Evaluation of a data assimilation technique for a mesoscale meteorological model used for air quality modeling

The magazine publisher is the copyright holder of this article and it is reproduced with permission. Further reproduction of this article in violation of the copyright is prohibited. To contact the publisher: http://www.ametsoc.org/AMS/

ABSTRACT
An observational data assimilation (ODA) technique was evaluated based on both its direct effect on meteorological model fields and its indirect effect on the results of two air quality models that input these meteorological fields: a Lagrangian particle model (LPM) and a photochemical model, the variable-grid version of the Urban Airshed Model (UAM-V). The purpose was to investigate the model performance improvements that are derived from using field-study observations with an ODA technique. The ODA technique, based upon Newtonian relaxation, was incorporated into the Colorado State University Regional Atmospheric Modeling System (RAMS). The technique was applied with rawinsonde, profiler, and sodar observations of winds, temperature, and moisture from an intensive field campaign during 3-7 August 1990 over the San Joaquin Valley in California. The RAMS meteorological fields, produced with and without the use of the ODA technique, and the results from the two air quality models using these two fields were evaluated and compared. The use of the ODA technique substantially reduced the gross errors of RAMS upper-air parameters but only produced minor reductions in the gross errors of RAMS surface-level parameters. The respective gross errors of RAMS upperair results with and without ODA were 0.23 and 1.1 m s[sup-1] for wind speeds and 1.1 and 1.9 K for temperatures. For RAMS surface-level results, the respective gross errors with and without ODA were 0.76 and 0.78 m s[sup-1] for wind speeds and 3.4 and 3.8 K for temperatures. In both cases, the RAMS vector wind biases near the surface and aloft were less than 0.9 m s[sup-1]. Both LPM simulations of the field-study tracer experiments produced particle distributions that were consistent with observations and that were similar to each other. However, the more accurate vertical temperature structure due to the use of ODA produced shallower planetary boundary layers and resulted in larger surface tracer concentrations. Both UAM-V simulations of ozone produced similar ozone results, with less than 22% normalized gross mean errors for observations greater than 40 ppb on the last two days of the simulations. The strong influence of the UAM-V boundary-condition values on the sensitivity of UAM-V to the two meteorological inputs is demonstrated. This influence suggests that errors in other UAM-V inputs may obscure improvements in ozone modeling results from the use of ODA in creating meteorological inputs.

1. INTRODUCTION
    The application of prognostic mesoscale meteorological models to provide wind fields for the study of regional tropospheric ozone is becoming routine (Seaman 2000). The calculations performed by these models satisfy momentum and energy-balance relationships and account for physical phenomena, such as moisture phase changes and radiative transport. In theory, these models are capable of producing results that are more realistic than those obtained by simply interpolating observations. However, models are only approximate descriptions of physical processes, and model inputs, such as initial and boundary conditions, are also approximations. Models consequently do not replicate observations exactly. Agreement between model solutions and observations can be increased by a data assimilation technique that adjusts model solutions to be more consistent with observations. A simple and computationally inexpensive technique based upon Newtonian relaxation has been under development over the past several years (Stauffer and Seaman 1994). Its effects on model solutions have been examined for different meteorological situations, such as nocturnal drainage flows (Fast 1995) and high-tropospheric-ozone episodes (Seaman et al. 1995). Tanrikulu et al. (2000) found that data assimilation improved performance in a meteorological model and a subsequent air quality model. Improvements in model performance due to this technique obviously depend on the original accuracy of these models. In addition, the measurement of their effect on subsequent air quality models may be obscured by inaccuracies in other factors, such as the specification of boundary conditions or pollutant emissions. Hence, the evaluation of the significance of this technique to the performance of the meteorological-air quality modeling system needs to consider other influential modeling factors. Such a broad evaluation would be useful in the design of field programs that are used for air quality modeling studies. Hence, we extend the investigation to consider the effects of meteorological data assimilation on the results of a meteorological-air quality modeling system. Because most surface observations are available routinely, we focus on the data assimilation of upper-air observations, which are only available in a spatially dense network during field studies. The objectives are 1) to examine the effects of data assimilation on the results of the prognostic meteorological model, 2) to evaluate the sensitivity of air quality models to meteorological inputs created with or without data assimilation, and 3) to contrast the sensitivity of the air quality models to the effects of meteorological data assimilation with their sensitivity to other air quality modeling factors.
    The period studied was 0000 UTC 3 August to 1200 UTC 7 August 1990, or 1600 LST 2 August to 0400 LST 7 August 1990 (Pacific standard time). This period is similar to one used by Seaman et al. (1995) and Tanrikulu et al. (2000) in their evaluation of a data assimilation technique in the Pennsylvania State University-National Center for Atmospheric Research (PSU-NCAR) Fifth-Generation Mesoscale Model. During this period, the state government of California together with other public and private parties conducted a collaborative field program known as the San Joaquin Valley Air Quality Study-Atmospheric Utility Signatures, Predictions, and Experiments Regional Model Adaptation Project (SARMAP; Ranzieri and Thuillier 1994) to observe the formation and transport of ozone in and around the San Joaquin Valley and to study its relationship with the region's meteorological conditions and pollutant emissions. The data collected contain meteorological observations from 165 surface and 51 upper-air sites and air quality observations from 150 surface sites. Nine aircraft sampled meteorological and air quality parameters aloft. A ground-based tracer experiment was conducted during the first two days of this period with tracer releases from two upwind locations in the northern part of the study domain near the San Francisco Bay Area.
    The observations assimilated into the meteorological model consisted of all rawinsonde, sodar, and profiler observations of winds, temperature, and moisture. The surface and aircraft observations are used as an independent dataset to measure model performance and the effectiveness of the data assimilation technique.
    Section 2 describes the meteorological-photochemical modeling system. Section 3 summarizes the meteorological conditions and the air quality of this study period. Section 4 presents our experimental design. Model performance and sensitivity results for each of the models are described in section 5.

