Wildfire smoke affects detection of birds in Washington State

Abstract

Wildfire smoke is likely to have direct health effects on birds as well as influence movement, vocalization, and other avian behaviors. These behavioral changes may affect if and how birds are observed in the wild, although research on the effects of wildfire smoke on bird behavior is limited. To evaluate how wildfire smoke affects detection of birds, we combined data from eBird, an online community science program, with data from an extensive network of air quality monitors in the state of Washington over a 4-year period. We assessed how PM_2.5, a marker of smoke pollution, affected the probability of observing 71 bird species during the wildfire seasons of 2015–2018 using bird observations from 62,908 eBird checklists. After accounting for habitat, weather conditions, seasonality, and survey effort, we found that PM_2.5 affected the probability of observing 37% of study species. The ambient concentration of PM_2.5 was negatively associated with the probability of observing 16 species and positively associated with the probability of observing 10 species, indicating that birds exhibit species-specific behavioral changes during wildfire smoke events that influence how they are observed. Our results suggest that wildfire smoke impacts the presence, availability, and/or perceptibility of birds. Impacts of smoke pollution on human observers, such as impaired visibility, may also influence detection of birds. These results provide a foundation for developing mechanistic hypotheses to explain how birds, and our studies of them, are impacted by wildfire smoke. Given the projected increase in large-scale wildfire smoke events under future climate change scenarios, understanding how birds are affected by wildfire smoke—and how air pollution may influence our ability to detect them—are important next steps to inform wildlife research and avian conservation.

Resumen

Es probable que el humo de los incendios silvestres tenga efectos directos sobre la salud de las aves, además de influir en el movimiento, la vocalización y otros comportamientos de ellas. Estos cambios de comportamiento pueden afectar la forma en que las aves son observadas en la naturaleza, aunque las investigaciones sobre los efectos del humo de los incendios silvestres en el comportamiento de las aves son limitadas. Para evaluar cómo el humo de los incendios silvestres afecta la detección de las aves, combinamos datos de eBird, un programa de ciencia comunitaria en línea, con datos de una extensa red de monitoreo de calidad del aire en el Estado de Washington durante un período de 4 años. Evaluamos cómo PM_2.5, un marcador de contaminación por humo, afectó la probabilidad de observar 71 especies de aves durante las temporadas de incendios silvestres de 2015–2018, utilizando observaciones de aves de 62,908 listas de verificación de eBird. Después de tener en cuenta el hábitat, las condiciones climáticas, la estacionalidad y el esfuerzo de muestreo, encontramos que PM_2.5 afectó la probabilidad de observar el 37% de las especies estudiadas. La concentración ambiental de PM_2.5 se asoció negativamente con la probabilidad de observar 16 especies y positivamente con la probabilidad de observar 10 especies, lo que indica que las aves exhiben cambios de comportamiento específicos para cada especie durante los eventos de humo de incendios silvestres, que influyen en la forma en que son observadas. Nuestros resultados sugieren que el humo de los incendios silvestres afecta la presencia, la disponibilidad y/o la perceptibilidad de las aves. Los impactos de la contaminación por humo en los observadores humanos, como la visibilidad reducida, también pueden influir en la detección de aves. Estos resultados proporcionan una base para desarrollar hipótesis mecanicistas que expliquen cómo las aves, y nuestros estudios sobre ellas, se ven afectados por el humo de los incendios silvestres. Dado el aumento proyectado de eventos de humo de incendios silvestres a gran escala en escenarios futuros de cambio climático, comprender cómo las aves se ven afectadas por el humo de incendios silvestres y cómo la contaminación del aire puede influir en nuestra capacidad para detectarlas, son los próximos pasos importantes para guiar las investigaciones sobre la vida silvestre y la conservación de las aves.

Lay Summary

• Wildfire smoke events threaten public health and are likely to impact the health and behavior of birds, influencing their perceptibility.

• In this study, we characterized how wildfire smoke affects detection of birds. We assessed how PM_2.5, a marker of wildfire smoke, affected the probability of observing 71 bird species in Washington State.

• We found that outdoor concentrations of PM_2.5 affected the probability of observing 37% of study species. PM_2.5 was negatively related to the detection of 16 species and positively related to the detection of 10 species. Our results suggest that wildfire smoke impacts the presence, availability, and/or perceptibility of birds.

• Given the projected increase in the frequency of large-scale wildfire smoke events, understanding how birds are affected by smoke—and how air pollution may influence our ability to detect birds—are critical next steps in avian conservation.

INTRODUCTION

Large-scale wildland fires pose direct threats to wildlife, including mortality, injury, and health effects from exposure to high ambient temperatures and the toxic gases and aerosols found in smoke (Engstrom 2010), such as fine particulate matter (suspended solid or liquid particles with an aerodynamic diameter <2.5 µm [PM_2.5]) (O’Dell et al. 2019). Although exposure to PM_2.5 from wildfire smoke has known, adverse impacts on human health (Cascio 2018, Wettstein et al. 2018), few studies have been conducted on how wildfire smoke affects the health and behavior of free-living, nonhuman animals (Cheyne 2008, Lee et al. 2017, Sanderfoot and Holloway 2017, Erb et al. 2018). Such responses may affect if and how wildlife can be observed, yet the impacts of wildfire smoke on the detection of wildlife species have not previously been explored.

Birds may exhibit behavioral changes during wildfire smoke events. Lee et al. (2017) showed that acoustic indices used to characterize bird activity declined for months at a location after it was inundated with smoke from wildfires, suggesting that birds altered their singing behavior, vacated the area, or died. Behavioral changes exhibited by birds and other animals in response to wildfire smoke may signal immediate, underlying health effects (Black et al. 2017). However, other changes in the physical environment during smoke pollution episodes, such as cooler air temperatures (Robock 1991) and reduced visibility (Haider et al. 2019), could also trigger behavioral responses in wildlife (Cheyne 2008).

