Western Africa: Mineralogical composition

14 In the last few years, several ground-based and airborne field campaigns have allowed 15 exploring the properties and impacts of mineral dust in Western Africa, one of the major 16 emission and transport areas worldwide. In this paper, we explore the synthesis of these 17 observations to provide with a large-scale quantitative view of the mineralogical composition 18 and its variability with source region and time after transport. 19 This work reveals that mineral dust in Western Africa is a mixture of clays, quartz, iron and 20 titanium oxides, representing at least 92% of the dust mass. Calcite ranged between 0.3 and 21 8.4% of the dust mass depending on the origin. Our data do not show a systematic 22 dependence of the dust mineralogical composition with origin, likely as in most of the cases 23 they represent the composition of the atmospheric burden after 1–2 days after emission, 24 when air masses mix and give raise to a more uniform dust load. This has implications for 25 the representation of the mineral dust composition in regional and global circulation models, 26 and satellite retrievals. 27 Iron oxides account for 58 ± 7% of the mass of elemental Fe, and between 2 and 5% of the 28 dust mass. Most of them are composed of goethite, representing between 52 and 78% of the 29 iron oxide mass. We estimate that titanium oxides account for 1-2% of the dust mass, 30 depending on whether the dust is of Saharan or Sahelian origin. 31 The mineralogical composition is a critical parameter to estimate the radiative and 32 biogeochemical impact of mineral dust. The results on dust composition have been applied 33 to estimate the optical properties as so as the iron fractional solubility of Saharan and 34 Sahelian dust. 35


Introduction
Mineral dust from wind-driven soil erosion is an important player in the climate system. It is emitted in large quantities from arid and semi-arid regions of the globe, mostly Africa, Asia, Australia and North America (Shao et al., 2011), accounting for about 2000 Mt to the annual global aerosol emissions. Once in the atmosphere, mineral dust has several climatic and environmental impacts, related to its ability to scatter and absorb radiation (both in the solar and the terrestrial spectrum) to act as giant cloud and ice nuclei, to alter the concentrations of some gaseous pollutants (such as ozone) and to provide nutrients to the ecosystems via dry and wet deposition, whereby also changing the surface albedo (Shao et al., 2011;Mahowald et al., 2011). The relevance of these phenomena can be expressed in terms of radiative forcing, whose magnitude, in terms of central values, is estimated to be as high as 0.5-1 W m −2 on the global scale . However, uncertainties remain important because of the spatial heterogeneity of concentrations,

Published by Copernicus Publications on behalf of the European Geosciences Union.
due to localised emission and short residence time in the atmosphere (1 week at the most, depending on particle size) but also because of the far from satisfactory knowledge of the underlying physico-chemical properties, composition, size and shape ruling the optical, chemical and physical interactions of mineral dust with radiation and the atmospheric, terrestrial or oceanic constituents (Formenti et al., 2011a). This is particularly the case with regard to composition. Dust is made of different minerals, whose proportions, at emission, depend on the mineralogy of the source region and surface wind speed relative to the soil roughness, which determines their size distribution (Marticorena and Bergametti, 1995;Alfaro et al., 1998). The various minerals -clays, quartz, potassic and sodic feldspars, calcium-rich carbonates and sulphates, iron and titanium oxides -have different mineralogical and crystallographic properties; therefore, they may act differently with respect to light extinction, absorption of water and surface reactivity (Pye, 1987). For example, iron oxides have been shown to control the interaction with light in the UVvisible spectrum (Sokolik and Toon, 1999), whereas quartz, clays and Ca-rich carbonates become important when looking at the infrared part of the radiation spectrum (Sokolik et al., 1998). Clays, but also feldspars, come in to play when it comes to investigating the capacity of dust to act as ice nuclei at low temperature (Hoose et al., 2008;Atkinson et al., 2013), whereas calcite (calcium carbonate) shows high surface reactivity with respect to various atmospheric pollutants (Crowley et al., 2010). Another clay family, the smectites, bears the most soluble iron which might become available as a nutrient to marine phytoplankton . The properties of mineral dust are size-dependent, and the dimensional spectrum of mineral dust particles at emission extends over various orders of magnitude from approximately 200-300 nm to 100 µm (Kandler et al., 2009), depending on wind speed at emission. As a consequence, the properties of mineral dust vary as a function of time after emission due to size-dependent deposition, mixing and gaseous condensation. Because these processes are not well established, the physico-chemical properties of mineral dust have to be determined experimentally at the various times of their life cycle. In this paper, we relate the mineralogical and chemical composition of mineral dust from western Africa at emission or at most within 1-2 days from emission. The southern Sahara is the location of two of the most persistent dust sources on the global scale: the Bodélé depression in between the Tibesti and the Ennedi mountain chains in Chad and the areas in between the east of Mauritania, the north of Mali and the south of Algeria (Prospero et al., 2002;Laurent et al., 2008;Washington et al., 2006;Shao et al., 2011). These sources are relevant on the global scale as their emissions are transported across the Atlantic Ocean towards South and Central America (Reid et al., 2003;Koren et al., 2006;Ben-Ami et al., 2009, 2010. The atmospheric content of mineral dust in western Africa shows a very pronounced seasonal cycle despite high vari-ability, both on the daily and interannual scales (N'Tchayi et al., 1994(N'Tchayi et al., , 1997Goudie and Middleton, 2001). This is largely explained by the alternance of two meteorological regimes. Throughout the year, the "Harmattan", a north-eastern dry wind, is responsible for intense dust emissions in the arid areas of the Sahara and very efficient transport towards the Atlantic Ocean and across the Sahel (Pye, 1987;Sokolik et al., 2001;Laurent et al., 2008). Wintertime is also the biomass burning season south of approximately 13 • N , and mixing between dust and biomass burning aerosols might occur during southward transport . During summer, the Intertropical Convergence Zone (ITCZ) displaces northward, allowing the surface monsoon flow to intrude into the continent from the south-west towards the continental Sahel. This shift induces a progressive organisation of convection from isolated cells to largescale, quickly propagating systems, which, particularly at the beginning of the monsoon season, often do not precipitate and are efficient in eroding bare soils and producing mineral dust (Marticorena et al., 2010;Abdourhamane Touré et al., 2011). The summertime column dust content over western Africa results, therefore, from the superimposition of local emission from the Sahel and remote transport of dust emitted in the Sahara (Tegen and Fung, 1994;Yoshioka et al., 2005;Marticorena et al., 2010).
The mineralogy of the Saharan and Sahelian source areas is different: the Sahara is mainly composed of Calcisols or Arenosols, whereas the Sahel is richer in weathered Plinthosols (Fe oxides, kaolinite, quartz) (Pye, 1987;Claquin et al., 1999;Caquineau et al., 2002;Nickovic et al., 2012;Journet et al., 2014). As a consequence, mixing or layering of dust transported from the Sahara and that emitted locally by convection over the Sahel should be detectable through differences in the composition over the atmospheric column.
Furthermore, western Africa is prone to land-use change because of the rapid demographic increase and desertification during drought periods (ECA, 2005;Dai, 2011), which could feed back as increased emissions (Carslaw et al., 2010). Rajot (2001) has shown that, in the Sahel, additional emission by wind erosion by mesoscale meteorological disturbances, such as organised convective systems, occurs on harvested soils only. These additional emissions, anthropogenic in origin, are not quantified to date (Boucher et al., 2013). To provide the quantitative estimate of the mineralogical composition of mineral dust on a regional scale, we report in this paper the synthesis of data from various aircraft and ground-based field campaigns which took place in western Africa in 2006 and 2007. These are the African Monsoon Multidisciplinary Analysis (AMMA) special observing period 0 (SOP0; Rajot et al., 2008), augmented by the airborne Dust and Biomass-burning Experiment (DABEX; Haywood et al., 2008); special observing periods 1 and 2 (SOP1-2; Reeves et al., 2010); the Dust Outflow and Deposition to the Ocean (DODO; McConnell et al., 2008) Haywood et al., 2011). Some of the data discussed in this paper have been published previously (Formenti et al., , 2011bRajot et al., 2008;Klaver et al., 2011). Here, they are revisited with a broader discussion of the mineralogical variability on the regional scale of western Africa and its implications for climate. This paper discusses whether, when mapped on the scale of western Africa, the regional variability in emission, deriving from the heterogeneity of the mineralogy of the parent soils, is still relevant or whether it becomes irrelevant when dust is airborne. This is important for developing a model parameterisation for dust in regional and global climate models or for satellite retrieval algorithms.

