A trend analysis was applied to a 14-yr time series of downwelling spectral infrared radiance observations from the Atmospheric Emitted Radiance Interferometer (AERI) located at the Atmospheric Radiation Measurement Program (ARM) site in the U.S. Southern Great Plains. The highly accurate calibration of the AERI instrument, performed every 10 min, ensures that any statistically significant trend in the observed data over this time can be attributed to changes in the atmospheric properties and composition, and not to changes in the sensitivity or responsivity of the instrument. The measured infrared spectra, numbering more than 800 000, were classified as clear-sky, thin cloud, and thick cloud scenes using a neural network method. The AERI data record demonstrates that the downwelling infrared radiance is decreasing over this 14-yr period in the winter, summer, and autumn seasons but it is increasing in the spring; these trends are statistically significant and are primarily due to long-term change in the cloudiness above the site. The AERI data also show many statistically significant trends on annual, seasonal, and diurnal time scales, with different trend signatures identified in the separate scene classifications. Given the decadal time span of the dataset, effects from natural variability should be considered in drawing broader conclusions. Nevertheless, this dataset has high value owing to the ability to infer possible mechanisms for any trends from the observations themselves and to test the performance of climate models.
Large uncertainties exist in the performance of general circulation models (GCMs) in predicting future climactic states. The projections of future global mean temperature vary by a factor of 2 between the most commonly used climate models worldwide (Solomon et al. 2007). These differences exist due to uncertainties in both climate forcing and climate sensitivity. Improvements in model uncertainty will require testing model performance against credible observations (Goody et al. 1998, 2002).
Thermal infrared spectra provide an effective tool to evaluate GCMs since they contain the signatures of both the forcing and response of the atmospheric climate system (Leroy et al. 2008b,a). The spectral signatures of CO2, CH4, N2O, and other well-mixed greenhouse gases reveal the long-term climate forcing. The infrared window region can be used to observe clouds and aerosols (e.g., Turner 2005, 2008). H2O and CO2 lines can be used for profiling water vapor and temperature (e.g., Feltz et al. 1998; Smith et al. 1999). Furthermore, optimal detection methods can be used to obtain longwave feedbacks of temperature, water vapor, and cloud (Leroy et al. 2008a; Huang et al. 2010). Infrared spectra can be generated from the state variable output of GCMs using radiative transfer models. This means that model performance can be effectively evaluated by comparing spectral infrared observations with spectra from the GCM output generated by a forward model. This eliminates the need to use retrievals on the observations, which would introduce additional uncertainties. The observations, however, must extend over a sufficiently long period, have high accuracy, and have known demonstrable uncertainty over the entire observing period for the comparison to be meaningful.
The Climate Absolute Radiance and Refractivity Observatory (CLARREO), a high priority recommendation of the U.S. National Research Council’s decadal survey (National Research Council 2007), is a satellite mission designed to produce a high-accuracy long-term time series of radiance and atmospheric refractivity that can be used to detect trends in the climate system and test the performance of climate models. The infrared component of CLARREO will use a novel onboard calibration and validation system to directly tie the measured radiances to fundamental physical standards. Thus, the observed infrared radiances will have high accuracy and known uncertainty that will be determined on orbit to demonstrate the veracity of the observations.
Until CLARREO becomes operational, we need to examine existing long-term high-accuracy datasets that are similarly credible. The Atmospheric Emitted Radiance Interferometer (AERI) is a ground-based spectrometer that was designed with a strong emphasis on accurate calibration. The AERI measures downwelling emitted spectral infrared radiance for the purpose of improving radiative transfer models, retrieving atmospheric temperature and water vapor profiles, and investigating the radiative properties of clouds and aerosols. Downwelling spectra have the advantage that the manifestation of temperature and water vapor signatures are additive in radiance, thus amplifying their signal, in contrast to upwelling spectra in which temperature and water vapor signatures have the opposite sign, thereby partially canceling each other out. In addition, downwelling radiances have cold space as a background, which simplifies radiative transfer compared to the upwelling case where surface temperature and surface emissivity need to be taken into account. The AERI employs a sophisticated calibration system that ensures highly accurate spectra. AERI instruments have been in near-continuous operation at several field sites worldwide for well over a decade. This makes the AERI radiance observations an excellent dataset for the evaluation of long-term atmospheric trends. Furthermore, the AERI radiances can be used to evaluate the performance of climate models on a local scale and to test the analysis methods that will be employed with the global-scale dataset from CLARREO in the future.
