1. Introduction
Mineral dust is generated by wind erosion over arid or semiarid land surfaces and is transported locally and over vast distances, causing adverse environmental and weather problems over broad areas. Dust particles reduce visibility and degrade air quality, thereby disrupting transportation and degrading health. As one of the major components of natural aerosols, dust modifies the radiation budget directly by influencing solar and infrared radiation and indirectly by modifying cloud properties.
Deserts are widely distributed in the southwest Asian countries of Afghanistan, Iran, Iraq, Pakistan, Saudi Arabia, and Syria. Blowing or suspended dust is a frequent weather phenomenon in these countries. The most severe dust storms usually occur in winter and spring and are associated with cold-air outbreaks from Europe and central Asia (Perrone 1979; Walters and Sjoberg 1988). These events can mobilize large amount of dust and transport it far beyond the sources. On regional or local scales, dust storms are associated with strong winds and severe turbulence (Xu et al. 2000; Liu et al. 2000). These factors influenced the outcome of the hostage rescue mission in Iran in 1980.
Real-time prediction of dust storms, especially quantitative forecasting of dust concentration and visibility, has become highly desirable as a meteorological service to the public and military activities. The development of advanced dust aerosol process-oriented numerical weather prediction models makes it possible to predict dust particle life cycles including emission, transport, and removal at high spatial resolution on short to medium time scales. Dust storm prediction models have been used operationally in Europe (Nickovic et al. 2001), Australia (Cope et al. 2004), and East Asia (Shao et al. 2003; Park and In 2003; Tanaka et al. 2003; Uno et al. 2004). At the U.S. Naval Research Laboratory, the U.S. Navy Mesoscale Prediction System’s Coupled Ocean–Atmospheric Mesoscale Prediction System (COAMPS), with an embedded dust aerosol model (Liu et al. 2003), has produced 3-day forecasts for southwest Asia in real-time runs since March 2003. The focus was on the Iraq region including the Arabian Gulf and the nearby areas, providing dust concentration, visibility, and optical depth in support of U.S. Department of Defense military operations. Real-time COAMPS dust forecasting has been carried on since then in a combined research and operational mode. This is the first mesoscale model being used in southwest Asia for operational dust forecasting.
In addition to the COAMPS dust model, the U.S. Navy has also developed the only high-resolution dust source database for southwest Asia that is compatible, in terms of resolution and accuracy, with COAMPS dust forecasting. The previous dust source databases have had a resolution of 1° or less. The work by Walker et al. (2003) is based on years of data from surface observations, satellites, geography, and desertification studies. A dust source database is the fundamental part of dust modeling and affects the modeling quality. COAMPS becomes complete only with the support of an appropriate database.
The U.S.-led coalition forces launched a military mission called Operation Iraqi Freedom (OIF) from 18 March to 30 April 2003. Weather maps, satellite observations, and postdeployment reports reveal that dust was one of the most important meteorological parameters in the Iraq region during OIF. Early in the mission, a strong dust storm swept across southwest Asia and visibility was reduced to less than 100 m. The dust storm dramatically impeded the military activities from the ground to the midtroposphere on 25 and 26 March. Overall, five large-scale dust storms were observed during the 43 days of OIF. The U.S. Navy’s Fleet Numerical Meteorological and Oceanographic Center (FNMOC) used these COAMPS real-time dust forecasts in their daily dust weather discussions.
Previous modeling studies have been either 1) short term (for a field campaign), 2) not truly operational, 3) forecasts of only surface concentration or aerosol optical depth, 4) lacking any statistical verification, 5) carried out at synoptic-scale resolutions, or 6) operated using a 1° resolution dust source database. The verification of COAMPS real-time dust forecasting for this special period is a requirement for a better understanding of the strengths and weaknesses of the COAMPS modeling system. Such verification will help to improve the model and meet the U.S. Navy’s needs for higher quality weather forecasting. In this paper, we apply, for the first time, the standard statistical measures employed in the verification of precipitation forecasts to the verification of the forecasts of dust storms and visibility. The following sections describe the model, observations, forecasts, comparisons, statistical analysis, evaluation, and conclusions.
2. Model description
The U.S. Navy’s mesoscale meteorological system, COAMPS, with embedded dust microphysics, is used to simulate dust storms during OIF. COAMPS is a nonhydrostatic and compressible dynamic model applied in a terrain-following sigma vertical coordinate. It predicts turbulent kinetic energy (TKE) for subgrid-scale diffusion, uses a force–restore method in the surface energy budget, and contains explicit cloud microphysics. The friction velocity, associated with the surface momentum flux, is calculated based on Monin–Obukhov surface-layer similarity theory. Ground wetness, an alternative of soil moisture, is calculated following the algorithm of Louis (1979) using precipitation, snow depth, ice coverage, and evaporation, as well as soil latent heat flux and moisture capacity. The U.S. Geological Survey (USGS) land-use 1-km resolution database is used to obtain surface roughness length for various land surfaces. The complete details of the model structure, dynamics, and physics can be found in Hodur (1997) and Chen et al. (2003).
The model domain extends vertically to an altitude of 35 km with 31 grid layers ranging in thickness from 10 m at the surface to 6 km at the top. There are 13 grid layers below 9 km. This resolution is used by FNMOC in operational runs and therefore is employed in this dust forecasting. In the horizontal, the model uses the three nested grid meshes shown in Fig. 1. The coarse 81-km resolution domain of 92 × 68 grid points is purposely chosen to cover the upstream deserts that might act as sources of dust. The middle 27-km grid nest with 127 × 109 grid points is a transition grid from coarse to fine resolution. The 9-km resolution grid nest with 181 × 181 grid points covers the focus area of Iraq and the eastern Persian Gulf. This combination of nested grid meshes allows COAMPS to capture both synoptic and mesoscale features of dynamical systems and dust storms.
