ARI Report No. RR-1330
ARI Contract No. 10066
Measurements and Analyses of
Urban Metabolism and Trace Gas Respiration
Prepared By
J.B. McManus, J.H. Shorter, M.S. Zahniser and C.E. Kolb
Center for Atmospheric and Environmental Chemistry
Aerodyne Research, Inc.
Billerica, MA 01821-3976
S.M. O’Neill, D. Stock, S. Napelenok, E.J. Allwine and B.K.
Lamb
Laboratory for Atmospheric Research
Washington State University
Pullman, WA 99164-2910
E. Scheuer and R.W. Talbot
Institute for the Study of Earth, Oceans and Space
University of New Hampshire
Durham, NH 03824-3525
F. San Martini, G. Adamkiewicz, B.K.-L.l Pun, C. Wang, and G.J. McRae
Department of Chemical Engineering
Massachusetts Institute of Technology
Cambridge, MA 01239-4307
L. Cao, A.A. Ismail, M. Kawabata, C.-H. Yeang, G. Narasimhan,
S. Humbad, M. Zhang, and J. Ferreira, Jr.
Department of Urban Studies and Planning
Massachusetts Institute of Technology
Cambridge, MA 01239-4307
Prepared For
Office of Earth Sciences
National Aeronautic and Space Administration
Washington, DC 20546-0001
May 2002
Table of Contents
1.0
PROJECT OVERVIEW .........................................................................................1
1.1
Motivation
................................................................................................2
1.2
Project Team Descrition ................................................................................2
1.3
Experiment Goals...........................................................................................2
1.4
Analysis Goals ...............................................................................................3
1.5
Project Accomplishments and Impacts ..........................................................4
2.0
FIELD MEASUREMENT STRATEGIES ............................................................6
2.1
Instrument Overview .....................................................................................6
2.1.1 Tunable Infrared Laser Sensors .........................................................6
2.1.1.1 Background ............................................................................6
2.1.1.2 Laser Sources .........................................................................7
2.1.1.3 Instrument Description...........................................................8
2.1.1.4 Signal Processing ...................................................................10
2.1.1.5 Instrument Operation .............................................................11
2.1.2 Commercial Licor CO2/UV Ozone/Eppley UV Sensors ...................12
2.1.3 Fine Aerosol Measurements (Condensation Particle Counter) ..........13
2.1.4 Tracer Release and Detection Instrumentation ..................................13
2.1.5 VOC Sampling and Analysis Instrumentation ...................................14
2.1.6 Sodar and Meteorological Instrumentation ........................................17
2.2
ARI Mobile Laboratory .................................................................................18
2.3
WSU Mobile Laboratory ...............................................................................19
2.4
Field Measurement Sites ................................................................................23
2.4.1 Manchester, NH .................................................................................23
2.4.2 Boston, MA ........................................................................................23
2.4.3 Cambridge, Massachusetts .................................................................23
2.5
Field Measurement Strategies ........................................................................25
2.5.1 Pollution Mapping .............................................................................25
3.0
FIELD DATA OVERVIEW ...................................................................................28
3.1
ARI Trace Gas Data Description ...................................................................30
3.2
UNH Total Particle Data Description ............................................................32
3.3
WSU Trace Data Description ........................................................................33
3.4
WSU VOC Data Description .........................................................................34
3.5
WSU Sodar and Meteorological Data Description ........................................34
4.0
DATA ANALYSIS STRATEGIES ........................................................................35
4.1
Motor Vehicle Pollutant from Mobile Measurements ...................................37
4.1.1 General Results of Mobile Measurements .........................................38
4.1.2 Special Issues for Mobile Measurements of ER’s .............................41
4.1.3 Separation of “Peaks” and “Local Background” ...............................45
4.1.4 Methods of Deriving Emission ratios ................................................45
4.1.5 NO Emission Ratio Results ...............................................................49
4.1.6 CO and CH4 Emission Results...........................................................62
iii
4.2
4.3
4.4
4.5
4.6
4.7
4.8
4.9
4.10
4.1.7 Discussion ..........................................................................................62
Background Pollutant Maps ...........................................................................69
Fixed Site Pollutant Measurement Analysis ..................................................72
Mesoscale Wind Field Modeling ...................................................................76
Turbulence Modeling of Urban Landscapes ..................................................80
4.5.1 Tracer Data Analysis..........................................................................80
4.5.2 Turbulence Modeling of an Urban Landscape...................................83
Urban Footprint Modeling .............................................................................85
Urban Emissions – Air Quality Relationships ...............................................92
4.7.1 Introduction ........................................................................................92
4.7.2 The Need for Better Emissions Inventories in Control
Strategy Design ......................................................................92
4.7.3 Inverse Modeling ...............................................................................93
4.7.4 Application to Atmospheric Systems.................................................94
4.7.5 Optimal Field Determination .............................................................95
4.7.6 Empirical Karhunen-Loeve Series Expansion ...................................96
4.7.7 Los Angeles Case Study ....................................................................99
4.7.8 Formulation of Optimization Problem ...............................................100
4.7.9 Pseudo-Data Inversion .......................................................................104
4.7.10 CO Inversion ......................................................................................104
4.7.11 Objective Function Modification .......................................................112
4.7.12 Summary of Results from Inverse Methods ......................................112
4.7.13 Role of Uncertainty ............................................................................114
4.7.14 Probabilistic Collocation Approach ...................................................114
4.7.15 Applications .......................................................................................117
4.7.16 Uncertain Parameters .........................................................................118
4.7.17 Results: Ozone Uncertainty and Variance Analysis .........................120
4.7.18 Observations/Conclusions..................................................................125
4.7.19 Conclusions for Inverse Modeling and Uncertainty Analysis ...........133
Model Inversion of Pollutant Maps: Diffusion Modeling of SF6
Release Experiments ...................................................................................134
4.8.1 Introduction ........................................................................................134
4.8.2 Observed Phenomena.........................................................................134
4.8.3 Experimental Data .............................................................................135
4.8.4 Model Description .............................................................................147
4.8.5 Results and Discussion ......................................................................151
4.8.6 Conclusions and Recommendations ..................................................168
Photochemical Steady State NOx Analyses ...................................................169
GIS Based Emissions Analyses .....................................................................175
4.10.1 GIS Analyses for Manchester, NH ....................................................175
4.10.1.1 Creating Base Maps ............................................................177
4.10.1.2 The Distributed GIS Architecture .......................................178
4.10.1.3 Visualizing Mobile Measurements of
Trace Gas Concentration..................................................182
iv
Table of Contents (Continued)
4.11
4.12
4.10.1.4 Geo-Processing Examples ..................................................183
4.10.1.5 Conclusions from Initial Work on
Distributed GIS Systems ..................................................186
4.10.2 GIS Analyses for Boston, MA ...........................................................187
4.10.2.1 Digital Orthophotography and GIS-Based Visualization ...187
4.10.2.2 Modeling Emission Sources ...............................................191
4.10.2.3 Stack Emission Model ........................................................193
4.10.2.4 Traffic Congestion Model ...................................................198
4.10.2.5 Emission Modeling Conclusions ........................................201
Comparing GIS-based Models and Trace Gas Observations ........................203
4.11.1 Spatial Regressions of the May 25, 1999, Trace Gas Observations ..205
4.11.2 System Architectures for Environmental Monitoring
and Modeling .................................................................................210
Fine Aerosol ................................................................................................212
5.0
PROJECT OUTPUT ...............................................................................................219
5.1
Symposia Presentations, Proceedings Papers and Journal Publications ........219
5.1.1 Presentations Without Published Proceedings Papers .......................219
5.1.2 Presentations With Published Proceedings Papers ............................220
5.1.3 Archival Journal Papers .....................................................................222
5.1.4 Graduate These Fully or Partially Supported.....................................222
5.2
Planned Archival Journal Papers ...................................................................223
5.3
Web Sites with Archived Project Data and Modeling/Analyses ...................223
6.0
SUMMARY
................................................................................................226
7.0
REFERENCES
................................................................................................228
v
LIST OF ILLUSTRATIONS
Figure
Page
2.1.1. Schematic for a Two-Laser TDL Instrument with 153 Meter Multiple
Pass Astigmatic Absorption Cell ...............................................................................4
2.1.2 Calculated Mirror Beam Spots for Two Patterns, Each with 182 Passes
Propogating in an Astigmatic Herriott Cell ...............................................................6
2.1.3 One-second TILDAS Spectra and Calculated Non-Linear Least Squares
Fit for 1 ppb Ambient Ethane and 1.7 ppm Methane ................................................6
2.1.4
SF6 Tracer Data Obtained May 25, 1998 in the South Boston Area ........................10
2.1.5
(a) Isoprene and (b) Toluene 1 Hour Integrated Canister Sample Results
for Manchester, NH on August 25, 1998 ...................................................................12
2.1.6
Comparison of MM5 and Sodar Wind Direction & Wind Speed at MIT ................13
2.3.1
(a) CO2, (b) CO Conentrations Measured by the WSU Van on May 25, 1999
in South Boston, MA ................................................................................................16
2.4.1 Boston Mobile Measurement Region, Comprising Principally of Dorchester
and Roxbury, MA
................................................................................................19
2.4.2 Stationary Field Measurement Site on the Campus of MIT in Cambridge, MA .......19
2.5.1 CO Data Collected at 1 Hz by the TDL System on June 16, 1998
in Manchester, NH ................................................................................................21
2.5.2 NO Data Collected at 1 Hz by the TDL System on May 25, 1999 in the
Boston, MA Area, Including the Routes to/from Billerica, MA ...............................22
4.1.1 Typical Segment of Mobile Concentration Data Showing Coincident Peaks
of CO2, NO and NO2 ................................................................................................38
4.1.2 Autocorrelations of Peak Segments of CO2 and NO Data, With Subtracted
Means
................................................................................................40
4.1.3 Cross-Correlation of Peak Segments (With Zero Mean) of CO2 and NO2
Showing the Strong Association Between These Gases ............................................40
4.1.4 Cosine of the Wind Direction as Sensed in the Truck, as a Function of
Truck Velocity, Assuming a General Wind from the North at 1 m/s. .......................42
vi
LIST OF ILLUSTRATIONS (Continued)
Figure
Page
4.1.5 Average CO2 Concentration for Velocity Bins During Traverses on
May 25, 1999.
................................................................................................43
4.1.6 Fourier Transforms of Data Records for CO2 and NO. .............................................44
4.1.7 A Data Sample with Local-Background Lines Derived from the
“Range-Minimum” Method .......................................................................................46
4.1.8 Example of CO2 and NO Data and the Emission Ratio as a Function of Time,
Determined Using Point Ratios and Linear Regression in a 5-Point
Sliding Window.
................................................................................................48
4.1.9 Histogram (Probability Density) of NO/CO2 ER from the SlidingWindow Regression Method......................................................................................49
4.1.10 Comparison of NO ER Distributions from Mobile Sampling in Boston on
5/25/99 and from a Cross-Road Remote Sensing Experiment Conducted in
California in 1996. ................................................................................................50
4.1.11 Measurement Route on 5/25/99 Color Coded by NO/CO2 ER..................................51
4.1.12 Histogram of NO/CO2 Emission Ratio as a Function of Sampling Speed ................52
4.1.13 Histogram of NO/CO2 Emission Ratio as a Function of Sampling Acceleration......52
4.1.14 Simple Model of Stop and Go Traffic, With (Raised) Sinusoidal Speed
and Circular Motion in the Speed-Acceleration Plane ..............................................53
4.1.15 City Driving Data Segment, Speed and Acceleration, with the Speed
Curve Color and Width Showing CO2 Concentration ...............................................53
4.1.16 A Segment of City Driving Data, with the Trace Color and Size
Indicating CO2 Concentration. ...................................................................................54
4.1.17 A Segment of City Driving Data, with the Trace Color and Size
Indicating CO2 Concentration. ...................................................................................55
4.1.18 Average CO2 Concentration as a Function of Speed and Acceleration
of the Mobile Lab, for City Driving in Boston on 5/25/99. .......................................56
4.1.19 Average ER (NO/CO2) as a Function of Speed and Acceleration of the
Mobile Lab, for City Driving in Boston on 5/25/99. .................................................57
vii
LIST OF ILLUSTRATIONS (Continued)
Figure
Page
4.1.20 Average CO2 Concentration as a Function of Driving Cycle Phase
for the Mobile Lab, for City Driving in Boston on 5/25/99.......................................58
4.1.21 Average ER (NO/CO2) as a Function of Driving Cycle Phase for the
Mobile Lab, for City Driving in Boston on 5/25/99. .................................................59
4.1.22 Scatterplot of N2O vs CO2 for City and Highway Driving in
Manchester, NH on 6/16/98, with Solid Line Showing the Linear
Regression Fit
................................................................................................60
4.1.23 Histogram of Pointwise Ratios for Mobile Peak Data (Dotted line),
N2O/CO2, City and Highway, and CO2 > 15 ppm. ..................................................61
4.1.24 Concentration Probability Distributions for NO and CO2 in Boston on 5/25/99. .....66
4.1.25 Computed Probability Distribution for a Gaussian Plume Model, with a
set of 20 Sources, each Emitting Two Different Gases at a Ratio of
Between 1 and 2.
................................................................................................67
4.2.1 Points Where CO2 Concentrations are Below the Maximum
Probability Concentration, in Boston on 5/25/99. .....................................................70
4.2.2 Interpolated Minima in CO2 Over a Rnage of +/-250 Meters During
Traverses on 5/25/99 in Boston. ................................................................................71
4.2.3 Interpolated Minima in NO Over a Range of +/-250 Meters During
Traverses on 5/25/99 in Boston. ................................................................................71
4.3.1 Measurement of Trace Gas Species at Harnett Park, Manchester, NH,
Site of an EPA Monitoring Station, and Surrounding Area on August 26, 1998. .....72
4.3.2 One Hour Averages and Standard Deviation of Each 1 Hour Data Set
of the 1 Hz NO, NO2, and CO2 Data from the ARI Mobile Laboratory. ...................74
4.3.3 Probability Distributions of NO, NO2 and CO2 as a Function of Time of Day. ........74
4.3.4 Probability Density of NO2, NO and CO2 for 12 1-hour Time Periods
Spanning from 22:00 EDT May 27 Until 20:00 EDT May 28. .................................75
4.4.1 (a) MM5 27 km Model Domain for Manchester, NH and
(b) MM5 3 km and 1 km Domains for Manchester, NH ...........................................77
viii
LIST OF ILLUSTRATIONS (Continued)
Figure
4.4.2
Page
(a) MM5 27 km Model Domain for Boston, MA and
(b) MM5 3 km and 1 km Domains for Boston, MA ..................................................77
4.4.3 MM5 Surface Layer Winds in the 1 km Manchester, NH
Domain at 6 pm EST on 11/11/97. ............................................................................78
4.4.4 Back Trajectories Calculated by RIP for the Period Ending
(a) August 28, 1998 at 9 am EDT and (b) August 30, 1998 at 9 am for
Manchester, NH and Boston, MA..............................................................................79
4.4.5 Comparison of MM5 and NCDC/FSL Radiosonde Archived
(a) Wind Direction and (b) Wind Speed at Chatham, MA on
May 25, 1999 at 7 pm EST. .......................................................................................80
4.5.1 Instantaneous Diffibusion Coefficients Calculated by the Centerline and
Moment Methods Versus Downwind Distance for Tracer Tests
Conducted in Manchester, NH, August 17, 28 and 30, 1998. ...................................81
4.5.2 (a) Logarithmic Wind Speed Profile Generated from MM5 Output, and
(b) Turbulent Kinetic Energy Estiamted from Stuff (1994), for
Manchester, NH 6:00 pm 11/11/97............................................................................83
4.5.3 TEMPEST Model Domain for a 2-D Idealized Urban Profile. .................................84
4.5.4 TEMPEST Solution for a 2-D Idealized Urban Profile. ............................................85
4.6.1 Nox (a) Point and (b) Area Emission Inventory Data for New England
Applied to May 25, 1999 at 12 pm EST. ...................................................................86
4.6.2 Upwind Source Area Influencing Boston, MA at 5 pm EST May 25, 1999. ............87
4.6.3 Source Contribution Calculation Results for Boston, MA at 5 pm EST
May 25, 1999.
................................................................................................88
4.6.4 Average Pollutant Source Travel Times in Relation to Boston, MA at
5 pm EST, May 25, 1999. ..........................................................................................89
4.6.5 Fractional Source Contributions of NOx on the Receptor Concentrations at
Boston, MA at 5 PM EST May 25, 1999. ..................................................................90
4.6.6 Fractional Source Contributions of NOx within Approximately
75 km of the Receptor, Boston, MA at 5 PM EST May 25, 1999. ............................91
ix
LIST OF ILLUSTRATIONS (Continued)
Figure
Page
4.7.1 Typical CIT Airshed Model Emissions Field ............................................................96
4.7.2 CIT Modeling Domain, UTM Coordinates ...............................................................99
4.7.3 Monitoring Stations Within SCAQS Region and CIT Modeling Domain ................100
4.7.4 ALKE and CO Emissions Eigenvalue Spectra for August 27-28, 1987....................101
4.7.5 Eigenvalue Representation Error for ALKE, CO and NO Emissions
(August 27-28, 1987). ................................................................................................102
4.7.6 Error Norm Reduction as a Function of Optimization Iteration ................................102
4.7.7 Flowsheet 1: Overall Optimization Strategy Flow Diagram ....................................103
4.7.8 Flowsheet 2: Search Algorithm and Associated Code ..............................................103
4.7.9 Flow Diagram for Objective Function Evaluator (Getfunc Shell Script) ..................104
4.7.10 Emissions Time Series for CO Emissions (August 27-28, 1987). .............................105
4.7.11 First Four ALKE Temporal Eigenfunctions ..............................................................106
4.7.12 Two-Dimensional Contour-Plots of First Five ALKE Eigenfucntions .....................107
4.7.13 Three-Dimensional Surface-Plots of First Four ALKE Eigenfunctions ....................108
4.7.14 Ozone Predictions for Base Case Simulation (August 28, 1987). .............................109
4.7.15 Optimized Result for August 28, 1987 (24 hour norm) .............................................110
4.7.16 Emissions Time Series for ALKE Emissions (August 27-28, 1987).........................110
4.7.17 Difference Between Optimized and Base Case ALKE Field (8 am) .........................111
4.7.18 Optimal Coefficients for Base Case, Optimized Field and 3 Error Runs
(ALKE Emissions) (August 27-28, 1987). ................................................................111
4.7.19 Optimized Result for August 28, 1987 (14.7.5 pm norm) .........................................112
4.7.20 Optimal Coefficients for Base Case and Optimized Fields for
Two Error Norms (August 27-28, 1987). ..................................................................113
x
LIST OF ILLUSTRATIONS (Continued)
Figure
Page
4.7.21 Emissions Time Series for ALKE Emissions (August 27-28, 1987).........................113
4.7.22 DEMM Flow Diagram [Wang, 1999]. .......................................................................117
4.7.23 Monitoring Stations within SCAQS Region and CIT Modeling Domain .................120
4.7.24 Uncertain Parameters With Increased Average Variance Contribution
Under at 1.5 x ROG Case ..........................................................................................123
4.7.25 ncertain Parameters with Decreased Average Variance Contribution
Under a 1.5 x ROG Case............................................................................................123
4.7.26 Morning Ozone Variance Contribution, CELA .........................................................126
4.7.27 Primary Variance Contributing Parameters at RIVR.................................................127
4.7.28 Ozone Time Series (1.5 x Base Case ROG) (8/27/87), Central Los Angeles ...........128
4.7.29 Ozone Variance Time Series (1.5 x Base case ROG) (8/27/87),
Central Los Angeles ................................................................................................128
4.7.30 Ozone Time Series (1.5 x Base Case ROG) (8/27/87), Claremont ...........................129
4.7.31 Ozone Variance Time Series (1.5 x Base Case ROG) (8/27/87), Claremont ............129
4.7.32 Ozone Time Series (1.5 x Base Case ROG) (8/27/87), Hawthorne...........................130
4.7.33 Ozone Variance Time Series (1.5 x Base Case ROG) (8/27/87), Hawthorne ...........130
4.7.34 Ozone Time Series (1.5 x Base Case ROG) (8/27/87), Pasadena .............................131
4.7.35 Ozone Variance Time Series (1.5 x Base Case ROG) (8/27/87), Pasadena ..............131
4.7.36 Ozone Time Series (1.5 x Base Case ROG) (8/27/87), Riverside .............................132
4.7.37 Ozone Variance Time Series (1.5 x Base Case ROG) (8/27/87), Riverside ..............132
4.8.1 Continuous Point Source............................................................................................135
4.8.2 Predicted v. Measured 10-Meter Wind Speed at Logan Airport on 5/25/99. ............136
4.8.3 Predicted v. Measured Wind Direction at Logan Airport for 5/25/99 .......................136
xi
LIST OF ILLUSTRATIONS (Continued)
Figure
Page
4.8.4 SF6 Release Site and Adjacent Four Grid Points for Which Wind Fields
Are Available.
................................................................................................138
4.8.5 Power-Law Fit for the Average Wind Speed at the SF6 Release Site for
16:00 GMT on 5/25/99. .............................................................................................139
4.8.6 Route for the Aerodyne (blue) and WSU Mobile Lab (yellow), and the
SF6 Release Site (Red) for 5/25/99. ...........................................................................140
4.8.7 Detail of the Route Followed by the Two Mobile Labs ............................................141
4.8.8 Shifted Coordinates Using the Wind Direction, where xo, yo are the
Coordinates of the SF6 Release Site...........................................................................142
4.8.9 Reduced Data Set for Aerodyne Data ........................................................................144
4.8.10 Rotated and Reduced Data Set from Aerodyne .........................................................144
4.8.11 Shifted WSU Data
................................................................................................145
4.8.12 Final WSU Rotated and Shifted Data Set. .................................................................146
4.8.13 Combined Rotated and Shifted Final WSU and Aerodyne Data ...............................147
4.8.14 Measured SF6 Concentrations Using the Data Set Combined3. ................................153
4.8.15 Measured SF6 Concentrations Using the Data Set ARI_SF6 Data. ...........................154
4.8.16 Measured vs. Predicted SF6 Concentration for the Combined3
Dataset Using the Best-Fit Parameters of Equation (4.8.32) .....................................156
4.8.17 Measured vs Predicted SF6 Concentration for the Combined3
Dataset Using the Best-Fit Parameters of Equation (4.8.33). ....................................156
4.8.18 Predicted Concentration of SF6 Assuming an Average Wind Speed
Of 7.2 m/s and p=0.077. ............................................................................................158
4.8.19 Individual Road Segments Considered from the Aerodyne Dataset .........................159
4.8.20 Detail of the Route for the Road Labeled Road_100 in Figure 4.8.20
(left) and the Observed and Predicted SF6 Concentrations (right). ...........................160
xii
LIST OF ILLUSTRATIONS (Continued)
Figure
Page
4.8.21 Road_250 in Figure 4.8.19.........................................................................................160
4.8.22 Road_600
................................................................................................161
4.8.23 Road_1000 in Figure 4.8.19.......................................................................................161
4.8.24 Road_1500 in Figure 4.8.19.......................................................................................162
4.8.25 Road_1500_WSU in Figure 4.8.19 ............................................................................162
4.8.26 Observed and Predicted SF6 Concentrations for Road_100 Using
Parameters Regressed Using Only Data from Road_100 ..........................................164
4.8.27 Observed and Predicted SF6 Concentrations for Road_250 ......................................164
4.8.28 Observed and Predicted SF6 Concentrations for Road_600 ......................................165
4.8.29 Observed and Predicted SF6 Concentrations for Road_1000 ....................................165
4.8.30 Observed and Predicted SF6 Concentrations for Road_1000 ....................................166
4.8.31 Observed and Predicted SF6 Concentrations for Road_1000 ....................................166
4.9.1 Relationship of NO2 to Total NOx.............................................................................171
4.9.2 Relationship of NO2 to total NOx ..............................................................................172
4.9.3 Relationship of NO2 to Total NOx.............................................................................174
4.10.1. Terrain Model of Manchester, NH, and ArcView Screen-Shot of Manchester.........177
4.10.2 MapCafe Screen-Shot Comparing CO2 Measurements at
Different Times of Day ..............................................................................................178
4.10.3 Conceptual Diagram of the Distributed GIS Architecture .........................................179
4.10.4 Detailed Diagram of Distributed GIS Architecture for Manchester Analyses ..........180
4.10.5 Query Box for Customized Querying and Mapping of Trace Gas Measurements ....183
4.10.6 Promixity-to-Road Model for Estimating the Spatial Distribution of
Vehicle Emissions
................................................................................................185
xiii
LIST OF ILLUSTRATIONS (Continued)
Figure
Page
4.10.7 Aerodyne Van Traversal Along I-93 and I-90 in Boston.. ........................................188
4.10.8 Eastern Mass Counties and Interstate Highways with a Boston Area Ortho .............189
4.10.9 Van Speed (Light Color is Slow) vs Concentration (Height) of NO (Left)
And NO2 (Right)
................................................................................................190
4.10.10 Urban Respiration Project – GIS Modeling ............................................................191
4.10.11 The Mechanism of Connecting GIS and RDBMS ..................................................193
4.10.12 Inverse Distance Weighted (IDW) Calculation of Trace Gas Concentrations .......194
4.10.13 Flow Chart of Method for Dispersion Model Calculations ....................................195
4.10.14 Estimated CO Concentration Distribution After the Stack Emission Model .........197
4.10.15 Estimated NO2 Concentration Distribution After the Stack Emission Model ........197
4.10.16 Find All the Road Intersections and Highway Exits ...............................................200
4.10.17 Estimated Trace Gas Concentration from Traffic Congestion Around Road
Intersections
................................................................................................201
4.11.1
Flow Chart of Emission Modeling Approach .........................................................203
4.11.2
Land Use in the Boston Study Area ........................................................................204
4.11.3
Population Density (total population per acre) for the Boston Study Area ............205
4.11.4
Average Observed Values of NO (ppb) During May 25, 1999 Mobile Van Runs .207
4.11.5
Typical Regression Results for May 25 NO Observations .....................................208
4.11.6
May 25 NO Residuals (as standard deviations) for the SW3 Model ......................209
4.12.1
Concentration of Fine Particles (particles cm-3) Along Boston Roads
and Highways on May 23, 1999. ............................................................................213
4.12.2
Selected Water-Soluble Composition of Urban Aerosols Sampled on the
MIT Campus During May 27-29, 1999. .................................................................214
xiv
LIST OF ILLUSTRATIONS (Continued)
Figure
Page
4.12.3
Time Series of CO2 and Fine Particles During Stationary Measurements
In Cambridge, MA ................................................................................................216
4.12.4
Summary Relationships of Total Fine Particles with NO and CO During
the Afternoon of May 28.........................................................................................216
4.12.5
Summary of Total Fine Particle and CO2 Relationships Observed During
the Mobile Laboratory Measurements in the Boston Area on Four different
Days in May 1999. ................................................................................................217
xv
LIST OF TABLES
Table
Page
Table 2.3.1 - Summary of WSU Van Instrumentation Operations .....................................20
Table 3.1 -
Mobile Campaign Data ..................................................................................28
Table 3.2.1 - Summary of Tracer Test Periods ...................................................................33
Table 4.5.1 - Instantaneous diffusion coefficients calculated by the
centerline method and moment method for tracer tests
conducted in Manchester, NH August 27, 28 and 30, 1998. .........................82
Table 4.6.1 - Percent Contribution of Emissions to the Boston Receptor at
Incremented Radial Distances........................................................................91
Table 4.7.1 - Properties of Well-Posed Inverse Problems ..................................................94
Table 4.7.2 - Percent of Variance Captured by First Five Eigenfunctions. ........................101
Table 4.7.3 - Lumped Organic Classes Used in the CIT Model. ........................................105
Table 4.7.4 - Typical Input Parameters used in Photochemical Models. ............................118
Table 4.7.5 - Uncertain mechanism parameters (based on Stockwell and Pun [1997])......119
Table 4.7.6 - Relative Error of Second Order Approximation for Ozone. ..........................119
Table 4.7.7 - Monitoring Site Description. .........................................................................120
Table 4.7.8 - Ozone Variance Contribution, CELA (percent)
(values > 5 are in bold, values < 0.1 not shown.............................................122
Table 4.7.9 - Ozone Variance Contribution, RIVR (percent)
(values > 5 are in bold, values < 0.1 not shown.............................................122
Table 4.7.10 - Relative Change in Ozone Percent Variance Contribution
1.5 ROG Case versus Base Case Analysis (values shown are
average factors from 12:00-16:00PST). .........................................................122
Table 4.7.11 - Effect of different scenarios on variance contribution (12 - 4 pm)
( A = Base Case; B = 1.5  ROG; C = 1.5  ROG and 1.5  NOx ). ............124
Table 4.7.12 - Temporal Differences in Ozone Concentration and Variance. ......................125
xvi
LIST OF TABLES
Table
Page
Table 4.7.13 - Ozone Variance Contribution for Selected Reactions at CELA
and RIVR (2pm). ...........................................................................................126
Table 4.7.14 - Primary Variance-Contributing Photolysis Reactions ...................................126
Table 4.7.15 - Dominant Variance-Contributing Parameters at RIVR. ................................127
Table 4.8.1 - Release Site Wind Field Parameters. .............................................................139
Table 4.8.2 - SF6 Release Site Coordinates. ........................................................................142
Table 4.8.3 - Summary of Parameters for Boston 05/25/99. ...............................................146
Table 4.8.4 – Summary of Utilized Data Sets......................................................................152
Table 4.8.5 - Best-fit parameters using Equation (4.8.32). .................................................154
Table 4.8.6 - Best-fit parameters using Equation (4.8.33). .................................................155
Table 4.8.7 - Best-Fit Parameters Using the Dataset Combined3 and
Allowing  to Vary Hourly. ..................................................................................................157
Table 4.8.8 - Theory v. Best-Fit Parameters. ......................................................................158
Table 4.8.9 - Individual Road Best-Fit Parameters. ............................................................163
Table 4.8.10 - Best-fit Parameters for Different Sized Data Sets. ........................................167
Table 4.8.11 - Best-fit Parameters, allowing for Offset in Wind Direction. .........................168
xvii
1.0 PROJECT OVERVIEW
1.1 Motivation
Human society has well defined metabolic processes that can be characterized and
quantified in the same way that an ecosystem’s metabolism can be defined and understood
[Fischer-Kowalski, 1998.] The study of “industrial metabolism” is now a well-established topic,
forming a key component of the emerging field of industrial ecology [Ayres and Simmonis,
1994; Fischer-Kowalski and Hüttler, 1998]. The fact that the metabolism of cities can be
analyzed in a manner similar to that used for ecosystems or industries has long been recognized
[Wolman, 1965.] However, the increasingly rapid pace of urbanization and the emergence of
megacities, particularly in the developing world, lends increased urgency to the study of “urban
metabolism.” A recent review by Decker et al. [2000] surveys energy and materials flow though
the world’s twenty-five largest metropolitan areas. In 1995 these cities had populations
estimated to range between 6.6 and 26.8 million people; all are expected to exceed 10 million
by 2010.
Urban metabolism, driven by the consumption of energy and materials, cannot take place
without respiration. Both combustion based energy sources and the human and animal
populations of cities consume atmospheric oxygen and expire carbon dioxide as well as a range
of other trace gases and small particles. While the detail content of these urban emissions are
generally not well known, there is no doubt that they are large and varied [Decker et al., 2000.]
There is growing recognition that airborne emissions from major urban and industrial areas
influence both air quality and climate change on scales ranging from regional up to continental
and global. Urban/industrial emissions from the developed world, and increasingly from the
megacities of the developing world change the chemical content of the downwind troposphere in
a number of fundamental ways. Emissions of nitrogen oxides (NOx), CO and volatile organic
compounds (VOCs) drive the formation of photochemical smog and its associated oxidants,
degrading air quality and threatening both human and ecosystem health. On a larger scale, these
same emissions drive the production of ozone (a powerful greenhouse gas) in the free
troposphere, contributing significantly to global warming. Urban and industrial areas are also
large sources of the major directly forcing greenhouse gases, including CO2, CH4, N2O and
halocarbons. Nitrogen oxide and sulfur oxide emissions are also processed to strong acids by
atmospheric photochemistry on regional to continental scales, driving acid deposition to sensitive
ecosystems. Direct urban/industrial emission of carbonaceous aerosols is compounded by the
emission of copious amounts of secondary aerosol precursors, including: NOx, VOCs, SO2, and
NH3. The resulting mix of primary (directly emitted) and secondary aerosols is now recognized
to play an important role in the climate of the Northern Hemisphere.
What is less widely recognized is the poor state of our knowledge of the magnitudes, and
spatial and temporal distributions, of gaseous and aerosol pollutants from urban/industrial areas.
While most cities in the developed world do have a few continuous fixed site monitoring stations
measuring point concentrations of regulated air pollutants; these measurements very poorly
constrain the patterns of pollutant measurements from the urban area as a whole. Most cities in
the developing world lack even these relatively sparse routine measurements. Air quality
agencies in the developed world have assembled urban/industrial emissions inventories for some
1
key pollutants, most notably NOx, CO, some VOCs, SO2, and some primary aerosols such as
soot and particulate lead. However, far too often these emission inventories are based on
engineering estimates rather than measured emissions. In addition, they often miss or poorly
quantify smaller fixed sources, mobile sources (motor vehicles, trains, boats, aircraft) and area
sources like landfills. Emissions inventories in developing countries, where they exist, are often
based on dubious extrapolations of those used for cities in the developed world.
This sad state of affairs is a serious problem. First, it is difficult to predict the impact of
poorly defined emissions and pollutant distributions on urban air quality and its impact on
citizen’s health and local ecosystem viability. Second, since the atmospheric chemistry which
drives processes like ozone or secondary aerosol production is highly nonlinear, the impact of
urban/industrial emissions on larger scales cannot be predicted without a relatively accurate and
detailed knowledge of the temporal and spatial distributions of their precursors. Since “business
as usual” is doing a poor job of specifying the real distributions of urban/ industrial atmospheric
pollutants, new tools and techniques need to be developed to more easily and accurately quantify
these emissions and allow accurate prediction of their subsequent chemical transformations and
transport to larger scales.
1.2 Project Team Description
Our NASA Earth Science Enterprise funded Urban Metabolism and Trace Gas
Respiration Project is an effort to better understand the distribution and emission patterns of
pollutants in urban areas. The project took place between February, 1997 and October, 2001 as
an Interdisciplinary Science (IDS) investigation associated with the Earth Observing System
(EOS) project. It involved a highly interdisciplinary collaboration between five research teams
from the Center for Atmospheric and Environmental Chemistry at Aerodyne Research, Inc.
(ARI), the Departments of Chemical Engineering and Urban Studies and Planning at the
Massachusetts Institute of Technology (MIT), the Institute of Earth, Oceans, and Space at the
University of New Hampshire (UNH), and the Laboratory for Atmospheric Research at
Washington State University (WSU).
The team included physicists, physical chemists, and environmental engineers expert in
atmospheric measurement techniques, chemical and environmental engineers skilled in
developing and utilizing models of atmospheric chemistry and dynamics, and urban planners
with a research focus on the development of geographical information systems (GIS) and their
innovative use in mapping and intercomparing urban characteristics, including pollutant
distributions. Graduate students from MIT, UNH, and WSU were involved in both the
measurement and modeling/analyses portions of the project.
1.3 Experimental Goals
Airborne platforms featuring fast response sensors have previously been deployed, with
dramatic effect, to measure stratospheric and free tropospheric processes (e.g. Anderson et al.,
1989) and even to follow urban emission plumes to quantify downwind pollution evolution
[Trainor et al., 1995; Nunnermacker et al., 1998]. Components of our team have also used
ground vehicles equipped with fast response trace gas sensors to quantify methane emissions
2
from urban (and rural) components of natural and town gas systems, urban landfills, and sections
of towns and cities [Lamb et al., 1995; Mosher et al, 1999; Shorter et al., 1996; 1997].
However, mobile fast response sensors had not been used previously to characterize
multi-pollutant distributions and source emissions within urban areas. For this project we
proposed to develop, deploy and demonstrate better urban atmospheric measurement techniques
based on sensitive, accurate, real-time trace gas and particulate sensors onboard a ground mobile
platform (a mobile laboratory.) We anticipated that the deployment of real-time (~1s response)
sensitive and specific trace pollutant instruments in a mobile laboratory would generate a wealth
of data on the distribution of both urban ambient pollutant levels and the distribution and nature
of both mobile and stationary (including point and area) emission sources.
As proposed, we first tested our instrumented mobile laboratory in two field missions in
Manchester, NH a compact urban area with a population of ~100,000 well isolated from other
urban centers. We then deployed our mobile laboratory in a intensive campaign in Boston, MA
at the center of a metropolitan area with ~3 million people. These field programs allowed us to
learn how to effectively deploy real-time mobile instruments in a major urban area and gain
valuable data on pollutant distributions and emission sources.
Our field measurement tools and strategies are presented in Section 2 of this report and
an overview of the urban field measurement data we obtained is presented in Section 3.
1.4 Analysis Goals
Since our real-time mobile measurements would generate copious amounts of data, a key
programmatic goal was to develop the data reduction and analysis methods that would allow us
to learn the most about pollutant distributions and emission sources. Further, since we proposed
to develop novel methods of investigating urban gaseous polluant and fine particle emissions and
distributions, we planned that analyses and evaluations of our initial field measurements would
be used to design better measurement strategies to collect and analyze trace gas and fine particle
concentration and flux data.
In order to analyze experimental strategies and field measurement data the MIT and
WSU groups have used state-of-the-art air quality models and developed new model analysis
techniques. The WSU team developed a two component approach to model the turbulent
atmospheric dynamics over urban landscapes. First, they used the Environmental Protection
Agency’s (EPA’s) state-of-the-art MM5 model to provide a mesoscale model of the regional
wind field and then applied TEMPEST, a 3-d turbulence model developed at the Pacific
Northwest National Laboratory (PNNL) that simulates the actual urban landscape. WSU also
developed a capability for predicting the downwind urban pollution footprint by combining
MM5 computed windfields, MCIP, the meteorological processor from EPA’s Models-3/CMAQ
model to invert the windfields, and the CALPUFF plume dispersion model. MIT used the MM5
windfields generated by WSU to test urban scale diffusion models by analyzing SF6 tracer
release experiments performed as part of our Boston field campaign. In addition, the MM5
output was used to input the California Institute of Technology (CIT) air quality model to assist
in analyses of the ozone and NOx trace gas distributions measured in Boston. Finally, MIT
3
investigated the use of air quality model inversion techniques to determine how well spatial
emissions distributions can be deduced from measured urban pollutant distributions.
The project also involved the novel use of geographic information systems (GIS) and
urban databases to correlate observed trace gas emission fluxes (urban respiration) with urban
and industrial activity and consumption factors (urban metabolism). Finally, correlations
between measured trace gas emissions and urban/industrial activity/ consumption factors are
used to identify parameters accessible to air- and satellite-borne remote sensing systems in order
to enable automated estimates of urban and industrial trace gas emissions relevant to global
change and regional pollution issues.
Data analysis strategies and modeling results are presented in report Section 4.
1.5 Project Accomplishments and Impacts
We believe that the research presented in this report and in the papers and symposia
presentations this project has stimulated or will produce demonstrate that our team has broken
new ground in the measurement and analyses of urban pollution distributions and emission
source characterization. The deployment of multiple fast response pollutant measurement
instruments on board a mobile ground vehicle was successfully accomplished. This
accomplishment has allowed the development of novel urban pollution measurement strategies
that have expanded our capability to more fully characterize the air quality and pollutant
emission sources in urban settings. The novel data sets obtained during our field measurements
have stimulated the development of innovative modeling approaches for urban atmospheric
processes. The interaction between atmospheric models /data analysts and urban planners using
GIS techniques to map and visualize urban processes has been fruitful, leading to interesting new
ways to present and interpret urban air quality and emissions data.
A summary of project related symposia presentations, symposia proceedings papers,
archival journal articles that have been presented or published to date is presented in Section 5 of
the report. This section also lists in preparation or planned journal articles and web sites where
data archives and modeling/analysis results can be accessed.
The project presented in this report is only a start. The measurement and
modeling/analysis methods developed during our project will only make an impact if they can be
deployed in a wide range of major cites worldwide. On that score, we can report good news.
The ARI mobile laboratory, supplemented by the addition of a novel, fast response, aerosol mass
spectrometer (AMS) developed at ARI [Jayne et al., 2000], has been incorporated into the
PMTACS-NY project, an EPA airborne particulate supersite program focused on New York City
and led by the Atmospheric Sciences Research Center of SUNY, Albany. The ARI mobile lab
was used to map pollutant distributions and measure mobile source emissions in two New York
City field campaigns conducted in October, 2000 and July/August, 2001. Preliminary data from
the first New York City measurement campaign are summarized in Shorter et al. (2001). The
ARI mobile laboratory and components of our WSU and MIT collaborations are also playing a
major role in the MIT led “Integrated Program on Urban, Regional and Global Air Pollution:
Mexico City Case Study” funded by the Comisión Ambiental Metropolitana (CAM), the
4
Mexican Agency in charge of improving air quality in the Mexico City metropolitan area, and by
MIT. With the help of Mexican research groups, ARI, MIT and WSU personnel deployed the
ARI mobile laboratory, supplemented with a proton transfer reaction mass spectrometer (PTRMS) for rapid response measurements of selected aromatic and oxygenated VOCs, and
additional pollutant measurement equipment supplied by WSU and MIT in exploratory Mexico
City field measurements in February, 2002. The PTR-MS was supplied and manned by Montana
State University (MSU). An second, more extensive field campaign by ARI, MIT, WSU, MSU
and numerous Mexican investigators is planned for the spring of 2003.
Since the NASA project developed methods presented in this report have already been
expanded and deployed in New York City and Mexico City, key examples of developed and
developing megacities, we are confident that they will be extensively utilized in the future to
better characterize urban respiration and determine its health, ecological and climate impacts on
all scales, from local to global.
5
2.0 FIELD MEASUREMENT STRATEGIES
Our field measurement approach is to combine real-time mobile measurements of multiple
trace gases and particulates with meteorological data collection. Our instrumented mobile
laboratory easily performs real-time, fast response, simultaneous measurement of multiple trace
gases under normal driving condition. Mobile measurement can identify the distribution of local
sources in an urban area and thus better correlate urban activity with emissions. Intensive
stationary data collection with our instrument suite can simulate a fast response monitoring site
for comparison to time averaged results from traditional air quality monitoring sites.
2.1 Instrument Overview
The mobile laboratory was deployed with a series of sensitive, specific, real-time (~1
second response) sensors for trace gases and fine particulates; a global positioning system
(GPS); and a central data logging computer. Specifically, the sensors include an ARI two-laser
tunable infrared laser differential absorption spectrometer system (TILDAS), capable of
measuring between 2 and 4 trace gases simultaneously; a Licor NDIR carbon dioxide (CO2)
instrument; A TSI condensation nuclei instrument for fine particulates detection; a uv
absorption ozone instrument; and an Eppley total ultraviolet radiometer. The real-time
instruments have been described in detail previously [Shorter, et al., 1998; 2000; Lamb, et al.,
1995; Zahniser, et al., 1995]. We summarize the instruments in the following
Sections (2.1.1 – 2.1.6)
2.1.1 Tunable Infrared Laser Sensors
2.1.1.1 Background
The mobile sampling work employed a dual tunable infrared laser differential absorption
spectrometer (TILDAS) for detecting gas phase urban pollutants and greenhouse gas emissions.
Essentially all gaseous combustion exhaust pollutants of interest have strong fundamental
vibrational/rotational transitions in the mid-infrared (mid-IR) spectral region between ~3 and
20 m. High resolution tunable mid-IR lasers can interrogate spectral micro-windows, where
trace pollutant absorption features can be detected and integrated between water and or carbon
dioxide lines. TILDAS methods for trace gas analysis generally operate in the linear absorption
regime of Beer’s law, where the fractional absorption between a molecular spectral feature and
the background baseline is proportional to the feature’s absorption cross section (), the
absorption pass length (L) and the species concentration (n):
I/I = nL
(2.1.1)

