The observation of changes in the earth’s spectrally resolved outgoing longwave radiation (OLR) provides a direct method of determining changes in the radiative forcing of the climate system. An earlier study showed that satellite-observed changes in the clear-sky outgoing longwave spectrum between 1997 and 1970 from the Infrared Interferometer Spectrometer (IRIS) and Interferometic Monitor of Greenhouse Gases (IMG) instruments could be related to changes in greenhouse gas composition. The authors present a new study that extends this to 2003, through the first use of a new, independent source of global atmospheric infrared spectra, from the Atmospheric Infrared Sounder (AIRS) experiment. AIRS is a dispersion grating spectrometer, while the other two were Fourier transform spectrometers, and this is taken into account in the analysis. The observed difference spectrum between the years 2003 and 1970 generally shows the signatures of greenhouse gas forcing, and also shows the sensitivity of the signatures to interannual variations in temperature. The new 2003 data support the conclusions found in the earlier work, though, interestingly, the methane (CH4) Q branch centered at 1304 cm−1 exhibits more complex behavior, showing a decrease in intensity in the difference spectrum between 1997 and 2003. Sensitivity analysis indicates that this is due to changes in temperature structure, superposed on an underlying increase in CH4. Radiative transfer calculations based on reanalysis data are used to simulate the changes in the OLR spectrum; limitations in such data and possible variations that could account for several observed effects are discussed.
The climate system has been studied extensively, and strong evidence linking surface temperature changes and greenhouse gas concentrations exists (Mitchell et al. 1995; Tett et al. 1999). The spectrally resolved outgoing longwave radiation (OLR) is a measure of cooling to space of the surface and atmosphere due to absorption and emission of greenhouse gases at characteristic wavelengths. Therefore it can be used to detect changes in radiative forcing and feedback processes due to the separation of the spectral signatures from different gases.
This study presents the use of a new data source, from the Atmospheric Infrared Sounder (AIRS), flying on the Earth Observing System (EOS)-Aqua satellite, launched in 2002, to extend the work of Harries et al. (2001). These authors compared data from 1970 [from the Infrared Interferometer Spectrometer (IRIS), which flew on Nimbus-4], and 1997 [from the Interferometic Monitor of Greenhouse gases (IMG), which flew on Advanced Earth Observing Satellite (ADEOS) I], and analyzed the differences between clear-sky spectrally resolved OLR. Changes were detected in the spectra that, through the use of a radiative transfer model, were attributed to known changes in radiative forcing. The new data from AIRS allows us to extend this analysis to include 2003, and to compare snapshots of the atmosphere in 1970, 1997, and 2003.
2. Observations and data
Spectrally resolved OLR observations at reasonably high spectral resolution with near-global coverage, are sparse, coming from only a small number of experiments.1 The data used in this study were recorded by the following instruments: IRIS (Hanel and Conrath 1970; Hanel et al. 1971); IMG (Kobayashi 1999; Kobayashi et al. 1999); AIRS (Aumann and Pagano 1994). The principal properties of the IRIS, IMG, and AIRS instruments are listed in Table 1.
In each case, the usable period of overlapping data is April, May, and June, and so we use this period throughout. The usable common spectral range is 700–1400 cm−1, defined by the increasing noise at lower wavenumbers in the IMG spectra, and at higher wavenumbers in the IRIS spectra. Note that the two Fourier Transform Spectrometer (FTS) instruments, IRIS and IMG, provide continuous spectral data from the Fourier transform process: for operational reasons, breaks in the spectral scan of AIRS were used. This means that gaps occur in the difference spectra involving AIRS. To clarify data processing and ease comparisons between the three possible difference spectra these gaps have been introduced to IRIS and IMG.
IRIS was a series of instruments launched by National Aeronautics and Space Administration (NASA) in the late 1960s and early 1970s. IRIS-D flown on Nimbus 4 provided all of the data that are available. Nimbus 4 was launched in April 1970 into an 1100-km altitude sun-synchronous polar orbit. It recorded data until January 1971 when it was switched off, having fulfilled its design brief. IRIS recorded spectra between 400 and 1600 cm−1 but wavenumbers above 1400 cm−1 suffer from high noise. The ground footprint was 95 km in diameter, and the apodized spectral resolution was 2.8 cm−1.
IMG was built by National Space Development Agency of Japan (NASDA) and launched in August 1996 onboard the ADEOS-1 satellite. Data were recorded between November 1996 and June 1997, when operations ceased because of satellite failure. IMG was developed to provide continuous coverage of the spatial distribution of the greenhouse gases and to measure detailed profiles of water vapor and temperature. IMG had a spectral resolution of 0.1 cm−1 and a spectral range of 600–3030 cm−1 (Kobayashi 1999). The detector used in this study recorded between 600 and 2000 cm−1. Usable data are available at wavenumbers above 700 cm−1 because of noise at lower wavenumbers. ADEOS orbited in a 797-km altitude polar sun-synchronous orbit giving a square ground footprint of 8 km by 8 km. IMG operated in a 4-day-on, 10-day-off cycle. A mirror alignment problem led to only 15% of the spectra recorded being available with acceptable noise levels.
