Livestock grazing, whether in intensive (feed lots and pastures) or extensive (free-range) systems is a traditional practice that has persisted in our modern society to cope with increasing food production demands (Michalk et al., 2019; Komarek et al., 2021). Within this tradition, the ever-expanding development of modern technology has allowed for the growth of precision livestock farming management and research (Eastwood et al., 2017). Indeed, this management approach focuses on the fine-scale monitoring of individuals’ health and food intake (Schellberg et al., 2008; Werkheiser, 2018). While it remains relatively easy to implement in barn and pasture settings, free-range farming presents additional challenges; herd supervision can become more difficult as cattle are not contained. Additionally, external factors affecting cattle such as exposure to climatic extremes, parasitic load, untreated diseases, accidents, and potential carnivore effects are complex to monitor (Hutchings et al., 2000; Silanikove, 2000; Sevi et al., 2009; Nedeva, 2020). Technological advances such as biosensors, camera-equipped drones and GPS collars have offered scientists, managers and farmers tools to address those challenges (Herlin et al., 2021).

Notably, modern GPS collars, which often contain accelerometry sensors, provide researchers with an opportunity for non-invasive, low maintenance and remote monitoring of livestock and wildlife. If properly calibrated, this allows for the study of fine-scale animal behavior, activity budgets and energy expenditure on an individual level (Vázquez Diosdado et al., 2015; O’Leary et al., 2020; Mulvenna et al., 2022). Even though accelerometry data has largely increased our understanding and knowledge of livestock behavior (Theurer et al., 2013; Uenishi et al., 2021), most behavioral studies using accelerometry sensors focus on a few key behaviors such as grazing, resting, and walking (Robért et al., 2011; Homburger et al., 2014; Benaissa et al., 2019), and they are conducted in controlled settings such as barns and pastures (Homburger et al., 2014; Hendriks et al., 2020; Rodriguez-Baena et al., 2020).

Very few studies examine the behavior of free-ranging cattle in remote areas, where cattle are difficult to monitor and observe due to them moving over large areas and using dense vegetation types (Tofastrud et al., 2019). Yet, in these conditions, behaviors such as vigilance, social behaviors or grooming can be representative of stress, as cows have been reported to increase vigilance when stressed (Welp et al., 2004), and to decrease milk production when separated from the herd and not allowed to perform social behaviors (Hedlund and Løvlie, 2015). In addition, self and allogrooming are reported to be frequent as maintenance behaviors and shown to be an important proxy for welfare (Kohari et al., 2007). While these behaviors are important to monitor, they are especially difficult to observe. Tofastrud et al. (2018) used two-axial accelerometry data at five-minute intervals on free-ranging cattle to study resting, grazing and movement activity patterns. Although this allows for insight into general free-ranging cattle behavioral habits in remote areas, it lacks the ability to precisely quantify additional, rarer behaviors and restricts the amount of information capable of being calibrated and further studied. Continuous, high resolution (10Hz) tri-axial accelerometry data can potentially increase the number of behaviors that can be classified and provide more detailed information about cattle behavior in large rugged, forested ranges (Hounslow et al., 2019). Such studies could contribute to the improvement of farmer monitoring systems related to high precision farming in outfields and offer the possibility to study effects from external factors, such as carnivores, on cattle behavior while monitoring welfare.

There are numerous ways to analyze accelerometry data, ranging from simple decision trees to complex neural networks (Riaboff et al., 2022). Most studies utilize supervised machine learning methods such as random forest (de Weerd et al., 2015; Williams et al., 2020; Riaboff et al., 2022), while others use unsupervised machine learning such as hidden Markov Models (Leos-Barajas et al., 2017; Chimienti et al., 2021; Rautiainen et al., 2022). Supervised methods provide the advantage of allowing for accelerometry data to be calibrated on actual behavioral observations, which then allows for prediction of behaviors based on collected data.

Feature selection is the first step in this process. Features can be calculated and extracted from raw accelerometry data and be used for behavioral classification. Orientation dependent features, such as the mean and variance of the raw x, y, z values, body pitch roll, yaw and dynamic acceleration ( Figure 1 ) remain sensitive to the sensor’s orientation (Abell et al., 2017; Benaissa et al., 2019). For example, even though collars are assumed to be stationary positioned on the animal, there can be noise related to rotation, collar deployment errors, and other causes. This may result in additional variability in orientation-dependent features, which may render these features unusable without correcting for orientation (Williams et al., 2017; Barker et al., 2018; Kamminga et al., 2018). This sensitivity can be challenging when standardizing accelerometer sensor placement during animal handling, and recapturing individuals to manually fix issues is difficult (Chakravarty et al., 2019; Cade et al., 2021; Rautiainen et al., 2022). To remedy this problem, orientation independent features can be utilized. Models can then account for displacement of the sensor’s orientation and make obsolete the need to correct for orientation.

