Geophysical Research Letters

Multiple dust sources in the Sahara Desert: The importance of sand dunes


Corresponding author: O. Crouvi, Department of Geosciences, University of Arizona, 1040 E. 4th St., Tucson, AZ 85721, USA. (


[1] We determine the current sources of dust in the Sahara Desert using quantitative correlation between the number of days with dust storms (NDS), derived from remote-sensing data of high temporal resolution, with the distribution of the soil types and geomorphic units. During 2006–8 the source of over 90% of the NDS was found to be sand dunes, leptosols, calcisols, arenosols, and rock debris. In contrast to previous studies, only few dust storms originated from playas and dry lake beds. Land erodibility was estimated by regressing the NDS to the number of days with high-speed wind events, and was found to be high for sand dunes. Clay and fine-silt grains and aggregates are scarce in sand dunes, which most likely produce dust particles through aeolian abrasion of sand grains. Thus, saltating sand grains impacting clay aggregates on playa surfaces cannot be the sole process for generating dust in the Sahara.

1. Introduction

[2] Mineral dust plays multiple roles in mediating physical and biogeochemical exchanges among the atmosphere, land and ocean, and thus is an active component of the Earth system [Harrison et al., 2001]. In order to estimate the past, current and future impacts of dust on the climate and on the environment, quantitative data on the chemical, physical and optical properties of the dust particles are necessary. Furthermore, as these properties vary among different source areas and affect susceptibility of surficial deposits to wind erosion (erodibility), the knowledge of the geomorphology of current sources of dust and their erodibility is critical [e.g., Webb and McGowan, 2009].

[3] The Sahara Desert is the major source of atmospheric dust on Earth [Prospero et al., 2002]. In the Sahara, current sources of dust have been identified as topographic depressions containing dry lake deposits and playas in extremely arid regions (≪100 mm yr−1), based on qualitative analysis of remotely sensed (RS) data with low temporal resolution (e.g., TOMS) [Prospero et al., 2002; Tegen et al., 2002]. Recent studies have shown that using RS data with low temporal resolution results in confusion of areas of transported dust with areas of dust emission [Schepanski et al., 2012], and have suggested other sources for dust (e.g., alluvial fans) [Mahowald et al., 2007; Schepanski et al., 2007; Bullard et al., 2011]. As the identification of dust sources is crucial for modeling dust emission and transport, the widely used identification of dust sources as topographic lows should be re-examined using quantitative analysis of RS data with high temporal resolution [Mahowald et al., 2007; Okin et al., 2011]. There is also a need to understand coarse-scale land erodibility better [e.g., Webb and McGowan, 2009]. Land erodibility is the susceptibility of the surficial deposits to wind erosion — it is a function of soil erodibility (e.g., soil texture), the fraction of the area that comprises non erodible elements, the arrangement of non-erodible elements, climate and land management [Okin, 2008; Webb and McGowan, 2009]. Although soil erodibility is well-known for different soil types from laboratory and field experiments, up-scaling soil erodibility to coarse-resolution land erodibility is complicated [Kurosaki et al., 2011].

[4] The goals of this study are to 1) identify which soil types and geomorphic units are currently the most frequent dust sources in the Sahara, and 2) quantify the land erodibility of these soil types and geomorphic units. To identify the dust sources composition, we spatially correlated the number of days with initiated dust storms per 1° × 1° grid cell for a given time period (hereafter NDS), derived from high temporal resolution RS data [Schepanski et al., 2009], with soil types and geomorphic units (hereafter map units), derived from a global digital dataset [Food and Agriculture Organization of the United Nations (FAO) et al., 2009] (Figure 1a). We further examined the erodibility of the map units through analysis of the covariance between the NDS, the number of days with high-speed wind events per grid cell for the same time period (hereafter NWE), and the map units. Although factors controlling soil and land erodibility operate over a range of spatial and temporal scales, we hypothesize that over a long time interval, regression between coarse-scale dust and wind data can be used to estimate the land erodibility of different map units. Here we define land erodibility as the susceptibility of the surficial deposits to wind erosion in terms of number of days with dust storms, not the common ‘physical’ erodibility that relates dust flux to wind speed.

Figure 1.

