SEARCH

SEARCH BY CITATION

Keywords:

  • Asian dust retrieval;
  • AOT;
  • dust height;
  • AIRS hyperspectral sounder

Abstract

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Construction of Training Data Sets
  5. 3 ANN Model Development
  6. 4 Validation
  7. 5 Case Studies
  8. 6 Summary and Conclusions
  9. Acknowledgments
  10. References

[1] In order to examine potential use of infrared (IR) hyperspectral measurements for dust monitoring, a statistical artificial neural network (ANN) approach was taken as an inverse method of retrieving pixel-level aerosol optical thickness (AOT) and dust height (zdust). The ANN model was trained by relating Atmospheric Infrared Sounder (AIRS) brightness temperatures across 234 channels, surface elevation, and relative air mass to collocated AOT derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) and zdust derived from Cloud Aerosol Lidar Infrared Pathfinder Satellite Observation (CALIPSO) observations for Asian dust cases. Results showing correlation coefficients of 0.84 and 0.79, and mean biases of 0.03 and about −0.02 km for AOT and zdust, respectively, suggest that dust retrievals from hyperspectral IR sounder measurements are comparable to MODIS-derived AOT and CALIPSO-measured zdust. The pixel-level retrievals of AOT and zdust during both day and night from IR hyperspectral measurements may offer great potential to improve our ability to monitor and forecast the evolving features of Asian dust.

1 Introduction

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Construction of Training Data Sets
  5. 3 ANN Model Development
  6. 4 Validation
  7. 5 Case Studies
  8. 6 Summary and Conclusions
  9. Acknowledgments
  10. References

[2] There has been growing concern about mineral dust originating in China and Mongolia (referred to as Asian dusts) because such dusts are transported to densely populated areas in East Asia and can have a significant negative influence on human activity. Occurrences of dust events have become more frequent in Korea and Japan in recent years [Kurosaki and Mikami, 2003; Kim, 2008], likely associated with an expansion of the arid area over eastern Mongolia [Park and Sohn, 2010; Lee and Sohn, 2011]. In order to reduce expected negative impacts on health and socioeconomic factors using an early warning system, accurate monitoring of the evolution of Asian dust after its outbreak is essential.

[3] Satellite observations provide an efficient tool to monitor the features of dust because of their extensive areal coverage and ability to quantitatively measure dust's optical properties. Techniques utilizing ultraviolet (UV) and visible (VIS) measurements have been successfully developed [e.g., Torres et al., 1998; Wald et al., 1998; Hsu et al., 2000; Sinyuk et al., 2003; Hsu et al., 2004; Remer et al., 2005], albeit only for daytime retrieval. Increasing daytime temporal coverage seems to improve dust-monitoring performance, and it is assumed that it would be beneficial if nighttime dust estimates were available. However, with no UV/VIS light available at night, increasing temporal coverage, especially during nighttime, should rely on passive infrared (IR) or active light sources.

[4] Some successes in detecting dust from satellite-based IR measurements have been reported [Ackerman, 1989; Ackerman, 1997; Legrand et al., 2001; Sokolik, 2002; Li et al., 2007]. For instance, the brightness temperature (BT) difference between thermal split window channels has long been used as an index representing the magnitude of dust loading [Ackerman, 1997; Legrand et al., 2001; Darmenov and Sokolik, 2005; Brindley and Russel, 2006; Merchant et al., 2006], although its use for the quantitative dust monitoring has been limited [e.g., Sohn et al., 2013]. Furthermore, quantitative physical parameters could be more beneficial than an ambiguous dust index if measurements were possible during the nighttime. Then, using the spectrally varying extinction efficiencies from the Mie calculation for the Asian dust [e.g., Han et al., 2012], it may be possible to convert IR aerosol optical thicknesses (AOTs) into more familiar VIS AOTs, if IR AOTs can be retrieved from IR measurements during the nighttime.

