lable at ScienceDirect

Environmental Modelling & Software 52 (2014) 19e37

Contents lists avai

Environmental Modelling & Software

journal homepage: www.elsevier .com/locate/envsoft

A computationally efficient model for estimating backgroundconcentrations of NOx, NO2, and O3

Sam Pournazeri a,1, Si Tan a,1, Nico Schulte a, Qiguo Jing b, Akula Venkatram a,*

aDepartment of Mechanical Engineering, University of California Riverside, Riverside, CA 92521, USAbBREEZE Software, Trinity Consultants, 12770 Merit Dr. Suite 900, Dallas, TX 75251, USA2

a r t i c l e i n f o

Article history:Received 20 March 2013Received in revised form5 October 2013Accepted 17 October 2013Available online 13 November 2013

Keywords:Air qualityBackground concentrationSpecies ageLagrangian modelLong-range transportUrban scale

Software availability

Program title: LBM-Lagrangian BackgroundContact address: Akula Venkatram, Departm

Engineering, University ofRiverside, CA 92521, USA.

Software: All platforms supporting Matlab.

* Corresponding author. Tel.: þ1 951 827 2195.E-mail addresses: venky@engr.ucr.edu,

(A. Venkatram).1 These authors contributed equally to the work.2 Performed the work at University of California, R

1364-8152/$ e see front matter � 2013 Elsevier Ltd.http://dx.doi.org/10.1016/j.envsoft.2013.10.018

a b s t r a c t

We formulate a Lagrangian model to supplement comprehensive Eulerian grid models such as CMAQ, toestimate concentrations of NOx, NO2, and O3 averaged over a spatial scale of the order of a kilometer overa domain extending over hundreds of kilometers. The model can be used to estimate hourly concen-trations of these species over time periods of years. It achieves the required computational efficiency byseparating transport and chemistry using the concept of species age. The model computes concentrationsby tracing the history of an air parcel reaching a receptor through back trajectories driven by surfacewinds. Chemical reactions within the parcel are modeled through the Generic Reaction Set (GRS)chemistry model, which approximates the photochemical processes that lead to the production of ozone.The model is evaluated with concentrations measured over two years, 2005 and 2007. Evaluation withdata measured at 21 stations distributed over the California South Coast Air Basin (SoCAB) during 2007indicates that the model provides an adequate description of the spatial and temporal variation of theconcentrations of NO2, and NOx. Estimates of maximum hourly O3 concentrations show little bias (lessthan 10%) compared to observations, and the scatter (sg2 � 2.56e95% confidence interval of the ratio ofpredicted to observed concentrations) is comparable to those associated with more computationallydemanding models. The model was also evaluated with data collected at monitors in the San JoaquinValley Air Basin (SJVAB) in 2005, and it shows similar performance to that at SoCAB. The paper alsoillustrates the application of the model to 1) screening regions for attainment of statistically based airquality standards, such as that for the daily maximum 8-h average O3, and 2) improving methods tointerpolate observations.

� 2013 Elsevier Ltd. All rights reserved.

Model.ent of MechanicalCalifornia Riverside,

1. Introduction

Exposure studies in urban areas require concentrations asso-ciated with emissions from a large number of sources, such as

venkatram@sbcglobal.net

iverside.

All rights reserved.

vehicles and power generators, distributed over the urban area.Computing the contributions from these sources necessitatescomputational resources that can become impractical even withcurrent computers, especially when it is necessary to conductsensitivity studies over long averaging times. The currentapproach to reducing these computational needs is to use separatedispersion models for different spatial scales so that sources atdifferent distances from the area of interest are treated withdifferent levels of source aggregation. The concentration at a re-ceptor has three components: a regional contribution computedfrom a long-range transport model with grid spacing of the orderof tens of kilometers, an urban “background” contribution fromsources aggregated over kilometer-sized grids, and a local contri-bution from models that estimate concentrations at meters from areceptor. This approach to estimate the contributions from emis-sions at varying distances from a receptor was pioneered byBrandt et al. (2001a, 2001b, 2001c, 2001d, 2001e, 2003) indeveloping an integrated operational air pollution forecast systemcalled THOR.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e3720

Urban background concentrations of species such as NO2 andozone can be computed using comprehensive models such asCMAQ (Community Multi-scale Air Quality Model; Byun and Schere,2006). Although the previous studies show that comprehensivemodels can provide reasonable estimates of background concen-trations of species such as NO2 and ozone, they can becomecomputationally cumbersome if concentrations are required forlong averaging times.

Most exposure studies avoid the use of dispersion modelsaltogether by analyzing data from monitoring networks throughland-use regression (LUR) statistical models (Brauer et al., 2003;Ross et al., 2007) that interpolate between observations to iden-tify small-scale gradients in concentrations. While such models areuseful for estimating current or past exposure, they cannot be usedto estimate the impact of changes in emissions on concentrationsand their performance is highly dependent on the measured datathat is used to calibrate them (Jensen et al., 2009). Furthermore,they cannot be used to estimate concentrations at short averagingtimes because they do not account for the impact of varyingmeteorological conditions (Gulliver and Briggs, 2011).

For the reasons given in the previous paragraphs, we need adispersion model that connects concentrations explicitly to emis-sions and meteorology, and at the same time is computationallyconvenient for long-term exposure studies. The outputs from sucha model can be used to interpolate observations or serve as back-ground concentrations for a model applied at a smaller scale of tensof meters. This paper describes such a model, which is designed toestimate hourly concentrations of selected pollutants. These con-centrations can be averaged over longer periods, relevant to healthstudies.

The simple Urban Background Model (UBM, Berkowicz, 2000)is typical of the response to the need for urban scale dispersionmodels with small computational demands. UBM achievescomputational efficiency through two simplifications: astraight-line steady dispersion model, and ozone chemistrybased on photo-stationarity, which neglects the role of hydro-carbons. In this paper, we focus on a model that is intermediatebetween comprehensive photochemical models and the simpleUBM. This model estimates urban “background” concentrationsof NOx, NO2, and O3, averaged over a scale of a few kilometers totens of kilometers. The lower limit on the grid size is determinedby the assumption that the concentration is well mixed throughthe depth of the mixed layer, and the upper limit depends on thevalidity of using surface winds to represent transport in theatmosphere. The model treats unsteady meteorological condi-tions with trajectories that reflect space and time varying sur-face winds, and it reduces the computational requirements ofphotochemical models by separating transport and chemistryusing a method described in Venkatram et al. (1998). Themodel is evaluated by comparing model estimates of relevantspecies with data from measurements made in the South Coast(SoCAB) and the San Joaquin Valley (SJVAB) air basins inCalifornia, USA.

2. The Lagrangian Background Model

The model described here is similar to the Lagrangian modelused in Europe to estimate long-range transport of sulfur (Eliassenand Saltbones, 1983). It estimates the concentration of a pollutantby tracing the history of the air parcel associated with the con-centration at the receptor of concern at a specified time. Back tra-jectories are calculated in 1-h time steps using the hourly averagedwind speed (at 10 m above the ground level), and wind directionfrom the meteorological station closest to the receptor. Each tra-jectory is extended backwards in time for 24 h, which assumes that

sources beyond this travel time make a negligible contribution toconcentrations at the receptor. This assumption was evaluated us-ing sensitivity studies. To facilitate the use of the model, meteo-rological inputs are taken directly from the surface input files usedby AERMOD (Cimorelli et al., 2005) which are generated by theAERMOD meteorological preprocessor (AERMET). In a single layermodel, the choice of the height of the wind used to compute tra-jectories is arbitrary; the choice of the 10 m wind is justifieda posteriori through comparison of model estimates withobservations.

