a. VDRAS

The Variational Doppler Radar Analysis System (VDRAS) uses four-dimensional variational data assimilation (4DVAR) to retrieve three-dimensional time-varying wind and temperature fields from a sequence of volume scans. The basic concept behind 4DVAR is to fit the output of a prognostic model to spatially and temporally resolved observations. An early prototype of VDRAS was developed and demonstrated by Sun et al. (1991) using simulated single-Doppler radar data. The technique was later applied to a dry gust front case using Doppler radar data (Sun and Crook 1994), and then adapted to study the structure and dynamics of convective storms (Sun and Crook 1998). More recently, versions of VDRAS have been run operationally for demonstration projects, including the 2000 summer Olympics in Sydney, Australia (Crook and Sun 2004), and the 2004 Pentagon Shield Field Program (Warner et al. 2007).

At a minimum, VDRAS requires radial velocity data from a single Doppler radar as well as temperature soundings. VDRAS can also be configured to assimilate radar reflectivity, surface mesonet data, and data from multiple Doppler radars. The model’s boundary conditions are prescribed, so that the initial conditions uniquely determine the model solution. Thus, the fit is performed by variational adjustment of the model’s initial state. The initial conditions are iteratively adjusted in order to optimize the agreement between the model solution and the available observations over a given assimilation cycle that typically includes either two or three complete volume scans. Optimization is achieved using the adjoint of the prognostic model.

For this study, an operational version of VDRAS is used so that the prognostic model simulates dry, shallow incompressible flow. Fourteen vertical levels were used up to 5.06 km. The height of the first level was at 187.5 m and a grid spacing of 375 m was used for subsequent levels. The coordinate system does not include variations in terrain; therefore, model physics and parameterization schemes in the operational version are simpler than those implemented in most mesoscale models. These simplifications are essential to reduce the complexity of the adjoint code and the resulting computational expense of the cost function minimization algorithm (Crook and Sun 2004). The VAD wind profile from the radar provided the first guess of winds, and a climatological potential temperature sounding was used for the first guess of temperatures.

The following two quality control (QC) procedures were performed on the radar scans: one was velocity dealiasing and the other was a generalized check to remove remaining noisy and spatially inconsistent data and to clean any clutter-contaminated data that had a small velocity value.

b. 2DVAR

The two-dimensional variational data assimilation (2DVAR) is a statistical interpolation technique (Daley 1991) developed for radar wind analyses by Xu et al. (2006) based on the previous radar velocity analysis technique described by Xu and Gong (2003). This 2DVAR can be regarded as an extension of the traditional velocity–azimuth display (VAD) technique (Browning and Wexler 1968). This method uses radar radial velocity data to apply an optimal correction to a prior background field using assumed analytic forms for the error covariance of the background field. The background field, in this case, is determined from a VAD analysis of the conical scan data.

The method employed by 2DVAR is appealing from an operational point of view because of its relative simplicity and computational speed, its ability to retrieve velocities on a single conical scan surface, and its limited data requirements. However, 2DVAR is relatively new and has not been run operationally like VDRAS. As with VDRAS, the algorithm’s ability to handle processes influenced by complex terrain has not been developed. Currently, the method assumes that the canonical form of the background error covariance tensor function is homogeneous and isotropic in the horizontal following a constant height above the sea level or a flat ground surface. Another consideration is with the sensitivity of the retrieved wind fields to changes in the prescribed decorrelation length parameters, although these parameters should and can be estimated in principle from radar innovation data (Xu et al. 2007).

