Introduction
After centuries of decline and subsequent legal protection, large carnivores are recolonizing parts of their historical European range (Trouwborst, 2010; Chapron et al., 2014). The multifunctional patchwork landscapes to which they are returning vary from moderate to high anthropogenic impact, and across Europe, large carnivores are settling in managed forests (Wabakken et al., 2001; Gurarie et al., 2011), forest-farmland mosaics (Sunde and Olsen, 2018), agro-ecosystems (Blanco and Cortés, 2007), and sometimes near human settlements (Carricondo-Sanchez et al., 2020) and urban areas (Basille et al., 2009; Bateman and Fleming, 2012; López-Bao et al., 2013). The distribution ranges of large carnivores are still increasing, and ecological niche models indicate that European human-dominated landscapes still provide ample space for further range expansions (Milanesi et al., 2017). Hence, although wilderness areas may be essential to keep favorable trends in the long term (Gilroy et al., 2015), the idea that large carnivores can only survive in remote or protected wilderness areas does not apply to today’s Europe (Chapron et al., 2014).
Among the European large carnivores, wolves (Canis lupus) have adapted to the most populated areas, with a mean human density of 36.7 inhabitants/km2 (range = 0–3,050) in areas of permanent wolf presence in Europe (Chapron et al., 2014). For instance, Europe hosts more wolves than the United States, despite being half the size and more than twice as densely populated (Chapron et al., 2014; Boitani, 2018). This land sharing between wolves and humans gives increased potential for direct wolf-human interactions (Bateman and Fleming, 2012).
The prospect of encountering wolves is a factor that can affect human attitudes toward the species (Bath, 2000; Røskaft et al., 2007). Fear of wolves is a recurring source of conflict, and has been associated with a lack of knowledge about the species (Bath, 2000; Bath and Majic, 2001), a perception that wolves are dangerous and unpredictable (Johansson et al., 2012), and a fear of the unknown (Zimmermann et al., 2001). Some people living in areas recolonized by wolves report that a concern for their own or their family’s safety results in diminished quality of life (Røskaft et al., 2007). In modern times, non-rabid wolf attacks on humans are very rare, and documented cases are usually linked to habituation to anthropogenic food sources (Linnell and Alleau, 2016; Reinhardt et al., 2020; Linnell et al., 2021; Nowak et al., 2021). In developed countries, recent focus has been on wolves developing fearless behavior, and on human behaviors that may enhance the risk of attacks (Linnell and Alleau, 2016; Penteriani et al., 2016; LCIE, 2019; Reinhardt et al., 2020). However, there is a lack of research to understand the processes that may lead to risky situations (Löe and Röskaft, 2004; Linnell and Alleau, 2016). Knowledge about wolf behavior toward humans should help mitigate fear and thus facilitate wolf-human coexistence.
Because fear of wolves is a common challenge across their recolonized European range (e.g., Bath, 2000; Bath and Majic, 2001; Røskaft et al., 2007; Johansson et al., 2012; Reinhardt et al., 2020), shared tools for evaluating wolf responses to humans, and comparing such responses across their range, can be valuable. Experiments in which human observers approach marked wolves can increase the knowledge about how wolves can be expected to respond to interactions with humans (Karlsson et al., 2007). Experimental human approaches have been done previously on species such as common buzzards (Buteo buteo) (Sunde et al., 2009a), reindeer (Rangifer tarandus) (Colman et al., 2012), red deer (Cervus elaphus) (Sunde et al., 2009b), moose (Alces alces) (Viljanen, 2019), polar bears (Ursus maritimus) (Andersen and Aars, 2008), brown bears (Ursus arctos) (Moen et al., 2012, 2018; Ordiz et al., 2013, 2019), lynx (Lynx lynx) (Sunde et al., 1998), and wolves (Karlsson et al., 2007; Wam et al., 2012, 2014). However, previous studies on wolves used VHF collars (Karlsson et al., 2007; Wam et al., 2012, 2014), limiting the level of detail at which the wolf response could be recorded without subsequent snow tracking (Karlsson et al., 2007). Today’s GPS collars provide movement data at much higher temporal resolution and spatial precision and accuracy, without the need for snow cover. Additionally, a standardized protocol will allow comparative studies across different study areas, e.g., comparing responses within and between populations, and among areas of varying human density. Hence, a baseline distribution of wolf response patterns can be established given a number of known predictors. Such knowledge about likely wolf responses to humans may increase predictability of wolf behavior to people living in wolf areas.
