A supervised model to identify wolf behavior from tri-axial acceleration

Abstract

Background

In wildlife studies, animal behavior serves as a key indicator of the impact of environmental changes and anthropogenic disturbances. However, wild animals are elusive and traditional GPS studies only provide limited insight into their daily activities. To address this issue, behavior classification models have increasingly been used to detect specific behaviors in wildlife equipped with tri-axial accelerometers. Such models typically need to be trained on data from the target species. The present study focuses on developing a behavioral classification model tailored to the grey wolf (Canis lupus) and encompassing a variety of ecologically relevant behaviors.

Methods

We collected data from nine captive wolves equipped with collar-mounted tri-axial accelerometers recording continuous acceleration at 32 Hz (“fine-scale”) and averaged acceleration over 5-min intervals (“activity”). Using simultaneous video observations, we trained Random Forest models to classify wolf acceleration data into specific behaviors. We investigated the potential limits to the generalizability of these models to unlabeled data through individual-based cross-validation.

Results

We present: (1) a model classifying fine-scale acceleration data (32 Hz) into 12 distinct behaviors (lying, trotting, stationary, galloping, walking, chewing, sniffing, climbing, howling, shaking, digging and jumping) with a class recall of 0.77–0.99 (0.01–0.91 in cross-validation), (2) a model classifying activity data (5-min averages) into 3 behavior categories (static, locomotion and miscellaneous) with a class recall of 0.43–0.91 (0.39–0.92 in cross-validation). Although classification performance decreased following cross-validation, recall scores for lying, trotting, stationary, galloping, walking and chewing individual behaviors (as well as static and locomotion categories) remained above 0.6. Classification performance was consistently poorer for rare behaviors, which constituted less than 1.1% of the training dataset.

Conclusions

We demonstrate the use of collar-mounted accelerometer to distinguish between 12 behaviors and 3 behavior categories in captive wolves, at fine-scale (32 Hz) and averaged 5-min resolutions, respectively. We also discuss the generalizability of the two models to free-ranging settings. These models can be employed to support future behavioral studies examining questions such as conflict mitigation, wolf responses to human disturbances, or specific activity budgets.

Background

In wildlife studies across the world, some recurring constraints consistently limit data collection: wild animals are elusive, accessibility and detectability are poor, and field conditions unpredictable [5, 97]. Data on physiology and behavior are especially difficult to obtain at fine temporal and spatial scales, yet they are key indicators of human impact on wildlife [8, 13, 89, 91], both in fundamental research and for implementation of conservation strategies [55]. Over the past decades, bio-loggers have been increasingly used to fill in this gap, with ever-growing precision and versatility, now allowing for the collection of high-resolution spatiotemporal information [2, 8, 15, 16, 65, 66].

In this context, the use of high-resolution data collected from animal-borne accelerometers to detect animal behavior has emerged as a promising solution to several problematics. Location data from GPS trackers have historically been used for behavioral studies, but limitations in device autonomy and available resolutions typically constricted data collection and resulting studies to large-scale habitat use and individual-scale speed [97]. Fine-scale accelerometers are also battery and storage-demanding, which can limit their use to one year at a time under the best scenarios: however, at a comparable scale (e.g., yearly, with resolution as high as 1 min or more), turning to raw acceleration allows for collection of more precise information than with GPS data, provided analysis methods are available. Once calibrated by associating acceleration signal patterns with observed body movements [47, 65], supervised machine learning algorithms such as Random Forests can be used to predict behaviors on unlabeled datasets from the target species [11, 20, 49, 73]. The availability of high-resolution acceleration data thus opens new possibilities for behavioral studies at finer scales [24, 60].

However, even though machine-learning potentially offers a high level of accuracy in such behavioral detection, and has been recently used in various species and fashions [6, 14, 29, 47, 49, 53, 56, 62, 64, 65, 67, 71, 73, 83, 92, 94], they most often remain constricted to few behaviors with a limited level of detail. The present study focuses on the grey wolf (Canis lupus), a species that has been extensively studied in various aspects of spatial ecology, from broad landscape use [61] to specific hunting strategies [78], often relying on extensive observations [58] or decades of GPS tracking [13, 79], and aims at identifying 12 ecologically significant behaviors.