The data assimilation is performed at 12-h incremental update cycles using the meteorological data from the world weather station network and satellite retrievals. The analysis and forecast fields of the U.S. Navy’s Operational Global Atmospheric Prediction System (NOGAPS) (Hogan and Rosmond 1991) are used for the initial conditions and for updating the lateral boundary conditions every 6 h during the forecast. However, there was no dust assimilation update: the model begins each new forecast cycle with the previous 12-h forecast of dust as the initial condition. The real-time dust forecasting started 15 March 2003 and has continued to the present, becoming operational at FNMOC in 2004. Each day COAMPS produces a 72-h forecast beginning at 0000 UTC (0300 Baghdad local time) and a 12-h forecast beginning at 1200 UTC (1500 local time).
A dust aerosol model is fully embedded in COAMPS as an in-line module of the prediction system using the model’s exact meteorological fields at each time step and at each grid point (Liu and Westphal 2001; Liu et al. 2003). The dust module has the same vertical grid structure, multiple nested grids, and grid nest interactions. The mass conservation equation contains source emission, advection, sedimentation, mixing, dry deposition at the surface, and wet removal by precipitation.
Coefficient A in (1) is the fraction of the area of each model grid box that is dust erodible, and thus capable of producing dust. It ranges from 0.0 to 1.0. The distribution of fractional erodibility for southwest Asia is shown in Figs. 2a and 2b. The erodibility for the 81-km grid is determined from a database that is built upon an analysis of the 1° resolution Total Ozone Mapping Spectrometer (TOMS) aerosol index (e.g., Prospero et al. 2002). The 81-km domain covers the deserts of Sahara, Libya, and Sudan in the west; Iraq and the Arabian Peninsula in the center; and Iran, Pakistan, Afghanistan, and India in the east. The erodibility for the 27- and 9-km grids is derived from a high-resolution database developed by Walker et al. (2003) described earlier. Figure 2b, showing the distribution of erodibility for the 9-km grid, reveals detailed structures (especially point sources) that the low-resolution TOMS aerosol-index analysis lacks. Examining Eq. (1), it is clear that the accuracy of dust production depends on both friction velocity and soil erodibility defined in the dust source database. The evaluations of dust forecasts in sections 4–6 will show that the dust erodibility currently used in SW Asia to be practical.
Dust emission is further restricted to the erodible areas that are predicted to have low values of ground wetness as predicted by COAMPS. In the real-time runs, COAMPS performs one-way interaction to the winds and other dynamics fields of grid nests; for example, the coarse-grid data are passed to the lateral boundaries of the nested fine grids. The same dust source function is used in all of the grid nests. Realizing that the dust source database is at very high resolution, a two-way interaction is enforced to the dust mass. The coarse-grid dust concentration is used as the lateral boundary condition of the nested grid, while the inner-grid dust concentration is passed to the coarse grids through averaging of the inner-grid values. On the other hand, dust mass has no impact on the COAMPS dynamics.
Dust is modeled as a monodispersed aerosol, that is, a single particle size, with a diameter of 2.0 μm and a density of 2650 kg m−3. This effective size of the particles was chosen because it provided the closest match to previous size-resolved simulations (10 size bins ranging from 0.05 to 35 μm), both in terms of optical properties and sedimentation fluxes (Liu et al. 2003). Applying Mie scattering theory, a specific extinction coefficient of 0.58 (m2 g−1) is derived for this effective particle size. The forecasted dust mass concentration is then converted to an extinction coefficient by multiplying this specific extinction. The forecasted dust mass load (mass vertical integral, in g m−2) is also directly converted to optical depth by multiplying this value.
Dust advection in both the horizontal and vertical directions uses a fifth-order-accurate flux-form scheme developed by Bott (1989a, b). The algorithm performs a polynomial fitting to the upstream-advected field in each grid box to make the fitting curves approach the data at the grid points through a weighting flux treatment. Therefore, mass conservation and positive definite conditions can be effectively achieved. The process of dust subgrid-scale turbulent mixing is also important to the dust transport, so to the visibility. The eddy diffusion coefficient is the same as that used for moisture scalars and temperature in the dynamics model. COAMPS solves the TKE equation explicitly, from which the eddy diffusion coefficients are generated. The turbulent mixing is then solved implicitly to maintain numerical stability. The details of the dust aerosol microphysics of gravitational sedimentation, dry deposition, and wet scavenging by convective and stable precipitation can be found in Liu et al. (2003).
3. Observational data
Miller (2003) has developed a dust enhancement product (DEP) that detects dust over surfaces in daytime using satellite radiances, provided the surface temperatures are not too cold. As an example, Fig. 3b shows the dust enhancement product using Terra Moderate Resolution Imaging Spectrometer (MODIS) radiances for 0745 UTC 26 March. The dust appears as pink shades in southern Iraq and northern Saudi Arabia with a dust front (leading edge) in the east. This product is useful in locating dust fronts and in qualitative verification of the dust forecasts. Figure 3b will be discussed more in section 4 below, along with other observations, to compare and qualitatively verify COAMPS modeled dust plumes.
The aerosol optical depth (AOD) can be retrieved from satellite radiances during daytime over cloud-free and glint-free oceans (Durkee et al. 2000). This type of retrieval can be used to quantitatively validate the model predictions over the Persian Gulf. As another example, Fig. 4a shows the retrieved optical depth over the Arabian Gulf at 1027 UTC 27 March. It reveals two areas of high optical depth in the Gulf with values of 1.6–2.6 in the north and 1.0–2.0 in the south. Figure 4a will be discussed more in section 4 below to compare with modeled optical depth for a quantitative verification. The AOD includes the effects of all aerosols, while the modeled AOD includes only the contribution by dust. This discrepancy can be neglected when spring dust events dominate the optical depth. The forecasted dust AOD is calculated with a mass extinction efficiency of 0.58 m2 g−1, based on Mie scattering theory for monodispersed dust particles of 2.0-μm diameter and previous size-resolved simulations (Liu et al. 2003).