Changes in avian behavior during wildfire smoke events could affect the probability of observing birds. For example, if wildfire smoke prompts a species to relocate, we would expect the probability of observing that species to be inversely related to the outdoor concentration of PM_2.5, a marker of smoke pollution. However, the probability of observing a species reflects both whether or not a species is present and if so whether or not it is observed (i.e. detected). Even if the presence of birds is unaffected by wildfire smoke, birds may change their behavior during smoke events in such a way that they are more or less likely to be observed. For example, if birds sing less or reduce their overall activity as smoke pollution from wildfires increases, they would be more difficult to observe due to a decline in the visual and auditory cues used to detect them. Alternatively, if wildfire smoke prompts birds to fly faster or congregate in larger groups, they would likely be easier to detect. Changes in the probability of observing a species during wildfire smoke events may therefore reflect behavioral responses to wildfire smoke that drive variation in presence, availability, and perceptibility, as well as any direct mortality.

Impacts of smoke pollution on human observers may also influence the probability of observing birds during wildfire smoke events. Detection of avifauna may be negatively impacted if wildfire smoke impairs the observational skills of birders, either due to acute health impacts or reduced visibility; this would lead to a decrease in the probability of observing avian species that is independent of bird behavior.

To explore how the probability of observing birds changes during wildfire smoke events, we combined data from eBird, a semi-structured community science program run by the Cornell Lab of Ornithology (https://ebird.org/home), with long-term monitoring data of outdoor concentrations of PM_2.5. eBird is a global community science program that collects bird observations submitted by volunteers in a checklist format (Sullivan et al. 2009). Due to the program’s popularity, the eBird database contains more than 600 million records, submitted by hundreds of thousands of participants from around the world. Community science programs like eBird facilitate the collection of ecological data at greater spatial and temporal scales than what is possible within a traditional science framework. However, eBird data are collected opportunistically by volunteers who differ in experience, training, skill, and objectivity, leading to spatial, temporal, taxonomic, and observer biases (Johnston et al. 2019). In our analysis, we used several variables to account for heterogeneity in sampling events and among observers, including time and duration of surveys and distance traveled by observers.

In this study, we estimated the effect of PM_2.5 on the probability of observing 71 bird species during the wildfire season in Washington, USA. Due to the potential effects of smoke pollution on avian behavior and perceptibility, we hypothesized that the concentration of PM_2.5 would be an important predictor of the probability of observing birds. More specifically, we expected that species with larger home ranges may be better able to emigrate from an area impacted by wildfire smoke pollution than birds with smaller home ranges, and therefore that the probability of observing them would be inversely related to the concentration of PM_2.5. We also expected that the probability of observing avian species primarily detected by ear would decrease as smoke pollution increased, due to previous research linking declines in bird vocalization to wildfire smoke exposure (Lee et al. 2017). Finally, we expected that detection of birds often observed at far distances, such as diurnal raptors and gulls, would decrease during smoke events due to reductions in visibility. Our study provides a foundation for developing mechanistic hypotheses to explain interspecific behavioral responses to wildfire smoke and highlights the need to consider how air pollution affects our observations of birds and other wildlife. Given the projected increase in large-scale wildfire smoke events under future climate change scenarios (Westerling et al. 2006, Jacob and Winner 2009), these efforts are critical to inform ornithological research, wildlife management, and avian conservation.

METHODS

Study Area

Washington, USA, is a large state with many habitat types, including wet and dry forests, shrubland, marshes, and urban and suburban areas (Washington Department of Fish and Wildlife 2015). Hundreds of wildfires occur in the state each year, burning tens of thousands of acres—more than 400,000 acres were burned by wildfires in 2018 alone (Washington State Academy of Sciences 2019). Smoke events driven by large-scale wildfires in 2015, 2017, and 2018 subjected both people and wildlife in Washington State to hazardous air quality, as indicated by high concentrations of PM_2.5 (Figure 1). Levels of PM_2.5 during these smoke pollution episodes were at times much higher than the National Ambient Air Quality Standard (NAAQS) of 35 µg m^–3 (U.S. Environmental Protection Agency 2016; Figure 1). Due to extensive wildfires, air pollution during the 2018 wildfire season was higher in some parts of Washington State than anywhere else in the world (Washington State Academy of Sciences 2019).

Figure 1.

Wildfire smoke events are marked by abrupt peaks in fine particulate matter (PM_2.5), suspended solid and liquid particles smaller than 2.5 µm in diameter that pose health risks to both people and wildlife. Plots show the daily mean concentration of PM_2.5 at air quality monitors in the state of Washington during the wildfire season (July, August, and September) in 2015 through 2018. Each point represents a measurement from one air quality monitor. Large-scale smoke events occurred in 2015, 2017, and 2018; note the sudden increases in ambient PM_2.5. The y-axis ranges from 0 to 500 µg m^–3. The National Ambient Air Quality Standard for PM_2.5 is only 35 µg m^–3. Plots illustrate that exposure to particle pollution during the wildfire season often exceeds levels deemed safe for public health.

Data Collection

Bird observations.

