New Inventory of Dust Emission Sources in Central Asia and Northwestern China Derived From MODIS Imagery Using Dust Enhancement Technique

The first inventory of dust emission sources in Central Asia and northwestern China (35–50°N, 50–100°E) derived from the twice daily MODIS imagery from 2003 to 2012 is presented. A high‐resolution (1 km) dust enhancement product was generated to produce maps of dust point sources (DPS), indicating geographical locations of the observed dust emissions, and gridded data sets of dust emission frequencies. About 13,500 DPS were detected over an area of ∼5 × 106 km2, however, their distribution was uneven. The highest frequency of DPS occurred in the northern and eastern Taklimakan, in the Aralkum, and in the regions, which were not widely reported in literature before, the Balkh delta in northern Afghanistan and the pre‐Aral, from the northern Caspian coast to the Betpak‐Dala desert in Kazakhstan. South of the Aral, DPS were mainly associated with fluvial features: drying lakes, dry river beds, alluvial deposits and agricultural activity which is closely linked to water availability in this arid region. In the pre‐Aral and Balkhash‐Junggar regions, land damaged by wildfire was the main source of dust. In China and eastern Kazakhstan dust emissions peaked in spring; in Central Asia and western Kazakhstan—in summer. The Aralkum was active throughout the year with a positive trend in emission frequency over the study period. Locations of DPS did not always correlate with high atmospheric optical depth (AOD) particularly west of the Aral and in the southern Taklimakan where few DPS were detected despite the high AOD values. This was attributed to dust transport from the upwind sources.

complex and spatially nonhomogeneous response of dust emissions to climate change and variability (Chin et al., 2014;Xi & Sokolik, 2015a).
There is extensive research on source regions, entrainment processes, and transportation of mineral dust in the atmosphere (Shao et al., 2011). The latter is most frequently characterized by Atmospheric Optical Depth (AOD) measured either in situ by the Aerosol Robotic Network (AERONET; Holben et al., 2001) or using remote sensing (e.g., Ginoux et al., 2010;. Other records of dust-related indices are available with a long-term global coverage, for example, the Total Ozone Mapping Spectrometer (TOMS) product (Washington et al., 2003). The Spinning Enhanced Visible and Infrared Imager (SEVERI) was used to develop regional inventories of dust emissions with very high temporal resolution in the regions covered by the Meteosat Second Generation (MSG) geostationary satellite, that is North Africa (Schepanski et al., 2009) and the Middle East (Hennen et al., 2019). There is a growing number of publications on mineral dust entrained from the deserts of China and Mongolia (e.g., Che et al., 2013;Shao & Dong, 2006;Qian et al., 2007;Wang et al., 2008Wang et al., , 2006Xuan & Sokolik, 2002).
Information on mineral dust in Central Asia (CA) and northwestern China (NWC), which accommodate vast deserts and dry steppes (Figure 1), including spatial distribution of sources, temporal trends in their activity, and atmospheric pathways of dust is more limited, although the importance of this region in the global dust cycle, and the role of changes in land cover leading to alterations in dust emissions, are recognized (Groisman et al., 2017;Indoitu et al., 2012;Monier et al., 2017;Sternberg, 2017;Wang et al., 2008). Regional economic activity, including agricultural production, and the health of millions of people are affected by dust storms and high concentrations of mineral aerosol (Orlovsky et al., 2005;Sternberg, 2017). Deposition of desert dust on the snow pack and glaciers in the Tien Shan, Pamir, and Hindukush significantly alters surface albedo, creating positive at-surface radiative forcing and enhancing snow and ice melt (Painter et al., 2012).
Research on dust storms began in the 1930s in CA and in the 1960s in NWC when regular monitoring of dust storms was introduced at regional meteorological stations. Analyses of changes in dust storm NOBAKHT ET AL. frequency, based on these data and covering the period between the 1960s and 2000s, were published by Indoitu et al. (2009) and Issanova et al. (2014;2015a;2015b) for CA and by Wang et al. (2006) for NWC. In the 1980s, when the desertification in the Aral Sea region was first reported, data acquisition campaigns were conducted, collecting a wide range of data on optical and radiative properties of desert dust (Hansen et al., 1993;Sokolik & Golitsyn, 1993) and meteorological conditions of dust storm formation (Smirnov et al., 1993). Following the breakdown of the USSR, data were mostly obtained through field campaigns in the Aral Sea region (e.g., Groll et al., 2013). There is a growing number of field studies in NWC (Wang et al., 2008). Currently, there are four operational AERONET sun photometers in Kyrgyzstan, Tajikistan, and the Muztag-Ata mountains in China. Ground-based lidar measurements (Hofer et al., 2017;2020) and regular PM 2.5 and PM 10 sampling (pers.com., Dr. S. Abdulaev, Physical Technical Institute of the Academy of Sciences of Tajikistan) are conducted in Tajikistan. China Aerosol Remote Sensing Network (CARSNET) began operating in NWC in 2002 (Che et al., 2015).
The use of remote sensing, particularly the Deep Blue AOD retrieval algorithm (Hsu et al., 2004), introduced in Collection 5.1 (C5) of MODIS aerosol products, enabled improved retrieval of aerosol properties over deserts, resulting in the expansion of dust AOD data collection globally. Ginoux et al. (2012) used this approach to characterize global distribution of desert dust in the atmosphere during the 2003-2009 period. However, the focus was on eastern Asia and the Middle East, while CA was represented to a limited extent. Further research showed that the Deep Blue AOD of mineral dust was underestimated in comparison with the AERONET measurements, especially over deserts (Shi et al., 2013) and, although this comparison was performed specifically for North Africa, the same limitation is likely to apply in CA (Xi & Sokolik, 2015a). The Enhanced Deep Blue (E-Deep Blue) aerosol product was introduced in Collection 6 (C6) of MODIS products as an improved version of the C5 Deep Blue product, enabling the extension of spatial coverage from the bright reflecting regions to the entire cloud-free and snow-free land area and reducing uncertainty of retrievals (Sayer et al., 2013). The C6 AOD was used by Xi and Sokolik (2015a) in combination with simulated dust emissions to analyze interannual variability and trends in dust load over CA, identifying a negative trend in dust activity attributed to decreasing surface wind speeds.
Several studies (Ginoux et al., 2012;Xi & Sokolik, 2015a;2015b) used AOD to characterize dustiness over CA and NWC highlighting the diversity of land-cover and land-use types contributing to dust emissions and varying sensitivity of dust fluxes to meteorological conditions. However, these and other studies using AOD were based on the retrievals of aged dust in the atmosphere, potentially displaced from its sources. There is a lack of detailed identification of actual dust sources, land-surface properties resulting in dust emissions, temporal trends and characterization of conditions of dust source activation (DSA) in CA in contrast to other major dust source regions for which this information is more widely available, for example, the Sahara (Schepanski et al., 2007;2009) and the Middle East (Hennen et al., 2019). Wang et al. (2006; discussed dust storm climatology in the Taklimakan in the context of land surface characteristics. A detailed analysis, characterizing erodible land surfaces and statistics of individual dust plume formation was produced by Walker et al. (2009) using MODIS imagery for the southwestern part of the region including Afghanistan, Iran and Pakistan. For the rest of CA, such information is not available limiting our understanding of the dust cycle in CA, our ability to model dust transport, to forecast dust storms, and manage land. This paper presents the first large-scale inventory of observed dust emissions and dust point sources (DPS) from which these emissions originated in CA and NWC, derived from twice daily MODIS imagery over the 2003-2012 period. We use the term DPS instead of "dust storm" because the latter can originate from multiple sources. DPS characterization of dustiness is different from that provided by AOD and discussed by Ginoux et al. (2012) and Xi and Sokolik (2015a;2015b) because DPS show actual emission sources while AOD shows the spread of the aged dust in the atmosphere. The two types of data are complimentary and in the future, DPS can be used as a control data set to resolve differences between the observed AOD and modeled dust entrainment (e.g., Xi & Sokolik, 2015a).
The objectives of this paper are to: (i) present DPS statistics including its locations and temporal variability; (ii) compare DPS locations with spatial distribution of AOD and Dust Optical Depth (DOD); (iii) briefly discuss landforms, land cover and land use types contributing to the formation of DPS. Meteorological conditions leading to the formation of dust storms and transportation of dust are addressed briefly and their detailed discussion will be presented elsewhere.