Visibility and current weather are routinely measured every 3 or 6 h at weather stations and are the only synoptically reported quantities useful for verification of model dust forecasts. In a dust-storm-dominant season such as the spring in the Iraq region, dust particles are the major factor in reduced visibility. Visibility reports are subjective but nevertheless clearly depict the passage of dust storms. For example, observed visibility from three stations of Arar, Hafr, and King Fahad in northern Saudi Arabia (whose locations are marked in Fig. 5) are shown in Figs. 6, 7 and 8. There is consistency among the observations at the three sites with similar timing at adjacent sites. Both Arar and Hafr show a dust frontal passage on 25 March following the cold front and surface temperature drop, while King Fahad shows it on early 26 March. The details of Figs. 6 –8 will be discussed in section 4 below for the model verification at these three stations.
Second, visibility can be used as a threshold for defining dust storm occurrence. For this analysis, an effective dust storm is assumed when the reported visibility is less than 3.5 km. This choice was based on the visibilities reported when the current weather was reported as any of the “dust storm” codes (e.g., 6–9 and 30–35), which are defined by the World Meteorological Organization for weather station reports. Based on this conditional test for a dust storm, we can calculate binary measures of model skill, such as storm frequency, threat scores, prediction rates, missing rates, and false alarm rates. Similar visibility comparisons in dust storm case studies, model verifications, and operational forecasts have been proven to be a successful approach in the dust modeling community in recent years (Shao et al. 2003; Chung et al. 2003; Leys et al. 2002; McTainsh et al. 1998, 2001). Other weather types may be the cause for reported visibilities less than the threshold. For the weather stations in the Iraq region, less than 10% of the observed visibility measured to less than 3.5 km was caused by precipitation, fog, smoke, and haze. Since COAMPS and Eq. (2) do not model these effects, a few nondust78 contributions would cause a small reduction in the skill scores for the forecasted visibilities.
4. Comparisons of model forecasts with observations
a. Verification for the 25–27 March dust storm
The strongest dust storm of the OIF period occurred on 25–27 March at the beginning, and is called the OIF dust storm. The storm grounded the U.S. Air Force for a day and impeded other military operations in and around Iraq. The ability to forecast this type of dust event is critical and the evaluation of COAMPS’ performance for this case is of particular significance.
On 24 March, a low pressure system and an associated cold front moved to the eastern Mediterranean Sea, where the cold air from Europe encountered the warm and dry air from the Sahara. The frontal system swept westward across Syria and Jordan and arrived in Iraq and northern Saudi Arabia on 25 March. The strong northwesterly winds behind the cold front raised dust storms as the front moved over the source regions in the deserts. Dust mobilization continued on 26 March, and the dust plumes followed the cold front, reaching the Arabian Gulf on 27 March.
Figure 3a shows the observed surface winds and dust storm observations at 0600 UTC 26 March. While the wind shift alone justifies the location of the analyzed cold front, we also overlay the COAMPS 6-h forecasts of sea level pressure and surface temperature to more clearly reveal the structure of the weather system. The Terra MODIS DEP for 0745 UTC 26 March (Fig. 3b) shows the dust covering most of Iraq and the northern Arabian Peninsula with the dust front at the leading, or eastern, edge just reaching the Arabian Gulf. For model verification, we show in Fig. 3c the 56-h forecast of dust mass load (the vertical integral of concentration) and surface wind vectors from the 9-km grid for 0800 UTC 26 March for comparison with the surface winds and the satellite DEP shown in Figs. 3a and 3b. The modeled plume covers most of the same large areas of Iraq and Saudi Arabia, except western Saudi Arabia. The location of the modeled dust front agrees well with that of the observed front (Figs. 3a and 3b). The forecasted winds also compare favorably with the observations.
Over the following 24 h, the front moved south-southeast, reaching the southern Arabian Gulf and the Straight of Hormuz by 1027 UTC 27 March, as seen in the satellite-retrieved optical depth of Fig. 4a. A further model verification is done by comparing this satellite AOD with the 58-h forecast of dust optical depth of the 9-km grid for 1000 UTC 27 March (Fig. 4b). The forecast shows maxima over the northern and southern Arabian Gulf, as is seen in the retrieved AOD (Fig. 4a). The spatial distribution agrees well with the retrieval. In particular, the modeled optical depths show two local maxima, as did the retrieval. The values of 0.8–2.0 at both ends of the Gulf agree with the retrieved values in the south but are 25% low in the north. The surface dust observations in Fig. 4b confirm the modeled dust plumes in the area.
Surface visibility observations at Arar and Hafr al Batin in northern Saudi Arabia and King Fahad on the Arabian Gulf (see Fig. 5 for the locations) are compared to the modeled visibility of the 9-km grid from a single 72-h model forecast in Figs. 6 –8 to examine the accuracy of the COAMPS dust forecast at specific sites. These sites were chosen because they are located along the path of the dust front. Lerner et al. (2004) conducted a study of the quality of weather station visibility reports in several southwest Asia countries for 2001–03. These three stations and the 16 other stations shown in Fig. 5 (and used in section 5) were found to report regularly and to provide consistent data. Iraq did not report their meteorological observations to the World Meteorological Organization (WMO) before or during the war.
The visibility observations and forecasted visibility as well as the forecasted winds and temperature from Arar for 0000 UTC 24 March to 0000 UTC 27 March are shown in Fig. 6 and depict the dust arriving on the morning of 25 March in the large-scale southerly winds that developed in the warm sector ahead the cold front. The main dust front arrived after the passage of the cold front that caused a large temperature drop in the afternoon. Strong westerlies across northern Saudi Arabia continued after the passage of the front until late on 26 March, in response to a surface low that developed over Syria. After a short period of moderate clearing early on 27 March, the second dust storm arrived. COAMPS accurately predicted the two events but the forecasted visibility during the clearing episode was higher than observed.
