1. Introduction
The quantification of mineral block supplement intake by individual cattle will be valuable for improving efficiency of supplement use in grazing systems [
1]. Exploiting automatic feeders such as GrowSafe
® [
2] or SmartFeed
® [
3] and incorporating chemical markers, such as lithium salts [
4] or fenbendazole [
5] into the mineral block supplements are common techniques used for estimating consumption by individual cattle. However, their use over a long period in a larger herd is considered impractical and technically prohibitive since not every animal has the chance to access to the feeder bin effectively [
3] and the necessity for extensive laboratory analysis of the marker [
5]. Hence, advancement of simpler more immediate methods of estimating supplement intake are required to assist managers in decision-making in order to improve efficiency of beef cattle production systems.
Wireless technology using animal-borne sensors allows individual animals to be physically monitored in real-time without interfering in their natural behaviour [
6,
7]. Tri-axial accelerometers have been routinely deployed to automatically record and classify behaviours of domesticated animals based on the acceleration movements over the three perpendicular axes [
8,
9,
10,
11]. Recent investigations have reported that tri-axial accelerometers were capable of categorising oral and intake behaviours of ruminants such as suckling [
12], ruminating, eating [
13], grazing [
14], chewing, biting [
11], and drinking [
15]. Apart from reducing observation time, the capability of accelerometers to discriminate feeding behaviours indicates the potential for developing algorithms to accurately predict feed intake [
16]. Greenwood et al. [
17] formulated a simple algorithm to predict pasture intake by individual cattle using accelerometers and Williams et al. [
15] reported that accelerometers could be used to predict water intake of grazing cattle based on prediction of visiting frequency and duration per visit to the water trough.
Tri-axial accelerometers have often been affixed to the body parts of beef cattle mainly on the ear (ear-tag) [
18], neck (collar) [
15] and muzzle (halter) [
12]. Several machine learning (ML) algorithms have also been applied to analyse the accelerometer data for developing behaviour classification models in cattle such as decision tree [
9,
13,
19], random forest [
20,
21], kernel support vector machine [
22,
23], discriminant analysis, and
k-nearest neighbours [
23,
24]. These algorithms generated diverse performances of the models depending mainly on the types of behaviour and sensor placement modes [
24,
25]. By using neck collar-based accelerometers, Williams et al. [
26] succeeded in differentiating drinking from standing (100% accuracy) and walking (92% accuracy) events. However, Kour et al. [
12] reported that fitting the accelerometer on a neck-collar was ineffective for classifying suckling behaviour in beef calves. Wolfger et al. [
18] found that the ear-tag based accelerometers were able to classify feeding behaviour of lot-fed cattle along with ruminating, active, and resting behaviours with 95% sensitivity and 98% negative predictive value.
Providing supplemental feeds for range cattle in the form of lick-block or loose-lick minerals containing urea during the dry season or phosphorus during the wet season is fundamental to successful cattle breeding in the tropical area of northern Australia [
27,
28]. The effectiveness of strategic supplementation is contingent upon the ability to decrease between- and within-animal (across days) intake variation [
1]. Because grazing cattle mostly ingest such supplements through licking [
29], identifying and monitoring this behaviour would be useful to determine whether or not individual animals can meet a targeted consumption, or to place an upper limit on access to a supplement. Simanungkalit et al. [
5] has previously shown that time spent at mineral blocks measured by an automatic supplement weighing unit was proportional to block intake on a herd basis. However, high deviation obtained from their linear association was found because of exploratory time before licking. Hence, identifying whether or not the animal is licking while visiting the block supplements is pivotal for improved accuracy of intake prediction. The capability of tri-axial accelerometers to classify behaviour in cattle may offer potential to quantify licking events and time spent licking for the prediction of mineral block supplement intake by individual cattle.
To the best of the authors’ knowledge, no studies have been reported to differentiate licking from other behaviours using tri-axial accelerometers in beef cattle. Hence, this pilot study aimed to determine the effectiveness of tri-axial accelerometers deployed on a neck collar and an ear-tag to characterise the licking behaviour of individually penned beef cattle at a mineral block supplement by distinguishing between licking and other observed (eating, standing, and lying) behaviours. To assess the performance of each deployment mode, four ML algorithms were used to develop behaviour classification models using three different sets of ethograms.
4. Discussion
Dependency upon integration of radio frequency identification (RFID) and automatic feeding systems to remotely monitor supplement intake of beef cattle has prompted the use of more efficient and accurate technologies for the collection of individual information in larger herds without disrupting their daily routines and natural behaviours. Tri-axial accelerometers have the capability of accurately differentiating mutually-exclusive behaviours of grazing ruminants [
6], and this is fundamental to predict individual feed intake based on time-spent feeding [
53]. For cattle offered mineral block supplements, licking events and time spent licking have to be appropriately distinguished from other behaviours to develop an algorithm for predicting individual mineral block consumption. Supplementing cattle with mineral blocks is usually conducted while cattle are grazing in the paddock. This current study was designed as a pilot study to examine the capability of tri-axial accelerometers to differentiate the signals associated with licking and other behaviours. Therefore, only a small number of cattle were used and closely monitored while housed in pens. Further studies would need to be conducted with more animals to test the suitability of the sensor and algorithms under field conditions.
