Seasonal dynamics of threshold friction velocity and dust emission in Central Asia.

An improved model representation of mineral dust cycle is critical to reducing the uncertainty of dust-induced environmental and climatic impact. Here we present a mesoscale model study of the seasonal dust activity in the semiarid drylands of Central Asia, focusing on the effects of wind speed, soil moisture, surface roughness heterogeneity, and vegetation phenology on the threshold friction velocity (u*t ) and dust emission during the dust season of 1 March to 31 October 2001. The dust model WRF-Chem-DuMo allows us to examine the uncertainties in seasonal dust emissions due to the selection of dust emission scheme and soil grain size distribution data. To account for the vegetation effects on the u*t , we use the Moderate Resolution Imaging Spectroradiometer monthly normalized difference vegetation index to derive the dynamic surface roughness parameters required by the physically based dust schemes of Marticorena and Bergametti (1995, hereinafter MB) and Shao et al. (1996, hereinafter Shao). We find the springtime u*t is strongly enhanced by the roughness effects of temperate steppe and desert ephemeral plants and, to less extent, the binding effects of increased soil moisture. The u*t decreases as the aboveground biomass dies back and soil moisture depletes during summer. The u*t dynamics determines the dust seasonality by causing more summer dust emission, despite a higher frequency of strong winds during spring. Due to the presence of more erodible materials in the saltation diameter range of 60-200 µm, the dry-sieved soil size distribution data lead to eight times more season-total dust emission than the soil texture data, but with minor differences in the temporal distribution. On the other hand, the Shao scheme produces almost the same amount of season-total dust emission as the MB scheme, but with a strong shift toward summer due to the strong sensitivity of the u*t to vegetation. By simply averaging the MB and Shao model experiments, we obtain a mean estimate (Exp_mean) of season-total dust emission of 255.6 Mt (megaton), of which 26.8%, 50.4%, and 22.8% are produced in spring (March-April-May), summer (June-July-August), and autumn (September-October), respectively. The Exp_mean estimate identifies the Ustyurt Plateau, dried seabed of Aral Sea (called Aralkum), Caspian Sea coast, and loess deserts as the strongest dust source areas in Central Asia. The spatial distribution and seasonality of the Exp_mean estimate are in general agreement with ground station dusty weather observations and satellite aerosol optical depth and absorbing aerosol index products. Compared to Cakmur et al. (2006), the Exp_mean estimate suggests Central Asia contributes 10-17% to the global dust emission in 2001. KEY POINTS: The WRF-Chem-DuMo model is used to study dust seasonality in Central Asia An accurate representation of u*t is critical for dust seasonality Multiexperiment mean dust emission estimate agrees with observations.

1 Introduction

Wind-blown dust aerosols from global dryland regions pose great threats to human by causing property damages, carrying pathogens, reducing visibility, and deteriorating air quality [e.g., Griffin et al., 2001; Pöschl, 2005]. There is also mounting evidence that dust plays crucial roles in the climate system through interactions with the energy, water, and biogeochemical cycles [e.g., Sokolik et al., 2001; Yin et al., 2002; Jickells et al., 2005; Shao et al., 2011]. Quantification of dust-induced environmental and climatic impact is subject to great uncertainties, which necessitates an improved understanding and model representation of the dust cycle of production, transport, transformation, and removal processes [Boucher et al., 2013]. Under the Aerosol Comparisons between Observations and Models (AeroCom) project, the global dust emission flux differs among 15 models by 1 order of magnitude in the range of 500–4000 Tg yr−1, with even larger discrepancies in the spatiotemporal distribution [Textor et al., 2006; Huneeus et al., 2011]. Despite the large model differences, it is difficult to identify and rank the sources of model errors in these models. To remedy this issue, unified dust model systems have been developed for a comprehensive evaluation of the dust emission uncertainties associated with relevant model parameterizations and input parameters [e.g., Darmenova et al., 2009; Kang et al., 2011].

Dust outbreak is known as a seasonal phenomenon at dryland regions [Goudie and Middleton, 2006]. Global drylands occupy about 41% of Earth’s terrestrial surface, out of which more than half are semiarid and dry subhumid areas [Mortimore et al., 2009]. Semiarid drylands span a wide swath of midlatitude regions and are dominated by temperate grassland, shrubland, and savanna biome [Safriel et al., 2005]. In these regions, wind erosion is regulated by the atmospheric circulation that determines the erosive force of wind and by the dynamic surface characteristics that affect the resistance of soils to wind erosion, or soil erodibility [Shinoda et al., 2011]. The fraction of exposed (and potentially erodible) surfaces and soil erodibility are strongly affected by the land and soil properties and state, such as soil texture, soil wetness, surface crust, and vegetation fraction, which in turn respond to precipitation and temperature changes and vegetation dynamics [Shao, 2008].

In addition to natural processes, the soil erodibility of semiarid drylands is subject to influence from human activities, given that drylands are widely exploited as pasture lands and croplands [Mortimore et al., 2009]. The area of exposed soils and the soil properties of agricultural lands vary strongly in response to land management practices, particularly the cropping and grazing cycles [Webb and Strong, 2011]. Improper land use can exacerbate the dryland vulnerability to wind erosion and lead to increased anthropogenic contribution to the global dust burden [e.g., Tegen et al., 2004; Ginoux et al., 2012].

The interplay between natural processes and human land use has long been known as a key player in the dryland landscape dynamics and dust variability in Central Asia [e.g., Orlovsky and Orlovsky, 2002; Gintzburger et al., 2005; Lioubimtseva et al., 2005]. Figure 1 shows the geopolitical map of Central Asia, overlapped with monthly normalized difference vegetation index (NDVI) and snow cover observed by the Moderate Resolution Imaging Spectroradiometer (MODIS) on board Terra. The region consists of five post-Soviet states (Kazakhstan, Uzbekistan, Turkmenistan, Kyrgyzstan, and Tajikistan) and extends from the Caspian Sea in the west to the Kuldzhuktau and Karatau mountains in the east and from the Kopet Dagh mountains in the south to the Kazakh steppe in the north. The region comprises a mixture of arid and semiarid deserts, Eurasian steppe, and mountain-plateau grasslands and can be broadly divided into two ecosystem zones: northern steppe and southern desert. In the past century, Central Asia has experienced drastic land cover/land use changes driven by institutional and socioeconomic reform, with far-reaching consequences on the region’s ecosystem and human welfare [Orlovsky and Orlovsky, 2002]. One notable example is the formation of the Aralkum dust source over the dried seabed of the Aral Sea as a result of agricultural expansion and over-irrigation since the 1950s [Micklin, 2007]. Figure 1 shows strong seasonal and latitudinal variations in the vegetation greenness and snow cover in Central Asia, suggesting potential effects of the land dynamics on the region’s dust seasonality. Indeed, ground station observations show the dust season typically lasts from early spring (March) to late autumn (October) and peaks in summer (June-July-August), in accordance with the observed snow-free period, and low summertime NDVI [Indoitu et al., 2012]. Therefore, an accurate representation of the dynamic surface characteristics is critical for model simulations of the seasonality of soil erodibility and dust emission in Central Asia.

Figure 1

The geopolitical map of Central Asia overlapped with MODIS monthly NDVI and snow cover in 2001. Shown in the July and August maps are locations of five post-Soviet states (Kazakhstan, Uzbekistan, Turkmenistan, Kyrgyzstan, and Tajikistan) and places mentioned in the paper: 1, rain-fed croplands; 2, Kazakh steppe; 3, Caspian Sea; 4, Kara-Bogaz-Gol gulf; 5, Mangyshlak Peninsula; 6, Caspian Sea coast; 7, Ustyurt Plateau; 8, Aral Sea; 9, Karakum Desert; 10, Kyzylkum Desert; 11, Amu Darya basin; 12, Syr Darya basin; 13, Muyunkum Desert; 14, Betpak-Dala Desert; 15, Lake Balkhash; 16, Saryesik-Atyrau Desert; 17, loess deserts; and 18, Taklimakan Desert.

There have been much efforts on developing physically based dust emission schemes to describe (1) the increase in the soil erodibility (quantified by the threshold friction velocity, u*) due to soil moisture and surface roughness and (2) the contribution to dust vertical flux by saltation sandblasting and disintegration of soil aggregates [e.g., Raupach et al., 1993; Marticorena and Bergametti, 1995; Shao et al., 1996; Fecan et al., 1999; Alfaro and Gomes, 2001]. Implementation of physically based schemes in regional and global dust models faces a scale-mismatch issue. That is, the land and soil parameters in large-scale models often do not meet the requirements of dust schemes, which characterize small-scale aeolian processes. For instance, the soil moisture simulated by land surface schemes often cannot be directly applied to the Fecan et al. [1999] u*-soil moisture relationship [e.g., Zender et al., 2003; Grini et al., 2005]. The aerodynamic roughness length in climate models characterizes atmospheric processes at a much larger length scale than saltation processes. To bridge this gap, dedicated surface roughness data sets are derived from satellite observations and models for dust schemes [Marticorena et al., 2004; Prigent et al., 2005; Pierre et al., 2012]. For simplicity and to save computational expense, global models tend to employ empirical dust schemes, which commonly adopt a fixed threshold wind velocity [e.g., Tegen and Fung, 1995; Ginoux et al., 2001; Uno et al., 2001; Takemura et al., 2009]. In these schemes, the dependence of soil erodibility on surface properties is described by empirical preferential dust source functions, which tune dust emission to easily erodible regions in order to maximize agreement with atmospheric dust burden observations [e.g., Ginoux et al., 2001; Kim et al., 2013].

