1. Introduction
The foundations of sustainable rangeland management practices lie in our understanding of the key rate processes of the system: plant growth and animal consumption. On the plant side, technologies such as remote sensing have transformed our ability to study rangeland vegetation [
1,
2]. On the animal side, in contrast, the study of grazing by herbivores on rangeland has lagged in terms of sensor technology, which in turn hinders progress in how we describe and conceptualize the process itself [
3]. Intake, in particular, remains challenging to study and quantify, despite its centrality to vegetation dynamics, animal nutrition and overall system stability [
4,
5].
One way of “eavesdropping” on intake is via the biting and chewing sounds it generates, as demonstrated in cattle decades ago (e.g., [
6]). The potential utility of acoustic monitoring becomes apparent from considering the basic properties of the intake process. Herbage intake is the outcome of a set of behaviors, of which a central one is movement (opening and closing) of the jaws to perform bite and chew actions, both of which generate sound. The potential scope of how an animal performs jaw movements is determined by anatomy and biomechanics [
7,
8,
9,
10], but its expression is profoundly influenced by forage conditions, including the abundance, spatial distribution and quality of the herbage [
11,
12]. Assuming for the moment that grazing is a binary state (the animal is either grazing or not grazing), a simple arithmetic definition of daily herbage intake rate is the product of daily active grazing time and intake rate during active grazing. Intake rate during active grazing can be expanded to the product of bite rate during active grazing and bite weight. Bite rate during active grazing can, in turn, be expanded to the product of jaw movement rate during active grazing and the proportion of jaw movements allocated to biting actions [
13].
While it is true that the measurement of bite mass is relatively intractable, the remaining terms in the above expression are tractable empirically, although not equally so. In light of the centrality of jaw movements, various sensors have been proposed to monitor them, as reviewed by Andriamandroso et al. [
14]. In acoustic monitoring, but probably more generally too, the detection of jaw movements is easier than their classification, especially when the chew–bite type of jaw movement is prevalent, as expected in cattle grazing [
15]. Crucially, by defining grazing in terms of the fundamental behavioral component of the process—the jaw movement—the functional connection that must exist between it and the mass component of intake, via chewing, can be leveraged [
16,
17]. Even without knowing the coefficient by which to convert chewing actions into units of mass, it is reasonable to argue that, under a given set of forage conditions, the short-term jaw movement rate is strongly coupled to the short-term bite rate, which is in turn coupled, perhaps less strongly, to short-term intake rate [
18,
19].
As a quantitative baseline for comparison, itself based on acoustic monitoring, consider the single-group study of Ungar and Rutter [
20], in which six dairy cows grazed under ideal conditions that placed no constraints on herbage abundance or quality. During a 10 min sampling interval of active grazing, the cows performed, on average, 69 jaw movements min
−1. Although there were sporadic pauses in the rhythm of the jaw activity, the point is that these animals grazed at a jaw movement rate that was close to its potential, unimpeded “cruising speed”, the entire (albeit limited) time.
Free-ranging livestock on extensive rangelands are probably near the opposite extreme of forage conditions in terms of the variability expected to be elicited in short-term grazing behavior, for two reasons. First, uninterrupted access to the forage resource may be associated with more grazing time at a lower behavioral intensity of grazing (presupposing that “grazing intensity” in this sense exists). Second, and perhaps more importantly, forage conditions on extensive rangelands are poorer than those encountered by dairy cows on pasture on every score: overall abundance may be an order of magnitude lower; nutritional quality would mostly be lower; and the spatial heterogeneity of abundance and quality might be an order of magnitude greater. Consequently, one might expect a more diffuse distribution of short-term jaw movement rate, as would be revealed by examining the timeline of jaw activity as rates.
A second, and more mechanistic, way of defining intake is functionally. The simplest functional definition of intake rate is based on the time budget of the process: the mass harvested in a bite in relation to the search and handling times invested [
21]. For the dairy cows above, performing 69 jaw movements min
−1, search time was effectively zero because sward uniformity was high and selectivity was close to zero. A high number of jaw movements performed per minute is indicative of unconstrained forage conditions, but the decisive factor is rhythm, i.e., the regularity of successive jaw movement events, as demonstrated by the time interval between them [
22,
23,
24]. Based on a companion, but smaller, acoustic monitoring study, conducted with goats under extensive forage conditions [
25], it was proposed that as the search time component of the intake process grows, the interval between successive jaw movements will be punctuated with larger values. Thus, a reduction in jaw movement rate below the unconstrained maximum is not the result of a general slowing down while maintaining a regular rhythm, but the fracturing of the unconstrained rhythm with interruptions. This may give rise to higher-level patterns, as would be revealed by interval-based analysis of the timeline of jaw activity.
