4.3.1. Location of the Most Frequent Dust Emission Sources
 Direct observations of dust storms reported by meteorological stations on synoptic reports have been widely used to establish the climatology of dust emissions in arid and semiarid regions. For example, Goudie  used dust storm observations to investigate their distribution in space and time. In northeastern Asia, he identified the Taklimakan desert as the most frequent source of dust storms, with 33 dust storms per year, and the northern deserts of China as the second source with a maximum of 19 dust storms in the Badain Jaran desert. On the basis of the analysis of 40 years (1960–1999) of synoptic observations from 174 stations located in China and Mongolia, Sun et al.  identified the region with the higher dust storm frequency as the region including the southern Gobi desert of Mongolia and China, the Badain Jaran, Tengger, Ulan Buh, Qubqi and Mu Us deserts, and a secondary maximum in the Taklimakan desert. On the basis of a similar time series (1960–2001) of measurements performed in Chinese meteorological stations, Sun et al.  confirmed the results from Goudie , indicating that the northwestern China, including the Taklimakan, Kumutaq, and Gurban Tonggut, is the region of China where dust storms are the most frequent, and that a secondary region with frequent dust storms is located in central and western part of Inner Mongolia. However, Sun et al. noted that the increase in the number of dust storms, especially observed in spring, during the last years mainly concerns the central and western part of Inner Mongolia. The difference between these two studies can be explained by two factors: (1) Sun et al.  focused on the period from March to May, while Sun et al.  mentioned that the dust storms frequencies in the Xianjiang Uygur Province, where the Taklimakan is located, are maximum from April to June; (2) both studies are based on interpolation from very sparsely distributed observations, but the study from Sun et al.  includes observations in Mongolia, while the study from Sun et al.  is restricted to Chinese stations. The inclusion of the Mongolian stations where frequent dust storms (>30 d yr−1) are observed [Natsagdorj et al., 2003] gives more importance to the area of the Gobi and northern deserts. Finally, we can conclude that the two regions where the dust storms are the most frequent are the Taklimakan desert and the Gobi and northern deserts with a comparable importance.
 As illustrated in Figure 8, this conclusion is in agreement with the simulated dust emission frequencies which are also the highest in the Taklimakan and the northeastern deserts. The maximum simulated dust emission frequencies for the northern desert are located in the Badain Jaran desert, in agreement with the analyses of the dust storm frequencies in Inner Mongolia [Gao et al., 2003]. However, the extent of the simulated source areas in the northern deserts is not comparable to the one deduced from the interpolated dust storms frequencies including the Mongolian stations. The simulations seem to underestimate the dust emission frequencies in the Gobi desert of Mongolia. This underestimation could be partly explained by the fact that the roughness map is incomplete (50%) over this region but it is more probably related to a bias in the surface wind fields. Indeed the predicted erosion thresholds (8–20 m s−1) are comparable to the wind velocities observed during dust storms in the meteorological stations (11–20 m s−1), but the surface wind velocity almost never exceeds such values in this area. Husar et al.  identified two severe dust storms on 15 and 19 April 1998, originating from the Gobi and northern deserts in which surface wind velocities were as high as 20 m s−1. For the less severe dust storm of the two (15 April 1998), our simulations indicate that the erosion thresholds are exceeded over the whole northern deserts where the surface wind velocities are higher than 15 m s−1, but only in a few pixels of the Gobi desert. For the second severe dust storm, the surface wind velocities do not exceed 15 m s−1. They are higher than the erosion thresholds only for a small area of the Gobi desert. It indicates that the ECMWF surface wind velocities are too low to simulate accurately some of the severe dust storms observed in the region of the Gobi and northern deserts.
 The simulated dust event frequencies can also be compared to the long-term (1960–2002) simulations of the dust emission from Asian sources performed by Zhang et al.  using the dust emission model described by Gong et al . From Zhang et al.'s simulation, ∼70% of the Asian dust emissions are produced in the Taklimakan desert, the region including the Badain Jaran, Tengger and Ulan Buh deserts and the Mongolia, with average respective contributions of 21%, 22%, and 29% for the period 1960–2002. However, compared to these averaged contributions, the dust emissions simulated by Zhang et al. for the period from 1995 to 1999 are significantly lower for the Mongolia (∼ −40%) and for the region including the Badain Jaran desert (∼ −15%), while a very slight increase of the emissions is simulated in the Taklimakan desert. Our results are thus in reasonable agreement with this study for the same period of simulation.