In a study from 2022, the Overall Dynamic Body Acceleration (ODBA) from nine captive wolves fitted with 32-Hz accelerometers was used to detect 5 behavior categories (resting, stationary, walking, highly active, and running [12], allowing to quantify day and night behavior patterns during breeding, pup rearing and recruitment seasons, in Denali National Park. Other studies targeted specific questions such as detection of kill sites [68], or use calibration based on domestic dogs [20]. Wolves are involved in human–wildlife conflict across their whole range [22, 25, 27, 31, 52, 80], which are often fueled by perceived lack of appropriate management response [45, 46, 57, 59, 76, 86, 93]. The ability to detect specific behaviors with ever-growing level of detail holds promises in improving tailored management responses: for instance, preventive prediction of conflict hotspots (which is often used in conservation management practice, see, e.g., [36, 48], has been refined over the last decades by the implementation of connected systems allowing, for instance, real-time alerts of raiding of agricultural areas [75]. Some of these systems now provide early detection through behavioral cues [40, 87] or even use detection-triggered repellents [1, 28, 38, 39, 90], and real-time treatment of acceleration data from an algorithm embedded in a collar to send out alerts based on behavior detection is achievable today [44]. The use of behavioral cues in early conflict detection has proven effective in preventing animal casualties [7], and unequivocal fine-scale behavior detection from acceleration data could further enhance such use, moving toward systems tailored to specific situations.

Provided an accurate and fine-scale behavioral detection algorithm is available, such systems could be used to bring up new insight into wolf ecology, but also adapted to alleviate conflicts related to wolf depredation on domestic animals, since automatic repellents and guarding systems have been demonstrated to be effective on short and long term, respectively [10, 25, 37]. Therefore, in this study we develop a behavioral classification model for detecting a maximum number of ecologically relevant behaviors to accommodate both behavioral research and conservation needs. We first present a model predicting 12 distinct behaviors, calibrated with supervised machine learning using video footage and tri-axial acceleration data sampled at 32 Hz from captive wolves.

However, this fine-scale setup is not the norm in most monitoring frameworks, where devices recording summarized bi- or tri-axial accelerometry (“activity”) are typically used. These devices have been available for longer, are less costly, less power- and storage-hungry, and align with most monitoring or spatial studies only requiring a few positions every day. Hence, we complementarily present a model predicting 3 behavior classes from calibration with such activity data, to be used on datasets commonly collected during routine monitoring, and discuss the need and hindrances to developing replicability between study sites, programs, species, and frameworks.

Methods

Data collection

Data collection took place at the Wildlife Science Center in Minnesota, from October to December 2018. We studied three family groups in separate enclosures, each consisting of a pair of adults and three to six younger wolves. We collected data from three 1.5-year-old wolves in each enclosure, one enclosure at a time (i.e., a total of 9 individuals, see Additional file A.I). The enclosures were flat with no vegetation. All subject animals were immobilized with ketamine and xylazine, weighed, their fur dyed on specific body parts for individual identification, and fitted with GPS collars featuring store-on-board tri-axial accelerometers (Vectronic VERTEX Plus V 2.1, Vectronic Aerospace Inc, here 1.74%–2.61% of body mass). The acceleration modules were set to record (1) tri-axial acceleration (surge, horizontal, X; sway, lateral, Y; and heave, vertical, Z, see Fig. 1) at 32 Hz and 4 g, and (2) summarized tri-axial acceleration over 300 s (Vectronic Aerospace Inc, 2024). Each enclosure was equipped with four IR cameras (Cromorc Wireless Security 1.3 MP, Cromorc Inc.) at the corners.

Fig. 1

Orientation of the accelerometer contained in the collar fitted on a subject wolf. X: surge; Y: sway; Z: heave; pitch and roll: vertical and lateral orientation across a 180° angle, respectively

Data processing

Processing of raw data

For our analysis we used both raw tri-axial acceleration continuously at 32 Hz (hereafter “fine-scale” datasets and corresponding models), as well as tri-axial acceleration averaged across 300 s (hereafter “activity” datasets and corresponding models, Vectronic Aerospace Inc., 2024), and continuous video footage for each enclosure for two to three days [see Additional file A.I]. Since most previously published ethograms are only relevant for wolves in free-ranging conditions [54], and to account for biologically relevant behaviors displayed by the subject wolves, we designed a specific ethogram based on observations from the video footage, which we used as a reference for labeling [see Additional file A.II]. One observer used software BORIS (v.7.4.1; [33]) to label behaviors with start and stop times, from video segments of 30 min every 3 h throughout the filming periods (total: 8 h per individual), with additional observations of rare behaviors at feeding time. To improve prediction accuracy during model training, we only classified clear, distinct, non-overlapping behaviors with unambiguous identification of the individual. We did not include state-transition behaviors.