Figure 7 shows the comparison of observed visibility with COAMPS forecasts at station Hafr al Batin for the same forecast period. The observations and COAMPS forecasts display the same events as observed at Arar, except the cold front arrived several hours later and the forecasted visibilities during the clearing event before the cold front are lower than observed. Additionally, there is a smaller event in the afternoon of 24 March that occurs in moderate southeasterlies. COAMPS predicted this event, but with lower visibilities than observed.
Figure 8 is similar to Figs. 4 and 5 but for station King Fahad and for a single 72-h forecast from 0000 UTC 25 March to 0000 UTC 28 March. Both observations and model forecasts show the low, prefrontal visibilities on 25 March, the cold front passage on 26 March followed by the low visibilities in the north and westerly winds. While the winds over northern Saudi Arabia diminish by the end of 26 March, dust remains in suspension over the Arabian Gulf and coastal areas throughout 27 March.
b. Overview of dust activity during OIF
Weather reports and satellite imagery indicate that there were numerous dust storms during OIF. For example, in Fig. 9a we show the time series of the total number of low-visibility (<3.5 km) reports for all the surface weather stations within the 9-km model domain (Fig. 1) from 16 March to 30 April. The data reveal five large events centered around 20 and 26 March, and 8, 16, and 27 April that were produced by frontal passages and deep low pressure systems with each episode lasting 2–3 days. In Fig. 9b we show the time series of COAMPS total dust mass load (Mt) of the 9-km grid in the entire 9-km domain. Though a different quantity, the simulated mass load shows good qualitative agreement in the timing and duration of the five large events of the observations in Fig. 9a. The relative magnitudes appear closely correlated, with the event on the 25–27 March being the largest.
The observations and COAMPS forecasts (Figs. 6 –9) reveal clear diurnal signals with the onset in lifting occurring soon after sunrise, or between 0000 and 0600 UTC, when vertical mixing of momentum occurs. The minima occur in the late-night hours, or before 0000 UTC, when nocturnal inversions minimize the vertical exchange and surface winds are lowest. The dust emission flux [Eq. (1)] is a function of u* and therefore accounts for both thermal stability and wind shear effects on the momentum exchange in the surface layer. The large-scale dynamic forcing modulates this diurnal pattern.
In summary, the direct comparisons of the modeled results with surface dust and visibility observations, satellite images, and optical depth retrievals illustrate that COAMPS accurately predicts the timing, strength, and spatial coverage of major dust events for the Iraq region. In the next section, we perform a more quantitative evaluation of model performance by conducting a statistical analysis of the visibility forecasts for the region.
5. Statistic analysis of model forecasts over the OIF period
a. Evaluation of statistical errors
Statistical analysis is an effective verification method widely used in the numerical weather forecast community and is applied here to evaluate the accuracy of COAMPS real-time dust forecasts throughout the OIF period in the focus area. Some representative statistical errors and forecast rates are calculated at 19 weather stations surrounding Iraq (see Fig. 5). These stations are chosen according to the data quality (Lerner et al. 2004), frequency of observations, and proximity to Iraq. In this analysis, the dust storm frequency is defined as the ratio of the number of observations of low visibility ≤3.5 km, to the total number of weather reports at each station in the OIF period. The stations having the highest dust storm frequency (i.e., ≥6%) are found to the south and southeast of Iraq (in Fig. 5) and at or downwind of dust source areas (Fig. 2b).
Figure 10 shows the bias error of the modeled visibility for the 9-km grid at each of the 19 stations throughout the OIF period (16 March–30 April). The stations are ordered by their locations in Fig. 5 from Siirt in northern Turkey, counterclockwise around the Iraqi border to Orumieh in northwestern Iran. The bias error is both positive and negative and varies from station to station, with nine stations showing positive bias, nine stations negative, and one station zero. The one pattern that emerges is for negative biases (forecasted visibility too low) along southern Iraq from Truaif to King Fahad and positive biases elsewhere. The negative biases are likely due to overestimation of the dust source strength. The positive biases may be due to the absence of other optically active aerosols in COAMPS, such as smoke, pollution, and dust generated by anthropogenic activities. The Aerosol Robotic Network (AERONET) measurement from Bahrain has shown persistently nonzero AOD (low visibility) even during dust-free conditions (Smirnov et al. 2002).
Figure 11 shows the rms and relative errors of modeled visibility for the 9-km grid at the 19 stations. Compared with Fig. 5, it is clear that the rms error depends on dust storm frequency with error of 3.0–3.5 km at the stations of high dust frequency in the south and error of 1.5–2.0 km at the stations of low dust frequency in the north. This is because the stations in the south experience more low-visibility days and are likely to have larger errors, while at the low-frequency stations, both the observed and predicted visibilities often reach the maximum reported values so individual errors are often small or even zero. The relative error has a similar dependency on dust frequency. It is about 20% high on average at the dust-frequent stations (e.g., an average 2.0 in Fig. 11) and 7% low on average elsewhere (e.g., on average about 0.7 in Fig. 11).
b. Evaluation of forecast rates
Another statistical tool used to quantify COAMPS accuracy is the calculation of forecast rates in four categories: dust storm prediction rate, missed dust storm rate, clear-sky (e.g., no dust) prediction rate, and dust storm false alarm rate. A linear threshold method is adapted from the evaluations of precipitation forecasts (Gyakum and Samuels 1987; Hamill and Colucci 1998; Colle et al. 1999; Colle and Mass 2000). A visibility threshold is required to delineate between a dust storm and a clear sky for a binary comparison between observations and the model. A threshold of 3.5 km again is chosen based on the visibilities reported during dust storms in the Iraqi region over the OIF period. A dust storm is presumed to have occurred when the observed visibility is less than the threshold. Dust-free, or clear-sky, conditions are assumed when the observed visibility is greater than the threshold. For a given station, COAMPS is assumed to have predicted a dust storm if any of the hourly, predicted visibilities fall below 3.5 km during the 3-h window centered on the observation time. Otherwise, it is considered a clear-sky prediction. In this way, each 3-h window is counted as one incident of either a dust storm or clear skies, based on the 3.5-km threshold. The forecast rates are calculated as follows:
dust storm prediction rate = number of correctly predicted dust incidents/number observed dust incidents,
dust storm false alarm rate = number of falsely predicted dust incidents / number of observed clear-sky, incidents
dust storm threat score = (number of predicted dust incidents)/(predicted dust + missed dust + false alarm dust incidents), and
total prediction rate = (number of correctly predicted dust incidents + correctly predicted clear-sky incidents)/(total observations).
