Inferring dust composition from wavelength-dependent absorption in Aerosol Robotic Network (AERONET) data



[1] Atmospheric mineral dust is an important component of the Earth system, affecting climate and biogeochemical cycling. We present an inverse method of inferring dust composition from wavelength dependence of light absorption and the complex refractive indices of aerosol components. Specifically, we separate absorption by black carbon from absorption by hematite and organic carbon. We apply this method to coarse-mode-dominated Aerosol Robotic Network (AERONET) observations in the “dust belt” region of the globe, and we identify differences between dust optical properties and composition in different locations. We solve for dust composition using two opposite, bracketing hypotheses of the state of mineralogical mixing: purely internal mixing and purely external mixing. We find that calculated absolute hematite concentrations are highly sensitive to mixing assumptions, while relative geographic and seasonal patterns are less sensitive to mixing assumptions. Internal mixing calculations appear to underestimate absolute hematite concentrations, while external mixing overestimates hematite concentrations. Inversions calculated assuming external mixing are better able to explain the wavelength dependence of dust absorption by only varying hematite concentration than inversions using internal mixing, which require substantial covariation between black carbon and hematite to match observations. Saharan and East Asian dust show higher hematite content than dust from Arabia. Saharan dust also shows seasonal variation in hematite content which may reflect seasonal shifts in dust source areas.

1. Introduction

[2] The climatic impact of aerosols remains a large uncertainty, due to their heterogeneity in space and time and their relatively poorly known optical properties. The role of weakly absorbing aerosols such as dust is particularly uncertain and may have either a warming or cooling effect on climate [Intergovernmental Panel on Climate Change, 2001]. The climatic impact of dust is primarily defined by its single scattering albedo (SSA), the ratio of scattering to extinction by the aerosol [Liou, 2002; Miller and Tegen, 1998]. The SSA of an aerosol is a function of particle size, shape, and composition. Single-particle analysis of dust has shown that individual particles are complex aggregates of smaller mineral grains, and that the composition of dust particles may vary from one source region to another [Reid et al., 2003].

[3] While dust is primarily composed of clay minerals and other silicates [Falkovich et al., 2001], most of these are very weakly absorbing at visible light wavelengths. Of the dust components that are derived from soil, hematite is strongly absorbing at wavelengths below 600 nm [Gillespie and Lindberg, 1992; Krekov, 1992; Sokolik and Toon, 1999]. Pollution aerosols mixed in with dust also absorb in the visible: organic carbon (OC) from biomass burning is absorbing at wavelengths below 600 nm [Kirchstetter et al., 2004], while black carbon (BC) from industrial burning is absorbing throughout the visible spectrum [Kirchstetter et al., 2004; Krekov, 1992].

[4] In this study, we focus on the hematite fraction of dust, and demonstrate a new method to infer its concentration from Sun photometer data. Specifically, we attempt to differentiate between absorption by hematite and black carbon (BC) in aerosols by utilizing observed profiles of absorption strength with respect to wavelength of light. The iron oxide concentration of dust is also important to determine because hematite and other iron oxides compose the bulk of iron in soils [Schwertmann, 1993] and therefore likely dominate dust as well. As dust is a major source of iron to remote, iron-limited ocean waters [Bishop et al., 2002], geographic variability in dust composition may have impacts on ocean biogeochemistry [Fung et al., 2000].

[5] The composition of dust has been derived from chemical analysis of samples collected at various locations and times [Falkovich et al., 2001; Reid et al., 2003; VanCuren and Cahill, 2002]. In this study, we build on these studies and examine whether variation in dust composition, in particular in the hematite fraction, is significant enough to effect dust optical properties in a systematic way. We apply a Gauss-Markov inversion method to derive a representative space-time climatology of dust compositions, using the set of aerosol observations made by the Aerosol Robotic Network (AERONET) of Sun photometers [Dubovik et al., 2002a].

[6] Remote sensing observations of dust SSA [e.g., Kaufman et al., 2001; Tanré et al., 2001] have shown that dust SSA is higher than had previously been thought, based on in situ measurements and dust composition [e.g., World Meteorological Organisation, 1983]. We hypothesize that this high SSA may be related to the state of mixing of dust, and therefore attempt our inversion using two sets of assumptions that bracket the state of dust mixing: internal mixing, in which individual dust particles are a mix of all components present; and external mixing, in which different components exist as separate particles. As is discussed below, an internal mixture assumes that the hematite is broken up into small units, which results in more effective absorption of radiation.

[7] In section 2, we describe the methods used in our inversions. In section 3, we compare the results of the two bracketing mixing hypotheses. We describe the geographic and seasonal variation of dust composition that we calculate from our inversions in section 4.

2. Methods

[8] We attempt to solve the inverse problem of estimating particle composition from AERONET refractive index and SSA measurements. The AERONET data are themselves inversion results solved for by inverting scattering and radiative transfer models constrained by angular measurements of light scattered by the atmosphere. We take as known properties the complex refractive indices of a limited set of materials, which we refer to as end-members, which make up dust. We then use a least squares inversion method to solve for the proportions of these components that best fit the AERONET observations.

