The dust top height and infrared optical depth over land are retrieved from the Atmospheric Infrared Sounder (AIRS) longwave infrared measurements by using a one-dimensional-variation retrieval algorithm for different Asian dust storms. By combining particle size measurements from a 10-year ground observation data set from the Dunhuang Skynet station located to the east of the Taklimakan Desert in China and the Optical Properties of Aerosols and Clouds data set of optical properties, the mineral dust scattering and absorption coefficients are obtained and then used to compute brightness temperatures with RTTOV 9.3. The retrieved dust thermal infrared optical depths are compared with the Ozone Monitoring Instrument and Moderate Resolution Imaging Spectroradiometer (MODIS) products. The retrieved dust top heights are compared against the extinction backscatter profiles obtained from the Cloud-Aerosol Lidar with Orthogonal Polarization lidar. Infrared optical depths from AIRS correlate favorably with visible optical depths from MODIS, and dust top heights agree reasonably with lidar observations for the single-layered dust storms over the Taklimakan Desert.
 Dust storms occur more frequently now due to climatic variability and land use change [Jickells et al., 2005; Lee and Sohn, 2011]. Asian dust generated mostly in the springtime in northwest China, one of the major sources of large-scale dust storms, is transported frequently toward the Pacific Ocean. Each year, thousands of tons of dust from the Taklimakan Desert are blown into the Pacific Ocean. In recent years, researchers have found that Asian dust storms could affect the radiation budget through the regulation of cirrus cloud formation and the marine ecosystem through nutrient supply [Eguchi et al., 2009; Uno et al., 2009]. Furthermore, the springtime aerosol load in the United States Pacific Northwest is related to Asian dust emissions [Fischer et al., 2009].
 High spectral resolution remote sensing provides the capability to detect dust storms and retrieve their heights and optical depths [Sokolik, 2002; Pierangelo et al., 2004; De Souza-Machado et al., 2006]. Sokolik  reported that the radiative signatures of dust are significantly different from those of clouds and the greenhouse gases in the atmosphere. De Souza-Machado et al.  investigated infrared spectral signatures of dust over ocean using the Atmospheric Infrared Sounder (AIRS) infrared measurements. Pierangelo et al.  developed an approach based on precomputed lookup tables to retrieve dust altitude and infrared optical depth over the Atlantic Ocean. The dust optical depth and the dust height in the Mediterranean are also retrieved by using the Thermal Infrared (TIR) observations from the AIRS [De Souza-Machado et al., 2010].
 Although great efforts have been devoted to extracting dust properties from high spectral resolution infrared measurements, certainly more work is needed on the dust retrieval from these data, especially over land [Peyridieu et al., 2010; De Souza-Machado et al., 2010]. Dust parameter retrievals over land are difficult mainly due to poor knowledge of the surface emissivity and inaccurate surface temperature as well as less accurate temperature of the underlying atmosphere. The retrievals may be complicated further by the fact that the single scattering properties of dust originating from different areas differ primarily due to different dust composition [Su and Toon, 2010].
 This study focuses on retrieving the altitude and thermal optical depth of Asian dust over land by using the TIR hyperspectral sounding data from AIRS. The mineral dust optical properties derived from a 10-year ground size distribution observation data set from the Skynet station [Kim et al., 2004] in Dunhuang (40.146°N 90.799°E), which is located to the east of the Taklimakan, are used in the radiative transfer model [Han et al., 2012]. A one-dimensional variation retrieval (1DVAR) algorithm is developed to extract dust parameters by using RTTOV 9.3 (Radiative Transfer for (A)TOVS) forward model. The retrieved dust TIR optical depths are compared with the Ozone Monitoring Instrument (OMI) aerosol index [Torres et al., 2007] and the Moderate Resolution Imaging Spectroradiometer (MODIS) Deep Blue product [Hsu et al., 2004]. The retrieved dust top heights are compared against the extinction backscatter profiles obtained from the Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) lidar [Winker et al., 2007].
 This paper is organized as follows: section 2 introduces the data used in this study, section 3 outlines the dust height and infrared optical depth retrieval method, section 4 investigates the accuracy of the dust property retrieval in different assumed conditions, and section 5presents the detailed analysis of dust top heights and optical depths retrieved from the AIRS measurements by comparison with other A-Train instruments for four different Asian dust storms. Finally, conclusions are presented insection 6.
 The A-Train constellation of satellites provides multi sensor observations to study cloud and aerosol properties, as well as weather and atmospheric chemistry [Stephens et al., 2002]. The measurements are made over the same geographic locations and separated by less than 15 min. The instruments include MODIS and AIRS on Aqua, OMI on AURA, and CALIOP on the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO). Since the satellites in this constellation fly closely to each other, the dust retrievals from AIRS can be easily collocated and compared to the products from the OMI, MODIS, and CALIOP in the study.
 The AIRS instrument [Chahine et al., 2006], onboard the Earth Observing System (EOS) Aqua satellite, was launched on 4 May 2002 and is the first of a new generation of high spectral resolution IR sounders. It has 2378 channels measuring outgoing radiances in three bands: 8.8–15.4, 6.2–8.2, 3.75–4.58 μm at a spatial resolution of 13.5 km at nadir. AIRS was primary designed to retrieve atmospheric temperature, water vapor, trace gases, and surface skin temperature with a 2000 km swath for weather prediction and climate studies [Chahine et al., 2006]. Its measurements also contain information about dust properties [Ackerman et al., 2004; Pierangelo et al., 2004; De Souza-Machado et al., 2006]. Additionally, it can be used to retrieve cloud microphysical properties by taking full advantage of high spectral resolution AIRS longwave cloud-sensitive radiance measurements [Li et al., 2005; Weisz et al., 2007].