I/I values down to10-5 are measurable in a few seconds with stable laboratory instruments,
although field conditions often restrict rapid measurements to minimum differential absorptions
of order 10-4 or larger.
The spectral specificity of TILDAS techniques make them particularly well suited to
detect small (2 to ~8) atom molecules which typically have resolvable vibrational/rotational lines
6
in the mid-infrared, or at least sharp absorption features such as highly structured Q-branches.
A high level of molecular symmetry, leading to simpler and intensity enhanced mid-IR
absorption features, also allows the effective measurement of some larger molecules like
benzene (C6H6). We have used TILDAS to detect CH4, N2O, C2H6, NO, NO2, SO2 and H2CO
emissions in our urban studies programs.
2.1.1.2 Laser sources
To date, most mid-infrared tunable laser trace gas instruments have employed lead salt
diode lasers, which have been commercially available for over twenty-five years. These tunable
diode lasers (TDLs) generally require cryogenic cooling, are relatively weak (typically, 0.1 mw
or less in single mode operation), and subject to multimode operation, usually requiring a
monochomator for mode selection. On the plus side, variations in composition have allowed the
production of lasers operating between ~2.5 and 25 m, although lasers operating between
3.5 and 15 m are more commonly available. Individual laser modes are typically tunable over
~2 cm-1 and via temperature selection of sequential modes each laser is typically piecewise
tunable over ~200 cm-1. Most of the trace gas pollutant measurements discussed in this report
were made with lead salt TDL systems.
With the advent of fiber optics telecommunications, near infrared (~0.8 to 2.5 m)
tunable diode lasers with III-V composition have also become widely available. These lasers
have the advantage of higher power and normally do not require cryogenic cooling. However,
near IR TDLs can access fundamental infrared transitions for very few molecules and thus must
exploit combination and overtone bands which are typically factors of 20 to 1000 less intense
than fundamental transitions, significantly compromising measurement sensitivities for trace
species.
More recently, other tunable mid-IR laser sources have become available. In our
laboratory, instruments to measure selected trace gases including CH4, CO, and N2O have been
based on Zeeman tuned rare gas discharge lasers [McManus et al., 1989; Kebabian and Kolb,
1993]. These lasers are limited in power and spectra coverage, but do offer an efficient, noncryogenic source when a rare gas plasma emission line is nearly coincident with an absorption
line of a relevant pollutant species.
Several laboratories have recently had significant success using difference frequency
generation (DFG) in nonlinear crystals driven by a near IR diode/NdYAG laser combinations or,
more recently, two near IR diode lasers. The most practical advanced systems developed to date
have exploited periodically poled lithium niobate (PPLN) driven by a two near IR diode lasers to
achieve usable intensities in the 3.3 to 4.3 m spectral range. Optical fiber amplifiers are used to
enhance the output of one or more pump diodes since output power scales as the product of the
input powers. While more complicated than single laser sources, DFG systems offer both noncryogenic operation and the promise of very wide spectral coverages. DFG systems based on
materials like phased matched GaAs may extend these systems much further into the infrared
than the 4.5-5.0 m opacity cutoff exhibited by PPLN.
7
Finally the advent of quantum cascade lasers has opened up a new source of
commercially available tunable mid-IR sources. Commercial versions of these lasers currently
require cryogenic cooling for continuous wave (cw) operation but can be used in pulsed
operation with thermoelectric cooling. The first functional instruments using this source
technology for non-cryogenic sub-part-per-billion detection of atmospheric trace gases have
recently been developed at Aerodyne for NO and NH3 [Nelson et al., 2002] and appear very
promising for mobile operation in future deployments.
2.1.1.3 Instrument Description
The system used in this study is an extended version of an instrument we first developed
in 1993 for measurements of methane and nitrous oxide source fluxes using the eddy correlation
method [Zahniser et al., 1995]. The instrument used in the mobile measurements has two
important modifications to the standard instrument: 1) The absorption path length is extended by
a factor of 4 (from 36 m to 150 m) to obtain the higher sensitivity, and 2) the system operates
with two lasers simultaneously sharing the same multiple pass cell. The dual system allows two
gases to be detected simultaneously without compromising the detection limit for either. The
dual system also provides a comfortable redundancy for field work in remote areas of the world
The dual-TILDAS system consists of two main modules: the optical bench apparatus,
including the diode lasers, optics, detectors, and reduced pressure multi-pass cell; and the
electronic module, containing two Pentium computers running control and data acquisition
software, a two-channel Laser Photonics laser control unit, and various related interface and
measurement electronics. The optical apparatus is constructed on a two by four foot aluminum
honeycomb table, surrounded by an aluminum cover. The aluminum cover and optical table
form a conductive enclosure to which thin film heaters are attached, allowing the temperature
inside to be closely controlled at 30o C. This design minimizes thermal gradients in the
instrument, which cause optical fringes to drift and add noise to the measurements. The optical
table and cover are contained within a roto-molded polyethylene shell with 6 cm of closed cell
foam insulation on all sides to assure temperature uniformity. The optical table is mounted
within an outer case using a coil spring suspension system to avoid vibration and shock during
shipping and transport in the back of the mobile van.
The dual TILDAS system employs separate diode lasers to produce distinct beams of two
different frequencies. The diodes are housed in one liquid nitrogen dewar, along with the
detectors for both the multi-pass cell sample beams and the reference cell beams. Figure 2.1.1
shows the schematic layout of the instrument.
The multi-pass absorption cell is an astigmatic Herriott type developed at ARI which
maximizes path length while keeping total volume small by effectively filling the volume
between the mirrors with the beams [McManus et al., 1995]. The small volume of the cell
insures a fast time response in the absence of wall effects. For a particular set of multi-pass cell
dimensions, with fixed mirror radii of curvature and base length, there are distinct configurations
of distance between and rotation of the mirrors for which the beam path exactly closes on itself
and exits the cell through the coupling hole by which it entered. The cell can easily be adjusted
8
PUMP
Path Length = 153.5 m
174 passes through the multi-pass cell
DEWAR
CH4 BEAM
2989
cm-1
Diode
#1
Diode
#2
Detectors
C2H6 BEAM
2990 cm-1
INLET
A-D
D-A
Converters
Display
Data
Analysis
Acquisition Line Fitting
Software
Software
Laser
Control
Module
REFERENCE CELLS
Figure 2.1.1. Schematic for a two-laser TDL instrument with 153 meter multiple pass
astigmatic absorption cell. Both lasers and infrared detectors are contained in the
same liquid nitrogen-cooled dewar.
to change path lengths by altering the mirror spacing and rotating the mirror axes to select for
these re-entrant patterns. Broad-band enhanced-silver mirror coatings provide a reflectivity of
99.2% in the infrared, so that with a base length of 88 cm between the mirrors the cell is capable
of supporting 334 passes for a total path length of 295 m in a volume of 5 L.
For the mobile measurements program, a 174 pass pattern with an overall path length of
153.5 m was pre-selected as a precaution against possible degradation in mirror reflectivity due
to the extremely dusty environments. Also, the lower pass pattern is less susceptible to
vibrational misalignment of the external optical path. This sacrifice of a factor of 2 in overall
sensitivity using a shorter path length seemed a small price to pay for increased confidence in the
system reliability under rigorous vibration of the vehicle since it would have been difficult or
impossible to readjust the mirrors once the field program was underway.
The two beam paths are multiplexed into the same cell at right angles, so that two
separate patterns propagate through the cell without interference. Two individual output beams
emerge and are subsequently transmitted to the detectors. The input beam defines the corner of a
rectangle, with the beam exiting the cell from the opposite corner; the input-output beam
directions define the coupling plane. This design orients the coupling planes of the two beams
9
orthogonally, so that one is horizontal and the other vertical. Simulations of the beam
reflections on the cell mirrors have been carried out so that the distinctive patterns formed by coaligned HeNe laser beams can be used to recognize proper cell alignment (Figure 2.1.2).
The pumping system consisted of rotary vane vacuum pump which provided a flow rate
of 300 liter/minute operating at a cell pressure of 30 Torr and giving a system response time of
about 1 second. The inlet to the mutipass sampling cell consisted of 5 meters of 4 mm i.d.
polyethylene and teflon tubing to go from the instrument to the front of the vehicle. The flow
regulating valve where the pressure dropped from 760 Torr to 30 Torr was located 2 meters
upstream of the cell. The time delay for the sample to transit this length of tubing was on the
order of .5 s.
2.1.1.4 Signal Processing
The TILDAS spectrometer measures absorption spectra directly using a rapid scan sweep
integration. This approach produces absorption spectra which are analyzed in terms of known
line strengths and positions to yield the absolute concentration of the trace gases. The control
module sweeps the current applied to the diodes in order to vary the wavelength over a number
of absorption lines and acquire a distinctive “fingerprint” for the trace gas. Compared to
monitoring at a single absorbing wavelength, this approach makes the retrieved concentrations
much less sensitive to potential interferences from other species absorbing in the same spectral
region, as well as those of weak etalon fringes inherent to the optical system when considering
fractional absorptions on the order of 10-5. The fast sweep integration also eliminates the need
for second harmonic detection while retaining information on the unabsorbed laser power so that
the technique remains an absolute measurement and does not need to be based on calibrations
from standard samples. Fingerprint fits are performed with an iterative nonlinear least squares
minimization routine which computes the Voigt profile for each line in the spectrum using the
HITRAN spectral database line parameters [Rothman et al., 1998], temperature, and pressure in
the cell. The TDL software developed by ARI allows up to 45 individual lines to be used in the
fingerprint fit for each species and can fit up to four species in each spectrum. The analysis of
spectra is done in real time and resulting concentrations are saved to disk, with the option of
archiving some or all background subtracted spectra for later review or analysis. A typical
spectrum for is shown in Figure 2.1.3.
A fraction of the beam from each laser is directed on a separate path through a reference
cell and onto a second detector. The reference cells have a 5 cm path length and are used to lock
the laser frequency to the proper frequency. The absorption in the reference leg can also be used
to confirm the mode purity of the laser and to allow corrections to be made to the ambient
measurements for variations in laser mode purity during the course of the field trials. The
reference optical path could also be arranged to pass through a monochromator when the
appropriate kinematically mounted mirror is inserted. This capability is used when
characterizing a new diode, both to determine its wavelength and its mode purity.
10
Figure 2.1.2. Calculated mirror beam spots for two patterns, each with 182 passes propagating
in an astigmatic Herriott cell. The two different wavelengths are shown as
different shades, and spot diameters are largest for the earliest reflections.
TRANSMISSION
1.0004
methane
.
1.0002
ethane 1 ppb
30 Torr, 300 K, 153.5 m
1.0000
0.9998
data
ethane
methane
fit
0.9996
2989.8
2989.9
2990.0
2990.1
WAVENUMBER (cm
2990.2
2990.3
-1
)
Figure 2.1.3. One-second TILDAS spectra and calculated non-linear least squares fit for 1 ppb
ambient ethane and 1.7 ppm methane. The top traces and the combined fit to the
data are calculated from tabulated molecular properties.
2.1.1.5 Instrument operation
Two well-known problems in tunable diode laser infrared spectroscopy set the
requirements for instrument operation and the degree of operator intervention required to achieve
good measurements. The first is the existence of interference fringes in the spectrum, which to a
greater or lesser extent will not be distinguishable from modulations of the laser intensity due to
molecular absorption lines. This can lead to uncertainty of a trace gas concentration depending
on whether a peak or a valley of the sinusoidal fringes is aligned with the peak of the absorption
line. If temperature changes in the apparatus lead to changes in the period of the fringes, this
leads to a systematic “baseline drift” over time, and if the temperature change continues long
enough in the same direction, the change in “baseline correction” can be approximately
sinusoidal as well. The second is the possibility that the laser is operating multimode rather than
single mode, so that only a fraction of the total detected light will be absorbed by a given
11
absorption line, no matter how strong. If this problem is not detected, it has the result that
concentrations derived from multimode spectra will be smaller than the true values.
To compensate for the first problem, baseline calibration procedures were carried out at
frequent intervals. A flow of dry nitrogen replaced the ambient air being pumped into the
multipass cell. This was done by opening a shutoff valve to a line to the gas vent of the liquid
nitrogen tank. The nitrogen was added through a “T” fitting located 10 cm from the end of the
sampling tube with a slight excess flow which vented through the sampling inlet without
changing the inlet pressure. The trace gases are completely flushed from the multipass cell in
about 5 seconds. The spectrum of the nitrogen-filled multipass cell was recorded. This
“background spectrum” contains all the information about the variation of diode intensity over
the scan, including any interference fringes. When the background spectrum is subtracted from
the ambient sample spectrum the resulting difference spectrum will contain only absorptions due
to the trace gases. The background spectrum also provides the absolute intensity of the laser
needed to apply Beer’s Law to calculate absorbance and molecular concentrations.
The second potential problem of mode purity was addressed in two ways: 1) observing
the depth of absorption in the reference cell and comparing it against the expected value in the
laboratory before deployment; and 2) periodic injection of high concentration of the trace gas at
the cell inlet to observe that the lines are actually fully adsorbed or “black”. The frequency of
this check is highly dependent upon the individual laser and its past history of mode drift and can
requires frequent operator attention.
In general, the current diode laser TILDAS system used in our urban respiration
measurements are able to quantify trace gaseous pollutants, including CH4, CH2O, CO, NO,
NO2, N2O, and O3 at the 0.5 to 0.5 ppbv level for one second measurement times.
2.1.2 Commercial Licor CO2/UV Ozone/Eppley UV sensors
Carbon dioxide mixing ratios were measured by sub-sampling ambient air from a
common sampling manifold connected to the shielded, forward-facing inlet installed in the ARI
mobile lab. This sub-sample was directed into a Licor model 6262 infrared analyzer calibrated
periodically with NOAA-CMDL certified CO2 standards. The CO2 sensor has a 1 sec response
time and a 1 ppm sensitivity.
Urban ozone mixing ratios were measured in the last campaign in Manchester (8/98) and
in the Boston campaign (5/99) with a UV Photometric ambient O3 analyzer/Calibrator (Model
49/49PS). The instrument measures ozone with a time lag of 10 seconds and response time of
20 seconds. Its minimum detectable limit is 2 ppb, with a precision of 2 ppb, and noise equal to
1 ppb.
The Eppley total ultraviolet radiometer was mounted on the roof of the ARI mobile van
during the field campaigns in 8/98 and 5/99. This radiometer has response between
290 – 385 nm; i.e., adhering closely to the generally accepted limits for solar ultraviolet
radiation reaching the earth’s surface. The sensor reports total uv as voltage, and has a response
of 2.01 mV/mW cm-2 (adjusted to ambient temperature of 25 °C).
12
2.1.3 Fine Aerosol Measurements (Condensation Particle Counter)
The number density of fine aerosols was measured with a TSI model 3022A particle
counter sensitive to fine aerosols in the 10 - 3000 nm aerodynamic diameter range. This size
range essentially corresponds to the PM2.5 designation by the U.S. EPA These are primarily
secondary aerosols composed of sulfates, nitrates, and organic material. Since these aerosols are
generally in the accumulation mode, they are not removed effectively from the atmosphere by
wet and dry deposition processes. These aerosols can be transported over long distances in the
troposphere and are usually hygroscopic making them effective cloud condensation nuclei
(CCN). These ambient aerosols also have significant light scattering capability (i.e., their size is
comparable to that of visible light; 520 nm).
The sampling inlet for fine aerosols was a tandem arrangement of rear-facing inlets
positioned about 1-foot above 8-inches horizontally offset from the ARI mobile lab. One flow
stream was directed directly into the instrument and the other incorporated a 3-foot heated length
that was maintained at 300 C. Either flow stream can be directed into the particle counter
providing information on the total/volatile/non-volatile fine aerosol fractions. At 300 C the
sulfates, nitrates, and a portion of the organic material are volatilized [Clarke, 1991]. The nonvolatile fraction (that remaining at 300 C) in an urban setting should be mostly representative
the black carbon component. The total fine particle number density is comprised of sulfates
(including H2SO4), nitrates, organic material and very fine crustal dust. Volatilization and
aerosol passing efficiency tests were conducted at the University of New Hampshire using
ambient aerosols and a TSI aerosol generator to verify quantitative volatilization of the nitrate
and sulfate components and passing of both fractions in our sampling inlet.
The data were collected using National Instruments hardware and LabView software on a
standard laptop PC. Temporal adjustment was accounted for an inlet residence time of
approximately 2.5 seconds. Bad data, generally caused by equilibration effects after switching
from one inlet to the other, were filtered out during post processing. This filtering resulted in a
loss of about 10 seconds of data every 2 minutes. Aerosol number density data was otherwise
reported at 1 hertz.
During the May 27 to May 29 1999 stationary intensive sampling campaign, bulk aerosol
composition measurements were made in addition to particle number density and CO2. These
measurements were accomplished using a standard Teflon filter exposure technique (see e.g.,
Talbot et al., 1992). Subsequent methanol/deionized water extraction and analysis by ion
chromatography was performed at the University of New Hampshire. Mixing ratios of aerosol
Na+, Cl-, Mg2+, NO3-, SO4=, NH4+, Ca2+, K+, PO43-, CH3COO-, HCOO- were reported for the hour
long (and half hour long during periods of high automobile traffic) sample integration period.
2.1.4 Tracer Release and Detection Instrumentation
During campaigns in Manchester, NH and Boston, MA, sulfur hexafluoride tracer gas
(SF6) was released to tag specific urban source areas and to obtain urban dispersion data.
Typically, tracer releases were conducted for several hours from a ground level location and real
time continuous analyzers were used in either the ARI Mobile Laboratory or in a WSU mobile
13
van to track the downwind movement of the tracer. The mobile operations included repeated
crosswind traverses along available roads at different downwind distances to obtain horizontal
tracer concentration profiles.
During each tracer release, SF6 was released at a steady rate from gas cylinders through a
stainless steel capillary flow restrictor and a dry gas meter. The gas cylinders and flowmeters
were housed in a van and tracer was released through tubing secured to the roof of the van. Dry
gas meter readings were manually recorded periodically throughout each release period, and the
tracer release rate was determined from sequential dry gas meter readings. Typical release rates
were approximately 1 g/s. The estimated variability and accuracy in the release rates are less
than 5%.
For all of the campaigns, the ARI van was fitted with a continuous SF6 analyzer and data
system. In Boston, a second analyzer was installed in a van operated by WSU. The SF6
analyzer, as developed by Benner and Lamb [1985], provides continuous detection of SF6 with a
detection limit less than 50 parts per trillion (ppt) and an instrument response time less than 1 s.
The WSU instruments were fitted with commercial 63Ni electron capture detectors (Valco, Inc.).
The instrument was connected to the van inlet manifold for sampling via a small external pump.
The SF6 signal was recorded on either the ARI data system or the WSU van data system at 1 Hz.
During tracer tests, the instrument was calibrated periodically using a series of commercial,
certified SF6/air standards (Scott-Marrin, Inc, 5% accuracy) over the range 300 ppt to 10 ppb.
Overall uncertainty in the measured SF6 concentrations is approximately 15%. Further
descriptions of the tracer release and analytical instrumentation are available in Lamb et al.
[1995].
Tracer tests were conducted during three of the four field campaigns. A summary of the
tracer test periods is given in Table 3.2.1. Figure 2.1.4 shows an example of the instantaneous
SF6 concentrations measured during one sample period from the Boston, MA May 25, 1998
tracer test period. Section 5.3 gives data file location information.
2.1.5 VOC Sampling and Analysis Instrumentation
During each field campaign, VOC whole air samples were collected at selected sampling
points within the urban area using portable canister samplers. Each sampler consisted of a
stainless steel inlet, aluminum-Teflon 12 DC sampling pump, flow restrictor, solenoid valve, six
liter electropolished canister and timer. Each unit was housed in a plastic cooler. The sample
inlet was at approximately 1 m above the surface. Approximately ten samplers were used in
each campaign. Samplers were deployed immediately prior to the sample period and collected
immediately following the sample period. Typically, 3 hr averaged samples were collected with
a pre-selected start time. Canisters were cleaned by heating overnight at 50 oC while flushing
with zero, humidified air. The canisters were pressurized with zero humid air for transport to the
field location. Prior to sampling, each canister was evacuated to less than 0.03 psia. During
sampling, the pump and solenoid valve were activated by the timer and air was pumped into the
14
Figure 2.1.4. SF6 tracer data obtained May 25, 1998 in the South Boston Area.
canister at a controlled rate. Typically, canisters were pressurized to approximately 18 psia
during the specified sample period. Canisters were returned to WSU for analysis after collection.
Additional details are given in Lamb et al. [1995].
The whole air samples were analyzed for individual VOC using cryogenic
preconcentration with capillary column gas chromatography with flame ionization detection
(GC-FID, HP5890) [Zimmerman and Westberg, 1992]. A specified volume of sample was
transferred from the canister to a cryogenic freeze-out loop filled with glass beads. The sample
loop was then heated quickly and the contents transferred to the head of the column. The column
was temperature programmed from –50 oC to 150 oC at 4 oC/min. The system was calibrated
periodically using a certified standard of neo-hexane (Scott Specialty Gases, Inc. 2% accuracy).
Individual VOC compounds were identified on the basis of peak retention time in comparison to
known mixtures of VOC and in comparison to results from GC-MS analyses using the same
chromatography parameters.
Processing has been completed for the following compounds for the August 1998 field
campaign: iso-pentane, n-pentane, isoprene, benzene, and toluene. Figure 2.1.5 shows
(a) isoprene and (b) toluene concentrations for August 25, 1998 at various locations throughout
Manchester, NH.
15
A u g u s t 2 5 , 1 9 9 8 : Is o p r e n e
1 4 .0 0
1 2 .0 0
H our
ppbC
1 0 .0 0
8 .0 0
12
14
16
6 .0 0
4 .0 0
2 .0 0
4
n
o
ti
d
ta
o
S
S
N
F
H
.S
A
ta
Q
ti
S
a
it
r
e
7
o
n
6
o
.S
F
F
.S
ta
ti
ta
ti
o
n
3
n
1
n
o
ti
ta
.S
F
F
C
.B
e
.C
m
h
e
u
ta
rc
ry
h
0
0 .0 0
S it e ID
August 25, 1998: Toluene
(b)
16.00
14.00
Hour
ppbC
12.00
10.00
12
8.00
14
16
6.00
4.00
2.00
n4
St
at
io
r
So
da
te
Si
NH
AQ
n7
F.
St
at
io
n6
St
at
io
F.
n3
St
at
io
F.
n1
0
St
at
io
F.
B.
Ch
ur
ch
F.
Ce
m
et
ar
y
0.00
Site ID
Figure 2.1.5. (a) Isoprene and (b) toluene 1 hour integrated canister sample results for
Manchester, NH on August 25, 1998 for hour 12 (12 pm EDT), hour 14
(2 pm EDT), and hour 16 (4 pm EDT).
16
2.1.6 Sodar and Meteorological Instrumentation
During the 6/98, 8/98, and 5/99 field campaigns, a mini Doppler acoustic sodar
(Aerovironment, Inc., Model 4000) was deployed at a fixed site and operated on a continuous
basis during the measurement periods. A sodar is designed to measure wind speed, wind
direction and other turbulence data of the atmosphere vertically at a fixed site. Typically, the
system was set to collect 15 minute average measurements from 14 m to 280 m for 40 layers.
The data were then processed into hourly average profiles of wind speed, wind direction,
standard deviation of the vertical wind component (sigmaW), and standard deviation of the wind
direction (sigmaWD). Figure 2.1.6 shows an example vertical profile of wind speed and
direction obtained in Boston, MA during the 5/99 field campaign. The sodar data are compared
with MM5 output data (see Section 4.4) and are also processed to remove the influence of a
building directly upwind that impacted the sodar measurements at approximately 80 m.
280
280
240
240
200
200
160
160
Height (m)
Height (m)
During the 5/99 Boston field campaign, the WSU mobile van was fitted with a horizontal
sonic anemometer to provide on-site surface wind speed and direction data. The unit was
mounted on the roof of the van and the signal recorded at 1 Hz on the van data acquisition
system. Due to the aerodynamics of the van, an accelerated wind field dominates measurements
when the van is moving, therefore data were only collected when the van was stationary. These
wind data are merged with the WSU van instrument data and GPS data
120
120
80
80
40
40
0
90
0
120
150
180
210
240
270
300
330
360
390
420
450
0
WindDirection(deg)
sodar
mm5
2
4
6
8
10
12
14
WindSpeed(m/s)
sodar avg
sodar
mm5
sodar avg
Figure 2.1.6. Comparison of MM5 and Sodar Wind Direction & Wind Speed at MIT on
May 25, 1999 at 5 PM.
17
2.2 ARI Mobile Laboratory
The Aerodyne Research Inc Mobile Laboratory is a step-van equipped with real-time trace
gas and particle instruments and a global positioning system. Our instrumented mobile
laboratory easily performs real-time fast response (1 sec), simultaneous measurement of multiple
trace gases under normal driving conditions. Fast response measurements allow us to detect
discrete changes with location and to react immediately to probe further into these changes. All
trace gases are sampled from a common shielded, forward-facing inlet installed on the front of
the van at a height of ~2 m. The particle instrument sampled from a separate inlet. The particle
inlet was a tandem arrangement of rear-facing inlets positioned about 1-foot above the roof of
the mobile laboratory, 8-inches horizontally offset. One flow stream was directed directly into
the instrument and the other incorporated a 3-foot heated length that was maintained at 300 C.
The mobile laboratory was equipped with sensitive, real-time sensors for greenhouse,
aerosol precursor and ozone precursor trace gases and fine particles; a global positioning system;
an Eppley total ultraviolet radiometer and a central logging computer. Trace gas sensors
deployed varied slightly over the course of the measurement campaigns of the program.
An ARI Zeeman HeNe tunable infrared laser differential absorption spectrometer
(TILDAS) for methane detection was deployed in the first survey measurements in Manchester,
NH in November 1997. The ARI two-color TDL system was deployed in the subsequent three
campaigns in June and August 1998 in Manchester and in May 1999 in Boston, MA. The TDL
monitored greenhouse trace gases (N2O, CO, CH4) in June 1998, and aerosol and ozone
precursors (NO and NO2) in August 1998 and May 1999. In addition the ARI mobile laboratory
had a Licor instrument to measure CO2 and a WSU fast response electron capture detector
(ECD) for SF6 detection in all measurement campaigns in 1997 – 1999. A TSI condensation
nuclei instrument for particulates detection was deployed in the van in the latter 3 measurement
campaigns (1998 – 1999). The real – time instruments have been described in more detail above,
and in previous publications [see Lamb et al., 1995; McManus, et al., 1991; Benner & Lamb
1985; Nelson, et al., 1996; Wormhoudt, et al., 1995; Zahniser, et al., 1995].
A global positioning system (Trimble Pro XR) with real-time differential correction
collected position information of the van at 1 Hz with an accuracy of ~0.3 m. Real-time
differential correction was possible by collecting radio beacon data from a stationary site
simultaneously with the gps satellite data. In addition, we performed post-corrections on the data
using local basestation data (available on the internet) for infrequent short drop-out periods in
the beacon signal. The travel speed and acceleration of the mobile laboratory was calculated
from the position data.
Data from the individual instrument were logged on a central computer, enabling all data
streams to be stored synchronously. Just prior to the May 1999 campaign, Trimble software
became available which allowed real-time output of the GPS position data to the data logging
computer. In all of the campaigns the GPS data was also stored on a separate laptop PC in
Trimble proprietary format. The clocks on the laptop and central logging computers were
synched at least 2 times a day (at start and end of data collection). The GPS data was later
exported into ASCII format and merged with the remaining data sets.
18
Raw data were archived daily on an ARI server and on zip disks. Post-processing
procedures merged the data onto a common 1 sec time stamp and eliminated fall-out periods in
the data stream. The result is a set of ascii data files; one for each day. The processed data was
maintained on the ARI server and posted on the Aerodyne FTP site for access by non-ARI
participants.
2.3 WSU Mobile Laboratory
During the 5/99 Boston campaign, a second mobile laboratory was set up by WSU. This
van, fitted with commercial gas pollutant analyzers and an SF6 analyzer, was used to
complement the data obtained with the ARI van. During selected study periods, the van was
used to map pollutant concentrations along roads, while in other situations, the van was parked
and data were collected over an extended time period at one location.
The van was a 15-passenger van with several seats removed. The van instruments
included commercial NOx (Teco Model 42), CO2 (Licor 6262) and CO (Monitor Labs Model
9830) analyzers mounted in an instrument rack secured in the van. An O3 instrument was also
installed; however, it did not perform well and the data were not utilized. A WSU real-time SF6
analyzer was also mounted inside the van (described previously). A Teflon sampling manifold
fixed to the roof of the van was used to drawn air into the van for analysis. A nephelometer
(DataRam PM2.5) fitted with a PM2.5 inlet was mounted on the roof of the van and used to
measure aerosol concentrations during fixed location sampling. Tests of the system showed that
the nephelometer inlet did not perform correctly when the vehicle was moving at speeds greater
than approximately 15 mph. In addition to the gas and PM analyzers, the van was fitted with
horizontal sonic anemometer on the roof (used only in fixed sampling locations) and a GPS
position recording system.
In the field, the NOx instrument and CO instruments were calibrated periodically using a
dynamic dilution system for NOx and commercial gas mixtures for CO. In the dynamic dilution
system, a primary standard (Scott Specialty Gases, 1% accuracy) was diluted in a glass mixing
chamber with scrubbed dilution air at rates controlled with Tylan mass flow controllers
(0 to 50 sscm for the standard, and 0 to 5 slpm for the dilution air). For CO, a single standard
mixture was used (Scott Specialty, Inc., 5% accuracy). Operating parameters for the various
van instruments are summarized in Table 2.3.1. In several cases, the WSU van and ARI van
were parked at the same location for comparison purposes.
Two laptops operated as the data acquisition systems for the WSU van. The first laptop
gathered data every 0.1 Hz from the trace gas instruments, nephelometer, and the sonic
anemometer. Data were taken every 1 Hz during the tracer release tests. The second laptop was
dedicated to the GPS system gathering time, latitude, longitude and elevation above mean sea
level (MSL) data every 1 Hz. Post-processing of the data was required to merge the two datasets
and remove anemometer data during periods of van movement. The final result is a set of ascii
data files; one for each day, with the following columns of data:
19
Time (hh:mm:ss)
CO2 (ppm)
CO (ppm)
NO (ppb)
NO2 (ppm)
Latitude
Longitude
Height above MSL
Wind speed (3 minute running mean)
Wind direction (3 minute running mean)
Table 2.3.1 - Summary of WSU Van Instrumentation Operations
Instrument
Concentration Range
Response or
Estimated accuracy (%)
integration time
0.5 ppb – 200 ppb
1 Hz
+/- 0.5 ppb precision
0.05 ppm – 200 ppm
20 seconds
+/- 0.1 ppm precision
WSU SF6 analyzer
300 ppt – 10 ppb
1 Hz
+/- 5%
DataRam
mg/m3 –
1 Hz
+/- 5% of reading
Nephelometer
400 mg/m3
Teco NOx
Model 42
Monitor Labs CO
Model 9830
Model PM2.5
Horizontal Sonic
0 to 25 m/s WS
Anemometer
0 to 360 deg WD
Trimble Pro XRS
Not Applicable
1 Hz
1 Hz
50 cm with 5 satellites
GPS System
CO2 Licor 6262
3 ppm – 3000 ppm
+/- 1 ppm
An example of data from the WSU van are shown in Figure 2.3.1 for CO2, CO, NO, and
NO2 obtained May 25, 1999.
20
1200
CO2 (ppm)
1000
800
600
400
200
0
12:00:00 13:00:00 14:00:00 15:00:00 16:00:00 17:00:00 18:00:00
Time (hr)
40
35
CO (ppm)
30
25
20
15
10
5
0
12:00:00
13:00:00
14:00:00
15:00:00
16:00:00
17:00:00
18:00:00
Time (hr)
Figure 2.3.1. (a) CO2, (b) CO concentrations measured by the WSU van on May 25, 1999 in
South Boston, MA [(c) NO, and (d) NO2.concentrations shown on next page].
21
600
NO (ppb)
500
400
300
200
100
0
12:00:00 13:00:00 14:00:00
15:00:00 16:00:00 17:00:00 18:00:00
Time (hr)
600
500
NO2
400
300
200
100
0
12:00:00 13:00:00 14:00:00 15:00:00 16:00:00 17:00:00 18:00:00
Time (hr)
Figure 2.3.1. (a) CO2, (b) CO, (c) NO, and (d) NO2 concentrations measured by the WSU van
on May 25, 1999 in South Boston, MA.
22
2.4 Field Measurement Sites
Two urban areas were selected as field measurement sites for the program. Manchester,
New Hampshire was chosen as the test city for the program. Our goal was to use the test city to
test and refine the instrumentation and measurement strategies before planning and
implementing a field campaign in the larger urban area of Boston, Massachusetts.
2.4.1 Manchester, New Hampshire
The test city for the study was Manchester, NH with a population of 100,000.
Measurements were performed there in 1997 and 1998. Located in Southcentral
New Hampshire, Manchester is a compact industrial city on the Merrimack River. Manchester
was selected for this study because of its location close to the ARI facilities, its isolation from
other urban areas, its mix of industrial and residential areas, and its road system which includes a
loop road surrounding the city. The loop road provided easy access to points upwind and
downwind of the city.
2.4.2 Boston, Massachusetts
The second urban area selected for the study was the Boston metropolitan area. Boston,
Massachusetts is located on the northeast coast of the United States and has a population of over
2,870,000. The area has a generally flat topography and is bordered on the east by the Atlantic
Ocean. Two general sections of Boston were chosen for our study in May 1999. We focused
our mobile measurements on the bordering Dorchester and Roxbury communities of Boston
(see Figure 2.4.1). Dorchester and Roxbury, two bordering urban regions, were chosen because
they contain a mixture of different kinds of residential and industrial land use. The areas
encompass both light and heavy traffic regions and the density of housing is variable. There is
also a large park area (Franklin Park) on the western edge of the region. The measurement
region covered about 4 x 6 km., approximately the area of one standard modeling cell. Attempts
to study the entire Boston region under a given set of meteorological conditions with the mobile
laboratory would have been impractical.
2.4.3 Cambridge, Massachusetts
The second measurement site chosen in the Boston metropolitan area was a fixed location
on the campus of the Massachusetts Institute of Technology (MIT) in Cambridge, MA. At the
end of the May 1999 campaign, about 46 hours of continuous, fixed location data were collected
at this site. The site, a campus parking lot along the south side of Main Street, was a few blocks
from the Charles River. The Charles River separates Cambridge and Boston. The box in
Figure 2.4.2 by Main Street gives the approximate location of the stationary test site. Main
Street, an artery leading from MIT, leads directly onto a bridge (the Longfellow Bridge), a main
connection between Cambridge and Boston. The site was downwind of Boston for a portion of
the measurement period.
23
Figure 2.4.1
Boston Mobile Measurement Region, comprising principally of Dorchester and
Roxbury, MA.
Figure 2.4.2
Stationary Field Measurement Site on the campus of MIT in Cambridge, MA.
24
2.5 Field Measurement Strategies
Our field measurement approach is to combine real-time measurements of multiple trace
gases and particulates with meteorological data collection. In this NASA program we
implemented both mobile and stationary measurement strategies, with the emphasis on mobile
measurements. Mobile measurements can identify the distribution of local sources in an urban
area and thus better correlate urban activity with emissions. Intensive stationary data collection
with the instrument suite can simulate a fast response monitoring site and compared to the results
of averaging such data.
Several measurement strategies were followed in the mobile measurements of the field
campaigns: pollution concentration surveying and mapping, determination of mobile source
emissions ratios, and area source tracer ratio method experiments. These methods have been
described in detail previously [Lamb, et al., 1995; Jimenez et al., 2001] and will be summarized
in the following Subsections (2.5.1- 2.5.3).
2.5.1 Pollution Mapping
There are two general approaches to collecting trace gas data as a function of location.
Concentration or pollutant surveys are a coarse set of traverses in an area to identify major
emission sources, while pollutant maps are a finer set of traverses designed for more detailed
identification of source location and trends in concentration with location. During “mapping” we
record trace gas concentrations on a dense grid within the city, together with precise position
data from the GPS, as a data set that can yield a map of trace gas concentrations. The data set is
also available for model-based inversion to determine source strengths. How fine the grid or
series of traverses is that one follows when mapping, is determined primarily by the time
required to traverse a section of the area. This in turn is governed by the size of the measurement
area, the time of day, the traffic patterns, and the time required to cover the area. Ideally, one
would like to complete the mapping entirely under the same meteorological conditions.
It is difficult to complete detailed pollutant maps of large urban areas such as Manchester
and Boston. The size and heavy traffic precludes attempting to traverse the area many times in a
fine grid within a reasonable period of time. Instead we collected general surveys covering the
entire Manchester area, and several more detailed pollutant maps in selected areas of
Manchester. An example of a map of CO mixing ratios in Manchester, NH is given in
Figure 2.5.1. The data points on the positions where they were measured are colored and sized
according to the CO mixing ratio measured at that point. The CO data was collected at 1 Hz by
the TDL system on June 16, 1998. The CO ranged from 192 ppb to 10.2 ppm, with the primary
source of CO being vehicle exhaust.
Boston covers an even greater land area than Manchester, with higher traffic levels. Our
approach was to select a limited area of Boston for the measurement campaign. The
Dorchester/Roxbury area, covering about 4 x 6 km, was manageable for surveys and coarse
25
10x10
3
43.04
8
Latitude (deg)
43.02
4
CO (ppb)
6
2
43.00
42.98
42.96
42.94
-71.48
-71.44
Longitude (deg)
-71.40
Figure 2.5.1. CO data collected at 1 Hz by the TDL system on June 16, 1998 in Manchester,
NH. The carbon monoxide mixing ratio ranged from 192 ppb to 10.2 ppm, with
the primary source of CO being vehicle exhaust.
mapping. The heavy traffic during the daytime measurements, prevented us from completing
fine maps of the area. However, the maps which we did generate still yields overall emission
characteristics of the area. An example of a map of NO mixing ratios is given in Figure 2.5.2.
We display the NO data from the routes into and out of Boston from Aerodyne Research, Inc.
located in Billerica, a suburb northwest of Boston, as well as NO levels on traverses in Boston.
In this figure the Dorchester/Roxbury area is in the lower right side of the map, where there is a
denser grid of data. Each data points is on the position where it was measured and is colored and
sized according to the corresponding NO mixing ratio. The data was collected on 5/25/99. The
NO color range is 0 to 500 ppb. Two large gaps in the data appear when the mobile van entered
a tunnel and lost GPS satellite coverage (e.g. at approximately 42.37 deg latitude and –71.06
degree longitude). The primary source of high NO was vehicle exhaust.
Additional examples and results of the pollutant mapping will be given in Section 4.2 of
this report.
26
42.50
latitude (deg)
42.45
42.40
42.35
42.30
0
100 200 300 400 500
NO (ppb)
-71.25
-71.20
-71.15
-71.10
longitude (deg)
-71.05
Figure 2.5.2. NO data collected at 1 Hz by the TDL system on May 25, 1999 in the
Boston, MA area, including the routes to/from Billerica, MA. The nitric oxide
mixing ratio in the figure ranges from 0 to 500 ppb.
27
3.0 FIELD DATA OVERVIEW
In this section we summarize the available data collected during the four mobile
measurement campaigns during the Urban Respiration project. The campaigns and the type of
experiments and data collected are summarized below in Table 3.1.
Table 3.1 - Mobile Campaign Data
Manchester Campaign, November 1997
Date & TOD
Weather
Data Collected
Exper. Type(s)
Note
--------------------------------------------------------------------------------------------------------------------11/10
CO2, CH4, canisters
Survey
-------------------------------------------------------------------------------------------- -----------------------------------------------11/11, 4-8pm
clear & cool
CO2, CH4, canisters
Survey, rush hour buildup
Vets day holiday
N/NW, 7-8 kt
------------------------------------------------------------------------------------------------------ -------------------------------------11/13
CO2, CH4, canisters
Survey
----------------------------------------------------------------------------------------------------------------------------- ---------------
Manchester Campaign, June 1998
Canister data from 6/14, 15, 16, 17, 19; Days with mobile data: 6/16, 17, 19 = 3
Date &TOD
Weather
Data Collected
Exper. Type(s)
Note
-------------------------------------------------------------------------------------------------------------------------- -----------------6/16, 2-11 pm
cloudy
N2O, NO CO2, CO
Extensive survey
Tracer test scrubbed
var wind,NE to SW
Particles
* Used in Chemosphere paper
-------------------------------------------------------------------------------------------------------------------------------------------6/17, 3-9 pm
cloudy
“
Survey
var. Wind, NE to E
“
+Balloon
----------------------------------------------------------------------------------------------------------------------------- --------------6/19, 7am- 2pm p. Cloudy
“
Tracer
rain later +
SF6
+Survey
----------------------------------------------------------------------------------------------------------------------------- ---------------
28
Manchester Campaign, August 1998
Date & TOD
Weather
Data Collected
Exper. Type(s)
Note
----------------------------------------------------------------------------------------------------------------------------- --------------8/22, 3p-12M clear, light wind NO, NO2, CO2,
multiple loops around city before sunset.
particles, uv (at sodar)
after sunset, 1 loop then cross town
----------------------------------------------------------------------------------------------------------------------------- --------------8/24, 5-8p hot & hazy, light wind NO, NO2, CO2,
Local measurements
High ozone day part’s, Van uv, O3
(Billerica/Burlington/Bedford)
------------------------------------------------------------------------------------------------------------------------ -------------------8/25, 10a-7p Hot and humid.
“
“
mapping &city traverses
NO laser drifting.
----------------------------------------------------------------------------------------------------------------------------- --------------8/26, 2-9p
“
“
neighborhood traverses
stationary measurements
by EPA monitoring station
----------------------------------------------------------------------------------------------------------------------------- --------------8/27, 2-8p west wind
“
“
tracer test
NO problematic again
----------------------------------------------------------------------------------------------------------------------------- --------------8/28, 12-5p S/SE winds
“
“
tracer test
-------------------------------------------------------------------------------------------------------------------------------------------8/30, 12-6p W/NW winds
NO2, CO2
tracer test
not collecting NO
part’s, Van uv, O3
response time
tests
Boston Campaign, May 1999
Date & TOD
Weather
Data Collected
Exper. Type(s)
Note
----------------------------------------------------------------------------------------------------------------------------- --------------5/21, 11a-7p
clear, breezy
NO, NO2, CO2
Mapping
Franklin Park bkgn
Fri
E wind
part’s, Van uv, O3
ARI + WSU mobile
----------------------------------------------------------------------------------------------------------------------------- --------------5/22, 11a-5p
clear, breezy
“
“
Mapping
JFK Lib Bkgn
Sat
NE wind
Some NO data lost
----------------------------------------------------------------------------------------------------------------------------- --------------5/23, 1p-6:30p overcast/drizzle “
“
Mapping
JFK Lib co-sample
Sun.
S wind
----------------------------------------------------------------------------------------------------------------------------- --------------5/25, 11aSunny/PC
“
“
Mapping
Tu
SW/W wind
+Tracer @ Ceylon Plg.
-------------------------------------------------------------------------------------------------------------------------------------------5/26
Cloudy to PC
“
“
Mapping
Wed
SW wind
+Early Tracer (~8 am)
-------------------------------------------------------------------------------------------------------------------------------------------5/27, 2p-...
Clear/PC
“
“
Stationary at MIT
Thur
warm
----------------------------------------------------------------------------------------------------------------------------- --------------5/28
“
“
“
Stationary at MIT Big Mem. Day Traffic Tie-up
Fri
With broken bridge
----------------------------------------------------------------------------------------------------------------------------- --------------5/29, ...-3p
“
“
“
Stationary at MIT
Sat
----------------------------------------------------------------------------------------------------------------------------- --------------TDL data from 5/21, 22, 23, 25, 26, 27, 28, 29 = 8 days
29
30
3.1 ARI Trace Gas Data Description
The trace gas data was collected at approximately 1 Hz. The data from the TDL, CO2
Licor, the GPS, and UV radiometer, were merged into a single file for each experimental day,
with data interpolated onto the 1 sec grid of the GPS data. The mixing ratios of the TILDAS
data (e.g. NO, NO2 in the May 1999 campaign) and , O3 when measured, and CO2 and uv
intensity (in August 1998 and May 1999) are combined in the data files with position data
measured with the GPS. The data files, in ASCII format, are stored on the ARI FTP site.
Each file contains tab delimited data. In general, the data is listed in the following order:
date&time
O3 CO2
latitude
uv
longitude
altitude
tdl species1
tdl species2 … tdl species n.
where ozone and uv level were only collected in the August 1998 and May 1999 campaigns, and
n is the total number of species measured with the TILDAS instrument. The time is the GPS
time is given in Greenwich Mean Time (GMT). The latitude and longitude are given in degrees
and the altitude is in meters (m). The mixing ratios of the tdl species and O3 are in parts per
billion by volume (ppbv); CO2 is in units of parts per million by volume (ppmv), and the uv
intensity is in mW/cm2.
There is data available for determining emission ratios from mobile combustion sources.
The data that are available is summarized in Table 3.1.1.
31
Table 3.1.1 - Data Available: Mobile Combustion Source Emission Ratios
Gas:
Date
Location
CH4
CO N2O NO
NO2
---------------------------------------------------------------------------------11/10/97
Manchester | x |
|
|
|
|
11/11/97
11/13/97
Manchester
Manchester
| x
| x
|
|
|
|
6/16/98
6/17/98
6/19/98
Manchester
Manchester
Manchester
|
|
|
| x
| x
| x
8/22/98
8/24/98
8/25/98
8/26/98
8/27/98
8/28/98
8/30/98
Manchester
Manchester
Manchester
Manchester
Manchester
Manchester
Manchester
|
|
|
|
|
|
|
5/21/99
5/22/99
5/23/99
5/25/99
5/26/99
5/27/99
5/28/99
5/29/99
Boston
Boston
Boston
Boston
Boston
Boston
Boston
Boston
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| x2 | x
| x | x
| x | x
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
x
x
x1
x
x1
x
|
|
|
|
|
|
|
x
x
x
x
x
x
x
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
x
x
x
x
x
x
x
x
|
|
|
|
|
|
|
|
x
x
x
x
x
x
x
x
|
|
|
|
|
|
|
|
Notes: 1: NO laser drifting & problematical
2: N2O data used in Chemosphere paper
32
3.2 UNH Total Particle Data Description
The fine particulate data were collected using National Instruments hardware and
LabView software on a standard laptop PC. Temporal adjustment was accounted for an inlet
residence time of approximately 2.5 seconds. Bad data, generally caused by equilibration effects
after switching from one inlet to the other (heated versus non-heated inlets), were filtered out
during post processing. This filtering resulted in a loss of about 10 seconds of data every
2 minutes. Aerosol number density data was otherwise reported at 1 hertz.
During the May 27 to May 29 1999 stationary intensive sampling campaign, bulk aerosol
composition measurements were made in addition to particle number density and CO2. These
measurements were accomplished using a standard Teflon filter exposure technique (see e.g.,
Talbot et al., 1992). Subsequent methanol/deionized water extraction and analysis by ion
chromatography was performed at the University of New Hampshire. Mixing ratios of aerosol
Na+, Cl-, Mg2+, NO3-, SO4=, NH4+, Ca2+, K+, PO43-, CH3COO-, HCOO- were reported for the hour
long (and half hour long during periods of high automobile traffic) sample integration period.
33
3.3 WSU Trace Data Description
Tracer tests were conducted during three of the four field campaigns. A summary of the
tracer test periods is given in Table 3.2.1. Then data is archived at WSU and is available as
described in Section 5.3.
Table 3.2.1 - Summary of Tracer Test Periods
SF6 Release
Location
Parking Lot
Downtown
Manchester, NH
Lat = 42.994 deg
Lon = -71.46 deg
<Same as above>
Date
Start Time Stop Time
(EDT)
(EDT)
8/27/98 4:06 pm
6:02 pm
Release
Rate (g/s)
0.926
8/28/98 2:58 pm
5:15 pm
0.756
<Same as above>
8/30/98 2:22 pm
5:06 pm
0.824
South Boston,
Ceylon Park
Lat = 42.3098 deg
Lon = -71.0751 deg
South Boston,
Playground south of
Gallivan St. and
Norfolk St.
Boston, corner of
Beacon St. &
Walnut St.
5/25/99 12:05 pm
5:40 pm
0.737
5/26/99 7:53 am
12 pm
0.787
5/28/99 3:46 pm
10 pm
0.633
34
Comments
Mobile test,
ARI Van
Mobile test,
ARI Van
Mobile test,
ARI Van
Mobile test,
ARI Van &
WSU Van
Mobile test,
WSU van SF6
analyzer not
functional
Stationary test,
WSU van SF6
analyzer not
functional
3.4 WSU VOC Data Description
During each field campaign, VOC whole air samples were collected at selected sampling
points within the urban area using portable canister samplers. Samplers were deployed
immediately prior to the sample period and collected immediately following the sample period.
Typically, 3 hr averaged samples were collected with a pre-selected start time.
3.5 WSU Sodar and Meteorological Data Description
During the 6/98, 8/98, and 5/99 field campaigns, a mini Doppler acoustic sodar
(Aerovironment, Inc., Model 4000) was deployed at a fixed site and operated on a continuous
basis during the measurement periods. A sodar is designed to measure wind speed, wind
direction and other turbulence data of the atmosphere vertically at a fixed site. Typically, the
system was set to collect 15 minute average measurements from 14 m to 280 m for 40 layers.
The data were then processed into hourly average profiles of wind speed, wind direction,
standard deviation of the vertical wind component (sigmaW), and standard deviation of the wind
direction (sigmaWD).
During the 5/99 Boston field campaign, the WSU mobile van was fitted with a horizontal
sonic anemometer to provide on-site surface wind speed and direction data. The unit was
mounted on the roof of the van and the signal recorded at 1 Hz on the van data acquisition
system. Due to the aerodynamics of the van, an accelerated wind field dominates measurements
when the van is moving, therefore data were only collected when the van was stationary.
35
4.0 DATA ANALYSIS STRATEGIES
In this long section, we present several different analysis strategies for the data collected
during the urban respiration field measurement campaigns. This data set is a new type of
window into urban air quality, and considerable effort has been spent during this project to
develop techniques to analyze this novel data. As of the writing of this report, the full data set
has not been thoroughly analyzed by all of the methods that we have explored. As we proceed
with the writing of planned papers [see Section 5.2] further analysis will be done. For the
analysis strategies presented below, we offer brief summaries and identify the lead institutions
responsible for those strategies.
Section 4.1, Motor Vehicle Pollutant Emissions from Mobile Measurements
[lead: Aerodyne Research]
We present methods of deriving motor vehicle pollutant emission indices from mobile
measurements. The emission index is a ratio of pollutant to CO2, which can be used to derive
aggregate vehicle fleet emissions, given an average mileage. We present emission indices for
NO, NO2, CO and CH4, with comparison to other published results.
Section 4.2, Background Pollutant Maps
[lead: Aerodyne Research]
We examine the lower concentrations of pollutants, between the high concentrations
produced by the many local sources [i.e. vehicles], to search for overall patterns in urban
pollutants.
Section 4.3, Fixed Site Pollutant Measurement Analysis
[lead: Aerodyne Research]
During a two day period in May of 1999 the mobile instrumentation was held stationary,
to mimic a high sensitivity fast response monitoring site. We observed the diurnal buildup of
pollutants and contributions due to local sources.
Section 4.4, Mesoscale Wind Field Modeling
[lead: Washington State University]
The MM5 meteorological model was used to simulate the wind field for the New
England coast during the three major field measurement campaigns. Prevailing winds are used
to estimate surface winds, which allows a back-propagation calculation of the pollutant footprint.
Section 4.5, Turbulence Modeling of Urban Landscapes
[lead: Washington State University]
Tracer data analysis is used to help understand turbulence and plume spread in an urban
boundary layer. Further, turbulence modeling of an urban landscape combines mesoscale
modeling of the urban windfield with a 3-d turbulence model of the urban landscape.
Section 4.6, Urban Footprint Modeling
[lead: Washington State University]
36
Application of plume diffusion theory along a back-trajectory yields the upwind source
distribution (source footprint) affecting a receptor at the starting point. Back trajectory
calculations were performed for the Boston campaign.
37
Section 4.7, Urban Emissions Air Quality Relationships
[lead: MIT Dept. Chemical Engineering]
We describe an improved, faster technique for solution of the inverse problem of
determining the spatial temporal and chemical form of pollutant emissions, given a sparse set of
measurements. The methods are demonstrated on an earlier set of measurements in Los Angeles.
Section 4.8, Model Inversion of Pollutant Maps
[lead: MIT Dept. Chemical Engineering]
Tracer release studies are analyzed in terms of a Gaussian plume model to help in
understanding of dispersion in the urban environment and how to better perform the inversion
techniques described in Section 4.7.
Section 4.9, Photochemical Steady State NOx Analysis
[lead: Aerodyne Research]
Mobile data for NO, NO2, O3 and ultraviolet are combined in a simple photochemical
steady state model linking these quantities. We found that the measurements were generally
consistent with photochemical steady state, except when very close to local sources.
Section 4.10, GIS Based Emissions Analysis
[lead: MIT Dept Urban Studies and Planning]
State-of-the-art geographical information system (GIS) techniques are presented that
allow the overlay of air pollutant concentration measurements on maps of urban population,
economic activities, and transportation infrastructures. Examples for the Manchester, NH and
Boston, MA metropolitan areas are presented.
Section 4.11, GIS Based Pollutant Activity Comparisons
[lead: MIT Dept Urban Studies and Planning]
The utility of correlations observed urban pollution ratios with available GIS based
activity factor distributions is examined. The capabilities of currently available GIS software is
critiqued and suggestions for improvements are presented.
Section 4.12, Fine Aerosol
[lead: University of New Hampshire]
We present results of measurement of fine aerosols (7 - 3000 nm diameter) from mobile
and stationary sampling periods, with correlation to NO and CO2.
38
4.1. Motor Vehicle Pollutant Emissions from Mobile Measurements
Measurements of pollutant emissions from motor vehicles under real world conditions
are useful for verification fleet emissions estimates based dynamometer measurements and fleet
composition models. One very direct way to validate fleet emissions models is by pollution
measurements on highway or city streets, for which several techniques have been reported. For
example, “tunnel studies” employ a roadway tunnel as an integration volume that allows
measurement of the average emissions for a large set of vehicles [Berges et al., 1999; 2000;
Sjödin et al., 1995; 1998; Becker et al., 1999; 2000]. Tunnel studies cannot easily provide the
statistical distribution of emissions, and the range of driving conditions in tunnels is limited.
Alternatively, individual emissions measurements can be conducted with commercially available
cross-road optical sensors, combined with license plate reading and vehicle identification
[Bishop et al., 1989; Zhang, 1996]. However, cross road measurements with current commercial
instrumentation suffers some restrictions on location and vehicle speed due to limited range and
sensitivity. New laser based cross-road instrumentation developed at Aerodyne Research, Inc. is
more sensitive than commercial lamp based instruments, and provides an alternative method of
measuring on-road vehicle emissions [Jiménez et al., 1997; 1999; 2000; Nelson et al., 1998;
1999].
Another way to measure aggregate motor vehicle pollution emissions is to continuously
sample exhaust plumes from a mobile platform that moves through traffic. Fast-response highsensitivity measurements of pollutant gases may be combined with simultaneous CO2
measurements so that ratios of elevations of pollutant and CO2 provides a molecular emission
ratio. Reporting data as a molar emission ratio, e.g. ER = NO / CO2 in the vehicle exhaust, is
a more basic measured quantity than other commonly reported measures, such as “grams per
kilometer”. The emission ratio can be converted to an extrapolated concentration as the gas
leaves the tailpipe, assuming stoichiometric gasoline combustion, which can be estimated to
good approximation by multiplying the emission ratio (e.g. NO / CO2) by an assumed tailpipe
CO2 mixing ratio of 0.13. Molecular emission ratio then can be put in terms of emission per unit
fuel consumption, based on average fuel use rate. With mobile instrumentation, emission ratio
data can be collected rapidly over a wide area and under a wide range of driving conditions.
The Urban Respiration project has provided an opportunity to generate statistics on
aggregate emission ratios (ER’s) measured with a mobile platform for the emitted gases: CO,
CH4, NO, NO2 and N2O. New methods to extract emissions ratios have been developed under
this contract. We derive average ER’s as well as their distributions. We observe variations in
ER’s in different situations, driving conditions, and road types. We derive statistical properties
of the emissions, but generally do not identify individual emitting vehicles. In follow-on work,
we identified specific vehicles by following closely while video recording and measuring their
emissions.
As of the writing of this report, the mobile data set has not yet been thoroughly analyzed
to extract the full set of emissions information. The methods of mobile data analysis have been
under development throughout the project, and some results are presented at a more advanced
39
state than others. We expect the analysis of the data to continue, and for further publications to
result.
4.1.1. General Results of Mobile Measurements
When we measure trace gases with our mobile laboratory on congested city streets, and
when the measured gases are typically found in motor vehicle exhaust, we generally record a set
of concentration peaks. We observe peaks in concentration for the major exhaust gas, CO2, as
well as minority species, which have included CO, NO, NO2, N2O and CH4. A representative
sample of data is shown in Figure 4.1.1. There generally is a strong correlation between CO2 and
the minority species, which leads us to believe that the minor species are co-emitted with CO2 by
local compact combustion sources, i.e. motor vehicles.
The data record contains a large number of peaks, of variable height and width in space
and time. It is not immediately clear from the data record if the peaks correspond to single
vehicles or groups of vehicles, but a consideration of the peak magnitudes gives some
indications. In general, the peaks in CO2 concentration range from ~10 to ~200 ppm above the
background level (of approximately 375 ppm). If we consider the CO2 peaks as the result of
complete combustion of hydrocarbons (generic formula (CH2)x), the initial CO2 concentration in
the exhaust gas is ~13 %, so the peak concentration range noted above represents exhaust
dilution factors of ~700 to ~13,000. The dilution factor in a plume will be approximately the
500
480
CO2, ppm
CO2
NO2
NO
140
1400
120
1200
100
1000
80
800
60
600
40
400
20
200
420
NO, ppb
440
NO2, ppb
460
5/25/99
Boston, MA
Dudley Sq
400
380
360
0
0
10.3
10.4
10.5
10.6
10.7
10.8x10
3
File point, (seconds)
Figure 4.1.1. Typical segment of mobile concentration data showing coincident peaks of CO2,
NO and NO2.
40
same as the increase in plume area as it spreads, so these dilution factors imply increases in
plume diameters by factors of ~25 to ~100. If the plume is from vehicle exhaust, where the
tailpipe diameter is ~5 cm, then the diluted plume diameter would be ~1.3 to ~6 m. Thus, the
smaller plumes at higher concentrations would likely be from single vehicles. The lower
concentration peaks could be from groups of vehicles, due to the larger effective plume size and