Large carnivores can show physiological and behavioral antipredator responses to humans (Ordiz et al., 2011; Støen et al., 2015), and may show a proactive fight or flight, or a reactive freeze or hide response (Koolhaas et al., 1999; Roelofs, 2017). The presence or absence of a flight response, and the distance from the disturbance at which the animal flees (flight initiation distance, hereafter FID), are therefore useful parameters to describe the responses of wild animals to human disturbance (for examples, see Colman et al., 2012; Moen et al., 2012; Ordiz et al., 2019; Viljanen, 2019). As shown in brown bears, extensive experimental approaches on GPS collared animals make it possible to compare flight initiation distances between individuals and demographic groups, and to study potential habituation and lasting behavior when approached repeatedly (Ordiz et al., 2013, 2019; Sahlen et al., 2015). For detecting flight initiation during human approaches, two statistical methods have been used previously, i.e., Upper Control Limit (UCL; Moen et al., 2012; Sahlen et al., 2015; Ordiz et al., 2019) and changepoint analyses (Killick et al., 2012; Viljanen, 2019; Græsli et al., 2020). However, it is not known if the methods differ in results for FID calculations.
In this paper, we present (1) a standardized field protocol for experimentally approaching GPS-collared wolves to assess their responses to encounters with humans, controlling for factors such as habitat parameters and number of approaching humans; (2) a descriptive comparison of two statistical methods (changepoint analysis and UCL) to detect the presence or absence of a measurable flight response, and to identify the time and location of both flight initiation and resettling; (3) a tutorial for using changepoint analyses to identify wolf flight initiation and resettling positions as a basis for extracting a number of flight parameters; and (4) a quantitative assessment of the effects on success rate and precision when extracting flight parameters from wolf GPS data at reduced temporal resolution. Lowering the GPS positioning frequency for the approach trials can be an alternative to increase collar battery life. The field protocol, a field data collection form, and a tutorial with the R code for extracting the time and coordinates for flight initiation and resettling are provided as Supplementary Presentations 1–3. This standardized protocol will facilitate studies of wolf responses to direct interactions with humans, and it is applicable at local, national, and international levels.
Discussion
We developed and successfully tested a standardized protocol for experimental human approach trials on wild, GPS-collared wolves. The wolves generally avoided the observers, as described in detail by Versluijs et al. (2022). The purpose of the experimental set-up was to create a realistic but standardized representation of human off-trail hikers in wolf habitat, to assess the wolf flight response at high temporal resolution. The protocol can be expanded, e.g., to involve observers that are jogging, skiing, or accompanied by a dog. Similarly, the purpose of the direct encounter criteria and the actions outlined in the direct encounter protocol, was to mimic the likely response of a hiker when meeting a wolf outside of an experimental setting. It provides information about potential wolf responses while leaving the opportunity for the wolf to retreat rather than provoking an aggressive response. In a non-trial situation, hikers may not detect a wolf that is not obviously visible ahead of them (criterion 2). Once seeing a wolf, we believe that most hikers would stop rather than continue walking toward it. In the presumably rare event that criterion 2 is met, but the wolf does not appear to have detected the observer (criterion 3 not met), we propose walking until detected, so that the wolf detects the observer approaching rather than suddenly counting out loud. The stepwise intensification of the actions in the direct encounter protocol provides the wolf the opportunity to assess and react upon increasing levels of human-induced stress, and its response can be scored accordingly. Notably, we were not able to test our direct encounter protocol as none of the trials fulfilled the direct encounter criteria.
Each approach trial is a large investment in terms of time as well as collar battery life. In the following, we discuss factors that can affect the likelihood of successful trials and the quality of the acquired data, weighed against collar battery usage. Our recommendations are summarized in Table 2 of the field protocol (Supplementary Presentation 1).
Consistent with previous studies of wolf circadian activity patterns in Scandinavia (Eriksen et al., 2011) and elsewhere (Merrill and Mech, 2003; Theuerkauf, 2009), most of the wolves approached in this study were mainly stationary in the middle of the day when the approach trials were initiated. Similar timing has been used in studies approaching brown bears, which have a similar circadian rhythm (e.g., Moen et al., 2012). In two wolf territories however, we were unable to conduct any successful trials due to wolves moving throughout the approach period. Prior to the first approach trial on an individual wolf, it is advisable to inspect the GPS data of the specific wolf. If the wolf shows consistently high activity during the day, the trial can be scheduled for the time of day when the wolf tends to be the least active. Any variation in the time of day of the trials will need to be considered when interpreting the results.