2. DESCRIPTION OF THE MODELING SYSTEM

A. METEOROLOGICAL MODEL
    The Colorado State University Regional Atmospheric Modeling System, version 3a (RAMS), a nonhydrostatic mesoscale meteorological model described by Pielke et al. (1992), was used in this study. RAMS prognostic equations predict east-west, north-south, and vertical winds (U, V, and W), temperatures, moisture levels, and turbulent kinetic energies (TKE). The prognostic second-order TKE level-2.5 scheme by Mellor and Yamada (1974) was selected to parameterize subgrid eddy diffusion processes. Precipitation as rain provided a sink for atmospheric moisture and was necessary to maintain realistic moisture levels in the surface stratus layer over the ocean and coastal regions. The longwave and shortwave radiation schemes found in version 3b of RAMS and described by Chen and Cotton (1987) were installed into our 3a version of RAMS.
    The RAMS grid configuration consists of coarse, medium, and fine interactive nested grids of 80-, 16-, and 4-km resolution and domain dimensions of 32 × 32, 64 × 72, and 114 × 158 grid cells, respectively. The outermost domain extends from the Rocky Mountains to 1200 km offshore and from the Canadian border to the southern tip of Baja California. The fine grid (see Fig. 1) includes the moutain ranges surrounding the San Joaquin Valley, a feature that strongly influences low-level circulations in the study area. RAMS uses a terrain-following sigma[subz] coordinate system. The thickness of the surface grid is set to 25 m, and each upper-level grid thickness is 1.25 times the thickness of the grid below, with a maximum thickness of 4 km. This configuration yields 11 layers in the lowest 1 km and places the top of the model at a height of about 20 km in 25 layers.
    Synoptic analyses from the National Centers for Environmental Prediction (NCEP) were interpolated isentropically to all RAMS grids to form background fields that are used to initialize and to specify boundary values for RAMS. Two assimilation methods were evaluated in this study. One method, the nonobservational data assimilation (NODA) method, relaxes, or nudges, model solutions to the background synoptic analyses only. The other method, the observational data assimilation (ODA) method, relaxes model solutions to the background synoptic analyses and to local upper-air observations.

1) ASSIMILATION OF SYNOPTIC ANALYSES
    Temporally interpolated synoptic-analyses background fields are assimilated into RAMS solutions by using weights that vary spatially:
    Weight(x, y, z)
    = lambda[sup-1](z / z[subtop])[1 - gamma[sup2] sin[sup2](pix / L[subx]) sin[sup2](piy / L[suby])], (1)
    where lambda is the minimum e-folding time for background tendencies or the time for the difference between a parameter and its background value to decrease by e[sup-1] (in this study lambda = 333 s); z is the height above ground level; z[subtop] is the height of the top of the model; gamma is 0.97, which provided a reasonable constraint of the solution near the model boundaries while allowing freedom of the solution in the model interior; x is the X grid location relative to coarsest grid (0 - L[subx]); L[subx] is the X width of the coarsest grid; y is the Y grid location relative to coarsest grid (0 - L[suby]); and L[suby] is the Y width of the coarsest grid. This weight structure has the following properties:
    1) It reduces the influence of the background field near the surface where synoptic analyses are least accurate and synoptic models are least compatible with our application of RAMS. For example, the terrain in RAMS is more detailed than in the NCEP synoptic analysis model.
    2) It reduces the influence of the background field on the interior of the model where it is desirable to allow the model to generate mesoscale features.
    3) It smoothly varies background field influences to reduce spurious model solutions caused by abrupt spatial changes in the background tendency terms.
    This weight scheme and its objectives are similar to one reported by von Storch et al. (2000).

2) ASSIMILATION OF OBSERVATIONS
    The goal of assimilating observations into model simulations is not only to improve the agreement with the data assimilated but also to improve agreement with observations in the surrounding regions by propagating the assimilated information through physically consistent model formulations. The data assimilation technique implemented into RAMS is similar to techniques successfully applied by Seaman et al. (1995) and Fast (1995). This technique relaxes, or nudges, model values toward observations by nonphysical tendency terms added to the model equations. The observational data assimilation tendency terms are of the form
    [df] / [df]tT[submodel](x, t, x[sub0], t[sub0]) = alpha[T[subobs](x[sub0], t[sub0]) - T[submodel](x, t)], (2)
    where T[submodel] (x, t) is the model parameter at grid position x and time t; T[subobs](x[sub0], t[sub0]) is the observed parameter at grid position x[sub0] and time t[sub0]; and alpha is the product of spatial, stability, temporal, and inverse e-folding-time factors.
    The spatial factor is given by [1 + beta[sup2](x - x[sub0])[sup2]][sup-1], where beta has units of inverse grid interval, and is set to 10/16 for potential temperature and to 10/8 for winds and moisture. These values for beta smoothly limit the perceptible range of influence of each data assimilation observation to within a radius of 16 grid cells for potential temperature and 8 grid cells for winds and moisture. Figure 2 shows the spatial distribution of alpha for potential temperature and shows that the data assimilation technique is most influential within the San Joaquin Valley and San Francisco Bay Area. To maintain thermal stability on sloping terrain, the tendencies are also scaled by (1 + deltaT[sup2]/2)[sup-1], where deltaT (K) is the difference between the assimilated observation and model potential temperature. The SARMAP rawinsonde observations of wind, temperature, and moisture were taken at 3-h intervals, covered the fine-grid modeling domain well, and were representative of moderate spatial extent. They were better suited for use in the data assimilation technique than were the profiler or sodar observations, which only measured winds, or the surface observations, which are generally representative of small spatial extents. Therefore, our technique was configured to be consistent with the rawinsonde observations. The temporal factors corresponded to linear interpolation between two observations spaced 3 h apart. The e-folding-time factor for alpha was set to 3 h. Furthermore, the computational burden of the data assimilation scheme was reduced by averaging all the observations that were contained in a single model grid cell and in the 3-h window surrounding the rawinsonde times. This consolidation reduced the unevenness in the weighting of the assimilated observations caused by variations in observational density while retaining the rawinsonde temporal details. However, some details of the 1-h profiler and sodar observations were lost.
    Surface observations were not assimilated, because they are generally representative of small, sometimes subgrid-length scales and may be incompatible with fine-grid model variables. Furthermore, a goal of this study is to determine the influence of the ODA method using field-campaign observations on meteorological and air quality models. Because surface observations are routinely available, only upper-air observations were used in the ODA method.
    To increase the effect of ODA, background-analysis field weights were scaled by 0.1 for locations more than four grid spaces interior to the coarse-grid boundaries. No remarkable effects appeared from this reduction.