with the opportunity to mix the exhaust plumes from multiple vehicles.
One manifestation of the “peakiness” of the concentration data is the high (positive)
statistical skew. Skew is the normalized third moment about the mean, which measures the
excursion of the data preferentially to one side of the mean. For Gaussian fluctuations, which are
symmetric about the mean, the skew is zero. For example, the statistics for three gases on
5/25/99 in Boston (city driving, ~14,000 seconds) were as follows:
Gas
CO2
NO
NO2
Average
402 ppm
76.7 ppb
17.8 ppb
STD
Skew
26.6
130
12.4
2.3
5.8
2.9
While we will later examine various methods of extraction of emission ratios,
correlations in the data records provide simple global measures of the associations between
gases, as well as the average width of the peaks. The correlation between two data records (A(t),
B(t)) is defined as:
C(A, B; ) = ∫ A(t+) B(t) dt
.
If A and B are the same data record, then the above expression is the autocorrelation. The
interpretation of correlations is simplified if we consider data with zero mean, i.e with the
average subtracted (labeled as “ac”, e.g. Aac(t)). If the data record consists of random
fluctuations the autocorrelation of the ac-data will have a sharp peak at zero offset (=0) and will
tend toward zero elsewhere. If the data consists of peaks of uniform width (w0), then the
autocorrelation will have a peak of width ~w0 2. In Figure 4.1.2, we show the autocorrelations
of the ac-parts of CO2 and NO concentrations recorded on 5/25/99 in Boston. The sharp
autocorrelation peaks are expected from the long data records with many separate and randomly
placed concentration peaks. The autocorrelation peak widths (FWHM) are 18 seconds for CO2
and 6 seconds for NO. Thus, the average concentration peak widths are ~13 seconds for CO2
and ~4 seconds for NO. The narrower autocorrelation peak width for NO versus CO2 is a robust
feature of the data and is not due to differing instrument response times.
The association between CO2 and NO is seen in the correlation between these two gases,
shown in Figure 4.1.3. If we assume that the pollutant gas is present in proportion to the CO2,
then the average emission ratio (ER) is given approximately by the ratio of the peaks of cross
correlation to autocorrelation:
ER = C(CO2ac, NOac; 0) / C(CO2ac, CO2ac; 0)
For the case of NO and CO2 on 5/25/99 in Boston, the ER determined by this method is
2.9 x10-3.
41
14x10
6
350x10
CO2_peaks_AC Autocorrelation
12
10
300
250
CO2acAutocor
NOacAutocor
8
200
6
150
4
100
2
50
0
0
-400
NO_peaks_AC Autocorrelation
Boston 5/25/99
CO2 & NO peaks
from Rangemin
Autocorrelations
-200
0
Points offset
200
6
400
Figure 4.1.2. Autocorrelations of peak segments of CO2 and NO data, with subtracted means.
NO_pks_ac & CO2_pks_ac Correlation
40x10
6
30
20
10
0
-400
-200
0
Points offset
200
400
Figure 4.1.3. Cross-correlation of peak segments (with zero mean) of CO2 and NO, showing
the strong association between these gases.
42
4.1.2. Special Issues for Mobile Measurements of ER’s
Vehicle Identification
Measuring emission ratios in a laboratory moving on city streets presents several special
issues and problems. The first issue to consider is that measuring emission ratios was not the
primary experimental objective of the Urban Respiration project. Rather, it was a foreseen but
side benefit of the pollutant mapping effort. In the collection of data, we did not attempt (except
in a few cases) to identify the specific vehicle being measured. Therefore, the data collected is
treated as an aggregate sampling of emission ratios. However, wind tunnel studies show that onroad sampling predominately reflects the vehicle in directly front [Clifford et al., 1997]. In a
later project designed to measure emissions from specific vehicles [Shorter et al., 2001] , we
used a video camera and a close following measurement strategy.
Possible tail-wind contamination
Another issue in mobile measurements is the possibility of sampling the exhaust from the
mobile laboratory itself. That is especially a concern because a gasoline-fueled generator
without emission controls is attached to the rear of the mobile laboratory. We need to be certain
that we sample with a relative head-wind in the mobile lab. When wind comes from behind the
lab, we may sense our own exhaust, either from the lab (truck) itself or the generator on the rear
bumper. In addition to tail-wind, other on-board wind directions may have measurement
significance. Head wind should be completely free of interference from the mobile lab exhaust.
Wind from the left may bring more plumes due to local traffic (in the opposing lanes of traffic),
as compared to wind from the right.
We can estimate the occasions when tail-wind sampling might be occurring by
calculating the wind direction as sensed in the mobile lab at each point in the measurement
trajectory, by assuming that the wind reported at a central location (e.g. the airport) applies to the
entire measurement area. A more direct measurement might be with a lab-mounted anemometer,
provided that it is mounted far enough forward that the air flow is unaffected by the vehicle
itself.
While data contamination by tail wind sensing is possible, we need to examine the data to
determine how much it actually occurs. One way to present the mobile data that shows the
extent of tail wind influence is to plot concentration as a function of X and Y components of
measurement velocity. Furthermore, if we rotate the velocity basis to align with the prevailing
wind and scale by the wind speed, then the dynamical regions for tail wind sampling are
apparent and also are put into a standard form. The distinction between head wind and tail wind
in such a plot can be seen from the cosine of the wind as sensed on the mobile lab. We show in
Figure 4.1.4, a calculation of the cosine of the wind as sensed in the truck, as a function of truck
velocity. The cosine of the on-board wind direction is 1 for head wind, -1 for tail wind, and 0 for
side wind. The contour in Figure 4.1.4 at cosine=0 separates head wind from tail wind. We have
assumed for the calculation in Figure 4.1.4 that the wind is from the north at 1 m/s.
43
In Figure 4.1.5, we show a compilation of the mobile lab sampling velocities experienced
during traverses on May 25, 1999. The average CO2 concentration is tabulated for each velocity
bin. The velocities are rotated so that the wind direction is from the top of the figure, in order to
match the geometry of Figure 4.1.4. Figure 4.1.5 indicates that we may have measured the
mobile lab cosine in the velocity regime corresponding to on-board wind direction cosine <0, the
tail wind sampling regime in Figure 4.1.4. However, the enhanced CO2 in the tail wind velocity
regime is not particularly strong for 5/25/99. Observing this possible sampling artifact allows us
to exclude the corresponding velocity regime from subsequent data analysis.
2
1
0 0.4
-0.6
0.2
0
-0.4
-0.8
Truck Northerly Speed
Truckwind Orientation
Contours of Cos(head-angle)
Wind sourc e north, speed 1
cos=0 divides head/tail
'Vxy =-0.8'
'Vxy =-0.6'
'Vxy =-0.4'
'Vxy =-0.2'
'Vxy =0'
'Vxy =0.2'
'Vxy =0.4'
'Vxy =0.6'
'Vxy =0.8'
-0.2
-1
0.6
0.8
-2
-2
-1
0
Truck Easterly Speed
1
2
Figure 4.1.4. Cosine of the wind direction as sensed in the truck, as a function of truck velocity,
assuming a general wind from the north at 1 m/s. The cosine of the on-board
wind direction is 1 for head wind, -1 for tail wind, and 0 for side wind.
44
Boston 5 /25/99
CO2 vs Tru ck Vel oci ty
Rotated for North Win d
ppm , Blu e375 -Re d440
15
10
Northerly Velocity, m/s
5
0
-5
-10
-15
-15
-10
-5
0
Easterly Velocity, m/s
5
10
15
Figure 4.1.5. Average CO2 concentration for velocity bins during traverses on May 25, 1999.
The velocities are rotated so that the ambient wind direction is from the top of the
figure, in order to match the geometry of Figure 4.1.4. We may have measured
the mobile lab cosine in the velocity regime corresponding to on-board wind
direction cosine <0, i.e. tail wind sampling
Time response and speed filtering
In order to derive accurate emission ratios from mobile data that is rapidly changing, the
instrument response times must be well matched. During the setup of the instrument suite, the
flow response times of the CO2 and TDL instruments were adjusted to be nearly equal. The data
rate of the instrumentation on board the mobile lab is one data point per second, which is well
matched to the gas flow response times through the instruments. A comparison of the Fourier
transforms of data records from the CO2 (LICOR NDIR) and NO (tunable diode laser)
instruments can reveal the similarity of the instrument responses. The Fourier transforms of NO
and CO2 data from 5/25/99 is shown in Figure 4.1.6, where the data first had their averages
subtracted and were then scaled by the STD. Both CO2 and NO FT’s have similar low frequency
45
slopes, and a high-frequency roll-off beginning at ~0.2 Hz, indicating that these instruments are
reasonably well matched. The CO2 rolloff slope is somewhat steeper than for NO. When we
smooth the NO data with a 1-second Gaussian filter before the Fourier transform, then the
resulting frequency rolloff is more similar to the CO2 rolloff.
A consequence of the fixed time response is a variable spatial resolution. When the lab is
stationary, the 1 second sample corresponds to the distance the wind moves in that time. When
the lab is moving at 30 m/s, the minimum spatial scale that is discernable is 60 m. When moving
at high speed, the concentration profiles will tend to be averaged out, and concentration peaks
may appear wider and lower than they really are.
Fourier Transform Amplitude
1000
100
Boston 5/25/99
NO & CO2 FT's
pre: *-avg, */STD
CO2city_FT
NOci ty_FT
NOci ty_FTs m
10
1
100µHz
1mHz
10mHz
Frequency
100mHz
Figure 4.1.6. Fourier transforms of data records for CO2 and NO. The light blue curve is for
NO data that has been smoothed with a 1-second Gaussian filter.
46
4.1.3 Separation of “Peaks” and “Local Background”
The data records may be viewed as consisting of many concentration peaks above a
slowly varying “local background”. Empirically, we observe that between cleanly separated
peaks, the concentration returns to a similar level, of near 380 ppm for CO2 and near zero for
NO. In computing emission indices, we would like to concentrate on the local sources (i.e.
vehicles), rather than the diffuse background which represents more varied sources. Thus the
peak data is used to calculate ER’s, while the local background can be used to estimate urban
accumulation of gases.
Several methods have been investigated for separating the peaks. The simplest is to
scroll through the data and identify local minima, then to interpolate between those minima to
estimate the local background. Subtracting the interpolated background leaves the peak data.
A more well defined method is to define the local background at each point as the minimum
concentration obtained within some distance range along the traverse, in the backward and
forward directions. We have employed this “range-minimum” method to automatically derive
background and peak data components, usually with a range of ± 500 m. We take the additional
step of smoothing the background level before subtracting. A data sample is shown in
Figure 4.1.7 with a local-background line derived from the “range-minimum” method. For the
city-driving data on 5/25/99, the local background average and STD gas concentrations are
383 ± 14 ppm for CO2 and 1.7 ± 3.2 ppb for NO.
4.1.4. Methods of Deriving Emission Ratios
The ratio of the excursions above background, i.e. (delta-pollutant)/(delta-CO2) gives the
molecular emission ratio (ER) of the local sources. We are interested in the average ER as well
as the statistical variability. The various methods we have examined for ER determination
include: correlations, area ratios, linear regressions, point-by-point ratios and sliding window
regressions. For the investigation of methodology, we have concentrated on a single data set,
NO and CO2 emissions on May 25, 1999 in Boston.
47
500
480
1200
460
1000
440
800
420
600
400
400
380
200
NO, ppb
CO2, ppm
1400
co2
CO2_RM500
no
NO_RM500
0
360
88
90
92
94x 10
3
Linear Distance Traveled, m
Figure 4.1.7. A data sample with local-background lines derived from the “range-minimum”
method.
Correlations
As we described in Section 4.1.1, the cross-correlation integral can be used to give an
emission ratio based on aggregating the data into large blocks. For the 5/25/99 city data
segment, we arrive at ER=2.90 x10-3. However, the correlation method has several problems.
The correlation ratio method operates on all values of concentration, so low concentration values
(with mixing) can have a significant effect, possibly lowering the average ER. This method
reports a single average value, and not a distribution of values. The unequal widths of the
correlation peaks adds uncertainty to the interpretation of the reported ER.
Area ratios
If we subtract backgrounds from the data waves for pollutant and CO2, the ratio of the
total local source concentrations above background will yield an aggregate ER. In the rangeminimum procedure, the background level for each point is the minimum concentration
encountered within ±500 m from the measurement point. We calculated area ratios for the city
segment of the 5/25/99 traverses, (points 2000 to 16500), giving the aggregate emission ratio
NO/CO2 = 3.45 x10-3.
48
Point by point ratios
If we subtract backgrounds from the data waves for P & C, the ratio of the individual
peak points (concentrations above background) will be a local ER. The ratios will be unstable
as the peak concentrations approach zero, so a limit on minimum elevation will be needed.
When we use the range-minimum subtracted data, and consider only points where CO2 is 10 ppm
above background, we get average ER 3.53 ± 3.03 x10-3.
Linear Regressions
Linear regressions assume the relationship P = A + B*C, and find the coefficients A & B
for the whole data set. A & B are determined by minimizing the Chi-squared for the data set,
assuming the linear relationship. The explicit relationships are:
A= [ ∑ i Ci2 ∑ i Pi - ∑ i Ci ∑ i (CiPi )] / [ N ∑ i Ci2 - ( ∑ i Ci)2 ]
B= [ N ∑ i (CiPi) - ∑ i Ci ∑ i Pi ] / [ N ∑ i Ci2 - ( ∑ i Ci)2 ]
.
The goodness of fit is related to the r-coefficient:
r = ∑ i(Ci - Cavg)(Pi - Pavg) / [ N ∑ i(Ci - Cavg)2 ∑ i(Pi - Pavg)2 ] 1/2
.
For the city segment of the data, the regression slope gives an emission index of
NO/CO2 = 3.12 x10-3. A regression analysis for the highway segment #2 (returning to ARI
through evening rush hour), shows a slightly higher correlation slope of 3.54 x10-3.
When linear regressions are performed on long data records, the r-coefficient for the fit
tends to be rather low, indicating a poor fit. Indeed, a plot of NO vs CO2 for a long data set
generally shows a large scatter. However, shorter segments of data tend to show less scatter and
better fits (higher r-coefficients). We believe the reason for the large scatter and poor fits for
long data sets is the variation of emission index through the sampling period. The problem of
variation in the regression slope is addressed with the sliding-window regression technique
described below.
Sliding window regressions
Linear regressions can be performed over relatively narrow “windows” in the data, giving
intercept (A) and slope (B) as a function of location in the data waves. This may be more stable
than the point-by-point ratios of background-subtracted waves. When the moving window
regression is performed on background subtracted data (i.e. the “peaks”) we hold A=0 (since the
intercept is presumably made zero by the background subtraction). Indeed, when we run a
windowed fit (5 points wide) with floating A, the value reported for A fluctuates with a large
amplitude. The results achieved by setting A=0 appear to be more consistent. For the city data
segment, a 5-point sliding window regression with A=0, and with the cut:
(CO2- background) > 10 ppm, gives ER 3.68 (± 2.81) x10-3 .
49
In Figure 4.1.8, we give show a data segment that compares the sliding-window
regression method with the point-by-point ratio method. The windowed regression result is
similar to the point ratio result, but smoother and less prone to large excursions. When we apply
the condition that CO2 be greater than 10 ppm (above local background) the large excursions in
ER are further reduced.
Summary of ER methods and results
In calculating the NO/CO2 ER by different methods for the same city segment of data
from 5/25/99, we have generated 11 different values. The average (± standard deviation) for
these values is 3.36 (± 0.19) x10-3. Thus, the different methods are in reasonably good
agreement. We prefer to use the last method described, the sliding window fit (with A=0 and
with a cut on CO2 level). The analysis path for that method is well defined and yields consistent
results. Also, that method allows us to examine the statistics of the ER, via moment analysis
(e.g. standard deviation and skew) as well as generating probability distributions for the ER.
25
20
15
10
100
CO2 peaks
NO peaks
NO/CO2 point ratio
Window ed Regres sion
Window Regr, CO2pks >10
500
400
300
5
NO Peaks, ppb
CO2 Peaks, ppm
150
600
ER, NO/CO2, ppb/ppm
200
200
0
50
100
-5
0 -10
5700
0
5800
5900
6000
File Points
Figure 4.1.8. Example of CO2 and NO data and the emission ratio as a function of time,
determined using point ratios and linear regression in a 5-point sliding window.
50
Boston 5 /25/99
ER NO/CO2 histogram s
All , city, h wy
0.1
Number per Bin
ER_hi st_a ll
ER_hi st_city
ER_hi st_h wy1
ER_hi st_h wy2
fi t_ER_hi st_a ll
0.01
Fit, sl ope= -.32 28
0.001
0.0001
0
5
10
15
ER (window fit), NO/CO2, ppb/ppm
20
25
Figure 4.1.9. Histogram (probability density) of NO/CO2 ER from the sliding-window
regression method.
4.1.5. NO Emission Ratio Results
The most recent and most extensive effort on the development of mobile ER analysis has
concentrated on NO emissions. Therefore, considerable results for NO emissions have been
presented earlier. In this section we consider the dependence of NO emissions on roadway type
(city vs. Highway) and on driving cycle within the city. Before those analyses are presented, we
first present a comparison between the mobile ER and ER data obtained in a cross-road remote
sensing experiment conducted by Aerodyne Research in California in 1996 [Jinenez et al., 1999].
In Figure 4.1.10, we show the distributions of ER’s obtained from the two experiments. The
agreement is quite good, with both distributions showing a similar exponential form and
exponential decay constant.
51
Emission Ratio: City vs Highway
We observed significant variations in NO/CO2 ER in different segments of highway and
city driving, as summarized in the map shown in Figure 4.1.11. On 5/25/99, the city driving
segment ER was 3.68 (± 2.81) x10-3. The highway driving segment toward Boston (Route 3
south to route 128 north to I-93 south) had an ER of 6.22 (± 2.91) x10-3. The return highway
segment (I-93 north, I-90 west, Route 128 north, Route 3 north) had a lower ER of 3.56 (± 2.19)
x10-3. The return highway segment traffic was generally slower than the first highway segment.
A particularly high ER [7.86 (± 2.11) x10-3] was observed on the first highway segment, on
Route 128 north where the speed was high and there was a long rising grade.
Bos ton 5/25 /99
ER NO/CO2 his togra ms
Mobi le: All & City
& CA Cro ss-Rd
Number per Bin
0.1
Mobi le, All
Fi t exp slop e= -.323
0.01
Cross Road, CA
Fi t exp slop e= -.258
0.001
Mob ile_ all
Exp fit_ Mo bile _all
Mob ile_ city
Cros s Road
Exp fit Cro ss Roa d
0.0001
0
5
10
15
20
25
ER, NO/CO2, ppb/ppm
Figure 4.1.10. Comparison of NO ER distributions from mobile sampling in Boston on 5/25/99
and from a cross-road remote sensing expermient conducted in California in 1996.
52
ARI
Rte 3
3
25x10
20
I-93
Northerly Distance, m
I-95 / Rte 128
15
10
I-90
5
0
Bos ton, 5/2 5/99
NO/CO2 Avg. ER Ma p
(w/ m oving wi ndow fit)
20 0 m cel ls
Blu e=0, Red=1 0 pp b/pp m
-10x10
3
Boston
-5
Easterly Distance, m
0
5
Figure 4.1.11. Measurement route on 5/25/99 color coded by NO/CO2 ER.
Driving cycle analysis:
Mobile data may be analyzed in terms of the variables of sampling speed and
acceleration, derived from the GPS record. Those variables apply to the mobile laboratory, and
not to the emitting vehicles. However, the mobile lab generally moves along with the other
traffic and experiences approximately the same dynamics, with some delay. In a simple course
of analysis, we can examine the CO2 concentration and NO/CO2 emission ratio as a function of
speed and acceleration. As shown in Figures 4.1.12 and 4.1.13, we observe higher ER’s at
higher speed and higher acceleration.
53
6.5
Boston city traffic, 5/25/99
Emis sion Ratio, NO/CO2
NO/CO2, ppb/ppm
6.0
5.5
5.0
4.5
4.0
3.5
3.0
0
2
4
6
8
Speed, m/s
10
12
14
Figure 4.1.12. Histogram of NO/CO2 emission ratio as a function of sampling speed.
Avg ER, NO/CO2, ppb/ppm
5.5
Boston city traffic , 5/25/99
Emiss ion Ratio, NO/CO2
vs Acceleration
5.0
4.5
4.0
3.5
3.0
-1.0
-0.5
0.0
Acceleration, m/s2
0.5
1.0
Figure 4.1.13. Histogram of NO/CO2 emission ratio as a function of sampling acceleration.
In a somewhat more complex analysis, we can examine the data as a function of both
speed (S) and acceleration (A), plotting the sampling data on the S-A plane. This way of plotting
the ER links the data to the driving cycle of stop and go city traffic. The driving cycle is
exemplified by a simple model of stop and go driving, with the speed as a function of time
proportional to a raised sinusoid, i.e. V(t) = Vo (1+ sin(2 t / ) ) , where the period of one cycle
is . If such motion is plotted in the speed-acceleration plane (Figure 4.1.14), a circle results.
54
Figure 4.1.14. Simple model of stop and go traffic, with (raised) sinusoidal speed and circular
motion in the speed-acceleration plane.
A segment of city driving data from 5/25/99 in Boston shows how the simple model
compares to real data. In Figure 4.1.15 we plot speed and acceleration, with the speed curve
color and width showing CO2 concentration. When a segment of city driving data is plotted on
the speed-acceleration plane (Figure 4.1.16), we see generally circular motion, within a speed
range of 0-15 m/s and acceleration range ±1 m/s2. With the curve color and width indicating
CO2 concentration, we see that most peaks in concentration extend over a fraction of a cycle.
That behavior is more apparent when we just plot the higher concentration points, as shown in
Figure 4.1.17.
20
Boston city traffic, 5/25/99
Speed trace color & s ize as CO2
Blue=380, Red=480 ppm
Speed_city
ac cel_city
1.5
1.0
Speed, m/s
0.5
0.0
10
-0.5
Acceleration, m/s2
15
5
-1.0
0
-1.5
14.0
14.2
14.4
Point number (sec onds )
14.6
14.8x10
3
Figure 4.1.15. City driving data segment, speed and acceleration, with the speed curve color and
width showing CO2 concentration.
55
1.5
Bos ton city traffic, 5/2 5/99
tra ce col or & s ize as CO2
Blu e=38 0, Red =480 ppm
city se gmen t pts 11 900-1 4875
Acceleration, m/s2
1.0
0.5
0.0
-0.5
-1.0
0
5
10
15
20
Speed, m/s
Figure 4.1.16. A segment of city driving data, with the trace color and size indicating CO2
concentration.
Next, we compute average CO2 concentrations as well as NO/CO2 emission ratio for bins
in the speed-acceleration plane. The results are plotted in Figures 4.1.18 and 4.1.19, with the
additional condition applied that more than 20 data points must be in each bin that is plotted.
The CO2 and ER patterns are subtle. More CO2 is seen in the sector at low speed and negative
acceleration, when approaching stopped traffic. Higher ER’s are seen at higher speeds and
higher accelerations.
The data can be further abstracted by averaging concentrations and ER’s as a function of
“phase angle” in the driving cycle. The phase angle is computed with respect to a center point in
the speed acceleration plane, at speed ~6 m/s, the median speed (excluding zero speed points)
and at zero acceleration. Such averaging (Figures 4.1.20 and 4.1.21) shows a clear pattern in
CO2 concentration and ER.
56
Boston 5 /25/99
City Tra ffic
Peaks Onl y, CO2 >420
Color & Size a s CO2
420 -520 ppm
Truck Acceleration, m/s2
0.5
0.0
-0.5
-1.0
0
2
4
6
8
10
12
14
Truck Speed, m/s
Figure 4.1.17: A segment of city driving data, with the trace color and size indicating CO2
concentration. For this plot, only the sections of the data corresponding to CO2 peaks are shown,
above concentrations of 420 ppm.
57
Boston 5 /25/99, city
CO2 avg vs S & A
B=38 0, R=4 10
onl y w/ >20 pts /ce ll
1.0
Acceleration, m/s2
0.5
0.0
-0.5
-1.0
0
2
4
6
8
Speed, m/s
10
12
14
Figure 4.1.18. Average CO2 concentration as a function of speed and acceleration of the mobile
lab, for city driving in Boston on 5/25/99. The concentration is color coded, from
blue=380 ppm to red=410 ppm.
58
1.0
Boston, city, 5/2 5/99
Avg NO/CO2 vs V & A
onl y w/ >20 pts per cel l
Blue =1, Re d=4 ppb/ppm
Acceleration, m/s2
0.5
0.0
-0.5
-1.0
2
4
6
8
Speed, m/s
10
12
14
Figure 4.1.19. Average ER (NO/CO2) as a function of speed and acceleration of the mobile lab,
for city driving in Boston on 5/25/99. The ER is color coded, from blue=1 x10-3
to red=4 x10-3.
N2O Emission Results
We present in this section results from a mobile measurement campaign in Manchester,
New Hampshire in 1998. The results summarized here have been published in a special issue of
Chemosphere on N2O emissions [Jiménez et al., 2000]. The results also have been cited in the
latest IPCC report on climate change [IPCC, 2001].
Nitrous oxide (N2O) is both a powerful greenhouse gas and the major precursor for
nitrogen oxides (NOx) in the stratosphere. Significant uncertainties remain in the atmospheric
budget of N2O, particularly in identifying sources to balance its stratospheric photolysis and
photochemical sinks [Cicerone, 1989; Khalil et al., 1992; NRC, 1993; IPCC, 1996]. Most
atmospheric N2O is believed to be produced by microbial action in soil, fresh water and marine
environments. The growing atmospheric burden of N2O is most likely due to the intensification
of agriculture, which deposits increased burdens of both synthetic and organic fixed nitrogen into
the biosphere [Kroeze et al., 1999].
59
1.5
Bos ton, ci ty, 5/25/9 9
Avg CO2 vs Drive Cycle Phas e
B=3 90, R=405 ppm
Acceleration, m/s2
1.0
0.5
0.0
-0.5
-1.0
0
4
8
Speed, m/s
12
Figure 4.1.20 Average CO2 concentration as a function of driving cycle phase for the mobile
lab, for city driving in Boston on 5/25/99. The concentration is color coded, from
blue=390 ppm to red=405 ppm.
Motor vehicle exhaust emissions are a second, non-agricultural, anthropogenic source of
N2O which is suspected to be increasing steadily. Nitrous oxide is known to be produced as a
byproduct of nitric oxide (NO) reduction and carbon monoxide/unburned hydrocarbon (CO/HC)
oxidation on noble metal three-way catalysts utilized to reduce pollutants in motor vehicle
exhaust emissions [Cant et al., 1998]. Attempts to quantify fleet emissions of N2O from motor
vehicle exhausts have faced difficulty because N2O emissions are dependent on driving cycle
variables, catalyst composition, catalyst age, catalyst exposure to variable levels of sulfur
compounds and other poisons in the exhaust, and to the fraction of the fleet equipped with
catalytic converters. Thus, measurements on small numbers of selected vehicles may not
represent fleet averages [e.g. Dasch, 1992], and fleet averages obtained from tunnel studies have
yielded disparate results [Berges et al., 1993; Sjödin et al., 1995, 1997; Becker et al., 1999;
Becker et al., 2000 ]. A recent driving-cycle study concluded that fleet emissions had been over
estimated [Michaels, 1998].
N2O Emissions Data Analysis
Two primary methods were used to derive emission ratios for N2O, linear regression and
point ratios. At the time of the N2O analysis, the sliding window regression method had not been
developed. Also, we separated the data into peaks and background components, but with the
earlier manual method.
60
1.5
Bos ton, ci ty, 5/25/9 9
Avg NO/CO2 vs Dri ve Cycle Phase
B=3 .5, R=5.5 ppb/ppm
Acceleration, m/s2
1.0
0.5
0.0
-0.5
-1.0
0
4
8
Speed, m/s
12
Figure 4.1.21. Average ER (NO/CO2) as a function of driving cycle phase for the mobile lab, for
city driving in Boston on 5/25/99. The ER is color coded, from blue=3.5 to
red=5.5 x10-3.
A linear regression fit to the peak component of the data gives an emission ratio of
(10.9±0.1) x10-5 (r2=0.36). A scatterplot of the data segment is shown in Figure 4.1.22. As
discussed earlier, the wide scatter in the points and the low regression coefficient is due to the
variation in the emission ratio. For shorter data segments the points fall closer to a line and the
regression coefficients are greater. A linear regression for the slowly varying trend line
(background) data shows a smaller slope, for a background N2O/CO2 ratio of (4.41±0.09) x10-5
(r2=0.15).
If each pair of N2O and CO2 data points is ratioed, we generate a set of ~104 emission
ratio samples, which then can be used to form distributions as well as averages. Selections and
conditions can be applied easily to the set of pointwise ratios, in order to find the best data subset
for emission index determination and to test dependencies. The first selection is to consider the
highway plus city roadway data set used in the scatterplot above [Figure 4.1.21]. Next, we select
data with CO2 greater than a minimum level (above the subtracted background). The minimum
CO2 elevation selection excludes data close to the background, which can be negative for either
CO2 or N2O. Thus, we reduce contamination of the ratio distribution with negative or
spuriously large values. We empirically set the CO2 cutoff to be 15 ppm (close to the average
for the whole “peaks” set), which eliminates 52% of the data points, with 5538 remaining
(from the 6/16/98 measurement). The cutoff value is selected by observing the effect of
increasing cutoff on the ratio distribution. There is a large change going from zero cutoff to 10
ppm, but quite small changes from 10 to 20 ppm.
61
12
Manchester, 6/16/98
Peak data, Hwy+City
N2 O peak, ppm x 10
-3
10
Regr. slope 10.9x10
-5
8
6
4
2
0
0
10
20
30
CO2 peak, ppm
40
50
60
Figure 4.1.22. Scatterplot of N2O vs. CO2 for city and highway driving in Manchester, New
Hampshire on 6/16/98, with the solid line showing the linear regression fit..
The selected normalized distribution of emission ratios (determined on a point by point
basis) from mobile measurements is shown in Figure 4.1.23. From this distribution, we derive
our reported average emission ratio of (12.8 ± 0.3) x10-5. The uncertainty in the mean is
primarily systematic, estimated from the variation in the mean with changes in CO2 cutoff. The
distribution has a similar shape as that observed in cross road measurements. The distribution is
skewed, with a peak at low values and an exponentially decreasing tail. The width of the
distribution in terms of standard deviation is ~10 x 10-5.
N2O emission dependencies
The data can be segregated to test the effect of potential controlling variables. For
example, if we divide the data contained in Figure 4.1.23 into highway vs city roads, then we
observe a difference in the emission ratio: (10.9 ± 0.3) x 10-5 for the highway versus (15.6 ± 0.3)
x 10-5 for the city roads. A separation of the data into two groups with speed less than or greater
than 16 m/s (36 mph) shows a similar difference in the emission ratio, reflecting the traffic speed
difference between city and highway roads. Thus, the grouping of data by city-highway or slowfast represents the same classification, with a robust difference in the N2O emission ratios. This
difference is probably due to the smaller fraction of vehicles in cold start on the highway, and
62
2
Normalized Histograms
Mobile Data
Cross Road Data
0.1
P(x) dx
6
5
4
3
2
0.01
6
5
4
3
2
0.001
6
5
-0.2
0.0
0.2
0.4
Ratio N 2 O/CO 2 , X10
-3
0.6
0.8
Figure 4.1.23. Histogram of pointwise ratios for mobile peak data (dotted line), N2O/CO2 ,
city and highway, and CO2 > 15 ppm. Also shown (solid line) is a histogram of
N2O emission ratios determined by cross-road sampling.
also to the higher catalyst temperatures at higher speeds, both of which result in lower N2O
emissions [Rabl et al., 1997; Odaka et al., 1998]. We tested other grouping methods, (positive vs
negative acceleration, positive vs negative vertical climb rate) and found much smaller
differences in emission ratios.
We have considered the question of possible dependence of our emission index
distribution on the method of grouping data. The data might be grouped into larger blocks, each
of which contains one (or a few) peaks. The set of minimum points used to identify the local
background provides a convenient way to segment the data into relatively small blocks. Some of
these segments are individual peaks, and some contain clusters of peaks. We calculated
regression slopes for each of the ~250 peak segments contained in the scatterplot [Figure 4.1.22].
The peaks have an average width in time of 28 ± 32 seconds, covering an average distance of
460 ± 680 m, so we expect that many vehicles contribute to each peak. Each sample point may
contain contributions from more than one vehicle, effectively averaging the measurement to
some degree. Averaging will tend to decrease the extremes of the distribution, at both high and
low values.
63
4.1.6. CO and CH4 Emission Results
During a field campaign in Manchester, New Hampshire in June of 1998 we had the
opportunity to measure emission ratios for carbon monoxide (CO) and methane (CH4). These
measurements were rather early in this research program, and the analysis methods applied were
relatively simple. We separated the data into peaks and local background components and
performed regression analyses on segments of peak data. The primary segmentation of the data
was between highway and city driving. Regression slopes for data taken on 6/16/98 are
presented in the table below, in units of ppbv/ppmv, or molecular ratio x10-3.
Data
Segment
All
CO vs CO2
Regr. Slope
Regr. Coeff.
r2
CH4 vs. CO2
Regr. Slope
Regr. Coeff.
r2
24.49
0.62
1.50 ± 0.44 (average of 7 segments)
38.02
0.70
0.990
0.58
0.54
0.924
0.58
0.65
1.80
0.82
25.18
0.47
2.31
0.68
City 1
23.28
0.71
1.48
0.75
City 2
29.89
0.74
1.38
0.48
City 4
46.92
0.78
-0.34
-0.061
City 5
41.74
0.67
1.60
0.32
Highway 1
(Rt. 3 N; I-495 N-E; I-93 N)
Highway 2
36.54
(Manchester I-93 I-293 loop)
Highway 3
16.92
(Manchester I-93 I-293 loop)
Highway 4
(Rt. 3 S)
4.1.7. Discussion
Summary of Observed Emission Ratios
The emission ratios for NO, N2O, CO, CH4 that we derived from mobile measurements
are summarized in the table below. In the table below, the ER’s have been derived by different
methods, which reflect our technique development over the length of the project. As we
discussed in Section 4.1.4, the various ER methods give similar average values.
64
Gas Data Source
Data Grouping
ER
ER Method
-------------------------------------------------------------------------------------------------NO
Boston, 5/25/99
City Roads
3.68 ± 2.8 x10-3
windowed regression
-3
Highway, 1
6.22 ± 2.9 x 10
“
Highway, 2
3.56 ± 2.2 x 10-3
“
N2O
Manchester, 6/18/98 City + H’wy
Highway
City
12.8 ± 0.3 x10-5
10.9 ± 0.3 x10-5
15.6 ± 0.3 x10-5
point ratio with CO2 cut
“
“
CO
Manchester, 6/18/98 City + H’wy
Highway
City
24.5 x10-3
29.2 ± 6.6 x10-3
35.5 ± 9.4 x10-3
linear regression
“ avg. of 4 segments
“ avg. of 4 segments
CH4
Manchester, 6/18/98 City + H’wy
Highway
City
1.50 ± 0.44 x10-3
1.51 ± 0.58 x10-3
1.49 ± 0.09 x10-3
linear regression, 7 segments
“ avg. of 4 segments
“ avg. of 3 segments
We also collected mobile data for NO2, but that data is not presented in the table. The
levels for NO2 were approximately one-tenth of NO levels. However, since we have not clearly
separated promptly emitted NO2 from that which is produced by O3 oxidation of NO, we do not
present an ER for NO2. A possible method to identify directly emitted NO2 is to consider the
photochemical model (Section 4.9) that links equilibrium concentrations of NO, NO2, O3 and UV
radiation. Points where NO2 is well in excess of the photochemical equilibrium concentration
should indicate direct emission. However, the question of direct emission of NO2 is better
answered by detailed studies of individual vehicles. In a follow-on study of vehicle emissions
employing a close following measurement strategy, we found that significant direct emissions of
NO2 are primarily due to HDDV’s, where NO2 may constitute 5 to 40 % of NOx [Shorter et al.,
2001].
The emission ratio analysis has been applied to a small fraction of the data collected
during the four field campaigns of the Urban Respiration project. During those field campaigns,
we collected data that could yield ER’s on at least 12 days. We hope to continue analysis of our
field data to generate a more complete picture of mobile emission ratios.
Comparison to other reported emission ratios
The above results may be compared to other work on the emission ratios of motor
vehicles. Numerous studies have shown that pollutant emissions are a strong function of vehicle
age, as newer vehicles reflect improved emissions control technologies and better maintenance
[e.g. Stephens et al., 1997; Sjödin & Andreasson 2000; Harley et al., 2001]. The average
emissions of a group of vehicles thus will depend upon the specific mix of vehicle age and type.
The comparison of our values to others can be expected to yield only approximate agreement.
65
In the technical literature, in line with EPA conventions, emission indices often are
reported in units of “grams/mile” or “grams/kilometer”. Our molecular ratios can be converted
to such units given an average fuel use rate and other physical parameters. For example, if we
use a recently published value, 6.89 g/mi CO at a fuel use rate of 21.5 MPG [Durbin et al.,
2002], the CO emission rate may be expressed as 0.246 moles/mi. The CO2 emission rate is:
( 21.5 MPG)-1 ( 2550 g/gal) (0.842 g carbon / g fuel) (12 g/mole carbon)-1
= 8.32 moles carbon/mi = 8.32 moles CO2/mi
,
using the density and carbon content of octane, and assuming that essentially all of the carbon
appears as CO2. The molar ratio (CO/CO2) then is 0.03.
We have presented a detailed comparison of our N2O results (both mobile measurements
and cross-road remote sensing) in an earlier publication [Jiménez et al., 2000]. We found that
the aggregate emission ratio for N2O was lower than many previous reports. Several
representative literature reports of other pollutant emission rates are summarized in the table
below.
Reported
ER,
Reference
Pollutant
Emission Rate
molar ratio to CO2
----------------------------------------------------------------------------------------------------------------CO
6.35 g/mi
0.036
Pierson et al., 1996
NO
~2 † - 0.2 # vol% exhaust
0.15 † - 0.015
~ 2 † - 0.15 #vol% exhaust
0.15 † - 0.01
63 g/L (fleet average)
0.047
Harley et al., 2001
6.89 g/mi
0.030
Durbin et al., 2002
0.72 g/mi
4.0 x10-3
Pierson et al., 1996
~0.13 † - 0.02 #vol% exhaust
CH4
#
#
0.01 † - 0.0001
Stephens et al., 1997
Sjödin & Andresson, 2000
#
Sjödin & Andresson, 2000
8.7±2.6 g/L (as NOx)
0.0061
Harley et al., 2001
0.573 g/mi
0.0023
Durbin et al., 2002
0.002 ± 0.0005
Heeb et al., 2001
13.8 ±2 mg/km, FTP Phase 3
† Older cars; # Newest cars
Peak Counting and Statistical Significance
We believe that we sampled a large number of vehicle exhaust plumes in our city wide
surveys, where we typically would drive in traffic for several hours, often under congested
conditions. However, the number of vehicles sampled and therefore the statistical significance
of the reported averages should be addressed quantitatively. The question of the statistical
significance of the data sets that we have analyzed may be addressed by counting the “peaks” in
the data. Each significant CO2 concentration peak can be viewed as a plume encounter or
66
“event”. As we discussed earlier, the stronger peaks probably represent individual vehicles,
while the smaller peaks may represent mixed plumes from several vehicles. However, it is not
clear from the data how many times we encounter the plume from a single vehicle. Counting the
peaks may give us only an order of magnitude estimate of the number of sampled vehicles.
In attempting to present a count of the peaks in a typical data set, we see that the number
depends upon how we define a “peak” and how the count is performed. The simplest definition
of a peak is as a local maximum. However, some local maxima are due to noise, and some are
due to fine structure within a broader Gaussian-like shape. Spurious fine structure local maxima
may be suppressed by smoothing the data. Low level peaks may be ignored with a concentration
cut. Another way to count peaks is to enumerate the (positive) crossings of a given
concentration level. By either method, the number of peaks depends upon both the degree of
smoothing and the concentration cut level.
Peak counting methods were tested with the CO2 data collected 5/25/99 in Boston, a data
record of 14,500 points (~4 hours at 1 Hz). We use the background-subtracted data set dervied
from the “range-minimum” procedure. The number of local maxima is 2059, of which 1409 are
above 10 ppm. When we smooth the data to a 4 second response (convolve with a Gaussian with
half width 4 seconds), the number of local maxima is 845, of which 693 are above 10 ppm.
When we count the number of crossings of 10 ppm level, we find 983 peaks in the raw data and
600 in the smoothed data. Thus, we count between 600 and 1000 peaks in this data set.
On the interpretation of ER probability distributions
A unique product of our mobile measurements is molar emission ratio data with high
spatial and temporal resolution, with statistical distributions instead of just averages. This type
of data relates to studies of the dispersion of pollutants in the urban roadway [e.g. Clifford et al.,
1997; Chan et al., 2001]. The interpretation of this new type of data leads to several questions.
First, one can ask how representative the measurements are of fresh tailpipe emissions versus
mixed air with possibly altered molar ratios. We have attempted to reduce the potential
influence of mixing by subtracting variable background levels and by limiting our attention to
data where the concentration is significantly above background.
The derived distributions of ER’s tend to have a similar shape, with a Gaussian-like
bump at low ER, and then an (approximately) exponentially falling tail at high ER’s. That shape
of curve raises a concern that mixing still is influencing the ratio statistics, since the
concentration statistics for individual gases have similar types of distributions. A typical
example of this type of concentration probability distribution is shown in Figure 4.1.24, for NO
and CO2 in Boston on 5/25/99. The histograms have overlaid fits, which are combinations of
Gaussians and exponential tails.
The type of concentration probability distribution shown in Figure 4.1.23 can be
explained in terms of mixing, as shown by a set of calculations using a simple Gaussian plume
model. In our model, we construct a set of sources each producing a Gaussian plume, and then
compute the probability distribution for downwind concentration sampling. We find that this
model robustly produces probability distributions with exponential tails, even when all the
67
source strengths are equal. An example of a model probability distribution is shown in
Figure 4.1.25, for a set of 20 sources, each emitting two different gases at a ratio of between 1
and 2. The sources are at random locations upwind of the sampling volume. The resulting
concentration distributions are very similar in form to those shown in Figure 4.1.24. The same
plume modeling that we used to generate Figure 4.1.25 indicated that distortions to the ER
probability distribution by mixing can be avoided if we do not include data with low CO2
concentration.
In balance, we do not believe that mixing effects dominate the ER distributions that result
from the moving window regression analysis. Empirical evidence for this conclusion is provided
by two observations. The distribution shapes remain stable as we increase the threshold
concentration of CO2 (above local background), for ∆CO2 greater than ~10 ppm. The mobile ER
distributions are similar to the cross-road ER distributions.
400
CO2 conc., ppm
450
500
550
Boston 5/25/99
City segment
NO bins 2 ppb
CO2 bins 2 ppm
0.1
P(C) dC
NO_ci ty_his t
CO2_city_hi st
CO2_ch_fi t
NO_ch _fit
0.01
0.001
0.0001
0
100
200
NO conc, ppb
300
400
Figure 4.1.24. Concentration probability distributions for NO and CO2 in Boston on 5/25/99.
The histograms have overlaid fits which are combinations of Gaussians and
exponential tails.
68
2
Comp uted Poll utan t
Concentratio n Prob abil ity
20 sou rces @ ran dom
lo cations , -13 <z<-3
P/C=1 or 2 , ran dom
0.1
8
7
6
5
Ph ist_ sc1
Ph ist_ sp1
4
P(C) dC
3
2
0.01
8
7
6
5
4
3
2
0.001
0.0
0.1
0.2
0.3
Concentration, arb units
0.4
0.5
Figure 4.1.25. Computed probability distribution for a Gaussian plume model, with a set of
20 sources, each emitting two different gases at a ratio of between 1 and 2. The
sources are at random locations upwind of the sampling volume.
Variable instantaneous pollutant emission rates also complicates the interpretation of
mobile sampled ER’s. Sampled ER’s in urban environments will reflect a combination of the
average emission characteristics of vehicles as well as high emission events for vehicles with low
average emissions. Recent studies have shown that pollutant emissions can vary strongly
through the driving cycle, with brief peaks in emissions of CO, NOx and HC [e.g. de Haan &
Keller, 2000]. A knowledge of the distribution of average vehicle emissions may be insufficient
to predict the observed distribution of urban ER’s. Despite that caution, the one available
detailed comparison between ER’s determined by mobile surveys and cross road remote sensing
(which we conducted, [Jiménez et al., 1999; 2000]), for NO and N2O shows good agreement
between the ER distributions, as shown in Figures 4.1.10 and 4.1.23.
69
The existence of brief high-emission events within the driving cycle would explain why
the autocorrelation for NO is narrower than for CO2, and why the number of NO peaks is greater.
Brief emission peaks within the driving cycle could give pollutant enriched local structure to the
CO2 plume from an individual vehicle.
Evaluation of the technique of mobile ER determination
The technique of assessing aggregate emission ratio by surveying concentrations of CO2
and pollutants in a mobile laboratory with fast response high sensitivity instruments can yield
average fleet emission factors and related statistics. The dependence of the ER on roadway type
and driving conditions can be evaluated. The observed ER’s can be compared to model
predictions based on fleet compositions and standardized emission factors. However, comparing
against detailed models of motor vehicle emissions needs inputs on the type of vehicle, age, and
detailed operating conditions. The techniques reported in this project do not supply that detailed
information on individual vehicles. We cannot define the composition of the vehicle fleet that
we sampled in the work reported here.
Our techniques of extracting ER’s from the mobile data has evolved throughout the
project. The latest method has a well defined procedure that appears to yield robust results, i.e.
the results do not depend significantly on the detailed settings used in the procedure. The latest
procedure contains the following steps:
1] Background subtraction: with background defined as the minimum value (greater than zero)
observed within a range (backward or forward) in the traverse, generally using a range of
±500 m. The background level thus derived is smoothed and then subtracted from the data
record to give the “peaks” segment.
2] Windowed linear regression: The peaks segment is fit to a linear form with the intercept
forced to zero (pollutant = A*CO2) over a moving square window of modest width, usually ~5
points. The slope (A) is the instantaneous molar emission ratio.
3] Cut on minimum CO2: We ignore data with CO2 (peak segment) less than some minimum,
usually 10 ppm.
The technique of evaluating aggregate emissions by mobile surveys could be improved in
several ways. First, a reliable method of measuring wind direction on board the mobile
laboratory could improve the reliability of gauging tail-wind contamination as well as identifying
the direction of the detected sources. An on-board wind sensing instrument must be set far
enough forward of the laboratory that the perturbed wind-flow around the laboratory does not
unduly influence the local wind reading. A video camera recording the vehicles in front of the
mobile lab would help to define the traffic conditions and the mix of vehicles being sampled.
Just such a video camera has been employed in follow-on work [Shorter et al., 2001].
70
4.2 Background Pollutant Maps
In spatial displays of pollutant concentrations along measurement traverses, the picture
typically is dominated by the many pollutant peaks due to local (on-road) sources. An
underlaying pattern of pollutant concentration may be better represented by the low levels
between the peaks. To observe general patterns of pollution from roadway traverses is a problem
analogous to seeing the forest floor between the trees. We have expected that lower
concentration points should be more typical of the well mixed urban air in the absence of local
sources. It has been an area of our investigation to determine the degree of correspondence
between local minima in concentration and the urban background.
The concept of an urban background pollutant level is based on measuring far from local
sources, so that the air within the boundary layer is well mixed with average concentrations that
vary slowly. Then the pollutant concentration typically would exhibit Gaussian statistics, with a
relatively constant standard deviation and negligible higher order statistical moments, such as
skew. Such a statistical description is quite distinct from much of our urban roadway pollutant
data, which generally is highly skewed. The background in the urban environment is somewhat
ill defined, since the pollutant concentrations can be expected to fall with greater distance from
sources or source complexes. The minimum concentrations encountered in an urban street
canyon may be relatively constant, while elevated from a nearby quiet neighborhood street or a
city park. Thus, the concept of background is connected to scale and structure within the city.
We have explored two methods of presenting urban background pollution concentrations.
The first is to consider the probability distribution of concentrations, which typically have a low
level Gaussian-like segment and a higher level exponential tail. We can select and plot the
points in the low level Gaussian regime to visualize spatial patterns. An example of this process
is shown in Figure 4.2.1, for CO2 concentrations in Boston on 5/25/99. We select the points that
are below the concentration corresponding to the maximum probability [Figure 4.1.24], where
the probability has a Gaussian form. The map shows little structure, except the segment in
Franklin Park with significantly lower concentrations (blue points) than the low points on
congested city roads.
The second method we have explored for presenting urban background pollution
concentrations is to plot the interpolated minima derived from the peak-background separation
process described in Section 4.1. The “Rangemin” wave is derived by replacing each point in
the data wave with the minimum concentration encountered within a certain linear distance of
that point. The resulting wave also is smoothed to remove transition effects. While the primary
purpose of the “Rangemin” procedure is to produce a background-subtracted peak segment for
emissions ratio analysis, the background wave may reveal spatial patterns. When the minimum
increases in a given location, the selection of lowest points (Figure 4.2.1) may leave that section
of the map blank, while the “Rangemin” map would show the higher minimums. However, we
regard such plots with some caution, since extrapolated and interpolated values are somewhat
removed from the actual data.
71
We show two examples of Rangemin maps in Figures 4.2.2 and 4.2.3, for CO2 and NO in
Boston on 5/25/99. From these maps we see that the minimum levels were quite low, near 380
ppm for CO2 and 0 to 5 ppb for NO. The significance of the patterns is not clear. In these plots,
the width of the trace decreases with distance traveled, to reveal different minima values derived
at different times. The minima generally are not constant between different traverses on the
same road. The relatively high level of NO seen on the northerly road (red in Figure 4.2.3) is not
observed in a subsequent traverse.
382 380 378 376
CO2 (ppm)
42.33
42.32
Lat
42.31
42.30
Boston 5/25/99
Lowest CO2 points
42.29
42.28
-71.10
-71.08
-71.06
-71.04
Long
Figure 4.2.1. Points where CO2 concentrations are below the maximum probability
concentration, in Boston on 5/25/99.
72
6000
400
390
380
CO2 (ppm)
Northerly distance, m
5000
4000
3000
2000
Boston 5/25/99
Interpolated Min-CO2
in 500 m windows
1000
0
0
1000 2000 3000 4000
Easterly distance, m
5000
Figure 4.2.2. Interpolated minima in CO2 over a range of +/-250 meters during traverses on
5/25/99 in Boston.
6000
5
4
3 2 1
NO (ppb)
0
Northerly distance, m
5000
4000
3000
2000
Boston 5/25/99
Interpolated Min-NO
in 500 m windows
1000
0
0
1000 2000 3000 4000
Easterly distance, m
73
5000
Figure 4.2.3. Interpolated minima in NO over a range of +/-250 meters during traverses on
5/25/99 in Boston.
74
4.3 Fixed Site Pollutant Measurement Analysis
Manchester, NH
In August 1998 we conducted an intensive measurement campaign in Manchester,
New Hampshire. During this campaign we had the opportunity to collect stationary data while
positioned next to an EPA monitoring station. The monitoring station was located in the
municipal parking lot at Harnett Park, located in the center of Manchester. The ARI van was
parked approximately 10-20 m from the monitoring station. This was the sole EPA station in
Manchester in 1998 that was measuring NO2 or O3. The site has since this time been replaced by
one in a different location. A time series of NO, NO2, CO2, and O3 mixing ratios from a
stationary period at Harnett Park on August 26, 1998 (21:33-22:05 UTC, or 5:33-6:05 pm EDT)
is presented in Figure 4.3.1, along with short periods before and after the stationary data during
which mobile data was collected on the surrounding roads. There were clearly lower levels of
NO, NO2 and CO2 at the station than on the streets in the flow of traffic just prior to and
following the stationary measurements. There were also fewer fluctuations in these gases at the
monitoring site. During the data collection at the lot winds were from the southwest at
approximately 2 m/sec. For this 32 minute segment, the average trace gas mixing ratios were:
6.8611.39 ppb NO; 3693 ppm CO2; and 4.721.75 ppb NO2.
60
CO2 (ppm)
40
30
480
440
400
NO2 (ppb)
500
400
300
200
100
0
*
25
20
15
10
5
0
9:30 PM
8/26/1998
9:45 PM
10:00 PM
NO (ppb)
*
Harnett Park
O3 (ppb)
50
10:15 PM
Time (UTC)
Figure 4.3.1. Measurement of trace gas species at Harnett Park, Manchester, NH, site of an
EPA monitoring station, and surrounding area on August 26, 1998. The data at
Harnett Park are bracketed by the two * in the figure.
75
Cambridge, MA
The instrumented ARI van, instrumented WSU van, and sodar were positioned at the
fixed site at MIT on May 27-29, 1999. A goal of these measurements was to analyze the
temporal variability of pollutants, without influence of spatial variability. The study provided a
data set for comparison of fast response (1 Hz) data with slow response data over an extended
period. Slow response data was simulated by averaging the fast response data over 1 hour
periods, the typical reporting period at fixed monitoring stations.
We collected approximately 48 hours of data, from ~19:00 EDT on Thursday May
27,1999 until ~15:00 EDT on Saturday May 29. In Figure 4.3.2 we report results of monitoring
NO, NO2, O3 and CO2 with the ARI instruments at the site. We show here the averages and
standard deviation of the 1 Hz data, for one-hour segments between 20:15 May 27 and 14:00
May 29 EDT. The standard deviation on the data reflects the numerous peaks in the fast
response data, presumably from vehicles on nearby roads. Late at night there are fewer local
sources, reflected in smaller deviations around the averages. However night-time inversion
conditions led to a buildup of chemically stable gases (e.g. CO2) and consequently higher
averages.
The variability of the pollutant mixing ratios was greater during the day than at night
because there were more local sources during the daytime hours, with the principal sources being
motor vehicles. The wind during the data set was initially from the north, but then switched to
the east/southeast. The latter direction corresponds to the measurement van being downwind of
the Southeast Expressway, a major artery in Boston. There were increases in NOx and CO2
around rush hour in the morning, particularly on Friday morning, followed by a slight drop in the
levels around mid-day. The change in NO2, an indirectly emitted pollutant, was less evident than
it was for CO2 and NO, directly emitted pollutants. The early onset of an increase in the
pollutants in the afternoon of May 27 may be linked to an early rush hour due to the upcoming
Memorial Day holiday weekend.
Much of the character of the variation in the data is suppressed by the averages and
standard deviations in Figure 4.3.2. The probability distributions of the statistical data, shown in
Figure 4.3.3, provides more insight into the characteristics of the pollutants within a specified
time period. We find that the distributions are similar to distributions of mobile data.
Distributions of two primary and one secondary pollutants are shown in Figure 4.3.3 for four
time periods previously averaged in Figure 4.3.2 -- late night, early morning, mid-day, and
afternoon rush hour. The time period of each analysis window in the figure was 1 hour. The
initial maximum in the distribution has a Gaussian shape that reflects the background level of the
pollutant, with an exponential tail resulting from high mixing ratios from local sources. The late
night and early morning curves have less of a tail because of the smaller number of local sources
during those time periods. The mid-day data has a lower background level in this data set due to
the meteorological conditions. There were clear, sunny skies which allowed for much vertical
mixing and therefore, dilution of the pollutants. We also observe that the probability distribution
of the secondary pollutant has little or no tail.
76
420
400
O3 (ppb)
380
50
40
30
20
10
20
0
NO2 (ppb)
40
NO (ppb)
CO2 (ppm)
440
100
0
00:00
06:00
12:00
18:00
00:00
06:00
Time (EDT)
Friday
12:00
Saturday
Figure 4.3.2. One hour averages and standard deviation of each 1 hour data set of the 1 Hz NO,
NO2, and CO2 data from the ARI mobile laboratory. Stationary measurements at
MIT campus in Cambridge, MA.
0.1
4
0.1
2
0.1
6
4
0.01
2
0.01
0.01
6
4
2
0.001
0.001
0.001
6
4
0
50
100
150
200
250
300
10
20
Probability
Figure 4.3.3
0
30
40
NO2 (ppb)
NO (ppb)
50
60
380
400
420
CO2 (ppm)
Probability Distributions of NO, NO2 and CO2 as a function of time of day. ---23:15-00:15 EDT, ---- 05:15-06:15 EDT, ---- 11:45-12:45 EDT, ---- 16:4517:45 EDT, from May 27 until May 28, 1999.
77
440
460
Another way of expressing the variation of a given trace species in the continuous data
set is through a plot of the normalized distributions of that species. In Figure 4.3.4 we display
the normalized distributions of NO, NO2, and CO2 for twelve 1 hour periods ranging from 20:00
EDT, May 27 through 20:00 EDT, May 28. These distributions have corresponding averages
plotted in Figure 4.3.2. The distributions are given with probability density as color, i.e. red
color corresponds to higher probability of occurrence of a given concentration (see color scale in
the figure). Buildup of background level is seen as an increase in the lowest mixing ratio band.
During high traffic periods, the numerous local sources produce a diffuse band toward higher
concentrations. This is especially true for the directly emitted pollutants, NO and CO2, while the
secondary pollutant, NO2, displays this feature to a lesser extent.
NO2, ppb
50
NO 2
40
30
20
0.0
10
-0.5
Log Probability, P(c) dc
0
NO
NO, ppb
300
200
. 100
CO2 , ppm
0
440
-1.0
-1.5
-2.0
-2.5
-3.0
CO2
-3.5
420
400
380
1
22:00
2
3
02:00
4
5
6
7
8
Time Period Number
06:00
09:00
9
14:00
10
11
12
18:00
Mid-Time (EDT)
Figure 4.4.4. Probability Density of NO2, NO and CO2 for 12 1- hour time periods spanning
from 22:00 EDT May 27 until 20:00 EDT May 28.
78
4.4 Mesoscale Wind Field Modeling
The MM5 meteorological model [Dudhia et al. 1994] was applied to simulate the wind field for
the Northern New England coast for periods during three of the field campaigns as follows:
November 10-13, 1997
August 26-31, 1998
May 24-26, 1999
-
Domain centered over Manchester, NH
Domain centered over Manchester, NH
Domain centered over Boston, MA
In all cases a four-nested domain was utilized with horizontal grid cell sizes of 27 km,
9 km, 3 km, and 1km. Thirty-three vertical layers of variable spacing were used for all
simulations and all domains.
A high degree of resolution is desired in order to capture the fine scale flows that are
critical for a correct characterization of an urban landscape. Other important features that play
an active role in New England meteorological conditions are both arctic air excursions from
Canada as well as moist air migration from the south. These characteristics are taken into
account by using a four-nested domain where the outer mesobeta scale grid is 27 km and extends
from Canada down through the Gulf of Mexico and across the Atlantic to a longitude of 50. The
27 km MM5 modeling domain for the Manchester, New Hampshire runs is shown in
Figure 4.4.1 and for the Boston, MA simulation in Figure 4.4.2(a). Inner domains are shown for
the two areas in Figures 4.4.1b and 4.4.2b. High resolution 30 second terrain and landuse data,
originally compiled by the USGS then manipulated by the University of Washington into a
format compatible with the MM5 pre-processors, is utilized in the second, third, and forth
domains.
Surface layer winds from the MM5 simulation 6:00 pm EST on November 11, 1997 for
Manchester, NH are shown in Figure 4.4.3 and illustrate the prevailing northwest winds. Such
winds were ideal for the field campaign allowing for clean background air to enter the region and
ensuring that sampling occurring across the southeastern portion of the city captured locally
generated pollutants.
Back-trajectories are calculated by the read-interpolate-plot (RIP) program developed by
the University of Washington to investigate pollutant source areas for Manchester, NH and
Boston, MA. Figure 4.4.4 (a) and (b) show surface layer winds from MM5 for the 3 km domain
and back trajectories calculated by RIP for Manchester, NH and Boston, MA for the August
1998 field campaign period. The trajectories indicate that similar source areas probably
influence both Boston and Manchester. Figure 4.4.4 (a) depicts a northerly flow and trajectories
for the 10 hr period up to 9 am August 28, 1998 EDT while Figure 4.4.4 (b) depicts a southerly
flow and trajectories for the 9 hr period up to 9 am August 30, 1998 EDT. Time in the figures is
in GMT where 98082600 is the MM5 simulation start time in YYMMDDHH format.
Vertical wind profiles from MM5 extracted from the Boston, MA simulation are
compared in Figure 4.4.5 with NCDC/FSL radiosonde archived wind data from Chatham, MA
on May 25, 1999 at 7 pm EST. These simulations are utilized in the turbulence modeling of
urban landscapes and urban footprint modeling sections described below.
79
Figure 4.4.1 (a) MM5 27 km model domain for Manchester, NH and (b) MM5 3 km and 1 km
domains for Manchester, NH.
Figure 4.4.2
(a) MM5 27 km model domain for Boston, MA and (b) MM5 3 km and 1 km
domains for Boston, MA.
80
Figure 4.4.3. MM5 surface layer winds in the 1 km Manchester, NH domain at 6 pm EST on
11/11/97.
81
(a
)
(b
)
Figure 4.4.4. Back trajectories calculated by RIP for the period ending (a) August 28, 1998 at
9 am EDT and (b) August 30, 1998 at 9 am for Manchester, NH and Boston, MA.
82
2000
1800
1800
1600
1600
1400
1400
1200
1200
Height (m)
Height (m)
2000
1000
1000
800
800
600
600
400
400
200
200
0
0
0
30
60
90
120
150
180
210
240
270
300
330
360
0
WindDirection(deg)
raob
5
10
15
20
25
30
Wind Speed (m/s)
mm5
raob
mm5
Figure 4.4.5. Comparison of MM5 and NCDC/FSL Radiosonde archived (a) wind direction and
(b) wind speed at Chatham, MA on May 25, 1999 at 7 pm EST.
4.5 Turbulence Modeling of Urban Landscapes
4.5.1
Tracer Data Analysis
The SF6 tracer data can be used to understand the nature of turbulence and plume spread in an
urban boundary layer. Advective processes causing plume meander and diffusive processes
causing plume spread determine plume growth. On a nearly instantaneous time scale the plume
is quite narrow, exhibits high concentrations and plume spread is determined by diffusion.
While over longer periods of time, plume meander has a greater influence, yielding a plume with
larger spread and lower average concentrations. The mobile vans were used to obtain horizontal
crosswind concentration profiles of the tracer plume at different downwind distances. Because
each individual traverse of the plume occurred relatively quickly since the plumes were narrow,
the mobile tracer data can be used to determine instantaneous plume diffusion coefficients.
Assuming that a plume spreads in a Gaussian manner and spreads equally in the horizontal and
vertical directions, two methods exist to calculate the instantaneous diffusion coefficient (Yi):
the Centerline method and the Moment method. Plume spread calculated by the centerline
method assumes that the plume centerline is successfully intercepted during the traverse, thus
plume spread is calculated by:
 Yi
 Q
 
