a. The nature of HIRS-12

The HIRS instrument is part of the TIROS-N Vertical Sounder (TOVS) flown on NOAA operational polar-orbiting satellites. Channel 12 is centered at 6.7 μm in the 6.3-μm vibration–rotation band of water vapor. HIRS-12 is sensitive to humidity and temperature over a broad layer in the upper troposphere. The exact height of this layer is determined by the humidity and temperature profile. Lower relative humidities will result in the emission coming from lower in the troposphere where temperatures are higher. Thus a lower relative humidity in the upper troposphere will result in higher brightness temperatures, and vice versa. Salathe and Smith (1996) have shown that the effects of absorbers other than H₂O are small in the 6.3-μm region. Gases O₂ and CH₄ combined are responsible for a uniform decrease in HIRS-12 brightness temperatures of less than 0.5 K. The effects of other absorbers are negligible. There are still some uncertainties in simulating HIRS-12 brightness temperatures due to the possible influences of cirrus clouds and uncertainties in the water vapor continuum (Salathe and Smith 1996).

Soden and Bretherton (1993, 1996) have shown that the relative humidity in the upper troposphere, when weighted by the sensitivity of HIRS-12 T_B to relative humidity perturbations, is strongly correlated to HIRS-12 T_B. This weighted upper-tropospheric humidity is referred to here as UTH. The sensitivity is defined as the response of HIRS-12 T_B to small perturbations in relative humidity at levels equally spaced in log pressure coordinates. It gives a measure qualitatively similar to the weighting function. Using zonally meaned 1865 and 2045 atmospheres from the model, the sensitivity profiles of HIRS-12 were derived as a function of latitude and log pressure. The sensitivity functions for 1865 are shown in Fig. 1, normalized in log pressure coordinates. It is clear that the HIRS-12 instrument is sensitive to a broad layer in the atmosphere, located between 200 and 550 mb at the equator and lower in the atmosphere at higher latitudes. The peak sensitivity is at about 325 mb at the equator and drops to about 450 mb at high latitudes. The true sensitivity function for a particular location will vary vertically in the atmosphere dependent upon the humidity profile so there is some uncertainty involved using a zonal mean sensitivity function.

One feature of the HIRS-12 channel is that the effects of temperature variations on T_B are very low. Typical variations in atmospheric relative humidity have 5–8 times the influence of typical variations in temperature at fixed relative humidity (Soden and Bretherton 1993), although the difference is lower at high latitudes.

b. Observational data

The processed and intercalibrated data used in this experiment (Bates et al. 1996) were obtained from Darren Jackson (University of Colorado) and John Bates (NOAA/Environmental Research Laboratory Climate Diagnostics Center). Monthly means of HIRS-12 brightness temperature (T_B) on a 2.5° × 2.5° grid have been produced for the years 1979–95, excluding January 1979. The data has been cloud cleared and “limb-corrected” to produce the equivalent nadir brightness temperatures. The data from different satellites has been intercalibrated as described below.

Since the data comes from instruments on successive NOAA satellites that are not totally identical, the weighting functions change slightly from satellite to satellite. A second problem with the HIRS-12 data is that of biases arising from the sampling times of the sun-synchronous orbits, which capture different aspects of the diurnal cycle. The orbits also tend to drift over the lifetime of a satellite and this might cause spurious trends as a result of the interaction with diurnal variability. Problems also occur where only one satellite is available, an instrument malfunctions, or the calibration drifts over time. These differences are crucial to the accuracy of a long-term climate dataset, so the data must in some way be intercalibrated.

Bates et al. (1996) describe the procedure used to intercalibrate the HIRS-12 data. Because the response of a radiometer channel is in general highly nonlinear, it cannot be assumed that there is a fixed bias between measurements from different satellites. Instead, they matched statistical distributions of measurements during periods of satellite overlap using “empirical distribution functions” (Weinreb et al. 1989). With this technique the response of one satellite may be normalized to that of another. They chose NOAA-7 as the baseline satellite and the brightness temperatures from other satellites were adjusted to match the response of this instrument. Brightness temperatures from different daily sampling times were adjusted to that of the evening orbit. Unfortunately, there was no overlap between NOAA-8 and NOAA-9 so Bates et al. (1996) used an ad hoc linear adjustment for brightness temperatures from NOAA-9 and later satellites.

