Cheetah (Acinonyx jubatus)- Data, Pictures & Videos

[Acinonyx sp.]

Categorising cheetah behaviour using tri-axial accelerometer data loggers: a comparison of model resolution and data logger performance


Abstract
Background
Extinction is one of the greatest threats to the living world, endangering organisms globally, advancing conservation to the forefront of species research. To maximise the efficacy of conservation efforts, understanding the ecological, physiological, and behavioural requirements of vulnerable species is vital. Technological advances, particularly in remote sensing, enable researchers to continuously monitor movement and behaviours of multiple individuals simultaneously with minimal human intervention. Cheetahs, Acinonyx jubatus, constitute a “vulnerable” species for which only coarse behaviours have been elucidated. The aims of this study were to use animal-attached accelerometers to (1) determine fine-scale behaviours in cheetahs, (2) compare the performances of different devices in behaviour categorisation, and (3) provide a behavioural categorisation framework.
Methods
Two different accelerometer devices (CEFAS, frequency: 30 Hz, maximum capacity: ~ 2 g; GCDC, frequency: 50 Hz, maximum capacity: ~ 8 g) were mounted onto collars, fitted to five individual captive cheetahs. The cheetahs chased a lure around a track, during which time their behaviours were videoed. Accelerometer data were temporally aligned with corresponding video footage and labelled with one of 17 behaviours. Six separate random forest models were run (three per device type) to determine the categorisation accuracy for behaviours at a fine, medium, and coarse resolution.
Results
Fine- and medium-scale models had an overall categorisation accuracy of 83–86% and 84–88% respectively. Non-locomotory behaviours were best categorised on both loggers with GCDC outperforming CEFAS devices overall. On a coarse scale, both devices performed well when categorising activity (86.9% (CEFAS) vs. 89.3% (GCDC) accuracy) and inactivity (95.5% (CEFAS) vs. 95.0% (GCDC) accuracy). This study defined cheetah behaviour beyond three categories and accurately determined stalking behaviours by remote sensing. We also show that device specification and configuration may affect categorisation accuracy, so we recommend deploying several different loggers simultaneously on the same individual.
Conclusion
The results of this study will be useful in determining wild cheetah behaviour. The methods used here allowed broad-scale (active/inactive) as well as fine-scale (e.g. stalking) behaviours to be categorised remotely. These findings and methodological approaches will be useful in monitoring the behaviour of wild cheetahs and other species of conservation interest.

