Aerosols have adverse health effects and play a significant role in the climate as well. The Multiangle Implementation of Atmospheric Correction (MAIAC) provides Aerosol Optical Depth (AOD) at high temporal (daily) and spatial (1 km) resolution, making it particularly useful to infer and characterize spatiotemporal variability of aerosols at a fine spatial scale for exposure assessment and health studies. However, clouds and conditions of high surface reflectance result in a significant proportion of missing MAIAC AOD. To fill these gaps, we present an imputation approach using deep learning with downscaling. Using a baseline autoencoder, we leverage residual connections in deep neural networks to boost learning and parameter sharing to reduce overfitting, and conduct bagging to reduce error variance in the imputations. Downscaled through a similar auto-encoder based deep residual network, Modern-Era Retrospective analysis for Research and Applications Version 2 (MERRA-2) GMI Replay Simulation (M2GMI) data were introduced to the network as an important gap-filling feature that varies in space to be used for missingness imputations. Imputing weekly MAIAC AOD from 2000 to 2016 over California, a state with considerable geographic heterogeneity, our full (non-full) residual network achieved mean R2 = 0.94 (0.86) [RMSE = 0.007 (0.01)] in an independent test, showing considerably better performance than a regular neural network or non-linear generalized additive model (mean R2 = 0.78-0.81; mean RMSE = 0.013-0.015). The adjusted imputed as well as combined imputed and observed MAIAC AOD showed strong correlation with Aerosol Robotic Network (AERONET) AOD (R = 0.83; R2 = 0.69, RMSE = 0.04). Our results show that we can generate reliable imputations of missing AOD through a deep learning approach, having important downstream air quality modeling applications.

Introduction

Estimating PM_2.5 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 PM_2.5 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 PM_2.5 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 PM_2.5 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 PM_2.5 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 PM_2.5 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 PM_2.5. 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 PM_2.5 has previously been challenging to predict. Prediction uncertainty estimates can also inform further model development and measurement error evaluations in exposure and health studies.

BACKGROUND: Wildfires in the Western United States are a growing and significant source of air pollution that is eroding decades of progress in air pollution reduction. The effects on preterm birth during critical periods of pregnancy are unknown. METHODS: We assessed associations between prenatal exposure to wildland fire smoke and risk of preterm birth (gestational age < 37 weeks). We assigned smoke exposure to geocoded residence at birth for all live singleton births in California conceived 2007-2018, using weekly average concentrations of particulate matter ≤ 2.5 µm (PM2.5) attributable to wildland fires from United States Environmental Protection Agencys Community Multiscale Air Quality Model. Logistic regression yielded odds ratio (OR) for preterm birth in relation to increases in average exposure across the whole pregnancy, each trimester, and each week of pregnancy. Models adjusted for season, age, education, race/ethnicity, medical insurance, and smoking of the birthing parent. RESULTS: For the 5,155,026 births, higher wildland fire PM2.5 exposure averaged across pregnancy, or any trimester, was associated with higher odds of preterm birth. The OR for an increase of 1 µg/m3 of average wildland fire PM2.5 during pregnancy was 1.013 (95 % CI:1.008,1.017). Wildland fire PM2.5 during most weeks of pregnancy was associated with higher odds. Strongest estimates were observed in weeks in the second and third trimesters. A 10 µg/m3 increase in average wildland fire PM2·5 in gestational week 23 was associated with OR = 1.034; 95 % CI: 1.019, 1.049 for preterm birth. CONCLUSIONS: Preterm birth is sensitive to wildland fire PM2.5; therefore, we must reduce exposure during pregnancy.