AIRS was launched on the EOS-Aqua satellite in May 2002. Aqua orbits in a 705-km polar sun-synchronous orbit. AIRS is a grating spectrometer. It uses this design to achieve similar resolution and spectral range to IMG at higher speeds. The spectral range of AIRS is from 650 to 2700 cm−1 measured by 2378 separate detector channels, which are separated into 17 modules of detectors, resulting in noncontiguous spectral coverage. Eight gaps occur in the 700–1400 cm−1 spectral range considered in this study, as described in Table 1. One hundred eighty channels either have failed or are too noisy to use in the 700 to 1400 cm−1 region. These spectral gaps are interpolated over in the analysis, except around 1226 and 1347 cm−1 where multiple neighboring channels failed and additional spectral gaps are created. The detectors in AIRS are 10 by 10 μm squares, giving a resolution of between 0.4 and 1 cm−1 depending on the wavenumber.
The sampling geometry for AIRS is very different to that of the other instruments. IRIS viewed nadir through a fixed aperture and image compensation mirror. However, the detector arrangement of IMG means that two of the three detectors viewed the earth from slightly off-nadir. The detector recording the data used here is one of these off-nadir viewing detectors. To enhance spatial coverage, AIRS scans to ±49.5° cross track as the satellite moves forwards taking 90 spectra with an instantaneous field of view of 1.1° in a row perpendicular to the direction of motion. This gives a ground footprint of 13.5 km diameter at nadir but closer to 41 km by 22.4 km at 49.5°. Only the central eight spectra are used in this study to ensure that off-axis beams with significantly longer atmospheric pathlengths than that recorded by IRIS are not included. We have concentrated our study on the central Pacific region (defined as 10°N–10°S, 180°–230°E) because this region is one of the better sampled regions in all three datasets.
3. Comparison method
a. Differing instrument characteristics
The differing characteristics of the three instruments require careful normalization to allow the detailed differencing of spectra. The reader should consult Harries et al. (2001) and Brindley and Harries (2003a, b) for extensive discussion of the details of these issues.
In the present work, we draw attention to some specific issues. First, to compare the spectra from the three instruments, it was necessary to degrade the IMG and AIRS spectra to the lower spectral resolution of IRIS. This was done in two steps. First, the spectral resolution of the IMG data was reduced to that of AIRS by multiplying the IMG interferograms (the Fourier transform of the spectra) by an appropriately sized Hamming function (Harris 1978). The Hamming window was chosen as this was the apodization function used on the IRIS data. The AIRS data and AIRS resolution IMG data cannot be reduced to the resolution of IRIS by multiplying their interferograms by an appropriately sized Hamming window because of the gaps in the AIRS spectrum. Therefore, their resolutions were reduced by convolving the spectrum with the Fourier transform of the Hamming window. The resolution reduction process uses a variable sized window to ensure the resolution of the AIRS data is correct across the entire spectrum. Spectral gaps were added to the IRIS and IMG spectra to match those in the AIRS spectra.
A second major difference in the instrument characteristics is the fields of view. In an FTS, the effect of the finite field of view is to broaden spectral features and shift them to lower wavenumbers, because off-axis rays travel a longer path through the interferometer than on-axis rays (Thorne 1988). Grating spectrometers do not exhibit such effects. The effect of spectral broadening is taken into account concurrently with the resolution degradation described above. Shifting to lower wavenumber is taken into account by shifting the IMG and AIRS data to the more highly shifted IRIS positions.
b. Identification of cloud-free spectra
To reduce the amount of variability seen in the spectrum, cloud-free spectra are used. Brindley and Harries (2003a) discuss the use of all-sky data to make similar studies, but conclude that IRIS and IMG do not have adequate sampling to provide the required accuracy. A two-stage process for cloud and dust identification was used. In the first stage, the equivalent blackbody brightness temperature in the most transparent part of the spectrum (1127.71 cm−1): transmittance from surface to space = 95% (Iacano and Clough 1996), is compared to the known sea surface temperature (SST), obtained from the National Centers for Environmental Prediction (NCEP)–National Center for Atmospheric Research (NCAR) 40-Year Reanalysis dataset (Kalnay et al. 1996). A threshold defining the permitted difference between the observed radiance and that calculated for the known SST was determined, taking account of the small amount of absorption present at this wavenumber. This threshold was determined by looking at the maximum difference between the brightness temperature at 1127.71 cm−1 and the sea surface temperature in a set of simulated clear-sky spectra created using a band model and the NCEP reanalysis data. Standard deviations were examined to ensure that no signature of cloud remains.
