Machine learning models accurately predict ozone exposure during wildfire events.

Epidemiologists use prediction models to downscale (i.e., interpolate) air pollution exposure where monitoring data is insufficient. This study compares machine learning prediction models for ground-level ozone during wildfires, evaluating the predictive accuracy of ten algorithms on the daily 8-hour maximum average ozone during a 2008 wildfire event in northern California. Models were evaluated using a leave-one-location-out cross-validation (LOLO CV) procedure to account for the spatial and temporal dependence of the data and produce more realistic estimates of prediction error. LOLO CV avoids both the well-known overly optimistic bias of k-fold cross-validation on dependent data and the conservative bias of evaluating prediction error over a coarser spatial resolution via leave-k-locations-out CV. Gradient boosting was the most accurate of the ten machine learning algorithms with the lowest LOLO CV estimated root mean square error (0.228) and the highest LOLO CV Rˆ2 (0.677). Random forest was the second best performing algorithm with an LOLO CV Rˆ2 of 0.661. The LOLO CV estimates of predictive accuracy were less optimistic than 10-fold CV estimates for all ten models. The difference in estimated accuracy between the 10-fold CV and LOLO CV was greater for more flexible models like gradient boosting and random forest. The order of estimated model accuracy depended on the choice of evaluation metric, indicating that 10-fold CV and LOLO CV may select different models or sets of covariates as optimal, which calls into question the reliability of 10-fold CV for model (or variable) selection. These prediction models are designed for interpolating ozone exposure, and are not suited to inferring the effect of wildfires on ozone or extrapolating to predict ozone in other spatial or temporal domains. This is demonstrated by the inability of the best performing models to accurately predict ozone during 2007 southern California wildfires.

Associations between respiratory health and ozone and fine particulate matter during a wildfire event.

Wildfires have been increasing in frequency in the western United States (US) with the 2017 and 2018 fire seasons experiencing some of the worst wildfires in terms of suppression costs and air pollution that the western US has seen. Although growing evidence suggests respiratory exacerbations from elevated fine particulate matter (PM2.5) during wildfires, significantly less is known about the impacts on human health of ozone (O3) that may also be increased due to wildfires. Using machine learning, we created daily surface concentration maps for PM2.5 and O3 during an intense wildfire in California in 2008. We then linked these daily exposures to counts of respiratory hospitalizations and emergency department visits at the ZIP code level. We calculated relative risks of respiratory health outcomes using Poisson generalized estimating equations models for each exposure in separate and mutually-adjusted models, additionally adjusted for pertinent covariates. During the active fire periods, PM2.5 was significantly associated with exacerbations of asthma and chronic obstructive pulmonary disease (COPD) and these effects remained after controlling for O3. Effect estimates of O3 during the fire period were non-significant for respiratory hospitalizations but were significant for ED visits for asthma (RR = 1.05 and 95% CI = (1.022, 1.078) for a 10 ppb increase in O3). In mutually-adjusted models, the significant findings for PM2.5 remained whereas the associations with O3 were confounded. Adjusted for O3, the RR for asthma ED visits associated with a 10 µg/m3 increase in PM2.5 was 1.112 and 95% CI = (1.087, 1.138). The significant findings for PM2.5 but not for O3 in mutually-adjusted models is likely due to the fact that PM2.5 levels during these fires exceeded the 24-hour National Ambient Air Quality Standard (NAAQS) of 35 µg/m3 for 4976 ZIP-code days and reached levels up to 6.073 times the NAAQS, whereas our estimated O3 levels during the fire period only occasionally exceeded the NAAQS of 70 ppb with low exceedance levels. Future studies should continue to investigate the combined role of O3 and PM2.5 during wildfires to get a more comprehensive assessment of the cumulative burden on health from wildfire smoke.

On the sources and sinks of atmospheric VOCs: an integrated analysis of recent aircraft campaigns over North America.

We apply a high-resolution chemical transport model (GEOS-Chem CTM) with updated treatment of volatile organic compounds (VOCs) and a comprehensive suite of airborne datasets over North America to (i) characterize the VOC budget and (ii) test the ability of current models to capture the distribution and reactivity of atmospheric VOCs over this region. Biogenic emissions dominate the North American VOC budget in the model, accounting for 70 % and 95 % of annually emitted VOC carbon and reactivity, respectively. Based on current inventories anthropogenic emissions have declined to the point where biogenic emissions are the dominant summertime source of VOC reactivity even in most major North American cities. Methane oxidation is a 2x larger source of nonmethane VOCs (via production of formaldehyde and methyl hydroperoxide) over North America in the model than are anthropogenic emissions. However, anthropogenic VOCs account for over half of the ambient VOC loading over the majority of the region owing to their longer aggregate lifetime. Fires can be a significant VOC source episodically but are small on average. In the planetary boundary layer (PBL), the model exhibits skill in capturing observed variability in total VOC abundance (R 2 = 0:36) and reactivity (R 2 = 0:54). The same is not true in the free troposphere (FT), where skill is low and there is a persistent low model bias (~ 60 %), with most (27 of 34) model VOCs underestimated by more than a factor of 2. A comparison of PBL: FT concentration ratios over the southeastern US points to a misrepresentation of PBL ventilation as a contributor to these model FT biases. We also find that a relatively small number of VOCs (acetone, methanol, ethane, acetaldehyde, formaldehyde, isoprene C oxidation products, methyl hydroperoxide) drive a large fraction of total ambient VOC reactivity and associated model biases; research to improve understanding of their budgets is thus warranted. A source tracer analysis suggests a current overestimate of biogenic sources for hydroxyacetone, methyl ethyl ketone and glyoxal, an underestimate of biogenic formic acid sources, and an underestimate of peroxyacetic acid production across biogenic and anthropogenic precursors. Future work to improve model representations of vertical transport and to address the VOC biases discussed are needed to advance predictions of ozone and SOA formation.

A Bayesian model for quantifying the change in mortality associated with future ozone exposures under climate change.

Climate change is expected to have many impacts on the environment, including changes in ozone concentrations at the surface level. A key public health concern is the potential increase in ozone-related summertime mortality if surface ozone concentrations rise in response to climate change. Although ozone formation depends partly on summertime weather, which exhibits considerable inter-annual variability, previous health impact studies have not incorporated the variability of ozone into their prediction models. A major source of uncertainty in the health impacts is the variability of the modeled ozone concentrations. We propose a Bayesian model and Monte Carlo estimation method for quantifying health effects of future ozone. An advantage of this approach is that we include the uncertainty in both the health effect association and the modeled ozone concentrations. Using our proposed approach, we quantify the expected change in ozone-related summertime mortality in the contiguous United States between 2000 and 2050 under a changing climate. The mortality estimates show regional patterns in the expected degree of impact. We also illustrate the results when using a common technique in previous work that averages ozone to reduce the size of the data, and contrast these findings with our own. Our analysis yields more realistic inferences, providing clearer interpretation for decision making regarding the impacts of climate change.

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.