Corresponding author: R. Liu, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, No11A, Datun Road, Beijing 100101, China. (email@example.com)
 The brightness temperature difference (BTD) between two thermal infrared bands is a common index for dust detection. However, the BTD is sensitive to the observed temperature, which hinders its use in automatic dust detection, especially over desert land surfaces. In this paper, a dynamic reference brightness temperature differences (DRBTD) algorithm was developed to detect dust by removing the influence of the observed temperature on the BTD. Using long-term MODIS observations, the algorithm establishes the clear-sky linear relationships pixel by pixel between the brightness temperatures (BTs) at 12 and 11 µm channels and the relationships between the BTs at 8.6 and 11 µm channels. From these relationships, the reference BTDs are dynamically generated according to the observed brightness temperatures. Next, the DRBTDI, which is the difference of the observed BTD and the reference BTD, is created and used to separate the dust from other observed objects. This algorithm is applied to MODIS observations to detect several dust events during the daytime and the nighttime over Mongolia and northwestern and northern China. The results are compared with Ozone Monitoring Instrument aerosol index (OMI AI), MODIS Deep Blue aerosol optical depth (AOD), and Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) observations. The comparisons indicate that the DRBTD algorithm can effectively distinguish dust from clouds and land surface. During the daytime, the DRBTDI is correlated with the OMI AI and MODIS AOD with a correlation coefficient of Pearson (r) of 0.79 and 0.77, respectively. At night, the DRBTDI is correlated with the CALIOP dust AOD with an r of 0.78.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 Mineral dust is an important factor in global climate, biogeochemical cycles, and air pollution [e.g., Mahowald et al., 2011]. It impacts climate directly by altering the radiation balance in the atmosphere through scattering and absorbing radiation [Haywood and Boucher, 2000] and indirectly by altering cloud properties [Wurzler et al., 2000]. Dust also increases ocean photosynthesis rates by carrying iron to the ocean [Krishnamurthy et al., 2010]. The characterization of dust properties and distributions at regional and global scales would help us better understand the role of dust in Earth's radiative budget and the global biogeochemical cycle [Nobileau and Antoine, 2005].
 Satellite remote sensing is advantageous in monitoring the spatial and temporal variations of dust events [Chiapello et al., 1999]. The satellite radiometer measurements in visible and ultraviolet channels have been used effectively to detect dust. For example, the ultraviolet measurements of the Nimbus 7 Total Ozone Mapping Spectrometer (TOMS) and the Ozone Monitoring Instrument (OMI) have been used to detect mineral dust [Chiapello et al., 1999; Torres et al., 2007], and the observations in the visible bands of the SeaWiFS and MODIS have been used to characterize the properties of dust aerosols [Hsu et al., 2006]. However, these algorithms, which are based on visible and ultraviolet channels, are applicable only during the daytime.
 Satellite radiometer measurements in thermal infrared channels provide another approach for dust detection during both the daytime and the nighttime [Shenk and Curran, 1974; Ackerman, 1989, 1997; Wald et al., 1998]. Although the brightness temperature (BT) of the dust layer is related to the physical properties of aerosol particles [Ackerman, 1997], the BT is not used as a direct indicator for automatic dust detection because it is sensitive to the temperature of the observed objects. Instead, the brightness temperature difference (BTD) between two thermal infrared spectral bands, which depends to a great extent on the mineral composition [e.g., Hudson et al., 2008], emissivity, and the particle sizes of the observed objects [Wald et al., 1998] and which usually shows strong contrasts among earth surfaces, dust, and clouds, is a more reliable index for detecting dust [Schepanski et al., 2007; Zhang et al., 2006]. For example, negative differences in the brightness temperature at 11 and 12 µm channels were observed for a dust-laden atmosphere in comparison to clear-sky conditions [Ackerman, 1997; Sokolik, 2002]. Airborne dust would also increase the BTD between the 8.6 and 11 µm wavelength bands [Wald et al., 1998].
 In addition to the atmospheric conditions, such as water vapor, aerosols, clouds, and temperature profiles, the underlying surface also makes considerable contributions to the BTDs. To eliminate the contributions of the underlying surface, several scientists have attempted to compose a reference BTD map representing the clear-sky conditions and extract the dust signals from the differences between the observed BTD and the reference BTD (hereafter referred to as the BTDanom method) [Ashpole and Washington, 2012]. Unfortunately, the BTD is slightly dependent on the observed temperature (section 5.1). Over water and dense vegetated surfaces, the temperature effect on the BTD is not obvious [Ackerman, 1997; Darmenov and Sokolik, 2005] because of the high contrast of the BTDs between the dust and the underlying surface. However, over desert and semidesert regions, the temperature effect is an impediment for dust detection due to the similarities of the emissivities of surface and airborne dust. Because the temperature of the clear-sky surface is usually much higher than that of airborne dust, the temperature effect still exists in the BTDanom method.
 In this paper, the dynamic reference brightness temperature differences (DRBTD) algorithm is proposed to remove the temperature effect on the surface reference BTD maximally to detect airborne dust during both daytime and nighttime. A highly linear relationship was found between the brightness temperature of two thermal bands under the same emissivity and atmospheric conditions (see section 3), and this relationship is independent of temperature. The linear relationship between the brightness temperatures of 12 (BT12) and 11 µm (BT11), as well as between the brightness temperatures of 8.6 (BT8.6) and 11 µm (BT11) wavelength channels under clear-sky conditions were established pixel by pixel from the time series MODIS observations from 2000–2008. From these relationships, the reference BTDs were dynamically generated at the pixel level to represent the clear-sky status according to the observed temperatures. The observed BTDs were later compared with the reference BTD to separate the dust from the clouds and the clear-sky surface. The desert regions in Mongolia and northwestern and northern China (33°N–54°N, 73°E–136°E), which are the main source of airborne dust in Asia [Prospero et al., 2002], were selected as the study area for the application of this algorithm.
 The paper is organized as follows. Section 2 presents a brief introduction to the data. The linear relationships between the brightness temperatures in the 8.6, 12, and 11 µm bands are demonstrated in section 3. The algorithm is described in detail in section 4. Section 5 describes the performance of the DRBTD algorithm on minimizing underlying surface effects, the results of dust detection efforts during several dust events, and the comparisons with OMI AI (aerosol index), MODIS Deep Blue AOD (aerosol optical depth), and CALIOP (Cloud-Aerosol Lidar with Orthogonal Polarization) observations during the daytime and nighttime. The influence of clouds, atmospheric water vapor, and surface changes are discussed in section 6. The final summary is presented in section 7.