4.3.2. Seasonal and Interannual Patterns
 The seasonal cycle derived from synoptic observations is clearly characterized by a maximum in spring with some minor differences depending on the region [Sun et al., 2001; Gao et al., 2003; Natsagdorj et al., 2003; Sun et al., 2003; Wang et al., 2003]. For example, a secondary maximum in the dust storm frequencies can be observed during winter (December and January) in the northern deserts of China or during the fall in some stations of Mongolia or Inner Mongolia [Gao et al., 2003; Natsagdorj et al., 2003; Sun et al., 2003]. The period of maximum dust storm frequencies in the Taklimakan extends from spring to early summer [Sun et al., 2003; Wang et al., 2003]. These two features are well reproduced in our simulations as illustrated in Figure 13.
Figure 13. Annual distribution of the relative dust event frequency simulated for 1997–1999 over the Taklimakan (black) and the region including the Gobi desert and the northern deserts of China (white).
Download figure to PowerPoint
 Finally, the simulated spatial and temporal variations agree with the general features of the dust storms frequencies observed in China and Mongolia. However, very few meteorological stations are located inside the deserts areas, where the maximum dust emission frequencies are simulated limiting the level of confidence of such comparisons. This does not allow any extensive and quantitative comparison. Satellite observations make a quantitative comparison with comparable spatial and temporal scales possible.
4.3.3. Comparison With TOMS Observations
 The Aerosol Index (AI) derived from the Earth Probes Total Ozone Mapping Spectrometer (TOMS) is the only available satellite aerosol products over the studied area. A detailed description of the TOMS AI is given by Herman et al.  and Torres et al. . The AI relies on the spectral attenuation of the Rayleigh scattering due to aerosol absorption. Positive values generally correspond to UV-absorbing aerosol (e.g., desert dust and carbonaceous particles) while negative values correspond to nonabsorbing aerosols (e.g., sulfate aerosols). The Absorbing Aerosol Index (AAI), defined as the positive values of the AI, has been used to investigate the distribution of the dust sources over the world [Prospero et al., 2002; Washington et al., 2003]. Washington et al.  restricted their analysis to positive residues of values greater than 0.7, to avoid contamination by noise resulting from surface signal or nonabsorbing aerosols. However, the TOMS AAI is only a semiquantitative indicator of the aerosol atmospheric content. Indeed, it can be influenced by the cloud cover and it is very sensitive to the aerosols layer altitude [Torres et al., 1998; Chiapello et al., 1999; Hsu et al., 1999]. This sensitivity to the aerosol layer altitude makes the discrimination between locally emitted dust and dust transported from remote sources difficult [Mahowald and Dufresne, 2004]. However, it is reasonable to assume that the atmospheric load (and thus the TOMS AAI) will be higher close to the region where dust emission occurs, in particular at the TOMS spatial resolution (1° latitude × 1.25° longitude). As a result, we rather expect a relative agreement on the trends of temporal variations at the seasonal and interannual timescales than a quantitative one between the simulated and observed dust emission frequencies.
 Daily global data of the Earth Probes TOMS are available in the TOMS website (http://toms.gsfc.nasa.gov/ftpdata.html). From the daily data, following Washington et al. , we computed the monthly dust occurrence frequency as the number of TOMS AAI higher than 0.7 related to the total number of observations. Daily observations are missing in November 1997, December 1997, December 1998, and January 1999. For these months, we computed the simulated dust emission frequencies only for the days of available observations.
 Figure 14 reports the simulated frequencies of significant dust emissions (Figure 14a) and the frequencies of TOMS AAI > 0.7 (Figure 14b). Both simulations and observations show higher dust emission frequencies in the Taklimakan than in the northern deserts. Moreover, the highest simulated dust emission frequencies (40°N, 87°E) and the highest frequencies of TOMS AAI > 0.7 (40°N, 84°E) are both located in the Taklimakan desert. However, near the south of Beijing eastern coast high frequency of TOMS AAI > 0.7 are observed whereas no dust emissions are simulated over this region, which is not known as a desert or desertified area. Since TOMS AAI is sensitive to UV absorbing aerosols, such high frequencies can be attributed to other aerosols types, and presumably to carbonaceous aerosols. Indeed, Herman et al.  indicated that from December to April, aerosols resulting from coal burning activities in northern China may be mixed with mineral dust and transported eastward over the Pacific Ocean and the southern portions of Japan. To avoid such intricate situations where there is a mixing with other absorbing aerosols, we focused the comparison on the Taklimakan desert (36°N–42°N; 77.5°E–90°E), where mineral dust is expected to be the dominant absorbing aerosol type contributing to the TOMS AAI.