 The OMI on the EOS Aura platform is a hyperspectral push-broom sensor measuring solar reflected and backscattered light in two spectral bands: the UV channel (270–350 nm) and the VIS channel (350–500 nm) [Levelt et al., 2006]. This instrument has a 2600 km wide viewing swath such that it provides daily global coverage at a spatial resolution 13 × 24 km at nadir.
 Arid and semiarid regions are problematic in retrievals using visible and near-IR measurements, as these surfaces have high reflectance in these spectral bands. However, low surface albedo in the UV spectral region for all land and water surfaces means that OMI can be used to retrieve aerosol properties [Torres et al., 2007]. There are two aerosol inversion algorithms applied to the OMI measurements: the OMI near-UV (OMAERUV) algorithm and the multiwavelength algorithm (OMAERO). The two algorithms make use of the sensitivity of UV measurements to aerosol absorption [Torres et al., 1998, 2007]. The OMAERUV is used to derive aerosol extinction and absorption optical depth with two UV wavelengths. This algorithm produces the Aerosol Index, and aerosol optical depth at 388 nm. OMI aerosol retrievals have been validated against AERONET observations [Torres et al., 2007; Curier et al., 2008; Brinksma et al., 2008]. The UV-absorbing aerosols, such as desert dust carbonaceous and volcanic ash, yield positive aerosol index values. The OMI aerosol index is also sensitive to the aerosol height.
 In this study, the OMI aerosol index-UV is used to evaluate the dust AOD (aerosol optical depth) retrievals from AIRS over land. The OMI aerosol index is calculated as a measure of the difference between the measured backscattered UV radiance from an atmosphere containing aerosols and that of a pure Rayleigh scattering atmosphere [Torres et al., 2007].
 The MODIS instrument on board the Aqua satellite in the A-Train has high spatial resolution and a near-daily global coverage. It provides observations of the Earth's surface and atmosphere using 36 channels centered in the visible (VIS), near-infrared (NIR), and infrared (IR) regions of the spectrum from 0.4 to 14.5 μm. The MODIS retrieval of AOD over land employs primarily three spectral channels (0.47, 0.66, and 2.1 μm) [Kaufman et al., 1997a]. However, the standard MODIS AOD product cannot provide information about aerosols over bright surfaces such as deserts because it is retrieved using NIR wavelengths (2.1 and 3.8 μm) [Kaufman et al., 1997b]. In this paper, the MODIS Deep Blue (DB) product (10 km resolution) is used to evaluate the dust AOD from AIRS. The DB algorithm has larger sensitivity to aerosols over bright surfaces because it primarily employs two blue channels (0.412 and 0.470 μm), for which surface reflectances are much darker, to infer aerosol properties [Hsu et al., 2004]. It is reported that the uncertainties of the MODIS DB product are around 20–30% by comparison with observations from AERONET sites in China and Mongolia [Hsu et al., 2006]. Li et al. use the ground-based remote sensing of AOD from Sun photometers at four sites in Xinjiang during the years 2002–2003 to validate aerosol products, including DB of MODIS. The results show that the DB algorithm is able to retrieve AOD at Tazhong in the Taklimakan desert, where the correlation coefficient between the DB product and ground-based Sun photometer measurements reaches 0.94.
 Operating at optical wavelengths, CALIOP is a space-based lidar on board the CALIPSO satellite platform, and provides unique measurements of the global vertical distributions of clouds and aerosols, including dust [Winker et al., 2007]. CALIOP transmits and receives backscattered light at laser wavelengths of 532 nm and 1064 nm. It has a typical horizontal averaging interval of 5 km for aerosols. The observations from CALIOP about 45 s behind AIRS will be used to validate the AIRS dust height retrievals. More information on CALIPSO instruments and products is available at http://www-calipso.larc.nasa.gov/.
3. Physical Basis and Method
3.1. Physical Basis
 The TIR observations could be used to detect dust over desert both day and night [Wald et al., 1998; Sokolik, 2002]. The impacts of the aerosol optical depth, aerosol layer altitude, surface emissivity, and size distribution on the AIRS observations were also investigated by Pierangelo et al. . Their study demonstrated that the longwave IR channels are very sensitive to the altitude and aerosol optical depth. As a result, it is possible to simultaneously retrieve the AOD and altitude of a dust layer using AIRS measurements simultaneously.
3.2. Forward Model
 In the physical retrieval, a forward model is used to model satellite observations as accurately as possible. The forward model F(x) relates the atmospheric state vector x to the brightness temperature measurement vector Y. F, therefore, has the same dimension as Y. In this study, the dust optical properties used are calculated with the Optical Properties of Aerosols and Clouds (OPAC) mineral refractive index (http://www.lrz.de/∼uh234an/www/radaer/opac.html). The particle size distributions are obtained by taking an average of all available dust observations at Dunhuang. Then we combine three modes for mineral dust components (nucleus, accumulation, and coarse modes) from OPAC database, and the number mixing ratios for the three modes are 0.862, 0.136, and 0.002, respectively. Considering the number size distribution, the mixing ratio of nucleus mode is higher than the other modes. However, for the volumetric size distribution considered, the coarse mode is dominant. It is a typical pattern of dust size distributions. For example, desert aerosol type of OPAC has a very low number mixing ratio of the coarse mode.