The Sahara desert. (a) The distribution of the dominant soil types and geomorphic units in the study area; only map units that discussed in the text are presented. (b) Number of days with dust storms (NDS) for the period 3/2006–2/2008 [Schepanski et al., 2009]. In purple are the 99th percentile NDS grid cells (n = 16), in red and purple are the 95th percentile grid cells (n = 77), and in grey, red and purple are the 90th percentile grid cells (n = 153). All the other NDS grid cells are shown as transparent. (c) Number of days with high-speed wind events (NWE) for the period 3/2006-2/2008 classified into 99th (purple), 95th (purple and red), and 90th (purple, red and grey) percentiles. Soil data are the same as in Figure 1a. (d) The Bodélé Depression (BD) in northern Chad. NDS classification is the same as in Figure 1a, but color appears only on the perimeter of the gird cells. Remotely sensed image is from USGS MDA Federal Landsat GeoCover ETM + 2000 Edition Mosaics tiles N-34-15, N-33-10, N-33-15, N-34-10. The diatomite outcrop, which appear in blue color in the middle of the figure, and the adjacent diatomite sediments to the SW (also blue) are mapped as sand dunes in the global soil map. Dominant wind direction is NE. The areas north and NE (upwind) to the diatomite outcrop are characterized by high NDS, and are covered by a mixture of sand dunes, leptosols and calcisols, and to lesser extent by rock debris and regosols.

2. Methods

[5] We studied the area enclosed between 20°W–45°E and 5°N–40°N and examined the data for a 2-yr interval (3/2006–2/2008; 703 days) on a 1° × 1° grid. Grid cells over southern Europe, Arabia, the Middle East and the oceans were excluded. Grid cells covering coastlines were included if the land fraction in the cell was >50%. In total, 1502 grid cells were included in the analysis.

[6] Location of dust storm initiation was determined by tracking back the dust plume to its point of first occurrence using MSG-SEVIRI images with a high temporal resolution of 15 minutes. From this information the NDS was derived for each 1° × 1° grid cell [Schepanski et al., 2007, 2009]. The NDS represents the counts of dust storms that originated from a specific grid cell during the 703-day time interval of the study (Figure 1b). Information on soil types and geomorphic units was taken from the Harmonized World Soil Database (HWSD) [FAO et al., 2009], a digital, GIS-based soil map (1-km spatial resolution) that uses the FAO soil classification system. The HWSD was compiled from global and regional soil maps, originally at scales of 1:1,000,000 to 1:5,000,000, and it holds information on the dominant soil type or geomorphic unit of each mapping unit. As soil taxonomy is determined partly by grain size and mineralogy, it is possible to infer physical characteristics of the soil type based on its classification. Moreover, several soil types can be related to specific geomorphic units (Table S1 in the auxiliary material). For example, solonetz and solonchaks are usually playa or sabkha soils, rich in soluble salts and clays; arenosols are quartz-rich sandy soils [FAO et al., 2009]. Similar interpretation can be done for the geomorphic units (e.g., sand dunes are rich in quartz sand grains). In the study area, 30 different map units are present (Figure 1a). As most 1° × 1° grid cells contain more than one dominant soil type and geomorphic unit, we chose the one with the highest areal coverage in the grid cell as the dominant soil type or geomorphic unit for the specific grid cell. This procedure narrowed down the number of map units to 24.

[7] To learn the importance of each map unit as a dust source we summed the total NDS in each map unit:

display math

where i is the index for the map unit, j is the index for the grid cells within each unit from 1 to ni. Then we calculated the following measures for each map unit:

display math

where NDSi(%) is the number of days with dust storms per map unit normalized to the total number of dusty days in the entire area and record,

display math

where Avg.NDSi is the average NDS per map unit, and ni is the number of grid cells per map unit.

[8] We also examined separately the grid cells that hold the highest values of NDS by selecting and summing the 95th percentile of the NDS data (defined as NDS95 and composed of 77 grid cells). We then calculated the following measure:

display math

where NDS95i(%) is the number of days with dust storms per map unit only for the most frequent NDS grid cells over the whole Sahara (>95th percentile), normalized to the total number of the most frequent NDS values (NDS95), and j is the index for the grid cells within each unit from 1 to mi.

