Volume 108, Issue D19
Aerosols and Clouds
Free Access

Saharan dust transport to the Caribbean during PRIDE: 1. Influence of dust sources and removal mechanisms on the timing and magnitude of downwind aerosol optical depth events from simulations of in situ and remote sensing observations

P. R. Colarco

P. R. Colarco

Laboratory for Atmospheric and Space Physics, Program in Atmospheric and Oceanic Sciences, University of Colorado, Boulder, Colorado, USA

Now at Goddard Earth Sciences and Technology Center, University of Maryland-Baltimore County, NASA Goddard Space Flight Center, Greenbelt, Maryland, USA.

Search for more papers by this author
O. B. Toon

O. B. Toon

Laboratory for Atmospheric and Space Physics, Program in Atmospheric and Oceanic Sciences, University of Colorado, Boulder, Colorado, USA

Search for more papers by this author
B. N. Holben

B. N. Holben

Laboratory for Terrestrial Physics, NASA Goddard Space Flight Center, Greenbelt, Maryland, USA

Search for more papers by this author
First published: 12 July 2003
Citations: 53
This is a commentary on DOI:

Abstract

[1] Dust emissions, transport, and deposition are simulated for the Puerto Rico Dust Experiment (June–July, 2000) with a three-dimensional aerosol transport model driven by assimilated meteorology. Sensitivity to the dust source formulation is tested by comparing simulations run with two different source modules to TOMS and AERONET measurements. Central and east African dust sources are more accurately simulated with the Ginoux et al. [2001] source model than with the Marticorena and Bergametti [1995] source model. The timing of high aerosol optical depth (AOD) dust events at Puerto Rico is largely independent of the timing of dust emissions, implying a persistent reservoir of suspended dust particles exists over Africa and that the timing of downwind dust events is more strongly coupled to transport dynamics than to the dust source model chosen. The downwind AOD and particle size distribution are accurately simulated using the dust particle size distribution produced by the Ginoux et al. [2001] source model or by fitting the Marticorena and Bergametti [1995] dust emissions to the d'Almeida [1987] background desert aerosol particle size distribution. The dust distribution determined by fitting the Marticorena and Bergametti [1995] emissions to the Schulz et al. [1998] initial particle size distribution produces aerosol optical depths that are consistently too high. Large dust particles observed far downwind of source areas may be partially explained if they are treated as flat disks rather than spheres in the particle fall velocity calculation. With the Ginoux et al. [2001] source model we estimate African and Arabian dust emissions during July 2000 to be 214 Tg for particles smaller than 10 μm radius. Nineteen percent of the emitted mass is transported west of Africa and over the North Atlantic. Twenty percent of the dust leaving Africa passes west of Puerto Rico.

1. Introduction

[2] Each year large quantities of Saharan dust are mobilized and transported across the tropical North Atlantic Ocean [Prospero and Carlson, 1972; Prospero, 1996]. Saharan dust is implicated in air quality issues in Florida [Prospero, 1999] and in coral reef mortality in the Caribbean basin [Shinn et al., 2000]. Mineral dust is ubiquitous in satellite imagery [Herman et al., 1997; Husar et al., 1997; Moulin et al., 1997b] and can have a substantial influence on the radiation budget within an atmospheric column [Li et al., 1996; Alpert et al., 1998; Maring et al., 2000; Quijano et al., 2000]. Radiative forcing by dust aerosols may have significant climate feedbacks [Perlwitz et al., 2001], but large uncertainties exist in such estimates due to uncertainties in dust radiative properties [Sokolik and Toon, 1999; Sokolik et al., 2001]. An optimal strategy for constraining dust radiative properties and potential climatic effects will require an integration of models, field data, and satellite observations which can trace the dust life cycle from sources to sinks.

[3] The Puerto Rico Dust Experiment (PRIDE) [Reid et al., 2003b] offers an opportunity to partially implement this integration strategy. PRIDE operated from 28 June to 24 July, 2000, and was based at Roosevelt Roads Naval Station on the eastern side of Puerto Rico (Figure 1). Six major dust events associated with Saharan dust sources were observed at Roosevelt Roads during PRIDE. The primary research aircraft during the mission was a twin-engine Piper Navajo operated by SPAWAR Systems Center San Diego which flew 21 science flights in the vicinity of Puerto Rico. Vertical profiles of aerosol optical depth (AOD) were measured from the Navajo with the NASA Ames Airborne Tracking Sunphotometer (AATS-6) [Livingston et al., 2003]. Particle size and number concentration measurements were made with a Forward Scattering Spectrometer Probe (FSSP-100X) and Passive Cavity Aerosol Spectrometer Probe (PCASP-100X) mounted on the Navajo's wing tips [Reid et al., 2003a]. The principal ground site for the mission was at Cabras Island, several hundred meters offshore of Puerto Rico. The University of Miami operated a suite of particle sampling instruments at this site, including a TSI Aerodynamic Particle Sizer (APS3310) [Maring et al., 2003a] and a high volume bulk aerosol sampler (D. Savoie et al., Vertical distribution of Saharan dust aerosols over the tropical North Atlantic based on ground-based dust measurements and AERONET aerosol optical depths, manuscript in preparation, 2002) (hereinafter referred to as Savoie et al., manuscript in preparation, 2002). A NASA Goddard Micropulse Lidar (MPL) provided vertical profiles of aerosol backscatter [Welton et al., 2001]. An AERONET Sun photometer measured spectral AOD at six wavelengths [Holben et al., 1998]. The purpose of these observations was to place some constraints on dust aerosol particle size distributions, vertical distributions, and optical properties. Such constraints are important for the analysis and interpretation of satellite and ground-based dust observations.

Details are in the caption following the image
Domain of the CARMA model. The top panel shows the G01 source function gridded at 1° ×1° resolution. The source function is the erodible percentage of the indicated grid cell, which is a function of the fraction of the grid cell covered by bare soil surface and is weighted by the topographical factor described in the text. The bottom panel shows the region covered by the MB95 source formulation. Here the shading indicates the threshold 10-m wind speed [m s−1] needed to cause dust emissions. The open (or unshaded) areas are not dust sources (NS).

[4] Our goal in this paper and its companion [Colarco et al., 2003] is to place the Saharan dust events observed during PRIDE into the context of large-scale Saharan dust mobilization, transport, and deposition patterns. We approach this through simulations with a three-dimensional dust transport model capable of simulating the mineral dust aerosol life cycle. In this paper we use the model to address the following questions related to the PRIDE data.

[5] 1. To what extent can the dust episodes observed during PRIDE be associated with large-scale dust lifting events in the Sahara desert? To address this question we implement two different dust lifting schemes in our transport model, and assess their capabilities at simulating the observed dust events, aerosol optical depths, and particle size distributions.

[6] 2. How do dust removal mechanisms play into the timing and magnitude of the dust events observed during PRIDE? To address this question we study the influence of the modeled sedimentation, dry deposition, and wet removal processes on the downwind distribution of dust.

[7] A companion paper [Colarco et al., 2003] addresses the dust transport pathway across the North Atlantic Ocean and its influence on the dust vertical distribution at Puerto Rico. We use the PRIDE data as much as possible to constrain our model, but where appropriate we rely on other data sets to answer these questions.

[8] In section 2 we discuss our transport model. In section 3 we discuss the dust source formulations used in this study. In section 4 we describe our sensitivity tests and show the results of comparison to various data sets. In section 5 we discuss our results and their implications in choosing a dust source scheme. Our conclusions are presented in section 6.

2. Model Description

[9] Our transport model is described by Colarco et al. [2002]. We derive our dynamical fields from the National Center for Atmospheric Research (NCAR) Model for Atmospheric Chemistry and Transport (MATCH) [Rasch et al., 1997]. MATCH is an offline chemical transport model which contains the basic physics packages of the NCAR Community Climate Model version 3 (CCM3) [Kiehl et al., 1996]. In order to simulate the PRIDE experiment we drive MATCH with assimilated meteorological data, using the National Center for Environmental Prediction (NCEP)/NCAR reanalyses. The NCEP/NCAR reanalyses are gridded at T62 horizontal resolution (approximately 1.875° × 1.875°) with 28 vertical layers extending to 35 km, and are available each day at 0000, 0600, 1200, and 1800 UTC. MATCH is run with an 1800 second time step, and the input fields are interpolated to the current step. At each time step MATCH diagnoses fields required to compute planetary boundary layer transport, convective mixing, and cloud and precipitation fields. Turbulent mixing in the planetary boundary layer is formulated using a nonlocal scheme [Holtslag and Boville, 1993]. This scheme provides eddy diffusion coefficients representative of local, small-scale mixing and countergradient transport terms representative of nonlocal mixing (when the size of the turbulent eddies approaches the depth of the boundary layer, typically under unstable and convective conditions). The convective transport parameterization employs a local (“shallow”) convective mixing scheme [Hack, 1994] in tandem with a penetrative (“deep”) convective parameterization [Zhang and McFarlane, 1995]. The cloud and precipitation parameterization follows the hydrologic cycle described by Rasch and Kristjánsson [1998].

[10] Rather than running MATCH for all of our simulations, we archive the dynamical fields it produces and feed those into a separate aerosol transport module. This module is a version of the University of Colorado/NASA Ames Community Aerosol and Radiation Model for Atmospheres (CARMA) [Toon et al., 1988]. CARMA is run at the same spatial and temporal resolution as MATCH, but we reduce the global fields provided by MATCH to the geographical region of interest (Figure 1). We are thus able to run multiple simulations without the computational burden of continuously regenerating MATCH's diagnosed dynamical products. We account for differences in the numerics of both models by, for instance, solving the air mass continuity equation in CARMA to derive appropriate vertical wind velocities.

