The main finding by Po-Chedley and Fu was that the University of Alabama in Huntsville (UAH) microwave sounding unit (MSU) product has a bias in its NOAA-9 midtropospheric channel (TMT) warm target factor, which leads to a cold bias in the TMT trend. This reply demonstrates that the central arguments by Christy and Spencer to challenge Po-Chedley and Fu do not stand. This reply establishes that 1) Christy and Spencer found a similar, but insignificant, bias in the UAH target factor because their radiosonde data lack adequate sampling and measurement errors were considered twice; 2) the UAH individual satellite TMT difference between NOAA-9 and NOAA-6 reveals a bias of 0.082 ± 0.011 in the UAH NOAA-9 target factor; 3) comparing the periods before and after NOAA-9 is not an adequate method to draw conclusions about NOAA-9 because of the influence of other satellites; 4) using the Christy and Spencer trend sensitivity value, UAH TMT has a cold bias of 0.035 K decade−1 given a target factor bias of 0.082; 5) similar trends from UAH and Remote Sensing Systems (RSS) for the lower tropospheric temperature product (TLT) do not indicate that the UAH TMT and TLT NOAA-9 target factor is unbiased; and 6) the NOAA-9 warm target temperature signal in UAH TMT indicates a problem with the UAH empirical algorithm to derive the target factor.
Po-Chedley and Fu (2012) worked toward reconciling a previously noted, long-standing difference between the University of Alabama in Huntsville (UAH) and the Remote Sensing Systems (RSS) Microwave Sounding Unit (MSU) midtropospheric temperature (TMT) datasets. This difference is also evident between the UAH and the National Oceanic and Atmospheric Administration (NOAA) MSU TMT dataset. The difference is related to the warm target calibration of the MSU measurements (Mears et al. 2003; Karl et al. 2006). UAH, RSS, and NOAA utilize a correction factor, referred to as a warm target factor, that is multiplied by the warm target temperature of the satellite to remove the residual signal of the warm target temperatures in the MSU measurements. For each satellite, the correction has the form , where TMSU and TMSU,C are the measured Earth brightness temperatures before and after correction, respectively; α is the target factor; C is a constant offset; and TWT is the warm target temperature of the satellite (Christy et al. 2000; Mears et al. 2003). Each team solves for this target factor differently, but in order to achieve the same goal: to remove warm target temperature contamination in the final MSU products (Christy et al. 2000; Mears et al. 2003; Zou et al. 2009).
Figure 1 shows the differences between the RSS and UAH TMT and lower tropospheric temperature (TLT) datasets over the NOAA-9 satellite time period versus the warm target temperature on board NOAA-9. We see that the differences in both TMT and TLT are significantly related to the warm target temperature of the NOAA-9 satellite, indicating that there are still large warm target temperature signals in one or both of the UAH and RSS MSU products. The differences between the UAH and the NOAA TMT products are also significantly related to the NOAA-9 warm target temperature (the slope is −0.053 ± 0.010 and the r2 value is 0.81). The difference between the RSS and NOAA TMT products versus the NOAA-9 warm target temperature has a slope of −0.010 ± 0.007 with an r2 value of 0.25, which is also statistically significant. NOAA does not yet produce a TLT dataset.
There are large differences in the magnitudes of the target factors from each group for the NOAA-9 satellite (Po-Chedley and Fu 2012). The UAH NOAA-9 target factor is more than 2 and 3 times larger than the target factors used by RSS and NOAA, respectively. We recognize that target factors may not be exactly the same because of different algorithms used in the satellite merging process, but there should be no detectable warm target signal in the final MSU product. Because the NOAA-9 warm target temperature explains most of the discrepancies between the datasets during the NOAA-9 era (Fig. 1), it is not possible that all target factors are equally valid (Po-Chedley and Fu 2012). Since the warm target temperature for NOAA-9 drifts over time, a target factor that is too large or too small would leave a spurious signal related to the warm target temperature drift in the final MSU TMT and TLT time series. The difference in the magnitude of the target factor has been well documented as an important discrepancy between the UAH and RSS datasets, contributing to about half of the trend difference between the groups (Mears et al. 2003; Karl et al. 2006). Po-Chedley and Fu (2012) sought to reconcile this specific, yet important, difference and not other differences between the three groups.
