An intercomparison study involving eight dust emission/transport models over Asia (DMIP) has been completed. Participating dust models utilize a variety of dust emission schemes, horizontal and vertical resolutions, numerical methods, and different meteorological models. Two huge dust episodes occurred in spring 2002 and were used for the DMIP study. Meteorological parameters, dust emission flux and dust concentration (diameter < 20 μm) are compared within the same domain on the basis of PM and NIES lidar measurements. We found that modeled dust concentrations between the 25% and 75% percentiles generally agreed with the PM observations. The model results correctly captured the major dust onset and cessation timing at each observation site. However, the maximum concentration of each model was 2–4 times different. Dust emission fluxes from the Taklimakan Desert and Mongolia differ immensely among the models, indicating that the dust source allocation scheme over these regions differs greatly among the various modeling groups. This suggests the measurements of dust flux and accurate updated land use information are important to improve the models over these regions. The dust vertical concentration profile at Beijing, China, and Nagasaki, Japan, has a large scatter (more than two times different) among the models. For Beijing, the scaled dust profile has a quite similar vertical profile and shows relatively good agreement with the lidar extinction profile. However, for Nagasaki, the scaled dust profiles do not agree. These results indicate that modeling of dust transport and removal processes between China and Japan is another important issue in improving dust modeling.
 Dust emission and transport modeling plays an important role in understanding the recent increase of Asian dust episodes. From the late 1990s, dust models have been developed at universities, research institutes, and government agencies, and then applied to the Asian region [e.g., Wang et al., 2000; Park and In, 2003; Gong et al., 2003; Liu et al., 2003; Shao et al., 2003; Uno et al., 2001, 2003]. These models have reproduced many important observational facts and retrieved valuable information to elucidate characteristics of Asian dust phenomena. Recently, the Asia Pacific Regional Aerosol Characterization Experiment (ACE-Asia) [Huebert et al., 2003] provided an excellent opportunity to examine a large-scale Asian dust episode using numerical dust transport models [e.g., Gong et al., 2003; Liu et al., 2003; Chin et al., 2003; Uno et al., 2004]. At least three regional dust models were applied for the ACE-Asia periods and reported reasonable simulated results. However, their estimated dust emission amounts ranged widely from 105 Tg/(2 months), 250 Tg/(3 months) to 640 Tg/(16 days). Similar scatter in the dust emission flux results is also evident in the modeling results from Park and In  and Shao et al. . Moreover, dust transport patterns from the emission source region are usually very similar, but the predicted surface level concentration sometimes shows a difference of more than two orders of magnitude. Such wide scattering of the estimated dust emission and concentration reflects differences in dust emission schemes, surface boundary data (such as soil texture, soil wetness, land use data including a recent desertification information), and atmospheric models (meteorological and transport models). Consequently, more detailed studies are required to reduce uncertainties related to Asian dust emission, transport and removal. In addition, a common understanding of the performance and uncertainty of dust erosion and transport models in the Asian region has become essential for developing dust storm forecast capacities in the region. At the 2nd Aeolian Dust Experiment on Climate Impact (ADEC [Mikami et al., 2002, 2006]) Workshop in Xi'an, China, January 2003, an urgent need for an intercomparison of recent dust models was discussed and an agreement was reached to pursue this activity. On the basis of this agreement, the dust model intercomparison (DMIP) activity over the Asia region was initiated as an activity of the ADEC project. This paper presents an outline of this DMIP project, an overview of its major findings, and its conclusions.
2. Framework of DMIP
 The DMIP project was designed to gain an understanding of dust model characteristics by intercomparing independent simulations made by each group. For the model intercomparison period, two major dust episodes were selected: period A, 15–25 March 2002 (ten days), and period B, 4–14 April 2002 (ten days) because of the very dense dust onsets observed over a wide area including China, Korea, and Japan [e.g., Chun and Lim, 2004; Shao et al., 2003; Sugimoto et al., 2003]. As shown later in section 4.2, the maximum TSP concentration in period A exceeded 12 mg/m3 at Beijing, and 2 mg/m3 at Seoul.