 Given the atmospheric parameter profiles, surface temperature, surface emissivity, and dust parameters, the AIRS observation vector has the form
where the individual elements are calculated by RTTOV 9.3, which is a development of the fast radiative transfer model originally developed at European Centre for Medium-Range Weather Forecasts (ECMWF) in the early 1990s [Eyre, 1991] for TOVS. The model allows rapid simulations of radiances for satellite infrared or microwave nadir scanning radiometers given atmospheric parameter profiles. This model, which includes a multiple scattering code, can be used to simulate radiances at infrared wavelengths in the presence of aerosols. Han et al.  has used this model to compare the simulated measurements to the AIRS observations in cases of dusty skies.
3.3. Surface Emissivity
 Dust retrieval over continental surfaces, such as the Taklimakan Desert, is possible because of the availability of surface emissivity data sets from MODIS, AIRS, and Infrared Atmospheric Sounding Interferometer (IASI) observations. These monthly mean data sets are derived from the clear sky observations, although they may be somewhat contaminated by dust and cloudy pixels which are misidentified as clear. This information paves the way for retrieving dust properties over land with infrared measurements.
Seemann et al.  derived a high spatial and moderate spectral resolution global database of Land Surface Emissivity (LSE) by using a procedure called the baseline fit (BLF) method. On the basis of an analysis of MODIS/UCSB and ASTER/JPL libraries, they extracted the most representative spectra (about 300) and identified 10 wavelengths (“hinge points”) situated within strategic regions of the emissivity spectrum. According to a conceptual model of LSE, the method adjusts a baseline emissivity spectrum based on the operational MODIS product MYD11 from Aqua or MOD11 from Terra. This approach captures the most important features of the emissivity spectrum. In their study, the impact of LSEs on atmospheric total precipitable water (TPW) retrievals with MODIS radiance measurements was evaluated over land. Substantial improvement was shown over traditional retrievals conducted with the typical assumption of constant LSEs. The surprisingly better results by using the MODIS products in the 11–12 μm range are reported by Li et al. , although the method could not provide the LSE quantitative accuracy due to the lack of the true emissivity. The original data set provides monthly averaged global IR LSE spectra from 3.7 to 14.3 μm at a 0.05° × 0.05° spatial resolution based on the analysis of MODIS data. To make this data set closer to AIRS observations in their spatial resolution, the data are degraded to a 0.25 × 0.25 degree latitude-longitude grid scale in the following study.
3.4. Channel Selection
 It is well known that the physical retrievals strongly depend on the accuracy of the forward model for selected channels. Significant discrepancies in the brightness temperatures could be caused by the differences in dust composition, which could strongly depend on geographic source regions [Claquin et al., 1999; Sokolik, 2002; Su and Toon, 2010]. As a result, the channels should be carefully selected after detailed radiative transfer simulations using the available dust composition in a particular area.
Han et al.  calculated the aerosol scattering and absorption coefficients by using the mineral aerosol size distribution data measured at the Dunhuang dust observation station located at 94.8°N and 40.1°E from October 1998 to January 2007 and the OPAC data set of optical properties. Then, these coefficients and the CALIPSO dust observations in 2008 are used to simulate the AIRS observations with RTTOV 9.3. Their results show that the discrepancies between the simulated and observed brightness temperatures are less than 0.5 K in the 10.2–11.5 μm band when the dust effect is taken into account. Consequently, only eight thermal channels in this band as shown in Table 1 are used to retrieve the dust top height and infrared optical depth. The sensitivity of this band to the dust has been extensively investigated by other studies [Pierangelo et al., 2004; De Souza-Machado et al., 2006; Peyridieu et al., 2010]. Channels 672 and 914 have been used by Peyridieu et al. . Although the shortwave infrared channels could provide more information on the dust optical depths, they could be contaminated by the effects of surface reflectance of downwelling solar radiance, degrading the retrievals during the daytime.
Table 1. AIRS Selected Channels Number, Wave Number, Wavelength, Surface Transmittance Using the Tropical Atmospheric Parameters in RTTOV 9.3
Wave Number (cm−1)
3.5. Retrieval Algorithm
 The general idea of the 1DVAR theory is to minimize a penalty function J(x), which measures the degree of fit of the dust measurements to the background information and possibly to other physical constraints. The 1DVAR or Optimal Estimation Method (OEM) tries to minimize the following penalty function [Rodgers, 1976]:
where xa is the background information vector of x, Sa is the assumed background error covariance matrix that constrains the solution, and Sεis the observation error covariance matrix, which includes instrument noise plus the assumed forward model error. To numerically solve the equation, a quasi-Newtonian iteration is used to obtain [Eyre, 1989]:
where , , and represents the linear or tangent model of the forward model F, which is used to calculate the AIRS observations from the inputs including the atmospheric parameter profiles, surface parameters, and dust properties. In the actual retrieval process, the iteration is stopped when the δYn is less than 0.5 K or the δYn difference from two successive iterations is less than 0.005 K.