2. Radiosonde comparisons
In Po-Chedley and Fu (2012), we attempted to detect artificial residuals related to the warm target calibration in MSU TMT products from each group using five global, homogenized, peer-reviewed radiosonde datasets as references. In particular, we examined whether there is a significant relationship between the difference of collocated MSU and radiosonde TMT measurements and the NOAA-9 warm target temperature. Note that this difference is actually the difference between MSU errors and radiosonde errors, and that the latter is unrelated to the satellite warm target temperature (Po-Chedley and Fu 2012). Thus, a significant relationship would indicate a detectable warm target temperature signal in the MSU datasets. We found that we could only detect a significant relationship using the UAH dataset with a mean slope of −0.051 ± 0.031. This value represents the magnitude of the bias in the target factor for UAH TMT (the target factor bias and the slope have the same magnitude but opposite signs) (Po-Chedley and Fu 2012). Christy and Spencer (2013) repeated our calculation using U.S. VIZ radiosondes (see their Fig. 1, top). They obtained a statistically insignificant slope of −0.042 ± 0.047. They further asserted that this calculation assumes no errors for the individual points. By adding an error bar as the “measurement error” in the VIZ radiosondes at each point, they produced a relationship of −0.042 ± 0.073. Below we demonstrate that their claim cannot be substantiated because they used a limited number of radiosonde stations and their consideration of measurement errors is incorrect.
Figure 2 shows the effect of the number of radiosonde stations on the bias detection threshold by randomly subsampling the Radiosonde Innovation Composite Homogenization (RICH) global radiosonde dataset (Haimberger et al. 2008). We obtain similar results from other radiosonde datasets utilized in Po-Chedley and Fu (2012). It shows that the uncertainty declines dramatically as we incorporate more radiosonde locations. There is a large uncertainty when few radiosonde stations are used, owing to the large error noise from individual radiosonde stations. For this reason, it is important to use a large number of radiosonde sites. The Christy and Spencer (2013) value is insignificant, owing to poor radiosonde sampling; Christy and Spencer (2013) utilized 31 radiosonde stations, while Po-Chedley and Fu (2012) used between 45 and 451 stations. While 31 stations are not enough to significantly quantify this bias, it is encouraging that Christy and Spencer (2013) find a similar target factor bias of 0.042 using a different subset of radiosondes [Po-Chedley and Fu (2012) measured a bias of 0.051 ± 0.031].
As discussed in Po-Chedley and Fu (2012), the difference of collocated MSU and radiosonde TMT measurements is actually the difference of measurement errors from the satellite MSU and radiosondes. By regressing this difference against the satellite warm target temperature, the effect of the target factor bias is separated from other errors. Its statistical uncertainty is thus determined by other unresolved errors in satellite measurements and errors in radiosonde measurements, and the latter dominates. Note that it is the radiosonde uncertainty that prevents us from detecting any significant target factor bias in the RSS and NOAA datasets using the radiosonde data, although the TMT difference between the RSS and NOAA datasets is still significantly correlated with the NOAA-9 warm target temperature (Po-Chedley and Fu 2012). The UAH target factor bias is large enough that we can detect a calibration bias. By adding an error bar as the measurement error at each individual point, Christy and Spencer (2013) considered the measurement errors twice, which is incorrect. Note that the radiosonde measurement error that Christy and Spencer (2013) reference is simply derived from the differences between collocated UAH MSU and radiosonde TMT measurements (Christy et al. 2011), which is actually the measurement error difference between satellite MSU and radiosonde measurements. Regardless, this uncertainty is already included in our estimation.
Furthermore, when we ignore the data for February and March of 1985 (influenced by NOAA-7) as done in Christy and Spencer (2013), we still measure a significant bias in the UAH NOAA-9 target factor. Christy and Spencer (2013) also showed that if they only used the TMT measurements during overlaps of NOAA-6 with NOAA-9 and NOAA-10 with NOAA-9 (14 months), then they found no relationship between the UAH error and the warm target temperature (their Fig. 1, bottom). This is an expected outcome of their calibration procedure. Our further comment on this finding is that regardless of the details of the methods used to derive the warm target factor, the final MSU product should not have a residual signal related to the warm target temperatures. Again, from Fig. 1, there is a residual related to the warm target temperatures and it is the UAH dataset that has a detectable bias (Po-Chedley and Fu 2012).
Christy and Spencer (2013) state that Po-Chedley and Fu (2012) calculated a target factor bias for RSS of 0.023. Po-Chedley and Fu (2012) did not calculate a target factor bias for RSS (there is no significant relationship between RSS TMT and the warm target temperature using radiosondes as a reference; the r2 value was 0.03). Christy and Spencer (2013) also state that Po-Chedley and Fu (2012) utilized RSS as a reference and then claim that the target factor bias for RSS (which we did not calculate) represents a drift in the radiosondes. Po-Chedley and Fu (2012) did not use RSS as a reference. Recall that there is a residual bias in RSS or NOAA, but any bias is too small to detect via radiosondes (Po-Chedley and Fu 2012). Furthermore, we do not expect that the target factors should be exactly the same because RSS, UAH, and NOAA perform other corrections that may alter the magnitude of the target factor (as noted by Christy and Spencer 2013). The derivation of the target factor of 0.071 in Christy and Spencer (2013) is a result of their unrealistic assumptions.