 The DMIP project accepted all original dust model results without specifying either model resolution or land use (desertification) information (i.e., each group can use their own land use category information for possible dust emission sources). Furthermore, we specified no meteorological model field (or driver), therefore allowing participants to use their own meteorological fields to execute dust simulation. Each group was compelled to use its own dust model and submitted a description of the model and dust concentration fields. Dust transport models were required to include the Gobi Desert and Mongolia. Preferably, they also included the Taklimakan Desert as a potential dust source region. The eastern boundary of all model domains was required to cover all areas of Japan. These relatively lax specifications were utilized in order to accommodate as many participants as possible to this project, which, however, necessitated a very careful scheme to perform intercomparison of model results.
 Nine dust modeling groups employed eight models in the DMIP project. Details of each dust model are summarized in Table 1. There were six regional dust models (horizontal resolutions of 27–80 km) and two global models (1° and 1.125°). In the vertical dimension, the resolution ranged from 20 to 46 levels. High-resolution models (horizontal grid size < 50 km) include the Coupled Ocean/Atmospheric Mesoscale Prediction System (COAMPS [Liu et al., 2003]), the Asian Dust Aerosol Model (ADAM [Park and In, 2003; In and Park, 2003]), the Dust Regional Atmospheric Model (DREAM [Nickovic et al., 2001]), the Northern Aerosol Regional Climate Model (NARCM [Gong et al., 2003]), and the Computational Environmental Modeling System version 5 (CEMSYS5 [Shao et al., 2002]). Low-resolution model groups (horizontal grid size > 50 km) are the Chemical weather Forecasting model System (CFORS [Uno et al., 2003]), the Navy Aerosol Analysis and Prediction System (NAAPS [Christensen, 1997; D. L. Westphal et al., Description of NAAPS: The Navy Aerosol Analysis and Prediction System, manuscript in preparation, 2006]), and Model of Aerosol Species in the Global Atmosphere (MASINGAR [Tanaka and Chiba, 2005]). A group from the Chinese Meteorological Agency (CMA) also submitted their results. Because their scheme and modeling concepts closely resemble those of the CEMSYS5 model, only the results from the CEMSYS5 model were shown in this paper.
Table 1. Description of DMIP Model Group
Naval Research Laboratory, Monterey, USA
Korean METRI, Korea
Hong Kong City University, China
Kyushu University, Japan
Naval Research Laboratory, Monterey, USA
N/A, not applicable.
It does not include the Taklimakan Desert.
Considers only those particles with diameters less than 10 microns.
Meteorological database or global model name
COAMPS (with NOGAPS for boundary and initial)
RDAPS (Regional Data-Assimilation and Prediction System/KMA)
RCM (Canadian Regional Climate Model with NCEP data)
NCEP/Eta (with NCEP data)
T213 forecast or NCEP reanalysis and CEMSYS4
RAMS (with ECMWF)
NOGAPS (global model)
MRI/JMA 98 GCM (with JMA GANAL)
Model domain (horizontal and vertical information)
Lambert projection (center 42°N, 106°E); horizontal 27km (with 301 × 221 grids), vertically 46 layers (up to 32 km)
Lambert conformal projection (center 38°N, 128°E); horizontal 30 km (with 191 × 171 grids), vertically 25 layer (up to 18 km)
north polar stereographic projection (center 35.6°N, 106.6°E) with a horizontal resolution of 45 km at 60°N (160 × 110 grids), vertically 22 layers (up to 30 km)
rotated lat-lon (center 39°N, 112.5°E) horizontal 0.5° (with 67 × 81 grids), vertically 23 layers (up to 16 km)
Lambert projection: horizontal 50 km, vertically 25 layers
rotated polar-stereographics (center 25°N, 115°E) horizontal 80 km (with 100 × 90 grids), vertically 23 layer (up to 23 km)
latitude-longitude grid; 1° horizontal resolution; 24 levels between surface and 100 mb
global; horizontal about 125 km (with 320 × 160 grids); vertically 20 layers (up to 45 hPa)
Thickness of first layer, m
Dust emission: threshold velocity or friction velocity
 Most of regional models are using 1km resolution land use data set (shown in Table 1), and they are using a mosaic type subgrid treatment for dust emission. This is done by sorting land surface according to soil and vegetation data, and wind erosion model is then run for the land surface subgrid. The dust emission rate is then an area-weighted sum of the dust emission rate from the subcells. For global models such as MASINGAR, their grid size coincides with the resolution of land and soil texture information; no mosaic type approach is used (NAAPS is using the similar mosaic type treatment based on the USGS land use type).
