Water vapor mixing ratios in the upper troposphere and lower stratosphere measured by the Aura Microwave Limb Sounder (MLS) version 2.2 instrument have been compared with Global Forecast System (GFS) analyses at five levels within the 300–100-hPa layer and North American Mesoscale (NAM) model analyses at six levels within the 300–50-hPa layer over the two years of 2005 and 2006 at four analysis times (e.g., 0000, 0600, 1200, and 1800 UTC). Probability density functions of the vapor mixing ratios suggest that both analyses are often moister than Aura MLS values, but NAM model analyses agree somewhat better with Aura MLS measurements than GFS model analyses over the same North American domain at the five common levels. Examining five subsets of the global GFS domain, the GFS model analysis is moister than Aura MLS estimates everywhere but at 150 and 100 hPa in all regions outside of the tropics. NAM model analysis water vapor mixing ratios exceeded the Aura MLS values at all levels from 250 to 150 hPa in all four seasons of both years and some seasons at 100 and 50 hPa. Moist biases in winter and spring of both years were similar at all levels, but these moist biases in summer and fall were smaller in 2005 than in 2006 at all levels. These differences may be due to the change in the NAM from using the Eta Model to using the Weather Research and Forecasting model (WRF) in June 2006.
Satellites have become an important source of atmospheric data in recent decades. They have been particularly useful in providing information in the upper troposphere and stratosphere, where measurements from other instruments are rare (e.g., Hegglin et al. 2008).
Although the amount of water vapor in the upper troposphere and lower stratosphere (UTLS) is small, water vapor in this region is important to the earth’s climate system. In addition to its role as one of the most important greenhouse gases in the atmosphere (e.g., Raval and Ramanathan 1989; Held and Soden 2000), water vapor in the upper troposphere is important in cirrus cloud formation (e.g., Eguchi and Shiotani 2004). Climate change prediction is strongly dependant on the background water vapor concentration (Forster and Shine 2002). Stratospheric water has two primary sources: oxidation of methane in the upper stratosphere and transport from the troposphere. In addition, water vapor is involved in many photochemical reactions such as its contribution to ozone depletion (Solomon et al. 1986). Research interest in improved understanding of the distribution and transport processes of water vapor in the upper troposphere and lower stratosphere is great.
Comparisons of water vapor observations in the upper troposphere and lower stratosphere with operational center analyses have been rare and generally limited to the European Center for Medium-Range Weather Forecast (ECMWF) analyses. By comparing the ECMWF operational relative humidity analyses with the Measurements of Ozone and Water Vapor by Airbus In-Service Aircraft (MOZAIC) data, Dethof et al. (1999) showed that the ECMWF operational analyses can be used to investigate the transport of moisture from the troposphere into the stratosphere. Dunkerton (1995) concluded from rawinsonde data and ECMWF analyses that the Asian and North American monsoons can transport significant mass into the lower stratosphere. Ovarlez and van Velthoven (1997) and Ovarlez et al. (2000) compared water vapor from ECMWF analyses with aircraft measurements from the Pollution from Aircraft Emissions in the North Atlantic Flight Corridor (POLINAT) experiments taken over a small area over the North Atlantic. They found that the ECMWF analyses in the upper troposphere underestimated the range of upper-tropospheric variations, being moister than the aircraft readings in dry environments and drier in wet environments. Although their comparisons were performed for only four days, discrepancies were attributed to the radiosonde water vapor measurements used in the model not being accurate in the troposphere and not used at all in the stratosphere.
In a similar manner, Oikonomou and O’Neill (2006) used the 40-yr ECMWF Reanalysis (ERA-40) ozone and water vapor reanalysis data to compare with independent satellite data from the Halogen Occultation Experiment (HALOE) and Microwave Limb Sounder (MLS) instruments on board the Upper Atmosphere Research Satellite (UARS) and with data from the MOZAIC program. They showed for water vapor that ERA-40 was drier than HALOE in the upper and middle stratosphere and moister than MOZAIC near the tropopause and upper troposphere. The dry bias in the upper and middle stratosphere was explained by the methane oxidation scheme used in the reanalysis. In a recent study, Luo et al. (2008) indicated an overall dry bias in the ECMWF analyses in comparison with the MOZAIC data, at least before a supersaturation adjustment was implemented in the ECMWF cloud parameterization.