[9] NDSi(%) reflects the importance of each map unit as a dust source for the entire Sahara; Avg.NDSi reflects the potential of each map unit as a dust source, and NDS95i(%) represents the contribution of the different map units to the most active dust hotspot.

[10] To assess the land erodibility of the map units, we also considered the surface wind speed. Schepanski et al. [2009] recently showed that most of the dust storms in the Sahara start during the morning, and are mainly related to the turbulent downward mixing of momentum from nocturnal low-level jets (LLJ). Thus, a simple annual mean of the surface wind speed cannot predict dust-emission well. In this study, a proxy for the LLJ occurrence was used by calculating the difference between wind speed at 900 hPa and 750 hPa levels at 0600 UTC, using meteorological re-analysis data (1° × 1° grid) [Berrisford et al., 2009]. If the difference was greater than 6 m s−1, a high-speed wind event was defined for this day. Following this approach, the NWE per grid cell was calculated (Figure 1c).

[11] To examine the land erodibility of different map units we used analysis of covariance (ANCOVA). This procedure includes features both of analysis of variance and of regression for continuous and for categorical variables [Milliken and Johnson, 2002]. The NDS and NWE data were square-root transformed to generate a normal distribution of the residuals. Out of the 24 map units, nine were found to be inappropriate for the analysis, either due to zero NDS values for all grid cells, or due to small number of grid cells. In the end, only 15 map units were suitable for the ANCOVA. The outcome of the ANCOVA is a set of linear regressions between NWE½ and NDS½, one for each map unit. As we use counts of occurrences we expect to find a linear relationship between the NWE½ and the NDS½ for each map unit, and the slope of the regression should generally represent their land erodibility. Theoretically, for an ideal transport-limited (i.e., supply unlimited) dust source, each high-speed wind event should cause one dust storm (NDS½ = NEW½, slope of regression = 1, variance = 0). Greater supply limitation for a dust source should be associated with lower slopes and higher variance. This is expected for supply-limited dust sources, since high NWE can cause both high and low NDS values, depending on temporal changes in dust availability at the source, such as due to water content, crust strength, vegetation coverage, and fluvial activity [e.g., Bullard et al., 2008]. Regression slopes near zero indicate that the map unit is not an important dust source.

3. What Soil Types and Geomorphic Units Are Currently the Most Frequent Active Dust Sources?

[12] During the 703 observation days, over 14,500 individual dust storms were initiated in the study area. Our spatial correlation of these counts of dust-storm initiation with the map units reveals that over 90% of the dust storms were initiated in areas where the dominant map unit was sand dunes (28%), leptosols (lithosols) (21%), calcisols (20%), arenosols (16%), and rock debris (7%) (Table 1, column 3). Gypsisols, solonchaks, solonetz and salt flats, which best represent playas and dry lake beds in the Sahara, serve as a source for only 1% of the total NDS. Normalizing these percentages to the number of grid cells of each map unit (Table 1, column 4) reveals that sand dunes and rock debris (e.g., slopes and pediments) have the highest Avg.NDSi values (22 and 20, respectively). Fluvisols, although not currently a major source of dust (only 2%), are ranked third, with an average of 12 NDS per grid cell. Dust hot spots (NDS95 data) are mostly located along the East-West dust belt at 15°–23°N, from Western Sahara to Sudan, with highest dust storm counts at the Bodélé Depression (BD) in Chad (16–19°N; 16–19°E) (Figure 1b) [Prospero et al., 2002; Schepanski et al., 2007]. Over 80% of the dust hot spots are located in grid cells characterized by map units of sand dunes (32%), leptosols (25%), calcisols (19%), and arenosols (9%) (Table 1, column 5). Sand dunes alone are responsible for 42% of the NDS95 data (Table 1, column 6). Playas and dry lake beds are not represented in the dust hot spots.