[11] CARMA contains our prescriptions for aerosol source, transport, and removal terms. The particle size distribution is treated using a number of discrete size bins, and the model transport processes affect each size bin independently (i.e., there are no microphysical interactions, such as coagulation, between particle size bins). Advective and diffusive transport is calculated using a piecewise parabolic scheme following Lin and Rood [1996]. We incorporate the MATCH planetary boundary layer diffusion coefficients and countergradient transport terms into our vertical transport equation. Dust sources are incorporated into the vertical transport equation through a lower boundary flux term.

[12] Particle transport by sedimentation and removal by dry deposition are incorporated into the vertical transport equation throughfall velocities and deposition velocities respectively. The sedimentation fall velocity is computed for each size bin following the treatment by Pruppacher and Klett [1997]. The surface layer dry deposition velocity is calculated with a two-layer dry deposition model which incorporates sedimentation, turbulent motion, and molecular diffusion across the lowest model layer [Shao, 2000].

[13] Wet removal (or “scavenging”) is the most important process for removing dust from the atmosphere far from its sources. MATCH provides the rates of cumulus precipitation and stratiform precipitation and evaporation. We treat scavenging as a first order loss process, as described by Barth et al. [2000]. All aerosol swept up in cumulus precipitation is assumed removed from the model. In grid-cells affected by stratiform precipitation, aerosol swept up in a higher model layer can be removed entirely from the model or returned to a lower layer where evaporation occurs. Following Balkanski et al. [1993], a 10% efficiency for below-cloud scavenging is assumed, which gives, as an example, an e-folding lifetime of 5 hours for aerosol under a typical stratiform precipitation rate of 2 mm hr−1. For in-cloud scavenging we assume that 100% of the aerosol present in the cloud fraction resides in the cloud water and is removed in proportion to the rate of precipitation formation. The wet removal calculation is uncoupled from the vertical advection equation and is applied after the transport calculation.

[14] Our wet removal formulation is aggressive for dust. Differences in the raindrop size distributions between convective and stratiform systems should yield different scavenging efficiencies for these processes [e.g., Guelle et al., 1998], where here these efficiencies are assumed to be the same. Additionally, we find that our assumptions about in-cloud scavenging make it by far the dominant wet removal sink. We recognize that pure silicate dust may be hydrophobic, and so the in-cloud scavenging process may be less efficient than it is treated here. On the other hand, dust particles are often coated with water soluble materials. As two extremes, we consider transport with and without the wet removal scheme discussed above. The assumptions made here certainly play into the vertical distributions of the dust as well as the magnitude of the downwind dust masses, but we show later that they do not greatly affect the timing of downwind dust events. We consider sensitivity of dust vertical distributions to our wet removal scheme in the companion to this paper [Colarco et al., 2003].

[15] Following advection and scavenging we apply the convective adjustments from MATCH. The Zhang and McFarlane [1995] parameterization uses updraft and downdraft plume models which entrain and detrain air mass laterally. The scheme is representative of an ensemble of convective elements. The Hack [1994] scheme is primarily representative of shallow subtropical convection. The convective transport of dust is calculated using these schemes in tandem for the fraction of aerosol in the cloudy volume. The vertical transport through convection is conceptually like transport through a “leaky pipe,” pumping the aerosols up vertically but allowing for entrainment and detrainment at any level within the convective column.

[16] CARMA is run for a number of simulations in order to test sensitivity to various parameterizations of sources and transport processes. Input dust sources are calculated off-line, as described in the next section. We archive the model fields every 6 hours, including the full three-dimensional dust particle size distribution and the size resolved dry and wet deposition fluxes.

3. Dust Sources

[17] Characterization of the dust mobilization process remains an active field of research [Sokolik et al., 2001]. In general, dust emission occurs when the wind speed over an erodible surface exceeds some threshold value. The threshold wind speed is dependent on a number of factors, including the surface soil particle size distribution, the presence and distribution of nonerodible elements (e.g., vegetation, rocks), the soil moisture content, and biogeochemical crusting of the surface [Gillette and Passi, 1988; Marticorena and Bergametti, 1995; Gillette, 1999]. For a dry, smooth surface devoid of nonerodible elements the threshold wind speed is a function of the soil particle size distribution, with the largest particles too massive to be easily mobilized and the smallest particles tightly bound to the surface or to one another as aggregates by cohesive forces. If the wind speed exceeds a particle's threshold, it is mobilized into a horizontal, bouncing motion called “saltation.” The optimal diameter for saltation is for particles around 75 μm [Iversen and White, 1982], but particles this size are too large to remain in suspension owing to fall speeds of several tenths of m s−1 (Figure 2). Instead, these particles fall back to the ground and impart their kinetic energy to aggregates on the surface. If the process is energetic enough then the aggregates can be broken, ejecting small particles (radius <10 μm) into the transport flow. This ejection process is called “sandblasting” [Alfaro and Gomes, 2001].

Details are in the caption following the image
Saltation threshold friction velocity (from [Marticorena and Bergametti, 1995], after Iversen and White [1982]) and spherical dust particle fall velocity [Pruppacher and Klett, 1997] as a function of particle diameter. The minimum in the saltation threshold is around 75 μm; the fall velocity for dust particles that size is 0.40 m s−1. Also shown is the fall velocity for disks-shaped dust particles (following Fuchs [1964]).

[18] Various parameterizations of these complicated processes exist in the literature. We are interested in testing how our modeled dust fields downwind of source areas vary depending on the dust mobilization scheme used. Our simulations are conducted using two different dust source formulations. The first is from [Ginoux et al., 2001] (hereafter referred to as the G01 formulation) and is based on an observed correlation between major dust source regions and large-scale topographical depressions. The second is from Marticorena and Bergametti [1995] and Marticorena et al. [1997] (hereafter referred to as the MB95 formulation) and is based on a physical treatment of the saltation process discussed above coupled with a detailed database of soil particle size distributions and nonerodible surface roughness elements.

[19] Both formulations are available on a 1° × 1° resolution horizontal grid and require the wind speed at 10-m above the ground as input. Dust emissions are sensitive to the magnitude of the surface winds, and so it is important to use a wind data set of comparable spatial resolution to the source model. The surface winds from the NCEP/NCAR reanalyses are at a coarse resolution relative to the source models, and have already been shown to be inappropriate for computing the emissions in the MB95 formulation without making assumptions about their sub-grid-scale variability [Colarco et al., 2002]. Accordingly, we have chosen to calculate dust emissions with the European Center for Medium-Range Weather Forecasting (ECMWF) Tropical Ocean and Global Atmosphere (TOGA) analyses, which are available four times per day on an approximately 1.125° × 1.125° horizontal grid. The wind fields are interpolated to the 1° × 1° resolution of the source models and the computed dust fluxes are then regridded to the CARMA grid for input to the aerosol transport module.

[20] Coupling the emissions calculated with one wind set (ECMWF TOGA analyses) to the transport calculated with another wind set (NCEP/NCAR reanalyses) introduces the possibility of some internal inconsistency in the overall results. Unfortunately, we do not have access to the full three-dimensional ECMWF TOGA dynamical fields. We do show later, however, that the downwind dust transport patterns are more strongly coupled to the actual transport dynamics chosen than to the precise cycle of dust emissions, so we anticipate that the effect of this inconsistency is relatively small. Nevertheless, the effect of different dynamical fields applied to the same numerical model is an interesting point that should be more fully explored in the future.

[21] Neither source model makes explicit reference to the surface air layer stability and shear in calculating dust emissions, although it is well known that dust emissions are frequently associated with unstable atmospheric conditions. In a neutrally buoyant surface layer the wind speed varies with height in a logarithmic profile:
equation image
where z is a reference height (e.g., 10-m), z0 is the surface roughness length (m), u* is the friction speed, and κ is the von Karman constant (0.4). Under the assumption of a neutrally buoyant surface layer, if the roughness length is known the friction speed may be calculated from the 10-m wind speed using equation (1). This is what is done in the MB95 scheme (see below) because this model includes a detailed database of surface roughness features. The G01 model, however, has no such database, and so effectively assumes a constant roughness length over the model domain (in which case there is a linear relationship between the friction speed and the 10-m wind speed). Liu and Westphal [2001] compared emissions calculated assuming a neutrally buoyant surface layer (using their 10-m wind speeds) with emissions calculated where this assumption is relaxed (using their model's friction speed). They concluded that calculating emissions with the 10-m wind speed misses some of the smaller dust lifting events. We do not follow their approach here, however, because it is unclear how to couple the friction velocity associated with the ECMWF TOGA analyses with the roughness lengths provided in the MB95 input data set. An alternative would be to archive stability parameters directly from the analyses and calculate the friction speed using the MB95 data set, following Westphal et al. [1988], but as we have not archived these variables at this time, we will leave this point for future study.