In this study, the QC package described in Gong et al. (2003), Zhang et al. (2005), and Liu et al. (2005) was applied to the level-2 raw data.

a. JU2003 case study

The NEXRAD wind retrievals from 2DVAR and VDRAS were evaluated using radar wind profiler data obtained during the 2003 Joint Urban (JU2003) field campaign in Oklahoma City, Oklahoma (Allwine et al. 2004). Our evaluation employs data between 3 and 14 July 2003 from the four radar wind profilers shown in Fig. 1a, along with the operational surface meteorological stations in the region, and radial velocities from the KTLX NEXRAD radar located ∼20 km southeast of Oklahoma City. The Argonne National Laboratory (ANL) profiler was located about 6 km north of downtown Oklahoma City and the Pacific Northwest National Laboratory (PNNL) profiler was located just south of downtown. The University of Oklahoma (OU) profiler was located in Norman, Oklahoma, in the vicinity of the National Weather Service (NWS) rawinsonde about 25 km south of downtown Oklahoma City. An existing radar wind profiler operated by the National Oceanic and Atmospheric Administration (NOAA) was located about 50 km south of Oklahoma City. The PNNL, ANL, and OU profilers measured 30-min-average wind speed and direction profiles up to ∼4 km AGL and the NOAA profiler measured hourly average wind speed and direction profiles up to ∼10 km AGL.

Prior to 10 July southeasterly-to-southwesterly ambient winds were usually observed over Oklahoma City during JU2003 as a result of the Bermuda high pressure system that was located over the southeastern United States. Downward momentum transfer resulted in southerly winds at the surface as well during most of the period (Allwine et al. 2004). Diurnal variations in wind speed and direction above the surface were largely the result of the dynamics associated with the nocturnal low-level jet (Bonner 1968). The meteorological conditions associated with the Bermuda high were disrupted only once during JU2003 as a result of a cold front that produced northerly winds and showers over the region on 10 July.

Both 2DVAR and VDRAS were set up to compute wind retrievals over an area centered over KTLX that encompassed Oklahoma City and the surrounding region using a grid spacing of ∼2 km, as shown in Figs. 1a,b, respectively. The spatial extent of the domains differs slightly; 2DVAR had 80 × 80 points over a 160 km × 160 km domain and VDRAS had 75 × 75 points over a 150 km × 150 km domain, and 2DVAR computed wind components along the radar elevation scans (Fig. 1c) while VDRAS computed wind components along constant-elevation surfaces (Fig. 1d).

b. Available wind retrievals

The lack of radar scans produced gaps in the wind retrievals during the 2-week period, but this was a small percentage of the total time as shown in Fig. 2. Even though both 2DVAR and VDRAS employ the same radar data, the data gaps are somewhat different, presumably because of the different QC techniques that the two models employ. For 2DVAR, winds are only obtained for those grid points with valid data. The fraction of valid data points in Fig. 2a reflects that the radar data quality varies in time and depends on the meteorological conditions and the scattering efficiency of the atmosphere. For this period, the fraction of data availability on the 2DVAR grid ranged from 0.35 to 0.85. Because VDRAS employs 4DVAR, it fills in information for all the grid points when pieces of the radar scans are missing so that the fraction of valid points for VDRAS is always 1.0 (Fig. 2b).

c. Computational requirements

The 2DVAR and VDRAS codes were run on a Red Hat Linux cluster that consisted of 2.4-GHz dual P4 processors, with Myrinet interconnect and 3 Gb of memory per node. The 2DVAR was compiled with version 6.0 of the Portland Group compiler and VDRAS was compiled using version 9.0 of the Intel compiler, but we expect differences in computer time not to be affected significantly by the compilers.

On average, 2DVAR retrievals were completed in ∼2 min while each retrieval from VDRAS was completed in ∼5 min. The combination of 2DVAR’s simpler statistical interpolation technique and computing retrievals only along the elevation scans results in less computer time than VDRAS. The computational time for both systems did not vary significantly during this period. Because a complete VCP is obtained in about 6 min, both codes can be run for real-time operational use.

a. Temporal and vertical variations

The radar wind profiler range gate closest to the wind retrieval levels was chosen for the evaluations of 2DVAR and VDRAS, rather than vertically interpolating either the wind retrievals to the profiler range gates or the profiler winds to the wind retrieval levels. For the PNNL, ANL, and OU profilers, the largest difference in the altitude of the wind profiler range gates and retrieval levels was 34 m. The difference at the NOAA profiler was as large as 146 and 97 m for 2DVAR and VDRAS, respectively. Both the radar wind profiler and the radar sample volumes of air, so that the height differences should not significantly affect the conclusions regarding the differences between measurements and wind retrievals. Radar wind profiler measurements also contain uncertainties up to ∼0.5 m s⁻¹ (Office of the Federal Coordinator for Meteorology Services and Supporting Research 1998).