 Except for the ADAM model, all model domains include the Taklimakan Desert area. Dust particle size ranges from 0.1 to 76 μm represented by a sectional approach of 1 to 12 bins. All dust emission schemes in DMIP are based on surface friction velocity (u*), and most dust emission flux is calculated by a function of the third or fourth power of u*. Three models (CEMSYS5, NARCM and MASINGAR) used the concept of saltation bombardment. One important point is that all dust models depend on regional or global meteorology models (e.g., NOGAPS, RDAPS, COAMPS, RCM, Eta, RAMS, and other AGCM, Table 1), which are based on objective analytical results from FNMOC, ECMWF, NCEP, JMA, KMA, etc. It is noteworthy that this DMIP project consists of various meteorological models, dust emission schemes, and land use information determined by each model development group and no adjustment is made to the submitted results from each model.
Table 2 shows the required DMIP model outputs. Because of the differences in model horizontal resolution and map projection, the model data are interpolated to a 1° longitude-latitude grid, for the low-resolution group, and 0.5° for high-resolution group, at 3-hour intervals. Two examination regions were chosen for the model comparison (Figure 1): Region A covered the domain of 75–125°E and 35–50°N for analyses of detailed dust emission processes and concentration and region B covered the domain of 75–150°E and 20–55°N for analyses of transport processes and concentration.
Table 2. Required Model Output for Dust Model Intercomparison
dust emission flux (d < 20 μm)
dust concentration at first model level (d < 20 μm)
dust concentration at 700 hPa (d < 20 μm)
dust column loading (height z < 10 km) (d < 20 μm)
wind speed at 10 m level
surface friction velocity
threshold surface friction for dust lift up
dust dry deposition (d < 20 μm)
dust wet deposition (d < 20 μm)
 We required that the initial dust concentration be set to zero and allow no dust inflow from the lateral boundary because the comparison starting date for each period was selected as a dust storm free day. Some global models did not meet these conditions (such as NAAPS). Consequently, we generally analyzed the submitted model output from the second day of each period.
3. Observation Data
 The purpose of DMIP is not to score (rank) the model performances of each group. However, it is important to examine some model characteristics using observation data. The selected dust episodes were with huge and strong dust storms. Therefore we were able to access many important observation data such as the NIES Mie lidar network data (lidar observation networks at Beijing, Seoul, Nagasaki, and Tsukuba) [Shimizu et al., 2004], WMO SYNOP report dust information (visibility and current weather report), and wind speed observations. Several surface measurements of Total Suspended Particle (TSP), PM10 and PM2.5 data in China, Korea, and Japan were also available for DMIP.
Figure 1 shows the location of observation sites, study regions A and B, and topography. Shaded areas indicate potential dust emission sources (pure desert, loess, and advancing desertification area), as reported by Gong et al.  and desert and semidesert areas, as reported in the USGS (United States Geological Survey) land use data set. The square boxes marked T (Taklimakan), G (Gobi), M (Mongolia), and I (inner Mongolia, China) are typical desert regions used for the detailed model comparisons in section 4.4. Please note that a specified area is a part of an important desert area but does not include the all region of traditional dust source area. The exact specifications of the regions and geographical locations of stations are described in Table 3.
Table 3. Geographical Locations of Selected Region and Stations
Region or Station
meteorological and dust statistics
meteorological and dust statistics
meteorological and dust statistics
meteorological and dust statistics
WS, visibility, PM10, TSP, lidar
WS, visibility, PM10
WS, visibility, PM10
 The SYNOP observations from Tazhong, Ejin Qi, Hohhot, Taiyuan, Shenyang, and Beijing were used for wind speed and dust concentration comparison. The SYNOP observation data were provided every 6 hours and wind speed was recorded with 1 m/s resolution. Note that the SYNOP site does not measure the dust concentration. However, Shao et al.  found an empirical relationship between visibility and dust concentration by fitting the near-surface TSP observations to visibility as the following:
where VCTSP is the dust concentration estimated by visibility in μg/m3 and Dv is visibility in km. We used this estimated TSP concentration only when real TSP (or PM) data were unavailable.