The threat score measures the forecast accuracy when the clear-sky incidents are removed from consideration. The total prediction rate is the overall accuracy of dust storm and clear-sky incidents that are correctly predicted.
Figure 12 shows the dust storm prediction rate and false alarm rate of the 9-km grid at the 19 stations for the period, as well as the dust storm threat score. As before, the forecast rates strongly depend on dust storm frequency with the high dust-frequency stations exhibiting high dust storm prediction rates of 50%–95%. (See the stations of high and low dust frequency in Fig. 5.) Due to the nature of dust particle sedimentation and the high mountains located in the north of the Iraq region, most of the dust mass is deposited in the south near the deserts, while a small amount of dust is transported downwind, mostly being lifted high up in the air for long-distance transport. Therefore, the surface stations in the north, far away from the source areas, experience much lower dust storm activity than those in the south. At these low- or no-dust incidents stations, this type of linear calculation method results in apparently poor forecast rates because the number of observed dust incidents is very small. As seen in Fig. 12, these stations in the north have a 20% or lower dust storm prediction rate, for example, an 80% or higher missed dust storm rate. On average, the dust storm false alarm rate is low at all 19 stations, for example, high clear-sky prediction rate, because the number of no-dust or clear-sky weather incidents is much larger than that of dust storm incidents. Therefore, the false alarm rate is low everywhere, even below 10% at high-dust-frequency stations located near the dust source areas. The dust storm threat score is accordingly high at high-dust-frequency stations due to the high dust prediction rate, having scores ranging from 0.3 to 0.55, whereas it is below 0.15 at the stations far from the dust source areas.
Figure 13 shows the total model prediction rate of the 9-km grid at the 19 stations. More than 85% of the observed weather incidents (dust storm and clear sky) were correctly predicted everywhere throughout the OIF period. COAMPS appears to have provided accurate real-time forecasts to the U.S. Navy and the U.S.-led coalition forces.
6. Statistical evaluation of 12–72-h forecasts for the 26 March dust storm
As described in section 4 and seen in Figs. 7 and 9, 25 and 26 March experienced the strongest dust storm in the 9-km domain during the OIF period. In this section, we calculate the forecast rates of dust storm and clear-sky prediction for the 12-, 24-, 36-, 48-, 60-, and 72-h forecasts at the single verification time of 1200 UTC 26 March and averaged over 95 weather stations in the 9-km domain of the Iraqi region. The same linear method as used above is used here. The forecast rates of dust storm prediction and clear-sky prediction (=1.0 − dust false alarm rate) for the 9-km grid averaged in the 9-km domain are plotted in Fig. 14 for the 12–72-h forecasts, valid at 1200 UTC 26 March.
COAMPS generates nearly constant high values of prediction rates of both dust storms (78%) and clear sky (95%) with little change in the forecast skill for the different length forecasts. This means COAMPS 3-day dust forecasts are as accurate as the 12-h forecasts. In addition to the forecast rates, the bias error, rms error, and relative error at the 9-km grid are calculated and averaged over the 9-km domain to have a further examination. A clear trend is also absent in these three statistical errors (Fig. 15). The 27- and 81-km grid nests also had nearly invariable dust forecast rates and statistic errors at this particular time of 1200 UTC 26 March over the 9-km domain (not shown).
Figure 16 shows the statistics of bias, rms, relative, and absolute errors, as well as vector error, at the 9-km grid for surface wind speed and direction calculated at the surface stations within the 9-km model domain, for 12–72-h forecasts valid at 1200 UTC 26 March. None of the errors show any trend. The relative error is between 25% and 28%, while the bias error is always positive, indicating COAMPS overpredicts wind speeds in this region. The 27- and 81-km grids in the 9-km domain also show no trend in forecast skill (not shown).
Figure 17 shows the bias, rms, relative, and absolute errors at the 9-km grid of surface temperature in the 9-km domain. Both rms error and absolute error decline very slightly from 12 to 72 h, but the differences are only about 5%–7% and are not significant enough to conclude that the COAMPS temperature forecast accuracy has changed. The bias error indicates a warm bias for a short forecast, changing to a cold surface temperature bias for longer forecasts. The same trends are found in the 27- and 81-km grids (not shown). The constant statistical scores in the 3-day forecasts of winds and temperature as forecast length increases in this case of 26 March implies a consistently accurate forecast, and that observational data (especially radiosondes) are insufficient in the Iraq region to show the benefit of data assimilation for COAMPS. Since observational data were consistently missing during the Iraq war period, it is likely that data assimilation did not contribute throughout the period and the performance of dust forecasting is similar to this case. The same model was evaluated in another study for the Middle East (Nachamkin and Hodur 2001). That study showed similar statistical scores of surface wind and temperature to the ones found in this study, but forecast skill improved with decreasing forecast length because of sufficient observations and effective data assimilation.
7. Summary and conclusions
Dust storms are a significant weather phenomenon in the Iraq region. We have conducted real-time dust forecasting for the U.S. Navy using the COAMPS Mesoscale Prediction System with an in-line dust aerosol model during the OIF period in March and April 2003. By simulating the dust life cycle of emission, transport, and deposition, and using a high-resolution dust source database for southwest Asia, we were able to forecast dust storms in 1–3 days in advance along with the conventional meteorological variables. Our daily 12–72-h forecasts were used in the weather discussion produced by the U.S. Navy Fleet Numerical Meteorological and Oceanographic Center during OIF. COAMPS real-time runs provided valuable products for the U.S. Navy and contributed to the U.S.-led OIF mission.