Data and Methods
The main data set produced by this study comprises DPS, which are defined as geographical locations of every single dust outbreak detected between January 2003 and December 2012. In addition, AOD values were derived as a supplementary data set to compare DPS locations with dust present in the atmosphere. MODIS Terra and Aqua acquisitions from overpasses between 10:30 and 13:30 local time (depending on location in the region), respectively, were used to generate both data sets. MODIS data were retrieved from NASA Level 1 and Atmosphere Archive and Distribution System (LAADS; https://ladsweb.nascom.nasa. gov) (Table S1).

Dust Optical Depth
Daily C6 MODIS Level-2 swath products MOD04 (Terra) and MYD04 (Aqua) datasets respectively were obtained as Hierarchical Data Format-Earth Observing System (HDF-EOS) granules for 2003-2012. Each granule covers 5 min of observation time along the satellite track or an area of 2,030 km along the path by 2,330 km along the swath (Savtchenko et al., 2004). Over 43,000 granules were obtained from the LAADS repository to produce daily raster mosaics for the study region at the ground resolution of 0.1° × 0.1°. The AOD retrievals by the Aqua satellite (MYD04) were used. Importantly, no AOD retrievals were available in MODIS C6 for the Aralkum due to the use of static surface reflectance model in AOD retrieval algorithm (pers. com., Dr. A. Sayer, MODIS AOD Developing Team, NASA Goddard Space Flight Center) and DOD was not calculated for this region.
To detect the location of dust hotspots (Gillette, 1999), frequency of occurrence of DOD (FO-DOD) and its spatial distribution were derived. The DOD datasets were constructed by applying three screening criteria to the AOD retrievals (Table S1): (i) Ångström Exponent (AE) less than 0.5; (ii) SSA at 412 nm less than 0.95; and (iii) increase of single scattering albedo with wavelength. Criterion (i), restricting the AOD observations to larger particles, is based on the inverse relationship between AE and dust particle size. Deposition of large dust particles is controlled by gravitational settling. They are likely to occur close to sources of dust emissions and indicate the presence of dust sources (Knippertz, 2014). However, this criterion can be satisfied in the absence of desert dust in coastal regions (e.g., the Caspian or Aral) or in proximity to other saline lakes (widespread in CA) where concentrations of coarse salt particles are high. To filter out salt particles, criterion (ii) was applied based on the higher SSA of sea salt (close to 1) in comparison with mineral dust. Criterion (iii) is based on a specific property of mineral dust -a positive difference between SSA at 650 nm and SSA at 412 nm (Ginoux et al., 2012). This three-step screening procedure was evaluated successfully by Ginoux et al. (2012) using AOD data from AERONET.
A ratio between number of days with DOD exceeding 0.2 (a threshold for background AOD; Ginoux et al., 2010) to the total number of AOD retrievals for the same season was calculated for each season and each grid cell and expressed as frequency of occurrence (%) of DOD values (FO-DOD). To avoid miscalculation over regions with scarce AOD retrievals, which can result from the complex background reflectivity, FO calculations for grid cells with less than 10% of AOD retrievals during the whole time period were discarded. The FO-DOD data were summarized as seasonal maps ( Figure S1).