[5] The use of hyperspectral IR measurements may provide such retrieval capability, since thousands of channels across the IR spectrum implicitly carry information on IR emission from surface and atmospheric gases and on radiative interactions with dusts. Indeed, recent studies have demonstrated that high-spectral-resolution IR sounders (e.g., the Atmospheric Infrared Sounder (AIRS) and the Infrared Atmospheric Sounding Interferometer (IASI)) are able to retrieve dust height and IR AOT [Pierangelo et al., 2004; DeSouza-Machado et al., 2010; Peyridieu et al., 2010; Yao et al., 2012]. Despite the successful use of hyperspectral IR measurements to retrieve dust properties, previous efforts have been limited to monthly retrieval over oceans [Pierangelo et al., 2004; Peyridieu et al., 2010] or to case studies [DeSouza-Machado et al., 2010; Yao et al., 2012]. Thus, there is a need for more in-depth studies of general retrieval accuracy or the possibility of using this information for dust monitoring or forecasting.

[6] Recognizing the potential use of hyperspectral IR measurements for retrieving VIS-like AOT and dust height regardless of the measurement time of day, we examine the degree of accuracy that can be achieved for AOT and dust height retrievals from hyperspectral IR measurements. In this study, AIRS hyperspectral IR BTs are related to Moderate Resolution Imaging Spectroradiometer (MODIS) AOTs and Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) dust height to study their pixel-level accuracies and potential applicability to dust monitoring. Using measured hyperspectral IR BTs, and based on the assumption that atmospheric and surface conditions can be described by those BTs, the relationship between hyperspectral BTs and aerosol parameters can be learned through artificial neural network (ANN) learning processes. If successful, the obtained temporal coverage of the dust and its heights at both daytime and nighttime may provide valuable information for better monitoring of the evolutionary features of dust aerosols.

2 Construction of Training Data Sets

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Construction of Training Data Sets
  5. 3 ANN Model Development
  6. 4 Validation
  7. 5 Case Studies
  8. 6 Summary and Conclusions
  9. Acknowledgments
  10. References

[7] We have adopted the ANN method as it enables learning about an arbitrary function through a training process, which can produce a model describing a nonlinear relationship [Krasnopolsky and Schiller, 2003; Mas and Flores, 2008]. For training of the ANN model, a set of collocated data is necessary to relate predictand (here, AOT or dust height) to predictor variables (here, AIRS BTs, surface topography, and relative air mass). The recent NASA A-Train mission [Stephens et al., 2002] provides a good means to construct these necessary collocated data sets, from which the ANN training can be accomplished.

[8] The AIRS, onboard the Aqua satellite as part of the A-Train, measures BTs on 2378 channels covering the IR spectral ranges 3.8–4.6 µm, 6.2–8.2 µm, and 8.8–15.4 µm [Chahine et al., 2006]. Although measurements are available for both day and night, only the longer wavelength spectral measurements over the 6.2–15.4 µm band are used to construct the collocated data, excluding the solar-affected band (3.8–4.6 µm) during the day time. The exclusion of shortwave band (3.8–4.6 µm) is also partly because dust height information should be linked to emission signals from the dust layer and scattering influence on the shortwave band should not be negligible. AIRS channels covering the wavelength range of 6.2–15.4 µm are roughly divided into four bands: window (WIN; 8.15–12.8 µm), CO2 (12.9–15.0 µm), water vapor (WV; 6.3–8.0 µm), and O3 (9.2–10.1 µm). Detailed information on those IR band absorptions can be found in Liou [2002]. Considering only channels with lower noise equivalent differential temperatures (NEΔT) and removing weak gaseous absorption channels over the WIN band, a total of 234 channels (130, 31, 44, and 29 channels for the WIN, CO2, WV, and O3 bands, respectively) are selected for the ANN model development. Spectral locations of the selected channels are shown in Figure 1. It is worthy of pointing out that the continuum absorption by water vapor in particular near the surface can give influences on the WIN channel BTs between conventional absorption bands. Although the water vapor continuum effect is not explicitly counted in this paper, it is expected to be implicitly explained through mapping 130 WIN channel TBs to the target parameter.

image

Figure 1. Selected AIRS 234 channels used for the ANN model development in this study. Dots in green, red, yellow, and blue represent selected channels over CO2, WIN, O3, and WV bands, respectively.