2.1 MODIS Data
 The MODIS, aboard the Terra and Aqua satellites, observes the Earth's entire surface every 1 to 2 days, crossing over the equator at 10:30 and 13:30 local time (LT) during the day and at 22:30 and 01:30 LT at night. The MODIS acquires data in 20 visible and shortwave infrared spectral channels and 16 thermal infrared channels. Among these channels, the 8.6 (Band 29), 11 (Band 31), and 12 µm (Band 32) channels are widely used in dust detection. The MODIS Level 1B Calibrated and Geolocated Data Set (MOD02) contains global-calibrated and geolocated at-aperture radiance values for all 36 bands. In this paper, the MOD02 radiances at wavelengths λ = 8.6, 11, and 12 µm at 1 × 1 km resolution at nadir for the period from February 2000 to December 2008 were gridded and converted into brightness temperature values and used in the DRBTD algorithm for dust detection. The MODIS Global Daily Level 2 Aerosol Product (MYD04_L2) provides global aerosol AOD datasets based on the Deep Blue algorithm with a spatial resolution of 10 km [Hsu et al., 2004]. These AOD products were compared with the DRBTD algorithm results. Additionally, the MODIS Precipitable Water product (MOD05_L2), which includes the column water vapor datasets with a 1 km resolution retrieved from the near-infrared algorithm, was used to illustrate the column water vapor distributions of the study area.
2.2 OMI AI
 The OMI aboard the Aura satellite flies in formation with five other satellites (including Aqua) in the international “A-Train” constellation for coincident Earth observations approximately 15 min behind the Aqua satellite. The OMI has collected radiative measurements of the Earth every day in near-UV and visible spectral channels since August 2004. The daily global-absorbing aerosol index (AI) was mapped from radiance measurements of the UV wavelengths and used to characterize the dust aerosols over deserts [Torres et al., 2007]. AI represents the signals of absorbing aerosols, including desert dust, carbonaceous aerosols from biomass burning, and weakly absorbing sulfate-based aerosols. AI is sensitive to many factors, such as the aerosol type and composition, aerosol layer height, and the single-scattering albedo. Generally, AI is positive for absorbing aerosols and negative for nonabsorbing aerosols. In this paper, the global daily 1.0° gridded UV AI from the Collection 3 OMI-Aura-OMTO3d product were compared with the algorithm results.
 The Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) satellite is also part of the NASA A-Train constellation, lagging Aqua by 1 to 2 min. The Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) aboard the CALIPSO satellite has provided high resolution information on the vertical structures and properties of aerosols and clouds across the globe since June 2006. CALIOP is a polarization-sensitive lidar with two wavelengths (532 and 1064 nm) from a near-nadir-viewing geometry during both day and night phases of the orbit. The level 2 aerosol layer products of CALIOP include the optical depth of all of the aerosols detected within a 5 km averaged profile at 532 and 1064 nm. The DRBTD algorithm results were compared with the nighttime AOD at 532 nm to evaluate the performance of the DRBTD algorithm for nighttime dust detection.
3 Simulation of the Linear Relationships Between TOA BTs in Channel Pairs of 8.6/11 and 12/11 µm Using an Atmospheric Radiative Transfer Model
 Based on Planck's law, it can be inferred that the BTs of the land surface at two thermal infrared bands have a linear relationship, with the slope and the offset being determined by the emissivity (ε) of these two channels. The linear relationships between BT8.6 and BT11, as well as between BT12 and BT11, are demonstrated in Figure 1. The BTs at the 8.6, 11 and 12 µm channels were calculated for land surface temperatures ranging from 260 K to 320 K and with the same surface emissivity set as that of bare soil from the SeeBor database [Borbas et al., 2005]. Similar linear relationships were also found for other surface types with different slopes and offsets.
 For the observed top of atmosphere (TOA) BTs at multithermal infrared bands, the relationships between the different channels would be modified by the atmospheric conditions. The effects of the atmospheric conditions on the relationships of the TOA BTs at thermal infrared bands (in this study, the 8.6, 11, and 12 µm channels) were simulated by the Santa Barbara DISORT (DIScreet Ordinate Radiative Transfer) Atmospheric Radiative Transfer (SBDART) model for bare soil with the standard atmospheric profiles of midlatitude winter (MLW) (Table 1) [Ricchiazzi et al., 1998]. The dust optical properties, including the single-scattering albedo and asymmetery parameter, were defined using tabulated values for mineral nucleation mode dust from OPAC (Optical Properties of Aerosol and Cloud) [Hess et al., 1998]. The surface emissivities of bare soil at the three channels were set based on the SeeBor database [Borbas et al., 2005].
Table 1. Parameters for BTs Relationships Simulation With SBDART
Midlatitude Winter (MLW)
Integrated water vapor amount (g/cm2)
0, 0.214, 0.427, 0.641, 0.854, 1.068, 1.281
View zenith angle
260 K–295 K, step = 1.0 K
Emissivity of bare soils
ε8.6 = 0.8597, ε11 = 0.9312, ε12 = 0.9433
Altitude of aerosol layer
Aerosol optical depth (at 0.55 µm)
0.2, 0.5, 1.0, 2.0, 3.0, 5.0
Altitude of cloud layer
Cloud optical depth (at 0.55 µm)
1, 10, 30, 50, 100
Cloud drop effective radius (µm)
 Figures 2a and 2b present an example of the simulated relationships between BT8.6 and BT11 and between BT12 and BT11. The results demonstrate the linear relationships between BT12 and BT11 as well as between BT8.6 and BT11 with a slope of approximately 1 in the standard atmospheric profiles without clouds, aerosol, or water vapor. This line is referred to as the ideal pristine line. When there is dust, clouds or water vapor in the atmosphere, the relationship departs from the ideal pristine line, but it is still linear under the same atmospheric conditions. For the relationships between BT8.6 and BT11, when a dust layer is present in the atmosphere, the simulated points are above the ideal pristine line of BT8.6/BT11. With an increase in the optical depth of the dust layer, the dust points are distant from the ideal pristine line. The linear relationship between BT12 and BT11 is less sensitive to dust aerosols with dust points slightly above the ideal pristine line of BT12/BT11. The presence of clouds has complex effects on the relationships among the BTs. In Figure 2, only the results of clouds with a cloud optical depth of 1 are shown to make the figure more readable. For thin clouds with a cloud optical depth at 0.55 µm of less than 10 µm, the simulated points are far below the ideal pristine line for BT12 and BT11, while they are above the ideal pristine line for BT8.6 and BT11. For thick clouds with a cloud optical depth greater than 10 µm, the BTs significantly decrease to approximately 220 K and are little related with the underlying surface temperature. These simulated points are below the ideal pristine line for BT12 and BT11; however, they are either below or above the pristine line for BT8.6 and BT11. When atmospheric water vapor is taken into account, the slopes of the BT12/BT11 and BT8.6/BT11 relationships decrease slightly with points below the ideal pristine line.
