1. Introduction
Virtual fencing represents an emerging technology that has the potential to fence and move livestock for enhanced grazing control and monitoring. Virtual fencing communicates the presence of boundaries to animals through audio and electrical pulse signals administered to individuals via a device worn by the animal, such as a neckband. The animals are trained to associate the audio warning tone with the negative stimulus to avoid virtual boundaries based on sound only [
1,
2,
3,
4]. There are several types of virtual fencing neckband devices that are currently being commercialised for use on beef and dairy cattle [
1,
2,
3,
4,
5], goats [
6], and sheep [
7,
8]. While there has been demonstrated success of the technology with cattle across a variety of paddock situations and different device designs [
1,
2,
3,
9,
10,
11], to date there has been limited success in applying automated technology to sheep. The fleece of sheep presents a barrier to successful administration of electrical pulses relative to cattle, and their smaller body weight places greater restrictions on device design. The automated Nofence devices (Nofence, AS, Batnfjordsøra, Norway) have been tested on sheep [
7,
8]. These devices used a 4s audio tone of increasing frequency when an animal reached the virtual border, followed by an electric shock if they kept moving forward [
7,
8]. However, the authors concluded the technology was a significant welfare concern due to high individual variation in reactivity to the electric shocks and learning rates that resulted in sheep not meeting predetermined learning criteria [
7,
8]. These results may have, in part, been due to the stage of the technology development at the time of the trials [
7,
8]. Earlier trials with dog collars and a fixed buried wire showed ewes did learn the concept of the virtual fencing cues after only a few experiences with the ‘line’, but the buried wire approach is limited in its commercial application [
12]. Audio-only devices have also been tested, where there is no electrical pulse, and aversive audio cues are administered [
13]. On average, 10% of animals across a small test group showed no response to the different audio stimuli tested, which may limit the system’s application [
13].
In the absence of suitable automated virtual fencing devices for sheep, manual electronic dog collars applying audio and electrical stimuli in a similar manner to the eShepherd
® automated cattle devices have been used across a range of studies to assess the learning capabilities of sheep, their responses to the stimuli, and potential commercial applications of the technology [
14,
15,
16,
17,
18,
19]. These manual trials have relied on visual observation by trial operators to apply a one second audio cue followed by an up to one second electrical pulse if the sheep kept moving forward, and as such have been limited to short-term trials during daylight hours only. When tested individually, sheep learned the association between the audio cue and the electrical pulse and exhibited the correct behaviour of stopping and/or turning around upon hearing the audio cue alone [
14,
15]. High individual variation in the rate of learning was reported, but all sheep during individual testing were able to learn the audio/electrical pulse association [
14,
15]. Sheep in a small group in a commercial paddock setting were excluded from a specific area and rapidly returned to the previously excluded area once the fence cues were no longer applied, indicating that they were learning to respond to the audio cue and not the location [
16]. A virtual fence was also comparable to an electric fence for targeted grazing within small plots and did not result in negative changes in grazing behaviour [
17]. Interestingly, there was evidence in the small (nine animal) short-term trial scenario that not all sheep were required to wear neckbands. Virtual fencing was effective when 66% of sheep were wearing devices, indicating social effects on responses to the fence cues; however, the fence was not effective when only 33% of animals were collared [
18]. Finally, a back fence was able to be implemented to prevent movement back over previously grazed areas [
19]. These studies all indicate the potential for virtual fencing technology to be applied to sheep, but any larger studies, including ones that run across the day and the night, are limited by the logistics of requiring personnel to manually operate the individual devices throughout the study period.
The eShepherd
® virtual fencing devices are currently being commercialised for predominantly beef cattle by Gallagher (Hamilton, New Zealand). The pre-commercial system comprises neckband devices, a base station on site to communicate with the devices, and an online user interface for animal and fence management. Previous research on beef cattle using varying iterations of the pre-commercial eShepherd
® prototypes have demonstrated that all tested cattle will learn to respond to the audio cue alone to avoid receiving electrical pulses, but with individual variation in learning rate and frequency of fence interactions [
9,
10,
11]. Beef cattle have been excluded from environmentally sensitive areas such as riparian zones [
10] and regenerating sapling plantings [
11] and have been comparably excluded from areas of pasture when using virtual fences or electric tape fences [
9]. There is potential for a modified version of this device to be applied to sheep to enable automated research trials.
The aim of this study was to apply modified eShepherd® virtual fencing neckbands to sheep to enable automated trials with small sheep groups across both day and night. These were the first trials conducted using these specific automated GPS-based devices on sheep. It was predicted that sheep would learn the association between the audio cue and electrical pulse and would remain within the inclusion zone.