B. LAGRANGIAN PARTICLE MODEL
    A Lagrangian particle model (LPM) simulates particles transported by time-varying three-dimensional advection and eddy-diffusion fields and is used to investigate transport properties of meteorological fields. Simulated winds and subgrid diffusion properties are interpolated to the location of each particle. Diffusive transport is determined from simulated TKE with the turbulent velocities being partitioned into the three spatial components based upon local atmospheric stability. The evolution of the turbulent velocity for each particle is governed by the Langevin equation (Reif 1965, p. 560) in which the new turbulent velocity is equal to the current turbulent velocity plus a random velocity. The random velocity is the product of the variance of the turbulent velocity and a computer-generated random number with Gaussian distribution. The particle's net velocity v is then the sum of the advective and the turbulent velocities. The particle's change in position over a time interval dt is given by vdt.
    The surface concentrations are estimated from an algorithm based upon local particle density relative to the distribution of all simulated particles. The fractional concentration (with units of "per volume") from a distribution of LPM particles at a model location x, y, and z is given by
    Concentration(x, y, z) = [Graphic Character Omitted]exp(-R[sup2[sub[subi]) / Npi[sup3/2]Vgamma[sup3],
    where N is the number of particles in the simulation, R[sup2[sub[subi] = R[sup2[sub[subxi] + R[sup2[sub[subYi] + R[sup2[sub[subZi], R[subXi] = deltaX[subi]/{Begin Greek}gD{End Greek}X, deltaX[subi] is the X distance of a particle to x, DeltaX is the maximum extent in the X direction of all particles, R[subYi] and R[subZi] = (X --> Y) and (X --> Z) in the three equations above, V = DeltaXDeltaYDeltaZ, and gamma is the range of influence in terms of fraction of the extent of the particle distribution (currently set to 0.1).

C. PHOTOCHEMICAL MODEL
    The variable-grid version of the Urban Airshed Model (UAM-V) was used to simulate the photochemistry of this episode. UAM-V uses the Carbon Bond Mechanism, version 4 (CB-IV; Gery et al. 1989; Dodge 1989; SAI 1993), to represent atmospheric photochemical reactions. The UAM-V grid was a 110 × 128 × 16 subset of the RAMS fine grid. The horizontal extent of the UAM-V domain is shown in Fig. 1.
    The UAM-V model was initialized from and its air quality boundary conditions (BC) were derived from a sparse collection of SARMAP surface and aircraft air quality observations. Locations far from observations were assigned concentration values used in previous modeling of this episode (DaMassa et al. 1996).
    Most of the UAM-V emissions inputs were obtained from the 3-6 August 1990 SARMAP modeling emissions inventory dataset developed by the California Air Resources Board (CARB; Magliano 1996). The extreme northern and southern areas were not available and were supplemented by county-total emissions data developed by CARB's emissions inventory group.

3. EXPERIMENTAL DESIGN
    The effects of ODA on meteorological and air quality model results were investigated by comparing results from ODA and NODA RAMS simulations and subsequent LPM and UAM-V simulations of the 2-7 August 1990 period. The ODA method used wind, temperature, and moisture observations from 28 rawinsonde soundings and wind observations from 12 Doppler sodar and 12 radar profilers. The minimum ODA e-folding time was selected to be 3 h for these observations. The twice-daily National Weather Service rawinsonde network observations were also assimilated into the medium and coarse grids with minimum e-folding times of 3 and 12 h, respectively.
    In both cases, RAMS was initialized with the 0000 UTC 3 August 1990 NCEP analysis and was run for 108 h to 1200 UTC 7 August 1990. Background fields, which are used for initial and boundary values, were developed from the twice-daily NCEP Global Tropospheric Analyses and the Nested-Grid Model North America Analyses, both available from NCAR.
    Temperature, wind, and moisture values from both RAMS simulations were compared with observations to determine the effects of ODA on RAMS performance. Then, both meteorological fields were used in the LPM and the UAM-V photochemical model to determine the effects of using ODA to generate meteorological inputs on each of these models. Differences between the LPM particle distributions in the simulations of the SARMAP tracer experiment (Tracer Technologies 1991) were used to evaluate the difference in accuracy of transport and the source-receptor relationships defined by the two meteorological fields. Differences in the UAM-V model results of SARMAP air quality observations were used to evaluate the influence of ODA on UAM-V model performance. Photochemical models are sensitive to variations in other inputs such as pollutant emissions and chemical boundary values. These inputs are often highly uncertain, and errors in them can affect the efficacy of ODA. Such an influence is demonstrated by comparing the efficacy of ODA on UAM-V results that used clean BCs and results that used the original observationally based BCs.
    The period for all UAM-V simulations was from 0400 LST 3 August to 2000 LST 6 August 1990, but statistical evaluations were based on the last two days of the simulation, minimizing the influence of initial conditions and other model spinup issues.