An example time series comparing the wind retrievals and radar wind profiler measurements near the surface at the ANL site is shown in Fig. 3. The 2DVAR 0.5° level over the ANL site was closest to the 357-m range gate (Fig. 3a) and the lowest VDRAS level was closest to the 192-m range gate (Fig. 3b). The measurements show a diurnal variation in wind speed and direction, primarily from the development of a low-level jet nearly every night. Both wind retrieval techniques reproduced this feature, although the retrieval nighttime wind speeds were frequently higher than observed. The daytime values were usually closer to the profiler measurements, perhaps because of the stronger backscattering during the day. At times, the radar wind profiler measurements exhibited strong variations in wind direction when the wind speeds were the lowest. Neither wind retrieval method reproduced this feature.

An example of qualitative comparison at a higher altitude is shown in Fig. 4. The 2DVAR 2.5° level over the ANL site was closest to the 1897-m range gate (Fig. 4a), and the fifth VDRAS level was closest to the 1677-m range gate (Fig. 4b). At this altitude, which is above the nighttime low-level jets, the radar wind profiler measurements and wind retrievals indicate much smaller diurnal variations in speed and direction. Both retrievals produced speeds that were in better agreement with the data and directions that differed from the data more frequently than the values closer to the surface (Fig. 3).

The average values, bias, root-mean-square error (RMSE), index of agreement (IA), and correlation coefficient (R) that quantify the performance (Willmott 1982) of the wind retrievals are listed in Tables 1 and 2 for 2DVAR and VDRAS, respectively. To be consistent with the averaging period of the radar wind profiler measurements, 30-min averages of the wind retrievals were computed for ANL, PNNL, and OU sites and 1-h averages of the wind retrievals were computed for the NOAA site. The number of retrieval measurement pairs decreased with height because radar wind profiler data were not as complete aloft as they were near the surface, and the range of the wind retrievals from 2DVAR varied in time (Fig. 2a) and did not always extend to the profiler sites.

As shown in Table 1, the wind speed bias for 2DVAR at the ANL, PNNL, and OU sites ranged from −1.4 to 2.7 m s⁻¹ and the RMSE ranged from 2.4 to 4.7 m s⁻¹. While the bias at the NOAA profiler was only 0.7 m s⁻¹ near the surface, a bias of ∼5 m s⁻¹ and an RMSE of ∼8 m s⁻¹ was produced above 2800 m AGL. These altitudes were higher than those available from the other radar wind profilers. The wind direction bias at all profiler sites ranged from −16° to 7° and the RMSE ranged from 18° to 67°. The IA and R for both speed and direction that was often greater than 0.8 suggests that the wind retrieval produced temporal variations that were similar to the radar wind profiler measurements. In general, IA and R were lower at the higher altitudes, so that the wind retrieval had more skill near the surface than aloft. One reason for this is that radar beamwidth is narrower when the range gate is closer to the radar. The statistics at the OU profiler (closest to KTLX) were generally good at all altitudes because the radar elevation angles passed over this site within ∼1700 m of the ground. As shown in Table 2, the wind speed bias for VDRAS at all profilers ranged from −1.9 to 2.4 m s⁻¹ and the RMSE ranged from 2.2 to 3.8 m s⁻¹. The wind direction bias ranged from −23° to 18° and the RMSE ranged from 15° to 85°. As with 2DVAR, the IA and R for both wind speed and direction was often greater than 0.8 and the values near the surface were greater than those at higher altitudes.