[22] Dust emissions are also sensitive to the surface soil moisture content. In general, as soil moisture increases the threshold wind speed to initiate dust production also increases owing to increased interparticle capillary forces. Selah and Fryrear [1995] and Shao et al. [1996] found that dust emissions are suppressed for volumetric soil moisture content in excess of 4–5%. Fécan et al. [1999] has parameterized the increase in the erosion threshold wind speed due to increased soil moisture for a number of soil types common to Saharan Africa. As discussed below, the two source models we use here make rather different assumptions about the soil type over this region, so it is not clear how to consistently apply the Fécan et al. [1999] parameterization in this study. Additionally, the soil moisture content reported in the ECMWF TOGA analyses for this region is almost always greater than 5% (i.e., dust emissions would be completely suppressed if we followed the observations of Selah and Fryrear [1995] and Shao et al. [1996]). Liu and Westphal [2001] empirically arrived at a threshold in which dust emissions are suppressed in regions where the soil moisture content exceeds 30%. This threshold was arrived at by comparing observations of dust storm occurrences in China with the climatological ground wetness data used in their model. We have chosen to adopt this threshold for our own simulations in order to put some constraint on the effects of soil moisture, but more work needs to be done in order to more fully explore the impact of this variable on computed dust emissions. Further, dust emissions are suppressed when the snow cover in a grid cell is nonzero. We now discuss how the G01 and MB95 source models are formulated.

3.1. G01 Source Formulation

[23] Ginoux et al. [2001] provide a source formulation which is based on observations made with the dust-sensitive Total Ozone Mapping Spectrometer (TOMS) satellite instrument. The TOMS Aerosol Index (AI) is a measure of the difference in the observed spectral contrast of an aerosol laden atmospheric column from what would be observed in a pure Rayleigh scattering atmosphere, and is especially sensitive to the presence of UV-absorbing aerosols like mineral dust [Herman et al., 1997]. Prospero et al. [2002] use long-term TOMS observations and geographical data sets to show that persistent dust sources are mostly associated with large-scale topographical depressions located in arid or hyper-arid regions. Despite the aridity of these source areas, the presence of ephemeral rivers and streams indicates previous and geologically recent fluvial activity, which has left deep and extensive deposits of fine-grained sediments suitable for aeolian erosion.

[24] Ginoux et al. [2001] empirically arrived at a source function, S, which emphasizes these preferential dust source regions, where S is the fractional coverage of a grid cell by bare soil weighted by a factor representing the degree of the local topographic depression relative to the surrounding region (i.e., the source is stronger when the depression is relatively deeper). We show the source function for the region our model covers in Figure 1. The data sets this formulation is derived from are gridded at 1° × 1° resolution and are assumed time invariant. S is interpreted in our model as the fraction of the grid cell available for erosion.

[25] The vertical dust flux equation from Ginoux et al. [2001] is
equation image
where C is a dimensional factor, S is the source function described above, sr represents the soil particle size distribution, u10m is the wind speed at 10-m above the surface, and ut,r is the threshold wind speed which must be exceeded in order for emissions to occur. The subscript r is to indicate terms which have a dependence on particle size.

[26] The functional form of equation (2) is similar to that used in other studies [e.g., Tegen and Fung, 1994; Tegen and Miller, 1998; Liu and Westphal, 2001]. Liu and Westphal [2001] do not consider the aerosol particle size distribution (implying sr = 1). Tegen and Fung [1994] and Tegen and Miller [1998] do consider the aerosol particle size distribution, and they adjust their prefactors (our C and sr) to arrive at their size dependent form of equation (2). In our case we follow Ginoux et al. [2001] and adjust the value of sr to account for the particle size distribution in the soil. The parameter C is used to tune the total emissions to match some desired annual total. Ginoux et al. [2001] give a value for C of 10−9 kg s2 m−5, which yields annual emissions in agreement with typical dust budget estimates (they set C to arrive at a global dust emission budget of particles with radius <6 μm of 1604 Tg for the year 1996).

3.1.1. Soil and Aerosol Particle Size Distributions

[27] There are large uncertainties in the global variability of soil particle size distribution, so Ginoux et al. [2001] make the assumption that for small silt sized particles (1 μm ≤ radius ≤ 6 μm) the size distribution follows the relationship dMass/d(log r) = constant. This means that for lognormally spaced size classes (dr/r = constant) sr has the same value for each class r and represents the mass fraction of the total soil that is in that class. Ginoux et al. [2001] chose three classes in the size range 1–6 μm, so sr = 1/3 for each class. Emissions of finer grained particles (0.1 μm < radius <1 μm) are computed for a single effective radius (reff = 0.75 μm) with the assumption that the fraction of soil mass in this fine particle class is 10% the total mass in the silt particle classes (that is, sr<1μm = 0.1). Particles in the smallest size class will not undergo significant fractionation during transport because of their small fall velocities, so it is sufficient to treat their transport as a single class. Dust radiative properties, however, are very sensitive to the particle size distribution across the small class size range, and so for purposes of computing those properties Ginoux et al. [2001] divided the small size class into four sub-classes (using the clay-sized fractionation of Tegen and Lacis [1996]).

[28] We have modified this structure somewhat to suit our own interests by adding a fourth silt sized class (radius extending from 6–10 μm) and assigning it sr = 1/3 (so that it has the same relative mass contribution as the other silt classes). With the sub-classing discussed above we now have eight size classes. We have used the term size classes to this point to emphasize that the dust fluxes are calculated as closely as possible to the description provided by Ginoux et al. [2001]. We use the term size bins now to distinguish the aerosol size distribution transported in CARMA. CARMA can be run with an arbitrary number of size bins, and for remote sensing applications it is of interest to have a high resolution in the aerosol particle size distribution. It is not straightforward to apply the G01 formulation to an arbitrary number of aerosol size bins, however, and so for calculations with this source we have run CARMA with eight aerosol size bins spaced from 0.1 to 10 μm radius following the same size classes used in the source calculation.

3.1.2. Emission Threshold Wind Speed

[29] The choice of the threshold wind speed, ut,r, is important. Liu and Westphal [2001] use a single threshold of ut = 6.3 m s−1. Tegen and Fung [1994] and Tegen and Miller [1998] do treat the aerosol particle size distribution, but they do not use a size-dependent threshold velocity. Tegen and Fung [1994] use ut = 6.5 m s−1. Tegen and Miller [1998] vary ut from 4 to 10 m s−1.

[30] In devising their source formulation, Ginoux et al. [2001] recognized that the threshold for emissions is size dependent. Initially they chose a threshold velocity which increased with increasing particle radius, but they realized that because of the cohesive forces which bind small particles more tightly to the surface a more realistic threshold is one which emphasizes that fact and increases with decreasing particle size. Ginoux (personal communication) has recommended using the Marticorena and Bergametti [1995] threshold function shown in Figure 2 to determine a threshold for fine particle emissions. It is difficult to justify use of this function to determine the threshold in equation (2) as it applies to the initiation of saltation and not the sandblasting process which produces aerosol. Furthermore, the Marticorena and Bergametti [1995] function applies to the friction wind speed which needs to be exceeded, not the 10-m wind speed used in equation (2). However, without a full treatment of the saltation/sandblasting process, it does capture one important point about dust emissions: a higher surface wind speed is needed to initiate emissions of finer grained particles than coarse grained particles. This point is backed up by observations which show an increase in fine particle emissions with increasing surface wind relative to coarse mode emissions [Alfaro et al., 1998].

[31] Realizing the assumptions discussed above, we calculate dust emissions in the G01 source using the Marticorena and Bergametti [1995] threshold wind speed function. In a sensitivity test we also calculate emissions using a constant threshold of ut = 6.5 m s−1 (after Tegen and Fung [1994]). This threshold is higher than what is calculated using the Marticorena and Bergametti [1995] function, and so we adjust our C factor to get the same total mass emissions. This choice of ut produces a relatively larger mass of submicron particles in the ut = constant case than when ut is allowed to vary. The larger mass occurs because the total mass is controlled by the large particles, and even though the threshold is higher in the ut = constant case, it is just as easy to get small particles as large ones (recall that in the Marticorena and Bergametti [1995] case it is harder to get small particles emitted). Additionally, although the threshold is higher in the ut = constant case, the wind speeds over the source regions are typically high enough that the distribution of active dust sources does not vary much between these two simulations. The net effect is that the aerosol optical depth is about 50% higher in the ut = constant case far from source regions. Our choice of ut is a possible source of error in our calculations, but does not greatly affect the timing of the downwind dust events simulated.

3.1.3. C Parameter

[32] Emissions are calculated with the G01 source following equation (2) and using the size binning discussed above and the Marticorena and Bergametti [1995] threshold wind speed function. Because of our assumptions and our use of a different wind data set to calculate emissions we have adjusted our C parameter to yield the same 1996 total emissions as Ginoux et al. [2001] (radius < 6 μm). For our simulations we set C = 0.43 × 10−9 kg s2 m−5.

3.2. MB95 Source Formulation

[33] Marticorena and Bergametti [1995] describe a dust emission formulation based on considerations of the small scale surface characteristics controlling the saltation process. This formulation requires detailed input parameters: the soil particle size distribution, the surface roughness length, and the distribution and spacing of nonerodible roughness elements. Because of this stringent data requirement, the source is only available over a limited area rather than globally (Figure 1). In earlier versions of the model this data set extended only over the western portion of the Sahara [Marticorena et al., 1997]. More recently the domain has been extended to include the Arabian peninsula and southwest Asia [Marticorena et al., 1999]. The data are gridded at 1° × 1° resolution and allow for up to 5 different surfaces within a single grid cell.