 Daily averaged PM10 observations made by the State Environmental Protection Agency, China (SEPA) at Beijing, Taiyuan, Hohhot and Shenyang were also used. Beijing Normal University provides the TSP (hereafter BNU TSP) and PM10 for the March dust episodes [Sun et al., 2004]. The Korea Meteorological Administration (KMA) provided 30 min interval TSP concentration at Gwanak-san and Gunsan. The East Asia Monitoring Network (EANET [Network Center for EANET, 2003]) provided hourly PM10 and PM2.5 data for Oki, Sado, and Rishiri Islands.
 The National Institute for Environmental Studies (NIES) Mie lidar observation data from Beijing, China and Nagasaki, Japan [Shimizu et al., 2004] were also used to compare vertical profiles of dust. This study used the dust extinction coefficient based on the lidar signal separated by a depolarization ratio. At Beijing, we also used time series data of the boundary layer averaged dust extinction coefficient for comparison.
4. Results and Discussion
4.1. Analyses of Dust Concentration, Wind Fields, and Dust Emission Flux
Figures 2 and 3 show the instantaneous surface dust concentration in region B at the onset of both major dust storms. Distributions of the TOMS Aerosol Index (AI) and SYNOP dust reports are superimposed to allow comparison with the modeled dust distributions. Figures 2 and 3 clearly portray that the dust snapshot distribution appears quite similar in each model. Notwithstanding, the concentration levels are quite different. For example, the difference of dust concentration over the Beijing area on 20 March (Figure 2) is more than 10 times greater when comparing the minimum and maximum concentration models. Similar differences are apparent in the April dust snapshot (Figure 3). Results from low-resolution model groups generally produce smooth dust distributions, whereas high-resolution groups yield many patches of high dust-concentration regions. A great difference is evident in the dust concentrations over Mongolia and inner Mongolia: some models show dense dust concentrations.
Figures 3 and 4 show an average of dust emission flux (DFLX, see Table 2) and wind fields for periods A and B. The averaged dust flux distributions indicate that the dust source region allocation differs greatly among modeling groups. For example, only two dust models allocate Mongolia and inner Mongolia region as high dust emission sources. Another area of large differences is the Taklimakan Desert. The total dust emission for region A (Table 1) ranged from 27 to 336 Tg, with a mean of 120 Tg for period A and 18 to 103 Tg, with a mean of 36.3 Tg for period B. Most models predicted less emission during period B than period A.
 The averaged 10-m wind field (Figures 4 and 5) shows fundamentally similar patterns. However, the wind flow over the Taklimakan area and Tibetan Plateau differ considerably between models. Some models indicate very calm conditions in the Taklimakan Desert, whereas other models give a systematic easterly wind. These differences in wind speed in the source regions partially account for the differences in the DFLX. There are small differences over the Beijing-Shandong peninsula region. Detailed analyses of wind speeds and other meteorological parameters are addressed in the following subsections.
4.2. Time Series Comparisons of Wind Field
 The time variation of the modeled 10-m wind speed and SYNOP observation is shown in Figure 6. Five SYNOP observation sites are selected for this time series comparison. To avoid an extreme value from some of the models, wind speeds from different models are sorted statistically and the minimum, quartile values (25%, 50% = mode and 75% percentile values), and maximum wind speed are shown in Figure 6. To measure the scatter from each model's results, we define the nondimensional scatter ratio S as
where Qn% is the n% percentile value of the model variation of variable Q. In Figure 6, we also show the averaged value of SWS and root-mean-square error (RMSE) for each site. Note that S = 1 means that a model variation range of Q75% and Q25% have the same value of mode (Q50%), which implies a large scatter of model output.