Background
Global biodiversity loss is one of the biggest crises currently threatening the natural world [1,2,3]. Approximately 40% of plant species, 23% of invertebrates, and 18% of vertebrates assessed are considered to be threatened [4]. For mammals, 22% of described species [4] and 26% of assessed carnivores are considered to be threatened [5]. Some of the primary threats to carnivores include reductions in prey [6] and habitat [7], human-wildlife conflict (primarily in terms of livestock losses) [8,9,10,11], and illegal trade in animals or animal parts [11,12,13,14,15]. Conservation is therefore at the forefront of policy-making decisions worldwide [16,17,18,19].
To implement effective conservation and species management strategies, an understanding of target species populations, ecology, and behaviour is important to provide an insight into the status of the species, its needs, and putative causes of decline. When monitored over time, population data can indicate trends in particular groups of animals or in species as a whole [20, 21] or the efficacy of conservation efforts in comparison to control areas [22]. Remote sensing technologies such as camera traps have aided in species population assessments [23,24,25,26], as well as in our understanding of species ecology, and behaviour [23]. Advances in Global Positioning System (GPS) devices have further contributed towards understanding species ecology by providing insights into movements and habitat use. Knowledge of behaviour in space and time can provide insights into the importance of particular habitats and microhabitats for a species. For example, Wege et al. [27] identified novel foraging sites used by fur seals, contributing towards conservation policy-making as these sites were heavily utilised during the winter and were not previously considered when making assessments for potential marine protected areas (MPAs), where summer use is considered to be more important.
Accelerometer data loggers have been used independent of [28, 29] and in combination with [30,31,32] other remote sensing technologies such as GPS devices, magnetometers, and gyroscopes. Tri-axial accelerometers measure acceleration in three orthogonal axes (heave, surge, and sway), providing information on omnidirectional dynamic movement of an animal, as well as its posture (via static acceleration) [33, 34]. When accelerometers are used alongside devices such as GPS loggers, detailed behaviour patterns in space and time can be elucidated (e.g. [35, 36]). Unlike other remote sensing technology such as camera traps, loggers are fitted to the animals of interest (either directly or via collars or harnesses), providing data on the individual for the entire deployment period, not simply when activated. This feature is particularly useful for assessing the behaviours of cryptic species with large home ranges or that utilise difficult-to-monitor habitats (e.g. dense forests/bush, burrows, or expansive deserts). Although, the relative affordability and ease with which loggers can be deployed has led to their widespread use, less consideration appears to be given to device selection and subsequent downstream data processing. Most applications of animal-borne accelerometers have been to examine behaviours (e.g. [31, 37,38,39,40]), with several resulting in the categorisation of coarse-scale descriptions (i.e. three or four different behaviours) [28, 33, 40, 41]. While several studies have categorised behaviours manually by coarsely examining the acceleration traces generated (e.g. [33, 42]), others have implemented machine-learning techniques (many described in [43]), including random forests (RFs) [29, 31, 37,38,39,40, 43, 44], to classify behaviours to datasets using training and test data. Other approaches, such as the use of magnetometers, have proven successful in the determination of specific behaviours (e.g. biting and chewing in grazing herbivores) [45].
Cheetahs (Acinonyx jubatus) are medium-large felids inhabiting Africa and Iran [46,47,48]. They are classified as ‘Vulnerable’ by the IUCN [46] with the most recent population assessment (2014) suggesting just under 7100 adolescent and adult cheetah remain in the wild [47]. Population strongholds exist in southern and eastern Africa [46, 47]. Whilst conservation measures such as confiscation of traded animals and parts and reducing conflict with humans have been put in place, cheetah populations continue to decrease, with habitat loss, persecution, and illegal hunting and trade comprising major threats [7, 11, 47]. As such, detailed monitoring of cheetah movements, habitat use, and behaviour can assist with conservation efforts to ensure stringent monitoring of frequently used areas to reduce poaching and the adequate provision of resources to meet the needs of the species. To date, only coarse behaviours (active, inactive, and feeding) have been defined for cheetahs using remote sensing technology (accelerometers) [33, 41]. However, other ecological information such as different hunting strategies they may adopt and the associated costs of chasing prey [30, 32, 49, 50] (using GPS and accelerometers) have also been elucidated. However, while fine-scale behaviours, such as stalks (which may not result in a hunt), different movement gaits (e.g. walking vs. sprinting), and resting, have yet to be described for cheetahs, such data are available in other species (e.g. [31, 37,38,39, 43, 44]). Cheetahs are considered to be “extreme” movers, potentially reaching top speeds of 64mph (103kph) in a matter of seconds [51]. Therefore, the ability to distinguish between fine-scale behaviours may help to define the ecological needs of cheetahs, including hunting success rate, and, thus, contribute to conservation efforts. However, due to the high power and accelerations attained by cheetahs, monitoring their behaviour remotely may be limited by the capacity of individual devices.
