Wildfire activity is driving summertime air quality degradation across the western US: a model-based attribution to smoke source regions

The spatial information, temporal information, and usage of datasets included in this study are presented in table 1.

2.2.1. STILT

STILT is a receptor-oriented Lagrangian particle dispersion model developed for time-reversed simulations that produce back trajectories for an ensemble of fictitious air parcels arriving at the receptor (site of observation in urban centers; see SI section S1). Production of back trajectories is performed through combination of time reversed mean wind fields and a stochastic turbulence parameterization represented as a Markov chain process [34]. From air parcel backward trajectories, STILT generates a gridded surface flux footprint that quantifies the influence of upwind grid cells on the receptor at a particular time (e.g. figure 1) [34]. When convolved with wildfire PM_2.5 emissions estimates, the STILT footprint generates a 4-dimensional, quantitative description of the PM_2.5 contributions from upwind source fires to the urban center receptor (e.g. figure 1). Further details regarding STILT are provided in SI section S2.

Zoom In Zoom Out Reset image size

In this work, STILT ingests mean wind fields from WRF simulations (see SI section S2). At 3 h intervals throughout August and September of 2003–2020, we simulate 1200 back trajectories (see sensitivity tests in SI section S3) for each of the 33 western US urban centers considered. Trajectories were run 120 h back in time, as this should capture the majority of transport over the western US and Canada [32] before exiting over the Pacific Ocean. Wildfire PM_2.5 emission estimates used for convolution with STILT output were obtained from the Quick Fire Emissions Dataset (QFED) [43] and were downscaled to 3 h emissions profiles on the basis of monthly aggregated satellite active fire detections sourced from the Global Fire Emissions Database (GFED; see SI section S2) [44, 45]. Modeling was limited to the months of August and September due to the computational/storage expense of producing regional-decadal scale WRF simulations. The years 2003–2020 were selected for consistency in the underlying input datasets. A limited number of STILT simulations, 120 out of 289,872, have been excluded from trend analyses as a result of model performance. A brief explanation of these excluded simulations may be found in SI section S2.

2.2.2. STILT smoke module

Trend analyses of the observed PM_2.5 at urban centers possessing adequate observations and wildfire derived PM_2.5 (modeled) arriving at each of the 33 urban centers were performed using 98th quantile regression. Quantile regression is effectively a linear fit to the data that incorporates an asymmetric weighting for the least-squares minimization [40, 41], such that the fit may preferentially weight extreme data points (as is the case at the 98th quantile).

Daily averaged PM_2.5 observations at EPA observation sites (corresponding to model receptor sites) were aggregated to monthly (August and September) averaged values for the years 2003–2020. Sites containing >10 daily averaged PM_2.5 observations per month for >15 years over the span 2003–2020 were considered for trend analyses. Similarly, model outputs were aggregated to monthly average values for each of the 33 urban centers considered. Observed and modeled time series for the months of August and September were then subjected to quantile regression. Statistical significance was determined via a bootstrapping procedure that resampled each time series 100,000 times, with replacement, and performed 98th quantile regression to see if the actual regression was distinguishable from the scatter of the resampled regressions at p < 0.05.

Identification of particularly relevant smoke source regions was similarly performed using 98th quantile regression, but in this case time series were constructed to reflect the PM_2.5 mass contributed to a particular urban center from fires in a specific ecoregion/ecoprovince. Ecoregions/ecoprovinces, sourced from the US EPA and Government of Canada [38, 39], define areas of similar ecosystem composition (climate, vegetation, hydrology, etc.), such that they may adhere to a similar fire regime. In the context of this analysis, the use of ecoregions/ecoprovinces is intended to allow spatial aggregation for the benefit of statistical relationships while maintaining the possibility of ecosystem scale factors influencing wildfire activity and subsequent air quality. Construction of time series specific to a particular ecoregion/ecoprovince was performed by spatially subsetting STILT-simulated PM_2.5 contributions to individual ecoregions/ecoprovinces during convolution of wildfire emissions and STILT model output. Subsetting was performed specifically for US level 3 ecoregions and Canadian ecoprovinces (figure 2(a)) because they provided a reasonable balance between spatial resolution, illuminating informative transport pathways and smoke source regions, and valuable statistical relationships. At finer spatial resolutions (e.g. US level 4 ecoregions) many of the ecoregions are spatially discontinuous, and thus often lack additional information about atmospheric transport while simultaneously yielding less robust statistics given increased inter annual variability in wildfire activity.