 The model results have scatter with a large deviation. There was no systematic difference between the high-resolution and low-resolution horizontal grid model groups (not shown in Figure 6). A great difference was revealed in the Tazhong site (the center of the Taklimakan Desert), whereas the wind speed difference in Ejin Qi (located in the Gobi Desert) was relatively small. At Hohhot, all model results overestimated the wind speed, but they captured the day-to-day variations quite well. One reason for this systematic difference might be the site location. Hohhot is located within a relatively steep valley. Therefore there could be a difference in elevation between the observation site and the model grid point. The observed wind speeds at Beijing and Taiyuan are sometimes higher than the maximum of the modeled values, but the reason is not certain at this moment. Nevertheless, it is clear that modeled range of WS25%–WS75% percentiles generally reproduce the observed wind variation.
 It is important to mention here that all DMIP dust models are based on the output of global or regional meteorological models using objective analysis results (such as NCEP, ECMWF, JMA, and NOGAPS); in addition, the horizontal resolution of models is 27–100 km. This resolution can be considered as high, and topographical resolution might not exert a great difference among the models. It was also found that the highest-resolution model is not always that which best fits the observations (not shown in Figure 6). The dust emission flux is fundamentally proportional to the third or fourth power of the surface friction velocity (u*). Therefore even small differences (say, 2–3 m/s) will engender a factor of 2 or 3 times difference in dust emission flux. This fact indicates that the difference in model results may lie within the meteorological parameters. Improvement of the meteorological model is a key issue to reduce the differences among dust models.
 The meteorological fields from the various models differ, significantly at times, as was shown in Figure 6, thereby complicating the comparison of the dust emission. The necessity of the unified meteorological condition for dust model comparison was argued within the DMIP community. However, each dust model is strongly connected to its own meteorological driver (model) and the usage of unified meteorological conditions was considered impractical by the participating groups. Consequently, unified meteorological data were not specified for the DMIP groups. Further statistical analyses and discussion of the wind fields are presented in section 4.4.
4.3. Time Series Comparisons of Surface Dust Concentration
Figures 7–10 show a comparison of modeled surface dust concentration with observations. Dust concentration results from China (Tazhong, Ejin Qi, Lanzhou, Hohhot, Taiyuan, Shenyang; Figures 7 and 8) are compared with the estimated TSP concentration (VCTSP) and SEPA PM10 concentrations (daily averaged values). In Korea (Gwanak-san and Gunsan; Figure 8), the modeled dust concentrations are compared with the KMA TSP observations. In Japan (Oki, Sado and Rishiri; Figure 9) the comparison is made with the EANET PM10 and PM2.5 data. Geographical locations of these observation sites are shown in Table 3 and Figure 1. As shown in Figure 6, the surface dust concentrations from each model are sorted statistically and show the minimum, quartile values (25%, 50% and 75% percentile values), and maximum concentration.
 It is clearly seen in Figures 7 and 8 that the modeled dust concentration captured the time variation of VCTSP and SEPA PM10 quite well. Some models yield extremely high or low concentrations. Nevertheless, most of time, the observed values are distributed within the model range of C25%–C75%.
 There exists a very wide scatter of dust concentrations at Tazhong (Taklimakan Desert). Sometimes, the instantaneous value of nondimensional model variation SDC1B exceeded 20. Averaged over the period A SDC1B is 6.18 and period B it is 4.42, which is the highest in this comparison. Figure 7a portrays the large range of modeled dust concentrations at Tazhong, which is very reasonable because of the widely varying wind speeds. This result reinforces the importance of dynamical forcing to dust emission and transport modeling over the Taklimakan Desert, which has been pointed out by Shao and Wang  and Uno et al. .
 At Shenyang, model predicted concentrations for April agreed quite well with SEPA PM10 but were less than VCTSP values. The good correlation between VCTSP and SEPA PM10 at Lanzhou and Hohhot (Figures 7c and 7d) supports the validity of equation (1). The poorer correlation (with VCTSP > SEPA PM10) at Taiyan and Shenyang (Figures 8a and 8b) points to an increasing contribution by aerosols other than dust in visibility degradation.