[34] The vertical dust flux equation from Marticorena and Bergametti [1995] is
equation image
where i indexes from 1–5 the soil types in a single grid cell, αi is the sandblasting efficiency of the soil type (as derived from wind tunnel studies), and pi is the erodible fraction of the soil. The parameter α relates the ratio of the horizontal saltation flux to the vertical dust flux, and is treated here as a constant dependent only on the soil type. The index j is over the soil particle size distribution and Δsi,j is relative contribution of soil type i to the jth size class. ρair is the air density at the surface and g is the acceleration of gravity. u* is the wind friction speed, derived from the input 10-m wind speed as a function of the surface roughness length as in equation (1). u*t(j) is the saltation threshold friction speed (as shown in Figure 2). feff is called the “efficient fraction.” It is a function of the distribution of nonerodible roughness elements on the surface and indicates that in the presence of these obstacles a larger threshold velocity is needed to initiate saltation. Other than the surface feature data set, the only input required to this model is the 10-m wind speed. By integrating equation (3) over the soil particle size distribution, we can thus determine the minimum 10-m wind speed needed for dust emission to occur (Figure 1).

[35] Equation (3) is the vertical mass flux of dust particles of radius <10 μm, but does not provide an emitted aerosol particle size distribution. Schulz et al. [1998] implemented the MB95 source formulation in their own transport model and concluded that good agreement between modeled and observed aerosol optical depths were obtained using a modified form of the Shettle [1985] particle size distribution for background desert aerosols as the initial particle size distribution. We test this conclusion in our simulations through sensitivity tests using two different initial particle size distributions: (1) the [Shettle, 1985] background desert aerosol distribution as modified by Schulz et al. [1998] and (2) the d'Almeida [1987] background desert aerosol size distribution. The parameters of these distributions are shown in Table 1. Since in both cases we are distributing the dust mass across a continuous distribution we can choose the number of size bins in the transport model to obtain any desired resolution. For calculations carried out with the MB95 source we run our model with 16 radius bins spaced from 0.1 to 10 μm.

Table 1. Lognormal Distribution Mass Median Radii (ri), Geometric Standard Deviation (σi), and Mass Fraction (fi) for Two Models of the Initial Aerosol Particle Size Distributiona
r1 σ1 f1 r2 σ2 f2 r3 σ3 f3 Fraction
Schulz et al. [1998]b 0.0011 2.13 2.6 × 10−6 1.26 2.00 0.781 21.15 1.89 0.219 80%
d'Almeida [1987]c 0.41 2.10 0.036 2.41 1.90 0.957 9.69 1.60 0.007 98%
  • a Each distribution has three modes. The fraction column indicates the fraction of the distribution mass in the radius range 0.1–10 μm. Here ri are in μm.
  • b Modified from Shettle [1985].
  • c Taken from Balkanski et al. [1996].

3.3. Comparison of the Dust Source Formulations

[36] We have briefly described two dust emission formulations used in the simulations presented in this paper. The G01 formulation is empirically based and associates regions of dust lifting with large-scale topographical depressions. The correlation of dust source regions with topographical lows comes from observations with the TOMS AI product [Prospero et al., 2002], and so we expect agreement between simulated dust events and high AI events observed by TOMS over dust source areas. The MB95 formulation is physically based on a description of the saltation process and includes a database of surface features over North Africa and the Arabian Peninsula, including the soil particle distribution and surface roughness. The MB95 formulation makes no explicit connection between source location and topographic depressions. Neither scheme treats sandblasting, and there are a number of assumptions about the initial aerosol particle size distribution (as discussed above). Alfaro and Gomes [2001] have developed a dust production model which incorporates a physically based treatment of the sandblasting process into the Marticorena and Bergametti [1995] source formulation. This dust production model may offer an improvement in the source formulation because it provides an emitted aerosol particle size distribution, but we have not had the opportunity to evaluate it in the context of our own simulations.

[37] Figure 3 shows the initial dust particle size distribution calculated for MB95 and G01 dust sources for a dust emission flux of 10−7 kg m−2 s−1, a typical value for the emission flux. The dominant mass modes of the Schulz et al. [1998] and d'Almeida [1987] particle size distributions are apparent in the MB95s and MB95d examples (Table 1). The MB95d and G01g initial particle size distributions are similar for radius <2 μm, but at larger sizes the G01g simulation maintains an almost constant mass distribution while the MB95d simulation peaks and then drops off rapidly. The greater initial mass of large particles in the G01g simulations means there is more deposition of dust near the source regions owing to the greater fall velocity of these particles. In contrast, both the MB95s and MB95d simulations have most of their mass in particles smaller than a few μm radius. These smaller particles have smaller fall velocities and so are more effectively transported over long distances, which has implications for the aerosol optical depth simulated downwind of dust source regions.

Details are in the caption following the image
Initial particle size distribution for a dust flux of 10−7 kg m−2 s−1 for particles of radius <10 μm radius. We show the initial particle size distribution for the G01 source (G01g) and the MB95 source with the Schulz et al. [1998] (MB95s) and d'Almeida [1987] (MB95d) initial particle size distributions.

4. Sensitivity Tests and Results

[38] In this section we present a series of sensitivity tests conducted to understand the influence of the dust source and removal mechanisms on the model results. Saharan dust takes about 5–7 days to propagate from west Africa to Puerto Rico [Reid et al., 2003b]. In order to allow a proper spin-up time of dust sources we have run our model for the period 1 June to 31 July 2000, but show only results after 20 June.

[39] We summarize our sensitivity tests in Table 2. We assign each test a name which reflects the source formulation used and the description of the test. MB95s and MB95d are the CARMA runs using the MB95 source formulation and the Schulz et al. [1998] and d'Almeida [1987] initial particle size distributions, respectively. MB95snw and MB95dnw refer to simulations with the Schulz et al. [1998] and d'Almeida [1987] initial particle size distributions and wet removal of dust turned off. The MB95off run uses the d'Almeida [1987] initial particle size distribution, but offsets the dust fluxes from the model dynamics (as discussed below). The UNIF case uses the d'Almeida [1987] initial particle size distribution but a temporally- and spatially-uniform dust source over the same geographic region covered by the MB95 source. The G01g run uses the G01 dust source with the Marticorena and Bergametti [1995] threshold function over the entire model domain. In the G01gnw run we turn off wet removal of dust. In the G01af run we turn off the Arabian and Asian dust sources and restrict emissions to Africa only. In G01waf we only allow emissions from west Africa. In the G01gdisk run we use the entire source model domain, but we calculate particle fall speeds as if the dust particles behaved aerodynamically like flat disks (following Fuchs [1964]).

Table 2. Sensitivity Tests Using Different Realizations of the Dust Source Formulationa
Test Source Description Sal Dakar Roosevelt Roads Barbados Mass Ratio
MB95s MB95 Schulz et al. [1998] distribution 2.42 0.97 0.98 1.27 1.32
MB95snw MB95 Schulz et al. [1998] distribution no wet removal 3.42 1.95 3.03 5.00 5.91
MB95d MB95 d'Almeida [1987] distribution 1.09 0.43 0.40 0.54 0.85
MB95dnw MB95 d'Almeida [1987] distribution no wet removal 1.51 0.85 1.27 2.12 3.76
MB95off MB95 d'Almeida [1987] distribution dynamics offset from fluxes 1.15 0.54 0.36 0.57 0.76
UNIF time/space uniform source d'Almeida [1987] distribution 1.03 1.20 0.58 0.75 0.95
G01g G01 all dust sources 1.12 0.89 0.54 0.74 1.32
G01gnw G01 no wet removal 1.79 1.54 1.68 2.66 5.41
G01af G01 African sources only 1.09 0.86 0.53 0.72 1.28
G01waf G01 West African sources only (west of 4 W) 0.58 0.23 0.18 0.25 0.44
G01gdisk G01 particle fall speed 1.31 1.04 0.67 0.90 1.84
  • a Shown are the ratio of the time-averaged model AOD to the time-averaged AERONET AOD (670 nm) over the period 25 June to 31 July 2000, at Sal, Dakar, Roosevelt Roads, and Barbados. The last column shows the ratio of the time-averaged model surface dust mass concentration to the time-averaged measured dust mass concentration at Roosevelt Roads for the period 2–24 July 2000.

[40] We compute the model AOD at Sal, Dakar, Roosevelt Roads, and Barbados for each case, and calculate the ratio of the time-averaged model AOD to the time-averaged AERONET AOD (670 nm) for model results which are within 3 hours of AERONET measurement times. AERONET data were cloud-screened and quality assured according to Smirnov et al. [2000]. We calculate the model AOD from the column integrated particle size distribution with a Mie scattering code [Wiscombe, 1980] assuming a visible wavelength refractive index of 1.50 − 0.01i (after Moulin et al. [1997a]). In Table 2, a ratio <1 means the model underpredicts the AOD relative to the Sun photometer observations. Results are shown for the period 25 June to 31 July 2000.

[41] Figure 4 shows a comparison of the model AOD to the AERONET observations at Roosevelt Roads, Barbados, Sal, and Dakar. For clarity we show only the MB95s, MB95d, and G01g simulations. There is a similar time evolution in the AOD for all three model realizations. The magnitude of the model AOD differs greatly with the choice of source model and the initial particle size distribution. For example, at Dakar on 5 July the AOD peak in the AERONET data is approximately matched with the G01g simulation (τ670 ≈ 0.75), but the MB95s simulation has an AOD that is about a factor of four higher. There are similar differences in the dust magnitude at Sal, Barbados, and Roosevelt Roads.