The purpose of the present study is to determine how well the National Centers for Environmental Prediction (NCEP) Global Forecasting System (GFS) and North American Mesoscale (NAM) model analyses of water vapor compare with Aura MLS data in the upper troposphere and lower stratosphere on both global and regional scales. In addition, results from this study have value in diagnosing possible errors in both the GFS and NAM model initializations and can be used in future studies of physical mechanisms for transport of moisture between the upper troposphere and lower stratosphere. The results support the need for assimilating satellite retrievals into models, potentially improving forecasting ability.
2. Data and methodology
Water vapor mixing ratios from two years, 2005 and 2006, during four seasons, denoted in this study as winter [January and February 2005 (JF) and December 2005–February 2006 (DJF)], spring [March–May (MAM)], summer [June–August (JJA)], and fall [September–November (SON)] have been compared between Aura MLS measurements and GFS and NAM analyses. Water vapor volume mixing ratios [parts per million by volume (ppmv)] in the model analyses were computed using temperature and relative humidity data from the following equations:
where qυ is water vapor volume mixing ratio (ppmv), e is the water vapor pressure (hPa), T is the temperature (K), RH is the relative humidity (%), es(T ) is the saturation water vapor pressure (hPa), and p is pressure (hPa).
Comparisons were performed at five levels for both the NAM and GFS analyses: 300, 250, 200, 150, and 100 hPa. An additional comparison was done at 50 hPa for the NAM analyses, because NAM data were available at that level. Aura MLS water vapor volume mixing ratio data were interpolated using a log–log interpolation to these levels.
The GFS includes a medium-range forecast model (MRF) and a global data assimilation system (GDAS). The GFS was developed experimentally (MetEd 2007) during the late 1970s and implemented as the global forecast model at the National Meteorological Center (NMC; now NCEP) on 18 March 1981. During the years evaluated in the present study, the GFS was run four times per day. The analysis scheme used during 2005 and 2006 was a three-dimensional variational data assimilation (3DVAR) scheme referred to as the Spectral Statistical Interpolation (SSI) algorithm (Derber et al. 1991; Parrish and Derber 1992; Derber and Wu 1998). The analysis system integrated all of the observational information (including radiances from several satellites, surface temperatures, radiosonde data, aircraft winds, temperatures, and other observations; Caplan et al. 1997). Above 300 hPa, it used directly only radiances from satellites for analysis variables (e.g., temperature, humidity; J. C. Derber 2009, NCEP, personal communication). The radiance data came from three instruments: the High Resolution Infrared Radiation Sounder (HIRS), the Microwave Sounding Unit (MSU), and the Stratospheric Sounding Unit (SSU). The Joint Center for Satellite Data Assimilation (JCSDA) Community Radiative Transfer Model (CRTM) was incorporated into the SSI to improve radiance assimilation. Derber and Wu (1998) noted that most of the errors in the data were from the ground processing (e.g., cloud clearing, correction to nadir, etc.) and radiative transfer errors. The direct use of radiances in the analysis showed considerable improvement in NCEP’s forecast skill, especially in the Southern Hemisphere. GFS data interpolated to a 1° grid and initialized at the four analysis times (e.g., 0000, 0600, 1200, and 1800 UTC) were used for the comparisons in the present study. The GFS data covered the globe, and comparisons were made in five subregions defined as tropics (TP) restricted by latitudes ranging from 30°S to 30°N, northern midlatitude (NM) with latitudes from 30° to 60°N, northern polar (NP) with latitudes from 60° to 90°N, southern polar (SP) with latitudes from 90° to 60°S, and southern midlatitude (SM) with latitudes from 60° to 30°S (Fig. 1).
The NAM analyses covered North America and nearby ocean regions with 12-km grid spacing. The comparisons between NAM analyses and the Aura MLS over the NAM domain (from 12.19° to 59.5132°N and from 133.459° to 63.9548°W) are performed at four model analysis times (0000, 0600, 1200, and 1800 UTC). The NAM analyses were from the Eta Model (Mesinger et al. 1988; Janjic 1994) during the January 2005 through 20 June 2006 portion of our study period, but the Eta was replaced on 21 June 2006 with the Weather Research and Forecasting model using the Nonhydrostatic Mesoscale Model dynamic core (WRF-NMM; Janjic 2003) and Gridpoint Statistical Interpolation (GSI) analysis. It is important to note that there are some differences between the two models. The new WRF uses hybrid sigma-pressure layers, which replace the step-mountain eta layers in the Eta Model. In the new GSI analysis, two of the main new features are the background-error covariance generation and humidity analysis variable. The background-error covariances, which were previously generated from lagged forecast differences, are now generated from a Monte Carlo method. The humidity analysis variable, which formerly was pseudorelative humidity, is now normalized relative humidity. The domain, grid spacing, and output grid geometry did not change. By comparing data from the three-month summer (JJA) and three-month fall (SON) seasons between 2005 and 2006, differences resulting from the use of the two different models may be identified.