Development of DPS Data Set: Dust Enhancement
Detecting dust over the land surface, and particularly over deserts, using visible-shortwave infrared (solar spectrum) optical satellite imagery is challenging because mineral aerosol and the background desert surface have similar reflectivity in this part of the spectrum (Baddock et al., 2009). This problem can be resolved by using dust enhancement techniques based on a brightness temperature difference (BTD) between the warmer ground surface and cooler elevated dust plumes, derived using emitted infrared (IR) and thermal infrared (TIR) bands (Baddock et al., 2009;Miller, 2003). This approach, however, can lead to potential misclassification of dust in the presence of cirrus clouds, which are cooler than the surface, and can be falsely identified as dust when IR/TIR channels only are used. To avoid this problem, we used the dust enhancement method (Miller, 2003) which combines a conventional BTD technique with information from reflective solar radiation bands (Table S1) to improve visual detection of dust events (Figure 2).
The dust enhancement algorithm is based on the following principles: (i) elevated dust produces lower brightness temperature against the hot land surface background; (ii) dust can be discerned from clouds having the same radiometric temperature based on its spectral properties; (iii) dust often produces a positive reverse BTD between MODIS bands 32-31 (12-11 μm); and (iv) information from the additional infrared MODIS Band 26 (1.38 μm) enables distinguishing between dust and cirrus clouds. Normalization bounds were determined by Miller (2003) experimentally from a wide variety of dust case studies over land and values were selected to optimize dust contrast while maintaining an enhancement appearance consistent with the over-water algorithm. The algorithm ( The definition of DE thresholds, discriminating between dust and nondust, involved compilation and examination of pixel histograms ( Figure 2). In each histogram, peaks were found and attributed to the specific scene components such as clouds or aerosol and threshold values, which represent the presence of dust in the scene best, were determined by the operator. The thresholds are sensitive to varying atmospheric conditions, surface reflectance, dust concentrations and mineralogy and need to be calculated for relatively small regions where dust sources have similar properties or even specifically for each dust event (Baddock et al., 2009;Darmenov & Sokolik, 2005). The use of the scene-or event-specific thresholds was shown to improve the recognition of dust significantly (Baddock et al., 2009). In this study, thresholds were derived for individual scenes (e.g., 0.9 for September 4, 2006 and 1.3 for March 30, 2008; Figure 2). If preliminary inspection indicated that multiple sources of dust were present, pixel histograms were resampled for dust-emitting areas within the scene to adjust dust/nondust threshold and enhance dust visualization. Therefore, DE thresholds varied temporally and between regions.   Table S1. BT is brightness temperature; R is reflectance; and DE is dust enhancement values over land. Terms for DE are from Miller (2003). Dust event observed on 30 March 2008 is shown in Figure S3.