Behaviors selection and formatting

The ethogram was composed of 12 behaviors (Fig. 2) based on ecological significance and potential for practical use in wildlife research and management: locomotion (walking, trotting, galloping), interaction with the environment (jumping, climbing, shaking, digging), and with food and conspecifics (chewing, sniffing, howling). We partitioned static states into lying (lying down immobile), and stationary (sitting, lying down while moving the head, or standing), since all of the latter presented high similarity in acceleration metrics. Hereafter, we call “rare” behaviors with a prevalence inferior to 1%: galloping, sniffing, climbing, howling, shaking, digging and jumping. Acceleration data and behaviors were matched using their time stamps.

Fig. 2

Acceleration signatures of the 12 fine-scale wolf behaviors targeted for classification using machine-learning algorithms. Raw surge (x, red), sway (y, yellow) and heave (z, blue) are displayed on a main 15-s intervals grid with colored bars indicating the corresponding observed behavior

In “fine-scale” models, we compared a model featuring all 12 behaviors (model A, Table 1), to a model in which rare behaviors were grouped into a “miscellaneous” category (model A.1, Table 1). When analyzing the 300-s averaged activity dataset, we tested different ways of grouping the 12 behaviors into fewer ecologically and practically meaningful categories to account for the comparably short duration of individual behaviors (Table 1). For instance, we included a 5-category model (Model C), intending to preserve explicit detection of certain behaviors of ecological interest despite low classification success (namely, separating running and walking, and including chewing). Each 300-s interval was labeled according to the comparative duration of all fine-scale behaviors or categories observed within it, with the label corresponding to the behavior or category that occupied the longest duration.

Table 1 Results of Random Forest model building for classifying wolf behaviors from acceleration data

To correct for sampling imbalance caused by the inclusion of rare behaviors, we incorporated a relative weight calculated as the proportion of each behavior relative to the overall dominant behavior (fine-scale: P_lying = 0.637, and activity: P_stationary = 0.237, detail of sampling proportions is provided in Additional file A, Table III).

Decision variables

In line with Shepard et al. [83], we used biologically meaningful variables derived from the raw acceleration data, characterizing the orientation and movement of the neck [99]. Aside from standardized mean tri-axial acceleration values (x, y, z), we included mean pitch and roll representing vertical and lateral orientation across a 180° angle, respectively (Fig. 1), dynamic tri-axial accelerations (D_x, D_y and D_z), mean overall dynamic body acceleration (ODBA), mean vector of dynamic body acceleration (VEDBA), mean magnitude of acceleration (AMAG) and variance of all previous variables (V) [detail of equations in Additional file B.I]. To calculate these acceleration-derived variables, we tested several running means (S) ranging from 1 to 30 s. Based on performance during model fitting, the comparative duration of behaviors (e.g., Fig. 2), and the high mobility of the wolf as a species, we used a 1-s running mean for all analyses.

Model building

We built a Random Forest (RF) multi-class classification algorithm [11, 23, 51, 63], using R package h2o [34]. We trained a behavioral classifier with 200 trees to balance explanatory power and computational cost [69], using an 80–20 randomized split keeping behavior class proportions for the training and testing sets, respectively [26, 30] and a one-versus-all approach [74]. To alleviate potential overfitting and gain a better understanding of the model’s potential in classifying unlabeled data, we performed a leave-one-individual-out cross-validation, with a final performance score averaged across the 9 wolves [4, 100]. All analyses were performed on R version 4.3.3 [70].

Model performance

As we aimed at achieving the best identification of behaviors displayed by unseen free-ranging individuals, we mainly based our performance assessment on individual class recall (i.e., proportion of true positives predicted out of each class) and precision (i.e., proportion of true positives out of all positive predictions) [5]. We also investigated log-loss (quantifying the uncertainty of a classifier by penalizing false classification) and relative variable importance (based on how much the overall squared error decreased when a variable was selected to split on during the tree building process). We visualized decision variables using packages plotly version 4.10.4 [85] and ggplot2 version 3.5.0 [95], and model results with package networkD3 version 0.4 [35]. The final results incorporate general training and cross-validation into what we will call the “main models” (“MM”). However, we will focus our discussion on extracted cross-validation results (in the following, “cross-validation” or “CV”) as a reflection of situations with high individual variation, which can limit the potential of the main model.