Field campaigns
AMMA was an international project to improve our knowledge and understanding of the West African monsoon (WAM), its variability on daily-to-interannual timescales and its effects on the state of the atmosphere . Because of their links to precipitation, wind speed, soil surface state and drought at different timescales, mineral dust aerosols play a role in this comprehensive scientific framework. The experimental strategy of the AMMA program was based on embedded multiyear, seasonal and intensive observation periods (Janicot et al., 2008;Lebel et al., 2010). The observations used in this paper were conducted during the intensive observations periods, called special observing periods (SOPs). The wintertime SOP (SOP0), was dedicated to the investigation of mineral dust and biomass burning, and their mixing; it took place between 13 January and 13 February 2006. The summertime observations (periods SOP1 and SOP2) started on 31 May 2006 and ended on 17 July 2006 and were dedicated to the detailed description of the interactions between aerosols and convection.
The ground-based observations were conducted at the AMMA supersite of Banizoumbou (13.5 • N, 2.6 • E; 250 m above sea level), situated in a remote location about 60 km east from the capital of Niger, Niamey. This site has been operational since the early 1990s, when the first measurements of soil erosion were performed on a cultivated field and a fallow (Rajot, 2001). Since 1995, the site has also been an Aerosol Robotic Network (AERONET) station, measuring columnar aerosol optical properties. The ground-based site operations are fully described in Rajot et al. (2008).
The aircraft observations were conducted from Niamey, Niger. The wintertime airborne campaign, called AMMA SOP0-DABEX, was conducted onboard the Facility for Airborne Atmospheric Measurements (FAAM) BAe-146 aircraft . The aircraft performed 13 research flights in the area between 8 and 18 • N, including three dedicated excursion to the north of Niger, toward the Aïr Mountains and the Ténéré desert, to probe specifically mineral dust which was being transported. In summertime, the observations were conducted onboard the Service des Avions Français Instruments pour la Recherche en Environnement (SAFIRE) ATR-42, which operated 20 research flights in the area between 6 and 20 • N, thus crossing the ITCZ to the north; this was done in order to contrast the Sahara and the Sahelian dust properties in their respective boundary layers (Reeves et al., 2010;Formenti et al., 2011b).
The GERBILS derived its scientific motivation from the discrepancy that had been found between the satellite observations and the numerical weather model predictions of the top-of-the-atmosphere outgoing long-wave radiation over desert areas in western Africa (Haywood et al., 2005). This discrepancy could be reconciled by including mineral dust in the model. The GERBILS experiment, whose full rationale and operating details are described in Haywood et al. (2011), was therefore designed to test this hypothesis, and to determine the properties of the mineral dust to be included in the model. The experiment consisted of 10 research flights onboard the FAAM BAe-146, performed between Nouakchott (18.10 • N, 15.94 • W; Mauritania) and Niamey (13.48 • N, 2.18 • E; Niger) on an almost constant route along the 18 • N meridian. Flights were performed between 18 and 29 June 2007, a period characterised by a persistent and widespread dust load.
Finally, the DODO airborne experiment took place at two different periods in 2006: in wintertime (7-16 February) and in summertime (22-28 August). The aim of DODO was to characterise dust as a possible nutrient for the marine ecosystems. For this reason, the FAAM BAe-146 aircraft was based in Dakar, in Senegal, and most of the 12 research flights (6 in the winter period and 6 in the summer period) were conducted over the Atlantic Ocean in the outflow from the African continent. However, two flights, one per season, were carried out inland above Mauritania in order to sample dust at or very close after emission (McConnell et al., 2008).

Ground-based
Aerosol sampling at the ground-based AMMA supersite of Banizoumbou was performed using two identical, purposebuilt, wind-oriented inlets designed for the AMMA field campaigns . By using the standard formulae of particle losses in inlets and tubings reported in Baron and Willeke (2001) and Hinds (1999), the cut-off diameter (50 %) efficiency of these inlets has been calculated to be approximately 40 µm. Each inlet leads to a sampling chamber containing seven different sampling lines, each of them dedicated to a different instrument. Multiple sampling lines are very convenient as they allow collecting various filter samples in parallel (differentiated and optimized in terms of sampling medium), exposure interval and flow rate as a function of the analytical technique to be applied. This also results in minimal manipulation of the filters after sampling. Samples were collected on acid-washed, 42 mm polycarbonate Nuclepore filters (0.4 µm pore size) mounted on plastic rings. Samples were stored in Petri dishes after sampling. Sampling time was of the order of hours and was varied depending on the aerosol mass concentration measured on-line by a tapered element oscillating microbalance (TEOM; Rupprecht and Patashnick, Albany, USA) .

Airborne
During the AMMA SOP0-DABEX, DODO and GERBILS campaigns, samples were collected onboard the FAAM BAe-146 aircraft. The aircraft is capable of carrying 2 crew, 18 scientists and a total scientific payload of up to 4000 kg for a distance of 3700 km with a ceiling of 35 000 feet and has a typical science speed of 110 m s −1 . For the typical scientific payload, flight patterns, ambient meteorological conditions and the proximity of diversion airports encountered during the campaigns, the average endurance of the BAe146 aircraft was approximately 5 h. Sampling was performed using a set of two parallel thin-walled inlet nozzles with a curved leading edge; the design was based on criteria for aircraft engine intakes at low Mach numbers (Andreae et al., 1988). The sampling system was operated at flow rates that averaged 120 L min −1 (at ambient pressure and temperature). The flow was adjusted to maintain slightly subisokinetic sampling conditions. The passing efficiency of the inlets has not been formally quantified. During AMMA SOP-0/DABEX, Chou et al. (2008) showed that the number size distributions of the aerosols collected on the filters (counted by electron microscopy) extended up to 10 µm diameters and were of comparable magnitude to those measured by wingmounted optical counters.
Aerosol particles were sampled by filtration through two stacked-filter units (SFUs) mounted in parallel. Each SFU can hold a maximum of three filters on sequential 47 or 90 mm diameter polyethylene supports. Only one stage was used during the campaigns, i.e. samples represented the aerosol bulk composition. Samples were collected only during horizontal flight legs lasting not less than 20 min in order to guarantee sufficient loading of the filter samples. As a consequence, at the aircraft cruise speed, each sample had, at best, a spatial resolution of 180 km. Each SFU consisted of a Nuclepore filter (Whatman) of nominal pore size 0.4 µm.
During the AMMA SOP1-2 campaigns, samples were collected onboard the SAFIRE ATR-42. With the necessary payload for the campaign, the aircraft had a maximum endurance of 4 h. Minimum flight altitude was 300 m over land. The ceiling was approximately 7 km. The aircraft was equipped with basic sensors for measuring the radiative, dynamic and thermodynamic properties of the atmosphere (Saïd et al., 2010).
Aerosol sampling was performed using the AVIRAD aerosol sampling system, newly developed for AMMA (Formenti et al., 2011b). It consists of an isoaxial and isokinetic inlet whose 50 % passing efficiency has been estimated to be 9 µm in diameter. At the cruise speed of the ATR-42 (93 m s −1 ) the delivered volumetric flow rate is 350 L min −1 , which was distributed to various instruments, including two SFU for collecting bulk aerosol samples. Samples were obtained by parallel filtration through 42 mm diameter polycarbonate membranes (nominal pore size 0.4 µm, Nuclepore, Whatman) on polyethylene supports. Samples were collected only during horizontal flight legs lasting no less than 20 min. Immediately after each flight, the loaded filters were stored in Petri dishes.
Ground-based operation with multiple sampling lines allowed for parallel sampling where the sampling duration was adapted to the characteristics (detection limit, saturation level) of the technique to be employed for analysis. On the aircraft, the two parallel samples available had to be manipulated further to allow the number of analyses one wanted to perform, also taking into account the fact that some analyses are destructive of the sample. Therefore, the handling protocol for aircraft samples consisted of analysing entire filters or portions thereof, depending on their load of mineral dust. A dedicated tool in polyethylene has been developed for cutting the 47 or 90 mm filters into halves, quarters or eighth without touching and, therefore, contaminating them.

Elemental composition
The total elemental concentrations for elements from Na to Pb were measured by wavelength dispersive X-ray fluorescence (WD-XRF) for samples collected on the ground and on the ATR-42 and particle-induced X-ray emission (PIXE) for the aircraft samples collected onboard the BAe-146. The reason for using different techniques for ground-based and aircraft samples is the fact that the WD-XRF sample holder set-up could not, at the time, host filters which are not supported on polyethylene frames. The details of the analytical protocols are provided in Formenti et al. (2008Formenti et al. ( , 2011b, McConnell et al. (2008), Rajot et al. (2008) and . Elemental concentrations are estimated with a 5 % error. An intercomparison exercise on samples generated in the laboratory from geo-standards of known and certified composition has shown that the two techniques yield comparable results within the error bars .