3. Results
The six sheep at Chiswick were contained within the inclusion zone for 99.8% of the trial period (
Figure 2). Primarily, the incursions into the exclusion zone were on the first day during learning (
Figure 2). Across the entire trial period, the mean audio-only percentage (±SD) was 74.93% ± 4.6%. Across all individuals, the mean number (±SD) of audio cues received was 7.8 ± 8.3, and the mean number (±SD) of electrical pulses received was 1.6 ± 2.2.
Following the 90-min training period, the 10 sheep at Lameroo were contained within the inclusion zone for 92.2% of the trial period (
Figure 3 and
Figure 4). They did cross over into the exclusion zone on the third night at approximately 01:00. Downloaded GPS logs indicated all animals moved from the inclusion to exclusion zone within a 10-min window (
Figure A1), but observation of animal movement in the user interface the following morning indicated it was a rapid flock movement within a few minutes. The online user-interface enabled more precise tracking of the animals’ movements, as additional data were transmitted when an animal interacted with the fence line. However, the data delivery system had limitations in the amount of data it could send; thus, the downloaded cloud logs were the primary datasets used. The animals remained in the exclusion zone until they were walked out by personnel for their daily yard check around 09:00 the morning of the fourth day (
Figure 5). During the time in the exclusion zone, the animals moved around in one corner of the paddock initially, then settled as two groups in proximity, before combining as one group (
Figure 5) until the morning. Once the animals were walked back to the inclusion zone following their yard check, they correctly remained in the inclusion zone until the conclusion of the trial (
Figure 4 and
Figure 5).
During the 90-min training period before personnel intervened, the mean percentage of total audio cues that were audio only (±SD) was 76.2 ± 7.4%, during the incursion period on the third night it was 87.4 ± 6.7%, and for the remainder of the trial it was 82.2 ± 22.9%. Across all individuals, the mean number (±SD) of audio cues received during the training period was 77.3 ± 32.1, and the mean number (±SD) of electrical pulses received during training was 19.1 ± 10.3. For the remainder of the trial period (including the incursion event), the mean number (±SD) of audio cues received was 31.4 ± 57.2, and the mean number (±SD) of electrical pulses received was 4.2 ± 7.9. The mean daily audio cue and electrical pulse values are presented in
Table 1, including the training period and the incursion period, when the animals crossed over into the exclusion zone. The highest numbers of stimuli were received during training and the incursion period (
Table 1). Daily audio cues and electrical pulses per individual animal are displayed in
Figure 6, which indicates high individual variation in received stimuli and thus interactions with the fence line.
4. Discussion
This study applied modified eShepherd
® virtual fencing neckbands to sheep using harnessed units wired to collars to enable automated trials with two small groups of sheep across both the day and night. The results showed the sheep were able to appropriately respond to the virtual fencing cues and learn to respond to the audio cue alone with similar audio-only percentages to cattle wearing the eShepherd
® devices [
9,
11]. The animals were predominantly kept in the inclusion zone across the five-day trial periods, but there was one instance where sheep rapidly broke through the boundary at night, possibly due to an animal scare, and remained within the exclusion zone for approximately 8 h until they were walked back to the yards by personnel. It is unknown when they would have returned to the inclusion zone without personnel intervention. The results of these preliminary trials are promising for the use of automated technology on sheep, but suitable devices still need to be designed that can be applied longer-term, with algorithms specific to sheep behaviour that will successfully herd them back into the inclusion zone. The necks of these reduced-wool sheep were clipped to ensure consistency of contact; thus, a collar device would need sufficient electrode contact and may only be applicable to some sheep breeds. A device that uses aversive audio cues may suit multiple sheep breeds, but habituation to the signals longer-term is probable, as sheep would likely learn to ignore the cue if there are no meaningful consequences. Alternatively, an ear tag could be viable if the required components can be reduced to a small-enough weight.
The relatively successful results of the two automated trials were similar to what has been previously shown using manually operated dog collars on sheep [
14,
15,
16]. Sheep predominantly remained in the inclusion zone and showed responses to the audio cues alone, with variation between individuals in learning rate as well as fence interaction frequency. As described by Lee and Campbell, 2021 [
24], the inclusion of a metric such as the relative proportions of audio cues and electrical pulses is informative with regard to the animal welfare impacts of virtual fencing. This metric indicates if animals are learning to avoid the electrical pulse and respond to the audio cue alone [
25], which has been shown to minimise stress responses to virtual fencing cues in sheep [
26]. However, the high individual variation indicates some sheep may have been experiencing different degrees of welfare impact during the learning process. The audio-only cue percentages were similar to what has been observed in previous cattle trials using the eShepherd
® devices ([
9]: 71.5%; [
11]: 74.5%). Early work applying the Nofence automated devices concluded that virtual fencing may not be suitable for sheep due to limited containment success and welfare concerns [
7,
8]. In the first trials with Nofence, only 37.5% of the test ewes reached the pre-established learning criterion of responding to the audio cue alone after three repetitions of the audio-pulse sequence [
7]. However, this is a strict learning criterion, designed to minimise negative impacts of the electrical pulses on animal welfare. Previous work with manual collars has shown 22% of sheep responded to the audio cue alone after their third interaction and 67% of sheep did so by the fifth interaction [
14], similar to early studies with cattle, where 56% of animals correctly responded by the fifth interaction [
27]. These calculations could not be made in the current study with the way the device data were recorded, but all animals showed responses to the audio cue alone within the approximate 90-min training period.