2.1. AERONET Observations

[9] AERONET is a network of 150 Sun photometers distributed around the globe. The Sun photometers are Cimel CE 318N, and measure at 4 wavelengths of 440, 673, 873, and 1022 nm. The Sun photometers are highly calibrated to allow for comparison between observations at different locations and times, and data are automatically and manually checked to remove errors resulting from clouds [Holben et al., 1998]. The AERONET algorithm uses measurements of direct transmission and diffuse reflection at multiple scattering angles to determine optical properties of the scattering aerosol [Dubovik et al., 2002b; Dubovik and King, 2000; Dubovik et al., 2000]. This algorithm simultaneously solves for the real and imaginary refractive indices and size distribution that best match the observed aerosol angular scattering phase function. All the observations we use were made prior to 2004 and were processed using the version 1 AERONET algorithm.

2.2. Optical Properties of End-Members

[10] Table 1 lists real and imaginary refractive indices for major components of dust at the four AERONET wavelengths, and Figure 1 shows imaginary refractive indices over the visible and near-IR spectrum. Note that Figure 1 has a logarithmic scale for imaginary refractive indices, and that only hematite, BC, and OC have appreciable imaginary refractive indices in this region of the spectrum. The silicate minerals that compose the bulk of dust mass have relatively uniformly low imaginary refractive indices in the visible part of the spectrum, all of which are less than or equal to 10−4.

Figure 1.

Imaginary refractive indices of dust components, from Table 1. Vertical grey lines represent the four principal AERONET wavelengths.

Table 1. Complex Refractive Indices of Dust Components at Four AERONET Wavelengths as Noted in the Literature
Hematite3. [1992]
Hematite2.802.702.552.510.920.0860.0590.041Sokolik and Toon [1999]
Hematite0.640.042Gillespie and Lindberg [1992]
Soot (BC)1.811.851.871.910.740.720.690.68Krekov [1992]
BC0.710.73Kirchstetter et al. [2004]
Soot (BC)1.561.571.581.610.490.480.490.53Nilsson [1979]
Smoke (OC)0.0720.00310.0010.001Kirchstetter et al. [2004]
Kaolinite1. and Toon [1999]
Illite1. and Toon [1999]
Quartz1.541.531.531.520.00050.00050.00050.0005Krekov [1992]
Silicate1.481.481.471.460.000240.000290.000360.00043Krekov [1992]
Sulfate1.431.431.431.4410−610−710−810−8Palmer and Williams [1975]
Water1.341.331.331.322 × 10−93 × 10−84 × 10−71 × 10−6Krekov [1992]
Water1.341.331.331.3310−92 × 10−83 × 10−73 × 10−6Hale and Querry [1973]

[11] Three major uncertainties are inherent in any method to derive dust concentration and composition from a mixture of aerosols. These relate to (1) the difficulty in separating hematite from other wavelength-dependent absorbers, particularly anthropogenic OC which, while an order of magnitude weaker at absorbing than hematite, has a similar spectral profile; (2) uncertainty in the absorption profile of anthropogenic BC related to variability of combustion sources [Bond, 2001]; and (3) lack of knowledge about the mixing state of the aerosol. We address the first two uncertainties by limiting our analysis to those AERONET observations that are highly dominated by coarse-mode aerosols; this limits the influence of combustion aerosols, which we expect to reside primarily in the fine mode. We address the last uncertainty by bracketing the range of expected mixing states using opposite assumptions about aerosol mixing, and discussing the implications of the different assumptions on our data analysis.

[12] In our simplified model of dust optical properties, we treat dust as a mixture of nonabsorbing silicate, hematite, and BC. Multiple different spectra exist in the literature for hematite [e.g., Gillespie and Lindberg, 1992; Krekov, 1992; Sokolik and Toon, 1999]. For our inversion we use the values corresponding to the E ray data reported by Sokolik and Toon [1999]; this spectrum shows slightly less wavelength sensitivity to absorption than the O ray data of Sokolik and Toon [1999]. We do not use the spectrum of Gillespie and Lindberg [1992] because it only covers the two shorter-wavelength AERONET bands; nor do we use the spectrum of Krekov [1992], which shows similar wavelength sensitivity but an order of magnitude less absorption across wavelengths than the other two studies.

[13] To separate absorption by BC, we use a spectrum of BC complex refractive indices reported by Krekov [1992]. This spectrum agrees well with the BC spectrum of Kirchstetter et al. [2004] but covers the entire AERONET wavelength range. As discussed above, we cannot explicitly account for absorption by OC, because the OC absorption spectrum is similar to that of hematite, given the spectral resolution (4 wavelength bands) of the AERONET instruments. Instead we try to minimize the effect of OC (and BC) on our analysis by applying a particle size threshold on the observations we use; we discuss the effectiveness of this below by looking at the seasonal cycle of our results in mixed dust-biomass burning regions.