Both the GFS and NAM analyses used data from satellites for operational assimilation at NCEP. The data sources currently used and expected to be implemented in the near future in the assimilation include Geostationary Operational Environmental Satellite, Aqua, and Terra for atmospheric wind vectors; Special Sensor Microwave Imager (SSM/I) surface wind speeds; scatterometers; GPS radio occultation, SSM/I and Tropical Rainfall Measuring Mission (TRMM) precipitation estimates; Solar Backscatter Ultraviolet (SBUV) ozone profiles; and radiances from Advanced Microwave Sounding Unit-A [AMSU-A; National Oceanic and Atmospheric Administration-15 (NOAA-15), NOAA-16, NOAA-18, Meteorological Operation (MetOp), and Earth Observing System (EOS) Aqua], AMSU-B/Microwave Humidity Sensor (MHS; NOAA-15, NOAA-16, NOAA-17, NOAA-18, MetOp) HIRS (NOAA-16, NOAA-17, NOAA-18, MetOp), Atmospheric Infrared Sounder (AIRS; EOS Aqua), GOES sounders (1 × 1–4 detectors, GOES-11, and GOES-12), and imagers [Advanced Very High Resolution Radiometer (AVHRR), GOES, Meteosat, etc.; J. C. Derber 2007, NCEP, personal communication]. Thus, the Aura MLS data are not assimilated in either the NAM or the GFS analysis systems and so the Aura MLS dataset is useful for comparison with GFS and NAM analyses.
The MLS is one of four instruments on the NASA Aura satellite, launched in July 2004 in a sun-synchronous polar orbit. The first version of the MLS dataset was version 1.5 (v1.5; Livesey et al. 2005). The present study uses version 2.2 MLS water vapor mixing ratios in the upper troposphere and stratosphere, even in the presence of cirrus, where observations by other techniques (infrared, visible, and ultraviolet) could be flawed. The v2.2 water vapor mixing ratios were filtered before they were used in the comparisons. The filtering was based on recommended criteria of the data having positive precision values, quality values greater than 0.9, and an even profile status. A positive precision value indicates that the retrieved water vapor is mostly from radiance information. The quality values refer to the goodness of the residual between the measured and calculated radiances. Larger values of quality generally indicate good radiance fits and therefore trustworthy data. The status values are integers indicating circumstances where profiles are not to be used. An odd value of status implies that profiles should never be used in any scientific study. Some nonzero even values of status indicate the retrieval algorithm detected cloud signatures in some radiances (Read et al. 2007; Livesey et al. 2007).
Aura MLS measurement locations for a 24-h period include tangent points for individual limb scans with 200-km along-track separation between adjacent limb scans and 7-km across-track spacing and a vertical resolution of about 1.5–3.5 km from 316 to 4.6 hPa (Livesey et al. 2007). Each satellite data point thus represents an area 200 km long and 7 km wide, areas much different from the GFS and NAM grid boxes. Thus, water vapor mixing ratios in the model analyses were averaged over these same 200 km × 7 km blocks for comparison with the Aura MLS data. However, because the horizontal grid spacing of the GFS and NAM are 1° and 12 km, respectively, the average is performed only in the latitudinal direction, over 2 grid points for GFS and 17 grid points for NAM. In addition, Aura MLS data points (each measurement is taken roughly 25 s apart) were collected within a 6-h period (±3-h window) centered on the model analysis time for the comparison. The ±3-h window used for the Aura MLS observations was chosen to be consistent with the data window used in the model analyses (e.g., Caplan et al. 1997). The differences in spatial resolution between the model analyses and Aura MLS data may contribute to some differences seen in the comparisons. A deeper understanding of these differences is beyond the scope of the present study.