 At Beijing (Figure 10), the time variation between VCTSP and BNU_TSP value shows very good agreement (except for the absolute concentration). The models also captured this variability well. Dust extinction coefficients by NIES lidar also show a very sharp increase of dust extinction coefficients between Julian days 79 and 80, as simulated by the models. In March, the modeled and observed TSP (PM10) concentrations had a common single peak. In April, the modeled dust concentrations showed twin peaks in China (first peak at Julian day 96 and second peak at 97–98 at Beijing), which agree well with lidar measurement. At this point, we want to point out that the greatest scatter in model dust concentration at Julian day 79 (20 March) and Julian days 96–97 (6–7 April).
 The TSP and PM10 observations in Korea and Japan (Figures 8 and 9) are also well reproduced by the DMIP models. The model variation range of C25%–C75% shows excellent agreement with PM10 and TSP observations when we consider the onset timing of dust. The modeled dust concentrations in Japan (Figure 9) are usually between the observed PM10 and PM2.5 values. One important point is that modeled and observed dust concentrations in Korea and Japan had a twin peak in March (around Julian days 76–77 and 80–81) except at Rishiri. We can see the similar twin peak in Ejin Qi (located in Gobi source region) and Lanzhou stations (and other stations). However, the observation of this twin dust peak strongly depend on the location of observation site and dust transport path. Another point is that the C75% value shows good agreement with the Japanese PM10 observation (which means the averaged model concentration tends to underestimate the dust concentration in Korea and Japan).
 On the basis of analyses presented for Figures 7 and 10, the observed TSP and PM10 concentrations are generally located within the modeled C25% and C75% range in China, whereas C75% has a better agreement in Korea and Japan. We emphasize that the model dust concentration captured the onset and cessation timing of dust phenomena well. However, the prediction of the absolute concentration level itself still presents scattering in the model output.
4.4. Area-Specified Statistics
 As shown in previous sections, there exists a wide range of scatter in modeled meteorological parameters and dust concentrations. The time-averaged statistics from all model outputs within the four subdomains in Figure 1 was examined. The specified areas are the Taklimakan desert (T), the Gobi desert (G), inner Mongolia (I), and southern Mongolia (M), together with several examination areas over Beijing, Qingdao, Seoul, Fukuoka, and Sado (the 2° longitude × 2° latitude area centered over each city). Table 3 lists the longitude-latitude of each area and city. Because of the large scatter in the results, the quartiled statistics were used to exclude those extreme values.
 A large difference is shown for DFLX (see Figure 11a). For example, over the Taklimakan area (T) in April, the averaged dust emission flux ranges from DFLX25% = 16.7 to DFLX75% = 78.2 mg/m2/h with SDFLX = 2.04. The averaged dust concentration ranges from DC1B25% = 181 to DC1B75% = 1120 μg/m3 with SDC1B = 1.83. For the Gobi area (G) in April, the dust emission flux ranges from DFLX25% = 81.8 to DFLX75% = 190.0 mg/m2/h with SDFLX = 0.81. The dust concentration ranges from 371 to 1848 μg/m3 with SDC1B = 1.67. The largest nondimensional scatter ratios of DFLX (SDFLX) were found in areas I (inner Mongolia) and T (Taklimakan), indicating that the dust emission from these two areas differs greatly among the DMIP models. It should be noted that SDFLX is greater than 0.47 in every area.
Figure 12 also shows that SDC1B in areas T and G have the highest values in both March and April. Ranges of concentration variation tend to decrease downwind of the sources to the down stream cities (i.e., toward Korea and Japan). Within the dust source areas of T and G, SDC1B is always greater than 1.2 and reduces to about 0.3–0.8 downwind (except for Sado in April). These tendencies imply that the range of dust concentration differences is almost 120% of its average value within the dust source areas.
 Comparing Figures 12a and 12b we see that the DC1B falls off with distance from the source more quickly than DCLN owing to the fallout of large particles (the ratio between DC1B and DCLN (based on the mode value) is 1:(0.7–1) for Mongolia and Gobi, 1:1.5 for inner Mongolia, and 1:3 for other areas (not shown in Figure 12)). This pattern is quite reasonable because the dust particles over the dust source area include coarse (large) particles, which are removed by gravitational settling and other deposition processes during transport. However, the individual slope between DC1B and DCLN over the source region for each model has a wide large scatter of 1:(0.4–2), which is highly dependent on the treatment of dust emission strength and dynamical forcing.