We analyzed bird observations collected by the public and submitted to eBird (eBird Basic Dataset 2021). eBird supports opportunistic data collection by volunteers who submit online checklists of species they observe and other related information (e.g., date, time, duration and type of survey, distance traveled, number of observers, etc.) (Sullivan et al. 2009). We extracted all bird observations documented in Washington State from July 1 to September 30, 2015–2018 from the eBird Basic Dataset (2021). We used the auk package (Strimas-Mackey et al. 2018) in R (R Core Team 2020) to filter bird observations to include only those from stationary, traveling, and area counts. Stationary surveys take place within 30 m of a point location, traveling surveys are conducted along a route of known distance, and area surveys are carried out in an area of known size. We chose to include all 3 types of surveys in our analysis because wildfire smoke events are of short duration and we wanted to ensure that we included bird observations on days when air quality was poor. Stationary checklists were assigned a distance traveled of 0 m. To estimate a distance traveled for area checklists, we assumed the area surveyed was a square and the distance traveled was 2 times the length of one side of the square. We also restricted our dataset to only include bird observations from complete checklists—those for which observers indicated that they recorded all of the species that they detected during a survey. By excluding incomplete checklists, we were able to infer when species were not detected, as opposed to simply not recorded by an observer, thereby strengthening our inference (Johnston et al. 2019).

Next, we used the geosphere package (Hijmans 2019) in R (R Core Team 2020) to spatially filter the data to include only bird observations from checklists with locations within 32 km of an active PM_2.5 monitor in Washington State. Data from ground-based air quality sensors are often used to inform public health; however, measurements from air quality sensors are only representative of exposure to air pollution in close proximity. We therefore restricted our dataset to observations near air quality monitoring sites to ensure that the measurements of PM_2.5 at those locations would serve as reasonable estimates of exposure to this pollutant where birds were observed. We specifically chose a distance of 32 km to align with the spatial resolution of the weather data used in our analysis. This distance is also within the range of threshold distances used when characterizing exposure to air pollution in public health research. If the distance an eBirder traveled exceeded 32 km, they may have been birding outside a 32-km radius from the nearest air quality monitor; to negate this possibility, we excluded any checklists with a distance traveled >32 km.

We focused on commonly observed species in our analysis because we wanted to compare responses to wildfire smoke across a large set of species, but we needed to ensure that there were a sufficient number of detections of those species to fit our models. We used the filtered data to select our study species. We included all species with at least 750 detections in 2015, the baseline year. Seventy-one species met this criterion (see Table 1 for a complete list of study species). We used the auk package to generate detection/non-detection data for each study species (Strimas-Mackey et al. 2018). We then linked each of the checklists to the daily concentration of PM_2.5 at the monitoring station closest to the checklist location. We eliminated checklists for which we could not assign a daily concentration of PM_2.5 from the nearest ground-based air quality monitor. Data on PM_2.5 were not available from every monitor on every date of our study period because not all monitors measure particle pollution continuously.

Table 1.

List of the 71 study species included in this analysis. Species are listed by common name, in order of how often they were detected in the checklists included in our final dataset. Complete model results for all study species are reported in Supplementary Material Appendix B.

Common name	Scientific name
American Robin	Turdus migratorius
Black-capped Chickadee	Poecile atricapillus
American Crow	Corvus brachyrhynchos
Song Sparrow	Melospiza melodia
Northern Flicker	Colaptes auratus
Great Blue Heron	Ardea herodias
Barn Swallow	Hirundo rustica
Mallard	Anas platyrhynchos
American Goldfinch	Spinus tristis
Spotted Towhee	Pipilo maculatus
European Starling	Sturnus vulgaris
House Finch	Haemorhous mexicanus
Cedar Waxwing	Bombycilla cedrorum
Steller’s Jay	Cyanocitta stelleri
Anna’s Hummingbird	Calypte anna
Canada Goose	Branta canadensis
Dark-eyed Junco	Junco hyemalis
White-crowned Sparrow	Zonotrichia leucophrys
Glaucous-winged Gull	Larus glaucescens
Belted Kingfisher	Megaceryle alcyon
Red-breasted Nuthatch	Sitta canadensis
Bewick’s Wren	Thryomanes bewickii
Killdeer	Charadrius vociferus
Chestnut-backed Chickadee	Poecile rufescens
California Gull	Larus californicus
Osprey	Pandion haliaetus
Rock Pigeon	Columba livia
Red-winged Blackbird	Agelaius phoeniceus
Double-crested Cormorant	Phalacrocorax auritus
Caspian Tern	Hydroprogne caspia
Eurasian Collared-Dove	Streptopelia decaocto
Violet-green Swallow	Tachycineta thalassina
Common Raven	Corvus corax
House Sparrow	Passer domesticus
Savannah Sparrow	Passerculus sandwichensis
Bald Eagle	Haliaeetus leucocephalus
Common Yellowthroat	Geothlypis trichas
Mourning Dove	Zenaida macroura
Red-tailed Hawk	Buteo jamaicensis
Downy Woodpecker	Dryobates pubescens
Ring-billed Gull	Larus delawarensis
Swainson’s Thrush	Catharus ustulatus
Pied-billed Grebe	Podilymbus podiceps
Least Sandpiper	Calidris minutilla
Bushtit	Psaltriparus minimus
Brown Creeper	Certhia americana
Western Sandpiper	Calidris mauri
Pigeon Guillemot	Cepphus columba
Golden-crowned Kinglet	Regulus satrapa
Yellow-rumped Warbler	Setophaga coronata
Western Wood-Pewee	Contopus sordidulus
Heermann’s Gull	Larus heermanni
Turkey Vulture	Cathartes aura
Black-headed Grosbeak	Pheucticus melanocephalus
Yellow Warbler	Setophaga petechia
Brown-headed Cowbird	Molothrus ater
Orange-crowned Warbler	Leiothlypis celata
Western Tanager	Piranga ludoviciana
Wood Duck	Aix sponsa
Marsh Wren	Cistothorus palustris
Greater Yellowlegs	Tringa melanoleuca
Tree Swallow	Tachycineta bicolor
Gadwall	Mareca strepera
Pelagic Cormorant	Phalacrocorax pelagicus
California Quail	Callipepla californica
Green-winged Teal	Anas crecca
Brewer’s Blackbird	Euphagus cyanocephalus
Spotted Sandpiper	Actitis macularius
Rhinoceros Auklet	Cerorhinca monocerata
American Coot	Fulica americana
Northern Harrier	Circus hudsonius

Environmental data.