Download figure to PowerPoint

[9] The MODIS sensor onboard the same Aqua satellite measures radiances at 36 spectral bands spanning the VIS-IR region (0.415–14.235 µm) with a spatial resolution of 250 m, 500 m, or 1 km at nadir, depending on the wavelength. The MODIS Level 2 aerosol product provides aerosol properties with a spatial resolution of about 10 km based on VIS measurements [Remer et al., 2005]. In this study, AOTs based on Hsu et al. [2004] are collocated with AIRS BTs over land because the deep blue algorithm (using channels at wavelengths shorter than 0.5 µm) is known to more efficiently detect dust aerosols over bright arid and semiarid regions. Over the ocean, AOTs based on a look-up table approach [Levy et al., 2003] are collocated with AIRS BT data.

[10] The CALIPSO satellite, also part of the A-Train, has a two-wavelength (532 and 1064 nm) lidar instrument (the Cloud-Aerosol LIdar with Orthogonal Polarization; CALIOP) that offers a cross-sectional view of aerosol and thin cloud along the satellite track [Winker et al., 2007]. The CALIPSO Lidar Level 2 aerosol layer top and base heights and AOT at 532 nm are collected to specify vertical distribution of dust layers and to construct a training data set for the ANN model. In order to identify dust-laden pixels, data flagged as dust or polluted dust aerosols of the CALIPSO aerosol product are used for collocation with AIRS BTs. Total attenuated backscatter profiles at 532 nm, measured by the CALIOP, are used for validating AIRS-derived height for case studies, which will be discussed in section 5. Table 1 summarizes the data sets used in this study.

Table 1. Description of Data Sets Used in This Study
ProductSpatial ResolutionParametersDescription
AIRS Level 1B13.5 km at nadirRadiancesRadiances for 234 AIRS channels
TopogSurface topography
SatzenSatellite zenith angle
MODIS Level 210 kmOptical Depth Land and OceanAOT over ocean
Deep Blue Aerosol Optical Depth 550AOT over land
CALIPSO Level 25 kmLayer Top AltitudeDust top height
Layer Base AltitudeDust base height
Feature Optical Depth 532Aerosol optical thickness at 532 nm
Feature Classification FlagScene classification flag
CALIPSO Level 1B1/3 kmTotal Attenuated Backscatter 532Total attenuated backscatter at 532 nm

[11] Some might concern about influences of uncertainties in MODIS AOTs on the ANN retrievals. To reduce those uncertainty influences, information on best quality assurances for MODIS AOT over ocean and CALIPSO scene classification flags are used. However, MODIS AOTs over the land (from “deep-blue” algorithm) and CALIPSO dust altitude are kept as they are, since quality control information is not available.

[12] In addition, land surface topography (zsfc) and relative air mass from the surface to the sensor (1/μ; μ = cosθ where θ is the satellite zenith angle) are included as inputs to the model. The land surface topography and satellite zenith angle are also obtained from AIRS Level 1B product. Brightness temperature differences between 11 µm and 12 µm (BTDs) are also included as an input to examine the sensitivity of the ANN model to the given inputs. For the BTD calculation, the difference is taken between BTs averaged over the 11.8–12.2 µm and 10.8–11.2 µm bands.

Construction of AIRS Versus MODIS Collocated Data

[13] For a given dust loading in an atmospheric column, the spectral variation of optical thickness can be theoretically expressed by spectrally varying extinction efficiencies that can be obtained from the Mie calculation for given dust properties (see Liou [2002] for the general relationship between optical thickness and extinction efficiency). That is,

  • display math(1)

where Qext,λ1 and Qext,λ2 represent extinction efficiencies at wavelengths λ1 and λ2, respectively. When λ1 and λ2 are assumed to be the VIS and IR bands, respectively, IR-based AOT can be converted to VIS-based AOT using equation (1). When dusts are a dominant aerosol species, equation (1) can be accepted as a first-order approximation for converting IR-based AOT at one band to VIS-based AOT at a different band. However, in this study, instead of explicitly converting BTs to IR-based AOTs, and then to VIS-based AOTs through equation (1), IR BTs are implicitly related to VIS-based AOTs (here MODIS-derived AOTs) via a training process of the ANN model. It should be noted that equation (1) is valid only if loaded aerosols are also sensitive to the IR spectra.

