1. Theoretical Basis for Converting Satellite Observations to Ground-Level PM 2.5 Concentrations
Training Workshop in Partnership with BAAQMD
Santa Clara, CA September 10 – 12, 2013
Applied Remote SEnsing Training (ARSET) – Air Quality
A project of NASA Applied Sciences

2. Objective
OBJECTIVE
Estimate ground-level PM2.5 mass concentration (µg m-3) using satellite derived AOD

3. What are we looking for?
Sensitive groups should avoid all physical activity outdoors; everyone else should avoid prolonged or heavy exertion
Sensitive groups should avoid prolonged or heavy exertion; everyone else should reduce prolonged or heavy exertion
Sensitive groups should reduce prolonged or heavy exertion
Unusually sensitive people should consider reducing prolonged or heavy exertion
None
Cautionary Statements
51-100
0-50
Index
Values
PM10
(ug/m3)
PM2.5
Category
55-154
0-54
0-15.4
Good
Very Unhealthy
Unhealthy
Unhealthy for Sensitive Groups
Moderate

4. What Do Satellites Provide?
Sun
Column
Satellite
Measurement
Ozone
Rayleigh Scattering
10km
Water vapor + other gases (absorption)
Seven MODIS bands are utilized to derive aerosol properties: 0.47, 0.55, 0.65, 0.86, 1.24, 1.64, and 2.13 µm
10X10 km2 Res.
Aerosol
Surface
Satellite retrieval issues - inversion
(e.g. aerosol model, background).

5. Measurement Techniques
AOD – Column integrated value (top of the atmosphere to surface) - Optical measurement of aerosol loading – unitless. AOD is a function of shape, size, type and number concentration of aerosols
PM2.5 – Mass per unit volume of aerosol particles less than 2.5 µm in aerodynamic diameter at surface (measurement height) level

6. MODIS-Terra True Color Images
May 11, 2007
May 12, 2007
May 13, 2007
May 14, 2007
May 15, 2007
May 16, 2007

7. MODIS-Terra AOD May 11, 2007 May 12, 2007 May 13, 2007 May 14, 2007

8. AOD – PM Relation Composition Size distribution Vertical profile
 – particle density
Q – extinction coefficient
re – effective radius
fPBL – % AOD in PBL
HPBL – mixing height
Composition
Size distribution
Vertical profile

9. Support for AOD-PM2.5 Linkage
Current satellite AOD is sensitive to PM2.5 (Kahn et al. 1998)
Polar-orbiting satellites can represent at least daytime average aerosol loadings (Kaufman et al., 2000)
Missing data due to cloud cover appear random in general (Christopher et al., 2010)

10. Assumption for Quantitative Analysis
When most particles are concentrated and well mixed in the boundary layer, satellite AOD contains a strong signal of ground-level particle concentrations.

11. Modeling the Association of AOD With PM2.5
The relationship between AOD and PM2.5 depends on parameters hard to measure:
Vertical profile
Size distribution and composition
Diurnal variability
We develop statistical models with variables to represent these parameters
Model simulated vertical profile
Meteorological & other surrogates
Average of multiple AOD measurements
No textbook solution!

12. Methods Developed So Far
Statistical models
Correlation & simple linear regression
Multiple linear regression with effect modifiers
Linear mixed effects models
Geographically weighted regression
Generalized additive models
Hierarchical models combining the above
Bayesian models
Artificial neural network
Data fusion models
Combining satellite data with model simulations
Deterministic models
Improving model simulation with satellite data

13. Simple Models from Early Days
Chu et al., 2003
Wang et al., 2003 13

14. AOT-PM2.5 Relationship
Gupta, 2008

15. Limitation: Major Unsolved Issue
All three Statistical models (TVM, MVM, ANN) underestimate high PM2.5 loadings

16. Vertical Distribution
Al-Saadi et al., 2008
Engel-Cox et al., 2006

17. What Satellites can provide for vertical information? - CALIPSO
Aug 17
Aug 18
Aug 19
Aug 20
Aug 21
Aug 22
Aug 23
Aug 25
Aug 28
5 km
(courtesy of Dave Winker, P.I. CALIPSO)

18. Advantages of using reanalysis meteorology along with satellite
TVM
MVM

19. Spatial Comparison MODIS-Terra, July 1, 2007
TVM
MVM
ANN
Spatial Comparison
Satellite-derived PM2.5 fills the gap in surface measurements
All three methods underestimate the higher PM2.5 concentrations.