To directly compare the statistics of 2DVAR and VDRAS on the same scale, the biases in the wind speed and direction are shown in Fig. 5. At the OU profiler, the wind speed and direction bias from 2DVAR was closer to zero than from VDRAS. At the PNNL and ANL sites, both retrievals usually overestimated the wind speeds near the surface and underestimated the wind speeds between 0.5 and 2 km AGL, but mean bias from 2DVAR was usually closer to zero. Farther aloft, both retrievals had higher wind speeds than the profiler measurements. At the NOAA site, the bias from VDRAS was closer to zero than from 2DVAR at all altitudes. While both retrievals produced wind speeds larger than the profiler measurements above 2.5 km AGL, the bias from 2DVAR was much larger than from VDRAS. Both retrievals produced a larger range in the bias for speed and direction at higher altitudes, indicating a larger uncertainty aloft.

The differences in the performance of the two retrieval techniques can be attributed to the different approaches employed by 2DVAR and VDRAS. In general, the performance of VDRAS was better aloft because it is based on a 3D mesoscale model that permits the dynamic and thermodynamic fields to adjust realistically. The first guess in 2DVAR is based solely on the VAD wind profile. Despite the differences in the level of complexity in the techniques employed by 2DVAR and VDRAS, both retrievals produced similar results within 1–2 km of the surface. While one model may have performed better than another at a specific place or time, one model did not outperform the other in terms of statistical measures.

b. Spatial variations

We also evaluated how well the wind retrievals reproduced the spatial variability among the radar wind profiler sites at three altitudes: 550, 900, and 1400 m. The radar wind profiler data were interpolated to the three altitudes and to the time of the retrieved wind field. Then, the retrieved wind profile, nearest to the radar wind profilers, was interpolated to the three altitudes.

An example of the u and υ wind component differences between the PNNL and NOAA sites at the 550-m level is shown in Fig. 6 for the 30-min-averaged data over the 2-week period. Differences in the profiler wind components between the sites were usually within 3 m s⁻¹. The differences from 2DVAR and VDRAS usually did not follow the temporal variations in the wind profiler differences when the wind profiler differences were less than 3 m s⁻¹. There were a few instances in which the radar wind profilers indicated strong horizontal gradients (e.g., 10, 12, and 13 July). On 10 July, a front moved through the region that produced differences between the PNNL and NOAA sites as large as 10 m s⁻¹. The wind retrievals from 2DVAR and VDRAS both produced larger spatial variations, and their temporal evolution was not exactly the same as the measurements. In general, the peak differences in the wind components were larger than observed, especially after 10 July. The 2DVAR produced larger peak spatial variations in both the u and v components than VDRAS.

The correlation coefficients between the measured and retrieved velocity differences (Table 3) are usually higher with larger separations, that is, between PNNL and NOAA or between ANL and NOAA (Fig. 1a). By contrast, the distance between the PNNL and ANL profilers is only ∼7 km. The observed velocity differences between the PNNL and ANL sites were very noisy, and thus the retrieved and observed velocity differences are poorly correlated. One would expect the velocity differences between the sites to be small (especially aloft) because of their proximity. VDRAS performed somewhat better than 2DVAR in terms of representing the spatial wind variations; however, the correlation coefficients from both methods were low.

a. Evaluation of CALMET winds

Before examining the effect of the six cases on the predicted dispersion, we first present comparisons of the observed winds and the winds derived from CALMET. As with any diagnostic model, interpolation techniques and adjustment for mass conservation will result in wind fields that are similar to, but not exactly the same as, the input data. Figure 7 shows an example of the CALMET wind fields for cases 1–4 on 5 July compared to the ANL wind profiler data interpolated to the CALMET vertical levels.

For case 1 (Fig. 7a), the simulated winds from CALMET were qualitatively similar to the radar wind profiler because the observed winds did not change significantly on this day. Therefore, the temporal interpolation of the twice-daily rawinsonde wind profiles was a reasonable approximation of the observed variation. However, the simulated wind speeds for the low-level jet during the morning before 0600 LST were lower than those observed, and the directions aloft differed by as much as 50°. For case 2 (Fig. 7b), the CALMET winds closely followed the profiler winds as expected because those winds were employed by CALMET, indicating that the interpolation and mass adjustment did not lead to a wind field that deviated significantly from the observed winds. For cases 3 and 4 (Figs. 7c,d) the wind speeds and directions were close to the measurements within 2 km AGL and are similar to case 2, suggesting that the wind retrievals provided information comparable to the radar wind profiler. However, the wind speeds and directions above 2 km AGL were often very different from the profiler.

