Machine Learning-Based Integration of High-Resolution Wildfire Smoke Simulations and Observations for Regional Health Impact Assessment

Abstract

Large wildfires are an increasing threat to the western U.S. In the 2017 fire season, extensive wildfires occurred across the Pacific Northwest (PNW). To evaluate public health impacts of wildfire smoke, we integrated numerical simulations and observations for regional fire events during August-September of 2017. A one-way coupled Weather Research and Forecasting and Community Multiscale Air Quality modeling system was used to simulate fire smoke transport and dispersion. To reduce modeling bias in fine particulate matter (PM_2.5) and to optimize smoke exposure estimates, we integrated modeling results with the high-resolution Multi-Angle Implementation of Atmospheric Correction satellite aerosol optical depth and the U.S. Environmental Protection Agency AirNow ground-level monitoring PM_2.5 concentrations. Three machine learning-based data fusion algorithms were applied: An ordinary multi-linear regression method, a generalized boosting method, and a random forest (RF) method. 10-Fold cross-validation found improved surface PM_2.5 estimation after data integration and bias correction, especially with the RF method. Lastly, to assess transient health effects of fire smoke, we applied the optimized high-resolution PM_2.5 exposure estimate in a short-term exposure-response function. Total estimated regional mortality attributable to PM_2.5 exposure during the smoke episode was 183 (95% confidence interval: 0, 432), with 85% of the PM_2.5 pollution and 95% of the consequent multiple-cause mortality contributed by fire emissions. This application demonstrates both the profound health impacts of fire smoke over the PNW and the need for a high-performance fire smoke forecasting and reanalysis system to reduce public health risks of smoke hazards in fire-prone regions.

Keywords: PM2.5 air pollution; fire smoke modeling; health impact assessment; machine learning-based data fusion.

Figure 1

Fire hotspots and smoke over the PNW region on 9/5/2017. (a) satellite imagery detected by the NASA/NOAA Suomi NPP satellite at 20:36 UTC (13:36 PDT) (credit: Image by the NASA’s Land, Atmosphere Near real-time Capability for EOS (LANCE/EOSDIS) Rapid Response team). (b) WRF-CMAQ PM_2.5 simulation based on the SENS experiment with all sources. PNW: Pacific Northwest. NASA: National Aeronautics and Space Administration. NOAA: National Oceanic and Atmospheric Administration. NPP: National Polar-orbiting Partnership. WRF: Weather Research and Forecasting model. CMAQ: Community Multiscale Air Quality model. PM_2.5: Fine particulate matter. SENS: sensitivity experiment. UTC: Coordinated Universal Time; PDT: Pacific Daylight Time; EOS: Earth Observing System.

Figure 2

Comparison of the MAIAC AOD at 550 nm and model simulated AOD results at 20:15 UTC (13:15 PDT) on 9/7/2017. (a) the MAIAC AOD product onboard the MODIS Aqua satellite; (b) CMAQ simulated AOD; (c) bias-adjusted AOD by the RF method; (d) same as (c) but by the GBM method. The r-values on top-left corners of the subplots (b)–(d) denote spatial correlation coefficients between each model result and corresponding MAIAC AOD product. MAIAC: Multi-Angle Implementation of Atmospheric Correction. AOD: aerosol optical depth. MODIS: Moderate Resolution Imaging Spectroradiometer. CMAQ: Community Multiscale Air Quality model. RF: Random forest. GBM: Generalized boosting model.

Figure 3

Comparisons of the aerosol vertical distribution at 11:30 UTC (04:30 PDT) on 9/7/2017 between the CATS satellite retrievals and the CMAQ simulation. (a) CATS total attenuated backscatter (unit: km⁻¹ sr⁻¹) at 1064 nm; (b) CATS aerosol types; (c) 2-D cross section of PM_2.5 concentrations (unit: μg m⁻³) from the WRF-CMAQ SENS experiment along the CATS satellite track. Please note that the X-axis range in subplot (c) is different due to the smaller model domain. (credit: The CATS products in subplots (a) and (b) were produced and distributed by NASA Goddard Space Flight Center). CATS: Cloud-Aerosol Transport System. SENS: sensitivity experiment.

Figure 4

Statistical evaluation results of the raw WRF-CMAQ SENS simulated PM_2.5 surface concentrations during 8/15–9/14/2017: (a) monthly averaged PM_2.5 surface concentrations (unit: μg m⁻³). The black circles and red numbers denote 16 selected ground sites for the time series evaluation in Figure 5; (b) temporal correlations (unitless) between the SENS simulation and AirNow observations at each ground site; (c) fractional biases (unit: 100%) based on the SENS simulation and AirNow observations; (d) RMSE (unit: μg m⁻³) based on the SENS simulation and AirNow observations. RMSE: root mean squared error.

Figure 5

Time series of daily PM_2.5 concentrations at the 16 selected AirNow ground sites in Figure 4a. The red solid lines denote AirNow ground measurements. The blue dashed lines denote CMAQ simulation results. The green dashed lines denote the MLR results. The golden dashed lines denote the RF results. The purple dashed lines denote the GBM results. The numbers on top-left corners of each subplot denote temporal correlations between these model results and observations. MLR: Multi-linear regression.

Figure 6

Comparisons of daily PM_2.5 model simulation before and after data integration with AirNow ground observations during the fire smoke pollution episode. (a) a scatter plot of PM_2.5 concentrations on a log scale based on the raw WRF-CMAQ SENS simulation and AirNow observations; (b) same as (a) but based on the MLR data integration method; (c) same as (a) but based on the RF method; (d) same as (a) but based on the GBM method. The color shading in all subplots denotes the PM_2.5 data density in terms of sample counts.

Figure 7

Comparison of PM_2.5 concentrations and speciation between the IMPROVE ground measurement network (left pie charts with site numbers) and the WRF-CMAQ SENS model simulation (right pie charts without site numbers). The colors in each pie chart denote PM_2.5 chemical compositions and the size of each pie chart denotes bulk PM_2.5 concentrations as shown by the reference numbers in the legend.

Figure 8

PM_2.5 pollution exposure evaluation and attribution. (a) multiple-cause mortality (unit: ^#) at the county level over the PNW in August-December of 2017; (b) county-level deaths (unit: ^#) attributed to the PM_2.5 exposure during the 2017 fire pollution episode; (c) PM_2.5 concentrations (unit: μg m⁻³) contributed by fire sources (CMAQ_RF minus CMAQ_CTRL); (d) PM_2.5 concentrations (unit: μg m⁻³) contributed by non-fire sources (CMAQ_CTRL).