It must be noted that the adjustments did not, however, eliminate the possible effect of drifts in sensor calibration or orbital sampling time over the lifetime of a single instrument. Second, the intercalibrated dataset is self-consistent but does not have an absolute calibration. Radiosonde measurements of upper-tropospheric moisture are poor (Elliot and Gaffen 1991), so they cannot be used to validate the absolute calibration of the instrument. Since the pattern similarity statistics compare relative changes rather than absolute magnitudes, this should not be a problem.

Only data between 70°N and 70°S were used, since HIRS-12 data from polar regions is very sparse due to persistent low clouds, snow and ice cover, and strong inversions that prevent the derivation of cloud-cleared radiances (Wu et al. 1993). Wu et al. (1993) estimated the error on a single HIRS-12 measurement due to instrument noise, angular correction, and errors in the cloud clearing algorithm to be 2.48 K. The standard error s on a mean of HIRS-12 data is given by s = 2.48/(N)^1/2, where N is the number of observations used in the mean. The number of measurements used in the 5° × 5° HIRS-12 yearly means is on average 2044, which results in an estimate of the standard error of 0.05 K. This estimate does not include the systematic error resulting from undersampling in cloudy areas, which results in a slight positive bias in the data estimated to be of the order of 0.7 K (Soden and Lanzante 1996). If changes in cloud cover are seen over time, an effect on satellite sampling could have some influence on the HIRS-12 trends.

A small amount of missing data needed to be estimated by bilinear interpolation before the data could be used to produce yearly mean brightness temperatures on a 5° × 5° grid for comparison to the HADCM2 data. Data for January 1979 were not available and May 1986 had 72% missing data. Apart from these periods, there were 1% missing data points per month. A monthly mean climatology was created for each grid point using the good data over the 17-yr period. For May 1986 equal weight was given to the climatology and the results of a bilinear interpolation from the data available. January 1979 was filled with the climatology values, since any inaccuracies arising should be outweighed by the benefits of an extra year of data. It must, however, be remembered that the yearly means in 1986 and 1979 are biased slightly toward the 17-yr mean. Yearly means of the anomalies, ΔD(x, t), were created from the monthly mean HIRS-12 data and a selection is plotted in Fig. 3. The spatial correlation statistics essentially evaluate the similarity of the pattern of these fields to the predicted model changes ΔM(x) (Fig. 7a).

The interannual variability of the HIRS-12 brightness temperatures is exemplified by the variations in Fig. 3. The anomalies are strongly affected by the ENSO events in 1983, 1987, and 1992–93. Bates et al. (1996) show that an ENSO event results in lower brightness temperatures (upper-tropospheric moistening) over the central and eastern Pacific, with a corresponding increase (upper-tropospheric drying) in the brightness temperatures over Indonesia, the northern half of South America, and in the subtropics in the central Pacific. The anomalies in 1983 show this pattern more clearly than the other ENSO years, but the pattern of equatorial moistening and subtropical drying in the central Pacific is common to all. The year 1989 shows some evidence of anomalies in the Pacific of opposite sign to those seen in ENSO years. There is also much interannual variability that cannot be linked to ENSO; for example, the contrast between 1979 and 1986 is quite marked.

Figure 4 shows the temporal standard deviation of the yearly mean HIRS-12 brightness temperatures. A complex structure is apparent, with generally higher standard deviations in the Tropics and subtropics than the midlatitudes. The Tropics and subtropics show standard deviations mostly in the range 0.4–0.8 K but in much of the Pacific they are higher than this, peaking at 1.55 K at 20°N in the eastern Pacific. ENSO is largely responsible for the high variability in the Pacific.

