Estimating Wildfire Smoke Concentrations during the October 2017 California Fires through BME Space/Time Data Fusion of Observed, Modeled, and Satellite-Derived PM2.5.

Exposure to wildfire smoke causes adverse health outcomes, suggesting the importance of accurately estimating smoke concentrations. Geostatistical methods can combine observed, modeled, and satellite-derived concentrations to produce accurate estimates. Here, we estimate daily average ground-level PM2.5 concentrations at a 1 km resolution during the October 2017 California wildfires, using the Constant Air Quality Model Performance (CAMP) and Bayesian Maximum Entropy (BME) methods to bias-correct and fuse three concentration datasets: permanent and temporary monitoring stations, a chemical transport model (CTM), and satellite-derived estimates. Four BME space/time kriging and data fusion methods were evaluated. All BME methods produce more accurate estimates than the standalone CTM and satellite products. Adding temporary station data increases the R2 by 36%. The data fusion of observations with the CAMP-corrected CTM and satellite-derived concentrations provides the best estimate (R2 = 0.713) in fire-impacted regions, emphasizing the importance of combining multiple datasets. We estimate that approximately 65,000 people were exposed to very unhealthy air (daily average PM2.5 ≥ 150.5 µg/m3).

Machine Learning Approach To Estimate Hourly Exposure to Fine Particulate Matter for Urban, Rural, and Remote Populations during Wildfire Seasons.

Exposure to wildfire smoke averaged over 24-hour periods has been associated with a wide range of acute cardiopulmonary events, but little is known about the effects of sub-daily exposures immediately preceding these events. One challenge for studying sub-daily effects is the lack of spatially and temporally resolved estimates of smoke exposures. Inexpensive and globally applicable tools to reliably estimate exposure are needed. Here we describe a Random Forests machine learning approach to estimate 1-hour average population exposure to fine particulate matter during wildfire seasons from 2010 to 2015 in British Columbia, Canada, at a 5 km × 5 km resolution. The model uses remotely sensed fire activity, meteorology assimilated from multiple data sources, and geographic/ecological information. Compared with observations, model predictions had a correlation of 0.93, root mean squared error of 3.2 µg/m3, mean fractional bias of 15.1%, and mean fractional error of 44.7%. Spatial cross-validation indicated an overall correlation of 0.60, with an interquartile range from 0.48 to 0.70 across monitors. This model can be adapted for global use, even in locations without air quality monitoring. It is useful for epidemiologic studies on sub-daily exposure to wildfire smoke and for informing public health actions if operationalized in near-real-time.

Spatiotemporal prediction of fine particulate matter during the 2008 northern California wildfires using machine learning.

Estimating population exposure to particulate matter during wildfires can be difficult because of insufficient monitoring data to capture the spatiotemporal variability of smoke plumes. Chemical transport models (CTMs) and satellite retrievals provide spatiotemporal data that may be useful in predicting PM2.5 during wildfires. We estimated PM2.5 concentrations during the 2008 northern California wildfires using 10-fold cross-validation (CV) to select an optimal prediction model from a set of 11 statistical algorithms and 29 predictor variables. The variables included CTM output, three measures of satellite aerosol optical depth, distance to the nearest fires, meteorological data, and land use, traffic, spatial location, and temporal characteristics. The generalized boosting model (GBM) with 29 predictor variables had the lowest CV root mean squared error and a CV-R2 of 0.803. The most important predictor variable was the Geostationary Operational Environmental Satellite Aerosol/Smoke Product (GASP) Aerosol Optical Depth (AOD), followed by the CTM output and distance to the nearest fire cluster. Parsimonious models with various combinations of fewer variables also predicted PM2.5 well. Using machine learning algorithms to combine spatiotemporal data from satellites and CTMs can reliably predict PM2.5 concentrations during a major wildfire event.

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

Smoke impacts from large wildfires are mounting, and the projection is for more such events in the future as the one experienced October 2017 in Northern California, and subsequently in 2018 and 2020. Further, the evidence is growing about the health impacts from these events which are also difficult to simulate. Therefore, we simulated air quality conditions using a suite of remotely-sensed data, surface observational data, chemical transport modeling with WRF-CMAQ, one data fusion, and three machine learning methods to arrive at datasets useful to air quality and health impact analyses. To demonstrate these analyses, we estimated the health impacts from smoke impacts during wildfires in October 8-20, 2017, in Northern California, when over 7 million people were exposed to Unhealthy to Very Unhealthy air quality conditions. We investigated using the 5-min available GOES-16 fire detection data to simulate timing of fire activity to allocate emissions hourly for the WRF-CMAQ system. Interestingly, this approach did not necessarily improve overall results, however it was key to simulating the initial 12-hr explosive fire activity and smoke impacts. To improve these results, we applied one data fusion and three machine learning algorithms. We also had a unique opportunity to evaluate results with temporary monitors deployed specifically for wildfires, and performance was markedly different. For example, at the permanent monitoring locations, the WRF-CMAQ simulations had a Pearson correlation of 0.65, and the data fusion approach improved this (Pearson correlation = 0.95), while at the temporary monitor locations across all cases, the best Pearson correlation was 0.5. Overall, WRF-CMAQ simulations were biased high and the geostatistical methods were biased low. Finally, we applied the optimized PM2.5 exposure estimate in an exposure-response function. Estimated mortality attributable to PM2.5 exposure during the smoke episode was 83 (95% CI: 0, 196) with 47% attributable to wildland fire smoke.Implications: Large wildfires in the United States and in particular California are becoming increasingly common. Associated with these large wildfires are air quality and health impact to millions of people from the smoke. We simulated air quality conditions using a suite of remotely-sensed data, surface observational data, chemical transport modeling, one data fusion, and three machine learning methods to arrive at datasets useful to air quality and health impact analyses from the October 2017 Northern California wildfires. Temporary monitors deployed for the wildfires provided an important model evaluation dataset. Total estimated regional mortality attributable to PM2.5 exposure during the smoke episode was 83 (95% confidence interval: 0, 196) with 47% of these deaths attributable to the wildland fire smoke. This illustrates the profound effect that even a 12-day exposure to wildland fire smoke can have on human health.