The air parcel has horizontal dimensions of the grid squareused to represent emissions of NOx and VOC (volatile organiccompounds) over the domain. The height of the air parcel corre-sponds to the local mixed layer height. In order to account forhorizontal dispersion, we examined the approach used in UBM(Berkowicz, 2000), in which the concentration at a receptor istaken to be the average of the concentrations corresponding toslightly different wind directions ðDf ¼ 3

� � 5� Þ centered on the

average wind direction. We found that perturbing the back tra-jectories using this approach made little difference to the results.Consequently, this approach was dropped to improve computa-tional efficiency.

Emissions are injected every hour into the box at the gridstraced by the back trajectory, and then mixed through the volumeof the box. The concentrations are stepped from the (i � 1)th to theith time step through

Ci ¼ Ci�1min�zi�1zi

;1�þ Dmi

zi(1)

where Ci is the concentration of the species at time i, Dmi is themass of pollutant injected into the air parcel, and zi is the mixedlayer height. The termwithin the parenthesis on the right hand sideof the equation ensures that the concentration does not increasewhen themixed layer decreases during a time step. The loss of masswhen the mixed layer height decreases ensures that the near sur-face concentration is affected primarily by material that is less than24 h old. The mixed layer heights used in this model are generatedthrough AERMET, which calculates the height of the convectiveboundary layer (CBL) through a simple one-dimensional energybalance model (Carson, 1973). The height of the stable boundarylayer (SBL) is based on the formulation described in Venkatram(1980).

Currently the model does not account for dry deposition,although this process can be readily incorporated into Equation (1).The mass of pollutant injected per unit surface area of the air parcelis Dmi ¼ qiðr;!sÞDt, where qið r!; sÞ is the emission density at thelocation of the parcel, r!, injected at a time from the initiation of thetrajectory, s, and Dt is the time step of the trajectory calculation.

The incremental concentration during the last hour of the airparcel’s path is computed with a steady state dispersion model(Venkatram and Cimorelli, 2007) that accounts for incompletevertical mixing,

DCi ¼ffiffiffiffi2p

rqsw

ln�1þ swDt

h

�(2)

where q is the emission rate per unit area, sw is the standard de-viation of the vertical velocity fluctuations, and h is the initialvertical spread of surface emissions which is taken to be 1 m. Theequation is modified if the pollutant is well mixed through theboundary layer during the last time step before the parcel reachesthe receptor.

Once the concentrations of the primary pollutants are esti-mated, the model calculates the effective age of each species in the

Fig. 1. Gridded daily average NOx emissions and monitoring stations located in the South Coast Air Basin (Left panel). The right panel shows the assumed temporal profile of NOx

emissions.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e37 21

box (Venkatram et al., 1994,1998). The effective age of a molecule isthe time taken for the molecule to travel from source to receptor.We can build upon this simple idea to formulate a conservationequation for species age that accounts for complex flows andemissions in a Eulerian grid model. This equation allows thecalculation of age in addition to concentration of a species at everyreceptor.

In this simple Lagrangian model the formulation for the speciesage, Ai, reduces to

Ai ¼ Ai�1

�1� Dmi

mi

�þ Dt

�1� 1

2Dmi

mi

�(3)

In the absence of fresh emissions, that is Dmi ¼ 0, we obtain theexpected result: Ai ¼ Ai�1 þ Dt. Note that fresh emissions alwaysdecrease the effective age of the species within the box. Then, thechemical transformation of these species is estimated by reactingthem with other species in a box with initial concentrations cor-responding to those in the absence of chemistry. The time periodfor chemical calculations is specified by the end time correspondingto the time of interest and a start time that is the end time minusthe species age. The chemical calculation is performed over themaximum of the ages of the species in the air parcel. In the sub-sequent discussion, we refer to the proposed Lagrangian Back-ground Model as LBM.

The chemistry, which accounts for the variation of photolysisrates with time of day, uses the Generic Reaction Set (GRS) chemicalscheme proposed by Azzi et al. (1992). This scheme approximatesthe reactions leading to the formation of ozone using seven re-actions among seven species:

ROC þ hn/RP þ ROC (R1)

Fig. 2. Daily average NOx (left panel) and VOC (right p

RP þ NO/NO2 (R2)

NO2 þ hn/NOþ O3 (R3)

NOþ O3/NO2 (R4)

RP þ RP/RP (R5)

RP þ NO2/SGN (R6)

RP þ NO2/SNGN (R7)

whereROC ¼ reactive organic compounds.RP ¼ radical pool.SGN ¼ stable gaseous nitrogen product.SNGN ¼ stable non-gaseous nitrogen product.The reactions and the corresponding reaction rates are:R1. Radical production from photo-oxidation of ROC

k1 ¼ 0:0067 k3 fðTÞwhere fðTÞ

¼ exp�� 1000g

�1T� 1316

��; g ¼ 4:7

R2. Oxidation of nitric oxide by radicals

k2 ¼ 3:58� 106=T ppm�1 min�1

R3. Photolysis of nitrogen dioxide to nitric oxide

k3 ¼ exp�� 0:575sinðqÞ

�where q is the sun elevation angle:

anel) emissions (g s�1) in SoCAB from Samuelsen.

Fig. 3. Comparison of NOx and O3 estimates based on GRS and CBM IV at stations in the SoCAB. The upper panels show the monthly averaged NOx and O3 concentrations and thelower panels show daily variations of NOx and O3 concentrations at the San Bernardino monitoring station from January to December.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e3722

R4. Nitric oxide-ozone titration reaction

k4 ¼ 9:24� 105 T�1exp��1450

T

�

R5. Radical pool sink through recombination to stable products

k5 ¼ 10200 ppm�1 min�1

R6. Sink for nitrogen dioxide to stable gaseous nitrates

k6 ¼ 120 ppm�1 min�1

R7. Sink for nitrogen dioxide to stable non-gaseous nitrates

k7 ¼ 120 ppm�1 min�1

Reactions R3 and R4 represent chemically exact mechanisms,while the rest approximate reactions of generic chemical counter-parts. Reaction R1 is a semi-empirical representation of all theprocesses that lead to radical production from VOCs through photo-oxidation. Notice that ROC is conserved in the reaction; thus, ROCbecomes a surrogate for the products of the initial oxidation of theemitted VOCs. Reaction R2 represents the conversion of NO to NO2by radicals. Notice that, unlike the reactions in the actual mecha-nism, it leads to the termination of generic radicals, RP. Reaction R5represents another sink for the radical pool. Reactions R6 and R7lead to the formation of organic and inorganic nitrates. The rates of

these “pseudo” reactions have been determined empirically byfitting the ozone obtained from the GRS to smog chamber data.Reaction R1 is the most important reaction in the semi-empiricalGRS. The rate of this reaction has been calibrated against the rateat which radicals are produced by the different types of VOCs. Thereactivity coefficient, 0.0067, in R1 was derived by Johnson (1984)for a mixture of VOCs dominated by automobile emissions, and isincorporated by Hurley et al. (2003) in a Eulerian air pollutionmodel. Venkatram et al. (1994) used a slightly different approach byconverting VOC emissions to equivalent ROC by calibrating the GRSmechanism with a more complete chemical mechanism.

The current version of LBM does not simulate aerosol chemistryor other pollutants such as Persistent Organic Pollutants (POPs) orHeavy Metals, which are simulated by more complex models suchas CMAQ (Matthias et al., 2008) and CIT airshed (Dabdub et al.,2008). As demonstrated earlier (Venkatram et al., 1998), GRS canbe extended to include reactions to generate hydrogen peroxide,sulfuric and nitric acids, organic nitrates, and secondary organicaerosols.

LBM requires concentrations of NOx, VOC and O3 at the bound-aries of the domain where the back trajectory is terminated.Currently, their values are specified, but they could be derived froma larger scale model. A simple schematic of the LBM structure isshown in Fig. S1. The next section evaluates the performance of thesimplified chemistry in LBM by comparing two versions of LBM: onewith GRS and the other with the more complete Carbon Bond IVchemistry.