[14] In training, the predictand (or target output) in the AOT retrieval model is defined as the AOT equivalent to that retrieved from MODIS at 550 nm, such that AIRS BTs can be implicitly related to AOTs at 550 nm. Because of AIRS target sizes (13.5 km at nadir) are similar to 10 km resolution of MODIS AOT products, the MODIS AOT whose target location is closest to the AIRS field of view (FOV) is chosen for collocation. Training data for AOT are constructed for a period from February to May 2007 over the East Asian region (15–55°N and 70–150°E), resulting in enough samples because of frequent dust outbreaks during the spring time. In this procedure, only dust-laden AIRS pixels are retained after applying the AIRS-based dust detection method by DeSouza-Machado et al. [2010], which is referred to as dust-detection algorithm (DDA). The DDA is based on the BTD approach, but with five AIRS channels in the window spectral bands (i.e., 8.12, 8.86, 10.40, 11.11, and 12.6 µm) which were chosen after accounting low instrumental noise and IR dust absorption and avoiding water absorption bands.

Construction of Collocated CALIPSO Versus AIRS Data

[15] For a given atmospheric column, dust is determined to be present if the CALIPSO feature classification flag indicates dust or polluted dust. Then, vertical extent of dust layer is determined from CALIPSO-derived top and bottom altitudes of single or multiple dust layers. For the ANN training model for dust height, all dust layer heights within one CALIPSO pixel are weighted by AOTs of the corresponding layers to obtain a single dust height at a given location, i.e.,

  • display math(2)

where z(i) and AOT(i) are mean height above sea level and AOT at the ith dust layer, respectively, and N is the total number of dust layers from CALIPSO products. Thus, in equation (2), AOT(total) is a simple summation of AOTs in N individual dust layers. Because of the weighted mean, zdust can be considered to be “effective dust height,” but we keep “dust height” in this study for the simplicity.

[16] Due to different spatial resolutions (i.e., 13.5 km versus 5 km at nadir for AIRS and CALIPSO, respectively), multiple CALIPSO pixels can be within one AIRS FOV. If at least one of the CALIPSO pixels is determined to be dust laden, all CALIPSO aerosol measurements within one AIRS FOV are averaged to construct the one-to-one matching data. To take an average of the CALIPSO pixels within an AIRS FOV, the dust heights are weighted by their total column AOT in the same manner as in equation (2). The collocation is performed over the same geographical domain used for MODIS AOT versus BT collocation However, the same training period of 4 months used for the AOT retrieval is not enough for construction of sufficient zdust matchup data because of along-tracking CALIPSO measurement. To obtain the enough sample size for the training, 2-year period (January 2007 to December 2008) is applied.

3 ANN Model Development

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Construction of Training Data Sets
  5. 3 ANN Model Development
  6. 4 Validation
  7. 5 Case Studies
  8. 6 Summary and Conclusions
  9. Acknowledgments
  10. References

[17] Remotely sensed IR radiances from satellites carry information related to atmospheric parameters, but have a strong nonlinearity. For example, the top-of-atmosphere (TOA) IR radiance is contributed by atmospheric emission expressed by combination of emissivity with layer mean temperature and by surface emission. Thus, it should be worthwhile to examine to what extent dust information may be gleaned from hyperspectral measurements, using a tool that can explain the nonlinear relationship between dust loading and IR radiances.

[18] The ANN provides a convenient and powerful means of performing nonlinear classification and regression by partitioning the input space into sub-domains, which may then be associated with a desired set of output values [Freeman and Skapura, 1991]. We use a so-called multilayer perceptron (MLP) ANN model, which allows a feed-forward network to link either MODIS AOT at 550 nm or CALIOP zdust with AIRS BTs over the WIN, CO2, WV, O3 bands, and with zsfc, 1/μ, and BTD. The multilayer feed-forward ANN model used in this study consists of three layers—an input layer, a hidden layer, and an output layer—that are connected by synaptic weights. The model has a single hidden layer with five neurons, and a hyperbolic tangent sigmoidal function is used to link their nonlinear responses. More detailed explanation of training procedures is found in Lee and Sohn [2012].

[19] Using the collocated data sets, ANN models are developed by first examining how well the ANN model performs with given input parameters. Although weights of channels derived by ANN approach can be directly used for relating all AIRS BTs to dust properties, we intend to examine physical reasoning of why certain inputs perform better. In doing so, various input parameters are related to a total of 46,073 collocated MODIS AOTs, and take sensitivity tests demonstrating the model performance. The tests are conducted by adding parameters in an orderly manner (i.e., BTs over the WIN band first, followed by BTs over the CO2, WV, and O3 bands, and zsfc, 1/μ, and BTD last, in a cumulative fashion). The number of hidden neurons was kept fixed throughout the experiments regardless of different input layers.