A second example, shown in Fig. 8, is for 10 July when the observed winds were stronger and a front moved through the region so that the westerly winds during the morning shifted to northeasterly by the afternoon. In contrast to Fig. 7a, the CALMET and observed winds aloft were very different for case 1 (Fig. 8a) because the temporal interpolation did not work well for the rapidly changing meteorological conditions associated with the front. The simulated winds were in better agreement with the profiler data close to the surface because CALMET employed the hourly surface observations. As expected, a good agreement between the observed and simulated winds was produced for case 2 (Fig. 8b). The CALMET fields that employed the wind retrievals for cases 3 and 4 (Fig. 8c,d) were better than those from case 1, but relatively large errors in both speed and direction still occurred at the time of the frontal passage. On this day, VDRAS produced better winds above 2 km AGL than 2DVAR.

A stationary front was located north of Oklahoma City on 12 July for the third example shown in Fig. 9. The behavior of the CALMET wind fields was similar to the 10 July period (Fig. 8). While the observed upper-air directions remained westerly throughout the period, the wind speeds decreased so that the temporal interpolation between the twice-daily rawinsonde profiles did not work well (Fig. 9a). The cases that employed the NEXRAD wind retrievals (Figs. 9c,d) were better than case 1, but not as good as case 2 (Fig. 9b).

Cases 5 and 6 produced winds that were very similar to cases 3 and 4 in general (not shown), and the incorporation of radar wind profiler data did not reduce the large differences between the observed and simulated winds above 2 km AGL. The similarity between cases 5 and 6 with cases 3 and 4 occurred because CALMET weights all of the input data equally, and there were far more wind retrieval profiles near the radar wind profiler site.

For case 1, which employed only the standard observations, the wind speed bias, RMSE, IA, and R over all altitudes and sites range from −1.5 to 0.99, from 1.38 to 3.91, from 0.76 to 0.91, and from 0.66 to 0.88, respectively. As with the graphical depictions, case 2, which employed the radar wind profiler data, had the best performance, as expected. The index of agreement and correlation coefficient was usually close to 1, and biases and root-mean-square errors were usually less than a few tenths of a meter per second within 2 km of the ground. The statistics for cases 3–6 that include the NEXRAD wind retrievals were better than that from case 1. The differences in the statistics between cases 3 and 5 and between cases 4 and 6 were small, but the statistics for cases 5 and 6, which included the radar wind profiler data, were slightly better.

Because of the different vertical coordinates employed by 2DVAR and VDRAS (Figs. 1c,d), it was also useful to identify problems in the retrievals by directly comparing the horizontally varying wind fields generated by CALMET. An example when the wind field derived from 2DVAR varied more spatially than VDRAS is shown in Fig. 10 at 0400 LST (1000 UTC) 5 July. Both models produced higher wind speeds at 450 m AGL (Figs. 10a,b) than aloft at 2600 m AGL (Figs. 10c,d) as a result of the nocturnal low-level jet. Wind speeds over the domain at 450 m AGL ranged from 3 to 10 m s⁻¹ from 2DVAR and from 7 to 10 m s⁻¹ for VDRAS. The locations of the peak wind speeds were different as well; however, the wind speed and direction at the profiler sites were quite similar. The wind fields farther aloft at 2600 m AGL were qualitatively similar, except that the wind speeds derived from VDRAS were 1–2 m s⁻¹ higher than from 2DVAR.