In this present study, MVA and SD
X were the top two features used to classify the licking behaviours of beef cattle by the ML algorithms employed on the tri-axial accelerometer data. This trend was consistent across five out of six ethograms (3 for each deployment mode). Gao et al. [
8] explained that MVA is the variability of waveform length aggregate of amplitude, frequency and duration over the
X-,
Y- and
Z-axes values while SD
X represents distribution of the signal within the
X-axis values. Hence, the differentiation of
X-axis values was evidence of apparent dorso-ventral moving direction recorded by neck-collar and ear-tag accelerometers when the event changed from licking to other behaviours. A recent study using an ear-tag accelerometer configured at 12.5 Hz with a 10 s time interval reported MVA and SD
X as the two most important features to classify grazing, lying, standing and walking events of sheep [
39]. The presence of MVA and SD
X in our study indicated that the ML algorithms discriminated the behaviours based on the difference of movement patterns between behaviours.
For the neck-collar deployment, AVG
Z was the first important feature in ethogram 3 followed by SMA and MAX
Z and is the most consistent feature within the top three features in all ethograms. González et al. [
35] found that SD of the vertical (up-down) acceleration from neck-collar accelerometer was more sensitive for differentiating grazing behaviours in cattle because of its ability to capture head positions. The change in
Z-axis values in the present study signified that the neck-collar accelerometers captured the distinction of longitudinal (anterior-posterior) movements of the head when the cattle were licking. During licking the head is lowered and as the tongue protrudes, the head moves back and forth in the longitudinal plane. This might relate to the high accuracy of the neck-collar accelerometer in a situation where similar head orientation was captured from licking and biting behaviours. Also, SMA is a suitable measure to differentiate static and dynamic activities from the accelerometer signals [
8,
19,
54]. Hence, the presence of SMA in ethogram 3 is indicative of the neck-collar accelerometer’s capability to distinguish between licking and inactive behaviours.
By using random forest ML algorithm, two deployment modes (neck-collar and ear-tag) of tri-axial accelerometers were capable of classifying licking by contrast with eating, standing, and lying behaviours with high accuracy (>90%;
Table 4). The behaviour classification model for the RF algorithm was superior to that of SVM, kNN, and DT algorithms across all ethograms within the two deployment locations. Compared to other ML classifiers, RF has the capability to rank the most important predictor variables and to model multifarious interactions among variables to improve prediction accuracy [
55]. Hence, instead of using all variables, RF randomly selects subsets of variables to determine the best split of each junction of the tree [
43]. A study using a neck-collar accelerometer on dairy cows found that the RF algorithm was able of categorising grazing, ruminating, walking, and resting with an overall accuracy and kappa of 0.97 and 0.95, respectively [
45]. The high accuracy of RF is mainly because of its robustness to noisy data and ability to handle non-linear correlated data [
56].
The lower performance of the DT algorithm in this present study might be because of over-fitting the model and the hierarchical partitioning of each tree that reduces (1) the ability to categorise relationship between variables and (2) the effective sample sizes causing a difficulty in identifying rules and trends in each subsample [
43]. It should be noted that in ethogram 2, inactive behaviour combined standing and lying while in ethogram 3, non-licking behaviour combined eating consisting of biting (head lowered) and chewing (head raised), standing (head raised) and lying (resting). Therefore, it was likely that the accelerometer signals from licking and biting when the cattle lowered the head would be misclassified, as the feeding bucket and mineral block supplement were positioned at a relatively similar height from the floor. This might be responsible for the moderate sensitivity of ear-tag deployment in ethogram 3 (<90%) and may have affected overall accuracy of the algorithm. In addition, lower PPV and sensitivity of the ear-tag accelerometer may have occurred because of a more flexible attachment of the sensor to the ear that increased the false positive rate. A lower ear-attached accelerometer (SensOor) performance was reported by Wolfger et al. [
18], where negative predictive value and sensitivity of feeding class were 97% and 93%, respectively, with low specificity (70%) and poor PPV (54%). This was because of a high proportion of rumination that was categorised as feeding in their model.
In this current study, the behaviour classification model for the neck-collar tri-axial accelerometer was more accurate than the ear-tag tri-axial accelerometer, with Cohen’s Kappa coefficient for the neck-collar deployment model being also superior to the ear-tag deployment. The substantial agreement between actual and model-predicted behaviour was higher in the present study than studies with dairy cows by Bikker et al. [
57] and dairy calves by Roland et al. [
16] who found 0.77 and 0.68 of Cohen’s kappa value for eating and drinking using an ear-attached accelerometer. The lower kappa coefficient for the ear-tag accelerometer compared to that for the neck-collar was affected by complex and repetitive ear movements. Barwick et al. [
6] reported that a possible interdependency of ear-tag acceleration signals from body movements might cause uniformity of the signals from different behaviours. Hence, rigid attachment of the sensors would maintain their orientation and consistent signal to generate accurate behaviour classification.
Apart from the lower performance of ear-tag based accelerometers compared to the neck-collar accelerometers, the practicalities of adoption in commercial contexts favour ear-attached sensors. The smaller size makes it less invasive to the cattle and costs less to implement per individual. Therefore, classification algorithms must be capable of dealing with interdependent dynamic accelerations. The potential of an ear-tag based sensor to accurately discriminate licking would be an improvement enabling measuring mineral block supplement intake based on time spent licking by individual cattle. It also offers versatility and is an efficient way to monitor and harness individual information particularly in an extensive environment. Advancements in remote monitoring systems using internet technology are required to remotely transmit the data from the ear-tag sensor to a central database system for improving production efficiency by reducing time of mustering for individual data collection. However, in commercial systems where cattle are already fitted with neck-collars for other purposes, measuring licking with neck-collar accelerometers would be ideal due to the greater accuracy with this deployment.