Identification of DPS and Compilation of the Data Set
Daily DEP images were generated from MODIS Terra and Aqua L1B Top of the Atmosphere (TOA) radiance data for the period between January 1, 2003 and December 31, 2012 for the study region ( Figure 1). For both Terra and Aqua satellites, on average, seven granules were required to generate a daily mosaic covering the study area between 35-50°N and 50-100°E. Six spectral bands were extracted from each MODIS granule (Table S1; Figure 2). The data were re-projected from the MODIS native swath projection to WGS84 system at a 0.01° (∼1 km) resolution using the associated MODIS geolocation products (MOD03) and MODIS Re-projection Tool (MRT; https://lpdaac.usgs.gov/tools/modis_reprojection_tool) and used to create two daily mosaics of 5,000 × 1,500 km encompassing the study region at 0.01° (∼1 km) resolution.
The ∼2-hour time difference between the two overpasses enabled the investigation of the development of individual dust plumes between the two snapshots and establishing locations of DPS at the head of each dust plume with greater precision. While the DE method distinguishes between clouds and dust more efficiently than other methods, it is known to be sensitive to changes in ground surface reflectance. Baddock et al. (2009) demonstrated potential misclassification of ephemeral lakes and fire scars (areas de-vegetated by wildfire; Figure S4), which are potential sources of dust, as emitting dust sources. Walker et al. (2009) showed that false enhancement is possible over the cold terrain, particularly in winter and in the mornings. Inspection of two consecutive MODIS scenes for dust plume dispersion as well as additional imagery such as DOD and visible MODIS (e.g., Figure S2) helped to reduce this uncertainty.
DPS were identified manually and archived using the following procedure: (i) The DOD and DEP images were visually inspected. Although DOD, which has 10 km spatial resolution, does not necessarily capture point locations of DPS, it confirmed the occurrence of dust events pointing at the need for pixel histogram resampling and potential adjustment of thresholds for different events within a scene, and reduced risk of dust misclassification (ii) Dust emitting areas were visually determined and DPS marked as DSA events on the twice-daily DEP images. Manual mapping was successfully used before to map dust sources in the Sahara (Schepanski et al., 2007) and the Middle East (Hennen et al., 2019). To avoid bias created by increasing observer skill over time, the operator (M. Nobakht) trained on a year of data prior to the compilation of the DPS data set. Two control steps were introduced. (1) The upwind location of the dust plume had to be unchanged in both Terra and Aqua images. In majority of the events, the operator could visually observe the dust plume expansion between the two overpasses, while the head of the plume (recorded DPS) was at the same location in both snapshots.
(2) If DPS were visible on a single overpass image, auxiliary information about dust emitting surfaces (v) was used to confirm DPS location (iii) Locations (latitude, longitude) of all activated DPS and date of emission were recorded. If DPS appeared on both Terra and Aqua images on the same day, they were only recorded once. If emissions from DPS continued for several consecutive days, these DPS were recorded for each day. There was no discrimination between the intensity of dust storms (iv) In order to characterize uncertainty in the DPS data, "Quality Flags" (QF) on a scale of 1-3 were assigned to every recorded DPS. Criteria for QF assignment are shown in Table S2. QF1 corresponded to the most reliable data whereby DPS was clearly visible and there were no clouds or overpassing dust plumes (e.g., Figure S2b) and QF3 corresponded to the DPS partially obscured by meteorological or dust clouds ( Figure S3) (v) Auxiliary data were examined to increase confidence in locating DPS and gather information about the nature of the eroding surfaces including GTOPO30 global DEM at 1 km spatial resolution (available from the U.S. Geological Survey) and high-resolution satellite optical imagery for dust-free days (e.g., Worldview, GeoEye, Landsat, SPOT, and Ikonos) obtained from various sources (vi) The ERA Interim reanalysis surface wind data were recorded for the time and location of dust events and later updated to ERA5 Locations of all identified DPS, timing of dust events, auxiliary data and meteorological variables formed an inventory of DPS in CA and NWC. DPS locations were plotted against FO-DOD to compare the data sets. In addition, number of days with DSA were calculated and binned in 1° × 1° grid. Days with at least one DPS activation in a grid cell were considered as DSA days. The DSA maps do not account for the number of DPS in the grid cells and characterize frequency but not intensity of dust events.

Dust Point Sources
Overall, 13,500 DPS were registered in the study area in the 2003-2012 period. Spatial distribution of the identified DPS is summarized for two sub-regions: (i) CA (north and south of the Aral Sea) and (ii) NWC and eastern Kazakhstan (Figure 3) comparing DPS locations with the summer and spring FO-DOD values, respectively.  Degradation or removal of vegetation cover, predominantly by wildfire but also by overgrazing, were identified as the main processes leading to the formation of dust storms in the pre-Aral and Betpak-Dala. Fire scars remaining on the burned land were detectable on satellite imagery and dust emissions were traced back to the fire scars, identifying them as activated DPS (Figures S2 and S3). The amount of time required for the vegetation and biological soil crust to be restored over the burned areas is determined by a combination of burn severity, soil type and climatic conditions (Strong et al., 2010). In CA, fire scars remained active sources of dust for several years except in a few areas which regained their vegetation cover shortly after the onset of the wet season.