Details are in the caption following the image
Comparison of the modeled AOD to AERONET measured AOD [670 nm] at Roosevelt Roads, Barbados, Sal, and Dakar (see Figure 1). Model results are shown for the MB95s (dashed line), MB95d (shaded line), and G01g (solid line) model runs (see Table 2). The AERONET measurements are indicated by the circles.

[42] The variability in the model AOD for the different simulations is also apparent in the ratios shown in Table 2. At Sal the ratio of model AOD to AERONET AOD is 2.42 with the MB95s simulation, while it is only 1.09 with the MB95d simulation and 1.12 with the G01g simulation. The model shows high dust AOD at Sal during periods when AERONET observations are sparse, so it is difficult to determine how well the model actually does at simulating the dust AOD at Sal. At Roosevelt Roads, however, there are more AERONET observations than at Sal, and the model and observations are better correlated in terms of identifying high AOD dust events. The relative differences between the simulations persists, however, and the ratios of model AOD to AERONET AOD are 0.98, 0.40, and 0.54 for the MB95s, MB95d, and G01g simulations at Roosevelt Roads.

[43] The AERONET AOD is a column measurement can be contaminated by the presence of other aerosol species (e.g., sea salt), so it is useful to compare the model with an unambiguous dust measurement. Table 2 also shows a comparison of the model surface dust mass concentration at Roosevelt Roads to the surface dust mass concentration measured with the University of Miami high volume bulk aerosol sampler. The bulk aerosol sampler measurements were made at Roosevelt Roads during PRIDE for the period 2 July to 24 July (Savoie et al., manuscript in preparation, 2002). In Table 2 we show the ratio of the time-averaged model dust mass to the time-averaged measured dust mass for each of our sensitivity tests. A ratio >1 means the model overpredicts the dust mass at the surface relative to the bulk aerosol sampler measurements.

[44] In Figure 5 we compare the model surface dust mass concentration at Roosevelt Roads to the bulk aerosol sampler dust measurements for our MB95s, MB95d, and G01g simulations. Also shown are the measured sea-salt mass concentrations. All three model realizations exhibit a similar time evolution, and are in good agreement with the measured time evolution until about 19 July. The model does not see the relatively high dust concentration measured on 21 July, a case which is discussed in some detail in the companion to this paper [Colarco et al., 2003]. Additionally, the G01g and MB95s runs overestimate the surface dust mass concentration by a factor of three during the 5–6 July dust event while getting the magnitude of the 9–10 July and 15 July events about right.

Details are in the caption following the image
Comparison of the MB95s, MB95d, and G01g modeled surface dust mass concentrations at Roosevelt Roads to measurements with the University of Miami high volume bulk aerosol sampler. Also shown are the bulk aerosol sampler measured sea-salt surface mass concentrations.

[45] It is interesting that the MB95s and G01g simulations have similar surface mass concentrations while differing greatly in column AOD at Roosevelt Roads. The AOD is controlled primarily by submicron particles, which do not contribute greatly to the total mass concentration. As we will see later, the MB95s simulation has many more submicron particles than the G01g simulation, which we expect from source considerations (Figure 3). The simulations with wet removal turned off greatly overpredict the amount of dust at the surface.

[46] Figure 6 shows the July 2000 total dust emissions calculated with each source model. Over the entire domain the total emissions of particles with radius <10 μm is 214 Tg in the G01 source and 112 Tg in the MB95 source. The Ginoux source has larger emissions over central Africa, particularly in Mali, Algeria, and northern Libya, while the MB95 source has stronger emissions in southwest Asia, particularly in Iran and Afghanistan. The July 2000 emissions over Africa are 152 Tg in the G01 source and 56 Tg in the MB95 source. Both formulations have a strong source on the west African coast. The dust emissions from west African sources (north of 20° N and west of 4° W) are 54 Tg with the G01 source and 52 Tg with the MB95 source during July 2000. The predicted emissions for July 2000 for this region are comparable to the values reported by Marticorena et al. [1997] for west African dust sources in an earlier version of their model (61 and 65 Tg for July 1991 and 1992 respectively). We further note that almost all of the African emissions calculated with the MB95 source are in west Africa.

Details are in the caption following the image
Dust emissions per unit area [g m−2] computed for July 2000 with the (top) G01 and (bottom) MB95 source formulations. The results have been regridded from the 1° × 1° source grid to the CARMA resolution.

[47] Figure 7 shows the time evolution of dust mobilization for the period 20 June to 25 July 2000. We show the total dust lifting, dust lifting over west Africa, and dust lifting over Arabia and Asia. Also shown is 1200 UTC 10-m wind speed from the ECMWF surface analyses averaged over west Africa. The dust lifting in west Africa is generally well correlated with the higher surface wind speeds from the analyses; conversely, when the 10-m wind speed is low the dust fluxes are much smaller. Dust mobilization with the MB95 source is more episodic, with periods where almost no mobilization occurs, while with the G01 source mobilization occurs all the time. Both the G01 and MB95 source formulations have about the same timing of dust lifting events for the west African sources, although the amplitudes are larger in the MB95 source. The total emissions are almost the same over west Africa with both source models because the G01 source generally has emissions continuing during periods when the MB95 emissions are very small. We also see in Figure 7 that almost no dust comes from central and east Africa with the MB95 source, while there is a substantial contribution from this region in the G01 source.

Details are in the caption following the image
Time evolution of dust mobilization for (top) the G01 source and (middle) MB95 source. The totals shown are the emissions per six hours (corresponding to the update frequency of the wind speeds). The shaded bars are the total emissions over the entire model domain. The thick solid line shows emissions from west Africa. The thin solid line shows emissions from Arabian and southwest Asian dust sources. (bottom) The 1200 UTC 10-m wind speed from the ECMWF analyses averaged over west Africa.

[48] Table 3 summarizes the total dust emission and deposition masses during July 2000 for our sensitivity studies. Also shown are the July 2000 dust masses passing west of the African coast and passing west of Puerto Rico. From these values we calculate two transport factors: 1) a source transport factor, which is the percentage of the dust emission mass that leaves Africa to the west, and 2) an Atlantic transport factor which is the percentage of dust leaving Africa which passes west of Puerto Rico. Note that these transport factors refer only to the westward transport of dust, not to the net westward transport (westward - eastward).

Table 3. Total Dust Emissions and Deposition During the Period 1–31 July 2000 for the Sensitivity Tests Shown in Table 2a
Test Emissionsb Dry Deposition, Tg Wet Deposition, Tg AF PR Source Transport Factor, % Atlantic Transport Factor, %
MB95s 112.1 28.6 62.1 43.4 9.2 39 21
MB95snw 112.1 38.3 0 56.7 37.3 51 66
MB95d 111.5 51.5 45.4 31.5 5.7 28 18
MB95dnw 111.5 64.5 0 39.7 23.5 36 59
MB95off 115.7 49.9 45.6 32.4 5.1 28 16
UNIF 149.7 65.0 70.9 31.5 8.0 21 25
G01g 214.2 133.8 66.8 41.4 8.5 19 20
G01gnw 214.2 159.3 0 61.7 37.0 29 60
G01af 151.6 95.5 50.4 40.0 8.1 26 20
G01waf 53.6 32.6 18.1 20.4 3.0 38 15
G01gdisk 214.2 107.3 88.5 55.2 12.1 26 22
  • a Also shown is the mass [in Tg] of dust leaving Africa to the west (AF) and the mass of dust passing west of Puerto Rico (PR). The Source transport factor is the percentage of the total emissions leaving Africa to the west (= AF/Emissions). The Atlantic transport factor is the percentage of dust leaving Africa which passes west of Puerto Rico ( = PR/AF).
  • b Ideally, the dust emission mass totals for the MB95s and MB95d test should be the same. Some roundoff in the specification of the particle size distribution makes them slightly different.

[49] To illustrate the meaning of the transport factors we consider comparisons between various simulations. The source transport factor is 39% in the MB95s simulation versus 19% in the G01g simulation. This difference is because of differences in the initial particle size distributions simulated in each run, and reflects the influence of the source term on the downwind transport. The Atlantic transport factors for these two simulations are almost the same, however, indicating that the removal processes affecting the dust transported across the North Atlantic Ocean are similar in both simulations. The source and Atlantic transport factors are higher in the cases run with wet deposition turned off. The G01gdisk simulation has a larger source transport factor than the G01g simulation because the fall velocity of disk-like particles is smaller than the fall velocity for spheres.

5. Discussion

[50] In this section we discuss the sensitivity test results shown in Tables 2 and 3. We consider data from AERONET and TOMS in evaluating (1) the geographic distribution of dust sources in Africa which contribute to dust in Puerto Rico, (2) the cause of the timing of major dust events, and (3) the influence of removal mechanisms and the dust particle size distribution at the source in determining the magnitude of the dust AOD downwind of sources.