[14] We define the third compositional end-member, which makes up the bulk of the dust in the inversion, to be a relatively nonabsorbing silicate with real refractive index of 1.5 and imaginary refractive index of 10−4 at all four wavelengths. The choice of silicate imaginary refractive index, which we varied between 10−3 and 10−4, the range of clay particles in Figure 1, has a small effect on the calculated mean BC concentration, but little effect on the calculated hematite values.

[15] Dust particles are highly complex, aspherical, nonuniform aggregates of different minerals. Nonetheless, to model dust optical properties, we must make simplifying assumptions. To investigate the effect of mixing state on our results, we attempt to bracket the possible range of mixing states by calculating two sets of dust compositions using opposite mixing models. We first hypothesize that individual particles are composed of an internal mixture of the three constituents: a matrix of silicate and randomly distributed spherical inclusions of hematite and soot. In this hypothesis, we calculate an effective bulk complex refractive index using the Bruggeman effective medium approximation [Bohren and Huffman, 1983] and compare this with the refractive indices calculated from AERONET observations. Our second, contrasting hypothesis is that dust is composed of a pure external mixture of three components, i.e., each particle is purely one constituent, and we calculate total scattering and absorption by varying the proportions of particles of each type.

[16] Aerosols at any given site are a heterogeneous mixture of the fundamental types of aerosols: industrial pollution, dust, smoke, etc. To extract information about the hematite fraction in mineral dust, we first select observations in regions where mineral dust is known to predominate the aerosol loading; these correspond to sites adjacent to the “dust belt” described in detail by Prospero et al. [2002]. We only consider observations of aerosol optical thickness greater than 0.4 at 440 nm, because the accuracy of AERONET observations degrades when the aerosol optical thickness is less than 0.4 [Dubovik et al., 2000].

[17] To minimize the impact of pollution aerosols (BC+OC) on our calculation, we also utilize the size distributions provided by the AERONET inversion. Desert dust dominates the coarse-mode aerosol size distribution, while pollution aerosols dominate the fine mode [Dubovik et al., 2002a]; therefore we only consider AERONET observations where the ratio of coarse-mode aerosol to fine-mode aerosol is above a threshold value.

[18] To select this size ratio threshold, we examined the relationship of equation imagei,obs, the imaginary refractive index averaged over the four AERONET wavelengths, with coarse-mode to fine-mode concentration ratio for sites in East Asia, where dust from western China mixes with industrial BC (Figure 2). Our a priori expectation is that BC concentration should increase with increasing fine-mode aerosol concentration, and therefore the mean imaginary refractive index should also increase with increasing fine-mode particle content. This is what we observe, and we find that absorption stabilizes somewhat above a critical coarse to fine ratio of 7. As a result, we limit our analysis to observations with coarse-mode to fine-mode concentration ratios greater than a threshold value of 7. This skews our analysis away from a “typical” dust aerosol and toward a “clean” dust aerosol. This also limits the number of AERONET sites with enough observations meeting our criteria to the ones discussed below.

Figure 2.

Mean imaginary refractive index versus ratio of coarse-mode to fine-mode aerosol for East Asian AERONET stations. Fine mode is dominated by strongly absorbing soot. Dust optical properties appear to dominate above coarse to fine ratio of 7 (dashed line), which we use as threshold for dust-dominated aerosol. Note also that coarse-dominated, less absorbing aerosols are more often observed during the dusty spring months (red crosses), while highly pollution dominated aerosol are more often observed outside the Asian dust season (blue diamonds).

[19] The climatology of BC in aerosols has been investigated by Schuster et al. [2005], who use absorption measurements in AERONET data to determine the amount of BC present in aerosol. Our approach is similar, in that we are inferring composition of aerosols based on their scattering and absorption properties. Our approach is also similar in concept to that of Fialho et al. [2005], who used wavelength-dependent absorption in aethelometer data to decouple absorption effects of black carbon and desert dust in the Azores. However, they do not explicitly incorporate mineralogical effects into their model of dust optical properties, rather they fit the wavelength dependence of absorption by the two aerosol types to exponential curves, and fit the aethelometer measurements to a combination of these exponential curves.

2.3. Internal Mixtures

[20] The optical properties of a homogeneous internal mixture of components can be approximated using the Bruggeman effective medium approximation [Bohren and Huffman, 1983], generalized to an n component system:

equation image

where fj is the volumetric fraction of component j, ɛj is the complex dielectric function of component j, and ɛeff is the complex dielectric function of the mixture. The dielectric function of a material is a function of wavelength, and is the square of the material's complex refractive index n:

equation image
equation image

[21] The Bruggeman effective medium approximation assumes that all components of the aerosol are randomly internally mixed. This assumption is likely to be violated in most cases for the various dust components; nonetheless we use this as one end-member of possible dust mixing states, and also consider an opposite extreme below.