The second stage was designed to remove residual cloudy spectra not removed by stage 1, which might be contaminated by thin cirrus (Ackerman et al. 1990). This is based on the fact that the effect of cirrus cloud on the spectrum is to introduce a gradient in brightness temperature across the atmospheric window. Differences between brightness temperatures at 913.57 and 1250.08 cm−1 were compared to an acceptance threshold chosen by the same method as above. A high difference indicates a high tilt across the atmospheric window, indicative of cirrus cloud in the spectrum. In this study, slightly different wavenumbers had to be chosen than those used in Harries et al. (2001), because of the gap in the AIRS spectra at 909.8 cm−1, which was used in the second part of the cloud identification in Harries et al. (2001). The effect of this was that slightly different spectra were included in the IMG and IRIS datasets and so the 1997–1970 difference spectra presented here may show small differences from that presented in Harries et al. (2001).
4. Observational spectral OLR differences
a. Sampling considerations
The numbers of spectra before and after removing cloud in each dataset are listed in Table 2. A larger proportion of spectra are removed in the case of IRIS than AIRS and IMG, respectively, due to a higher probability of cloud in the larger fields of view. The spatial and temporal sampling of the clear-sky spectra is shown in Fig. 1.
The irregularities in the sampling of IMG are visible in Fig. 1d reflecting the 4-day-on, 10-day-off power sequence, and the effect of removing noisy spectra. The sampling of AIRS is good, with the exception of the gap in temporal sampling between 26 May and 12 June 2003, due to the data in this period being corrupted. Otherwise, the sampling of all datasets shows no obvious biases. The standard deviation of all datasets are very similar (see Figs. 2 and 3 below): if either the IRIS or IMG sampling were inadequate, the standard deviation would be a different magnitude to that of the well-sampled AIRS dataset.
The AIRS data were subsampled to match the sampling pattern of IRIS. For each IRIS spectrum, the closest geographical AIRS spectrum was selected on any given day. If no AIRS spectrum existed for that day, no spectrum was selected. To confirm that the imperfect sampling of IRIS did not have a significant effect on our analyses, we calculated the mean of the subsampled and the full AIRS datasets. These could not be shown to be statistically significantly different at the 95% confidence level. Further tests with many random spatial realizations of the AIRS data supported this conclusion. Brindley and Harries (2003a) have also shown that the sampling patterns of IRIS and IMG result in average spectra within ±0.3 K of the true mean, which would be obtained with a fully sampled dataset.
b. Average spectra and spectral differences in the central Pacific
Figures 2 and 3 present the average brightness temperature spectra obtained using the above procedures. These are averages for April, May, and June for IRIS (1970), IMG (1997), and AIRS (2003), respectively, in the central Pacific region (10°N–10°S, 180°–230°E). The spectral resolution is 2.8 cm−1, and the data gaps correspond to those in the AIRS data. Figures 2a and 3a show the three average spectra for the 700–1150 and 1200–1400 cm−1 spectral regions respectively, superposed to show the consistency in the spectral resolution and position of individual features.
This indicates that the resolution and field of view corrections have been correctly applied. The principal features that appear in the spectra are as follows. The water vapor ν2 vibration-rotation band in the 1200–1400-cm−1 region; the ozone ν3 band centered at 1043.4 cm−1; the edge of the carbon dioxide ν2 band observed from 700 cm−1 (the limit of the IMG data) to 800 cm−1; and the methane ν4 band with a strong Q branch centered at about 1304 cm−1 are all observed. Weaker bands due to CFCl3 (850 cm−1) and CF2Cl2 (915 cm−1) were also shown in Harries et al. (2001) to be just visible against the noise. Figures 2b and 3b show the three spectra from Fig. 2a and 3a separated by an offset of 10 K, for greater clarity. The high degree of reproducibility between the three independent cases is indicative of the high accuracy of the measurements and the techniques used to normalize them. It also reminds us that the spectral signatures due to climate change are very small, demanding high accuracy to be detected.
Figures 2c and 3c show the standard deviations (σ) of the three averages for the spectral regions 700–1150 and 1200–1400 cm−1, respectively. Previous similar analyses for the IRIS data have been reported, for example by Iacano and Clough (1996), though this is the first comparison for all three instruments. The similarity of the standard deviation should be noted, indicating that the three average spectra have been equally well cloud-cleared. Working from lower to higher wavenumbers, we note the following:
Very low values of σ toward the center of the CO2 band: the increased optical depth leads to higher layers of the atmosphere, into the stratosphere, being sampled. The stratosphere is more stable than the troposphere, leading to lower values of σ.
The σ values between ±1.0 and 1.7 K in the atmospheric window are dominated by variability of surface temperature and emissivity, and to a lesser extent lower atmosphere temperature and humidity.
There is little evidence for any significant variability of O3, known to be stable in the Tropics.
The σ increases up to ±5.0 K in the ν2 band of water vapor, which in the lower troposphere is highly variable. We shall see this effect dramatically in the analysis of spectra below.
The brightness temperature difference spectra between AIRS and IMG (top of Fig. 4), AIRS and IRIS (middle of Fig. 4), and IMG and IRIS (bottom of Fig. 4) have been separated by an arbitrary 5 K for clarity. An initial inspection indicates that the processing of the data has not caused any major artifacts. In all cases the difference spectra are seen to have consistent and reproducible features. The only sign of asymmetry (which could indicate a mismatch of wavenumber scales between the spectra) is in the CO2 (0110 → 1000) band at 720 cm−1, which may be due to its position on the very steep high frequency wing of the CO2 fundamental centered at 667 cm−1.