Mineralogical composition
The identification of the major minerals making up mineral dust (quartz (SiO 2 ), calcite (CaCO 3 ), dolomite (CaMg(CO 3 ) 2 , gypsum (CaSO 4 q 2H 2 O), clays (kaolinite, illite, smectite, chlorite), feldspars (orthoclase KAlSi 3 O 8 and albite NaAlSi 3 O 8 )) was performed by X-ray diffraction (XRD) analysis at the Institut de Recherche pour le Développement (IRD) in Bondy, France. The diffractometer, a Siemens D500 with Ni-filtered Cu-K α radiation, was operated at 40 kV and 30 mA. Samples were scanned from 2 to 70 • (2θ) in steps of 0.02. The analytical procedure and semi-quantitative treatment are fully described by Caquineau et al. (1997), who adapted the sample preparation to lowmass mineral aerosol (load deposited on filter > 800 µg). The calibration protocol that was applied to the XRD spectrometer in order to quantify the mineralogical composition is described in Klaver et al. (2011). The calibration curves were improved after publication of the Klaver et al. (2011) paper by calibrating the feldspars (orthoclase and albite). The calibration factors K i , representing the ratio between the peak surface area S i in the diffraction spectra and the mass m i of the i mineral, are reported in Table 1. The error in the mineral masses is estimated as the sum of the relative error in the peak area in the diffractograms and in the calibration factor K i : m i m i = K i K i + S i S i . The term K i /K i is estimated from the error in the slope of the calibration curve (Table 1), and it varies between 2 and 10 %. The term S i /S i depends on the abundance of the mineral in the samples and has been estimated from the counting statistics obtained.

Iron oxide content and speciation
The iron oxide content, defined as the fraction of iron which is not in the crystal lattice of silicates (Karickhoff and Bailey, 1973), was determined with the adapted citrate-bicarbonatedithionite (CBD) extraction method developed by Lafon et al. (2004). This method is an adaptation for aerosol filters (with typical material mass less than 500 µg) of the classical method of Mehra and Jackson (1960) for soil analysis. The method uses a reagent to dissolve iron oxides selectively via reduction. Details of this analysis are given in Formenti et al. ( , 2011b and Klaver et al. (2011). The percent error in the iron oxide content is estimated to be 10 % as its assessment is based on two XRF analyses (prior to and after dissolution of the iron oxides from the sample).
The speciation of iron oxides was performed by X-ray absorption (XAS) in the Fe K range: XANES (X-ray absorption near-edge structure) and EXAFS (extended X-ray absorption fine structure) at the SAMBA (Spectroscopies Applied to Materials based on Absorption, http://www.synchrotron-soleil.fr/portal/page/portal/ Recherche/LignesLumiere/SAMBA) beam line SOLEIL synchrotron facility in Saclay, France. XAS spectroscopy is based on the analysis of the position and shape of the K pre-edge and edge peaks, depending on the oxidation state of iron but also on the atomic position of the neighbouring atoms, mostly O + and OH − . The quantification of the Fe status was based on the analysis of five standards of Fe(III)-bearing minerals which can be found in mineral dust. Full details of the experimental protocols and the data analysis are provided in Formenti et al. (2014).

Source identification
The identification of source regions has been performed using a variety of ancillary products: 3. back trajectories and air mass dispersion calculations using the Met Office Numerical Atmosphericdispersion Modelling Environment (NAME; Cullen, 1993) and the NOAA HYbrid Single-Particle Lagrangian Integrated Trajectory Model (HYSPLIT) as described in Chou et al. (2008), Rajot et al. (2008) and Klaver et al. (2011).

Calculation of the total mass
In the following, the composition of mineral dust will be presented as a percentage of the total dust mass (TDM). According to Pye (1987), the TDM can be decomposed as where m clays is the mass of the major clay species (kaolinite, illite, smectite, chlorite); m quartz is the mass of quartz (SiO 2 ); m Ca-rich = m calc + m dolom + m gypsum is the mass of calcium carbonates and sulfates in the form of calcite (CaCO 3 ), dolomite (CaMg(CO 3 ) 2 ) and gypsum (CaSO 4 ); m feldspars is the mass of feldspars, such as albite and orthoclase (NaAlSi 3 O 8 and KAlSi 3 O 8 , respectively); m iron oxides is the mass of hematite (α-Fe 2 O 3 ) and goethite (α-FeOOH); m tit oxides is the mass of titanium oxides (TiO 2 ); and m organic is the mass of organic matter which is present in the soils as biological debris and complex organic molecules (humus). In practical terms, the TDM can be approximated by the TEDM (total estimated dust mass) as the sum of oxides corresponding to the major elements in mineral dust (Al 2 O 3 , SiO 2 , Fe 2 O 3 , CaCO 3 , K 2 O and Ti 2 O) whose masses are measured by XRF analysis. The TEDM estimated in this way can be compared to the direct measurement of the total dry gravimetric mass (TDGM) (which is available for some of the data sets (ground-based AMMA SOP0 and SOP1-2 and GERBILS)), either by weighing or by on-line Tapering Element Oscillating Microbalance (TEOM, model 1400a, Rupprecht and Patashnick Albany, New York, USA) (see Rajot et al., 2008;Klaver et al., 2011).
The slope of the least-squared linear regression between the TEDM and TDGM, calculated for the 81 samples for which the extraction of iron oxides was performed (see Sect. 3.3), is 0.98 and the coefficient of determination (R 2 ) is 0.88. This is very satisfactory, taking into account that, because the mass is dominated by coarse particles, which have a large mass but small number, differences might arise through the loss of some coarse particles between the time at which the filter was weighed and the time at which it was analysed by XRF or by small differences between the filter sampling duration and the TEOM integration times. As a matter of fact, most of the scatter in the data is observed for samples which corresponded to local erosion or transported emission by convective systems and which had the largest fraction of coarse particles and the shortest integration times. This becomes apparent the ground-based AMMA SOP0 and SOP1-2 data ( Fig. 1) are plotted separately.
Therefore, we consider that the TDM is equal to the TDGM for samples for which the TDGM is measured. When this is not available, the TDM is approximated via the TEDM, calculated as the sum of oxides. The percent uncertainty is calculated as the squared sum of the errors in the calculation of the TEDM from the measured Al, Si, Fe, Ca, K and Ti elemental masses. In this way, the percent uncertainty in the TDM comes to 12 %.

Calculation of the mineralogical composition
The mass apportionment presented in Eq. (1) has been performed as follows: 1. The masses of quartz, calcite, dolomite, gypsum and feldspar (m quartz , m Ca-rich , m feldspars ) were quantified by XRD using the mass/intensity calibration curves as described in Sect. 2.3.
2. The mass of iron oxides (m iron oxides ) was measured directly by CBD analysis. The division of m iron oxides into its main components (hematite and goethite) was performed by XANES spectroscopy.
3. The mass of the clay species, kaolinite, illite, smectite and chlorite (m clays ) was estimated by the difference between the TDM and the sum of the masses of m quartz , Atmos. Chem. Phys., 14, 10663-10686, 2014 www.atmos-chem-phys.net/14/10663/2014/ m Ca-rich , m feldspars and m iron oxides . The division of m clays into its constituent species is difficult to achieve without a direct calibration. This calibration is highly uncertain as there is basically no mineralogical reference with the same crystallographic and chemical status as in the actual aerosol samples because of the exposure to the environment, weathering in particular, that they experience before and after becoming airborne. Furthermore, smectite clays are actually a family of species with different chemical characteristics. When only illite and kaolinite are present in the samples and no traces of chlorite or smectite are shown by XRD analysis, their mass can be apportioned readily using the illite-to-kaolinite ratios (I / K) established by Caquineau et al. (2002) as a function of dust source region as When smectites or chlorite are also detected, Eq. (2) does not hold true. In the following sections, we will evaluate the uncertainties induced by this approximation.
4. The mass of titanium oxides m tit oxides was estimated by calculating the TiO 2 contribution as the measured elemental Ti multiplied by 1.67, the ratio between the atomic weight of TiO 2 and that of elemental Ti.
5. The mass of organic compounds (m organic ) in mineral dust was generally ignored on the basis of the work by Lepple and Brine (1976); it should not exceed 3 % of the TDM.
6. We also did not take into account the mass of diatomite fragments which have been observed in samples originating from Bodélé . Diatomite fragments are silicate skeletons of algae and they do not have a crystalline structure which can be quantified by XRD analysis. Their presence can be proven by elemental analysis showing excessive values of the Si-to-Al ratio as high as 4 (Formenti et al., 2011b). The diatomite mass can, therefore, be estimated by comparing the excess SiO 2 mass calculated from elemental Si concentrations with respect to the mean ratio to Al to the mass of quartz estimated by XRD analysis. When doing so, the excess SiO 2 mass is comparable to that of quartz except for the few samples for which dust originated from the Bodélé depression; in this case their contribution to the TDGM is in the range 6-13 %. This contribution is accounted for in the error bars of the clay fraction for these samples.