Model analyses were compared with Aura MLS v2.2 data primarily using two techniques, the first being conditional probability density functions (PDFs) of individual water vapor measurements and the second being box-and-whisker diagrams of relative differences. The PDFs show the fraction of the observations that measure a specified value. This powerful tool is ideal for use on large datasets and supplies detailed information on the variability and bias of the data under a wide variety of atmospheric states and geophysical locations (e.g., Sparling 2000). The PDFs are used not only to compare the model analyses with satellite observations but also to compare the analyses themselves. Box-and-whisker diagrams are particularly useful for comparing distributions between several seasons and between models. The diagrams show the spread of a set of data with the upper quartile, lower quartile, and median. The median is indicated by a line dividing the box into two parts. The whiskers are straight lines extending from the ends of the box to the maximum and minimum values, thus showing “outliers,” which may indicate inaccurate data. Because of the large number of plots generated by these comparisons, a supplemental Web site has been established (available online at http://www.meteor.iastate.edu/~lvthien/ISU_MLS.htm) to complement the limited amount of plots discussed here. Also, because differences between the models and Aura MLS observations are similar at the four analysis times, the results discussed here are restricted to 1200 UTC for GFS and 1800 UTC for NAM.
a. GFS analyses compared to Aura MLS observations
1) Tropics (30°S–30°N)
The tropical region is an area with strong upper-tropospheric moistening and the deepest convection on the earth (e.g., Alcala and Dessler 2002). For the tropical region in the present study, the Aura MLS and GFS analysis data points covered both continental (Africa) and oceanic (central Pacific) regions at the 1200 UTC analysis times (Fig. 1) and at 0000 UTC (not shown). At 0600 and 1800 UTC, the data points covered southern America, central Atlantic, southern Asia, and eastern Indian Ocean (not shown). Although the data are at different parts of the globe, the PDFs for this area were similar in shape in all four seasons of both 2005 and 2006 at all analysis times (more details at the supplemental Web site). Results from summer 2005 at 1200 UTC are shown in Fig. 2. The majority of GFS data points were moister than Aura MLS values at all levels from 300 to 150 hPa. This trend is found to be the same at all four analysis times in all seasons in 2005 and 2006 (more details at the supplemental Web site). Box-and-whisker diagrams (Fig. 3) show clearly a moist bias at the tropical tropopause level in winter and spring and throughout the tropical upper troposphere in all seasons. The tropical tropopause layer (TTL) in this region, sometimes also called the tropical transition layer, is usually around 100 hPa. The medians of the differences between GFS analyses and Aura MLS observations are roughly zero at 100 hPa in summer and fall (Fig. 3) so that the GFS analyses do not appear to have a dry or moist bias compared to Aura MLS observations during these seasons. In winter and spring, however, GFS analyses are overall moister than Aura MLS data. Although one might suspect problems in the MLS data around 100 hPa because of the averaging kernel used in this region, where substantial moisture gradients might exist across the tropopause, Read et al. (2007) show that MLS data at 100 hPa agreed well with balloon data and that there was almost no contribution to the 100-hPa value from the 147-hPa level, where conditions would likely be more moist. The air at this region is driest in the winter and wettest in the summer (e.g., SPARC 2000), and a moist bias during winter and spring may indicate problems in the GFS assimilation of observations that clearly depict seasonal variations or a lack of such observations.
In general, at all levels below the TTL in all seasons, the GFS is moister than Aura MLS, and the mean values of GFS mixing ratios are greater than those of Aura MLS. It is worth noting that over these regions a maximum in convection is consistent with a maximum in upper-tropospheric moisture. The deep convection over these regions may lead in the GFS model to creating excessive moisture in the analyses in the tropical upper troposphere. In addition, Spichtinger et al. (2003) used UARS MLS data, which showed that ice supersaturation occurred most frequently over these tropical regions in the upper troposphere, and the ice supersaturated regions were colder and moister than other nearby regions. The excessive supersaturation in the analysis associated with the moist bias may be related to errors in the estimate of humidity background errors, which in practice are not known and must be modeled (Dee and da Silva 2003).