Trials also ceased in [
8], as the majority of sheep exceeded the pre-set criterion of a maximum of four shocks on the second day of the trial. Mean electrical pulse values in the current study far surpassed the criterion in [
8] during the training period and when the group of animals broke through into the exclusion zone. However, it is unclear if every delivered pulse in the current study was aversively felt by the individual. In the study in [
7] using Norwegian white and Spæl sheep, many individuals were removed because of severe running reactions, but conversely, other individuals were removed due to no visible reaction to the pulse. In a follow-up study with the same breeds, similar individual variation in sensitivity to the same pulse strength was observed; however, animals in the test groups were not preselected [
8]. This variation in sensitivity to the electrical pulse could hinder learning for the individual receiving the electrical pulses as well as for other individuals within the flock through social facilitation. Partial records of electrical pulse delivery in the current trial indicated some electrical pulses were likely not of sufficient contact to be felt, but these datasets were not complete given the limitations in data transfer capacities. The studies with the Nofence devices used earlier prototypes that had several technical issues; this could have contributed to the poor learning outcomes. More recent trials have successfully applied Nofence devices to cattle for an extended period of time [
1] and to small goat groups for five days [
6]. If animals are receiving a high number of aversive electrical pulses, then this would need to be addressed before an automated system can be considered a welfare-friendly fencing alternative. A managed training period on-farm could mitigate some of the risks with this individual variation. However, individual variation in sensitivity, which could be exacerbated by variation in wool coverage and/or patchy wool shedding, could increase the inconsistency in electrode contact and animal responses in the long term.
The current trials were the first to apply the eShepherd
® algorithm by leaving devices on to run continuously across the day and night in sheep (minus daily yard checks), confirming the on-farm feasibility for this technology. However, there was an instance during the Lameroo trial where the flock crossed over the boundary at night and remained within the exclusion zone until the following morning. The eShepherd
® system operates so that animals still receive audio cues and electrical pulses if moving farther into the exclusion zone (but not if moving in the direction out of the exclusion zone), unless they reach a predetermined limit and initiate a temporary cue timeout. There are no different cues emitted to communicate the breach to the animal. The animals spent the time in one corner of the exclusion zone, predominantly stationary, presumably lying down. It is unclear precisely what led to the sheep running through the boundary. Regardless of the reason, the system is designed to herd cattle back in such situations, so it is also unclear if the sheep did not know how to return, or if the location was preferred for nighttime resting. Sheep are highly gregarious, with some individuals and even breeds more gregarious than others [
28,
29]. Furthermore, social isolation is a significant stressor for sheep [
30]. This could present issues if there are one or two individuals that may cross through the boundary and attract others to follow. Similarly, once the sheep are in the exclusion zone, the automated algorithm may struggle to herd them back if it is working against the flocking instinct. Cattle trials have shown the virtual fence is less effective when there is close visual contact with other cattle groups across a boundary [
31]. Trials with automated dog collars using a buried wire system showed that if naive sheep crossed the boundary, their trained peers would occasionally follow after them [
12]. Similar issues were observed when only some animals had functioning Nofence collars [
8]. When investigating whether controlling only some sheep within a flock with virtual fencing devices was effective at containing sheep, Marini et al. (2020) [
18] showed that a minimum of 66% of sheep needed to be virtually fenced to contain the whole group. It should be noted that this was for a short period with only nine animals; longer time frames and larger animal numbers may reduce the effectiveness. Conflict between the aversive electrical pulses and a desire to reunite with conspecifics have also been demonstrated in cattle [
19]. Thus, there needs to be further work to understand how sheep may be encouraged to return to the inclusion zone in situations where the entire group has crossed over. This further work would also need to extend to fencing situations where there are multiple virtual boundaries (i.e., a virtual enclosure), rather than a single virtual boundary combined with physical fences. Algorithms that are specific to sheep behaviour may need to be developed for the automated technology to operate successfully. More precise GPS movement detection may accommodate the smaller and potentially more rapid direction changes of sheep. In the current study, it is unknown if the sheep would have eventually returned if they had not been walked back by personnel for their yard checks. Similarly, personnel intervened at the conclusion of the training period at Lameroo to minimise any welfare compromises to the few sheep that were separated from the group and remained within the exclusion zone.