So far, there have been no regional model studies to address the dust variability in Central Asia. In light of this, the goal of this study is to examine the seasonal dust activity in Central Asia using a coupled regional dust model system WRF-Chem-DuMo and a set of dust observations from ground stations and satellites. The WRF-Chem-DuMo model incorporates multiple options of dust emission schemes and terrestrial input data sets (e.g., soil size distribution) for Asian dryland regions, which allows us to quantify the dust emission uncertainties associated with model parameterization and input data [Darmenova et al., 2009]. This study focuses on two questions: (1) How does the u* change in space and time under the influence of soil moisture, surface roughness, and vegetation phenology? (2) How does dust emission vary in response to the seasonality in the surface wind and u*? We conduct model simulations of the u* and dust vertical flux for the dust season of 1 March to 31 October 2001, which is influenced by a prolonged drought event over Central Southwest Asia [Barlow et al., 2002]. The simulated dust emissions are compared to ground-based dusty weather records and satellite aerosol observations in order to evaluate the model representation of dust seasonality in Central Asia.

2 Data and Methodology

2.1 Model Descriptions

The dust model system WRF-Chem-DuMo is a modified version of the community WRF-Chem model (v3.4) through incorporation of a dust emission module DuMo. Darmenova et al. [2009] provided in-depth descriptions of DuMo and made recommendations on the input parameters for Asian dust source regions. WRF-Chem-DuMo includes two physically based dust schemes of Marticorena and Bergametti [1995] and Shao et al. [1996] with some new developments (hereafter referred to as MB and Shao schemes, respectively) and one empirical dust scheme of Tegen and Fung [1995] (hereafter referred to as TF scheme). The model provides two options of soil grain size distribution based on either the U.S. Department of Agriculture (USDA) soil texture classification or in situ dry-sieved soil size measurements. The model employs the following physics options: Noah land surface scheme, Mellor-Yamada-Janjic planetary boundary layer scheme, Janjic Eta surface layer scheme, Thompson microphysics scheme, and Kain-Fritsch cumulus scheme. National Center for Atmospheric Research/National Centers for Environmental Prediction reanalysis data are used to provide initial and boundary conditions. According to Todd et al. [2008], the model is configured 42 vertical layers with 10 layers below 1 km and the lowest layer at about 20 m in order to improve simulations of near-surface winds. Model simulations are performed using a fine grid spacing of 10 km × 10 km to accommodate the region’s complex topography and resolve the small-scale dust sources, such as the Aralkum and agricultural lands.

2.2 Data

To account for the vegetation effects on dust emission, we use the MODIS/Terra NDVI monthly composite level 3 0.05° × 0.05° Climate Modeling Grid product (MOD13C2) to derive the erodible fraction and surface roughness parameters used by the MB and Shao schemes, which is described in section 3.1. To evaluate the simulated 10 m wind (U10), we obtain the land and marine surface stations data from the Met Office Integrated Data Archive System of the British Atmospheric Data Centre. The data set contains 3-hourly U10 (speed and direction) observations at worldwide meteorological stations from 1853 to date. Dust observation is also available in the present weather (PW) records from manned stations (WMO code 4677). We do not use horizontal visibility due to the effects of haze and fog. We find these nondust events cannot be excluded using the temperature and dewpoint difference criteria suggested by Mahowald et al. [2007]. We divide the dust-related PW events into severe (PW = 33, 34, 35), moderate (PW = 30, 31, 32, 98), and weak (PW = 06, 07, 08, 09) dusty weather categories. The dusty weather categories are assigned different weights according to O’Loingsigh et al. [2014] to compute the dust frequency index (DFI): where SD, MD, and WD represent the number of severe, moderate, and weak dusty weather events, respectively. The denominator is the total number of PW observations adjusted by the weighted dust observations, so that DFI falls in the range of 0–1.0. The DFI can be calculated for a single or multiple stations (Stn) for a given time period (T). Compared to the commonly used dusty day count or dust frequency, the DFI is a better measure of dust source strength by accounting for the dust event intensity.

Aside from DFI, we obtain satellite products of deep-blue aerosol optical depth (AOD) at 550 nm from MODIS/Terra and SeaWiFS (Sea-viewing Wide Field-of-view Sensor) and the absorbing aerosol index (AAI) from TOMS (Total Ozone Mapping Spectrometer) on board Earth Probe. To compare with dust emission on a monthly basis, all satellite products are processed into monthly means. Best quality pixels from MODIS collection 5 level 2 granules are mapped onto a regularly spaced latitude-longitude grid to create daily AOD fields, which are then averaged over a month to compute the monthly mean AOD. The SeaWiFS monthly mean AOD is computed by averaging the v004 level 3 0.5° daily global gridded product. The TOMS monthly mean AAI is computed by averaging the v008 1° × 1.25° daily global gridded product. Download links for all data are provided in the acknowledgments section.

2.3 Computation of Threshold Friction Velocity

The MB and Shao schemes explicitly account for the effects of soil moisture and surface roughness on the size-dependent u*. The u* parameterization follows the concept of adding a series of correction terms to the threshold friction velocity over an idealized dry and smooth surface (u*): where D is the soil grain diameter; f and f are the correction terms for soil moisture and nonerodible surface roughness elements, respectively; and f and f are equivalent to the ratios of u* with and without the soil moisture and surface roughness effects, respectively. Generally, as the wind velocity increases, soil grains with a diameter of 60–200 µm are the first to enter saltation and eject dust particles to the surface layer. Both schemes produce the lowest u* within the diameter range of 60–200 µm. In the saltation size range, the Shao scheme produces a slightly higher u* than the MB scheme. For a diameter of 100 µm, the u* by the Shao scheme (0.24 m s−1) is 0.03 m s−1 or 11% higher than the MB scheme (0.21 m s−1). The u* difference between the two schemes is further enlarged by differences in the f and f terms.

Soil moisture affects the soil erodibility by increasing the interparticle cohesive force between soil grains. This effect is known to act on short time scales (e.g., daily), because of the sporadic nature of precipitation and high potential evaporation in drylands [Ravi et al., 2004]. To account for the soil moisture binding effect, we use the approach of Fecan et al. [1999] in both the MB and Shao schemes. Fecan et al. [1999] assumed that the interparticle capillary force starts to enhance the u* after the soil water content (w) reaches the maximum water amount (w) held by adsorptive forces. Water adsorption on soil grains at low moisture levels is assumed to have negligible effect on the u*. The soil moisture correction term f is expressed as where w is a function of soil clay content, w = 0.0014 × (% clay)2 + 0.17 × (% clay), and w is the gravimetric soil moisture of the topmost 0–2 cm soil layer [Darmenova et al., 2009]. Measurement of the 0–2 cm soil moisture is difficult, and we are unaware of any such measurements in Central Asia. In the WRF-Chem-DuMo model, the Noah land surface scheme simulates the volumetric soil moisture at four soil layers: 0–10, 10–40, 40–100, and 100–200 cm. We find the Noah 0–10 cm soil moisture leads to complete suppression of dust emission in both schemes by overestimating the f. Koster et al. [2009] pointed out the soil moisture predicted by land surface schemes is highly model specific and suggested the modeled soil moisture be treated as a wetness index of the true moisture state with a dynamic range defined by the land surface scheme. Therefore, it may be inappropriate to directly transfer the soil moisture from a certain land surface scheme into other models for different purposes, such as dust emission schemes. On the other hand, the f parameterizations are usually derived from wind tunnel experiments and based on a limited number of soil samples, and therefore may require recalibrations when being implemented in climate models. These reasons may explain why the reanalyzed soil moisture was found to be too high for the Fecan et al. [1999] parameterization [Zender et al., 2003; Grini et al., 2005]. Koster et al. [2009] suggested the true value of model-simulated soil moisture lies in its temporal variations, but not in the absolute magnitude. This suggests that necessary scaling could be applied to the simulated soil moisture to use in other models without losing the information of the soil moisture dynamics. Here we use a factor of 0.3 to multiply the Noah-predicted 0–10 cm soil moisture in order to approximate the superficial 0–2 cm soil moisture used by the MB and Shao schemes. This scale factor is selected on the basis of no soil moisture effect (f = 1) at the central Karakum Desert in summer, when the desert is under extremely dry hot conditions [Orlovsky et al., 2005]. According to our model simulations and ground observations, the summertime monthly precipitation is less than 5 mm in the southern desert area. Under such low precipitations, the soil moisture is likely below the critical moisture level (w) in equation 3. Besides, the soil moisture can be depleted very quickly in sandy soils during dust outbreaks and therefore may have limited effects on the u* [Cornelis and Gabriels, 2003].

The effect of surface roughness on the u* depends on the partition of wind shear stress between the erodible fraction of land surface and the nonerodible roughness elements. In the MB scheme, the original drag partition of Marticorena and Bergametti [1995] was modified by MacKinnon et al. [2004] to accommodate the high roughness of partially vegetated surfaces: where z0 is named aeolian roughness length to avoid confusion with the aerodynamic roughness length in mesoscale models. Here z0 is an integrative measure of nonvegetation (e.g., gravels and pebbles) and vegetation roughness elements [Marticorena and Bergametti, 1995; Menut et al., 2013]. The z0 is the aeolian roughness length of an idealized smooth surface and can be calculated as 1/30 of the coarse-mode mass median diameter (MMD) in the lognormal soil size distribution.