Depending on the collar’s mode of communication, poor coverage can result in a lack of updated positions during the preparation period. The duration and positioning frequency of the preparation period, and the data transmission frequency, can be adjusted to suit each project or study area. Increasing any of these will increase the number of times at which the collar tries to send positions and hence the chances of receiving updated positions in time to start the trial, albeit at battery cost. After we increased the preparation period from 2 h in 2018 to 4 h in 2019–2021, we did not miss any trials due to not receiving updated positions in time. Poor coverage can also result in the collar needing multiple attempts before picking up the approach schedule. This should be considered when planning a trial, making sure to send the programming long enough in advance. In areas with poor GSM coverage, other communication options can be considered. In the case of our study, the four trials that failed due to collars not sending positions were all conducted in territories where one wolf had Iridium and one had GSM communication. Hence, neither of these two communication methods is a guarantee against this problem.
The challenge of reaching the start position in time will depend on accessibility, territory size, and wolf activity during the preparation period, i.e., whether the start position needs to be changed due to the wolves moving. In our Scandinavian study area, the high road density ranging from 0.5 to 1.1 km/km2 allowed us to access the start position by car for most of the trials, although access was sometimes restricted by gated or snow-covered roads. Reaching the start position may be a bigger challenge in more remote study areas.
We expected a starting distance of one kilometer to be far enough that the wolves would not immediately respond to our presence, as previous studies have reported average FIDs of 106 m (Karlsson et al., 2007) and 248 m (Wam et al., 2014), and average alert distances of 293 m (Wam et al., 2014). In our trials, flight initiation distance never exceeded 400 m, hence we believe that a 1 km starting distance will be sufficient in most cases. A shorter starting distance can be considered in areas of high human impact, where it is not feasible to find a starting position at 1 km without other sources of disturbance (e.g., roads) between the wolf and the observer. However, given the range in FIDs and alert distances, the starting distance should not be less than 500 m. We used a passing distance of 50 m rather than walking straight toward the wolf to give the wolf an opportunity to either stay or flee. A 50 m passing distance has also been used during human approaches on Scandinavian brown bears (Ordiz et al., 2013). Even when planning a passing distance of 50 m, the real passing distance will vary due to GPS error, wolves moving between the last acquired wolf position and the passing time, and the difficulty of walking in a perfectly straight line in rugged terrain. However, keeping the passing distance as standardized as possible within and between studies is important for comparing the tendency for the wolves to stay or flee, as it is reasonable to assume that wolves are less likely to stay when passing distances are shorter.
The somewhat arbitrary choice of analyzing the first 10 min of the flight at 1-min resolution (“initial flight”) does not reflect an observed change in flight parameters after 10 min of fleeing (Supplementary Figure 1). However, considering the battery cost of the high positioning frequency, we allocated 2 h of 1-min positioning to each approach. This was assumed to be enough to confirm the wolves being stationary, walk the approach route and have a minimum of ten 1-min positions for analysis. The duration of the post-disturbance positioning (10-min intervals) is set to 3 h in this protocol, but most importantly, it should be long enough to include the entire flight period as defined in the instruction manual for data analyses (Supplementary Presentation 3).
In the vast majority of trials with a detected flight, it seemed convincing from visual inspection that the wolf movement was in fact a response to the approaching human. Only one case was unclear, as visual inspection of the GPS positions revealed two movements, the first of which was detected as the flight initiation by the changepoint analysis (Supplementary Presentation 4, Panel 3D). Importantly, in any given trial and regardless of the definition criteria, one can never exclude the possibility that what we interpret as a flight response is just coincidental movement. Nevertheless, without ever being able to unequivocally identify a flight response, we believe that the frequency of incorrectly inferred flights will be low enough to justify general inference about wolf flight response using our protocol.
The collars were programmed to transmit the maximum available batch size of seven positions. During data transmission the GPS is turned off, and in areas of low network coverage, transmission might exceed 1 min resulting in data gaps on high-frequency schedules. This will result in an earlier detected flight initiation if the flight starts during a gap, since the flight initiation will be the last position before the gap. Excluding approach trials with missing data around the time of the suspected flight initiation is a way to avoid a bias in FID (Moen et al., 2012). However, the potential bias should be weighed against the reduction of the sample size if such gaps are frequent. Imputation as an alternative method to account for gaps is inappropriate in this case, as it is usually based on an assumption of constant speed. Furthermore, different imputation techniques can differ in results (Moritz et al., 2015).