Fig. 4. Comparison of modeled and measured monthly averaged NOx, NO2, and O3 concentrations at 21 sites in the SoCAB from January to December 2007. Left panels correspond toUBM and right panels to LBM.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e37 23

3. Evaluation of LBM using SoCAB data

We used LBM to estimate NOx, NO2, and O3 concentrations in theSoCAB at the monitors depicted in Fig. 1, maintained by the SouthCoast Air Quality Management District (SCAQMD). Fig. 2 shows thedaily average NOx and VOC emissions at the 994, 5 km � 5 km gridsused by Samuelsen et al. (2005) for the SoCAB emission inventory.These emissions are projected into the future considering thegrowth in population, industrial and commercial activity, vehiclemiles traveled, and other factors. This inventory includes stationarypoint and nonpoint, non-road mobile, mobile, biogenic, and geo-genic emission sources as required by Consolidated EmissionsReporting Rule (CERR, 2002). We assumed that the diurnal varia-tion of NOx emissions corresponds to traffic volume, shown in the

right panel of Fig. 1. Fig. 2 shows the daily (24-hr) averaged NOx andVOC emissions in the SoCAB. As seen in Fig. 2, the maximum NOx

emissions occur near the port of Long Beach where ships, trucks,trains, and cargo-handling equipment emits about 48 tons of NOx

per day (http://www.polb.com). In addition, high levels of NOx andVOC emissions occur near the city of Ontario where highway 10 and60 as well as the Ontario international airport are the maincontributors.

The transport model was run with surface meteorological datacorresponding to 2007, measured at 21 meteorological stationsoperated by the SCAQMD. Model estimates are compared with NOx

concentrations measured at 21 monitoring stations numbered inthe left panel of Fig. 1. First, we compared the results from the LBMthat incorporates GRS chemistry with those from the LBM with the

Fig. 5. Monthly averaged NOx and NO2 concentrations compared with observations at two sites, in West LA (left panels) and San Bernardino (right panels), in the SoCAB.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e3724

more complete Carbon Bond Model IV (CBM IV) mechanism. In theCBM IV chemistry module, the volatile organic compounds (VOC)are taken to be amixture typical of ambient measurements made inLos Angeles: the VOC is distributed among eight surrogate speciesand one inert species. In these simulations, background ozone istaken to be 20 ppb.

We first compared the monthly averaged concentrations of NOx

and O3 from the two models at the 21 receptors located in SoCAB.Fig. 3 indicates that estimates of concentrations obtained from GRSand CBM IV follow the 1:1 line, except at small concentrations,where CBM IV predicts higher concentrations than GRS. Themaximum bias between the estimates from GRS and CBM IV is 7%.Fig. 3 shows that the models predict similar diurnal variations ofhourly NOx while O3 estimates based on GRS are slightly lower thanthose from CBM IV.

Next we examined the performance of UBM (Berkowicz, 2000)and LBM in describing the monthly (JanuaryeDecember 2007)averaged background concentrations of NOx, NO2, and O3 over 21receptors (Fig. 4). Model performance is described in terms of thegeometric mean and standard deviation, mg; sg, of the ratio of theestimated to the observed concentrations (Venkatram, 2008). FAC2refers to the fraction of the model estimates within a factor of twoof the corresponding observations. The values of normalized meanbias (NMB) and normalized mean error (NME) are also provided sothat the resultsmay be comparedwith the evaluation studies foundin the literature. The formulas used to compute NMB and NME are

NMB½%� ¼P�

Cpi � Coi�

PCoi

� 100 (4)

NME½%� ¼Cpi � CoiP � 100 (5)

P Coi

where Cpi is the ‘ith’ predicted concentration and Coi is the corre-sponding observed concentration.

We see that UBM overestimates NOx (top left panel) and NO2(middle left panel) concentrations while it provides relatively un-biased estimates of the O3 (bottom left panel) concentrations asindicated by the value of mg close to unity (w0.78). Fig. 4 showsthat LBM provides unbiased estimates of concentrations: all theNOx and NO2 estimates are within a factor of two of the observa-tions with a 4e13% bias. O3 concentrations from both models showsimilar comparisons with observations. One of the reasons for thepoor performance of UBM in estimating the NOx concentration forthis specific case study could be the domain size. UBM is designedto provide urban background concentrations for domains with anarea of about 50 km � 50 km, in which straight-line trajectories area reasonable approximation. The computational domain used inthis study covers a region of about 400 km� 150 km. For such largedomains, Brandt et al. (2000a,b, 2001a,b,c,d) developed a coupledsystem that consists of a Eulerian model, DEOM (Danish EulerianOperational Model) for air pollution forecasts at a regional scale(grid resolution of 50 km � 50 km), which is combined with theUBM model (Berkowicz, 2000) at a 2 � 2 km2 grid resolution. Theoutput from UBM is fed into the operational street pollution model(OSPM; Berkowicz, 1999) to estimate street level pollution con-centrations at a resolution of meters. This air pollution modelingsystem yields results that compare favorably with measurementsfrom the Danish urban monitoring network (Brandt et al., 2001c,2003). The correlation coefficients (r2) between model estimates

Fig. 6. Daily maximum O3 (top panels) and NO2 (bottom panels) concentrations compared with observations at two sites, in West LA (left panels) and San Bernardino (right panels),in the SoCAB.

Table 1Statistical performance parameter of LBM for calendar year 2007 at two sites (WestLA and San Bernardino) in the SoCAB.

Species NMBa (%) NMEb (%) r2

W LAc SBd W LA SB W LA SB

NO2 Daily max 14 18 36 36 0.18 0.11Monthly averaged �2 0 27 31 0.64 0.37

O3 Daily max �19 �2 34 36 0.09 0.43Monthly averaged �44 �24 48 27 0.57 0.92

a Normalized mean bias.b Normalized mean error.c W LA: West LA.d SB: San Bernardino.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e37 25

and observations of NO, NO2 and CO range from 0.72 to 0.96 and themaximum bias is a few ppb.

We tested the performance of LBM when NO2 is predictedthrough the photo-stationary method used in OSPM (Berkowicz,1999). The model was tested with different background NO2values (required to calculate the Ox ¼ NO2 þ O3). When the back-ground NO2 is set to 0 ppb, the model underestimates the NO2concentrations. Although the estimated NO2 falls within a factor oftwo of the observations (Fig. S2) whenwe use a background NO2 of15 ppb, the estimated NO2 does not show any correlation with theobserved data.

Next we examined the performance of LBM in estimating NO2and NOx concentrations at two sites, one in the west and the otherin the east of the Los Angeles basin, which is downwind of themajor sources of ozone precursors. The top two panels of Fig. 5show that the modeled NO2 and NOx concentrations, averagedover a month, are correlated well with the corresponding obser-vations. However, themodel underestimates NO2 during thewintermonths at the San Bernardino site. The discrepancies betweenmodel estimates and observations during the winter months couldbe related to the uncertainty in predicting the stable boundary layerheight as described in Pournazeri et al. (2012).

Fig. 6 compares the modeled and observed maximum daily O3(top panels) and NO2 (bottom panels) concentrations at these sta-tions. Statistics of the model performance reveal that the bias, asindicated by mg, ranges from 10 to 20%, and the scatter (95% con-fidence interval for the ratio of predicted to observed concentra-tions) is less than 2.56. More than 87% of the estimated values arewithin a factor of two of the observations. The performance of LBM

in estimating the background concentrations of O3 and NO2 in theSoCAB is expressed in terms of the statistical parameters NMB,NME, and the correlation coefficient r2 (Table 1).

Fig. 7 shows the daily variation of NOx and NO2 concentrationsaveraged over March and September in the SoCAB compared withmodel estimates. We see that the model overestimates NO2 andNOx concentrations in the early morning hours at the San Ber-nardino site. This might be related to the uncertainty in estimatingthe mixed layer height and the possibility of NO2 dry depositionduring these hours. It could also be associated with the assumedtemporal profile of NOx emissions. Fig. 8 indicates that the modelprovides a satisfactory description of the spatial variation of theconcentrations of these species across the 21 stations in theSoCAB.