4. DESCRIPTION OF THE 3-7 AUGUST 1990 PERIOD
    During the period of 3-7 August 1990, a high pressure region moved onshore and strengthened over California and Nevada. Figure 3 shows the 500-hPa heights and winds from the NCEP Global Tropospheric Analyses interpolated to the RAMS coarse-grid domain for 0000 UTC 5 August 1990. The resulting anticyclonic flow tended to bring warmer continental air off the high desert areas cast of the Sierra Nevada down over California, strengthening the inversions and lowering mixing heights. The warming over California caused by this synoptic flow pattern is consistent with the observation of increasing temperatures aloft shown in Fig. 4. A detailed description of the meteorological conditions associated with this SARMAP episode can be found in Seaman et al. (1995).
    The synoptic meteorological conditions are conductive to local stagnation conditions needed for air pollution episodes in California. Figure 5 shows that average surface ozone observations (see Fig. 1 for station locations) did increase over the 4-day episode and that the daily maximum average ozone values (solid line) increased by approximately 5 ppb day[sup-1].
    In addition, the spatial extent of sites within the study area that exceeded the California 1-h ozone standard (90 ppb) broadened as the episode progressed. The highest observed ozone values occurred in the San Joaquin Valley and near Sacramento. The evolution of ozone and its precursors during this period is summarized by DaMassa et al. (1996).

5. RESULTS

A. EFFECTS ON THE METEOROLOGICAL MODEL
    We evaluated the effects of ODA on RAMS results for the 3-7 August 1990 period. RAMS-simulated values were compared statistically with observations (see Table 1 for the definitions of the statistical measures). Observations that were contained within a single model grid cell were averaged to a single value for comparison with the modeled value. Modeled parameter values were interpolated to the location of the observation. The parameter values of lowest model level were used in situations in which observation locations were below the lowest model level. The vertical thickness of the lowest model level is 25 m, which corresponds to a nominal mean parameter height of 12 m above ground level (AGL). Wind observations, taken generally at 10 m AGL, are comparable to model winds. Temperatures and moisture observations, which generally are taken at lower heights (typically 2 m AGL), may not be entirely comparable to model parameters, which represent values at 12 m AGL. However, no adjustments were made to account for these kinds of incomparability. Furthermore, the conclusions of this study obviously relate more to data-dense areas, such as the San Joaquin Valley and urban coastal areas, than the data-lean areas, such as the Sierra Nevada and the remote coastal areas (see distribution of surface observations in Fig. 2).

1) WINDS
    Wind comparison statistics for surface and upper-air observations from both meteorological cases are tabulated in Table 2. For the NODA case, surface and upper-air wind biases are low, less than 1 m s[sup-1] in absolute value, indicating that there are no significant imbalances between the model configuration and the synoptic background. With the addition of ODA, upper-air vector wind difference (VWD) decreased significantly from 4.31 to 2.34 m s[sup-1], whereas surface VWD decreased from 2.75 to 2.67 m s[sup-1]. The small decrease in surface VWD suggests that surface winds are influenced less by upper-air wind ODA than by surface forces. Seaman et al. (1995) and Fast (1995) observed similar decreases in upper-air VWD with their data assimilation technique. Fast (1995), in his control, or no-data assimilation, case, obtained surface VWD similar to our NODA results, but, in his data assimilation case in which he assimilated surface wind observations, he obtained substantial decreases in surface VWD.
    The vector wind biases and wind speed gross errors for each surface station for the ODA case are shown in Fig. 6. The results for the NODA case (plot not shown) are similar to the ODA case. In both cases, vector wind biases and gross errors greater than 4 m s[sup-1] are found near the San Francisco Bay. Inspection of the hourly vector winds indicates that these localized biases probably are due to lower-than-observed simulations of seabreeze wind speeds. The sea breeze in this area is channeled by the coastal hills and shallow marine boundary layer through the natural opening at the San Francisco Bay. This channeling produces higher wind speeds than the RAMS model, with its 4-km grid spacing, is able to reproduce.

2) TEMPERATURES
    Comparative temperature statistics for both meteorological cases are summarized in Table 2. The correlation coefficients for both cases were relatively high: 0.88 for the NODA case and 0.91 for the ODA case. However, the bias, gross error, and rms difference statistics show that the ODA case agreed noticeably better with observations.
    Hourly surface temperature biases, shown in Fig. 7, generally increase over the simulation period. This increasing trend likely is due to a tendency for the surface soil layer to dry out from its initial value and an apparent positive temperature bias in the synoptic analyses relative to the SARMAP rawinsonde observations (discussed below). The bias histories for the NODA case and the ODA case are similar, but the NODA biases are slightly higher.
    The temperature biases at ground-level sites are shown in Figs. 8 and 9 for the ODA and NODA cases, respectively. The biases of the ODA case are about 5 K in the inland areas and become smaller along the coastal areas. The biases of the NODA case tend to be similar in the inland areas but remain large over the coastal areas. ODA produced noticeable reductions in temperature biases and gross errors in the coastal and San Francisco Bay sites and, surprisingly, produced the largest gross error reductions at the buoy sites that are distant from the assimilated observations. This result likely is due to the lower westerly wind bias for the ODA case (see U bias statistics in Table 2), which allows further advancement of the cooler marine boundary layer and limits the excursions of warmer land breezes over the ocean.
    The general positive bias of surface temperatures is a concern for RAMS model performance. Temperature gradients have a major influence upon winds and vertical mixing. Although the surface wind biases were not significant in either case, the issue requires further investigation.

3) MOISTURE
    There were only 217 surface moisture observations taken over the 108-h simulation period within the fine-grid modeling domain. Simulated mixing ratios are in reasonable agreement with observations in both cases (see Table 2). The slightly lower surface moisture bias in the ODA case is due to moisture added by ODA using the rawinsonde data. ODA significantly improved upper-air moisture statistics. Similar results were reported by Tanrikulu et al. (2000).
    The incomparability of moisture values between the typical observation height (2 m AGL) and the model parameter height (12 m AGL) may account for some of the negative bias. For this summer episode, soil and vegetation are the primary land sources of surface-level atmospheric moisture; therefore, 2-m-AGL observational values are expected to be greater than 12-m-AGL modeled values. However, no adjustments for the height differences were made in this comparison.

