Year-round monitoring reveals prevalence of fatal bird-window collisions at the Virginia Tech Corporate Research Center

Study area

The VTCRC is located in Blacksburg, Virginia, United States (Fig. 1). At the time of the study, the office park consisted of 28 buildings on approximately 230 acres. The first building was constructed in 1988 and future plans include the construction of an additional 16 buildings (VTCRC, 2016). Buildings were assigned identification numbers that correspond with a pre-existing VTCRC building map (VTCRC, 2016). The majority of the buildings are two-story and only two buildings in the office park (buildings 12 and 17) have three stories. There is one glass walkway connecting buildings 3 and 14. Two man-made ponds are located in the VTCRC, the largest of which is located adjacent to buildings 20 and 21. Scattered ornamental trees are present throughout the office park, with larger forested patches on the western boundary and northeast corner of the property. Highly reflective and mirror-like windows comprise the majority of windows at the VTCRC (Fig. 2). Twenty-two buildings were monitored when schedules and weather allowed. The remaining six buildings were not surveyed due to ownership, or because the type of work performed in the buildings prohibited surveys.

Collision surveys

We conducted year-round surveys opportunistically from October 19, 2013 until May 27, 2015. Methods were adapted from Hager & Cosentino (2014a) and Hager & Cosentino (2014b). A group of volunteers, comprised mostly of undergraduate students from Virginia Tech, carried out the surveys documenting all evidence of bird-window collisions. No carcasses were removed or collected; only photos were taken to document evidence of bird-window collisions (e.g., carcasses, feather piles). Each surveyor made one pass, walking slowly, around the perimeter of each building. The search area around each building was the width of the surveyor’s arms (approximately 2 m) held out horizontally from the building wall. Common substrates surveyed included lawn, pavement, and mulch. Surveyors also searched on top, behind, and inside of the shrubs next to the buildings. In addition to species information, date, start survey time, end survey time, building number, and building side were recorded. Surveys were not conducted during a specific time of day or in adverse weather conditions. Data was reported by volunteers every day or at minimum, once a week. We compared all findings, by photo or in the field, to previous findings to avoid double counting.

Data measurement and analysis

Land cover was classified within a 50 m buffer for each of the 22 buildings surveyed. Virginia Geographic Information Network (VGIN) data was used to classify surrounding land cover to the following classes: forest, lawn (including pasture and turfgrass), impervious, water, and tree. To better classify scattered ornamental trees present at the VTCRC, the ‘tree’ land cover class was used to distinguish between sparse tree coverage and forest. We estimated glass area for each building using the program ImageJ (Rasband, 2016) following methods detailed by Hager & Cosentino (2014b). Glass area was centered at 0 and scaled when we used it our Bayesian regression model (Supplemental Information). Meteorological seasons were used to report seasonality of our results. Fall was defined as September–November, winter as December–February, spring as March–May, and summer as June–August.

Linear regression analyses were performed in the program R Core Team (2017), with the RStudio Team (2015), to explore possible relationships between the number of bird deaths per building and predictor variables, glass area and surrounding land cover. We relied on negative binomial regression for this analysis because Poisson distribution led to overdispersion in the data. In order to compare the number of bird deaths across buildings despite uneven survey effort, Poisson regression was performed with an offset for the number of observations. These results informed the Bayesian hierarchical regression model we used to estimate the relative risk of fatal collisions at the 22 buildings across the 20 months. We assumed a zero-inflated Poisson (ZIP) distribution for the number of fatal collisions observed at each building within each month. A ZIP model is common for handling excess zero counts in data and assumes these zeroes are generated by a process separate than the count values (Lambert, 1992). The model contains two components, a binomial distribution for the excess zeroes and a Poisson distribution for the count values (which may contain zeroes). The Poisson distribution assumes a mean of the product of the relative risk θ_it and the expected count e_it at building i in month t. The expected counts are the product of the overall fatal collision rate r_t in each month and the number of site observations at each building in each month.

We modeled the log of the relative risk θ_it for building i at month t as: (1)${\displaystyle log\left({\theta }_{it}\right)={\beta }_1\times {\mathrm{PercentTree}}_i+{\beta }_2\times {\mathrm{PercentLawn}}_i+{\beta }_3\times {\mathrm{GlassArea}}_i+{\delta }_t}$ where δ_t is a temporal random effect. The parameter β₁ is the log relative risk associated with a percent increase in the tree area within 50 m of a building and, similarly, β₂ is the log relative risk associated with a percent increase in lawn area while β₃ is the effect of a standard deviation increase in glass area on a building (SD = 175.63 m²). We chose these variables based on known relationships and significant relationships found in the exploratory analyses. The temporal term δ_t captured variation in the collisions across time not explained by the variables.

For the binomial distribution, we specified the probability of excess zeroes at building i as: (2)${\displaystyle \text{logit}\left({\pi }_i\right)={\alpha }_1\times {\mathrm{PercentTree}}_i+{\alpha }_2\times {\mathrm{PercentLawn}}_i+{\alpha }_3\times {\mathrm{GlassArea}}_i}$

where α₁ is the increase in the log-odds of an excess zero for each percent increase in tree area, α₂ is the change in the log-odds for each percentage increase in lawn area, and α₃ is the increase in log-odds associated for each 175.63 m² of glass area on a building.