 uC cl



1/ 2
(4.5.1)
where Q is the release rate (g/s), u is the mean wind speed (m/s), and Ccl is the plume centerline
concentration (g/m3).
83
The second method, the moment method, involves integrating the plume profile along the
crosswind distance to calculate the instantaneous diffusion coefficient by:
  C ( y  y ) 2 y 
 Yi  


 Cy 
1/ 2
, where y 
 Cyy
 Cy
(4.5.2)
where C is the tracer concentration at the crosswind location y.
The moment method inherently contains greater uncertainty than the centerline method because
during a plume traverse the instantaneous plume is not fixed in space. If the plume meanders
with the traverse vehicle, plume spread may be over-estimated, and conversely, if the plume
meanders against the direction of the traverse vehicle, plume spread may be under-estimated
[Peterson and Lamb, 1995].
This is the first application of instantaneous plume dispersion theory to an urban
boundary layer. Previous applications of this methodology are detailed in Peterson and Lamb
(1995), and Peterson et al. (2001, 1999). Horizontal diffusion coefficients are summarized in
Table 4.5.1 for SF6 tracer data obtained during the August 1998 field study conducted in
downtown Manchester, New Hampshire. Figure 4.5.1 depicts the instantaneous diffusion
coefficients versus downwind distance. The centerline method demonstrates how typically
plume spread increases with downwind distance while the moment method predicts less plume
spread at the greater downwind distances.
Instantaneous Diffusion Coefficient (m)
1600
1400
1200
1000
Centerline Method
Moment Method
800
600
400
200
0
0
1000
2000
3000
4000
5000
6000
Downwind Distance (m)
Figure 4.5.1. Instantaneous diffusion coefficients calculated by the centerline and moment
methods versus downwind distance for tracer tests conducted in Manchester, NH
August 27, 28 and 30, 1998.
84
Table 4.5.1 - Instantaneous diffusion coefficients calculated by the centerline method and
moment method for tracer tests conducted in Manchester, NH August 27, 28 and
30, 1998.
Test
827c
827d
827d
827e
827e
827e
827f
827f
827f
828c
828e
828e
828f
828f
828f
828f
828f
828g
828g
830c
830d
830f
830f
830f
830f
830f
830f
830h
830h
830h
830h
830h
830i
830i
830i
Source-to-Receptor
Centerline Method Yi (m) Moment Method Yi (m)
Distance (m)
178
112
53
508
234
124
908
348
223
3217
616
138
3983
759
196
5211
681
174
4591
1232
349
4511
1011
223
4864
1092
157
400
205
44
510
409
196
1443
421
79
638
304
234
598
414
80
854
325
55
966
454
612
1410
479
256
538
202
17
469
138
12
3273
1398
224
515
116
31
1279
638
187
536
239
67
1189
624
132
605
257
151
1550
319
115
568
186
84
1273
281
203
1047
474
40
638
262
111
2387
840
109
1917
1201
42
556
196
64
370
155
63
537
201
65
85
4.5.2 Turbulence Modeling of an Urban Landscape
Turbulence modeling of an urban landscape involves two components: 1) mesoscale
modeling of the regional wind field, then 2) application of a 3-d turbulence model that simulates
the actual urban landscape. MM5 provides initial and boundary conditions to a fluid-dynamic
3-d code, TEMPEST [Trent et al. 1983], developed by Battelle Memorial Institute’s Pacific
Northwest Laboratory. TEMPEST simulates the complex flows of the urban landscape, such as
street canyon flow, the urban heat island effect, and building re-circulation zones. In TEMPEST
turbulence is modeled by solving the Reynolds averaged Navier-Stokes equations with a k-e
turbulence closure formulation. The focus is upon horizontal scales from 10’s to hundreds of
meters. TEMPEST has been applied to successfully simulate building flows in an arctic
industrial complex [Guenther et al. 1990] and isolated complex terrain features such as Steptoe
Butte in eastern Washington state [Dawson et al. 1991].
As an initial test of TEMPEST’s ability to capture urban flow features, a two-dimensional
idealized urban landscape is simulated. Two building obstacles are placed in the mean flow.
The first building is 10 m long and 10 m high. The second building is also 10 m long and 20 m
in height. An urban canyon of 20 m lies between the two buildings. The evening of
November 11, 1997 is a period when both TILDAS mapping data and city-wide VOC canister
sampling occurred, therefore TEMPEST is initialized with the wind profile shown in
Figure 4.5.2(a) extracted from MM5 output for 6:00 pm on 11/11/97. Turbulent kinetic energy
(TKE), also shown in Figure 4.5.2(b), is estimated from Stull (1995) for stable conditions with a
mixing height of 800 m, and is also input into the model. Section 4.4 describes the MM5 wind
field data used to drive the turbulence model.
1000
Height (m)
Height (m)
a)
100
10
1
0.00
2.00
4.00
6.00
8.00
10.00
b)
0.50
1.00
1.50
2.00
Turbulent Kinetic Energy (m^2/s^2)
Wind Speed (m/s)
Figure 4.5.2.
900
800
700
600
500
400
300
200
100
0
0.00
a) Logarithmic Wind Speed profile generated from MM5 output, and
b) turbulent kinetic energy estimated from Stull (1994), for Manchester, NH,
6:00pm 11/11/97.
86
Variable grid spacing is applied to minimize the grid spacing in the vicinity of the buildings and
maximize the grid spacing away from the buildings. A grid spacing of 2 m is applied to the
buildings, street canyon, and one building length upwind and downwind. Grid spacing is
thereafter doubled until reaching a value of 32 m where the remainder of the domain is covered
by 32 m grids. To indicate building location and grid spacing, Figure 4.5.3 only depicts a portion
of the model domain. The total domain dimensions are 660 m downwind and 804 m high
resulting in a 54 x 46 grid domain. In terms of the 20 m building height (HB), the upwind edge is
10 HB from any buildings, the downwind border is 24 HB, and the vertical domain extended to 40
HB [Guenther et al. 1990].
Figure 4.5.4 depicts a steady state convergent TEMPEST solution. In two dimensions the
model successfully captures the urban canyon recirculation patterns as well as the building leeside recirculation zone.
Figure 4.5.3. TEMPEST Model Domain for a 2-D Idealized Urban Profile.
87
Figure 4.5.4. TEMPEST Solution for a 2-D Idealized Urban Profile.
4.6 Urban Footprint Modeling
Plume diffusion modeling along a forward trajectory maps out the distribution of
pollutant concentrations due to an upwind source. Application of plume diffusion theory along a
back-trajectory yields the upwind source distribution (source-footprint) affecting a receptor at the
trajectory initial point. Thus, our approach is to use available modeling systems to derive a
detailed wind field, and then apply the CALPUFF model in reverse along back trajectories. The
results yield the upwind source distribution of sources affecting a downwind receptor.
Three components comprise this source-footprint modeling strategy: 1) Mesoscale
modeling of the regional wind field, 2) Application of MCIP, the Models-3/CMAQ
meteorological processor, then inversion of the resulting wind field, 3) Application of the
CALPUFF plume dispersion model to the inverted wind field. The MM5 prognostic
meteorological model provides detailed hourly wind fields to the MCIP [Byun et al. 1999] model
utilized here primarily to reformat the MM5 output for input into CALPUFF [Scire et al. 1999].
Finally, the winds are inverted and the CALPUFF model is applied. By applying CALPUFF to
the inverted wind field, the resulting plume trajectory and puff dispersion indicates the upwind
pollutant area affecting a downwind receptor. Correlation of this source-footprint with existing
emissions data yields the fractional contribution of the applicable upwind emissions to a
downwind receptor. The source-footprint is correlated with a gridded 1988 emission inventory
obtained from the Massachusetts Department of Environmental Quality. The meteorology is
from the May 24-26, 1999 MM5 run discussed in section 4.4.
88
The Massachusetts Department of Environmental Quality provided a 1988 5 km gridded
emission inventory suitable for the carbon-bond IV chemical mechanism. Since we are not yet at
the point where we can validate pollutant source strength, these data are used as an indicator of
the spatial distribution of emissions within Boson and its surrounding areas. The emission
inventory data are normalized by the total emission rate and the assumption made that pollutant
source locations and the relative strengths have not changed appreciably in the previous 11 years.
Data for June 21, 1988 (Tuesday) in particular were extracted from the emission inventory to
coincide as closely as possible with the May 25, 1999 (Tuesday) field data. Processing was
required to merge the area and point source emission data, and to re-grid the 5 km data to the
3 km MM5 grid. NOx (NO + NO2) was chosen as the pollutant of interest. Figure 4.6.1
illustrates the normalized distribution of pollutants across the domain from the gridded emission
inventory for point and area sources used to represent May 25, 1999 at 5 pm EST.
CALPUFF was applied to simulate the upwind source probability distribution for a
receptor by application of plume dispersion theory along a back trajectory. South Boston was
chosen as the receptor of interest because of the extensive mapping data available from the field
study. Figure 4.6.2 depicts a backward plume originating from South Boston May 25, 1999 at
5 PM EST. This backward plume depicts the probability source distribution for Boston for
pollutants undergoing advective and dispersive processes during the previous 5 hours. The
winds were steady from the southwest thus causing a narrow plume and indicating that pollutants
could have traveled from Connecticut and Rhode Island during the previous 5 hour period.
Figure 4.6.1. NOx (a) Point and (b) Area emission inventory data for New England applied to
May 25, 1999 at 12 pm EST.
89
Figure 4.6.2. Upwind source area influencing Boston, MA at 5 pm EST May 25, 1999.
As a proof of concept, the upwind source contribution area can be obtained by running
CALPUFF in the forwards mode (i.e. with the regular, non-inverted wind field), with every grid
acting as a source, and tracking each source’s contribution to the total concentration at Boston.
Each source’s contribution is then normalized by the total concentration at Boston. This should
yield a similar result as the source-footprint.
To achieve this source contribution calculation, CALPUFF was run 3600 (80 columns x
45 rows) times with each grid point acting as a source and the results are depicted in
Figure 4.6.3. Figures 4.6.2 and 4.6.3 compare well with each other demonstrating the usefulness
of the upwind source-footprint modeling technique.
90
Figure 4.6.3. Source contribution calculation results for Boston, MA at 5 pm EST May 25, 1999.
To correlate the source-footprint with an emission inventory, knowledge of travel time is
necessary because concentrations at a receptor at a particular time t are due to emissions upwind
at an earlier time (t-travel time). Furthermore, each grid in the domain can be impacted by more
than one puff; thus CALPUFF was modified to include a procedure to compute the average
travel time (tavg) for a puff, weighted by its concentration contribution, to travel from grid i, j to
the receptor for each grid of the domain, for every time t:
N
t avg (i, j , t ) 
 T (i, j, t , k ) * C (i, j, t , k )
k 1
(4.6.1)
CT (i, j , t )
Where,
N = Number of puffs emitted from the receptor from the beginning of the simulation to time t.
T(i,j,t,k) = Travel time of puff k from the receptor to the grid location i, j, at time t.
C(i,j,t,k) = Concentration that puff k contributes to grid location i, j, at time t.
CT(i,j,t) = Total concentration from all puffs at grid location i, j, at time t.
91
The gridded hourly average pollutant source travel times for 5 PM EST, May 25, 1999 for
Boston, MA is depicted in Figure 4.6.4.
Figure 4.6.4. Average pollutant source travel times in relation to Boston, MA at 5 pm EST,
May 25, 1999.
Once the average travel time from emission point i, j to the receptor is known (i.e. tavg), then the
emission rate at grid i, j that contributed to the receptor concentration is simply the emission rate
at t-tavg. In this way time varying emission inventories can be investigated with this method. The
final result is a two-dimensional footprint of the hourly emission inventory for the species of
interest in which each grid point contributed to some extent to the concentration at the receptor.
The fractional contribution of a particular grid point emission to the concentration recorded at the
receptor can then be calculated by
f (i, j, t ) 
Emis(i, j, t  tavg (i, j, t )) * Conc(i, j, t )
R
C

i 1 j 1
Emis(i, j, t  tavg (i, j, t )) * Conc(i, j, t )
Where,
C, R = Number of columns and rows in the domain.
92
(4.6.2)
Emis(i,j,t-tavg(i,j,t)) = Emission rate from the emission inventory contributing to the receptor
concentration at time t.
Conc(i,j,t) = Concentration (as an indicator of probability) from the backward CALPUFF plume.
Figure 4.6.5 depicts the fractional source contribution of NOx on the receptor
concentrations at Boston, MA at 5 pm EST May 25, 1999.
Further deductions regarding the area source contribution can be obtained by tracking
radial upwind areas contributing to the overall fractional contribution. In this manner, fractional
source contribution from 0-15 km, 15-30 km, 30-75 km, and 75-150 km can be determined.
Table 4.6.1 lists the percent contribution of emissions to the receptor concentration at Boston for
May 25, 1999 at 5 PM EST for each of the radial distance ranges. Within a five hour upwind
period, 93% of the emissions contributing to the receptor concentration are within 30 km of
Boston. Figure 4.6.6 shows the fractional source contribution of emissions within approximately
75 km of the receptor at Boston, MA.
Figure 4.6.5. Fractional source contributions of NOx on the receptor concentrations at
Boston, MA at 5 pm EST May 25, 1999.
93
Table 4.6.1 - Percent Contribution of Emissions to the Boston Receptor at Incremented Radial
Distances.
Radial Number
of Grids
Radial Distance
(km)
Contribution (%)
0-5
5 - 10
10 - 25
25 - 50
0 - 15
15 - 30
30 - 75
75 - 150
83.1
9.4
6.0
1.5
The source-footprint modeling system is a simple means to identify upwind source areas
responsible for downwind pollutant concentrations. By linking puff dispersion with emission
inventories, the picture of the resulting upwind source area takes into account distance and
source strength and thus yields a fine scale structure where specific source influences become
more obvious. In this application the system identified the upwind area extending from Boston
southwest to eastern Connecticut as the area responsible for regional pollutants impacting
Boston, MA when winds prevail from the southwest.
Figure 4.6.6. Fractional source contributions of NOx within approximately 75 km of the
receptor, Boston, MA at 5 pm EST May 25, 1999.
94
4.7 Urban Emissions – Air Quality Relationships
4.7.1 Introduction
In many areas of the United States and, indeed, throughout the world, the lack of good
emissions data is frequently cited as one of the key barriers in developing cost-effective
strategies for improving urban air quality [NRC, 1991]. Previous sections of the report have
clearly shown that new instruments can solve the problem of sparseness and lack of chemical
specificity in the routine monitoring networks. A key challenge and the opportunity provided by
the veritable flood of high quality measurement data is how to use this information to develop
more accurate emissions inventories. A key goal of the urban respiration project was to show
how advances in both instrumentation and air quality modeling could be used to better
understand source-receptor relationships. This section of the report describes the results of a
proof-of-concept study of how to develop more accurate emissions inventories by solving the so
called inverse problem: given a model of the transport and transformation processes occurring
determine the spatial, temporal and chemical form of the emissions that gives the best fit to the
air quality observations.
In the past the primary obstacle in solving this problem has been the computational
complexity. This section of the report presents a new approach for the solution of the inverse
problem that is orders of magnitude faster than existing approaches. Quite clearly there are many
potential sources of uncertainties in solving the inverse problem. Another contribution described
below is a new approach for identifying which of the many possible uncertainties in the analysis
have the greatest impacts on the emissions estimates. While the method was tested by
determining the CO and reactive organic gas (ROG) emissions in the Los Angeles basin it has
broad applicability to other regions. Los Angeles was chosen as a study site because of its
extensive data base, the richness of the air quality measurement data base and the opportunity to
do data withholding experiments. An additional practical reason was the need to test the inverse
and uncertainty analysis methods before data from the Massachusetts campaigns became
available. Subsequent sections present the need for more accurate inventories, describe the
character of the inverse problem, the new algorithms and the results. Further details can be found
in two Ph.D. theses supported by the project, [Pun, 1998] and [Wang, 1999].
4.7.2 The Need for Better Emissions Inventories in Control Strategy Design
There have been many obstacles to the development of control strategies, due to difficulties in
quantifying precursor emissions and their transport and reaction in the atmosphere. As with any
policy decision, control strategies are based on information that is uncertain at the time action is
required. The collection of data offers opportunities for added value, although these data incur a
cost measured in terms of both time and capital. In designing air pollution control strategies,
additional data collection and model development can provide valuable information on: ambient
chemical concentrations, meteorology, chemical kinetics, emissions, etc. To date, sampling
programs aimed at collecting this information have not been optimally designed, nor have the
data collected been fully utilized. The magnitude of capital that is invested in control strategies
warrants further work in the efficient collection and use of data that guide their design.
95
The need for improvements in control strategy design was noted by a 1991 National
Research Council (NRC) study [NRC, 1991], particularly in the quantification of emissions and
the atmospheric processes that lead to ozone formation. According to the study, the methods and
protocols used to develop anthropogenic and biogenic ROG emissions inventories should be
reassessed, since existing methodologies do not account accurately for all types of sources, nor
for the magnitude of ROG emissions. This problem has most likely resulted in smaller net
changes in ROG emissions due to past control technology. Also, future strategies are limited by
uncertainties in base-case inventories. The study noted that the use of ambient data to identify
precursor sources, to verify emissions algorithms, and to determine precursor relationships could
be used to verify model predictions that are based on existing inventories. Moreover, the
incorporation of observational data into the modeling of air quality offers many opportunities to
reduce uncertainties in air pollution policy decisions. Previous studies have suggested that
emission inventories currently contribute a large share of the uncertainty in air quality modeling.
Tunnel studies have indicated that mobile source emissions are underestimated by as much as a
factor of three to four. [Fujita, 1992; Pierson, 1990] Since these inventories are generally built
from the 'ground up,' the methodologies and models used to create them need reevaluation.
4.7.3 Inverse Modeling
Mathematical models are developed and used to predict the behavior of a system given a
set of known input parameters. The class of problems that fit this description are termed forward
problems (FP). Inverse modeling focuses on providing information on the values of model
parameters based on the measurement of observable quantities. These parameters can vary from
discrete numerical functions to continuous function of one or more variables. [Menke, 1989] The
relationship between a measured variable and the true value of a parameter function can be
expressed simply as follows:
z  F( )  v
(4.7.1)
where (x) lies in normed function space A, z, the measurement vector and v, the measurement
error vector, lie in an M-dimensional Euclidean vector space, and the forward operator F is a
functional that maps  to z. The goal of inverse estimation is to is find an inverse operator, G,
which maps the measurement vector to an estimate of :
  G(z)  G[F( )  v]
(4.7.2)
In the absence of measurement noise, provided that the forward operator is invertible, the inverse
operator is simply:
G  F 1
(4.7.3)
 ( x)  F 1 ( z )  F 1 ( F ( ))   ( x)
(4.7.4)
Inversion techniques have been studied extensively in geophysical applications, where
they have been classified into two categories: (1) direct or operator-based inversion and
(2) model-based inversion. [Sen and Stoffa, 1995] Direct inversion utilizes an explicit operator
96
to map the observed data to derive the parameter function. Due to the highly non-linear nature of
atmospheric advection-diffusion models, direct inversion of the forward operator is rarely
feasible. Model-based inversion methods utilize the forward operator and an assumed parameter
function to generate data that is compared with the observed data. If these data are not in
agreement, the parameter function is modified, and the process is repeated until the data sets are
comparable. The inversion process is thus transformed into an optimization problem, where a
search is conducted over the parameter space to minimize an error norm. The use of a lowdimensional representation of the parameter space and the application of an efficient search
algorithm are typically necessary to solve these problems.
Once the problem has been stated and structured mathematically, it is critical to assess
the validity and usefulness of the potential solution. To accomplish this for any mathematical
problem, it is necessary to determine if the problem is “well-posed.” An inverse problem is
“well-posed” if it satisfies the following requirements [Hadamard, 1952; Tikhonov and Arsenin,
1977; Sun, 1994]:
Table 4.7.1 - Properties of Well-Posed Inverse Problems
Existence
Uniqueness
Stability
For every measurement vector, there exists a parameter solution
Every parameter solution is unique
The inversion process should be stable on the spaces. Stated
another way, small changes in the measurement vector lead to small
changes in the parameter function
In addition, inverse methods should be robust, or insensitive to a small number of large errors in
the measurement vector [Sen, 1995].
4.7.4 Application to Atmospheric Systems
Inverse problems in atmospheric systems have been solved through the use of techniques
such as source-receptor modeling and Kalman filtering. Kalman Filtering has been widely used
in the estimation of non-stationary stochastic processes including signal processing, optimal
control, and aerospace problems [Daley, 1991]. The method utilizes an iterative inverse
algorithm which recursively updates estimates with only the most recent estimates [McLaughlin,
1995]. Hartley and Prinn [1993] applied an inverse method based on a linear Kalman filter to
determine surface emissions of trace gases using a global atmospheric transport model. Several
key conclusions are noted in their work, including: the high sensitivity of inverse methods to the
accuracy of model circulation and the potential use of inverse methods to identify locations for
future monitoring stations. Other studies of this type include Brown [1993] who solved a
linearized formulation of a chemical transport model to determine trace gas source emissions.
Brown's study, along with Newsam and Enting [1988] noted that, as the spatial and temporal
resolution of the transport model becomes continuous, the amplification of errors in the inversion
approaches infinity. This amplification is dependent on the advective and diffusive rates in the
atmosphere. As the transport and diffusion becomes stronger, the signal from the source will be
smoothed out more rapidly and will fall below the noise level in the observation. Thus,
information on the location and strength of the source will be lost.
97
In summary, several obstacles exist in the inversion of atmospheric transport and reaction
systems. Atmospheric processes of advection and diffusion can reduce the emissions “signal” to
below the noise level, allowing for the amplification of noise through the inversion [Mulholland
and Seinfeld, 1995]. The high dimensionality of grid-based airshed domains makes practical
implementation of emission inversion procedures difficult. What is needed is a low-dimensional
representation of the functional distribution of sources.
4.7.5 Optimal Field Determination
Since a detailed re-assessment of all emissions sources can be an expensive and difficult
task, the concept of inverse modeling can be applied to the problem of estimating urban-scale
emissions, using observational data to determine the actual emissions field.
Stated as a mathematical programming problem, the true emissions field can be found as
follows:
min || C m (E( x, t ), x, t )  C o ( x, t ) || n
(4.7.5)
E( x, t )
minimizing some norm of the error between observed concentration field Co and that predicted
by the model Cm using the emissions fields as the design variables. The goal of this problem is
to vary the emissions spatially and temporally until the difference between the observed and
predicted ozone fields is minimized. This approach assumes that the model itself does not
introduce additional errors. The difficulty in conducting such an optimization is that the
emissions fields in grid based airshed models consist of tens of thousands of data points for each
species of interest. With each iteration of a typical photochemical model requiring several hours
of CPU time, an optimization over all emissions is clearly unaffordable unless the number of
design variables is reduced by several orders of magnitude. For example, the CIT Airshed model
[Harley, 1993] uses an 80x30 cell domain for the Los Angeles basin, resolved into hourly
averaged fields. For a one-day simulation, optimization of this field requires a search of order
24x30x80. The order of the search also scales with the number of resolved species in the
chemical mechanism. The CIT model currently uses a modified version of the LCC
photochemical mechanism [Lurmann, 1987] that includes 35 chemical species, of which 16 are
directly emitted into the atmosphere.
One method of reducing the number of variables in a field is to represent the field by an
orthogonal expansion, such that the new design variables are the expansion coefficients. Such an
expansion should retain as much of the field structure as possible and require the fewest number
of terms possible. These requirements suggest the use of the Karhunen-Loeve procedure with
empirical eigenfunctions [Tatang, 1996]. The Karhunen-Loeve series expansion is widely
known for its optimality property in approximating fields, especially when there is a strong
correlation between points in the field. For a field with complicated structure, the closed form
Karhunen-Loeve series expansion does not give a good representation, therefore an empirical
type of that expansion can be used.
98
Figure 4.7.1. Typical CIT Airshed Model Emissions Field
4.7.6
Empirical Karhunen-Loeve Series Expansion
Consider a field E(x,t) which is to be approximated. The Karhunen-Loeve series
expansion of such a field has the form:
N
E( x , t )   c n  n ( t ) n ( x )
(4.7.6)
n 1
where cn is the square root of the nth eigenvalue, n(t) is the nth temporal eigenfunction, and
(x) is the nth spatial eigenfunction of correlation function of E(x,t). According to KarhunenLoeve series expansion properties, those eigenfunctions should satisfy the normality conditions:
2
(4.7.7)
2
(4.7.8)
 n ( t ) t  T  2n ( t )dt  1
 n ( x ) x  D  2n ( x )dx  1
and also the orthogonality conditions:
 E(x, t ) n ( t )d  c n  n (x)
(4.7.9)
T
 E(x, t ) n (t )d  c n  n (x)
(4.7.10)
D
99
The last two equations are obtained by considering the eigenfunction as an element of a complete
set of the orthonormal functions in the T  D space. From the above equations the relationship
between the eigenvalues and eigenfunctions may be derived as:
2
 C( t, s) n (s)ds  c n  n ( t )
(4.7.11)
T
2
 K( x, y) n ( y)dy  c n  n ( x )
(4.7.12)
D
where C(t,s) is the correlation matrix of the field in the temporal domain and K(x,y) is the
correlation matrix of the field in the spatial domain. They are defined by
C(t, s)   E(x, t )E( x, s)dx
(4.7.13)
D
K(x, y)   E(x, t )E( y, t )dt
(4.7.14)
T
When the spatial dimension is much larger than the temporal dimension, the first integral
equation is clearly preferable to the second one. Assuming the first integral equation is solved,
one next needs to obtain the spatial eigenfunctions. The answer to this actually comes from
Equation 4.7.9. This discussion follows a similar technique from Sirovich and Everson [1992]
who use a snapshot method to calculate the spatial eigenfunctions,
 n (x) 
1
 E( x, t ) n ( t )dt
TT
(4.7.15)
However, instead of using Equation (4.7.15), we use Equation (4.7.9) directly to calculate the
spatial eigenfunctions. This approach seems more appropriate if one wants to ensure the
Karhunen-Loeve series expansion properties in (4.7.9) and (4.7.10).
The empirical Karhunen-Loeve series expansion implies the use of empirical
eigenfunctions in the temporal and spatial domains, instead of closed forms. These empirical
eigenfunctions are matrices of values of the eigenfunctions at each point in T x D space, and
provides a further advantage: singular value decomposition may be used to obtain the temporal
eigenfunctions and eigenvalues and the first integral equation above need not be solved directly.
The Karhunen-Loeve series expansions can be considered as the orthogonalization of a field,
such that the terms in the resulting expansion are uncorrelated, analogous to the singular value
decomposition. Thus, the correlation matrix in the temporal domain is decomposed into
uncorrelated terms as follows:
N
C( t , s)   c 2n  n ( t ) n (s)
(4.7.16)
n 1
100
The values of cn2 and n(t) are the diagonal and orthogonal matrices resulting from application of
the singular value decomposition method to the correlation matrix:
C(t, s)  U  W  V T
(4.7.17)
where W  diag{Cn2 }nN1 and the n-th column of the orthogonal matrix U contains the values of
eigenfunction n(t) at times ti, i = 1,...,N. Thus, the eigenvalues and temporal eigenfunctions are
obtained from orthogonalization of the temporal correlation matrix.
The next step is to calculate the spatial empirical eigenfunctions. Using equation 4.7.9, we can
generate the spatial empirical eigenfunctions,
 n (x) 
1
 E( x, t ) n ( t )dt
cn T
(4.7.18)
which are self-normalized,
 n (x) 
 n (x)
(4.7.19)
2
  n ( x )dx
D
These spatial empirical eigenfunctions along with their temporal counterparts and related
eigenvalues can now be used as the Karhunen-Loeve series expansion of the field.
In general, the empirical Karhunen-Loeve series expansion can be used to approximate
any field accurately. Since the complete set of eigenfunctions from that expansion spans the
L2 space [Aubry, 1991], an arbitrary regular field in L2 space then can be theoretically
approximated using the same set of eigenfunctions. In other words, in practice, chemical or
physical phenomena with similar underlying structure may be represented by the same set of
empirical eigenfunctions [Aubry, 1993; Kramer, 1991; Krischer, 1993; Rowlands, 1992;
Urgell, 1990]. Based on this fact, we can incorporate our prior knowledge of a field and use it to
generate a better prediction.
Retaining only the coherent structures with higher energy as our prior information in the
optimization scheme suggests that we put more weight in keeping the structure of the field.
Similarly, we could also put more weight to a specific coherent structure by adding constraints to
the optimization problem, for example, the coefficient associated with the first empirical
eigenfunction should be greater than others. Other constraints that describe the limit of physical
or chemical phenomena could also be added to the problem. For example, if a chemical
concentration field is approximated, it can be specified that the posterior field will be greater
than zero for all grid points.
101
4.7.7 Los Angeles Case Study
The procedure described above has been applied to the problem of determining the
spatial and temporal structure of VOC and CO emissions within the Los Angeles basin. As in
most urban areas, the validity of the official emissions inventory for organic gases has been
questioned. This example uses data collected during August 27-29, 1987 as part of the Southern
California Air Quality Study (SCAQS). During this study, a variety of special meteorological
and air quality measurements were carried out to supplement the routine measurements made in
the Los Angeles area. The emissions inventory used in this study was received from the
California Air Resources Board. Mobile Source emission estimates were based on a travel
demand model and the EMFAC7E emission factor model. The South Coast Air Quality
Management District prepared stationary source emission estimates. These estimates include
day-specific power plant, aircraft, and refinery emissions. The detailed inventory includes
emissions from more than 800 source types, with the organic gas emissions broken down into
280 detailed chemical species. The inventory region includes the South Coast Air Basin plus
parts of the Southeast Desert Air Basin (see Figure 4.7.2).
The primary assumption of the method used is that errors in the predicted concentration arise
solely from an inaccurate emissions inventory. While it has been suggested that uncertainties in
emissions inventories contribute more to modeling uncertainty than differences in chemical
mechanisms, advection schemes and numerical methods, this is nonetheless a significant
assumption. Future studies using several models could test the robustness of the procedure and
the validity of this assumption.
Figure 4.7.2. CIT Modeling Domain, UTM Coordinates (dotted grid is the domain of CARB
emissions inventory)
102
4.7.8 Formulation of Optimization Problem
A critical step in the formulation of any norm-reduction optimization problem is
identification of error norm to be used as the algorithm’s objective function. For most
atmospheric modeling inverse problems, this step involves identifying a data set that is assumed
to capture the “true” concentration field, as well as a spatial and temporal domain that assures the
problem is well-posed. For the problem under study, monitoring data collected during the
SCAQS program were used as an estimate of the “true” ozone concentration field, assuming
normally distributed measurement errors. While the model domain encompassed over 60
monitoring stations, a subset of 37 stations was selected in order to reduce the weighting of
stations nearest to model boundaries.
Figure 4.7.3. Monitoring Stations Within SCAQS Region and CIT Modeling Domain
(Stations within dark border were used to determine error norms)
The temporal domain of the error norm was chosen to be the 24 hourly data sets for these
37 SCAQS stations on second day of the simulation, August 28, 1987. The elimination of the
first day’s data set is intended to reduce the effect of the model’s initial conditions. Since the
base-case simulation underestimates peak ozone concentrations during the afternoon, the
optimization procedure was also tested using an alternate norm, a measure of error during the
afternoon (the time period that ozone concentrations reach maxima during the SCAQS study).
The optimization was designed to perform a least squares minimization over the temporal and
spatial domains discussed above. In other words, the goal of the optimization is to minimize an
objective function of the following general form:
t n
Error Norm    O 3 (obs )  O 3 (pred) 2
(4.7.20)
1 1
where O3 (obs) are observed ozone levels, O3 (pred) are ozone levels predicted by the CIT
model, n the number of monitoring sites and t the number of hours sampled.
103
The next step is to determine the number of eigenfunctions to retain in the search. It is
clear that the dimensionality of the optimization depends only on the number of coherent
structures we use as our prior information. This determination should balance the incremental
energy of the added eigenfunction with the added computational cost of increasing the
dimensionality in the search algorithm. As is shown in Table 4.7.2, the first 5 eigenfunctions of
the decomposed ALKE and CO fields capture 95 percent and 94 percent, of the ensemble's
energy, respectively. Therefore, each emitted species in this study was represented by a fivedimensional KL expansion. The addition of more eigenfunctions does not significantly increase
the captured energy, allowing a search over the first five eigenfunctions to be acceptable.
Two additional representations of the high degree of compression available in these fields
are presented: (a) graphical depiction of the eigenvalue spectrum (see Figure 4.7.4) and (b) the
representation error (as measured by the ratio of total reduced emissions to total emissions)
(see Figure 4.7.5).
The Broyden-Fletcher-Goldfarb-Shanno (BFGS) variant of Davidon-Fletcher-Powell
(DFP) minimization method was used to obtain the optimal coefficients [Press et al., 1996].
Figure 4.7.6 shows a representative error norm reduction using DFP minimization in this study.
Notice that the most significant reduction is achieved after the first few iterations.
Table 4.7.2 - Percent of Variance Captured by First Five Eigenfunctions
Eigenfunction
First
Second
Third
Fourth
Fifth
ALKE
Variance
Captured
72.3 %
14.9 %
4.0 %
2.1 %
1.7 %
Cumulative
72.3 %
87.2 %
91.2 %
93.3 %
95.0 %
CO
Variance
Captured
79.0 %
5.55 %
4.01 %
2.95 %
1.89 %
Cumulative
79.0 %
84.6 %
88.6 %
91.5 %
93.4 %
Figure 4.7.4. ALKE and CO Emissions Eigenvalue Spectra for August 27-28, 1987
104
Figure 4.7.5
Eigenvalue Representation Error for ALKE, CO and NO Emissions
(August 27-28, 1987)
1.20
1.15
Error Norm
1.10
1.05
1.00
0.95
0.90
0.85
1
2
3
4
5
6
7
8
9
Iteration
Figure 4.7.6
Error Norm Reduction as a Function of Optimization Iteration
The overall optimization flowsheet is shown in Figures 4.7.7 and 4.7.8.
105
OPTIMIZATION FLOWSHEET 1
Assemble base
case input files
Determine fields to be
optimized, dimensionality, etc.
Create KL
basis files
power.c
Edit files used by main.c to
reflect fields to be optimized,
dimensionality, etc.
getfunc
join_direc
main.c
Compile and run main.c
(see Flowsheet 2)
Figure 4.7.7. Flowsheet
1: Overall
optimization
strategy flow diagram
OPTIMIZATION
FLOWSHEET
2
Initialize
Coefficients
main.c
Begin
Search
dfpmin.c
lnsrch.c
Evaluate
Objective
Function
getfunc
Finish
Search
Figure 4.7.8. Flowsheet 2: Search Algorithm and Associated Code
getfunc
Create emission fields
from coefficients
genfield.c
Create CIT emission file
using eigenfields, scaling
joinscale.c
106
Run CIT
CIT
Airshed Model
Finish
Search
getfunc
Create emission fields
from coefficients
genfield.c
Create CIT emission file
using eigenfields, scaling
joinscale.c
CIT
Airshed Model
Run CIT
Extract concentration
fields
getpred.c
Figure 4.7.9. Flow Diagram for Objective Function Evaluator (getfunc shell script)
4.7.9 Pseudo-Data Inversion
In order to test the inversion algorithm, searches were conducted using known solution and false
starting values. A base-case CO emissions field was used as input to the CIT model, producing
hourly averaged CO fields for every hour of a two-day run. CO concentrations at the stations of
interest were extracted and assumed to be the measured concentrations for the verification
simulations. As starting points for these runs, factors ranging from 1.25 through 2.00 were
applied to the true field, in an attempt to simulate inaccurate initial guesses. These searches
successfully reproduced the true field to within 5%.
4.7.10 CO Inversion
Mulholland and Seinfeld [1995] applied a recursive least-square technique similar to
Hartley, Prinn and Brown to estimate the necessary adjustments to the CO inventory in Los
Angeles in order to match observed concentrations. CO was selected to remove reaction effects,
since the consumption of CO by photochemical reactions over the one to two-day time scale is
relatively small. In the Mulholland and Seinfeld procedure, the ground level horizontal domain
of the solution volume is divided into smaller domains that comprise the entire region. Each one
of these source domains is used as the emissions field for the model, with a zero initial condition
and zero boundary condition. Two additional runs were conducted using (a) the initial condition
alone and zero emissions and (b) a time-varying boundary condition with zero emissions. For a
conservative specie, where superposition is valid, these solutions can be added together at any
time step, to construct the complete concentration distribution.
For the CO inversion in this study, the results were compared to the Mulholland/Seinfeld
study, which showed that the average factor by which the base case inventory must be multiplied
was found to peak at 3.0 at midday on weekdays. Their factor agrees with a general adjustment
on mobile emissions used by Harley [1993] to match ozone predictions and observations. On the
107
weekdays, the average factor then fell below unity from 9pm to midnight, indicating that the
current inventory may actually overestimate CO emissions at night. The study presented here
produced a overall adjustment of 1.41 to the base case inventory, reaching a maximum of 1.53 at
4pm, as shown in Figure 4.7.10.
VOC Inversion
Next, an inversion of the VOC emissions within the airshed was conducted. Although
urban air may contain hundreds of individual organic compounds, atmospheric models typically
use a simplified representation of these compounds. For example, the modified LCC mechanism
[Lurmann, 1987] used in the CIT model contains 9 lumped organic classes (see Table 4.7.3).
Figure 4.7.10. Emissions Time Series for CO emissions (August 27-28, 1987)
Table 4.7.3 - Lumped Organic Classes Used in the CIT Model
Class Code
ETHE
ALKE
ALKA
TOLU
AROM
HCHO
ALD2
MEK
MEOH
ETOH
LCC VOC group
Ethene
Alkenes
>C3 Alkanes
Mono-Alkylbenzenes
Di- and Tri-Alkylbenzenes
Formaldehyde
Higher Aldehydes
Ketones
Methanol
Ethanol and higher alcohols
108
Therefore, to test the procedure for the base-case organic fields, a full search over a subset of
coherent structures for each lumped category field would be conducted. In order to reduce the
computational burden of such a search, however, one of the VOC categories, the alkene field
(ALKE), was used as the test field. C3 and higher alkenes belong to this group, including
propene and trans-4.7.butene. The balance of lumped species were determined using the base
case ratio in each cell at each hour as follows:
E opt (Org)(x, t )
E opt (ALKE)(x, t )

E base (Org)(x, t )
E base (ALKE)(x, t )
(4.7.21)
Representations of the spatial and temporal ALKE eigenfunctions are shown in Figures 4.7.11
through 4.7.13.
As seen in Figure 4.7.14, the base case CIT simulation for the August 27-28 episode
under-predicts ozone levels in the Los Angeles basin. Four representative monitoring stations
are shown: Central Los Angeles (CELA), Pasadena (PASA), Claremont (CLAR) and Riverside
(RIVR).
0.3
EIG1
EIG2
0.2
0.2
0.0
0.1
-0.2
0.0
10
20
30
40
10
Time (hour)
20
30
40
Time (hour)
0.4
EIG3
EIG4
0.2
0.2
0.0
0.0
-0.2
-0.2
10
20
30
40
10
Time (hour)
20
30
Time (hour)
Figure 4.7.11. First Four ALKE Temporal Eigenfunctions
109
40
Figure 4.7.12. Two-Dimensional Contour-Plots of First Five ALKE Eigenfunctions
110
Total ALKE
Field at 8am
EIG1
EIG2
EIG3
EIG4
Figure 4.7.13. Three-Dimensional Surface-Plots of First Four ALKE Eigenfunctions
111
Figure 4.7.14. Ozone Predictions for Base Case Simulation (August 28, 1987)
The search resulted in a net increase of VOC emissions of 57 percent over the entire
airshed. If the entire field is scaled at the same factor, an increase of 61 percent is necessary to
minimize the error norm. A comparison between the base case predictions and the optimized
results is shown in Figures 4.7.15 – 4.7.17. An improvement in predictive capability of the
model is seen for all but the highest concentration ranges. The results of this study show that the
existing VOC emissions inventory developed for Los Angeles is underestimated.
In order to test the sensitivity of the method to measurement errors, three additional
optimizations were performed. These runs used the original data set of ozone observations, with
an artificial error term added to each measurement. As shown in Figure 4.7.18, the results from
these runs were in general agreement with the original optimized coefficients. However, it
should be noted that coefficients 4 and 5 show the greatest diversion for the base case
optimization. This is expected, since these eigenfunctions account for a very small fraction of
the total emission field; therefore, large fluctuations in their associated eigenvalues will have a
correspondingly smaller effect on the composite emission field than eigenfields 1 and 2.
112
Figure 4.7.15. Optimized Result for August 28, 1987 (24-hour norm)
Optimum Emissions
5
BASE
OPT
Total ALKE Emissions
4
3
2
1
0
0
4
8
12
16
20
24
28
32
36
40
44
48
Hour
Figure 4.7.16. Emissions Time Series for ALKE emissions (August 27-28, 1987)
113
Figure 4.7.17. Difference Between Optimized and Base Case ALKE Field (8am)
Error Runs
2
Magnitude
1
0
Base
Opt
Error1
Error2
Error3
-1
-2
1
2
3
4
5
Coefficient
Figure 4.7.18. Optimal coefficients for Base Case, Optimized Field and 3 error runs
(ALKE emissions) (August 27-28, 1987)
114
4.7.11 Objective Function Modification
In order to test the sensitivity of the optimization to an alternate objective function, an
additional optimization was conducted using a norm calculated from 14.7.5pm. These results are
shown in Figures 4.7.19 through 4.7.21. Interestingly, this search also resulted in a net increase
of VOC emissions of 57 percent over the entire airshed. Qualitatively, this simulation shows
that, in order to achieve the high ozone levels reached mid-day, over-estimation of early morning
and evening concentrations is necessary.
4.7.12 Summary of Results from Inverse Methods
The adjustments made to the ROG inventory in this section are potentially indicative of
the errors that exist in emissions inventories worldwide. In a region where the structure of the
emissions fields can be inferred from known information, this procedure can be applied to design
data collection studies using traditional sampling or remote sensing techniques. Inaccuracies in
base-case emissions have the potential to lead to inaccuracies in the control strategies developed
from them. As stated in Section 2.6, the primary assumption of the method presented here is that
errors in the predicted concentration field arise solely from an inaccurate emissions inventory.
Figure 4.7.19. Optimized Result for August 28, 1987 (14.7.5pm norm)
115
Objective Function Effect
2
Base
Opt (24 hour)
Opt (Afternoon)
Magnitude
1
0
-1
-2
1
2
3
4
5
Coefficient
Figure 4.7.20. Optimal coefficients for Base Case and Optimized Fields for Two Error Norms
(August 27-28, 1987)
Optimum Emissions
7
BASE
OPT
Total ALKE Emissions
6
5
4
3
2
1
0
0
4
8
12
16
20
24
28
32
36
40
44
48
Hour
Figure 4.7.21. Emissions Time Series for ALKE Emissions (August 27-28, 1987)
116
Thus, all additional factors that contribute to predictions and observations are assumed to be
correct, such as the chemical mechanism, advection scheme, mixing height algorithm and the
measurement vector. The following section addresses one potential source of uncertainty in
model predictions: the rate parameters used in the chemical mechanism. The incorporation of
model uncertainty into the methodology presented in this section, representing a relaxation of the
methodology's underlying assumptions, would likely alter the results presented here.
4.7.13 Role of Uncertainty
Uncertainty always exists in industrial and engineering systems that include reaction and
transport phenomena. Uncertainties can be introduced through computational sources, process
modeling and human factors. The critical problem, in practice, is to evaluate the effects of
uncertainties on predicted outcomes. This research has focused on quantifying uncertainties
introduced through the modeling of air pollution systems. For example, in the development of air
pollution control strategies, uncertainty can be introduced from many sources, such as chemical
mechanisms, numerical techniques, emissions inventories, advection schemes, etc.
Modeling uncertainty can originate from the structure and parameterization of the model,
or from uncertain input parameters. Structural uncertainty can be estimated by comparing the
output of several models of the same system. Parametric uncertainty has typically been estimated
using sensitivity analysis or sampling techniques such as Monte Carlo methods. For large
modeling systems, such as photochemical models, computational sampling techniques are not
efficient, since they require several thousand simulations.
4.7.14 Probabilistic Collocation Approach
A computationally efficient method can be used which allows parametric uncertainty to be
treated directly in the models of reaction and transport. This novel method, termed probabilistic
collocation approach, becomes a better alternative to Monte Carlo methods when the response
surface is smooth, regular, and polynomially approximable [Tatang, 1994]. Before discussing
the probabilistic collocation approach, it would be useful if the basis for the deterministic
variational approach is presented.
One of the applications of deterministic variational approach is to approximate a set of
outputs or response variables u(x) of a given model:
N(u,x)  u(x) = f(x)
(4.7.22)
where x are a set of inputs or independent variables, N(u,x) is a known operator which could be
linear or non-linear, f(x) is a known forcing function, by using a set of specified functions of
inputs {gi(x)}:
M
uˆ ( x)   u i g i ( x)
(4.7.23)
i 0
117
The problem has now changed from evaluating the outputs to calculating the coefficients in the
approximation {ui}. There are several available approaches that can be used to calculate those
coefficients, including the Galerkin and collocation approaches. They are commonly used for
solving partial differential equation models. The Galerkin approach can be described as follows:
Since the approximation of u(x) may not exactly satisfy the model, we define the residual of the
model as follows:
Rui , x   N (uˆ , x)  uˆ ( x)  f ( x)
(4.7.24)
The Galerkin approach minimizes the residual of the model with respect to a given set of
approximating functions {gi(x)}. It requires that the inner-product of residual and each member
of the set of approximating functions should be equal to zero (i.e., they are orthogonal to each
other):
 Ru , xg ( x)dx  0,
X
i
i
i  1,, M
(4.7.25)
From this equation the M unknown coefficients {ui} can be obtained. However, to solve for these
M unknown coefficients the form of residual function must be known and the integrals must be
solvable. As an example, a quadrature method may be used to evaluate these integrals:

X
Rui , x g i ( x)dx   v j Rui , x j g i ( x j ), i  1,..., M
K
(4.7.26)
j 1
where {vj} is the corresponding set of weights in the quadrature method. If vj \ gi(xj) has the
same sign and is not zero for all i and j, and K = M, then equation 4.7.26 can be approximated by
the following:
Ru i , x j   0,
j  1,..., M
(4.7.27)
Equation 4.7.26 can be rewritten in terms of the integral:
 Ru , x x  x dx  0,
X
i
j
j  1,..., M
(4.7.28)
This equation represents the use of the collocation approach for calculating the
coefficients ui, and it suggests that the collocation approach can be treated as an approximation
of the Galerkin approach. The main difference between the Galerkin and the collocation
approach lies in the application of basis functions in the Galerkin approach and the delta
functions in the collocation approach. This implies that the collocation approach can be applied
directly to a ``black-box" or implicit type model, in other words, the collocation approach does
not require one to have the form of residual function. If the value of the residual can be
calculated at several given values of inputs, the coefficients of approximation can be determined.
118
The concept of deterministic variational approach can be extended to the probabilistic
space. We wish to approximate the outputs or response variables of a model whenever the inputs
or independent variables are random variables with a specified joint probability density function.
If the inputs or independent variables are random variables then the outputs, the operator, and the
forcing function become random variables. As a result, the residual and the basis functions are
also random variables. Therefore, the orthogonal relationship defined by equation 4.7.27
between the residual and any basis function should be transformed to the probabilistic space by
incorporating the joint probability density function of inputs p x ( ) ( x( )) :

X ( )
p x ( ) ( x( )) Ru i , x( ) g i ( x( )) dx( )  0, i  1,..., M
(4.7.29)
where x(w) denotes random inputs. This relationship can also be defined in terms of the
expected value as the following:
E Rui , x( ) g i ( x( ))   0, i  1,..., M
(4.7.30)
Similarly, the equations representing probabilistic collocation approach become

X ( )
p x ( ) ( x( )) Ru i , x( )  ( x( )  x j )dx( )  0,
p x ( ) ( x j ) Ru i , x j   0,
j  1,..., M
j  1,..., M
(4.7.31)
(4.7.32)
If we choose xj such that p x ( ) ( x j ) is positive for all j then Equation (4.7.31) can be
applied to the cases where x are random variables. The first step in performing any uncertainty
analysis is to determine which uncertain variables are of interest. For example, if a chemical
mechanism is to be studied, the uncertain input variables could be the reaction rate constants or
initial concentrations, while the output variable of interest may the average or maximum
concentration of key species. The iterative approximation and estimation of error process makes
the collocation approach suitable to a "black-box" type model. The number of times a model has
to be solved will depend greatly on the shape and smoothness of response surface, and for some
problems, the polynomial chaos expansions may not give a good approximation. Nevertheless,
the generated solutions of the model at those collocation points can be reused if we want to
continue performing uncertainty analysis with Monte Carlo methods. This allows the results
from performing collocation to be useful whether a good approximation is achieved or not.
As presented by Tatang [1994] and Wang [1999], the general procedure for applying the
collocation method using DEMM is shown in Figure 4.7.22. Following model selection, the
uncertain input variables to be studied must be selected. It should be noted that, while collocation
methods generally provide a large computational advantage over Monte Carlo methods, an
uncertainty analysis of a large model on more than a few parameters may be computationally
expensive; therefore, an initial judgement must be made on the relative importance of input
parameters. Uncertainties in these parameters must then be represented, typically by assuming an
119
appropriate probability distribution. These distributions may be based on several methodologies,
including standard statistical estimation and expert elicitation (a full discussion of these methods
is presented by Morgan and Henrion [1990]). The order of the output polynomial chaos
expansion representation must also be decided and this is typically done iteratively, based on the
size of the representation error.
Iterative selection
of key parameters
Describe Uncertain Input Parameters
Determine Output PCE Representation
Yes,
Go to
Next
Order
Calculate Collocation Points
BLACK BOX MODEL
Reduce No. of
Uncertain Parameters
Based on VC
No
Acceptable Dimensionality?
Determine PCE Coefficients
Calculate No. of Terms of Next
Order of PCE Approximation
Calculate Error of Truncation
Calculate Variance Contributions
of Parameters
No
Acceptable Error?
Yes
Iteration to
increase the
accuracy of
the DEMM
approximation
Figure 4.7.22. DEMM Flow Diagram [Wang, 1999]
4.7.15 Applications
The relative contributions of several variables to uncertainties in predicted concentrations can be
determined through the application of the probabilistic collocation technique. This will serve as
the basis to assess priorities for data collection studies aimed at reducing these uncertainties.
The quantity of input data required to perform an urban scale photochemical simulation is
typically large. These data are either measured (e.g., temperature) or estimated (e.g., surface
roughness). Since measured data is collected generally on a sparse grid, interpolation or
smoothing techniques are used to generate a complete field. Table 4.7.4 includes a partial list of
the input variables that are used in photochemical models. Since all of these variables contain
some degree of measurement or estimation error, they all contribute to the uncertainty of model
predictions.
A first attempt will be made to assess the uncertainties in these variables and test the
applicability of collocation by estimating response surfaces. Previous studies have been
successful in applying the collocation method to a photochemical box model [Pun, 1998]. For
example, Pun's analysis of 19 photolysis rate constants in the SAPRC photochemical mechanism
showed that more than 50 percent of the uncertainty in predicted ozone levels originated from
only two of these rate constants.
120
4.7.16 Uncertain Parameters
The CIT Airshed model [Harley, 1993] currently uses a modified version of the LCC
photochemical mechanism [Lurmann, 1987] that includes 35 chemical species, of which 16 are
directly emitted into the atmosphere. This mechanism includes 95 reactions, each contributing
some degree of uncertainty to model predictions. While DEMM provides an efficient tool to
assess uncertainty in large models such as CIT, a full uncertainty assessment of these
95 reactions and their associated emissions is computationally prohibitive and impractical.
Fortunately, the photochemical box model simulations performed by Pun [1997] provide
guidance in limiting the scope of this study by focusing on variance-contributing inputs (VCIs),
rather than on all uncertain parameters. Table 4.7.5 lists the rate parameters that were selected
for the uncertainty analysis of the CIT model. These uncertainty factors were taken from
Stockwell's compilation for the SAPRC mechanism, a modified version of the LCC mechanism.
Based on the results of the inverse modeling study presented in Section 2, ROG emissions were
increased by a factor of 1.5.
As shown in Figure 4.7.22, one of the primary decisions in the application of the
collocation method is deciding the order of approximation. For this study, as in Pun [1998], a
second order approximation was used, requiring 97 model runs for a full evaluation and thirdorder error analysis. As shown in Table 4.7.6, this level of approximation was adequate,
producing relative errors generally below 0.05 (approximately equivalent to 5 % error) during
the time periods of interest. Errors above 0.1 were observed before 6 am and after 6pm,
however, ozone concentrations during these times are very low and these periods are not of
primary interest.
Table 4.7.4 - Typical Input Parameters used in Photochemical Models
Category
Chemical Mechanism
Emissions
Air Quality
Meteorological
Surface Removal
Variables
Rate Constants
Activation Energies
Ground Level Sources
Elevated Sources
Initial/Boundary Conditions
Upper Level Initial/Boundary
Conditions
Mixing Depth
Wind Fields
Temperature
Humidity
Solar Radiation
Surface Roughness
Land Use Categories
121
Table 4.7.5 - Uncertain mechanism parameters (based on Stockwell and Pun [1997])
RATE
PARAMETER
j(NO2)
k(O3+NO)
j(O3OSD)
k(OSD+H2O)
k(OSD+M)
k(HO+NO2)
j(HCHOR)
j(HCHOM)
k(ACO3+NO)
k(ACO3+NO)
k(PAN)
k(ALK+OH)
k(ETH+OH)
k(ETH+O3)
k(OLE+OH)
k(OLE+O3)
k(XYL+OH)
Factor (ENOx)
Factor (EROG)
UNCERTAINTY
FACTOR
(LOGNORMAL
DISTRIBUTION)
1.3
1.2
1.4
1.26
1.26
1.28
1.4
1.4
2.0
2.0
2.0
1.3
1.15
1.25
1.2
1.5
1.3
1.2
1.2
BASIS FOR
INCLUSION (E.G.)
VC (O3)
VC (O3)
VC (O3)
VC (OH)
VC (OH)
VC (O3, OH)
VC (O3)
VC (O3)
Secondary Organic VCI
Secondary Organic VCI
VC (NO2)
Primary Organic VCI
Primary Organic VCI
Primary Organic VCI
Primary Organic VCI
VC (O3, OH)
VC (O3)
Table 4.7.6 - Relative Error of Second Order Approximation for Ozone
Hour
6
7
8
9
10
11
12
13
14
15
16
17
18
CELA
3.61E-02
3.28E-02
5.87E-02
6.58E-02
5.03E-02
4.38E-02
2.82E-02
2.51E-02
2.45E-02
1.09E-02
2.57E-02
1.91E-02
3.78E-02
CLAR
1.17E-02
1.71E-02
2.57E-02
1.33E-02
1.44E-02
1.29E-02
1.33E-02
1.56E-02
2.33E-02
3.36E-02
3.23E-02
3.16E-02
2.61E-02
HAWT
2.92E-02
3.44E-02
4.09E-02
3.30E-02
2.18E-02
1.25E-02
2.28E-02
9.84E-03
1.84E-02
1.53E-02
1.35E-02
9.74E-03
1.14E-02
122
PASA
2.11E-02
1.78E-02
8.81E-02
5.04E-02
2.41E-02
3.19E-02
3.94E-02
3.90E-02
3.79E-02
2.97E-02
2.26E-02
1.85E-02
1.76E-02
RIVR
4.94E-02
3.11E-02
1.76E-02
1.44E-02
7.46E-03
7.41E-03
1.14E-02
9.74E-03
1.10E-02
1.83E-02
2.90E-02
3.24E-02
3.03E-02
4.7.17 Results: Ozone Uncertainty and Variance Analysis
In order to simplify the analysis and discussion presented here, several locations within
the modeled domain were selected as being representative of various pollution scenarios. These
locations correspond to the locations of five monitoring stations within the South Coast Air
Basin (See Table 4.7.7 and Figure 4.7.23). The Central Los Angeles location (CELA) was
selected as being representative of an urban location within the zone of highest emissions; ozone
concentrations within this area are generally lower than downwind sites due to local nitric oxide
emissions. Riverside-Rubidoux (RIVR) was selected as a representative downwind location; this
area typically experiences the highest ozone levels within the basin.
Table 4.7.7 - Monitoring Site Description
Site
Description
UTM East
HAWT
CELA
Hawthorne
Los-Angeles-NorthMain
Pasadena-Wilson
Claremont-College
Riverside-Rubidoux
373.4
386.9
UTM
North
3754.3
3770.1
396.1
435.1
461.5
3777.3
3773.5
3762.0
PASA
CLAR
RIVR
HAWT
CELA
PASA
CLAR
RIVR
Figure 4.7.23. Monitoring stations within SCAQS region and CIT Modeling domain
123
The time series of predicted ozone concentrations at the five selected locations and their
associated uncertainties are presented in Figures 4.7.24-37. As expected, ozone concentrations
and associated uncertainties are highest downwind of central Los Angeles. For example, the
peak ozone concentration at CELA is 109  33 ppb, while the peak at RIVR is 176  43 ppb. The
results of the variance analysis are shown graphically and these results are tabulated for two
sites, CELA and RIVR, and presented in Tables 4.7.8 and 4.7.9.
Table 4.7.8 -
NO.
Ozone Variance Contribution, CELA (percent)
(values > 5 are in bold, values < 0.1 not shown
1
5
UNCERTAIN REACTION RATE/
EMISSION FACTOR
NOx Emission Factor
ROG Emission Factor
NO2  h  NO  O
NO  O3  NO2
HOUR
4
10
95.1 32.3
11.8
13.3
0.5
4.7
14
16.9
15.5
10.5
3.7
18
21.5
2.6
21.9
4.8
22
34.4
0.3
38.7
0.3
16
O3  h  O2  O( 1D )
-
-
1.3
-
0.1
17
O(1D)  H 2 O  OH
-
-
0.3
-
0.4
18
-
-
0.3
-
-
22
O(1D )  O
NO2  OH  HNO3
0.6
13.7
8.2
2.5
0.6
39
40
47
HCHO  h  2HO2  CO
HCHO  h  CO
CH 3C(O)O2   NO  NO2  HCHO  RO2 R  RO2
-
4.9
0.1
5.3
6.6
0.3
13.1
0.5
0.1
19.9
0.4
0.1
13.0
48
CH 3C(O)O2   NO2  CH 3C(O)O2 NO2
0.5
4.4
9.8
5.5
1.4
51
CH 3C(O)O2 NO2  CH 3C (O)O2   NO2
0.2
5.6
12.2
20.3
9.8
57
71
72
ALKA  OH  HCHO  ALD2  MEK  RO2 N  R2O2  RO2
1.0
0.8
0.2
0.1
-
0.7
0.1
-
0.2
-
0.2
-
0.3
1.0
0.1
3.6
0.6
-
0.1
0.2
75
76
80
ETHE  OH  RO2 R  RO2  HCHO  ALD2
ETHE  O3  HCHO  HO2  CO
ALKE  OH  RO2  RO2  HCHO  ALD2
ALKE  O3  HCHO  ALD2  RO 2 R  RO 2  HO2  OH  CO
AROM  OH  CRES  HO2  RO2 R  RO2  DIAL  MGLY  CO
124
1
5
Table 4.7.9 - Ozone Variance Contribution, RIVR (percent)
(values > 5 are in bold, values < 0.1 not shown)
UNCERTAIN REACTION RATE/
HOUR
EMISSION FACTOR
4
10
NOx Emission Factor
96.5 5.1
ROG Emission Factor
0.5
8.5
11.8
NO2  h  NO  O
1.9
4.0
NO  O3  NO2
14
5.2
18.7
13.9
5.3
18
30.4
23.4
3.8
1.7
22
72.1
8.1
2.2
0.2
16
O3  h  O2  O( 1D )
-
0.2
0.9
1.4
0.4
17
O(1D)  H 2 O  OH
-
-
0.3
0.2
-
18
-
-
0.3
0.3
-
22
O(1D )  O
NO2  OH  HNO3
-
3.8
7.4
12.8
3.8
39
40
47
HCHO  h  2HO2  CO
HCHO  h  CO
CH 3C(O)O2   NO  NO2  HCHO  RO2 R  RO2
0.1
3.3
27.2
3.2
0.3
14.4
9.5
0.5
4.6
6.5
3.6
48
CH 3C(O)O2   NO2  CH 3C(O)O2 NO2
0.1
12.0
16.6
5.3
-
51
CH 3C(O)O2 NO2  CH 3C (O)O2   NO2
0.3
21.4
12.2
4.2
3.1
57
71
72
ALKA  OH  HCHO  ALD2  MEK  RO2 N  R2O2  RO2
0.2
-
0.5
-
0.8
0.1
-
0.9
0.1
-
0.1
-
0.2
0.1
0.2
2.1
0.5
0.1
0.9
-
RX
NO.
75
76
80
ETHE  OH  RO2 R  RO2  HCHO  ALD2
ETHE  O3  HCHO  HO2  CO
ALKE  OH  RO2  RO2  HCHO  ALD2
ALKE  O3  HCHO  ALD2  RO 2 R  RO 2  HO2  OH  CO
AROM  OH  CRES  HO2  RO2 R  RO2  DIAL  MGLY  CO
Table 4.7.10 - Relative Change in Ozone Percent Variance Contribution1.5 ROG Case versus
Base Case Analysis (values shown are average factors from 12:00-16:00PST)
RX
NO.
CELA
RIVR
1
5
UNCERTAIN REACTION RATE/
EMISSION FACTOR
NOx Emission Factor
ROG Emission Factor
NO2  h  NO  O
NO  O3  NO2
0.62
1.24
0.86
0.86
0.38
0.96
1.90
1.51
22
NO2  OH  HNO3
0.90
0.61
39
47
HCHO  h  2HO2  CO
CH 3C(O)O2   NO  NO2  HCHO  RO2 R  RO2
1.11
2.05
0.60
2.51
48
CH 3C(O)O2   NO2  CH 3C(O)O2 NO2
1.30
1.35
51
CH 3C(O)O2 NO2  CH 3C (O)O2   NO2
1.84
2.65
125
As expected, the increase in ROG emissions affected the variance analysis compared to
the base-case uncertainty analysis presented by Pun [1998]. Average changes in variance
contribution of the uncertain parameters during the hours when ozone levels are highest
(12 – 4 pm) were calculated (from hourly time series during this period for several parameters see Figures 4.7.24-25). Table 4.7.10 shows these average values at CELA and RIVR for
parameters that contributed more than 1 percent of ozone variance. At both sites, the
contributions of the NOx emission factor and the nitric acid formation reaction were reduced by
the increase in ROG rate [see Table 4.7.11]. The contributions of the oxidation of nitric oxide by
peroxyl acyl attack and the formation and destruction reactions for PAN were increased at both
sites. The contributions of the photolysis of nitrogen dioxide and the reaction of nitric oxide and
ozone were decreased at CELA and increased downwind at RIVR. The variance contribution of
the ROG emission factor was increased at CELA and slightly reduced at RIVR.
ROG Emission Factor Effect
CELA
ROG Emission Factor Effect
RIVR
4.0
5.0
kACO3NO
kACO3NO2
kPAN
FAC_ROG
2.0
1.0
VC(1.5ROG)/VC(BASE)
VC(1.5ROG) / VC(BASE)
4.0
3.0
jNO2
kO3NO
kACO3NO
kACO3NO2
kPAN
3.0
2.0
1.0
0.0
0.0
12
13
14
15
16
12
13
Hour
14
15
16
Hour
Figure 4.7.24. Uncertain parameters with increased average variance contribution under a
1.5xROG case
ROG Emission Factor Effect
CELA
ROG Emission Factor Effect
RIVR
2.0
3.0
kACO3NO
kACO3NO2
kPAN
FAC_ROG
2.0
1.0
VC(1.5ROG)/VC(BASE)
VC(1.5ROG) / VC(BASE)
4.0
0.0
kHONO2
jHCHOR
FAC_NOX
1.0
0.0
12
13
14
15
16
12
Hour
13
14
15
16
Hour
Figure 4.7.25. Uncertain parameters with decreased average variance contribution under a
1.5xROG case
126
Table 4.7.11 - Effect of different scenarios on variance contribution (12 - 4 pm)
( A = Base Case; B = 1.5  ROG; C = 1.5  ROG and 1.5  NOx )
NO.
UNCERTAIN REACTION RATE/
EMISSION FACTOR
NOx Emission Factor
1
NO2  h  NO  O
5
NO  O3  NO2
CELA
A
B
27. 16.
3
9
12. 15.
1
5
13. 10.
0
5
4.7 3.7
16
O3  h  O2  O( 1D )
1.8
1.3
1.0
1.5
0.9
1.2
17
O(1D)  H 2 O  OH
0.2
0.3
0.1
0.2
0.3
0.1
18
0.4
0.3
0.2
0.3
0.3
0.2
22
O(1D )  O
NO2  OH  HNO3
9.3
8.2
8.1
8.5
7.4
39
40
47
HCHO  h  2HO2  CO
HCHO  h  CO
CH 3C(O)O2   NO  NO2  HCHO  RO2 R  RO2
5.8
0.2
8.3
5.5
0.3
6.5
48
CH 3C(O)O2   NO2  CH 3C(O)O2 NO2
7.5
6.6
0.3
13.
1
9.8
51
CH 3C(O)O2 NO2  CH 3C (O)O2   NO2
8.5
6.6
57
71
72
ALKA  OH  HCHO  ALD2  MEK  RO2 N  R2O2  RO2
0.6
0.0
0.0
12.
2
0.7
0.1
0.0
0.6
0.0
0.0
4.9
0.2
14.
6
16.
5
12.
7
0.6
0.0
0.0
3.2
0.3
14.
4
16.
6
12.
2
0.8
0.1
0.0
14.
2
9.0
0.3
5.1
0.1
0.0
0.4
91.
8
24.
3
39.
4
20.
8
0.0
0.0
0.6
92.
8
35.
1
32.
4
18.
7
0.1
0.0
0.4
92.
7
18.
1
48.
9
18.
9
0.0
0.0
0.6
92.
1
43.
8
23.
8
17.
7
0.0
0.0
0.5
93.
7
43.
2
23.
9
18.
3
ROG Emission Factor
ETHE  OH  RO2 R  RO2  HCHO  ALD2
ETHE  O3  HCHO  HO2  CO
75
ALKE  OH  RO2  RO2  HCHO  ALD2
ALKE  O3  HCHO  ALD2  RO 2 R  RO 2  HO2  OH  CO
76
AROM  OH  CRES  HO2  RO2 R  RO2  DIAL  MGLY  CO
80
Total for highlighted parameters
Total for Peroxyacyl radical system (Rxns. 47, 48, 51)
Total for Emission Factor parametric uncertainty
Total for Photolysis Rate parametric uncertainty (Rxns. 1, 16, 39,
40)
127
C
35.
8
13.
1
12.
1
4.6
5.0
RIVR
A
B
8.8 5.2
15.
0
11.
1
4.3
18.
7
13.
9
5.3
C
29.
9
20.
0
3.7
1.9
6.7
5.1
0.8
0.1
0.0
0.1
0.0
1.6
90.
0
16.
9
49.
9
14.
2
4.7.18 Observations/Conclusions
Several observations can be made from the variance analysis:
(a) In a system where 95 reactions define the chemical mechanism, relatively few parameters
contribute a large portion of the total uncertainty for the compound of interest, ozone. In this
study, across different scenarios, approximately 8 of the 19 parameters studied contribute more
than 90% of the ozone variance. These 19 parameters (2 emission factors and 17 rate
parameters) were shown to be variance-contributing inputs (VCIs) by Pun (1997), and are thus
believed to contribute a large portion of the variance of the full model. Therefore, fewer than 10
parameters likely contribute most of the uncertainty of the full model.
(b) Before sunrise (05:25PST in the current simulation), as expected, uncertainties in the
photolysis rate constants do not contribute to the variance in ozone concentrations. While this
may be a trivial point, it is important to recognize that over different spatial and temporal ranges,
different reaction subsets will be active and thus, will contribute to uncertainties in model
predictions.
(c) Over much of the domain, the variance contribution of the NOx emission rate is larger than
that of the ROG emission rate. A notable exception is at RIVR at the time that corresponds to
peak ozone levels, where NOx and ROG emissions uncertainty contribute 5.2 and 18.7 %,
respectively.
(d) At the 'upwind' sites (CELA, PASA, HAWT), maximum ozone concentrations generally
coincided with the time when total ozone variance achieved a maximum, as given in Table
4.7.12. For the 'downwind' sites (CLAR, RIVR), the time at which maximum ozone levels were
achieved generally preceded the peak in total variance by several hours.
Table 4.7.12 - Temporal Differences in Ozone Concentration and Variance
Site
CELA
HAWT
PASA
CLAR
RIVR
Maximum Ozone
Concentration
14:00
13:00
15:00
13:00
14:00
Maximum Ozone
Variance
13:00
12:00
14:00
16:00
17:00
(e) Initially (following sunrise), in urban ('upwind') locations such as CELA, uncertainties in the
NOx photolysis rate, the formation of NO2 by reaction of nitric oxide and ozone, and NOx
emission rate, dominate the variance contribution to ozone [see Figure 4.7.26].
128
CELA Variance Contribution
100
Ozone Variance Contribution
FAC_NOX
kO3NO
80
jNO2
60
40
20
0
6
7
8
9
10
11
12
Hour
Figure 4.7.26. Morning Ozone Variance Contribution, CELA
(f) A few radical reactions that serve as sources and sinks for NOx were shown to contribute
significantly to ozone variance, including the three reactions that involve peroxyacyl radical [see
Table 4.7.13]. Two of these reactions control the formation and destruction of PAN. The
reaction between nitrogen dioxide and the hydroxyl radical was also an important contributor.
For example, these four reactions contribute more than 50% of the maximum ozone variance at
Riverside.
Table 4.7.13 - Ozone Variance Contribution for Selected Reactions at CELA and RIVR (2pm)
Reaction
NO2  OH  HNO3
Description
Nitric Acid formation
CELA
8.2 %
RIVR
7.4 %
CH 3C(O)O2   NO  NO2  HCHO  RO2 R  RO2
Oxidation of NO by
peroxy acyl radical
PAN source
13.1 %
14.4 %
9.8 %
16.6 %
PAN sink
12.2 %
12.2 %
43.3 %
50.6 %
CH 3C(O)O2   NO2  CH 3C(O)O2 NO2
CH 3C(O)O2 NO2  CH 3C (O)O2   NO2
TOTAL
(g) As one might expect in a photochemical mechanism, uncertainties in the photolysis rate
constants contribute an important fraction of the ozone variance. The four reactions in Table
4.7.14 contribute between 15 and 20 percent of peak ozone variance, depending on scenario:
Table 4.7.14 - Primary Variance-Contributing Photolysis Reactions
Reaction Rate
Photolysis Reaction
No.
1
NO2  h  NO  O
16
O3  h  O2  O( 1D )
39
40
HCHO  h  2HO2  CO
HCHO  h  CO
129
(h) At a downwind site like RIVR, where the highest ozone levels are typically reported for the
region, a small subset of the uncertain parameters analyzed in this study contributes more than
90% of the ozone variance throughout most of the day. (See Figure 4.7.27) These parameters are
given below in Table 4.7.15.
Table 4.7.15 - Dominant Variance-Contributing Parameters at RIVR
Uncertain Reaction Rate/Emission Factor
FAC_NOx
FAC_ROG
NO2  h  NO  O
NO2  OH  HNO3
Description
NOx Emission Factor
ROG Emission Factor
Nitrogen Dioxide photolysis
Nitric Acid formation
CH 3C(O)O2   NO  NO2  HCHO  RO2 R  RO2
CH 3C(O)O2   NO2  CH 3C(O)O2 NO2
Oxidation of NO by
peroxy acyl radical
PAN source
CH 3C(O)O2 NO2  CH 3C (O)O2   NO2
PAN sink
RIVR Ozone Variance Contribution
100
Variance Contribution
80
FAC_ROG
FAC_NOX
60
kPAN
kACO3NO2
kACO3NO
jHCHOR
40
kHONO2
jNO2
20
23
21
19
17
15
13
11
9
7
5
3
1
0
Hour
Figure 4.7.27. Primary Variance Contributing Parameters at RIVR
(i) Qualitatively, the results of the scenario analysis show that the few parameters that contribute
a large fraction of the mechanism uncertainty do not change across emissions scenarios.
However, as expected, the quantitative contribution of individual parameters is a function of
scenario.
130
CELA Ozone
[Ozone] (ppm)
0.16
0.12
0.08
0.04
0.00
0
4
8
12
16
20
Hour
Figure 4.7.28. Ozone Time Series (1.5x Base Case ROG) (8/27/87), Central Los Angeles
Figure 4.7.29. Ozone variance time series (1.5x Base Case ROG) (8/27/87),
Central Los Angeles
131
CLAR Ozone
[Ozone] (ppm)
0.20
0.16
0.12
0.08
0.04
0.00
0
4
8
12
16
20
Hour
Figure 4.7.30. Ozone Time Series (1.5x Base Case ROG) (8/27/87), Claremont
CLAR
FAC_ROG
FAC_NOX
2500
kXYLHO
kOLEO3
kOLEHO
Ozone Variance (ppb 2)
2000
kETHO3
kETHHO
kALKHO
1500
kPAN
kACO3NO2
kACO3NO
1000
jHCHOM
jHCHOR
kHONO2
500
kOSDM
kOSDH2O
jO3O1D
kO3NO
23
21
19
17
15
13
11
9
7
5
3
1
0
jNO2
Hour
Figure 4.7.31. Ozone Variance Time Series (1.5x Base Case ROG) (8/27/87), Claremont
132
HAWT Ozone
[Ozone] (ppm)
0.12
0.08
0.04
0.00
0
4
8
12
16
20
Hour
Figure 4.7.32. Ozone Time Series (1.5x Base Case ROG) (8/27/87), Hawthorne
HAWT
FAC_ROG
FAC_NOX
600
kXYLHO
kOLEO3
Ozone Variance (ppb 2)
500
kOLEHO
kETHO3
kETHHO
400
kALKHO
kPAN
kACO3NO2
300
kACO3NO
jHCHOM
200
jHCHOR
kHONO2
kOSDM
100
kOSDH2O
jO3O1D
kO3NO
23
21
19
17
15
13
11
9
7
5
3
1
0
jNO2
Hour
Figure 4.7.33. Ozone Variance Time Series (1.5x Base Case ROG) (8/27/87), Hawthorne
133
PASA Ozone
[Ozone] (ppm)
0.20
0.16
0.12
0.08
0.04
0.00
0
4
8
12
16
20
Hour
Figure 4.7.34. Ozone Time Series (1.5x Base Case ROG) (8/27/87), Pasadena
PASA
FAC_ROG
FAC_NOX
1800
kXYLHO
kOLEO3
1600
Ozone Variance (ppb 2)
kOLEHO
1400
kETHO3
kETHHO
1200
kALKHO
kPAN
1000
kACO3NO2
kACO3NO
800
jHCHOM
600
jHCHOR
kHONO2
400
kOSDM
kOSDH2O
200
jO3O1D
kO3NO
23
21
19
17
15
13
11
9
7
5
3
1
0
jNO2
Hour
Figure 4.7.35. Ozone Variance Time Series (1.5x Base Case ROG) (8/27/87), Pasadena
134
RIVR Ozone
[Ozone] (ppm)
0.24
0.20
0.16
0.12
0.08
0.04
0.00
0
4
8
12
16
20
Hour
Figure 4.7.36. Ozone Time Series (1.5x Base Case ROG) (8/27/87), Riverside
RIVR
FAC_ROG
FAC_NOX
3000
kXYLHO
kOLEO3
Ozone Variance (ppb 2)
2500
kOLEHO
kETHO3
kETHHO
2000
kALKHO
kPAN
kACO3NO2
1500
kACO3NO
jHCHOM
1000
jHCHOR
kHONO2
kOSDM
500
kOSDH2O
jO3O1D
kO3NO
23
21
19
17
15
13
11
9
7
5
3
1
0
jNO2
Hour
Figure 4.7.37. Ozone Variance Time Series (1.5x Base Case ROG) (8/27/87), Riverside
135
4.7.19 Conclusions for Inverse Modeling and Uncertainty Analysis
The goal for the work on inverse modeling and uncertainty analysis was to demonstrate a
proof of concept for the notion that measurement data could be used to infer the form of the
emission distribution. That object has been successfully achieved. More importantly, the new
algorithms for inverse modeling and uncertainty analysis now mean that it is feasible to perform
such calculations for realistic computational domains and typical urban areas. The
instrumentation systems that can provide high spatial, temporal and chemical resolution together
with the new inverse methods offer the promise of finally solving the “emission inventory
problem.” The techniques described above are currently being applied to mobile laboratory
mapping data from Boston and Mexico City.
136
4.8 Model Inversion of Pollutant Maps: Diffusion Modeling of SF6 Release Experiments
4.8.1 Introduction
Modeling of a tracer gas allows for the selection of a diffusion model for more complex
systems, such as those that include species that react in the atmosphere. The study of the
diffusion of an inert gas allows one to distinguish the gas source: either the normal background
concentration or levels produced by anthropogenic activities. By demonstrating methodologies
to accurately model plume dispersion, it is shown that model parameters can be estimated. Thus,
for example, the emission rate of a pollutant from a point source can be estimated and compared
with emissions inventory data.
Release experiments were performed by Aerodyne Inc. and Washington State University
(WSU) on August 25th 1999 in Boston. The tracer used was sulfur hexafluoride (SF6), which
was released from a stationary 2m-high port. Mobile data was collected, which include NO,
NO2, O3, CO2, and SF6 concentration measurements, as well as position measurements from
GPS. The tracer was release at a constant flow rate from a stationary location; two mobile labs
then measured the concentrations at different wind and crosswind directions.
4.8.2 Observed Phenomena
The phenomenon observed in the atmosphere as the gas is released is steady state
diffusion, where molecular diffusion is negligible compared to turbulent diffusion. Since SF6 is
inert, no gas-phase reactions need to be considered. Under steady-state conditions, turbulent
diffusion in the mean wind direction is usually neglected (slender plume appoximation). The
atmospheric diffusion equation is further simplified by assuming that horizontal homogeneity
exists, i.e., the mean wind u and lateral eddy diffusivity Kyy are independent of y. The SF6
release was a constant, stationary point source. Therefore, it is expected that the highest
concentration is in the centerline of the plume in the vertical and horizontal directions, as shown
in Figure 4.8.1. Thus it is expected that the peak observed concentration will decrease and the
plume will spread out as it is transported downwind of the release site.
It is well known that the configuration of the urban area where the release was performed
will create a departure from the expected Gaussian distribution. Mixing effects or increasing
turbulence in intersections, as well as channeling occurs in an area that consists of a series of
buildings and residential houses. A derivation of the model used here, along with the inherent
assumptions, is presented in Section 4.8.4.
137
Figure 4.8.1. Continuous point source.
4.8.3 Experimental Data
Wind Data
The wind field was not measured in the Mobile Lab. Earlier work (Duchini 2001) used
the wind data from Logan Airport as the best approximation for the wind at the release site. The
meteograms used for the Logan Airport meteorology were obtained from the University of
Wyoming and show the wind speed and direction as an hour average. Logan Airport, however,
is a significant distance from the release site (several kilometers), and is located directly on the
coast and will therefore likely experience different winds than the release site, an urban location
approximately two kilometers inland from the coast. Moreover, the dispersion model is very
sensitive to the accuracy of the wind field.
A more accurate description of the wind fields at the release site can be obtained from
MM5 predictions. O’Neil et al. [2001] applied the MM5 meteorological model to simulate the
wind field for the Northern New England coast for May 24-26, 1999. A four-nested domain was
utilized where grid resolutions were 27 km, 9km, 3 km, and 1 km. The 1 km domain was used in
this study.
O’Neill et al. (2001) compared the MM5 predictions to wind data taken from a sodar
operating near the Massachusetts Institute of Technology and found reasonable agreement for
both the wind direction and speed. In order to further validate the MM5 predictions, model
results for Logan Airport are compared here with available data. Figure 4.8.2 and 4.8.3 compare
the wind speed and direction for the duration of the release experiment. It should be noted that
the winds reported at Logan are from a 10-meter anemometer. Since the MM5 vertical grid
resolution does not coincide with 10 meters, the average 10-meter wind speed was calculated by
regressing the predicted wind profile according to the power law (see below):
u( ~
z )  az~ p
138
(4.8.1)
where ~z is the normalized height z – zo, where zo is the surface altitude (5.8 m), z is the altitude,
a is the wind speed at ~z = 0, and p is a parameter determined empirically. It should also be
noted that the lowest grid point in the MM5 domain for the period and location of interest is
approximately 50 m. In contrast to the wind speed, there is no easy way of correlating the
variation of wind direction with height; thus the predicted wind direction shown in Figure 4.8.3
is the wind direction at ~ 50 m. Figures 4.8.2 and 4.8.3 show that on average MM5 does a good
job of predicting the winds at Logan Airport for the period of interest.
13
12
Measured u 10
u10 [m/s]
11
10
9
Predicted u 10
8
7
6
15:00
16:55
18:50
20:45
Time [GMT]
Figure 4.8.2. Predicted v. measured 10-meter wind speed at Logan Airport on 5/25/99
Wind Direction [degrees]
235
230
Measured wind direction
225
220
215
210
205
Predicted wind direction
200
195
190
15:00
16:55
18:50
20:45
Time [GMT]
Figure 4.8.3. Predicted v. measured wind direction at Logan Airport for 5/25/99
139
Predicted Wind Fields at Release Site
Before presenting the MM5 predictions for the release site, some explanation of the
handling of the MM5 output file will be presented. The converter program gradsv3.deck
(available at http://www.ems.psu.edu/~bryan/mm5/grads/#conv) was used to convert the MM5
binary output file to a .ctl and accompanying .dat file that can be visualized using GRADS
(available at http://www.iges.org/grads/). GRADS was used for the data visualization and
extraction.
MM5 usually obtains and analyzes its data on pressure surfaces, but these have to be
interpolated to the model’s vertical coordinates. The vertical coordinate is terrain following,
meaning that the lower grid levels follow the terrain while the upper surface is flat. The model
levels are in terms of sigma levels, where each sigma level is defined by:

-ptop)/(po-ptop)
(4.8.2)
where p is the pressure, ptop is a specified constant top pressure, and po is the surface pressure.
Thus  is zero at the top and one at the surface. Note that in defining the  levels, full levels are
listed, but the values of the wind speed and direction are defined in the center of the cell, i.e., at
the  half-level.
The converter program gradsv3.deck interpolates from MM5’s sigma levels to pressure
levels (in mB). From the pressure level, the height (Z) of the data point can then be calculated
using the nonhydrostatic dyamic equation:
F
IJ  RT F
p I
ln
G
G
H K g Hp J
K
RA
p
Z
ln o
2g
poo
2
so
o
(4.8.3)
oo
where po is again the pressure at the surface, poo is sea-level pressure, R is the universal gas
constant, and Tso is the base temperature in the reference state idealized temperature profile
equation:
To  Tso  A log
F
p I
G
Hp J
K
0
(4.8.4)
oo
where A is a measure of lapse rate and taken to be 50 K. A value of 290 K for Tso was used; this
value is typical for midlatitude summer conditions. Both the surface and sea-level pressure were
taken to be 1003 mB.
While the use of MM5 results allows for a prediction of the wind field that is much closer
to the release site than the Logan Airport data, the wind field at the exact release site is not
available. Figure 4.8.4 shows a map of the release site (labeled SF6 release) and the four closest
grid points for which data is available (labeled wind location 1 through wind location 4). An
average of the four grid points was used to calculate the release site wind fields.
140
Figure 4.8.4. SF6 Release site and adjacent four grid points for which wind fields are available.
Note that the elevation of the release site is 31 m, which is high relative to the
surrounding area.
The mean wind speed was represented by the power-law function of height:
u( ~
z )  az~ p
(4.8.5)
where zo, the altitude of the release site, is 31 m. Figure 4.8.5 shows an example of the fit
obtained at the release site at 16:00 EDT.
141
2.38
2.36
ln(u)
2.34
2.32
2.3
y = 0.0871x + 1.9223
2.28
R2 = 0.9499
2.26
2.24
3.5
4
4.5
5
5.5
ln(z)
Figure 4.8.5. Power-law fit for the average wind speed at the SF6 release site for 16:00 GMT on
5/25/99. The functional form of the fit is ln( u ) = pln( ~z ) + ln(a).
Table 4.8.1 shows the wind field parameters for the duration of the release experiment.
Table 4.8.1 - Release Site Wind Field Parameters
Boston
Time
12:00 PM
1:00 PM
2:00 PM
3:00 PM
4:00 PM
5:00 PM
GMT
16
17
18
19
20
21
dir
227.3
225.0
226.7
227.6
209.4
208.7
142
a [m/s]
9.9
8.5
7.5
7.5
6.8
9.2
p
0.030
0.051
0.076
0.063
0.087
0.026
Data Handling
Figure 4.8.6 below shows the route taken by the Aerodyne van (blue) and the WSU van
(yellow). The SF6 release site is depicted in yellow. The winds were predominantly from the
southeast. Figure 4.8.7 shows a detail of the data set, focusing on the area close to the release
site.
Figure 4.8.6. Route for the Aerodyne (blue) and WSU mobile lab (yellow), and the SF6 release
site (red) for 5/25/99. The Aerodyne van terminated in Billerca at Aerodyne
Research Inc. headquarters.
143
Figure 4.8.7. Detail of the route followed by the two mobile labs. The average wind direction
is from the southwest and is depicted in red.
The next section describes the data processing that was performed to convert the data in a
format compatible with the dispersion model.
Data handling
The data is reported as time (GMT), longitude (degrees), latitude (degrees), altitude (m),
and SF6 concentration (ppt). Position variables obtained by GPS were transformed to rectangular
coordinates (m) in ArcView. The conversion was performed using the GetX and GetY scripts
and a Lambert Conformal Conic projection based on the 1983 State Plane. Given the rectangular
coordinates, the data points are shifted so that the release site is the new origin (0,0,0).
Table 4.8.2 lists the GPS coordinates of the release site and the corresponding rectangular and
shifted rectangular coordinates.
144
Table 4.8.2 - SF6 Release Site Coordinates
Longitude [o] -71.0751
Latitude [o] 42.3098
Altitude [m]
31
x [m] 235030.7
y [m] 895566.8
z [m]
31
xo [m]
yo [m]
zo [m]
0
0
0
The diffusion model assumes that the mean wind direction is along the x-axis, so the
resulting rectangular coordinates were rotated according to the wind direction as shown in
Figure 4.8.8. The angle  is the deviation from the west, which corresponds to 270o, forcing the
wind to flow from the negative x-axis.
y - yo
(y1-y0)



~
y1
~
x1

(x1-x0)
x-xo

Figure 4.8.8.
Shifted Coordinates using the wind direction, where xo, yo are the coordinates of
the SF6 release site.
By convention, a reported wind direction of 270 is wind coming from the west so that:
  ( 270  wind direction )
(4.8.6)
The final rectangular coordinates that agree theoretically with the slender plume
approximation are:
~
x
~
y
bx  x g by  y gcosF
G
Htan
bx  x g by  y gsinF
G
Htan
2
o
2
2
o
1
o
2
o
Fby  y gI   I
G
Hbx  x gJ
KJ
K
Fby  y gI   I
G
Hbx  x gJ
KJ
K
o
(4.8.7)
o
1
o
(4.8.8)
o
where xo, yo are the rectangular coordinates of the SF6 release site and the final (shifted and
x and ~y . The wind direction values given in Table 4.8.1 were used to
rotated) coordinates are ~
perform the coordinate rotation on an hourly basis.
145
Finally, the SF6 concentration is converted to mass concentration:


P M SF6
C   g 3  
 110 6  C  ppt 
m 

RT
(4.8.9)
where C is the SF6 concentration, P is the atmospheric pressure assumed at 1 atm, and T is the
atmospheric temperature (292 K).
Aerodyne Data
The raw Aerodyne data consisted of 19,651 data points, starting at 14:34 GMT
(10:34 EDT) and ending at 22:48 GMT (18:48 GMT). Thus the data collection started prior to
the beginning of the release experiment (approximately 12:15 EDT) and continued past the end
of the release experiment (approximately 17:49 EDT). First the data points that reported an SF6
concentration of zero and NAN (not a number) were removed, as well as all the data points prior
to 18:00 GMT (prior to the beginning of the release experiment). Next, all the data points below
the detection limit (10 ppt) were removed, reducing the data set 1,535 points. In addition, points
that were upwind of the release site and more than seven kilometers from the release site were
excluded. Figure 4.8.9 shows the reduced data set, and Figure 4.8.10 shows the rotated reduced
data set. The reduced data set has 1,460 data points, starting at 18:14 GMT and ending at
20:18 GMT.
It should be pointed out that concentrations above the detection limit were detected for
numerous points upwind of the release site. As previously mentioned, these points were
excluded from the analysis: a Gaussian plume model cannot predict concentrations upwind of
the source. A possible explanation for these points can be found from past measurement
campaigns, where small bursts of SF6 above the detect limit were detected in industrial areas,
likely from leaking high voltage electrical equipment where it is used to control arcing.
Moreover, it should be emphasized that the dispersion model assumes an hourly-averaged wind
direction (and speed). However, the wind can gust and rapidly change direction, effects that are
not captured well in an hour-averaged model, while the data are instantaneous. It is quite
possible that instantaneous wind directions could vary by a substantial degree from the average.
146
4500
Y - Y0 [m]
3500
2500
average wind direction
1500
500
-500
-200
0
200
400
600
800
1000
1200
1400
1600
X-X0 [m]
Figure 4.8.9. Reduced data set for Aerodyne data.
2000
Point 94
1500
Points 408, 409
Ytilda [m]
1000
500
hour-average
wind direction
0
0
200
400
600
800
1000
1200
1400
1600
-500
Xtilda [m]
Figure 4.8.10. Rotated and reduced data set from Aerodyne
147
1800
2000
The points circled in red (labeled 94, 408 and 409) were included in the regression analysis.
The model was also regressed excluding these three data points. See Section 4.8.5.
WSU Data
The WSU data starts at 12:29 EDT (16:29 GMT) and ends at 17:49 EDT (21:49 GMT),
and has 866 points. All data points that were below the detection limit were rejected.
A detection limit of 10 ppt was used for both data sets [Lamb, 1995]. In addition, points that
were upwind of the release site and more than seven kilometers from the release site were
excluded. The final WSU data set was 391 data points. The shifted and rotated WSU data set
are shown in Figure 4.8.11 and 4.8.12.
7000
5000
Y - Y0 [m]
average wind direction
3000
Above 10 ppt
Above 50 ppt
Release Site
1000
-300
200
700
1200
1700
2200
-1000
-3000
X-X0 [m]
Figure 4.8.11. Shifted WSU data
148
2700
3200
3700
3000
2000
1000
hourly averaged
wind direction
0
Ytilda [m]
0
1000
2000
3000
4000
5000
6000
7000
-1000
-2000
Point 2
-3000
-4000
Point 1
-5000
Xtilda [m]
Figure 4.8.12. Final WSU rotated and shifted data set. The hourly-averaged wind direction
coincides with the x-axis.
Summary
Table 4.8.3 summarizes the relevant parameters of the measurement campaign. The
height of the release port h is 2 m. Tracer concentrations were measured in parts-per-trillion
(ppt), and have an error of approximate 10% with a minimum detection limit of 10 ppt
[Lamb, 1995]. The tracer release rate was determined from sequential dry gas meter readings.
The estimated variability and accuracy in the release rate is less than  5% (WSU Draft Report).
Note that the data taken on May 26th was not used due to failure of the SF6 pump.
Table 4.8.3 - Summary of Parameters for Boston 05/25/99
Parameters
Predicted Average wind speed (m/s)
Predicted Average wind direction (o)
Predicted Average Temperature (oF)
SF6 flow rate (g/min)
Release port height (m)
Release Latitude (o)
Release Longitude (o)
Release Altitude (m)
Value
7.3
220
66
44.22
2
42.3098
-71.0751
31
The complete data set used for diffusion modeling is shown in Figure 4.8.13.
149
3000
2000
1000
0
Ytilda [m]
0
1000
2000
3000
4000
5000
6000
7000
-1000
WSU Data
Aerodyne Data
-2000
-3000
-4000
-5000
Xtilda [m]
Figure 4.8.13. Combined rotated and shifted final WSU and Aerodyne data.
4.8.4 Model Description
The species conservation equation is:
CA
    N A  RV , A  SV , A
t
(4.8.10)
where CA is the concentration of species A, the source terms RV,A and SV,A represent respectively
the rate of formation by chemical reaction and rate of addition of species A, per unit volume, and
 is the gradient operator. NA is the total molar flux of A relative to fixed coordinates and is
given by the sum of a convective and diffusive term:
NA = uCA + JA
(4.8.11)
where u is the wind vector (u,v,w) and JA is the molar diffusive flux of species A. For the case of
a dilute tracer gas in the atmosphere, Fick’s law is an excellent approximation to the diffusive
flux:
J A   DACA
150
(4.8.12)
where DA is the molecular diffusivity of the dilute species A in the carrier fluid (air). Thus for a
gas mixture where a single species is present in small concentrations, and the density and
diffusivity is constant, the conservation equation for the minor component A may be written as:
CA
   uCA  DA2CA  RV , A  SV , A
t
b g
(4.8.13)
where  2 is the Laplacian. SF6 is inert (RV,A = 0), so that for the steady state problem:
 uCA  DA2CA  SV , A
(4.8.14)
In the atmosphere, the wind field u is expected to be turbulent. Turbulent flows are
irregular and random, so that the velocity components at any location vary randomly with time.
Following the approach of Seinfeld and Pandis (1998), the instantaneous velocity component can
be represented as the sum of a mean ( u ) and fluctuating component (u’):
u  u  u'
(4.8.15)
Similarly, the instantaneous concentration of species A can be represented by:
CA  CA  CA '
(4.8.16)
where by definition the mean of a fluctuating quantity is zero (i.e., CA '  u'  0 ). Using
Equations (4.8.15) and (4.8.16) in Equation (4.8.14) and averaging over an infinite ensemble of
realizations of the turbulence yields (Seinfeld and Pandis 1998):
d i d
i
 u CA   u' CA '  DA2 CA  SV , A
(4.8.17)
The most common way of relating the turbulent fluxes u'CA ' to CA is through K theory:
u' CA '   K   CA
151
i = 1,2,3
(4.8.18)
where K is the eddy diffusivity tensor. Only the diagonal entries of K are non-zero when the
coordinate axes coincide with the principal axes of the eddy diffusivity tensor. Thus in Cartesian
coordinates Equation (4.8.18) becomes:
 CA
x
 CA
v ' CA '   K yy
y
u' CA '   K xx
w' CA '   K zz
(4.8.19)
 CA
z
Combining Equations (4.8.18) and (4.8.19) yields (in Cartesian coordinates):
d i  dv C i  dw C i   FK  C I   FK  C I   FK  C I 
x
y
z
x G
H x J
K y G
H y J
K z G
H z J
K
L C   C   C O
D M
(4.8.20)
P S
M
Nx y z P
Q
 u CA
A
A
A
A
xx
2
2
A
2
A
A
yy
zz
2
A
2
A
V ,A
2
In order to make Equation (4.8.19) tractable, two further assumptions are invoked:
1. Molecular diffusion is negligible compared with turbulent diffusion:
DA
 2 CA
xi
2

b g
F
G
H
 CA
 ui ' CA '


Kii
xi
xi
xi
I i =1,2,3
J
K
2. The atmosphere is incompressible:
u  v  w


0
x y z
With these two assumptions, Equation (4.8.20) becomes:
u
 CA
x
v
 CA
y
F
G
H
w
 CA

K xx
x
x
 CA
z

I   FK  C I   FK  C I  S
J
K y G
H y J
K z G
H z J
K
A
yy
A
zz
Equation (4.8.21) is the atmospheric diffusion equation.
152
V ,A
(4.8.21)
In order to find an analytic solution to Equation (4.8.21) further simplifications must be
made. The slender plume approximation allows one to neglect turbulent diffusion in the
direction of the mean wind (since it is expected to be small relative to the convective term). This
approximation holds identically when the wind direction coincides with one of the principal
axes, i.e., the reference frame was rotated such that the wind direction coincides with the x-axis
so that Kxx = 0. Moreover, in the rotated reference frame v = 0, and since u  w Equation
(4.8.21) can now be written as (where, using the same notation as Section 4.8.3, the tilda denotes
the shifted, rotated reference frame):
u
F
G
H
 CA
 CA