The Shao scheme uses a double drag partition method to account for the roughness effects of nonvegetation and vegetation elements separately. This method is based on the drag partition scheme by Raupach et al. [1993] for barren surfaces, which is extended to treat both the bare (B) and vegetated (V) fractions of land surface: where A is green vegetation fraction. The parameters σ, m, and β characterize the geometry and spatial distribution of the nonvegetation (B) and vegetation (V) roughness elements. These parameters are difficult to constrain due to lack of measurements. Here we use the values recommended by Darmenova et al. [2009]: σ = 1.45; m = 0.16; β = 202; σ = 1.0; m = 0.5; and β = 90. The roughness density (λ), also called lateral cover, is defined as the sum of the frontal areas of the obstacles present on the ground facing the wind direction divided by the ground surface area [Shao et al., 1996]. Hence, λ depends on the number and mean geometric dimensions (i.e., shape, height, and width) of the obstacles lying on the ground. We assume λ is time invariant and varies as a function of land type to facilitate the implementation of equation 5 in the WRF-Chem model. The λ values are shown in Table5. Determination of the monthly z0 and λ values is presented in section 3.1.

Table 5

Roughness Density (λ) and Geometric Height (h, cm) of Nonvegetation Elements and Monthly Geometric Heights (h, cm) of Vegetation Elements as a Function of Land Type

Land Cover	Vegetation Type	λB	hB	hV
				Mar	Apr	May	Jun	Jul	Aug	Sep	Oct
Dryland cropland/pasture	Wheat, barley	0.05	2.0	2	2	10	30	50	50	30	10
Irrigated cropland/pasture	Cotton, rice	0.05	2.0	2	5	15	30	50	50	40	15
Grassland	Temperate grasslands	0.04	2.0	5	10	20	15	12	12	10	8
Shrubland	Permanent rangelands	0.03	1.0	15	15	12	12	10	10	10	5
Barren or sparsely vegetated	-	0.02	0.5	2	2	2	2	2	2	2	2

2.4 Computation of Dust Fluxes

Darmenova et al. [2009] provided detailed descriptions and comparisons of the horizontal and vertical dust flux parameterizations in the MB and Shao schemes. Briefly, the MB scheme uses the White [1979] formulation for the horizontal flux (G): where the erodible fraction E is defined as the fraction of snow-free nonvegetated land area within a grid cell, u* is the friction velocity, ρ is the air density, g is the gravitational acceleration, and dSrel(D) is the relative surface covered by particles of diameter D and is derived from the soil size distribution. The dust vertical flux is calculated from G as a function of the size-dependent sandblasting efficiency, which depends on the balance of the kinetic energy of saltators and the binding energy of soil aggregates [Alfaro and Gomes, 2001].

In the Shao scheme, the horizontal flux is computed following the Owen [1964] transport-limited saltation approach: where p(D) is the soil size distribution. The dust vertical flux is computed from G using a sandblasting mass efficiency as a function of the sizes of saltating soil aggregates and suspended dust particles [Shao et al., 1996]. Equations 6 and 7 reveal that the MB and Shao schemes share similar mathematical formulations in the dust horizontal flux. In fact, Darmenova et al. [2009] showed that the two schemes produce similar (size-resolved or total) horizontal fluxes over dry smooth surfaces, because of small differences in the u* and dG(D). Their differences are enlarged over rough surfaces due to the different f formulations.

For both schemes, the friction velocity (u*) is computed from the model-predicted U10 assuming neutral atmospheric stability, so that the wind speed follows a logarithmic profile: where κ is the von Karman constant (0.41). The neutral stability assumption has minor effects on the u*, especially during strong dust events [Darmenova et al., 2009]. Based on equations 4, 5, and 8, surface roughness affects dust emission in two opposing ways: it reduces the available wind momentum for erosion and enhances the u*, while it also modifies the boundary layer wind profile causing an increase in the surface drag (u*). The effect of the u* increase generally dominates, such that an increase in the z0 leads to a decrease in dust emission.

Generally, there are two different approaches for representing the soil grain size distribution in dust schemes. Many global and regional models use the readily available soil texture data, which consists of the fractions of sand, silt, and clay-sized particles [e.g., Zakey et al., 2006; Heinold et al., 2007; Pérez et al., 2011]. Figure 2a shows the soil texture map in Central Asia based on the USDA textural classification. There are seven dominant soil texture classes: sand, sandy loam, loam, sandy clay loam, clay loam, clay, and bedrock. Each soil texture class is associated with a trimodal lognormal size distribution, z0, and clay content, as shown in Table1. We add the texture class of bedrock to the data of Zakey et al. [2006, Table 1] and assign to it the size distribution data of the sand texture, so that the northern part of the Karakum Desert (called Trans-Unguz Karakum) is treated as sandy desert, same as the other parts of Karakum Desert.

Figure 2

Maps of (a) USDA 16-class soil texture and (b) USGS 24-class land cover in Central Asia. Boxes encompass the dust source subregions described in Table2.

Table 1

Soil Size Distribution Data as a Function of Soil Texturea

Soil Texture	Mode 1			Mode 2			Mode 3			z0s	Clay
	n	MMD	σ	n	MMD	σ	n	MMD	σ
Sand	0.9	1000	1.6	0.1	100	1.7				3.3	3
Sandy loam	0.6	520	1.6	0.3	100	1.7	0.1	10	1.8	17.3	10
Loam	0.35	520	1.6	0.5	75	1.7	0.15	2.5	1.8	17.3	18
Sandy clay Loam	0.30	210	1.7	0.5	75	1.7	0.2	2.5	1.8	7.0	27
Clay loam	0.2	125	1.7	0.5	50	1.7	0.3	1	1.8	4.2	34
Clay	0.5	100	1.8	0.5	0.5	1.8				3.3	58
Bedrock	0.9	1000	1.6	0.1	100	1.7				3.3	3

Including trimodal lognormal parameters (mass fraction n, mass median diameter MMD (µm), and geometric standard deviation σ), smooth aeolian roughness length z0 (µm), and clay content (%).

Table 2

Soil and Land Characteristics of Dust Source Subregions in Central Asia

Subregion	Places	Land Cover	Soil Texture	Landscape
I	Ustyurt Plateau, Mangyshlak Peninsula	Shrubland, barren or sparsely vegetated	Clay loam	Gravelly and hilly desert plateau
II	Caspian Sea coasts, Kara-Bogaz-Gol gulf region	Barren or sparsely vegetated	Clay, sandy loam	Saline deserts and sand dunes
III	Karakum Desert	Shrubland, barren or sparsely vegetated	Sand, bedrock	Sand dune ridges and chains
IV	Kyzylkum Desert, Muyunkum Desert	Shrubland	Sand, clay loam	Sand dune ridges and chains
V	Loess deserts	Shrubland, grassland, barren or sparsely vegetated	Loam, sandy loam	Alluvial deposits of mountain ranges
VI	Aralkum	Barren or sparsely vegetated	Sandy loam	Saline desert, alluvial deposits, takyr soils
VII	Betpak-Dala Desert	Shrubland	Loam	Stony and hilly desert
VIII	Kazakh steppe	Grassland	Loam, sandy clay loam, clay loam	Temperate grassland and shrubland biome
IX	Cropland	Dryland cropland/pasture, irrigated cropland/pasture	Clay, loam, sandy loam	Rain-fed and irrigated croplands

Because soil texture is usually measured through wet sedimentation techniques (e.g., ultrasonic pretreatment and dissolution) and therefore may not reflect the natural state of soil aggregation, undisturbed dry soil size distribution measured by dry sieving the soil aggregates is recommended for dust schemes [Laurent et al., 2006]. Mei et al. [2004] conducted dry-sieved size measurements using soil samples collected from several Chinese deserts. Using these measurements, Laurent et al. [2006] derived the bimodal lognormal size distribution parameters. We are unaware of any such measurements for Central Asia. To overcome this problem, we conduct a comparison between the Central Asia dust sources and Chinese deserts, focusing on the similarities in the dryland landscape and dust source characteristics. Through the comparison, we attempt to map the dry-sieved soil size distributions of Chinese deserts to Central Asia.

Based on the soil size distributions reported by Laurent et al. [2006, Table 2], the Chinese deserts can be broadly divided into two groups: (1) sandy deserts located in topographic depressions, such as the Taklimakan Desert, and (2) the Gobi Desert comprising temperate grasslands, shrublands, and the small deserts of Badain Jaran, Mu Us, and Tengger. Sandy deserts are associated with a large MMD and small coarse-mode mass fraction, whereas the Gobi Desert has a larger coarse-mode mass fraction. The soil size differences between these two groups are reflected in their land cover and soil texture. The sandy deserts are dominated by barren or sparsely vegetated land type and sand soil texture, whereas the Gobi Desert is dominated by grassland and shrubland land types and comprises several soil texture classes, including sandy loam, loam, and sandy clay loam. Furthermore, the two desert groups differ greatly in the surface roughness. Based on the approach proposed by Marticorena et al. [2004], Laurent et al. [2005] derived a static z0 map for the Chinese deserts from the bidirectional reflectance products obtained from the Polarization and Directionality of the Earth’s Reflectance (POLDER) instrument on board the Advanced Earth Observation Satellite (ADEOS) for the period November 1996–June 1997. This z0 data set accounts for the presence of solid nonvegetation roughness elements and perennial vegetation on the surface. They reported that the z0 is less than 0.001 cm over the Taklimakan Desert and ranges from 0.01 to 0.5 cm over the Gobi Desert [Laurent et al., 2005].

