Ensemble-based deep learning for estimating PM2.5 over California with multisource big data including wildfire smoke
Lianfa Li,Mariam Girguis,Frederick Lurmann,Nathan Pavlovic,Crystal McClure,Meredith Franklin,Jun Wu,Luke D Oman,Carrie Breton,Frank Gilliland,Rima Habre,Luke D. Oman
DOI: https://doi.org/10.1016/j.envint.2020.106143
IF: 11.8
2020-12-01
Environment International
Abstract:<h3 class="u-h4 u-margin-m-top u-margin-xs-bottom">Introduction</h3><p>Estimating PM<sub>2.5</sub> concentrations and their prediction uncertainties at a high spatiotemporal resolution is important for air pollution health effect studies. This is particularly challenging for California, which has high variability in natural (e.g, wildfires, dust) and anthropogenic emissions, meteorology, topography (e.g. desert surfaces, mountains, snow cover) and land use.</p><h3 class="u-h4 u-margin-m-top u-margin-xs-bottom">Methods</h3><p>Using ensemble-based deep learning with big data fused from multiple sources we developed a PM<sub>2.5</sub> prediction model with uncertainty estimates at a high spatial (1 km × 1 km) and temporal (weekly) resolution for a 10-year time span (2008–2017). We leveraged autoencoder-based full residual deep networks to model complex nonlinear interrelationships among PM<sub>2.5</sub> emission, transport and dispersion factors and other influential features. These included remote sensing data (MAIAC aerosol optical depth (AOD), normalized difference vegetation index, impervious surface), MERRA-2 GMI Replay Simulation (M2GMI) output, wildfire smoke plume dispersion, meteorology, land cover, traffic, elevation, and spatiotemporal trends (geo-coordinates, temporal basis functions, time index). As one of the primary predictors of interest with substantial missing data in California related to bright surfaces, cloud cover and other known interferences, missing MAIAC AOD observations were imputed and adjusted for relative humidity and vertical distribution. Wildfire smoke contribution to PM<sub>2.5</sub> was also calculated through HYSPLIT dispersion modeling of smoke emissions derived from MODIS fire radiative power using the Fire Energetics and Emissions Research version 1.0 model.</p><h3 class="u-h4 u-margin-m-top u-margin-xs-bottom">Results</h3><p>Ensemble deep learning to predict PM<sub>2.5</sub> achieved an overall mean training RMSE of 1.54 μg/m<sup>3</sup> (R<sup>2</sup>: 0.94) and test RMSE of 2.29 μg/m<sup>3</sup> (R<sup>2</sup>: 0.87). The top predictors included M2GMI carbon monoxide mixing ratio in the bottom layer, temporal basis functions, spatial location, air temperature, MAIAC AOD, and PM<sub>2.5</sub> sea salt mass concentration. In an independent test using three long-term AQS sites and one short-term non-AQS site, our model achieved a high correlation (>0.8) and a low RMSE (<3 μg/m<sup>3</sup>). Statewide predictions indicated that our model can capture the spatial distribution and temporal peaks in wildfire-related PM<sub>2.5</sub>. The coefficient of variation indicated highest uncertainty over deciduous and mixed forests and open water land covers.</p><h3 class="u-h4 u-margin-m-top u-margin-xs-bottom">Conclusion</h3><p>Our method can be generalized to other regions, including those having a mix of major urban areas, deserts, intensive smoke events, snow cover and complex terrains, where PM<sub>2.5</sub> has previously been challenging to predict. Prediction uncertainty estimates can also inform further model development and measurement error evaluations in exposure and health studies.</p>
environmental sciences