 The dust layer attitude and atmospheric water vapor profiles would affect the satellite-observed BTs [e.g., Brindley and Russell, 2009]. The relationships between BT8.6 (BT12) and BT11 were simulated with various dust layer heights with atmospheric profiles of MLW. Figure 2 presents the relationships between BT8.6 and BT11 (Figures 2a, 2c, and 2e), as well as that between BT12 and BT11 (Figures 2b, 2d, and 2f), with a dust layer height of 2 km, 0.5 km, and 5 km, respectively. As the dust layer altitude increases, the BT8.6, BT11, and BT12 of dust observations decrease, especially for the heavy dust layer with an AOD greater than 2.0. However, the relationships between BT8.6 (BT12) and BT11 are similar for various dust layer heights, with dust points above the ideal pristine line and the distance between them increasing with AOD for BT8.6/BT11, and dust plots around the ideal pristine line for BT12/BT11.
 The atmospheric radiative transfer simulations were also performed with different standard atmospheric profiles to evaluate the effects of water vapor on the relationships between BT8.6 (BT12) and BT11, including MLW and MLS (Midlatitude Summer). The dust layer altitude is set to 2 km, and the surface temperature ranges from 280–315 K with a 1 K step for MLS simulations. The column water vapor content of MLW is approximately 0.854 g/cm2, which is much less than that of MLS (approximately 2.924 g/cm2). Figures 3a and 3b show the results for MLW, and the relationships for MLS are presented in Figures 3c and 3d. For the relationships between BT8.6 and BT11, the results are similar for MLW and MLS, with dust points being observed above the ideal pristine line, and the distance between them increasing with AOD. On contrast, differences are presented for that of BT12 and BT11 in two atmospheric profiles. For MLW, the dust points are close to the ideal pristine line, while they are below the ideal pristine line for MLS, indicating that uncertainties may be introduced by DRBTDI12 when applying the algorithm in humid regions and seasons. Therefore, the current approach is more appropriate to dry regions with low water vapor content in the atmosphere.
4 Description of the DRBTD Dust Detection Algorithm
4.1 Generation of Reference BT8.6/11 and BT12/11 Relationships From Observations
 From the simulation of Figure 2, the relationship between the BTs of thermal infrared bands is linear for the same atmospheric conditions. If BT11 is assumed to be the observed temperature, the clear-sky TOA BT8.6 and BT12 could be expressed as functions of BT11 with the slopes a8.6 and a12 and enhancements b8.6 and b12, respectively, which are related to the emissivity of the underlying surface, and the atmospheric conditions:
 The a8.6, b8.6, a12, and b12 for the ideal pristine line can be derived from the simulation of the atmospheric transfer model if the emissivity of the underlying surface is known. Unfortunately, retrieval of the emissivity of the underlying surface is difficult. Here, a line approximate to the ideal pristine line, which is defined as the most clear-sky line, was regressed from the time series MODIS observations pixel by pixel. Because the observations for the cloud or vapor conditions are below the most clear-sky line of BT12/BT11, a scatterplot between BT12 and BT11 for an ensemble of MODIS observations from the same position would have an upper envelope line with the slope b12 and enhancement a12 which represents the most clear-sky conditions of the observed site (hereafter referred to as the most clear-sky line of BT12/BT11). The observational data points for cloud conditions or humid atmospheric conditions fall below this most clear-sky line, while the points of dust conditions may lie above or below the most clear-sky line of BT12/BT11. In contrast, for scatterplots of the relationship between BT8.6 and BT11, the lower envelope line represents the clear-sky conditions of the slope a8.6 and enhancements b8.6 (hereafter referred to as the most clear-sky line for BT8.6/BT11). The BT8.6/BT11 data points for dust and cloud conditions are above the most clear-sky line, while those for humid atmospheric conditions are below this line. Because the columnar vapor amounts are usually low in these regions, the most clear-sky line in the desert regions should be very close to the ideal pristine line.
 The upper envelopes for BT12/BT11 and the lower envelopes for BT8.6/B11 were regressed with the MODIS observations for the 9 year period from 2000–2008. The following steps were performed to quantify the upper and lower envelopes, taking the upper envelopes for BT12/BT11 as the example. First, for a specific site, the MODIS observations from 2000–2008 were binned sequentially using BT11 values with a 5 K bin size. Second, the five highest BT12 values were kept in each bin, and a line was regressed through these points. Then, among the observations below the regression line, the one with the largest distance with respect to the regression line was discarded. The line was regressed using the left observations iteratively. If all the values in one bin had been removed, this bin was excluded in the next regression procedure. This regression procedure was iteratively performed until the valid bin was less than 10. The final regression line is the upper envelope line and was considered as the most clear-sky line of BT12/BT11. Because the BT values for clouds were much lower than that of land surface and dust layers and far away from the most clear-sky line, these points were discarded automatically during the regression procedure. The lower envelope line for BT8.6/BT11 was acquired by linear regression of points with the five lowest BT8.6 values in each BT11 bin, using a similar procedure for BT12/BT11. Finally, these regression lines were taken as the most clear-sky line for BT12/BT11 and BT8.6/11 and the values of a8.6, b8.6, a12, and b12 were mapped for each pixel at 1 × 1 km resolution over the study area. Although the dust observations lying above the clear-sky line would jeopardize the regression of the upper envelope for BT12/BT11, this effect is minor because these dust observations are mainly concentrated on a few bins and its number is small with respect to the number of clear-sky observations. Figure 4 shows several samples of scatterplots for several sites in China: the Taklamakan desert site, the Ulanqab semiarid site, the Qinlin forest site, the Beijing urban site, and the Qinhai Lake site. Both the upper and lower envelopes are well defined by the regression line with an R2 of approximately 1, which demonstrates that the linear relationship is appropriate for the given locations.