A least squares fit was used to obtain the 17-yr linear trend in the HIRS-12 anomalies at each grid point. The statistical significances of these gridpoint trends were calculated, and a description of the method is given in the appendix. Figure 5 shows the trends with areas of 90% statistical significance shaded. The figure shows some tropical moistening (T_B decrease) and subtropical drying (T_B increase), but these trends are mostly below the significance level so it would be unwise to draw any conclusions. Some drying trends at 30°S are, however, significant. In the Tropics and subtropics trends will in general be hard to identify because of the high interannual variability, especially over the Pacific.

At 60°S there is a significant general decrease in brightness temperature of approximately 0.025 K yr⁻¹. Given the high latitude, this could be a result of either upper-tropospheric cooling or moistening. There are some areas at 60°N that also show a statistically significant decrease in T_B, but at this northern latitude most areas show no real trend. A moistening trend off the equatorial east coast of Africa is significant but a similar-sized trend in the east Pacific is not significant, which is again most likely due to the high interannual variability in the Pacific region.

Peak trends equate to about 1 K brightness temperature over 17 yr. At fixed temperature, this equates to a drop in the relative humidity by a factor of roughly 0.12; that is, at 50% relative humidity, the change will be approximately −6%. These trends are relatively small compared to the high interannual variability in some areas. What Fig. 5 does indicate, however, is that the relative humidity of the upper troposphere does change significantly over this timescale at a regional level. In contrast, the area weighted 70°N–70°S mean change of T_B over the 17 yr of HIRS-12 data is only 0.02 K. The small magnitude indicates the difficulties inherent in searching for climate change using only changes in the spatial mean quantities.

a. Climate model and simulation of T_B

The data used for the brightness temperature simulations are taken from a transient climate change integration of the Hadley Centre Climate Model, referred to as HADCM2, which included the effects of increasing greenhouse gases and tropospheric aerosols (Mitchell et al. 1995). The model is a coupled ocean–atmosphere general circulation model using a model grid of 2.5° lat × 3.75° long with 19 atmospheric levels and 20 levels in the ocean (Johns et al. 1997). Comparisons with historical surface temperatures in Mitchell et al. (1995) and atmospheric temperatures (1961–95) from radiosonde in Tett et al. (1996) show reasonable agreement with HADCM2 simulations, giving confidence in the model. However, the HADCM2 integration used here does not include the effects of ozone changes. The incorporation of representative ozone scenarios improved agreement with observed atmospheric temperature changes in Tett et al. (1996). The main improvements were seen in the stratosphere and upper troposphere, where increased cooling was seen. The model also uses the “equivalent CO₂” approximation where enhanced CO₂ concentrations are used to represent the effects of the greenhouse gases methane, nitrous oxide, and the halocarbons. This causes an unrealistically large decrease in stratospheric temperatures, but it has only a minimal effect on the troposphere via downward flux from the stratosphere. The data used here are decadal means centered around 1865 and 2045 taken from the transient run. They were linearly interpolated onto the 5° × 5° grid used in this experiment. When 1865 and 2045 are used to describe the model data, we are referring to the decadal means.

The outgoing nadir radiances were simulated from the decadal-mean model profiles at each grid point using the MODTRAN radiation code (Wang et al. 1996). MODTRAN was used at 1 cm⁻¹ resolution between 1380 and 1565 cm⁻¹ with the effects of H₂O, O₂, and CH₄ included. The satellite-observed radiance was calculated by integrating the product of the simulated radiance data and the normalised NOAA-7 HIRS-12 spectral response. Bates et al. (1996) have removed the effect of variations in the instrument’s spectral response on the observed data by standardizing to that of the instrument on board NOAA-7, so this was used in all calculations. The satellite-observed radiance was then converted to brightness temperature using the Planck function.

One problem arises from the nonlinear dependence of HIRS-12 brightness temperatures on the upper-tropospheric humidity (Fig. 2). At low UTH there is a much stronger dependency of T_B on UTH than at higher UTH, so the brightness temperature simulated from a mean atmospheric profile will not be the same as the mean brightness temperature simulated from daily profiles. We examine the effect that using decadal-mean profiles will have on the predicted changes in T_B.

