Video recording and vegetation classification elucidate sheep foraging ecology in species‐rich grassland

Abstract Factors influencing grazing behavior in species‐rich grasslands have been little studied. Methodologies have mostly had a primary focus on grasslands with lower floristic diversity. We test the hypothesis that grazing behavior is influenced by both animal and plant factors and investigate the relative importance of these factors, using a novel combination of video technology and vegetation classification to analyze bite and step rates. In a semi‐natural, partially wooded grassland in northern Estonia, images of the vegetation being grazed and records of steps and bites were obtained from four video cameras, each mounted on the sternum of a sheep, during 41 animal‐hours of observation over five days. Plant species lists for the immediate field of view were compiled. Images were partnered by direct observation of the nearest‐neighbor relationships of the sheep. TWINSPAN, a standard vegetation classification technique allocating species lists to objectively defined classes by a principal components procedure, was applied to the species lists and 25 vegetation classes (15 open pasture and 10 woodland) were identified from the images. Taking bite and step rates as dependent variables, relative importance of animal factors (sheep identity), relative importance of day, and relative importance of plant factors (vegetation class) were investigated. The strongest effect on bite rates was of vegetation class. Sheep identity was less influential. When the data from woodland were excluded, sheep identity was more important than vegetation class as a source of variability in bite rate on open pasture. The original hypothesis is therefore supported, and we further propose that, at least with sheep in species‐rich open pastures, animal factors will be more important in determining grazing behavior than plant factors. We predict quantifiable within‐breed and between‐breed differences, which could be exploited to optimize conservation grazing practices and contribute to the sustainability of extensive grazing systems.

