Opportunities and risks in the use of drones for studying animal behaviour

In the last decade, drones have become an affordable technology offering highly mobile aerial platforms that can carry a range of sensory equipment into hitherto uncharted areas. Drones have thus become a widely applicable tool for surveying animal populations and habitats to assist conservation efforts or to study the behavioural ecology of species by monitoring individual and group behaviour. Here, we review current applications for drone surveys and the potential of recently developed computer algorithms for automatic species detection and individual tracking in drone footage. We further review which factors are reportedly associated with animal disturbance during drone presentations and how drones may be used to study anti‐predator behaviour. Drone surveys of species and their environments allow scientists to create digital terrain models of habitats, estimate species abundance, monitor individual behaviour and study the composition, spatial organization and movement of groups. As drones can influence the behaviour of many bird and mammal species directly, they also provide an experimental tool to study animal responses to novel situations, including the drone itself. We conclude that the combined use of drones and automated detection software can assist population estimates and opens new possibilities to study individual and collective behaviour. With regard to drone‐related disturbance and their potential use as predator models, we recommend to interpret results against the background of population‐specific predation pressure and sources of anthropogenic disturbance.

In the wake of this recent surge of drone usage, however, numerous reports indicate that animals may respond to drone surveillance with avoidance behaviour and other signs of disturbance including alarm calling, mobbing and displacement activities (Christie et al., 2016;Mulero-Pázmány et al., 2017) or less overt physiological responses (Ditmer et al., 2015). These reports have led to an increasing number of studies attempting to identify which factors during drone presentations typically lead to behavioural responses to minimize disturbance (Mulero-Pázmány et al., 2017;Singh & Frazier, 2018). Although most types of survey methods are associated with some form of animal disturbance, understanding how each method can affect the data collected is necessary to determine the optimal survey method and minimize sampling bias.
Here, we review drone applications that may be of interest for ecologists and behavioural scientists. We discuss how drones can be used for monitoring and as experimental stimuli to study group dynamics during movement, anti-predator behaviour and responses to novelty. Finally, we present a case study from our laboratory to illustrate how drones may be used to study alarm calling behaviour and auditory learning. While drones are generally applicable to investigate anti-predator behaviour, various challenges and limitations remain.

| SURVE YING ENVIRONMENTS AND P OPUL ATIONS
Monitoring environments, assessing the distribution of organisms and estimating fluctuating population sizes are common challenges for ecologists and conservationists. Complementing ground research and satellite imaging with standardized drone transects can quickly cover large areas. Drone-based assessments can be conducted independent of cloud cover with high spatial resolution and thus provide a valuable tool for ecosystem monitoring (Hughey et al., 2018).
The currently most common and commercially available models are fixed-wing and multi-rotor drones. Fixed-wing drones can cover larger areas, fly higher and typically support flight times of 1-2 hr, but have to circle continuously to remain above an area of interest. Multi-rotor drones can hover stationary above a target area, operate at lower altitudes and can rapidly adjust their position if surveilled animals on the ground start moving, but usually do not exceed flight times of 30-40 min. Aerostats, such as balloons or powered airships filled with gas lighter than air , have seen comparatively less use, but can provide long time aerial surveillance of designated areas. Flapping-wing drones and other 'avian-like' models may be used if drones should resemble raptors (Folkertsma et al., 2017) or need to be capable of specific flight manoeuvres (Ajanic et al., 2020). The increasing range of drone models allows scientists to tailor their choice of airframe and carried sensors to the specific needs of their study design, the targeted species or the weather conditions and temperature range of the environment in which the drone is to be operated (Anderson & Gaston, 2013;Linchant et al., 2015), see Figure 1.
To study nocturnal and elusive species or survey animals in habitats with poor contrast between individuals and background, drones can be equipped with infrared cameras .
Finally, drones can measure detailed structural characteristics of environments to create three-dimensional reconstructions of habitats (Hughey et al., 2018;Tuia et al., 2022). Such computer models can be used to study animal movement as a function of habitat structure (Strandburg-Peshkin et al., 2017) or help to asses long-and short-term fluctuations of habitat quality for different species (Olsoy et al., 2018).