The above background indicates that a description of grazing at the jaw movement level, as afforded by acoustic monitoring, should facilitate a deeper understanding of how herbivores interact with their foraging environment. We sought to advance this approach by monitoring the jaw activity of free-ranging beef cattle on Mediterranean herbaceous rangeland over multiple days. The detection of jaw movements acoustically is still in its infancy, and a commercially available sensor has yet to be developed. There were adequate resources to monitor all individuals in a single-group, observational study with no treatment structure. Obviously, jaw movement events are not independent and are highly serially correlated. Likewise, animals in a paddock are not independent in their behavior. For such reasons, the database is, strictly speaking statistically, a single data point or sample. But it is a well-grounded one, comprising ≈5 M jaw movement events from multiple animals, sampled continuously over multiple days, in two contrasting seasons. The rate-based and interval-based approaches introduced above were applied to this collection of events to profile how free-ranging animals interact with their foraging environment.
2. Materials and Methods
2.1. Study Site
The study was conducted on an area of Mediterranean herbaceous rangeland in the region south of Mt. Carmel, Israel, inside the UNESCO-recognized Biosphere Reserve of Megiddo (
https://www.unesco.org/en/mab/megiddo?hub=66369, accessed on 5 January 2025), and under the stewardship of kibbutz Ein HaShofet. The topography comprises rolling hills with a mean altitude of 300 m above sea level. The soils are predominantly shallow and of the rendzina type that developed over a soft chalky bedrock [
26]. The climate is strongly seasonal with hot, dry summers and cool, rainy winters, with mean annual rainfall of 600 mm (for a hydrological cycle starting 1 October) and average daily temperature of 18 °C. The species-rich vegetation of the rangeland is primarily herbaceous and dominated by annuals, interspersed with patches of low shrubs [
27]. The paddocks are bordered by planted windbreaks. The annual grazing cycle commences in December/January, approximately one month after the emergence of the annual herbaceous vegetation, a process triggered by the first major rains of the hydrological cycle. The rangeland paddock used in this study was of a prevailing southerly aspect with moderate slopes of 3–7%.
2.2. Animals and Their Management
The experimental animals were drawn from a mixed-breed, beef-cattle herd of 800 head. The primary calving season was in the summer (June–August), and calf weaning was at age 6–8 months. To minimize disturbance to the herd as a whole, 12 medium-sized mature cows (≈520 kg live weight), which did not display excessive agitation in the animal handling facilities, were separated and moved to the experimental paddock on 19 October 2012. The fenced paddock was equipped with troughs for water and supplementary feeding, as well as animal handling facilities. A bull was introduced two days later, at the start of the breeding season. The 12 cows had a number of months in which to form a stable social group and familiarize themselves with the paddock before the acoustic monitoring began. Poultry litter was provided ad libitum in the summer months as a source of non-protein nitrogen.
2.3. Acoustic Monitoring Periods and Environmental Conditions
The acoustic monitoring of the 12 experimental animals was conducted in the spring and in the summer of 2013, when the cows were in mid- and high pregnancy, respectively. Monitoring in the spring commenced on 21 March, at the height of the season in terms of herbage growth and quality. The summer, dry season monitoring commenced on 4 July, when the quality of the herbage was low, but before grazing had depleted it substantially (see photographs in
Supplementary Figure S1). Meteorological data for the relevant monitoring periods in the two seasons were obtained from a public database (Israel Meteorological Service, Beit Dagan, Israel;
https://ims.gov.il/en/data_gov, accessed on 5 January 2025). There was no rainfall during either monitoring period. The mean daily temperature in the spring and summer monitoring periods was 15.3 °C and 24.5 °C, respectively. Daily rainfall for the 2013/2014 hydrological year, and the 10-minutely timeline of environmental conditions during the two monitoring periods are shown in
Supplementary Figures S2–S7. The average sunrise time in the spring and summer monitoring periods was 05:40 h and 04:43 h, respectively, and the corresponding sunset times were 17:54 h and 18:47 h (times throughout are UTC+2).
