Identifying Birds’ Collision Risk with Wind Turbines Using a Multidimensional Utilization Distribution Method

Renewable energy plays a key role in reducing greenhouse gas emissions. However, the expansion of wind farms has raised concerns about risks for bird collisions. We tested different methods used to understand whether birds’ flight occurs over wind turbines and found kernel density estimators outperform other methods. Previous studies using kernel utilization distribution (KUD) have considered only the 2 horizontal dimensions (2D). However, if altitude is ignored, an unrealistic depiction of the situation may result because birds move in 3 dimensions (3D). We quantified the 3D space use of the Griffon vulture (Gyps fulvus) in El Estrecho natural park in Tarifa (southern Spain, on the northern shore of the Strait of Gibraltar) during 2012–2013, and, for the first time, their risk of collision with wind turbines in an area in the south of Spain. The 2D KUD showed a substantial overlap of the birds’ flight paths with the wind turbines in the study area, whereas the 3D kernel estimate did not show such overlap. Our aim was to develop a new approach using 3D kernel estimation to understand the space use of soaring birds; these are killed by collision with wind turbines more often than any other bird types in southern Spain. We determined the probability of bird collision with an obstacle within its range. Other potential application areas include airfields, plane flight paths, and tall buildings. © 2020 The Authors. Wildlife Society Bulletin published by Wiley Periodicals, Inc. on behalf of The Wildlife Society.

Wind farms have received public and government support as a clean source of renewable energy because they do not cause air pollution as does the burning of fossil fuels (Stigka et al. 2014, Yuan et al. 2015. The use of wind energy is therefore expanding rapidly worldwide. The Global Wind Energy Council reported that 2015 was another record-breaking year for the wind energy industry (Global Wind Energy Council 2016). However, wind farms may be causing a large number of fatalities to flying animals, affecting a large area of potential soaringhabitat around them , Zimmerling and Francis 2016, Marques et al. 2020. Therefore, the expansion of wind farms has raised concerns about their potential negative effects on wildlife populations and associated habitat quality.
Relatively high collision fatality rates have been recorded at several large wind farms in locations indicating that turbines pose a risk, especially to large raptors and other soaring birds (Marques et al. 2014). The Griffon vulture (Gyps fulvus) is one of the raptor species frequently killed by collision with wind turbines in southern Spain (Barrios and Rodríguez 2004;De Lucas et al. 2008Olea and Mateo-Tomás 2014). For instance, Carrete et al. (2012) found 342 dead Griffon vultures during a 10-year period (Jan 1998-Mar 2008 in an area of 34 wind farms with 799 turbines in the province of Cadiz, southern Spain. To develop efficient conservation practices, it is necessary to know where and when a threat of collision may occur (Wilcove 2010). The 2 most common approaches, namely home range and utilization distribution, have been used to depict and portray animal movements and their space use (Rutz and Hays 2009, Tomkiewicz et al. 2010, Monsarrat et al. 2013. The home range is the area "traversed by an individual [animal] in its normal activities of food gathering, mating and caring for young" (Burt 1943:351), whereas the utilization distribution reflects the animal's spatial use probability density (Van Winkle 1975, Signer et al. 2017. Recently, the home range has been viewed as one attribute of the animal's utilization