Central Asia: South of the Aral Sea
In this region, DPS were associated predominantly with fluvial features but also with agricultural landscapes which, for millennia, were developing in the proximity to water leading to the degradation of several water bodies, most notably the Aral Sea, in the 20th Century.
The highest density of DPS identified in this study characterized a newly developed desert, the Aralkum (Figure 3a). Approximately 59% of the DPS registered in the Aralkum occurred in the northern and eastern sectors of the desiccated lake where loamy deposits are the prevailing soil type and dry solonchaks are widespread (Kozhoridze et al., 2012). Around 20% of the observed DPS occurred in the southern and southeastern parts of the lake where lighter sandy and sandy-loamy soils prevail.
The remaining 21% were registered on the exposed lake bed between the former Amu Darya delta and Vozrozhdeniya Island which was a site for biological and chemical weapons testing in the 1950-1990s. Decontamination was carried out in 2002 but there is potential risk of contamination of dust emitted from this region by remaining toxic pathogens (Micklin, 2007). Importantly, very few DPS were detected on the Myngyshlak peninsula and virtually none on the Ustyurt plateau, west of the Aralkum, which were characterized by the highest AOD and FO-DOD values in this (Figures 3a and S1) and in previous studies (Xi & Sokolik, 2015a).
Further south, FO-DOD values peaked over the Kyzylkum desert between the Syr Darya and the Amu Darya Rivers (Figures 3 and S1). Here, DPS were recorded in relatively narrow zones on the floodplains in the middle courses of both rivers. Over 160 DPS were recorded along the nearly 400 km stretch of the Syr Darya and its tributary, the Arys, in southern Kazakhstan where a decline in the extent of irrigated land in the 1990s resulted in erosion of the abandoned fields (Baitulin, 2001). Similarly, agricultural lands in the lower reaches of the Zeravshan River and in the Samarkand region in Uzbekistan were also vulnerable to wind erosion, leading to the formation of an active cluster of DPS (Figures 3a and S5).
In the Karakum, DPS were associated with fluvial features, both modern and ancient (Figure 3). A large number of DPS occurred along the 750 km long dry channel of the ancient River Uzboy ( Figure S5a) which in the past connected Lake Sarygamysh with the Caspian Sea, but has been dry for at least 800 years (Velichko & Spasskaya, 2002). A cluster of multiple DPS, accounting for about 17% of all DPS in the Karakum-Kyzylkum region, was registered in the Uzboy and Atrek delta (Figure 3a). A high concentration of DPS occurred in the Sarygamysh depression (Figures 3a and S5b) particularly east of Lake Sarygamysh on a flat lowland composed by the alluvial deposits left by the migrating Amu Darya channel . It accounted for ∼15% of all DPS registered in the Karakum-Kyzylkum. The lake dried in the 18th Century following the natural migration of the Amu Darya course toward the Aral Sea (Létolle et al., 2007) but was restored in the 1960s to collect irrigation wastewater. The influx of water varies interannually and seasonally, causing strong fluctuations in the lake's level exposing both old and modern alluvial sediments (Kostianoy et al., 2013;Orlovsky et al., 2012).
DPS clusters were frequently associated with alluvial fans forming in river deltas. Two particularly active clusters, associated with these landforms, were identified in the Balkh delta in northern Afghanistan and in the deltas of the Tejen and Murghab Rivers in southern Turkmenistan (Figures 3a and S5e). The former region was characterized by the second highest identified DPS density, after the Aralkum and, together with the Aralkum and Sarygamysh regions, can be viewed as area (rather than points) source of dust.

Northwestern China
The Taklimakan desert in NWC produced the largest number of DPS in the region, accounting for 35% of all dust events registered across the entire study area (Figure 3b). These were located predominantly along the Tarim River and around Lake Lop Nur in the east. Prior to the construction of reservoirs and irrigation systems in the mid-20th Century, the Tarim emptied into Lake Lop Nur. Now the Lop Nur is a salt-encrusted lake bed (Middleton, 1986) producing a major cluster of DPS (Figures 3b and S6a). The eastern margin of the Taklimakan, including the Lop Nur, was the most active and persistent area source of dust where over 2,200 DPS (a half of all DPS in this region) were recorded ( Figure 3b) and activated most frequently ( Figure 4).
The Tarim's runoff is formed by snow and glacier melt in the Tien Shan and Kunlun and 60%-80% of the runoff occurs between June and August, carrying over 80% of the annual sediment load (Rittner et al., 2016;Ye et al., 2014). In the remaining months, high air temperature and falling river stage allow the alluvial fans and flood plains to dry out, becoming active sources of dust. The alluvial fans of the River Konqi and Rivers Aksu and Muzart account for 13% and 9% of the Taklimakan DPS, respectively (Figures 3b, S6b, and S6c). These fans support agricultural production, and dust events occur primarily at the margins of the agricultural land and at the abandoned or degraded farmland. Similarly, a cluster of DPS on the Yarkand River, west of the Aksu (Figure 4b Figure S1), there were few DPS in its central and southern sectors (Figure 3b). This discrepancy is addressed in Section 4.2.
A large number of DPS were recorded in the Hexi corridor and in the Qaidam Basin where FO-DOD values were also high. By contrast, in the Junggar region, merging into the Balkhash depression in the west, a relatively large number of DPS was detected overall but because of their low density and infrequent activation, FO-DOD values were low (Figures 3b and S1). Thus in the Hexi Corridor, DPS were activated 26-45 times per year and in the Balkhash-Junggar region, 10-18 times increasing to 32-34 in 2004-2006 (Figures 5 and6). In contrast to other sub-regions, Balkhash-Junggar is predominantly dry steppe characterized by higher threshold friction velocity than any other sub-region (Xi & Sokolik, 2015b).  (Figures 5c and 5i). A weak positive trend in annual precipitation and, probably, pasture abandonment and decline in livestock reverting overgrazing may explain these trends (Jiang et al., 2019).

Seasonal Cycles of DPS Activation
Seasonal statistics of DSA days are summarized in Figures 4 and 6. Similarly to FO-DOD ( Figure S1), a spring maximum characterizes NWC and the Balkhash-Junggar region. Dust storms are observed in the northern Taklimakan and Hexi Corridor almost throughout the year from February to October, starting and peaking earlier in the east (in February and March-April, respectively) and about a month later in the west. A secondary late summer-autumn peak is observed in the Balkhash-Junggar region in August-October when vegetation dies down and frequency of depressions increases (Figure 6c).
In CA, in all regions except the Aralkum, dust storms occurred in June-September but dust activity peaked in different months, starting earlier in the south, for example, in June-July in northern Afghanistan and southern Turkmenistan and in August-September elsewhere ( Figure 6). Dust activity weakened significantly between November and February although a positive trend in DSA days was observed in the Karakum-Kyzylkum region in winter. In northern Afghanistan, where dust storms were more frequent in summer, they were probably more intense between October and March when a larger number of DPS was activated per day (Figure 6g).
In the Aralkum, dust sources were activated throughout the year and, while the maximum frequency was observed in April-May, it remained high in June-October and in February-March (Figure 6e). The average number of DPS per dusty day, however, was highest during the colder season from November to May (fivesix DPS per day) indicating more intense dust storms. Between June and October, 3-4 DPS per day were activated. This finding is in contrast to previous studies, suggesting a summer peak in dust activities in this region (Groll et al., 2013.;Orlovsky et al., 2005;Xi & Sokolik, 2015b).
the exposed lake bed. In spring, in the absence of vegetation, lower threshold wind speeds are required to dislodge dust particles (Knippertz, 2014) while in summer, salt crusts prevent dust entrainment (Argaman et al., 2006).

