A multi-analysis approach for estimating regional health impacts from the 2017 Northern California wildfires

Figures & data

Figure 1. (a) Visible satellite imagery from the VIIRS instrument aboard Suomi-NPP and fire hot spot detections (red) from the VIIRS instrument aboard the Suomi-NPP satellite for October 9, 2017. The image is downloaded from NASA Worldview website. (b) Fire perimeters of the Atlas, Tubbs, Nuns, Redwood Valley, and Pocket wildfires (red). Other prescribed fires and wildfires occurring during the Oct 8–20, 2017 time period are shown in blue. Fire perimeters are from the GEOMAC system and hot spot locations are from the MODIS and VIIRS instruments aboard the Terra, Aqua and SUOMI-NPP satellites

Figure 2. Locations of PM_2.5 air quality monitors. Circles are permanent monitors from the EPA AQS System. Triangles are temporary monitors deployed for wildfires. The circles and triangles are color-coded by the Air Quality Index by the maximum measured 24-hr average PM_2.5 value during the October 6–20, 2017 time period

Table 1. WRF-CMAQ model simulation summary

Settings	No Fires	Baseline	GTP
Period	Oct 2–20, 2017	Oct 2–20, 2017	Oct 2–20, 2017
Resolution	Horizontal: 4-km Vertical: 37 layers	Horizontal: 4-km Vertical: 37 layers	Horizontal: 4-km Vertical: 37 layers
Meteorology	WRF v3.7	WRF v3.7	WRF v3.7
Chemistry	CMAQv5.2, SAPRC07, AERO6	CMAQv5.2, SAPRC07, AERO6	CMAQv5.2, SAPRC07, AERO6
Fire Emissions	–	BlueSky v3.5.1	BlueSky v3.5.1
Fire Activity	–	Five Wine Country Fires: GOES-16 Other Fires: MODIS/VIIRS	Five Wine Country Fires: GOES-16 Other Fires: MODIS/VIIRS
Fire Diurnal Profile	–	Five Wine Country Fires: CMAQ (default) Other Fires: CMAQ (default)	Five Wine Country Fires: GOES-16 Other Fires: CMAQ (default)
Non-Fire Emissions	CARB area and non-road, EMFAC2017 on-road, BAAQMD facility-level point source emissions, BEIS3.61	CARB area and non-road, EMFAC2017 on-road, BAAQMD facility-level point source emissions, BEIS3.61	CARB area and non-road, EMFAC2017 on-road, BAAQMD facility-level point source emissions, BEIS3.61

Table 2. Data fusion and machine learning approaches

Method	Modeling Dataset	Resolution	Surface Data	Meteorological Data	Remotely Sensed Data
Data Fusion (DF)	–	3-km	AQS	–	Standard MODIS AOD
Multi-linear Regression (MLR)	WRF-CMAQ-Baseline	4-km	AirNow	WRF, NARR	MAIAC AOD
Generalized Boosting (GB)	WRF-CMAQ-Baseline	4-km	AirNow	WRF, NARR	MAIAC AOD
Random Forest (RF)	WRF-CMAQ-Baseline	4-km	AirNow	WRF, NARR	MAIAC AOD

Figure 3. Illustration of the components involved in data fusion. (a) Surface PM_2.5 monitoring data from EPA AQS and (b) MODIS AOD merged to create (c) a surface of PM_2.5 concentrations on a 3-km grid for October 9, 2017

Table 3. Definitions of quantitative analysis metrics. M = modeled data. O = observed data

Metric	Equation
Mean Bias	$\frac{1}{n}\sum _1^n\left(M-O\right)$
Median Absolute Difference	$Median\left(\left|M_i-O_i\right|,i=1,n\right)$
Fractional Bias	$\frac{1}{n}\left(\frac{\sum _1^n\left(M-O\right)}{\sum _1^n\left(\frac{\left(M+O\right)}{2}\right)}\right)$
Pearson Correlation (r)	$\frac{{\sum }^{\left(M-\bar{M}\right)}\left(O-\bar{O}\right)}{\sqrt{{\sum }^{{\left(M-\bar{M}\right)}^2}{\sum }^{{\left(O-\bar{O}\right)}^2}}}$
Root Mean Square Error (RMSE)	$\sqrt{\frac{\sum _1^n{\left(M-O\right)}^2}{n}}$

Figure 4. Hourly PM_2.5 emissions from wildfires and prescribed fires in California, October 8–20, 2017. Black line (Baseline): diurnal profile of the emissions from the five Wine Country wildfires, calculated using BSF and allocated hourly based on the default profile in CMAQ. Dashed line (GTP): hourly emissions from the five Wine Country wildfires, calculated using BSF and allocated hourly based on GOES-16 FDC data. Gray line (Other Fires): hourly emissions from all other fires, calculated using BSF and allocated hourly based on the default profile in CMAQ

