Comparison of data fusion techniques using observations, chemical transport models, and satellite retrievals for air pollution exposure estimates in health studies

Tuesday, 16 December 2014: 2:55 PM
Heather Holmes1, Howard H. Chang2, James A. Mulholland3 and Armistead Russell3, (1)University of Nevada Reno, Atmospheric Sciences Program, Department of Physics, Reno, NV, United States, (2)Emory University, Department of Biostatistics and Bioinformatics, Atlanta, GA, United States, (3)Georgia Institute of Technology Main Campus, School of Civil and Environmental Engineering, Atlanta, GA, United States
As the spatial resolution of heath data continues to improve (i.e., from aggregates within census tracts to point-level residential addresses from patient records), the spatial resolution and coverage of air pollution exposure estimates must also improve for studies that aim to investigate the impact of air pollution on health outcomes. As part of the Southeastern Center for Air Pollution and Epidemiology (SCAPE) air quality models are being developed to provide enhanced exposure metrics for spatially resolved time-series epidemiologic studies. To generate spatiotemporal air quality metrics unique data fusion approaches are developed that combine ground level observations from regulatory monitoring networks, chemical transport model results (CTM), and aerosol optical depth (AOD) from satellite retrievals. The objectives of this presentation are to compare data fusion approaches and assess the ability of the different data fusion models to predict ground level PM2.5 concentrations specific for use in exposure assessments. We present data fusion results from statistical downscaling of gridded CTM simulations and AOD data, to predict ambient ground level PM2.5 concentrations. The downscaling framework will focus on four model configurations using CTM results (CMAQ) or satellite AOD as a spatial predictor, with and without additional land use and meteorology predictors. Daily, spatially resolved (12-km) PM2.5 concentrations for the southeastern United States from 2003 to 2005 will be shown. Data withholding and monitoring data from an independent dataset are used to calculate prediction performance statistics for comparison. The independent dataset will also be used to compare these results with results from previously published data fusion models that are being increasingly utilized for health analyses in the United States. It is expected that the downscaling approach with land use regression terms will improve prediction performance, especially for the CTM spatial predictor where the temporal data coverage is more complete than it is for satellite AOD.