4) UPPER-AIR OBSERVATIONS
    We examine the effect of ODA on the average vertical structure of RAMS winds and temperatures. Average wind and temperature biases and gross errors relative to rawinsonde observations as a function of height are presented in Fig. 10. This plot supports the following findings:
    1) Above 3000 m, at which level the background analysis nudging dominates, the 2 m s[sup-1] wind biases and the 1-K temperature biases are due to inconsistencies between the NCEP analyses used to create the RAMS background fields and the SARMAP rawinsonde observations. These temperature biases aloft contribute to surface heating rates and result in higher simulated surface temperatures.
    2) ODA reduces the temperature gross errors at all heights, with the maximum gross error reduction of about 2.7 K in the lowest 250 m. Above 500 m, the average temperature gross error reductions are less than 1 K.
    3) A steplike feature appears at 500 m in the ODA case. This feature likely corresponds to the average top of the boundary layer and is produced by excessive surface-layer heat, which diffuses throughout the boundary layer, increasing its temperature.
    4) The NODA case shows a 4-K temperature bias at the surface that diminishes to about 0 K at 500 m. This overprediction of near-surface air temperatures indicates excessive surface heating. The -0.7-K bias between 700 and 1200 m is probably related to the northwesterly wind bias over the same elevations that results in more cool marine air being brought into this level.
    5) The wind biases below 500 m are less than 1 m s[sup-1] for both cases. The ODA wind biases are less than 1 m s[sup-1] up to 3000 m, where they begin to increase to about 2 m s[sup-1]; the NODA wind biases begin to increase at 500 m to a similar value of 2 m s[sup-1]. The close agreement of the NODA near-surface winds is due to reasonable simulation of the boundary layer forces, which produce sea breezes, slope flows, and orographic channeling, and to the general uniformity of the winds within the boundary layer. The NODA result suggests that the forces that dominate the near-surface ODA winds are not the ODA forces, but are the same boundary layer forces that produced agreement with observed winds in the NODA case.
    To illustrate specific differences between the two RAMS results, we present, in Figs. 11 and 12, comparisons between rawinsonde observations and simulated values for 2300 UTC 5 August 1990 near Visalia. Examination of these plots supports the following findings:
    1) The temperature and wind profiles from the ODA case are in better agreement with observations, as expected.
    2) The temperature profiles of the ODA case tend to reproduce the general features found in the observations, whereas the temperature profiles of the NODA case are more characteristic of idealized boundary layers and may differ significantly from observations. The RAMS vertical grid spacing, which is about 100 m at 500 m AGL, does not resolve the detailed structure in the observed temperature profiles.
    3) The modeled boundary layer heights, inferred from nonzero TKE values, often differed from heights inferred from observed temperature profiles. The ODA heights were often lower than observed estimates, which may be due to the restriction of TKE by stable regions in the ODA temperature profiles. In the NODA case stable regions were not forced to develop, and the heights were generally higher than the ODA heights.
    4) Near 2000 m, surface-layer and synoptic-background forces are both weak, and the NODA winds there often did not agree with observations. The disagreement in this region is not easily reconciled. Indeed, there is some association between atmospheric stability and development of shear layers, as can be seen in these examples. This result suggests that assimilation of temperatures is as important to the determination of winds as is assimilation of winds, if not more so. In these simulations, this elevation seems to have benefited most from ODA.

B. EFFECTS ON THE TRACER SIMULATIONS
    We evaluate the differences between the LPM simulations of the SARMAP tracer experiments (Tracer Technologies 1991) caused by the differences in transport properties between the two fine-grid meteorological fields. Figure 13 shows particle distributions and surface-concentration estimates 12 and 24 h after the 1400-1700 UTC 3 August 1990 release of tracer material from Pittsburg, California. Nonzero observations (fL L[sup-1]) are shown as numerals, the planar projection of the particles is shown as dots, and the estimated surface concentrations from the particle distribution are shown as isopleths. In the first 12 h, the LPM results for both cases show similar development of the particle distribution and show qualitative agreement with ground-based observations. After 12 h, the disparity increases between the planar distribution of the LPM particles and the location of significant surface concentrations. This suggests that particles have been transported by winds different from boundary layer winds. This scenario is likely after early evening, at which time the atmosphere, becoming more stable, is able to maintain increased wind shear.
    Simulated surface concentrations from the ODA case compare well with observations in their distribution and values. However, note that these simulated values are sensitive to the value of gamma selected in the above formula for computing concentrations. Nevertheless, NODA surface concentrations are noticeably lower than the ODA case, which is the consequence of its deeper boundary layers (cf. TKE of Fig. 11 and Fig. 12). Yet, because the differences between the two average winds become significant above the boundary layer (see Fig. 10), the similarity in the resulting surface concentrations suggests that the source-receptor relationships from the San Francisco Bay Area through the San Joaquin Valley are primarily defined by boundary layer winds.
    The simulation of the tracer release from Pittsburg agreed better with observations than the simulation of the simultaneous tracer release from San Jose (not shown). The trajectories from San Jose tended to travel southeastward down the western part of the San Joaquin Valley where the simulated winds disagreed with observations more than in the center of the San Joaquin Valley. Furthermore, there are large terrain features directly east and generally downwind of San Jose that complicate the flow and increase the sensitivity of tracer trajectories to the details of winds around San Jose.