To compare with Chinese deserts, we divide the study domain into nine dust source subregions shown in Figure 2, according to their geomorphology and lithology of parent soils, which are described in Lioubimtseva [2002]. Each subregion has a dominant land cover and soil texture as summarized in Table2. Subregion I comprises the gravelly and hilly desert of Ustyurt Plateau. Subregion II comprises the northern and southeastern Caspian Sea coasts and Kara-Bogaz-Gol gulf region. The northern Caspian Sea coast lies below the mean sea level and is dominated by solonchak soils with high salt accumulation (i.e., saline deserts), while the Kara-Bogaz-Gol region and southeastern Caspian Sea coast are covered by drifting sand dunes and solonchaks. Subregions III and IV cover the Karakum, Kyzylkum, and Muyunkum sandy deserts of aeolian-alluvial origin. Subregion V comprises the loess deserts formed on the alluvial deposits and clayey sediments. Parts of the loess deserts are transformed into rain-fed and irrigated croplands, as shown in Figure 2b. Subregion VI comprises the dried seabed of Aral Sea, or Aralkum, which is covered by solonchaks, takyr soils, and alluvial deposits [Orlovsky and Orlovsky, 2002]. Based on Landsat images, we modify the land/water mask in WRF-Chem-DuMo to reflect the areal extent of the Aral Sea in 2001. Subregion VII covers the stony Betpak-Dala Desert used as pasture lands. Subregion VIII comprises the Kazakh steppe. Subregion IX comprises the rain-fed cropland belt (e.g., wheat and barley) in north Kazakhstan and the irrigated croplands (e.g., cotton and rice) located in the Amu and Syr Darya river valleys and mountain alluvial plains. Due to the heterogeneous geomorphology and lithology, the dust source subregions differ greatly in surface roughness. Using the same method as Laurent et al. [2005], we derive the z0 for Central Asia from POLDER data. Figure 3 shows that the POLDER-derived z0 ranges from 0.05 to 1.0 cm in Subregions I, V, VII, and IX and from 0.001 to 0.01 cm over Subregions III and IV.

Figure 3

The static aeolian roughness length derived from POLDER bidirectional reflectance product. Boxes encompass the dust source subregions described in Table2.

Table3 summarizes the mapping of the dry-sieved soil size distributions of Chinese deserts to the dust source subregions in Central Asia. Similar to the Gobi Desert, Subregions I and VII are associated with shrubland land type, loamy soil texture, and high z0. Therefore, we assign the soil size distribution of Gobi Desert to Subregions I and VII. The sandy deserts in Subregions III and IV are assigned with the soil size distribution of Taklimakan Desert. For the saline deserts in Subregions II and VI, we use the dry-sieved soil size distribution measured by Argaman et al. [2006] from Aralkum soil samples. The loess deserts in Subregion V are assigned with the soil size distribution of Hexi Corridor, given that both are formed on alluvial deposits. The steppe in Subregion VIII is assigned with the soil size distribution of Horqin Desert, considering that both were used primarily as grazing lands and have undergone serious desertification in recent decades [Su et al., 2005]. Cultivation practices are known to increase the soil susceptibility to wind erosion by destroying the soil aggregates and generating large amounts of fine materials [e.g., Tegen and Fung, 1995]. Therefore, we assign the soil size distribution of sandy deserts to the rain-fed croplands in Subregion IX to account for the easily erodible fine materials on disturbed soils.

Table 3

Dry-Sieved Soil Size Distribution Data for the Dust Source Subregions in Central Asiaa

Subregion	Mode 1			Mode 2			z0s	Clay
	n	MMD	σ	n	MMD	σ
I	0.58	457	1.74	0.42	86	1.38	15.2	15.8
II	0.28	271	1.37	0.72	14	1.17	9.03	13.4
III	0.03	442	1.42	0.97	84	1.34	2.8	3.0
IV	0.03	442	1.42	0.97	84	1.34	2.8	3.0
V	0.6	386	1.59	0.4	97	1.26	12.9	11.5
VI	0.28	271	1.37	0.72	14	1.17	9.03	13.4
VII	0.58	457	1.74	0.42	86	1.38	15.2	15.8
VIII	1.0	315	1.29	-	-	-	10.5	15.6
IX	0.03	442	1.42	0.97	84	1.34	2.8	3.0

Including bimodal lognormal parameters (mass fraction n, mass median diameter MMD (µm), and geometric standard deviation σ), smooth aeolian roughness length z0 (µm), and clay content (%).

Unlike the MB and Shao schemes, the empirical TF scheme assumes a fixed U10 threshold (U10). A value of U10 = 6.5 m s−1 has been widely used [e.g., Tegen and Fung, 1995; Uno et al., 2001; Takemura et al., 2009]. The dust vertical flux (F) in the TF scheme is defined as follows: where C is a tuning constant used to minimize model-observation differences. Here C is computed by tuning the TF season (March-October) total dust emission to match the physically based schemes. S is a dust source function used to shift dust emission toward easily erodible regions. According to aerosol index observations from the Total Ozone Mapping Spectrometer (TOMS), topographic low basins in arid regions containing loose alluvial deposits are a major contributor to the global dust cycle [Prospero et al., 2002]. Based on the TOMS AAI, a static dust source function was derived to map these preferential source regions in the Goddard Chemistry Aerosol Radiation and Transport (GOCART) model [Ginoux et al., 2001]. Recently, Kim et al. [2013] introduced vegetation variability into the GOCART dust source function based on the Advanced Very High Resolution Radiometer (AVHRR) NDVI products. We use this newly developed dynamic dust source function in the TF scheme.

Using the three dust schemes (MB, Shao, and TF) and two soil size distribution data sets (soil texture and dry-sieved), we conduct four model experiments for the dust season of 1 March to 31 October 2001: MB_Dry, MB_Wet, Shao_Dry, and TF_U10, as listed in Table4. These experiments aim to quantify the dust emission uncertainty from the selection of dust emission scheme and soil size distribution data and to evaluate the differences between physically based and empirical dust schemes.

Table 4

Experiment Design of Dust Emission Simulations

Experiment	Dust Scheme	Soil Size Distribution Data	Wind
MB_Dry	MB	Dry-sieved	u*
MB_Wet	MB	Soil texture	u*
Shao_Dry	Shao	Dry-sieved	u*
TF_U10	TF	-	U10

3 Results

3.1 Incorporation of Vegetation Dynamics Into Threshold Friction Velocity

Vegetation affects dust emission by decreasing the erodible fraction and enhancing the surface roughness and u*. Figure 1 shows the northern steppe area (north of 46°N) is mostly covered by steppe vegetation from April to October. The southern desert area has a low NDVI (e.g., less than 0.2) during most time of the dust season. Driven by precipitation and temperature changes, the temperate grassland belt between 46°N and 49°N displays a vegetation growth-decay cycle of spring greening and summer senescence. In the Karakum and Kyzylkum deserts, the spring peak rainfall promotes the growth of ephemeral plants causing an increase in NDVI (0.15–0.2). The NDVI decreases after the ephemerals enter dormancy in summer. Unlike the natural lands, croplands (e.g., cotton) located in the Amu and Syr Darya river valleys display a NDVI peak in summer due to irrigation and subsequent decrease upon the start of harvest season in September.

To account for the effects of herbaceous species on Sahelian dust, Pierre et al. [2012] used a vegetation model to simulate the vegetation structural parameters, including leaf area index and canopy height, from which they estimated the z0 used by the MB scheme. Here we adopt a similar approach but use MODIS NDVI to derive the roughness parameters (z0, A and λ) for the MB and Shao schemes in a consistent manner. NDVI is the ratio of the difference between the near-infrared and visible reflectance to their sum and theoretically ranges from −1.0 to 1.0 [Glenn et al., 2008]. Healthy green vegetation produces a large positive NDVI, because the chlorophyll content in leaves strongly absorbs visible light while the plant cell structure strongly reflects near-infrared light [Tucker, 1979]. NDVI has been widely used as a measure of the amount and state of photosynthetic vegetation and as a proxy for deriving various land and vegetation parameters in climate models [Glenn et al., 2008].

To derive A from NDVI, we use the linear method from Gutman and Ignatov [1998]: where NDVI(t) is the observed time-varying NDVI value. NDVI and NDVI represent the values for dense vegetation and bare soils, respectively. We use the values of NDVI = 0.93 and NDVI = 0.06, based on the A-NDVI relationship obtained by Kimura and Shinoda [2010] for the Mongolian and Chinese grasslands. After calculating the A, we derive the λ using the empirical method by Shao et al. [1996]:

Figure 4a shows that the monthly A values exhibit strong variability driven by vegetation phenology. In March, the northern steppe is covered by snow. As temperature warms, A increases rapidly and exceeds 30% in the steppe area. At the grassland belt (46°N–49°N), A displays a decreasing trend from spring to summer as precipitation decreases. Driven by the spring rainfall, growth of ephemeral plants causes A to exceed 12% over parts of the Karakum and Kyzylkum sandy deserts and loess deserts. As the ephemerals die back during summer, A drops below 9%. In the Amu and Syr Darya river valleys and loess deserts, irrigated lands display a crop cycle: green onset in April (A < 15%), green peak in August (A > 30%), and defoliation in October (A < 24%). The barren deserts of Ustyurt Plateau, Caspian Sea coast, Aralkum, and central Karakum Desert have a low A (<6%) during the entire dust season.