To increase collar battery life, it may be desirable to conduct the trials using GPS data at lower temporal resolution. Furthermore, reducing the positioning frequency may allow enough time for data transmission between fixes, thus reducing the number of gaps. In this study, 23% of the positions on a 1-min schedule were missing versus 4% for 2-min schedule. Calculating flight initiation from GPS data at 2- and 3-min resolution did not reduce the number of detected flights compared to the 1-min positioning data. With the exception of initial speed, we did not detect significant differences in the response variables calculated at 2-min resolution. However, at 3-min intervals, the calculated minimum wolf-observer distance, flight initiation distance, as well as initial speed, differed significantly from those calculated from the original 1-min positioning data. At 5-min resolution, all initial flight variables differed significantly from those calculated from the original 1-min data. Notably, with a larger sample size, significant differences may also be found at 2-min intervals, as the same tendency was observed for all variables also at this resolution. Similarly, down-sampling from 10- to 20-min intervals did not result in loss of detection of resettling, but it did result in increased flight duration and overall straightness, and reduced flight displacement and total distance travelled. Reducing the GPS positioning frequency can be considered in future wolf approach trials to increase collar battery life. However, flight parameters cannot be directly compared across datasets collected at different temporal resolution, and consequently, comparative studies will need to down-sample all data sets to match the one with the lowest resolution. Additionally, with a reduced positioning frequency, it may not be feasible to wait for the first batch of positions from the approach period before making the final corrections to the approach route and starting the trial, increasing the risk of unsuccessful trials. All things considered, when possible, we recommend keeping the 1- and 10-min positioning frequency for the approach period and post-disturbance period, respectively. Nevertheless, our results show that the protocol can generate useful data also at lower temporal resolution.
Changepoint analyses require data point segments of a certain length to detect a change in mean and variance. Hence it is important to have enough data before and after the change (Killick et al., 2012). If the flight happens too early or too late within the 2-h timeslot of 1-min positions, there is a chance that none or inaccurate changepoints are found. As a rule, we intended to reach the passing position (PP) minimum 10 min before the end of the approach period to ensure fine-resolution data for the first 10 min of the flight. However, since gaps in the data exist with an average length of 3 ± 1.4 positions per gap, we rather recommend reaching the passing position (PP) at least 15 min before the end of the approach period. Using 15 min as a buffer will likely create a minimum of ten 1-min positions after flight initiation, which is unlikely to create problems with the changepoint analyses.
Changepoint analysis uses a penalty structure to test the likelihood of a change happening. Different penalty methods are available, and in certain cases it is known which penalty gives the most accurate results in changepoint location (Killick et al., 2012; Truong et al., 2020). The AIC penalty is prone to overestimation as it increases the probability of detecting changepoints (Lavielle, 2005). However, in approach trials with a short flight duration, changepoint analyses with MBIC penalty may not be able to detect changepoints. When visual inspection shows a flight that is not detected using MBIC penalty, AIC penalty may be used, as changepoint analyses are done independently for every individual interaction. However, flight initiations identified with this method should always be checked visually.
In most cases, changepoint analysis detected the resettling with 10-min positioning intervals. Exceptions were if the resettling did not take place within the post-disturbance period or when the flight was too short. Both exceptions can be detected by visual inspection of the data. When the flight is short, detecting the resettling with the 1-min positions might be considered.
The obtained UCL of 23 m/min is comparable to the UCL found in brown bears approached by humans (15.1 m/min, Ordiz et al., 2019). With this UCL, flight initiation could be identified for all interactions where a flight was confirmed by visual inspection. However, filtering the control data on intervals between 30 and 90 s was necessary to calculate the UCL. As the UCL was used as a defined threshold between stationary and non-stationary behavior, a high UCL resulted in the inability to identify flight initiation. Therefore, omitting the deviating time intervals resulted in a better estimate for the UCL, as we were interested in the baseline speed that is detected due to GPS measurement error even when wolves are not moving. As described by Montgomery (2007), a process is out-of-control when exceeding the UCL, which in this case means that the wolf exceeds the speed at which it can be considered stationary. In the brown bear approach studies, the UCL was set and checked visually before defining the FID (Moen et al., 2012; Ordiz et al., 2019). Occasionally the UCL was exceeded with only one position, and visual inspection showed no spatial movement when this occurred. Therefore, we added the condition that the UCL is exceeded for more than one position.