Results
The whole data set (ground-based and airborne AMMA SOP0 and AMMA SOP1-2, GERBILS and DODO campaigns) consists of 704 samples, for which we measured the total elemental composition. Out of this total, 54 samples have been selected to provide data on the major mineral composition, 86 have been analysed to provide the content of iron oxide and 12 to determine the iron oxide speciation. 31 samples combined sufficient information to determine the full mineralogical composition. The sample selection has been based on the filter loading and on the possibility of identifying the samples' provenance and attributing a source region to them.

Identification of source regions
Identifying the origin of the mineral dust collected on the samples implies that one is able to make the link between the knowledge of the position and the mineralogy of the source regions and that of the aerosol samples. To do so, one must know (1) the transfer function, at emission, linking the soil mineralogical/chemical composition to that of the aerosols and (2) the transfer function describing how its composition is modified during transport. The first transfer function describes the mineralogical fractionation that occurs between the soil and the aerosol as a consequence of the size segregation that happens when the soil grains saltate on the soil surface (Rahn, 1976). The fractionation can alter the elemental ratio between Si and Al as a result of the depletion, at emission, in quartz and feldspars with respect to clays. Quartz and feldspars have higher Si content than clays and are more abundant in the sand than in the silt/clay fractions of the soil (Mason, 1966;Rahn, 1976;Chatenet et al., 1996). The latter are enhanced during emission by saltation/sandblasting (Alfaro et al., 1998). Whereas there is still no systematic assessment of the transfer function between the soil and the aerosol composition, Claquin et al. (1999) have shown that some mass ratios between minerals (for example, that of illite to kaolinite and that of calcite to quartz) are conserved from the soil to the aerosol if the comparison is restricted to roughly the same size fractions (clays or silt). Lafon et al. (2006) also suggested that the free-to total-iron ratio is preserved. Jeong (2008) has shown that the calcite and the K-feldspar fractions are very similar in Asian dust and the corresponding silty soil fraction. The percent of phyllosilicate is higher in the dust aerosols than in the soil, whereas the quartz and plagioclase fractions are lower.
The second transfer function describes how the dust composition is modified during transport. Our data set consists of data which have been collected at most 2 days after emission. We expect that only particles of a diameter larger than 10 µm would be lost in this short time (Pye, 1987). We therefore assume that the chemical/mineralogical composition of To identify the pathway of air masses during transport and trace them back to an active dust source, we used back trajectory and dispersion modelling and various satellite products (e.g. Chou et al., 2008;Klaver et al., 2011). This approach does not apply to samples collected in the boundary layer during the monsoon season south of the ITCZ. In this case, the dominant flow is from the south-west: dust is present in the boundary layer only as a result of erosion by organised convection or under particularly stable conditions when high temperatures are established and the top height of the thermal turbulence in the boundary layer rises to or above the shear level, bringing dust from the free troposphere to the surface level (Flamant et al., 2009;Formenti et al., 2011b).
Even when this is the case, this approach has various caveats: first, if their horizontal velocity is high and exceeds the local threshold velocity, air masses within the boundary layer might get loaded with dust along their way before reaching the sampling point. This happens, for example, when emission is induced by large-scale frontal systems as the one of 3-7 March 2006 (Tulet et al., 2008). Secondly, ground-based sampling lasts several hours, and the direction of air masses might change in between. Finally, aircraft samples cover large areas to which different air masses from different source regions might contribute. This problem might be overcome, or at least minimised, by increasing, as much as possible, the sampling frequency to match the variability of transport. We did that by following the variability of concentrations, in time or space, with ancillary measurements at higher temporal resolution (mass concentration by the TEOM balance or the scattering coefficient by a TSI nephelometer).
In spite of this, we cannot exclude the possibility that our samples might represent larger emission areas than single "hot spots" (see, for example, Fig. 1 in Klaver et al., 2011). We do not necessary regard this as a problem, as the atmospheric dust load, both in short-and large-scale transport, is often due to a variety of sources active at the same time (Glaccum and Prospero, 1980;Reid et al., 2003;Formenti et al., 2003;Schepanski et al., 2007;Laurent et al., 2008;McConnell et al., 2008;Marticorena et al., 2010). Also, this supra-imposition, both in the vertical and in the horizontal, is what is actually seen by spaceborne sensors. The results of this paper might then be used to build a dust composition model on a scale relevant to satellites.
When doing so, we found that samples presented in this paper are representative of major African sources which, following  and Scheuvens et al. (2013), are indicated as PSAs (potential source areas) in Fig. 2: -The PSA2, including arid areas in Western Sahara, Mauritania and Morocco, was sampled during the DODO and the GERBILS campaigns. With respect to the extent of this source, the northern part is underrepresented as it was sampled on one occasion only.
-The PSA3 represents an emitting area in northern Mali and southern Algeria. This is the dominant source in terms of frequency of sampling. The air masses travelling from PSA3 often show a recirculating pathway over northern Niger. By doing so, they may at times pick up dust from the desert areas close to the Aïr massif, which is not depicted in Fig. 2, before reaching the sampling site. These occurrences, when identified, are labelled as PSA3-NN.
-The PSA4 was sampled on a few occasions during wintertime. As for PSA3, samples originating from PSA4 which involved the air masses coming into contact with the surface in northern Niger when satellite images indicated emissions have been labelled as PSA4-NN.
-A few samples also represent dust originating from the Bodélé depression (PSA5). In summertime, a mixture of Sudan and Chad sources, including Bodélé, may have been sampled (Flamant et al., 2009). Finally, a number of episodes of erosive emissions by convection in the Sahel were also sampled and characterised (referenced as Sahel). These occurred both over Mali and over Niger.
The PSA1 source is not discussed in this paper as it was not sampled during the field campaigns. To provide a first link between geographical origin and chemical composition, we made a scatter plot of elemental Si / Al versus Fe / Ca, which we colour-coded according to the source region and to satellite, back trajectories and dispersion modelling information. These ratios have been shown previously to be fairly robust indicators of the origin of mineral dust on a large scale (Formenti et al. (2011a) and references therein). The result is shown in Fig. 3.

Atmos
There are some clear distinctions that can be made. Data points having Si / Al > 3 are unambiguously identified as samples from the Bodélé depression, i.e. PSA5. Their Fe / Ca ratio is about 1. Data points having Fe / Ca > 3 are related to emission from erosive events in the Sahel. There are various sub-ranges in this group. Data having Fe / Ca in the range of 3-5 correspond to transported dust aerosol from the Sahel, sampled 1-2 days after emission and corresponding to mixed sources, such as PSA3 + Sahel. Data having Fe / Ca > 5 correspond to samples collected immediately or at a very early stage after emission by convective events, mostly locally at the ground-based Banizoumbou site (when the Fe / Ca ratio exceeds 10) but occasionally from uplifted dust onboard aircraft.
Data with Si / Al in the range of 2.5-2.9 and Fe / Ca in the range of 1-3 have their origins mostly at latitudes in the range of 16-27 • N, i.e. PSA3. Whereas the variability of the Si / Al ratio is low, that of Fe and Ca can be used to make some distinctions amongst sources. Fe / Ca higher than 1 is measured in data points from northern Mali, Mauritania, Algeria, northern Niger and Libya whereas Fe / Ca < 1 is measured in dust from Western Sahara (PSA2). For this source, there are two clear outliers corresponding to samples of dust originating from Morocco and showing the lowest Fe / Ca ratio (∼ 0.4).

Mineralogical composition
The investigation of the mineral composition is based on 51 samples, 15 of which were collected onboard an aircraft. This data set represents a diversity of sources and emission conditions.
The most evident peaks, for all samples, are those of illite, kaolinite and quartz. Less evident, and not ubiquitous, are the peaks corresponding to feldspars (albite and orthoclase) and calcium-rich minerals (calcite, dolomite, gypsum). Feldspars are found in the samples from local erosion and southern Algeria, northern Niger, Mauritania and Western Sahara. This last set might also contain calcium-rich minerals. Smectite clays are detected as a large and rather weak peak at angles comprised between 4 and 7 • . This is indicative of the fact that smectites are a family of weathered clays of different chemical composition (e.g. montmorillonite, nontronite), whose crystalline structure can be heavily modified in the environment. Smectite clays were only evident in samples originating from southern Algeria, northern Niger, Mauritania and Western Sahara.
We can obtain the mass fractions of minerals which can be detected by X-ray diffraction by using the calibration factors as described in Sect. 2.3.2 (Fig. 4).
Regardless of origin, clays and quartz account in average for 87 % (±6) and 10 % (±6) of the diffracting mass. There is only one clear outlier in an episode of intense local erosion sampled at the Banizoumbou supersite (sample SOP1-8) for which the percent clay fraction is reduced (56 %) and the quartz and feldspars fraction enhanced (41 and 3.3 %, respectively). This is consistent with the fact that, at erosion, the size distribution shows an enrichment of coarse particles whose composition should be closer to that of the soil silt fraction (Gillette and Walker, 1977;Chatenet et al., 1996). The analysis of the XRD spectra also show that kaolinite and illite are very often the only clay species detected. A broad peak corresponding to smectites was detected on samples from Algeria and, at times, Bodélé, consistent with the finding of Moreno et al. (2006). No chlorite has been detected in our samples. As explained in the subsection "Calculation of the mineralogical composition" of Sect. 2.4.2, the relative proportions (by mass) of illite and kaolinite have been estimated using the origin-dependent ratios established by Caquineau et al. (2002). For the sources considered in this data set, the illite-to-kaolinite ratios could vary between 0.1 and 2.3. The central and western Saharan sources (PSA2 and PSA3) display the highest values (1.6 to 2.3) for PSA5; northern Niger and Sahel display the lowest values (0.1 to 0.7).