| ALGORITHMS FOR AUTOMATED DE TEC TI ON OF S PECIE S , IND IVIDUAL ORIENTATI ON AND BODY S IZE
While early studies primarily evaluated drone footage manually, recent years have seen an increase in the development of machinelearning algorithms for automated detection of target species from visible-spectrum and thermal video recorded during aerial surveys (Corcoran, Winsen, et al., 2021;Eikelboom et al., 2019;Hollings et al., 2018;Kellenberger et al., 2018;Lyons et al., 2019), see aspects and issues regarding double counting can be accounted for is therefore critical for the reliability of drone survey-based population estimates.
In addition to automated detection of species, software has been developed to track the movement, body orientation and approximate visual fields of individuals from drone footage (Graving et al., 2019;Mathis et al., 2018;Walter & Couzin, 2021), see Figure 2d. Originally developed for laboratory studies, these methods may now be applicable in field settings. Although body size of targeted species and habitat structure constrain the use of these tools, they possess unprecedented potential for the study of collective movement and behaviour (Hughey et al., 2018). Drones may even track and monitor groups independently, as demonstrated by attempts to equip drones with radio-telemetry sensors that automatically find and follow tagged individuals (Cliff et al., 2018;Hui et al., 2021). Finally, video and photogrammetry methods have been developed that use visible-spectrum and thermal drone footage of marine mammals to estimate individual body size or physiological parameters (Dawson et al., 2017;Horton et al., 2019;Krause et al., 2017).
While automated species detection has been shown to be a promising method for population estimates (Corcoran, Winsen, et al., 2021;Hodgson et al., 2018;Kellenberger et al., 2021), automated tracking of individual pose and movement still requires more empirical testing. Nevertheless, automated evaluation of drone recordings combined with other forms of remote sensing or biologgers has tremendous potential for behavioural ecology and conservation research (Hughey et al., 2018;Tuia et al., 2022). Implementing software for automatic tracking, drones may now also provide the necessary data on individual position. Our own attempts to monitor Guinea baboon, Papio papio, movements, were thwarted however, as the baboons mostly remained under tree cover as long as the drone was hovering above (D. Montanari, pers. commun.).

| MONITORING G ROUP MOVEMENT, S PATIAL ORG ANIZ ATION AND INDIVIDUAL B EHAVIOUR
The first studies that used drones to gather data on movement and inter-individual distances focused on larger herbivores. Using drones to track caribou during their migration, Torney et al. (2018)) investigated whether movement choices of individuals belonging to different age classes depended on the movement of neighbouring animals and how individual decisions in turn affected group-level trends, see Figure 2b. Drones were also used in conjunction with ground observers to estimate inter-individual distances and individual position relative to the group's centre in feral horses (Inoue et al., 2019(Inoue et al., , 2020. The authors employed the same approach to create inter-individual proximity networks estimated from drone footage to make inferences about the social organization of the species (Maeda et al., 2021). While they evaluated drone footage manually, future field studies could incorporate tracking tools that have been tested on drone footage from zebras (Graving et al., 2019;Walter & Couzin, 2021), see Figure 2d.
In marine habitats, drone monitoring was applied to measure individual behavioural events and states in cetaceans or detect and track movements of animals in shallow waters. The results of these studies suggest that drone monitoring increases effective observation time and can provide data on habitat use, social and foraging behaviours that are difficult to collect during conventional boat surveys (Giles et al., 2021;Oleksyn et al., 2021;Torres et al., 2018).