distribution. Animal space use or home range has been quantified using different methods such as minimum convex polygon (Mohr 1947), bivariate normal method (Jennrich Figure 1. The study area during 2012-2013 was part of the El Estrecho natural park in Tarifa (southern Spain) on the northern shore of the Strait of Gibraltar. The Griffon vulture colony (red square) was located at an escarpment close to wind turbines (colored circles with numbers, each color represents a group of identical turbines). and Turner 1969), grid square method (Siniff andTester 1965, Macdonald et al. 1980), population utilization distribution (Ford and Krumme 1979), and kernel density (Worton 1995). Several methods have recently been developed for the time-explicit estimation of animal space use, such as the dynamic Brownian bridge movement model and bivariate Gaussian bridges (Kranstauber et al. 2012(Kranstauber et al. , 2014. The kernel density estimator method has low bias, and greater flexibility in handling complex location patterns and in assuming location independence (Worton 1989(Worton , 1995Seaman and Powell 1996;Fieberg 2007;Benhamou and Cornélis 2010).
So far, most studies estimating animal home ranges or utilization distributions have only considered the 2 horizontal dimensions (Katajisto and Moilanen 2006, Powell and Mitchell 2012, Fleming et al. 2015. If altitude-the third dimension-is neglected, an unrealistic depiction of reality may be attained for species moving in 3-dimensional (3D) space, such as birds, bats, fish, or climbing species (Belant et al. 2012, Monterroso et al. 2013. However, few studies have quantified space use patterns in 3D. For instance, Koeppl et al. (1977) presented a model based on an ellipsoid of a particular size, shape, and orientation in space. It was one of the first models used to compute home range in 3D. Hindell et al. (2011) quantified the 3D space use of 5 different species (2 mammals and 3 bird species). They highlighted that the greatest concentrations of locations of southern elephant seals (Mirounga leonina) occurred within the 1,000-m bathymetric contour. Simpfendorfer et al. (2012) calculated the utilization distribution of European eels (Anguilla anguilla) using 2D and 3D kernel density. They emphasized that the 2D analysis overestimated the amount of movement overlap between individuals by 13-20%. Recently, Cooper et al. (2014) studied the 3D space use and overlap of American redstarts (Setophaga ruticilla) using a direct observation method for data collection and kernel density estimator. Their study was confined to observing focal territories throughout each sampling period, with birds located visually and the altitude estimated by observers. The number of locations for each observed bird was also limited. Nevertheless, their findings concurred with a former study on the overestimation by 2D analysis compared with the 3D method. They also found that American redstarts may avoid areas of overlap, presumably to limit interactions with neighbors. We collected locations of an individual Griffon vulture with the use of a bio-logger, quantified the bird's 2D and 3D utilization distributions, and, for the first time, its collision risk with wind turbines using kernel utilization distribution (KUD). We demonstrate that volumetric analysis (3D) is more informative than planar analysis (2D) in utilization distributions. We show that neglecting the third dimension would provide incomplete depiction of the aerial species' space use, whereas 3D kernel estimators cannot only be used to improve our understanding of the bird's movements, but they also can be considered as a way to determine wildlife collision risk with an obstacle in the territory or home ranges in conservation plans.