Zoom In Zoom Out Reset image size

Evaluation of model outputs against EPA observed PM_2.5 was performed on the basis of individual sites (urban center receptors) and sub-regions of the western US (Pacific Northwest, Northern/Central California, Intermountain West, Southwest). With the intent of testing the capacity of this modeling framework to capture observed PM_2.5 time series, we used the statistical analysis software R [51] to calculate the correlation between modeled and observed PM_2.5. It was required that sites have greater than 5 years of monthly averaged data, requiring >10 daily observations per month, and that the most recent year be at least as recent as 2017. A cutoff of 2017 was selected to ensure that presented correlations reflect periods relevant to our trend analyses given that many western US ecoregions present a possible acceleration in PM_2.5 emissions trends since 2017/2018 [19]. Of the 33 urban centers considered, 17 met these criteria, with average, minimum, and maximum duration of monthly average data being 14.7, 5, and 18 years, respectively. Given the lack of non-wildfire PM_2.5 sources in the model, it should be recognized that even if the model perfectly captured fluctuations in wildfire derived PM_2.5, the presented correlations to the observations would be less than unity.

We find that urban centers across the western US present positive and statistically significant (p < 0.05) 98th quantile trends in wildfire-emitted PM_2.5 for the months of August and September over the years 2003–2020 (figure 2). In accordance with the empirical, observation-driven findings of Wilmot et al [12], wildfires exert a significant impact on air quality at a larger number of urban centers in August than September, and the Pacific Northwest and Northern/Central California demonstrate the most robust trends in wildfire PM_2.5 degraded air quality. At many urban centers throughout the Pacific Northwest and Northern/Central California, modeled trends exceed 1 μg m⁻³ yr⁻¹. For context, the median observed summertime PM_2.5 over the years 2003–2020 ranges from ∼4 μg m⁻³ (Portland, OR) to ∼12 μg m⁻³ (Bakersfield, CA) for the 33 urban centers considered. In the most extreme cases, modeled trends exceed 2 μg m⁻³ yr⁻¹, with such results preferentially occurring during September. These extraordinary September trends are largely tied to 2020 values that exceed 45 μg m⁻³, with the available observations suggesting model overestimation by a factor of ∼2 at Fresno, California, and model underestimation by a factor of ∼2 at Salem, Oregon. While further investigation is needed, the shift toward wind-driven fire activity from summer to fall offers a possible explanation for the large September trends modeled at sites in Oregon and California [22, 52]. Trends toward increasing temperatures and atmospheric aridity enhance the probability of coincident dry fuels and strong winds, allowing for extreme rates of fire spread that may have severe impacts on air quality [22, 52]. Attribution to wildfires at the ecoregion scale indicates that September 2020 PM_2.5 values for Fresno (California), Fort Collins (Colorado), Salem and Eugene (Oregon) are driven by nearby wildfires in the Sierra Nevada, Southern Rockies, and Cascades ecoregions, respectively. Further, for Eugene and Salem, 2020 PM_2.5 values can be attributed directly to the Labor Day fires (figure 1(a)), which burned an estimated 11% of the Oregon Cascades following a downslope wind event that drove rapid fire spread [53].

Trend analyses of time series spanning 2003–2019 indicate that the extreme fire activity of summer 2020 substantially influenced the 98th quantile PM_2.5 trends reported at sites along the Colorado Front Range during August, within the Pacific Northwest during September, and within portions of California during both August and September. For urban centers along the Colorado Front Range, August trends spanning 2003–2019 are statistically significant, but are, on average, 49% smaller than the corresponding 2003–2020 trends. At urban centers within the Pacific Northwest (excluding Spokane, WA), September trends spanning 2003–2019 are only significant at p < 0.1, and they are, on average, 53% smaller than the corresponding trends for 2003–2020. Within California, trends spanning 2003–2019 at Bakersfield, Sacramento, and Santa Rosa during August and Fresno, Sacramento, Lake Elsinore, and San Jose during September, are 41% (p = 0.27), 27% (p < 0.05), 12% (p < 0.05), 69% (p < 0.05), 29% (p < 0.1), 55% (p = 0.25), and 87% (p = 0.43) smaller than the corresponding 2003–2020 trends, respectively. These findings highlight the sensitivity of identified trends to the years considered, while also drawing a possible connection to the hypothesized acceleration of trends in western US wildfire emissions and plume top heights over recent years [19].