[20] The obtained correlation coefficient and root mean square error (RMSE) for each set of input parameters are shown in Figure 2a for AOT. It is noted that training results based only on WIN band BTs shows good agreement with MODIS AOT retrievals, with a correlation coefficient larger than 0.86, demonstrating that hyperspectral WIN channel measurements carry rich information about dust AOTs. With more channels from other bands (i.e., CO2, WV, and O3) added for training, statistics steadily improve with each addition. zsfc and 1/μ also exert a positive influence on the correlation, resulting in a highest correlation coefficient of around 0.90 and lowest RMSE of 0.34 when combined with BTs over various bands. However, poorer statistics are noted for BTD, probably because BTD is an averaged WIN band that contains information that has already been used in 130 WIN channels.

image

Figure 2. Correlation coefficients and RMSEs (a) between AIRS-derived and MODIS-estimated AOTs and (b) between AIRS-derived and CALIPO-measured dust height. Solid and dashed lines are for correlation and RMSE, respectively. WIN, CO2, WV, and O3 on the abscissa represent inputs of AIRS BTs at window, CO2, water vapor, O3 bands, respectively. zsfc: land surface topography, 1/μ: relative air mass from the surface to the sensor, BTD: brightness temperature difference between 11 µm and 12 µm. Inputs are arranged in a cumulative fashion.

Download figure to PowerPoint

[21] Using 50,976 pairs of collocated data, the same procedures used for AOT are applied to examine how sensitive zdust retrieval is to given input parameters. Statistics obtained from these sensitivity tests are given in Figure 2b. A correlation coefficient of 0.68 for only the WIN channel BTs suggests that zdust is also most sensitive to WIN channels, although the correlation is weaker than that in the AOT test. It is noted that a substantial variance is explained by adding CO2 channel BTs, as shown in correlation coefficient increased to about 0.74 and reduced RMSE from about 0.94 km to 0.86 km. The substantial improvement in statistics resulting from the addition of the CO2 channels suggests that the synergistic use of WIN channels with CO2 channels can improve zdust retrieval from hyperspectral sounder measurements. This may be because information about temperature profiles from CO2 channels can be added to dust emissivity information, likely from WIN channels, resulting in better description of TOA radiances.

[22] In contrast, other input parameters such as WV, O3, 1/μ, and BTD appear to have little impact on zdust retrieval, despite slightly improving statistics with addition of further input parameters. Again, the addition of BTD tends to detract from the overall performance of zdust retrieval.

[23] Based on the results of the sensitivity tests, it can be concluded that WIN channels are most important for both AOT and zdust retrieval. However, adding CO2 channels results in substantial improvements in zdust retrieval. Nevertheless, in this study, we use all parameters except BTD in the development of ANN models for both AOT and zdust because of the tendency for resultant statistics to improve with the addition of parameters.

4 Validation

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Construction of Training Data Sets
  5. 3 ANN Model Development
  6. 4 Validation
  7. 5 Case Studies
  8. 6 Summary and Conclusions
  9. Acknowledgments
  10. References

[24] In order to validate the performance of the retrieval models discussed above, we compare AIRS pixel-based AOT and zdust retrievals with MODIS AOT and CALIPSO zdust estimates. For the validation of AOT retrieval from AIRS measurements, AIRS AOT retrievals during the period from February to May 2008 are compared with collocated MODIS-derived AOTs. For the dust height validation, AIRS zdust retrievals over a 1-year period (January to December 2009) are compared with collocated CALIPSO measurements. Both comparisons are conducted for the same geographical domain used for constructing the training data set.