2) Northern midlatitudes (30°–60°N)
For northern midlatitudes, the Aura MLS and GFS analysis data points mainly covered Europe and the North Pacific Ocean at the 1200 UTC analysis time (Fig. 1) and at 0000 UTC. At 0600 and 1800 UTC, the data covered much of the northern Atlantic Ocean and Asia (not shown). Despite differences in the regions of coverage, the PDFs comparing GFS analyses and MLS data were similar at all four analysis times (for more details, see supplemental Web site). At 1200 UTC, the GFS PDFs for this region have the modes at higher water vapor values than Aura MLS PDFs at levels from 300 to 200 hPa throughout all seasons. On the other hand, at 150 and 100 hPa the GFS PDFs have the modes at lower water vapor magnitudes than Aura MLS PDFs at all seasons except at 150 hPa in JF and MAM 2005. This can be seen in the PDF comparison for summer 2006 (Fig. 4). Also, the mean values found in this figure are greater for GFS than for Aura MLS at all levels except at 100 hPa. Figure 5 shows box-and-whisker diagrams of relative differences between GFS and Aura MLS values, which allows for comparison of different seasons both in magnitudes and percentages. The moist bias percentage at 150 hPa was higher during the winter and spring of 2005 compared to 2006. A GFS dry bias existed at 100 hPa in all seasons. This pattern of biases relative to height remains relatively the same in all four seasons, despite the fact that the average tropopause height varies over these seasons in the region. The reversal from general moist biases in the upper troposphere to dry biases in the lower stratosphere could be a result of cold biases in the model that enhance the dehydration so that GFS analyses are too dry near and above the tropopause. Verification data from NCEP (available online at http://www.emc.ncep.noaa.gov/gmb/ssaha/) did show a cold bias in the GFS analyses over all regions in parts of the stratosphere during 2005 and 2006 compared to rawinsonde observations, although it was most pronounced and persistent around 50 hPa and only present at 100–150 hPa during winter and spring 2005. It should be noted that the tropopause level in this region is lower than in the tropics. The dry bias in the midlatitude stratosphere, where the water vapor concentrations are controlled by horizontal transport and downwelling, is consistent with the idea that too vigorous convection in the tropics could lead to enhanced dehydration of the tropical stratosphere.
3) Northern pole (60°–90°N)
The number of Aura MLS data points in the northern polar region was smaller than in the tropics and north and south midlatitudes, and most of the data points lay within the Western Hemisphere (Fig. 1) at 1200 and 0000 UTC. At 0600 and 1800 UTC, data covered most of Greenland and the Arctic Ocean (not shown); however, despite a shift in location, PDFs at the four analysis times did not differ substantially. The PDF curves in the northern polar region for the GFS analyses and Aura MLS data strongly resemble those for the northern midlatitudes. The PDFs and the mean values from fall 2005 of GFS analyses are greater than those of Aura MLS, except at 100 and 150 hPa (not shown). As in the northern midlatitudes, the pattern of biases does not seem to be affected much by the changing of seasons and average height of the tropopause. However, in this region the differences are greater, especially between the modes. The pronounced dry bias and moist bias are larger than in the subpolar regions. A moist bias is found at 300, 250, and 200 hPa. In addition, a dry bias is also found at both 150 and 100 hPa in both years, with extremely large percentage differences found in spring, summer, and fall 2006 and summer and fall 2005 (not shown).
4) Southern pole (60°–90°S)
In general, PDFs for the southern polar region (not shown) were similar to what was found in the northern polar region, with the exception that a moist bias was present at 150 hPa in the GFS data instead of a dry bias in JJA. This reversal of biases at 150 hPa between the two polar regions was true only in JJA. The cause of this reversal is not obvious, but it may be related to strong dehydration in the winter in the southern polar region (SPARC 2000) observed by MLS. GFS may lack the ability to correctly depict the dehydration. This effect is only pronounced in winter in the Southern Hemisphere. Thus, MLS data appear to successfully capture this feature as other observations have.