Christy and Spencer (2013) also compare the differences in the TMT temperature before and after the NOAA-9 time period for a number of datasets, claiming that radiosondes have a spurious shift relative to RSS and UAH and that NOAA-9 differences between MSU groups have little impact on the long-term time series. It is important to note that Christy and Spencer (2013) did not collocate radiosonde and MSU measurements in this comparison. Spatial sampling alone can account for very large differences between datasets (Mears et al. 2011), so the comparisons between radiosonde and MSU datasets in Christy and Spencer (2013) are actually comparisons of very different metrics. Christy and Spencer (2013) suggest that the bias we detect is caused by radiosonde warming relative to RSS and UAH. This is unlikely because Po-Chedley and Fu (2012) found a bias in the UAH target factor relative to five global, homogenized radiosonde datasets but did not find a bias in the RSS or NOAA target factor. If this were simply an issue of radiosonde quality, then we would expect to measure a bias in each of the three MSU datasets. It is also important to note that the slope of the UAH minus RSS (NOAA) TMT versus the NOAA-9 warm target temperature, which is independent of radiosondes, is −0.043 (−0.053). This is consistent with a UAH target factor bias of 0.051 as derived from radiosondes.
Christy and Spencer (2013) argue that there are no significant differences between RSS and UAH before and after NOAA-9. Unfortunately, the Christy and Spencer (2013) comparison is confounded by the influence of satellite problems in other time periods. For example, UAH TMT spuriously cools relative to RSS and NOAA between 1983 and 1985. It is unclear which dataset has the artificial warming or cooling, but it leads to spurious agreement between UAH and RSS (NOAA) during the time intervals utilized in Christy and Spencer (2013). In general, we object to the Christy and Spencer (2013) method of comparing the periods before and after NOAA-9 because it introduces the influence of other satellites, which makes it difficult or impossible to draw conclusions about NOAA-9. This is also a less direct method than that of Po-Chedley and Fu (2012), who use the NOAA-9 time period to make conclusions about NOAA-9. It is clear that the NOAA-9 target factor is responsible for long-term trend differences, based not only on previous studies (e.g., Mears et al. 2003; Karl et al. 2006) and Po-Chedley and Fu (2012), but also on Christy and Spencer (2013), who find that their TMT trend increases when the target factor value is reduced. The fact remains that there must be a bias in one or more MSU TMT products (Fig. 1) and when we attempt to detect the bias, UAH is an outlier relative to two complementary MSU datasets, five radiosonde datasets and, as we will demonstrate, its own co-orbiting satellite measurements.
One important point made in Christy and Spencer (2013) is that the bias measured in Po-Chedley and Fu (2012) may be influenced by other co-orbiting satellites during the NOAA-9 time period. Here, we check the UAH NOAA-9 TMT time series for a residual bias using the UAH warm-target-corrected global mean TMT values for the NOAA-6 and NOAA-9 satellites. The warm target factor of NOAA-6 is near zero for all three MSU teams and thus NOAA-6 has little warm target temperature contamination. We find that the warm target temperature of NOAA-9 significantly explains the differences between the UAH warm-target-corrected NOAA-9 and NOAA-6 TMT measurements. In Fig. 3a, we show UAH NOAA-9 minus UAH NOAA-6 TMT versus the warm target temperature for NOAA-9. We find that the UAH NOAA-9 target factor has a bias of 0.082 ± 0.011, which is larger than the target factor bias measured in Po-Chedley and Fu (2012). One reason for this difference, as Christy and Spencer (2013) point out, is that the influence of satellites other than NOAA-9 may diminish the signal measured in Po-Chedley and Fu (2012). In Fig. 3b, we show the NOAA-9 minus NOAA-6 differences over time and the fit to these data using the NOAA-9 warm target temperature as a predictor. Figure 3 clearly illustrates that the warm target bias has not been removed from the UAH NOAA-9 TMT time series. Furthermore, since the NOAA-9 constant offset is calculated relative to NOAA-6 in late 1985 and 1986, NOAA-9 will start off too warm and spuriously cool because of this bias. NOAA-9 is incorporated into the TMT time series in late 1984 and overlaps with NOAA-7 during this time (when NOAA-9 was artificially warm), so we would expect large positive residuals for NOAA-9 minus NOAA-7, since both had constant offsets calculated relative to NOAA-6. We indeed find that there are large residuals between NOAA-9 and NOAA-7 (the average is 0.205). We show the residuals for NOAA-9 minus NOAA-7 in Fig. 3b, even though NOAA-7 was not used in determining the bias in the UAH NOAA-9 target factor. The large residual between NOAA-9 and NOAA-7 is further evidence that NOAA-9 has a calibration problem related to the warm target factor.