To characterize the ambient concentration of PM_2.5 at checklist locations, we used data from the Environmental Protection Agency (EPA) Air Quality System (AQS) (U.S. Environmental Protection Agency 2019; https://www.epa.gov/outdoor-air-quality-data). We downloaded daily PM_2.5 concentrations (24-hr averages) available from all ground-based PM_2.5 monitors in the state of Washington for the years 2015 through 2018. (For more information on how we processed air quality data, see Supplementary Material Appendix A.) Only one checklist was linked to a PM_2.5 concentration >300 µg m^–3. The PM_2.5 measurements at this location the day before and after the observation date of this checklist were much lower; we therefore removed this value as an outlier. In our final dataset, the daily concentration of PM_2.5 ranged from 0 to 295.8 µg m^–3, with a mean value of 8.6 µg m^–3.

To account for how weather may influence the probability of observing birds, we used data from the North American Regional Reanalysis (NARR) to determine daily mean air temperature and daily accumulated precipitation for each eBird checklist. NARR data were provided by the National Oceanic and Atmospheric Administration Physical Sciences Laboratory in Boulder, Colorado, USA from their website at https://psl.noaa.gov/data/gridded/data.narr.html. The spatial resolution of NARR data is ~32 km. We used the ncdf4 package (Pierce 2019) in R (R Core Team 2020) to extract the weather data.

We also incorporated land cover as a predictor because we expected that the 71 study species would have different habitat requirements and preferences. We assigned a land cover type to each checklist using data from the 2016 National Land Cover Database (NLCD; https://www.mrlc.gov/data), developed by the Multi-Resolution Land Characteristics (MRLC) Consortium and made available as a layer in ArcMap by the Environmental Systems Research Institute (ESRI 2019, 2020). Land cover classifications were grouped into 9 categories: open water, perennial ice/snow, developed, barren, forest, shrubland, herbaceous, planted/cultivated, and wetlands. Prior to analysis, we removed checklists for which a land cover type could not be determined because the checklist locations fell outside the bounds of the NLCD. We also eliminated checklists that were assigned a land cover class of “perennial ice/snow” because very few checklists fell into this category.

Our final dataset included detection/non-detection data for 71 species from 62,908 eBird checklists, each of which represented a single survey. These checklists were submitted by 4,865 unique eBird observers and linked to air quality data from a total of 71 PM_2.5 sensors. A map showing the locations of eBird checklists and air quality monitoring sites included in our analysis is shown in Figure 2. This map was created using the rnaturalearth (South 2017) and sf (Pebesma 2018) packages in R (R Core Team 2020).

Figure 2.

Map of Washington showing the locations of 62,908 eBird checklists included in our analysis as gray dots. The locations of the 71 air quality monitoring sites included in our analysis are marked as black triangles.

Statistical Analysis

We used generalized linear mixed models (GLMM) with a binomial distribution to model the probability of observing each of the 71 study species during the 2015–2018 wildfire seasons. Each model included fixed effects of 8 numeric variables (day of year, day of year squared, time observations started, duration of survey, distance traveled, daily mean air temperature, daily accumulated precipitation, and daily concentration of PM_2.5), 2 categorical variables (year and land cover class), and a random effect of individual observer. For each of the 71 study species, we modeled $p_i$⁠, the probability that a species is observed in checklist i, as:

$$\begin{array}{l}\text{logit}(p_i)={\alpha }_0+{\alpha }_1\times {\text{year}}_{\text{i}}+{\alpha }_2\times {\text{land cover class}}_i+{\alpha }_3\times {\text{day}}_i+{\alpha }_4\times {\text{day}}_i^2+\\ {\alpha }_5\times {\text{time}}_i+{\alpha }_6\times {\text{duration}}_i+{\alpha }_7\times {\text{distance}}_i+{\alpha }_8\times {\text{temp}}_i+{\alpha }_9\times {\text{precip}}_i+\\ {\alpha }_{10}\times {\text{PM}}_{{2.5}_i}\end{array}$$

We allowed the intercept α ₀ to vary by unique observer as a random effect to account for differences in skill, knowledge, and experience among eBird observers as well as a lack of independence in their checklists. In addition to the environmental variables (land cover, temperature, precipitation, and PM_2.5), we included day of year and day of year squared to account for seasonal variation in species presence, which may take a linear or quadratic shape. We included time of day because it is well known that birds exhibit daily activity patterns, which directly influences detection of many species (Robbins 1981a, Johnson 2008). We also included the duration of a survey in minutes because we expected that the more time eBirders spent birding, the more likely they would be to observe birds. We included distance traveled to capture additional variation in surveys. To account for annual variation in population demographics, we included a year effect term. All numeric variables were standardized. Prior to analysis, we compared all continuous variables for potential collinearity (see Supplementary Material Appendix C for the full correlation matrix).

Each model was fit using the lme4 package (Bates et al. 2015) in R (R Core Team 2020). We calculated conditional R² values to assess model fit (Lüdecke et al. 2019). We used an information-theoretic approach (Akaike’s information criterion [AIC]) to first evaluate the overall importance of PM_2.5 by comparing the full model to the same model without PM_2.5. We anticipated that AIC might select the model with PM_2.5 even when the effect was not statistically significant due in part to the large size of the dataset. We therefore used P-values and 95% confidence intervals (CI) on the effect of PM_2.5 in the full model to determine if this predictor was statistically significant (P < 0.05) for any of the 71 study species.