| DIS TURBAN CE AND ANTI -PREDATOR B EHAVIOUR
The first descriptions of drone induced disturbance behaviour were reported in studies that were primarily concerned with using drones for population estimates. As it was apparent that aversive reactions of animals would have a negative impact on census quality, scientists began to investigate this phenomenon directly. While this led to a vast literature covering many taxa, the methods employed in these studies are very diverse, which hinders systematic comparisons, as has been pointed out in previous reviews on this subject (Christie The noise produced by drones and the increase in sound pressure level that is associated with shorter distances between drones and surveyed animals has also been identified as a relevant factor for disturbance (Bennitt et al., 2019;Rümmler et al., 2021;Schroeder et al., 2020;Vas et al., 2015;Weimerskirch et al., 2018). Duporge et al. (2021) reviewed this subject specifically and suggested a method to calculate the minimum drone altitude based on the audiogram of the target species and the noise profile of the drone at different heights. However, even if species-specific sensory abilities, prevalent background noise levels and the sound propagation characteristics of the environment are known or can be estimated with reasonable certainty, the potential effects of drone noise should also be evaluated with regard to similar anthropogenic noise that animals may experience and could associate with the noise profile of the drone. Another important aspect in this context is that the flight behaviour and approach speed of the drone will affect the noise profile. In marine habitats, drone noise appears to be less problematic since it does not penetrate deep into the water, meaning submerged animals are less likely to be affected than those close to the surface (Christiansen, Rojano-Doñate, et al., 2016). Bird colonies with high levels of background noise may also show higher tolerance for drone sound (Weimerskirch et al., 2018).
Especially for species facing aerial predation, the shape and movement of novel flying objects might determine whether they categorize them as potential threats or not. Evidently, the more drone behaviour and silhouette resemble that of an aerial predator, the more likely it is that drones provoke escape responses (McEvoy et al., 2016). To elicit anti-predator responses experimentally and avoid habituation of subjects, scientists may consider testing flapping-wing drones (Folkertsma et al., 2017) or other more 'avian-like' models (Ajanic et al., 2020) that resemble aerial predators and are capable of flight manoeuvres observed in raptors. Noteworthy, species-and population-specific predation pressure have been reported to affect the occurrence of disturbance responses in mixed ways. Bennitt et al. (2019) reported that ungulates with higher assumed levels of general predation pressure were less responsive to drones, whereas Brisson-Curadeau et al. (2017) reported that seabird colonies with aerial predators on site responded more strongly than those where aerial predators were absent. We therefore suggest that the extent of populationspecific predation pressure should be reported and considered when interpreting anti-predator responses.

Variation of response intensity between populations and
individuals is another aspect requiring further investigation.
Differences between populations could arise due to variation in predation pressure or the extent of anthropogenic disturbance (Ditmer et al., 2019;Schroeder et al., 2020). Inter-individual differences were found to be associated with age categories and the breeding status of individuals (Brisson-Curadeau et al., 2017;Pomeroy et al., 2015;Weimerskirch et al., 2018). Future research should consider social and demographic factors to evaluate how these aspects of group composition affect group-level trends during anti-predator behaviour (Schroeder & Panebianco, 2021;Torney et al., 2018).
Finally, group size can affect the likelihood, intensity or distance at which responses occur (Giles et al., 2021;Ramos et al., 2018;Schroeder et al., 2020;Schroeder & Panebianco, 2021). Although quantification of group-level responses is challenging under field conditions, disturbance or flight behaviour can frequently be observed to spread contagiously within groups (Pomeroy et al., 2015;Vas et al., 2015;Weimerskirch et al., 2018). While it has been suggested that larger groups may be able to detect drones faster (Schroeder & Panebianco, 2021), it appears that animals often detect drones acoustically before they can be assumed to be able to spot them visually. Since the predator detection advantage of larger groups is mostly understood as a function of the larger areas they can cover visually by collective vigilance, this mechanism alone would not explain different response characteristics of larger groups if drone detection first occurs acoustically. Simulations and laboratory experiments that investigate collective decisions suggest that the increased sensitivity of groups that can be observed in the form of cascading escape responses is a function of how individuals contribute and respond to the dynamic spatial organization of the group itself (Sosna et al., 2019;Sridhar et al., 2021). Furthermore, increasing group size can reduce the relative proportion of individuals with preferred movement directions that is needed to guide group movement (Couzin et al., 2005). Thus, in the context of drone avoidance, the first individuals to respond likely affect subsequent decisions of other group members who may respond stronger to the reconfiguration of a larger group's spatial organization than to the drone itself.