[25] Comparison results from a total of 57,384 collocated pairs of AOTs are given in a two-dimensional density histogram (Figure 3a). It is indicated that AIRS AOTs obtained using the ANN method are in good agreement with MODIS AOTs, with a correlation coefficient of 0.84, mean bias of 0.03, and RMSE of 0.37. The obtained statistics suggest that MODIS-derived dust AOTs at 550 nm can be fairly accurately estimated from hyperspectral IR measurements, probably through the relationship between wavelengths and extinction efficiencies expressed in equation (1). In other words, IR-based dust AOTs can be obtained quite accurately from hyperspectral IR sounder measurements and can, therefore, be related to the VIS AOTs through equation (1).

image

Figure 3. Two-dimensional histograms of (a) MODIS-derived versus ANN-retrieved AOTs for February 2008 to May 2008 and (b) CALIPSO-derived zdust versus ANN-retrieved zdust for January 2009 to December 2009 in a 0.1 bin size.

Download figure to PowerPoint

[26] For the validation of zdust retrievals, AIRS-retrieved zdust data at a pixel level are collocated with CALIPSO measurements over a 1-year period. Validation results based on a total of 20,272 AIRS dust pixels are presented in Figure 3b. It is shown that the retrieved zdust is in agreement with CALIPSO estimates, with a correlation coefficient of 0.79, mean bias of about −0.02 km, and RMSE of 0.54 km. It is of interest to note that a large portion of dust loadings are found below 3 km in weighted height. Note that the AIRS FOV along the CALIPSO track is about 13.5 km at the nadir, resulting in less than half of the pixel area being filled with CALIPSO pixels, which are not necessarily all dust pixels. Thus, it can be expected that better retrievals may be possible when exact collocated data sets with complete dust pixels are employed.

5 Case Studies

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Construction of Training Data Sets
  5. 3 ANN Model Development
  6. 4 Validation
  7. 5 Case Studies
  8. 6 Summary and Conclusions
  9. Acknowledgments
  10. References

[27] In addition to the validation of the ANN models using global statistics, it is also of interest to examine how ANN retrievals are compared against MODIS and CALIPSO measurements on a two-dimensional basis. In this attempt, we select two cases of dust outbreak that occurred on 17 and 23 April 2009 over the Taklimakan and Inner Mongolian Gobi deserts, respectively. It should be noted here that the ANN model is retrained for zdust without 1/μ. This is because the CALIPSO targets are within a zenith angle of 10–19° from the AIRS nadir view; thus, AIRS targets outside this range were not defined. Note that the following AIRS-derived zdust data (in Figures 5 and 7) are retrieved from BTs on 234 channels and zsfc.

The 17 April 2009 Case

[28] Figure 4 shows the MODIS true color image, AIRS-retrieved AOT, and MODIS AOT distribution for the Asian dust outbreak that occurred on 17 April 2009 over the Taklimakan desert area. In the true color image (Figure 4a), a bright yellowish dust area is clearly visible in the north of the Tibetan Plateau, which is located in the southern part of the image and is covered with broken clouds above its darker surface. AIRS pixel-based AOT retrievals (Figure 4b) show the presence of dust extending from southwest to northeast, with maximum AOTs up to 4.0 centered over the area around 39°N and 87°E (an eastern part of the Taklimakan desert). General patterns and magnitudes shown in AIRS retrievals are quite similar to those found in MODIS AOTs based on the deep blue algorithm, although maximum areas and magnitudes are rather large. The MODIS deep blue algorithm detected weak AOTs north of about 43.5°N, where a clear area is seen on the MODIS true color image. Since CALIPSO measurements (Figure 5b) indicate that no dust is present north of 43.5°N (around an eastern tail of the Tien Shan mountain range), low AOT areas diagnosed by the deep blue algorithm are not likely to be dust related. A correlation coefficient of 0.86 obtained from collocated pixels between AIRS-derived AOTs and MODIS AOTs (not shown here) again indicates that the pixel-based results are consistent with results obtained from the validation of the ANN model (Figure 3).

image

Figure 4. (a) MODIS true color image, (b) AOT retrieved from AIRS measurements, and (c) MODIS-derived AOT on 17 April 2009 over the Taklimakan desert area.

Download figure to PowerPoint

image

Figure 5. (a) Two-dimensional distributions of dust heights retrieved from AIRS measurements and (b) vertical view of CALIOP-measured total attenuated backscatter at 532 nm on 17 April 2009 over the Taklimakan desert area, along the CALIPSO track (red line in Figure 5a). The backscatter in unit of km−1 sr−1 is multiplied by 1000, and values smaller than 5.0 × 10−3 km−1 sr−1 are masked out in Figure 5b. The blue cross mark and black solid line in Figure 5b indicate AIRS-retrieved zdust and surface topography, respectively.