1. Factors influencing grazing behavior in species-rich grasslands have been little studied. Methodologies have mostly had a primary focus on grasslands with lower floristic diversity.
2. We test the hypothesis that grazing behavior is influenced by both animal and plant factors and investigate the relative importance of these factors, using a novel combination of video technology and vegetation classification to analyze bite and step rates. 4. Taking bite and step rates as dependent variables, relative importance of animal factors (sheep identity), relative importance of day, and relative importance of plant factors (vegetation class) were investigated. The strongest effect on bite rates was of vegetation class. Sheep identity was less influential. When the data from woodland were excluded, sheep identity was more important than vegetation class as a source of variability in bite rate on open pasture. 5. The original hypothesis is therefore supported, and we further propose that, at least with sheep in species-rich open pastures, animal factors will be more important in determining grazing behavior than plant factors. We predict quantifiable within-breed and between-breed differences, which could be exploited to optimize conservation grazing practices and contribute to the sustainability of extensive grazing systems.
At a more local level, within areas dominated by conventional agriculture, areas of high floral and faunal biodiversity may survive, such as many of the nature reserves in the UK (https://jncc.gov. uk/) and elsewhere. Therefore, at both landscape and local levels, conservation management is required, often with support from agrienvironment schemes. Conservation grazing may be applied, which is the use of grazing livestock to promote habitat and species diversity by the removal of excess herbage and the management of vigorous, invasive species. Management is usually empirical and sitespecific (Rook, Dumont, et al., 2004) and needs to be underpinned by ecological and behavioral research on how grazing livestock exploit these landscapes.
In the global context, local breeds of livestock are often claimed to possess local adaptations, which may be highly relevant to the sustainability of pasture-based livestock farming under conditions of global climate change (Boettcher et al., 2015). Animal genotypes that possess relevant behavioral adaptations, such as appropriate patterns of selectivity (Meuret & Provenza, 2015;Provenza et al., 2003), would therefore be important genetic resources, but the underlying genetics are incompletely understood. Specific regions of the sheep genome have been shown to have been under postdomestication selection (e.g., Fariello et al., 2014;Paim et al., 2018), but no such region directly relating to grazing behavior has yet been identified.
However, it is clear from quantitative genetics that some behavioral components of grazing behavior are heritable (Fogarty et al., 2009;Snowder et al., 2001), supporting the contention that grazing behavior is an outcome of genotype-environment interaction (Osoro et al., 1999;Rook, Dumont, et al., 2004). Nongenetic animal factors such as individual learning and social interactions Sibbald et al., 2008) modulated by traditional herding practices (Meuret & Provenza, 2015) will also be important. As a result, there is considerable, and poorly understood, individual variation in the structure of grazing behavior (Searle et al., 2010).
There have been many studies of how sheep under extensive husbandry range among plant communities (e.g., Baumont et al., 2000;Hester et al., 1999;Michelena et al., 2009), and of the nutritional consequences (e.g., Mobaek et al., 2012), but the structure of grazing behavior in semi-natural species-rich pastures of conservation interest has received relatively little attention. The types of grasslands where most experimental and modeling work has been conducted have been either relatively simple swards, or mosaics, where patches of favored plants have occurred within a matrix of less attractive vegetation. Little formal experimentation on the grazing of semi-natural pastures of conservation interest has been made (but see, e.g., Hellström et al., 2003), although some general principles exist. For example, there is evidence that plant species richness itself enhances foraging motivation, from both field studies (Arzak et al., 2016) and cafeteria feeding trials (Feng et al., 2016), suggesting the influence of vegetation on grazing behavior is not simply attributable to available biomass or to cover abundance of favored species (Parsons et al., 1994).
From a long-term study in a species-rich alpine grassland, Mysterud and Austrheim (2016) hypothesized that the "dietary composition" of sheep arises from a combination of "sheep preferences" and "immediate environment." We have tested this hypothesis in a similarly diverse vegetation environment, but with a more behavioral approach. Bite and step rates were taken as surrogates for dietary composition, and the focus has been on a single component of the immediate environment, namely, the vegetation within which bites are prehended. Sheep preference is clearly a multifactorial term, which was not defined by Mysterud and Austrheim (2016). As its surrogate, we have adopted sheep identity (the serial code of each animal) a term which encompasses all morphological, behavioral, experiential, and other variables in which one sheep might differ from another. Formally, we predict that bite and step rates are influenced by the categorical factors sheep identity, vegetation, day, and time of observation. Social behavior and grazing behavior are known to interact (e.g., Sibbald et al., 2008), so we investigated whether variation in social behavior, expressed as nearest-neighbor relationships, also varied according to vegetation.
In a novel development, we are taking as our "vegetation" term an objective multivariate assessment of the species composition of the area being grazed by the animal, without reference to the cover abundance (extent of ground covered) of individual plant species.
Assessment was by a standard technique (TWINSPAN: Kent & Coker, 1992), which, though well known to plant ecologists (Hearn et al., 2011), seems not to have been widely adopted by animal ecologists. This investigation of a hypothesis therefore also evaluates the use of a novel combination of ecological and behavioral approaches.

| Study area and animals
The study took place on semi-natural species-rich neutral Agrostis capillaris-Festuca rubra pasture with spatially variable hydrology and unfenced adjoining woodland, on a farm within Lahemaa National Park, in northern Estonia (59 o 30'N, 25 o 40'E), previously described by Hall et al. (2020). Most plant species present are not uncommon in Estonia, but the grass Holcus mollis, which was fairly abundant,