Iron oxide content and speciation
The quantification of the content of iron in oxide form (Fe ox ), that is, free iron which is not in the crystal lattice of aluminosilicates, was performed for 81 samples of different origins and load (33 of these were collected onboard aircraft). On average, Fe ox accounts for 58 ± 7 % of the mass of elemental Fe. The range of variability is illustrated in Fig. 5.
The lowest values (down to 44 %) are found for samples of dust originating from the PSA4 (Chad/Egypt) area and for a set of samples collected above the PSA2 areas. The highest values (up to 71 %) are obtained for samples of dust emitted by local erosion in the Sahel, including one episode in which uplifted dust was collected onboard an aircraft after emission. Other subsets of samples, such as those of dust originated from the southern Algeria/northern Niger areas, display equally high values when, during long-range transport, dust from Algeria mixes with locally produced dust in Niger. It should also be noted that, because of the internal variability of each group, there is no statistically significant difference between the iron oxide fractions of the total iron for long-range-transported dust originating in latitudes north of 16 • N. A broad distinction between the Fe ox -to-Fe ratio of Sahelian samples with respect to those of Saharan samples can be made by sorting the data points by their corresponding Fe / Ca ratio. Mineral dust emitted from the Sahel (Fe / Ca > 3) is characterized by a mean Fe ox -to-Fe ratio value of 65 % (±5 %, standard deviation of the mean), whereas episodes of transport from the Sahara (Fe / Ca < 3) are characterized by a mean ratio of 56 % (±6 %). No further distinction based on origin is possible in this second group of samples. This is illustrated in Fig. 6.
X-ray absorption spectroscopy on 12 samples of differing origins indicated that goethite is, overall, the dominant species of iron oxides. Goethite accounts for 48 to 73 % of the iron oxide mass, whereas hematite accounts for 27 to 52 %. These values do not show a clear trend with respect to the sample origin. The relative proportions of iron in the form of hematite and goethite with respect to total iron are shown in Table 2.

Titanium oxide content
Titanium oxides deserve attention as they are involved in photocatalytical heterogeneous reactions with various atmospheric constituents, including volatile and semi-volatile organic compounds, which are found in transport regions of mineral dust due to biomass burning emissions (Gustafsson et al., 2006).
We therefore examined our data set in order to provide estimates of the content of titanium oxides in mineral dust from various sources in western Africa. The titanium oxide Atmos. Chem. Phys., 14, 10663-10686, 2014 www.atmos-chem-phys.net/14/10663/2014/ content has been estimated from the measured elemental Ti concentration (obtained by XRF), which is converted to TiO 2 using the appropriate molar mass conversion factor (1.67). This should be considered as an upper-limit estimate as some of the titanium in mineral dust is present as a substitution element in aluminosilicates (Dolcater et al., 1970). Figure 7 presents the correlation between TiO 2 and the TDM. Two populations can be identified: one corresponding to samples collected in relation to erosion events and one corresponding to advection events. For both populations, the correlation is excellent (R 2 equal to 0.97 and 0.92, respectively) and the scatter of points is minimal. The TiO 2 content is higher in erosion events from the Sahelian region (average 1.2 % (±0.1) of the mass) than for advection events of Saharan dust, for which TiO 2 represents, on average, 0.71 % (±0.01) of the mass.
Looking at the elemental ratios between Ti, Fe and Fe ox provides further insights (Fig. 8). The ratio of Ti to Fe is insensitive to changes in the proportions of Fe ox to Fe up to 60 % and then starts increasing linearly with increasing Fe oxto-Fe ratio. This suggests that, in Sahelian dust, Ti is associated with iron oxides but that is not the case for Saharan dust.

Mass apportionment of the mineralogical composition
By combining the information presented in Sects. 3.1 to 3.4, we can estimate the mineralogical composition of samples of mineral dust of local or distant origin. The mass apportionment is shown in Table 3. Note that we provide, here, the composition of the bulk mineral dust, without explicitly taking into account the dependence on size.
There are clear similarities in the composition of the dust from the different origins, first of all in the fact that the clay fraction represents the largest fraction of the dust mass. Also note the absence of Ca-rich minerals (calcite, dolomite, gypsum) in dust originating from the Sahel compared to that originating elsewhere.
Despite the overall similarities, there is indeed a large internal variability within each source group, expressed by the standard deviation around the mean; this might be related to either the time or height after transport.
An example is provided by the composition of the four samples in the Bodélé/Sudan group: one sample was collected in the dry season, whereas the remaining three are sequenced samples collected in the wet season, a time characterised by the absence of local emission but also by a progressively increasing entrainment of dust transported from the free troposphere to the surface layer (J. L. Rajot, personal   . This variability corresponds to differences in the volume size distribution, normalised to the total volume for sake of comparison ( Fig. 9). As a matter of fact, the volume size distribution of the dry-season sample presents a large coarse mode peaking around 4-5 µm, whereas the wet-season samples have a bimodal volume size distribution with an additional mode centred at 8 µm. Within this group, the sample having the largest quartz fraction is also the one displaying the larger fraction of particles above 10 µm. A second example of the links between time after emission, size distribution and composition is given in Fig. 10. This shows the normalised particle size distributions (by volume) for one case of Sahelian dust by local erosion (sample SOP1-8), two samples representing the background dust composition in the absence of local erosion (SOP1-15 and SOP1-60) and one case representing the composition of Sahelian dust advected at the sampling site after having been emitted by local erosion nearby (SOP1-17). There are clear differences between these cases: sample SOP1-8 has a mode at 8 µm and the largest coarse-particle fraction; Samples SOP1-15 and SOP1-60 have a bimodal size distribution with modal diameters at 3-4 and 8 µm; finally, only the mode at 3-4 µm remains for sample SOP1-17. As for the case presented in Fig. 9, it is evident from Fig. 10 that the coarse mode at 8 µm bears most of the quartz and feldspar mass fractions and that the dynamic evolution of this mode controls the content of quartz and feldspar in dust. The comparison of Figs. 9 and 10 also suggests that, once the mode at 8 µm is deposited, differences in the mineralogical composition of mineral dust will be more independent of the size dynamics and reflect better the differences in the mineralogy of the parent soil.