5) Southern midlatitudes (30°–60°S)
The data used for southern midlatitudes at 0000 and 1200 UTC covered regions near southern Africa and the central Pacific. At 0600 and 1800 UTC, the data partly covered southern South America, Australia, and the Southern Ocean. Despite differences in the regions of coverage, the PDFs comparing GFS analyses and MLS data were similar at all four analysis times (more details at supplemental Web site). The shapes of the PDFs for the Aura MLS data and GFS analyses for the southern midlatitudes were similar to those for the northern midlatitudes at all seasons and levels (not shown). In particular, in the southern midlatitudes, a dry bias was found to be larger in summer and fall than in winter and spring at 100 hPa in the two years, and the moist bias was smaller at 300 and 150 hPa than at other levels.
b. NAM analyses compared to Aura MLS observations
The area over which the Aura MLS observations and NAM analysis data points could be compared is mainly over North America at all four analysis times (0000, 0600, 1200, and 1800 UTC). Figure 6 shows the shapes of the NAM PDFs and Aura MLS for fall 2005 at 1800 UTC. In general, the PDFs of Aura MLS and NAM agree very well with each other, especially at 300 hPa. The vertical dashed lines show that the mean values of NAM are closer to those of Aura MLS at 300 hPa. At other levels, the mean values of NAM are greater than those of Aura MLS. Results for other times and seasons can be found at the supplemental Web site. The PDFs valid for the NAM analyses in the summer seasons of 2005 and 2006 were most consistent with the Aura MLS data at 300 hPa in both years. Like in the fall, at all levels, the mixing ratios of greatest frequency in the PDFs for the NAM and Aura MLS were closer in 2005 than in 2006 (not shown), and the mean values for NAM are greater than those for Aura MLS at all levels. In the winter, the PDFs of Aura MLS and NAM agree very well with each other. The vertical dashed lines show that the mean values of NAM are greater than those of Aura MLS. Although the PDFs of the two datasets show similar shapes, the NAM PDFs have the mode at a higher water vapor mixing ratio than the Aura MLS PDFs. As in winter, the PDFs of Aura MLS and NAM agree very well with each other, especially at 300 hPa during spring. The PDFs in the two years are similar, with the NAM PDFs having the mode at higher values at all levels (not shown). Also, the mean values in the NAM analyses are greater than those of Aura MLS at all levels.
Here, a comparison of the fall and summer seasons is important, because the NAM changed from using the Eta Model to using the WRF model on 20 June 2006, and a fall and summer comparison thus allows one to see what impact the change in model may have had in representing water vapor at these high levels. During these seasons, the mixing ratios of greatest frequency in the NAM PDFs were more consistent with those from the Aura MLS PDFs in 2005 than in 2006 at all four analysis times. Figure 7 compares the NAM and MLS data over the full two-year period at 1800 UTC by using box-and-whisker diagrams. A moist bias typically exists at almost all levels but at 100 hPa in winter 2005 and 2006, spring 2006, and fall 2005 and at 50 hPa in winter 2005 and 2006. Although there are similar moist bias percentages in both winters and springs during 2005 and 2006, bigger differences are noted between summers and falls of the 2 yr, with smaller moist bias percentages in 2005 than in 2006. The same results are found at other analysis times in the supplemental Web site. This result suggests a negative impact due to the change in the NAM from using the Eta to using the WRF.
c. NAM and GFS analyses compared with Aura MLS observations over the NAM’s domain
Figure 8 shows the PDFs for the Aura MLS, NAM, and GFS analyses in the NAM’s domain at 1800 UTC in fall 2005. Compared to the GFS, the NAM PDFs agreed better with the Aura MLS PDFs at all levels over this domain. The vertical dashed lines also showed that the mean values of the NAM analyses were closer to those of MLS than the GFS analyses. Also, Fig. 9 shows box-and-whisker diagrams of differences between NAM and GFS analyses and Aura MLS water vapor mixing ratios at 1800 UTC in summer 2005. Although both NAM and GFS are wetter than Aura MLS at levels from 300 to 150 hPa and at 100 hPa in NAM, the differences in NAM are overall smaller than in GFS at all levels. The PDFs for the Aura MLS, NAM, and GFS analyses in the NAM’s domain in 2005 and 2006 at all four analysis times are shown in detail in the supplemental Web site. Overall, the PDFs for the NAM analysis matched the Aura MLS PDFs better than the GFS did. Both the NAM and GFS PDFs had modes at higher water vapor values and had broader shapes than the PDFs for the Aura MLS observations at all levels between 300 and 200 hPa. In contrast, at 100 hPa the NAM PDF peaks look about the same as those of the MLS data in fall 2005 and higher in fall 2006, whereas the GFS PDFs peaked at a lower water vapor value compared to the Aura MLS PDFs in both of these two seasons. Also, although the GFS PDFs for 2005 and 2006 stayed basically the same, the NAM PDFs changed between these two years, being more similar to Aura MLS PDFs in 2005 than in 2006 at all levels. The box-and-whisker diagrams of differences between NAM and GFS model analyses and Aura MLS observations in the two years are also shown in the supplemental Web site. Both the NAM and GFS analyses over the smaller NAM domain have the same moist bias tendency at all levels but at 100 hPa, where the GFS has a drier bias in 2005 and 2006 compared to Aura MLS and the NAM has a moist bias in 2005 and 2006.