High resolution (> 1Hz) accelerometry data is usually smoothed with a running average over a given time window (Shepard et al., 2008a). The chosen window length can result in the loss of certain less frequent and shorter behaviors that might not be detectable, while longer lasting, more common behaviors increase the accuracy and precision of the predictions (Mansbridge et al., 2018; Chang et al., 2022). Therefore, clear study goals become essential in accelerometry analysis, as the study aim will determine different trade-offs and decisions for modeling (Chang et al., 2022).

For instance, some authors suggest that averaging values in the sensor provides an opportunity to increase data collection capacity, and thus allows for live monitoring of behaviors with computationally simple and cost-efficient features and algorithms, which reduce battery usage (Kamminga, 2020; Nuijten et al., 2020). Study questions will thus determine data collection methods, which resolution the data needs to be collected in, as well as which steps should be included in the pre-processing. This will induce trade-offs which will determine what type and number of behaviors that can be analyzed (Kamminga et al., 2018; Riaboff et al., 2022).

In this paper, we aim to calibrate high-resolution accelerometry data collected by commercially available livestock collars deployed on free-ranging cattle in remote areas. Additionally, we aim to investigate the effect of collar placement during deployment of collar and how it might affect accelerometry data. Finally, we attempt to address the gap in knowledge concerning the classification of less frequent cattle behaviors to accelerometry data, as these behaviors can represent behavioral changes or even be indicators of stress.

Our classification study contributes to potential research on cattle social interactions, behavioral responses to carnivores, and energy expenditure of free-ranging cattle in remote forested areas. Additionally, this study has the potential to develop tools for improved monitoring systems for farmers, and therefore to contribute to the practice of agroforestry and precision livestock farming.

We first hypothesized that the model performance to predict general behaviors such as walking (locomotion), foraging, vigilance, standing, laying/resting, or ruminating, would be affected by the choice of features included in the models (orientation-dependent versus orientation-independent features) and by the direction of the collar position on the neck of the animals (as we wanted to account for collar placement variation by the farmers on cattle) (H1). Models with orientation-independent features might perform better than those using orientation-dependent features, as orientation-independent features can account for potential rotation or tightness differences of collars when individuals navigate through rugged terrain. Additionally, models including orientation-dependent features corrected for direction might perform better than those without correction.

Secondly, we hypothesized that less frequent behaviors such as social interactions and body care movements would be impacted by the smoothing of the data (H2). Data smoothed with a short running mean would allow for the detection of rarer behaviors, but at the cost of a loss in the overall model’s performance as a short running window mean can induce more noise during the analysis.

We annotated 6 898 behavioral observations based on video analysis and matched them with accelerometry data, resulting in a total of 1 240 588 observations for 31 behaviors ( Supplementary Table 2 ). Visual inspection showed appropriate matching, i.e. the accelerometry data showed different signatures for different behaviors ( Figure 3 ).

Overall, the models with 20 seconds running mean performed better than those with shorter running means ( Table 2 ). Orientation-independent models showed a lower performance than orientation-dependent models. Corrected orientation-dependent models performed similar to uncorrected models ( Table 2 ). Both orientation-dependent models with 20 seconds running mean (models C and F) had average accuracy, precision and recall > 0.96. The orientation-independent model I had an average precision of 0.88, recall 0.93 and accuracy 0.99 ( Table 2 ). Precision averaged 0.62 and 0.63 for the 1 second running mean in the orientation-dependent and corrected orientation-dependent models, respectively (models A and D, Table 2 ), and averaged 0.37 and 0.64 for the 1 second and the 5 seconds running means, respectively in the models with orientation-independent features (models G and H, Table 2 ). Similarly, the recall averaged 0.69 for the 1 second running mean in the orientation-dependent and corrected orientation-dependent models (models A and D, Table 2 ). The orientation-independent models G and H had an average recall of 0.39 and 0.72 for 1 and 5 seconds running mean, respectively ( Table 2 ). Prediction performance varied across behaviors for recall, accuracy and precision ( Figure 4 , and Supplementary Figures 2 , 3 , respectively). Behaviors in models with a shorter running mean varied more in recall than in models with a longer running mean, e.g., ‘foraging_high’ and ‘other’ in the orientation-dependent models A and D were below 0.4 ( Figure 4 ). Those behaviors were correctly classified in less than 40% of their occurrence. Orientation-dependent models using 5 seconds running means performed better with lowest recall of 0.8 (model B, Figure 4 ). For models with 20 seconds running mean, the recall was never below 0.96, showing that the lowest performing behavior had a classification success of 0.96. The performance of the orientation-independent models showed overall lower recall values for each behavior. Still, increased running means in the orientation-independent models helped to improve prediction and decreased the number of classification errors.

