Abstract View
Correcting Biases in Speciated PM2.5 Simulations Using a Geographically Weighted Regression
CARLOS HERNANDEZ, Ksakousti Skyllakou, Pablo Garcia, Brian Dinkelacker, Spyros Pandis, Allen Robinson, Peter Adams, Carnegie Mellon University
Abstract Number: 423
Working Group: Aerosol Exposure
Abstract
The ability to provide speciated and source-resolved PM2.5 estimates make chemical transport models an alluring alternative, or complement, to empirical models for exposure assessments. Correcting characteristic biases in simulated PM2.5 could encourage the adoption of chemical transport models in this area. We use geographically weighted regression, similar to van Donkelaar et al. [Environ. Sci. Technol., 2015, 49, 17, 10482-10491], to predict the bias between speciated PM2.5 simulations and observations in the continental U.S. The regression model is trained by comparing simulated annual averages of NH4, NO3, SO4, EC and OA in 2010 and 2001 to corresponding observations from the CSN and IMPROVE networks. Biases for speciated PM2.5 simulations are predicted at U.S. census tract geographies. Corrected simulations are then aggregated to population-weighted averages for metropolitan statistical areas (MSAs). Correlations between MSA-level simulated total PM2.5 and empirical estimates [Kim et al. PLOS ONE. 2020, 15, 2, e0228535] improve from 0.70 to 0.89 in 2010, and from 0.62 to 0.90 in 2001. On average, errors in fractional composition between speciated monitors and simulations were reduced by 40% in 2010, and 50% in 2001. A leave-one-out cross-validation shows a reduction of biases across all simulated components at speciated monitors. For OA and NO3, normalized mean biases ranged from -20% to -35% before correction, and 0% to 3% after correction. For EC, normalized mean biases ranged from -10% to 25% before correction, and 2% to 5% after correction. For SO4 and NH4, normalized mean biases ranged from 3% to 10% before correction, and -2% to 1% after correction. Results from a ten-fold cross-validation are consistent with those from the leave-one-out cross-validation.