In this paper, we discuss a trend detection exercise carried out with 14 years of AERI observations at the U.S. Department of Energy (DOE) Atmospheric Radiation Measurement (ARM) program site in the Southern Great Plains (SGP) in north-central Oklahoma (Ackerman and Stokes 2003). Our aim is to demonstrate analysis methods that can be used to detect long-term trends in observations. We show that significant trends exist in the 14 years of AERI data. While a complete explanation and attribution of the observed long-term trends will be a task for climate theorists, we present some hypotheses regarding the causes of the trends. In the following section, we describe the AERI instrument and the data analysis method used to obtain trends in spectral radiance. Section 3 presents the discussion of the trend results and their implications. We summarize our conclusions in section 4.
The Atmospheric Emitted Radiance Interferometer (AERI) is a ground-based instrument that was developed at the University of Wisconsin—Madison for the ARM program (Knuteson et al. 2004a). Over a dozen AERI units have been deployed and operated successfully from 1988 to present, providing measurements of absolutely calibrated downwelling spectral radiance worldwide. The AERI measures infrared radiance between 520 and 3020 cm−1 (3.3–19 μm), at a resolution of 1 cm−1, using two detectors cooled to cryogenic temperatures with a Stirling cycle cooler. A gold-coated rotating fold mirror allows the AERI to selectively view atmospheric scenes or calibration targets. The AERI uses two high-emissivity blackbodies for radiometric calibration, using the method of Revercomb et al. (1988). The blackbody employs a light-trapping cavity geometry and is painted with a highly emissive diffuse black paint. Thermistors are embedded in the blackbody cavity to measure its temperature. The temperature probes are calibrated against National Institute of Standards and Technology (NIST) standards and measure the absolute temperature of the AERI blackbodies within an uncertainty of 0.05 K (3σ).
The performance of each AERI instrument is evaluated in the laboratory and a suite of diagnostics is carried out, including nonlinearity characterization, electronics calibration, and field of view tests. The combined uncertainty of each AERI spectrum is ensured to be better than 1% (3σ) of the ambient radiance by the design of the instrument, its calibration subsystem, and laboratory diagnostics (Fig. 1a) (Knuteson et al. 2004b). Expressed in terms of scene radiance, the combined uncertainty has typical values of 1%–2%, though it can get as high as 8.5% for very cold and dry scenes (e.g., clear skies in the wintertime) (Fig. 1b). Over the span of a 14-yr observing period, these values correspond to a 3-σ upper bound for the instrument uncertainty in a trend of 0.1%–0.3% yr−1 in the typical case and 1.2% yr−1 in the worst case (wintertime clear sky). We can show, however, that the trend uncertainty is not dominated by either random (type A) or systematic (type B) instrument uncertainty. Seasonal means are used in this trend analysis, thus many hundreds of spectra are averaged for each data point, thereby reducing the random instrumental uncertainty (type A) to negligible levels. The systematic bias (type B uncertainty) of the AERI-01 instrument deployed at the SGP site has been validated by another AERI instrument during the water vapor intensive observation period field campaigns in 1996, 1997, and 2000 (Revercomb et al. 2003), by the collocated AERI-E14 instrument from 2004 to present, and through a full set of laboratory diagnostics in 2005. The instrument performed within the original error budget in each case. In Fig. 1c we show the radiance difference between the collocated AERI-01 and AERI-E14 instruments during the overlap period; the difference measurements all lie within the expected value. Furthermore, the difference measurements exhibit no observed year-to-year secular trend. In addition, routine views of the hot calibration blackbody are used to monitor the responsivity of the instrument; therefore, both the calibration and the sensitivity of the AERI are known throughout the course of the observations. Since the combined uncertainty in the AERI observations is low, is evaluated throughout the observing period, meets the original error budget, and shows no evidence of any secular drift, any statistically significant trend in the observed data over this time can be attributed to changes in atmospheric properties and composition and not to changes in the sensitivity or responsivity of the instrument. This is critical for any long-term trend detection.