2.4. Sensor Design Considerations
Methodologically, obtaining acoustic data from free-ranging beef cattle is much more challenging than doing so in short-duration trials conducted with relatively docile dairy cows (as in [
20]). Beef cattle need to be corralled and then coaxed through a cattle chute and into a cattle squeeze for safe handling. Aside from being time-consuming and disturbing to the daily behavioral rhythm of a free-ranging animal, all the steps of this process are stressful for the animal. That necessitates incorporating into the sensor design sufficient power to sustain continuous monitoring for multiple days. Furthermore, makeshift arrangements for microphone placement on the head of the animal, such as on the forehead as used by Ungar and Rutter [
20], would not suffice under free-ranging conditions. A halter-based sensor design is feasible, but has disadvantages; not least, the microphone would be on the nasal bridge, making it susceptible to physical damage. The approach chosen for the present study was to attach the sensor to one horn of the animal. This microphone location was found by Tani et al. [
28] to yield a high-quality signal, which we confirmed in a preliminary field test of the sensor that was assembled here.
2.5. Acoustic Sensor Components
The acoustic sensor was built around a commercially available MP3 device (Sansa Clip+, SanDisk, Milpitas, CA, USA), which had a record mode and a slot for a MicroSD memory card. The device was small and of rectangular shape (54.5 × 34.5 × 10.4 mm; rear clip removed), lightweight (24 g) and found to be robust under field conditions. A number of modifications were made to the device in house. The built-in microphone was detached and discarded, and two wires, of 4 cm length, were soldered to the motherboard connectors in its place. To the other end of the wires was attached a 2-pin Molex-type female connector. In order to connect to an external voltage source, an additional pair of wires, of 4 cm length, were soldered to the motherboard connectors to which the internal battery was, and remained, connected. A 2-pin Molex-type male connector was attached to the other end of the wires. An aperture was cut into the device casing to accommodate both pairs of wires and enable the casing to be fully shut. An external vibration-type piezoelectric microphone (Model wcp500, Cherub Technology Co., Nanshan, China) provided the source signal to be recorded. The original cable was shortened to 10 cm and terminated by a 2-pin Molex-type male connector. The external voltage source was a rechargeable 3.7-V battery pack (Meircell, Ashdod, Israel) of dimensions 67.1 × 55.2 × 18.5 mm, weight 145 g and capacity 7800 mAh, sufficient to power the recording device for approximately eight days of uninterrupted recording. The battery pack was fitted with a 2-pin Molex-type female connector in order to connect to the recording device. The device operated under RockBox software 1.2.7 (
https://www.rockbox.org/, accessed on 5 January 2025) which added the critical functionality of recording to the MicroSD memory card, and the automatic, gapless saving of files of specified size or duration, in a choice of formats. The device was configured to store recordings on a 32-GB MicroSD card as WAV-format files of 90 min duration, at a sampling frequency of 48 kHz. In field tests, the inward-facing microphone recorded sound vibrations generated by jaw movements and transmitted via the skull to the horn of the cow, with virtually no surrounding noise.
For protection, the recording device was housed in a customized aluminum casing (A. Braun Metals, Tel-Aviv, Israel) of dimensions 61.6 × 41.4 × 15.6 mm and weight 73.6 g, with an aperture for the two pairs of cables. This was stacked on the external battery and placed inside a larger customized casing of dimensions 74.2 × 61.1 × 35.2 mm and weight 179.5 g, with an aperture for the microphone cable. Following assembly and the initialization of recording, the casing was waterproofed with heavy-duty duct tape. The total weight was 425 g.
2.6. Acoustic Sensor Deployment
Prior to installation on an animal, and after removal, a set of taps was recorded indicating the precise date–time, allowing later correction for clock drift. At installation, each animal in the group was restrained in a cattle squeeze, with its head immobilized for operator safety. The box-shaped protective casing containing the recording device and battery was taped around one horn approximately 4 cm from its base and securely fastened to the horn using a turn-key metal hose clamp (Ideal Tridon, Smyrna, TN, USA; diameter 152.4 mm). The external microphone was then taped to the horn in the space between the casing and its base, with the vibration pad facing inward. The entire assembly was wrapped well with heavy-duty duct tape (
Supplementary Figure S8). At the end of the monitoring period (after approximately one week), the animals were similarly restrained for the removal of the equipment.
It was imperative that the sensor should not detectably alter the animal’s behavior or impinge on its welfare by virtue of its size or weight. Three measures were taken to verify this. First, on release from the squeeze, the animals were retained in a small holding area for approximately 30 min in order to observe them closely for any indication that the apparatus was bothering the animals (e.g., head-shaking, pacing, general irritation). Second, all animals were inspected and observed during active grazing every day of the monitoring periods. Third, a series of 21 video recordings of 5 min duration were made of a random selection of the experimental cows during active grazing to verify that behavior was normative.