Comparison with other data sets
All comprehensive papers relating to the composition of mineral dust state that the comparison of compositional data from different data sets depends critically on the experimental conditions, in particular the duration and location of sampling (integration time, near-source location) (Formenti et al., 2011a;Scheuvens et al., 2013). These considerations are particularly relevant to aircraft sampling, where the difficulty in controlling the size cut-off of samplers generally does not favour the coarse-particle fraction in an often nonquantified way (Formenti et al., 2011a). Furthermore, as the aircraft is moving during sampling, a 20 min aircraft sample will represent a distance of the order of 120 km, which might Atmos. Chem. Phys., 14, 10663-10686, 2014 www.atmos-chem-phys.net/14/10663/2014/ combine the contribution of various sources . As a consequence, the integration and the comparison of different data sets of the composition of mineral dust is not a straightforward exercise.
Overall, our results indicate the dominance of clays in the dust composition when only the aerosol fraction is measured. In this respect, they are in accordance with previously published papers for dust collected in Africa (Kandler et al., 2009) or after long-range transport (Delany et al., 1967;Kiefert et al., 1996;Glaccum and Prospero, 1980;Caquineau et al., 2002). These measurements, however, tend to indicate that the clay fraction accounts for ∼ 60-70 % of the dust weight, whereas our measurements suggest higher values (80-90 %). It should be remembered that the determination of the mineralogical composition will depend on the crystallography of the standard minerals used for calibration, and on the normalisation to the total dust mass. In our work the percentage of clays in dust depends directly on the normalisation factor, regarding which we have an uncertainty of the order of 17 % as we estimate clays from the difference between the total dust mass and the calibrated mass of non-clay compounds. As an example, Glaccum and Prospero (1980), using standard clays minerals for calibration, estimate that the crystalline mass represented 70 to 110 % of the weighed mass. These values reflect, at least partially, the uncertainty regarding the clay fraction.
Nonetheless, in some cases differences might be due to the sampled aerosol size distribution. As a matter of fact, the lowest detected percentage of clays corresponds to sample SOP1-8, collected during an intense episode of local erosion. In this case, clays account for around 50 % of the measured mass, in accordance with measurements by Caquineau et al. (1997Caquineau et al. ( , 2002 in similar conditions. The influence of size on composition is also clear, for example, by looking at the data published by Kandler et al. (2009), who documented the dust composition in a dust storm (total suspended particulate concentrations higher than 300 000 µg m −3 ) and in low-dust periods (total suspended particulate concentrations around 100 µg m −3 ) in Morocco. Quartz accounts for as much as ∼ 65 % in dust storm conditions, suggesting that soil grains, and not only aerosols, were sampled. Conversely, low-dust conditions are characterised by only about 20 % of quartz. These authors showed that, for locally emitted dust, illite is dominant over kaolinite (up to 30 and 5 % by volume, respectively). These authors also detected clay in the form of chlorite (less than 10 % by weight) and calcite (up to 15 %).
Few authors have published data on the iron oxide content. Lafon et al. (2004) evaluated the free-to-total iron content in samples collected at the Banizoumbou ground-based site and found mean values of 0.44 (±0.11) and 0.65 (±0.04) for dust transported in the Harmattan flow (from the Sahara) and dust emitted locally by convective erosion. Lazaro et al. (2008) found slightly lower values than ours ranging from 0.26 to 0.63 but only exceeding 0.56 on three occasions when back trajectories indicated transport from the Sahel or a mixture of dust originating from the Sahara and the Sahel. Our values are higher than those of these authors. Nonetheless, the back trajectories presented in the paper seem to indicate that the Saharan dust originated from higher latitudes than our samples -and mostly from Tunisia and Algeria. It will be interesting in the future to sample those source areas to confirm these rather low values.
Finally, our results on the iron oxide speciation are consistent with findings of different authors using XANES speciation (Wilke et al., 2001;Ohta et al., 2006;Prietzel et al., 2007;Majestic et al., 2007) and visible spectroscopy (Arimoto et al., 2002), indicating that Fe(III) is the dominant redox form of iron oxides and that goethite is the predominant form of iron oxide over hematite.

Atmospheric implications
The results showed in this paper suggest that, when restricted to the mass fraction below 20 µm, the dust composition in western Africa may apparently be simplified to a few minerals, clays, quartz, calcite, iron and titanium oxides. Feldspars and other calcium-containing minerals such as dolomite and gypsum are very minor components, close to detection limits. The geographical distribution of the dust composition is relatively unimportant in terms of mass fractions, in particular in comparison to the importance of the size distribution, whose variability with time seems to be able to explain some differences observed between samples from the same source region at different times after emission. Nonetheless, it is important to evaluate whether differences in composition on the regional scale, even when minor, might give rise to differences in the impacts of mineral dust.  (1987) 2. 5-200 Glotch et al. (2007) 5-100 Kaolinite Egan et Hilgeman (1979) 0.185-2.6 Roush et al. (1991) 5-25 Glotch et al. (2007) 5-100 Calcite Querry et al. (1978) 0.2-25 Long et al. (1993) 2.5-300 Dolomite Barthelmy (2007) 0. 185-2.5 Querry (1987) 2.5-40 Albite Barthelmy (2007) 0.185-2.5 Orthoclase Barthelmy (2007) 0.185-2.5 Gypsum Barthelmy (2007) 0.185-2.5 Long et al. (1993) 2.5-300 Goethite Bedidi and Cervelle (1993) 0.46-0.7 Glotch and Rossman (2009)  To do so, we explore the influence of dust composition on the spectral complex refractive index ñ and the fraction iron solubility.

Implications for optical properties
As a mean of evaluating the effect of the variability of composition on dust optical properties, the volume-average refractive index has been calculated as where f j is the volume fraction of each individual mineral in the dust sample (as reported in Table 3) and ñ i is its complex refractive index. Equation (4) assumes that minerals in dust are internally mixed. In reality, minerals are present in dust as external mixtures, with the exception of iron oxides, which may be found as inclusions in clays (Sokolik and Toon, 1999). However, the internal mixing hypothesis is frequently used in experiments and climate models (e.g. Balkanski et al., 2007;Kandler et al., 2009;McConnell et al., 2010;Hansell et al., 2011).
Values of the complex refractive index ñ j of the individual minerals detected in our samples are listed in Table 4. We have restricted the calculations to two spectral domains: the near UV-visible between 370 and 700 nm and the thermal IR from 8 to 15 µm, where values for all individual minerals are available.
Results are presented in Fig. 11, where they are compared to the values of the "desert" dust aerosol model in the OPAC (Optical Properties of Aerosols and Clouds) database (Hess et al., 1998), often used in satellite retrieval algorithms in the thermal infrared (see Table 1 in Klüser et al., 2012) but also in radiative transfer calculations Haywood et al., 2005). The "desert" model in the OPAC database represents the properties of mineral dust in the source region, with the dust considered to be a mixture of quartz and clays in different size fractions.
We restrict Fig. 11 to the visible and infrared imaginary parts of the complex refractive index ñ i . In the visible, the imaginary part of the complex refractive index is relatively constant with wavelength. Mean values are of the order of 0.003, with standard deviation of the order of 0.001, with the exception of data from the PSA4 source, which average at 0.006 and present a standard deviation of 0.003. It is difficult to judge this variability as it is induced by one sample only, representing an outstanding iron oxide fraction. In the infrared, the imaginary part shows a band-type structure relating to the major absorption bands of clays and quartz, mostly found in the atmospheric window between 8 and 11 µm. There is very little sample-to-sample variability in the position of the bands or in the relative proportions of the band peaks, with the exception of the Sahelian dust samples, which present a larger standard deviation due to the enhancement of the band at 9.2 µm corresponding to the outstanding quartz fraction of sample SOP1-8. There is very little difference in the spectra of sources PSA3, PSA4 and, to a lesser extent, PSA5, which have the highest absorption values at 9.6 µm. However, the spectral dependence for dust from source PSA2 is different (peak at 9.2 µm more pronounced than that at 9.6 µm).
The composition-based refractive indices calculated in this study are very different to those of the OPAC database. Below 0.6 µm, the volume-averaged values are lower than the OPAC values, with the exception of values for the PSA4 samples, due to their larger standard deviation. Above 0.6 µm, the two data sets converge. The spectral dependence is also inconsistent and reflects the reduced spectral variability of the values for iron oxides (both hematite and goethite) in Bédidi and Cervelle (1993) with respect to other commonly used databases, such as that of Shettle and Fenn (1979) for hematite on which OPAC is based.
The calculated volume-average infrared values in the absorption band between 8 and 10 µm are two times higher than in the OPAC database. The main reason for this is the volume-mixing rule, which enhances absorption. Additionally, for single-mineral species, various authors (Hudson et al., 2008a, b;Mogili et al., 2007Mogili et al., , 2008 found strong differences in peak positions and spectral shape between extinction Atmos. Chem. Phys., 14, 10663-10686, 2014 www.atmos-chem-phys.net/14/10663/2014/ Figure 11. Imaginary part of the complex refractive index calculated from the mineralogical composition of dust originating from sources PSA5, PSA4, PSA3, local erosion in the Sahel, and PSA2 (from top to bottom). Panel (a) represents the real part in the 0.3-0.7 µm, calculated assuming the refractive index from Table 3. Grey-shaded curves represent results obtained when using the data from Bedidi and Cervelle (1993) for hematite, whereas blue-shaded areas correspond to values from Shettle and Fenn (1979). Panel (b) represents the imaginary part of the complex refractive index calculated in the 6-16 µm spectral domain. The black line with white diamonds represents data from the OPAC database (Hess et al., 1998). spectra measured by Fourier transform infrared spectrometry and those calculated from published optical constants. All these studies show residual differences suggesting that optical constants might vary depending on whether they had been determined using powders or bulk samples, on the degree of water association of minerals and on their chemical form (substitutions, impurities). As this variability might be even larger when looking at actual dust samples, the comparison clearly suggests that new measurements of IR extinction spectra on real dust samples representative of the mineralogical composition of the different source areas are required.

