Date post: | 03-Jan-2016 |
Category: |
Documents |
Upload: | ariel-moss |
View: | 15 times |
Download: | 1 times |
Theoretical Basis for Converting Satellite
Observations to Ground-Level PM 2.5 Concentrations
Applied Remote SEnsing Training (ARSET) – Air Quality
A project of NASA Applied Sciences
Training Workshop in Partnership with BAAQMD
Santa Clara, CASeptember 10 – 12, 2013
Estimate ground-level PM2.5 mass concentration (µg m-3) using satellite
derived AOD
OBJECTIVEObjective
2
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
201-300
151-200
101-150
51-100
0-50
Index
Values
PM10
(ug/m3)
PM2.5
(ug/m3)
Category
355-424
255-354
155-254
55-154
0-54
150.5-250.4
65.5-150.4
40.5-65.4
15.5-40.4
0-15.4Good
Very Unhealthy
Unhealthy
Unhealthy for Sensitive Groups
Moderate
What are we looking for?
3
Surface
Aerosol
Rayleigh Scattering
Water vapor + other gases (absorption)
Ozone
10km
Satellite
Sun
ColumnSatelliteMeasurement
Satellite retrieval issues - inversion (e.g. aerosol model, background).
Seven MODIS bands are utilized to derive aerosol properties: 0.47, 0.55, 0.65, 0.86, 1.24, 1.64, and 2.13 µm10X10 km2 Res.
dzAOD ext
What Do Satellites Provide?
4
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
5
May 11, 2007 May 12, 2007 May 13, 2007
May 14, 2007 May 15, 2007 May 16, 2007
MODIS-Terra True Color Images
6
May 11, 2007 May 12, 2007 May 13, 2007
May 14, 2007 May 15, 2007 May 16, 2007
MODIS-Terra AOD
7
AOD – PM Relation
• – particle density• Q – extinction
coefficient• re – effective radius
• fPBL – % AOD in PBL
• HPBL – mixing height
AODH
f
Q
rC
PBL
PBLe 3
4
Composition
Size distribution
Vertical profile
8
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)
2 4 6 8 10 12
9
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.
10
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! 11
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 data12
Simple Models from Early Days
13
Chu et al., 2003
Wang et al., 2003
13
Gupta, 2008
AOT-PM2.5 Relationship
14
Limitation: Major Unsolved Issue
All three Statistical models (TVM, MVM, ANN) underestimate high PM2.5 loadings
15
Vertical Distribution
Engel-Cox et al., 2006Al-Saadi et al., 2008
16
Aug 17Aug 18Aug 19
Aug 20Aug 21Aug 22Aug 23
Aug 25Aug 28
5 km
(courtesy of Dave Winker, P.I. CALIPSO)
What Satellites can provide for vertical information? - CALIPSO
17
TVM MVM
Advantages of using reanalysis meteorology along with satellite
18
TVM
MVM ANN
Spatial Comparison
MODIS-Terra, July 1, 2007
Satellite-derived PM2.5 fills the gap in surface measurements
All three methods underestimate the higher PM2.5 concentrations.
19
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?
20
MODIS AODCensus / traffic data
Predicted PM2.5 surface
land use
EPA air monitorsStatistical prediction model
NLDAS meteorology
Exposure Modeling Approach
Basic idea:
PM2.5 = f(AOD, meteorological variables, land use variables, etc.) + ε
21
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
22
Satellite-derived PM2.5 = x satellite AOD
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.
ModelAOD
PM
5.2
23
Scaling approach can be applied wherever there are satellite retrievals, but prediction accuracy can vary a lot.
24
Comparison of two approaches
Statistical 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
Scaling model
• 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)
25
Study Domain Design
Number of monitoring sites: 119 Exposure modeling domain: 700 x 700 km2
26
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
erForest_Cov Wind_Speed Temp
RH PBL AOD ~][PM
y)(x, 6y)(x, 5y)4(x,
),(3y)(x, 2),(1),(0y)(x,2.5
yxyxyx
27
PM2.5 (ug/m3) PBL (m)
RH (%) TMP (K)
Summary Statistics
28
U-Wind (m/sec) V-Wind (m/sec)
Forest (%) AOD
29
Model Fitting Results
>=10 matched data records is needed to stabilize the model
No residual spatial autocorrelation was found in 78% of the daily GWR models.
Local R2 values vary in time – daily model is necessary.
Max Obs. Per Day 101
Model Days 137 (37.5%)
Total Obs. 4,477
30
Model Performance Evaluation
Mean Min Max
Model R2
0.86 0.56 0.92
CV R2 0.70 0.22 0.85
Mo
de
l
CV
Putting all the data points together, we see unbiased estimates
31
Spatial Pattern of Model Bias
Model Fitting Cross Validation
Negative and positive model / CV residuals are randomly distributed.
32
Predicted Daily Concentration Surface
33
Model Predicted Mean PM2.5 Surface
Note: annual mean calculated with137 days 34
Comparison with CMAQ
General patterns agree, details differ35
Caveats• Data coverage
– MISR has the least coverage
MODIS
GOES
36
Data Quality - MISR
r = 0.93, mean absolute diff = 32%
urban biomass burning dusty
r = 0.92, mean absolute diff = 26%
Source: Kahn et al. 2010Source: Liu et al. 2004
r = 0.87, mean absolute diff = 51%
Expected error globally: ±(0.05 or 0.20t) In the US: ±(0.05 or 0.15t) 37
Data Quality – MODIS DDV Product
Source: Levy et al. 2010
Expected error: ±(0.05+0.15t)
38
Data Quality - GOES
Source: Prados et al. 2007r = 0.79 over 10 northeastern sites, slope = 0.8
39
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?40