Discussion
The deserts of CA and NWC are among the largest source regions of mineral dust in the world by area and by dust emission flux supplying up to 25% of desert aerosol to the atmosphere (Ginoux et al., 2012;Shao & Dong, 2006). Regional sources of dust are located in proximity to the major glacierized regions of High Asia and play an important role in the snow/ice-albedo feedback ( (e) (f) potentially affecting snow melt in the mountains and provision of water to the arid lowlands. The development of this unique 10-years data set of DPS, which retains the 1-km resolution of satellite data, for the first time enabled a detailed assessment of dust sources in this region, using consistent observations at an appropriate spatial resolution. This data set expands dust storm and dust deposition climatologies, derived from regional historical weather station records. While these climatologies span several decades, they are sparse and many were discontinued in the 1990s in CA (Groll et al., 2013;Indoitu et al., 2012;Issanova et al., 2015b) although observations continue in the Taklimakan (Wang et al., 2006;. It provides data for regions such as northern Afghanistan, where there were no previous observations. The high-resolution mapping of dust sources improves characterization of spatiotemporal variability in dust emissions derived from the AOD data by Ginoux et al. (2012) and Xi and Sokolik (2015a) by localizing dust sources within the broad dust hotspot areas defined by AOD and thus contributing to improvements in numerical dust and weather forecast models in the region (Li & Sokolik, 2017;Walker et al., 2009).

Uncertainties and Limitations of the DPS Data Set
There are several sources of uncertainty associated with this DPS data set: (i) availability of only two MODIS images per day limiting observations to a period between 10:30 and 13:30 local time; (ii) obscuration of DPS either by clouds or by pre-existing dust plumes generated upwind; (iii) limitations of the DEP algorithm; (iv) potential operator's errors of manual DPS detection.
Availability of two fixed-time acquisitions per day is a significant and recognized limitation of using MODIS data for mapping of dust sources (Miller, 2003) particularly in comparison with the application of geostationary SEVIRI data (Hennen et al., 2019;Schepanski et al., 2007). Dust mobilization is forced by meteorological processes varying in time or following diurnal cycles. In other dust source regions, for example, in the Sahara (Schepanski et al., 2007;2009) and the Middle East (Hennen et al., 2019), dust emissions occur frequently after sunrise and breakdown of nocturnal inversions. Thus in the Middle East, dust emissions are most frequent between 6:00 and 12:00 with considerably fewer activation in the afternoon and at night. Hennen et al. (2019) assessed bias in the detection of dust emissions in the Middle East by degrading temporal resolution of SEVIRI data from 15 to 60 min showing reduction in the number of observed events particularly in summer when it is caused by the mesoscale rather than synoptic-scale processes. Using coarser temporal resolution, however, did not significantly affect the derived geographical distributions of sources, that is fewer point sources were detected in the same regions and smaller dust events were missed.
There are no data on timing of DSA in CA except the data produced by this study. However, diurnal cycles of the occurrence of dust storms, as registered by regional meteorological stations, show that their frequency peaks between 9:00 and 15:00 which is later in the day in comparison with the Middle East. Minimum is observed between 18:00 and 6:00 (Indoitu et al., 2009;. We used diurnal cycles of surface wind speed, derived from ERA5 reanalysis for the study period and averaged over the subregions and meteorological seasons, as a proxy for timing of DSA ( Figure S7). Across most of the study region, maximum wind speed is observed between 10:00 and 14:00 with exception of the northern and eastern sectors of the Taklimakan where wind speed peaks at 10 a.m. in summer and spring. The timing of MODIS overpasses, therefore, coincides with the diurnal maxima in wind speed ( Figure S7) and dust storms occurrence (Indoitu et al., 2009; partly alleviating the problem although DSA in the afternoon may be underestimated more significantly. In the absence of geostationary data, it is impossible to assess the proportion of underestimated sources. Following Hennen et al. (2019), we suggest that statistics of DPS activation per day is underestimated but it is unlikely that spatial clustering of DPS, linked to specific landforms (Table S3), is affected to a significant extent, that is data with higher temporal resolution are expected to show more frequent DPS activations, especially with regard to smaller sources, in the same locations.
It is not possible to detect dust beneath the meteorological clouds using this dust enhancement technique (Miller, 2003) although it is more successful in discriminating between dust and clouds than other techniques (Baddock et al., 2009). The obscuration of DPS by clouds affected the northern part of the study area to a greater extent, as low-pressure systems are more frequent there, particularly in March-April and October-November. The obscuration of DPS by dust clouds affected their detection particularly in the Taklimakan, Kyzylkum, Karakum and in northern Afghanistan. Thus in the Taklimakan, where dust plumes, forming in the north and east, circulate within the closed basin for several days, the highest quality (QF1) DPS observations account for 7%-20% of the detected DPS between March and August and annually, between 11% (2003) and 26% (2004 and 2009) (Figures 5a and 6a). In the Karakum and Kyzylkum, less than 4% of observations accounted for QF1 DPS in 2003 and2007 (Figure 5f). Here, higher quality observations were made in May-June (13%-22% QF1 DPS) and lower quality in July-August (5%-9% QF1 DPS) (Figure 6f). By contrast, in the north, where DPS density is lower, QF1 DPS accounted for 50%-90% of all observations (Figures 5, 6c, and 6d).
A key problem limiting the effectiveness of the DEP technique is that defining a universal threshold distinguishing between dust and nondust proved highly problematic. The DE thresholds varied between sub-regions and seasons and, in some instances, individual dust events (Figure 2). Resampling of pixel histograms to adjust the dust/nondust thresholds helped to achieve effective dust enhancement and visualization in this and other studies (Baddock et al., 2009) but made the development of the DPS data set extremely time-consuming. Varying thresholds make the development of any automated DEP algorithm problematic and compromises between quality and efficiency will be required.