Implications for Fe fractional solubility
The solubility of dust iron is a function of iron speciation, i.e. oxidation states, bonding environments and mineralogy Schroth et al., 2009). Thus, due to higher solubility of iron trapped in the crystal lattice of aluminosilicates and the large abundance of clay minerals, Journet et al. (2008) suggested that the use of hematite as a single source of DFe in biogeochemical models might cause an underestimation of the DFe supply to the ocean. On this assumption, Paris et al. (2011) showed that the iron fractional solubility in mineral dust can be estimated as the sum of the Fe solubility of various minerals compounding iron in the dust matrix. We propose to apply this calculation to this data set to estimate the solubility of dust in the source region. For this purpose, we use the iron solubility from minerals measured by Paris et al. (2011) at pH = 4.7, the typical pH of African rainwaters (Desboeufs et al., 2010), i.e. SFe equal to 0.17 % for illite and 0.006 % for free iron from goethite and hematite. Paris et al. (2011) made the approximation that all the iron associated with clay is present as illite. Journet et al. (2008) emphasized that even the iron in impurities in kaolinite or feldspars matrices present a high solubility. Our measurements show the predominance of kaolinite in comparison to illite, so we consider also the dissolution of Fe from kaolinite due to the large proportion of this mineral in the studied samples. We use SFe equalling 0.63 % for kaolinite (P. Paris, personal communication, 2013). The percent quantity of feldspar being usually inferior to 1 % in the samples, we neglect the iron associated with these minerals. Smectites, such as montmorillonite, are also minerals with a high content of iron with a higher solubility . However, due to the difficulties in quantifying the smectite fraction, this other source of structural Fe is not well constrained. Moreover, Shi et al. (2011) mention the presence of highly soluble nanoparticles of ferrihydrite or poorly crystallised iron in the fine fraction of soils. However, the quantification of this iron phase is obtained by sequential extraction and is difficult to apply on field samples due to the low mass. Therefore, we are not able to consider this species in our calculation. Among the samples present in our database, the mineralogical composition including goethite, hematite, illite and kaolinite is evaluated for 27 samples, enabling us to calculate the iron fractional solubility (Table 3). In this case, the iron content in illite is assumed to be equal to 4.0 ± 0.9 %, which corresponds to the mean stoichiometric value of two types of illites (Fithian, Illinois, and Rochester, Minnesota, as shown in Journet et al., 2008). The iron content in illite can vary from 0.8 to 8.4 % (Murad and Wagner, 1994); this large range is due to the fact that Fe is present as a substitution of Al or K in the sheet-like structure of illite, depending on the environmental conditions that the clay has experienced in the soil fraction (weathering, etc.). In goethite and hematite, we estimated the iron content based on the analysis of Journet  Mestdagh et al. (1980) showed that the iron content is related to the crystallinity of the mineral and ranges from 0.02 to 0.81 % with an average of 0.4 % ± 0.2 %, which is used for our calculation.
In order to gain a wider view of the variability of iron solubility as a function of emission source, we can also calculate iron solubility from the partitioning between structural and free iron as obtained by the CBD extraction analysis (Table 2). This is doable for 85 samples, among which the 27 samples for which the mineralogical composition is estimated. In this case, we consider the structural iron (%Fe structural ) to be associated with illite (%Fe illite ) and kaolinite (%Fe kaolinite ) in the proportion of the illite-tokaolinite ratio corresponding to the origin of the samples (see Sect. 3.2), such that %Fe kaolinite − %Fe structural 1 + 10 × I K (4) and %Fe illite − %Fe structural − %Fe kaolinite , with 10 representing the illite-to-kaolinite iron content ratio (4.0 % / 0.4 %). Results are summarised in Fig. 12. The mean value of iron fractional solubility is 0.10 % (±0.02 %), whatever the methods of calculation, i.e. mineralogical composition or CBD extraction analysis. There is little variability in iron solubility as a function of the source; nonetheless, this is the direct opposite of the iron oxide content, i.e. the lowest values (0.08 % ± 0.01 %) obtained for samples of Sahelian dust with the highest iron oxide content. Inversely, the samples with the highest solubility (0.13 % ± 0.01 %) issued from PSA3, i.e. southern Algeria/Mali, present a low percentage of free iron (49 %). Data show that 54 to 87 % of soluble iron is structural iron associated with illite, meaning that this mineral is the major provider of soluble iron. The determining factor for this parameter is the quantification of illite and, hence, the hypothe-sis made on the illite-to-kaolinite ratio as a function of emission source, which is estimated from Caquineau et al. (2002). Thus, the regional variability in iron solubility obtained here is mainly associated with this ratio, which should be better constrained. A rapid calculation for the PSA3 samples, the richest in structural iron, shows that iron solubility is divided by a factor of 2 when I / K varies from 0.1 to 2.3, the extreme values observed by Caquineau et al. (2002). This means that the solubility between the various emission sources varies, at best, by a factor of 2. Another source of error could be the estimation of iron oxide, which is affected by errors of 30 %. Sample-to-sample comparison shows that there is a rather good correspondence between the two methods of calculation (within 15 %). However, the calculation based on mineralogy yields values lower by a factor of 1.5 for the samples from PSA3 alone or for samples originating from a mixture of sources, meaning an underestimation of structural iron in comparison to the direct measurements by CBD analyses. Smectites are identified for several samples of this source; we, therefore, underestimate the calculated solubility for the samples where these minerals are present.
The mean fractional iron solubility obtained by calculation is consistent with, though rather inferior to, the observations obtained for transported Saharan dust over the Atlantic, which ranges from 0.12 to 4.1 % (Sarthou et al., 2003;Baker et al., 2006;Sholkovitz et al., 2012). However, our calculations are exclusively based on mineralogical parameters, since our samples have been collected at most 2 days after emission and do not integrate the various chemical and physical processes increasing dust solubility during long-range transport, such as cloud processes (Desboeufs et al., 2001); photochemical processes (Hand et al., 2004) or organic complexation (Paris et al., 2013). Thus, the range of calculated values could be representative of iron solubility for the various sources of African dust before transport. Moreover, the limited regional variability estimated here in comparison to the variability of measurements over the Atlantic Ocean shows that the emission source of dust is less critical when estimating the fractional iron solubility of dust than the process after transport. Nevertheless, new direct measurements of the iron fractional solubility for sources of dust in western Africa are necessary to validate these conclusions.