Figure 13 presents the correlation between averaged model parameters for each area and each model. In Figure 13 we show WS10 versus surface friction velocity u*.
 The results shown in Figure 13 are complicated. There is a relatively clear relationship between wind speed and surface friction velocity within the DMIP group (Figure 13). Two groups can be identified in each area. For example, the model groups 3, 6 and 8 are separated from the others and take different slopes of u* level, even for the same surface wind speed. This separation is apparent for all four areas indicating that the treatment of surface roughness length, wind shear, and surface heat flux is highly dependent on the DMIP model type. A weak positive correlation exists between WS10 and u*, except for the Mongolia area (M).
4.5. Vertical Profile Comparisons
 In this section, the differences in the vertical concentration profiles is discussed. As described in section 3, NIES lidar observation results provide some verification of the model results. Beijing and Nagasaki are selected for the comparison of dust vertical profiles.
Figures 14 and 15 show the time-height (TH) cross section of the potential temperature and dust concentration at Beijing (Figures 14a and 15a) and Nagasaki (Figures 14b and 15b). The NIES lidar dust extinction coefficient is also shown in Figures 14 and 15.
 The NIES lidar observation captured the dust onset to both Beijing and Nagasaki very clearly. Figures 2, 3, and 10 show that the big dust onset to Beijing occurred on Julian days 79–81. The lidar measurement shows that the vertical extent of this dust layer is from the surface to approximately 2 km level (i.e., most dust is captured within the boundary layer), but the highest concentration is near the surface. For Nagasaki, the lidar operation started mainly from Julian day 79. They observed the major onset of dust around Julian day 81–82 from the surface to 2 km altitude. It should be noted that lidar sometimes cannot retrieve the dust signal correctly when the dust layer is so dense to penetrate into the upper layer as pointed out by Uno et al. .
 As Figure 14a shows, every model captured a large-scale dust onset in Beijing around Julian days 79–81 (shown in Figure 2). Every TH profile of dust and potential temperature appears similar. For the vertical dimension, most DMIP models reproduced the BL dust layer quite well. For Nagasaki (Figure 15b), most of the models reproduced the correct onset timing of dust that arrived around Julian days 81–82, but they show different vertical profiles and dust concentration levels. It is interesting to point out that the time-height variation of the potential temperature from the vertically high resolution model shows the fine structure (e.g., COAMPS).
Figure 16 compares the time averaged vertical dust profile. Modeled dust concentrations from the typical dust episodes days are averaged and plotted in Figure 16 for Beijing (Figure 16a) and Nagasaki (Figure 16b). As in Figure 6, modeled results are sorted to show the quartiled values (shaded zone is the range of 25% and 75% percentiled value). The small figure inserted in the upper right section shows the scaled dust profile (C*) by column dust loading (height below H = 10 km). The NIES lidar dust extinction coefficient is also time averaged and shown in Figure 16 (dashed line).
 Beijing (Figure 16a) shows a wide scatter of dust concentration level of each model (SC is larger than 0.35), which is mainly attributable to the difference of dust concentration level, as shown in Figure 12. However, the shapes of the scaled dust concentration profiles have similar profiles, indicating that the main body of dust concentration is captured within the BL for both March and April episodes. It is important to point out that boundary dust layer in Beijing is so dense that the vertical profile by lidar dust tends to underestimate the upper level dust profile, and this may be a reason for the difference between lidar and dust models.
 At Nagasaki (Figure 16b), the vertical dust profiles differ greatly among models (even after by scaling). Some models show an elevated dust peak and others show a surface layer peak. The nondimensional scatter ratio S, as defined by equation (2), was also calculated. The S value for dust concentration below 2 km is greater than 0.35 for Beijing and greater than 1.0 for Nagasaki, which indicates that Nagasaki has a much larger scatter among the DMIP models. This suggests that modeled dust transport/removal processes between China and Japan must be an important issue in improving the dust modeling. Furthermore, it is important to correct transport and removal processes after the dust departs the continent.