Table 1. Measures of the Importance of Different Map Units as a Dust Sourcea
Map UnitbAnalysis of the Entire DataAnalysis of the 95th Percentile of Total NDS
Number of Grid Cells per Map Unit (ni) (Counts, % in Brackets)Number of Days With Dust Storms per Map Unit (NDSi) (counts, % in Brackets)Average Number of Days With Dust Storms per Map Unit (Avg.NDSi) (Counts, ±stdev)Number of Grid Cells per Map Unit (mi) (Counts, % in Brackets)Number of Days With Dust Storms per Map Unit (NDS95i) (Counts, % in Brackets)
  • a

    Data are sorted by number of days with dust storms per map unit (NDSi). See text for details on the calculation procedure of the measures. Linear regression was not performed for the lower nine map units (appear in italic) due to zero or very low number of days with dust storms.

  • b

    Geomorphic units are marked with (G), all the other units are soil types.

Sand dunes (G)186 (12.4)4,051 (27.7)21.8 ± 38.925 (32)2,480 (42)
Leptosols270 (18.0)3,129 (21.4)11.6 ± 21.519 (25)1,419 (24)
Calcisols314 (20.9)2,987 (20.4)9.5 ± 14.515 (19)884 (15)
Arenosols235 (15.6)2,332 (15.9)9.9 ± 15.97 (9)536 (9)
Rock debris (G)50 (3.3)1,020 (7.0)20.4 ± 18.55 (6)299 (5)
Fluvisols27 (1.8)317 (2.2)11.7 ± 20.43 (4)172 (3)
Luvisols35 (2.3)294 (2.0)8.4 ± 17.32 (3)124 (2)
Regosols27 (1.8)119 (0.8)4.4 ± 8.50 (0)0 (0)
Cambisols11 (0.7)116 (0.8)10.5 ± 16.51 (1)44 (1)
Gypsisols14 (0.9)99 (0.7)7.1 ± 13.10 (0)0 (0)
Vertisols64 (4.3)64 (0.4)1.0 ± 3.10 (0)0 (0)
Gleysols6 (0.4)40 (0.3)6.7 ± 8.50 (0)0 (0)
Solonetz6 (0.4)26 (0.2)4.3 ± 9.20 (0)0 (0)
Solonchaks8 (0.5)19 (0.1)2.4 ± 1.60 (0)0 (0)
Lixisols93 (6.2)11 (0.1)0.1 ± 0.40 (0)0 (0)
Planosols2 (0.1)3 (<0.1)1.5 ± 0.70 (0)0 (0)
Acrisols41 (2.7)0 (0)00 (0)0 (0)
Alisols2 (0.1)0 (0)00 (0)0 (0)
Andosols1 (0.1)0 (0)00 (0)0 (0)
Ferralsols37 (2.5)0 (0)00 (0)0 (0)
Kastanozems2 (0.1)0 (0)00 (0)0 (0)
Nitisols36 (2.4)0 (0)00 (0)0 (0)
Plinthosols33 (2.2)0 (0)00 (0)0 (0)
Inland water (G)2 (0.1)0 (0)00 (0)0 (0)
Total1502 (100)14,627 (100) 77 (100)5,958 (100)

[13] The BD, the most active dust source in the world [Prospero et al., 2002; Washington et al., 2006], is an east-west elongated dry lake basin (134 × 103 km2) [Bristow et al., 2009] (Figure 1d). Previous studies emphasized the importance of a unique combination of erodible material availability and prevalence of high-speed winds as responsible for the high dust-emission rates there [Washington et al., 2006; Bristow et al., 2009]. The major source of dust at the BD was identified as diatomite dry lake beds and sand dunes of diatomite pellets, located in the middle of the basin [Bristow et al., 2009]. Yet, the diatomite sediments cover only 18% (24 × 103 km2) of the BD basin itself, whereas quartz-rich aeolian sands are much more abundant (65%; 86 × 103 km2) [Bristow et al., 2009] (Figure 1). The presence of the diatomite sediments explains the highest NDS in the Sahara (Figure 1, mapped as sand dunes in the HWSD), but the presence of high NDS values up to several hundreds of km upwind (NE) of the diatomites indicates that these sediments cannot be the sole source of dust at the BD. Our results suggest that the surrounding quartz-rich sand dunes serve as an additional and major source of dust in the BD, and are not only the blasting agents for the dry lake fine deposits downwind.