STUDY AREA
The study area was located in the natural park of El Estrecho, in Tarifa (southern Spain) on the northern shore of the Strait of Gibraltar ( Fig. 1; 36°07′−36°06′N, 5°45′−5°46′W). This area was the most southern protected area in Europe. It was a maritime-terrestrial park along 54 km of coastline in Andalusia and an Important Bird Area (Guerra García et al. 2009, BirdLife International 2017. In this area, Ferrer et al. (2012) reported the greatest collision rates ever published for birds (1.33/turbine/yr) with the Griffon vulture being the most frequently killed species (0.41 deaths/turbine/yr). There were several Griffon vulture colonies in the area, consisting of approximately 320 breeding pairs in total. We focused on a colony at an escarpment running north-south, 4 km from the Strait of Gibraltar; with approximately 65 breeding pairs (Del Moral 2009). Our analysis was constrained to the space used by one tagged Griffon vulture; the space encompassed an area of 152 km 2 and included 20 wind farms with 269 operational turbines (Table 1).

METHODS
We captured a Griffon vulture using a foot-snare trap. We attached a bio-logger as a backpack using a harness made of Teflon (Chemours, Wilmington, DE, USA) ribbon with one strap fitted across each wing and another strap below the crop (Kenward 2000). The capture and release took place on 11 September 2010. We attached distinctive yellow patagial markers, with a unique combination of numbers and letters (i.e., 9FJ) to both wings. This method was shown to be harmless to the bird and led to no detectable changes from its normal behavior (Reading et al. 2014). Our Griffon vulture was a male, subadult, and with a body mass of approximately 7 kg. We used the Bird Tracking System biologger developed at the University of Amsterdam . The key features of this bio-logger are solar rechargeable batteries, light weight (45 g, <0.6% of body mass), 2-way datacommunication, 4-megabyte flash memory (capable of  . We used the GPS fixes and their properties to quantify the Griffon vulture's 3D movement to determine the overlap of air space use between the bird and wind turbines. We retrieved the GPS fixes of our Griffon vulture for 18 months (Feb 2012-Jul 2013. These data comprised 169,778 locations at 5-minute intervals. The procedures of this research, including the bird trapping and bio-logger tagging were conducted with permission from the Consejería de Medio Ambiente of the Junta de Andalucía (Regional Council for the Environment).

Data Analysis
The Griffon vulture is a diurnal species; therefore, we considered only data points collected during daytime and filtered out stationary locations (speed <4 m/sec). We used the remaining 12,611 locations to quantify KUDs (50% and 95%) in 3D (Benhamou and Cornélis 2010). The multivariate kernel density estimate is defined by is an independent and identically distributed sample of a random variable X, h is the bandwidth, and K is the kernel function of dimensions d.
We used a plug-in bandwidth selector to estimate the smoothing factor matrix. This method provides adequate results in the utilization distribution estimation and requires less intensive computation compared with other methods, such as least-squares cross-validation or the reference method. We quantified monthly 2D and 3D KUDs (50% and 95%) for the Griffon vulture using the "ks" package (Duong 2007) in Program R Statistical environment (version 3.2.3, www.r-project.org, accessed 5 Sep 2016). At the location of each turbine, we extracted the value of probability density generated by the 2D kernel function. In a similar fashion, we extracted the density values generated by the 3D kernel, but this time the total height of a turbine (i.e., turbine height plus blade length) was considered. This is the sum of the turbine's height, the length of a blade, and the land elevation determined using a digital elevation model. In addition, we considered the value for the probability density in both 2D and 3D KUD as a proxy of the plausible collision risk. Then we used the Mann-Whitney-Wilcoxon test to examine any significant differences in the extracted values. In 2D and 3D KUD, values above the third quartile (i.e., the greatest 25% of the values) were selected as a proxy for plausible collision with high risk. We calculated the frequency distribution of plausible risk pertaining to the turbines to determine which turbines might be relatively dangerous in the course of data gathering.

RESULTS
The 2D KUD of our tagged Griffon vulture showed that all the wind turbines were located in the core and extended home range, where the KUD values were relatively high (Fig. 2). Values extracted from the 2D KUD were greater than for the 3D model at the turbine locations (Mann-Whitney-Wilcoxon Test W = 72,351, P < 0.001). The 3D kernel estimation of the Griffon vulture's occurrence covered a large area of space use (Fig. 3): 50% and 95% KUD were estimated to be 476 km 3 and 11,120 km 3 , respectively (more situations related to Fig. 3 are available in Supporting Information, available online). Values extracted from the 3D kernel estimation showed a high probability density in the winter and early spring of 2012 and 2013. The results showed that there was no sign of collision risk in May and June in both years because the Griffon vulture was not in the vicinity of the turbines. However, the concentration of collision risk increased in March and April of both years (Fig. 4).
In 3D space, just 3 turbines had a relatively high risk (i.e., above the third quartile) in 12 out of 17 months, whereas 7 turbines appeared to have such a risk in 2D space. The maximum number of turbines that had a relatively high risk in 3D and 2D space were 55 and 62, respectively (Fig. 5). To be specific, the high risk occurred in 2 months (Feb and Apr) for 3D and in 3 months (Feb, Apr, and Jun) for 2D space use. Turbines located in the southern part of the study area, in the vicinity of the Griffon vulture colony, had a relatively high risk of collision in both the 3D and 2D analyses (Fig. 6).