These discrepancies could be related to several key differences between the two models. Among the most important is that GFS used SSI and NAM used GSI in their assimilation systems during 2005 and 2006. The improvements in GSI over SSI include incremental noise reduction and a balanced analysis increment improvement, which had an immediate impact on the quality of the short-term forecasts that are used in the analyses (e.g., Kleist et al. 2009). GFS used a cloud-top temperature below −15°C to set a threshold for ice saturation, whereas NAM set the temperature to −30°C. In addition, the vertical resolution of the NAM analyses (25 hPa) is higher than in the GFS (50 hPa). In the models themselves, different convective schemes were used along with different horizontal resolutions, which could affect the analyses through the use of forecasts as backgrounds or first guesses in the assimilation systems. The higher horizontal resolution in the NAM analyses may result in a more realistic depiction of the moisture fields in the UTLS than that with the lower horizontal resolution in the GFS analyses. A deeper understanding of the sensitivity of horizontal resolution within GFS and NAM analyses to the water vapor in the UTLS may help resolve these differences. We leave this work to a future study.
Comparisons between GFS and NAM water vapor analyses and measurements from Aura MLS have been performed at all four model analysis times (0000, 0600, 1200, and 1800 UTC) in the 2 yr of 2005 and 2006 and show some substantial differences between the analyses and the MLS data. Some possible explanations for the differences are offered later. In all regions in the upper troposphere, the GFS analyses are moister than MLS observations in all seasons and at all levels. In the stratosphere outside of the tropics, the GFS is drier than MLS. The reason for this switch could be due to a cold bias near the tropopause/lower stratosphere, which leads to enhanced dehydration in the stratosphere. Such a cold bias could occur if the model’s convective scheme is too vigorous, for instance, with the tropopause height being too high. Moreover, water vapor mixing ratios from MLS on the Aura satellite itself also showed a dry bias in the upper troposphere compared with mixing ratios estimated from the AIRS on the Aqua satellite and frost point sondes (e.g., Fetzer et al. 2008; Read et al. 2007) in the upper troposphere. A dry bias in the MLS observations would at least partially explain why the GFS analyses would be moister at these levels. Near the tropical tropopause (100 hPa), there is also some seasonality to the bias. The seasonal bias might reflect a deficiency in the GFS assimilation of observations that clearly show seasonal variations in moisture. In polar regions, a very pronounced dry bias is found in the stratosphere, which again might be due to enhanced dehydration caused by too low temperatures.
NAM model analysis water vapor mixing ratios exceeded the Aura MLS values at all levels from 250 to 150 hPa in all four seasons of both years and in some seasons at 100 and 50 hPa. It is worth noting here again that a dry bias in the MLS measurements was found over this area compared with AIRS and frost point sondes at these levels, and this dry bias might explain why NAM analyses would be moister than the MLS values. Moist biases in winter and spring of both years were similar at all levels, but these moist biases in summer and fall were smaller in 2005 than in 2006. These differences may be due to the change in the NAM from using the Eta Model to using the WRF model in June 2006. One important change made in the assimilation system at this time was in the humidity analysis variable, which was pseudorelative humidity in the Eta Model but became normalized relative humidity.