Overall, both changepoint analysis and UCL performed similarly well for detecting flight initiation. Hence, as previous studies have shown, both can be used successfully to detect flight initiation (Moen et al., 2012, 2018; Ordiz et al., 2019; Viljanen, 2019). However, both methods have limitations, and neither performed flawlessly in our study. It might be necessary to calculate the UCL separately for different areas as GPS measurement error can vary between collar type, location and environmental factors. Additionally, UCL calculations rely on control data. Using control data from when the wolves were not moving, based on visual inspection, can result in interpretation errors. In this study, UCL detected all flights which were also confirmed by visual inspection. However, Moen et al. (2012, 2018) reported cases in which visual inspection indicated a flight, but speed did not exceed the UCL. Changepoint analysis does not require control data, and it requires less data preparation, as it can handle various data distributions. However, limitations such as the bias toward the start and the end of the time series, and the probability that a short flight might not result in a changepoint, should be considered. Furthermore, changepoint analyses can be done in various ways, with a broad palette of possible requirements. This makes it customizable to many types of data (Killick et al., 2012, 2016; Truong et al., 2020). However, there is a risk of adjusting changepoint analyses to a desired result. (Le Corre et al., 2014; Edelhoff et al., 2016; Gurarie et al., 2016; Barry et al., 2020). We used visual inspection of the GPS data to confirm the flight initiations detected using automated methods, which we recommend doing given that future studies of this kind will have a limited sample size.
We suggest using changepoint analysis in combination with visual inspection for studying the flight response of wolves during experimental human approaches, as it increases reproducibility and comparability for this type of studies, does not need control data, and can be used in a variety of situations. This is consistent with the objective of the protocol, which is to provide a standardized method which increases reproducibility and is applicable across different study areas.
Although we found that speed alone is a simple and suitable variable for flight detection, other, more spatially explicit tools (such as First Passage Time) have been explored for identifying changes in animal movement patterns (McKenzie et al., 2009; Le Corre et al., 2014; Edelhoff et al., 2016; Gurarie et al., 2016; Barry et al., 2020), and could also be considered.
The standardized protocol developed in this study (Supplementary Presentations 1–3) will facilitate studies of wolf behavioral responses to direct interactions with humans, applicable at a local, national, and international level. From a scientific perspective, data compiled in a standardized way from multiple study areas can be used to quantify the variation in wolf responses to humans within and between populations, and in relation to predictors such as social or reproductive status of the wolves, landscape factors, human population density, and proxies of anthropogenic impact. Once standardized data with a sufficient range in the above-mentioned predictors have been collected, the variation in wolf responses can be used to quantify the degree to which results can be generalized for comparable areas and circumstances. This is of particular importance, as a likely limitation for the extensive application of this protocol is the cost associated with equipping wolves with programmable GPS collars. Hence, from a management perspective, results might be used to establish a baseline distribution of wolf response patterns given a number of known predictors. This information can contribute to the knowledge base for management authorities also in areas where GPS studies are not feasible, e.g., for predicting the range of likely wolf behaviors, assessing the likelihood of wolf-human encounters, and complementing existing tools for evaluating reports of wolves that are perceived as bold (Karlsson et al., 2006; LCIE, 2019; Reinhardt et al., 2020). From a dissemination perspective, showing how wolves in general, as well as in specific cases, respond to human encounters, should help demystify the behavior of wild wolves toward humans in their shared habitat. In fact, to date, neither brown bears nor wolves have shown any aggressive reaction to standardized approach trials in Scandinavia (Karlsson et al., 2007; Wam et al., 2012, 2014; Ordiz et al., 2019; Versluijs et al., 2022), which is by itself an important message for managers and the general public. Fine-scale GPS data from experimental approach trials gives ample opportunities for displaying the results graphically, e.g., in videos or animations. However, finding effective ways of disseminating such information to the general public is a challenge, as the scientific literature reports mixed results from efforts to use dissemination to reduce fear of carnivores (Johansson et al., 2016a). Some studies propose animal or habitat exposure to increase predictability of animal behavior (Johansson et al., 2016a), and in Scandinavia, bringing people along for guided brown bear approach trials resulted in reduced self-reported feelings of fear among the participants (Johansson et al., 2016b). However, Johansson et al. (2016a) points out that any intervention aimed at reducing fear toward large carnivores should be accompanied by thorough evaluations of the effects of the intervention, preferably as part of an adaptive management scheme. From a conservation perspective, human approach trials generate important information on the behavioral reactions of large carnivores inhabiting human-dominated landscapes, such as the time it takes to resume their regular circadian activity patterns after being disturbed. Altering the time budgets from more profitable activities (foraging, resting) to increased vigilance after disturbance may have fitness costs (Ordiz et al., 2013). Therefore, quantifying how human activities can affect carnivore behavior, and the potential implications for energy budgets (e.g., Bryce et al., 2022), can elucidate the role of wilderness areas for large carnivore conservation (Gilroy et al., 2015).