COAMPS real-time forecasts have been analyzed and verified by comparing with available observations of enhanced satellite images and retrievals, surface weather reports, and visibility measurements. We show that COAMPS predicted the five major dust storms that occurred in the OIF period in good agreement with observations in terms of timing, strength, duration, and spatial coverage. With grid nesting (triple nested) and high-resolution modeling (9 km), COAMPS predicted the dust front passage of the strong dust storm on 25–27 March. Three weather stations located along the Iraqi border and the Arabian Gulf were specifically verified for the 3-day forecasts of visibility, revealing consistent forecasts of the passage of dust storms with accurate timing and intensity of dust plumes and visibility.
We use 19 weather stations surrounding the Iraqi border to calculate statistical errors and forecast rates of modeled dust visibility for a quantitative evaluation of COAMPS dust forecasting for the OIF period. The bias error varies from positive (due to possible overestimation of dust fluxes) to negative (due to the lack of other aerosols in the model). It was found that both rms and relative errors depend on the frequency of dust storms and the locations of the stations. The stations located in the south near the source areas experience more dust storms and a larger range in visibilities, and tend to have higher statistical errors. While the stations in the north often have visibilities reaching the maximum reported and a smaller range of visibilities, so individual statistical errors are smaller.
Both the dust storm prediction rate and threat score were high at the stations of high dust frequency in the south, reaching 50%–90% accuracy and a 0.3–055 score for the OIF period, respectively. The false alarm rate is low everywhere because the number of clear-sky weather incidents is much larger than that of dust storms. Overall, COAMPS correctly predicted more than 85% of the observed dust storm and clear-sky weather at all stations in the Iraq region.
Dust forecasts of various lengths (12–72 h), averaged over 95 stations in the 9-km domain of the Iraqi region, were evaluated for the strongest dust event on 26 March to examine the model forecast skill. COAMPS generates nearly constant high values of prediction rates of dust storms (78%) and clear sky (95%) for all forecast lengths. We also found that both the statistical errors of visibility and the errors of the corresponding dynamic forcing of winds and temperature present little change in different length forecasts. The constantly high forecast rates and the lack of improvement in forecast performance imply that the model forecasts were consistent, and the lack of radiosondes data in the Iraqi region eliminates the potential benefit of data assimilation to COAMPS dust modeling.
This paper has presented the new capability of COAMPS in dust modeling and the new skill in operational forecasting. The study of real-time dust forecasting and verification has given us several insights into operational dust forecasting. 1) Dust source specification and resolution are a fundamental part of dust modeling and greatly affect the accuracy of forecasting. Work is on going on high-resolution dust source identifications for the rest of the globe. 2) Dust aerosol modeling with a single particle size is practical and accurate in operational runs where computational resources are limited. 3) The surface weather station observations and satellite retrievals are invaluable to the dynamics and dust modeling verification. 4) The improvement of forecast accuracy may depend on the data assimilation of the dust component using satellite retrievals in the future. The COAMPS dust forecast model became operational at FNMOC for the Iraq region in 2004.
Acknowledgments
We thank Dr. Jason Nachamkin for the discussions about statistical analyses. We also thank Mr. Arunas Kuciauskas for providing archived satellite images. We thank FNMOC for valuable discussions on post-OIF dust forecasting. The support of the Office of Naval Research and the Naval Research Laboratory through Program PE-0602435N are gratefully acknowledged. This work is also supported in part by a grant of HPC time from the Department of Defense Shared Resource Center.
REFERENCES
Alfaro, S. C., Gaudichet A. , Gomes L. , and Maillé M. , 1997: Modeling the size distribution of a soil aerosol produced by sandblasting. J. Geophys. Res., 102 , 11239–11249.
Bott, A., 1989a: A positive definite advection scheme obtained by nonlinear renormalization of the advective fluxes. Mon. Wea. Rev., 117 , 1006–1015.
Bott, A., 1989b: Reply. Mon. Wea. Rev., 117 , 2633–2636.
Chen, S., and Coauthors, 2003: COAMPS 3.0 model description—General theory and equations. NRL Tech. Note NRL/PUB/7500-0-3-448, 143 pp.
Chung, Y. S., Kim H. S. , Park K. H. , Jhun J. G. , and Chen S. J. , 2003: Atmospheric loadings, concentration and visibility associated with sandstorms: Satellite and meteorological analysis. Water Air Soil Pollut.: Focus, 3 , 21–40.
Colle, B. A., and Mass C. F. , 2000: The 5–9 February 1996 flooding event over the Pacific Northwest: Sensitivity studies and evaluation of the MM5 precipitation forecasts. Mon. Wea. Rev., 128 , 593–617.
Colle, B. A., Westrick K. J. , and Mass C. F. , 1999: Evaluation of MM5 and Eta-10 precipitation forecasts over the Pacific Northwest during the cool season. Wea. Forecasting, 14 , 137–154.
Cope, M. E., and Coauthors, 2004: The Australian air quality forecasting system. Part I: Project description and early outcomes. J. Appl. Meteor., 43 , 649–662.
Durkee, P. A., and Coauthors, 2000: Regional aerosol optical depth characteristics from satellite observations: ACE-1, TARFOX and ACE-2 results. Tellus, 52B , 1–14.
Gillette, D. A., 1978: A wind tunnel simulation of the erosion of soil: Effect of soil texture, sandblasting, wind speed, and soil consolidation on dust production. Atmos. Environ., 12 , 1735–1743.
Gillette, D. A., and Passi R. , 1988: Modeling dust emission caused by wind erosion. J. Geophys. Res., 93 , 14233–14242.
Gyakum, J. R., and Samuels K. J. , 1987: An evaluation of quantitative and probability of precipitation forecasts during the 1984–85 warm and cold seasons. Wea. Forecasting, 2 , 158–168.
Hamill, T. M., and Colucci S. J. , 1998: Evaluation of Eta–RSM ensemble probabilistic precipitation forecasts. Mon. Wea. Rev., 126 , 711–724.