Figure 4

Monthly surface input parameters in the MB and Shao schemes: (a) vegetation fraction, (b) total roughness density, and (c) aeolian roughness length.

Figure 4b shows the monthly total roughness density (λ) with the contribution of vegetation estimated using equation 11. The lowest values (∼0.02) are found over the barren surfaces (Ustyurt Plateau, Caspian Sea coast, Aralkum, and central Karakum Desert). The growth-decay cycle of natural vegetation, including steppe and desert ephemerals, causes a spring peak (λ > 0.1) and subsequent summer decrease (λ ∼ 0.05–0.07). Agricultural lands become erodible in spring, while they are completely protected in summer due to high vegetation cover. The contribution of vegetation to the total roughness density, namely, the ratio λ/λ, is found to vary from 60% to 70% over the steppe area to 50–60% (spring) and 40–50% (summer) over the desert area. In comparison, Marticorena et al. [2006] reported that λ ranged from 0.025 to 0.233 in the semiarid drylands of south Tunisia. They showed λ tended to increase with vegetation, consistent with our findings.

To incorporate the vegetation dynamics into the MB scheme, we use an empirical relationship between the aeolian roughness length (z0), the mean geometric height (h), and roughness density (λ) of all roughness elements. Based on wind tunnel experiments, Marticorena et al. [1997] found that on a logarithmic scale the ratio z0/h varied linearly with λ until a critical value (0.11), beyond which z0/h became constant. Marticorena et al. [2006] revisited the z0/h − λ relationship by adding to curve fitting in situ measurements over a range of dryland landscapes in south Tunisia. The critical λ value (0.045) in the Marticorena et al. [2006] formula causes too high z0 for our study domain, because the NDVI-derived λ is higher than 0.045 in most regions. On the other hand, the z0/h and λ measurements in Marticorena et al. [2006, Figure 10] reveal that the ratio z0/h keeps increasing with λ when λ is below 0.1. Therefore, we modify the critical λ value to 0.1 in the Marticorena et al. [2006] formula to accommodate the high λ values in our study: where λ is the sum of the roughness densities of nonvegetation and vegetation elements: λ = λ + λ; h is the effective geometric height by weighting the heights of nonvegetation (h) and vegetation (h) elements by their contributions to the total roughness density: h = (h · λ + h · λ)/λ. This implicitly assumes that the aerodynamic contributions of nonvegetation and vegetation elements are identical when weighted by their respective roughness densities, which is valid only for low-porosity (<20%) vegetation [Marticorena et al., 2006]. For nonvegetation elements, we prescribe λ and h values as a function of land cover, shown in Table5. The use of land type is to facilitate the model development through lookup tables, similar to those used in the land scheme to map the soil and vegetation parameters. Marticorena et al. [2006] reported the λ varied from 0.01 to 0.16 and the h varied from 0.8 to 2.0 cm at nine selected sites. They showed that degraded pasture lands tended to have a lower λ. Based on that, we use lower λ and h values for the land types of grassland, shrubland, and barren or sparsely vegetated than croplands. For vegetation elements, we use monthly h values to represent the vegetation growth-decay cycle and crop calendars. We realize it is impossible to accurately represent the vegetation mean geometric height in a 10 km × 10 km grid on a monthly basis. The h values listed in Table5 aim to capture the vegetation temporal dynamics and difference among land types (e.g., vegetation species). For example, cropland types (e.g., cotton and wheat) tend to have higher h values than the grassland and shrubland land types. Also, the cropland types have stronger temporal variations in the h with peaks in summer months, whereas the natural land types peak in spring.

Figure 10

The experiment mean (Exp_mean) monthly dust emissions obtained by averaging the MB_Dry, MB_Wet, and Shao_Dry model experiments.

To derive the monthly z0 affected by vegetation, we divide the study domain into three regimes according to NDVI. Exposed soils or barren surfaces correspond to very low NDVI values, commonly NDVI ≤ 0.1 (or A ≤ 4.6%) [e.g., Zou and Zhai, 2004]. Over the barren regions, λ and z0 are contributed by nonvegetation elements only and can be assumed constant with time. We assign the POLDER-derived static z0 values for the barren regions, given that the POLDER-derived z0 adequately captures the roughness heterogeneity of the study domain. Kimura et al. [2009, Figure 2] reported that the dust frequency dropped to nearly 0 when NDVI exceeded 0.3 over the loess plateau of China. Based on that, we use a criteria of NDVI > 0.3 (or A > 27.6%) to separate the regions that are completely protected against wind erosion. For regions in the range of 0.1 < NDVI ≤ 0.3, λ and z0 are contributed by both nonvegetation and vegetation elements. Their z0 is derived according to equation 12 from the monthly total roughness density (Figure 4b). Therefore, we are able to derive the dynamic z0 for the study domain by merging the POLDER-derived static z0 with MODIS monthly NDVI.

Figure 4c shows among the barren regions, the gravelly Ustyurt Plateau has a higher z0 (0.2–1.0 cm) than the sandy Karakum Desert (0.001–0.01 cm). Similar to A and λ, z0 displays strong seasonal variations driven by vegetation phenology in the steppe area, desert ephemerals, and agricultural lands. The presence of vegetation substantially increases the z0 (>1.0 cm). Our derived z0 values are in good agreement with the model estimates by Pierre et al. [2012] for the Sahelian grasslands.

3.2 Seasonality of Threshold Friction Velocity

Using the derived surface input parameters, we compute the u* and dust vertical fluxes at 30 min steps for the model experiments listed in Table4. Figure 5 shows the monthly mean u* for a soil diameter of 100 µm and the associated f and f for the MB_Dry experiment. The u* varies strongly in space and time. In the steppe area and grassland belt of 46°N–49°N, u* exceeds 0.8 m s−1 during spring and drops to 0.4–0.6 m s−1 during summer. In the sandy and loess deserts, u* exceeds 0.7 m s−1 during spring and drops to 0.3–0.5 m s−1 during summer. In the barren regions of Ustyurt Plateau, Caspian Sea coast, and Aralkum, u* exceeds 0.6 m s−1 during spring and drops to 0.3–0.5 m s−1 during summer. The lowest u* (0.2–0.3 m s−1) is found in the central Karakum Desert during summer.

Figure 5

Monthly (a) threshold friction velocity and corresponding (b) soil moisture and (c) surface roughness corrections for the MB_Dry experiment.

The spatiotemporal variability of u* is driven by the combined effects of soil moisture and surface roughness. Figure 5b shows the spring soil moisture causes u* to be 1.5–2.0 times as large as dry soils in the desert area and 2.0–2.5 times as large in the steppe area. As precipitation decreases and soil moisture depletes, the soil moisture effect diminishes during summer, especially in the southern deserts. Figure 5c shows the steppe vegetation causes u* to be 2.5–4.0 times as large as smooth surfaces during spring and 2.0–3.5 times as large during summer. In the sandy and loess deserts, ephemeral plants cause the u* to be 2.5 times as large as smooth surfaces during spring. For the barren regions, u* is 1.5–2.0 times as large as smooth surfaces. The roughness effect is smallest over the central Karakum Desert, where the z0 is close to z0.

Apparently, the u* dynamics is largely controlled by the seasonal precipitation in Central Asia. The spring peak rainfall causes higher soil water content and favors the growth of steppe vegetation and desert ephemeral plants, both leading to a higher u*. Due to the sporadic nature of precipitation and high potential evaporation in dryland areas, a post-rain increase in the soil water content may affect the u* for a short time (up to a day or two) before the soil moisture drops below the critical level [Ravi et al., 2004]. In comparison, the vegetation effect is continuous and lasts much longer through the growth-decay cycle. Figures 5b and 5c reveal the surface roughness causes stronger enhancement effect than soil moisture on the u*. In particular, the seasonal f values over the southern desert area are 2.13 (spring), 2.09 (summer), and 1.99 (autumn), while the f values are 1.75 (spring), 1.22 (summer), and 1.38 (autumn). For the entire dust season, the surface roughness effect is 30% stronger than the soil moisture effect in the southern desert area.

For a given soil grain size, the selection of soil size distribution data affects the u* through the clay content and z0, which affect the f and f, respectively. From equations 3 and 4, f and f are inversely related to the clay content and z0, respectively. Figure 6a shows that, compared to the dry-sieved soil data, the soil texture data cause small differences (±5%) in the u* over most desert areas. Figure 6b shows that the soil texture data produce smaller u* (by more than 30%) over the Ustyurt Plateau in spring and October, due to a weaker soil moisture effect than generated by the dry-sieved data. In fact, the soil texture data produce no soil moisture effect (f = 1.0) over the Ustyurt Plateau because of a higher clay content. On the other hand, Figure 6c shows that the soil texture data lead to a higher u* over the grassland belt (46°N–49°N) and southern mountainous piedmont area, because of a stronger roughness effect. This is due to the smaller z0 in the soil texture data.

Figure 6

Ratios of the MB_Wet to MB_Dry experiments in the monthly (a) threshold friction velocity and (b) soil moisture and (c) surface roughness corrections.

Compared to the MB scheme (MB_Dry), Figure 7a shows that the Shao scheme (Shao_Dry) produces a lower u* over the barren surfaces of Ustyurt Plateau, Caspian Sea coast, and Aralkum, whereas the Shao scheme produces a higher u* over vegetated areas. This is consistent with Figure 7b, which shows the Shao scheme produces a stronger (weaker) roughness effect over vegetated (barren) surfaces than the MB scheme. On the other hand, at some parts of barren regions, the smaller f is partly compensated by the higher u* in the Shao scheme, thereby leading to a higher u*. The comparison reveals a stronger u* sensitivity to vegetation in the Shao scheme than the MB scheme, which may lead to large differences in the dust emission seasonality in response to vegetation phenology.