In all three difference spectra, the brightness temperature difference in the atmospheric window between 800 and 1000 cm−1 is zero or slightly positive, indicating a warming of surface temperatures between the three time periods. Indeed, although at the limit of uncertainty due to noise, the positive anomaly in this region seems to increase with time duration between observations, in the order AIRS–IMG, IMG–IRIS, and AIRS–IRIS. This is in agreement with the trend in sea surface temperatures in the NCEP reanalysis. A negative brightness temperature difference is observed in the CO2 band at 720 cm−1 in the IMG–IRIS (1997–70) and the AIRS–IRIS (2003–1970) difference spectra, indicating increasing CO2 concentrations, consistent with the Mauna Loa record (Keeling et al. 1995). However, this channel in the difference is also sensitive to temperature, and we note that in the 2003–1997 difference, despite a growth in CO2 between these years, there is no signal at 720 cm−1. A signature is observed in the ozone band around 1043 cm−1 in all difference spectra, though this is obscured by an AIRS data gap. The methane band centered at 1304 cm−1 is perhaps the most obviously interesting feature of the entire difference spectra. A strongly negative brightness temperature difference is observed for the AIRS–IRIS (2003–1970) difference spectra, supporting the previously reported result for IMG–IRIS (1997–70), but a signature in the opposite sense is seen in the AIRS–IMG (2003–1997) spectrum. This would seem to indicate either that the methane concentration in 2003 was less than that observed in 1997 or that changes in temperature have dominated over an increase in methane. The methane concentration measurements presented in Dlugokencky et al. (2003) would suggest the latter.
To introduce a more quantitative idea of uncertainty in these results, we present the same data in Fig. 5, though here we have added two further pieces of information. First, the thickness of the difference spectrum line in each case indicates the standard error on the mean σm = σ/N, where σ is the standard deviation (see Figs. 2c and 3c), and N is the number of spectra in each average. The standard error on the means for each average spectrum, σmi are combined in quadrature to give the standard error on the difference, σmd = (σ2m1 + σ2m2)1/2. We see that σmd is significantly larger when the numbers of spectra are smaller. The small size of the standard errors indicates that the distribution of brightness temperatures at all wavelengths is well sampled. Second, we have indicated the statistical significance of the difference spectra, which was tested against the null hypothesis that the true difference spectrum at each wavenumber is zero, using a two-sided Student’s t test. Regions with greater than 95% confidence for rejecting the null hypothesis are indicated by the vertical gray shading. The statistical significance of the majority of the difference spectra is a further indication that the data sampling is adequate.
We note that we performed similar analyses (not reported here) for other locations around the globe and the major features in the difference spectra seen in Figs. 4 and 5 were observed in all cases. The central Pacific region was chosen in this work to allow direct comparison with the results of Harries et al. (2001). Studies of other regions will be published later.
c. Key results in the observed difference spectra
We refer to Fig. 5, and consider the three difference spectra.
The CO2 band at 720 cm−1, though asymmetric for the reasons stated earlier, nevertheless shows some interesting behavior, with strong negative brightness temperature difference features for 1997–1970 and 2003–1970: whereas, the 2003–1997 (a much shorter period, of course) shows a zero signature. Since we know independently that the CO2 concentration globally continued to rise between 1997 and 2003, we must conclude that the 2003–1997 result must be due to changes in temperature that compensate for the increase in CO2. This would mean a warming of the atmosphere at those heights that are the source of the emission in the center of this band. This is somewhat contrary to the general (small) cooling of the stratosphere at tropical latitudes.
In the statistically significant regions of the atmospheric window, excluding the O3 band centered at 1043 cm−1, there is some evidence (e.g., between 850 and 900 cm−1) that the larger the time difference in the difference spectrum, then the larger the positive difference signal. However, elsewhere (e.g., the narrow statistically significant strip between 990 and 1000 cm−1) this is not so clear. However, it is certainly true overall that the shortest time difference, 2003–1997, produces the window difference that is closest to zero. Clearly, we are at the very limit of the accuracy of these results.
It is difficult to reach any conclusions about the O3 band centered at 1043 cm−1, largely due to the AIRS data gap in the middle of this band. Future investigation using datasets with continuous spectra, such as the Tropospheric Emission Spectrometer on EOS-Aura (Beer et al. 2001), may allow further insight into the changes in this spectral region.
The CH4 Q branch centered at 1304 cm−1 presents one of the most interesting results. The strong negative difference spectrum feature in 1997–1970 reported in Harries et al. (2001) is repeated here, and also the 2003–1970 signature here is strong and negative (though not as strong as for 1997–1970). However, the 2003–1997 result shows a clear positive feature in the difference spectrum. In our processing, this feature statistically was a very robust feature. The possible causes could only be a reduction in atmospheric CH4 between 1997 and 2003, or a warming between these years of the atmospheric layers giving rise to the CH4 emission, or both.