 ~ K yy ~
~
x
y
y
I   FK  C I  S
J
K ~z G
H ~z J
K
A
zz
V ,A
(4.8.22)
where the source term SV,A can be expressed as:
SV , A  q ( ~
x ) ( ~
y ) ( ~
z  h)
(4.8.23)
where q is the source strength (g/s),  is the Dirac delta function, and h is the height of the
release port above the ground (m). The mobile experiments released the tracer at a port height h
= 2 m. In summary, Equation (4.8.22) describes the steady-state diffusion from an elevated point
source of an inert species in a turbulent shear flow.
An analytic solution of Equation (4.8.22) can be obtained by expressing the lateral eddy
diffusivity as the rate of change of the lateral mean square particle displacement (Huang 1979):
1 d y
K yy 
2 dt
2
(4.8.24)
By assuming flow with homogeneous turbulence in the lateral direction at each horizontal
plane Taylor’s hypothesis can be invoked, and thus Equation (4.8.24) becomes [Huang, 1979]:
K yy 
1 d y
u
2 dx~
2
(4.8.25)
The mean wind speed and vertical eddy diffusivity may be represented by a power law
with respect to height:
u( ~
z )  a~
zp
(4.8.26)
bg
(4.8.27)
K zz ~
z  b~
zn
153
For a source near the ground (i.e., h in Equation (4.8.23) is small), and assuming
Equations (4.8.25) – (4.8.27) hold, the solution to Equation (4.8.22) is given by [Huang, 1979]:
b g
C ~
x, ~
y, ~
z 
q
2 y ~
x
1 p

F ~y I Fuc~z  h hI
expG JexpG
J
b ~
x
F1  p I 2
K
u c
b h  G J H K H
H K

2

2
1 p
2
y



2
(4.8.28)
where  and  are defined as:
  2 pn

(4.8.29)
1 n
(4.8.30)

The lateral standard deviation y is a function of downwind distance; the plume will
spread out as it is transported downwind, i.e., the lateral standard deviation increases with
downwind distance. Based on the recommendations of [Briggs, 1974], the functional form of the
lateral standard deviation for urban conditions can be approximated as:
b
 y  sy ~
x 1  0.0004 ~
x
g
1
2
(4.8.31)
The parameter sy depend on the wind stability class, and Briggs (1974) found that sy varied
between 0.11 (very stable) to 0.32 (very unstable).
Least Squares Minimization
Using Equation (4.8.28), the distribution with position of SF6 concentration downwind of
the release site can be calculated based on four parameters: the release source strength q, the
lateral standard deviation parameter sy, and the vertical eddy diffusivity parameters b and n (see
Equation (4.8.27)). These parameters were estimated using least squares minimization, i.e.,
c
Minq ,s y ,b,n  CAmeasured
 CApredicted
,i
,i
h
2
(4.8.32)
i
In order to judge model performance, the predicted and measured concentrations are
compared, and the values of the best-fit parameters compared with those predicted from theory.
4.8.5 Results and Discussion
The least squares minimization was run numerous times using different data sets in order
to investigate the sensitivity of the results to the data included in the minimization. Table 4.8.4
summarizes the characteristics of the different data sets utilized.
154
Table 4.8.4 – Summary of Utilized Data Sets
Name
WSU_Set1
WSU_Set2
WSU_Set2
ARI_SF6
ARI_SF6v2
Combined1
Combined2
Combined3
Combined4
Figure
4.8.12
4.8. 12
4.8.12
4.8.10
4.8.10
4.8.13
4.8.13
4.8.13
4.8.13
Comments
All of WSU’s data, total of 391 data points
Removed point 1 in 4.8.12
Removed points 1 and 2 in Figure 4.8.12
All of Aerodyne’s data, total of 1460 data points
Removed points 94, 408, and 409 in Figure 4.8.10
ARI_SF6+WSU_Set1
ARI_SF6v2 + WSU_Set3
Combined2 – all points with SF6 < 50 ppt
Combined3 – all points more than 5 km downwind of site
Figure 4.8.14 shows the dataset Combined3 as plotted using the Matlab interpolating
function ‘griddata’. Figure 4.8.14 clearly shows a secondary peak in the positive lateral
direction; the model used will be unable to capture this peak because it predicts the highest
x to occur at ~y = 0, i.e., the predicted peak concentration with
concentration at a given ~
downwind distance will be at the centerline of the plume. This secondary peak indicates that the
assumed wind direction may be off; specifically, that the assumed wind angle may be too large.
This will be investigated in Section 4.8.4.
Comparing Figure 4.8.13 with Figure 4.8.14 shows there is an apparent bias in terms of
where the measurements were taken: more measurements were taken in the positive lateral
direction (i.e., ~y > 0). Figure 4.8.7, however, shows that data was collected in both the positive
and negative lateral directions. The reason this data does not appear in Figure 4.8.14 is that it
was below the instrument detection limit, thus supporting the hypothesis that the wind direction
may be off.
The Matlab function ‘griddata’ had to be used to plot the data due to the high amount of
scatter in the dataset, i.e., small changes in position often showed a significant change in
measured SF6 concentration. This high amount of scatter may be due to a variety of sources,
e.g., rapid fluctuations in wind speed and direction that are not captured due to the ensemble
averaging of the model (i.e., going from Equation (4.8.14) to Equation (4.8.17)), mixing effects
or increasing turbulence in intersections, as well as channeling that occurs in an area that consists
of a series of buildings and residential houses. While it is possible that the data scatter is due to
measurement noise, it is unlikely that the instrument quality is so poor and therefore likely that
the observed scatter is real. The high amount of scatter in the data also points to the fact that the
resolution of the data is likely higher than the model resolution (the real-time data collected by
Aerodyne was collected ~ 1 s-1).
155
Figure 4.8.14. Measured SF6 concentrations using the dataset Combined3.
Figure 4.8.15 shows the dataset ARI_SF6. Comparing Figures 4.8.14 and 4.8.15, it is
clear that the secondary peak off the centerline shown in Figure 4.8.14 is due to the WSU data.
The prediction of higher concentrations occurring along the plume centerline is reflected more
clearly in Figure 4.8.15. However, still evident in Figure 4.8.15 is that most of the measurements
remain in the positive lateral direction.
For the model parameter regression, two key questions must be considered:
1.
What is the ‘best’ data?
2.
What is the ‘best’ fit?
The importance of the first question should be readily apparent by comparing Figures 4.8.14
and Figure 4.8.15 and the discussion in Section 4.8.2. The second question highlights the
importance of carefully considering the method selected to perform the parameter regression.
Specifically, the selection of the objective function to be minimized affects the value of the bestfit parameters. Thus there is an appreciable difference between minimizing the sum of the
square of the residuals versus minimizing the sum of the absolute values of the residuals, i.e.,
instead of non-linear program (NLP) being Equation (4.8.32) the NLP is:
Minq ,s y ,b,n 
cC
measured
A ,i
i
156
 CApredicted
,i
h
2
(4.8.33)
Figure 4.8.15. Measured SF6 concentrations using the dataset ARI_SF6 data.
Tables 4.8.5 and 4.8.6 show the values of the best-fit parameters for the different data
sets using Equation (4.8.32) (i.e., the objective function is the sum of the squares) and
Equation (4.8.33) (i.e., the objective function is the sum of the absolute values of the residuals).
Table 4.8.5 - Best-fit parameters using Equation (4.8.32)
Data Set
ARI_SF6
ARI_SF6v2
Set1
Set2
Set3
Combined1
Combined2
Combined3
Combined4
q [g/s]
352870
352810
247870
247680
247680
312040
311990
376030
375840
n
1.3947
1.3947
1.8150
1.8152
1.8152
1.4150
1.4151
1.1852
1.1857
sy
0.6741
0.6742
0.1838
0.1838
0.1838
0.5748
0.5748
0.7511
0.7510
157
b [m2/s]
0.0468
0.0468
0.0871
0.0871
0.0871
0.0483
0.0482
0.0363
0.0363
Objective Function
46098
46098
2458.5
2458.5
2458.5
49515
49515
42262
42261
Table 4.8.6 - Best-fit parameters using Equation (4.8.33)
Data Set
ARI_SF6
ARI_SF6v2
Set1
Set2
Set3
Combined1
Combined2
Combined3
Combined4
q [g/s]
513020
513630
145940
145970
145970
439270
439170
533720
533380
n
1.6107
1.6131
1.7849
1.7853
1.7853
1.6252
1.6253
1.5224
1.5228
sy
0.6772
0.6725
0.1204
0.1203
0.1203
0.5573
0.5573
0.7516
0.7515
b [m2/s]
0.0628
0.0631
0.0812
0.0812
0.0812
0.0652
0.0652
0.0531
0.0531
Objective Function
5743.3
5743.0
333.9111
333.8455
333.7818
6195.9
6195.6
5367.9
5364.6
Tables 4.8.5 and 4.8.6 illustrate that including what appears to be outliers in the
individual data sets has a small effect on the value of the parameters estimated, i.e., the
difference between the results for ARI_SF6 and ARI_SF6v2 is small, as is the difference
between Set1, Set2, and Set3, and Combined1 and Combined2. Tables 4.8.5 and 4.8.6 also
clearly show a significant difference between the parameter estimates based on the Aerodyne
versus the WSU mobile laboratory data. The Aerodyne dataset is a richer dataset – it has more
than three times the number of data points than the WSU dataset. However, to my knowledge
there is no reason to believe that one dataset is more reliable than the other. It is therefore
unclear why there is such a large difference between the two datasets. Moreover, the difference
between the best-fit parameters for Combined1 and Combined2 versus Combined3 and
Combined4 indicates that the data points below 50 ppt (but above 10 ppt) may be unreliable, and
that a detection limit of 50 ppt may be more appropriate than 10 ppt.
Contrasting Table 4.8.5 and 4.8.6, it is clear that the selection of objective function has a
significant affect on the value of the best-fit parameters. For example, the predicted source
strength q is always higher when Equation (4.8.33) is used. The question is what is a ‘better’ fit?
Figures 4.8.16 and 4.8.17 show respectively the measured versus predicted SF6 concentration for
the Combined3 dataset using the best-fit parameters based on Equation (4.8.32) and (4.8.33).
Comparing the two figures it is not readily apparent that one approach is ‘better’ than the other.
Judging by the different scales of the two figures, however, it appears that using
Equation (4.8.32) tends to underpredict the higher SF6 measurements. One important feature that
can be observed in both Figures 4.8.16 and 4.8.17 is the large amount of scatter from the ideal y
= x line of the measured vs. predicted scatter plots. This feature was consistently evident in all
the datasets and regression methods used, and reflects the high amount of scatter in the data.
158
Figure 4.8.16. Measured vs. predicted SF6 concentration for the Cominbed3 dataset using the
best-fit parameters of Equation (4.8.32).
Figure 4.8.17. Measured vs. predicted SF6 concentration for the Cominbed3 dataset using the
best-fit parameters of Equation (4.8.33).
159
Based on the high amount of scatter in the data, concerns about the detection limit of the
SF6 analyzer, and the significant differences between the predictions based on only the Aerodyne
versus the WSU data (and the fact that there is no reason to believe one over the other), the
dataset Combined3 was selected as the most reliable dataset. For the parameter regression so far,
an average value of p was used when calculating  in Equation (4.8.29). The regression was
performed again, but this time the hourly value of p was used when calculating . Table 4.8.7
shows the values of the resultant best-fit parameters, for both the case where the objective
function is the sum of the square of the residuals (Equation (4.8.32)) and the sum of the absolute
value of the residuals (Equation (4.8.33)).
Table 4.8.7 - Best-Fit Parameters Using the Dataset Combined3 and
Allowing  to Vary Hourly
Equation
(4.8.32)
(4.8.33)
q [g/s]
388420
582180
n
1.1836
1.5144
sy
0.7790
0.8500
b [m2/s]
0.0356
0.0508
Objective Function
41976
5335.2
Figure 4.8.18 shows the predicted SF6 concentration using the parameters calculated
using Equation (4.8.33) and an average wind speed of 7.2 m/s and p = 0.077. As can be seen in
Figure (4.8.14) the shape of the predicted distribution is symmetric about the ~y = 0 line. The
peak predicted SF6 concentration of 20.2 g/m3 agrees well with the maximum observed SF6
concentration of 22.7 g/m3. If the parameters estimated using Equation (4.8.32) are used
instead, the maximum predicted SF6 concentration is 17.8 g/m3.
While Figure 4.8.18 compares reasonably well with the measured data
(see Figure 4.8.14), it is critical to compare how the best-fit parameters compare with theory.
Table 4.8.8 compares the best-fit parameters used in Figure 4.8.18 to those predicted by theory.
The source strength was 44.22 g/min = 737 mg/s. The predicted source strength was 582 mg/s;
thus while the predicted source strength is underestimated, given the model error that is likely
present due to the effect of traffic, buildings, etc. on the plume dispersion, the agreement is good.
Based on diffusion data curves, the lateral standard deviation parameter sy was estimated by
Briggs [1974] for urban conditions to vary between 0.11 and 0.32, depending on the atmospheric
stability class. Given that the 10m surface wind speed for the site was predicted to be greater
than 6 m/s and that the release experiment was performed during the summer where strong solar
radiation can be expected, the Pasquil stability class is likely to be C (slightly unstable). The
standard deviation parameter should therefore be order 0.2. For a slightly unstable atmosphere,
we expect the Monin-Obukhov to be negative. The roughness length of the terrain should be of
order 5 m [McRae et al., 1982]. Based on the Pasquil stability class, the surface roughness, and
the estimated wind power-law parameter p (see Table 4.8.1) the Monin-Obukhov length L is
likely to be ~ -1 m. Huang [1979] argues that the diffusion parameter n can be estimated using
the conjugate power law or from the Monin-Obukhov similarity theory. Using the conjugate
power law, n ~ 1 – p,
160
Figure 4.8.18. Predicted concentration of SF6 assuming an average wind speed of 7.2 m/s and p
= 0.077. The parameter values used are: q = 582 mg/s, n = 1.5144, sy =
0.8318, and b = 0.0508. Note the symmetry of the predicted plume in the lateral
direction.
and thus for the conditions in Boston n ~ 0.9. Using Monin-Obukhov similarity theory, and
assuming L ~ -1 m, n is predicted to be ~ 1.2. To my knowledge there is no easy way of
estimating the vertical eddy diffusivity parameter b.
Table 4.8.8 - Theory v. Best-Fit Parameters
Parameter
q [mg/s]
n
sy
b [m2/s]
Best-fit
571
1.5152
0.8318
0.0508
Theory
737
~1
~ 0.2
N/A
Data Withholding Exercise
The analysis above considered the available data together, and hence the minimization
results are for an average of all the data points used, and are therefore likely to ‘average out’
some of the observed variance. An alternative approach is to perform the minimization on
individual road sweeps, and then average these results. Figure 5.8.19 shows five roads from the
161
Aerodyne data set that were examined individually, and Figures 4.8.20 – 4.8.24 show a detail of
the route and compare the predicted and observed SF6 concentrations, where the parameters used
are those in Table 4.8.8, i.e., based on all the data. Figure 4.8.25 shows a detail of the route
followed by the WSU van (designated Road_1500_WSU) and the predicted and observed tracer
concentrations. A common theme seen in Figures 4.8.20 – 4.8.25 is that the predicted tracer
concentrations do not capture the variability of the observations.
2000
Ytilda [m]
1500
1000
Road_1500
500
Road_1000
Road_250
0
0
500
Road_100
1000
1500
2000
Road_600
-500
Xtilda [m]
Figure 4.8.19. Individual road segments considered from the Aerodyne dataset.
162
20
25
Observed SF6
Predicted SF6
0
0
100
200
300
20
400
SF6 [ug/m3]
-20
Ytilda [m]
-40
-60
15
10
-80
5
-100
-120
0
0
-140
200
Xtilda [m]
400
600
800
1000
Data Point
Figure 4.8.20. Detail of the route for the road labeled Road_100 in Figure 4.8.20 (left) and the observed
and predicted SF6 concentrations (right). We expect the tracer concentration to increase as the mobile lab
approaches the plume centerline, then decrease as it travels downwind of the release site (approximately)
along the plume centerline. When the road turns away from the plume centerline, the predicted tracer
concentration decreases with distance from the centerline, until the road turns and approaches the release
site, at which point the lateral distance continues to increase while the downwind distance decreases.
Note the high amount of scatter in the data.
25
300
2:14 PM (EDT)
3:33 PM (EDT)
SF6 Concentration [ug/m3]
250
Ytilda [m]
200
150
100
50
0
200
-50
250
300
350
20
15
2:14 PM
3:33 PM
10
Predicted SF6
(3:33 PM)
Predicted SF6
(2:14 PM)
5
400
0
-100
0
Xtilda [m]
100
200
300
400
Position [m]
Figure 4.8.21. Road_250 in Figure 4.8.19. Note that there were two passes along the road, one at 2:14
pm (EDT) and 3:33 pm (EDT); the observed tracer concentrations (right) show good consistency. As can
be seen from Table 4.8.1, there is a negligible difference in wind speed between these two times. The
difference between the predicted SF6 concentrations for 2:14 PM and 3:33 PM is due to the difference in
stability parameter p (0.063 v. 0.076). Also, note that as the road travels further downwind and closer to
the plume centerline, there is an observed decrease in tracer concentration, while the predicted tracer
concentration remains quite high. This observation further supports the hypothesis that the assumed wind
direction is off. The excellent agreement between the 2:14 and 3:33 pm data also raises the question of
whether the results can be expected to be highly consistent, whether the scatter in Figure 4.8.20 or good
agreement in Figure 4.8.21 represents more typical results.
163
100
8
Observed SF6
Predicted SF6
7
50
Ytilda [m]
6
0
600
5
800
1000
1200
4
-50
3
2
-100
1
-150
0
Xtilda [m]
0
50
100
150
200
Figure 4.8.22. Road_600
600
6
2:19 PM
2:59 PM
500
2:19 PM
2:59 PM
3:39 PM
Predicted SF6
Series5
Series6
5
3:39 pm
SF6 [ug/m3]
Ytidla [m]
400
300
200
100
0
900
-100
4
3
2
1
1000
1100
1200
1300
0
0
-200
200
400
600
800
Xtilda [m]
Xtilda [m]
Figure 4.8.23. Road_1000 in Figure 4.8.19. Note that there are three passes along the road that
peak and show low tracer concentrations at different points, and that the predicted
SF6 concentrations are a good average of the three passes.
164
900
3
Observed SF6
800
600
SF6 [ug/m3]
2
500
Ytilda [m]
Predicted SF6
2.5
700
400
300
200
1.5
1
100
0.5
0
-1001600
1650
1700
1750
1800
0
0
-200
50
Xtilda [m]
100
150
200
250
Data Points
Figure 4.8.24. Road_1500 in Figure 4.8.19. Again, note that the while the predicted tracer
concentration matches the observed concentrations well on average, the predicted
tracer concentration does not capture the variation in observed concentrations.
2
600
1.8
400
1.6
Ytilda [m]
200
0
1000
1.4
1.2
1500
2000
2500
3000
1
-200
Observed SF6
0.8
Predicted SF6
0.6
-400
0.4
-600
0.2
-800
0
Xtilda [m]
0
20
40
60
80
Figure 4.8.25. Road_1500_WSU in Figure 4.8.19. Note that the trend in the observed tracer
concentrations does not match the predicted trend.
165
Table 4.8.9 shows the resultant best-fit parameters for the individual road sweeps. The
maximum predicted source strength q is 5.77 x 109 g/s, four orders of magnitude greater than
the theoretical result. The value of n varies from 2 to –10, while theory predicts n should be of
order unity. The standard deviation parameter sy is expected to be ~ 0.2, while it is found to vary
between 0.3 and 2,600. It is important to understand that while the regressed parameters may
deviate significantly from the values predicted by theory, the deviation between observation and
theory is less than if the parameters of Table 4.8.18 are used (see Figures 4.8.26 – 4.8.29 for a
comparison of observed and predicted tracer concentrations using the parameters of Table 4.8.9).
This means that only using data from an individual road is likely to result in unreliable estimated
model parameters.
There are two likely explanations for this: first, there simply may insufficient data points
to estimate the four parameters; second, data is required over a sufficiently wide range of
downwind and lateral distances as well as times. In order to further investigate this, the
following additional data withholding exercises were performed. The data from the six roads
was combined and the regression performed using only every other, every third, and every fourth
data point from this combined data set. The results of this data withholding exercise are shown
in Table 4.8.10. While there is some variation in the value of the best-fit parameters, this
variation is orders of magnitude smaller than found in Table 4.8.9, indicating that in order to
reliably estimate plume dispersion parameters, data from a range of downwind and lateral
distances as well as times is required. The model uses a mean wind speed and direction, while
the actual wind speed affecting the plume at the time of the traverse could easily be different by a
factor of two or more, and the wind direction could also be significantly different. Thus it is
critical to ensure that the time resolution of the measurements is broad enough to capture the
variability of the wind fields.
Table 4.8.9 - Individual Road Best-Fit Parameters
# Points
Road
Road_100
787
Road_250
79
Road_600
183
Road_1000
214
Road_1500
193
Road_1500_WSU
61
q [g/s]
7.50E+05
2.07E+06
5.77E+09
3.67E+06
2.60E+05
2.67E+06
166
n
1.68
1.98
2.03
1.88
1.91
-9.86
sy
0.93
0.73
2627.20
1.45
0.31
179.90
b [m2/s]
0.07
0.43
0.22
0.11
0.05
0.23
30
Observed SF6
Predicted SF6
25
SF6 [ug/m3]
20
15
10
5
0
0
200
400
600
800
1000
Data Point
Figure 4.8.26. Observed and predicted SF6 concentrations for Road_100 using parameters
regressed using only data from Road_100 (q = 7.50E+05 g/s, n = 1.68, sy =
0.93, b = 0.07 m2/s).
SF6 Concentration [ug/m3]
25
2:14 PM
3:33 PM
20
Predicted SF6
(2:14 PM)
Predicted SF6
(3:33 PM)
15
10
5
0
0
100
200
300
400
Position [m]
Figure 4.8.27. Observed and predicted SF6 concentrations for Road_250 (q = 2.07E+06 g/s, n
= 1.98, sy = 0.73, b = 0.43 m2/s). This figure should be compared with
Figure 4.8.21 (right). Using the new parameters, there now is a significant
difference between the 2:14 and 3:33 pm predicted tracer concentrations, i.e.,
using the new parameters causes the different p values used to have different
predicted tracer concentrations for the two times, despite the fact that it the
predictions are for the same road and very similar wind speeds.
167
6
Observed SF6
Predicted SF6
5
4
3
2
1
0
0
50
100
150
200
Figure 4.8.28. Observed and predicted SF6 concentrations for Road_600 (q = 5.77E+09 g/s, n
= 2.03, sy = 2627.2, b = 0.22 m2/s). Comparing this figure with Figure 4.8.22
(right), it is clear that using the new parameters dramatically improves the data fit.
However, the parameters are orders of magnitude different than predicted by
theory.
6
2:19 PM
2:59 PM
3:39 PM
Predicted SF6 (2:19 PM)
Predicted SF6 (2:59 PM)
Predicted SF6 (3:39 PM)
5
SF6 [ug/m3]
4
3
2
1
0
0
200
400
600
800
Position [m]
Figure 4.8.29. Observed and predicted SF6 concentrations for Road_1000 (q = 3.67E+06 g/s, n
= 1.88, sy = 1.45, b = 0.11 m2/s).
168
3
Observed SF6
Predicted SF6
2.5
SF6 [ug/m3]
2
1.5
1
0.5
0
0
50
100
150
200
250
Data Points
Figure 4.8.30. Observed and predicted SF6 concentrations for Road_1000 (q = 2.60E+05 g/s, n
= 1.91, sy = 0.31, b = 0.05 m2/s).
0.6
Observed SF6
Predicted SF6
0.5
0.4
0.3
0.2
0.1
0
0
20
40
60
80
Figure 4.8.31. Observed and predicted SF6 concentrations for Road_1000_WSU (q = 2.67E+06
g/s, n = -9.86, sy = 179.90, b = 0.23 m2/s). Note that the peak concentrations
observed occur as the mobile van is moving away from the plume centerline in a
positive lateral direction (see Figure 4.8.25 (left)), a phenomenon that the model
cannot capture, and further evidence to support the hypothesis that assumed wind
directions in Table 4.8.1 are too large.
169
Table 4.8.10 - Best-fit Parameters for Different Sized Data Sets
Data Set
Every other
Every third
Every fourth
q [g/s]
6.02E+05
5.33E+05
6.27E+05
# Points
758
505
379
n
1.61
1.60
1.63
sy
0.83
0.73
0.84
b [m2/s]
0.06
0.06
0.06
Wind Direction
The hypothesis that the assumed wind direction was off was investigated by
reformulating the optimization problem as follows:
Minq,s y ,b,n,offset 
cC
measured
A ,i
 CApredicted
,i
h
2
(4.8.34)
i
x cor ,
where offset is the deviation of the wind direction, such that the corrected coordinates ~
~y are given by:
cor
~
y
~
xcor  ~
x2  ~
y 2 cos tan 1 ~  offset
(4.8.35)
x
~
ycor  ~
x2  ~
y2
F
G
H
Ftan
sin G
H
1
F
IJ IJ
G
HK K
~
yI
I
F
 offset J
G
J
~
Hx K K
(4.8.36)
Given the higher than expected tracer concentrations observed in a positive lateral direction, we
expect the offset to be positive.
The results of the optimization are shown in Table 4.8.11, which also includes for
comparison the parameter values for when the optimization was performed assuming the wind
direction was correct and the values predicted by theory. The predicted wind offset is 6.4o. As
expected, the offset is positive, i.e., the predicted offset would rotate the x,y- plane in the
clockwise direction, bringing the observed peak concentrations closer to the plume centerline.
While the predicted offset is quite small (~ 6 o), the effect of allowing for the uncertainty in the
wind direction on the source strength is significant, bringing it within experimental error of the
expected value. Including the wind direction offset also has a small effect on the lateral
deviation parameter sy.
Preliminary analysis of sodar data taken from WSU indicates that for the time period
where data is available (13:00 – 14:00 EDT), the standard deviation in direction is approximately
50o. This value is quite high; given the expected stability class (see Section 4.8.5), a likely
standard deviation of wind azimuth is ~ 15 o. Future work could use a more complex model that
directly incorporates the expected variation in wind direction; however, by allowing for a small
uncertainty in wind direction and folding this into the optimization routine, the simpler model
given in Equation (4.8.28) can be used effectively to estimate the source strength and other
model parameters.
170
Table 4.8.11 - Best-fit Parameters, allowing for Offset in Wind Direction
Parameter
q [mg/s]
n
sy
2
b [m /s]
Theory
737
~1
~ 0.2
N/A
Offset = 0o
580
1.5
0.85
0.05
Offset = 6.4o
640
1.5
0.95
0.05
4.8.6 Conclusions and Recommendations
Using data collected in Boston on 5/25/1999, the dispersion of a tracer gas from a
continuous point source was modeled. The predicted plume dispersion matched the measured
SF6 concentrations well. The model parameters regressed from the data are in reasonable
agreement with theory. The predicted plume dispersion did not reflect the fluctuations observed
in the measured SF6 concentrations. This is to be expected since the model represents ensemble
average diffusion, but the measurements reflect the instantaneous plume behavior.
The data withholding exercise showed that a single pass along a road is unlikely to
provide data of sufficient quality to reliably estimate a point-source strength and other model
parameters. In order to have confidence in the estimated parameters, it is necessary to have data
at different downwind and lateral distances as well as over a sufficiently long time scale, so that
the average wind values used in the model are applicable to the measurements that are influenced
by the turbulent variations in the wind fields.
The model is sensitive to the wind direction, and it is likely that the predicted wind
directions are off (on the order of a couple degrees). By incorporating the uncertainty in wind
direction into the optimization problem, the tracer release source strength is recovered within
experimental error. Thus, by using a relatively simple model, downwind measurements of tracer
concentration, wind field predictions from MM5 and by allowing for a small uncertainty in wind
direction into the optimization routine, the release rate of tracer and other model parameters were
reliably estimated. Clearly, given the sensitivity of the model to wind direction, measuring the
wind direction (and speed) at the release site will increase the reliability of the estimated model
parameters.
Several important lessons were learned simply by examining the data. The original
dataset was approximately 20,000 data points. Once the data points below the instrument
detection limit and the points that were upwind of the source were removed, more than 90% of
the data could not be utilized. In order to maximize the use of data and the reliability of the
regressed model parameters, an increase in the source strength should be considered.
Meteorological predictions can be used to plan the mobile laboratory route to maximize data
usage.
171
4.9
Photochemical Steady State NOx Analyses
The mobile, fast response instrumentation, which has been developed and deployed in the
Urban Respiration program, allows us to address a number of scientific issues. These include a
need to study the impact of urban air pollutants on regional viability and on global change issues.
A specific problem is the need to determine the significance of the strength and variability of
levels of nitrogen oxides and ozone, urban air pollutants, and how they impact global change
issues. We have studied nitrogen oxides (NOx) and ozone (O3) concentration dynamics in an
urban environment during field campaigns in Manchester, NH and Boston, MA in 1998 and
1999.
In August 1998 we measured smog precursors in Manchester, NH on several clear, sunny
days. Nitrogen oxides (NO and NO2), ozone, carbon dioxide (CO2), and particles were measured
throughout the city with high sensitivity instruments mounted in a mobile platform. All species,
except for O3, were measured at a rate of 1 measurement per sec (1 Hz). The measured mixing
ratios of these species result from contributions from background air, area and point sources, and
their interaction in the troposphere.
Ozone formation in the troposphere is governed by a basic photolytic cycle:
1
NO2 + h NO + O(3P)
2
O(3P) + O2 + M  O3 + M
3
O3 + NO  NO2 + O2
k1[ NO2 ]
. In urban areas the major form of NOx is NO (nitric
k 3 [ NO]
oxide) and it results from combustion processes. To understand tropospheric urban ozone we
recognize that nitrogen dioxide formation is controlled by different limiting forces during the day
than at night. In the night time the limiting step to NO2 formation is O3 + NO  NO2 + O2 (step
3 ) because there is no light to drive step 1. If [O3]b  [NO] all of the NO will react to form NO2
and thus [NO2] will equal the initial NO concentration, where [O3]b is the background
concentration of ozone. If, however, [O3]b [NO], then the concentration of NO2 will reach
[O3]b. During the daytime, NO2 formation is limited by the photostationary state, and
k [ NO2 ]
, with the rate of NO to NO2 conversion on the order of 4 to 6 hours.
[O3 ]  1
k 3 [ NO]
In steady state, [O3 ] 
We use a simple model to describe nitrogen dioxide formation, particularly in response to
sources of nitrogen oxides. This model helps us to understand the air quality impact of such
sources. Background air is assumed to pass by point and area sources, which create plumes that
add to the background air. The total air eventually reaches a monitoring site. The observed NOx
at the site, [NOx]o, is equal to the sum of the observed NO and NO2, i.e., [NO]o + [NO2]o, and is
also equal to the sum of background, area sources, and point sources:
[NOx]=[NO]b + [NO2]b + [NO]a + [NO2]a + [NO]p + [NO2]p.
172
(4.9.1)
The rate equations associated with reactions 1 - 3 are manipulated to obtain
[NO2]=[NO2]0 + [O3] 0 - [O3]
(4.9.2)
and
1
2
2
1
k1  1 
k1 
k1
[O3 ]    [ NO]0  [O3 ]0     [ NO]0  [O3 ]0    4 [O3 ]0  [ NO2 ]0 
2
k3  2 
k3 
k2


(4.9.3)
where [NO2]0, [NO]0, and [O3]0 are the initial concentrations at the measurement point of NO2,
NO, and O3, respectively. The result of this analysis is that we can easily examine the
relationship of NO2 to total NOx in our NOx and O3 data.
In Figure 4.9.1 we have displayed NO2 versus NOx (NO + NO2) collected by our mobile
laboratory on August 28, 1998 in Manchester, NH, a small industrial city in New Hampshire.
Lines 1 and 2 in the figure are the night time limits:
Line 1: Nighttime limit where all NOx is in the form of NO2
Line 2: Nighttime limit where all O3 has been converted to NO2, and this is
the maximum NO2 limit.
[NO2 ] 
Line 3 is the day time limit determined by the photostationary state limit,
k 3 [NO][O3 ]
k1
.
The rate constants were set to k1= 8.9 x 10-3 sec-1 and k3=4.7 x 10-4 ppb-1sec-1 as determined from
the experimental conditions during the measurements (uv intensity and air temperature). The
initial concentrations of NO, NO2, and O3 at the measurement point are based on the data, and
were set to 8.4, 3.4 and 15 ppb, respectively, for the determination the day time line in
Figure 4.9.1. These values are based on the average background levels from the previous day.
The O3 in the plume from the source is assumed to be zero. NO in the plume is varied from 0 to
1000 ppb. For each NO value we calculate the expected NO2 and O3 mixing ratios.
In Figure 4.9.1 there is a wide range of NO2 and NOx levels. Most measurements lie in
the area bracketed by the three limits. A number of data points also lie above the “max NO2”
line. We can discriminate the data by CO2 level, where CO2 serves as an indication of a direct
combustion source. For urban measurements at street level, the local combustion source is a
vehicle in the vast majority of cases. Points are separated into two groups: those with CO2
below 400 ppm and those with CO2 greater than or equal to 400 ppm. All of the points above
line 3 are associated with high CO2 and therefore local sources. Local sources are a critical
173
70
2
1
60
NO2 (ppb)
50
40
30
20
3
10
NO2 with CO 2 <400 ppm
0
100
200
300
400
500
600
NO x (ppb)
Figure 4.9.1.
Relationship of NO2 to total NOx. Data from August 28, 1998. Manchester, NH
• CO2  400 ppm • CO2 < 400 ppm Line 1: All NOx in form of NO2. Line 2:
Maximum possible NO2 where all O3 converted to NO2. Line 3: Photostationary
state limit where k1=8.9 x 10-3 sec-1; k3=4.7 x 10-4 ppb-1sec-1. Background NO,
NO2 and O3 : 8.4, 3.4, and 15 ppb, respectively.
component in the understanding of urban NOx and O3. This is also illustrated in Figure 4.9.2 in
which similar data from August 27, 1998 is shown. On this day there were fewer high NO2
points, but many of the higher levels are associated with high CO2. The data in Figures 4.9.1 and
4.9.2 were collected on consecutive days in Manchester, NH, but they display different
distributions of NO2 versus NOx. Both data sets contain both city and highway data and were
collected during the daytime in warm weather (mid 80s F). The wind direction during the
measurements, however, differed on the 2 days. It was from the southeast on 8/28 and varying
between northwest and southwest on 8/27. The different patterns of the data in the 2 figures
174
1
60
NO 2 with CO 2 < 400 ppm
2
50
NO2 (ppb)
40
30
20
10
3
0
Figure 4.9.2.
100
200
NOx (ppb)
300
Relationship of NO2 to total NOx. Data from August 27, 1998. Manchester, NH
• CO2  400 ppm • CO2 < 400 ppm Line 1: All NOx in form of NO2. Line 2:
Maximum possible NO2 where all O3 converted to NO2. Line 3: Photostationary
state limit where k1=8.11x10-3 sec-1; k3=4.38 x 10-4 ppb-1sec-1. Background NO,
NO2 and O3: 5, 3 and 3 ppb, respectively.
probably reflect the influence of different sources of nitrogen oxides during the measurements.
More local combustion sources contributed to the ambient air during the measurements on
August 28 than on August 27. The significance of local sources on regional air quality is thus
reflected in our measurements and subsequent analysis.
We now examine the diurnal data that we collected at the fixed site in Cambridge, MA on
May 27-28, 1999. We analyze the 24 hours of data covering May 27 at 20:00 until May 28 at
20:00 EDT in a similar manner as we did the mobile daytime data from Manchester. One of the
first things that we observe in Figure 4.9.3 is the richness of the data and how most of the NO2
levels appear to be approaching a maximum NO2 level with increased NOx. The latter
observation was not as apparent in the Manchester data, probably because it included almost
exclusively daytime measurements. In Figure 4.9.3, the nighttime limits to NO2 as related to
total NOx, are given by lines 1 and 2. As described previously, Line 1 represents the NO limit,
175
where [O3]b[NO] and thus NO2 will equal the initial NO concentration. There are just a few
points in the figure where NOx<NO2. They are result of periods of very low NO mixing ratios
during which some points fall below zero statistically. Line 2 represents the ozone limiting case,
i.e. [O3]b[NO] and NO2 can only be as high in concentration as the concentration of [O3]b. In
the data of Figure 4.9.3, line 2 appears to be too low. It is based on the maximum observed
ozone at the site on 5/27 and 5/28 including prior to the measurement period. The data however,
seems to be approaching a level of NO2 of 55 ppb. This is the level of ozone that we expect in
the background air. The maximum observed ozone on 5/26 while driving in Boston was about
52 ppb, similar to that of line 2. One possibility is that this air is contributing to the background
air the next day. The other possibility is that background air with high ozone originates from
outside the measurement area and was perturbed by the time we observed it.
Line 3 is the daytime photostationary state limit for the diurnal data. The photostationary state
limit is calculated with the following parameters: k1=9.74 x 10-3 sec-1; k3=3.6 x 10-4 ppb-1sec-1
and background NO, NO2 and O3 equal to 7, 2.5 and 9.5 ppb, respectively. All of these
parameters are based on observed experimental factors, including ultraviolet level, air
temperature, and trace gas levels. They appear to do an excellent job representing the daytime
limit. Only a few number of nitrogen dioxide points fall below this limit.
As we did in the analysis of the Manchester photochemistry data, we indicate in
Figure 4.4.3 the points with associated elevated carbon dioxide levels. The points in the figure in
the blue shades are those points for which CO2 < 400 ppm (i.e., not elevated). The other points
may be linked with direct combustion sources, such as vehicle emissions. Many of the high NOx
points occurred with high carbon dioxide emission.
176
1
20:00-08:00
8:00-20:00
20:00-08:00 CO2<400ppm
8:00-20:00 CO2<400ppm
60
NO2 (ppb)
50
2
40
30
20
3
10
0
0
100
200
300
400
500
NOx (ppb)
Figure 4.9.3.
Relationship of NO2 to total NOx. Data from May 28-29, 1999. at the MIT
stationary site in Cambridge, MA.+,+ all points; •, • CO2 < 400 ppm
Line 1: All NOx in form of NO2. Line 2: Maximum possible NO2 where all O3
converted to NO2. Line 3: Photostationary state limit where k1=9.74 x 10-3 sec-1;
k3=3.6 x 10-4 ppb-1sec-1. Background NO, NO2 and O3: 7, 2.5 and 9.5 ppb,
respectively.
177
4.10 GIS Based Emissions Analyses
One goal of the project involved the use of geographic information system (GIS)
technologies and geo-spatially referenced datasets. By using modern GIS tools and datasets, we
sought to identify readily replicated methods to facilitate the planning of field experiments and to
assist in correlating the observed trace gas emission fluxes (urban respiration) with urban and
industrial activity and consumption factors (urban metabolism).
Increasing standardization and spatial disaggregation of metropolitan databases is
providing detailed data about land use, land cover, topography, traffic congestions, and other
urban activity. If such data could be correlated sufficiently well with observed trace gas
emission fluxes, then it would greatly simplify the task of calibrating and using models of
atmospheric chemistry and dynamics in additional metropolitan areas. It would also facilitate the
use of inverse models to identify the location and nature of land uses and urban activities that are
significant contributors to adverse atmospheric conditions.
4.10.1 GIS Analyses for Manchester, NH
Initial efforts to assemble and test suitable GIS tools and data layers focused on
Manchester, NH, and the field campaigns of November, 1997, and June, 1998. The methods and
results are described in further detail in one of the project papers [Yeang, et al., 1999]. We
collected a series of data layers from a variety of sources (mostly online) that are representative
of datasets describing land use, land cover, terrain, demographics, meteorology, and known
point-sources of air pollution. Although they are not an exhaustive list of determining factors for
air pollution, they constituted a meaningful and reasonably comprehensive set of data that are
relevant to the generation and dissemination of air pollutants. Moreover, they are data that can
be assembled in a reasonably standardized fashion for most U.S. metropolitan areas in order to
help us understand the urban respiration phenomenon. These data layers included:
Land use/land coverage
The land use/land cover (LULC) data files from US Geological Survey (USGS) describe
the vegetation, water, natural surface, and cultural features on the land surface. Original data
sources include high-altitude aerial photographs and earlier land use maps and field surveys.
They are stored in a Geographic Information Retrieval Analysis System (GIRAS) format and are
often available online through State-supported Web sites that archive environmental
management data. The scale of the LULC maps is 1:250,000.
EPA monitoring sites
The Aerometric Information Retrieval System (AIRS) and the AIRS Facility Subsystem
(AIRS/AFS) are online services of the EPA’s Envirofact database – a large database of
environmental data maintained by U.S. Environmental Protection Agency. It comprises the
identification information, spatial coordinates, and emission inventory of EPA-monitored sites.
The chemicals monitored by EPA include CO, NO2, particluate matters, lead, SO2, and volatile
178
organic compounds. The database is open to the public and can be accessed through the World
Wide Web (http://www.epa.gov/enviro). Since the Envirofact data warehouse is stored in Oracle
and supports SQL queries from the public using ODBC protocols, our distributed GIS
architecture can include online queries of emission information for sites in the Envirofact
database. [Recently, in late 2001, security concerns have prompted EPA to limit direct access to
its Oracle data warehouse from outside of EPA. Special arrangements involving registration and
the use of virtual private networking are now needed for such access. Such security concerns
and procedures complicate the use of a distributed computing model to allow rapid development
and calibration of urban respiration models. But they will be with us for the foreseeable future
and will require added coordination and technical complexity both for those who model urban
respiration and for those, such as EPA and State GIS Data Centers, who archive relevant data in
online repositories.]
Terrain data
Terrain data for Manchester, NH, comes from the New Hampshire Resource Net, a GIS
warehouse supported by the state of of New Hampshire and maintained at UNH
(http://nhresnet.sr.unh.edu/). The archived data are stored in a 7.5-minute digital elevation
model (DEM) format and represent elevation estimates for use with map scales of 1:24,000 or
1:25,000. The elevation estimates are 30 meters apart for an x-y grid that is based on a UTM
(Universal Transverse Mercator) projection method and the 1927 North American Datum (NAD
1927).
Surface water hydrography
The data for surface water hydrography also comes from the New Hampshire Resource
Net. It contains the vector representations of the boundaries and/or centerlines of various surface
water bodies (lakes, ponds, rivers, wetlands, reservoirs, etc.) The scale is also
1:24,000/1:25,000. It uses New Hampshire State Plane Feet projection and 1983 North
American Datum (NAD, 1983).
Road networks and railroads
The data about road and railroad networks also come from the same New Hampshire
Resource Net and have the same projection, horizontal datum, scale, and file format as the
hydrography data.
Demographics
Demographic data about population characteristics come from 1990 U.S. Census Bureau
datasets (STF3a). They are georeferenced by place of residence down to the blockgroup level
and are linked to vector-based maps of census tracts, block groups, and streets using the Census
Bureau’s TIGER files (and other street centerline datasets supplied by third-parties and derived
from and/or compatible with TIGER data).
179
4.10.1.1 Creating Base Maps
By using the various data layers described above, we were able to create and overlay a
rich collection of data relevant to the modeling and analysis of urban respiration in and around
Manchester, NH. To facilitate viewing and analysis that is consistent with NH supported GIS
data, all the georeferenced data were converted to New Hampshire State Plane coordinates
(NAD-83, feet). While this conversion process is well understood, many GIS packages do not
yet support robust enough methods of on-the-fly coordinate projection. Among the few data
sources that we assembled for Manchester, NH, we encountered four different coordinate
systems, 2 different NAD reference sets and both ‘meters’ and ‘feet’ accounting. Since there are
good reasons why one ‘size’ does not fit all purposes, we regard this phenomenon as evidence
that robust, on-the-fly methods for coordinate conversion are a key element of building useful
distributed GIS systems for supporting multi-disciplinary modeling and analysis. For this work,
and for subsequent analyses of the Boston, MA, metropolitan area, we used GIS software (such
as ArcView and ArcInfo) from the Environmental Systems Research Institute. This vendor,
ESRI, is the market leader in industrial strength GIS software and is used by most US Federal
and State Agencies (such as EPA, USGS, Census, and the GIS data centers in all six New
England states). Using this software was a good test of off-the-shelf GIS capabilities and
facilitated the use of public datasets in the manner that could be most easily replicated in other
metropolitan areas.
Several of the Manchester data layers are illustrated below. Figure 4.10.1 shows, on the
left, a lattice mesh terrain model of Manchester, NH, with major roads and hydrography draped
over the lattice. The graphic on the right is a screen-shot of the ArcView GIS software showing
a plan view of the Manchester metro area with county boundaries, major roads, and hydrography
visible.
Figure 4.10.1. Terrain model of Manchester, NH, and ArcView screen-shot of Manchester.
180
Figure 4.10.2 shows another screen shot of ArcView with the mobile van measurements
of CO2 from November 10 and 11, 1997, overlaid on top of a land use map. The two maps
compare readings for two different times of day. In addition, the maps are viewable in an
ordinary web browser using a Java applet called, MapCafe, that is downloaded to the client’s
browser by ESRI’s ArcView-IMS (internet map service) software. Using this distributed GIS
architecture, maps can be viewed and explored by team members who need not have the
underlying datasets stored locally on their machines.
4.10.1.2 The Distributed GIS Architecture
Figure 4.10.3 provides a conceptual diagram of the system architecture for the distributed
GIS approach that we first implemented in order to permit team members to access maps and
data via theWeb from their respective locations. Since the Internet and World Wide Web are so
pervasive, we used TCP/IP and Web browser protocols for connectivity. On the client side,
individual users access the data through the Internet without requiring proprietary software. A
Web browser can function as the Graphical User Interface (GUI) at the client side. Users can use
this interface to send requests about the characteristics of the GIS data they want to process or
view. For example, they can identify the layers to be turned on, the spatial range of the map
display, and the specific subsets of certain data layers that they need. The browser interface can
also be used to display the resulting maps or GIS data files and/or to bundle subsets of data for
downloading.
Figure 4.10.2. MapCafe Screen-Shot comparing CO2 Measurements at Different Times of Day.
181
Figure 4.10.3. Conceptual diagram of the Distributed GIS Architecture
On the server side, a Web server communicates with the client-side Web browsers.
Unlike a conventional Web server, this one has an additional function of bridging to a GIS
server. It translates commands from the Web browsers into the function calls identified by the
GIS software. Going in the other direction, it also converts GIS maps and datasets into the
formats which can be displayed on the Web.
A GIS server sits behind the Web server. It functions as conventional GIS software
except with distributed computing capability. In other words, it will be able to handle multiple
requests from different clients. Data consistency among multiple users is one of the major
problems for distributed data sharing. If two users modify the same data simultaneously, the
race conditions may lead to unexpected changes in the data. A fully-fledged database system has
various protection mechanisms to avoid these conflicts (such as priority queuing and transactionbased writeback). Nevertheless, due to the nature of this project, a concrete database system for
complicated data sharing purposes (such as a distributed corporate database) is not required.
This is because it is practical to limit client-side processing so that the ‘core’ maps and datasets
that are shared are viewed as ‘read-only’. Most users are satisfied with this unidirectional data
access and recognize the simplified data management that results. They can generate subsets of
data and map the results, and the can downloading maps and filtered datasets. But they cannot
use the interface to upload new data or otherwise alter the ‘core’ datasets
A relational database management system (RDBMS) sits behind the GIS server and
stores all the tabular data from the measurement and monitoring of weather and trace gases. The
major advantage of using a DBMS rather than the GIS per se to store these data is the capability
to extract (and aggregate, filter, or otherwise process) specific subsets from the data in response
to users’ requests. For example, a single evening of mobile van readings throughout a
metropolitan area might acquire tens or hundreds of thousands of observations. A researcher
may wish to extract and plot a small subset of these data after, say, a smoothing operation that
aggregates them in time and space to a hundred yards and/or ten seconds. These requests are
readily handled as queries (and stored procedures) using the industry-standard Structured Query
Language (SQL). Access to the RDBMS data is directly available to client-side users via normal
distributed DBMS protocols (such as ODBC). But the system also allows the RDBMS datasets
to be accessed from the GIS server in response to user-defined queries entered by the user
through their GIS interface.
182
Figure 4.10.4 explains the specific implementation of the distributed GIS framework that
was diagrammed conceptually in Figure 4.10.3. For the GIS server, we chose ArcView with the
Internet Map Server (IMS) extensions (Environmental System Research Institute, ESRI). There
were several reasons for the choice. The research team had substantial experience using ESRI
software (both ArcView and ArcInfo) and, as indicated earlier, many land use and demographic
datasets of interest to the project were readily available in ArcView readable formats. In
addition, ArcView has both a user-friendly interface, data formats compatible with ArcInfo, and
significant market share as a standalone desktop mapping package. The former advantage makes
it easier to build the graphical user interface for project participants whose expertise is not GIS.
The second advantage allows us to perform sophisticated GIS operations using ArcInfo and then
display the resulting map layer in Arcview without the need for additional data conversion.
However, ArcView plus its IMS extension cannot satisfy all our data access and
management needs. If we stored the monitoring and measurement data as GIS layers (in this
case, as ArcView shape files), end users would have little freedom to select and display specific
subsets of the entire data set. Excluding the display of unwanted subsets of the shapefile data
would be too complicated a task that is not possible with the reduced GIS functionality that is
available via the ArcView-IMS graphical interface. Therefore, we decided to use a database
management system (DBMS) to store the air quality data, and allow users to access these data
through Arcview IMS. We chose Oracle since it supports industry-standard SQL and distributed
DBMS protocols such as ODBC, it is compatible with ArcView, and it is widely used to store
enterprise-level datasets for environmental planning and metropolitan management. To illustrate
the need for a DBMS, suppose a user wants to display the measurement samples obtained during
2pm to 4pm among all data collection trips. The user can input his or her queries using the
graphical user interface at the client side. The server-side of the ArcView-IMS application
passes the request (as a standard SQL query) to the DBMS, which then returns the results as an
ArcView table. ArcView then converts the data table into a GIS point data layer and display it
on client’s map.
Figure 4.10.4. Detailed Diagram of Distributed GIS Architecture for Manchester Analyses.
183
The client side of the implemented system is a generic web browser which runs a Java
applet called MapCafe that comes with the IMS extensions to ArcView. MapCafe sends
commands generated by the user interaction (in http format) to the web server, and it receives
and displays image files, tables, and annotations generated by the GIS Server and packaged (in
http format) for client-side viewing. The look of the GUI at client side is similar to ArcView,
except more limited in its toolbars and interactivity. (Figure 4.10.5 will provide an example later
in this paper). Basically, the MapCafe acts as a thin-client application providing the user with a
‘keyboard extension’ through the Web so that the user can run (some of) the ArcView
functionality that they would have if they were sitting directly at the GIS server. Users can turn
specific data layers on/off, zoom in/out, pan, get information about specific data elements, and
print. The added functionality of executing (and then mapping) logical DBMS queries of air
quality datasets required customized ‘Avenue’ scripts and JAVA programming.
At the Web server side, ESRI provides a CGI program (a dynamically linked library,
esrimap.dll, in the case of a PC server) which converts the incoming http commands into the GIS
commands (Avenue scripts) required by the GIS server (Arcview). This ‘relay’ code on the Web
server also takes the images and texts produced by ArcView and packages them for client-side
display.
The ArcView application runs on a GIS server that can be – but need not be – the same
machine as the Web server (or the DBMS server). In our case we use either a PC or a Unix
workstation as the GIS server. The GIS server listens for avenue script commands from the relay
program, performs the requested manipulations as if the user were requesting them at the
keyboard, and sends the updated display back to the relay program. For example, suppose the
client-side user zooms in by dragging a rectangle across the map showing in their browser
window. The relay program will send to the GIS server the commands that reset the zoom level
of the map (the command will include the four vertices of the zoom rectangle expressed in
viewing coordinates). ArcView performs the zoom, redraws the map on its local map display,
and then sends the map – as a GIF image – back through the Web server to the client. In our
scheme, ArcView also functions as a relay to database server. Through the Open Database
Connection (ODBC), it can send SQL commands to one (or more) Oracle database servers and
receives the query results. It then converts the resulting table into a temporary shape file, opens
this shape file as a new layer of the map, and sends the map to the client as explained above.
By using ArcView-IMS, we chose a thin-client solution for our distributed GIS. No GIS
‘smarts’ exist on the client-side. The GIS server packages all maps as graphic images and passes
them to the client-side browser. All spatial data processing – e.g., buffering, point-in-polygon
computations, and the like must be done on the server side and the map ‘themes’ (i.e., data
layers) and associated data the ArcView displays must be explicitly loaded and prepared for
viewing by the manager of the GIS server. In effect, the Web is used as a long (and somewhat
cumbersome) keyboard extension to allow the client to act as if they were sitting at the GIS
server examining prepared ArcView ‘projects’. An alternative (thick-client/thin-server)
approach would be to use the servers only as a data repository and move the GIS processing to
the client-side. A simple example would be to run ArcView on each client and use a network
file server to share the raw data over the net.
184
4.10.1.3 Visualizing Mobile Measurement of Trace Gas Concentration
Figure 4.10.2 above illustrates the use of the distributed GIS architecture. The maps
show the CO2 concentration data collected on November 10th and 11th 1997. The background
shading represents land use data from the USGS-LULC coverage. The mobile measurements are
point data that form the linear paths that crisscross the downtown area and evident in the map
foreground. Please note that the printed figure is far less readable than an on-screen map in
color. The narrow and lighter portions of the paths represent CO2 readings at or near the
background level. The wider and brighter portions of the paths represent successively higher
readings. The trip on November 10th was conducted during the late evening (from 8pm to 1am),
whereas the trip on November 11th was during the late afternoon and early evening (4pm to
9pm). The 10 hours of mobile readings produced 36,000 readings (one per second) for each of
the two gases that were monitored (CO2 and CH4). The time stamp, GPS location, and trace gas
readings for the data are stored in Oracle and converted into mappable data using ArcView.
The contrast of CO2 concentration between the two time periods shown in Figure 4.10.2
is striking. The November 11th data contains more high concentration spots than the November
10th data. Except at the southern tip of the city (the intersection between routes 3 and 293), all
other high concentration points are not overlapping. This reveals the possibility that either (1)
most CO2 sources are non-stationary and time-varying, which strongly point to motor vehicles;
and/or (2) the measurement data are very sensitive to local traffic conditions, especially if the
van is measuring vehicle exhaust from cars close in front of it.
The land use classification data from USGS LULC are too coarse (both spatially and in
terms of land use categories) to demarcate precisely those areas that have higher CO2
concentrations (for reasons other than car emissions). Although almost all the high CO2
concentration spots are located within the industrial, commercial or residential areas, there are
many low concentration spots which are located within these areas, too. Terrain and
meteorological conditions (e.g. being upwind or downwind of emission sources) are likely to
have significant impacts even if little atmospheric chemistry is occurring (at the surface).
These observations indicate that the data processing and ‘reverse engineering’ needed to
link trace gas measurements to emission sources requires considerable data filtering, modeling
and analysis. But the ArcView-IMS user is limited to simple zoom, pan, and query options for
exploring a limited number of data layers. Complex filtering and querying of the mobile
measurements is either not possible or too awkward and time consuming with the standard
ArcView-IMS tools. To address some of these issues, we added a customized query capability
to ArcView-IMS so that end users can access the Oracle database directly in order to query and
map selected subsets of the monitoring data. Figure 4.10.5 shows the graphical user interface for
these queries of the Oracle database. The user must specify a username in order to distinguish
(and hide) their customized ArcView ‘themes’ from those that other users create. Users also
need to specify the name of the table to be queried and the text label they wish to use when
mapping their customized datasets. These text labels are color-coded for added clarity.
185
Figure 4.10.5. Query Box for Customized Querying and Mapping of Trace Gas Measurements
Users can use this dialogue box to select the subsets of trace gas readings that they wish
to extract and map. They can select by type of gas, time period (including day and time of day),
and measurement levels. The five selection boxes and text fields below the check boxes allow
users to specify a range of conditions. In this example, only one condition is specified: the CO 2
level must be between 550 ppm and 600 ppm. Query results are the intersection of all specified
conditions. After these fields are filled, users can either click the SUBMIT button to submit the
query command to the Arcview server, or click the QUIT command to close the GUI window.
Hitting the ‘quit’ button after entering only one’s user name will load all previously generated
user-defined themes into the viewing window without generating a new theme. If a new query is
specified, ArcView passes the query (in SQL format) to Oracle and automatically generates and
maps a new shapefile comprising the selected datapoints.
The look of the modified system is very similar to the original IMS user interface
(Figure 4.10.2) except that a new button is added to the topmost button list. New layers of userselected CO2 and CH4 observations show up as additional themes in the viewing legend. Since
these layers are created from individual user queries, different users can view and map different
subsets of the entire trace gas measurement database.
4.10.1.4 Geo-Processing Examples
The previous example illustrated the use of distributed GIS tools to facilitate the
visualization, filtering and aggregation of trace gas observations. The same approach can be
used to handle a range of mobile (and fixed) measurement data (for meteorological data and a
suite of trace gases). But such processing and mapping is only the first step in the modeling and
analysis process. The broader goals are (a) to use the measurement data to calibrate the surface
186
conditions for running (and improving) the volumetric models of atmospheric chemistry and
fluid dynamics, and (eventually) (b) to ‘reverse-engineer’ the model predictions in order to
identify the locations and types of emissions that are most influential in producing adverse air
pollution conditions.
The distributed GIS architecture can also be helpful in supporting these goals. We
illustrate one such use involving the estimation of certain baseline surface conditions for
calibrating the model. Vehicular traffic is a key source of trace gases (such as CO2). The
location and level of these emissions is, of course, dependent upon road location, traffic
congestion, cold-start effects, vehicular emission controls, and the like. The GIS basemaps
described earlier provide a rich source of data that can be used with the GIS tools to build useful
models of the spatial pattern of vehicle emissions. For example, Figure 4.10.6 shows the road
network in and around Manchester in gray and solid lines. The buffers surrounding the roads are
shaded lighter and lighter to the extent that they are closer and closer to more (or more major)
roads. (A simple inverse distance model is used with weightings based on road class and number
of lanes). Further adjustment could be done to reflect time of day traffic congestion and/or
estimated differences in vehicle mix and emission levels depending upon the proximity of the
roads to residential neighborhoods with different demographic profiles.
Such models can be used to create contour maps for the expected (surface) levels of trace
gas concentration due to emissions from motor vehicles. And, trace gas measurements from,
say, morning rush hour periods could be used to calibrate the parameters of these surface-level
emission models. The calibrated models could then be used in estimating the road contribution
to trace-gas emissions throughout the metropolitan area. They could also be used to extrapolate
estimates of aggregate emission levels (from vehicles) that are generated within each grid cell
across the metropolitan area. Surface level ‘initial conditions’ such as these are needed to
calibrate and seed the air pollution models. Similar analyses and spatial data processing could be
done with land use and terrain data (e.g., to estimate surface ‘roughness’), and with demographic
data.
GIS tools such as ArcView have the buffering, rasterization, and map algebra capabilities
needed to make these estimation, spatial interpolation, and spatial aggregation steps reasonably
automatic. A limited amount of coding – similar to the oracle queries discussed earlier – can add
such functionality to the distributed GIS capabilities of our system. Subsequent project work
experimented with these approaches.
187
Figure 4.10.6. Proximity-to-Road Model for Estimating the Spatial Distribution of
Vehicle Emissions
188
4.10.1.5 Conclusions from Initial Work on Distributed GIS Systems
This section described the intial prototyping and use of a distributed GIS system for
modeling, analyzing, and monitoring urban respiration. A thin-client approach was used to
distribute limited access to spatial data sets and monitoring data using Internet & Web-based
protocols. Off-the shelf GIS and RDMBS tools were used to provide Internet-accessible spatial
data processing and together with querying services with a minimal amount of customized
programming. Initial experience with these tools indicates that:

It is relatively easy to assemble and standardize key datasets of spatially referenced data that
can serve as ‘basemaps’ for visualizing and analyzing trace gas monitoring data.