Figure 14. (a) Map of the annual frequencies of significant dust emissions (dust flux >10−10 g cm−2 s−1) and (b) map of the annual frequencies of TOMS AAI > 0.7 averaged on 3 years (1997–1999).
Download figure to PowerPoint
 To obtain a more quantitative comparison between the simulations and the frequencies derived from TOMS AAI, we aggregated our results at the TOMS AAI spatial resolution (1° latitude × 1.25° longitude), i.e., 30 simulated pixels. If one pixel out of the 30 exceeds the erosion threshold, the corresponding TOMS resolution pixel is considered as dusty. TOMS AAI are produced from daily instantaneous observations at ∼1120 LT (i.e., ∼0320 UT), while the simulated dust emission frequencies were initially estimated with a 6-hour time step. We thus considered that if a pixel exceeds the erosion threshold once during the 24h previous to the TOMS observation, the corresponding TOMS resolution pixel is considered as dusty.
 Figures 15 represents the distributions of monthly simulated frequencies of significant dust emissions (Figure 15a) and of TOMS AAI (AAI > 0.7) (Figure 15b), averaged in the Taklimakan for 1997, 1998, and 1999. The maxima are globally simulated and observed during the same period, i.e., in late spring from April to May. Both the TOMS AAI and the simulations show a more pronounced seasonal cycle in 1998 and 1999 than in 1997; that is, the frequencies are higher in spring (March, April, and May).
Figure 15. Monthly frequency (a) of significant simulated dust emissions (dust flux >10−10 g cm−2 s−1) and (b) of TOMS AAI > 0.7 over the Taklimakan desert for 1997, 1998, and 1999.
Download figure to PowerPoint
 There is a general agreement between the simulated and observed frequencies, except during some months in autumn: for example, in September 1997 the simulated frequencies are proportionally lower than the TOMS AAI frequencies, whereas in September 1998 and September 1999 the simulated frequencies are higher than the TOMS AAI ones.
 Figure 16 presents the monthly frequencies derived from our simulations of significant dust emissions averaged over the Taklimakan region as a function of the monthly averaged frequencies of TOMS AAI > 0.7 for the 3 studied years. The two data sets are significantly correlated (r between 0.94 and 0.95) for the 3 tested years, with no significant difference in the slope as a function of the year.
Figure 16. Monthly frequencies of significant simulated dust emissions (dust flux >10−10 g cm−2 s−1) as a function of the monthly frequencies of TOMS AAI > 0.7 averaged on the Taklimakan desert for 1997, 1998, and 1999.
Download figure to PowerPoint
 We then examined the correlation between the monthly frequencies of simulated significant dust emissions and the monthly frequencies of TOMS AAI > 0.7 for each pixel (1° latitude × 1.25° longitude) and for the 3 years (Figure 17). The range of frequencies derived from TOMS AAI (from 0 to 97%) has been divided in classes 5% wide. For each TOMS AAI frequency class, the mean simulated dust emission frequency and the associated standard deviation have been computed. This comparison involved 2160 cases, with a number of tested cases in the different classes varying from 20 to 124, except for the lowest-frequency class (1065 cases), and for the highest one (6 cases), the latter which is not statistically representative. Figure 17 presents the results of this computation for 1997, 1998, and 1999 in the Taklimakan area. We obtain a significant correlation (r = 0.98, slope = 0.44), with similar slope and origin than for the individual pixels. Similar correlations and slopes are obtained on an annual database (1997: r = 0.93, slope = 0.38; 1998: r = 0.98, slope = 0.47; 1999: r = 0.92, slope = 0.38). The simulated dust emission frequencies are systematically lower than those derived from TOMS AAI > 0.7 by a factor from 2 to 2.5. Such a difference may be explained by the fact that both local dust emissions and intense dust plume transported from neighboring pixels can produce high TOMS AAI.
Figure 17. Monthly frequencies of significant simulated dust emissions (dust flux >10−10 g cm−2 s−1) as a function of the monthly frequencies of TOMS AAI > 0.7 over the Taklimakan desert for the 3 years 1997, 1998, and 1999. Small dots represent individual data; circles represent the averaged frequency of simulated dust emissions for classes (5% width) of frequency of TOMS AAI > 0.7; the solid line represent the linear fit of the averaged data (without accounting for the last class which is not representative).
Download figure to PowerPoint