Prior distributions were required to complete the Bayesian hierarchical model. All parameters for both the Poisson and binomial model were given non-informative normal priors centered at 0 with variance 10 and the temporal term δ_t followed a random walk prior of order one. This prior allows each δ_t term to be correlated with the previous δ_t−1 term with variance 1,000 and assumes that fatal collisions are more associated in consecutive months than collisions in non-consecutive months.

We estimated the model parameters using Markov chain Monte Carlo (MCMC) in WinBUGS (Spiegelhalter et al., 2002) (Supplemental Information). We ran one MCMC chain for 30,000 iterations with a burn-in of 15,000. The remaining sample was thinned to every 3 iterations, yielding a final posterior sample of size 5,000 for computing posterior summaries. We assessed convergence through visual inspection of trace and density plots for each parameter and Geweke’s diagnostic (Geweke, 1992). For parameter estimates, we sampled from the joint posterior distribution and report the mean posterior estimates and the 95% credible intervals on the relative risk scale. We also plotted the estimated relative risks for each building month to month.

Glass area and surrounding land cover

Seven buildings were responsible for the majority (57.1%) of window collision deaths (Fig. 4). Buildings 11 and 12 had the highest avian mortality and accounted for a quarter (24.6%) of all documented collision deaths. Although we were unable to obtain window specifications on all buildings to make quantitative comparisons of reflectivity and glass type, observations from our surveys suggest that window reflectivity plays a large role in the frequency of collisions at the VTCRC. The amount of glass on buildings ranges from 18 m² (building 8) to 693 m² (building 9). Based on our model, this feature played a significant role in fatal collisions. Building 8, with the least amount of glass, had only one recorded collision death. The two buildings that killed the most birds, 11 and 12, have an estimated 451 m² and 492 m² of glass, respectively. Buildings 1 and 2 are almost identical buildings located adjacent to each other. Building 1, with 219 m² glass area, killed 15 birds; building 2, with 191 m² glass area, killed nine. A notable difference between the two buildings is the amount of forest area (840 m²) on one side of building 1, while building 2 has no surrounding forest area.

Impervious area (51%) consisting of parking lots, sidewalks, and roads dominate the surrounding 50 m of all buildings in the VTCRC. Lawn (35%), tree (10%), forest (3%), and open water (1%) comprise the remaining area. Our model indicated that lawn area mitigated the relative risk of fatal collisions while tree area was associated with a slight increase in risk of collision. Building 11 has the highest percentage of surrounding forest cover (25.2%), while building 12 has no surrounding forest area. Building 11 also has the least amount of lawn area compared to the other buildings.

Seasonality

July, May, and September had the highest mortality respectively when correcting for uneven survey effort (Fig. 5). Collision deaths were lowest in the winter months. The species with the most collision deaths in the fall were Ruby-throated Hummingbird (n = 15), Golden-crowned Kinglet (Regulus satrapa) (n = 7), Mourning Dove (Zenaida macroura) (n = 5), and White-throated Sparrow (n = 5). In the winter, White-throated Sparrow (n = 2), Mourning Dove (n = 2), Golden-crowned Kinglet (n = 2), and Dark-eyed Junco (Junco hyemalis) (n = 2) were the most frequent colliders. American Robin (Turdus migratorius) (n = 11), Gray Catbird (Dumetella carolinensis) (n = 10), and Ruby-throated Hummingbird (n = 7) dominated in the spring. In the summer, American Robin (n = 8), Mourning Dove (n = 3), Gray Catbird (n = 3), Ovenbird (n = 3), and Song Sparrow (Melospiza melodia) (n = 3) were most frequently detected.

Species

The Ruby-throated Hummingbird was the most common window collision species and accounted for 10% of all deaths recorded during our entire study. The majority (29.2%) of hummingbird deaths were at building 11. The next three most common species, American Robin, Gray Catbird, and Mourning Dove, accounted for 21% of identifiable deaths combined. Nine of the 55 species we documented are disproportionately vulnerable to collisions at all building types (Loss et al., 2014). Six Virginia Species of Greatest Conservation Need (Virginia Department of Game and Inland Fisheries, 2015) were documented as window kills, including American Woodcock (Scolopax minor), Eastern Towhee (Pipilo erythrophthalmus), and Field Sparrow (Spizella pusilla) (Table 1).

The most common families were Emberizidae (13% of collision deaths), Turdidae (12%), Trochilidae (10%), Mimidae (8%) and Parulidae (8%). Emberizidae collisions were mostly Dark-eyed Junco (n = 9) and White-throated Sparrow (n = 9). American Robin (n = 21) represented the majority of Turdidae collisions. Ruby-throated Hummingbird was the only species within Trochilidae. Within Mimidae, Gray Catbird (n = 15) was the most common species. Thirteen species comprised Parulidae, with Common Yellowthroat (Geothlypis trichas) and Ovenbird (n = 3) comprising most of the collisions.