[14] On regional and continental scales, fine dust (<10 μm) originates from a mixture of several sources that makes geochemical fingerprinting difficult. Nevertheless, quartz is globally the most abundant (+50%) mineral in local uplifted dust, gradually decreasing to 15–30% with transport distance from the sources as a result of size sorting and gravitational deposition [Lawrence and Neff, 2009]. Phyllosilicates (mainly clay minerals) dominant the long-distance travelled dust, with abundance of 20–60% [Lawrence and Neff, 2009]. We propose that sand dunes (and to lesser extent arenosols) provide some of the fine quartz particles in dust (and most of the coarse quartz [Crouvi et al., 2010]), as well as some of the clay minerals that may have coated the sand grains [Bullard et al., 2004, 2007]. These grains, in turn, mix with fine dust from other sources (leptosols, calcisols, rock debris) that add fine grains of quartz and other common minerals found in transported dust (usually calcite, feldspars and more clay minerals). These three map units are relatively rich in quartz, especially in areas in the Sahara composed of magmatic and metamorphic rocks (Table S1).

[15] Our quantitative analyses show that sand dunes, leptosols, calcisols, arenosols, and rock debris are major dust sources in the Sahara (>90% of the current dust storms), whereas the map units that best represent playas and dry lake beds are less important in terms of number of dust storms (1%). We acknowledge that the latter map units are often localized, cover small areas, and might be under-represented as a dominant map unit in our analyses, even though they might be important dust sources. Therefore, we repeated the previous analyses assuming that gypsisols, solonchaks, solonetz and salt flats are the dominant map unit in each grid cell that they appear in, regardless of their areal extent. The results indicate that these map units are present in a maximum of 284 grid cells (19% of the data). Summing the NDS that appear in these cells reveals that these potential dust sources are responsible for maximum 18% and 20% of the total NDS and NDS95 data, respectively. This observation strengthens the suggestion that, alone, playas and dry lake beds cannot explain the overall distribution of frequency of dust sources and additional dust sources must be considered.

4. The Land Erodibility of the Sahara's Soil Types and Geomorphic Units

[16] The results of the ANCOVA show that each of the two examined covariates, and their interaction, control the NDS (P < 0.001; Table S2). These results demonstrate that the NWE, used here as a proxy for the occurrences of sudden and strong surface wind speeds, explains well the dominant atmospheric mechanism for dust emission over most of the Sahara. Ranking the map units according to their regression slope can be interpreted as ranking the erodibility of these units: high slope values are attributed to sand dunes (0.53) and gypsisols (0.45), and intermediate slope values were found for leptosols (0.23), rock debris (0.20), calcisols (0.18) and arenosols (0.14). Units with the lowest slopes are vertisols (0.08), and lixisols (0.03) (Table S3 and Figure 2). These results are in general agreement with field and laboratory experiments for estimating the soil erodibility of different soil types and soil textures [e.g., Shao, 2008]. We emphasize that the land erodibility of map units should not be confused with the current major dust sources. For example, gypsisols are a potentially important dust source as their land erodibility is one of the highest found in this study, however, their spatial distribution over the Sahara is relatively limited, which make these soils an unimportant dust source for the entire Sahara (NDSgypsisols = 0.7%) (Tables 1, S3).

Figure 2.

Land erodibility curves (NDS½ vs. NWE½) for 11 of the studied map units for which the regression significance level is <0.1 (see Table S3). Regression line in black, prediction lines in grey.