The DOE operates a large suite of atmospheric instruments in addition to the AERI at the SGP ARM site. Some of the operational remote sensing instrumentation at the site includes shortwave and longwave spectrometers and radiometers, microwave radiometers, and a Raman lidar. Several studies have been conducted with long-term data from these instruments. Turner et al. (2001) investigated the vertical distribution of water vapor mixing ratio and aerosol extinction as measured by the operational Raman lidar at the site. Long et al. (2009) have examined trends in broadband shortwave irradiance. Michalsky et al. (2010) studied the multiyear trends in aerosol optical depth and its wavelength dependence.
At least one AERI has been deployed at the SGP site since 1995, resulting in more than 14 years of downwelling infrared radiance observations. Between 1996 and 2010, more than 800 000 spectra have been collected. For the current analysis, we have selected the 14 yr of AERI observations in the period between June 1996 and May 2010. The AERI-01 instrument has operated nearly continuously at the SGP site over this entire time span. The instrument suffered a malfunction of the Stirling cycle cooler and was taken offline for a 13-month period between January 2005 and January 2006. Furthermore, due to optical contamination of the scene mirror, as well as anomalous behavior of the Stirling cycle cooler, the AERI-01 spectra exhibited a warm bias between June 2008 and June 2009. Fortuitously, a second AERI instrument, the AERI-E14, was deployed at the SGP site in 2004 and has been operating continuously alongside the AERI-01. Data from the AERI-E14 was used to fill the observation gap in 2005/06 and to replace the anomalous data in 2008/09 so as to create the combined AERI dataset that was used in this study. Although the AERI-01 averages sky radiance for a 3-min period every 8 min to obtain a fully calibrated atmospheric spectrum, the AERI-E14 operates in rapid sample mode, averaging sky radiance spectra for 12 s every 30 s (Turner et al. 2006). Therefore, the AERI-E14 data had to be temporally downgraded to match the AERI-01, by averaging observations over 8-min intervals. The combined dataset spans 14 yr, contains 803 112 spectra, and has observations in every month between June 1996 and May 2010. Figure 2 shows the number of observations in each month of the dataset. The number of monthly observations varies, owing to occasional downtime from routine maintenance and because the instrument only takes valid observations in the absence of precipitation; otherwise, a hatch is closed over the field of view to protect the instrument. Most months contain more than 4000 observations. Only 3 of the 168 months contain less than 2500 observations, and may not accurately capture the synoptic variability for that month; however, in our seasonal analysis presented below, we use 3-month averaging periods, and the synoptic variability in all seasons is adequately sampled. Thus, overall, the time series provides good representation of annual, seasonal, and diurnal variability over the 14-yr period.
Very minimal quality control was performed on the dataset for two main reasons. First, the AERI is a very robust instrument and inherently produces continuous reliable operational data. Second, the very large number of reliable observations dominates statistically over any brief periods of anomalous data. The only data flagged and removed from the raw dataset were single point outliers that were unphysically high or low above the baseline radiance, and scenes with significant variability in the view of the hot blackbody. Overall, more than 98% of the spectra were retained for this analysis from the raw observations.
To gauge the broad characteristics of the dataset, a histogram of the radiance temperature for 985 cm−1 (10.1 μm) is shown by the black curve in Fig. 3. This region of the atmospheric spectrum is fairly transparent in clear skies, containing only weak emission from water vapor. The resulting histogram for the 14 years of AERI observations exhibits a trimodal distribution. It was hypothesized that the three modes correspond to different scene types observed at the SGP site, comprising clear-sky, thin cloud, and thick cloud scenes.