K E Y W O R D S
animal-borne video, grazing behavior, plant-animal interactions, sheep behavior is nationally rare (Kull et al., 2002). The woodland was not closedcanopy and the ground flora was relatively dense, such that when observing individual sheep in this area it was often not possible to see many other flock members.
Field work took place on five days (5-9 July 2019). The weather was cool with occasional rain. The sward condition was not quantified, but there was clearly adequate forage, substantially more than in the previous year (Hall et al., 2020, and personal observation).
The study flock, accompanied by four Maremmano guard dogs, in view of an apparent risk of predation from bears and wolves, was in permanent occupation and numbered approximately 50 ewes of the Estonian Native breed (Michelson, 2013). At the time of our study, the sheep were healthy, non-pregnant, non-lactating, and in good body condition. Sheep were handled by specialized personnel in accordance with European Union Directive No.609/1986, and the study was in full compliance with the ethical requirements of the Estonian University of Life Sciences.
The flock was housed in a barn at night and customarily foraged in an adjacent field for about 3 hr or 4 hr in the morning. Some foraging took place near the barn during the afternoon, but the main grazing period, typically of about 100 min, was in early evening when the dogs would accompany the flock across the field to a 5-ha pasture about 250 m from the barn. Four ewes were selected on the basis of diversity of body size, as we intended to maximize between-animal variability. Body weights were 29, 32, 42, and 46 kg and corresponding incisor breadths (measured from the marks made in a slab of dental impression material) were 29.5, 29, 32, and 34 mm. On each, a GoPro® Hero 5 Session camera (GoPro Inc., San Mateo, CA, USA) was fitted to a breast-bone harness so as to have a field of view, while harvesting vegetation, from the forefeet to the lower forward extremity of the muzzle. The area of the field of view was therefore not standardized but remained constant for each individual sheep.
The flock was confined briefly before the morning grazing period in order to attach the monitoring equipment. After about 2 hr, the flock and dogs returned spontaneously to the barn and the equipment was removed. Sheep were confined again, the equipment was downloaded, and batteries recharged. The process was repeated in late afternoon.

| Video, sound, and behavior monitoring
We set the cameras to record video and sound continuously. As well as providing good images of the vegetation and of steps, they also enabled registration of the bites that prehended vegetation (harvesting bites). Sounds of chewing (Galli et al., 2011) were not discerned.
The data storage capacity and battery capacity of these devices enabled generation of a maximum of six time-stamped consecutive MP4 files each of 17 min 42 s duration (total duration approximately 104 min). It was not possible to start the cameras remotely. After release, a period averaging 29 min (range 20-60 min) elapsed before the sheep had commenced feeding. Each focal (camera-fitted) sheep was watched by an observer. Our social behavior protocol was modified from that of Sibbald et al. (2005), every two minutes activity (grazing defined as prehension of vegetation, standing, walking, other behaviors) was noted. Social variables were also recorded, those used in the present analysis were the head-to-head distance (meters) to the nearest neighbor, estimated by eye, and whether the nearest neighbor had changed since the previous observation.
Observations were made during morning and afternoon on each of the five days. On some occasions, video records were absent or incomplete, mainly due to malfunctions of the harness or expiry of the camera battery, although behavior data continued to be collected. Observations ceased either when sheep returned apparently spontaneously to the barn, or at battery expiry. On some occasions, sheep were out of sight so only partial social behavior data were obtained.

| Merging behavior and vegetation data
Behavior observations yielded scan data (Martin & Bateson, 1993

| Vegetation
A list of species was recorded for each of the selected images, in the form of presence/absence data. In some cases, species name was assigned on the basis of apparent similarity to species that were familiar from experience in England and Scotland. In practice, none of the species thus designated were recorded in sufficient numbers to affect the results. We defined occurrence of fallen wood as a distinct species, indicating tree cover. Species richness (the number of plant species registered) was tabulated for each of the images.
When plants were only partially visible in the selected video frame, the video would be scrolled back or forwards by up to 3 s so the species could be accurately recorded. This often involved examining the area peripheral to the immediate field of view at the instant of the bite. No attempt was made to identify the actual plant species prehended. The bites and steps taken during 10 s before and 10 s after the instant of the video image were scored using Cowlog software (Pastell, 2016).