C. EFFECTS ON THE OZONE SIMULATIONS
    We evaluate the performance of the UAM-V photochemical model for each of the two meteorological inputs. We also demonstrate that the relatively strong sensitivity of the UAM-V model to BC variations can obscure identification of performance improvements from ODA.
    Both UAM-V simulations met California state- and federal-recommended performance criteria. The State of California performance criteria (DaMassa et al. 1992) evaluate the performance of a given photochemical modeling application relative to other earlier applications. The performance of these simulations met or exceeded the state's "typical" performance criteria and satisfied the less-stringent performance evaluation criteria of the U.S. Environmental Protection Agency (U.S. EPA 1991). Figure 14 compares simulated and observed surface-level ozone for the ODA case, with a correlation coefficient of 0.78. The slope of 1.0 is somewhat accidental in the sense that it is due to the balanced overprediction of low observations and the underprediction of high observations. The same balancing of over- and underpredictions helps to yield the small bias of -4.41 ppb.
    To characterize further these simulations, Fig. 5 presents the histories of mean hourly surface-level ozone (MHSO) from the observations and from the simulation of the ODA case. Although the two histories are generally similar, there are notable differences. Observed diurnal maximum MHSO increases by approximately 5 ppb day[sup-1] over the 4-day period. The corresponding simulated parameter, after the first day, similarly increases, although at one-half of the rate. The simulation on the first day, a model spinup day, overpredicts the maximum MHSO by 12 ppb, a result that likely is due to inaccurate initial conditions. Diurnal minimum MSHO, seen as short-duration dips, is overestimated consistently by 5-11 ppb on the last two mornings. These short-duration minima may correspond to the destruction of residual nighttime ozone by the injection of nitric oxide into the 25-m surface layer from morning commuter traffic. Tanrikulu et al. (2000), in their modeling of this period, reported similar but larger nighttime overpredictions of ozone (14-17 ppb) that they attributed to excessive dilution of nitric oxide and subsequent insufficient ozone destruction in their 60-m-thick surface layer.
    Gross errors (Fig. 15) are not spatially skewed, and most values are less than 25 ppb. The highest gross errors generally occur where ozone values themselves are highest, such as in the San Joaquin Valley. There were differences between the surface ozone patterns produced by the two simulations, but Fig. 16 shows mostly small changes in gross error. In this figure, a star indicates improved performance with ODA and a circle indicates worse performance with ODA. In most locations, the difference in gross error is less than 5 ppb, with no apparent spatial pattern in these differences and with little indication that one simulation was better than the other.
    In Fig. 10, we examine vertical profiles of gross error and bias of the two cases relative to aircraft-spiral ozone observations for elevations from surface to 1500 m. We find that both cases are biased negatively above 500 m, with values sometimes exceeding -15 ppb. In general, the biases of the ODA case are shifted 5 ppb in the positive direction relative to the NODA case. The negative biases above 500 m likely are due to incorrect BC values. We find no consistent pattern in gross errors of vertical profiles.
    The interpretation of the effects of ODA is complicated by the uncertainty in other photochemical modeling inputs such as BCs and pollutant emissions. To investigate the possible consequences of this issue, we repeat both UAM-V simulations with alternate BC values, values set to the U.S.-EPA-recommended clean background concentration values (DaMassa et al. 1996), values that are acceptable in model applications. In the boundary layer, the clean BCs contained 4 ppb ozone, 2 ppb oxides of nitrogen, 19 ppbC nonmethane hydrocarbons, and 3 ppbC aldehydes in contrast with 4 ppb ozone, 3 ppb oxides of nitrogen, 26 ppbC nonmethane hydrocarbons, and 9 ppbC aldehydes for the observationally based BCs. Figure 17 presents daily normalized gross mean error (NGME; defined in Table 1) of ozone for the two original simulations and the two additional simulations with clean BCs. These NGME statistics were formed with a cutoff value of 40 ppb. The clean BCs caused significant underprediction of ozone by the fourth day of the simulation, on which NGME for ozone increased from 21% to 28% in the ODA case and from 20% to 24% in the NODA case. In neither case is ODA beneficial. However, it is significant that when observations are used for setting BCs, not only does the performance improve with either meteorological input, but the ODA cases switch from a clear disadvantage to neutral. To be specific, for the last two days of simulation, with clean BCs the NODA case performs better than the ODA case; however, with BCs based upon observations, the two meteorological cases are nearly equivalent.
    In summary, we found that the use of ODA meteorological fields has a small effect on UAM-V ozone performance relative to changes made by reasonable variations of BCs. The NGME for ozone with a 40 ppb cutoff changes less than 4% between the two meteorological cases, and, by comparison, it changes by nearly 8% between the use of observationally based and clean BCs. Errors and uncertainties in other modeling factors similarly may influence the evaluation of ODA.

6. SUMMARY AND DISCUSSION
    The effectiveness of the ODA technique in improving RAMS results depends upon the disparity between observations and the results obtained without ODA. For example, surface and upper-air wind biases were small in the NODA case and, therefore, did not change significantly with ODA. As expected, substantial reductions in gross error of upper-air winds, temperature, and moisture values were obtained by applying ODA to these observations; however, reductions in bias and gross error of surface parameters were small. This result suggests a weak coupling between surface parameters and the ODA of upper-air observations. The similarity of results in the two cases of boundary layer parameters, particularly winds, suggests strong influences in this layer from surface features, primarily orography and land-water interfaces.
    The LPM simulation of the tracer experiments using ODA meteorological inputs resulted in higher ground-level tracer concentrations and better agreement with observations than did the results obtained without ODA. The larger concentrations are due to the shallower boundary layers of the ODA meteorological inputs. Although details of the LPM surface particle distributions were slightly better with ODA, the general features of both cases were similar because of similarities in the boundary layer winds. The differences in winds aloft did not have a significant influence on LPM surface concentrations, which suggests that, for these trajectories, transport within the boundary layer is the primary mechanism governing source-receptor relationships.
    UAM-V results were not improved by the use of the ODA meteorological inputs. We propose four possible reasons for this negative result: 1) winds above the planetary boundary layer (PBL), at which height large decreases in gross error were obtained with ODA, were not influential in the formation of ozone; 2) PBL winds were not sufficiently improved by ODA to make noticeable improvements to UAM-V results; 3) UAM-V results for the NODA case, already having acceptable performance according to CARB and U.S.-EPA criteria, is near optimal, and further improvements may not be possible; and 4) the difference in UAM-V performance from the different meteorological inputs was demonstrated to be a function of BC values. This result implies that improvements in photochemical model performance from improved meteorological inputs could be obscured by errors in other air quality model inputs. For example, the emissions inventory is another influential input that is very uncertain.
    We must qualify the results of this study. First, this was a modeling exercise for a single period; other periods may yield different results. Second, the ODA method in this study, the objective of which was to determine the benefits of field-study data on model performance, did not employ routinely available surface observations nor did it use sodar or profiler observations at their full time resolution. The gross errors and biases of boundary layer meteorological parameters could be reduced further by doing so (Fast 1995; Tanrikulu et al. 2000). These additional ODA refinements may improve LPM results, given that LPM trajectory simulations of tracer releases are very sensitive to the details of the flow patterns at the time and location of the release. Tracers that traverse areas with significant terrain are subject to flows that reflect the details of the terrain that low-resolution simulated flows capture poorly. Indeed, LPM performance in these situations likely will benefit from ODA of surface observations and from increasing the time resolution of ODA data. However, UAM-V simulations of ozone are less likely to benefit from more detail in the ODA method because ozone is a secondary pollutant and the sources of ozone precursors are widely distributed.
ADDED MATERIAL
    TAKATO UMEDA AND PHILIP T. MARTIEN
    Bay Area Air Quality Management District, San Francisco, California
    Corresponding author address: Takato Umeda, Bay Area Air Quality Management District, 939 Ellis St., San Francisco. CA 94109. E-mail: tumeda@baaqmd.gov
    Acknowledgments. This work was partially funded by the Monterey Bay Unified Air Pollution Control District. We thank Vernon Hughes and Martin Johnson of CARB for providing emissions estimates.
    TABLE 1. Definition of terms.