DISCUSSION
We used kernel utilization distribution (KUD), for the first time, to understand the plausible collision risk between wind turbines and bird occurrence. Our results demonstrate the advantage of 3D KUD for modeling 3D space use by birds and, in particular, the comparative risk of a bird colliding with a turbine. Although 2D analyses are useful to summarize information on the location of individuals (Simpfendorfer et al. 2012), volumetric analyses (i.e., with altitude added as a third dimension) provide a more detailed depiction of species' occurrence. Using a 3D KUD, we show that the most dangerous times and greatest-risk turbines can be identified. This information can be used to reduce the mortality rate caused by bird collisions with turbines and offers wildlife managers insights on how to minimize the probability of such collisions (Belant et al. 2012).
To date, the probability of collision has been studied by analyzing a range of complex factors such as the species' flight behavior, topography, and weather (De Lucas et al. 2008). Those studies were conducted with the aim of reducing the avian mortality rate at wind farms, particularly of raptors (Barrios and Rodríguez 2004, DeVault et al. 2005, Drewitt and Langston 2008, Tellería 2009, Bellebaum et al. 2013. However, 3D space use was not considered and we show that including 3D space may influence results. A trial mitigation measure was instigated by Regional Council for the Environment in 2008-2009 that power companies selectively shut down some wind turbines when raptors were observed in their vicinity. This measure reduced the Griffon vulture fatality rate by 50% . The trial also demonstrated that the distribution of Griffon vulture mortality was not uniform, which is consistent with our results from the 3D KUD approach. Additionally, this approach can be considered in turbine selection if mechanisms are implemented to shut down the turbines when birds are in the vicinity. In Europe, an environmental impact assessment (EIA) is required prior to the construction of new wind farms. The anticipated effects of development on the site's bird populations are included in the EIA (Environmental Impact Assessment Directive 97/11/EC). Ferrer et al. (2012) ascertained that risk assessment studies had erroneously assumed a linear relationship between the frequency of observed birds and fatalities. They concluded that the correlation between predicted and actual fatalities can be improved by changing the scale of studies and concentrating on the location of each proposed turbine. Our findings, with the focus on the location of the turbines, support this conclusion and offer a new tool for performing such calculations. Specifically, the proxy of plausible collision risk per turbine can be estimated by deriving the values generated by 3D KUD. This 3D model can assist wind farm developers to calculate the risk of installing a turbine at a specific location. Prior to calculating risk, preconstruction data are needed for target species. Our 3D approach could also be used during the postconstruction and operational phases of wind farms, helping management to predict periods of high risk and reduce the number of bird collisions by selectively curtailing certain wind turbines. We purposely used recorded movement data for a tagged Griffon vulture rather than simulated data to depict the real situation. Griffon vultures have similar flight and foraging behavior (Bosè and Sarrazin 2007, Xirouchakis and Andreou 2009, Mateo-Tomás and Olea 2011, so these results could be generalized to other individuals. Although this new application of 3D KUD can be used for identifying collision risk between obstacles and species in 3D space, some aspects of the method need to be investigated further. For example, more research into producing easily interpretable results with confidence measures is needed. Figure 6. Spatial distribution of wind turbines with a relatively high collision risk in 12 out of 17 months (the entire study period) for a Griffon vulture in 2D (above) and 3D (below) space use in El Estrecho natural park in Tarifa (southern Spain) on the northern shore of the Strait of Gibraltar during 2012-2013.
In addition, spatiotemporal autocorrelation in movement data is an important issue because this yields an underestimation of an individual's space use (Fleming et al. 2015). So far, in animal movement research, many studies have focused on autocorrelated data in 2D, whereas 3D data studies might well be required (Fieberg 2007, Fleming andCalabrese 2017). We expect this new application of 3D KUD to offer exciting opportunities for exploring the process of volumetric analysis in animal movement research, such as spatial autocorrelation in estimating risk and the need to develop methods for 3D kernel density estimators.