[22] Determining the fractions of different components in the dust mixture is an inverse problem: given observations of ɛeff and assumptions about ɛj of a discrete set of predefined, pure end-member components, we seek to determine fj. While it is possible to solve for fj in the above equation directly, doing so does not take full advantage of the information about aerosol absorption, because both the real and imaginary parts of the dielectric function are dependent on both the real and imaginary parts of the refractive index. Since the imaginary part of dust refractive index is typically orders of magnitude smaller than the real part of the refractive index, much of this information gets lost in the uncertainty of the calculation. In order to best utilize information on light absorption, we solve for a linear approximation of the above formulation, spanning the range of dust compositions commonly observed, for the imaginary part of the refractive index rather than the complex dielectric function. This linear approximation to the Bruggeman equation introduces a <1% error for the range of hematite concentrations we expect in our analysis based on our prior studies of dust composition.

[23] The real part of the refractive index, nr, has little spectral dependence for various possible components in dust, therefore it provides little constraint for our analysis. Furthermore, variation in the magnitude of real part of the refractive index could be due to multiple physical, rather than compositional properties of the dust particles, including hygroscopic growth, particle porosity, and variations in the concentrations of nonabsorbing minerals such as clays or quartz. Ignoring variations in the real part of the refractive index may lead us to underestimate the hematite content of dry dust in regions where considerable hygroscopic swelling has occurred, since we are calculating the volumetric concentration of hematite in the aerosol as it exists rather than as it would be dry. Nonetheless, because our results calculated considering both real and imaginary refractive indices are qualitatively similar to those calculated using only imaginary refractive indices, we present here the simplest method.

[24] The values of ni,eff are derived by AERONET. The AERONET inversion algorithm uses direct sun and reflected sky radiances at multiple wavelengths to simultaneously solve for aerosol optical depth, size distribution, and complex refractive index. The method involves inverted scattering and radiative transfer models, and solves for both spherical and spheroidal particles, therefore making the method relatively accurate even for nonspherical dust particles [Dubovik et al., 2002b; Dubovik and King, 2000]. AERONET has several locations located at the margins of the active dust-producing deserts of the world. We have examined data from all of these stations and analyzed those with more than a minimum number of valid observations. Because these criteria for valid observations are somewhat restrictive, we have very few data outside of the global dust belt identified by Prospero et al. [2002].

[25] We distinguish absorption by hematite from absorption by BC using the sharp dropoff in the hematite spectrum between 440 and 673 nm. For our internal mixture hypothesis, we use a simplified method that considers only two pieces of information: the mean imaginary refractive index at all four AERONET wavelengths, and the difference between the imaginary refractive indices at 440 and 673 nm wavelengths. We then use Gauss-Markov inversion on the following 3 × 3 matrix:

equation image

Here equation imagei,obs is the mean imaginary refractive index at four AERONET wavelengths, equation imagei,obs is the difference between imaginary refractive index at 440 and 673 nm, and equation imagei,comp and equation imagei,comp are the mean and the difference between 440 and 673 nm, respectively, for slopes of the linearized Bruggeman mixing lines for component comp. The final row of the matrix ensures that the fractions of all components sum to a total fraction of 1.

[26] To solve the above, we first multiply through a prior estimate of dust composition, which we take to be 96% silicate, 2D% Fe, and 2% BC. We choose these prior estimates to be consistent with laboratory observations of dust [e.g., Falkovich et al., 2001; Reid et al., 2003; VanCuren and Cahill, 2002]; however the results are only very loosely tied to the prior guesses, whose main function is to make unphysical inversions (i.e., with negative concentrations) less likely. We then subtract the expected observations based on this prior from our AERONET observations, and solve for the residual composition using Gauss-Markov inversion [Tarantola, 2005; Wunsch, 1996]:

equation image

where F is the vector of dust residual compositions, equation image is the linearized Bruggeman mixing coefficients matrix, Y is the observation matrix, and Rxx and Rnn are specified noise and prior covariance matrices. We assume that the intrinsic error in the AERONET imaginary refractive index measurements is 30%, as reported by Dubovik et al. [2000]. We assume that the errors in the four wavelengths are uncorrelated, resulting in an expected error based on Monte Carlo calculations of 1.4 times this 30% for the equation imagei,obs, and 0.7 times 30% for equation imagei,obs; this assumption of uncorrelated error is likely a best-case scenario for equation imagei,obs and a worst-case scenario for equation imagei,obs, but has little effect on the results. We then set the diagonal elements of the noise covariance matrix Rnn to the expected errors in the measurements. We allow the diagonal elements of Rxx, the prior covariance matrix, to be large in order to minimize the effect of our prior estimate on the outcome. Lastly, we add back our prior estimate to F to get the resulting best fit dust composition.

[27] For the internal mixture case, the Gauss-Markov approach is not strictly necessary because the problem as posed is not ill-defined; however, we choose this method to be consistent in methodology with the underdetermined external mixture case. We find that the solution is relatively invariant to our choice of priors or uncertainty, and <1% different from a result using linear least squares inversion; this is due to the fact that we have already transformed the data to highlight the aspects that give the most information on dust composition.