To separate the various scene types a neural network was implemented to distinguish between clear and cloudy scenes based on AERI radiances alone. The observed radiance in selected AERI microwindows, as well as the running mean and standard deviation computed over 70-min intervals, was used as the network input. A “truth” dataset, consisting of 17 months of Raman lidar observations at the SGP site, was used to train the neural network. The network output is bimodal: the majority of AERI observations are unambiguously identified as either clear or cloudy. A small number of observations fall in the intermediate region. A threshold was selected at the minimum point of the network output between the two regimes to demarcate the clear/cloudy classification. Based on an independent testing dataset of Raman lidar observations, the network correctly classifies clear scenes 98% of the time and cloudy scenes 93% of time. Adjusting the threshold has only a minor impact on the overall radiance trend results: changing the threshold within a 30% range results in a maximum trend shift of 0.2% yr−1, though a more typical value is 0.05% yr−1. Further details of the neural network classification method are given in Turner and Gero (2011).
The scenes identified by the neural network as cloudy were further separated into “thin” and “thick” cloud cases. A scatterplot of 985 cm−1 radiance observations clearly shows a clustering of high opacity scenes that have radiance greater than 250 K in radiance temperature [41.5 mW m−2 sr−1 (cm−1)−1]. All such points are classified as thick cloud scenes. The remaining points, which are identified by the neural network as cloudy, are classified as thin cloud scenes. Typical clear-sky, thin cloud, and thick cloud spectra are shown in Fig. 4. The colored curves in Fig. 3 show the resulting radiance temperature histograms for this classification scheme. The thick cloud scenes are separated very clearly into a single mode. The clear-sky and thin cloud scenes exhibit multimodal behavior, which is due to the seasonal variation of the emission temperature, as well as the presence of clear-sky scenes with high water vapor content. The behavior and impacts of this classification scheme, as well as the seasonal distribution of clear sky and cloudiness determined from the AERI observations, are studied in more detail by Turner and Gero (2011).
To simplify the analysis while retaining high information content, 21 microwindows between 520 and 1200 cm−1 were selected from the AERI spectra (Table 1, Fig. 4). The microwindows, which are narrow spectral regions chosen to not include any strong gaseous emission lines, range in width from 2 to 13 cm−1. The microwindows at 531 and 560 cm−1 have relatively strong absorption by water vapor but are semitransparent in clear-sky scenes. These two microwindows are also sensitive to clouds when the water vapor amount is less than approximately 1 cm (Turner 2005); at higher water vapor amounts, these two microwindows become opaque. The 678 and 703 cm−1 microwindows are in a CO2 absorption band and, thus, are fairly opaque and sensitive primarily to the near-surface atmospheric temperature. The microwindows between 810 and 1160 cm−1 lie in the main infrared window region and are mostly transparent in clear skies with weak sensitivity to water vapor, but strong sensitivity to clouds and aerosols. For each microwindow, the mean of the AERI spectral points that lie within the microwindow is calculated and is used in the determination of trends.
Trends in the AERI observations are calculated on a seasonal basis. A seasonal time series is constructed by averaging each AERI observation within a given season for each scene type. The observations are also further subdivided into diurnal components. The trend of the time series is calculated using a linear regression weighted by the ratio of the number of observations divided by their variance, for each season.
Two methods are used to determine the significance of the trends. A two-sample t test is used to test the hypothesis that the observations from the first 5 yr and the last 5 yr of the AERI dataset have equal means to 95% significance. Therefore, data that do not meet this test contain significant mean radiance differences over the observing period. We also used the formulation of Weatherhead et al. (1998) to calculate the trend uncertainty to a 95% confidence level, based on the length of the time series (n), its autocorrelation (φ), and its standard deviation (σN), given by
It is a known caveat that this formulation assumes a univariate autoregressive process of order 1, whereas the climate system is multivariate. Thus, this formulation underestimates correlation time, biasing trends toward larger significance. The assumptions in the formulation, however, become more valid as the time step of the time series increases, and variability on smaller time scales becomes less relevant. Since seasonal (i.e., 3 month) means are used in this study, the Weatherhead et al. formulation can be reasonably applied. Furthermore, by combining the trend estimates from the Weatherhead et al. formulation with the hypothesis testing of the t test, we believe that a credible measure of the significance of a trend in the AERI time series can be produced. In the results presented in Figs. 5–9, the trend error bars (95% confidence level) are given by the trend uncertainty from Eq. (1). Trends are flagged as significant if the absolute value of the trend is greater than its uncertainty and the two-sample t test of the time series indicates unequal means corresponding to the sign of the trend.