Figure 7

Ratios of the Shao_Dry to MB_Dry experiments in the monthly (a) threshold friction velocity and (b) surface roughness correction.

3.3 Analysis of Surface Winds

To evaluate the model-predicted U10, we focus on the high tail of the wind speed distribution, which is highly relevant for dust emission and one of the largest sources of errors in dust emission models [Knippertz and Todd, 2012]. We compute the frequencies of strong wind speeds from 30 min predicted U10 and 3-hourly observed U10. Figure 8a shows that a high frequency (more than 30%) of U10 > 6.5 m s−1 occurs over the vast steppe and desert areas driven by the springtime cyclonic activity. Strong winds tend to be more frequent over the desert area than the steppe area. In particular, strong winds constantly blow over the Ustyurt Plateau, Caspian Sea coast, Aralkum, and loess deserts throughout the dust season. In the southern desert, the most frequent U10 > 6.5 m s−1 events occur in April (36%), followed by June (29%), March (28%), July (27%), October (25%), September (24%), May (22%), and August (18%). In addition, April reports 11% wind events exceeding 10 m s−1, more than twice as frequent as in other months. Such extreme winds may lead to enormous dust emissions, because of the power law relationship between the dust flux and wind speed. Compared to ground observations (Figure 8b), the model generally captures the spatial pattern of the strong wind frequency, in other words, the locations that experience high winds. Also, the model is consistent with the observed wind speed seasonality, which shows more frequent strong winds in spring.

Figure 8

Frequency of U10 > 6.5 m s−1 based on (a) 30 min model predictions and (b) 3-hourly ground station observations.

3.4 Seasonality of Dust Emission

Following the model experiments in Table4, we compute the dust vertical flux (g cm−2 s−1) and monthly dust emissions (megaton (Mt)) using a cutoff particle diameter of 35 µm. We tune the TF_U10 season-total dust emission to match the mean of the MB_Dry, MB_Wet, and Shao_Dry experiments. Because the dust source function used in the TF scheme is available at 1° × 1° resolution, the dust emissions in all experiments are aggregated onto a 1° × 1° grid for comparison. Figure 9a shows that the MB_Dry experiment produces strong emissions from the Ustyurt Plateau, Aralkum, Caspian Sea coast, and loess deserts. The Betpak-Dala and Kyzylkum deserts emit large amounts of dust in April and July. In contrast, the Karakum Desert appears to be a weak source. Despite more frequent strong winds in spring, spring dust is less than summer because of a higher u*. Table6 shows the spring, summer, and autumn dust account for 35.7%, 44.8%, and 19.5% of the season-total emission (365.0 Mt), respectively.

Figure 9

Monthly dust emissions from the (a) MB_Dry, (b) MB_Wet, and (c) Shao_Dry experiments.

Table 6

Dust Emissions of All Model Experiments

Experiment	Season-Total Emission (Mt)	Monthly Emission (%)
		Mar	Apr	May	Jun	Jul	Aug	Sep	Oct
MB_Dry	365.0	7.5	15.1	13.1	20.9	18.0	5.9	9.6	9.9
MB_Wet	39.8	5.6	17.2	12.1	17.0	21.4	7.3	9.9	9.6
Shao_Dry	361.9	0.4	11.4	5.2	22.7	24.8	9.2	20.1	6.1
Exp_mean	255.6	4.0	13.5	9.3	21.5	21.4	7.5	14.6	8.2
TF_U10	255.6	13.4	13.3	8.9	14.3	16.0	10.1	11.9	12.1

Compared to dry-sieved soil size distribution data (MB_Dry), Figure 9b shows that using soil texture data (MB_Wet) leads to substantial decreases in dust emission. Although the MB_Wet experiment identifies the same strong source areas as MB_Dry, the dust emission magnitude is much smaller. The season-total emission for MB_Wet is 39.8 Mt, eight times less than MB_Dry. Nonetheless, there is very small change in the seasonal distribution: spring and summer account for 34.9% and 45.7% of the total emission, respectively. Given that both experiments use the same u*, the large reduction in the MB_Wet emission can be explained by either an increase in the u* or the difference in the soil size distribution data. According to Figure 6a, the soil texture data lead to minor changes and even decreases in the u* over some strong source areas (such as Ustyurt Plateau and loess deserts), thereby ruling out the u* factor. Meanwhile, we find that the dry-sieved soil data generate a larger fraction of soil grains than the soil texture data within the saltation range of D ∼ 60–200 µm over all source regions. The presence of more erodible materials in the dry-sieved soil data may explain the much stronger dust emission in the MB_Dry experiment. Given that the dry-sieved data are considered more accurate in depicting the natural state of soil aggregation, using the soil texture data may underestimate the dust emission from Central Asia.

Figure 9c shows the monthly dust emissions for the Shao_Dry experiment. A striking feature is the strong contrast between barren and vegetated surfaces. That is, most dust is generated from regions with low roughness density (Figure 4b), including the Ustyurt Plateau, Aralkum, and Caspian Sea coast. The spatial pattern is also very similar to the u* difference between the Shao and MB schemes (Figure 7a), suggesting that the dust emission difference may result from the different u* parameterizations in the MB and Shao schemes. Compared to MB_Dry, Shao_Dry produces nearly the same amount of season-total emission (361.9 Mt). However, more dust is shifted toward summer (56.7%) compared to spring (17%) and autumn (26.3%). This is consistent with the Shao scheme producing a lower (higher) u* over barren (vegetated) surfaces than the MB scheme. That means temporally, the Shao scheme may overestimate summer dust and underestimate spring dust. And spatially, the Shao scheme may overestimate (underestimate) dust emission from barren (vegetated) surfaces. Indeed, Kang et al. [2014] found that the Shao scheme overestimated the dust emission from the Mongolian grasslands after the drying of aboveground vegetation. They noted that introducing additional roughness effect of dead plant residues led to improved predictions of the dust concentration.

To this end, we have identified two key sources of dust emission uncertainty in the WRF-Chem-DuMo model for the 2001 dust season in Central Asia: (1) the selection of dry-sieved or soil texture-based soil size distribution data, which strongly affects the season-total emission but has a minor effect on its seasonal distribution, and (2) the selection of MB or Shao schemes, which strongly affects the dust seasonal distribution but has a small effect on the total emission amount. As an attempt to bracket the uncertainties and biases of individual experiments, we calculate an experiment mean (Exp_mean) dust emission by averaging the monthly emissions of the MB_Dry, MB_Wet, and Shao_Dry experiments, shown in Figure 10. Based on the Exp_mean estimate, strong dust sources include the Ustyurt Plateau, Aralkum, Caspian Sea coast, and loess deserts, whereas the sandy deserts are weak sources. The Exp_mean season-total emission is 255.6 Mt, of which 26.8% is produced in spring, 50.4% in summer, and 22.8% in autumn.

Unlike the MB and Shao schemes, the TF scheme relies on a preferential dust source function to empirically represent the dependence of the erosion threshold on surface properties. Figure 11a shows that the GOCART dynamic dust source function (S) displays an increasing trend from spring to summer as the vegetation fraction decreases. Large S values highlight the dust source areas of the Karakum, Kyzylkum, Betpak-Dala, and Saryesik-Atyrau deserts (S > 0.7), as well as the Ustyurt Plateau and Caspian Sea coast (S ∼ 0.6–0.7). According to equation 9, the dust source function combined with the strong wind frequency determines the dust emission seasonality and spatial distribution in the TF_U10 experiment. Figure 11b shows that, compared to Exp_mean, the TF scheme produces more spatially evenly distributed dust emissions and identifies the Karakum and Kyzylkum sandy deserts, Ustyurt Plateau, and Caspian Sea coast as strong source areas. The dust source function is able to shift more dust emission to summer, despite more frequent strong winds in spring. Spring and summer account for 35.6% and 40.4% of the TF_U10 season-total emission, suggesting that the dust source function partly captures the soil erodibility dynamics in the physically based schemes. The missing Aralkum source in the TF_U10 experiment suggests the dust source function fails to capture the formation of Aralkum, due to an inaccurate land/water mask used in the AVHRR NDVI product. This points out the empirical dust source function derived from satellite climatology of a given time period may cause systematic dust model errors in case of drastic land cover/land cover changes.

Figure 11

The monthly (a) GOCART dust source function and (b) dust emissions in the TF_U10 experiment.