Fig. 7. Monthly averaged daily variation of NOx and NO2 compared with observations at two sites, in West LA (left panels) and San Bernardino (right panels), in the SoCAB duringtwo different months.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e3726

How do the performance measures of LBM compare with thoseof more comprehensive models such as CMAQ? Table 2 comparesthe performance of LBMwith those of comprehensive models. Ederand Yu (2006) compared a full year of concentrations from CMAQsimulations to data from four nationwide monitoring networks(IMPROVE, STN, CASTNET, and ARIS-AQS) in the United States.CMAQ’s performance at estimating criteria pollutants and particu-late matter (PM) concentrations varies significantly. Estimates of 1-hr and 8-hr peak O3 concentrations compare well with data, with

Fig. 8. Comparison of modeled and measured annually averag

correlation coefficients, r2, of 0.46 and 0.47 respectively. However,CMAQ does not show comparable performance in estimating ni-trate (NO3

�) concentrations with r2 ranging from 0.13 to 0.38. CMAQ(Smyth et al., 2006) captures the spatial and temporal distributionof O3 concentrations measured during the Pacific 2001 experiment(Li, 2004). Liu et al. (2010) applied CMAQ to study the formation andseasonal variations of major pollutants such as SO2, NO2, and PM10

in China during January, April, July, and October of 2008. The pre-dicted surface NO2 concentrations were significantly smaller than

ed NOx and NO2 concentrations at 21 sites in the SoCAB.

Table 2Statistical parameter of various air quality models.

Study Region/Database Species Model Data type Simulation period NMBa (%) NMEb (%) r2

Eder and Yu (2006) IMPROVE, STN, CASTNET,ARIS-AQS

O3 CMAQg Daily Maximumof 1-hr average

AprileSeptember,2001

4.0 18.3 0.46

Liu et al. (2010) Eastern China NO2 CMAQh Monthly averageof daily average

Jan, Apr, Jul, and Oct2008

�6.5w�32.0d 47.1e66.6d 0.09e0.36d

Liu et al. (2010) Eastern China O3 CMAQh Daily maximumof 1-hr average

Jan, Apr, Jul, and Oct2008

1.1e12.0d 16.9e36.6d 0.5e0.7

McNair et al. (1996) Southern California, US NO2 CIT Airshedi 1-hr Average June 25th, 1987 e 69c 0.1McNair et al. (1996) Southern California, US NO2 CIT Airshedi 1-hr Average August 28th, 1987 e 44c 0.52McNair et al. (1996) Southern California, US O3 CIT Airshedi 1-hr Average June 25th

1987e 38c 0.83

McNair et al. (1996) Southern California, US O3 CIT Airshedi 1-hr Average August 25th1987

e 29c 0.82

Smyth et al. (2006) Vancouver, Canada O3 CMAQj Daily maximumof 1-hr average

August 9th �20th2001

�2.2 24.3 e

Vijayaraghavanet al. (2006)

Central California, US O3 CMAQk 1-hr Average Jul 30theAug 1st 2000 �3.9w�61.1e 15.5e61.1f e

Zhang et al. (2009) IMPROVE, STN, CASTNET,ARIS-AQS, SEARCH,NADP

O3 CMAQg Daily maximumof 1-hr average

2001 0.1w�11.6 19.8e20.5 e

AIRS-AQS: Aerometric Information Retrieval System-Air Quality Subsystem. Contains 1161 sites located primarily in cities and towns in the U.S.CASTNET: Clean Air Status and Trends Network. Contains 83 sites located primarily in remote/rural areas in the U.S.IMPROVE: Interagency Monitoring of Protected Visual Environments. Contains 134 sites located primarily in remote areas in the western U.S.NADP: National Acid Deposition Program. Contains 250 sites nationwide in the U.S.SEARCH: Southeastern Aerosol Research and Characterization. Contains 8 sites located in the urban/suburban areas in the southeastern U.S.STN: Speciated Trends Network. Contains 139 sites in urban areas in the U.S.

a Normalized mean bias.b Normalized mean error.c Normalized gross error.d Based on surface concentrations using 36 km horizontal grid spacing.e Normalized bias.f Normalized error.g 36 km horizontal resolution.h Nested 12 km horizontal resolution within 36 km outer grid.i 5 km horizontal resolution.j Nested 4 km horizontal resolution within 12 km outer grid.k 4 km horizontal resolution.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e37 27

the corresponding observed concentrations for all fourmonths, andthe mixing ratios of the maximum O3 concentration in January andJuly are over-predicted. Vijayaraghavan et al. (2006) evaluatedCMAQ’s performance against an episode from the Central California

Fig. 9. Daily average NOx emission in the San Joaquin Valley in 2005. Red dots indicatemonitoring site locations, and the numbers are the corresponding site numbers. Note:Emissions from Highway 101 are excluded in the simulations. (For interpretation of thereferences to colour in this figure legend, the reader is referred to the web version ofthis article.)

Ozone Study (CCOS) in July and August 2000. This study found thatCMAQ underestimates 1-hr O3 concentrations. Zhang et al. (2009)conducted a comprehensive study to evaluate CMAQ’s ability toreproduce the long-term variation of pollutants such as O3. Thestudy compared results from a CMAQ simulation of a full year toboth ground-based and satellite measurements; the normalizedmean bias ranges from 11% to 0.1% for the annual maximum 1-hr O3mixing ratio (Zhang et al., 2009).

The California Institute of Technology (CIT) Airshed model isanother comprehensive air quality model that is widely used(Cohan et al., 2008; Ensberg et al., 2010; Carreras-Sospedra et al.,2010) for regional air quality studies in Southern California. McNairet al. (1996) evaluated the CIT Airshed model against the SouthernCalifornia Air Quality Study (SCAQS; Lawson, 1990) database andconcluded that the CIT Airshed model is able to predict the diurnalvariation of the reactive species and the transport of the relativelynon-reactive species.

A comparison between the values listed in Table 1 and those ofTable 2 shows that the performance of LBM is comparable to that ofcomplexmodels such as CMAQ and CIT Airshed. TheNMB is less than45% and the NME is within 25e50%.

4. Evaluation of LBM using San Joaquin Valley data

The SJVAB is one of the 15 air basins located in California, USA.The SJVAB has an area of approximately 60,900 square kilometersand is surrounded by the Coastal Range Mountains to the west, theSierra Nevada Mountains to the east, the Transverse RangeMountains to the south, and the Sacramento Valley to the north.These mountain ranges give the Valley a bowl-shaped topography

Fig. 10. Annual average concentration of the 19 monitoring sites in the SJVAB. Upper left panel: NOx. Upper right panel: NO2. Lower panel: O3. Note: Statistics shown are calculatedexcluding site 9 and 11.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e3728

that retains air pollutants generated by the activities of the Valley’sthree million residents and their twomillion vehicles. The presenceof two major highways, CA 99 and Interstate 5, adds high vehicularemissions to the existing NOx emissions in the valley. The SanJoaquin Valley does not meet the 2008 8-h averaged ozone nationalambient air quality standards (NAAQS) of 75 ppb (Jin et al., 2011).The California Air Resources Board (CARB) has proposed a new stateimplementation (SIP) plan to attain the 1997 80 ppb 8-h averagedozone NAAQS in the San Joaquin Valley by June 15, 2024 (CARB,2007).