Consider a series of atmospheric profiles, P_i=1,N, with a mean profile, P. The brightness temperature simulated from this mean profile will differ from the correctly simulated mean brightness temperature by a certain error, E:

View Expanded

In general for HIRS-12, T_B(P) will be lower than the correct mean T_B, since the drier (higher T_B) profiles will have relatively more influence on the mean when simulated individually. The greater the variability of UTH in the series of profiles, and the lower the mean UTH, the larger this difference will be. We examine only the change of brightness temperature in this study, so it is the difference in the error, E, between the 1865 and 2045 mean atmospheres that will affect the model-simulated change.

To evaluate the difference between a change in brightness temperature simulated from mean atmospheres (ΔT_B) and that simulated correctly from the individual profiles (ΔT^*_B), we use a sample of points from the European Centre for Medium-Range Weather Forecasts (ECMWF) analyses for 1997. Fifteen points were chosen, equally spaced between 70°N and 70°S, along a line of longitude at 35°E, giving a sample of the Tropics, subtropics, and midlatitudes. This line intersects the Egyptian dry region (Fig. 8b), which shows very high brightness temperatures (i.e., very low relative humidities) during the summer and where a large meaning error E might be expected. Atmospheric profiles at 0000 and 1200 UTC were extracted for the whole of 1997. To justify the use of only a year of data, rather than a decade of data, we observe that the intra-annual variability of UTH is much larger than the interannual variability. Since it is the variability in UTH that causes the meaning error, the greatest contribution to this will come from the intra-annual variability. In terms of HIRS-12 brightness temperature, the intra-annual standard deviation in the ECMWF samples ranged from 3.5 to 6 K. In contrast, the interannual standard deviation of HIRS-12 T_B peaks at 1.55 K but is generally much lower (Fig. 4).

Brightness temperatures were simulated over 1997 for each point and a mean was calculated (i.e., T_B). For each point the yearly mean atmospheric profiles were derived and from this another brightness temperature was calculated [i.e., T_B(P)]. Although the water vapor variable used here is always relative humidity, the means in the ECMWF profiles were calculated using specific humidities, since the HADCM2 mean atmospheres were also created this way. We found that meaning in relative humidity leads to similar conclusions to those below. The relative humidities in the troposphere (below 200 mb) in the ECMWF profiles were then changed by values representative of the model-predicted changes (Fig. 7b) and brightness temperatures were again simulated. Using these humidity changes, we derived a set of true mean changes in brightness temperature (ΔT^*_B) to compare to those simulated using yearly mean profiles (ΔT_B).

A plot of ΔT_B against ΔT^*_B is shown in Fig. 6, and a different symbol is used for each set of 15 points corresponding to a particular magnitude of relative humidity change. A linear least squares fit was performed between the two sets of data. Figure 6 reveals that the main effect of using yearly mean profiles is that the magnitude of ΔT_B compared to ΔT^*_B is typically reduced by a factor of 0.59, that is, the overall pattern of changes simulated from yearly mean atmospheres will be reduced in amplitude compared to the true pattern of changes. The pattern itself will be relatively unchanged;there is a good correlation of 0.98 between ΔT^*_B and ΔT_B, but the linear fit reveals a slight bias of −0.04 K. The largest deviations from the straight line (i.e., from the pattern) occur for points corresponding to subtropical dry regions when large relative humidity changes are made, as might be expected.

The main effect on the patterns of brightness temperature change appears, therefore, to be a reduction in the overall pattern amplitude. This is not very important given the nature of the pattern correlation statistics and the experimental setup. If future studies compare the evolution of absolute model and data values, rather than looking for patterns of change, more accurate mean radiances will be needed. One way to make this easier would be to generate spectrally resolved top-of-the-atmosphere radiances as part of the climate model.

b. Relative humidity, temperature, and T_B changes in the model

This section analyzes the humidity and temperature changes predicted by the model and their effect on the predicted change in brightness temperatures.