It is relatively easy to store and cross-reference the GPS-referenced trace gas measurements
so that they can be overlaid on the basemap layers.

It is useful to provide (via ArcView-IMS) a minimal level of desktop mapping capability,
with some consistency and user flexibility, to the various research teams.

It is still difficult (within the limitations of a thin-client approach like ArcView-IMS) to
provide sufficient flexibility and analytic capability to avoid the need for ftp exchange of raw
datasets among the research teams.

Standardized, industrial strength RDBMS services are needed to store manage and query the
measurement and monitoring data with sufficient flexibility and power.

The performance issues, reliability, and Java code requirements of a tool like ArcView-IMS
are sufficiently complex and non-standard to warrant continued reliance on simpler map
distribution strategies (such as static web pages and PDF-formatted maps) for some project
purposes.

More complex (and customizable) strategies are warranted for supporting (a) the filtering,
interpolation, aggregation, and visualization of the meteorological and trace gas
measurements [Figures 4.10.2 and 4.10.5], and (b) the spatial data processing needed to
estimate surface level emissions, roughness, and other air pollution model parameters
[Figure 4.10.6].
189
4.10.2 GIS Analyses for Boston, MA
Subsequent efforts to assemble and test suitable GIS tools and data layers focused on
Boston, MA, and the field campaign of May, 1999. This section (4.10.2) describes the additional
GIS data layers and tools that were available for Boston and explains the more detailed models
that were developed to estimate the spatial distribution of trace gas emissions resulting from land
use and urban activity. The next section (4.11) describes our efforts to calibrate the emission
models using the observations from the Boston field experiments in May 1999. The methods
and results are described further in one of the project papers [Cao, et al., 2001] and in a project
website [see http://metro.mit.edu/urbanair/overview ].
The research team had access to GIS data layers for the Boston metro area that were
similar to those used in Manchester. These datasets included the EPA inventory of AIRS/AFS
sites and land use, hydrography, census, terrain and road network data from various State and
Federal agencies. As was the case in Manchester, much of these data are now available online
from the state’s lead GIS agency, MassGIS, (see http://www.state.ma.us/mgis ). For
Massachusetts, the landuse data was much more detailed than it is in the USGS topographic
maps that we used for Manchester, NH. For several decades, the Massachusetts has periodically
funded a Resource Mapping Project at the University of Massachusetts, Amherst. The project
classifies land use into a few dozen categories based on photointerpretation of 1:40,000 scale
color infrared photos. We were also able to use high resolution, digital orthophotos of the Boston
metro area. These one-byte, grayscale orthophotos have a ground resolution of one-half meter
and were developed from aerial photography at 1:30,000 scale that was orthorectified and
registered to Mass State Plane coordinates (NAD 1983).
4.10.2.1 Digital Orthophotography and GIS-based Visualization
The digital orthophotography was quite helpful both for planing the field experiments and
for interpreting the results. Since the orthos have a meaningful geographic coordinate system,
they can provide a useful visual underlay for other GIS data layers such as census demographics
and the trace gas measurements recorded by our GPS-equipped mobile van. Figure 4.10.7
shows a ‘zoomed-in’ portion of the Mass orthophotos cover a 1.5x1.5 km portion of Boston near
the I-93 (Southeast Expressway) and I-90 (Mass Turnpike) interchange. The red dots
superimposed on the orthophoto mark the GPS-recorded location of our mobile van during one
of our Boston field experiments. The dots are shaded darker red when the van was traveling fast
and lighter pink when traveling slowly. The GPS-recordings were every second and they are far
enough apart to be distinguished from one another along the Turnpike and for the southbound set
of points along the Southeast Expressway. The Southbound run was made at 10 AM and the
Northbound run (and the loop onto the Turnpike) was made during rush hour at 5 PM. The
‘bunching-up’ of the lighter colored dots - i.e., slow speeds - is quite noticeable during rush hour
(until the van gets past the bottleneck and onto the free-flowing part of the Turnpike). Also
visible at the lower left of the image are some registration errors. The van is displayed about
30 meters west of the roadway at this point. Both the orthorectification and the GPS estimates
are subject to some error but only the GPS readings are likely to be off by this amount.
190
Figure 4.10.7. Aerodyne Van Traversal along I-93 and I-90 in Boston. GPS readings at onesecond intervals are shaded darker for higher speeds and overlaid on a digital
orthophoto.
Onscreen, the orthos are even sharper and it is possible to zoom in further to take
advantage of the half-meter resolution. However, the digital ortho files are quite large and
cumbersome. The raw imagery for Eastern Mass is more than 50 Gigabytes and the USGS
countrywide series (of DOQQs at 1 meter pixel resolution) will consume 3 Terabytes of storage.
Even with modern compression methods, these files are unwieldy to distribute and use.
Beginning in 1996, the MIT research team pioneered Web services for ‘just-in-time’ delivery of
customized snippets of digital orthos. Our methods and protocols have become part of the Open
GIS Consortium (www.opengis.org) standards for interoperable and distributed GIS components.
For the Boston field experiments and subsequent analysis, we used our Web server for Boston
metro orthophotos (ortho.mit.edu) and an extension that we wrote for one of our GIS software
packages (ArcView). The extension added two buttons to the normal ArcView mapping
window. One button allows the user to select a web server that uses our MITOrthoTools to
provide ortho snippets. The other ‘fill up’ button requests an ortho from the web server that is
just big enough to fill the mapping window (using the lowest resolution for good on-screen
viewing). The web server prepares and sends the customized ortho snippet within a few seconds
and the returned files are relatively small (usually under 150 KB). There’s no need to archive a
191
complete set of orthos locally and it is easy to click the ‘fill up’ button to slip an ortho under
one’s GIS map whenever it is useful. Figure 4.10.8 illustrates this use by showing an ArcView
mapping window of Eastern Massachusetts with the Mass County boundaries and the Interstate
highways turned on. The dark gray rectangle in the middle of the map is the ortho layer – a pie
slice is cut out on the right side of the rectangle since that part is all water and was not
photographed. I have zoomed out, after clicking the ’fill up’ button, to illustrate that only the
orthos needed to fill the ArcView mapping window are obtained. At this scale, the ortho isn’t
very detailed and interesting. But, if we zoom into the intersection of the Turnpike and I-93 –
and then click the ‘fill up’ button again - we’ll get a fresh ortho snippet that has the detail of the
image in Figure 4.10.7 above.
The use of such a web service was very helpful at project meetings to select field
experiment routes and to interpret and discuss results. It is also a good illustration of the general
trend toward Web-based mapping and GIS. If such tools were widespread, a researcher could
obtain needed datasets from distributed repositories only as they are needed. If the server has the
right smarts, the repository data can be transformed as needed – e.g., projected to a new
coordinate system, filtered using pre-established ‘business rules,’ clipped to match the view
window, transparently overlaid with other data, etc. Moving as many such geoprocessing
services as possible to the server side can free up the researcher to dig deeper in exploring and
understanding the data and can speed up the interactions among interdisciplinary teams with
different expertise. Later on, after discussing our modeling and analyses, we will have more to
say about appropriate distributed GIS architectures for environmental monitoring and modeling.
Figure 4.10.8. Eastern Mass Counties and Interstate Highways with a Boston area Ortho
192
Figure 4.10.9 illustrates some of the GIS-based visualization that we used to help
interpret trace gas readings from the Boston field experiments. The two graphics trace the route
of the Aerodyne mobile van through North Dorchester in May of 1999. The basemap is a fairly
coarse portion of the Boston orthos draped on a terrain model. The GPS-based locations of the
van route has been color coded and extruded above the surface. The color coding indicates
speed with light pink being slow and dark red being fast. Each bar represents one observation of
the concentration of NO or NO2. The height of the bar is proportional to the trace gas
concentration of NO (on the left) and NO2 (on the right). Notice how the trace gas
concentrations tend to rise as the mobile van slows down (light colors) at intersections or in
traffic. The pattern is quite visible and we can use the GIS to zoom in on trouble spots and
develop a sense of neighborhood characteristics and strategies for improved filtering of the data.
Other types of visualizations (not illustrated here) involved the use of wind arrows
(rather than extruded bars) to show the wind direction at the time of each observation. The color
of the wind arrow is proportional to the trace gas concentration and the size and direction of the
arrow reflect the wind speed and direction.
Figure 4.10.9. Van Speed (light color is slow) vs. Concentration (height) of
NO (left) and NO2 (right).
193
4.10.2.2 Modeling Emission Sources
The purpose of the GIS-based modeling and analysis is to translate readily available GIS
datasets about land use and urban activity into plausible baseline estimates of surface-level
concentrations of trace gases. The procedure builds upon the distributed GIS and modeling
experiments described in section 4.10.1 above. The whole modeling process can be divided into
three stages: trace gas emission modeling, trace gas concentration (surface level) modeling, and
air pollution modeling. Our work is focused on the modeling of trace gas concentration at surface
level. The process is illustrated in Figure 4.10.10. We assume that the primary sources of trace
gas emissions are point sources (such as smoke stacks in the EPA databases), vehicle emissions,
other non-point source emissions that can be correlated with land use, land cover, and population
density. We wish to estimate the amount and location of these emissions and then combine the
estimates together by modeling wind and dispersion. To calibrate the surface-level trace gas
concentration estimates from our combined model, we can use the observed concentrations of
trace gases on days when we do not expect many complications from atmospheric chemistry. At
this stage the focus is on a proof-of-concept regarding the feasibility of developing GIS
processing ‘pipelines’ to carry out the necessary steps using readily available data. We would
like this ‘pipeline’ to consist of standardizable GIS modeling components and standard ways to
integrate them. In this case, models can easily be replaced with newer, better ones and new
components can be added as needed. We would also like to distribute these components over the
network so that they can be shared readily but maintained and calibrated by the agencies and
experts most familiar with the relevant domain.
Trace Gas
Emissions
Land Use
(non-point
source
emissions)
Stacks
(point
source
emissions)
Vehicle
Emissions
Simple Surface Level
Concentration
GIS ModelsModeling
Land Use
Model
Stack
Model
Wind
Model
Road Model
Trace Gas
Concentration
(surface
level)
Traffic Congestion
Model
Figure 4.10.10. Urban Respiration Project – GIS Modeling
194
Air
Pollution
Modeling
Methodology
To illustrate the modeling approach, we will focus on two of the five models needed in
Figure 4.10.10 - the stack emission model and traffic congestion model. The methodology,
modeling process, and modeling results will be discussed in the following parts. Surface level
trace gas concentrations are the result of integrating all the models. To simplify the spatial
aggregation, we create a regular grid across the study area – a 20x20 km square surrounding the
Dorchester and lower Roxbury areas that were the location of our field experiments. We chose a
grid cell size of 200m x 200m so the resulting cell matrix would be reasonably fine grained
without being computationally unwieldy. This cell matrix contains 100 rows and 101 columns
and the grid cells are overlaid on top of the data layers for each model component. The approach
is to develop a grid cell value for each grid cell and for each data layer that represents the
concentration estimate that we expect in that grid cell from each source. We also have trace gas
concentration measurements for some of these grid cells. The trace gas measurement is obtained
from a mobile van that travels around the city and measures the concentration of trace gases in
real-time at the rate of one measurement every one to six seconds. This actual measurement can
be viewed as the integrated result of all the models: That is:
trace gas measurement = f (land use, stacks, roads, traffic, wind).
Spatial regression analyses are performed to model observed trace gas concentrations as a
function of each model’s result. The parameters of this estimated function can then be used to
interpolate a baseline concentration surface for all surface level grid cells in the urban area and
these estimates can be used to initialize volumetric atmospheric models.
Integrating GIS and RDBMS in the Modeling Process
As indicated in Section 4.10.1 for the Manchester experiments, we found it necessary to
use relational database management systems (RDBMS) to manage many of the requisite data
layers. The RDBMS not only facilitates the data management, but it makes it easier to systemize
the modeling process and make the models more repeatable when some modeling parameters
change. For example, SQL scripts and stored procedures can be written and saved to filter the
trace gas observations and even to run the wind and dispersion models discussed below. At
present the GIS software is rather weak at streamlining the modeling process and high-speed
calculation. For example, we have tried to use ArcView’s “Model Builder” extension to
automate the modeling process but gave up because of its limited functionality. (The Summer,
2002, release of a much improved ‘model builder’ by ESRI is claimed to have the desired
features).
In this project, the GIS software we used was ESRI’s ArcView and Arc/Info and the
RDBMS is Oracle 8i. The mechanism to connect them includes two parts as shown in Figure
4.10.11: from ArcView to Oracle; from Oracle to ArcView. When we transfer a table from
ArcView to Oracle, first the table is exported as an Arc/Info table. Then the info table is
transferred to Oracle through the Arc/Info command “infodbms”. This transferring process has
195
ArcView
table
Arc/Info
table
Oracle
table
Oracle
table
Automated by an Avenue script
Database
Access
extension ArcView
table
Figure 4.10.11. The mechanism of connecting GIS and RDBMS
been automated by an Avenue script. When we pull an Oracle table into ArcView, the ArcView
extension “Database Access” is used to directly connect to the database and get the table
(Figure 4.10.11).
4.10.2.3 Stack Emission Model
Stack emission data are obtained from the Environmental Protection Agency’s (EPA)
Envirofacts data warehouse in the same manner that we used for Manchester (see section 4.10.1).
The data we get are for each stack within the study area. We aggregate the data at the facility
level (since more precise x-y coordinates for each stack are often missing), We then consider all
the stacks within the same facility as one point and add up all their emissions. Two kinds of trace
gas emissions are taken into account: NO and CO2. Then we pull the resulting table back into
ArcView and convert it to a point theme called “AFS_PLANTSTACK_PRJ.SHP”.
Data processing in ArcView involves the following steps: First, we perform a "point-inpolygon" operation to join the attribute table of AFS_PLANTSTACK_PRJ.SHP to that of the
base grid cell matrix. We need to do this in order to assign a matrix cell ID (gridcode) to each
stack, therefore making it possible (in the RDBMS) to aggregate stack emissions within one cell
to a single value for the model calculation. Once this is done, we select all the records in the
joined table whose gridcode > 0 and export the selected records to a new dBase table,
stacks_sa3gd.dbf , for use in ArcView. We used an Avenue script to transfer this dBase table
back to Oracle.
The Simple Dispersion Model (at surface level)
A simple dispersion model – an Inverse Distance Weighted (IDW) model, is applied to the table
to simulate the distribution of trace gas concentration at surface level. For each cell of the matrix,
we consider the adjacent 11x11 square cells (approximately 2000m x 2000m areas) in any
direction surrounding it as its “neighbor cells”. The formulae of the IDW calculation for the
center cell and its neighbor cells are as given in Figure 4.10.12:
196
Concentration of Cell(i, j):
Center cell(i, j) = (1/N2) x Eij x (1/0.05)2
E: emissions, i = 1 … N, j = 1 … N
Neighbor cell(a, b) = (1/N2) x Eab x (1/(0.2 x Dab))2
Dab = Sqrt ((i-a)2 + (j-b)2)
Concentration of cell(i, j) = center cell(i, j) + SUM
(all the neighbor cells)
Note:
1. Distance unit: kilometer
2. Center cell distance = 50m = 0.05km
3. Grid size = 200m x 200m = 0.2km x
0.2km
Figure 4.10.12. Inverse Distance Weighted (IDW) Calculation of Trace Gas Concentrations
The dispersion modeling is calculated in Oracle. The whole process can be illustrated by
the flow chart in Figure 4.10.13.
A cell’s final value equals its value at the center cell plus the sum of all its neighbor cells’
value. The scope of the neighbor cells can be adjusted according to different types of pollutant
and conditions to reflect the real dispersion effect. For the stack emission model, we define the
neighbor cells as 11x11 cells around a center cell. In other words, we think that the dispersion
effect on a cell from a stack more than 1000 meters away can be ignored. Before the model
calculation the original cell value is the total emission amount of all the stacks located within that
cell. If there is no stack in a cell, the cell value is zero.
197
In ArcView, perform "pointin-polygon" joining to get
the table
stacks_sa3gd_jn.dbf
Dump the stacks_sa3gd_jn.dbf
to Oracle via Arc/Info, using
Avenue Script: dbf2ora.ave. Get
table: in_ora
Dbf2ora.ave
in_ora (table)
 Containing fields: emission_co,
emission_no2, gridcode
 Containing incomplete cellids, only
cells having stacks appear in this
table
 Existing duplicated cellids since
there may be more than one stack
within a cell Sql_1
gd200mtri_sum (view)
 Summing up the emission_co and
emission_no2 respectively of all
the stacks within the same cell
 Containing fields: emisCO,
emisNO2, gridcode
Oracle
URBANRESP Database
gd200msa3 (table)
 Containing fields:
cellid, x_coordinate,
y_coordinate , rowi,
colj
Simple Join two tables
Insert records whose cellids don’t appear
in “gd200mtri_sum” to the joined table
Sql_2
gd200mtri (table)
 Containing fields: cellid, x_coordinate,
y_coordinate, rowi, colj, emisCO,
emisNO2
 Containing complete cellids from 1 to
10100
IDW Calculation
IDW Calculation in Oracle (11x11 cells
square neighborhood)
Get table: “gd200mtri11”
Transfer the tables to ArcView
Transfer the resulting Oracle table to
ArcView via “SQL Connect” extension.
Join this table to the attribute table of
"sa3gd.shp"(base grid matrix) by common
cellid. Then display the distribution of CO
andChart
NO2 in
Figure 4.10.13. Flow
of ArcView.
Method for Dispersion Model Calculations.
198
The following is part of the SQL queries in the “IDW Calculation” step:
/* Center cell computation: (weight = 1 for 11x11 window)
create view gd200mcenter as
select g.id, g.emisco, g.emisno2, g.rowi, g.colj,
(1/121) * 400 * g.emisco emisco_c,
(1/121) * 400 * g.emisno2 emisno2_c
from gd200mtri g;
/* neighbor cells computation:
create table gd200mtri_wt11 as
select g.id, g.rowi, g.colj,
sum(h.emisco *(1/121) * 25/(power((g.rowi-h.rowi),2) + power((g.colj-h.colj),2)) ) emisco_s,
sum(h.emisno2*(1/121) * 25/(power((g.rowi-h.rowi),2) + power((g.colj-h.colj),2)) ) emisno2_s
from gd200mtri g, gd200mtri h
where g.rowi > 5 and 100 – g.rowi >= 5
and g.colj > 5 and 101 – g.colj >= 5
and abs(g.rowi – h.rowi) <= 5
and abs(g.colj – h.colj) <= 5
and (g.rowi <> h.rowi or g.colj <> h.colj)
group by g.id, g.rowi, g.colj;
/* Join the two tables to get total for each cell:
create index grid200id on gd200mtri_wt11(id);
create view gd200mtri11 as
select g.id, g.rowi, g.colj, g.emisco, g.emisno2, emisco_c + emisco_s emisco_wt11,
emisno2_c + emisno2_s emisno2_wt11
from gd200mcenter g, gd200mtri_wt11 w
where g.id = w.id;
The final Oracle table is pulled back to ArcView via the “Database Access” extension.
Then this table is joined to the attribute table of the base grid matrix via the common cell ID in
order to display the concentration distribution of CO and NO2. The modeling results are
illustrated in Figure 4.10.14 and 4.10.15. The estimated concentrations are classified based on
the standard deviation of the Z-score of CO and NO2 concentrations respectively.
199
Figure 4.10.14. Estimated CO Concentration Distribution After the Stack Emission Model
Figure 4.10.15. Estimated NO2 Concentration Distribution After the Stack Emission Model
200
4.10.2.4 Traffic Congestion Model
Vehicle emissions are estimated using the proximity-to-road approach described earlier
for the Manchester experiments (see section 4.10.1) plus a model that estimates the location of
traffic congestion. The traffic congestion model is constructed by scaling and combining spatial
features of the road network in ways that are computable using GIS and RDBMS tools and
readily available data. We identify the locations of road intersections and highway exits and
estimate the traffic congestion level at these locations. Ultimately, vehicle counters and other
monitors will enable these models to be improved. We assume that traffic congestion tends to
depend on the number of traffic lanes converging at a specific intersection or exit. The
hypothesis is that part of the surface level trace gas concentration (P) is proportional to the level
of traffic congestion at the road intersections and highway exits: P = W * N (W is the weighting
coefficient; N is the number of lanes. W can be calculated through the later spatial regression
analysis which combines all the models.). The traffic congestion model has been divided into
two parts: road intersections and highway exits. It doesn’t take into account the traffic on the
roads appart from the intersections because that has been simulated by the road model explained
in Section 4.10.1.
As for the intersections, three types of road have been identified, as well as their lane numbers:
 Class 2 - Multi-lane Highway, not limited access; 3 lanes
 Class 3 - Other numbered route; 2 lanes
 Class 4 - Major road – connector; 1 lane
When two roads intersect, the intersection is assigned a value equal to the total number of
lanes of these two roads. The same principle applies to three-road intersections and so on. We
handle exits from Class 1 - Limited Access Highways, separately and weighted them as if they
were 4 intersecting lanes.
The data layer of major roads is obtained from MassGIS website:
http://www.state.ma.us/mgis/majrdmhd.htm. It is a statewide arc coverage called MAJRDMHD,
containing all four classes of roads mentioned above. In ArcView we process this data layer by
removing Class 1- Limited Access Highway from it and clipping it to the scope of our study
area. Then we get a shape file called “ROAD234_IN3.shp” which contains all the class 2, 3, and
4 roads within the study area. In order to calculate the traffic congestion distribution at road
intersections, first we must identify all the intersections and their total converged lane numbers.
This process is completed in Arc/Info and Oracle. The criteria to judge a node as an intersection
is that its appears in the attribute table of ROAD234_IN3 as a FNODE# or a TNODE# more than
twice1. The first step of this process is building the topology of ROAD234_IN3.shp in Arc/Info
to get the arc coverage ROAD234_IN3. Then dump its attribute table (AAT) to Oracle through
the connection between Arc/Info and Oracle.
1
ROAD234_IN3 is an arc coverage built from ROAD234_IN3.shp in Arc/Info. Its attribute
table contains the node information of each arc. The column of “FNODE#” in that table means
the internal sequence number of the from-node. The column of “TNODE#” means the internal
sequence number of the to-node.
201
Based on this attribute table, a series of SQL queries are performed to find all the road
intersections and the number of lanes converging there. The process can be illustrated as the
following flow chart:
Create table: Fnodecount
Counting the frequency
of FNODE#
Create table: Tnodecount
Counting the frequency of
TNODE#
1.1.1.1.1.1
Join
two
tables
Create view: Totalcount
GET THE X, Y COORDINATE
INFORMATION FOR EACH
INTERSECTION NODE
Counting the total frequency
of a node as a FNODE or a
TNODE
1.1.1.1.1.2
Find
Create view: Internodes
intersAssign lane
Selecting nodes whose total ectionnumbers to
frequency is greater than 2 s
each node
Join two tables and sum up lane
numbers for each intersection by
“group by NODE#”
Get
“road234_in3c.NAT” in
Arc/Info by commands:
Arc: BUILD
road234_in3c NODE
Arc:
ADDXY
Pull
this
road234_in3c
table to
NODE
Get table: Nodescoord
1.1.1.1.2
Oracle
Create table: Allintersections
Containing columns: NODE# -- node number
TOTALLANES – lane numbers for an intersection
1.1.1.1.2.1
Containing columns:
NODE#, X coordinate, Y
coordinate
Join
two
tables
Create table: intersections
Containing columns: NODE#,
TOTALLANES, X_COORD,
Y_COORD
After creating the table “intersections”, we pull it into ArcView via the "Database
Access" extension. Then in ArcView, we create an event theme based on this table, using
X_COORD and Y_COORD as X and Y coordinate fields. This point theme contains all the road
intersections and their traffic congestion information. A visual display of the identified
intersections and highway exits is shown in Figure 4.10.16
The data layer of highway exits locations is also obtained from MassGIS website:
http://www.state.ma.us/mgis/majrdmhd.htm. This statewide point coverage includes the exit
number and the highway route number associated with each exit. This data layer is processed in
ArcView by being clipped to the scope of the study area and having a weight equivalent to 16
traffic lanes assigned to each exit in the attribute table of the coverage.
202
Figure 4.10.16. Find all the Road Intersections and Highway Exits
Data Processing in ArcView and Oracle
The data processing in ArcView is similar to that of the stack emission model:




Join the attribute table of road intersections to that of the base grid matrix;
Join the attribute table of highway exits to that of the base grid matrix;
Export the selected records to two dBase tables: intersections_sa3gd.dbf and
exits_sa2gd.dbf;
Run the Avenue script to transfer these two dBase tables to Oracle.
An IDW dispersion model is applied to the above two Oracle tables, following the similar
formulae and procedure of the stack emission model. The neighbor cells are also defined as
11x11 square cells around the center cell. The difference is to substitute the stack emission value
of each cell with the traffic congestion value measured by the number of traffic lanes. After the
calculation two tables are generated, describing the traffic congestion distribution around all the
intersections and exits locations.
203
These two final tables are pulled back to ArcView via its “Database Access” extension.
Then they are joined to the attribute table of the common grid matrix respectively. The resulting
contributions to trace gas concentrations that result from the dispersion of road intersection
congestion is shown in Figure 4.10.17. Once again, the shading is based on standardized z
scores of the concentration estimates.
Figure 4.10.17. Estimated Trace Gas Concentration from Traffic Congestion
around Road Intersections
4.10.2.5 Emission Modeling Conclusions
These stack emission, road, and traffic congestion models are three of the five components of the
surface level trace gas concentration model diagramed in Figure 4.10.10 above. The land use
model also used the 200mx200m grid cells and was based on a combination of population
density and the percentage land in the neighborhood of each cell that was use for residential,
commercial, transportation, and open space (or water). The wind model simply shifted all
estimates 1, 2, or 3 grid cells in the direction that the prevailing wind was blowing. All the
models were computable from readily available data using standard GIS procedures plus SQL
queries in the relational database manager. (We ended up doing all the IDW weightings in the
RDBMS after encountering problems with the equivalent procedure in ArcView when there were
204
large neighborhoods or many cells). The major advantage of incorporating RDBMS in the
modeling process is to improve the accuracy, efficiency, flexibility and repeatability of the data
processing pipeline. The RDBMS has good tools complex filtering and querying of the data and
for storing and automating the procedures. The modeling procedures involve data preparation
and transferring, dispersion modeling, and results mapping. All five models are combined with
several parameters to explain and estimate the surface level trace gas concentrations. By
correlating the estimates with trace gas measurements obtained from suitably equipped mobile
vans, we can estimate the model parameters. The next section explains the spatial regression
analyses use to calculate the influence of each factor on the urban air quality at the surface level.
Once such calculations are made, the composite modeling results then can be used to interpolate
surface level concentrations throughout the study area which can, in turn, be used as input for the
3-dimentional ambient air pollution modeling work of other groups in our research team. While
our study focused on metro Boston, similar raw data layers for other urban areas are increasingly
available on the Internet and the methodology we propose could be applied in a reasonably
standardized way for many metropolitan areas. The next section discusses the practicality of
such GIS modeling efforts as well as the spatial regression modeling.
205
4.11 Comparing GIS-based Models and Trace Gas Observations
This section compares the trace gas observations from the Boston field experiments with
the estimates developed from the GIS modeling described in Section 4.10. In addition, we use
the modeling experience to develop conclusions about the limits and practicality of using GIS
modeling tools and distributed GIS architectures to support complex environmental monitoring
and modeling.
In experimenting with the GIS-based models, we are trying to prototype methods and
system architectures that can be good templates for improved models. Toward this end, we seek
models that are modular in nature and able to capitalize on emerging technologies and
improvements in environmental monitoring. We have decomposed the emission modeling task
along the lines suggested in Figure 4.11.1. We have focused on the most significant and
straightforward activities that can be well-captured by GIS modeling.
Our GIS modeling has generated, for each 200mx200m grid cell, a parametric equation
estimating the emissions contribution of traffic, stack, and land use activities. In addition, we
developed a simple wind model (in SQL) that models wind by shifting the grid cell values in the
direction of the wind. If we assume no atmospheric chemistry or vertical dispersion, then the
emission estimates (for any particular trace gas) should be additive and, in principle, we should
Figure 4.11.1. Flow Chart of Emission Modeling Approach.
206
now be able to calibrate the model parameters from the mobile van’s trace gas observations
during the Boston field experiments. Of course there is some chemistry, there are many degrees
of freedom in our models, and the linear assumptions and simple dispersion estimates are not
ideal. Nevertheless, if we picked an observation day with clear skies, a healthy southwest
(prevailing) wind, and an early morning start, then we should have a shot at finding reasonable
model parameter to explain surface level trace gas concentrations at the start of a daily cycle.
The limited time available and the logistics of handling the several mobile vehicles
prevented us from getting pre-rush hour runs through the study area. But the research team did
make several afternoon and evening passes and the weather conditions were reasonably
consistent that day and suggested little atmospheric chemistry. In the next sub-section, we
describe the equations and techniques used to calibrate and test the model.
Figure 4.11.2 maps the land use pattern in the study area. The map shows the land use at
the center of each 200mx200m grid cell. The variables used in the model are the percentage of
land in the neighborhing 25 grid cells that are of each land use type. Figure 4.11.3 shows the
population density pattern in the study area (total population per acre).
Figure 4.11.2. Land Use in the Boston Study Area
207
Figure 4.11.3. Population Density (total population per acre) for the Boston Study Area
4.11.1 Spatial Regressions of the May 25, 1999, Trace Gas Observations
Which set of trace gas observations should we compare with the model estimates?
Extensive study of the spatial and temporal patterns of trace gas observations by the Aerodyne
team were reported earlier. We used the ‘shaved’ data from May 25 for our model tests. These
shaved data filtered out observations that, due to the NO/CO2 ratio and other tests, were
considered either representative of local point sources and not the general pollutant levels, or
were otherwise considered unreliable. The remaining NO and CO2 observations were averaged
within their 200mx200m grid cells (for the selected time periods). These calculations yielded
about 300 grid cells in the study area with mobile van observations. The average NO
concentrations are shown in Figure 4.11.4. The observed average concentrations were then
regressed against a linear function of the model components (after scaling the road/congestion
model components to standard z scores). Instead of running ordinary least squares, we used
S*Plus Spatial Statistics to run a ‘spatial’ regression that tried to account for lack of
independence in the observation because of spatial autocorrelation. The ArcView extension for
Spatial Statistics made it easy to move the grid cell values into S*Plus for the statistical analysis
and then back into ArcView for mapping.
208
The right hand side variables for the regression are
May25sw3f2shv.rm2scsa3$PCT.OT7FIX
May25sw3f2shv.rm2scsa3$PCT.W7FIX
May25sw3f2shv.rm2scsa3$PCT.C7FIX
May25sw3f2shv.rm2scsa3$PCT.I7FIX
May25sw3f2shv.rm2scsa3$PCT.OP7FIX
May25sw3f2shv.rm2scsa3$LANESWT.ZSCORE
May25sw3f2shv.rm2scsa3$EXITSWT.ZSCORE
May25sw3f2shv.rm2scsa3$RD.ZSCORE
=
=
=
=
=
=
=
=
landuse ‘other’ percent
landuse ‘water’ percent
landuse ‘commercial’ percent
landuse industrial percetn
landuse ‘open space’ percent
intersection ‘lane’ z-score
highway ‘exit’ z-score
road density z-score
The percentage of land that is in residential use is the base case and is not included in the
regression. The prefix to the variable names indicates the use of May25 data with a southwest
wind model of 3 grid cells and the runs used the ‘shaved’ observations. A typical set of 4 runs is
shown in Figure 4.11.5. The four are for ‘no wind shift’ and southwest wind shifts of 1, 2, or 3
grid cells. That day, the prevailing wind was southwest. The one-cell southwest wind shift
produces the lowest residual standard error – although not by much. The high-low difference is
only about 10%. The second best model is no-wind. In both cases, the commercial/industrial
landuses have statistically significant coefficients and signs in the expected direction – positive,
indicating higher observed concentrations (compared to the base residential case) in area that are
more commercial or industrial. But the percent of land that is water also has a positive
coefficient for the SW1 case (and an insignificant coefficient in the no-wind case. Likewise, one
of the road congestion terms (highway exits) has a significant wrong sign in the SW1 case. The
no-wind case looks more plausible.
Figure 4.11.6 plots the NO residuals (as standard deviations) for the SW3 model. Red
values show overestimated cells and blue values should underestimated. The spatial regression
tends to avoid substantial spatial clustering of residuals, but the predictions still leave a lot to be
desired.
209
Figure 4.11.4 Average Observed Values of NO (ppb) during May 25, 1999 Mobile Van Runs
210
sw3_conges2_baseR.txt
Value Std. Error t value Pr(>|t|)
(Intercept) 9.4991 0.9796
9.6968 0.0000
May25sw3f2shv.rm2scsa3$PCT.OT7FIX -0.0183 0.0430
-0.4245 0.6715
May25sw3f2shv.rm2scsa3$PCT.W7FIX 0.0116 0.0653
0.1772 0.8595
May25sw3f2shv.rm2scsa3$PCT.C7FIX 0.0703 0.0284
2.4728 0.0140
May25sw3f2shv.rm2scsa3$PCT.I7FIX 0.0687 0.0452
1.5210 0.1293
May25sw3f2shv.rm2scsa3$PCT.OP7FIX -0.0708 0.0219
-3.2252 0.0014
May25sw3f2shv.rm2scsa3$LANESWT.ZSCORE 0.7255 0.2731
2.6560 0.0083
May25sw3f2shv.rm2scsa3$EXITSWT.ZSCORE -0.0123 0.1722
-0.0714 0.9431
May25sw3f2shv.rm2scsa3$RD.ZSCORE -0.1396 0.2803
-0.4980 0.6188
Residual standard error: 2.77583 on 302 degrees of freedom
sw2_conges2_baseR.txt
Value Std. Error t value Pr(>|t|)
(Intercept) 8.9155 0.9703
9.1884 0.0000
May25sw2f2shv.rm2scsa3$PCT.OT7FIX 0.0710 0.0469
1.5118 0.1316
May25sw2f2shv.rm2scsa3$PCT.W7FIX -0.0086 0.0568
-0.1509 0.8801
May25sw2f2shv.rm2scsa3$PCT.C7FIX 0.0162 0.0277
0.5852 0.5589
May25sw2f2shv.rm2scsa3$PCT.I7FIX 0.0597 0.0483
1.2359 0.2175
May25sw2f2shv.rm2scsa3$PCT.OP7FIX -0.0781 0.0242
-3.2340 0.0014
May25sw2f2shv.rm2scsa3$LANESWT.ZSCORE 0.1238 0.1863
0.6644 0.5069
May25sw2f2shv.rm2scsa3$EXITSWT.ZSCORE 0.5695 0.1286
4.4293 0.0000
May25sw2f2shv.rm2scsa3$RD.ZSCORE 0.1801 0.2622
0.6869 0.4927
Residual standard error: 2.82466 on 302 degrees of freedom
sw1_conges2_baseR.txt
Value Std. Error t value Pr(>|t|)
(Intercept) 7.6975 1.0700
7.1941 0.0000
May25sw1f2shv.rm2scsa3$PCT.OT7FIX 0.0156 0.0473
0.3295 0.7420
May25sw1f2shv.rm2scsa3$PCT.W7FIX 0.1045 0.0464
2.2531 0.0250
May25sw1f2shv.rm2scsa3$PCT.C7FIX 0.0691 0.0263
2.6292 0.0090
May25sw1f2shv.rm2scsa3$PCT.I7FIX 0.1291 0.0482
2.6759 0.0079
May25sw1f2shv.rm2scsa3$PCT.OP7FIX -0.0330 0.0259
-1.2737 0.2038
May25sw1f2shv.rm2scsa3$LANESWT.ZSCORE 0.1481 0.1336
1.1085 0.2685
May25sw1f2shv.rm2scsa3$EXITSWT.ZSCORE -0.1951 0.0893
-2.1838 0.0297
May25sw1f2shv.rm2scsa3$RD.ZSCORE -0.2092 0.2104
-0.9943 0.3209
Residual standard error: 2.62654 on 302 degrees of freedom
nowind_conges2_baseR.txt
Value Std. Error t value Pr(>|t|)
(Intercept) 7.1760 1.1088
6.4717 0.0000
May25nowindf2shv.rm2scsa3$PCT.OT7FIX 0.0326 0.0486
0.6711 0.5027
May25nowindf2shv.rm2scsa3$PCT.W7FIX 0.0599 0.0369
1.6257 0.1051
May25nowindf2shv.rm2scsa3$PCT.C7FIX 0.0588 0.0270
2.1774 0.0302
May25nowindf2shv.rm2scsa3$PCT.I7FIX 0.0924 0.0451
2.0477 0.0415
May25nowindf2shv.rm2scsa3$PCT.OP7FIX -0.0158 0.0270
-0.5877 0.5572
May25nowindf2shv.rm2scsa3$LANESWT.ZSCORE -0.1028 0.0926
-1.1095 0.2681
May25nowindf2shv.rm2scsa3$EXITSWT.ZSCORE 0.0980 0.0376
2.6082 0.0096
May25nowindf2shv.rm2scsa3$RD.ZSCORE 0.1828 0.2131
0.8577 0.3917
Residual standard error: 2.70482 on 302 degrees of freedom
Figure 4.11.5. Typical Regression Results for May 25 NO Observations
211
Figure 4.11.6. May 25 NO Residuals (as standard deviations) for the SW3 model
There are a number of reasons why we might expect the regressions to be of limited
value. In general the mobile van made two traversals of the study area several hours apart with
the return trip during rush hour. The conditions are likely to be different later in the day when a
seabreeze from the east began competing with the prevailing southwesterly in some parts of the
study area. The WSU wind instruments travelled with the mobile van during parts of the day and
confirmed that the local wind was not always southwest.
Before trying the spatial regressions, we did a number of more exploratory analyses. For
example, we identified, by overlaying van traversals on orthos land use and density maps, places
where the van made parallel runs, close together in time, but through local streets and
neighborhoods with different land use (e.g., residential for one and commercial/industrial for the
other). In some cases, we were able to observe the expected patterns – lower NO and CO2 values
in the areas downwind of open space and residential and higher values in the areas downwind of
commercial/industrial. But the shifting local winds – and the limited number of van traversals
and wind readings - that day seemed to limit our ability to capture the active relationships.
212
4.11.2 System Architectures for Environmental Monitoring and Modeling
Within the research team, we have discussed but not pursued in detail more complex
model linkages between the GIS models of urban activity and the volumetric models of
atmospheric chemistry and diffusion. For example, the land use data can be used to estimate
surface roughness – a key factor in modeling atmospheric chemistry. At this stage, however, we
did not want to address chemical interactions in the GIS modeling. Nevertheless, the land use
data that we do include have breakdowns of land use at the sub-kilometer scale. These
percentages can be used to develop surface rouhgness measures at 3 km and greater scales.
Other studies (such as the EPA-sponsored Georgia Tech experiments in Atlanta) are developing
detailed experiments to monitor and measure vehicle emissions by age of vehicle, slope of road,
temperature of engine/air, etc. During the next few years, such studies and improved traffic
congestion information systems will provide better vehicle emission data and models that can
easily be used within the modeling architecture envisioned in this study.
These frustrations with our limited ability to do more data mining of the observations
highlight one of the important findings of the study. If we had been able to collect and analyze
even a sample of all our observations within a few hours, if not in realtime, then we could have
adjusted the mobile van traversals to make observations in the times and places that were most
likely to be relevant. By the end of the study, the various teams had streamlined most of the
analysis procedures to where this was conceivable. At the same time, GIS technologies have
begun to appear with sufficient modeling tools, distributed computing capabilities, and
interoperability standards that it will soon be practical to do near realtime data acquisition and
analysis using off-the-shelf tools without limited amounts of custom programming.
The difficulties that we encountered in automating the GIS ‘pipeline’ and utilizing web
mapping technologies are worth enumerating and were sufficient to slow down the interaction of
the research teams when we weren’t all together for the field experiments. When we were doing
the experiments, the overly complex analysis ‘pipelines’ forced us to archive most of the data for
later use. Listing a few of the impediments is instructive in explaining just how many
calculations and technical glitches would bog down the process. The ArcView-IMS software
used to provide Web mapping in the first and second year required too much Java download
and/or client PC capability and tended to cause research team member’s PC to lock up. In one
case, the user had too small a display and too limited a color table to make the maps usable. The
current ‘model builder’ tool is a long ways from a universal modeling language (UML) tool that
could greatly facilitate the integration of GIS tools with complex engineering models. The
current version can not mix and match vector and raster operations of sufficient complexity to be
helpful in automating our data processing pipeline.
One example of specific difficulties involved the use of different coordinate systems.
Local data tended to be in a State Plane coordinate system (to preserve areas). The
meteorological models use a Lambert conical projection – and the raster grid cells in that
coordinate system are no longer square grid cells when projected to state plane systems.
Moreover, the grid cell boundaries in the MM5 models are based on a starting center point and
not easily registered to any grid cell boundaries that might be generated in a typical GIS system.
213
The tens of thousands of data points collected each observation day need considerable
filtering and analysis before they are sufficiently robust to be used in modeling. Dumping them
directly into a networked RDBMS database engine would make a lot of sense. Then stored
procedures could be written to automatically check and transform the data – transform
coordinates to state plane, adjust GMT time, filter out bad NO/CO2 readings, detect subtle
system anomolies, etc. Current wireless technologies now make it practical for the mobile van to
do live reporting of GPS and trace gas readings. Once this step is taken, any number of users
could study the data in near real-time. Additional team members back at the office could
examine the data and plot afternoon traversals. At the same time, metropolitan areas are
improving their base mapping of parcels land fills and other large-scale features that could
greatly assist in calibrating the emission models. New datasets are also coming online for traffic
congestion reporting, air quality monitoring stations, and the like. With these datasets for ground
truthing and the right system architecture to support complex but flexible engineering modeling,
near realtime analysis and modeling will become practical. When this occurs, it will be come
practical to use satellite imagery for the mobile component of environmental monitoring in order
to significantly increase the scope and replicability of air quality models.
Modeling the environmental implications of land use and transportation is a complex,
multidimensional and multidisciplinary undertaking that has begun to involve a growing number
of researchers and Federal agencies in the US and around the world. Progress in this area will
depend on the gradual improvement, standardization, and interoperability of many model
elements, data layers and geoprocessing methods. For this reason, we have focused on exploring
the capabilities and limitations involved in using readily available datasets and current GIS and
RDBMS technologies. The idea is to see how easily we can assemble and interconnect relevant
data layers and models into a data processing 'pipeline' that can be easily replicated and return
with improved datasets and modeling components. The results are promising but highlight the
need for improved modeling languages and 'process modeling' tools (such as UML and ESRI's
'model builder'). Current versions of these tools are too limited and inflexible in their handling
of modeling complexity, interoperability, network transparency, and user interface.
Improvements in metropolitan information infrastructures along with increased interest in
environmental monitoring and real-time traffic information systems is beginning to provide the
data needed to make such modeling pipelines practical and reliable. But the model elements are
too complex, multidisciplinary, and evolutionary to be manageable, using today's GIS tools, as a
collection of distributed and interoperable modeling components. As the GIS technology evolves
from standalone programs to distributed geoprocessing components the type of modeling
systems described in this paper will become increasingly practical and portable.
214
4.12 Fine Aerosol
The concentration of fine particles (~7 – 3000 nm diameter) was measured through nonheated and heated (300ºC) inlets using a TSI model 3022A condensation particle counter. Tests
showed that the heated inlet (non-volatile fraction) volatilized nitrates, sulfates, and organic
materials. The non-volatile fraction in an urban environment is typically composed of fine
crustal dust, soot carbon, and high molecular weight organic materials.
The 1999 Boston campaign included both mobile and stationary sampling regimes.
Mobile samples were collected while traveling through and around major city thoroughfares
while stationary sampling was conducted over a several day period on the Massachusetts
Institute of Technology campus in Cambridge. In Figure 4.12.1 we show typical data for fine
particles collected during highway runs around the Boston area.
The largest concentrations of fine particle were observed in the vicinity of traffic lights or
congested merging areas, suggesting that mobile sources were a significant source of both total
and non-volatile particles. The variability and differences in total and non-volatile particle
sources is evident from the fact these two parameters were not always correlated positively.
“Hot spots” were apparent for both parameters, and these were usually not located
concomitantly. Interestingly, the total particle, but not the non-volatile, concentration was quite
elevated in the area of the “Big Dig” in Boston. This can be identified as the green colored
highway area in the upper right corner of the left panel. These particles could be related to the
extensive use of heavy machinery for highway excavation and construction in the area. We
speculate below that our aerosol filter collections during stationary measurements in Cambridge
actually sampled suspended particles from the “Big Dig” construction activities.
During the stationary sampling on the MIT campus during May 27-29, 1999 we observed
high concentrations of particulate calcium (Figure 4.12.2). Calcium is a tracer of soil dust,
cement, and other alkaline materials. Calcium and sulfate values shown in Figure 4.12.2 have
been corrected for contribution from sea salt (i.e., non-sea-salt concentrations shown only). In
the late afternoon on May 28 we observed what appeared to be a sea breeze front move into the
greater Boston area. This is evident by the very high levels of sodium and chloride. Coincident
with these were elevated concentrations of non-sea-salt calcium. It is highly likely that the sea
breeze brought suspended fine dust and other alkaline materials into the Cambridge area that
were associated with “Big Dig” activities to the east of our sampling site. This was a very well
defined event, which ended around 18:00 local time, just about the time the sea breeze would be
expected to retreat back toward the ocean as the land surface began to cool.
Throughout the night and into the early morning hours of May 29 a major pollution
episode influenced the air quality at the Cambridge site. Here ppbv levels of ammonium and
nitrate were observed with lesser amounts of sulfate aerosol. It is highly likely that the high
levels of nitrate were directly related to the significant mobile sources of reactive nitrogen in the
area. It also points to large urban or mobile sources of ammonia. Mobile sources have been
proposed to be a significant source of ammonia in the Los Angeles basin leading to production of
215
Total Counts
Non-Volatile Counts
283001-1350000
45000-803000
137001-283000
30000-45000
56301-137000
15000-30000
5450-56300
2680-15000
Figure 4.12.1. Concentration of fine particles (particles cm-3) along Boston roads and highways
on May 23, 1999. The left panel the concentration of total fine particles
(unheated) while the right one is the non-volatile fraction (heated). The Ferriera
group at MIT generated these GPS-GIS based maps.
fine light scattering aerosols after reaction with nitric acid [Fraser and Cass, 1998]. Our data for
the Boston area tends to support their findings. On a global scale, biogenic and agricultural
sources dominate the ammonia budget [Deneter and Crutzen, 1994]. This may not be the case in
urban areas, and this issue clearly requires closer attention in future studies.
216
Multiple mechanisms are possible for formation of nitrate aerosols in an urban
environment, including direct reaction of nitric acid and ammonia, reaction/uptake of N2O5 and
HONO at nighttime, and uptake of nitric acid on basic particles such as crustal and highway
abraded dust (e.g., cement) [Roberts, 1995]. The relative amount of nitrate and sulfate aerosols
in this urban area is the opposite of what has been observed upwind of the Boston area in central
Massachusetts. At Harvard Forest near Petersham, MA, Lefer and Talbot [2001] nearly always
observed much higher concentrations (factors of 2 –20) of sulfate compared to nitrate aerosols.
Only during periods of high anthropogenic influence did nitrate concentrations exceed 100 pptv.
These results appear to be very consistent with an urban source for the enhanced levels of
ammonium nitrate aerosol occasionally observed at Harvard Forest.
1600
Mixing Ratio, pptv
Sodium
Chloride
Magnesium
nss-Calcium
1200
800
400
0
12:00:00
20:24:00
4:48:00
13:12:00
21:36:00
6:00:00
14:24:00
6:00:00
14:24:00
2000
Mixing Ratio, pptv
Ammonium
nss-Sulfate
Nitrate
1500
1000
500
0
12:00:00
20:24:00
4:48:00
13:12:00
21:36:00
Local Time
Figure 4.12.2. Selected water-soluble composition of urban aerosols sampled on the MIT
campus during May 27-29, 1999. About 95% of total calcium and sulfate were of
non-sea-salt origin.
217
The aerosol composition data collected during this project emphasizes the complex
nature of urban aerosols, and the diversity of sources and mechanisms influencing the
distribution and concentration of fine particles. Since many large cities in the U.S. tend to occur
along the coastline, they too are likely to have a strong but variable marine influence. This
probably has important implications for gas phase chemistry, particularly that of reactive
nitrogen, which have yet to be realized. This project provided a first look at the complexities
offered by studying heterogeneous atmospheric chemistry in a coastal urban environment.
A summary of the CO2 and fine particle data for May 28, 1999 is shown in Figure 4.12.3.
The data were initially collected at 1 Hz, averaged to a one-minute interval, and then binned as
hourly segments. The peak that occurs around 16 hours after midnight is probably due to rush
hour traffic. This was a Friday afternoon going into Memorial Day weekend, so the traffic was
unusually heavy. There was close correspondence between elevated CO2 and fine particles
during this time interval, again suggesting that mobile sources were dominant at this time. Note
that the non-volatile fraction represents about 50% of the total fine particle concentration. This
was typical of our Boston area measurements, including the mobile-based ones. During the
instrument-testing phase of this study we sampled in rural and small communities in New
Hampshire and on the University of New Hampshire campus and rarely observed more than 30%
of the total as the non-volatile fraction. The total particle concentrations were surprisingly quite
similar between rural New Hampshire and Cambridge (i.e., on the order of 104). It would be
interesting to see if other urban areas have characteristically a larger proportion of non-volatile
particles compared to the surrounding region.
We explored additional inter-relationships between fine particles and trace gases using
data collected on May 28. Figure 4.12.4 illustrates some of these using the NO and CO data.
The correlation of fine particles and NO was slightly better than that with CO. This may reflect
the importance of mobile sources for both of these species. The sources of fine particles, NO,
CO, and even CO2 are extremely diverse in an urban environment. These plots therefore reflect
to a large extent the net effect, so a close correspondence would not necessarily be expected
between fine particles and trace gases. In addition, fine particles have both primary and
secondary sources, which adds significant confounding problems to interpreting the data. It is
possible that these relationships are better defined in the downwind sector from an urban plume.
For the Boston area this is not easy to sample, as the downwind area is usually extends out over
the North Atlantic.
218
475
CO2, ppmv
450
425
400
375
Unheated, Particles cm
-3
350
1e+5
8e+4
6e+4
4e+4
2e+4
Heated, Particles cm
-3
0
1e+5
8e+4
6e+4
4e+4
2e+4
0
0
10
20
30
Hours Since
Midnight May
1999 28.
Hours since
Midnight
on28,May
160
Particles per cubic centimeter (x 10-3)
Particles per cubic centimeter (x 10-3)
Figure 4.12.3. Time series of CO2 and fine particles during stationary measurements in
Cambridge, MA. The individual data are hourly averages and the solid line
represents a spline fit to the data.
140
120
100
80
60
40
20
0
140
120
100
80
60
40
20
0
0
1
2
3
4
0
CO, ppmv
20
40
60
80 100 120 140 160 180 200
NO, ppbv
Figure 4.12.4. Summary relationships of total fine particles with NO and CO during the
afternoon of May 28.
219
To identify trace gas – fine particle relationships for the greater Boston area,
measurements of fine particles and CO2 from the mobile runs on May 22, 23, 25 and 26 were
examined. As before, the data were collected at 1 Hz, but were averaged to one minute to look
for these inter-relationships. The highway paths taken were typical to the ones depicted in
Figure 4.12.5 (the first one shown here). Summary plots of fine particles and CO2 are shown in
Figure 4.12.6. We purposely choose to conduct measurements over a weekend and then on two
weekdays. May 22 and 23 were a Saturday/Sunday pair, and May 25 and 26 a
Tuesday/Wednesday pair. This was followed up by the stationary measurements discussed
previously that occurred on the following Thursday to Saturday interval.
May 23, 1999
May 22, 1999
300
Particles per cubic centimeter (x 10-3)
Particles fper cubic centimeter (x 10-3)
300
200
100
0
d Particles/d CO2 = 1575
350
400
450
500
550
0
d Particles/d CO2 = 1698
350
600
400
450
CO2, ppmv
May 25, 1999
May 26, 1999
500
700
-3
Particles per cubic centimeter (x 10 )
Particles per cubic centimeter (x 10-3)
100
CO2, ppmv
300
200
100
0
d Particles/d CO2 = 1712
350
200
400
450
600
500
400
300
200
100
500
350
CO2, ppmv
d Particles/d CO2 = 1829
0
400
450
500
550
600
CO2, ppmv
Figure 4.12.5. Summary of total fine particle and CO2 relationships observed during the mobile
laboratory measurements in the Boston area on four different days in May 1999.
220
We found was a surprisingly robust relationship between fine particles and CO2. The
slopes of the plots were very similar, average 1704 ± 104 particles cm-3 per ppmv CO2. We are
exploring various ways to use these data to better estimate regional Boston fine particle
emissions and determine emission factors for urban areas.
221
5.0 PROJECT OUTPUT
While we have presented or published only a small portion of our project’s output to date,
we will be preparing additional publications and presentations as the data are fully analyzed and
modeled. The following subsection, 5.1, lists presentations, as well as published symposia
papers and archival journal papers published to date. Section 5.2 lists archival journal articles
that are currently planned or in preparation. Section 5.3 lists and describes web sites where
project data archives or model/analysis results can be accessed.
5.1 Symposia Presentations, Proceedings Papers and Journal Publications
5.1.1 Presentations Without Published Proceedings Papers
Presentations at the Urban Emissions and Atmospheric Chemistry Symposium, Spring American
Geophysical Union Meeting, Boston, MA, May, 1999.
J.B. McManus, J.H. Shorter, C.E. Kolb, B.K. Lamb, E. Allwine, S. O’Neill, G McRae,
J. Ferreira, R. Talbot, P. Crill and E. Scheuer, Correlations of Pollutant Gases From RealTime Measurements in an Urban Area, Paper A21D-01.
G. Adamkiewicz, G.R. McRae and J.H. Shorter, Determination of Urban-Scale
Emissions Using Inverse Air Quality Modeling, Paper A21D-02.
J.H. Shorter, G.J. McRae, J.B. McManus, C.E. Kolb, B.K. Lamb, S. O’Neill
and E. Allwine, Measurement and Interpretation of Nitrogen Oxide and Ozone
Concentration Dynamics in an Urban Environment, Paper A21D-03.
S.M. O’Neill, B.K. Lamb, D. Stock, E. Allwine, J.B. McManus, J.H. Shorter, G McRae
and J. Ferreira, An Urban Emissions Footprint Model, Paper A22A-013.
A.A. Ismail and J. Ferreira, Jr., Distributed GIS Tools for Integrating Urban Land Use
and Demographic Data Into Air Quality Models, Paper A22A-14
Other Presentations Without Published Proceedings Papers
S.M. O’Neill, 1999, An Urban Emissions Footprint Model. EPA Science To Achieve
Results (STAR) Graduate Fellowship Conference Abstracts, Arlington, VA, p. 192.
S.M.,O’Neill, B.K. Lamb, D. Stock, R. Villasenor, E. Allwine, J.H. Shorter,
J.B. McManus, 1998. Turbulence Model of an Urban Landscape for use in an Urban
Footprint Model. Graduate and Professional Student Association (GPSA) paper and
poster competition, Washington State University, Pullman WA.
222
C.E. Kolb, J.B. McManus, J.H. Shorter, D.D. Nelson, M.S. Zahniser, G. Adamkiewicz,
G.J. McRae, and J. Ferreira, 1997, Urban Metabolism and Trace Gas Respiration, NASA
12th Mission to Planet Earth/Earth Observing System/Investigators Working Group
Meeting, San Diego, CA (February, 1997).
C.E. Kolb and J.B. McManus, 1997, Tools to Characterize Urban Respiration, NASA
13th Mission to Planet Earth/Earth Observing System/Investigators Working Group
Meeting, Atlanta, GA (November, 1997).
C.E. Kolb, J.B. McManus, J.H. Shorter, D.D. Nelson, M.S. Zahniser, G. Adamkiewicz,
G.J. McRae, and J. Ferreira, 1998, Urban Metabolism and Trace Gas Respiration, NASA
14th Mission to Planet Earth/Earth Observing System/Investigators Working Group
Meeting, Durham, NH (October, 1998).
C.E. Kolb, J.B. McManus, J.H. Shorter, D.D. Nelson, M.S. Zahniser, G. Adamkiewicz,
G.J. McRae, and R.C. Harriss, 1998, Innovative Tools and Methods to Characterize
Urban Emissions, Workshop on Urban Air Quality, Houston, TX (April, 1998).
C.E. Kolb, 2002, Results of Mobile Laboratory Measurements of Urban Emissions
Sources and Airborne Pollution Distributions in Boston and New York City, Fifth
Workshop on Mexico City Air Quality, Ixtapan de la Sal, Mexico (January, 2002).
During the Spring of 1999 and 2000, project data and models were used for group projects on
GIS-based urban environmental modeling in MIT subject 11.521, Spatial Database Management
and Advanced Geographic Information Systems. The class projects tested the robustness and
practicality of GIS-based methods developed by the urban respiration project.
5.1.2 Presentations With Published Proceedings Papers
S. Napelenok, S. M. O'Neill, B. K. Lamb, E. J. Allwine, D. Stock, Modeling the Upwind
Pollutant Source Footprint Along Backward Trajectories Using the
MM5/CALMET/CALPUFF Modeling System, Preprint Volume of the 24th Conference
on Agricultural & Forest Meteorology, 14th Conference on Biometeorolgy and
Aerobiology, and 3rd Urban Environment Symposium American Meteorological Society,
Davis, CA (2000).
J.H. Shorter, J.B. McManus, C.E. Kolb, B.K. Lamb, S.M. O’Neill, E.J. Allwine,
R.W. Talbot, E. Scheuer, P.M. Crill, J. Ferreira, Jr., ad G.J. McRae, Understanding the
Influence of Local and Regional Sources on the Temporal and Spatial variability of
Pollutants in Urban Environments, Preprint Volume of the 24th Conference on
Agricultural & Forest Meteorology, 14th Conference on Biometeorolgy and Aerobiology,
and 3rd Urban Environment Symposium American Meteorological Society, Davis, CA
(2000).
223
S.M. O'Neill, B.K. Lamb, J. Chen, S. Napelenok, E. Allwine, D. Stock, J.B. McManus,
J.H. Shorter, C E. Kolb, , Correlating an Upwind Source-Footprint with Urban Emissions
Data Using the MM5/MCIP/CALPUFF Modeling System. Preprint Volume of the
International Emission Inventory Conference, US EPA (2001)
(http://www.epa.gov/ttn/chief/conference/ei10/).
J.H. Shorter, J.B. McManus, C.E. Kolb, B.K. Lamb, S.M. O'Neill, E.J. Allwine, R.W.
Talbot, E. Scheuer, J. Ferreira and G.J. McRae, "Understanding The Influence Of Local
And Regional Sources On The Temporal And Spatial Variability Of Pollutants In Urban
Environments", Workshop on Atmospheric Composition, Biogeochemical Cycles, and
Climate, Aspen Global Change Institute, http://www.agci.org/cfml/programs/eoc.cfm
(August, 2000).
J. H. Shorter, J. B. McManus, C. E. Kolb, E.J. Allwine, S.M. O’Neill, B.K. Lamb,
E.Scheuer, P.M. Crill, R.W. Talbot, J. Ferreira, Jr., and G.J. McRae, 1998, "Recent
measurements of urban metabolism and trace gas respiration," Preprint Volume of the
23th Conference on Agricultural & Forest Meteorology, 13th Conference on
Biometeorolgy and Aerobiology, and 2rd Urban Environment Symposium American
Meteorological Society, Albuquerque, NM, 49-52, (1998).
S. O’Neill, R. Villasenor, B. K. Lamb, D. Stock, E. Allwine, J. H. Shorter, J. B.
McManus, 1998, Turbulence model of an urban landscape for use in an urban footprint
model, Preprint Volume of the 23th Conference on Agricultural & Forest Meteorology,
13th Conference on Biometeorolgy and Aerobiology, and 2rd Urban Environment
Symposium American Meteorological Society, Albuquerque, NM, p. 180-183.
C-H Yeang and J. Ferreira, Jr., “Distributed GIS for Monitoring and Modeling Urban Air
Quality”, Proceedings of 6th International Conference in Urban Planning and Urban
Management, September, 1999, Venice. (Also translated into Italian and subsequently
published in the Journal URBANISTICA, n.114, October, 2000.)
L. Cao and J. Ferreira, Jr., “Integrating GIS and RDBMS to Model Traffic Congestion
and Urban Air Pollutants”, Proceedings of Urban and Regional Information Systems
Association, 39th Annual Conference, October 2002, Long Beach, CA.
(This paper won the prize for the best student-authored paper at
this conference).
224
5.1.3 Archival Journal Papers
Jiménez, J.L., J.B. McManus, J.H. Shorter, D.D. Nelson, M.S. Zahniser, M. Koplow,
G.J. McRae and C. E. Kolb, 2000, Cross Road and Mobile Tunable Infrared Laser
Measurements of Nitrous Oxide Emissions from Motor Vehicles, Chemosphere: Global
Change Sci. 2, 397-412.
5.1.4. Graduate Theses Fully or Partially Supported
Adamkiewicz, B., An Integrated Study of Photochemical Air Pollution- From Emissions
to Health Effects, Ph.D. Thesis, Department of Chemical Engineering, Massachusetts
Institute of Technology, 2000.
Alameh, N., Scalable and Extensible Infrastructures for Distributing Geographic
Information Services on the Internet, Ph.D. Thesis, Civil and Environmental Engineering,
Massachusetts Institute of Technology, 2001.
Ismail, A., A Distributed System Architecture for Spatial Data Management to Support
Engineering Modeling, MCP Thesis, Department of Urban Studies and Planning,
Massachusetts Institute of Technology, 1999.
O’Neill, S., Modeling Ozone and Aerosol Formation and Transport in the Pacific
Northwest and Calculating Fractional Source Contributions to Downwind Receptors,
Ph.D. Thesis, Department of Civil and Environmental Engineering, Washington State
University, 2002.
Pun, B.K.-L., Treatment of Uncertainties in Atmospheric Systems: A Combined
Modeling and Experimental Approach, Ph.D. Thesis, Department of Chemical
Engineering, Massachusetts Institute of Technology, 1998.
Wang, C., Parametric Uncertainty Analysis for Complex Engineering Systems, Ph.D.
Thesis, Department of Chemical Engineering, Massachusetts Institute of Technology,
1999.
225
5.2 Planned Archival Journal Papers
Talbot, R. W., et al., 2002, An assessment of fine particles in the Boston metropolitan
atmosphere, Atmos. Environ., in preparation.
Shorter, J.H., J.B. McManus, C.E. Kolb, Spatial and Temporal Variability of Pollutants
in an Urban Area, Environ. Sci. and Technol., in preparation.
McManus, J.B., et al., Pollutant Emission Indices from Mobile Measurements, Environ.
Sci. and Technol., in preparation.
J. Ferreira, Jr., Liou Cao, and Mizuki Kawabata, Spatial Data Infrastructures for Urban
Environmental Modeling, [Forthcoming paper for journal submission to: International
Journal of GIS or Environment and Planning (B)].
C.E. Kolb, S.C. Herndon, J.B. McManus, J.H. Shorter, M.S. Zahniser, D.D. Nelson, J.T.
Jayne, M.R. Canagaratna, and D.R. Worsnop, Mobile Laboratory for Realtime
Measurements of Urban and Regional Trace Gas and Particulate Distributions and
Emission Source Characterization, Environ. Sci. Technol., in preparation.
5.3 Web Sites With Archived Project Data and Modeling/Analyses
MIT Urban Studies GIS Web Site
http://metro.mit.edu/urbanair/overview - Website for GIS-based modeling and data analyses of
the Boston metro area. The site documents the GIS-based data processing 'pipeline' for linking
observed, surface-level trace gas concentrations to readily available data about local land use,
traffic conditions, and major urban activities. The web site also provides links to underlying
datasets and to detailed information about the modeling components and experimental results.
The site will be maintained for the foreseeable future as a repository for project-related
information.
226
WSU Field Campaign Archived Data
Field
Data
Campaign Type
11/97
MM5
06/98
Sodar
08/98
VOC
08/98
Sodar
08/98
SF6
Tracer
Data
08/98
MM5
05/99
Sodar
05/99
SF6
Tracer
Data
MM5
05/99
05/99
WSU
Van
Data Location
NCAR Mass Storage System, directory:
/SMO/MM5V2/manchesterNH/nov10-13.1997
lar.ce.wsu.edu, directory:
/users/lar/susan/UrbanResp/jun98/sodar
(one file for each day)
lar.ce.wsu.edu, directory:
/users/lar/susan/UrbanResp/aug98/VOC
lar.ce.wsu.edu, directory:
/users/lar/susan/UrbanResp/aug98/sodar
(one file for each day)
lar.ce.wsu.edu, directory:
/users/lar/susan/UrbanResp/aug98/POST_tracer
files for tests conducted: 8/27/98, 8/28/98,
8/30/98
NCAR Mass Storage System, directory:
/SMO/MM5V2/ManchesterNH/aug26-31.1998
lar.ce.wsu.edu, directory:
/users/lar/susan/UrbanResp/may99/sodar
(one file for each day)
lar.ce.wsu.edu, directory:
/users/lar/susan/UrbanResp/may99/Post/052599_
sf6.prn
NCAR Mass Storage System, directory:
/SMO/MM5V3/boston/may 24-26.1999
lar.ce.wsu.edu, directory:
/users/lar/susan/UrbanResp/may99/Post
(one file for each day)
Comments
Ascii data Hourly
averages WS, WD,
sigmaW, sigmaWD
MS excel file
Ascii data Hourly
averages WS, WD,
sigmaW, sigmaWD
Ascii data
Date, Time, Lat,
Lon, MSL, SF6 (ppt)
Ascii data Hourly
averages of WS,
WD, sigmaW,
sigmaWD
Ascii data
Time, SF6 (ppt), Lat,
Lon, MSL
Ascii data, CO2, CO,
NO, NO2, Lat, Lon,
MSL, WS, WD
ARI Field Campaign Archived Data
The data from the TDL, CO2 Licor, the GPS, and UV radiometer, were merged into a
single file for each experimental day, with data interpolated onto the 1 sec grid of the GPS data.
The data files, in ASCII format, are stored on the ARI FTP site, ftp.aerodyne.com. This is a
read-only ftp site, accessible via anonymous ftp. The data files are in the gps_data folder in
ftp.aerodyne.com, with a separate folder for each field campaign: man_june contains 6/98 data,
man_august contains 8/98 data, bos_may99 contains 5/99 data. A read-me file is located within
each folder and has information about the data file format, and specific information with respect
to the specific field campaign.
227
Each data file contains tab delimited data. In general, the data is listed in the following
order:
date&time
O3 CO2
latitude
uv
longitude
altitude
tdl species1
tdl species2 … tdl species n.
where ozone and uv level were only collected in the August 1998 and May 1999 campaigns, and
n is the total number of species measured with the TILDAS instrument. The time is the GPS
time is given in Greenwich Mean Time (GMT). The latitude and longitude are given in degrees
and the altitude is in meters (m). The mixing ratios of the tdl species and O3 are in parts per
billion by volume (ppbv); CO2 is in units of parts per million by volume (ppmv), and the uv
intensity is in mW/cm2.
228
6.0 SUMMARY
Our Urban Metabolism and Trace Gas Respiration project was a very ambitious and
highly multi-disciplinary project. Its goals included changing the basic way we measure and
assess urban trace gas and fine particle emissions, the way we measure and analyze ambient
urban air pollutant concentration distributions, and the way we associate pollutant emissions and
concentration distributions with underlying urban activities and processes. While we cannot
claim full success in all of these areas, we believe we have demonstrated substantial progress in
each. Despite initial difficulties in communicating across disciplines, we believe that the results
obtained are a clear indication that progress in urban air quality issues in particular, and the
atmospheric sciences in general, will require the type of cross disciplinary efforts measured in
this report.
In summarizing our project’s accomplishments we would like to highlight the eight
following achievements:
1)
We have demonstrated that a mobile laboratory equipped with sensitive, specific,
fast response instruments can be used to gather temporally and spatially resolved
distributions of key ground level pollutants in urban areas.
2)
We have shown that by using CO2 as an internal tracer of combustion source
emission plumes that the trace gas and fine particle data collected by the mobile
laboratory can be analyzed to yield accurate values of pollutant emission factors for both
mobile and fixed combustion sources.
3)
We have determined that areal distributions of background urban pollutant levels
can be determined by using correlations among measured pollutants to subtract out the
effects of strong local sources.
4)
We have demonstrated that an inert chemical tracer like SF6 can be used to trace
urban air motion and to test models used to determine how large point and extended
emission sources influence urban pollution distributions.
5)
We have developed a novel way, using inverse diffusion modeling, to determine
the fractional contribution of upwind sources to concentrations measured at downwind
receptors.
6)
We have shown how coupled models of regional meteorology and atmospheric
chemistry can be used to predict the regional impact of urban pollutants, and to possibly
identify targets for evolving remote sensing satellite instruments designed to monitor
tropospheric pollutants.
7)
We have demonstrated that coupled urban meterology/photochemistry models can
be inverted to deduce if measured distributions of pollutants are consistent with assumed
pollutant and pollutant precursor distributed emissions inventories. Coupled with the
ability to determine better urban pollutant distributions using the mobile laboratory
229
techniques noted above, this advance promises to allow a much more incisive test of
current urban emission inventory models and assumptions.
8)
We have mapped measured urban pollution distributions onto sophisticated GIS
representations of urban activity factors, demonstrating new methods of presenting and
evaluating the connections between urban pollution emissions and distributions, and the
distributions of urban population, economic activities, and transportation infrastructure.
The symposia presentations and proceedings papers, published and planned archival
papers, graduate student theses, and web archives listed in Section 5 of this report have presented
or will present the specific results of our measurements and analyses of pollutant emissions and
distributions in the Manchester, NH and Boston, MA metropolitan areas. More importantly, we
feel that this project has clearly demonstrated innovative methods of gathering and understanding
urban air pollutant emission and distribution data.
We believe that the measurement and analyses techniques we have pioneered during this
project will profoundly influence the way complex urban air pollution issues are engaged in the
future. This is a bold assertion. Our best evidence of its correctness is that the results of our
Manchester and Boston measurements have been impressive enough that major elements of our
team have been funded to use these methods for in depth studies of air quality in New York City
and Mexico City. Our recent measurement campaigns in these megacities have been very
successful and our initial analyses indicate that we have identified important underreported
pollutant emissions sources and unexpected pollutant distribution features in both cities. The
freedom to develop new urban air pollutant measurement and analysis techniques afforded by the
NASA EOS/IDS program made these subsequent investigations possible. We think the legacy of
this NASA/IDS program is a greatly enhanced set of tools to investigate the impact of trace gas
and fine particle emissions from the world’s cities on the atmosphere, from the local to the global
scale.
230
7.0 REFERENCES
Anderson, J.G., W.H. Brune, and M.H. Proffitt, 1989, Ozone Destruction by Chlorine Radicals
Within the Antarctic Vortex: The Spatial and Temporal Evolution of ClO-O3 Anticorrelation
Based on in Situ ER-2 Data, J. Geophys. Res. 94, 11.465-11.479.
Aubry, N., R. Guyonnet, and R. Lima, 1991, Spatio temporal analysis of complex signals:
Theory and applications. J. Stat. Phys., 64, 683-739, 1991.
Aubry, N., W.-Y. Lian, and E.S. Titi., 1993, Preserving symmetries in the proper orthogonal
decomposition. SIAM J. Sci. Comput. 14, 483-505.
Ayres, R.U. and U.E. Simonis, eds. 1994, Industrial metabolism: Restructuring for Sustainable
Development, New York, United Nations University Press.
Batterman, S.A., 1992, Optimal estimators for ambient air quality levels. Atmos. Environ. 29,
497-516.
Becker, K.H., J.C. Löizer, R. Kurtenbach, P. Wiesen, T. Jensen, T.J. Wallington, 2000,
Contribution of vehicle exhaust to the global N2O budget, Chemosphere 2, 387.
Becker, K.H., J.C. Löizer, R. Kurtenbach, P. Wiesen, T. Jensen, T.J. Wallington, 1999,
Nitrous Oxide (N2O) Emissions from vehicles, Environ. Sci. Technol. 33, 4134 - 4139.
Benner, R., B. Lamb, 1985, A Fast Response Continuous Analyzer for Halogenated Atmospheric
Tracers, J. Atmos. Ocean. Technol. 2, 582.
Bennett, R.J., R.P. Haining, and D.A. Griffith, 1984, The problem of missing data on spatial
surfaces. Ann. Ass. Am. Geograph., 74:138-156.
Berges, M.G.M., R.M. Hofmann, D. Scharffe, P.J. Crutzen, 1993, Nitrous Oxide Emissions
from Motor Vehicles in Tunnels and Their Global Extrapolation, J. Geophys. Res. 98, 18,527 18,531.
Bishop, G.A., J. R. Starkey, A. Ihlenfeldt, W. J. Williams, D. H. Stedman, 1989, IR Long Path
Photometry: A Remote Sensing Tool for Automobile Emissions, Anal. Chem. 61, 671 - 677A.
Braddock, J.D., 1981, Impact of Low Ambient Temperature on 3-Way Catalyst Car Emissions,
Society of Automotive Engineers Paper 810280.
Brown M, 1993, Deduction of emissions of source gases using an objective inversion algorithm
and a chemical transport model. J. of Geoph. Res. 98, 12,639-12,660.
Byun, D., J. Pleim, R. Tang, A. Bourgeois, 1999, Science Algorithms of the EPA Models-3
Community Multiscale Air Quality (CMAQ) Modeling System, Chapter 12. EPA/600/R-99/030.
Chan, T.L., G. Dong, C. S. Cheung, C. W. Leung, C. P. Wong, W. T. Hung, 2001, Monte
Carlo Simulation of Nitrogen Oxides Dispersion from a Vehicular Exhaust Plume and Its
Sensitivity Studies, Atmos. Environ. 35, 6117-6127.
Cicerone, R.J., 1989, Analysis of Sources and Sinks of Atmospheric Nitrous Oxide (N2O)”,
J. Geophys. Res. 94, 18,265 - 18,271.
231
Clarke, A. D., 1991, A thermo-optical technique for in situ analysis of size-resolved aerosol
physicochemistry, Atmos. Environ., 25, 635-644.
Clifford, M.J., R. Clarke, S. B. Riffat, 1997, Local Aspects of Vehicular Pollution, Atmos.
Environ. 31, 271-276.
Daley R., 1991, Atmospheric Data Analysis. Cambridge University Press, Cambridge.
Dasch, J.M., 1992, Nitrous Oxide Emissions from Vehicles, J. Air Waste Manage. Assoc. 42, 63
- 67.
Dawson, P.J., D.E. Stock, and B.K. Lamb, 1991, The Numerical Simulation of Airflow and
Dispersion in Three-Dimensional Atmospheric Recirculation Zones. J. of Appl. Meteorol., 30,
1005 - 1024.
Decker, E.H., S. Elliott, F.A. Smith, D.R. Blake and F. S. Rowland, 2000, Ann. Rev. Energy
Environ. 25, 685-740.
DeHaan, P., M. Keller, 2000, Emission Factors for Passenger Cars: Application of Instantaneous
Emission Modeling, Atmos. Environ. 34, 4629-4638.
Deneter, F. J., and P. J. Crutzen, 1994, A Three-Dimensional Model of the Global Ammonia
Cycle, J. Atmos. Chem., 19, 331-369.
Dudhia, J., G.A. Grell, D.R. Stauffer, 1994, A description of the Fifth-Generation Penn
State/NCAR Mesoscale Model (MM5). NCAR Technical Note NCAR/TN-398+STR.
Durbin, T.D., et al., 2002, Estimates of the Emission Rates of Ammonia From Light-Duty
Vehicles Using the Standard Chasis Dynamometer Test Cycles”, Atmos. Environ. 36, 14751482.
Elbert, T. F., 1984, Estimation and Control of Systems. Van Nostrand Reinhold, New York.
EPA, 1995, Air Quality Trends Report.
Fischer-Kowalski, M. and W. Hütter, 1998, Society’s Metabolism – The Intellectual History of
Materials Flow Analysis, Part II, 1970-1998, J. Industr. Ecology 2, 107-136.
Fischer-Kowalski, M., 1998, Society’s Metabolism – The Intellectual History of Materials Flow
Analysis, Part I, 1860-1970, J. Industr. Ecology 2, 61-78.
Fraser, M. P., and G. R. Cass, 1998, Detection of Excess Ammonia Emissions from in-Use
Vehicles and the Implications for Fine Particle Control, Environ. Sci. Technol., 32, 1053-1057
Fujita, E.M., B.E. Croes, C.L. Bennett, D.R. Lawson, F.W. Lurmann, and H.H. Main., 1992,
J. Air Waste Management Assoc. 42, 264-276.
Garfinkel, R., S. Kunnathur, and A.S. Liepins., 1986, Optimal imputation of erroneous data.
J. Oper. Res. 34, 744-751.
Guenther, A., B.K. Lamb, and D.E. Stock, 1990, Three-Dimensional Numerical Simulation of
Plume Downwash with a k-e Turbulence Model. J. of Appl. Meteorol., 29, 633 - 643.
Hadamard, J., 1952, Lectures on Cauchy’s problem in Linear Partial Differential Equations.
Dover, New York.
232
Harley, R.A., A. G. Russell, G. J. McRae, G. R. Cass, and J. H. Seinfeld, 1993, Photochemical
modeling of the southern California air quality study. Environmental Sci. Technol. 27, 378-388.
Harley, R.A., et al., 2001, Analysis of motor vehicle emissions during the Nashville/middle
Tennessee ozone study, J. Geophys. Res. 106 (D4) 3559-3567.
Hartley D. and R. G. Prinn, 1993, Feasibility of determining surface emissions of trace gases
using an inverse method in a three-dimensional chemical transport model. J. Geophy. Res. 98,
5183-5197.
Heeb, N.V., et al., 2001, Comparison of methane, benzene, alkyl benzenes: cold start emission
data of Euro I and Euro II three-way catalyst passenger cars and light duty vehicles from time
resolved gas analysis, Proceedings of the 10th International Symposium Transport and Air
Pollution, p. 93, Boulder, Colorado, p. 93-100.
Horii, C.V., M.S. Zahniser, D.D. Nelson, J. B. McManus and S.C. Wofsy, 1999, Nitric Acid and
Nitrogen Dioxide Flux Measurements: a New Application of Tunable Diode Laser Absorption
Spectroscopy, Application of Tunable Diode and Other Infrared Sources for Atmospheric
Studies and Industrial Process Monitoring, SPIE Proceedings, V. 3758, 152-161.
IPCC, Climate Change 1995: The Science of Climate Change, Contribution of Working Group I
to the Second Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge
University Press, Cambridge, U.K (1996).
IPCC, 2001: Climate Change 2001: The Scientific Basis. Contribution of Working Group I to the
Third Assessment Report of the Intergovernmental Panel on Climate Change, [J. T. Houghton,
Y. Ding, D. J. Griggs, M. Noguer, P. J. van der Linden, X. Dai, K. Maskell and C. A. Johnson
(eds.)]. Cambridge University Press, Cambridge,.U.K.
Jiménez, J.L., D.D. Nelson, M.S. Zahniser, J.B. McManus, C. E. Kolb, M. D. Koplow, S.E.
Schmidt, 1997, Remote Sensing Measurements of On-Road Vehicle Nitric Oxide Emissions and
of an Important Greenhouse Gas: Nitrous Oxide, in Seventh CRC On-Road Vehicle Emissions
Workshop, San Diego, CA.
Jiménez, J.L., J.B. McManus, J.H. Shorter, D.D. Nelson, M.S. Zahniser, M.Koplow, G.J.
McRae, C.E. Kolb, 2000, Cross road and mobile tunable infrared laser measurements of nitrous
oxide emissions from motor vehicles, Chemosphere: Global Change Science 2, 397 - 412.
Jiménez, J.L., M.D. Koplow, D.D. Nelson, M.S. Zahniser, S.E. Smith, 1999, Characterization
of on-road vehicle NO emissions by a TILDAS remote sensor, J. Air Waste Manage. Assoc. 49,
463 - 470.
Kebabian, P.L. and Kolb, C.E., 1993, The Neutral Gas Laser: A Tool for Remote Sensing of
Chemical Species by Infrared Absorption, In Sampling and Analusis of Airborne Pollutants,
edited by E.D. Winegar and L.H. Keith, Boca Raton, FL: Lewis Publishers.
Khalil, M.A.K., R.A. Rasmussen, 1992, The global sources of nitrous oxide, J. Geophys. Res.
97, 14,651-14,660.
Kramer M.A., 1991, Nonlinear principal component analysis using autoassociative neural
networks. AIChE J. 37, 233-243.
233
Krischer, K., R.R. Martinez, I.G. Kevrekidis, H.H. Rotermund, G. Ertl, and J.L. Hudson, 1993,
Model identification of a spatiotemporally varying catalytic reaction. AIChE J. 39, 89-98.
Kroeze, C., A. Mozier, L. Bouwman, 1999, Closing the global N2O budget: A retrospective
analysis 1500-1994, Global Biogeochem. Cycles 13, 1-8.
Lamb, B., J.B. McManus, J.H. Shorter, C.E. Kolb, B. Mosher, R.C. Harriss, E. Allwine, D.
Blaha, T. Howard, A. Guenther, R.A. Lott, R. Siverson, H. Westberg, and P. Zimmerman, 1995,
Measurement of Methane Emissions from Natural Gas Systems Using Atmospheric Tracer
Methods, Environ. Sci. Techn, 29, 1468-1479.
Lamb, B., J.B. McManus, J.H. Shorter, C.E. Kolb, B. Mosher, R.C.Harriss, E. Allwine,
D. Blaha, T. Howard, A. Guenther, R.A. Lott, R. Siverson, H. Westberg and P. Zimmerman,
1995, Development of Atmospheric Tracer Methods to Measure Methane Emissions from
Natural Gas Facilities and Urban Areas, Environ. Sci. Technol. 29, 1468-1479
Lefer, B. L., and R. W. Talbot, 2001, Summertime measurements of aerosol nitrate and
ammonium at a northeastern U.S. site,, J. Geophys. Res. 106, 20,365-20,373.
Little, R.J., D.B. Rubin, and W.E. Strawderman, 1989, Statistical Analysis with Missing Data.
John Wiley, New York.
Lorenc A.C., 1986, Quart. J. R. Met. Soc. 112, 1177-1194.
Lurmann, F.W., W.P.L. Carter, and L.A. Coyner, 1987, A surrogate species chemical reaction
mechanism for urban scale air quality simulation models. Volumes I and II. ERT Inc., Newbury
Park, CA and Statewide Air Pollution Research Center, University of California, Riverside, CA,.
McLaughlin, D. and L. Townley, 1995, A Reassessment of the Groundwater Inverse Problem,
Water Resources Res. submitted.
McManus, J.B., Kebabian, P.L., and Kolb, C.E., 1989, Atmospheric Methane Measurement
Instrument Using a Zeeman split He-Ne Laser, Appl. Optics. 28, no. 23, pp. 5016-5023.
McManus, J.B., P. L. Kebabian, and M. S. Zahniser, 1995, Astigmatic Mirror Multiple Pass
Absorption Cells for Long Pathlength Spectroscopy, Appl. Opt., 34, 3336.
McRae, G.J., Goodin, W.R., and Seinfeld, J.H., 1982, Development of a Second-Generation
Mathematical Model for Urban Air Pollution 1. Model Formulation. Atmospheric Environment,
16, 679-696.
Menke, W., 1989, Geophysical Data Analysis: Discrete Inverse Theory, Academic Press, Inc.,
San Diego.
Michaels, H., Emissions of Nitrous Oxide from Highway Mobile Sources: Comments on the
Draft Inventory of U.S. Greenhouse Emissions and Sinks 1990-96 (March 1996), U.S.
Environmental Protection Agency: Ann Arbor, Michigan (1998).
Milford, J. B., A. G. Russell, and G. J. McRae, 1989, A New Approach to Photochemical
Pollution Control: Implications of Spatial Patterns in Pollutant Responses to Reductions in
Nitrogen Oxides and Reactive Organic Gas Emissions. Environ. Sci. Technol., 23, 1290-1301.
234
Mosher, B.W., P.N. Czepiel, R.C. Harriss, J.H. Shorter, C.E. Kolb, J.B. McManus, E. Allwine
and B.K. Lamb, 1999, Methane Emissions at Nine Landfill Sites in the Northeastern United
States, Environ. Sci. Technol. 33, 2088-2094.
Mulholland, M. and J. H. Seinfeld, 1995, Inverse Air Pollution Modelling of Urban-Scale
Carbon Monoxide Emissions. Atmos. Environ. 26A, 113-123.
Nahi, N., 1969, Estimation Theory and Applications. John Wiley and Sons, New York.
Napelenok, S., S. M. O'Neill, B. K. Lamb, E. J. Allwine, D. Stock. 2000, Modeling the Upwind
Pollutant Source Footprint Along Backward Trajectories Using the MM5/CALMET/CALPUFF
Modeling System. Preprint Volume of the 24th Conference on Agricultural & Forest
Meteorology, 14th Conference on Biometeorolgy and Aerobiology, and 3rd Urban Environment
Symposium American Meteorological Society, Davis, CA.
Nelson, D.D. and coauthors, 1996, Recent Improvements in Atmospheric Trace Gas Monitoring
Using Mid-Infrared Tunable Diode Lasers, Proc. SPIE 2834, 148-159.
Nelson, D.D., J. H. Shorter, J. B. McManus, M. S. Zahniser, 2002, Sub part-per-billion
Detection of Nitric Oxide in Air Using a Thermoelectrically Cooled Mid-Infrared Quantum
Cascade Laser Spectrometer, Appl. Phys. B., submitted.
Nelson, D.D., J.B. McManus, M. Zahniser, C.E. Kolb, 1999, Laser system for cross road
measurement of motor vehicle exhaust gases, U.S. Patent No. 5,877,862.
Nelson, D.D., M.S. Zahniser, J.B. McManus, C.E. Kolb, J.L. Jiménez, 1998, A tunable diode
laser system for the remote sensing of on-road vehicle emissions, Appl. Phys. B 67, 433-441.
Newsam, G.N. and I.G. Enting, 1988, Inverse problems in atmospheric constituent studies:
I. determination of surface sources under a diffusive transport approximation. Inverse Problems,
4, 1037-1054.
NRC, 1993, Understanding and Predicting Atmospheric Chemical Change, National Research
Council, National Academy Press, Washington, DC.
NRC, 1991, Rethinking the Ozone Problem in Urban and Regional Air Pollution. National
Research Council, National Academy Press, Washington D.C.
Nunnermacker, L.J., D. Imre, P.H. Daum, L. Kleinman, Y.-N. Lee, J.H. Lee, S.R. Springston,
L. Newman, J. Weinstein-Lloyd, W.T. Luke, R. Banta, R. Alverez, C. Senff, S. Sillman, M.
Holdren, G.W. Keigley, and X. Zhou. 1998, Characterization of the Nashville urban plume on
July 3 and July 18, 1995, J. Geophys. Res. 103, 28,129-28, 148.
O’Neill, S. M., 1999, An Urban Emissions Footprint Model. EPA Science To Achieve Results
(STAR) Graduate Fellowship Conference Abstracts, Arlington, VA, p. 192.
O’Neill, S. M., B. K. Lamb, D. Stock, E. Allwine, J. H. Shorter, J. B. McManus, G. McRae,
J. Ferreira, 1999, An Urban Emissions Footprint Model, American Geophysical Union 1999
Spring Meeting, Published as a Supplement to EOS, p. S35.
O’Neill, S. M., B. K. Lamb, D. Stock, R. Villasenor, E. Allwine, J. H. Shorter, J. B. McManus
(1998). Turbulence Model of an Urban Landscape for use in an Urban Footprint Model.
235
Graduate and Professional Student Association (GPSA) paper and poster competition,
Washington State University, Pullman WA.
O’Neill, S. M., R. Villasenor, B. K. Lamb, D. Stock, E. Allwine, J. H. Shorter, J. B. McManus
(1998). Turbulence Model of an Urban Landscape for use in an Urban Footprint Model.
Preprint Volume of the 23rd Conference on Agricultural & Forest Meteorology, 13th Conference
on Biometeorolgy and Aerobiology, and 2nd Urban Environment Symposium, American
Meteorological Society, Albuquerque, New Mexico, 180-183.
O’Neill, S.M., et al., 2001, Correlating an Upwind Source-Footprint with Urban Emissions Data
Using the MM5/MCIP/CALPUFF Modeling System.
Odaka, M., N. Koike, H. Suzuki, 1998, Deterioration effect of three-way catalyst on nitrous
oxide emission, Society of Automotive Engineers Paper 980676.
O'Neill, S. M., B. K. Lamb, J. Chen, S. Napelenok, E. J. Allwine, D. Stock, J. B. McManus,
J. H. Shorter, C. E. Kolb (2001). Correlating an Upwind Source-Footprint with Urban Emissions
Data Using the MM5/MCIP/CALPUFF Modeling System. Preprint Volume of the International
Emission Inventory Conference, US EPA (http://www.epa.gov/ttn/chief/conference/ei10/).
Peterson, H. G., D. Mazzolini, S. M. O’Neill, B. K. Lamb, 1999, Instantaneous Spread of
Plumes in the Surface Layer, J. of Appl. Meteorol. 38, 343-352.
Peterson, H., and B. Lamb, 1995, An Investigation of Instantaneous Diffusion and Concentration
Fluctuations. Journal of Applied Meteorology, 34, 2724-2746.
Peterson, H.G., T. Donovan, S. M. O’Neill, B. K. Lamb,2001, Concentration Fluctuations and
Short-Term Peaks for Plumes in the Surface Layer, Journal of Applied Meteorology (in press).
Pierson, W.R., A.W. Gertler, and R.L. Bradow, 1990, J. Air Waste Management Assoc. 40,
1495-1504.
Pierson, W.R., et al., 1996, Real world automotive emissions- summary of studies in the Fort
McHenry and Tuscarora Mountain tunnels, Atmos. Environ. 30, 2233-2256.
Press, W.H., B.P. Flannery, S.A. Teukolsky, and W.T. Vetterling, 1992, Numerical Recipes in C:
The Art of Scientific Computing, Cambridge University Press.
Pun, B.K., 1998, Treatment of Uncertainties in Atmospheric Systems: A Combined Modeling
and Experimental Approach, Ph.D. thesis, Massachusetts Institute of Technology.
Rabl, H.-P., A. Gifhorn, R. Meyer-Pitroff, 1997, The formation of nitrous oxide (N2O) over
three-way catalysts, in 4th International Conference on Technologies and Combustion for a
Clean Environment, Lisbon, Portugal.
Roberts, J. M., 1995, Reactive odd-nitrogen (NOy) in the atmosphere, in Composition,
Chemsitry, and Cliamte of the Atmosphere, edited by H. B. Singh, pp. 176-215, Van Nostrand
Reinhold, New York.
Rothman, L.S. et al., 1998, The HITRAN Molecular Spectroscopic Database and HAWKS
(HITRAN Atmospheric Workstation): 1996 Edition, J. Quant. Spectrosc. Radiat. Transfer, 60,
665-710.
236
Rowlands, G. and J.C. Sprott, 1992, Extraction of dynamical equations from chaotic data.
Physica D, 58, 251-259.
Scire, J.S., D.G. Strimaitis, and R.J. Yamartino, 1999: A user’s guide for the CALPUFF
dispersion model (version 5.0). Earth Tech Inc., 462 pp.
Sen, M. and P. Stoffa, 1995, Global Optimization Methods in Geophysical Inversion, Elsevier
Science, Amsterdam.
Shorter, J. H., J. B. McManus, C. E. Kolb, E.J. Allwine, S.M. O’Neill, B.K. Lamb, E.Scheuer,
P.M. Crill, R.W. Talbot, J. Ferreira, Jr., and G.J. McRae, 1998, Recent Measurements of Urban
Metabolism and Trace Gas Respiration, American Meteorological Society Proceedings, Second
Urban Environment Symposium, Albuquerque, NM, pg. 49-52.
Shorter, J.H., J.B. McManus, C.E. Kolb, E.J. Allwine, B.K. Lamb, B.W. Mosher, R.C. Harriss,
V. Partchatka, H. Fisher, G.W. Harris, P.J. Crutzen, and H.J. Karbach, 1996, Methane Emission
Measurements in Urban Areas in Eastern Germany, J. Atmos. Chem. 24, 121-140.
Shorter, J.H., J.B. McManus, C.E. Kolb, B.K. Lamb, S.M. O’Neill, E.J. Allwine, R.W. Talbot,
E. Scheuer, P.M. Crill, J. Ferreira, Jr., ad G.J. McRae, 2000, Understanding the Influence of
Local and Regional Sources on the Temporal and Spatial Variability of Pollutants in Urban
Environments, American Meteorological Society Proceedings, Third Symposium on the Urban
Environment, 90-91.
Shorter, J.H., J.B. McManus, C.E. Kolb, E.J. Allwine, R. Siverson, B.K. Lamb, B.W. Mosher,
R.C. Harriss, T.Howard and R.A. Lott, 1997, Collection of Leakage Statistics in the Natural Gas
System By Tracer Methods, Evironon. Sci. and Technol. 31, 2012-2019.
Shorter, J.H., S. C. Herndon, M. S. Zahniser, D. D. Nelson, J. J. Jayne, C. E. Kolb, 2001,
Characterization of heavy-duty vehicle exhaust in dense urban environments, Proceedings of the
10th International Symposium Transport and Air Pollution, Boulder, Colorado, p. 33-40.
Sirovich, L. and R. Everson, 1992, Management and analysis of large scientific datasets.
Int. J. Supercomp. Appl. 6, 50-68.
Sjödin, Å. and K., 2000, Andreasson, Multi-year remote sensing measurements of gasoline light
duty vehicle emissions on a freeway ramp, Atmos. Environ. 34, 4657-4665.
Sjödin, Å., D.A. Cooper, K. Andréasson, 1995, Estimates of real-world N2O emissions from
road vehicles by means of measurements in a traffic tunnel, J. Air Waste Manage. Assoc. 45,
186-190.
Sjödin, Å., K. Persson, K.I. Andréasson, B. Arlander, B. Galle, 1997, On-road emission
factors derived from measurements in a traffic tunnel, 4th International Symposium on Transport
and Air Pollution, Avignon, France.
Stephens, R.D., et al., 1997, An analysis of Michigan and California CO remote sensing
measurements, J. Air Waste Manage. Assoc. 47, 601-607.
Stull, R.B, 1995, Meteorology Today for Scientists and Engineers. West Publishing Co. p.166.
237
Sun, N-Z, 1994, Inverse Problems in Groundwater Modeling, Kluwer Academic Publishers,
Dordrecht.
Talbot, R. W., A. S. Vijgen, and R. C. Harriss, 1992, Soluble species in the Arctic summer
troposphere: Acidic gases, aerosols, and precipitation, J. Geophys. Res., 97, 16,531-16,543.
Tarantola, A., 1987, Inverse Problem Theory, Elsevier Science Publishers, New York.
Tatang, M.A., 1987, Direct Incorporation of Uncertainty in Chemical and Environmental
Engineering Systems. PhD thesis, Massachusetts Institute of Technology.
Tatang, M.A., T.J. Resch, P. Wyckoff and G.J. McRae, 2002, A note on field approximation by
empirical karhunen-loeve series expansions. In preparation.
Tikhonovand, A.N., V.Y. Arsenin, 1977, Solution of Ill-Posed Problems. Halsted Press/Wiley,
New York.
Trainer, M., B.A. Ridley, M.P. Buhr, G. Kok, J. Walega, G. Hübler, D.D. Parrish, and
F.C. Fehsenfeld, 1995, Regional ozone and urban plumes in the southeastern United States:
Birmingham, a case study, J. Geophys. Res. 100, 18,823-18,834.
Trent, D.S., L.L. Eyler and M.J. Budden, 1983: TEMPEST a Three-Dimensional TimeDependent Computer Program for Hydrothermal Analysis - Volume I: Numerical Methods and
Input Instructions. Pacific Northwest Laboratory, PNL - 4348 Vol. 1, 249.
Urgell, L.T. and R.L. Kirlin, 1990, Adaptive image compression using Karhunen-Loeve
transform, Signal Process. 21, 303-313.
Wang, C., 1999, Parametric Uncertainty Analysis for Complex Engineering Systems,
Ph.D. Thesis, Massachusetts Institute of Technology.
Wolman, A., 1965, The Metabolism of Cities, Scientific American 213, 178-193.
Wormhoudt, J. et al., 1995, Infrared Tunable Diode Laser Measurements of Nitrogen Oxide
Species in an Aircraft Engine Exhaust, Proceedings SPIE 2546, Optical Techniques in Fluid,
Thermal and Combustion Flows, 552 – 561.
Yeang, C-H and J. Ferreira, Jr., 1999, Distributed GIS for Monitoring and Modeling Urban Air
Quality, Proceedings of 6th International Conference in urban Planning and Urban Management,
Venice, Italy. (Also translated into Italian and subsequently published in the Journal
URBANISTICA, n.114, October, 2000.)
Zahniser, M.S. and coauthors, 1995, Measurement of Trace Gas Fluxes Using Tunable Diode
Laser Spectroscopy, Phil Trans. R. Soc. Lond. A. 351, 371-382.
Zahniser, M.S., et al., 1995, Measurement of Trace Gas Fluxes Using Tunable Diode Laser
Spectroscopy, Phil. Trans. R. Soc. Lond. A. vol. 351, no. 1696, pp. 371-382.
Zhang, Y., D. H. Stedman, G.A. Bishop, S. P. Beaton, P. L. Guenther, I. F. McVey, 1996,
Enhancement of Remote Sensing for Mobile Source Nitric Oxide, J. Air & Waste Manage.
Assoc. 46, 25-29.
Zhang, Y., G.A. Bishop, D. H. Stedman, 1994, Automobile emissions are statistically gamma
distributed, Environ. Sci. Technol. 28, 1370-1374.
238
239