The orientation-dependent model with 20 seconds running mean had a few occasions where behaviors were misclassified ( Table 3 ). However, the confusion matrix for the orientation-independent model with 20 seconds running mean showed a higher degree of misclassification ( Table 4 ), especially for behaviors that were more similar in their acceleration signature (e.g. ‘foraging_low’ and ‘walking’, or ‘ruminating_standing’ and ‘ruminating_laying’). The confusion matrices for all other models are presented in Supplementary Tables 4–10 .

The individual-based cross validation model using a 20 second running mean and orientation-dependent data indicated that rarer behaviors were more difficult to predict for individuals with less observations, but overall, the model performed similar to the model using the 80-20 data split, because there were enough observations of these rare behaviors across all individuals. The averaged accuracy was 0.997, the averaged precision was 0.965, and the averaged recall was 0.987 (See Supplementary Table 11 for the confusion matrix).

The model including more behaviors had an average accuracy of 0.99, precision of 0.71, and recall of 0.96. The classification success for the behaviors that were also present in the other models performed similarly well ( Tables 3 – 5 ). However, rarer behaviors had a high classification success with almost no misclassifications, e.g., in the behaviors ‘stretching’ and ‘throw_head’ ( Table 5 ).

In this study, tri-axial accelerometer signatures were assigned to observed behaviors of free-ranging cattle based on supervised machine learning algorithms and using different accelerometer features and running mean smoothing windows. We found that (1) using a long running mean (20 seconds) translated to best model performance across all model categories, and (2) model performance remained excellent when using orientation-dependent instead of orientation-independent features, or when adding more behaviors (with accurate classification even with minority class behaviors i.e., allogrooming, suckling calf).

How tight and in which direction collars were deployed, and terrain ruggedness through which the cows navigated, varied widely in our study. We therefore expected that orientation-independent features derived from accelerometer data, would lead to better prediction performance than orientation-dependent features (hypothesis H1). However, models based solely on orientation-dependent features performed better than those based on orientation-independent features.

While our study assessed the outcome and performance of either orientation-dependent or -independent features, previous studies have combined up to 60 features of both types to increase model performance (Lush et al., 2018; Riaboff et al., 2020; Riaboff et al., 2022). However, as our best model with solely orientation-dependent features had excellent performance (0.997 accuracy, 0.961 precision, 0.985 recall), we did not need to extend the model by including additional orientation-dependent and -independent features. We believe our model performed so well in part because the accelerometer was placed along with the battery and other collar electronics on the low side of the collar. The combined weight of the unit seems to hold the accelerometer in place, independently from collar tightness. In comparable studies on marine mammals, accelerometer placement is highly variable (Shepard et al., 2008b).

Moreover, and contrary to our predictions, we found that correcting for collar orientation did not improve the model performance. Indeed, both corrected and uncorrected models performed similarly. While placing the accelerometer backwards can impact axes values, the accelerometer unit remains below the neck of the animal and hangs similarly for all individuals. In fact, backwards placement of the accelerometer affects only the mean of the X and Y axes, with a more pronounced effect for the X axis, as well as the mean pitch values; the variance of the Z axis and the pitch are not affected by orientation (Wang et al., 2015; Chang et al., 2022). Despite variations in feature values, the amplitude and pattern of movement remain consistent, which might explain why orientation-corrected models did not outperform uncorrected models. Additionally, the sample size in our study is large enough to rule out noise in the data due to collar deployment error, making our models robust (Riaboff et al., 2022). Furthermore, cattle are large and slow animals, and behavior-specific body movements can be more easily differentiated compared to small-sized animals. This might be more challenging for smaller, faster moving species, and placement of accelerometers in those species, and placement of accelerometers in those species is likely more important (Grünewälder et al., 2012; Brewster et al., 2018).