| DRONE S A S NOVEL AERIAL THRE ATS
In previous work from our group, Wegdell and colleagues used drones as novel, aversive aerial stimuli to investigate the mechanisms that guide predator recognition and alarm call production in free ranging monkeys (Wegdell et al., 2019). They presented drones to West African green monkeys Chlorocebus sabaeus to test whether the monkeys would categorize the drones as potential aerial predators that would elicit the production of alarm calls. Although researchers at the site had rarely heard aerial alarm calls in this population before (presumably due to the absence of aerial predation), the monkeys responded strongly to the drones and the produced calls closely resembled aerial alarms of a congener, the East African vervet monkey Chlorocebus pygerythrus (Wegdell et al., 2019).
Wegdell and colleagues followed up the drone presentations by playing back the drone sound to individuals and found that monkeys showed more vigilance during test trials compared to control trials.
Since the monkeys did not only show general vigilance behaviour but specifically began to look up and scan the sky, this suggested that they had learned to connect the sound with the presence of an aerial threat. This design could be applicable to other taxa as well that produce distinct alarm calls for different threats to probe whether animals categorize novel flying objects as potential aerial predators. As long as the target species has a form of vigilance behaviour specifically associated with aerial predation avoidance, playbacks could be used to study whether a species is capable of rapid auditory learning in a predator context.
Future studies that aim to use drones to investigate alarm calling behaviour should aim to gather information on the natural frequency and duration of aerial alarms typically produced by their study species. Such data allow for a more efficient planning of the study and the set-up of the recording devices. Given the uncertainty which animals will be calling, it is necessary to have a sufficient number of researchers on the ground. Another problem is the inevitable habituation to the drone and we strongly recommend to keep the presentation of the drone to a minimum. Previous experience with drones may completely derail a study devoted to responses to entirely 'novel' stimuli. It is thus necessary to gauge the prior experience before commencing such a study. We also recommend considering to what degree the potential study population is exposed to airplanes, helicopters and other forms of anthropogenic disturbance that have a remote resemblance to drone overflights.

| CON CLUS IONS
Utilizing the potential of aerial monitoring with automatic detection of species and individual orientation and movement, drones may soon be able to conduct semi-autonomous surveys and collect highly standardized data on habitat structure, species abundance and the demographic composition and movement patterns of groups. This new source of data could improve our estimates of population sizes, the spatial and temporal distribution of species and also allow inferences about the mechanisms that guide individual and collective behaviours under natural conditions.
In addition to monitoring, drones can be used as novel aerial threats to probe the cognitive abilities of species. In some cases, they can elicit alarm calling and other anti-predator behaviour or may be used in combination with playbacks to investigate the auditory learning capabilities of species. Unless there is a direct need to use drones as predator models, it may therefore in many cases be better to habituate subjects to drones first and subsequently study their responses to more realistic predator models while using drones for monitoring during such experiments. In conclusion, drones are a promising tool for experimental and long-term data collection that can be used alongside other forms of remote sensing and conventional field observations.

AUTHOR S' CONTRIBUTI ON S
Both authors conceived the review; L.S. conducted the literature search and wrote the first draft. Both authors edited the manuscript.

CO N FLI C T O F I NTE R E S T
None declared.

PEER R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1111/2041-210X.13922.

DATA AVA I L A B I L I T Y S TAT E M E N T
This paper does not contain any data or code.