5.1. Geographic Distribution of Dust Sources

[51] In Figure 8 we show the geographic distribution of the model AOD from the G01g and MB95d simulations for 28 June, 7 July, 15 July, and 21 July. These are fairly typical days from the model simulations which illustrate both dust lifting and long-range transport features. Also shown is the Earth Probe TOMS Aerosol Index for these days. As discussed in section 3.1, the TOMS AI is sensitive to the presence of UV-absorbing aerosols like dust. Because the UV reflectivity is low over both land and water, the AI is sensitive to the presence of these aerosols over both surface types in the absence of clouds [Herman et al., 1997]. AI increases with increasing AOD, and so gives an indication of the amount of dust present, but we do not attempt a quantitative comparison between the AI and the model AOD because of the sensitivity of the AI to the height of the aerosol layer and the single scatter albedo of the aerosol [Torres et al., 1998; Colarco et al., 2002]. Note that the high AI values south of the equator are associated with smoke aerosols not modeled in our simulations. Note also that the low AI values in the North Atlantic north of the latitude of Florida could be correlated with modeled dust distributions; although this region appears devoid of dust in the model simulations there is some dust there, but its amount is below the lowest AOD contour value shown.

Details are in the caption following the image
Comparison of the model AOD to the EP-TOMS Aerosol Index for 28 June, 7 July, 15 July, and 21 July. Gaps in the satellite data are due to the spacing of the satellite orbit and cloud cover. The panels show (top) the EP-TOMS AI, (middle) the model results for the G01g simulation, and (bottom) the model results for the MB95d simulation.

[52] Figure 8 shows high AI values over much of central and east Africa, indicating the importance of these regions as dust source areas. These areas are not represented as strong dust sources in the MB95 source formulation, as evidenced by the low AOD values present in the MB95d simulation (τ670 < 0.1 over most of this region). A similar figure showing the results for the simulation with the Schulz et al. [1998] (MB95s) initial particle size distribution instead of the MB95d results would have somewhat higher AOD values over central and east Africa owing to the different initial particle size distribution assumed, but the interior of Africa would still not appear as a strong dust source because very little dust gets mobilized in this region (Figure 6). In the G01g model run, however, there is typically a band of high AOD (τ630 > 0.3) across central and east Africa at about 15° N.

[53] We perform two sensitivity tests to constrain the importance of various source regions to the downwind AOD. In the G01af test we considered only dust sources in Africa, turning off the dust sources in Arabia and Asia. In G01waf we considered only dust sources in west Africa (west of 4° W), turning the sources off everywhere else. In Table 2 we see that the ratio of the model AOD to the AERONET observations at Sal, Dakar, Barbados, and Roosevelt Roads is almost the same regardless of the presence or absence of the Arabian and Asian dust sources, as is the average surface mass concentration at Roosevelt Roads. Similarly the Atlantic transport factors for the G01g and G01af simulations are the same. This means that the Arabian and Asian dust sources are relatively unimportant to our study of dust transport over the North Atlantic Ocean. We found this conclusion to also be true for the MB95d and MB95s simulations (results not shown).

[54] In contrast, the amount of dust transported west of Puerto Rico in the G01waf simulation is about one third the mass in the G01g run, and there is a comparably large difference in the AOD ratios and surface mass concentration ratio shown in Table 2. Figure 9 compares the modeled AOD from the G01g and G01waf simulations with the AERONET observations at Sal and Roosevelt Roads. The timing of the dust events is the same in both model realizations, but the difference in the AOD magnitude between the two simulations indicates that dust mobilized in central and east Africa is transported over the North Atlantic Ocean. The importance of central and east African dust sources is also suggested by comparing the Atlantic transport factors of the G01g and G01waf simulations. This factor is 21% in the G01g run, but only 15% in the G01waf run, indicating that the transport of dust from west Africa across the Atlantic is less efficient than the transport of dust from central and east Africa. These results indicate that the central and east African dust sources contribute significantly to the G01 source in estimating the correct AOD at Roosevelt Roads.

Details are in the caption following the image
Comparison of the model AOD to the AERONET measured AOD at (left) Roosevelt Roads and (right) Sal. Model results are shown the G01g simulation (solid, all G01 dust sources) and the G01waf simulation (shaded, only west African dust sources).

[55] The TOMS observations show the presence of central and east African dust sources, which appear much more prominently in our simulations with the G01 source formulation than in our simulations with the MB95 source. The results of the G01waf simulation shows that those central and east African dust sources are important to explaining the long-range transport of dust over the North Atlantic Ocean for simulations carried out with the G01 source formulation. In Figure 9 and Table 2 we see that by considering dust sources in west Africa only the model greatly underestimates the amount of dust transported to Roosevelt Roads. From the G01af simulation, however, we see that dust sources in Arabia and Asia are relatively unimportant in determining the long-range transport of dust over the North Atlantic Ocean. In Figure 4 we see that despite the absence of central and east African dust sources in the MB95 source formulation we are still able to get about the right AOD at Roosevelt Roads with both the MB95d and MB95s model runs when compared with AERONET. We recall from section 3.3 that the total mass of dust emissions in west Africa is about the same in both the G01 and MB95 sources. As we will show later, the explanation for better agreement in the MB95 source simulations' AOD with AERONET at Roosevelt Roads than in the G01waf simulation's AOD is because of differences in the initial particle size emitted.

[56] As a final point, we consider that the differences in dust emissions in central and east Africa between the G01 and MB95 source formulations could be related to the surface roughness features included in the MB95 source, which may inhibit dust emissions. However, the threshold velocity for dust emissions in this region is comparable to the threshold in west Africa (7–9 m s−1, Figure 1), which suggests that the wind speeds provided by the ECMWF analyses may be too low in central and east Africa. This low bias may be the result of poor data sampling in this region for the assimilation model used to generate the analyses. If the wind speeds in this region are biased too low that may affect the MB95 source more than the G01 source because the threshold for dust emissions in the G01 source is lower, being controlled only by the saltation threshold velocity (Figure 2) and not the surface roughness features. We performed a series of sensitivity tests (not shown) in which the MB95 source is recalculated with the surface wind speed increased by 1, 2, and 3 m s−1 in the different tests. Predictably, an increase in the wind speed enhanced the total dust emissions. The distribution of central and east African dust sources was largely unchanged until the wind speed was increased by 2 m s−1, at which point the sources identified in the TOMS imagery became apparent in the simulation. The total emissions in this case were about a factor of six higher than in the baseline cases, however, and so the downwind transport was unrealistic. Another possible explanation for the low emissions in the MB95 scheme over this region is that we have not included the effects of surface layer instability in formulating the friction speed used to calculate emissions. As suggested by Liu and Westphal [2001], it is possible that allowing for instability to affect the friction speed could enhance weak dust events over this region. Further elaboration on possible errors in the surface wind fields is beyond the scope of this study.

5.2. Timing of Dust Events

[57] Figures 4 and 7 suggest an association between major dust lifting events and the downwind peaks in the model dust AOD. For example, there is a peak in the dust lifting with both the G01 and MB95 sources on about 22 June; there is a subsequent peak in the model AOD at Sal on 25 June and at Roosevelt Roads on 28 June. A similar association can be made for the other high AOD dust events observed at Roosevelt Roads during PRIDE. We summarize the timing of west African dust episodes and the downwind AOD peaks at Sal and Roosevelt Roads in Table 4 for the MB95d simulation. We realize that the “timing” of these dust events is somewhat subjective because these events are really large-scale features that may persist over extended time periods. Also shown in Table 4 is the time between a peak in west African dust emissions and a subsequent AOD peak at Sal and Roosevelt Roads. We will show later, however, that the timing AOD peaks at Roosevelt Roads is largely uncorrelated with the timing of west African dust emissions.

Table 4. Timing of Peaks in West Africa Dust Production and the Subsequent Model AOD Peaks at Sal and Roosevelt Roads During PRIDE as Determined From the MB95d Simulationa
Peak of Dust Emissions Sal AOD Peak (Model) ΔtSal Roosevelt Roads AOD Peak (Model) ΔtRR
21–22 June 24–25 June 3 28–29 June 7
28–29 June 1–2 July 3 5–6 July 7
2–3 July 5–6 July 3 9–10 July 7
8–9 July 10–13 July 3 15 July 7
13–14 July 15–16 July 2 21 Julyb 8
17–18 July 19–20 July 2 23–24 July 6
  • a ΔtSal and ΔtRR are the times [days] between a peak in west Africa dust emissions and a downwind AOD peak at Sal and Roosevelt Roads.
  • b On July 21 the model had the dust placed too far south of Roosevelt Roads (see Figure 8), and so for this point we show the peak AERONET AOD at that position.

[58] To test the sensitivity of the timing of downwind AOD peaks to the timing of dust lifting episodes we run two simulations which play with the timing and distribution of dust lifting events. The first is a simulation in which the dust sources are offset by two days from the model dynamics (e.g., for the 3 July model dynamics we use the 1 July dust source). For this test we use the MB95 source because the dust lifting is more episodic in that source than in the G01 source, with periods of almost no dust lifting between major events (Figure 7). Offsetting by a longer period of time runs the risk of aliasing the dust lifting events, which occur with a frequency of about five days. In the second simulation we consider a dust source which is spatially and temporally uniform. For this simulation the dust source region is considered to extend over the entire domain identified in the MB95 source, and the flux from each grid cell is time-independent and tuned so that the dust mass leaving Africa is the same as in the MB95d case (Table 3). We use the d'Almeida [1987] initial particle size distribution for both of these cases.

[59] Figure 10 shows the modeled AOD for the uniform (UNIF), offset (MB95off), and baseline nonoffset (MB95d) cases at Roosevelt Roads and Sal. All three cases show a similar timing of high AOD dust events at Roosevelt Roads, all of which are generally similar to the timing of dust events observed by AERONET. At Sal the three cases do not look much alike, although the MB95d and MB95off cases do have some similarly timed dust events predicted. The uniform dust lifting in the UNIF case is clearly unrealistic, so the similarities between the three simulations at Roosevelt Roads indicates that far downwind of the dust sources the timing of dust events is controlled mainly by the transporting dynamics and is uncoupled from the timing of dust lifting events. Dynamics is less important closer to the sources, as evidenced by the model comparisons at Sal. It is interesting to note how similar the three models' AODs are at Roosevelt Roads despite their large differences at Sal.