Because mixing ratio (and specific humidity) exhibit extreme variability and changes in the scale of errors and in the fields themselves, the use of these variables causes difficulties in assimilation systems resulting in large extrapolation errors (e.g., Dee and da Silva 2003). Relative humidity is spatially and temporally more coherent, such that error statistics are easier to obtain. But, the use of relative humidity can lead to unrealistic and unstable stratospheric accumulation of moisture when model temperatures are biased. Pseudorelative humidity and normalized relative humidity both solve some of these problems. For pseudorelative humidity, the mixing ratio is scaled by background saturation mixing ratio, and this scaling effectively is a flow-dependant transformation of the observed mixing ratio. It has similar statistical properties to relative humidity but preserves specific humidity in the absence of moisture observations. Normalized relative humidity is another representation of the moisture content that avoids the problems when mixing ratio or specific humidity is used during assimilation. It is a statistically normalized version of the relative humidity and gives background-error statistics that are more homogeneous and Gaussian than for specific humidity. It also effectively eliminates the possibility of supersaturation and negative humidity being generated by the analysis and is multivariated related with temperature and pressure (available online at http://www.wmo.int/pages/prog/www/DPFS/ProgressReports/2005/UnitedStates.pdf). The main difference between the two variables, pseudorelative humidity and normalized relative humidity, is that, in the GSI assimilation system, the relative humidity control variable can only change via changes in specific humidity when pseudorelative humidity is used. With normalized relative humidity, humidity can change via changes to surface pressure, temperature, or specific humidity.
In a comparison between NAM and GFS analyses in North America, the NAM results compared well with MLS and better than the values from the GFS. The better agreement could be related to several key differences between the two models. The GFS used SSI and the NAM used GSI in their assimilation systems during 2005 and 2006. Kleist et al. (2009) found that several changes in GSI over SSI had an immediate impact on the quality of the short-term forecasts used in the analyses. The two models also differed in their horizontal resolutions and the vertical resolution of the output datasets, with GFS being coarser than NAM and thus less able to resolve strong temperature gradients. Other differences were present in their microphysics and convective parameterizations, which could affect the analyses through the use of forecasts as backgrounds or first guesses in the assimilation systems.
5. Summary and conclusions
We have presented the first comparisons between Aura MLS satellite-based water vapor measurements and GFS and NAM model analyses in the upper troposphere and the lower stratosphere. The GFS analyses generally agreed better with satellite observations in the tropics and both northern midlatitudes and southern midlatitudes than in both the northern and southern poles with regard to the magnitude and to all seasonal distributions.
NAM water vapor analyses were generally more consistent with Aura MLS measurements at all levels than GFS analyses. Of note, the NAM water vapor analyses in two seasons, summer (JJA) and fall (SON) 2005, agreed better with the Aura MLS data than in the same two seasons in 2006. The poorer performance may be related to the change in the NAM from using the Eta Model to using the WRF model in June 2006. In particular, this change included a change in the vertical coordinate from using step-mountain eta layers to using hybrid sigma-pressure layers, a change in the background-error covariance calculations from a lagged forecast differences method to a Monte Carlo method and a change in the humidity analysis variable from pseudorelative humidity to normalized relative humidity. The analysis in an atmospheric data assimilation system is constructed by combining a model-generated background estimate with observations. The analysis will extrapolate information from the observations into the analysis. Regarding moisture in the analyses, Rabier et al. (1998) noted that the inaccurate extrapolation of information from upper-tropospheric observations can contribute a small but significant accumulation of excess water vapor in the lower stratosphere. It is unknown if the differences in the analyses at these levels have affected the accuracy of the NAM forecasts overall. Also in the fall seasons of both 2005 and 2006, the NAM analyses agreed better with the Aura MLS data than in the summer seasons. The moisture differences between the NAM analyses and Aura MLS observations are similar in winter and spring in the two years.
Overall, a moist bias was found in all four seasons at all six levels evaluated in the GFS analyses for the tropics and in the NAM analyses for the northern American domain but with less severity in some seasons at 300 and 100 hPa. In other regions within the GFS analyses, the moist bias was present in all four seasons at 300, 250, and 200 hPa, with a dry bias at 150 and 100 hPa.
Comparisons between GFS and NAM analyses and Aura MLS satellite data may help diagnose possible errors in the model initializations or deficiencies in the algorithms applied to the satellite data. In addition, the differences found in the mixing ratios between the analyses and MLS observations also support the need for assimilating satellite retrievals into models to potentially improve forecasting ability.
The authors thank Eric A. Aligo for his assistance with the computational work. Comments from two anonymous reviewers substantially improved the paper. This research was funded by the National Aeronautics and Space Administration (NASA) under Grant NNX06AH91G.
* Current affiliation: University of Hawaii at Manoa, Honolulu, Hawaii.
Corresponding author address: Le Van Thien, University of Hawaii at Manoa, 2525 Correa Rd., HIG 350, Honolulu, HI 96822. Email: firstname.lastname@example.org