5. Concluding Remarks
 A dust model intercomparison project (DMIP) over Asia with eight dust emission/transport models has been accomplished. Two large dust episodes occurred in spring 2002 and were used as the focus for the DMIP study. The submitted model results of DMIP are all based on the original dust emission, transport, and deposition schemes, and have a wide variety of dust emission distribution, surface concentration, and transport processes. Each model has a different dust size bin range and horizontal grid resolution. The meteorological parameters, dust emission flux, and concentration (d < 20 μm) are compared within the same domain. The major findings from this DMIP activity can be summarized as follows:
 1. The modeled surface dust concentrations are compared with SYNOP visibility, Chinese SEPA PM10, Korean KMA TSP, and Japanese EANET PM10 concentrations at several sites. It is found that a modeled concentration ranging between 25% and 75% percentile generally agreed with PM observation. The model results correctly captured the major dust onset and cessation timing at each observation site. However, the peak concentration level of each model was 2–4 times different.
 2. The dust concentration during the major dust onsets shows a quite similar distribution, except for its concentration level. However, the averaged dust emission flux and wind field show visible differences. In particular, the dust emission fluxes from the Taklimakan Desert and Mongolia differ immensely among the models, indicating that the possible dust source allocation scheme and modeled winds over these regions differ greatly by each modeling group. Measurement of surface winds, dust flux and accurate or updated land use information are important in these regions.
 3. The regional averaged statistical analyses clearly indicate that (1) meteorological parameters (wind speed and friction velocity) over the Taklimakan area are highly uncertain, (2) the dust emission flux from inner Mongolia has the greatest scatter, and (3) surface dust concentrations over Mongolia and Gobi have large scatter. The differences decrease downwind over Korea and Japan, even though the nondimensional scattering ratio (equation (2)) is larger among the model results. This reduced discrepancy means that the influence of the difference in dust emissions over the source area diminishes during long-range transport.
 4. The comparison of the dust vertical concentration profile of Beijing and Nagasaki show large scatter (more than twice the scatter in concentration level). For Beijing, the scaled dust profile has a similar vertical profile and relatively good agreement with the lidar extinction profile. For Nagasaki, the scaled dust profile did not have a single profile, indicating that modeled dust transport/removal processes between China and Japan is an important issue in improving the dust modeling. It is important to correctly model transport and removal processes after the dust clouds depart the continent.
 Results of the current DMIP project provide several important directions for the future study of Asian dust modeling and observations. Here we want to point out several directions: (1) further study of dust emission and transport from the Taklimakan Desert is important; (2) dust emissions from Mongolia and inner Mongolia have been only sparsely measured and are highly model-dependent and require more observations and consensus within the dust modeling community; and (3) statistical analyses demonstrate that the scatter of dust concentration among the different dust models diminishes after long-range transport. This last point implies that measurement near the source region is more important, especially for the dust emission flux, dust size distribution, and vertical profile.
 Numerical dust forecast are now operationally used in JMA (Japan Met. Agency) and KMA (Korean Met. Agency), however, their forecast accuracies depend on several conditions. This is because the dust emission processes are sometimes not properly forecasted (or identified). At this moment, as shown in this paper, reliable surface land use conditions and soil/surface information are more important than the complexity of the dust emission scheme or model horizontal resolution. While the accurate information over the desert area in China and Mongolia is very important, the availability of most updated Chinese/Mongolia databases depends on the modeling group's interest and connection. Despite these differences, most of the observed dust concentrations are within the 25–75% percentile of model forecasts. We conclude that an operational ensemble dust forecast, made up of the different numerical models based on the several sources of land use and soil information, dust emission schemes and meteorological model outputs, would provide useful forecasts of dusty weather.
 This work was performed through the project work of Aeolian Dust Experiment on Climate Impact (ADEC) supported by the Ministry of Education, Culture, Sports, Science and Technology of Japan and National Natural Science Foundation of China (project 40305018). The authors wish to acknowledge the Chinese Meteorological Agency (CMA) for their preparation of the SYNOP data set used in this study and Yasunori Kurosaki of the Meteorological Research Institute (currently at Center for Environmental Remote Sensing, Chiba University) for his analyses of SYNOP observation data. We would also like to acknowledge the Chinese SEPA for their PM10 measurement, Beijing Normal University for PM10 and TSP, KMA for their TSP, and EANET for PM10 data for the use of model comparison.