Several studies have been conducted to examine the formationof ozone and particulate matter in the SJVAB (See Table 2 forrelevant studies). Here we describe the results from the applicationof LBM to the SJVAB for the year 2005. The emission inventory,provided by Samuelsen et al. (2010), consists of an 80 � 89 grid of4 � 4 km squares (Fig. 9). Since the total NOx emission in this in-ventory was slightly different from that reported by CARB as theofficial inventory for the year of 2005, we scaled this inventory tomatch the 594.6 tons per day of total NOx emission in the SJVAB asreported by CARB (http://www.arb.ca.gov/ei/emissiondata.htm).High emissions occur primarily alongmajor roads such as Interstate5 and Highway 99, and at large cities such as Fresno and Bakers-field. The background ozone concentration is taken to be 30 ppb,and the VOC concentration is assumed to be a multiple of the NOx

concentration:

½VOC� ¼ ½NOx� � Ratioþ ½VOC�Background (6)

where the ratio is taken to be 6. This ratio is consistent withmeasurements of VOC/NOx ratios measured in the Los Angelesbasin (Fujita et al., 2003). We then add a background VOC con-centration of 40 ppbC for the simulation of NOx, NO2, and O3 in theSJVAB. This value is consistent with the minimum daily averagedconcentrations of non-methane hydrocarbons (NMHC) in the SJVABreported by the Air Quality and Meteorological Information System(AQMIS) of the CARB for the year, 2005 (http://www.arb.ca.gov/aqmis2/aqmis2.php).

There are 28 ambient monitoring sites in the San Joaquin Valleystudy domain as shown in Fig. 9. The red dots indicate the locationof the 28 sites and the numbers are the corresponding sitenumbers. We simulated the concentrations at the 21 monitoringsites that are located in the valley. The 2005 hourly NO2, NOx, andO3 data at the 21 sites were obtained from the CARBwebsite (CARB,2012). Sites 9 and 11 were excluded from this evaluation studybecause the concentrations were overestimated by a large amountand the evaluation at these two sites is not representative of themodel performance. The reason for this overestimation might bedue to the uncertainties in the gridded emission inventory. Bothsites 9 and 11 are located in the suburbs of Fresno and Bakersfieldand might have much lower emissions than that indicated by thegridded emission inventory. Wind speeds and directions measuredat 11 meteorological stations in the SJVAB were used to computeback trajectories. The GRS mechanism was used to model thechemical processes.

Fig. 10 shows model performance in describing annually aver-aged 1-hr NOx, NO2, and O3 concentrations of 19monitoring sites in

Fig. 11. Comparison of modeled daily-average concentrations with the corresponding observed concentrations. Upper panels: NOx, middle panels: NO2, and lower panels: O3. Leftpanels correspond to Site 5, and right panels correspond to Site 14.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e37 29

the San Joaquin Valley. More than 85% of the modeled NOx and 90%of the NO2 concentrations are within a factor of two of the corre-sponding observed concentrations. The predicted annually aver-aged 1-hr O3 concentrations for all 19 monitoring sites are within afactor of two of observations.

Fig. 10 shows that LBM provides a good description of the spatialvariation of NOx and NO2 at regional scales; the correlation co-efficients (r2) are 0.55 and 0.67, respectively. Fig. 10 also indicatesthat the observed annual averaged O3 in the SJVAB varies over anarrow range (20e30 ppb), except for stations 12, 17, and 20 (therelative standard deviation of the observed O3 concentrations isabout 12%). This suggests that the annual averaged O3 is mostlydriven by the background O3, and that the high ozone events arerelatively infrequent. With mg close to unity (¼1.02, 1.11, and 1.04

for NOx, NO2 and O3, respectively), the model shows little or no biasin estimating the annual average NOx, NO2 and O3 concentrations.

We next examine the performance of LBM in the San JoaquinValley in more detail at two of the 19monitoring sites: Site 5, whichis located near Bakersfield, and Site 14, which is located in Shafter, atown 26 kmnorthwest of Bakersfield. Fig.11 compares themodeleddaily averaged NOx, NO2, and O3 concentrations to the corre-sponding observed concentrations at these two sites. The leftpanels of the figure show that the model provides an adequatedescription of the daily average NOx, NO2, and O3 concentrations atBakersfield, while it slightly underestimates and overestimates theNOx and NO2 concentrations in Shafter, respectively. More than 80%of the modeled NOx and NO2 concentrations are within a factor oftwo of the corresponding observed concentrations. Even though

Fig. 12. Comparison of modeled daily maximum concentrations with the observed concentrations for sites 5 and 14 from JanuaryeDecember 2005. Upper panels: NOx, middlepanels: NO2, and lower panels: O3. Left panels correspond to Site 5, and right panels correspond to Site 14.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e3730

the correlation (r2) between the modeled and observed dailyaverage NOx and NO2 ranges from 0.25 to 0.4, the modeled dailyaverage O3 concentration correlates well with the observed O3concentration as seen in the lower panels of Fig. 11 (r2 is 0.58 and0.79 at site 5 and 14, respectively).

Fig. 12 compares the modeled daily maximum NOx, NO2,and O3 concentrations to the observed concentrations atsites 5 (Bakersfield) and site 14 (Shafter) for thewhole year of 2005.We see that the model performs reasonably well at site 5 while atsite 14 it slightly under/overestimates NOx and NO2 concentrations,respectively. At both sites 5 and 14, about 90% of the modeled dailymaximum O3 concentrations are within a factor of two of the cor-responding observed concentrations. Statistics of the model per-formance show that the scatter (sg2) of the predicted dailymaximumO3 is less than 2.3 (sg2¼ 2.28 at site 5), and the bias (mg) is20% and 11% at site 5 and 14, respectively. The correlation of

observed and predicted daily maximum O3 is relatively high at site14 (r2 ¼ 0.61) while it is 0.25 at site 5.

The ability of LBM in reproducing the seasonal variations of NO2,NOx and O3 concentrations is depicted in Fig. 13. Table 3 providesperformance measures of LBM in describing the daily average andmaximum NO2 and O3 concentrations at two sites in the SJVAB.

In order to compare the performance of LBM to comprehensivemodels such as CMAQ, we analyzed two O3 episodes in site 5 (July10e14, 2005) and site 14 (June 11e15, 2005) as shown in Fig. S3,and evaluated the performance of the model for these two epi-sodes. Results from this evaluation were compared to those fromVijayaraghavan et al. (2006) presented in Table 2. This comparisonrevealed that the performance of LBM in predicting the dailymaximum O3 is comparable to that of the comprehensive regionalphotochemistry models such as CMAQ. The NMB values for dailymaximum O3 from Vijayaraghavan et al. (2006) are between 3.9%

Fig. 13. Modeled monthly averaged concentration compared with corresponding observed concentration for each month of the year at site 5 and 14. Upper panels: NOx, middlepanels: NO2, and lower panels: O3. Left panels correspond to Site 5, and right panels correspond to Site 14.

Table 3Statistical performance measures of the LBM model for calendar year 2005 at two sites (Bakersfield and Shafter) in SJVAB.

Species NMBa (%) NMEb (%) r2

Bakersfield Shafter Bakersfield Shafter Bakersfield Shafter

NO2 Daily max 17.9 25.6 46.8 48.4 0.01 0.06Daily averaged 13.5 �0.6 40.4 33 0.26 0.39

O3 Daily max 17.5 4.9 32.3 18.3 0.25 0.61Daily averaged 2.3 7.7 25.5 19.4 0.58 0.79

a Normalized mean bias.b Normalized mean error.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e37 31

Fig. 14. Modeled annual averaged daily variation of NOx, NO2 and O3 concentration compared with corresponding observations. Upper panels: NOx, middle panels: NO2, and lowerpanels: O3. Left panels correspond to Site 5, and right panels correspond to Site 14.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e3732

and 61.1%, while the NMB of LBM are 9% and 8% for Site 14 and 5,respectively. Similarly, the NME values from LBM are in the range of8%e9%, while CMAQ shows values of 15%e60% (Vijayaraghavanet al., 2006).