The overall aim of the current study was to ground-truth behaviours performed by cheetahs against data collected using tri-axial accelerometers. Specifically, we wanted to (1) determine the accuracy with which a suite of behaviours in a cheetah’s repertoire could be defined; (2) determine whether this could be affected by the technical specifications of two different accelerometer devices, and; (3) provide a framework in the form of a vignette containing “R” code to develop behaviour categorisation models for other species of policy or conservation interest.
Methods
Study animals and collar preparation
This study was carried out in October 2012 at the Cheetah Conservation Fund (CCF) research centre near Otjiwarongo, Namibia (− 20.447763° N, 16.677918° E). Five resident adult cheetahs (three males and two females) were fitted with their own neck collars (nylon dog collars with plastic clip buckle: mass = 75 g, length = 570 mm, width = 20 mm) equipped with two tri-axial accelerometer data loggers: 1. G6, CEFAS Technology Limited, Lowestoft, UK (maximum = 2.3 g, size = 40 × 28 × 15 mm (L × B × D), mass = 18 g including urethane encasement, recording frequency = 30 Hz); 2. X8M-3, Gulf Coast Data Concepts (GCDC), LLC, Waveland, MS, USA (maximum = 8.6 g, resolution = 0.001 g, size = 50 × 30 × 12 mm (L × B × D), mass = 21.6 g including epoxy encasement, recording frequency = 50 Hz). To ensure the collar remained centred on the ventral side of the neck, an additional weight comprising four steel nuts (120 g) was added. The total weight of the fully equipped collars was approximately 235 g (see Additional file 3: Figure S1a for constructed collar design). Prior to being fitted to cheetahs, collars were hung on a metal rail with the accelerometers located at the bottom of the collar to allow for the devices to be calibrated (see “Data processing—accelerometers” below).
Exercise arena and video capture
Cheetahs were exercised by chasing a lure (cloth rag) attached to ~ 285 m of cord around a pre-determined track. The lure machine, powered by an electric motor, was remotely controlled by a keeper, such that the speed and direction of the lure could be altered at will. The keeper changed the direction of the lure strategically to attempt to outwit the chasing animals and prevent capture of the lure. The chasing animals were thereby encouraged to employ different strategies to try to catch the lure, including stalking behaviour and high-speed pursuits. Each cheetah was exercised individually and behaviour was recorded using a video camera (Canon PowerShot SX230 HS; Canon, Japan). Typically exercise bouts lasted 10–15 min and consisted of three or four active chases (e.g. running, stalking) punctuated by two or three lower intensity rest periods (e.g. lying down, walking, standing). Collars were retrieved when the animal had finished exercising. As exercise bouts comprised periods of activity and inactivity, data associated with both hunting and resting were collected and ground-truthed against video footage.
Data processing—accelerometers
Following exercise bouts, data loggers were removed from collars and data were downloaded. The data collected for both devices were calibrated to correct for non-centred mounting of the devices on the collars using the region of the dataset where the collars had been attached to the metal rail (see Additional file 1: Study details, collar calibration, and calculations). The data corresponding to the times of captured video footage were selected and the rest of the data were removed. Static acceleration (acceleration due to gravity; Additional file 5: Figure S2, static acceleration diagram) was derived for each axis from the corrected heave (acceleration in vertical axis), surge (acceleration in longitudinal axis), and sway (acceleration in transverse axis) data by calculating a rolling mean over a two-second window [52]. Dynamic acceleration was then calculated for each axis as the absolute result of subtracting static acceleration for a particular axis from its raw acceleration. Vectorial Dynamic Body Acceleration (VeDBA), Vectorial Static Body Acceleration (VeSBA), animal static acceleration (Anim.stat), pitch, and roll were also determined (Additional file 1: Study details, collar calibration, and calculations).
Data processing—video footage
All video footage (approximately 58 min; 103,869 CEFAS logging events; 174,185 GCDC logging events) was synchronised with its complementary accelerometer datasets. Video footage was assessed frame-by-frame (Avidemux software; Developer: Mean) and cheetah behaviour was matched with the accelerometer data. Initially, 22 behaviours and behaviour combinations were identified (Table 1). Any other behaviour was recorded as ‘other’ and instances where behaviour could not be assigned (e.g. if an object obstructed a clear view of the animal) were removed from the dataset as we could not be certain of categorisation, resulting in a loss of approximately six minutes’ worth of data. Each labelled dataset was amalgamated to give two master spreadsheets of labelled accelerometer data; one for each model of accelerometer device (CEFAS and GCDC).