 The variance Sa approaching zero implies that the constraint is a very good estimate of the true parameters and the solution approaches the constraint xa. Conversely, as Sa is approaching infinity, the constraint is a very poor estimate of the true structure and the solution tends toward that specified by the exact inversion of the forward model.
 Although the observations are sensitive to atmospheric and surface state, dust height, and dust amount as well as its composition, the dust retrieval will be unstable if all parameters are varied simultaneously. Consequently, in this study, the vector x only contains the dust top height and density number (corresponding to the optical depth) that need to be solved, and other inputs for the forward model calculation are assumed to be known. The vector Ym is the observed AIRS brightness temperatures used in the retrieval process. The vector Y(xn) is the calculated AIRS brightness temperatures from a dust state xn by using a forward transfer model. The superscripts n and n + 1 indicate the iteration number and indicates an estimate or retrieved value of x.
De Souza-Machado et al.  illustrated that the dust optical depth retrieval is less sensitive to the geometrical thickness of the dust layer than to the dust layer height. To reduce the retrieval uncertainty, the dust geometrical depth is set to five layers, which is about 1 km in the lower troposphere. The initial dust top is set to log (617 hPa) with an uncertainty of 0.1. The initial density number is set to 2000, which result in an optical depth of 1 in the longwave infrared band, with an uncertainty of 2000. The brightness temperature uncertainty from the model and the observation is assumed to be 0.5 K at each selected channel. The matrix Sεis a fixed diagonal matrix in which each diagonal element is the square of the assumed uncertainty of the forward model and AIRS observation, while the off-diagonal elements are set to zero. Similarly, in the matrix Sa, each diagonal element is the square of the assumed uncertainty of the initial guess and all off-diagonal elements are set to zero.
 The optical depth at each channel could be calculated in the RTTOV model from the retrieved density number. As with Peyridieu et al.  and De Souza-Machado et al. , the 10.4 μm wavelength, which roughly corresponds to the maximum of the IR emission from Earth and the absorption of the dust, is chosen as the reference wavelength for infrared optical depth.
4. Simulation Analysis
4.1. Retrieval Accuracy
 For a simple and robust evaluation of AIRS dust retrievals, a series of simulation experiments are conducted. The simulation analysis could take advantage of the fact that the true values are known. The simulation experiments are conducted as follows: (1) specify true dust parameters xtrue; (2) simulate observations ysim using the forward model; (3) specify the constraint xa, its error covariance Sa and observation error covariance Sε; (4) retrieve xret; (5) calculate the bias xret − xtrue. The final retrieval is a weighted average of two quantities: an a priori estimate of parameters and those for which the predicted brightness temperatures from a forward model exactly match the observations. The weights are obtained from the covariance of the error constraint on the one hand and the effective accuracy of the observations on the other hand. In addition to the accuracy of the radiometer itself, effective accuracy depends on the sensitivity of the brightness temperatures to changes in the dust parameters.
Figure 1 shows the results of the simulation experiments for different dust top heights and optical depths. For a true vector of dust parameters identical to the constraint values, the retrievals produce identical results with zero bias. As the true vector is allowed to vary above or below the constraint, the retrieved parameters are always biased somewhat toward the constraint because the result is a compromise between the constraint and the observations. Our findings show that better results could be obtained in the case of optical depths greater than 0.5 and dust heights greater than 2.5 km. The worst retrievals occur when the optical depth is close to zero. In this case, the dust signature in the observations could be comparable with the forward and observation errors. When the true dust top height is between 3 and 5 km, the bias in the retrieved dust infrared AOD is less than 25%.
4.2. Retrieval Sensitivity to Other Parameter Errors
 The retrieval uncertainty not only depends on the sensitivity of TIR observations to dust parameters to be retrieved but also on the error associated with each of these measurements from the instrument itself and from the assumptions of the surface parameters and the atmospheric state. Instrument error primarily results from calibration issues and is about 0.2 K [Chahine et al., 2006]. The assumed error from the forward model used to simulate radiances, however, is generally larger. The remainder of this section is intended to test the algorithm's ability to estimate dust parameters in the presence of additional sources of error. In this study, “reference” values (AOD = 1, dust top = 3.93 km (at 617 hPa), surface emissivity = 0.98, atmospheric profiles = tropical climate) are used in the following simulation analysis.
 The surface emissivity uncertainty is considered as the primary error source for the dust property retrievals, especially for semitransparent dust. Pierangelo et al.  shows that if the dust optical depth is 1.5 at an altitude of 2.4 km, an increase in the surface emissivity form 0.8 to 0.9 changes the observations by 2 K or less. In this study, when the emissivity bias is set to −0.1, the dust top height retrieval error changes by +0.24 km, and the dust optical thickness error changes by −18%. Similarly, when the emissivity bias is −0.02, the retrieval errors for dust height and optical thickness change by +0.05 km and −4%, respectively. Furthermore, when a bias of −2 K is added to the surface temperature, the retrieval errors for dust height and optical thickness change by +0.15 km and −12%, respectively, in the reference case.
