Estimating PM2.5 using Fusion of Satellite Remote Sensing, GEOS-Chem, and other Parameters • Joel Schwartz • Harvard University
Critical Issues in Modeling/Fusion • Missing Data • Surrogates for emissions have time varying impacts • Aerosol Optical Depth has a time varying relationship to ground level particle concentrations • Nonlinearities • Many factors have multiplicative impacts on particle concentrations: High order interactions
Aerosol Optical Depth • Is based on direct physical measures at every square kilometer of earth twice per day • This provides high spatial and temporal resolution, BUT • It is based on scattering and absorption of light in the entire column of air, not just ground level • The scattering and absorption varies with particle color and size
Key Issues in Using AOD to Model PM2.5 • The correlation between AOD and PM2.5 is not that high • Differences in vertical profiles and particle composition that change extinction explain a good bit of that • These factors have more day to day variation, but vary spatially smoothly • Many grid-cell-days are missing
A density plot exhibiting the daily variation of AOD slopes between 2000-2008 during the stage 1 calibrations
Low pollution day November 15, 2003 AOD
Land Use Regression • Offers the possibility of highly geographically resolved estimates of exposures • Standard Models have limited temporal resolution • Many are calibrated with intensive monitoring campaigns with limited duration • This raises issues of spatial-temporal error
Concentrations Change with Time BC over time in Boston 1.4 BC 1.0 0.6 1995 2000 2005 2010 Year
What is worse is the decline varies depending on where you are
Combining with Land use improves The model
Other Approaches • Chemical Transport Models – Also provide daily (or better) time resolution • Hybrid Models (CTM + Land Use + Weather)
GEOS-Chem • GEOS-Chem is the most widely used chemical transport model in the world. • It incorporates detailed emissions inventories, meteorology, and nonlinear chemistry to estimate transport of aerosols and gases, and formation of secondary pollutants, including organic aerosols. • The individual species can be added to estimate PM2.5
Advantages • There is no missing data • It gives us PM components, not just PM2.5 • Disadvantages • It has a lower resolution (25km) • It has a lower R 2 • Uses modeled meteorology from Reanalysis data not real • Emission inventories are not perfect, and poorly resolved spatially and temporally • But: The errors from this model derive from completely different sources
Why not combine them all?
Put it All Together • MAIAC AOD from Aqua and Terra • AAI, O3 and NO2 from OMI • GEOS-Chem output • Land use and Meteorology • Monitoring Data • Neural Network Algorithm • Entire US, Daily 2000-2012
Out of Sample R2 • 0.85 for PM2.5 • 0.76 for Ozone • Daily predictions for each of 11 million 1km cells in the Continental US for each day Jan 1 2000-Dec 31 2012.
Recommend
More recommend