Figure 11 shows the winds 6 h later at 1000 LST (1600 UTC) 5 July. At this time, the winds at 450 m AGL (Figs. 11a,b) between 2DVAR and VDRAS were quite similar. In contrast with Fig. 10, spatial variations from 2DVAR were only ∼1 m s⁻¹ greater than from the VDRAS at this time. The shift from large to small spatial wind speed variations between 0400 and 1000 LST was also evident at the ANL profiler for the time series shown in Fig. 6. The quasi-symmetrical perturbation pattern in the wind speed and direction at 2600 m AGL (Fig. 11c) is an example of an artifact produced by 2DVAR that was usually produced only at high altitudes. This artifact also explains the larger errors in 2DVAR wind speeds and directions when compared with the radar wind profiler data above 2 km AGL (Table 1). The 2DVAR wind field at this time was unrealistic because there were no sudden shifts in the ambient synoptic conditions, and one would expect horizontally homogenous winds aloft like the wind field produced from VDRAS (Fig. 11d). A rapid decrease in the available data occurred between 0400 and 1000 LST 5 July, as shown in Fig. 2; therefore, the variational approach in 2DVAR may produce unrealistic results when relatively few NEXRAD radial velocity data points are available.

b. CALPUFF sensitivity simulations

CALPUFF simulations were performed based on the six CALMET wind field cases. Instead of using the JU2003 tracer experiment release site and intermittent release periods, we employed release characteristics in the model that would better illustrate the differences between the wind fields. A constant release rate (10 g s⁻¹) of material between 0300 and 1500 LST was made near the OU profiler site (Fig. 1) from a stack. The release height, exit velocity, and temperature were 40 m AGL, 5 m s⁻¹, and 355 K, respectively. This site was chosen because the southerly-to-southwesterly winds usually observed during the 2-week period would transport the plume through a large fraction of the modeling domain. Time-integrated concentrations (i.e., exposure) defined as the time integral of the surface concentration at a point, plume centerline locations, and plume widths were computed to illustrate differences in dispersion.

Figure 12 shows the 12-h-accumulated concentration of the predicted plumes on 5 July for cases 1–4. The plume for case 1 was transported to the northeast during the period (Fig. 12a). When the radar wind profiler data were included in CALMET (Fig. 12b), the plume was transported farther to the east, and the footprint’s bifurcation indicates a wind shift during the period that transported material toward the north for a short period of time. The northward transport occurred between 1130 and 1230 LST when winds from case 1 were southwesterly (Fig. 7a) and the observed winds and winds from case 2 were southerly (Fig. 7b). The wind retrievals in CALMET resulted in bifurcations in the footprints for cases 3 and 4 (Figs. 12c,d) that were qualitatively similar to the footprint from case 2. The plume centerlines (defined in the appendix) from cases 2, 3, and 4 differed by as much as 10° from case 1, which included only the standard observations. The differences in the plume width (defined in the appendix) among the cases were small near the release site and became as large as 2 km at 50 km downwind of the release site.

The dispersion of the downwind plume during the frontal passage on 10 July is shown in Fig. 13. The footprint of the plume from case 1 (Fig. 13a) shows that the variable winds transported the tracer over a wide region toward the east and south. Before 0900 LST, the simulated winds were southwesterly so that the plume was transported toward the northeast. As the front moved through the region, new material, as well as previously released material over the northeastern portion of the domain, was transported to the south and southwest. The footprint from case 2 (Fig. 13b) was remarkably similar to that of case 1, even though there were large differences in the wind fields as shown in Figs. 8a,b, suggesting that transport was governed more by the winds within a few hundred meters of the surface than the winds aloft. After a frontal passage, the air would likely be more stable and would result in less coupling with the wind aloft. The simulated winds from 0900 to 1500 LST from case 2 better represented the northeasterly near-surface winds and transported more material farther to the southwest than that from case 1. Cases 3 and 4 produced footprints (Figs. 13c,d) that were qualitatively similar to both cases 1 and 2; however, the 2DVAR wind fields in case 3 produced some transport to the north that was not seen in cases 1, 2, and 4. The plume centerline from cases 3 and 4 with the wind retrievals was located farther to the south than cases 1 and 2, especially near the release site. In contrast to 5 July, differences in the plume width among cases 1–4 on 10 July were much larger—as much as 8 km at 22 km downwind of the release site.