 The errors in temperature profiles from ECMWF reanalysis data should be far less than 2 K, which was reported by Eyre et al. . The impacts of atmospheric temperature errors on the brightness temperature calculation at these window channels are significantly less than those at sounding channels. Given that the Numerical Weather Prediction (NWP) error for the dust storm is generally worse, a bias of −2 K is added to the temperature profile to investigate the dependence of dust retrievals on the temperature profile errors. The results show that the dust height error changes by −0.65 km and the optical thickness error changes by 3%. It suggests that the temperature profile accuracy has significant impact on the dust top height retrieval.
 More than 20% of the globe is covered by cirrus clouds [Liou, 1986]. Consequently, dust observations may be contaminated by high-level cirrus.Hong et al. simulated high spectral resolution infrared observations under cirrus only, dust only, and cirrus and dust combined. Their study shows that the simulated observations are sensitive to cirrus in the case of the low-level dust simultaneously covered by cirrus. In theory, the bias in the dust top height decreases as the cirrus approaches the low-level dust. Further uncertainties associated with multilayered dust and the variation in dust composition, although they are certainly important, are beyond the scope of this paper and will not be discussed.
5. Application Results
 One of the world's largest shifting-sand deserts and a main source of atmospheric dust in northern China [Xuan and Sokolik, 2002], the Taklimakan Desert lies in the Tarim Basin, between the Tien Shan Mountains in the north and Kunlun Mountains in the south. The basin's lowest point is roughly 154 m below sea level, and salt collects in the basin due to a lack of drainage (http://www.eoearth.org/article/Taklimakan_desert). Because of its aridity and abundant sand, this desert is a regular source of dust storms in Asia. To evaluate the retrievals simply and robustly, we have excluded cases for which multiple aerosol layers are detected and measured by the lidar. Three dust storms in the Taklimakan Desert are analyzed in detail by using the 1 DVAR retrieval algorithm in the following study. Additionally, one storm originating from Mongolia is also selected to evaluate the dependence of the retrievals on the dust sources. In the retrieval process, the maximum number of iterations is limited to 10. In the iteration, the log of the dust top pressure must be 0.1 greater than that of the surface pressure, which is about 1 km in height above the surface level.
 Before retrieving the dust parameters, dust detection is performed by the algorithm described by De Souza-Machado et al. , which uses a summed score from brightness temperature differences from five AIRS channels (822.4, 900.3, 961.1, 1129.0, and 1231.3 cm−1) in the 10 μm atmospheric window region in order to identify dust pixels. In this study, comparisons against visible images show that this algorithm produces few false dust pixels for the selected cases. To further reduce false detection, the pixels with brightness temperatures at 822.4 cm−1 greater than their brightness temperatures at 900.3, 961.1, and 1129.0 cm−1are excluded. The results show that these additional criteria could produce more consistent results with the MODIS visible composite images. In the retrieval, the atmospheric profiles and surface temperature are assumed to be known and taken from the ECMWF reanalysis data. The surface infrared emissivity spectrum is taken from the University of Wisconsin-Madison (UW) baseline fit (BLF) emissivity data sets [Seemann et al., 2008].
5.1. Case 1: Dust Storm of 13 March 2006
 On 13 March 2006, a dust and clouds combination lingered over the Taklimakan Desert. In Figure 2a, an Aqua's MODIS visible composite image shows that the dust storm sweeps over most of the desert. From the northeast toward the center, the desert remains relatively dust free and its floor is clearly shown. White clouds fringe the peripheral edge of the desert.
 Because OMI, MODIS, and AIRS view the same area within 15 min, the AOD distribution should be very similar. Figures 2b and 2c also show the AIRS retrieved infrared AOD and OMI aerosol index for the dust storm. The heavy dust areas detected by AIRS are consistent with those shown in Figure 2. The aerosol distribution pattern from AIRS agrees well with that from the OMI aerosol index. Additionally, the MODIS DB visible AOD is also shown in Figure 2d and its distribution pattern also agrees very well with the AIRS AOD retrievals.
Figure 3a illustrates the relationship between the infrared AOD from AIRS and the aerosol index from OMI. In the collocation, the distance between the AIRS and OMI pixel centers is less than 7 km, and the AIRS pixels with retrieved AOD differences greater than 0.5 from the adjacent ones are excluded. As expected, a favorable correlation of 0.91 between the AIRS dust thermal optical depth and OMI aerosol index is obtained for this particular dust storm because the two instruments could be used to extract the similar features of the dust storm. It is noted that there are some pixels below the regression line. In the case of low OMI index values, the AIRS underestimation could be caused by the surface parameter errors due to high atmospheric transparence as mentioned above.
 A clear correlation is also illustrated between the AIRS infrared AOD and the MODIS visible AOD in Figure 3b. The slope of the linear regression between the MODIS visible AOD and the AIRS infrared AOD is about 1.35. It is noted that the ratio between the MODIS AOD and the AIRS AOD increases with the decrease of the AIRS AOD. It means that the relationship between the visible and infrared AOD values varies with the dust optical thickness and the MODIS is more sensitive to the light dust than the AIRS.
5.2. Case 2: Dust Storm of 10 May 2007
 A springtime dust storm covering most of the Taklimakan Desert in western China was captured by Aqua's MODIS at 07:30 UTC on 10 May 2007, which is shown in Figure 4a. In this image, the dust appears thickest in the west and even more concentrated at the western edge of the desert. North of the desert, snow cover caps the Tian Shan Mountains.