20. Questions to Ask: Issues
How accurate are these estimates ?
Is the PM2.5-AOD relationship always linear?
How does AOD retrieval uncertainty affect estimation of air quality
Does this relationship change in space and time?
Does this relationship change with aerosol type?
How does meteorology drive this relationship?
How does vertical distribution of aerosols in the atmosphere affect these estimates?

21. Exposure Modeling Approach
Basic idea:
PM2.5 = f(AOD, meteorological variables, land use variables, etc.) + ε
Census / traffic data
land use
MODIS AOD
NLDAS meteorology
Statistical prediction model
EPA air monitors
Predicted PM2.5 surface

22. Requirements for this job
PCs with large hard drives and good graphics card
Internet to access satellite & other data
Some statistical software (SAS, R, Matlab, etc.)
Some programming skill
Knowledge of regional air pollution patterns
Ideally, GIS software and working knowledge

23. Scaling approach
Basic idea: let an atmospheric chemistry model decide the conversion from AOD to PM2.5. Satellite AOD is used to calibrate the absolute value of the model-generated conversion ratio.
Satellite-derived PM2.5 = x satellite AOD

24. Scaling approach can be applied wherever there are satellite retrievals, but prediction accuracy can vary a lot.

25. Comparison of two approaches
Statistical model
Scaling model
Flexible modeling structure
Ability to provide estimates at high spatial resolution with the help of other covariates
Higher accuracy
Requires extensive ground data support
Models are often not generalizable to other regions
Mathematically more elegant
Applicable in areas with extensive ground data support
Requires ability to conduct CTM runs
Resolution of predictions dependant on satellite data
Lower accuracy (no error terms allowed in the model)

26. Study Domain Design Number of monitoring sites: 119
Exposure modeling domain: 700 x 700 km2

27. Geographically Weighted Regression Model
GWR allows model parameters to vary in space to better capture spatially varying AOD-PM relationship – major advantage over global regression models.
Model Structure
Datasets (2003):
PM2.5 – EPA / IMPROVE daily measurements
AOD – MODIS collection 5 (10 km)
Meteorology – NLDAS-2 (14 km)
Land use: NLCD 2001

28. Summary Statistics
PM2.5 (ug/m3)
PBL (m)
RH (%)
TMP (K)

29. U-Wind (m/sec)
V-Wind (m/sec)
Forest (%)
AOD

30. Model Fitting Results
Max Obs. Per Day
101
Model Days
137 (37.5%)
Total Obs.
4,477
Local R2 values vary in time – daily model is necessary.
No residual spatial autocorrelation was found in 78% of the daily GWR models.
>=10 matched data records is needed to stabilize the model

31. Model Performance Evaluation
Mean
Min
Max
Model R2
0.86
0.56
0.92
CV R2
0.70
0.22
0.85
Putting all the data points together, we see unbiased estimates
Model CV

32. Spatial Pattern of Model Bias
Model Fitting
Cross Validation
Negative and positive model / CV residuals are randomly distributed.

33. Predicted Daily Concentration Surface

34. Model Predicted Mean PM2.5 Surface
Note: annual mean calculated with137 days

35. Comparison with CMAQ
General patterns agree, details differ

36. Caveats
Data coverage
MISR has the least coverage
MODIS
GOES

37. Data Quality - MISR Source: Kahn et al. 2010
urban biomass burning dusty
r = 0.92, mean absolute diff = 26%
r = 0.93, mean absolute diff = 32%
r = 0.87, mean absolute diff = 51%
Expected error globally: ±(0.05 or 0.20t)
In the US: ±(0.05 or 0.15t)
Source: Kahn et al. 2010
Source: Liu et al. 2004

38. Data Quality – MODIS DDV Product
Expected error: ±( t)
Source: Levy et al. 2010

39. Data Quality - GOES r = 0.79 over 10 northeastern sites, slope = 0.8
Source: Prados et al. 2007

40. The Use of Satellite Models
Currently for research
Spatial trends of PM2.5 at regional to national level
Interannual variability of PM2.5
Model calibration / validation
Exposure assessment for health effect studies
In the near future for research
Spatial trends at urban scale
Improved coverage and accuracy
Fused statistical – deterministic models
For regulation?