| TWINSPAN classification
Species lists were subjected to the TWINSPAN procedure in Community Analysis Package version 5 (Henderson & Seaby, 2014), which is based on an algorithm comparable with Principal Components Analysis. TWINSPAN places species lists on principal component axes and identifies the species, known as indicator species, which are the most influential in this placement. A divisive key is generated with indicator species at each division of the key. How the divisive key operates is not critical to the interpretation of our results, and we include it because reference to the indicator species helps with interpretation of the vegetation-animal relationships we have identified.
The first two axes, interpreted with reference to the known ecologies of the indicator species, enabled the underlying environmental gradient to be identified as being related to soil dampness, against a background of increasing shade. Species lists cluster on these axes, and these clusters are defined as vegetation classes. In this study, 30 of the 95 species recorded were identified as indicator species.
TWINSPAN then automatically allocated each image to a vegetation class. In principle, a divisive key would identify 32 classes at the fifth division, but in this study only 25 classes were identified as some of the divisions generated classes with too few members for further classification (classes 47-49, 52-55). For some subsequent analyses, we combined classes to give composite vegetation categories. These were designated Pasture1 (classes 32-35), Pasture2 (classes 36-46), and Woodland (classes 50-63).
As vegetation classes represent relative location on a principal components axis, we defined the first axis as the x-axis in graphical comparisons of their properties.
TWINSPAN is most familiar as a tool for characterizing the vegetation of a relatively large area. The divisive key is generated from an initial sample of relatively small plots, and this key is then applied throughout the area to be surveyed. Our procedure differed in that while the video images provided the initial sample, the resulting key was not then applied more widely. If we were to extend the study, by making a further sample from the video records, we would classify each image by using the key.

| Statistical analysis
Statistical tests were conducted in r (R Core Team, 2015, Crawley, 2013). Correlations were expressed as Pearson's r statistic.

| Grazing behavior
In a preliminary analysis, bite and step rates during the time period surrounding the video image were found not to be normally distributed so nonparametric statistical tests were used. Taking numbers of bites and numbers of steps during the 20 s surrounding the instant of each video image as the dependent variables, and considering images recorded in each vegetation class separately, the Kruskal-Wallis test was used to detect differences in the effects on bite and step rates of the two factors sheep identity, and day of observation.
Some patterns were only discernible when composite vegetation categories (Pasture1, Pasture2, Woodland), rather than vegetation classes, were used, so tests were repeated on this basis.
In the 2-min interval between observations, sheep could move from one vegetation class to another and this could influence the reliability of the assignment of bite and step rates, which were assessed over 20 s, to specific vegetation classes. Pairs of consecutive images were therefore characterized as to whether they revealed short-term change or constancy in the vegetation classes and composite vegetation categories grazed by the sheep.

| RE SULTS
From 41 animal-hours of video observation, 1,002 images were coincident with a behavior observation. All were used in the TWINSPAN analysis, but 27 subsequently proved unsuitable for analysis of bite and step rates, which are therefore based on the 975 images for which both rates were obtained (Table 1). Not all behavior variables could be tabulated for every image. Numbers of bites in the two main divisions (Pasture1 with Pasture2, and Woodland) were, respectively, 22,128 and 2,357, and of steps, 7,037 and 1,645, totaled over the 20-s periods associated with each of the 975 images.

| Vegetation
Ninety-six taxa of flowering plants, ferns, lichens, and mosses were distinguished. The mean number of species recorded per image was 4.96 (median 5, range 1-9). In addition, there was fallen wood (twigs and branches not assigned to species).

| Objective classification
Fallen wood and 29 species were identified by TWINSPAN as indicators. The classification, derived from 1,002 images ( Figure 1 and Figure 2), is presented in Figure 3 as a divisive key, with species name abbreviations in Table 2   Bites and steps were negatively correlated for most of the vegetation classes in the Pasture1 and Pasture2 composite vegetation classes (Table 4).