 The vertical cross section along the CALIPSO track is also shown in Figure 4c. It presents the CALIOP backscatter values (km−1 sr−1) from 0 to 12 km, with the color bar showing the backscatter times 1000. The surface level is represented by the solid black line. For the sake of clarity, CALIOP backscatter data less than 0.75 × 10−3 km−1 sr−1 are masked. The dashed line indicates the AIRS retrieved dust top heights. It is noted that a single dust layer is visible on the right half of the Taklimakan Desert. A thick layer of dust is present from 1 to 2 km, with a tenuous layer between 2 and 4 km. Its top height is about 3 km in the south and then gradually increases to 4 km in the north. It is encouraging that the AIRS retrievals could accurately capture the dust top height variation between 39° and 42° latitude. Also, it is noted that the AIRS retrieved top is significantly underestimated near 38° latitude. This should be caused by the relatively thin aerosol shown in Figure 4d, which is consistent with the simulation results in the case of the AOD less than 0.5 and the aerosol top greater than 5 km in Figure 1b. Between 41° and 42° latitude, the CALIOP dust top is about 1 km greater than the AIRS retrieval because the aerosol near the top is very thin as shown in Figure 4c.
 For this dust storm the optical depth generally decreases as the dust top height increases. From the horizontal distribution of the dust top height shown in Figure 4b, it is noted that some retrievals are near 5 km, which are significantly larger than neighboring values, thus these pixels could be contaminated by high-level cirrus. The optical depth as a function of latitude along the CALIPSO track is also shown inFigure 4d. There are some discrepancies between the CALIOP retrieved optical depths and those retrieved from the passive instruments, which is similar to the case presented by De Souza-Machado et al. . However, the variation trend between the AIRS retrieved infrared AOD and the MODIS visible AOD along the track between 38.5° and 41° latitude is consistent.
 From Figures 4e, 4f, and 4g, it can be seen that the horizontal distribution patterns of the retrieved AIRS AOD and the OMI index as well as the MODIS AOD are very similar. Owing to the relatively higher resolution, the MODIS AOD image could capture detailed variation as shown in the visible composite image, especially in the west of the dust storm. Figure 5b shows the comparisons of the infrared AOD retrievals with the optical products from OMI and MODIS. The correlation coefficients are very similar to those for the first case. It is noted that the offset of the regression between MODIS visible and AIRS infrared AOD values is about 0.81. This positive offset means that the visible AOD from MODIS may be overestimated and/or the infrared AOD from AIRS may be underestimated, especially for the light dust areas. The underestimation of the AIRS retrieval may result from the underestimation of the ECMWF surface temperature and/or the surface emissivity used in this study, especially in the case of semitransparent dust, as presented in the previous simulation studies. In general, AOD retrievals are mainly influenced by two factors: surface albedo and aerosol type. At low aerosol loading, the surface albedo is the primary factor causing errors; at high aerosol loading, aerosol model including the radiative transfer model and the OPAC database would affect the AOD retrieval. The first case is a heavy dust storm, while this one is a relatively light dust storm.
Figure 6 demonstrates the brightness temperature differences (BTD) between the observed and the simulated measurements before and after the iterations in the retrieval. Before the iteration, the BTDs are dramatically larger in areas with more transparent dust in the middle of the Taklimakan Desert as shown in the MODIS composite image, which means that the initial AOD is too large for these areas. However, it clearly shows that the iterative process significantly decreases the brightness temperature discrepancies, which could contribute to the final retrieval results.
5.3. Case 3: Dust Storm of 8 April 2007
 On 8 April 2007, dust plumes blew out of the Taklimakan Desert toward the east. They are captured by the MODIS image shown in Figure 7a. In this image, the dust seems more concentrated in the east of the desert. White clouds exist along the outside of the desert edge. The CALIOP cross section is also shown in Figure 7. Compared to the two previous storms, this one is relatively more transparent. Along the CALIPSO track, the high cirrus clouds and low level dust are all captured by the lidar, which is shown in Figure 7c.
 According to the lidar, a dust layer with a top height between 5 km and 4 km and dust base between 2 km and 3 km is present between 38° to 42° latitude. A cirrus layer with geometrical depth ranging from 500 m to 1 km is also present between 39° to 41.5° latitude, separated from the dust layer below by about 6 km. The comparison of the AIRS dust top and the CALIOP backscatters demonstrates that the dust top is retrieved accurately between 39° and 41° latitude, although the high level cirrus, which has a visible optical depth of less than 0.5 according to Figure 7d, are located above the dust layer with a visible optical depth of greater than 0.5. However, the AIRS dust top near 41.5° latitude is overestimated due to cirrus with a visible optical depth of greater than 0.5 overlapping the dust with a visible optical depth near 0. Between 37° and 38° latitude, the retrieved height seems dominated by the lower level clouds covered by optically thin dust. Additionally, the underestimation near 38.5° latitude may be caused by inaccurate surface and atmospheric parameters used in the retrieval. From Figures 7a and 7b, it is illustrated that the dust detection method could not mask the pixels in the presence of this optically thin cirrus. Some areas covered by thin cirrus are identified as dusty pixels and the retrieved dust top heights are significantly greater than those over the nearby areas. Over this desert, the dust could be lifted up above the surrounding mountains by the strong surface wind accompanying a cold front. As a result, it is difficult to use the negative slope between BT820 and BT960 as well as low BT820 values to indicate thin high cloud over the dust due to the fact that the high level dust could also generate a similar spectrum. Consequently, it is nontrivial to differentiate the high level dust from the high cirrus clouds overlapping a dust layer.