2.4. External Mixtures

[28] In their analysis of individual dust particles, Reid et al. [2003] observe, using single-particle scanning electron microscopy with energy dispersive analysis with X rays (EDAX), that much of the iron in dust is associated with individual, highly Fe-enriched particles, rather than forming a uniform background in all particles. We consider this as an alternate hypothesis to that of internal mixing by repeating our analysis making the assumption that hematite exists as a purely external mixture of particles that are either silicate, hematite, or BC.

[29] Making the assumption of an external mixture results in less absorption, and less wavelength dependence of absorption, than does the assumption of an internal mixture for small amounts of hematite. The basic reason for this is due to the extremely high imaginary refractive index for hematite at short wavelengths. In the absence of scattering, the intensity of radiation at a point inside a material is [Bohren and Huffman, 1983]

equation image
equation image

where I is the intensity of radiation with initial intensity I0 and wavelength λ at a distance x into the medium.

[30] For the case of hematite at 440 nm wavelength, the e-folding attenuation distance 1/λ is approximately 0.05 μm, which is smaller than the mean size of a dust particle. Therefore if hematite aggregates approach the size of a whole particle, then the centers of the particles are effectively not illuminated and do not interact the incoming radiation. This effect is described by Gillespie and Lindberg [1992], who note in their laboratory analysis is that hematite particles must be ground to submicron sizes in order for the complete absorption effects to be observed.

[31] Figure 3 shows forward calculation of SSA values for silicate-hematite mixtures using internal and external mixture assumptions as a function of wavelength and hematite fraction. We use as a size distribution for these calculations the mean size distribution of all our coarse-mode-dominated AERONET observations (Figure 4). Note that the aerosol SSA drops much more sharply at 440 nm for the internal mixture calculation, compared to the case with the external mixture. In the inverse problem of calculating hematite content from observations, this difference between the external and internal mixing assumptions leads to an order of magnitude lower calculated hematite concentration for a given amount of aerosol absorption.

Figure 3.

SSA at four AERONET wavelengths of silicate-hematite (E ray) mixtures as a function of the volumetric hematite fraction. Dashed lines are for internal mixture; solid lines are for external mixture. SSA computations assume spherical particle size distributions of Figure 4.

Figure 4.

Mean size distribution of AERONET coarse-mode-dominated observations, averaged from all observations used in this work.

[32] For our inversion of aerosol composition based on an external mixture assumption, we use the same Gauss-Markov method described above, but instead of solving for imaginary refractive indices, we solve for total scattering and absorption optical thicknesses at the four AERONET wavelengths:

equation image

where τsca and τabs are the scattering and absorption optical thickness, defined below by τext, the extinction optical thickness, and SSA, the single scattering albedo:

equation image
equation image

[33] To compute the elements of the coefficient matrix, we first calculate Mie scattering and absorption coefficients of each material for each size bin in the AERONET size distribution using the method of Bohren and Huffman [1983]. To prevent the fine ripple structure of the Mie calculation results from impacting our results, we take a center-weighted mean of several particle sizes within each of the 22 AERONET size bins. Then, for each AERONET observation, we calculate total optical thicknesses for each material by multiplying the Mie coefficients by a cross-sectional area computed from the observed volume concentration of each size bin and summing over all size bins.

[34] We assume spherical particles in the Mie scattering computations of scattering and absorption cross sections, rather than the spheroidal T matrix calculations used in the AERONET algorithm itself. This should not introduce significant error, however, as we are inverting for SSA, which is less influenced by particle nonsphericity than angular measurements [Mishchenko et al., 1997].

[35] The matrix is inverted, as for the internal mixture, using equation (4). For a priori error estimates, we use a 0.02 absolute uncertainty for SSA measurements, and 1% relative uncertainty for the extinction optical thickness measurements, as discussed by Dubovik et al. [2000], and we propagate these errors through to calculate the a priori errors for the scattering and absorption optical thicknesses.

3. External Versus Internal Mixture Results

[36] Figures 5 and 6 show calculated dust hematite and BC fractions, for internal and external mixing hypotheses, respectively. Two primary differences can be observed between the results calculated assuming these hypotheses. The first is that the amount of hematite and BC calculated using internal mixing hypothesis is <1% and is at least an order of magnitude lower than the concentrations calculated using external mixing. The second is that the ratio of hematite to BC calculated is generally <1 for the internal mixture, and >1 for the external mixture. The reason for these two effects is as follows: the internal mixing hypothesis leads to higher absorption for small amounts of hematite compared to the external mixture (see Figure 3); therefore, for a given absorption observation, smaller amounts of hematite and BC are required to reach that absorption. Because this effect is strongest at short wavelengths for hematite, the internal mixing hypothesis also overestimates the wavelength dependence of absorption by hematite relative to the external mixing hypothesis; therefore more BC is needed to compensate for the wavelength-independent absorption.