Data analysis

Data analysis was carried out in ‘R’ version 3.4.3 [54] using the ‘h2o’ package version 3.16.0.2 [53]. RF analysis (Additional file 2: Code) was conducted on the datasets labelled with behaviours. The datasets were split into three, such that 60% of cases were selected at random to entrain models (training dataset), 20% of cases were selected at random to validate the model (validation dataset), and the remaining 20% were used to test model performance (test dataset). The training data were used to entrain the RF model to categorise specific behaviours (see Table 2 for behaviour list). The validation dataset was then used to assess the performance of the model via model accuracy, (root) mean square error (RMSE and MSE), and r2. The validation data were also used to refine the model by altering model parameters and comparing the metrics listed above. The test data were only used once at the end of the process to compare model accuracy after validation to the outputs of the training dataset.


Model structure
Initially, models were entrained to categorise 17 behaviours (Table 2, fine-scale). The predictor variables were: heave, surge, sway, static heave, static surge, static sway, dynamic heave, dynamic surge, dynamic sway, VeDBA, VeSBA, Anim.stat, pitch, and roll. A stopping criterion (stopping-rounds = 2) was implemented to optimise the duration for which models were run. A stopping criterion of two stops fitting the model when the two-tree average is within 0.1% accuracy of the previous two-tree average. If this criterion is increased, the average is taken over the specified number of trees. Models were refined by changing their depth and comparing their overall accuracy (percentage of correctly categorised behaviours divided by percentage of incorrectly categorised behaviours). The model with the highest accuracy was retained. Models were re-run using coarser behavioural categories (Table 2). For each model, cross-validation was performed using five folds and comparing mean accuracy, RMSE, MSE, and r2 to the training dataset. ‘R’ code for RF model constructs and additional model information are provided in the supplement (Additional file 2: Code). Logger performances were compared for the categorisation of each individual behaviour using chi-squared tests.



Results

There was no indication of significant overfitting when cross-validation of models was carried out (Table 3).


CEFAS loggers

In the first model, behaviours were categorised on a fine scale. The behaviours sought to be categorised were: crouch, lie, sit, stand, head movement, crouching stalk, lying stalk, sitting stalk, standing stalk, walking stalk, trotting stalk, walk, trot, canter, gallop, pounce, and other. The overall accuracy of the model was 83.3% (MSE = 0.18, RMSE = 0.42, r2 = 0.99). However, as ‘other’ was an uninformative category, which didn’t require correct positive categorisation in the training dataset as it comprised a ‘rag bag’ of various movements, its categorisation could be disregarded (but the variable still remained in the model). Once disregarded, the accuracy of the model increased to 84.2%. Sitting stalk, lying stalk, lying, and standing were categorised with over 90% accuracy. Behaviours with < 50% categorisation accuracy included pouncing, crouching, trotting, and trotting stalk (see Table 4 ; Fig. 1 for full description of classification accuracy). Crouching behaviour was most often confused with lying (14.5%), other (27.6%), and standing (47.4%). Trotting and trotting stalk were most often confused with cantering (trotting: 21.1%; trotting stalk: 19.4%), galloping (trotting: 6.1%; trotting stalk: 13.2%), other (trotting: 42.2%; trotting stalk: 45.8%), and walking (trotting: 19.7%; trotting stalk: 9.7%). In addition, trotting stalk was confused with walking stalk (6.3%) (Fig. 2A). In terms of predictor variables, static acceleration in all three axes was most important in categorising behaviours (heave: scaled importance (improvement of MSE relative to maximum improvement across all predictors) = 100%, explanatory power = 14.4%; sway: scaled importance = 80.0%, explanatory power = 11.5%; surge: scaled importance = 71.1%, explanatory power = 10.2%), followed by VeDBA (scaled importance = 53.3%, explanatory power = 7.7%), roll (scaled importance = 52.0%, explanatory power = 7.5%), and heave acceleration (scaled importance = 51.6%, explanatory power = 7.4%). In all, these six predictors explained 58.7% of the RF model variance.