3. Results and discussion
The overall radiance trends for the 14-yr AERI time series are shown in Fig. 5 for all observations (all sky) as well as for the component clear-sky, thin cloud, and thick cloud scenes, for each of the microwindows listed in Table 1. Statistically significant trends are identified in bold. The trends in the fraction of clear-sky, thin cloud, and thick cloud scenes are also shown to aid the interpretation of the all-sky results. Overall, small negative radiance trends are observed in most of the microwindows during the all-sky, clear-sky, and thick cloud scenes. In these scenes the AERI data are showing that the downwelling infrared radiance over the SGP has been decreasing significantly over this period. This decrease in downwelling radiance is largely governed by the decrease in the amount of time when thick clouds are overhead and to a lesser extent by the decrease in the radiance within the different cloud classifications. By looking at seasonal and diurnal trends, more prominent trends reveal themselves, which may be countered by each other in the overall time series.
Figure 6 shows the seasonal trends for all-sky scenes as well as the trend in the fraction of clear-sky, thin cloud, and thick cloud scenes. There is a definite trend in all-sky radiance for all seasons. This is primarily attributed to the changes in the fractions of scene types (i.e., cloudiness). In the winter, summer, and autumn (Figs. 6a, 6c, and 6d), the fraction of clear-sky scenes is increasing and the fraction of thick cloud scenes is decreasing; both of these effects lead to a negative trend in all-sky radiance. The situation is reversed in the spring (Fig. 6b), when the fraction of clear-sky scenes is decreasing while the fraction of cloudy scenes is increasing; this leads to a positive all-sky radiance trend. These trend results indicate that there is a definite change in the energy budget of the atmosphere at the SGP site.
Figure 7 shows the trends for clear-sky, thin cloud, and thick cloud scenes separated seasonally. Figure 8 further separates the seasonal results into daytime and nighttime diurnal components. Figure 9 shows the difference between daytime and nighttime trends, separated seasonally. Note that the vertical axes for Figs. 5–9 are all on the same scale.
The most distinct result from these plots is that clear-sky scenes are getting colder (i.e., less downwelling radiance) for all seasons and spectral regions (Fig. 7). Since the downwelling infrared radiance is very sensitive to changes in precipitable water vapor (PWV) (Turner et al. 2004), this almost certainly indicates a decrease in PWV at this site over this period. While the reason for this drying of the atmosphere cannot be determined from AERI data alone, it may be due to a decrease in the evapotranspiration and drying of soils in the past decade (Jung et al. 2010).
Note that the magnitude of the clear-sky trend in the main infrared window region in the wintertime (Fig. 7a) is on the same order as the worst-case 3-σ instrument uncertainty with respect to scene radiance (Fig. 1b). As discussed in section 2, we can show that the instrument uncertainty does not dominate the trend uncertainty, thus the combination of the Weatherhead et al. formulation [Eq. (1)] and the t test is a realistic indicator of the trend uncertainty. For clear-sky trends in other seasons (Figs. 7b, 7c, and 7d), even the worst-case 3-σ instrument uncertainty would not affect the significance of the trends in the infrared window.
The hypothesis of a negative trend in PWV over the SGP site can be tested using observations from the microwave radiometer (MWR). The ARM SGP site has a two-channel MWR that observes the downwelling microwave radiance at 23.8 and 31.4 GHz. The MWR is calibrated by an automated routine that performs “tip curve” calibration scans when the sky is determined by the onboard logic to be cloud free and homogeneous (in a plane-parallel sense); this routine is able to maintain the calibration in these channels to better than 0.3 K rms (Liljegren 2000). The observed brightness temperature data are then used in a physical retrieval , which is able to retrieve PWV with an accuracy of 0.07 cm (Turner et al. 2007). As the MWR is collocated with the AERI at the SGP site (within 200 m), the PWV retrieved from the MWR can be used to investigate the hypothesis that the PWV is decreasing from 1996 to 2010. Unfortunately, the MWR did not operate continuously during this period. Notably, there are large gaps in the observations for winter 1997/98, winter 1999/2000, autumn 2005, and autumn 2007. Nevertheless, the available MWR time series from September 1996 to August 2008 was used to calculate trends in PWV. The observations were temporally matched up with the AERI dataset and sorted into clear-sky, thin cloud, and thick cloud scenes using the AERI-based neural network. The resulting trends are shown in Fig. 10. There is a consistent negative trend in the PWV observed by the MWR in clear-sky scenes for all seasons. This provides supporting evidence to the hypothesis that PWV is decreasing in clear-sky scenes over the SGP site during this 14-yr period.