 From Figures 7e and 7g, it can be seen that the coverage of the dust from AIRS is similar to that from MODIS and the dust variation patterns are also very similar. Figure 7f shows that the heavy dust area from the OMI index agrees well with the results from AIRS. Figure 8 shows the comparisons of the infrared AOD retrievals with the optical products from OMI and MODIS. Although the AIRS infrared AOD correlates well with the MODIS visible AOD and slightly worse with the OMI index, the AIRS AOD and dust top height should be overestimated due to the presence of higher level cirrus, which could be easily seen from the MODIS image, over the northern part of the desert according to the analysis in section 4. The values greater than 3 in the MODIS visible AOD may also be contaminated by the cirrus. The correlation between the results from different instruments will be more complicated to explain due to their different sensitivities to the cirrus overlapping the dust. For this case, the majority of the AIRS infrared AOD values are less than 1, and the ratio of the MODIS visible AOD and the AIRS infrared AOD is greater than those for the two previous cases. This further confirms that ratio is greater for the light dust than for the heavy dust over this area.
5.4. Case 4: Dust Storm of 27 May 2008
 On 27 May 2008, a springtime dust storm spreading from Mongolia through northern China was captured by Aqua's MODIS image shown in Figure 9a. In this image, dust and clouds form a giant swirl at the borders of Mongolia, Russia, and China. Although clouds mask much of the dust and the arid landscape below, isolated tan streaks mixing with the clouds hover over eastern Mongolia. Dust plumes blow in a counterclockwise direction. Clear skies at southwest of the clouds allow a view of the plumes' origin. A large amorphous mass of uniform tan formed by the dust exists in the southeast.
 As the CALIOP backscatter shows in Figure 9c, a dust layer with a top height between 4 km and 3 km and dust base between 1 km and 3 km is present from 41° to 45.5° latitude. The geometrical thickness of the dust layer near 42° latitude reaches 3 km. From this location to 44.5° latitude, the geometrical thickness gradually decreases to 0.5 km. The AIRS retrieved dust top heights agree well with the lidar observations between 41.5° and 44.5° latitude. It seems that the retrieved dust top is insensitive to the bottom of the dust layer. The retrieved results near 45° latitude are about 1 km higher than the lower dust layer top. This difference could be explained by the higher thin dust or cirrus over the dust layer near the height of 3 km. It demonstrates that the algorithm works reasonably for the dust top height retrieval for this particular dust storm originating from Mongolia.
 For the AOD retrieval, the results reveal that the correlation between the AIRS retrieval and the OMI and MODIS products is remarkably worse than those for the three previous cases. Although the distribution pattern of the AIRS AOD appears to agree well with that of the OMI aerosol index shown in Figure 9e, the poor correlation between the AIRS infrared AOD and the OMI aerosol index partly resulted from the unreasonable variations between the adjacent observation lines along the satellite track. The significant variations of the horizontal distribution in dust details in the distribution may have been caused by the relatively poor horizontal resolution of the OMI. Additionally, although the heavy dust areas and the detailed variation of the dust storm are clearly shown in the MODIS visible composite image, they are not illustrated well by the MODIS visible AOD image as shown in Figure 9f, in which most of the values is significantly lower than the AIRS retrievals. It appears that the MODIS product is underestimated for this case.
 Also, it can be seen that the AIRS infrared AOD results are more consistent with the MODIS composite image than OMI and MODIS aerosol products, especially for the three heavy dust areas. As a result, it is difficult to conclude that the AIRS AOD retrieval is poor for this particular case. Strictly speaking, it is also hard to say that the optical properties from the dust originating from the Taklimakan Desert could be applicable for the dust originating from other areas.
5.5. Summary of Results for All Cases
 The relation between the AIRS retrieved AOD and the OMI aerosol index is summarized in Table 2. For the three single-layered dust storms over the Taklimakan Desert. The offset and slope of the regression between the AIRS infrared AOD and OMI aerosol index are stable. It is also noted that the UV AI is greater than 2 when the infrared AOD is near zero. Actually, the ultraviolet could detect small absorbing aerosol particles, which could not be sensed by the longwave infrared channels. Consequently, some dust pixels with an OMI aerosol index less than 2 could not be correctly identified by using the AIRS longwave infrared band. In other words, this algorithm works well for the optically thick dust storms evaluated in this work but not for tenuous dust.
Table 2. Correlation Coefficients (R), Regression Offsets, and Slopes Between AIRS Infrared AOD and OMI Aerosol Index for Different Asian Dust Storms
AIRS Infrared AOD Versus OMI Index
AIRS Infrared AOD Versus MODIS DB AOD
 The relation between the AIRS infrared and MODIS DB AOD values is also shown in Table 2. For the three single-layered dust storms, the AIRS infrared AOD is correlated very well to the MODIS DB AOD. The slopes are comparable to the ratio of 1.5–2.5 reported byPeyridieu et al.  for three regions of the tropical North Atlantic and one region of the northwestern Indian Ocean. They also find a slight increase of this ratio when the distance to the sources increases, which may be caused by a small loss in coarse size particles due to gravitational settling. It is also noted that the ratio for the third case is greater than those for the two former cases. From Figures 3b, 5b and 8b, it is shown that the ratio is about 2 when the AIRS AOD is less than 1. It means that the relationship between the AIRS and MODIS DB AODs for the light dust is stable over the Taklimakan desert. With the increase of the AIRS AOD, the ratio will decrease because of the increase of the coarse size particles.