4.2 The Dynamic Reference Brightness Temperature Differences Index (DRBTDI)
 Based on the relationships of clear-sky BT8.6/BT11 and BT12/BT11, for a pixel with an observed BT11 (BT11obs), its corresponding reference BT8.6 (RBT8.6) and reference BT12 (RBT12) can be inferred as follows:
 The reference BTD for 8.6 and 11 µm (RBTD8.6) and for 12 and 11 µm (RBTD12) for each pixel, which represents the BTD at the most clear-sky condition without airborne dust, was dynamically calculated from the observed BT11 as follows:
 The difference between the observed BTD and the reference BTD for 8.6 and 11 µm (DRBTDI8.6) and for 12 and 11 µm (DRBTDI12) can be expressed as follows:
where BT8.6obs, BT11obs, and BT12obs are the observed BT in 8.6, 11 and 12 µm bands, respectively.
 Because DRBTDI8.6 can distinguish dust and clouds from underlying surfaces and DRBTDI12 can distinguish dust from clouds, the combination of these two indices can distinguish dust from both clouds and land surfaces. The DRBTDI is defined as the summation of DRBTDI8.6 and DRBTDI12:
 When the observed BTD is compared with the reference BTD dynamically derived from the same temperature, the effects of temperature on the underlying BTD can be mitigated. Because the contributions of land surface have been removed by the linear relationships between BT8.6 and BT11 and between BT12 and BT11, the DRBTDI8.6 and DRBTDI12 are signals of the atmospheric conditions, especially the dust aerosols and clouds. For clear-sky surfaces, the reference BTD is close to the observed BTD such that the DRBTDI8.6 and DRBTDI12 are approximately 0, and the DRBTDI is also approximately 0. The DRBTDI8.6 would increase rapidly and the DRBTDI12 would somewhat increase with the increase of the dust AOD, resulting in positive values of the DRBTDI. For cloud conditions, the DRBTDI12 is usually negative, while the DRBTDI8.6 is generally greater than zero. Generally, the DRBTDI12 is a large negative value for clouds. When the DRBTDI8.6 is a large positive value, the DRBTDI would be a positive value. In this case, the negative DRBTDI12 is used to identify the cloud. For a thin cloud, the DRBTDI12 is usually a small negative value and the DRBTDI would be a small positive value, resulting in a nearly zero value of DRBTDI. Therefore, the dust pixels were identified by positive DRBTDI values and DRBTDI12 and DRBTDI8.6 values greater than −0.5. The pixels with DRBTDI12 or DRBTDI8.6 values of less than −0.5 were classified as clouds based on the analyses of the observations over the study area. Although the DRBTDI8.6, DRBTDI12, and DRBTDI values are sensitive to many factors, such as atmospheric water vapor, dust single-scattering albedo, dust layer height, and viewing geometry, they are strongly positively correlated with the AOD, which means that this index can be used to semiquantitatively indicate the dust optical thickness.
5.1 Evaluation of the Effects of Temperature on the Reference BTD
 The reference BTD, which represents the BTD at the most clear-sky condition without airborne dust and clouds, relates to surface temperature. The effects of temperature on the reference BTD were evaluated based on a8.6, b8.6, a12, and b12 using equations (5) and (6). The reference BT8.6 and BT12 were inferred from the BT11, and the RBTD8.6 and RBTD12 were calculated from the reference BT8.6 and the reference BT12.
 Figure 5 shows the RBTD8.6 and RBTD12 over the study area with BT11 = 260 K (Figures 5a and 5c) and BT11 = 280 K (Figures 5b and 5d). The RBTD12 and RBTD8.6 show significant variation from site to site. Generally, the RBTD12 is positive in clear-sky conditions in desert regions. For a given BT11, the RBTD12 for desert surfaces in the northeast part of the study area is greater than that for vegetated regions, which is approximately zero. Negative values of RBTD8.6 are observed for desert surfaces, which have also been demonstrated by Hansell et al. . The RBTD8.6 is greater for vegetative pixels than for desert surfaces. These differences are attributed mainly to the different emissivities of various land cover types. Noticeable variations also exist within the same land cover type, which is attributed to the differences in emissivity due to different surface components and water content. With the larger variability in the emissive nature of the 8.6 µm band, the RBTD8.6 shows a more significant spatial variability than the RBTD12.
 The reference BTDs are highly affected by the temperature. With an increase of BT11 from 260 K to 280 K, the RBTD12 generally decreases by approximately 0.5 K, especially in sparsely vegetated regions in the northwest part of the study area where the RBTD12 decreases from 1.5 K to 1.0 K. The RBTD8.6 shows more significant decreases, decreasing from approximately 0 K to −2 K when BT11 increases from 260 K to 280 K in the vegetated regions in the northern and eastern parts of the study area, whereas RBTD8.6 decreases from −3.5 K to −5.5 K in the desert surface regions. These results are consistent with those of Ashpole and Washington , who suggested that the reference BTD11-8.7 is higher and the reference BTD11-8.7 is lower at nighttime than during the daytime. Thus, the bias due to temperature for the RBTD12 and the RBTD8.6 should not be neglected. The dynamic generation of clear-sky reference BTDs as proposed here should help improve the identification of the dust signals.
5.2 Evaluation of the Ability of the DRBTD Method to Remove Variability of the Underlying Surface
 For those most clear observations with low water vapor and low aerosol in the atmosphere, the atmospheric effect is small. In these conditions, if a dust detection method works well, the variations from the underlying surface should be small. That is, an ideal method to remove the underlying surface variations should cause a clear-sky observed scene to appear homogeneous, and a method to remove the temperature effect should cause two clear-sky scenes with different temperatures in the same region to be identical. Otherwise, the spatial and surface temperature variations may result in great variations of the dust index even in clear-sky conditions without airborne dust loading, making it hard to separate dust from clear-sky surfaces. Two clear-sky cases were selected to evaluate the ability of the DRBTD algorithm to remove the spatial variations and temperature effects of underlying surfaces. One case is of an inhomogeneous region over northwest China and Mongolia, evaluating the removal of spatial variations of the underlying surface, and the other case is of a small homogeneous region over the Taklimakan Desert during different seasons, evaluating its performance under different surface temperature conditions. We also seek to test the performance of our new DRBTD method to remove variability in observations against traditional BTD and BTDanom methods, which also focus on elimination of the contributions of the underlying surface by composing a reference BTD map representing the clear-sky conditions and extracting the dust signals from the differences between the observed BTD and the reference BTD [Ashpole and Washington, 2012]. The BTDanom value was calculated from the difference between the observed BTD and the composite reference BTD map, which is the mean BTD from 15 days of rolling clear-sky observations [Ashpole and Washington, 2012]. The standard deviation (SD) and mean values were calculated to evaluate the homogeneity and the distribution of the results.