Paired comparisons of vegetation classes were possible for 875
images and the one immediately following (Table 5). In 28% of cases, during the 2 min between observations, the vegetation class being grazed remained the same. When these comparisons were on the basis of composite vegetation category, the proportion was 65%.

| Social behavior in relation to vegetation
Social behavior data were incomplete due to animals being temporarily out of sight, and the actual numbers of observations obtained are given in the tables. Distances to the nearest neighbor (n = 844, range 0-20 m) varied between sheep and days, but not between composite vegetation categories (Table 6).
F I G U R E 3 TWINSPAN classification. At each division, the relevant indicator species are listed (abbreviations: Table 1). The values in the square brackets are, respectively, the maximum indicator score for a sample to be directed upward, and the minimum indicator score for it to be directed downward. For example, regardless of what nonindicator species are present, a sample featuring the indicator species Agr cap, Rhy squ, and Ach mil would be directed upward at the first (net score −2), second (−2), third (−1), and fourth (−2) divisions, and downward at the fifth, to be placed in vegetation class 33. A sample with Fallen wood, Oxa ace and Fes rub would be directed downward at the first (net score +2) and second (+2) divisions, upward at the third (−2), downward at the fourth (+1), and fifth (+2) to be placed in vegetation class 59.

| D ISCUSS I ON
We emphasize the potential value to animal ecology of well- These approaches may not always be appropriate in the case of highly complex mosaics like our study area, which included pasture that varied spatially in relatively subtle ways, woodland, and boundary areas with at least 96 species in approximately 5 ha. Neither would they have enabled us to combine detailed behavioral studies with fine-scale study of vegetation.
We found that differences in vegetation were highly influential on bite rates and step rates when the woodland data were included, TA B L E 4 Variation among vegetation classes in the correlation between vegetation class and bite and step rates. Rates are expressed as the numbers of bites and steps recorded during the 20-s period centered on each video image

62-63
Woodland 11.5, 10 9.5, 2 −0.032 14 NS A possible alternative hypothesis on relationships between grazing behavior and vegetation is that the sheep are responding to plant species richness per se. This can be predicted from nutritional considerations (e.g., Provenza et al., 2003). Experimental support has come from Wang et al. (2010Wang et al. ( , 2011  One process by which vegetation influences grazing behavior could be that, in some environments, which could in principle be identified by TWINSPAN, favored and palatable plant species are unable to form large patches because of interspecies competition or a harsher microenvironment (Grime et al., 2007). In our case, Agrostis capillaris and Festuca rubra that are both palatable to sheep could be such species. This would be consistent with the bite rates in our study (median 78 per minute) being lower than those found by other workers, on less diverse pastures. For example, Penning et al. (1991) recorded rates of 86 and 96 bites per minute from ewes (79 kg body weight) on clover and grass, respectively, while , studying ewes (97 kg body weight) with video recording, reported 87 bites per minute on a ryegrass-clover sward.