 Furthermore, cases of lower level dust overlapped by cirrus and double-layered dust are difficult to address due to their complexities. When multiple layers coexist in a vertical column and a single-dust layer assumption is imposed on a retrieval of the dust properties, the resulting dust top tends to be incorrect. The followup effort of the present study will focus on differentiating dust from the cirrus-contaminated observations over the typical dust regions.
 Although the method could be applied to dust storms occurring over other areas, the quality of the retrievals could not be guaranteed due to the discrepancies in dust composition, which could lead to differences in the absorption and scattering properties. The correlation between the AIRS retrievals and OMI and MODIS optical products significantly degraded for a dust storm originating from Mongolia, which may be partly caused by the different dust composition although the offset is still comparable with those from other cases.
5.6. Impact of Surface Temperature Uncertainty Assumption on Retrievals
 Simulation analysis shows that surface temperature error has a strong impact on the AIRS AOD retrieval. In addition, The ECMWF surface temperature may not be trusted during the day since solar warming of the surface is so variable. Consequently, one may have to simultaneously adjust the surface temperature in the retrieval. To investigate this impact, the experiments are also conducted by adjusting the surface temperature, in which the surface temperature as a variable is added to the vector x described in section 3.5 and other corresponding vectors are also changed easily. In theory, the sensitivity of the AOD retrieval to the surface temperature error increases with the decrease of the dust optical thickness. So, this analysis will mainly focus on the case 3, which is a relatively light dust storm and in which the majority of the visible AOD value are less than 1.5.
 Comparisons of the AIRS AOD values from different surface temperature uncertainty assumptions against the MODIS DB visible AOD products are shown in Figure 10. It shows that the correlation is gradually improved with the increase of the surface temperature uncertainty. It also clearly reveals that the offset of the linear regression decreases as the surface temperature uncertainty increases. Further analysis shows that the improvements mainly result from the AOD increase in the light dusty pixels, which should be caused by the increase of the surface temperature in the retrieval. The results indicate that the inclusion of the surface temperature as a variable could make the dust AOD retrieval slightly better for this dust storm. For the second case, which is a relatively heavy dust storm, the correlation coefficients of the AOD values do not change for different assumptions as the surface is mainly blacked out under conditions of heavy dust loading.
6. Conclusions and Discussion
 It is important to detect dust and estimate its properties to evaluate the impact of dust on regional weather and climate systems as well as improve atmospheric parameter retrievals. However, the quality of satellite dust retrievals over land critically depends upon the modeling accuracy of surface properties and dust composition. Also, comparing aerosol data sets from different satellite-based instruments is challenging, and large discrepancies may occur because each retrieval algorithm has its own limitations and shortcomings [Liu and Mishchenko, 2008]. Good correlations between the AIRS infrared AOD and MODIS DB for different one-layered dust storms over the Taklimakan Desert reveal that hyperspectral infrared radiances in the longwave infrared band could be used to retrieve the dust storm optical property over land. The work reported here could help develop an operational dust retrieval algorithm for future AIRS processing. Particularly, it could provide the dust height in the day and night that other passive instruments cannot. The algorithm could also be used to study day/night differences of the optical depth and height of dust close to its sources. Furthermore, the synergetic use of the AIRS and MODIS for the improvement of the dust properties will be conducted in the near future by using the stable relation between the AIRS AOD and MODIS DB. Additionally, the retrievals could be used to validate the dust forecast results. Also, assimilation of the results are expected to help dust forecast.
 Although the AIRS observations contain abundant dust information and the simulation retrievals by using more channels show that the dust properties could be accurately extracted, the retrieval capability from AIRS in shortwave IR channels is limited by the reflected solar radiation. In particular, the simulated observations are generally over 10 K less than the observations in the shortwave band in the daytime. Further experiments also confirm that the iteration could not converge due to very large brightness temperature values in the shortwave band. Consequently, dust properties could not be effectively extracted by using short-wave infrared channels in the day. Additionally, to extract dust information globally, the dust composition and their optical properties over different areas should be investigated by in situ measurements, which will significantly benefit the retrievals.
 The authors would like to thank Timothy J. Schmit for his suggestions on improving the manuscript. Elisabeth Weisz is also thanked for her help on using AIRS data. This work is partly supported by basic research program 2010CB950802, the National Oceanic and Atmospheric Administration (NOAA) GOES-R Algorithm Working Group (AWG) and GOES-R Risk Reduction programs NA10NES4400010 and the National Natural Science Foundation of China (NSFC) grant 41075021. Hyo-jin Han and B. J. Sohn were supported by the National Space Lab program funded by the National Research Foundation of Korea (S10801000184-08A0100-18410) and the Korea Meteorological Administration Research and Development Program under grant CATER 2012-2060. The authors would like to specifically thank the AIRS, MODIS, OMI, and CALIPSO science teams for providing satellite data and ECMWF for sharing the model output data. We appreciate the anonymous reviewers for their valuable comments and suggestions to improve the quality of this paper.