Figure 5.

Calculated hematite and soot values for AERONET stations using internal mixture assumption. (a) Hematite, summer-fall season. (b) Hematite, winter-spring season. (c) BC, summer-fall season. (d) BC, winter-spring season. Diamonds are means, bars are standard deviations, and boxes are median and quartiles. Above station names are number of observations used. Bands group together stations from similar geographic regions: west Sahara and Sahel, north Sahara/Mediterranean, Middle East, and East Asia.

Figure 6.

Same a Figure 5 but for external mixture hypothesis. Calculated hematite and soot values for AERONET stations using external mixture assumption. (a) Hematite, summer-fall season. (b) Hematite, winter-spring season. (c) BC, summer-fall season. (d) BC, winter-spring season. Diamonds are means, bars are standard deviations, and boxes are median and quartiles. Above station names are number of observations used. Vertical bands group together stations from similar geographic regions.

[37] To further illustrate the differences between internal and external mixing hypotheses, Figure 7 shows the relationship between mean SSA at the four AERONET wavelengths and the difference between the SSA at 440 and 673 nm. The two lines represent calculated ideal mixing curves for a pure silicate-hematite mixture. The upper line, representing an internal mixing curve, shows stronger wavelength dependence to absorption than the line for external mixing.

Figure 7.

Plot of the difference between the SSA measured at two shortest AERONET wavelengths versus the mean SSA averaged over all four AERONET wavelengths. Blue diamonds are observations during summer-fall months, while red crosses are observations during winter-spring months. The two lines represent the idealized mixing curves for a pure silicate-hematite dust using the two extreme mixing assumptions: the upper line is for internal mixing, while the lower line is for external mixing. (a) Data points are from AERONET stations recording Saharan dust. (b) Data points are from AERONET stations recording Middle Eastern dust. The tail of points falling below the main population for the Middle Eastern sites in Figure 7b are likely soot-impacted observations from the Bahrain site.

[38] The data points in Figure 7 are separated into two geographic regions: West Sahara and Sahel stations (Figure 7a), and Middle Eastern stations (Figure 7b), and are identified by seasonality. With the exception of the summertime Middle Eastern points, which are from the Bahrain site and appear to be BC impacted, all of the data points cluster much closer to the external hematite-silicate mixing line than they do to the internal line. This suggests that the external mixture better captures the wavelength dependence of absorption than does the internal mixture.

[39] Aside from the above quantitative differences between the internal and external mixing assumptions, the qualitative climatological patterns are relatively unchanged between the two sets of inversions. The two sets of inversions represent bounds on the amount of hematite: measured Fe content of dust is approximately 2–3% [Falkovich et al., 2001; Reid et al., 2003; VanCuren and Cahill, 2002], with the internal mixture an underestimate and the external mixture an overestimate. We discuss the results of the external inversion below, because it emphasizes the wavelength dependence of absorption by dust solely as a function of varying hematite content, and provides a baseline for examining mixtures with BC and other absorbers.

[40] Sato et al. [2003] also discuss the relative accuracies of internal vs. external mixtures for BC in the AERONET data, and find that an internal mixture is probably closer to the true situation than an external mixture. They base this on the observation that total light absorption by BC is higher than would be expected, given present estimates of total BC emissions, if BC was externally mixed with other pollution aerosols. We argue that, while this may be true for BC, hematite mixing with other minerals in dust appears to be closer to an external mixture, based on the wavelength dependence of absorption.

[41] The interpretation of dust as an external mixture of hematite and silicates is consistent with single-particle observations of iron concentrated in separate particles [Reid et al., 2003]. Furthermore, Bullard and White [2005] propose a mechanism by which relatively pure iron-rich particles could be produced from sand dunes: saltating sand grains scour each other's weathered outer rinds to produce iron oxide-rich particles. Our results support these observations, and suggest that much of the hematite in dust is occurring as separate, hematite-rich particles.

4. Climatology of Dust Composition

4.1. Geographic Variation

[42] For Figures 5 and 6, we use shaded bands to group observations regionally. These bands correspond to, from left to right: west and south Sahara, north Sahara, Middle East, and East Asian. These groupings allow us to discuss some of the broader regional differences in dust. Figure 8 shows, in map form, the geographic distribution of the median data values from Figure 6 (external mixing) to more clearly identify these regional differences. No results are reported when sites had fewer than three valid observations during either season. In addition to Figures 5, 6, and 8, Table 2 lists the mean values and standard deviations (external mixture hypothesis) for all observations aggregated over these geographic regions.

Figure 8.

Maps of calculated hematite and soot for coarse-mode-dominated aerosols at several AERONET sites, using external mixture assumption. (a) Map of mean calculated summer-fall hematite content. (b) Mean winter-spring hematite. (c) Mean summer-fall BC content. (d) Mean winter-spring content.