Both daytime and nighttime trends in the clear-sky downwelling radiance are significant during all seasons (Fig. 8). The diurnal difference in the clear-sky downwelling radiance is statistically significant and negative (daytime drier than at night) for the spring, summer, and autumn seasons (Figs. 9b, 9c, and 9d). This may counter the hypothesis of reduced evapotranspiration because the negative trend in clear-sky diurnal difference (less downwelling radiance during the daytime than at night) implies that the daytime boundary layer is not drying as rapidly as at night, perhaps owing to local influences. Therefore, a more detailed analysis of the water vapor advection into the region over this period will be needed. Unlike the other three seasons, the clear-sky diurnal difference in the wintertime (Fig. 9a) is positive but was not deemed to be significant in the infrared window region (800–1200 cm−1) based on the t test. However, a statistically significant positive trend is identified in the far infrared (531 and 560 cm−1) where the water vapor absorption is stronger.
The microwindows sensitive to the near-surface air temperature (678 and 703 cm−1) show distinct seasonal trends (Fig. 7). The clear-sky scenes have consistent negative trends, which indicate that the temperature of the atmosphere close to the surface is cooling with time. The AERI data show that the lowest layers of the atmosphere are getting colder in all conditions during the winter (Fig. 7a), while the near-surface air in thick cloud scenes is getting warmer in the spring, summer, and autumn (Figs. 7b, 7c, and 7d).
Thick cloud scenes exhibit a distinctly negative trend in the infrared window region (800–1200 cm−1) in the winter and autumn (Figs. 7a and 7d). This could be caused by the thick clouds getting colder, or possibly higher in altitude, over time. In the autumn the thick cloud trends are more negative during the daytime than at night (Figs. 8g, 8h, and 9d). In the spring and summer, this diurnal trend is reversed and the daytime scenes are warmer; however, only the diurnal difference trends (Figs. 9b and 9c) are significant and not the diurnal trends themselves (Figs. 8c, 8d, 8e, and 8f).
The spectral shapes in Fig. 7 may also yield clues about changes in microphysical cloud properties. The smooth curves in the infrared window region for the thick cloud trend (e.g., autumn, Fig. 7d) and thin cloud trend (e.g., spring, Fig. 7b) are characteristic of the shape of the Planck function. The results indicate, however, that more energy is contained in the 1000–1200 cm−1 region than the 800–1000 cm−1 region. This may be a result of cloud particle size changes that can cause the spectral slope to change, effectively pivoting about a spectral point.
There are few significant trends in the thin cloud scenes. A positive trend exists in the spring (Fig. 7b), with the nighttime trends being warmer than the daytime ones (Figs. 8c, 8d, and 9b). A small negative trend exists in the winter (Fig. 7a), with daytime scenes being significantly negative (Fig. 8a), but the diurnal difference is not significant (Fig. 9a). In the autumn, nighttime data (Fig. 8h) have a small negative statistically significant trend and the diurnal differences indicate significantly lower nighttime radiances (Fig. 9d). The interpretation of the thin cloud scenes, however, is more problematic since such a scene could result from one of (i) optically thin stratiform clouds, (ii) clouds that fill the field of view of the instrument for only a fraction of the time of the sky integration period (i.e., for less than 3 min), or (iii) some combination of the two. Contributions from these various effects may be canceling each other in a way as to not produce large distinct trends.