To the best of our knowledge, only the Soviet scientists have so far conducted in situ monitoring of dust emission from Central Asia, specifically from Aralkum. Their original studies were published in Russian and summarized in the United Nations report by Orlovsky and Orlovsky [2002]. Based on the land/water mask in WRF-Chem-DuMo, we estimate the size of Aralkum as 42,900 km2 in 2001, compared to a full sea size of 67,500 km2 [Micklin, 2007] and a reduced size of 40,300 km2 in 1999 [Breckle et al., 2001] and 42,000 km2 in 2000 [Singer et al., 2003]. The Aralkum season-total dust emissions are 2.02, 0.17, and 36.3 Mt based on the MB_Dry, MB_Wet, and Shao_Dry experiments, respectively. Orlovsky and Orlovsky [2002] reported that the Aralkum annual dust emission was estimated as 7.3 Mt yr−1 for the period of 1966–1979. They further showed the Aralkum dust emission was most severe during two periods, 1970–1971 and 1983–1985, during which severe drought anomaly occurred in Central Asia according to the Palmer Drought Severity Index [Dai, 2011]. Drought is known to exacerbate wind erosion by depleting the soil water and reducing vegetation cover [e.g., Cook et al., 2009]. Barlow et al. [2002] reported a prolonged drought event in Central and Southwest Asia during 1998–2001, which would most likely lead to above-normal dust emission. That means compared to the multiyear average (7.3 Mt yr−1), the MB scheme (MB_Dry and MB_Wet) may have underestimated the dust emission from Aralkum. On the other hand, it is likely that the Shao scheme (Shao_Dry) has overestimated dust emission from the barren surfaces of Aralkum. By simply averaging the three experiments, we obtain a more reasonable estimate by offsetting the positive and negative biases of the MB and Shao schemes. The Exp_mean estimate yields a season-total dust emission of 12.8 Mt from the Aralkum.

To compare the dust emission strength of Central Asia to the neighboring East Asia, we refer to the study by Laurent et al. [2006], who estimated the total dust emission as 459 Mt from the Chinese and Mongolian deserts in 2001. Their study is comparable to our MB_Dry experiment in that they used the same u* formulation and POLDER-derived static z0, although they used different dust vertical flux parameterization and wind data. We find the MB_Dry and Exp_mean estimates are 20% and 44% less than the dust emission from East Asia, respectively.

Global dust emission estimates are usually obtained by constraining simulations of dust burden or deposition using ground-based and satellite observations [e.g., Cakmur et al., 2006]. Because the observation data sets used in global models are heavily weighted toward the most important dust sources such as Sahara Desert, less constraint is applied to Central Asia where dust observations are scarce. To provide a perspective on the importance of Central Asia to global dust budget, we refer to the study by Cakmur et al. [2006, hereinafter C06], who obtained an optimal range of global annual dust emission (1533–2608 Mt yr−1) through minimizing the difference between a global model and a comprehensive collection of dust observations. They reported even larger uncertainty (147–650 Mt yr−1) for a region bounded by 25°E–90°E and 32°N–53°N (also referred to as Central Asia in their study but larger than our study domain, hereafter referred to as CentralAsia_C06). The uncertainties in the C06 dust emission estimates resulted from four model experiments using different dust source functions. It implies the dust source function alone can lead to substantial model errors in dust emission. Using equal weights, the C06 experiments yield an average estimate of 429 Mt yr−1 for CentralAsia_C06, 67% larger than our Exp_mean estimate, and 2184 Mt yr−1 for global source areas. Accordingly, our Exp_mean estimate suggests Central Asia contributes 10–17% (on average 12%) to the global dust emission in 2001.

3.5 Comparison With Observations

Because there are no direct measurements of dust emission, we attempt to evaluate the dust model using ground station dusty weather observation and satellite aerosol products. To compare the performance of different model experiments in representing the dust seasonality, we use the root mean square difference to describe the closeness (cl) between the modeled dust emission (M) and dust observations (O): where M and O are normalized by their season totals so that ∑M = 1 and ∑O = 1. Equation 13 does not intend to provide an absolute measure of the model performance, because a perfect agreement between (normalized) dust emission and dust loading is unlikely, given that total dust emission is mainly contributed by nonsuspendable particles. Rather, the purpose of cl is to serve as a relative measure of how accurately the different experiments capture the dust seasonality reflected in a given data set. Furthermore, because the GOCART dust source function was originally tuned to match the TOMS AAI, the TF scheme may inherently make a better comparison (i.e., lower cl) against satellite observations than the MB and Shao schemes.

At dust source regions, dust loading is mainly controlled by emission process while the effects of transport and removal processes are relatively small. This allows using visibility or dusty weather data from meteorological stations to examine the dust source variability [e.g., Orlovsky et al., 2005; Indoitu et al., 2012]. Figure 12a shows the monthly DFI derived from the dusty weather records using equation 1. Because the present weather observation is limited to manned stations only, most stations at the heart of major deserts have discontinuous or even no dust records. Figure 12a reveals that frequent dusty weather occurs at stations located at the Caspian Sea coast, Aralkum, Amu, and Syr Darya basins through the entire dust season and at the loess deserts during summer and autumn. The Exp_mean dust emission (Figure 10) agrees with the observed DFI by showing strong dust emissions from these areas.

Figure 12

(a) Monthly dust frequency index (DFI) and (b–e) comparisons between normalized monthly dust emissions and DFI at four hot spot regions: Aral Sea, Syr Darya basin, Caspian Sea coast, and loess desert.

To evaluate the model performance using equation 13, we focus on four hot spot regions: Aral Sea, Syr Darya basin, Caspian Sea coast, and loess deserts. Each region is represented by the enclosed or nearby ground stations with available DFI data: Aral Sea—Aralskoe More (Kazakhstan, 46.78 N, 61.65°E); Syr Darya basin—Dzhusaly (Kazakhstan, 45.5°N, 64.08°E) and Kyzylorda (Kazakhstan, 44.85°N, 65.5°E); Caspian Sea coast—Akkuduk (Kazakhstan, 42.97°N, 54.11°E) and Krasnovodsk (Turkmenistan, 40.03°N, 52.98°E); and loess deserts—Termez (Uzbekistan, 37.23°N, 67.27°E) and Karshi (Uzbekistan, 38.8°N, 65.72°E). For each hot spot region, the monthly DFI of all stations is averaged and normalized as O in equation 13. The counterpart M is computed from the dust emissions from the model grids corresponding to the station sites, except for the Aral Sea hot spot where dust emission from the entire Aralkum is used. The results are shown in Figures 2b–2e. Apparently, dusty weather is more frequent in spring (especially April) in the Aral Sea and Syr Darya basin regions. In contrast, the Caspian Sea coast and loess deserts experience maximum frequency of dusty weather in summer and autumn. This difference might be due to different dominant factors—the Aral Sea and Syr Darya basin are largely influenced by strong winds during spring (Figure 8), whereas low vegetation and dry soils may cause more severe soil erosion at the Caspian Sea coast and loess deserts during summer (Figure 5). The closeness between the Exp_mean dust emission and DFI falls within the range of the closeness between DFI and individual experiments, because of the compensation of model biases in the Exp_mean estimate. Furthermore, the Exp_mean and TF_U10 experiments show mixed performance at different regions. The TF_U10 dust emission has a better agreement with DFI in the Syr Darya basin, whereas the Exp_mean emission has a better agreement with DFI in the Caspian Sea coast. They show similar agreement with DFI in the loess deserts. Note that TF_U10 is excluded from the Aral Sea hot spot because the TF scheme does not produce dust emission from the Aralkum (Figure 1b).

Compared to ground stations, satellites offer much larger spatial coverage and revisiting capability for continuous monitoring of dust aerosols. Satellite aerosol products have been used to map global dust source regions, based on the rationale that active sources are associated with frequent occurrences of large AOD or AAI values [Prospero et al., 2002; Ginoux et al., 2012]. Draxler et al. [2010] developed an empirical dust scheme based on multiyear MODIS AOD statistics, implying that long-term dust loading observations near source areas can be used to represent the emission strength. Furthermore, Indoitu et al. [2012] showed the daily dust activity peaks between local time 9:00 and 13:00 in Central Asia, which coincides with the MODIS/Terra overpass. Therefore, satellite aerosol observations may provide valuable information to evaluate the dust emission seasonality.

Figure 13 shows that the monthly MODIS and SeaWiFS deep-blue AOD and TOMS AAI. Large MODIS AOD (>0.7) values are found over the Ustyurt Plateau, Aralkum, Caspian Sea coast, and loess deserts, reaching a maximum of 2.8 at the Ustyurt Plateau in July. The Karakum and Kyzylkum deserts have relatively lower AOD (<0.5). The domain-averaged AOD indicates more dust in summer (0.52) than spring (0.38), with a peak in June (0.62). Seemingly, the AOD spatial and temporal patterns are in good agreement with the Exp_mean dust emission. Interestingly, Figure 13a shows that the Taklimakan Desert has the largest AOD in April. In this hyperarid desert, the effects of vegetation or soil moisture are so weak that the dust seasonality is mainly controlled by the atmospheric circulation. The contrasting dust seasonality in Central Asia reflects the importance of the surface characteristics in controlling the dust activity in semiarid drylands.

Figure 13

Monthly observations of dust aerosol: (a) MODIS deep-blue AOD, (b) SeaWiFS deep-blue AOD, and (c) TOMS AAI. The domain-averaged values are shown in each panel.

Compared to MODIS, Figure 13b shows large SeaWiFS AOD at the same regions but with much smaller magnitude. The underestimation is very likely due to the SeaWiFS cloud screening that may misclassify strong dust plumes as clouds (A. Sayer, personal communication, 2014). In addition, SeaWiFS has no AOD retrieval over the Aralkum, which is incorrectly treated as water body in the land/water mask used in the retrieval algorithm (A. Sayer, personal communication, 2014).