Table 2. Statistics for Geographic and Seasonal Groups of AERONET Observations
Sites and SeasonNumber of ObservationsMean Hematite, vol %SD HematiteMean BC, vol %SD BC
West and south Sahara summer2697.765.030.683.55
West and south Sahara winter1066.285.310.462.93
North Sahara summer4811.974.290.813.67
Middle East summer1284.102.702.404.88
Middle East winter1083.752.570.051.24
East Asia summer55.893.426.592.55
East Asia winter406.254.181.883.10

[43] Our external mixture hypothesis results show that the AERONET sites west and south of the Sahara see dust with relatively similar optical properties. The mean summertime (the primary dust season) hematite content of dust in this region, from Table 2, is 7.8%. We calculate higher hematite values in the Mediterranean sites north of the Sahara, with a mean of 12.0%. Middle Eastern AERONET stations show calculated hematite values much lower, with mean summertime values of 4.1%. East Asian sites show mean winter-spring dust hematite values of 6.3%, slightly below those in the main Saharan plume.

[44] We assess the statistical significance of the differences in mean hematite content using a Tukey test for the four regions during their dominant dust seasons. We find that the Middle Eastern stations are significantly different from both the Asian and Saharan dust stations at 95% confidence, but that the Saharan and Asian regions could not be definitively separated.

[45] Figures 6c, 6d, 8c, and 8d, the external mixture hypothesis BC results, show that the East and South Asian sites at all seasons and Bahrain during the summer show an appreciable BC influence. The ubiquity of BC in Asian dust is expected based on well-known characteristics of Asian aerosols [Chuang et al., 2003; Huebert et al., 2003]. That this is observable in our results even after size-filtering the data suggests that in these highly polluted sites, we are not able to completely remove the influence of these pollution aerosols on our inversion.

[46] A robust result from our analysis is that the Saharan and Sahelian sites show much more hematite than the Middle Eastern sites, particularly the Solar Village site. Claquin et al. [2003] predict high hematite values for dust generated along the southern margin of the Sahara, which is where Figure 9a shows the highest aerosol optical thicknesses, and therefore major dust source regions, to lie. A fundamental explanation for this could be that the Sahara and the Middle East have different climate and weathering histories, especially during the Holocene, and especially along the southern margin of the Sahara, which was much wetter during the recent past [Hoelzmann et al., 1998]. Since hematite is produced as a weathering product in soils, this may explain why Saharan dust may on the whole have high hematite values. The difference between dust along the main Saharan dust plume and dust advected northward over the Mediterranean may reflect different source characteristics between these two dust transport pathways.

Figure 9.

Mean aerosol optical thickness as observed by MISR instrument. (a) Mean annual aerosol optical thickness (AOT) over dust belt region. (b) Seasonal cycles for AOT over two major dust-producing regions in Sahara. Note that eastern, Bodele-centered source (red) has higher AOT during winter and spring months relative to western source region (blue), but this reverses in summer.

[47] The differing optical properties between Saharan and Middle Eastern dust has been described by Kubilay et al. [2003], who used back trajectories to compare dust arriving from different locations to a single AERONET site in the northeastern Mediterranean. In addition to attempting to quantify the mineralogical differences between the regions, we show that the climatology of the network supports this observation from a single site.

[48] The East Asian sites show high variability in hematite between stations, with the Dunhuang site having much lower calculated hematite content than the sites further to the east. As discussed above, all sites east of the Dunhuang station show high amounts of BC. Given the possibility for variation in BC optical properties [Bond, 2001], we have less confidence in the uniqueness of our hematite inversion results at these sites than in less polluted regions. However, it is possible that the variability here reflects a greater diversity of dust sources than is evident, for example, in the Sahel sites. This would follow from the highly diverse landscapes of the region, where several geologically distinct dust-producing regions have been identified [Bory et al., 2003; Sun et al., 2001].

[49] In addition to the broad regional trends discussed above, we observe some variability between sites within regions. This suggests that local differences in the mineralogy of dust-producing regions play a role in influencing aerosol optical properties. While the ultimate cause of this variation is unknown, likely factors include differences in the extent of soil weathering, sediment production rates, parent rock type, and geomorphic environments of the dust sources.

4.2. Seasonal Variation

[50] In Figures 5 and 6, we have separated our observations at each AERONET site into two seasons: summer-fall (June–November) and winter-spring (December–May). We do this for several reasons. First, we are interested in the effect of expected seasonal mixing of dust and anthropogenic pollution aerosol on our results. Second, this allows us to examine if other seasonal changes in dust optical properties are taking place. For sites west and south of the Sahara, we expect to see more of a pollution influence during the winter months during the West African biomass-burning season [Kaufman et al., 2005]. For Middle Eastern sites, we expect pollution to be more present during the summer, as noted by Kubilay et al. [2003]. For East Asian sites, the prime dust season is in the spring months (March–May) [Bory et al., 2003; Husar et al., 2001; VanCuren and Cahill, 2002] therefore we expect fewer clean dust observations outside of this period.