Figure 5. Panel of visible satellite imagery and WRF-CMAQ runs at 11:00 am PDT October 9, 2017. (a) Visible GOES-16 satellite imagery and surface 24-hr average PM_2.5 concentrations (circles) from EPA AirNowTech color-coded by air quality index (Figure: NOAA AerosolWatch) (b) the Baseline total column PM_2.5, (c) the Baseline surface 1-hr average PM_2.5 concentration, and (d) the GTP surface 1-hr average PM_2.5 concentrations (same scale of PM_2.5 concentrations as in 5c)

Figure 6. (a) Locations of PM_2.5 monitors grouped by region. (b) Time series comparison of measured and modeled 24-hr average PM_2.5 concentrations by region. Black lines: observations; colored lines: WRF-CMAQ model simulations (NoFires, Baseline, and GTP), data fusion (DF), and three machine-learning analyses (GB, MLR, and RF). See key at top

Table 4. Permanent monitor location analysis. Numbers in bold and italics are the best and second-best model performer respectively. 1179 data points (mean and standard deviation) or data pairs (other metrics) for all cases except in the DF case, where there are 1327 data points/pairs

Model	Mean (µg/m^3)	Standard Deviation (µg/m^3)	Pearson Correlation	RMSE (µg/m^3)	Fractional Bias	Mean Bias (µg/m^3)	Median Absolute Difference (µg/m^3)
NoFires	11.33	6.37	0.03	16.20	−0.23	6.71	5.76
Baseline	24.95	25.71	0.65	12.25	0.16	6.91	6.08
GTP	26.64	38.19	0.64	12.65	0.19	8.60	5.82
DF	16.81	15.26	0.95	5.73	−0.04	−1.23	0.71
GB	17.25	12.25	0.81	9.60	−0.02	−0.79	3.45
MLR	16.66	11.19	0.68	11.95	−0.04	−1.38	4.11
RF	17.63	11.84	0.86	8.32	−0.01	−0.41	2.66
Measured	18.04	15.48

Table 5. Temporary monitor location analysis. Numbers in bold and italics are the best and second-best model performer respectively. 406 data points (mean and standard deviation) or data pairs (other metrics) for all cases

Model	Mean (µg/m^3)	Standard Deviation (µg/m^3)	Pearson Correlation	RMSE (µg/m^3)	Fractional Bias	Mean Bias (µg/m^3)	Median Absolute Difference (µg/m^3)
NoFires	5.59	3.98	−0.05	24.87	−0.61	−17.16	9.85
Baseline	26.84	44.16	0.34	22.93	0.08	4.10	6.80
GTP	32.06	69.81	0.29	23.58	0.17	9.31	6.95
DF	11.98	7.00	0.38	23.23	−0.31	−10.76	4.96
GB	18.24	15.86	0.50	21.62	−0.11	−4.51	5.11
MLR	17.55	18.59	0.35	22.92	−0.13	−5.19	4.75
RF	18.49	15.55	0.49	21.74	−0.10	−4.25	4.85
Measured	22.75	24.84

Figure 7. Comparison of measured (x-axis) and modeled (y-axis) 24-hr average PM_2.5 concentrations at all monitoring locations. Red circles: permanent monitors; blue circles: temporary monitors. The top panels are the three WRF-CMAQ modeling results and the four bottom panels are the data fusion (DF) and machine learning (MLR, GB, and RF) results

Figure 8. Daily model performance at the permanent (red) and temporary (blue) monitor locations in terms of (a) root mean square error (RMSE) and (b) mean bias. Vertical blue lines indicate October 8, 2017, the start of the wildfires

Figure 9. PM_2.5 bias by Air Quality Index (AQI) category. The numbers at the bottom of each panel are the number of model-monitor pairs in the AQI category

Figure 10. Multiple-cause mortality attributable to PM_2.5 exposure using a relative risk of 1.1% (95% CI: 0, 0.26%) per 10 μg/m³ increase of surface PM_2.5 concentration (Johnston et al. 2012). (a) Mortality related to PM_2.5 exposure from the NoFires case and (b) the additional mortality due to smoke from the wildland fires (the five Wine Country wildfires and other smaller wildland fires) from the RF case

Johnston, F. H., S. B. Henderson, Y. Chen, J. T. Randerson, M. Marlier, R. S. DeFries, P. Kinney, D. M. J. S. Bowman, and M. Brauer. 2012. Estimated global mortality attributable to smoke from landscape fires. Environ. Health Perspect. 120:695–701. doi:10.1289/ehp.1104422.

 PubMed Web of Science ®Google Scholar