We were concerned that the number of checklists submitted may decline during smoke events. To test this, we used a GLMM to assess how the number of checklists varied with the level of risk posed by particle pollution. We binned PM_2.5 into 3 categories based on current public health standards: good to moderate (<35.5 µg m^–3), unhealthy for sensitive groups (≥35.5 and <55.5 µg m^–3), and unhealthy to hazardous (≥55.5 µg m^–3). We then ran a negative binomial GLMM with the total number of checklists as the response variable, PM_2.5 as a fixed effect, and air quality monitor as a random effect. We included air quality monitor as a random effect because some monitors are located in more densely populated areas.

RESULTS

Based on our model selection using AIC, the top model included the effect of ambient concentration of PM_2.5 for 37 of the 71 study species (52%). Of those 37 species, PM_2.5 had a statistically significant effect on the probability of observing 26 of the 71 study species (36.6%) (Figure 3). The probability of observing birds decreased with elevated concentrations of PM_2.5 for 16 of these 26 species, including Great Blue Heron (Ardea herodias), Canada Goose (Branta canadensis), Killdeer (Charadrius vociferus), California Gull (Larus californicus), Osprey (Pandion haliaetus), Double-crested Cormorant (Phalacrocorax auritus), Caspian Tern (Hydroprogne caspia), Savannah Sparrow (Passerculus sandwichensis), Bald Eagle (Haliaeetus leucocephalus), Red-tailed Hawk (Buteo jamaicensis), Bushtit (Psaltriparus minimus), Heermann’s Gull (Larus heermanni), Turkey Vulture (Cathartes aura), Marsh Wren (Cistothorus palustris), Greater Yellowlegs (Tringa melanoleuca), and Northern Harrier (Circus hudsonius) (Figure 3). The concentration of PM_2.5 was positively related to the probability of observing 10 species: Spotted Towhee (Pipilo maculatus), Cedar Waxwing (Bombycilla cedrorum), Red-breasted Nuthatch (Sitta canadensis), Yellow-rumped Warbler (Setophaga coronata), Black-headed Grosbeak (Pheucticus melanocephalus), Yellow Warbler (Setophaga petechia), Brown-headed Cowbird (Molothrus ater), Orange-crowned Warbler (Leiothlypis celata), Western Tanager (Piranga ludoviciana), and California Quail (Callipepla californica) (Figure 3). No consistent pattern was observed in the direction of other temporal and environmental predictors included in our models for these 26 species. We did not find evidence that air quality influenced the total number of eBird checklists submitted, which suggests that the frequency of checklist submissions by eBirders was not affected by outdoor concentrations of PM_2.5 during our study period in Washington State.

Figure 3.

Fine particulate matter (PM_2.5) influenced the detection of 26 of the 71 study species included in our analysis. Here we show the predicted probability of observing a species at daily concentrations of PM_2.5 ranging from 0 to 300 µg m^–3 in the year 2015 in developed areas, assuming average levels of all other predictors. Species with a negative association are shown in panel (A) and those with a positive association are shown in panel (B). Each color represents predictions for 1 species. Solid lines indicate median predictions and ribbons illustrate bootstrapped 95% CIs. Color palette provided by Pedersen and Crameri (2021).

The effect of year, land cover class, day of year, day of year squared, time observations started, daily mean air temperature, daily accumulated precipitation, duration of survey, and distance traveled varied by species (Figure 4). As expected, duration of survey was consistently a positive predictor of the probability of observing birds—the longer eBirders were in the field, the more likely they were to observe each species in the study (Figure 4). This effect was statistically significant (P < 0.05) for all species included in our analysis except for Brewer’s Blackbird (Euphagus cyanocephalus). Detection of most species was positively related to distance traveled; however, distance was a statistically significant, negative predictor (P < 0.05) of the probability of observing 9 of the 71 study species. These species included common backyard birds (e.g., American Goldfinch [Spinus tristis], House Finch [Haemorhous mexicanus], Anna’s Hummingbird [Calypte anna]), as well as shorebirds (e.g., Western Sandpiper [Calidris mauri], Least Sandpiper [Calidris minutilla], and Greater Yellowlegs). In addition, the time observations started was a statistically significant, negative predictor (P < 0.05) of the probability of observing 58 of the 71 study species, indicating that the probability of observing most species decreased throughout the day. However, the probability of observing Western Sandpipers, Least Sandpipers, Turkey Vultures, Western Yellowlegs, and Pelagic Cormorants (Phalacrocorax pelagicus) increased throughout the day.

Figure 4.

Although small, the effect of PM_2.5 on the probability of observing birds sometimes exceeded the effect of weather and other temporal and environmental predictors known to influence detection of birds. These violin plots show the probability density of the effect sizes (coefficient estimates on the logit scale) for the 8 numeric temporal and environmental predictors included in the logistic regression analysis for all 71 study species. The box plots within each violin plot denote the median and interquartile range. Predictors include day of year (Day), day of year squared (Day²), time observations started (Time), duration of survey (Duration), distance of survey (Distance), daily mean air temperature (Temp), daily accumulated precipitation (Precip), and daily concentration of fine particulate matter (PM_2.5). Color palette provided by Pedersen and Crameri (2021).