The model-predicted change of HIRS-12 T_B between 1865 and 2045, ΔM(x) is shown in Fig. 7a, and the T_B simulated for 1865 is shown in Fig. 8a. Figure 7a shows a complex pattern of changes, with some latitudinal structure. In the higher latitudes, the change is small and typically less than +0.5 K. Large areas of the subtropics and Tropics show increases in brightness temperatures of up to 2.5 K, which would be expected to come from decreasing upper-tropospheric humidities. Decreasing brightness temperatures are seen in the subtropics mainly over the South Atlantic and in a band over the South Pacific. The decrease in the South Atlantic is quite strong, peaking at approximately −3 K over a large region. These brightness temperature changes will be related to relative humidity and temperature changes in this subsection. Because of the logarithmic dependence of brightness temperature on UTH, one cannot associate a particular magnitude change in T_B with a change in UTH unless the background level of UTH is specified.

The area-weighted mean change in T_B over the region studied (70°N–70°S) is 0.35 K, which is small compared to the peak changes seen in Fig. 7a. It should not be taken to indicate a slight overall drying of the upper troposphere, as the combined effects of temperature and relative humidity must be evaluated.

In order to determine the effect of changes in relative humidity and temperature between 1865 and 2045 on the pattern of HIRS-12 T_B, quantities must be extracted in the atmospheric layer that influences HIRS-12. Using the relevant zonal sensitivity functions (e.g., Fig. 1) layer weighted humidities (UTH) and temperatures (UTT) can be extracted from the 1865 and 2045 HADCM2 atmospheres. For clarity we have chosen to look at a fixed layer in the atmosphere using the 1865 sensitivity function to extract UTT and UTH for both 1865 and 2045. Figures 7b and 8b show respectively the difference UTH₂₀₄₅ − UTH₁₈₆₅ and the 1865 UTH distribution. The visible differences between UTH plots for 1865 and 2045 are small, so a plot for 2045 is not shown. Figures 7c and 8c show similar plots for UTT. We could have chosen instead to look at changes in UTH and UTT by following the contributing layer as it moves in the atmosphere rather than looking at changes in a fixed layer. Changes in the humidity profiles and increases in tropospheric temperature result in a small upward shift in the zonal mean sensitivity functions between 1865 and 2045. The choice of either of these methods has little effect on the pattern of change of UTH but the pattern of change of UTT is markedly affected. Figure 7c shows large temperature increases in the fixed layer of the atmosphere, as expected. However, if we extract UTT₂₀₄₅ using the 2045 sensitivity functions, there is a much-reduced change in UTT, since the sensitivity functions move up in the atmosphere with the increasing temperatures. Essentially, the effect of UTT on HIRS-12 brightness temperatures comes partly from its effect on the sensitivity functions. This points to a more subtle effect of UTT on the HIRS-12 brightness temperatures than that of UTH and means that the clearest way of looking at the changes in temperature is using a fixed layer in the atmosphere.

Figure 8b shows a distribution of UTH with strong zonal differences and regional features in the Tropics and subtropics. Subtropical dry regions can be seen at around 20°N and 20°S, and the equatorial regions of higher UTH are associated with the intertropical convergence zone (ITCZ). A feature of high UTH extending across the South Pacific is thought to be associated with the convection of the South Pacific convergence zone (SPCZ). The UTH increases to 60%–70% in the region of the midlatitude depressions.

In contrast, the distribution of UTT (Fig. 8c) shows a mainly zonal distribution with the greatest departures from the zonal pattern over a region at about 160°E. The changes in UTT (Fig. 7c) also show a mainly zonal pattern. The greatest tropospheric temperature increases in the model are found in the tropical upper troposphere, in exactly the layer sensed by HIRS-12. The midlatitudes show smaller UTT increases of 1 to 2.5 K with less warming in the Northern Hemisphere.