Comparison
Bite rates are relatively easy to measure directly, but it is bite mass that is of pivotal importance in the understanding of foraging ecology (Shipley, 2007;Spalinger & Hobbs, 1992;Ungar, 1996), and the wide range of incisor breadths exhibited by our study animals will have led to corresponding differences in bite mass. While practical demonstration would be difficult, we predict that bite mass will also vary among vegetation classes.
Sheep grazing behavior is the result of the interaction of many factors, including individual variation, the behavior of other sheep, and the vegetation. Implying a prioritization of social behavior over vegetation type, Sibbald et al. (2008) interpreted this interaction in Scottish Blackface sheep on a heather-grass mosaic, as leading sheep to favor patches of vegetation that enabled them to graze at their preferred spacing. The mean nearest-neighbor distance in that study was 4.9 m, much greater than those observed in our study  1-2 m). The close spacing evident in our study could be a breed characteristic or a consequence of management, although perceived predation risk and presence of guard dogs (Webber et al., 2015) might be influential.
Differences in bite rate among sheep were greater than differences in step rate. The relative uniformity of step rates between the vegetation classes of the Pasture1 and Pasture2 composite vegetation categories suggests that step rates are influenced, to some extent, by reference to the rest of the visible flock, and may indication a motivation towards flocking behavior (Dwyer, 2009). There was a negative correlation overall between step rate and bite rate consistent with the sheep maintaining step rate presumably in response to a motivation to remain close to other sheep, with their bite rates then varying accordingly. The direct relevance of bite rate and step rate to conservation outcomes does not appear to have been investigated, but Adler et al. (2001) suggested that spatial heterogeneity of vegetation would be expected to increase with patch grazing and to decrease with homogeneous grazing (in other words, to be related to the patterns of interspersion of steps and bites; Hall et al., 2020).
Classification at the level of vegetation class enabled the relative importance of animal and plant factors on bite and step rates to be assessed, but the statistically significant results in relation to social behavior mostly derive from comparisons of composite vegetation categories, which are aggregates of vegetation classes. The period of 20 s that we used could have reduced the distinctiveness, in terms of bite and step rates, of the short-term response of sheep to vegetation differences. We consider that detailed methodological development is beyond the scope of this report, but suggest that the evident sensitivity of our approach to the mosaic structure of grasslands may well be of value.
Our finding of the substantial influence of animal individuality in relation to bite rate parallels the earlier findings of Melin et al. (2005), that 84%-98% of the variation in feeding patterns of housed dairy cows was due to individual differences. It also emphasizes how elucidating the well-established effect on foraging behavior of animal individuality is critical to the scientific underpinning of conservation grazing (Rook, Dumont, et al., 2004). For this understanding to be applied, the genetics of grazing behavior need to be characterized (Rook, Dumont, et al., 2004). Currently, the characterization extends little further than a general understanding that sheep breeds may differ from each other in this respect, with body size being particularly important (Rook, Dumont, et al., 2004). The findings of Provenza et al. (2003) suggest that this diversity could have arisen from a diversity of grazing practices. Those based on relatively low stocking rates on pasture that is maintained at relatively high levels of available biomass may have had the effect of selecting for genotypes that graze selectively. A more episodic type of grazing at a high stocking rate would have promoted evolution for lower selectivity.
Our findings have relevance to the practical management of conservation grazing initiatives. At the technical level, the value of objective classification for assessing vegetation change has already been shown (Hall & Bunce, 2019) and practical characterization of foraging ecology could be made against this background. At the management level, this characterization could inform breeding policy within the flock or herd, or could help with the development of alternative approaches, for example specific management practices such as training animals (Thorhallsdottir et al., 1990).
There are also consequences for policymakers and for the direction of research activity. Sheep husbandry can sustain rural livelihoods but can also damage the environment and therefore poses significant challenges to policy makers (see, e.g., Bunce et al., 2018aBunce et al., , 2018b. While sheep grazing can promote floristic diversity (Metera et al., 2010), it can also damage sensitive species (Kose et al., 2019). A better understanding of the comparative importance of animal and of vegetation factors in modulating grazing behavior could contribute to resolving such conflicts.
On the broader scale, the sustainability, under conditions of global climate change, of pasture-based livestock farming may depend on the availability of animal genotypes that possess relevant behavioral adaptations. Development of new breeds for sustainable commercial exploitation of rangelands is conceivable, and in such a situation, artificial selection could be applied to behavioral components of the grazing process (Beausoleil et al., 2012). Indeed, breeding for behavior traits may be preferable to breeding for anatomical adaptations. For example, selecting for bite mass, which is strongly associated with incisor arcade breadth, would also tend to select for body size (Milner et al., 2000). A change in body size could be undesirable for commercial and husbandry reasons, and selection for a behavioral trait could avoid the consequences of unfavorable genetic correlations of this kind.
Quantitative field studies of foraging ecology, in association with objective vegetation classification, have potential to elucidate the adaptability to extensive husbandry of livestock breeds. Use of appropriate genetic resources, whether as pre-existing breeds or as the result of breeding programs within current commercial breeds, could contribute to the sustainability of extensive husbandry and to the conservation of biodiversity in species-rich and semi-natural pastoral agroecosystems.

ACK N OWLED G M ENTS
Imbi and Taavi Jaetma kindly made available their animals and facilities. Giovanni Marcone, Maria Soonberg, and Mathilde Frances Lind helped with field observations and data analysis. We are deeply grateful to reviewers of a previous version of the manuscript for their care and sympathetic consideration.

CO N FLI C T O F I NTE R E S T
None declared.