVS processing. When looking at Figs. 5 and 7 we do not see further coincidences between an apparent breakpoint and a change from two to three satellites, or vice versa. Thus, we do not expect that this breakpoint can be explained by sampling or Kriging effects. Consequently, we extended the analysis by comparing the AIRS data record to the ATOVS products for each NOAA and Metop satellite separately. During this exercise the Kriging routine is not applied to the ATOVS data. Figure 10 shows the bias between the ATOVS products derived from each NOAA and Metop satellite and the AIRS data record. During overlapping periods, data from Metop-A, NOAA-16 and NOAA-19 exhibit similar biases relative to AIRS, with values around 2 kg m-2. Also noticeable is the increase in bias for NOAA-16 between 2003 and 2009 and the decrease in bias after the maximum in 2009. All NOAA satellites typically have different Equator-crossing times and exhibit a drift in Equator-crossing time (see, for example, John et al., 2012, their Fig. 4). NOAA-16's orbital drift is the strongest and ranges from 14:00 local time in 2003 to 17:30 local time in 2009 and to 19:30 in 2011. The AIRS orbit is stable, with an Equator-crossing time of 13:30. Thus, at the beginning of the bias time series, the difference in temporal sampling is minimal. If the difference in Equator-crossing time were the dominant contributor to the bias, the maximum in bias can be expected at 19:30. It seems again that the diurnal cycle of deep convection in combination with differences in temporal sampling impacts the bias between AIRS and ATOVS. In addition to such a sampling error, the retrieval uncertainty might also be affected by cloud handling, which then results in a diurnal cycle of the retrieval error in the presence of convective clouds. What is most obvious is that the bias for the data derived from NOAA-15 is systematically smaller than the bias for the data derived from any of the other satellites. The average difference between the NOAA-15 and AIRS biases and between those of NOAA-16 and AIRS is almost 1 kg m-2. Consequently, with the removal of NOAA-15 from the retrieval, the bias relative to AIRS and GUAN data records increases. We recomputed the bias between the TPW from ATOVS and AIRS data records with NOAA-15 data being removed from the retrieval in June 2008. The results are shown in Fig. 11. The breakpoint now clearly appears in June 2008 instead of May 2009.