AAI has been used to identify UV-absorbing aerosols such as dust [Herman et al., 1997; Prospero et al., 2002]. Figure 13c shows that the TOMS AAI falls in the range of 1.0–1.4 over most desert areas. The highest AAI (>1.4) values occur at the Ustyurt Plateau, Aralkum, Caspian Sea coast, and loess deserts, consistent with the MODIS AOD. These AAI values are lower than the Taklimakan Desert where AAI exceeds 2.0. This can be explained by several factors: (1) lower spring dust loading from Central Asia compared to Taklimakan Desert (Figure 13a), (2) reduced instrument sensitivity to dust plumes over elevated areas such as the Ustyurt Plateau [Mahowald and Dufresne, 2004], and (3) the rich salt content in the dust aerosols that may reduce the dust absorption capability [Singer et al., 2003].

In addition to monthly averages, we compute the frequency of occurrence (FOO) of daily MODIS AOD > 0.7 (FOOAOD) and TOMS AAI > 1.0 (FOOAAI), which can be used as a measure of the dust source strength [Prospero et al., 2002; Draxler et al., 2010; Ginoux et al., 2012]. We exclude regions with less than 9 or 30% daily observations in a month. Figure 14a shows that the FOOAOD reaches 80% in the Ustyurt Plateau, Aralkum, and Caspian Sea coast from April to September. In other words, the daily AOD is greater than 0.7 within at least 24 days out of a month. In contrast, the FOOAOD is less than 20% in the Karakum and Kyzylkum deserts. Compared to FOOAOD, Figure 14b shows very similar spatial pattern in the FOOAAI, which reaches a maximum (>80%) in the Ustyurt Plateau, Aralkum, Caspian Sea coast, and loess deserts. The Exp_mean dust emission agrees with the FOOAOD and FOOAAI by capturing the distribution of strong and weak source areas.

Figure 14

Frequency of occurrence (FOO) of (a) MODIS daily AOD > 0.7 and (b) TOMS daily AAI > 1.0.

Furthermore, we compute the closeness using equation 13 between the monthly dust emissions and the MODIS AOD and TOMS AAI, shown in Table7. We select three hot spot regions that, when combined, account for over 80% of the Exp_mean season-total emission: Ustyurt Plateau and Caspian Sea coast (50°E–58°E, 37°N–48°N), Aralkum and Amu, and Syr Darya basins (59°E–65°E, 41°N–47°N), and loess deserts (60°E–70°E, 37°N–41°N). Overall, the dust model shows a better agreement with AOD and AAI than with the DFI, thanks to the larger spatial coverage of satellite observations. The MB scheme is in a better agreement with the AOD and AAI than the Shao scheme, indicating a potential bias in the Shao scheme. The bias may result from the strong u* sensitivity to vegetation and is partially compensated in the Exp_mean estimate. The TF scheme generally shows a similar degree of agreement with the AOD and AAI as the MB scheme and a better agreement than the Shao scheme. For the hot spot of Aralkum and Amu and Syr Darya basins, the GOCART dust source function does not capture the Aralkum source, which worsens the comparison of TF_U10 against satellite observations.

Table 7

Closeness (cl) Between Normalized Monthly Dust Emissions of Different Model Experiments and Satellite AOD and AAI Observations

		Ustyurt Plateau and Caspian Sea Coast	Aralkum, Amu, and Syr Darya Basins	Loess Deserts	Entire Domain
AOD	MB_Dry	0.03	0.01	0.06	0.03
	MB_Wet	0.03	0.01	0.04	0.03
	Shao_Dry	0.09	0.01	0.01	0.07
	Exp_mean	0.04	0.01	0.03	0.05
	TF_U10	0.01	0.02	0.01	0.03
AAI	MB_Dry	0.02	0.01	0.06	0.05
	MB_Wet	0.02	0.01	0.04	0.06
	Shao_Dry	0.1	0.01	0.01	0.1
	Exp_mean	0.05	0.01	0.03	0.08
	TF_U10	0.02	0.02	0.01	0.04

4 Conclusions

We present a model study on the seasonal threshold friction velocity (u*) and dust emission in Central Asia for the dust season from 1 March to 31 October 2001, focusing on the effects of atmospheric circulation (i.e., wind speed), soil moisture, surface roughness, and vegetation phenology. Three dust schemes, including two physically based schemes (MB and Shao) and one empirical scheme (TF), and two soil grain size distribution data sets, including soil texture and dry-sieved, are used in multiple model experiments to investigate the dust emission uncertainty due to the selection of dust scheme and soil size distribution data. Ground station and multisensor satellite observations are used to evaluate model predictions of surface winds and dust seasonality.

We find the u* significantly varies in space and time, in response to the soil moisture variation, surface roughness heterogeneity, and vegetation phenology. The u* is greatly enhanced by the green vegetation of temperate steppe grasslands and desert ephemeral plants and, to less extent, the wet soils due to the spring peak rainfall. As the aboveground vegetation dies back and soil moisture depletes, the u* decreases in the dry hot summer. For regions with low vegetation, the u* depends on the presence of solid roughness elements and displays a higher value over the gravelly Ustyurt Plateau than the central Karakum sandy desert. In contrast to natural lands, irrigated croplands in the Amu and Syr Darya river basins are protected against wind erosion during summer due to a high vegetation cover.

As a result of the u* dynamics, the MB and Shao schemes produce more dust during summer, despite more frequent strong winds during spring. Due to the presence of more erodible soil grains within the diameter range of 60–200 µm, the dry-sieved soil size distribution data lead to eight times more season-total dust emission (365.0 Mt) than the soil texture data (39.8 Mt), but with minor changes in the seasonal distribution. Compared to the MB scheme, the Shao scheme produces nearly the same season-total dust emission (361.9 Mt), but with a shift toward summer due to the strong sensitivity of the u* to vegetation. Averaging three model experiments reduces the biases of individual experiments in the total dust emission and seasonal distribution. The resultant experiment mean (Exp_mean) estimate yields a total dust emission of 255.6 Mt for the 2001 dust season, of which 26.8% is produced in spring, 50.4% in summer, and 22.8% in autumn.

Comparison of the Exp_mean dust emission to Soviet-era dust monitoring at the Aralkum suggests that the Exp_mean produces a reasonable dust emission estimate from the Aralkum in 2001 (12.8 Mt). Compared to Laurent et al. [2005], the Exp_mean estimate is 44% less than the dust emission from East Asia in 2001. Using Cakmur et al. [2006] as a reference, we find that Central Asia contributes 10–17% (on average 12%) to the global dust emission. The Exp_mean dust emission is in general agreement with dusty weather frequency and satellite AOD and AAI observations in terms of the spatial distribution and seasonality. Compared to Exp_mean, the TF scheme generates more spatially evenly distributed dust emission and partly captures the summer dust peak, thanks to the dynamic dust source function.

Through this study, we demonstrate that the dust seasonality in semiarid drylands is regulated by concomitant changes in the atmospheric circulation and surface characteristics. It is imperative for coupled dust-climate models to account for the effects of key controlling factors, such as soil moisture, surface roughness, and vegetation dynamics, on the threshold friction velocity for an accurate representation of the dust seasonality. This study also highlights the challenges in applying physically based dust models to semiarid dryland regions, such as the availability of dry-sieved soil size distribution data and treatment of soil moisture and vegetation effects. We map the dry-sieved soil size measurements of Chinese deserts to Central Asia based on the similarities in the soil (i.e., soil texture) and land (i.e., land type and surface roughness) characteristics. The mapping method is based upon detailed comparisons on the dust source origin and landscape environment between China and Central Asia, both of which are located in the midlatitude cold desert belt and influenced by similar seasonal weather systems and land use disturbance. Given that the soil texture-based soil size distribution data cause substantial model bias in dust emission, there is a great need for dry-sieved soil size measurements of global dust sources for an accurate representation of the saltation diameter range (60–200 µm).

To account for the soil moisture effect on dust emission, we use a scale factor of 0.3 to adjust the Noah land scheme-predicted 0–10 cm soil moisture to use in the Fecan et al. [1999] parameterization. This scale factor is selected based on the assumption of no soil moisture effect at the Karakum Desert in summer and at the same time meets the need of keeping the spatiotemporal variability of predicted soil moisture in the dust model. The need for such empirical tuning of soil moisture reflects an important issue in dust models, that is, the mismatching scale between dust schemes and host climate models. This necessitates the development of soil erodibility and dust flux data sets at appropriate climate model scales that can be used to develop and test dust schemes in large-scale models and possibly be used to refine other model components, such as the land surface scheme, in order to meet the needs of dust schemes [King et al., 2012].

By using MODIS NDVI to derive the surface roughness properties, we implicitly neglect the effects on dust emission of nonphotosynthetic vegetation (NPV), including dead plant residues and brown vegetation. The NPV is known to inhibit dust emission by reducing soil exposure and increasing the erosion threshold [e.g., Zou and Zhai, 2004; Abdourhamane Toure et al., 2011]. So far, the NPV effects on the dust cycle are poorly understood and receive little attention from dust modelers. It remains unclear if the model parameterizations designed for green or photosynthetic vegetation (PV) can be readily used for NPV, which may interact in different ways with wind erosion and saltation processes. Okin [2010] proposed a spectral mixture analysis method to separate the NPV and PV signals in the satellite reflectance measurements. It is however unclear how the NPV spectral signal can be validated and be translated into the input parameters of dust models. Further, vegetation may affect dust emission in more complicated ways than described by the MB and Shao parameterizations. Okin [2008] introduced a wind shear partition scheme accounting for the spatial structure of vegetation distribution that affects the air flow and saltation flux of unvegetated gaps. A better understanding of the importance of these factors to the dust variability at semiarid dryland regions requires more observational and physically based model studies in the dust community.