Dust Point Sources versus Dust Optical Depth
Comparison of DPS locations with FO-DOD above the 0.2 threshold and with previously published AOD climatologies (Li & Sokolik, 2017;Xi & Sokolik, 2015a), showed that in many regions, there was a close match between high FO-DOD values and DPS count despite a high proportion of QF3 observations. Thus in the Karakum-Kyzylkum region, the QF1:QF2 and QF2:QF3 ratios were 0.25 and 0.83, respectively (Figures 5f and 6f) but the dust enhancement method successfully recognized DPS in the areas with the highest FO-DOD, for example, in the Syr Darya valley, over the southeastern Caspian shores and in the lower reaches of the Zeravshan (Figure 3a). DPS were successfully detected in the northern Taklimakan where QF1:QF2 and QF2:QF3 ratios were 0.58 and 0.60, respectively, and also in the eastern Taklimakan (Figure 3b), northern Afghanistan, and in those regions of the pre-Aral where atmospheric dust load was high (Figure 3a). While obscuration by dust clouds inevitably reduced accuracy of DPS mapping, these examples show that the dust enhancement technique discerns DPS in regions with high AOD. However, there was a notable lack of agreement in several regions. In the Balkhash-Junggar region (Figures 3b and S1), the discrepancy between DPS and FO-DOD can be attributed to the low density ( Figure 7) and infrequent activation (Figures 5c and  6c) of sources, which did not enable build-up of dust in the atmosphere. Here, the number of DPS (dusty days) declined from 80-115 (20-32) in 200480-115 (20-32) in -200880-115 (20-32) in to 16-50 (10-16) in 200980-115 (20-32) in -2012). However, more often the observed discrepancy was due to the displacement of dust in the atmosphere.
A relatively low number of DPS and high FO-DOD values were identified in the central and southern Taklimakan ( Figure 3b). Previous field studies showed that the sands of the central Taklimakan have low (below 4%) content of dust-sized particles and while dust sources were reported, field studies did not identify it as a major dust-producing region (Wang et al., 2006;. The Dust Enhancement method successfully identified a relatively small number of aeolian DPS in the central and eastern Taklamakan (Figure 3b; Table S3) confirming these conclusions. In the southern Taklimakan, it successfully identified clusters of DPS associated with fluvial sources and alluvial piedmont fans (Figure 3b) where high frequency of dust storms was previously reported (Wang et al., 2006). However, fluvial, alluvial and lacustrine sources are less common in the south than in the north. By contrast, gravel cover and stone pavements are more widespread reducing the availability of fine material (Wang et al., 2008). We suggest that while the identified sources do generate dust locally, atmospheric dust observed over the southern Taklimakan is predominantly transported from the northern and eastern sectors of the desert and banked against the Kunlun Mountains. On 73% and 62% of all days, when dust sources were activated in the Taklimakan in spring and summer, respectively, north-easterly, north-westerly and northerly winds dominated over its southern sector indicating dust transport from the Tarim and Lop Nur regions. Modeling confirms this conclusion (Uno et al., 2005). The Ustyurt plateau and Mangyshlak peninsula, previously identified by Xi and Sokolik (2015a) as one of the main dust hotspots in CA on the basis of high AOD and low NDVI values, were another region with high FO-DOD and very few DPS (Figures 3a and S1). The Ustyurt is predominantly a gravelly gypsiferous desert where coarse gravel, either loose or sparsely covered by Anabasis salsa or Artemisia, dominates land cover (Lioubimtseva, 2002;Viktorov, 1971). There is a notable absence of landforms associated with accumulation of dust-size material and its availability is low across the Ustyurt. Exceptions are a series of depressions containing salt pans or solonchaks on the border with the pre-Aral region and sands in the north and west of the plateau (Viktorov, 1971) where the dust enhancement method successfully identified DPS despite the high FO-DOD values (Figures 3a and S1). Modeling of dust entrainment by Li and Sokolik (2017) showed low dust emission from the Ustyurt and Mangyshlak. These results were interpreted by the authors as erroneous because they contradicted the observed high AOD but they are in agreement with the absence of DPS in this region. Analysis of the ERA5 wind fields showed that, easterly or north-easterly flow was frequent over the Aralkum throughout the year, accounting for 45% and 34% of the DSA days in spring and summer, respectively, and over the Sarygamysh region in spring, transporting dust toward the Caspian and the Ustyurt.