Hodur, R. M., 1997: The Naval Research Laboratory’s Coupled Ocean/Atmosphere Mesoscale Prediction System (COAMPS). Mon. Wea. Rev., 125 , 1414–1430.
Hogan, T. F., and Rosmond T. E. , 1991: The description of the U.S. Navy Operational Global Atmospheric Prediction System’s Spectral Forecast Model. Mon. Wea. Rev., 119 , 1786–1815.
Lerner, J. A., Westphal D. L. , and Reid J. S. , 2004: Quality controlled surface visibility observations used to validate predicted surface aerosol concentration for southwest Asia. Preprints, 20th Conf. on Weather and Forecasting, Seattle, WA, Amer. Meteor. Soc., CD-ROM, P4.3.
Leys, J., McTainsh G. H. , Shao Y. , and Tews K. , 2002: Testing of regional wind erosion models for environment auditing. Proc. Joint Meeting of the Int. Conf. on Aeolian Research and the Global Change and Terrestrial Ecosystem–Soil Erosion Network, Lubbock, TX, Texas Tech University, USA Publ. 02-2, 168–172. [Available online at http://www.csrl.ars.usda.gov/wewc/icar5/individuals/50.pdf.].
Liu, M., and Westphal D. L. , 2001: A study of the sensitivity of simulated mineral dust production to model resolution. J. Geophys. Res., 106 , 18099–18122.
Liu, M., Westphal D. L. , Holt T. R. , and Xu Q. , 2000: Numerical simulation of a low-level jet over complex terrain in southern Iran. Mon. Wea. Rev., 128 , 1309–1327.
Liu, M., Westphal D. L. , Wang S. , Shimizu A. , Sugimoto N. , Zhou J. , and Chen Y. , 2003: A high-resolution numerical study of the Asia dust storms of April 2001. J. Geophys. Res., 108 .8653, doi:10.1029/2002JD003178.
Louis, J. E., 1979: A parametric model of vertical eddy fluxes in the atmosphere. Bound.-Layer Meteor., 17 , 187–202.
McTainsh, G. H., Lynch A. W. , and Tews E. K. , 1998: Climatic controls upon dust storm occurrence in eastern Australia. J. Arid Environ., 39 , 457–466.
McTainsh, G. H., Leys J. , and Tews K. , 2001: Wind erosion trends from meteorological records. State of the Environment Technical Paper Series (Land), Series 2, Dept. of Environment and Heritage, Canberra, Australia. [Available online at http://www.deh.gov.au/soe/techpapers/wind-erosion/.].
Miller, S. D., 2003: A consolidated technique for enhancing desert dust storms with MODIS. Geophys. Res. Lett., 30 .2071, doi:10.1029/2003GL018279.
Nachamkin, J. E., and Hodur R. M. , 2001: Sensitivity of short-term forecasts from the navy COAMPS to grid configuration and data assimilation. Preprints, 18th Conf. on Weather Analysis and Forecasting/14th Conf. on Numerical Weather Prediction/Ninth Conf. on Mesoscale Processes, Fort Lauderdale, FL, Amer. Meteor. Soc., 77–80.
Nickling, W. G., and Gillies J. A. , 1993: Dust emission and transport in Mali, West Africa. Sedimentology, 40 , 859–868.
Nickovic, S., Kallos G. , Papdopoulos A. , and Kakaliagou O. , 2001: A model for prediction of desert dust cycle in the atmosphere. J. Geophys. Res., 106 , 18113–18129.
Park, S. U., and In H. J. , 2003: Parameterization of dust emission for the simulation of the yellow sand (Asian dust) event observed in March 2002 in Korea. J. Geophys. Res., 108 .4618, doi:10.1029/2003JD003484.
Patterson, E. M., and Gillette D. A. , 1977: Measurements of visibility vs mass concentration for airborne soil particles. Atmos. Environ., 11 , 193–196.
Perrone, T. J., 1979: Winter Shamal in the Persian Gulf. NAVENVPREDRSCHFAC, TR 79-06, Naval Research Laboratory, 160 pp. [Available from Naval Research Laboratory, Monterey, CA 93043.].
Prospero, J. M., Ginoux P. , Torres O. , Nicholson S. , and Gill T. , 2002: Environmental characterization of global sources of atmospheric soil dust identified with the NIMBUS7 Total Ozone Mapping Spectrometer (TOMS) absorbing aerosol product. Rev. Geophys., 40 .1002, doi:10.1029/2000RG000095.
Shao, Y., and Coauthors, 2003: Northeast Asian dust storms: Real-time numerical prediction and validation. J. Geophys. Res., 108 .4691, doi:10.1029/2003JD003667.
Smirnov, A., and Coauthors, 2002: Atmospheric aerosol optical properties in the Persian Gulf region. J. Atmos. Sci., 59 , 620–634.
Tanaka, T. Y., Orito K. , Sekiyama T. T. , Shibata K. , Chiba M. , and Tanaka H. , 2003: MASINGAR, a global tropospheric aerosol chemical transport model coupled with MRI/JMA 98 GCM: Model description. Paper Meteor. Geophys., 53 , 119–138.
Tews, E. K., 1996: Wind erosion rates from meteorological records in eastern Australia 1960–92. Honours thesis, Griffith University, Nathan, Australia, 99 pp.
Uno, I., and Coauthors, 2004: Numerical study of Asian dust transport during the springtime of 2001 simulated with the CFORS model. J. Geophys. Res., 109 .D19S24, doi:10.1029/2003JD004222.
Uno, I., and Coauthors, 2006: Dust model inter-comparison (DMIP) study over Asia—Overview. J. Geophys. Res., 111 .D12213, doi:10.1029/2005JD006575.
Walker, A. L., Richardson K. , Miller S. , and Westhpal D. , 2003: Revised land use characteristic dataset for southwest Asia for NAAPS and COAMPS dust. Proc. Battlespace Atmospheric and Cloud Impacts on Military Operations Conf., Monterey, CA, Defense Research and Engineering, 1.14. [Available online at http://www.nrlmry.navy.mil/BACIMO/2003/bacimo.html.].