Christy and Spencer (2013) point out that a major difference between RSS and UAH is the degree of temporal smoothing used to determine satellite target factors. UAH smoothes MSU daily data over relatively long periods (61–121 days), while RSS uses 5-day averages (pentads) when computing target factors (Christy et al. 2000; Mears et al. 2003). Note that our result using UAH NOAA-9 and UAH NOAA-6 daily data (Fig. 3) is very similar to that derived using 5-day average data. Christy and Spencer (2013) explain that without sufficient temporal smoothing, intersatellite difference trends remain. But it is not necessarily true that these trends are caused by the warm target calibration for NOAA-9. Mears et al. (2003) utilized the UAH NOAA-9 target factor and found that the UAH coefficient reduced the intersatellite trend between NOAA-6 and NOAA-9, but that it significantly increased intersatellite trends for other satellites, especially satellite overlaps ignored by UAH. Mears et al. (2003) also found that processing decisions other than temporal smoothing are important in determining the NOAA-9 target factor. When RSS utilized the UAH diurnal drift correction and merging procedure, but used pentad data, they found a target factor of 0.075—nearly as large as the actual UAH value (Mears et al. 2003). In general, smoothing, diurnal drift corrections, and the selection of satellite overlaps all seem to influence the NOAA-9 target factor. Most importantly, UAH is the only MSU dataset containing detectable residual warm target signals compared to independent radiosonde datasets.
3. Impact of target factor bias on TMT trend
We made a simple estimate of the impact this bias would have on the trend, warning that “while our estimate of the UAH TMT trend sensitivity to the warm target factor is similar to estimates by the UAH and RSS teams, our study indicates that the UAH team will need to incorporate an optimal NOAA-9 target factor into their merging procedure for an accurate trend estimate” (Po-Chedley and Fu 2012, 650–651). When the UAH team does incorporate the target factor that we suggest into their algorithm, they find that their TMT trend increases by 0.022 K decade−1 (Christy and Spencer 2013). Christy and Spencer (2013) report a lower trend sensitivity value, because the drift of NOAA-9 due to the target factor bias only introduces a cold bias between the NOAA-6/NOAA-9 overlap and the NOAA-9/NOAA-10 overlap (NOAA-9 connects NOAA-6 to NOAA-10). In Po-Chedley and Fu (2012), we had assumed that the drift caused a cold bias throughout the entire NOAA-9 time series.
The Christy and Spencer (2013) estimate of the trend sensitivity is still significant (it would increase the UAH TMT 1979–2009 trend by about 50%) but, importantly, it may not be appropriate to simply fix the NOAA-9 target factor to estimate the impact of this bias on the trend. The NOAA-9 target factor influences other satellite target factors and the constant offsets between satellites. It is important to optimize all of the target factors and constant offsets. The UAH NOAA-9 target factor may retain a residual bias because of differences in the treatment of the diurnal cycle drift bias, smoothing, or because UAH uses fewer satellites to constrain the NOAA-9 target factor compared to RSS and NOAA. It is important to isolate the cause of this bias.
Finally, given the fact that the final MSU TMT product during the NOAA-9 period is a combination of measurements from NOAA-9 with other co-orbiting satellites, the central estimate of the warm target factor bias for UAH TMT as estimated from Po-Chedley and Fu (2012) may be too low. Since the UAH TMT trend varies linearly with the NOAA-9 target factor (J. Christy 2012, personal communication), we can estimate the impact of the NOAA-9 target factor bias on the UAH TMT trend. Using a warm target factor bias of 0.082 (from the NOAA-9 and NOAA-6 comparisons in this note) and the Christy and Spencer (2013) value for the trend sensitivity, the UAH trend would increase by 0.035 K decade−1.