Download figure to PowerPoint

[29] The two-dimensional distribution of the AIRS-derived zdust data for this case is shown in Figure 5a, where the red solid line indicates the CALIPSO along-track direction from south to north. A vertical view of CALIOP total attenuated backscatter at 532 nm is given in Figure 5b, along with collocated zdust (+) retrieved from AIRS measurements. Considering that the backscatters are included only if larger than 5.0 × 10−3 km−1 sr−1, the close agreement shown in Figure 5b suggests that AIRS retrieval well represents the altitude of the dense dust layer. It is of interest to note that the lower zdust (<3 km) areas in Figure 5a roughly correspond to the lower desert basin area lower than 1000 m above sea level.

[30] While Figure 5b shows a general agreement between AIRS zdust and the CALIPSO total attenuated backscatter, difference is clear around 42.5–43°N latitude band, where a single strong dust layer appears to split into multiple layers. For the ANN zdust training model, all CALIPSO dust layers at a given pixel location have been weighted with their respective layer AOTs to convert into a single AOT-weighted height. Consequently, for multiple dust layers, difference may become larger. In addition to retrieval errors, difference between AIRS zdust and the CALIOP backscatter in Figure 5b is at least in part due to the different definition of dust height, in particular associated with multiple dust layers.

The 23 April 2009 Case

[31] The AOT and zdust for another dust outbreak, which occurred on 23 April 2009 over the Mongolian Gobi desert, are retrieved from AIRS measurements and compared with MODIS AOT and CALIOP zdust (Figures 6 and 7). In the MODIS true color image (Figure 6a), dust features are represented by a bright yellow color over the northwest of Shanxi Province (located in the southeast corner of the domain), likely following a low-pressure system causing a vortex-shaped cloud over the Gobi desert. In the AIRS retrieval (Figure 6b), the high AOT area over 42°N, 109°E appears to be in the vicinity of an area of vortex-shaped cloud. In this retrieval, because cloud screening is not done separately but by applying the algorithm of DeSouza-Machado et al. [2010] for determining the presence of dust, some cloud areas remain and are determined to be areas of dust; this results in high dust AOT in spite of the presence of cloud near the vortex center. It is interesting to observe a narrow dust band on the MODIS true color image, along the narrow zone off the Qinghai high-terrain area (located in the southwest quarter of the MODIS image domain). The narrow dust band, probably moving from the eastern Taklimakan desert, is captured well by both AIRS and MODIS AOT. Again, at the pixel level, a high correlation coefficient of 0.88 between the two retrievals indicates that the ANN-based AIRS dust AOT values are comparable to those from the MODIS deep blue algorithm.

image

Figure 6. Same as Figure 4, except for the dust outbreak on 23 April 2009 over the Gobi desert.

Download figure to PowerPoint

image

Figure 7. Same as Figure 5, except for the dust outbreak on 23 April 2009 over the Gobi desert.

Download figure to PowerPoint

[32] Two-dimensional distributions of the AIRS-derived zdust for this case are shown in Figure 7a, and a comparison is made using the vertical profiles of CALIOP-measured total attenuated backscatter along the track from south to north in Figure 7b (expressed as a red line in the Figure 7a). It is shown that CALIPSO-measured height distribution along the track is generally captured well by AIRS retrievals. However, overestimation is apparent over the high terrain area around 38°N. Although it is not clear whether the complex terrain features give rise to this overestimation, it is not surprising to note that remote sensing over the high complex terrain area has been always problematic. A closer look suggests that the high attenuated backscatter region south of 37°N appears not to be related to dust presence, but to cloud (see the MODIS true color image). The dust layer is found to be substantially diffused vertically in comparison to the rather compact dust layer in the 17 April case (Figure 5b). By nature, because AIRS zdust represents the layer mean AOT-weighted height, retrieval results may be more different to CALIOP measurements if the total dust layer is deeper or more diffused.

