Ensemble-based deep learning for estimating PM2.5 over California with multisource big data including wildfire smoke.

INTRODUCTION: Estimating PM2.5 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. METHODS: Using ensemble-based deep learning with big data fused from multiple sources we developed a PM2.5 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 PM2.5 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 PM2.5 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. RESULTS: Ensemble deep learning to predict PM2.5 achieved an overall mean training RMSE of 1.54 µg/m3 (R2: 0.94) and test RMSE of 2.29 µg/m3 (R2: 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 PM2.5 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/m3). Statewide predictions indicated that our model can capture the spatial distribution and temporal peaks in wildfire-related PM2.5. The coefficient of variation indicated highest uncertainty over deciduous and mixed forests and open water land covers. CONCLUSION: 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 PM2.5 has previously been challenging to predict. Prediction uncertainty estimates can also inform further model development and measurement error evaluations in exposure and health studies.

US particulate matter air quality improves except in wildfire-prone areas.

Using data from rural monitoring sites across the contiguous United States, we evaluated fine particulate matter (PM2.5) trends for 1988-2016. We calculate trends in the policy-relevant 98th quantile of PM2.5 using Quantile Regression. We use Kriging and Gaussian Geostatistical Simulations to interpolate trends between observed data points. Overall, we found positive trends in 98th quantile PM2.5 at sites within the Northwest United States (average 0.21 ± 0.12 µg·m-3·y-1; ±95% confidence interval). This was in contrast with sites throughout the rest of country, which showed a negative trend in 98th quantile PM2.5, likely due to reductions in anthropogenic emissions (average -0.66 ± 0.10 µg·m-3·y-1). The positive trend in 98th quantile PM2.5 is due to wildfire activity and was supported by positive trends in total carbon and no trend in sulfate across the Northwest. We also evaluated daily moderate resolution imaging spectroradiometer (MODIS) aerosol optical depth (AOD) for 2002-2017 throughout the United States to compare with ground-based trends. For both Interagency Monitoring of Protected Visual Environments (IMPROVE) PM2.5 and MODIS AOD datasets, we found positive 98th quantile trends in the Northwest (1.77 ± 0.68% and 2.12 ± 0.81% per year, respectively) through 2016. The trend in Northwest AOD is even greater if data for the high-fire year of 2017 are included. These results indicate a decrease in PM2.5 over most of the country but a positive trend in the 98th quantile PM2.5 across the Northwest due to wildfires.

Evaluation of the KCl denuder method for gaseous oxidized mercury using HgBr2 at an in-service AMNet site.

During the summer of 2013, we examined the performance of KCl-coated denuders for measuring gaseous oxidized mercury (GOM) by calibrating with a known source of GOM (i.e., HgBr2) at the North Birmingham SouthEastern Aerosol Research and Characterization (SEARCH) site. We found that KCl-coated denuders have near 95% collection efficiency for HgBr2 in zero air (i.e., air scrubbed of mercury and ozone). However, in ambient air, the efficiency of KCl-coated denuders in capturing HgBr2 dropped to 20-54%. We also found that absolute humidity and ozone each demonstrate a significant inverse correlation with HgBr2 recovery in ambient air. Subsequent laboratory tests with HgBr2 and the KCl-coated denuder show that ozone and absolute humidity cause the release of gaseous elemental Hg from the denuder and thus appear to explain the low recovery in ambient air. Based on these findings, we infer that the KCl denuder method underestimates atmospheric GOM concentrations and a calibration system is needed to accurately measure GOM. The system described in this paper for HgBr2 could be implemented with existing mercury speciation instrumentation and this would improve our knowledge of the response to one potentially important GOM compound.