We initially hypothesized that the detection of less frequent behaviors, such as social interactions and body care movements, would be impacted by the smoothing of the data, as these behaviors might not be detected by long running mean windows. Contrary to our predictions, we found an increase in model performance across all categories (orientation-dependent, orientation-dependent with correction, orientation-independent) with increasing smoothing window, with highest performance when using a 20 second running mean. Other studies have shown that smoothing of the data increases the classification success by reducing noise, and that larger animals often require a longer running mean as their movements are generally slower (Shepard et al., 2008a). However, this often results in a decrease in classification success for behaviors which are rare and short in time (Mansbridge et al., 2018; Chang et al., 2022). In our study we did not find such a decrease. An explanation might be that we excluded all behaviors shorter than 5 seconds, resulting in a large enough difference in mean and variance of the features for those behaviors for successful classification. Furthermore, the large number of observations in the minority classes (rarer behaviors) across individuals might have contributed positively during the training of the model, resulting in high model performance.

Finally, we predicted that adding more behaviors would affect model performance negatively and expected that the addition of behaviors would lead to loss of model performance (H2) as it has been shown in previous studies (Vázquez Diosdado et al., 2015; Lush et al., 2018). Contrary to our prediction, we found that models performed well with an increase of behaviors and could accurately predict less frequent behaviors such as head butting, throwing head and shaking. Less frequent behaviors were observed for fewer individuals than more common behaviors, which may lead to a stronger impact of individual-specific accelerometer signatures in the random forest models. We addressed this through class weighting, and individual-based cross validation. Even though there was variation in prediction success across individuals and behaviors, the overall model performed similarly well as with the 5-fold (i.e. with random 80/20 data split) cross-validation model.

This study’s sample size of individuals is larger than most accelerometry classification studies of free-ranging cattle (Chapa et al., 2020; Kour et al., 2021). Riaboff et al. (2022) recommended using a minimum of 10 animals and emphasized more robust analysis with at least 25 animals and a variety of breeds and farms. While we exceeded this recommendation (n = 38 individuals distributed on four breeds and four farms), we did not specifically account for breed, farm or individual characteristics such as body weight or age, as this would require an even larger sample.

Interestingly, we were able to differentiate between behaviors that we expected to have similar accelerometry signatures, such as laying ruminating and standing ruminating. When looking at posture, cattle laying causes the angle of the accelerometer to vary slightly compared to when they are standing, as the electronic housing often leans against the individual’s chest. Similarly, we could identify vigilance behavior, which is likely due to our decision rule defining individuals as being vigilant only when no other behavior happened ( Supplementary Figure 1 ).

Often, scientific studies develop models and tools that are appropriate for experimental settings, but too expensive or impractical to be used for commercial settings. Our results based on the accelerometers contained in the commercial Nofence collars open up for a range of end-user applications. For example, these could be Nofence tools for easy handling by the customers, such as a built-in algorithm in the collar converting accelerometer data directly to behavioral states or to time budget summaries, which could be continuously transmitted to the farmer. This could allow for an easy and fine-scale supervision and monitoring of free-ranging cattle in remote areas or dense habitats. Furthermore, the success in classification of ruminating, vigilance and social behaviors could contribute to the study of free-ranging cattle welfare, stress and productivity, through the quantification of precise nutrient intake and energy expenditure.

In conclusion, our study succeeded in categorizing high resolution accelerometer data into behaviors for free-ranging cattle in rugged terrain of the boreal forest. Not only were collar deployment errors confirmed to not significantly impact model performance, but our models also showed success in detecting more behaviors than previously published, including less frequent behaviors other than resting, grazing, ruminating and walking. Calibrating such data with an array of different behaviors makes a valuable contribution to livestock precision farming in extensive rangeland systems. It may allow farmers to monitor the welfare of their animals continuously and to detect non-normal behaviors caused by e.g. diseases or carnivore attacks.

The data and R scripts to replicate the models of this study are openly available in Dataverse NO at https://doi.org/10.18710/ND4CLL.

Ethical review and approval was not required for the animal study because this study used commercially available GPS/virtual fence collars in Norway and is approved by the Norwegian authorities for use on cattle. Written informed consent was obtained from the owners for the participation of their animals in this study.

EV, LJN, MS, BZ, and ALE conceived the study. BZ and ALE secured the funding. LJN conducted the filming and video analysis. EV and MS conducted the accelerometry analysis, with additional support of OD. EV and LJN drafted the manuscript. MS, BZ, AH, MT, OD, ALE reviewed and commented on the initial drafts. All authors contributed to the ideas and edits to the manuscript and approved the submitted version.