Extrapolation of these trend analyses into the future supports the notion that many western US urban centers may soon face severe summertime air quality degradation. Though wildfire events are deemed 'exceptional events' and are thus not considered with regards to NAAQS compliance, the 35 μg m⁻³ limitation imposed by the EPA on 98th quantile PM_2.5 provides concerning context for the trends modeled here. While extrapolation of short-term trends into the future has significant uncertainties, evidence for climate driven trends toward enhanced wildfire activity in relevant source regions [15, 16, 22] suggests that such an extrapolation may be reasonable. Further, modeling based on population, climate, and vegetation trends suggests an expectation that late 21st century western North American burned area and wildfire emissions will exceed a 1990–2010 baseline by 50%–120% and 110%–250%, respectively, under a moderate emissions mitigation scenario (Shared Socioeconomic Pathway 2–4.5) [54]. In terms of PM_2.5 pollution, model estimates suggest that the moderate emissions mitigation scenario will translate to August-September 95th quantile PM_2.5 values ranging from 85 to 125 μg m⁻³ at sites throughout the Pacific Northwest [54].

Comparison of modeled and observed 98th quantile trends demonstrates that modeling reasonably reproduces the observational record, particularly in August, and allows for varying levels of attribution of observed western US air quality degradation to wildfire sources (table 2). Eugene, Oregon demonstrates particularly strong alignment between modeled and observed trends during the month of August, with values of 0.549 μg m⁻³ yr⁻¹ and 0.547 μg m⁻³ yr⁻¹, respectively. The generally strong agreement between observed and modeled trends points towards the utility of this modeling framework and supports interpretation of modeled trends for urban centers. Beyond model errors and uncertainty, it is possible that disagreement between observed and modeled trends stems from uncertainty in wildfire emissions estimates [55] and/or a decreasing trend in anthropogenically emitted PM_2.5 superimposed upon the increasing trend in wildfire derived PM_2.5 [11, 14].

Trend analyses of the wildfire derived PM_2.5 contributions from specific ecoregions/ecoprovinces to western US urban centers indicates regionally varying smoke source regions (figure 2). Overall, the Pacific Northwest is impacted by smoke from numerous source regions throughout the western US and Canada, while urban centers in California and along the Colorado Front Range primarily see influences from wildfires in California and the Southern Rockies, respectively. Urban centers along the Wasatch Front of Utah, in Southern California, and in western Texas are characterized by contributions of varying magnitude from the Wasatch-Uinta mountains, Southern California/Northern Baja Coast, and Southern Rockies ecoregions, respectively. These findings largely follow the relatively coarse atmospheric transport pathways identified by Brey et al [32], though we identify transport pathways at a refined spatial resolution and find a proportionally greater contribution of smoke from wildfires in the Rocky Mountains (as defined by Brey et al [32]) to air quality in Colorado and along the Wasatch Front. This increased relevance of the Rocky Mountains to air quality along the Colorado Front Range may be due, in part, to intense 2020 wildfire activity in the nearby Southern Rockies ecoregion. Time series of PM_2.5 arriving at urban centers along the Colorado Front Range as contributed from all wildfires and from wildfires within the Southern Rockies present a greater correlation when 2020 data is included as compared to analyses limited to 2003-2019 (SI figure S1). Dominance of the Southern Rockies in terms of wildfire-emitted PM_2.5 contributions to the Colorado Front Range further agrees with prior work relating the time averaged 700 hPa flow to correlations between aridity/fire area burned in the Southern Rockies and organic aerosol loading at sites to the east of the Southern Rockies [30].