Results

Behavioral characteristics

Of the 12 behaviors selected for classification, two (stationary and lying) contributed 87.7% of all 32-Hz acceleration datapoints. The subsequently most represented behaviors were trotting (5.7%), walking (2.9%) and chewing (2.6%). The remaining behaviors were rare and made up 1.1%. Prevalence in activity categories followed the same pattern. Behavior prevalence varied between individuals, and jumping, climbing, and digging, were not displayed at all by 1, 2, and 3 individuals, respectively. Due to technical limitations during footage analysis in one enclosure, 3 individuals contributed less than the others to the overall datasets [see Additional file A.III].

Fine-scale models (32 Hz)

While results of the main 12-behavior model returned a mean per-class recall of 0.88, with scores ranging from 0.77 (shaking) to 0.99 (lying), the cross-validation results taken separately showed a mean per-class recall of 0.42, ranging from 0.01 (digging) to 0.91 (lying) (Model A, Fig. 3, details in Table 1). The behaviors classified with the best recall score were also the most represented in the dataset, with lying, stationary, and trotting scoring above 0.75, and walking, galloping and chewing scoring between 0.60 and 0.70. The remaining rare behaviors scored a recall lower than 0.1, except for sniffing (0.25). Between individuals, mean recall varied from 0.71 to 0.93. Precision scores were always lower than or comparable to recall, varying from 0.006 (digging) to 0.98 (lying). Finally, the most important variables to predict behaviors were mean pitch, x, ODBA, VEDBA, and z (Fig. 3 and Additional file C.VIII). Grouping rare behaviors together in a reduced model (Model A.1, details in Table 1) did not significantly improve recall or precision for the remaining behaviors.

Fig. 3

Sankey diagram representations of cross-validation confusion matrices for models A and B, and variable importance. Observed behaviors are featured on the left with relative prevalence indicated in associated circles. Classification results are featured on the right, alongside performance metrics for each classified behavior. Internal links represent classification confusions. The length of external bars on each side and the thickness of internal links represent relative prevalence in sampling, confusions and classification. MM: main model; CV: cross-validation. (all confusion matrices are available in Additional files C.I–VII)

Plotting head pitch against overall dynamic body acceleration (ODBA), which are among the most important variables, and useful for representing behaviors in a 2D space, shows variable levels of overlap between acceleration datapoints for each behavior (Fig. 4).

Fig. 4

Distribution of the main decision variables (pitch, ODBA) for each behavior in the fine-scale model. Distribution of calculated mean pitch (vertical angle of the head of a wolf) and mean ODBA (overall dynamic body acceleration) for static (a–c), locomotion (d–f) and miscellaneous (g–l) behavior categories, for each 32-Hz datapoint. Each behavior within a category is represented in a consistent shade of grey, with target behavior in its associated color, to better highlight specificities and overlaps

Activity model (300 s)

During classification training for activity data, we compared models with various behavior groupings. The simplest 3-category model (static, locomotion and miscellaneous, Model B, Table 1) was the most performant, with per-class recall scores of 0.92, 0.77 and 0.44, respectively (0.92, 0.73 and 0.39 in cross-validation) (Fig. 3). Precision scores were lower than recall, the most prevalent behaviors in the dataset were classified best, and the most important variables for activity models were mean activity, AMAG, and mean sway (y). Model B also strongly relied on mean surge (x) [Additional file C.VIIi]. Amongst individuals, we obtained mean CV recalls ranging from 0.48 to 0.88. We tentatively performed a leave-one-out approach to enrich this simple 3-category model B with more behaviors, in a 5-category mixed model (Model C, stationary, running, walking, chewing, miscellaneous, Table 1). Albeit not improving walking scores compared to the full 12-behaviors model F, model C slightly improved chewing scores (per-class recall scores of 0.92, 0.81, 0.26, 0.51 and 0, respectively) (0.93, 0.77, 0.16, 0.5 and 0 in cross-validation) [Additional file C.IV], and remained similar to model B in regard to performance metrics.