5.2.1 Performance for Removal of Spatial Variation of the Underlying Surface
 A region in northwestern China and Mongolia with an area of 624,900 km2 (represented by the pink rectangle in Figure 6) was selected to evaluate the ability of the DRBTD algorithm to remove the spatial variations from the underlying surface. The region includes diverse surface types with different physical components, including the Tenggeli Desert, the Gobi Desert, and the semiarid region in the Gansu and Inner Mongolia provinces in China. The clear-sky case that was selected was from 10 October 2008. Figures 6c, 6d, and 6e show typical surface emissivities for October 2008 in this region in 8.3, 10.8 and 12.1 µm channels, as derived from a combination of MODIS observations and laboratory-measured emissivity spectra [Seemann et al., 2008]. Though the wavelength regions do not overlap exactly with those of MODIS, the patterns should be comparable. It can be observed that the surface emissivity varies spatially, especially for 8.3 µm channels, which affects the satellite-observed BTs and make it difficult to deduce the presence of dust using thermal infrared satellite observations. The DRBTD, BTD, and BTDanom methods were applied to the thermal infrared bands in the 8.6 and 12 µm channels to the 11 µm channel from the Terra/MODIS observations.
 The RGB (red, green, and blue) color composite image from the Terra/MODIS TOA reflectance and column water vapor amount are shown in Figures 6a and 6b, respectively. Figures 7a–7f show the maps of BTD8.6-11, BTD12-11, BTDanom8.6,11, BTDanom12,11, DRBTDI8.6, and DRBTDI12, and their corresponding distribution frequencies are shown in Figure 8. Table 2 presents the mean and SD of these indices. Generally, the DRBTDI values show much stronger homogeneity than those of the BTD and they are also slightly more homogenous than those of BTDanom. This comparison suggests that the DRBTD algorithm is more effective in the removal of the underlying surface variations. The BTD8.6-11 values range from −11.7 K to 3.5 K with a mean of −5.71 K and a SD of 1.73 K. The BTDanom8.6,11 has a lower SD of 0.70 K with a mean value of 0.61 K. In contrast, using our DRBTD algorithm, the values of DRBTDI8.6 are more concentrated around zero, with a lower SD of 0.41 K and a mean value of 0.32 K. Similar results were observed for the 12 µm band. The SD of DRBTDI12 (0.16 K) is smaller than those of BTD12-11 (0.32 K) and BTDanom12,11 (0.22 K). The mean of DRBTDI12 (0.28 K) is similar to that of BTDanom12,11 (0.24 K). The spatial variabilities of BTD12-11, BTDanom12,11, and DRBTDI12 are much smaller than those of BTD8.6-11, BTDanom8.6,11, and DRBTDI8.6, respectively, which may be due to the more spatially variable emissivities of the different surface materials around the 8.6 µm band compared to those around the 12 µm band (Figures 6c and 6e).
Table 2. Statistic of the BTD, BTDanom, and DRBTDI in North China and Mongolia on 10 October 2008
5.2.2 Performance for Removal of the Temperature Effect on the BTD
 The performances for the removal of the temperature effect on the BTD by the DRBTD, BTD, and BTDanom algorithms were evaluated with two clear-sky Terra/MODIS scenes over the Taklimakan Desert, a relatively homogeneous region with an area of 40,350 km2 (represented by the pink rectangle in Figure 9), taken in autumn (28 September 2008, DOY272, Figure 9a) and winter (4 January 2007, DOY4, Figure 9b). The BT11 of this region is approximately 310 K for 28 September 2008 (Figure 9c) and 275 K for 4 January 2007 (Figure 9d), which represents the effects of variable surface temperatures on the algorithms. The clear-sky reference maps for the BTDanom algorithm were comprised of the observations around those 2 days. The column water vapor was shown with the MOD05_L2 of the same area on the two dates (Figures 9e and 9f).
 Figures 10a–10d show the frequency of the six indices on these 2 days. Table 3 presents their mean and SD for the September images: DRBTDI8.6 has a smaller SD (0.18 K) when compared to BTD8.6-11 (0.32 K) and BTDanom8.6,11 (0.30 K); likewise, the SD of DRBTDI12 (0.07 K) is far smaller than those of BTD12-11 (0.18 K) and BTDanom12,11 (0.19 K). These results are similar to the results described in section 5.2.1. The values of both DRBTDI8.6 and DRBTDI12 are concentrated approximately 0, with means of 0.10 and 0.02, respectively. Although the mean value of BTDanom8.6,11 (0.05 K) is also close to 0, the BTDanom12,11 values diverge more from 0, with a mean value of 0.20 K. The BTD8.6-11 and BTD12-11 values show significant differences in January compared to September. With the decrease in surface temperature, the mean value of BTD8.6-11 increases by approximately 3.10 K (from −7.83 K to −4.73 K), while BTD12-11 increases by 0.64 K (from 0.57 K to 1.21 K). In the evaluation of the application of the BTDanom method to the image from 4 January 2007, the BTDanom8.6,11 and the BTDanom12,11 were calculated from the clear-sky reference maps on 28 September 2008 and 4 January 2007, respectively. When the clear-sky reference map from September is employed, the mean value of BTDanom8.6,11 shifts by 3.1 K from January (3.15 K) to September (0.05 K), while for BTDanom12,11, the value shifts 0.64 K from the mean value of 0.20 K in January to 0.84 K in September. When the clear-sky reference map on 4 January 2007 is used, the BTDanom8.6,11 and BTD12,11 show better consistency with the results in September, with differences of a mean value of 0.67 K and 0.33 K, respectively. This result suggests that the clear-sky reference BTD map produced by the BTDanom method in one period should not be applied to another period due to the change in the observed temperature. However, it may be difficult to generate the clear-sky reference BTD for some regions or seasons with abundant presentation of clouds in a short period, which will introduce uncertainties in dust detection for this method. For the DRBTD algorithm, the results in January are consistent with those in September, with differences in mean values of 0.25 K for DRBTDI8.6 and 0.03 K for DRBTDI12. Likewise, the SD values are also the smallest (0.17 K and 0.06 K) when compared to those of the BTD and BTDanom. Additionally, the atmospheric water vapor in the study area on 28 September 2008 (approximately 1.2 cm) is higher than that on 4 January 2007 (approximately 0.3 cm). With this differences in atmospheric water vapor amount and surface temperature, the results of the DRBTD algorithm show less variation on the two dates, indicating that the DRBTD algorithm is more effective in eliminating the effects of temperature and low atmospheric water vapor on the BTD.
