A hybrid prediction model for PM2.5 mass and components using a chemical transport model and land use regression

摘要

GEOS-Chem, a chemical transport model, provides time-space continuous estimates of atmospheric pollutants including PM2.5 and its major components, but model predictions are not highly correlated with ground monitoring data. In addition, its spatial resolution is usually too coarse to characterize the spatial pattern in pollutant concentrations in urban environments. Our objective was to calibrate daily GEOS-Chem simulations using ground monitoring data and incorporating meteorological variables, land use terms and spatial-temporal lagged terms. Major PM2.5 components of our interest include sulfate, nitrate, organic carbon, elemental carbon, ammonium, sea salt and dust. We used a backward propagation neural network to calibrate GEOS-Chem predictions with a spatial resolution of 0.500 degrees x 0.667 degrees using monitoring data collected during the period from 2001 to 2010 for the Northeastern United States. Subsequently, we made predictions at 1 km x 1 km grid cells. We determined the accuracy of the spatial temporal predictions using ten-fold cross-validation and "leave-one-day-out" cross-validation techniques. We found a high total R-2 for PM2.5 mass (all data R-2 0.85, yearly values: 0.80-0.88) and PM2.5 components (R-2 for individual components were around 0.70-0.80). Our model makes it possible to assess spatially- and temporally-resolved short- and long-term exposures to PM2.5 mass and components for epidemiological studies. (C) 2016 Elsevier Ltd. All rights reserved.