Fig. 14 shows that the modeled daily variations averaged over ayear correlate well with the corresponding observed variations. Atsite 5, the modeled NOx concentrations clearly show a diurnalvariation with the maximum occurring during rush hours, whichdiffers slightly from the observed variations. At site 14, the modelslightly overestimates NOx concentration from midnight to 5am.The lower panels of Fig. 14 show that the model can estimate thedaytime O3 concentration well, but that it slightly overestimatesduring nighttime. Although emissions are relatively low during

nighttime, the high nighttime NOx concentrations are related to thechoice of the boundary layer height, the estimation of which ishighly uncertain (Pournazeri et al., 2012). The LBM program isimplemented in MatLab, and the one year simulation of hourlyconcentrations at 28 receptors in the SJVAB took approximately145 min to run on a machine with a 4 core Intel i7-920 2.67 GHzprocessor and 6 GB of ram.

5. Comparison of O3 time series

Here we examine the performance of LBM in simulating highozone episodes from the two air basins (SoCAB and SJVAB). In theepisode selected for the SoCAB during September 1e4, 2007, a

Fig. 15. O3 episodes during September 1e4, 2007 in Azusa in the SoCAB (upper panel)and July 14e17, 2005 in Fresno in the SJVAB (lower panel).

Fig. 16. Comparison of modeled and observed annual fourth-highest daily maximum8-h average ozone concentrations for 19 stations in the SoCAB (upper panel) and 12stations in the SJVAB (lower panel). The monitoring sites were selected in order to haveuniform spatial distribution over the air basin.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e37 33

maximum hourly averaged O3 of 158 ppb was recorded in the cityof Azusa located on the northeast of Los Angeles County. In theSJVAB, one of the ozone episodes occurred during July 14e17, 2005,when the hourly averaged O3 concentration reached 131 ppb inFresno.

Fig. 15 shows that for the O3 episode in the SoCAB, the modeledconcentrations are in good agreement with the observed concen-trations, except that the model slightly underestimates themaximum O3 concentration that occurred on September 3. In theSJVAB, the estimated concentrations are in good agreement withobserved data during daytime, but the model fails to explain therelatively high O3 concentrations during nighttime. Estimated O3concentrations decrease rapidly to a value close to zero and staylow throughout the night while the observed O3 concentrationsdecrease at a much slower rate throughout the entire night toapproximately 20 ppb just before sunrise. This again could berelated to uncertainties in estimating the nighttime boundaryheight. The modeled boundary layer height decreases rapidly aftersunset, which leads to an increase in NO concentrations and resultsin rapid destruction of O3. An overestimation of the nighttime NOx

emission could also contribute to the discrepancy between theobserved and modeled O3 destruction rate. In addition, the modelslightly overestimates the daily peak O3 concentrations on July14th, 16th, and 17th.

6. Application of LBM in air quality analysis

The results from the evaluation studies described in thepreceding sections indicates that LBM provides a description ofthe spatial and temporal variation of NOx, NO2, and O3 overregional scales that is comparable to that from more compre-hensive models. In this section, we provide examples of appli-cations that can take advantage of the computational efficiencyof LBM.

6.1. Attainment of O3 National Ambient Air Quality Standard

The National Ambient Air Quality Standards (NAAQS) defined in1979 required that the 1-h averaged ozone concentrations cannotexceed 120 ppb for more than one hour per year. In July 1997, theozone standard was replaced with a more conservative 8-h stan-dard that was set at a level of 80 ppb. Later in March 2008, thisstandard was set at the more stringent level of 75 ppb averagedover an 8-h period. In order to meet this standard, the annualfourth-highest daily maximum 8-h average ozone concentration ateach monitoring station should be less than or equal to 75 ppb.

In this study, we examined the performance of LBM in esti-mating the annual fourth-highest daily maximum 8-h averageozone concentrations in both the SoCAB and the SJVAB. Resultsfrom this assessment, shown in Fig. 16, reveal that, except forcertain receptors in the SoCAB and the SJVAB, model estimatesagree well with those obtained from observed data. As shown inFig. 16, both modeled and observed data in the SoCAB indicate thatozone concentrations are low near the coast and increase at re-ceptors located further inland. Furthermore, the model performsvery well in determining the non-attainment areas where dailymaximum 8-h average ozone concentrations exceed 75 ppb. Thiscan be seen in Table 4, which shows the number of areas classified

Fig. 17. Cross-validation of Kriging interpolation in the SoCAB. Left panel: Simple Kriging of NOx and NO2 concentrations. Right panel: Residual Kriging of NOx and NO2

concentrations.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e3734

as attainment/non-attainment by LBM that are also classifiedattainment/non-attainment based on observations. Table 4 andFig. 16 also show that the model over-estimates ozone in thenorthwest of the SJVAB, leading to a prediction of non-attainmentof NAAQS at five sites where observations show attainment; how-ever, these sites are near non-attainment, and there is only one sitewhere the model falsely predicts attainment of NAAQS. Thisassessment shows that LBM can be employed as a useful tool toscreen NAAQS attainment demonstrations or uniform rate ofprogress assessments.

6.2. Improving Kriging interpolation using LBM for exposure studies

In epidemiological studies, exposure to air pollution is esti-mated using concentrations measured at several sites within a re-gion of interest. It is difficult to capture the spatial variation of theconcentration field without a dense network of monitoring sites(Johnson et al., 2010). Methods such as LUR (Brauer et al., 2003;Ross et al., 2007) use information on factors such as land use that

Table 4Comparison of attainment (A) and non-attainment (NA) of the O3 NAAQS based onobservations and LBM prediction.

Modeled Modeled

SoCAB NA A SJVAB NA A

Observed NA 12 0 Observed NA 5 1A 4 3 A 5 1

might influence concentration gradients to enhance the interpo-lation of observed concentrations. Here we show how LBM can beused to provide the information required to add value to theinterpolation of observations.

Geo-statistical interpolation techniques such as Kriging(Matheron, 1971) are commonly used to interpolate observationsmade at irregularly spaced locations, such as air quality monitoringsites. Simple Kriging assumes that observation fields can be rep-resented as the sum of a spatially constant mean and a local fluc-tuation. The statistics of the fluctuating component are oftenassumed to be spatially isotropic and homogeneous. These as-sumptions are not likely to be true for concentration fields whoseunderlying spatial structure is governed by those of emissions andmeteorology. We can improve on simple Kriging by using adispersion model to remove the underlying structure in the con-centration field (Venkatram, 1988). The residuals between modelpredictions and observations are then likely to meet the isotropyand homogeneity assumptions required by simple Kriging. Weillustrate this approach to interpolation, which we refer to as re-sidual Kriging, by using LBM to construct the underlying trend ofNOx and NO2 concentrations. Isakov et al. (2012) presented a similarapproach, where they used a hybrid regional-local air quality modelto improve the LUR technique.

We evaluated the efficacy of LBM in enhancing the results ofsimple Kriging using data from the SoCAB and the SJVAB. There areseveral techniques for evaluating the performance of interpolation,the details of which are in EPA (2004). Here we used the mostcommon technique, known as leave-one-out cross validation(LOOCV): the observation at a selected location is left out in the

Fig. 18. Cross-validation of Kriging interpolation in the SJVAB. Left panel: Simple Kriging of NOx and NO2 concentrations. Right panel: Residual Kriging of NOx and NO2

concentrations.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e37 35

interpolation, and the interpolated value at this location is thencompared with the actual observation. Fig. 17 and 18 compare thecross-validation of simple Kriging and residual Kriging. We see thatresidual Kriging improves the correlation between observed NOx

and NO2 concentrations over simple Kriging. In the SoCAB, thecorrelation coefficient (r2) increases from 0.06 to 0.42 and from 0.17to 0.29 for NOx and NO2, respectively. We find a similar result forthe SJVAB data: r2 increases from 0.17 to 0.57 for NOx, and from 0.16to 0.45 for NO2. Scatter between the observed and estimated con-centrations also decreases as indicated by the smaller sg (except forNO2 in the SoCAB). These results suggest that LBM can be used toremove the underlying trend in data, and thus enhance interpola-tion methods, such as Kriging.