Discussion

Model performance

From the Random Forest trainings, we obtained models able to discriminate between several combinations of ecologically and practically significant behaviors. We selected two of them for discussion and general use: one model (A) using 32-Hz data to identify 12 behaviors with recalls above 0.77 (of which 6 behaviors with a recall score above 0.6 in cross-validation), and one model (B) using activity data to identify 3 behavior categories, with static and locomotion behaviors scoring recalls above 0.73 in both general model performance and cross-validation. As we want to explore more specifically the potential caveats related to individual variation, in the following, we will only use cross-validation recalls unless otherwise specified.

Fine-scale models

In the fine-scale model, recall scores per behavior were overall positively correlated with prevalence in the training set (Fig. 3), except for the running behaviors (trotting, 0.90 and galloping, 0.71) which, albeit constituting 6% and 0.4% of the dataset, respectively, scored similarly to static behaviors (lying, 0.91 and stationary, 0.79) constituting 64% and 24% of the dataset. Other behaviors making for less than 3% of the dataset returned mid-range recall scores (between 0.6 and 0.3). Notably, howling (prevalence 0.4%) scored below 0.1, and was frequently confused with being merely stationary, implying that the distinctive back and forth of the head on the z-axis when a wolf howls (Fig. 2) was overridden in the model by the comparably upright position of the head. This is consistent with mean pitch and ODBA being the most important decision variables in the model. Behaviors involving little body movement (e.g., lying or standing in different positions) have been reliably classified in previous studies and various species [29, 41, 47, 65, 67, 92, 94], but can be difficult to differentiate without reducing overall classification performance (e.g., by increasing the time window for smoothing running means, [47, 82]). This phenomenon seemed to occur between the three static behaviors, lying, stationary and howling.

Despite the use of a weighting factor, all rare behaviors (prevalence under 0.1%, i.e., climbing, shaking, digging and jumping) were poorly classified and frequently confused with other behaviors, which is a common issue in imbalanced datasets. These behaviors also displayed very short duration without distinctive pattern (e.g., climbing and jumping, Fig. 2). Although it is recommended for the sampling frequency to be at least twice the frequency of the fastest behavior classified [14, 42], which is the case here (0.217 s for jumping), these behaviors might be too short compared to the 1-s time window chosen for the running mean and fail in providing the model with enough variation in the decision variables.

On the contrary, running behaviors, albeit displayed sometimes just as rarely (0.28% for galloping), showed high recall and precision in classification. These behaviors have longer duration and present more distinctive features (e.g., acceleration patterns or ranges of ODBA, see Figs. 2, 4) than comparatively more poorly classified behaviors. The overlaps in pitch/ODBA ranges between static behaviors or rare behaviors such as chewing, sniffing and digging (Fig. 4g, h, k) seem correlated to the confusions observed in classification (Fig. 3A). The low prevalence and lack of clear characteristics in behaviors such as shaking or jumping (Fig. 4i, l) is correlated to poor recall. On the other hand, a clear distinction between ODBA ranges is visible in the better-performing locomotion behaviors. We theorize that, while higher prevalence improves classification power by increasing the available information on behavioral variation during training [47], a longer duration of the target behavior and more distinctive or stereotyped features might have a similar effect to that of a larger sample size. While classification of highly variable behaviors will benefit from larger sample sizes, classification of short behaviors could be improved by combining a classifier with automated detection relying on techniques such as Fast Fourier transform to filter out small body movements [29], KNN clustering identification [5], or techniques such as changepoint analysis [9] or custom Boolean indicators [98] to better detect significant shifts in parameters pointing towards a change in behavior.

Activity models

As the 300-s duration of activity intervals largely exceeds our recommended duration threshold, we did not expect to be able to predict short behaviors with high accuracy from this dataset. Activity models, however, remain important in that they allow us to make the most of long-term data series, even when data collection started with lower-resolution equipment or when resources such as money, time, battery or storage are limited. This makes interpretation and comparison possible between different types of data, equipment and protocols. As grouping behaviors will affect classification performance and behavior inference [47], we chose to retain the simplest but most performant activity model (Model B), able to adequately classify locomotion states versus static states, as well as another model with clear differentiation between behaviors of ecological interest with the highest potential for reliable classification (Model C, Table 1). Provided that predictions from this last model are backed-up with sensible, contextual interpretation, they can allow for the detection of feeding events from chewing behavior, and for the detection of potentially stressful or physically demanding events from running behaviors. Similarly, we chose to keep the full fine-scale model (Model A) for applications in free-ranging animals and future improvements rather than performing extensive grouping to improve recall scores (Model A.1, Table 1).