We have analyzed a time series of downwelling emitted spectral infrared radiance from 1996 to 2010 measured by the AERI instrument at the ARM SGP site. The 14-yr time series contains more than 800 000 spectra, each measured with an accuracy better than 1% of the ambient radiance due to a well-designed calibration subsystem and instrument diagnostics. Motivated by the trimodal distribution of radiance temperature for a given frequency of observation, a neural network method was used to classify AERI observations as clear or cloudy, with a further distinction being made between “thin” and “thick” clouds. Weighted linear trends were calculated for each scene type for the overall time series, as well as for seasonal and diurnal components. The statistical significance of the trends was ascertained through the union of hypothesis testing with a two-sample t test and an autocovariate method of trend uncertainty determination.
Significant long-term trends are obtained from the AERI radiance dataset when looking at the data on an annual, seasonal, and diurnal time scales. Many trends have emerged from this dataset after only 14 years of observations, in part due to the low instrumental uncertainty of the AERI. The seasonal all-sky radiance shows statistically significant decreasing trends in the winter, summer, and autumn, with values greater than 1% yr−1 in the winter and autumn, and a trend of increasing downwelling radiance in the spring. These trends in all-sky radiance are primarily caused by changes in the fraction of scene types (i.e., cloudiness) over the SGP site. Overall, there are more clear-sky scenes and fewer thick cloud scenes in the winter, summer, and autumn, thus leading to a negative all-sky radiance trend, whereas the opposite is true in the spring. Furthermore, clear-sky radiance is decreasing in all four seasons, which we hypothesize is due to a decrease in the precipitable water vapor in all seasons. Thick cloud radiance is decreasing in autumn and winter. Thin cloud radiance is increasing in spring and decreasing in winter. Diurnal as well as diurnal difference time series contain further significant trends. The trend spectra reveal changes in cloud characteristics that may be attributed to changes in cloud height, temperature, and particle size.
Further work needs to be done to analyze the trends, test the hypotheses presented here, and attribute physical mechanisms to the observed trends. The AERI dataset can be further studied by including other spectral data, such as in the gaseous absorption bands, and by looking at the data in specific time domains. Attribution can be improved through better separation of cloud types. With the inclusion of additional spectral channels, the neural network could be trained for cloud height detection. Making use of the rapid-sampled AERI-E14 dataset allows processes on shorter time scales (~30 s) to be resolved and improves discrimination for partly cloudy scenes that contribute to the thin cloud category. The trends in the AERI data in different synoptic conditions, as determined using an approach such as Marchand et al. (2009), may also provide insight into the mechanisms that are driving the change in the downwelling infrared radiance. Combining the AERI dataset with other observing instruments at the ARM SGP site can yield insights into specific atmospheric processes. Using observations from AERIs deployed at other field sites, such as the North Slope of Alaska or the tropical western Pacific, will reveal trends in diverse climactic regions.
Given the decadal time span of the dataset, natural variability must be taken into account when drawing broad conclusions. Nonetheless, the high accuracy of these spectral observations and the ability to infer possible mechanisms for any trends from the observations themselves makes this a valuable dataset that only increases in value over time.
Highly accurate decadal-scale observations, such as those from the AERI instruments, can be compared with global and regional climate models to evaluate their performance. A climate model should reproduce the major modes of variability of observed data on annual, seasonal, and diurnal time scales to be credible. Given the geography of the Southern Great Plains and the use of seasonal averages, the AERI observations presented here are representative of a large area, on the order of 100 × 100 km2 (Li et al. 2005), that is comparable to gridbox sizes for global and regional climate models. Thus, the AERI dataset can be a valuable tool in evaluating the performance of climate models on a local scale. Furthermore, this work lays the foundation for the use of global infrared radiance measurements from satellite instruments to ascertain global climate trends and test general circulation models.
The data used in this analysis were obtained from the Atmospheric Radiation Measurement Program (ARM) sponsored by the U.S. Department of Energy, Office of Science, Office of Biological and Environmental Research, Climate and Environmental Sciences Division. This work was supported by NASA Grant NNX08AP44G as part of the CLARREO program and by DOE Grant DE-FG02-06ER64167 as part of the ARM Program. The authors acknowledge contributions and discussions with our colleagues, Drs. Y. Huang, R. Knuteson, S. Leroy, H. Revercomb, and D. Tobin.