[33] The case studies have demonstrated that AIRS retrievals agree well with MODIS-derived AOTs and CALIPSO-measured total attenuated backscatters over the major dust source regions, where dominant mineral dust aerosols are expected. However, away from the source regions, dust particles may undergo changes in optical properties because dust particles tend to be mixed with other aerosol species during the transport [e.g. Sohn et al. 2007]. Over the ocean, dust particles are mixed with sea-salt particles whose sizes are generally larger than dust particles, inducing potential problems of separating dust particles from mixed forms. Caution should be placed on interpreting the IR-based dust retrievals in particular when the dust aerosol is mixed with other types of aerosols.

6 Summary and Conclusions

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Construction of Training Data Sets
  5. 3 ANN Model Development
  6. 4 Validation
  7. 5 Case Studies
  8. 6 Summary and Conclusions
  9. Acknowledgments
  10. References

[34] An ANN algorithm retrieving AOT and zdust of Asian dust from AIRS hyperspectral IR measurements has been developed. In developing the algorithm, two training data sets were prepared: one for dust AOT based on collocated AIRS BTs (130, 31, 44, and 29 channels for WIN, CO2, WV, and O3 bands, respectively) and MODIS AOTs at 550 nm, and another for zdust based on AIRS BTs on the same channels with CALIPSO zdust. Values of zsfc and 1/μ associated with satellite-viewing geometry were also added to the collocated data sets. From the various sensitivity tests relating AIRS BTs to either AOT or zdust in the ANN model, it was found that WIN channel BTs in the hyperspectral sounder measurements (here, AIRS measurements) are far more important in predicting either AOT or zdust. It is because emitted radiation from the surface under dust conditions should interact with dust layers, involving absorption and re-emission of IR radiation by the dust layer. Thus, surface temperature, surface emissivity, and IR emission from the dust layer should be important contributors to the TOA radiance, and these are best resolved by near-transparent window channels. On the other hand, substantial contributions can be made by BTs in CO2 channels to zdust retrievals. Based on sensitivity results, ANN models for AOT and zdust retrievals were established with BTs using a total of 234 channels, zsfc, and 1/μ as inputs.

[35] To assess the performance of the developed ANN model for AOT retrieval, AOTs retrieved from AIRS measurements were compared with MODIS-derived AOTs over a period different to that of the training dataset. The pixel-level AOT comparison shows a correlation coefficient of 0.84, mean bias of 0.03, and RMSE of 0.37, suggesting that the ANN-derived AOTs appear to be comparable to MODIS AOTs. The comparison of pixel-level zdust retrieved from AIRS measurements also showed good agreement with the collocated CALIPSO zdust, with a correlation coefficient of 0.79, mean bias of about −0.02 km, and RMSE of 0.54 km.

[36] Two-dimensional dust retrievals for two dust events over Chinese desert regions show good agreement with collocated MODIS AOT and CALIPSO zdust, demonstrating that the developed algorithm can be applied to dust monitoring. Considering the fact that retrievals are available regardless of the timing of AIRS measurements, the use of hyperspectral sounder measurements has great potential to improve our ability to monitor or forecast evolutionary dust features. Further expansion to other sensors such as IASI and the Cross-track Infrared Sounder (CrIS) of the National Polar-orbiting Operational Environmental Satellite System (NPOESS) Preparatory Project (NPP), and to other dust source regions such as the Sahara and Arabian deserts, is anticipated.

[37] It should be noted, however, that this is a pilot study demonstrating capability of hyperspectral IR measurements to estimate dust properties (here AOT and dust height). In fact, for any other practical use of this method such as dust monitoring in different seasons and different geographical regions, more specific trainings reflecting temporal and regional characteristics should be preferred.

Acknowledgments

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Construction of Training Data Sets
  5. 3 ANN Model Development
  6. 4 Validation
  7. 5 Case Studies
  8. 6 Summary and Conclusions
  9. Acknowledgments
  10. References

[38] The authors would like to thank three anonymous reviewers for their constructive and valuable comments, which led to an improved paper. This work was funded by the Korea Meteorological Administration Research and Development Program under Grant CATER 2012–2092.

References

  1. Top of page
  2. Abstract
  3. 1 Introduction
  4. 2 Construction of Training Data Sets
  5. 3 ANN Model Development
  6. 4 Validation
  7. 5 Case Studies
  8. 6 Summary and Conclusions
  9. Acknowledgments
  10. References