In the second, coarser model, several behaviours from the previous model were combined in an attempt to reduce the error rate. ‘Pounce’ was entered as ‘other’ as it could not be categorised reliably and was often confused with several other behaviours. Behaviours in this model included: Sedentary (‘crouch’, ‘lie’, ‘sit’, ‘stand’), head movement, moving stalk (‘trotting stalk’, ‘walking stalk’), crouching stalk, sitting stalk, lying stalk, standing stalk, gallop, canter, trot, and walk. The overall accuracy of this model was 84.7% (MSE = 0.16, RMSE = 0.40, r2 = 0.98), which increased to 86.6% when ‘other’ behaviours were removed. Sedentary, sitting stalk and lying stalk were the only behaviours where the prediction accuracy surpassed 90%. The prediction accuracy for two behaviours was lower than 50%; cantering and trotting (see Table 4 and Fig. 1 for full description of classification accuracy). Cantering was most often confused with galloping (43.8%), other (34.4%), and trotting (8.0%), while trotting was most often confused with cantering (20.1%), other (40.9%), and walking (20.8%) (Fig. 2B). Once again, static accelerations (heave: scaled importance = 100%, explanatory power = 13.0%; surge: scaled importance = 87.8%, explanatory power = 11.5%; sway: scaled importance = 76.2%, explanatory power = 9.9%), VeDBA (scaled importance = 73.7%, explanatory power = 9.6%), and VeSBA (scaled importance = 50.8%, explanatory power = 6.6%) featured among the important predictor variables. Overall, the top five predictors explained 50.7% of the RF model variance.

The final, coarsest model comprised a simplified RF where behaviours were either deemed to be active (‘walk’, ‘trot’, ‘canter’, ‘gallop’, ‘walking stalk’, ‘trotting stalk’, ‘pounce’), inactive (‘crouch’, ‘lie’, ‘sit’, ‘stand’, ‘crouching stalk’, ‘lying stalk’, ‘sitting stalk’, ‘standing stalk’), head movement, or other. The accuracy of this final model was 88.2% (MSE = 0.11, RMSE = 0.32, r2 = 0.89). When ‘other’ was removed, model accuracy increased to 92.7%. Inactivity was predicted to the highest degree of accuracy and head movement with the lowest. Activity was predicted with 86.9% accuracy (see Table 4; Fig. 1 for full description of classification accuracy). Whilst head movement was most often confused with all three behaviour categories; inactivity (47.3%), activity (28.8%), and other (24.0%), activity was most often confused with inactivity (58.9%) and ‘other’ behaviours (40.2%) (Fig. 2C). The most important predictors in this model were VeDBA (scaled importance = 100%, explanatory power = 15.1%), static acceleration in the heave axis (scaled importance = 84.1%, explanatory power = 12.7%), and dynamic acceleration in the sway (scaled importance = 63.0%, explanatory power = 9.5%), heave (scaled importance = 56.8%, explanatory power = 8.6%), and surge (scaled importance = 52.5%, explanatory power = 7.9%) axes as well as static acceleration in the surge axis (scaled importance = 50.4%, explanatory power = 7.6%). In total, the top six variables explained 61.6% of the model variance.

GCDC loggers
The models outlined above were repeated for the GCDC data loggers. The model containing the finest-scale behaviours was 85.5% accurate (MSE = 0.17, RMSE = 0.41, r2 = 0.99). Accuracy increased to 85.8% when the category ‘other’ was omitted. The sedentary behaviours of lying, lying stalk, and sitting stalk were categorised with > 90% accuracy. Crouching was the only behaviour that was categorised with < 50% accuracy (see Table 4 and Fig. 1 for full description of classification accuracy); it was most often confused with standing (41.1%) and other behaviours (41.1%) (Fig. 3A). Static acceleration in all three axes was the most important predictor for behaviour (heave: scaled importance = 100%, explanatory power = 16.0%; sway: scaled importance = 80.1%, explanatory power = 12.8%; surge: scaled importance = 72.8%, explanatory power = 11.6%). In total, static acceleration variables explained 40.4% of the model variance.