Finally, the behavior of the difference spectrum in the ν2 H2O band covering the 1200–1400 cm−1 region (which is largely statistically significant) shows a difference between 2003 and 1997 and the others: for example, strong difference features at 1365, 1375, and 1389 cm−1, which are negative in 1997–1970 and 2003–1970 are (more weakly) positive in 2003–1997. This could suggest either a reduction of humidity or an increase in temperature between 1997 and 2003. The source for these lines is in the lower and middle troposphere.
Our conclusion from this analysis so far is the difference spectra are well behaved, not affected by sampling, and contain statistically significant spectral signatures at 720 cm−1; in parts of the atmospheric window; at 1304 cm−1; and between 1200 and 1400 cm−1 in the water vapor band. The nature of the signatures in the three difference spectra leads us to conclude that in the short period between 1997 and 2003, in addition to the known concentration changes of CO2 and CH4, changes in atmospheric temperature seem to have dominated over increases in concentration. Simulation of the observed spectra using the best available information about the state of the atmosphere can be used to aid interpretation of these results.
5. Simulation of spectra
a. Description of modeling and sensitivity studies
We considered the use of two independent reanalyses, the NCEP reanalysis (Kalnay et al. 1996) and the European Centre for Medium-Range Weather Forecasting (ECMWF) reanalysis (Uppala et al. 2005), to simulate the difference spectra presented above. A comparison of the two provides a useful measure of uncertainty in these reanalysis data. We have also performed sensitivity analyses to investigate areas of the spectrum where disagreement between observation and simulation occur.
Spectra were simulated using the MODTRAN version 3.7 band model (Berk et al. 1989) at a spectral resolution of 1 cm−1. The resolution was then reduced to match the observational value of 2.8 cm−1 using the Hamming window. MODTRAN was run with user-defined profiles constructed using humidity, temperature, and sea surface temperature from the reanalyses, and other atmospheric composition information from a number of sources: these will be referenced as they are considered below. Profiles were defined for temperature, water vapor, CO2, CH4, O3, N2O, CFCl3, and CF2Cl2. Spectroscopic data from the 1996 version of HITRAN (Rothman et al. 1998) were used. We note that the reanalyses provide only the water vapor concentration and temperature profiles, and surface temperature. The effect of the other trace gases on the temperature and humidity of the atmosphere at the time of the reanalysis are automatically included in the reanalysis temperature and humidity data. However, no trace gases are provided from the reanalyses, but are introduced into the calculation from other sources. So the radiative forcing of these other trace gases is included only once.
The approach was to simulate the average spectrum using the average atmospheric state for the region of interest, rather than to simulate many spectra for many atmospheric grid boxes, and average those in the region for all days in the analysis period, which would be extremely computationally expensive. Tests showed that the average atmosphere approach produce difference spectra, which, outside the 1200–1400 cm−1 region, are within 0.2 K of those using the more computationally intensive method. Within the water vapor band between 1200 and 1400 cm−1, using the average spectrum produced a slightly colder (<1 K) spectrum. Bearing in mind the intrinsic high variability of the spectrum in the water band (see Fig. 3), this is acceptable. We also have examined the effect of using the MODTRAN band model, rather than a line-by-line code. Consistent with the work of Thériault et al. (1993), this showed only minimal differences, the largest (1 K) being in the methane band. Both the speedier averaging and the band model were used.
1) NCEP/Climate Monitoring and Diagnostic Laboratory simulation
Temperature and water vapor profiles, and surface (skin) temperature were taken from the NCEP reanalysis. Since water vapor profiles were available only up to 300 hPa, values above this level were adopted from the ECMWF reanalysis described below. Tests showed that this introduced insignificant differences from, for example, simply using a standard model stratosphere. However, the depiction of the water vapor average distribution in the troposphere presents special problems (see below). Ozone profiles were taken from the STOCHEM chemical transport model forced by realistic emissions scenarios in the troposphere (Collins et al. 1997), and used measured Stratospheric Processes and Their Role in Climate (SPARC) trends in the stratosphere (Randel and Wu 1999). For 2003, these model runs were not available so the 1997 profile was scaled by the TOMS total column abundance (Heath et al. 1975; McPeters et al. 1998). Surface concentrations of the other gases are taken as the average of the Climate Monitoring and Diagnostics Laboratory (CMDL) flask measurement system values between 10°N and 10°S for methane (Dlugokencky et al. 1994) and globally for N2O and the CFCs (Conway et al. 2003; Elkins et al. 1993). To obtain vertical profiles, these concentrations were used to scale the standard U.S. tropical atmospheric profiles (Anderson et al. 1986) in order to obtain vertical profiles. For 1970, the flask values were extrapolated back in time from the earliest available measurement as in Boden et al. (1994). The profiles used in this simulation are shown in Tables A1, A3 and A5 in the appendix.