Table 3. Statistic of the BTD, BTDanom, and DRBTDI in Taklamakan Desert on 28 September 2008 (DOY272) and 4 January 2007 (DOY4)a
aBTDanom8.6,11_Ref2007004 and BTDanom8.6,11_Ref2008272 refer to the BTDanom8.6,11 with reference BTD of DOY4 in 2007 and DOY272 in 2008, respectively. Similarly, BTDanom12,11_Ref2007004 and BTDanom12,11_Ref2008272 are the BTDanom12,11 with reference BTD of DOY4 in 2007 and DOY272 in 2008.
5.3 Dust Detection and Comparison With OMI/MODIS/CALIPSO Observations
 The DRBTD algorithm was applied to two dust events that occurred during the daytime (case 1, 22 April 2007) and the nighttime (case 2, 12 April 2007) over the Taklimakan Desert, respectively. The nighttime results were compared with the CALIOP AOD products, and the daytime results were compared with the OMI AI and MODIS Deep Blue AOD products.
5.3.1 Daytime Case: 22 April 2007
 A heavy dust storm swept over the Taklimakan Desert on 22 April 2007. The MODIS aboard NASA's Aqua satellite captured an image of the thick dust plume between 37°N and 42°N on that day (Figure 11a). Dust was being generated around the periphery of the Taklimakan Desert basin. The dust plume was also observed in central Gansu Province (37°N–40°N, 100°E–105°E). The DRBTD method was applied to the Aqua/MODIS data and those pixels with DRBTDI8.6 or DRBTDI12 less than −0.5 K were labeled as clouds. The DRBTDI, OMI AI, and MODIS AOD images are shown in Figures 11b–11d. All three of the images showed the two thick dust plumes, with values larger than 3 for DRBTDI, 2.5 for OMI AI, and 2 for MODIS AOD, while the DRBTDI image was able to detect the detailed spatial pattern due to its fine resolution of 1 km (OMI AI and MODIS AOD have resolutions of 1.0º and 10 km, respectively). Most cloud pixels could be identified by negative DRBTDI values, whereas DRBTDI values of slightly greater than zero denoted the clear-sky nondusty pixels.
 Figure 12 shows the corresponding BTD8.6-11, BTD12-11, BTDanom8.6,11, BTDanom12,11, DRBTDI8.6, and DRBTDI12 images. In heavy airborne dust conditions with AOD greater than 2 according to the MODIS deep blue AOD map (Figure 11d), the values of BTD12-11, BTDanom12,11, and DRBTDI12 are larger than 1.5 K. For those clear-sky pixels with AOD less than 0.5 and cloud conditions, their values are approximately equal to or less than zero. The BTD8.6-11, BTDanom8.6,11, and DRBTDI8.6 showed ambiguous relationships for dust and clouds, with positive values for most cloud and dust pixels; however, the desert surfaces in eastern Taklimakan Desert are found to have smaller values than those of dust. Therefore, the combination of observations of the 8.6, 11, and 12 µm bands is helpful in the separation of dust signals from cloud and clear-sky pixels over desert surfaces.
 The images produced by the BTD, BTDanom, and DRBTD methods were resampled to 1.0º and 10 km by averaging the valid values in each grid and then comparing them with OMI AI and MODIS AOD, respectively. The pixels with AI less than 2 or DRBTDI8.6 or DRBTDI12 less than −0.5 K were excluded to eliminate the effects of clouds. Figure 13 presents the scatterplots of all seven indices compared to OMI AI, and Figure 14 presents the comparison to MODIS AOD. All seven indices show positive values of correlation coefficients r with OMI AI and MODIS AOD, indicating that these indices could capture the signal of airborne dust with various magnitudes of correlation. For the 8.6 µm band, the plots are dispersed for BTD8.6-11 and OMI AI with an r of 0.52, while it shows better correlation with MODIS AOD with an r of 0.77. For BTDanom8.6,11, the values of the index increase with OMI AI and MODIS AOD, with the r approximately 0.70. The DRBTDI8.6 shows good correlation with OMI AI (MODIS AOD) with an r of 0.70 (0.75). For the 12 µm band, the plots are dispersed and no significant relationships are observed for BTD12-11, BTDanom12,11, and DRBTDI12, with an r of approximately 0.65 (0.55) for OMI AI (MODIS AOD). Generally, the indices based on the 8.6 µm band show better correlations than that of the 12 µm band. It may be because the 8.6 µm channel helps to discriminate airborne dust over such a bright surface. Once DRBTDI8.6 and DRBTDI12 are combined, the plots of DRBTDI show good correlations with OMI AI and MODIS AOD, with the values of r being 0.79 and 0.77, respectively.
5.3.2 Nighttime Case: 12 April 2007
 Another heavy dust storm swept over the Taklimakan Desert at night on 12 April 2007. The CALIOP observed the dust plume. The blue line in Figure 15 shows the CALIPSO track at 2038–2124 UTC, approximately 45 s before Aqua/MODIS passed overhead. Figures 16a–16d present the CALIOP observations and the BTD, BTDanom, and DRBTDI results along the track. The column optical depth for aerosol and clouds at 532 nm shows the thick dust layer located between 36.2°N and 42°N with an AOD of over 0.5 (Figure. 16a). Several dust plumes were also present between 42°N and 50°N. Clouds were observed at approximately 36°N, 40.2°N, 45°N to 48°N, and 50°N to 52°N with a cloud column optical depth of over 0.5. Generally, the contrast between BT8.6 and BT11 shows the opposite response of that of BT12 and BT11. For the BTD method, BTD8.6-11 is positive for clouds and negative for dust, while BTD12-11 shows opposite signals (Figure 16b). BTDanom8.6,11 is positive and BTDanom12,11 is negative for clouds (Figure 16c). However, it appears difficult to use the BTDanom method to detect airborne dust against the desert background during night hours, as BTDanom8.6,11 and BTDanom12,11 are approximately zero; this result was also demonstrated by Ashpole and Washington , who suggested that nighttime pristine sky spectral characteristics are similar to that of dust, and dust effects on brightness temperatures decrease at nighttime. The DRBTD method could track the dust plume along the CALIPSO track, especially over the Taklimakan Desert, with a DRBTDI of approximately 2 K (Figure 16d). The presence of clouds was inferred where DRBTDI12 is negative, at approximately 40.2°N, 45°N to 48°N and 50°N to 52°N.