Fig. 19 shows the 4-Day average NO2 concentration map in theSoCAB using both simple and residual Kriging to interpolate theNO2 concentrations at the 994 grid points. We see that residualKriging predicts that the majority of the high NO2 concentrationsoccur near major cities such as Los Angeles and Ontario, as well asthemajor roads. Simple Kriging shows reasonable spatial variationsof NO2 in the central regionwhere most of the monitoring sites arelocated, but it fails to explain the spatial variation of NO2 near themajor roads and in regions far from emission sources, such as thesoutheast corner and the central-north region, where limitednumbers of observations are available. The results of residualKriging are more reasonable because they show greater variationnear emission sources such asmajor roads. The concentration mapsfurther demonstrate the value of LBM in residual Kriging andassessing exposures.

7. Summary and conclusions

Wehave formulated a simple Lagrangianmodel that can be usedto estimate “background” concentrations of NOx, NO2, and O3 overspatial scales of a kilometer in a domain extending over hundredsof kilometers. The model achieves computational efficiency byseparating transport and chemistry using the concept of speciesage. Evaluation with measurements made in the SoCAB during2007 indicates that the model provides an adequate description ofthe spatial and temporal behavior of NOx and NO2. Model estimatesof maximum hourly ozone concentrations are unbiased relative toobservations and the 95% confidence interval ðzs2gÞ of the ratios ofobserved to estimated concentrations is slightly over a factor oftwo. The model was also evaluated against measurements made inthe SJVAB in 2005. Although the model shows slight under/over-estimation of NOx and NO2 concentrations respectively, it providesa reasonable description of the temporal (daily and seasonal) aswell as spatial variation of NO2 and O3 concentrations in the region.

The results presented in this paper show that a relatively simplemodel can be used to relate concentrations of NOx, NO2, and O3averaged over several kilometers to emissions of NOx and VOC. Thismodel can be used to make first order estimates of the impact ofprecursor emission changes on the concentration of these pollut-ants. The study also shows that the estimated ozone concentrationsfrom LBM can be used to judge attainment of statistically basedambient air quality standards.

LBM is computationally efficient and relatively easy to imple-ment, which makes it an appropriate candidate for long-term

Fig. 19. 4-Day average NO2 concentration map in the SoCAB. Upper panel: Interpolated using Simple Kriging. Lower panel: Interpolated using residual Kriging.

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e3736

exposure studies. We have also demonstrated the use of LBM toenhance interpolation of observations. Stein et al. (2007) andIsakov et al. (2009) combined results from CMAQ with a short-range transport model such as AERMOD to assess personal expo-sure; similarly, results from LBM can also serve as inputs to ashort-range dispersion model to estimate the impact of a source ofNOx on NO2 and ozone at a scale of tens of meters from the source.The model can provide hourly concentrations of these species overtime periods of a year, which is required in human exposurestudies.

Acknowledgments

This work was supported by the California Energy Commission,contract number 500-08-055.

Appendix A. Supplementary data

Supplementary data related to this article can be found at http://dx.doi.org/10.1016/j.envsoft.2013.10.018.

References

Azzi, M., Johnson, G.M., Cope, M., 1992. An introduction to the generic reaction setphotochemical smog mechanism. In: Proceedings of the 11th InternationalConference of Clean Air Soc. of Aust. and N.Z., Brisbane, pp. 451e462.

Berkowicz, R., 1999. OSPM e a parameterised street pollution model. In: 2nd In-ternational Conference on Urban Air Quality, Madrid.

Berkowicz, R., 2000. A simple model for urban background pollution. Environ.Monit. Assess. 65, 259e267.

Brandt, J., Christensen, J.H., Frohn, L.M., Geernaert, G.L., Berkowicz, R., 2000a.Development of a new operational air pollution forecast system on regional andurban scale. In: Schiermeier, G. (Ed.), Millennium NATO/CCMS InternationalTechnicalMeeting on Air PollutionModelling and its Application XIV,15e19May2000, Boulder Colorado,USA. KluwerAcademic/PlenumPublishers, pp. 573e581.

Brandt, J., Christensen, J.H., Frohn, L.M., Berkowicz, R., Palmgren, F., 2000b. TheDMU-ATMI THOR Air Pollution Forecast Systemeesystem Description. NERITechnical Report No. 321. National Environmental Research Institute, Roskilde,Denmark, 60 pp.

Brandt, J., Christensen, J.H., Frohn, L.M., Palmgren, F., Berkowicz, R., Zlatev, Z., 2001a.Operational air pollution forecasts from European to local scale. Atmos. Environ.35 (1), 91e98.

Brandt, J., Christensen, J.H., Frohn, L.M., Berkowicz, R., 2001b. Operational airpollution forecast from regional scale to urban street scale. Part 1: systemdescription. Phys. Chem. Earth (B) 26 (10), 781e786.

Brandt, J., Christensen, J.H., Frohn, L.M., 2001c. Operational air pollution forecastfrom regional scale to urban street scale. Part 2: performance evaluation. Phys.Chem. Earth (B) 26 (10), 825e830.

Brandt, J., Christensen, J.H., Frohn, L.M., Zlatev, Z., 2001d. Operational air pollutionforecast modelling by using the THOR system. Phys. Chem. Earth (B) 26 (2),117e122.

Brandt, J., Christensen, J.H., Frohn, L.M., 2001e. Traffic Scenarios for the City ofAalborg Using the DMU-ATMI THOR Model System. Technical Report. NationalEnvironmental Research Institute, Frederiksborgvej 399, P.O. Box 358, DK-4000Roskilde, Denmark, p. 20.

Brandt, J., Christensen, J.H., Frohn, L.M., Berkowicz, R., 2003. Air pollution fore-casting from regional to urban street scale e implementation and validation fortwo cities in Denmark. Phys. Chem. Earth 28, 335e344.

Brauer, M., Hoek, G., van Vliet, P., Meliefste, K., Fischer, P., Gehring, U., Heinrich, J.,Cyrys, J., Bellander, T., Lewne, M., Brunekreef, B., 2003. Estimating long-term

S. Pournazeri et al. / Environmental Modelling & Software 52 (2014) 19e37 37

average particulate air pollution concentrations: application of traffic indicatorsand geographic information systems. Epidemiology 14, 228e239.

Byun, D., Schere, K.L., 2006. Review of the governing equations, computational al-gorithms, and other components of the models-3 community multiscale airquality (CMAQ) modeling system. Appl. Mech. Rev. 59, 51e77.

CARB, 2007. 2007 San Joaquin Valley 8-hour Ozone Plan. Air Resources Board,Sacramento, CA. http://www.arb.ca.gov/planning/sip/2007sip/sjv8hr/sjvozone.htm.

CARB, 2012. California Air Quality Data for the Years 1980e2009. Air ResourcesBoard, Sacramento, CA. http://www.arb.ca.gov/aqd/aqdcd/aqdcddld.htm.

Carreras-Sospedra, M., Vutukuru, S., Brouwer, J., Dabdub, D., 2010. Central powergeneration versus distributed generation e an air quality assessment in theSouth Coast Air Basin of California. Atmos. Environ. 44 (26), 3215e3223.

Carson, D.J., 1973. The development of a dry inversion-capped convectively unstableboundary layer. Q. J. R. Meteorol. Soc. 99, 450e467.

CERR, 2002. Consolidated Emission Reporting Rule, Federal Register: 10 June 2002,Vol. 67. No. 111, Proposed Rules, pp. 39602e39616.

Cimorelli, A.J., Perry, S.G., Venkatram, A., Weil, J.C., Paine, R., Wilson, R.B., Lee, R.F.,Peters, W.D., Brode, R.W., 2005. AERMOD: a dispersion model for industrialsource applications. Part I: general model formulation and boundary layercharacterization. J. Appl. Meteorol. 44 (5), 682e693.