Generalizability

Behavior classification from biologging in free-ranging animals is inevitably confronted with several caveats. First, the definition of a behavior is rather fluid: prior to model training, we defined and classified only non-overlapping behaviors to maximize classification power. However, many behaviors are highly variable in their amplitude, duration, or pattern [18] and are often displayed with additional organic and individual variability, such as partial displays or mixing (e.g., walking and sniffing simultaneously). Realistically, acceleration signatures are also affected by the placement and type of accelerometer device used (e.g., a collar will be biased towards head movements and might be more mobile compared to devices attached to the animal’s core such as backpacks or body-mounted tags), the transition between different behaviors, as well as by variations in environmental features. We stress that the data used in this study were recorded in a captive setting, from wolves provided with deer carcasses and kept in flat enclosures with no vegetation and little enrichment (mostly concrete pipes). Notably, behavior prevalence reflects overall low activity levels in captive animals, feeding events, and patrolling of the enclosure, the most common behaviors displayed by social canids in captivity [19, 32, 84]. Consequently, our models lack information on movements influenced by features such as vegetation, topography, environmental obstacles (e.g., rocks or logs), movements through different media (e.g., water or snow), and entire behavior sequences (e.g., hunting and killing of prey). This can lead to poorer performance when using the model on unseen data. Knowledge about wolf biology, environment, as well as the likely behavior confusions as quantified in this study, are key to ensure proper interpretation of behavior classifications in situ.

From the classifier itself, it would be useful to, for instance, extract true class probabilities in terminal leaves and base decisions on the target behavior distribution compared to confused behaviors through a more quantitative approach [21, 72, 101]. Another solution to bridge the gap between model training on captive data, and predictions on unlabeled free-ranging data, is to use the information obtained from models built in captive studies to guide parameter selection, or interpret and generally validate behavior classification obtained from clustering with unsupervised machine learning approaches or Hidden Markov Models [3, 16, 17, 44, 50, 77].

Improvements and recommendations for future work

Continuous development and optimization of existing models in ecology, conservation, and monitoring are an integral part of model building frameworks. In the case of the present study and in behavior classification models in general, increasing the training sample with new comparable data from new individuals and rare behaviors would provide useful information on behavioral variation to improve model recall. Given the vast and variable datasets collected with accelerometer tags available today, it is even more important to carefully plan, select and tailor protocols to the research question, the study species, environment, device location on the body, type of data available, and potential for future studies [81, 82, 92]. Protocols might also be improved, and model validated by using variables derived from other sensors [96], such as conductivity [65], body temperature [43], depth [16], or even footage from collar-mounted cameras or acoustic recordings from embedded microphones. As previously stated, with typically low sample sizes, a supervised model will most likely be undertrained on rare behaviors and on environmental and individual variation, resulting in poorer prediction power in free-ranging settings. Such caveats can only be reliably estimated by multiple trainings on unlabeled data, however, it is noticeable that only training and validation results are consistently reported in the literature. We therefore argue that such models should be systematically cross-validated to avoid overestimation of performance in free-ranging settings, and that individual cross-validation results should be discussed when evaluating performance [4, 44, 100], as they allow an estimation of the potential bias caused by individual variation and can give perspective on the real predictive ability of a model.

Conclusions

Automated behavior identification in free-ranging animals, both in real-time and in delayed studies, has demonstrated usefulness and potential throughout the past decades [7, 39, 40, 48, 87, 88]. In this study, we were able to develop two behavior identification models for wolves, returning adequate classification for common behaviors when tested on captive data. These models can be used to predict behaviors in free-ranging wolves, provided attention is given to the likeliest behavior confusions, especially in the behaviors for which these models show poor predictive ability. It is also necessary to remain aware of the variation in prediction reliability introduced in free-ranging data. The impact of such limitations can be mitigated using contextual clues and knowledge on wolf biology, or the use of complementary models tailored specifically to target more efficiently specific rare behaviors of interest. In further refinement studies, complementary algorithms such as unsupervised clustering models could be used to more adequately identify similar acceleration signatures in highly variable free-ranging situations.