 The results from the BTD, BTDanom, and DRBTD algorithms were compared with the CALIOP AOD observations along the track in scatterplots in Figure 17. The pixels covering the center of each CALIOP measurement were selected for this comparison. Pixels with CALIOP column optical depth for clouds at 532 nm larger than 0.15 or AOD = 0 or DRBTDI8.6 < −0.5 K or DRBTDI12 < −0.5 K were excluded. For the 8.6 µm band, BTD8.6-11 and BTDanom8.6,11 exhibit worse correlation with negative relationships to CALIOP AOD, with the dots being greatly dispersive. The spectral characteristics of land surface and airborne dust at nighttime differ from those during daytime due to many factors, such as the different surface temperature and atmosphere temperature profiles. Thus, the relationships between the indices and AOD may differ from that in daytime. The DRBTDI8.6 shows better correlation with CALIOP AOD with an r of 0.80. For the 12 µm band, none of the three indices are well correlated to CALIOP AOD. No significant correlation is found for BTD12-11, which has an r of 0.60, as observed in Figure 17d. The absolute values of r for BTDanom12,11 and DRBTDI12 are less than 0.5. This poor correlation may be because that the 12 µm band is not very efficient in the discrimination of airborne dust over such a bright surface. With the combination of DRBTDI8.6 and DRBTDI12, the DRBTDI results are better correlated with CALIOP dust AOD observations than the BTD and BTDanom results are, with an r of 0.78. Overall, although the absolute values of r are above 0.5 for most indices, the dots for BTD, BTDanom, and DRBTD are greatly dispersive in comparison with daytime results, indicating a worse correlation with CALIOP AOD at nighttime, especially for the 12 µm band. This result may be attributed to the reduced thermal contrast of the dust with the surface during nighttime due to such factors as the subsidence of the dust layer induced by surface cooling and a lower lapse rate [e.g., King et al., 1999; Kluser and Schepanski, 2009]. Additionally, the provisional quality stage of Version 3.01 CALIOP AOD products may also introduce uncertainties. It is still challenging to discriminate airborne dust during nighttime.
5.4 Tracking Airborne Dust Movement
 One example was presented here to show the ability of the DRBTD algorithm to track airborne dust movement. The Taklimakan Desert and the Gobi Desert are the main source regions for Asian dust [Shao and Dong, 2006]. Dust aerosols that originate near these two regions are generally transported by westerly jets over eastern Asia and the Pacific Ocean and may even reach North America [Huang et al., 2008]. The transportation of dust from the Gobi and Taklimakan Deserts across Northeastern Asia on 6–7 April 2001 was tracked using our algorithm and presented in this section.
 Two large dust events occurred in northern China and southern Mongolia on 6 April 2001 (DOY096), including the dust storms generated in the Gobi Desert and the Taklimakan Desert. Figure 18a depicts these dust events using a color composite image from the Terra/MODIS on that day. In this image, the dust event was observed to be composed of two dust plumes. The dust storm whipped up sand in the Gobi Desert and blew eastward along the border of Inner Mongolia and Mongolia. At the same time, another dust plume originated in the Taklimakan desert and blew across the Hexi Corridor (37°N–42°N, 93°E–102°E) from its east gate and then reached the middle Gansu Province and Ningxia Province. The DRBTD algorithm was applied to the Terra/MODIS observations, and those pixels with DRBTDI8.6 or DRBTDI12 less than −0.5 K were labeled as clouds. Figure 18c shows the MODIS DRBTDI image that corresponds to Figure 18a. Two dust plumes with DRBTDI greater than 2 K were detected by MODIS DRBTDI, while the clouds in the north and southeast parts of the image were detected by DRBTDI12 or DRBTDI8.6 as values less than −0.5 K. Figure 18d presents the MODIS DRBTDI image at night on 6 April 2001. The dust plumes in Mongolia and northern China curl downwind towards the northeast to the Loess Plateau and east Inner Mongolia province, while the airborne dust in the Taklimakan Desert subsided.
 On 7 April 2001 (DOY097), the airborne dust was enhanced and transported to northeastern China by the developing synoptic-scale cyclones. Figure 18b is the color composite image from the Terra/MODIS sensor. The beige dust swirls in a clockwise direction around a hub of white clouds. The light brown-colored dust can be distinguished easily from the brighter, white clouds in this scene. The thick dust plume appears to be blowing from the Gobi Desert in Inner Mongolia and the Loess Plateau and then across northeast and north China on its eastward trek. MODIS DRBTDI could separate the airborne dust from the clouds and land surface (Figure 18e). The dust appears thickest in the center of the dust plume, which is indicated by the DRBTDI greater than 5 K. In addition, the suspended dust in northern and western Taklimakan Desert was detected. At night on 7 April, the dust aerosol in northeast China was transported northeastward to the Heilongjiang Province and the dust plume in northern China was transported over the Bohai Bay (Figure 18f).
 The performance of the DRBTD algorithm may be affected by several factors, including cloud presence, atmospheric water vapor, atmospheric temperature profile, change in surface conditions, and viewing geometry. Generally, a negative value of DRBTDI12 implies clouds. In the dust event cases of 22 April 2007 and 6–7 April 2001, most clouds were distinguished by DRBTDI12 values of less than −0.5. However, clouds are very complicated with varied properties of composition, form, temperature, and altitude. The characteristics of DRBTDI8.6 and DRBTDI12 for thin cirrus coupling with the barren land surface led to difficulty in distinguishing them from airborne dust. For example, for thin cirrus over barren or salt shores of lakes, the DRBTDI12 value is approximately zero and DRBTDI8.6 has a relatively large positive value, which results in large DRBTDI values and leads to misclassification of thin cirrus as airborne dust. In Figures 18c and 18e in section 5.4, the DRBTDI of thin cirrus over the numerous lakes in the Tibet Plateau is approximately 0.5. For very thick dust, DRBTDI12 would decrease to less than zero and DRBTDI8.6 would be large so this type of dust would be misidentified as clouds.
 The atmospheric water vapor has a significant effect on the detection of airborne dust. Atmospheric absorption by water vapor may mask the dust signal in the thermal longwave infrared channels [Chaboureau et al., 2007]. The atmospheric radiative transfer simulations and experiments have demonstrated that the column water vapor has a great influence on the thermal infrared measurements used for dust detection, and it is suggested that a thick dust plume with an AOD in 550 nm up to 0.8 may even be masked in SEVIRI dust products if the column water vapor exceeds 20–25 mm [Brindley et al., 2012]. The simulation in section 3 also suggests that the relationships between the BTs may be different for various atmospheric water vapor profiles, especially for those between BT12 and BT11. There are two adverse effects of water vapor on the DRBTD algorithm. In the regression procedure, the nondusty observations with high humidity will be below those with low humidity, which would lead the regression line to deviate from the ideal pristine line. And, the high atmospheric water vapor may misidentify dust observation as a cloud during the detection procedure due to the large negative DRBTDI12 in this situation. Therefore, large amounts of atmospheric water vapor may introduce high uncertainties in the performance of the DRBTD algorithm, especially in those regions (e.g., western Sahara) where dust events are strongly associated with deep convection and high levels of atmospheric moisture [Marsham et al., 2008]. Currently, the most clear-sky line regression approach is more appropriate to dry regions with low water vapor content in the atmosphere.