Details are in the caption following the image
Comparison of the modeled AOD at Roosevelt Roads and Sal for the spatially and temporally uniform source (dashed line, UNIF), the time offset source (shaded line, MB95off), and nonoffset baseline case (solid line, MB95d).

[60] In Figure 11 we show the computed AOD over the model domain on 7 July for the MB95d, MB95off and UNIF model runs. Near Africa the three model runs have different dust distributions; not surprisingly, the rather pathological UNIF run differs especially from the other two. Far from the dust sources, however, the differences in the timing of the dust lifting events is only really apparent as small differences in the magnitude of the dust AOD. A typical example of this is 7 July. Note that all three simulations show a V-shaped dust plume around Roosevelt Roads with its vertex over Barbados. That the downwind dust plume evolves similarly in these three runs indicates that the dynamics and dust removal processes are more important to determining the downwind plume shape and magnitude than the mechanics and timing of the dust lifting process.

Details are in the caption following the image
Modeled AOD [670 nm] for 7 July for the (left) nonoffset MB95d, (middle) offset MB95off, and (right) the spatially and temporally uniform source (UNIF) model runs.

[61] The AOD ratios shown in Table 2 are similar for the MB95d and MB95off simulations, as are the dust emissions masses and transport factors shown in Table 3. The differences between the modeled AODs at Sal shown in Figure 10 occur mainly at times for which there are no AERONET observations, and so do not show up in Table 2, but it is clear that the dust AOD at Sal is sensitive to the timing of dust lifting events. Figures 10 and 11 show that the timing of downwind dust events and the shape of the transported dust plume are similar for the MB95d and MB95off simulations, and so we conclude that the timing of the dust pulses reaching the Caribbean is largely independent of the timing of the dust lifting. This result can be generalized to the G01g simulation by noting the similarity of the dust plume shape when the G01g simulation is compared to the MB95d run as in Figure 8.

[62] As a final comparison, we consider the effects of wet removal on the timing of downwind dust events. Figure 12 shows the modeled dust AOD at Roosevelt Roads and Sal for the G01g and G01gnw cases. Although the magnitude of the AOD is quite different in both cases, the timing of the high AOD dust events is about the same regardless of whether we have wet removal turned on or not. Considering the aggressive nature of our wet removal scheme, we consider these two cases as bounds on the importance of this process. On the basis of Figure 12 we conclude that wet removal cannot have a significant effect on the timing of downwind dust events far from sources.

Details are in the caption following the image
Comparison of the modeled AOD at Roosevelt Roads and Sal for model runs with wet removal turned on (solid line, G01g) and turned off (shaded line, G01gnw).

[63] These results imply that the dust source cannot be validated by considering only the far downwind distribution simulated. They also imply that there must be a reservoir of fine grained dust aerosol particles continuously available for transport. This reservoir is likely related to both the large-scale spatial distribution of the dust sources (over several hundred km) and to the fact that the finest grained particles do not sediment out of the model rapidly and so can reside suspended for at least two days. Finally, we find that the wet removal mechanism has little importance in the timing of downwind dust events far from sources.

5.3. Magnitude of Dust AOD

[64] The MB95s and MB95d results in Figure 4 show that where the dust source is the same the choice of initial dust particle size distribution plays an important role in determining the magnitude of the downwind AOD. This is in part because of variations in the Mie extinction efficiency with particle size parameter, but also because size dependent removal processes alter the amount of dust transported over long distances in each of these simulations. In particular, differences in the sedimentation and dry deposition between these two simulations are reflected in the transport factors shown in Table 3, where the source transport factor is 39% for the MB95s simulation and 28% for the MB95d run. In this section we consider the effects of the model dust removal processes and the variations in the Mie extinction efficiency with particle size on the downwind AOD simulated.

[65] Figure 13 shows a scatterplot comparison of the AERONET AOD with the model AOD at Sal and Roosevelt Roads for model output times within 3 hours of AERONET measurement time. The model results shown are for the MB95s, MB95d, and G01g simulations, as well as for the same source runs but with wet removal turned off (MB95snw, MB95dnw, and G01gnw, respectively). Also shown are the regression line and correlation coefficient for the comparison of the model to AERONET.

Details are in the caption following the image
Scatterplot of the model AOD versus AERONET AOD measurements at (top) Roosevelt Roads and (bottom) Sal. Results are for model output times within 3 hours of AERONET measurements. We show model results for the (left) MB95s, (middle) MB95d, and (right) G01g simulations. The solid triangles are for simulations with wet removal turned on; the shaded triangles are for simulations with wet removal turned off. The solid line is the 1-to-1 line for comparison. Also shown are regression lines and correlation coefficients. Note that for the MB95s simulations at Sal the AOD gets as high as 5; we have cut the graph axes down for clarity, although these high AOD values are included in the regression analysis.

[66] The model results in Figure 13 differ markedly from one another. In Table 2 we showed that the ratio of the model AOD to the AERONET measurements at Sal was 2.42 for the MB95s simulation. This bias toward high AOD is apparent in Figure 13, where most of the points lie above the 1-to-1 line for this case. We can also see the bias toward low AOD at Roosevelt Roads for the MB95d simulation (ratio = 0.40) and G01g simulation (ratio = 0.54).

[67] All of the simulations generally underpredict the AOD at Roosevelt Roads. Possibly this is due to an AOD contribution from another aerosol type that we are not modeling. The likely candidate for this aerosol is sea salt. We show the measured sea-salt mass concentrations at Roosevelt Roads during PRIDE in Figure 5. Generally, the surface mass concentration was fairly uniform throughout the experiment, and by looking at days when the dust concentration is small we estimate the 670 nm AOD contribution from sea salt to be about 0.1. Given the uniformity in sea-salt concentrations it is likely that its contribution will be only to elevate the total AOD without greatly affecting the slope shown in Figure 13.

[68] The correlation of the model with the AERONET measurements is not especially good with any of the simulations, although it is somewhat better with the G01g simulation. The correlation is stronger at Sal for all of the simulations than at Roosevelt Roads. When we turn off the wet removal scheme the model AOD increases while the timing of events remains largely unchanged (Figure 12). The correlation strengthens at Sal in all the simulations when the wet removal is turned off, but is gets worse at Roosevelt Roads. The model greatly overpredicts the AOD at Roosevelt Roads when the wet removal scheme is turned off.

[69] Figure 14 shows the average modeled dust volume particle size distribution at Roosevelt Roads and Sal over the period 25 June to 31 July 2000. For purposes of computing the average we retain only the model profiles which are within 3 hours of AERONET retrievals of particle size distribution made during dusty conditions (τ670 > 0.20). AERONET inversions of particle size distribution are available from the Sun photometer almucanter scans where the solar zenith angle of the observation is <30° [Dubovik and King, 2000]. Note that the AERONET retrieval is a column integrated size distribution, and we have calculated the model size distribution accordingly. The model results shown are for the MB95s, MB95d, and G01g runs discussed earlier. Also shown is the G01gdisk run, in which the particle fall velocities are calculated as if the dust particles behaved aerodynamically like flat disks (following the treatment of Fuchs [1964]).

Details are in the caption following the image
Average dust particle size distribution at (top) Roosevelt Roads and (bottom) Sal. Model results are for the MB95s, MB95d, G01g, and G01gdisk simulations. Also shown is the particle size distribution retrieved by AERONET. To emphasize the dust contribution to the size distribution, we retain AERONET inversions only for cases where τ670 > 0.2. The model results are averaged from model output within 3 hours of the AERONET measurement.

[70] In Figure 14 we also show the AERONET retrieved effective radius for coarse mode particles (r > 0.6 μm). For comparison we have computed the model effective radii over the same size range. The MB95d (reff = 1.32 μm) and G01g (reff = 1.49 μm) simulations evolve to a similar particle size distribution. The MB95s distribution has a somewhat smaller effective radius (reff = 1.05 μm), consistent with the initial particle size distribution shown in Figure 3. The G01gdisk simulation (reff = 1.67 μm) differs from G01g because the disk-like particles have a smaller fall velocity than spheres, and so the largest particles are transported further in the G01gdisk simulation than in the G01g run.

[71] The best agreement between the models and the AERONET particle size distributions is for the G01gdisk simulation, although it somewhat overpredicts the volume of large particles at Sal. The MB95d and G01g simulations agree reasonably well with AERONET at Sal. The MB95s simulation appears to be shifted toward too small of particle sizes to agree with the retrievals at Sal. At Roosevelt Roads the largest particles have already fallen out in the model, and none of the model distributions agree particularly well with the AERONET retrieval. We can partially explain “missing” large particles if the dust particles fall like disks instead of spheres, but the model still misses the volume at 10 μm radius by more than an order of magnitude. The effect of calculating the fall velocity as if the particles behaved like disks has little effect for the smallest particles because they do not fall out of the model. Missing large particles might also be associated with other aerosol types (e.g., sea salt) or with uncertainties in the retrieval of the AERONET size distributions [Reid et al., 2003a]. In particular, AERONET retrievals assume spherical particles, which may bias the retrieved particle size distribution in cases of nonspherical particles (e.g., dust). This bias is typically evident as a spurious enhancement in the fine mode of the aerosol particle size distribution [Dubovik et al., 2000].