2.7. Processing of Acoustic Data
The large database of audio files was analyzed automatically by an algorithm that was developed in house and described in Navon et al. [
29]. The algorithm reliably identifies sound-generating jaw movements but does not classify them. Algorithmically, detecting jaw movements is easier and more precise than classifying them [
23,
30,
31,
32]. The algorithm was designed to be as general as possible in terms of animal species, foraging environment and recording equipment, and to require no calibration. The algorithm identifies jaw movements according to key features in the time domain that are defined in relative terms and uses a machine-learning approach to separate true jaw movement (JM) sounds from spurious background noise. One feature of true JMs is that the signal intensity undergoes a sharp change at the start and end of the sound burst. Second, true JMs rarely occur in isolation as “singles” but occur as part of a sequence. Third, the sound signal intensity of bites and chews is strong compared to any background ambient noise that might be detected. Fourth, in cattle, there is a limited range of duration within which all normative sound-generating jaw movements fall. Combining these criteria enables low rates of false positive and false negative events to be achieved, especially when working with vibration-type microphones, as in the present study. Preprocessing of the signal includes correction for the clock drift of the recording device, high-pass filtering (60 Hz), amplification and down-sampling to reduce computational load. The algorithm outputs the timeline of unclassified JM events. To monitor the performance of the algorithm, the waveform of the sound signal was overlaid with the timeline of JM events identified by the algorithm, and at least 30 validation segments per season–animal, of 10 min duration each, were examined closely for false negative and false positive events.
2.8. The Core Datasets
The raw dataset for the spring season (March) deployment contained ≈1.6 M JM events obtained from 11 of the 12 recording devices that each yielded ≈100 h of continuous recording from installation. (The twelfth device failed to record soon after deployment.) The recording duration was shorter than expected due to poor battery performance. The signal quality was consistently poor in one device, and its data were excluded. The stream of JM events from the remaining 10 devices/cows was trimmed: data from the day of installation (early morning until midnight) were excluded, as were all data after midnight 72 h later. The remaining three complete 24 h cycles, starting from midnight, were retained for analysis. The dataset contained 1,203,568 JM events, including rumination.
The raw dataset for the summer season (July) deployment contained ≈3.9 M JM events obtained from 11 recording devices that each yielded ≈180 h of continuous recording. (One device failed early, but battery performance was in the expected range for the remainder.) By way of trimming, data from the early morning until 14:00 h on the day of installation were removed, as were all data after 14:00 h, seven complete 24 h cycles later. In one device, the data for the last 24 h cycle were incomplete; thus, the data for that entire day were excluded. The resultant dataset, drawn from 11 cows, contained 3,652,708 JM events, including rumination.
2.9. Designation of Rumination Bouts
The most important distinction that needed to be made in the JM timelines was that between rumination JMs and the remainder, with all the latter being regarded as ingestive JMs. This can be performed manually by scanning the timeline of JM events for the distinctive pattern generated by rumination: a long, repetitive sequence of highly regular events (within-bolus chews) interspersed at regular intervals by short, few-second (inter-bolus) gaps. However, when using screens of a fixed time span, the number of screens to scan is high and many of them contain few, if any, events. It was found to be more efficient and informative to scan interval-based plots, whereby JM events are arranged sequentially along the x-axis, and the time interval between successive events (previous to current) is shown on the y-axis. For each season–device combination, the intervals were computed from the second JM event until the last. For visual-scanning purposes only, the data were arbitrarily divided into 1000-event groupings that were displayed graphically in a series of plot panels, each showing 1000 values across the full width of the screen, with sufficient resolution to identify individual events. The time interval (y-axis) was displayed on a logarithmic scale.
For the most part, the visual identification of rumination bouts was rapid and unequivocal. It is inconceivable that such distinctively regular rhythms of sound bursts and interruptions could be generated by chance during grazing; thus, aural verification that all the JMs were exclusively pure chews was not required. Small deviations or transient fluctuations in the pattern, or occasional longer breaks in the order of 1 min, were tolerated as part of a single rumination bout. The interval associated with the first chew action of a bout was designated as part of grazing and not rumination. Using features of the graphic interface of JMP v16 (SAS Institute Inc., Cary, NC, USA), all such clearly identifiable bouts of rumination (comprising multiple boluses) were designated “rumination” in a dedicated pass of the 1000-event plot panels. Individual boluses were not delineated. Runs of JM events to be designated as rumination could begin or end at any position in a plot panel and could span multiple panels.