4. The lower tropospheric temperature time series
Christy and Spencer (2013) present the agreement between UAH and RSS for global TLT trends as evidence that there is no UAH target factor bias. This is not true; it is possible that the UAH trend could exceed that of RSS when this bias is accounted for with trend differences resulting for different reasons. In other words, the current trend agreement may be due to compensating biases in one or both datasets. Po-Chedley and Fu (2012) never claimed that RSS should be used as a reference. The NOAA-9 target factor bias in UAH TMT would affect the UAH TLT product because UAH uses the TMT target factor for TLT. Figure 1b clearly shows that the differences in the TLT measurements between RSS and UAH over the NOAA-9 time period are explained by the warm target temperature of NOAA-9. We could not detect a bias in the warm target factor for TLT using radiosondes because the noise in the TLT product is amplified by a factor of 2–3 compared to TMT (Hurrell and Trenberth 1998; Christy et al. 2000; Mears and Wentz 2009).
5. Summary and discussion
The main finding in Po-Chedley and Fu (2012) was that UAH has a bias in its NOAA-9 warm target factor. To reiterate some of the key points that indicate that UAH has a bias in the NOAA-9 warm target factor: 1) We show that the NOAA-9 warm target temperatures and thus target factor differences explain the differences between the MSU teams during the NOAA-9 time period (Fig. 1), which has important implications for the TMT trend (Mears et al. 2003; Karl et al. 2006); 2) this indicates that one or more teams has a bias in its warm target factor; 3) Po-Chedley and Fu (2012) found that UAH has a bias in its target factor using five peer-reviewed, homogenized radiosonde datasets; 4) Christy and Spencer (2013) found a similar, though insignificant, bias in the UAH target factor, but we show in this work that their radiosonde data lack adequate sampling to detect this bias and that their uncertainty is overestimated; 5) we show from individual UAH satellite records that a large, significant bias exists in its NOAA-9 target factor; and 6) the large residuals between UAH NOAA-9 and UAH NOAA-7 are further evidence of a UAH NOAA-9 target factor bias. The results 1, 2, 5, and 6 are independent of the radiosonde data.
As suggested in Po-Chedley and Fu (2012), UAH will need to determine the source of this bias and correct it for an accurate trend estimate in TMT. Since their TLT product utilizes the TMT target factor, a bias should also exist in their global TLT trend. UAH will need to remove the target factor bias for more reliable UAH TMT and TLT trend estimates. Using the UAH trend sensitivity (Christy and Spencer 2013) with a bias of 0.051 and 0.082 in the UAH NOAA-9 warm target factor, the trend in the UAH TMT product increases by 0.022 and 0.035 K decade−1, respectively, accounting for a majority of the global TMT trend differences between UAH and RSS for 1979–2009.
Christy and Spencer (2013) describe other documented reasons for differences between the RSS and UAH datasets. This information pertains mainly to differences in the 1990s and the diurnal cycle drift bias corrections. This information is not relevant to Po-Chedley and Fu (2012), though we agree that other differences between MSU datasets, particularly in the tropics, need to be resolved (e.g., Fu and Johanson 2005; Karl et al. 2006; Santer et al. 2008; Christy et al. 2010). The scope of “A bias in the midtropospheric channel warm target factor on the NOAA-9 Microwave Sounding Unit” (Po-Chedley and Fu 2012) pertained to the NOAA-9 warm target bias, though different biases that apply to other satellites during different time periods are still important.
The UAH group has been a pioneering force in temperature measurements from the MSU. In fact, it was the UAH team that developed the warm target calibration (Christy et al. 1998; Christy et al. 2000) that is employed by RSS and NOAA. Po-Chedley and Fu (2012) demonstrate that the UAH warm target correction for the NOAA-9 satellite is too large, which artificially reduces the UAH TMT trend. The cause of this bias should be explored further to understand, in more detail, the effect of the diurnal cycle drift correction, the effect of smoothing, and the selection of satellites used in determining the NOAA-9 target factor. Regardless of the cause of the bias, it leaves NOAA-9 warm target temperature signals in the UAH MSU products and leads to a long-term cooling bias in the UAH dataset.
We thank Dr. John Christy (UAH), Dr. Carl Mears (RSS), and Dr. Cheng-Zhi Zou (NOAA) for their useful comments on our work, for providing us additional data related to the MSU, and their continued willingness to make clarifications about MSU data. We also thank Pu Lin for her helpful suggestions on this manuscript. This work was supported by the National Science Foundation Graduate Research Fellowship (Grant DGE-0718124) and Grant NESDIS-NESDISPO-2009-2001589 (SDS-09-15).
The original article that was the subject of this comment/reply can be found at http://journals.ametsoc.org/doi/abs/10.1175/JTECH-D-11-00147.1.