When the second model outlined above was reproduced for the GCDC logger, it had an accuracy of 86.2% (MSE = 0.15, RMSE = 0.39, r2 = 0.98), which increased to 87.4% when the ‘other’ category was removed. Behaviours categorised with > 90% accuracy were sedentary, lying stalk, and sitting stalk. No behaviour categorisation was < 50% accurate; those behaviours that were most difficult to categorise were trotting, cantering, and head movement (see Table 4, Fig. 1 for full description of classification accuracy, and Fig. 3B for confusion matrix). In terms of predictor variables, static acceleration in all axes (heave: scaled importance = 100%, explanatory power = 14.7%; surge: scaled importance = 78.6%, explanatory power = 11.6%; sway: scaled importance = 78.2%, explanatory power = 11.5%), VeSBA (scaled importance = 51.8%, explanatory power = 7.6%), and Anim.stat (scaled importance = 51.5%, explanatory power = 7.6%) were most important in determining behaviours in this model. The top five predictor variables explained 53.1% of the model variance.

In the final RF model for the GCDC data loggers, activity, inactivity, head movement, and other behaviours were categorised. This model performed with a categorisation accuracy of 90.2% (MSE = 0.09, RMSE = 0.31, r2 = 0.90), which increased to 92.9% when the ‘other’ behaviour category was omitted. Inactivity was most easily classified and no behaviour had a classification accuracy below 61.3% (see Table 4, Fig. 1 for full description of classification accuracy, and Fig. 3C for confusion matrix). In this model, static acceleration in the heave (scaled importance = 100%, explanatory power = 12.7%) and surge (scaled importance = 84.8%, explanatory power = 10.8%) axes were most important for categorisation, followed by VeDBA (scaled importance = 83.2%, explanatory power = 10.6%), static acceleration in the sway axis (scaled importance = 75.8%, explanatory power = 9.6%), Anim.stat (scaled importance = 61.1%, explanatory power = 7.8%), VeSBA (scaled importance = 57.4%, explanatory power = 7.3%), and dynamic acceleration in the heave (scaled importance = 52.0%, explanatory power = 6.6%) and sway (scaled importance = 52.0%, explanatory power = 6.6%) axes. Together these eight variables explain 79.1% of the model variance.

Logger comparison
Generally, the higher capacity loggers, with higher recoding frequency (GCDC) outperformed their lower capacity (CEFAS) counterparts when determining cheetah behaviour (Table 5; Fig. 1). This was particularly evident during medium- and fine-scale behaviour categorisation. Of the 30 different behaviour-model combinations run, the CEFAS loggers significantly outperformed the GCDC loggers only four times: standing (χ2 = 11.63, df = 1, p < 0.001) and galloping (χ2 = 19.46, df = 1, p < 0.001) in the fine-scale model, galloping (χ2 = 23.15, df = 1, p < 0.001) in the medium-scale model, and inactivity (χ2 = 4.42, df = 1, p = 0.035) in the coarse model (Table 5; Fig. 1). Conversely, the GCDC loggers were significantly better than the CEFAS loggers at behaviour categorisation on 11 occasions (Table 5; Fig. 1), notably when defining trotting, moving stalks, head movement, pouncing, and, in the fine-scale model, most sedentary behaviours. Whilst the GCDC loggers were better overall at defining behaviours on a fine- and medium-scale, there was no significant difference between the two devices when categorising behaviours on a coarse-scale (Table 5).

*This image is copyright of its original author

Reply