The Comprehensive Fire Information Reconciled Emissions (CFIRE) inventory: Wildland fire emissions developed for the 2011 and 2014 U.S. National Emissions Inventory.

Wildland fire emissions from both wildfires and prescribed fires represent a major component of overall U.S. emissions. Obtaining an accurate, time-resolved inventory of these emissions is important for many purposes, including to account for emissions of greenhouse gases and short-lived climate forcers, as well as to model air quality for health, regulatory, and planning purposes. For the U.S. Environmental Protection Agency's 2011 and 2014 National Emissions Inventories, a new methodology was developed to reconcile the wide range of available fire information sources into a single coherent inventory. The Comprehensive Fire Information Reconciled Emissions (CFIRE) inventory effort utilized satellite fire detections as well as a large number of national, state, tribal, and local databases. The methodology and results for CONUS and Alaska were documented and compared against other fire emissions databases, and the efficacy of the overall effort was evaluated. Results show the overall spatial pattern differences and relative seasonality of wildfires and prescribed fires across the country. Prescribed burn emissions occurred primarily in non-summer months were concentrated in the Southeast, Northwest, and lower Midwest, and were relatively consistent year to year. Wildfire emissions were much more variable but occurred primarily in the summer and fall. Overall, CFIRE represents a third of total emitted PM2.5 across all sources in the National Emissions Inventory, with prescribed fires accounting for nearly half of all CFIRE emissions. Compared with other wildland fire emissions inventories derived solely from satellite detections, the CFIRE inventory shows markedly increased emissions, reflecting the importance of the multiple national and regional databases included in CFIRE in capturing small fires and prescribed fires in particular. Implications: Wildland fire emissions inventories need to incorporate multiple sources of fire information in order to better represent the full range of fire activity, including prescribed burns and smaller fires. For the 2011 and 2014 U.S. National Emissions Inventory, a methodology was developed to collect, associate, and reconcile fire information from satellite data as well as a large number of national, regional, state, local, and tribal fire information databases across the country. The resulting emissions inventory shows the importance of this type of integration and reconciliation when compared against other emissions inventories for the same period.

Source apportionment of fine particulate matter in Phoenix, AZ, using positive matrix factorization.

Speciated particulate matter (PM)2.5 data collected as part of the Interagency Monitoring of Protected Visual Environments (IMPROVE) program in Phoenix, AZ, from April 2001 through October 2003 were analyzed using the multivariate receptor model, positive matrix factorization (PMF). Over 250 samples and 24 species were used, including the organic carbon and elemental carbon analytical temperature fractions from the thermal optical reflectance method. A two-step approach was used. First, the species excluding the carbon fractions were used, and initially eight factors were identified; non-soil potassium was calculated and included to better refine the burning factor. Next, the mass associated with the burning factor was removed, and the data set rerun with the carbon fractions. Results were very similar (i.e., within a few percent), but this step enabled a separation of the mobile factor into gasoline and diesel vehicle emissions. The identified factors were burning (on average 2% of the mass), secondary transport (7%), regional power generation (13%), dust (25%), nitrate (9%), industrial As/Pb/Se (2%), Cu/Ni/V (7%), diesel (9%), and general mobile (26%). The overall contribution from mobile sources also increased, as some mass (OC and nitrate) from the nitrate and regional power generation factors were apportioned with the mobile factors. This approach allowed better apportionment of carbon as well as total mass. Additionally, the use of multiple supporting analyses, including air mass trajectories, activity trends, and emission inventory information, helped increase confidence in factor identification.

High-resolution MODIS aerosol retrieval during wildfire events in California for use in exposure assessment.

Retrieval of aerosol optical depth (AOD) from the Moderate Resolution Imaging Spectroradiometer (MODIS) using the Collection 5 (C005) algorithm provides large-scale (10 × 10 km) estimates that can be used to predict surface layer concentrations of particulate matter with aerodynamic diameter smaller than 2.5 µm (PM2.5). However, these large-scale estimates are not suitable for identifying intraurban variability of surface PM2.5 concentrations during wildfire events when individual plumes impact populated areas. We demonstrate a method for providing high-resolution (2.5 km) kernel-smoothed estimates of AOD over California during the 2008 northern California fires. The method uses high-resolution surface reflectance ratios of the 0.66 and 2.12 µm channels, a locally derived aerosol optical model characteristic of fresh wildfire plumes, and a relaxed cloud filter. Results show that the AOD derived for the 2008 northern California fires outperformed the standard product in matching observed aerosol optical thickness at three coastal Aerosol Robotic Network sites and routinely explained more than 50% of the variance in hourly surface PM2.5 concentrations observed during the wildfires.