Land cover was also an important predictor for all 71 species. Day of year and day of year squared were significantly (P < 0.05) associated with the probability of observing birds for 64 and 59 species, respectively. This indicates that seasonality is an important source of variation in the probability of observing most species included in this analysis. The direction of this effect varies both between and within families. Temperature had a significant effect (P < 0.05) on the probability of observing 58 species. Temperature was negatively related to detection of 27 species, across a wide range of taxa, including small aerial insectivores (e.g., Barn Swallow [Hirundo rustica], Tree Swallow [Tachycineta bicolor]), and diurnal birds of prey (e.g., Bald Eagle, Turkey Vulture). Temperature was positively related to detection of 31 species, also ranging widely in taxa, including waterbirds (e.g., Great Blue Heron, Canada Goose, and Mallard [Anas platyrhynchos]) and passerines (e.g., Black-capped Chickadee [Poecile atricapillus] and House Sparrow [Passer domesticus]). Precipitation had a significant effect (P < 0.05) on the probability of observing 36 species, and the effect was negative for 15 species and positive for 21. This environmental predictor was also important in determining the detection of a diverse assortment of species. For example, detection of several passerines was negatively associated with precipitation, including American Robin (Turdus migratorius), American Goldfinch, and Dark-eyed Junco (Junco hyemalis), yet precipitation was also inversely related to detection of several waterbirds, including Pigeon Guillemot (Cepphus columba) and Rhinoceros Auklet (Cerorhinca monocerata). The list of species for which precipitation was positively related to detection also included songbirds (e.g., Savannah Sparrow and Common Yellowthroat [Geothlypis trichas]) and birds associated with aquatic habitats, including Wood Duck (Aix sponsa) and Gadwall (Mareca strepera). Complete model results are provided in Supplementary Material Appendix B.

DISCUSSION

Our study shows that ambient concentrations of PM_2.5, a marker of smoke pollution, during the 2015–2018 wildfire seasons affected the probability of observing 26 of the 71 most commonly sighted bird species in Washington, USA. These 26 species included waterbirds, raptors, and passerines, suggesting that PM_2.5 is an important driver of presence, availability, and/or perceptibility across avian taxa.

Wildfires create and maintain important habitat for birds. For example, some bird species, including woodpeckers and flycatchers, will use early post-fire habitat, whereas others, such as the Black-backed Woodpecker (Picoides arcticus), are almost entirely dependent on recently burned forests (Hutto 1995, Saab et al. 2004). While wildfires play a crucial role in generating high-quality habitat for some wildlife species, large-scale wildland fires also pose direct threats to animals, including mortality, injury, and health effects from exposure to extreme heat and smoke (Engstrom 2010). Such health effects may affect bird behaviors, including movement or vocalization. Birds may also adjust their behavior during smoke events in response to cooler air temperatures or reduced visibility. Any behavioral changes exhibited by birds during smoke pollution episodes may affect if and how birds are observed in the wild. Furthermore, impacts of smoke pollution on human observers, such as impaired visibility, may also influence detection of birds. Our results demonstrate a clear impact of PM_2.5 during the wildfire season on the detectability of birds. Researchers should recognize the potential for smoke and other types of air pollution to affect our observations of birds and other wildlife, which has the potential to change the inferences that might be made in observational studies.

Logically, the effect of air pollution on the probability of observing birds may be especially important during wildfire smoke events, when concentrations of PM_2.5 are often above air quality standards (Figure 3). The average daily concentration of PM_2.5 for the checklists included in our analysis was 8.6 µg m^–3 (SD = 14.5 µg m^–3). However, during smoke events in Washington State between 2015 and 2018, concentrations of PM_2.5 spiked to well above 150.5 µg m^–3, the breakpoint between “unhealthy” and “very unhealthy” air quality (Figure 1). This is a common occurrence during large-scale wildfires (Laing and Jaffe 2019). Only 83 checklists included in our analysis were submitted on days when the PM_2.5 concentration exceeded 150.5 µg m^–3, highlighting the difficulty of capturing these uncommon events. To ensure that limited availability of bird observations at the most extreme levels of particle pollution did not affect our inference about the species-specific relationship between PM_2.5 and detection of birds, we ran our models on a subset of our dataset that excluded checklists with PM_2.5 concentrations above 150.5 µg m^–3. We found that the relationship between PM_2.5 and detection of birds remains unchanged for all but 1 of the 26 species for which we report a statistically significant effect (California Quail). Furthermore, we found that PM_2.5 became a statistically significant predictor of detection for 6 additional species, with positive relationships for Steller’s Jay, Western Sandpiper, Golden-crowned Kinglet, and Pigeon Guillemot, and negative relationships for Ring-billed Gull and Spotted Sandpiper. None of the species with positive relationships were observed in more than 5 of the 83 checklists with PM_2.5 concentrations above 150.5 µg m^–3. This post hoc analysis highlights the need to collect data on wildlife during poor air quality conditions to better understand species-specific responses, which could be nonlinear and may be related to other weather and habitat variables. In fact, PM_2.5 concentrations vary by land cover, which is another reason why it was important to consider habitat type when assessing the potential effects of PM_2.5 on detection of birds. It is inherently challenging to collect data on birds and other wildlife during smoke pollution episodes because they are difficult to forecast and often of short duration. We recommend focusing future studies on bird communities within particular habitat types during the wildfire season to better understand how species-specific habitat selection influences avian responses to smoke, especially in landscapes more likely to experience smoke events.