 Atmospheric temperature profiles would also affect the performance of the algorithm, especially when there is the presence of temperature inversion. When a near-surface temperature inversion is present, the presence of a dust layer may cause opposite BTD signals compared to that in normal atmospheric temperature situations, resulting in difficulties in dust detection at night [Ashpole and Washington, 2012]. However, this temperature inversion usually occurs on windless clear-sky nights, and severe dust events rarely happen in this windless situation [Chen et al., 2012].
 Surface change is another source of uncertainty. In our algorithm, the clear-sky relationships between the BTs were generated by assuming that the land surface is stable throughout the processing period. Here, all MODIS observations for the period from February 2000 to December 2008 were used to generate reference BT relationships. If the composition or physical form of the surface had changed significantly during this period, the emissivity might change and introduce uncertainties in the reference BT relationships. For example, dew, ice accretion, the seasonal variation of vegetation in semiarid regions, and the seasonal variation of lake shores would result in changes in surface emissivity and, consequently, in the clear-sky BT reference relationships. For arid regions, the land surface is relatively stable and should introduce little uncertainty. To assemble sufficient numbers of observations to regress the envelope line for establishing the clear-sky BT relationship, the DRBTD method requires long-term observation as input. It is only appropriate in regions with little change in surface conditions, such as deserts, evergreen forests, and oceans. For semiarid regions, the regression envelope line is apt to detect these clear-sky pixels in dry seasons and with less vegetation. Because most dust events occur over arid or semiarid regions in dry seasons, which coincide with the status of the BT relationship establishment, the uncertainty in these regions should be lower.
 The performance of the algorithm may be affected by the geometry of the satellite observations. Due to limited satellite observations for a specific site for each day, the MODIS observations with high viewing zenith angles were not subsampled here. Thus, these observations may be subject to limb darkening which will alter the magnitude of the BTD signals observed.
 The BTDanom method uses the reference BTD map from 15 day rolling clear-sky observations, which ensures that most of the land surface is stable because the period is notably short. Conversely, the DRBTD method requires long-term stability for the underlying surface. Although BTD is dependent on the observed temperature, which impedes the determination of thresholds for automatic dust detection, the fact that low temperature enhances the contrast between dust and the underlying surface amplifies the dust signal, which is favorable to the visual interpretation of dust from RGB composite images [Schepanski et al., 2007]. Many studies have attempted to retrieve the aerosol optical thickness from thermal infrared bands of satellite data quantitatively [e.g., Legrand et al., 2001; Brindley and Russell, 2006, 2009; Merchant et al., 2006; De-Souza Machado et al., 2006; Li et al., 2007; De Paepe and Dewitte, 2009; Kluser et al., 2011]. Currently, the DRBTDI only describes the dust aerosol thickness semiquantitatively. Conduct more simulations with consideration of various real atmospheric and surface conditions is required to build the quantitative relationship between dust aerosol thickness and the DRBTDI. However, this study provides a new path to remove the contribution of the underlying surface to the satellite data, which is a key step in the quantitative retrieval of atmospheric components.
 An algorithm was developed to detect airborne dust during both daytime and nighttime using satellite thermal infrared imagery. The DRBTD dust detection algorithm uses dynamic reference brightness temperature differences to eliminate the effects of temperature based on the significant linear relationships among the BTs at thermal infrared bands. The clear-sky relationships of brightness temperature between the 12 and 11 µm bands, as well as between the 8.6 and 11 µm wavelength channels, were mapped pixel by pixel from MODIS observations during the period of 2000–2008. Based on these relationships, the reference BTD12-11 and the reference BTD8.6-11 were dynamically generated pixel by pixel with the observed BT11. The observed BTDs were subsequently compared with the reference BTDs to separate dust from clear-sky surface and clouds.
 The algorithm was applied to identify the dust events at both daytime and nighttime over Mongolia and northwestern and northern China using MODIS data. The results were compared with OMI AI, MODIS Deep Blue AOD, and CALIOP AOD products. The cases examined in this paper show that the DRBTD algorithm effectively distinguishes mineral dust from clouds and land surfaces. During the daytime, with the combination of DRBTDI8.6 and DRBTDI12, DRBTDI corresponded with OMI AI and MODIS AOD with an r of 0.79 and 0.77, respectively. At night, DRBTDI correlated with the CALIOP dust AOD with an r of 0.78. Several clear-sky cases were also examined to evaluate the method's performance regarding the elimination of varying underlying surface effects. The results suggest that the DRBTD algorithm exhibits better spatial homogeneity and a mean value much closer to zero than do the BTD and BTDanom methods, with a mean value of 0.32 K (0.28 K) and a SD of 0.41 K (0.16 K) being observed for DRBTDI8.6 (DRBTDI12). These results indicate that the DRBTD algorithm effectively eliminates the effects of spatial variable emissivity. Additionally, the results in January are consistent with those from September with a difference in mean value of 0.25 K for DRBTDI8.6 and 0.03 K for DRBTDI12, indicating that the DRBTD algorithm also eliminates the effects of temperature. The algorithm can be used to track the transportation of airborne dust. Because the algorithm is based on thermal infrared data, the algorithm provides a dust indicator both during the daytime and during the nighttime, which is extremely valuable for global dust transport monitoring and for assessing diurnal cycles of dust emission.
 This research was funded by the China 973 program (2010CB950701), the National Natural Science Foundation of China (41171285), and the China Postdoctoral Science Foundation (2012 M510343). MODIS data were obtained from the Land Processes Distributed Active Archive Centre. The OMI AI data used in this effort were acquired as part of the activities of the Science Mission Directorate of NASA and are archived and distributed by the Goddard Earth Sciences (GES) Data and Information Services Center (DISC). The CALIPSO data were obtained from the NASA Langley Research Center Atmospheric Science Data Center. The authors also would like to thank the three anonymous reviewers for their insightful and critical comments, which improved this paper.