Figure 14 shows the footprint of the plume predicted on 12 July. The stronger and near-constant wind directions in the CALMET fields using the standard observations in case 1 (Fig. 14a) produced a downwind plume that was straight and narrow. The footprint produced by case 2 (Fig. 14b) was wider as a result of transport to the east from 0700 to 1200 LST (Fig. 9b). While cases 3 and 4 also produced downwind transport toward the northeast (Figs. 14c,d), both cases had more meandering of the plume and a distinct bifurcation in the plume with the smaller branch transported to the east. On this day, the plume footprint using the wind retrievals was significantly different than case 2, which employed the radar wind profiler data. Cases 3 and 4 were more similar to one another than to case 2. The centerline of the plume predicted by cases 3 and 4 were again farther south of cases 1 and 2 on 12 July. On this day, the plume width from case 1 was narrow and only 3 km wide, and 50 km downwind of the release site. However, the plume widths from the other cases were much larger (between 6 and 13 km).

In general, the shapes of the plume from 2DVAR and VDRAS differed somewhat, but the results suggest that the choice of wind retrieval would not significantly affect the predicted dispersion for the meteorological during JU2003. The plume footprint from cases 5 and 6 were similar to those of cases 3 and 4 for all the dispersion runs (not shown) because the wind fields in CALMET were also very similar. The similarity of 2DVAR and VDRAS in terms of dispersion is also based on the predicted exposure for other days not shown in Figs. 12 –14. There was only one day—10 July (Fig. 13)—for which the exposure using the VDRAS wind retrievals performed better than 2DVAR. As discussed previously, the mesoscale model in VDRAS is able to better represent the complex winds, and thus the dispersion, associated with the frontal passage on that day. Following this line of reasoning, VDRAS may be more appropriate to simulate dispersion for other sites and time periods with rapidly changing meteorological conditions; however, additional tests comparing VDRAS and 2DVAR are necessary for verification.

For all of the CALPUFF simulations, the plume widths from case 1 were narrower than those from case 2, which included radar wind profiler data in addition to the standard observations. The inclusion of hourly wind profiles introduced more temporal variability into the CALMET wind fields that spread out the plume. The plume widths were even wider for cases 3 and 4, demonstrating that the additional spatial variations in the winds derived from the NEXRAD data resulted in increased dispersion downwind of the release site that was not produced by inclusion of the radar wind profiler data alone in CALMET.

The way in which the NEXRAD retrievals increase the downwind dispersion of a plume is reasonable, but would need a tracer experiment to adequately quantify the improvement in the representation of dispersion. The SF₆ tracer experiment conducted during JU2003 obtained measurements only up to 5 km downwind of the downtown release site, and therefore would not be suitable to examine how the spatial variations in the NEXRAD wind retrievals would affect dispersion.

Other parameters that drive ADMs, such as temperature gradients, mixing depth, and turbulence, are often not observed at sufficient spatial and temporal scales near the surface and within the boundary layer. Uncertainties in these quantities will also affect the spatial distribution and concentration of predicted atmospheric contaminates.

c. Effect of temporal resolution of meteorological inputs

To be consistent with most operational dispersion model simulations, CALMET wind fields were produced at hourly intervals, and wind retrievals were averaged over each 1-h period. The NEXRAD wind retrievals can be computed in near–real time; therefore, it is possible to obtain wind retrievals at a rate of approximately once every 10 min.

To examine the effects of the time-varying NEXRAD wind retrievals on the simulated dispersion patterns, the CALPUFF simulations were repeated for 5, 10, and 12 July, which were based on CALMET wind fields at 15-min intervals. The spatial distribution and magnitude of the predicted surface instantaneous concentrations and exposure that employed the 15-min wind fields were very similar to the 1-h wind fields. The difference in the maximum exposure was less than ∼6%. The similar dispersion patterns are consistent with Figs. 3 and 4, which show gradual changes in wind speed and direction from both wind retrievals. The number of periods with rapidly changing winds over subhourly intervals was relatively few for the present cases. Nevertheless, there are likely to be meteorological conditions in which subhourly wind fields would significantly improve the transport and mixing of atmospheric contaminants.