We found that the probability of observing birds was negatively correlated with particle pollution during the wildfire season for 22.5% of the species in our study; however, the probability of observing birds was positively related to PM_2.5 for some species (14.1%). This suggests that birds exhibit species-specific behavioral changes during air pollution episodes that ultimately influence whether or not they are present and, if so, detected. It is important to note that we could not differentiate whether a bird species was present but not detected or truly absent from a survey location in this analysis. That said, we hypothesized 2 possible mechanisms that could explain declines in bird observations, specifically (1) smoke compromised the observational skills of birders (e.g., impaired visibility) or (2) smoke triggered behavioral changes in birds that made them less likely to be observed (e.g., reduced vocalization). We do not expect that poor air quality or limited visibility would enhance a birder’s observational skills. Therefore, if the probability of observing birds was driven only by changes in observer ability, we would not expect to find a positive effect of PM_2.5 for any species. Given that the probability of observing birds increased with elevated concentrations of PM_2.5 for 10 species in our study, our results suggest that at least some species must exhibit behavioral changes during smoke events that make them easier to detect or move into areas that are smoky. We found that species more likely to be observed at higher concentrations of PM_2.5 included birds often observed in the upper branches of trees, including Cedar Waxwing, Yellow-rumped Warbler, Yellow Warbler, Orange-crowned Warbler, and Western Tanager. If reduced visibility forces these species to forage closer to the ground, they may be more readily observed. Future studies should investigate whether birds use habitat differently during smoke events. Diminished visibility may also prompt birders to focus their attention on species closer to the ground, increasing the probability of observing species often sighted at ground level, such as Spotted Towhee and California Quail.

However, we cannot rule out impaired visibility as a driver of the negative correlation between PM_2.5 and the probability of observing birds for some study species (Figure 3). We expected that smoke impacts on visibility would be most important for birds that are observed at far distances, such as diurnal raptors and gulls. Our results support this notion, as PM_2.5 was negatively correlated with the probability of observing birds of prey (e.g., Bald Eagle, Red-tailed Hawk, Osprey, Turkey Vulture, Northern Harrier) and gulls (e.g., California Gull, Glaucous-winged Gull [Larus glaucescens], Heermann’s Gull, and Ring-billed Gull [L. delawarensis]), although this effect did not always meet our threshold for statistical significance (i.e. P < 0.05).

Species-specific behavioral responses to wildfire smoke may also explain the inverse relationship between PM_2.5 and detection of some study species (Figure 3). While the presence of observers can affect bird behavior, recent research suggests that this alone does not significantly influence the singing rates of birds (Hutto and Hutto 2020). Previous research has shown that wildlife acoustic activity may decline during large-scale smoke events following wildfires. Cheyne (2008) reported that gibbons (Hylobates albibarbis) do not sing as often when it is smoky and Lee et al. (2017) documented declines in bird vocalization during a prolonged wildfire smoke event. If birds call and sing less during smoke pollution episodes, they may be more difficult to detect in the field. We expected that the probability of observing avian species primarily detected by ear, such as chickadees, kinglets, and other woodland species (Robbins 1981a), would decline with increasing ambient concentrations of PM_2.5. Our results suggest that detection of Bushtits decreases at higher concentrations of PM_2.5 (Figure 3); however, PM_2.5 was not a statistically significant, negative predictor of the probability of observing other woodland species included in this analysis. In addition to reduced vocalization, animals may reduce their activity levels when exposed to toxic particle pollution. Exposure to aerosols has been linked to declines in spontaneous activity (e.g., walking or preening) and reduced water and food intake in Rock Doves (Columba livia) (Sterner 1993a, 1993b), and a recent study showed that orangutans (Pongo pygmaeus wurmbii) are less active during smoke events, resting more and traveling less (Erb et al. 2018). Less activity could result in lower detection rates. However, racing pigeons exhibited faster homing rates at higher outdoor concentrations of particulate matter (Li et al. 2016). If birds fly faster or spend more time in their territories, they may be more readily detected in and around their activity centers.

While our methods did not allow us to explicitly test if particle pollution during the wildfire season drives changes in the presence or availability of birds, it is not unreasonable to expect that some avian species may seek refuge from wildfire smoke if they are experiencing severe adverse health impacts. We expected that species with larger home ranges may be better able to emigrate from an area impacted by wildfire smoke pollution than birds with smaller home ranges. Our results support this possibility, as many of the species (e.g., Great Blue Heron, Double-crested Cormorant) that were detected at lower rates with increasing concentrations of PM_2.5 do have larger home ranges compared to other species analyzed. In addition, smoke degrades visibility and therefore may negatively impact the hunting success of birds of prey, prompting them to relocate. Our results show that the probability of observing several birds of prey, including Bald Eagle, Red-tailed Hawk, Osprey, Turkey Vulture, and Northern Harrier, declined with greater concentrations of PM_2.5. In the absence of predatorial birds, prey species may be more active and therefore more detectable (MacLeod et al. 2005), which may also explain why the species more readily observed as PM_2.5 increases includes several smaller songbirds. Future studies should investigate how home range size, territoriality, vigilance, predator-prey dynamics, and daily activity patterns influence behavioral responses to smoke pollution from wildfires.

It is well established that the detection and presence of birds are variables that are influenced by temporal and environmental predictors. Migration and breeding phenology drive seasonal patterns in species presence and bird activity (e.g., vocalization), which in turn influences detection (Skirvin 1981, Johnson 2008). To capture how seasonality drives changes in the probability of observing birds, ecologists often include day of year as a linear and quadratic covariate (e.g., Broms et al. 2014, 2016, Purves et al. 2019). We found that day of year and day of year squared were important predictors of the probability of observing 64 and 59 species, respectively, suggesting that seasonality influences detection of a wide range of species. Time of day and weather are also known to have an effect on bird activity (Johnson 2008). Many bird species are more active in the early morning hours (Robbins 1981a), which is consistent with our results that show 58 of the 71 study species were more likely to be detected earlier in the day. Some avian species may be less active when exposed to extreme temperatures or during precipitation events (Robbins 1981b), potentially reducing the probability of observing them. We found that daily average temperature had a significant effect on the probability of observing 58 species whereas precipitation had a significant effect on the probability of observing 36 species, indicating that temperature was overall a more important predictor of detection than precipitation across species included in the analysis.