During August, urban centers in the Pacific Northwest are uniquely subject to smoke from Canadian wildfires, and experience further contributions from the mountainous Pacific Northwest and California, a result that aligns with the observational findings and hypotheses of Wilmot et al [12]. The range of smoke source regions contributing to aerosol loading at urban centers in the Pacific Northwest reflects the unique challenge to air quality that results from the Pacific Northwest's proximity to wildfires in Canada, Washington, Idaho, Oregon, and California (e.g. figure 1(b)). In extreme cases, urban centers in the Pacific Northwest may experience unhealthy levels of PM_2.5 pollution regardless of wind direction as a result of being surrounded by wildfires that are regionally relevant smoke sources (figure 1(b)). The presence of general north-south gradients in PM_2.5 contributions from wildfires in British Columbia and California to urban centers in the Pacific Northwest indicates the utility of urban center—smoke source proximity as a first estimate of the relevance of that smoke source to local air quality. Similar gradients are seen for California urban centers in terms of August PM_2.5 contributions from the Klamath Mountains/California High North Coast Range and the Central California Foothills and Coastal Mountains ecoregions. Given the known connections between the Klamath Mountains region and western US wildfire activity/air quality [12, 13, 22], it is of note that this is the sole ecoregion that presents a statistically significant (p < 0.05) impact on August PM_2.5 loading at each of the 33 western US urban centers considered.

While regional contribution patterns are largely consistent from August to September, we find a distinct shift of the dominant contributing ecoregions. Most notably, September features the emergence of the Cascades, Sierra Nevada, and Wasatch-Uinta Mountains ecoregions as the dominant PM_2.5 contributors to urban centers in the Pacific Northwest, California, and along the Wasatch Front, respectively. Relative to August, September presents comparatively large contribution trends from individual ecoregions. As indicated previously, these extraordinary September trends are largely the result of unprecedented wildfire activity in September of 2020, and we hypothesize a possible connection to wind-driven fire activity during the fall for sites in Oregon and California. The diminished importance of emissions from Canadian wildfires during September is in agreement with earlier findings [12].

In terms of monthly aggregated PM_2.5 values, the modeling framework evaluates well against EPA observed PM_2.5. For sites with an adequate number of observations for evaluation, results indicate strong correlation coefficients between modeled and observed PM_2.5 (figure 3). At the level of individual sites, maximum and minimum values in the correlation coefficient of 0.99 and −0.04 are found at Salem, Oregon in September and Palm Springs, California in August, respectively. Given that this framework neglects PM_2.5 from non-wildfire sources, we would expect a well-constructed model to present stronger correlations to observations in regions and months with greater wildfire contributions to PM_2.5. When comparisons are averaged by region, we find that the model presents stronger observation-model correlation coefficients in the Pacific Northwest and California, while demonstrating weaker correlations in the Southwest (table 3). Correlations between modeled and observed PM_2.5 were stronger in August relative to September. These results are in line with the expectation of enhanced correlations where wildfires play a greater role in PM_2.5.

Zoom In Zoom Out Reset image size

In looking at figure 3, it is clear that the model framework consistently underestimates observed PM_2.5 values. However, as this model framework neglects PM_2.5 from non-fire sources, a consistent low-bias is to be expected. For a well-constructed model, it would be expected that this low-bias reflects background, non-smoke impacted, PM_2.5 values within the urban centers considered. Comparison of median observation-model differences to median observed August and September PM_2.5 values at the regional scale largely supports the notion that background, non-smoke, PM_2.5 explains the consistent observation-model discrepancy (table 4). The similar magnitude of median PM_2.5 observations and observation-model differences, as well as their regional variability, suggests that PM_2.5 from non-fire sources comprises much of the observation-model differences.

Further, linear fit trend analyses for the observation-model differences suggest only minor implications for the modeled 98th quantile trend analyses reported here. Negative trends, indicating a decrease in the observation-model discrepancy over time were found at Seattle during August (−0.23 μg m⁻³ yr⁻¹; p < 0.05) and September (−0.24 μg m⁻³ yr⁻¹; p < 0.05), Denver during August (−0.12 μg m⁻³ yr⁻¹; p < 0.05) and September (−0.14 μg m⁻³ yr⁻¹; p < 0.05), Palm Springs during August (−0.32 μg m⁻³ yr⁻¹; p < 0.05), Salt Lake City during September (−0.15 μg m⁻³ yr⁻¹; p < 0.1), and Bakersfield during September (−0.29 μg m⁻³ yr⁻¹; p < 0.05). While these negative trends suggest the possibility of somewhat inflated modeled PM_2.5 trends at these urban centers, in particular at Denver in August, and at Seattle, Salt Lake City, and Denver in September, we note that this observation-model convergence may alternatively be the result of trends toward decreasing anthropogenic emissions of PM_2.5 [14].