[17] Many field and laboratory observations suggest that the sandblasting model is the major process that accounts for dust emission worldwide [e.g., Gillette and Walker, 1977; Alfaro et al., 1997; Shao, 2008]. In its most basic form, sandblasting is defined as the process of dust generation through saltating sand grains that either disaggregate clay aggregates or remove clay coatings from sand grains [Gillette and Walker, 1977]. Previous studies partly adopted sandblasting to model dust emission and transport worldwide, and particularly in the Sahara, assuming that dust emission is proportional to the clay content of the soil [Marticorena et al., 1997; Laurent et al., 2008]. Thus, these studies accounted for only the disaggregation of clay-rich soils (i.e., in playas) through the bombardment of sand grains from adjacent sand dunes. Whereas these studies successfully predicted observed dust emission in central and northern Sahara, they underestimated dust emission in the southern Sahara (18–21°N), an area for which satellite and field observations indicated high dust emission. Our results can explain this discrepancy: As sand dunes contain only small amounts of silt and clay grains and aggregates, the application of these specific sandblasting models to these areas is problematic. A recent study that assumed that dust originates from highly reflective areas (i.e., sand dunes), showed a much better agreement with the observed dust hot spots in the southern Sahara [Grini et al., 2005]. We suggest that the widespread distribution of active sand dunes in the southern Sahara is the main reason for the underestimation of dust emission in these previous studies, as dust emission from sand dunes cannot be modeled accurately assuming that dust emission rate is proportional to clay content. We propose that sandblasting that removes sharp corners and clay coatings from sand grains is a dominant dust-emission mechanism for the vast sand-rich regions of the Sahara. This process, which can be also termed as aeolian abrasion, has been proven to produce fine dust from sand grains in experimental studies [Whalley et al., 1982; Bullard et al., 2004] and is the most probable explanation for the formation of coarse silts from active dune fields [Crouvi et al., 2008, 2010].

5. Applications to Past and Future Dust Emission Models

[18] Our results allow a new interpretation of the extreme dustiness of the last glacial period, estimated as 2–4 times higher than during the Holocene [e.g., Mahowald et al., 2006; McGee et al., 2010]. Previous studies related the increase in dustiness either to increased aridity or to increased wind gustiness [e.g., McGee et al., 2010]. We suggest that the wide extent of active sand dunes over the Sahara during the last glacial period [Sarnthein, 1978], together with the increase in gustiness and possibly in LLJ strength, could result in frequent dust storms and could explain increased dustiness then. The observations of concurrent high accumulation rates of dust in Atlantic Ocean cores off northwest Africa and of continental loess sequences along the edges of the Sahara, downwind of the sand fields [Crouvi et al., 2010], support our hypothesis. We further suggest that considering aeolian abrasion of sand grains in dust-emission and transport models will improve model results for past, current and future scenarios.

[19] The results of this study enable us to predict future change in dust emission for the Sahara. Previous studies by Tegen et al. [2004] and Washington et al. [2009] have pointed out that dust emission over the Sahara will increase under projected future conditions. Various scenarios simulate a strengthening pressure gradient over North Africa that ultimately will lead to stronger low-level wind speeds [Washington et al., 2009]. Using the linear regressions presented in this study (Table S3 and Figure 2) we can predict the increase in NDS for a given increment of NWE. For example, if the NWE in a given area will increase by 125 days (e.g., from 100 days to 225 days over a 2-yr time period), then sand dunes are predicted to produce on average additional 24 days with dust storms. On the other hand, for a similar increment in NWE, an area covered with vertisols will add on average only 0.5 dusty days.

6. Summary and Conclusion

[20] We used data of number of days with dust storms, number of days with high-speed wind events, and detailed soil types and geomorphic units map to examine the composition of the dust sources, and to estimate the land erodibility of the different map unit in the Sahara. We show that active sand dunes are currently the most frequent dust sources, due to their large spatial extent in regions with high-speed winds, together with their high land erodibility. Several mechanisms are responsible for dust emission in the Sahara, and we point out to the important role of aeolian abrasion as the main process for generating fine dust particles in active sand dunes. We suggest that our results can explain the increased dustiness during the last glacial period, when active sand dunes were probably more common than during the Holocene. Our results have the potential to improve the performances of regional scale dust-transport models for the assessment of future effects of dust on the environment, as they provide valuable information on dust-source identification and the erodibility of Sahara's soils.


[21] The authors thank A. Trakhtenbrot, A. Mushkin, J. Pelletier, V. Baker, and D. Yaalon for fruitful discussions which improved ideas and research; the statistics consultancy course at the U. of Arizona and D. Billheimer for statistical advice; G. Okin, J. Bullard, and an anonymous reviewer for improving an earlier draft of this paper; and the ECMWF for providing ERA-Interim re-analysis data. Funding was provided by grants from the US-Israel Binational Science Foundation (2006-221) and US-ARO (DAAD19-03-1-0159). OC appreciates a post-doctoral fellowship from J. Pelletier (U. of Arizona).

[22] The Editor thanks an anonymous reviewer for assisting in the evaluation of this paper.