      Term                             Definition
Bias                  1 / N [Graphic Character Omitted] (Prediction[subi] - Observation[subi]
Gross error           1 / N [Graphic Character Omitted] |Prediction[subi] - Observation[subi]|
Rms difference        [square root]1 / N [Graphic Character Omitted] (Prediction, Observation[subi])[sup2]
Veetor wind dif-
  ference (VWD)       [square root]1/N [Graphic Character Omitted] (U[subi] - U[suboi])[sup2] + (V[subi] - V[suboi])[sup2].
                where
                U[suboi] = U component of observed wind,
                U[subi] = corresponding U component of model wind,
                V[suboi] = V component of observed wind, and
                V[subi] = corresponding V component of model wind
Normalized gross      1 / N [Graphic Character Omitted] |Prediction[subi] - Observation[subi] / Observation[subi]|
  mean error
  (NGME)
                        for Observation[subi] > cutoff,
                where
                              N = No. of terms in sum,
                    Observation[subi] = ith observation, and
                     Prediction[subi] = prediction corresponding
                                       to Observation[subi]

    TABLE 2. RAMS ODA and NODA statistics.

                                                                 Sample
                                             ODA      NODA        No.
                      Surface observations
U bias (m s[sup-1])                          -0.755    -0.768      7708
V bias (m s[sup-1])                           0.028     0.150      7708
Wind speed error (m s[sup-1])                 0.755     0.783      7708
U rms difference (m s[sup-1])                 2.010     2.017      7708
V rms difference (m s[sup-1])                 1.764     1.873      7708
VWD (m s[sup-1])                              2.674     2.753      7708
Temperature bias (K)                         2.89      3.24       8581
Temperature gross error (K)                  3.37      3.83       8581
Temperature rms difference (K)               3.21      3.63       8581
Temperature correlation coef                 0.91      0.88       8581
Moisture bias (g kg[sup-1])                  -0.90     -1.11        217
Moisture gross error (g kg[sup-1])            1.077     1.131       217
Moisture rms difference (g kg[sup-1])         1.148     0.901       217
                      Upper-air observations
U bias (m s[sup-1])                          -0.024    -0.666    21 789
V bias (m s[sup-1])                          -0.227    -0.884    21 789
Wind speed error (m s[sup-1])                 0.228     1.107    21 789
U rms difference (m s[sup-1])                 1.601     2.869    21 789
V rms difference (m s[sup-1])                 1.706     3.221    21 789
VWD (m s[sup-1])                              2.340     4.313    21 789
Temperature bias (K)                         0.883     1.084    10 418
Temperature gross error (K)                  1.120     1.925    10 418
Temperature rms difference (K)               1.207     2.382    10 418
Moisture bias (g kg[sup-1])                  -0.252    -2.204    10 271
Moisture gross error (g kg[sup-1])            0.612     2.433    10 271
Moisture rms difference (g kg[sup-1])         0.780     1.906    10 271