Our results suggest that for some species, air pollution may be a more important source of variation in the probability of observing birds than standard temporal and environmental predictors such as seasonal variation, time of day, or weather conditions. For example, we found that increased concentrations of PM_2.5 had a negative effect on the probability of observing Great Blue Herons, and the magnitude of this effect exceeded that of precipitation. Detection of Spotted Towhees was positively related to PM_2.5, while the effects of both temperature and precipitation on the probability of observing this species were not statistically significant.

Our analysis was limited to observations of birds closer to urban areas because we relied on data from the EPA network of ground-based air quality monitors to characterize particle pollution. These monitors are primarily located in cities and towns (Diao et al. 2019; Figure 2). Atmospheric models and measurements from satellite instruments could be leveraged to expand the spatial scope of air pollution estimates in future ecological studies, although both approaches have their own limitations that should be carefully considered (Diao et al. 2019). In addition, eBird supports opportunistic data collection, which may result in spatial, temporal, and taxonomic biases (Boakes et al. 2016). Still, our results suggest that, at least in locations near ground-based air quality monitoring sites in Washington where eBirders were active in July, August, and September of 2015–2018, detection of 36.6% of the most commonly observed bird species was impacted by outdoor concentrations of PM_2.5, a marker of smoke pollution. Furthermore, by including year as a categorical predictor in our models we may have underestimated the true effect of PM_2.5 on the probability of observing birds. We included year to account for annual variation in population demographics that would influence the probability of observing birds; however, year is also a significant predictor of the daily concentration of PM_2.5 due to annual variation in the frequency and intensity of smoke events (Figure 1).

Our study suggests that wildfire smoke may impact the presence, availability, and/or perceptibility of birds, which has major consequences for both ornithological research and avian conservation. Many field studies take place during the summer and fall, months that increasingly overlap with an extended wildfire season and large-scale smoke events. Our model selection results suggested that the top model included PM_2.5 for over 50% of the species; thus, by not including air quality as a predictor in ecological models, we may be missing an important source of variation in the detection of birds. Failing to model heterogeneity in detection would bias our inference about bird activity and population demographics (Kéry et al. 2010). Occupancy models would be useful in teasing out the effect of wildfire smoke on the perceptibility of birds while accounting for the influence of habitat in determining where species are present. eBird data can be used to build occupancy models, although doing so requires careful consideration of the definition of a site and the covariates used to account for heterogeneity in surveys (Johnston et al. 2019). Targeted sampling before, during, and after wildfire smoke events may be most useful in assessing how smoke affects avian behavior and movement or the detectability of birds. However, air pollution episodes are often unpredictable, and it may be difficult for researchers to capture these events even in the most well-designed studies. Researchers may consider implementing before-after-control-impact studies around prescribed burns as an alternative strategy to characterizing how smoke affects bird behavior, although it is worth noting that the smoke from prescribed burns is less toxic than smoke from wildfires (Prunicki et al. 2019). Avian behavioral responses to particle pollution could signal underlying health effects, serve as effective strategies to limit exposure to toxic gases and aerosols, and/or stem from changes in the physical environment (e.g., visibility) or other stressors (e.g., predators) during smoke events. Regardless, behavioral responses to wildfire smoke may ultimately impact fitness, survival, and reproductive success. Although the long-term health consequences of smoke exposure for wildlife could be substantial (Black et al. 2017), research is limited (Lee et al. 2017, Erb et al. 2018). It is therefore critical to rapidly expand existing knowledge on how animals are affected by wildfire smoke and identify which species may be most at risk.

CONCLUSIONS

Our study shows that the ambient concentration of PM_2.5, a marker of smoke pollution, was an important source of variation in the probability of observing birds in Washington State during the wildfire seasons of 2015–2018. Incidence of large-scale wildfire smoke events is expected to increase under future climate change scenarios (Jacob and Winner 2009); understanding avian behavioral responses to wildfire smoke and how perceptibility of birds changes during smoke events is critical to inform wildlife research and avian conservation. Finally, our study demonstrates that community science programs could be an important source of data in future studies looking at the impacts of smoke events or other air pollution episodes on birds. We found that eBirders not only survey birds across vast expanses, but they also continued to conduct surveys during large-scale smoke events. Community science data could therefore be a valuable resource for ecologists looking to characterize broad-scale impacts of major pollution events on wildlife populations.

ACKNOWLEDGMENTS

We would like to thank the 4,865 volunteers whose eBird checklists were used in this analysis—this study would not have been possible without their efforts. We would also like to thank Staci Amburgey, Briana Abrahms, Hannah Sipe, Nick Etzel, Sarah Bassing, Robbie Emmet, Trent Roussin, and Becky Reese for reviewing code and drafts of the manuscript, as well as the anonymous reviewers whose feedback greatly improved our analysis. Insightful conversations about our research goals with members of the Quantitative Ecology and Quantitative Conservation labs at the University of Washington were immensely helpful in developing this manuscript.

Funding statement: This work was supported by the National Science Foundation Graduate Research Fellowship Program [grant no. DGE-1762114] and the McIntire-Stennis Cooperative Forestry Research Program [grant no. NI19MSCFRXXXG035/project accession no. 1020586] from the USDA National Institute of Food and Agriculture. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation or USDA.

Author contributions: O.V.S.: conceptualization, methodology, formal analysis, validation, investigation, data curation, writing (original draft), visualization, project administration, and funding acquisition; B.G.: conceptualization, methodology, formal analysis, validation, writing (review and editing), visualization, supervision, and funding acquisition.

Ethics statement: No live birds or other animals were handled during this study.

Conflict of interest statement: We have no conflicts of interest to declare.

Data depository: Analyses reported in this article can be reproduced using the data provided by Sanderfoot and Gardner (2021).