[51] Most of the sites that have enough AERONET observations to appear in both seasons have relatively similar hematite fractions between the two seasons. This is not to say that typical aerosol properties are invariant between different seasons at all of these AERONET stations, rather that our size-based filtering appears to be able to remove much of the mixing between mineral and anthropogenic aerosols. The relatively aseasonal hematite fraction argues for common summer-fall and winter-spring source region for most sites.

[52] One area where inferred dust hematite content does vary weakly with season is the region to the west and south of the Sahara. These stations show lower calculated hematite content during the mixed dust-biomass burning winter season (6.3% hematite) than during the cleaner summer dust season (7.8% hematite). Binning all AERONET observations for the western Saharan and Sahel regions for the two seasons results in statistically different mean hematite values for the separate seasons using a two-tailed t test at 95% confidence. As discussed above, it is not possible to separate the effects of absorption by OC from the effects of absorption by hematite using our inverse method because of the similar spectral absorption profiles of the two materials. However, if this seasonal variation were caused by light absorption by OC mixed with dust, we would expect to see the opposite seasonality, since OC is more dominant during the winter months.

[53] Another possibility for the seasonal variation in dust absorption in the Sahel is that dust source areas are different for winter and summer seasons, resulting in an inherent mineralogical difference in the absence of pollution aerosols. Figure 9a shows mean aerosol optical thickness observations from the MISR instrument [Diner et al., 1998] during the primary summertime dust season; Figure 9b shows the seasonal cycle of mean aerosol optical thickness in the two boxed regions. These two regions show highest mean summer aerosol optical thickness, suggesting that they are dominant centers of dust activity over the Sahara: one is centered over the Bodele Basin [Washington and Todd, 2005; Prospero et al., 2002], while the other is located further west. During the winter-spring months, Figure 9b shows that the western source area is weakened relative to the eastern, Bodele-centered source region, which peaks during the winter [Washington and Todd, 2005].

[54] The Bodele Basin has an extremely high albedo so it is likely that dust emitted from the region is also highly reflective and low in hematite. Therefore a possible explanation for the seasonal decrease in calculated iron oxide content in the Sahel region could be the observed seasonal shift in strength between the two primary centers of dust activity in the Sahara.

5. Conclusions

[55] This study applies the Gauss-Markov method to wavelength-dependent aerosol absorption measurements from the AERONET network to solve for the hematite content in dust. We show two sets of calculated hematite content, using different sets of assumptions about the mixing state of hematite within the aerosol. We discuss below the geographic variation of the results calculated assuming the external mixture hypothesis.

[56] The calculated hematite content varies from 4% to 12% of the dust by volume, with lowest hematite content in the Arabian region (4%), higher hematite in the west Sahara/Sahel region (7%), and highest hematite in the north Sahara/Mediterranean region (12%). This variation is most likely due either to differences in climatic and hence weathering history of the deserts or to different weathering conditions in the specific geomorphologic regions responsible for dust formation. In East Asia, the Dunhuang site, which is the least affected by BC contamination and therefore the most robust site for determining hematite content, shows hematite content of 4%, close to the Arabian values.

[57] For south and west Saharan sites where observations were available during both the summer and winter seasons, calculated hematite values were lower during the winter (6.3%) than during the summer (7.8%). This difference is statistically significant and may reflect mineralogical differences between the Bodele Basin and western Saharan regions where dust production is strongest during the two respective seasons. Seasonal variation of hematite fraction was weaker at Arabian sites.

[58] Our two separate hypotheses of internal and external dust mixing are extremes intended to bracket real dust mixtures and to illustrate the effect that mixing state has on dust. The observed spectral absorption profile of dust is better matched by the external mixture case, therefore we prefer the external mixing hypothesis as a better fit to the observations. This is consistent with the single-particle observations of Reid et al. [2003]; in addition, externally mixed hematite in dust is able explain the relatively nonabsorbing dust described by Kaufman et al. [2001] and Tanré et al. [2001] relative to prior estimates, since the light absorption of externally mixed dust is much less than that for an internal mixture. We do not suggest that all hematite exists as externally mixed particles, but rather that hematite particles or aggregates in dust are large enough that their centers are not effectively illuminated at the shortest AERONET wavelength and therefore contribute less to absorption per unit mass than they would if we assume small hematite particles.

[59] Calculating absolute amounts of hematite from light absorption measurements is not possible without better understanding of the size distribution and clustering of individual minerals constituting dust. Nonetheless, comparison between sites indicates enough variability among dust observations to reject the hypothesis that dust can be considered to be a globally uniform aerosol.


[60] We thank the AERONET PIs and their staff for establishing and maintaining the 21 sites used in this investigation and Brent Holben for permission to use the data. We thank the MISR science team for aerosol optical thickness data. We thank an anonymous reviewer for helpful comments. This work was supported by NASA Headquarters under the Earth System Science Fellowship grant NNG04GQ76H. We would like to dedicate this paper to the memory of Yoram Kaufman.