Spatiotemporal resolved models were developed predicting daily fine particulate matter (PM2.5) concentrations across Switzerland from 2003 to 2013. Relatively sparse PM2.5 monitoring data was supplemented by imputing PM2.5 concentrations at PM10 sites, using PM2.5/PM10 ratios at co-located sites. Daily PM2.5 concentrations were first estimated at a 1 × 1km resolution across Switzerland, using Multiangle Implementation of Atmospheric Correction (MAIAC) spectral aerosol optical depth (AOD) data in combination with spatiotemporal predictor data in a four stage approach. Mixed effect models (1) were used to predict PM2.5 in cells with AOD but without PM2.5 measurements (2). A generalized additive mixed model with spatial smoothing was applied to generate grid cell predictions for those grid cells where AOD was missing (3). Finally, local PM2.5 predictions were estimated at each monitoring site by regressing the residuals from the 1 × 1km estimate against local spatial and temporal variables using machine learning techniques (4) and adding them to the stage 3 global estimates. The global (1 km) and local (100 m) models explained on average 73% of the total,71% of the spatial and 75% of the temporal variation (all cross validated) globally and on average 89% (total) 95% (spatial) and 88% (temporal) of the variation locally in measured PM2.5 concentrations.
Keywords: Air pollution; Exposure assessment; Fine particulate matter; Satellite; Spatiotemporal models.
Copyright © 2017 Elsevier Ltd. All rights reserved.