Walters, K. R., and Sjoberg W. F. , 1988: The Persian Gulf region—A climatological study. USAFETAC TN-88/002, USAF Environmental Technical Application Center, Scott Air Force Base, IL, 62 pp. [Available from USAF Environmental Technical Application Center, Scott Air Force Base, IL 62225.].
Westphal, D. L., Toon O. B. , and Carlson T. N. , 1987: A two-dimensional investigation of the dynamics and microphysics of Saharan dust storms. J. Geophys. Res., 92 , 3027–3049.
Westphal, D. L., Toon O. B. , and Carlson T. N. , 1988: A case study of mobilization and transport of Saharan dust. J. Atmos. Sci., 45 , 2145–2175.
Xu, Q., Liu M. , and Westphal D. L. , 2000: A theoretical study of mountain barrier jet over sloping valleys. J. Atmos. Sci., 57 , 1393–1405.
Model domains of nested grids of 9-, 27-, and 81-km resolution used in COAMPS dust forecasting for the Iraq region. (The number of grid points in the x and y directions are given in parentheses)
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
(a) Dust source distribution for the 81-km domain expressed as a grid erodible fraction (0.0–1.0) using the U.S. Navy’s high-resolution dataset in Iraq and north Saudi Arabia and the TOMS 1° database elsewhere. (b) Dust source distribution for the 9-km domain expressed as a grid erodible fraction (0.0–1.0) using the navy’s high-resolution dataset.
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
(a) Weather map of observed surface wind (blue, full tick = 10 m s−1), dust observation (red symbols), and cold front (heavy black dashed line) for 0600 UTC 26 Mar 2003. Simulated sea level pressure (green contours) and surface temperature (yellow contours) are also added to show synoptic forcing. (b) Satellite image of dust plumes (pink) in the Iraq region retrieved by the U.S. Naval Research Laboratory at 0745 UTC 26 Mar 2003. Heavy dashed line indicates the dust front. It is the same area, coverage is as in (a). (c) COAMPS 56-h forecasts of dust mass load (vertically integrated, mg m−2) and surface wind (full tick = 10 m s−1) on the 9-km grid valid at 0800 UTC 26 Mar 2003.
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
(a) Satellite-retrieved optical depth in the Arabian Gulf by U.S. Naval Research Laboratory at 1027 UTC 27 Mar 2003. Two dashed circles indicate the maximum areas of optical depth. (b) COAMPS 58-h forecast of dust optical depth in the 27-km domain valid at 1000 UTC 27 Mar 2003 and surface dust observations (dollar symbols) in the Arabian Gulf.
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
A map showing the locations of 19 weather stations surrounding Iraq. Stations Arar, Hafr al Batin, and King Fahad are marked in dashed circles. Stations marked with solid dots have observed high frequency of dust storms and are near dust source areas. Solid triangles are the radiosondes in the Iraq region. The order of the stations in the statistics plots (Figs. 10 –13) from left to right begins at Siirt in eastern Turkey and continues counterclockwise around Iraq, ending with Orumieh in northern Iran.
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
Time series of (a) observed visibility, (b) COAMPS 0–72-h forecasted dust visibility on the 9-km grid, and (c) 0–72-h forecasts of surface temperature and wind barbs (a full tick = 10 m s−1) at Arar weather station from 24 to 27 Mar 2003. Note the inverted visibility (y) axis.
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
Same as in Fig. 6 but for station Hafr al Batin.
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
Same as in Fig. 6 but for station King Fahad from 25 to 28 Mar 2003.
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
(a) Time series of the total number of weather stations that observed dust storms (defined as visibility < 3.5 km) in the 9-km model domain. (b) Time series of modeled total dust mass load in the air (Mt) in the 9-km domain for the 9-km grid. Maximums are circled to show the timing of storms between (a) and (b).
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
Bias error (km) of modeled visibility on the 9-km grid at 19 stations from 16 Mar to 20 Apr 2003. The order of the stations begins with Siirt in eastern Turkey and moves counterclockwise around Iraq, ending with Orumieh in Iran (see Fig. 5).
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
The rms error (km) and relative error (%) of the modeled visibility on the 9-km grid at 19 stations from 16 Mar to 20 Apr 2003. Heavy arrow bar covers the stations of high dust frequency. Relative error is multiplied by 10 to better show station-dependent distribution.
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
COAMPS dust storm prediction rate, false alarm rate, and threat score on the 9-km grid at 19 stations from 16 Mar to 20 Apr 2003. Heavy arrow bar covers the stations of high dust frequency.
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
COAMPS total prediction rate of correctly predicted dust storm and correctly predicted clear sky on the 9-km grid at 19 stations from 16 Mar to 20 Apr 2003.
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
Dust storm forecast rate and clear-sky (=1.0 − false alarm rate) forecast rate of the 12-, 24-, 36-, 48-, 60-, and 72-h forecasts valid at 1200 UTC 26 Mar 2003, averaged in the Iraq region of the 9-km domain for the 9-km grid.
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
Surface visibility statistics for the 9-km grid of bias error (km), rms error (km), and relative error (%) of 12–72-h visibility forecasts valid at 1200 UTC 26 Mar 2003 in the Iraq region for the 9-km domain.
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
Surface winds statistics on the 9-km grid of bias error (m s−1), rms error (m s−1), absolute error (m s−1), vector error (m s−1), and relative error (%) for the 12–72-h forecasts valid at 1200 UTC 26 Mar 2003 in the Iraq region for the 9-km domain.
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
Surface temperature statistics on the 9-km grid of bias error (K), rms error (K), absolute error (K), and relative error (%) for the 12–72-h forecasts valid at 1200 UTC 26 Mar 2003 in the Iraq region for the 9-km domain.
Citation: Weather and Forecasting 22, 1; 10.1175/WAF971.1