2) ECMWF/CMDL simulation
The ECMWF have produced a 40-yr reanalysis dataset (ERA-40) for September 1957 to August 2002. This was used to simulate the spectra observed in 1970 and 1997. To simulate the 2003 spectrum, the ECMWF current operational analysis is used. The minor gas concentration profiles are the same as the NCEP/CMDL simulations. The SST is not determined as a separate parameter, and therefore the temperature at 1000 hPa was used as a proxy for SST. The profiles used in this simulation are shown in Tables A2, A4 and A6 in the appendix. The water vapor average profile presents special difficulties, which we turn to now.
3) Water vapor
This is a special case, for two reasons. First, the average profiles of water vapor obtained from NCEP and ECMWF reanalyses are very different: at some heights and times, the two analyses are different by as much as a factor of 2 (see appendix). Second, the difference spectrum is very sensitive to the absolute values and the vertical distribution of water vapor. This means that the description of the water vapor field is the most critical problem (for clear skies) in trying to make accurate simulations of the changes in the atmospheric emission spectrum. The tables in the appendix indicate that the data for 1970 show the largest differences between NCEP and ECMWF, indicating that the atmosphere is not so well sampled in that year as later. It was decided to adopt the NCEP profile up to 300 hPa, for consistency with the temperatures used (see next section), and above 300 hPa to use the ECMWF profile. Tests showed that spectra simulated using ECMWF at the top of the profile did not significantly differ from those simulated using, for example, the tropical standard atmospheric water profile. It has to be recognized, however, that the limitations on our knowledge of the atmospheric humidity, in particular, sets a limit on the accuracy of the difference spectrum simulations.
b. Simulation and interpretation
1) Simulations of absolute spectra
We compare the results of the simulations by focusing on the atmospheric window between 800 and 950 cm−1, and the CH4/H2O bands between 1280 and 1400 cm−1. The appendix lists the values of temperature and composition parameters used as inputs to these calculations. Figure 6 presents the brightness temperature spectra for three cases: simulations using the NCEP/CMDL and using the ECMWF/CMDL input data, plus the observed spectra, over these two spectral regions.
The top curves are for 1970, the middle pair for 1997 and the lower pair for 2003. The data gaps are those in the AIRS measurements. Note the difference of brightness temperature scales in the two columns.
The 800–950 cm−1 region shows up to about 2-K difference between the observations and simulations, always in the sense of the simulation being lower than the observation. This tendency could be associated with incorrect surface temperature, or incorrect simulation of the lower, humid layers of the atmosphere, where continuum absorption is important. Our earlier tests of the effects of uneven sampling indicate that errors as large as 2 K from this source are unlikely. In 1970, the two simulations indicate close agreement, within ∼0.5 K, but are ∼1.5 K lower than observed. In the two later cases, however, the NCEP/CMDL simulation is within 1 K or better of the observation, while the ECMWF/CMDL simulation is colder by 2 K, a significantly poorer agreement. In this frame, therefore, it seems that the NCEP/CMDL simulation is more accurate.
The 1280–1400-cm−1 region shows differences between the observations and the simulations of up to 10 K in a few places, but within 2 K more generally, especially for 1997 and 2003. The quality of agreement between the simulations and observations is poorer in 1970 than in later years, possibly reflecting the poorer state of atmospheric and surface observations at that time. In 1970, the ECMWF/CMDL simulation is generally closer to the observations, though still with considerable differences from the observed spectrum. The agreement between simulations and observation is much better in 2003 than in earlier years. The differences between simulations and observations are both positive and negative. This spectral region is heavily dominated by water vapor and temperature, so any errors in knowledge of these fields will cause significant errors in the spectrum. These simulations and comparisons indicate that there is no clear difference in accuracy between the NCEP/CMDL and ECMWF/CMDL simulations: for simplicity and clarity, we therefore arbitrarily selected the NCEP data to perform the simulations.
To examine the detailed comparison of observation against simulation for the three years 1970, 1997, and 2003, we have formed the difference between the observed average spectra and the corresponding NCEP-based simulation (Fig. 7). The three frames are remarkably similar to each other. Clearly, some systematic, consistent messages are present. At 700 cm−1, the negative difference signal indicates that in the midtroposphere the temperatures used in the simulations are consistently too high, leading to too high a simulated radiance. The characteristic asymmetry in the 720-cm−1 band of CO2 still exists, and has been discussed earlier. The window between 800 and 1000 cm−1 is in all cases remarkably flat, punctuated by weak CO2, H2O, and CFC lines, though the difference Obs-NCEP is uniformly positive, indicating that surface/lower atmosphere temperatures in the simulations are too low, or that some source of absorption in the simulation is not well represented, for example perhaps aerosol absorption or incorrect water vapor continuum absorption. The ozone band centered at 1043 cm−1 shows remarkable similarity for all differences: since the ozone column amounts and profiles used in the three simulations are similar, this accords with the known low variability of tropical ozone.
The CH4 band centered at 1304 cm−1 shows strongly in each year. The sense is that the simulated concentration of CH4 is too low, or the temperature of the emitting methane in the simulation is too high. This supports the conclusion drawn above about mid/upper tropospheric temperature being too high in the simulation from the CO2-band behavior. There is also a suspicion, discussed further below, that the spectroscopy of the CH4 Q branch may be causing some of this disagreement. At higher wavenumbers, there is more variability and generally fairly large disagreement between the observation and simulation reflecting the large uncertainties in humidity profiles mentioned earlier.