Final remarks, conclusions and perspectives
In this paper we have presented an insight into the mineralogical composition of mineral dust from western Africa from the synthesis of the airborne and ground-based observations of field campaigns conducted between 2006 and 2007. These observations have been made at emission or within 2 days of transport into the atmosphere. The strength of this synthesis resides in the fact that the observations were made in a consistent way from the methodological point of view (both for sample collection and analysis) and that they span a large Atmos. Chem. Phys., 14, 10663-10686, 2014 www.atmos-chem-phys.net/14/10663/2014/ geographical region covering some of the major African dust sources.
The composition data presented in this paper are being made public as part of a data set called DUST-MAP, to date still in its trial version. Meanwhile, data are available upon request. The prospective evolution of this data set will be the inclusion of data from other field campaigns, both close to source regions and at some distance. Amongst these are the Fennec and the SAMUM (Saharan Mineral Dust Experiment) field campaigns (Heintzenberg et al., 2009;Ansmann et al., 2011;Washington et al., 2012), which were conducted at downwind sources in Algeria and Morocco and were relevant to large-scale transport (Caquineau et al., 2002). This will require a close-up examination of experimental protocols and a careful evaluation of their influence on the results.
The presentation of this work deserves some final considerations and opens various perspectives.
The bulk mineralogical composition of dust in western Africa can be described by a simple model where kaolinite and illite account for 80 % or above of the mass, and quartz, iron and titanium oxides for the remaining 20 %. Calcite is a distinctive element of Saharan sources, for which it ranges between 0.3 and 8.4 % of dust mass. For the remaining minerals, the variability is due to source type (e.g. the proportions of illite and kaolinite as evidenced by Caquineau et al., 2002) or to differences in the size distribution, particularly the increase of the quartz fraction corresponding to a coarse mode centred around 8 µm in diameter. To the best of our knowledge, we are the first to provide a large data set of titanium oxide data, which can be relevant to atmospheric photochemistry (Ndour et al., 2008). We estimate that Sahelian dust is composed of 2 % (by mass) of titanium oxides, that is, double the mass fraction of dust from sources in the PSA2, PSA3, PSA4 and PSA5 areas, with very little sample-to-sample variability. Iron oxides, also important for atmospheric chemistry processes and radiation, display higher values in Sahelian than in Saharan dust.
The discussion of the regional variability of the mineralogical composition is based on our ability to detect source areas based on satellite remote sensing or on soil property maps (e.g. Claquin et al., 1999;Schepanski et al., 2007;Laurent et al., 2008;Ginoux et al., 2012;Nickovic et al., 2013;Journet et al., 2014) and to attribute the provenance of an air mass to it via back-trajectory calculations (e.g. . Beside the uncertainties inherent in the detection of sources, there are two additional factors of ambiguity in source attribution: first, air masses, especially during wintertime, mostly travel in the boundary layer and might uplift dust aerosols at various times before reaching the sampling site; secondly, aircraft samples might combine transport from various sources active at the same time due to the fact that an aircraft moves during sampling and that sample collection lasts at least 20 min, which corresponds to about 120 km at typical aircraft cruise speed (see Fig. 1 in Chou et al. (2008) and Fig. 1 in Klaver et al. (2011)). It is there-fore clear that some fine features of the mineralogy specific to individual sources might be lost. Nevertheless, our data represent the composition of the regional dust load well as it results from the natural mixing occurring during transport. They are also of comparable spatial resolution to those of various state-of-the-art global and regional circulation models including dust mineralogy models (Balkanski et al., 2007;Hoose et al., 2008;Huneeus et al., 2011). Henceforth, the composition data presented in this paper may serve to evaluate the regional and global model estimates of the dust mineralogical composition in the western African atmosphere and to calculate dust properties relevant to their direct, indirect and biogeochemical impacts. Claquin et al. (1999) have discussed some comparisons between the mineralogy of dust atmospheric data and that of the soil fraction predicted by a surface mineralogy table constructed from maps of the arid soil mineralogy (updates of this table have been recently published by Nickovic et al. (2012) and Journet et al. (2014)). To do so, these authors have calculated weight ratios between some minerals (calcite, feldspar and hematite, where hematite is meant to represent the sum of the content of goethite and hematite) with respect to quartz. Our data are consistent with the Claquin et al. (1999) data set regarding the variability of the calcite-to-quartz ratio but display higher and lower ratios for hematite and feldspar, respectively. Our ratio of hematite-to-quartz ranges between 0.03 and 1 (0.09 and 3 when the sum of goethite and hematite is considered) instead of 0.009-0.09, whereas our ratio feldspar-to-quartz is in the range of 0.03-0.11 instead of that of 0.2-2 in Claquin et al. (1999). These differences illustrate the fact that our data are closer to representing the clay soil fraction, whereas the data of Claquin et al. (1999) best represent the silt fraction, richer in quartz particles. When limited to sample SOP1-8 (the local erosion sample enriched in quartz and with a particle fraction larger than 20 µm), ratios are 0.09 for hematiteto-quartz (0.03 if hematite only is considered) and 0.08 for feldspar-to-quartz; these are the ranges expressed by Claquin et al. (1999).
Finally, we have provided the estimate of parameters relevant to the direct impact of mineral dust on radiation and on ocean productivity: the complex refractive index ñ and the iron fractional solubility SFe.
Assuming internal mixing, we have shown that the complex index of refraction based on mineralogy has some variability, mostly related to the separation of illite and kaolinite and to the quartz fraction higher than 40 % of the dust volume. The spectral behaviour of the estimated complex index of refraction is quite different from the OPAC database. The OPAC approach overestimates absorption in the visible. On the other hand, a very recent study presents values of the complex refractive index in the shortwave based on a global mineralogy soil map (Smith and Grainger, 2014). At 550 nm, these authors predict a real part of the complex refractive index around 1.56, that is, higher than ours which ranges between 1.45 and 1.51. Their imaginary part averages www.atmos-chem-phys.net/14/10663/2014/ Atmos. Chem. Phys., 14, 10663-10686, 2014 at 6.2×10 −4 , that is, almost an order of magnitude lower than our predicted values (average 3 × 10 −3 ). This is the consequence of the fact that optical properties estimated from the soil fraction will reflect more closely the optical properties of quartz, which is enriched in the soil fraction with respect to the aerosol fraction due to the size fractionation at emission. An additional confirmation is given by the fact that, in our data, the highest value of the real part of the complex refractive index is obtained for sample SOP1-8, representing the Sahelian erosion case with the highest quartz fraction (41 % by weight), whereas the highest values of the imaginary part are obtained for samples representing long-range transport from Libya through northern Niger and having a low or even the lowest quartz percentage by weight (1.5-6 %) but having the highest calcite fraction (4-8 % by weight). As a matter of fact, quartz is transparent in the visible, whereas calcite and clays are absorbing. These considerations lead to the conclusion that not only soil mineralogy but also the transfer function between the soil and the aerosol fractions must be implemented in models estimating the dust optical properties.
In the long-wave spectrum, our calculations suggest higher absorption than the OPAC model in the atmospheric window. The 10-12 µm band, which is sensitive to differences in the mineralogical composition amongst sources, is relevant to the retrieval of various spaceborne sensors, including IASI, AIRS and MODIS (De Souza Machado et al., 2006;Hudson et al., 2008a;Klüser et al., 2011;Laskina et al., 2012). Some of the discrepancies between our values and the OPAC database are, at least partially, due to differences in the optical constants of the individual minerals. Dedicated in situ measurements of the optical constants of the dust samples would allow us to confirm this hypothesis.
Nonetheless, a limitation of this study, impacting our estimate of the complex refractive index, is the quantification of smectite-type clays. For the purposes of the present study, we have chosen to ignore the contribution of smectites, which were detected only in a few samples and which we consider as a trace component of the aerosol. Also, smectites are dominant in the 0.2 µm size fraction, which is poor in mineral dust in our samples (Klaver, 2012).
In the visible, Egan and Hilgeman (1979) indicated that two different montmorillonite types (one from Wyoming and one from Mississippi) have refractive indices very similar, to those of either kaolinite or illite, both in spectral dependence and magnitude. However, below 500 nm the imaginary part of the complex refractive index of the montmorillonite from Mississippi is higher by a factor of 2 than that of illite, reaching 0.004 between 200 and 360 nm. In the infrared, various authors (Toon et al., 1976;Glotch et al., 2007;Hudson et al., 2008) have shown that montmorillonite has similar optical properties to those of illite, albeit displaying a higher intensity of extinction around 9.5 µm in the Si-O stretching band. Köster et al. (1999) examined five types of nontronites and Fe-rich smectites and found that the position and the intensity of this band vary depending on tetrahedral Fe(III) content. Also, the speciation of clays, in particular that of montmorillonite, has implications for the cloud and ice nucleation properties of mineral dust and the estimate of iron fractional solubility (Hoose et al., 2008;Journet et al., 2008). Ultimately, these facts depend on the compositional heterogeneity in Al-rich smectite in the natural soils and sediments (Christidis and Dunham 1993;Christidis, 2006), due to various processes, including weathering and/or hydrothermal alteration of basic rocks (Köster et al., 1999) but also due to reduction reactions of structural Fe by microorganisms (Pentrakova et al., 2013). This makes the quantification extremely dependent on the environmental conditions. Similar considerations will hold for chlorite-type clays, which were not detected in our samples but which might be present in dust from northern African deserts, such as in Morocco and Tunisia (Paquet et al., 1984;Kandler et al., 2009). It should be noted that, to date, no measurements of the complex refractive index of chlorites are available in the shortwave. In the long wave, chlorite has significantly different optical properties than illite and kaolinite, in particular concerning their major absorption bands, which are located at 9.6µm for illite and kaolinite and at 10.2 µm for chlorite (Egan and Hilgeman, 1979;Mooney and Knacke, 1985). Should chlorite be detected in the samples, its lack of quantification will be treated as an additional source of error to be evaluated.
Regarding the iron fractional solubility, our data are not conclusive in identifying a clear regional difference with respect to the variability observed for dust collected over the Atlantic Ocean after short-and long-range transport. This limited regional variability suggests that a single reference value (0.1 %) of the fractional solubility SFe of Saharan and Sahelian dust before atmospheric transport could be used in biogeochemical models. Direct field measurements of SFe close to emission sources, unavailable at the moment, are needed to confirm this finding. Finally, the SFe value estimated for dust near source regions is in the lowest range, lower than values obtained for dust collected over the Atlantic Ocean, pointing to the need to take atmospheric processing into account when iron solubility is to be described on a global scale.
Acknowledgements. This research was funded by several institutions. Based on a French initiative, AMMA was built by an international scientific group and is currently funded by a large number of agencies, especially from France, the United Kingdom, the United States and Africa. It has been the beneficiary of a major financial contribution from the European Community's Sixth Framework Research Programme. Detailed information on scientific coordination and funding is available on the AMMA International Web site at www.amma-international.org. Financial support by the API-AMMA and LEFE (project BIRD) French national funding programs is also acknowledged. The participation of P. Formenti in the DODO and GERBILS field campaigns was supported by the UK NERC and the UK Met Office; this support is gratefully acknowledged.
Atmos. Chem. Phys., 14, 10663-10686, 2014 www.atmos-chem-phys.net/14/10663/2014/ Thanks are also due to the LISA personnel who participated in field campaigns and helped with sample collection, to the BAe-146 and SAFIRE air and ground crews as well as the FAAM and Met Office observers.