There were patterns that retained features of rumination but deviated significantly from the idealized level of regularity, perhaps caused by transient periods of poor signal quality that degraded event identification. There were also rumination-like patterns that were not sustained sufficiently to make the designation definitive, as they may have occurred by chance during grazing. Both these types of pattern were designated “possible rumination” in a separate, dedicated pass of the 1000-event plot panels. All jaw activity not defined as rumination or possible rumination was assumed to be associated with ingestion.
2.10. Designation of Additional Patterns
In the process of scanning the interval-based plots to designate rumination, it became clear immediately that “grazing”, i.e., anything not designated as rumination or possible rumination, does not have a universal “signature” pattern as does rumination. Although the same baseline rhythm of interval values of ≈1 s dominated grazing, the interruptions inserted into it were highly erratic. Nevertheless, three fairly distinct patterns were recognizable among the non-rumination sections, representing opposite extremes of JM intensity and something in the middle. Since these patterns emerged organically from the data and were not a planned part of the methodology, they are described in detail in the Results (
Section 3). Bouts of the two extreme patterns were marked up in separate, dedicated passes of the 1000-event plot panels. All remaining events, representing the majority of non-rumination JM events, then received a single designation. Note that certain patterns included long intervals (in the order of 10
3 s), which would conventionally be called “rest”. Rather than drawing an arbitrary line between “rest” and “graze” intervals, we treated the entire (non-ruminating) day as a set of intervals that can be considered collectively as “grazing” in a broader sense of the term, even though they span orders of magnitude.
2.11. Rate-Based Analysis
The timeline of JM data was summarized as minutely counts (within season, device and day number). This generated a file of 109,761 records, which was then zero-padded to account for “quiet” minutes and ensure a record for every minute in the time range of the data for each season and device (n = 152,640 records). Non-zero counts were subdivided according to the five-way designation of patterns (two for rumination and three for grazing), and a minutely designation was assigned based on the maximum. Quiet (zero-padded) minutes were designated as such. The resulting file was used for all rate-based analyses and to generate empirical cumulative distribution function (CDF) curves.
5. Conclusions
Much can be learned from just the timeline of unclassified jaw movement events, especially when acoustic monitoring is conducted continuously over multiple days. The main insights gained can be summarized as follows.
1. Jaw activity reveals rhythms, patterns and styles. Rumination bouts were unique in combining strong rhythm—shown by the within-bolus runs of chew events all falling within a narrow interval band of ≈1 s—with a strong pattern created by the regular insertion of fairly constant inter-bolus intervals, many multiples of the baseline interval, at a larger time scale.
2. There was no single “signature” jaw movement pattern for grazing (i.e., all non-rumination JM events). Although a similar underlying “natural” rhythm seen during rumination dominated the non-rumination population of intervals, it was intermittently and irregularly interrupted by longer intervals whose size scaled logarithmically.
3. There was evidence of further substructure showing a degree of separation between “grazing” and “resting” in the conventional sense of the terms. This revealed itself as two zones of relatively steep ascent in the time accumulation curve: one in the region of the natural, unimpeded RJM, with an interval of ≈100 s, and one in the region of 103 s, which would appear to correspond to rest in the conventional sense.
4. The distribution of non-zero ingestive JM rates shows strong bimodality. The primary modal region was centered near the frequency corresponding to the natural rhythm of ≈60 JM min−1, demonstrating that jaw movement rate on natural rangelands can match that achieved on abundant sown pastures. The secondary modal region was centered at ≈15 JM min−1, suggesting a different mode of grazing. Beyond speculation, it is unclear how to explain the bimodal nature of grazing intensity, and what benefit low-intensity grazing brings.
5. The notion of behavioral grazing intensity is supported strongly by the rate-based results. This calls into question the approach of viewing grazing as a binary state or expecting measures of grazing time to be indicative of intake rate.
Rate- and interval-based analyses of information at the jaw movement level have furnished a useful way of profiling how an animal interacts with its foraging environment, culminating in the time accumulation curve. The classification, and not just identification, of jaw movements would be expected to yield greater insights, which is a major reason why acoustic monitoring holds much promise. These findings would seem to justify further development of the acoustic monitoring technology to enable its wider adoption in grazing research and management.