2) Simulation of difference spectra
The simulations for 1970, 1997, and 2003 have been treated like the observations, and differences AIRS–IMG, IMG–IRIS, and AIRS–IRIS formed from the simulations described above. These simulated difference spectra (NCEP) are shown in the lower frame of Fig. 8: the upper frame shows a repeat of the observations shown in Fig. 4.
As in Harries et al. (2001), the simulated difference spectra show general agreement with those observed, indicating that observed changes in the spectrally resolved OLR can be attributed to the known changes in greenhouse gases. The 1997–1970 difference is the same comparison as in Harries et al. (2001), and reference to that paper will show a very similar result, though with small differences, by virtue of the AIRS data gaps and slightly different frequencies used for cloud clearing as noted before. These differences are not large enough, however, to affect the conclusions, in 2001 or here.
We should note that our simulations have shown that the difference spectra simulated using reanalysis data are particularly sensitive to small changes in the concentration and vertical distribution of temperature, methane, and particularly water vapor. Thus there is considerable uncertainty involved in the use of reanalysis data to simulate the outgoing spectrum, and this fact must be borne in mind. For example, the water vapor profiles from NCEP and ECMWF for the tropical region studied show disagreement at some heights by a factor of 2.
An additional source of uncertainty in the water vapor and temperature reanalysis fields is due to the presence of clouds. The profiles may be significantly changed in the vicinity of a cloud and a significant bias may arise between the actual and reanalyzed water and temperature fields. Griggs (2005) showed that in the water vapor ν2 band this could give an uncertainty in the difference spectrum 1997–1970 of the order of ±4 K. Thus, a comparison with reanalysis based simulations of the difference spectra, while of great value in interpretation, must be viewed with some caution.
In Fig. 8, both the observations (Obs) and simulation (NCEP) show a rather flat difference spectrum, close to zero except in the ν2 band of H2O. This would be expected, since in the short time interval between 1997 and 2003 little growth of greenhouse gases occurred. The concentration and ν2 band intensity of H2O is highly variable. CH4 shows an unexpected positive difference in the observations. Investigation using sensitivity studies in which temperature and methane were varied showed that, given the known small growth in CH4 concentration between 1997 and 2003 (1.73 to 1.75 ppmv) the observed difference must be due to a warmer mid/upper troposphere in 2003 than 1997. This would also account for the smaller observed methane difference signal in 2003–1970 than in 1997–1970. The reanalysis data, however, do not show a clear change in temperature between 1997 and 2003, and the simulation does not pick up this positive difference feature, but rather produces a weak negative brightness temperature difference signature at 1304 cm−1.
This difference spectrum closely accords with the difference spectrum observed in Harries et al. (2001) for 1997–1970. The water vapor ν2 band exhibits very similar behavior to previous, although we note the ±4-K uncertainty in absolute accuracy that can arise in this part of the difference spectrum as discussed above, and the difference between observation and simulation is easily accounted for by this. In our sensitivity studies, we have varied the temperature and humidity structure of the upper troposphere and find that the difference spectrum between 1300 and 1400 cm−1 can be moved in the simulation between +3 and −3 K using different, realistic humidity profiles.
The observed magnitude of the negative methane signal at 1304 cm−1, smaller than in 1997–1970, supports a relative warming in 2003 in the mid/upper troposphere. The simulation of this methane feature shows a characteristic smaller intensity at the center, and a larger width than in the observations. This feature of the simulation has been noticed before (Harries et al. 2001; Griggs 2005) and may suggest some anomaly in the band model representation of this complex Q branch and its rotational line substructure, for example, due to the influence of line mixing. We are currently studying this question, which will be discussed in a future publication. Despite these differences between observation and simulation, the observed feature of the difference spectrum is clearly captured by the simulations, despite the uncertainties largely associated with the reanalysis data.
Another significant feature is the absence in the simulation of any significant feature at 720 cm−1, where a weak CO2 Q branch exists in the observations, albeit an asymmetric feature as discussed above. We investigated this feature in a sensitivity study, by introducing a variation of the temperature of the upper troposphere/lower stratosphere. We found that cooling the 2003 (or warming the 1970) lower stratosphere by 2 K or more produced a very clear feature at 720 cm−1, thereby calling into question again the precise accuracy of the reanalysis data, since this feature is very clearly seen in the observations. Interestingly, this small adjustment also produces a feature in the ozone band difference spectrum at 1040–1050 cm−1, which is more similar to the observation. We conclude that the reanalysis temperature data have uncertainties of at least ±1 K in both the 1970 and the 2003 temperatures in the lower stratosphere: this seems quite possible from our earlier discussion of the reanalysis data.