FIG. 1. RAMS fine-grid domain with the UAM-V domain delineated with a dashed line. Relevant geographic landmarks are shown. Ozone observation sites used in this study are indicated with circles.
FIG. 2. Meteorological observation sites (B = rawinsoders, S = surface observations, D = sodars, P = profilers) and the spatial distribution of data assimilation weights for potential temperature. The contour lines for the weights indicate relative values from zero to the maxima, which are located at the rawinsonde sites (B).
FIG. 3. The 500-hPa heights and winds for 0000 UTC 5 Aug 1990, interpolated from the NCEP Global Tropospheric Analyses onto the RAMS coarse-grid modeling domain.
FIG. 4. Average temperature histories from rawinsonde observations near 850 hPa (solid), surface inland-area observations (dashed), and surface coastal-area observations (dotted).
FIG. 5. History of surface SARMAP ozone observations for the episode. Plus symbols indicate individual hourly observations. The solid line plots the average of these observations for each hour. The dashed line plots the average of the corresponding simulated values for the ODA case.
FIG. 6. Surface vector wind bias at each observation site for the ODA case. Wind speed gross error is indicated by the diameter of the circle.
FIG. 7. History of temperature bias for the ODA (dashed) and NODA (solid) simulations.
FIG. 8. Temperature bias at each surface observation site for the ODA case. Temperature bias is indicated by symbol size; positive bias indicates overprediction.
FIG. 9. Same as Fig. 8, but for NODA case.
FIG. 10. Temperature, wind, and ozone gross errors and biases for the simulation period from both cases are shown as a function of height above sea level (ASL). The ODA comparisons are shown as dark lines, and the NODA comparisons are shown as gray lines. The mean observed wind vectors are shown along with the vector wind biases. Wind and temperature observations are from rawinsonde soundings. Ozone observations are from aircraft spiral measurements.
FIG. 11. Vertical profile of observed and simulated potential temperatures and winds near Visalia at 2300 UTC 5 Aug 1990 from the ODA case. Simulated TKE is also plotted. The location of Visalia is shown on inset map.
FIG. 12. Same as Fig. 11, but for the NODA case.
FIG. 13. Nonzero tracer observations (numerals in femtoliters per liter) from the 1400 UTC 3 Aug 1990 Pittsburg experiment 12 and 24 h after release. LPM surface concentrations (isopleths in femtoliters per liter) are shown using the (a), (b) NODA and (c), (d) ODA meteorological fields. Planar view of LPM particles (black dots) and the tracer release location (X) are shown.
FIG. 14. Observed vs simulated surface ozone for the ODA case on 5 and 6 Aug 1990 LT. For 4217 observations, the bias is -4.41 ppb, the slope is 1.00, and the correlation coefficient is 0.78. The horizontal stripes are caused by some ozone measurements being reported only to the nearest 10 ppb.
FIG. 15. Gross errors of ozone at each observation site for the ODA case.
FIG. 16. The change in gross error of ozone at each observation site from ODA relative to NODA.
FIG. 17. Daily normalized gross mean error for ozone for four modeling assumptions: meteorological fields--ODA (thick lines) or NODA (thin lines); BCs--observationally based (solid) or clean (dashes). Ozone observations of less than 40 ppb were not included in these statistics. The dates refer to LST.

REFERENCES
    Chen, C., and W. R. Cotton, 1987: The physics of the marine stratocumulus-capped mixed layer. J. Atmos. Sci., 44, 2951-2977.
    DaMassa, J., P. Allen, T. McGuire, D. McNerny, A. J. Ranzieri, and K. Wagner, 1992: Technical guidance document: Photochemical modeling. California Air Resources Board Tech. Rep., 128 pp. [Available from California Air Resources Board, P.O. Box 2815, Sacramento, CA 95812.]
    DaMassa, J., S. Tanrikulu, K. Magliano, A. J. Ranzieri, and L. Niccum, 1996: Performance evaluation of SAQM in central California and attainment demonstration for the August 3-6, 1990 ozone episode. California Air Resources Board Tech. Rep., 109 pp. [Available from California Air Resources Board, P.O. Box 2815, Sacramento, CA 95812.]
    Dodge, M. C., 1989: A comparison of three photochemical oxidant mechanisms. J. Geophys. Res., 94, 5121-5136.
    Fast, J. D., 1995: Mesoscale modeling and four-dimensional data assimilation in areas of highly complex terrain. J. Appl. Meteor., 34, 2762-2782.
    Gery, M. W., G. Z. Whitten, J. P. Killus, and M. C. Dodge, 1989: A photochemical kinetics mechanism for urban and regional scale computer model modeling. J. Geophys. Res., 94, 12 925-12 956.
    Magliano, K., 1996: Performance evaluation of the SAQM in central California and attainment demonstration for the August 3-6, 1990, ozone episode. Appendix C, summary of SARMAP emissions modeling. California Air Resources Board Tech. Rep., 68 pp. [Available from California Air Resources Board, P.O. Box 2815, Sacramento, CA 95812.]
    Mellor, G. L., and T. Yamada, 1974: A hierarchy of turbulence closure models for planetary boundary layers. J. Atmos. Sci., 31, 1791-1806.
    Pielke, R. A., and Coauthors, 1992: A comprehensive meteorological modeling system--RAMS. Meteor. Atmos. Phys., 49, 69-91.
    Ranzieri, A., and R. Thuillier, 1994: SJVAQS and AUSPEX: A collaborative air quality field measurement and modeling program. Planning and Managing Regional Air Quality Modeling and Measurement Studies: A Perspective through the San Joaquin Valley Air Quality Study and AUSPEX, P. A. Solomon, Ed., Lewis Publishers, 1-28.
    Reif, F., 1965: Fundamentals of Statistical and Thermal Physics. McGraw-Hill, 651 pp.
    SAI, 1993: Users' guide to the Urban Airshed Model (UAM-V). Systems Applications International Tech. Rep. SYSAPP-93/153, 177 pp.
    Seaman, N. L., 2000: Meteorological modeling for air-quality assessments. Atmos. Environ., 34, 2231-2259.
    Seaman, N. L., D. R. Stauffer, and A. M. Lario-Gibbs, 1995: A multiscale four-dimensional data assimilation system in the San Joaquin Valley during SARMAP. Part I: Modeling design and basic performance characteristics. J. Appl. Meteor., 34, 1739-1761.
    Stauffer, D. R., and N. L. Seaman, 1994: Multiscale four-dimensional data assimilation. J. Appl. Meteor., 33, 416-434.
    Tanrikulu, S., D. R. Stauffer, N. L. Seaman, and A. J. Ranzieri, 2000: A field-coherence technique for meteorological field-program design for air quality studies. Part II: Evaluation in the San Joaquin Valley. J. Appl. Meteor., 39, 317-334.
    Tracer Technologies, 1991: Revised San Joaquin Valley Air Quality Study atmospheric tracer data, appendix C, TEST 2--August 3-5, 1990. Tracer Technologies Tech. Rep., Contract ARB-88-33, 104 pp. [Available from California Air Resources Board, P.O. Box 2815, Sacramento, CA 95812.]
    U.S. EPA, 1991: Guideline for regulatory application of the Urban Airshed Model. U.S. Environmental Protection Agency Tech. Rep. EPA-450/4-91-013, 108 pp.
    von Storch, H., H. Langenberg, and F. Feser, 2000: A spectral nudging technique for dynamical downscaling purposes. Mon. Wea. Rev., 128, 3664-3673.