The changes in UTH (Fig. 7b) are the result of both changes in the patterns of the mean circulation and changes in the humidity of the regions associated with this. The dry region of the South Atlantic moistens by about 5% in relative humidity, causing a strong decrease in T_B (Fig. 7a). In comparison, increases in UTH of similar magnitude in the SPCZ region have a smaller influence on the brightness temperatures, again due to the lower sensitivity of HIRS-12 T_B to changes at high UTH. The Pacific shows a drying of the ITCZ but a slight moistening and shift to the northeast of the SPCZ. There is a drying of the subtropical dry regions of the Indian Ocean but some of the strongest drying is seen where the dry region over Australia enlarges to the northeast following the move in the SPCZ. A strong moistening is seen in the northeast Pacific. The origin of these changes is unfortunately obscured by the decadal mean nature of the data, and analysis using the output of the model on seasonal scales might in future give clearer and more understandable trends. The area weighted 70°N–70°S mean UTH change is only −0.15% in relative humidity, which shows that on a nearly global scale the model predicts relatively constant UTH. This should be contrasted with the strong changes on regional scales.

It is clear from Fig. 7 that the HIRS-12 T_B changes between 1865 and 2045 in the Tropics and subtropics are dominated by the changes in UTH, despite the UTT increase of typically 4 K. Increases in UTH result in a decrease in the brightness temperature, and vice versa. However, at latitudes higher than 40°N and 40°S, despite small predicted UTH increases, some areas show an increase in brightness temperature. Increases in UTT would be expected to increase the T_B and clearly this must be influencing the changes at high latitudes.

To assess the influence of the change in humidity (ΔUTH) on the change in T_B (ΔM), scatterplots of these two quantities were produced. Two regions were considered: between 30°N and 30°S and 30°–70°N and 30°–70°S, which are plotted in Figs. 9a and 9b, respectively. In both cases the expected inverse relationship between ΔUTH and ΔM is seen. It is clear from Fig. 9b that most points in the extratropics show T_B increases despite increases in UTH. To obtain an estimate of the influence of the temperature changes, a linear fit was performed for each region and the best fit lines plotted in Fig. 9. A linear dependence of ΔM on ΔUTH would not be expected for larger ΔUTH, as is clear from the scatter in the plots, but the displacement of this line from the origin will give an estimate of the effect of the mainly zonal temperature changes, if the effect of temperature is assumed to be independent of UTH and ΔUTH. The gradient will give an estimate of the influence of ΔUTH on ΔM. Table 1 gives the y intercepts and gradient of the linear fits.

Two main conclusions can be drawn from this table. First, the effect of the ∼4 K UTT increase in the Tropics is extremely minimal, compared to the strong influence of UTH changes. At high latitudes, where the UTT change is of order 2 K, the UTT influence on HIRS-12 brightness temperature is actually greater, while the effect of ΔUTH on ΔM is approximately halved. The halving of the response to ΔUTH is due to the higher-background UTH at midlatitudes than in the Tropics. The effect on T_B of UTH changes is opposed by the UTT changes, and with the lower response to ΔUTH most points reveal slight T_B increases, showing that temperature changes have dominated. Only stronger increases in UTH of the order of 3%–5% cause decreases in T_B at these latitudes, as can be seen comparing Figs. 7 and 9b.

The above analysis shows that the Tropics and subtropics are the most promising regions for the detection of climate change as regards its effect on HIRS-12 T_B and on upper-tropospheric humidities. At higher latitudes, the changes in UTH are smaller and the competing effects of temperature and relative humidity allow only a small signal in brightness temperature. In the Tropics and subtropics, changes in UTH dominate the signal despite the large UTT increases also seen in this region. This result gives confidence that in the real HIRS-12 data, a reasonable change in T_B in the Tropics can be attributed to UTH change rather than temperature change. The changes in T_B in this region are quite large, of up to ±2.5 K.

It is important to notice that in this climate simulation, the change in upper-tropospheric relative humidity is minimal on a global scale compared to the significant changes on regional scales. On a global scale it seems a fair assumption that very little extra cooling of the earth can occur in this part of the 6.3-μm water vapor band, though clearly one cannot draw inferences from this to the outgoing longwave radiation as a whole. The temperature increases of up to 4 K in the tropical upper troposphere show little effect on the brightness temperature compared to that of changes in UTH. Again the importance of looking for spatial patterns in the detection of climate change is emphasized.