Finally, there exists a marked gradient in the simulated spectrum between 800 and 700 cm−1, which is absent in the observations. This coincides with the far wings of the strong CO2 band centered at 667 cm−1. In sensitivity tests, this gradient showed sensitivity to the amount of CO2, and is therefore related to the strong CO2 band, and may reflect reanalysis uncertainties in temperature. The gradient also showed sensitivity of water vapor amount. The mechanism by which water vapor could affect this part of the spectrum is not clear, but is perhaps associated with the continuum absorption representation in MODTRAN. This will be further examined in future work.
The observed and simulated difference spectra in the water vapor ν2 band show similar spectral features to each other, but as already pointed out, uncertainties here can be up to ±4 K. The absence in the simulation of the observed 720 cm−1 weak CO2 Q branch, which is clearly observed in the measured difference spectrum arises for the same reasons as given above for 2003–1970. Cooling the 1997 (or warming the 1970) lower stratosphere by 2 K or more produced a clear feature at 720 cm−1. Since the observations quite clearly show the 720 cm−1 feature, we conclude that the reanalysis temperature data have uncertainties of at least 1 K in both the 1970 and the 1997 temperatures in the lower stratosphere.
The difference in the magnitude and width of the negative methane signal at 1304 cm−1 between observation and simulation has been commented on before. We note that the cooling of the lower stratosphere in 1997 also improves the simulation of the methane band slightly. This difference between observation and simulation was previously noted (Harries et al. 2001; Griggs 2005), and may be associated with uncertainties in the detailed representation of the methane Q branch. This is being investigated in a future paper.
The gradient in the simulation between 700 and 800 cm−1 is similar to that in the 2003–1970 case, and does not appear in the observations. It is sensitive to CO2 amount and water vapor. The mechanism for this latter sensitivity may involve continuum absorption in water vapor.
The principal conclusions of this work are
The new AIRS infrared spectral data have been successfully added to an ongoing study of spectral signatures of climate change. The 2003 April–June tropical AIRS data compare extremely well with spectra recorded by IRIS and IMG in 1970 and 1997, suggesting considerable stability in the data, and successful normalization of the different instrumental characteristics.
The AIRS, IRIS and IMG spectra have variability statistics that are very similar characteristics, indicating that the variability is dominated by the earth system, not by the characteristics of the different instruments.
Observed emission spectra for 1970, 1997, and 2003 have been compared with simulations based on reanalysis data. Differences between observation and simulation are seen, of about 1 K in the atmospheric window, up to +10 K in the ν2 band of H2O, about +5 K in the ozone band at 1043 cm−1, and up to −8 K in the CH4 Q branch. Much of these larger differences may be due to uncertainty in reanalyzed water vapor.
Using the AIRS data with data from the IRIS project allows a difference spectrum to be generated for the period 2003–1970, a period of 33 yr. Changing spectral signatures due to CH4, CO2, and H2O on decadal time scales are observed using the new AIRS data, thus adding confidence to the previous 1997–1970 study.
The 2003–1997 difference spectrum shows only weak signals of greenhouse gas changes because of the short time interval. It covers a period where the annual increases in atmospheric methane concentrations were much smaller than in the previous three decades. The difference spectrum shows how, on short time scales, temperature changes can mask effects due to growth in greenhouse gases, since the 2003–1997 CH4 observations can only be explained by a small warming (circa 1 K) of the upper troposphere between 1997 and 2003.
The use of reanalysis data to simulate the average and difference spectra in 1970, 1997, and 2003 is useful, but limited by the significant uncertainties in reanalysis data (e.g., in humidity). The simulations of both CO2 and CH4 spectra indicate that the upper troposphere and lower stratosphere in the reanalysis datasets are 1–2 K warmer than observed by the spectrally resolving instruments. Simulations of the atmospheric window indicate reanalysis surface temperatures to be too cold by up to 1 K, or perhaps a deficiency in the modeling due to aerosol or continuum effects.
The simulated difference spectra, 1997–1970, 2003–1970, and 2003–1997, show a clear signal in CH4, through it is difficult to reproduce the precise observed shape of the methane feature. Reasons for this may include the influence of adjacent water vapor lines, the complexity of the Q branch (future work), and small errors in the reanalysis upper-troposphere temperature.
This work was supported by an NERC studentship. We thank H. Brindley and R. Bantges for advice, L. Chen and R. Goody for access to the IRIS data, and H. Kobayashi for the IMG data. AIRS data were provided by the GES DISC DAAC and NCEP data by the NOAA-CIRES Climate Diagnostics Center, Boulder, Colorado (ftp://ftp.cdc.noaa.gov/). ECMWF data were obtained via the BADC. CMDL data were provided by the Climate Monitoring and Diagnostics Laboratory.
Gas Concentration and Atmospheric Temperature Profiles Used in Simulation of Observed Spectra
* Current affiliation: Bristol Glaciology Centre, University of Bristol, Bristol, United Kingdom
Corresponding author address: Dr. J. A. Griggs, Bristol Glaciology Centre, University of Bristol, Bristol BS8 1SS, United Kingdom. Email: email@example.com
A strong case could be made for the development of a sensor to monitor the spectrum on a continuous basis, since the spectrum contains so much information about the state of the climate.