[72] Figure 15 shows the same comparison to the AERONET retrieved particle size distribution as Figure 14, but for the MB95snw, MB95dnw, and G01gnw simulations. Over the same averaging period, the simulations with no wet deposition have a larger total particle volume than the simulations which include wet deposition. This is consistent with the higher AOD ratios shown in Table 2 and the larger transport factors shown in Table 3. Our wet removal scheme is not size-dependent, however, so differences in the shape of the particle size distribution between the wet- and no wet-deposition runs will be because of other loss processes that occur along the aerosol transport path. A treatment which incorporated a particle size-dependent wet removal would probably lead to a relative enhancement of 0.5–1 μm sized particles because this is where the rain droplet collection efficiency curve reaches a minimum [e.g., Seinfeld and Pandis 1998]. The model effective radii are slightly higher at Roosevelt Roads for all the simulations when wet removal is turned off, but the difference between the wet and no-wet runs remains much smaller than differences between the different source realizations.

Details are in the caption following the image
As in Figure 14 but for model runs with wet removal turned off.

[73] Another way to consider the particle size distribution is in terms of its effect on the calculated AOD. Figure 16 shows the model averaged size distribution of AOD at Sal and Roosevelt Roads for the MB95s, MB95d, G01g, and G01gdisk simulations. What we are showing is essentially the column-integrated particle area size distribution weighted by the Mie extinction coefficient, which illustrates the contribution of different sized particles to the total column AOD. For comparison we include the AERONET retrieved particle size distribution weighted by the Mie extinction coefficients. Again, we retain only AERONET measurements for which τ670 > 0.20. Figure 16 shows that in the MB95s simulation the dominant contribution to the AOD is from submicron dust particles. This is especially apparent at Sal. MB95s differs greatly from the MB95d, G01g, and G01gdisk simulations, where the dominant contribution to the AOD is from particles of radius >1μm. As anticipated from Figure 14, the AOD contribution at large particle sizes is greatest for the G01gdisk simulation. The MB95d and G01g results are in better agreement with the AERONET distribution shown than the MB95s simulation is, but we do not intend the comparison to AERONET to be especially robust as we are not assured of closure between our Mie scattering code and the AERONET inversion.

Details are in the caption following the image
Average size distribution of AOD at (top) Roosevelt Roads and (bottom) Sal. Model results are for the MB95s, MB95d, G01g, and G01gdisk simulations. Also shown is the AOD size distribution inferred from the AERONET retrieved particle size distribution.

[74] We have shown that differences in the model AOD occur because of the initial dust particle size distribution assumed and the removal processes affecting the dust. Our comparisons of model particle size distribution and AOD to data from AERONET lead us to believe that the MB95s simulation is unrealistic because of the small dominant mass mode in the initial particle size distribution. Wet removal of aerosols has an effect on the total AOD, but does not modify the particle size distribution as we have implemented the process. None of the model runs capture the large particles AERONET retrieves at Roosevelt Roads, although our best agreement came for the simulation in which the particles were treated as disks instead of as spheres in the fall velocity calculation (G01gdisk). We note that Reid et al. [2003a] point out some considerations such as particle shape which may bias the AERONET retrieved particle size distribution.

[75] Ginoux [2003] considered the effect of particle shape on sedimentation in a three-dimensional dust transport model. They found insignificant differences in the modeled dust concentrations and particle size distributions between cases where the particles were treated as spheres and where they were treated as prolate ellipsoids of aspect ratio λ = 1.5 (based on observations). Ginoux [2003] point out they are working with a limited set of observations, based on dust collected in China, and that other observations support aspect ratios much larger to λ = 1.5 with a sensitivity to source region of the dust. More significant differences were found when the particles were treated as very elongated spheroids (λ = 10), qualitatively similar to our treatment of the particles as disks.

[76] Maring et al. [2003b] compared the observed PRIDE dust particle size distributions at Roosevelt Roads with measurements made previously at Izaña (in the Canary Islands, near the dust sources). They found that large particles fall out less rapidly in transit from Izaña to Puerto Rico than gravitational settling alone would suggest. Since their measurement technique is based on the aerodynamic properties of the dust particles they cannot attribute the slower fallout to particle shape, as we have, but instead hypothesize a small upward vertical velocity due to turbulence or ascent caused by radiative heating. We do not explicitly account for radiative heating of the dust layer in our model, although it may be accounted for implicitly in the analyses driving the simulation [e.g., Alpert et al., 1998]. We do not pursue this possibility further here.

6. Conclusions

[77] We have shown results of model runs simulating the emissions, transport, and deposition of dust aerosols during the time period of the PRIDE field campaign. The model has been shown to capture the timing of most of the dust events observed during the campaign. Our sensitivity tests allow us to place some constraints on the importance of source and removal parameterizations on the timing and magnitude of downwind dust events. We point out that our model runs were only for the two-month period bounding the PRIDE campaign, so further work will need to address the generality of those results to other time periods and longer durations.

[78] Simulations were run with two different dust emission schemes. TOMS imagery suggests that there are important dust sources in central and east Africa (Figure 8). Those source regions are generally not active in the MB95 source but are in the G01 source. We find that the G01 source requires those dust sources to be active in order to get close to the observed magnitude in AOD downwind of Africa (Table 2 and Figure 9). We also find that the magnitude of the downwind AOD peaks in the tropical North Atlantic Ocean are largely independent of Arabian and Asian dust sources. In our G01g simulation we found July 2000 total dust emissions of 214 Tg. 19% of the emitted mass is transported west of Africa and over the North Atlantic Ocean. 20% of the dust leaving Africa passes west of Puerto Rico.

[79] The timing of modeled downwind dust events is similar for both the MB95 and G01 sources, and generally compares well with the timing of AOD peaks observed by AERONET (Figure 4). In the MB95off simulation we offset the timing of dust emissions from the timing of the transport dynamics. We determine that the timing of the downwind AOD peaks is relatively unaffected by the timing of dust lifting episodes (Figures 10 and 11). The timing is also unaffected by the presence of central and east African dust sources (Figure 9). These results imply that a persistent reservoir of suspended dust particles exists over Africa and that the timing of downwind dust events is more strongly coupled to transport dynamics than to the dust source model chosen. Additionally, the timing of downwind dust events is also independent of our treatment of wet removal (12).

[80] The magnitude of the model AOD is very sensitive to the initial aerosol particle size distribution and the inclusion of the wet removal scheme. The MB95d and G01g runs evolve a similar AOD at Sal and Roosevelt Roads despite the much larger total dust emission in the G01g simulation. This is because the MB95d run puts more of the emitted dust mass in smaller sized particles that can be transported over large distances, which can be seen by comparing the source transport factor for each run (28% for MB95d, 19% for G01g). The Atlantic transport factors for these simulations are similar (18% for MB95d, 20% for G01g), although the slightly smaller value for the MB95d run reflects the less efficient west African transport we saw in the G01waf simulation.

[81] In the MB95s simulation the average AOD is much higher than the AERONET observations at Sal (Table 2 and Figure 4). There is a large contribution to the AOD in that simulation from submicron particles, and a comparison of that model run to the AERONET size distribution suggests this contribution is unrealistically large (Figure 16). Because these particles do not fall out rapidly there is significant transport out of Africa (source transport factor = 39%) and across the Atlantic to Roosevelt Roads (Atlantic transport factor = 21%). At Roosevelt Roads the ratio of the model AOD to AERONET is about one, suggesting that all of the AOD at Roosevelt Roads can be explained by dust. This is unrealistic as we have not included other optically significant aerosols (e.g., sea salt) in our model. We have attributed the high AOD from the MB95s simulation to the small mass median radius of the dominant aerosol mode, and we suggest that mode is unrealistically small.

[82] Further work will need to focus on the specification of the emitted particle size distribution. Our choice of the threshold wind speed for emissions in the G01 source does have an effect on the initial aerosol particle size distribution, which in turn will have an impact on the aerosol optical properties far from dust sources. A starting point for dealing with this issue is to incorporate a physical treatment of the sandblasting process which produces the dust aerosol [Alfaro and Gomes, 2001]. Further work will also need to be done to incorporate a more realistic treatment of the effects of soil moisture on dust emissions. Additionally, it interesting that the G01 source gives much more satisfying central and east African emissions than the more physically based MB95 source. We have attributed this to a combination of poorly prescribed wind fields in this region and a higher required threshold in the MB95 source, although from the work of Liu and Westphal [2001], we feel it is worth considering the effects of surface layer stability and buoyancy on dust emissions (as discussed in section 3). We have shown that the timing of downwind dust events is largely uncoupled from the specifics of the dust lifting scheme, so we feel that future work to improve the characterization of the dust sources will need to focus on regions near the sources themselves. From our own work here and from that of Prospero et al. [2002] we are optimistic that satellite sensors (e.g., TOMS) can be a useful tool in this regard, particularly in places where ground-based data are difficult to obtain.

Acknowledgments

[83] We acknowledge support from NASA ESS fellowship NGT-30155 and NASA TOMS NAG5-11069. We also acknowledge D. Tanré, B. Chatenet, and F. Lavenu for kindly providing AERONET data, and O. Dubovik for assistance with the AERONET particle size distribution retrievals.