Cohan, A., Chang, W., Carreras-Sospedra, M., Dabdub, D., 2008. Influence of sea-saltactivated chlorine and surface-mediated renoxification on the weekend effectin the South Coast Air Basin of California. Atmos. Environ. 42 (13), 3115e3129.

Dabdub, D., Carreras-Sospedra, M., Chang, W., Cohan, A., 2008. How New ChemistryFindings Affect Our Understanding of the Weekend Effect e a Modeling Study.In: Final Report to the California Air Resources Board. ARB Contract 04-333.

Eder, B., Yu, S., 2006. A performance evaluation of the 2004 release of models-3CMAQ. Atmos. Environ. 40, 4811e4824.

Eliassen, A., Saltbones, J., 1983. Modelling of long-range transport of sulphur overEurope: a two-year model run and some model experiments. Atmos. Environ.17 (8), 1457e1473.

Ensberg, J.J., Carreras-Sospedra, M., Dabdub, D., 2010. Impacts of electronicallyphoto-excited NO2 on air pollution in the South Coast Air Basin of California.Atmos. Chem. Phys. 10 (3), 1171e1181.

EPA, November 2004. Developing Spatially Interpolated Surfaces and EstimatingUncertainty. U.S. Environmental Protection Agency, Office of Air and Radiation,Office of Air Quality Planning and Standards, Research Triangle Park, NC. EPA-454/R-04-004 http://www.epa.gov/airtrends/specialstudies/dsisurfaces.pdf.

Fujita, E.M., Stockwell, W.R., Campbell, D.E., Keislar, R.E., Lawson, D.R., 2003. Evo-lution of the magnitude and spatial extent of the weekend ozone effect inCalifornia’s South Coast Air Basin, 1981e2000. J. Air Waste Manag. Assoc. 53 (7),802e815.

Gulliver, J., Briggs, D., 2011. STEMS-Air: a simple GIS-based air pollution dispersionmodel for city-wide exposure assessment. Sci. Total Environ. 409 (12), 2419e2429.

Hurley, P., Manins, P., Lee, S., Boyle, R., Ng, Y.L., Dewundege, P., 2003. Year-long,high-resolution, urban airshed modelling: verification of TAPM predictions ofsmog and particles in Melbourne, Australia. Atmos. Environ. 37 (14), 1899e1910.

Isakov, V., Johnson, M., Touma, J., Özkaynak, H., 2012. Development and evaluationof land-use regression models using modeled air quality concentrations. AirPollut. Model. Its Appl. XXI, 717e722.

Isakov, V., Touma, J.S., Burke, J., Lobdell, D.T., Palma, T., Rosenbaum, A., Özkaynak, H.,2009. Combining regional- and local-scale air quality models with exposuremodels for use in environmental health studies. J. Air Waste Manag. Assoc. 59(4), 461e472.

Jensen, S.S., Larson, T., Deepti, K.C., Kaufman, J.D., 2009. Modeling traffic air pollu-tion in street canyons in New York City for intra-urban exposure assessment inthe US multi-ethnic study of atherosclerosis and air pollution. Atmos. Environ.43 (30), 4544e4556.

Jin, L., Harley, R.A., Brown, N.J., 2011. Ozone pollution regimes modeled for asummer season in California’s San Joaquin Valley: a cluster analysis. Atmos.Environ. 45 (27), 4707e4718.

Johnson, G.M., 1984. A simple model for predicting the ozone concentration ofambient air. In: Proceedings of the Eighth International Clean Air Conference,vol. 2, pp. 715e731.

Johnson, M., Isakov, V., Touma, J.S., Mukerjee, S., Özkaynak, H., 2010. Evaluation ofland use regression models used to predict air quality concentration in an urbanarea. Atmos. Environ. 44 (30), 3660e3668.

Lawson, D.R., 1990. The Southern California air quality study. J. Air Waste Manag.Assoc. 40, 156e165.

Li, S.M., 2004. A concerted effort to understand the ambient particulate matter inthe Lower Fraser Valley: the Pacific 2001 air quality study. Atmos. Environ. 38,5719e5731.

Liu, X.-H., Zhang, Y., Cheng, S.-H., Xing, J., Zhang, Q., Streets, D.G., Jang, C., Wang, W.-X., Hao, J.-M., 2010. Understanding of regional air pollution over China usingCMAQ, part I performance evaluation and seasonal variation. Atmos. Environ.44, 2415e2426.

Matheron, G., 1971. The Theory of Regionalized Variables and Its Applications, vol. 5.Ecole des Mines de Paris, Fontainebleau.

Matthias, V., Aulinger, A., Quante, M., 2008. Adapting CMAQ to investigate airpollution in North Sea coastal regions. Environ. Model. Softw. 23 (3), 356e368.

McNair, L.A., Harley, R.A., Russell, A.G., 1996. Spatial inhomogeneity in pollutantconcentrations, and their implications for air quality model evaluation. Atmos.Environ. 30, 4291e4301.

Pournazeri, S., Venkatram, A., Princevac, M., Tan, S., Schulte, N., 2012. Estimating theheight of the nocturnal urban boundary layer for dispersion applications.Atmos. Environ. 54, 611e623. http://dx.doi.org/10.1016/j.atmosenv.2012.02.024.

Ross, Z., Jerrett, M., Ito, K., Tempalski, B., Thurston, G.D., 2007. A land use regressionfor predicting fine particulate matter concentrations in the New York City re-gion. Atmos. Environ. 41, 2255e2269.

Samuelsen, S., Dabdub, D., Brouwer, J., Carreras-Sospedra, M., Vutukuru, S., 2010. AirQuality Impacts of Distributed Generation in the South Coast Air Basin and theSan Joaquin Valley. California Energy Commission. PIER Energy-Related Envi-ronmental Research, CEC-500-2009-070.

Samuelsen, S., Dabdub, D., Brouwer, J., Medrano, M., Rodriguez, M., Carreras-Sosedra, M., 2005. Air Quality Impacts of Distributed Generation. CaliforniaEnergy Commission. PIER Energy-Related Environmental Research, CEC-500-2005-069-F.

Smyth, S.C., Jiang, W., Yin, D., Roth, H., Giroux, É., 2006. Evaluation of CMAQ O3 andPM2.5 performance using Pacific 2001 measurement data. Atmos. Environ. 40,2735e2749.

Stein, A.F., Isakov, V., Godowitch, J., Draxler, R.R., 2007. A hybrid modeling approachto resolve pollutant concentrations in an urban area. Atmos. Environ. 41 (40),9410e9426.

Venkatram, A., 1980. Estimating the Monin-Obukhov length in the stable boundarylayer for dispersion calculations. Boundary Layer Meteorol. 19, 481e485.

Venkatram, A., 2008. Computing and displaying model performance statistics.Atmos. Environ. 42, 6862e6868.

Venkatram, A., Cimorelli, A.J., 2007. On the role of nighttime meteorology inmodeling dispersion of near surface emissions in urban area. Atmos. Environ.41, 692e704.

Venkatram, A., Du, S., Hariharan, R., Carter, W., Goldstein, R., 1998. The concept ofspecies age in photochemical modeling. Atmos. Environ. 32, 3403e3413.

Venkatram, A., Karamchandani, P., Pai, P., Goldstein, R., 1994. The development andapplication of a simplified ozone modeling system (SOMS). Atmos. Environ. 28,3665e3678.

Venkatram, A., 1988. On the use of kriging in the spatial analysis of acid precipi-tation data. Atmos. Environ. 22 (9), 1963e1975.

Vijayaraghavan, K., Karamchandani, P., Seigneur, C., 2006. Plume-in-grid modelingof summer air pollution in central California. Atmos. Environ. 40, 5097e5109.

Zhang, Y., Vijayaraghavan, K., Wen, X.-Y., Snell, H.E., Jacobson, M.Z., 2009. Probinginto regional ozone and particulate matter pollution in the United States: 1. A 1year CMAQ simulation and evaluation using surface and satellite data.J. Geophys. Res. 114, D22304.