Comparison of the Dust Source Regions
In all, just under 13,500 DPS were detected in the study region between 2003 and 2012. The largest numbers of sources, accounting for 35% of all DPS, were registered in the Taklimakan followed by the pre-Aral region (16%) and the Aralkum (15%) (Figure 7). While the Taklimakan and the pre-Aral are vast areas occupying about 0.95 × 10 6 km 2 and 1.29 × 10 6 km 2 , respectively, the Aralkum is a smaller desert characterized by the highest average density of 12 DPS per 1,000 km 2 . The density of DPS in the Taklimakan was comparatively low (5 DPS per 1,000 km 2 ) because most DPS were concentrated along the Tarim and its tributaries and Lake Lop Nur while the central and southern Taklimakan were considerably less active (Figure 3b). An even lower density characterized the Karakum and Kyzylkum deserts where DPS were also associated with fluvial landforms and concentrated along the modern or palaeo river channels. A very high density of DPS characterized the Balkh delta (upper Amu Darya), where 8% of all DPS developed over an area which is 10 times smaller than the Taklimakan. Although meteorological aspects of DPS activation are outside the scope of this paper, we note that the threshold wind speed for dust entrainment was the lowest in the Balkh delta, with about 80% of all dust storms initiated by wind speeds of 5 m s −1 or less. It is also in these areas-northern and eastern Taklimakan, Aralkum, and Balkh delta-that the highest frequency of dusty days was observed (Figure 4a).
Many studies have investigated links between land surface characteristics and dust storms (e.g., Bullard et al., 2011Bullard et al., , 2008Xi & Sokolik, 2015b). We attempted to establish links between DPS activity and landforms using the geomorphological classification of dust-emitting surfaces developed by Bullard et al. (2011) and historical archives of high-resolution satellite imagery. A lack of detailed information about land surface properties in CA, going beyond a broad Land-Cover and Land-Use Change (LCLUC) scheme (e.g., Xi & Sokolik, 2015a;2015b) did not allow us to conduct an in-depth analysis within the framework of this study.
In most cases, there was a clear link between land surface characteristics and development of DPS. A notable example is the formation of DPS on fire scars ( Figure S4) which were the predominant sources of dust in the north, from western Kazakhstan to the Junggar basin in China (Table S3). However, attribution was more difficult in the areas without a predominant surface type, for example, at the margins of alluvial fans where cultivated farmland, fully or partially abandoned agricultural fields and natural alluvial plains exist in proximity. Under 2% of all DPS formed in such areas.
The largest number of DPS-40% of all registered in the study area-were associated with fluvial systems and alluvial deposits south of the Aral Sea and in the Taklimakan (Table S3). Dry lakes accounted for the second highest share at 29.7% of all DPS. This was typical not only of the Aralkum but also of the Taklimakan (Table S3) where DPS concentrated near the desiccated Lake Lop Nur. While dry lakes are prominent sources of dust in many regions (Bullard et al., 2011;Ginoux et al, 2012;Hennen et al., 2019;Schepanski et al., 2009;Shao et al., 2011), in CA and NWC the contemporary lake desiccation is particularly widespread (Liu et al., 2019) due to fluctuations of lake levels on a decadal timescale caused by tectonic activity (Koronkevich, 2002), excessive water abstraction for irrigation and poor water management, and the observed climatic warming (Bai et al., 2011;Liu et al., 2019;Micklin, 2007;Middleton, 2002;Yapiyev et al., 2017). Drying lakes were identified as major sources of dust in the wider area including the Sistan basin in Afghanistan and Iran (Rashki & Kaskaoutis, 2019) and Lake Urmia basin in Iran (Ghomashi & Khalesifard, 2019). Many lakes-from Iran and Afghanistan to Kazakhstan-are saline, and the contribution of drying lakes to the formation of dust and salt storms, damaging to crops and natural ecosystems (Ge et al., 2019;Goudie et al., 2016), requires further investigation.

Conclusions
This study presented the first data set of dust sources in CA and NWC derived from MODIS imagery, at a spatial resolution of ∼1 km which made detection of individual dust emissions possible. We compiled seasonal 1° × 1° gridded data sets of frequency of DSA days. These data helped to improve the localization of dust sources in comparison with AOD data, highlighting the eastern and northern Taklimakan, the Aralkum, northern Afghanistan and Amu Darya, Syr Darya and Uzboy valleys (but not the wider Karakum and Kyzylkum) as the most persistent sources of dust. The region between the Aral and Caspian Seas, as well as the central and southern Taklimakan, were shown to have few dust sources despite the high AOD which forms due to dust transport from the upwind sources. It is envisaged that the new data set will complement the existing meteorological station and AOD data, help to improve regional-scale dust emission and transport models and to resolve discrepancies between modeling and observations of dust in the atmosphere. We attempted a preliminary-and limited-analysis of links between dust storms and landforms in the study area. This analysis should be expanded, focusing on the role of the wildfire in the north, changing lake levels across the region, water use for irrigation and agricultural practices in the south. However, this will require detailed land surface data, which is currently limited in CA.