Behavior Classification and Analysis of Grazing Sheep on Pasture with Different Sward Surface Heights Using Machine Learning

Abstract

Simple Summary
The monitoring and analysis of sheep behavior can reflect their welfare and health, which is beneficial for grazing management. For automatic classification and the continuous monitoring of grazing sheep behavior, wearable devices based on inertial measurement unit (IMU) sensors are important. The accuracy of different machine learning algorithms was compared, and the best one was used for the continuous monitoring and behavior classification of three grazing sheep on pasture with three different sward surface heights. The results showed that the algorithm automatically monitored the behavior of grazing sheep individuals and quantified the time of each behavior.

Abstract
Behavior classification and recognition of sheep are useful for monitoring their health and productivity. The automatic behavior classification of sheep by using wearable devices based on IMU sensors is becoming more prevalent, but there is little consensus on data processing and classification methods. Most classification accuracy tests are conducted on extracted behavior segments, with only a few trained models applied to continuous behavior segments classification. The aim of this study was to evaluate the performance of multiple combinations of algorithms (extreme learning machine (ELM), AdaBoost, stacking), time windows (3, 5 and 11 s) and sensor data (three-axis accelerometer (T-acc), three-axis gyroscope (T-gyr), and T-acc and T-gyr) for grazing sheep behavior classification on continuous behavior segments. The optimal combination was a stacking model at the 3 s time window using T-acc and T-gyr data, which had an accuracy of 87.8% and a Kappa value of 0.836. It was applied to the behavior classification of three grazing sheep continuously for a total of 67.5 h on pasture with three different sward surface heights (SSH). The results revealed that the three sheep had the longest walking, grazing and resting times on the short, medium and tall SHH, respectively. These findings can be used to support grazing sheep management and the evaluation of production performance.

Full text

Citation: Jin, Z.; Guo, L.; Shu, H.; Qi,
J.; Li, Y.; Xu, B.; Zhang, W.; Wang, K.;
Wang, W. Behavior Classification and
Analysis of Grazing Sheep on Pasture
with Different Sward Surface Heights
Using Machine Learning. Animals
2022,12, 1744. https://doi.org/
10.3390/ani12141744
Academic Editors: Guoming Li, Yijie
Xiong and Hao Li
Received: 9 May 2022
Accepted: 5 July 2022
Published: 7 July 2022
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
animals
Article
Behavior Classification and Analysis of Grazing Sheep on
Pasture with Different Sward Surface Heights Using
Machine Learning
Zhongming Jin 1, Leifeng Guo 1, *, Hang Shu 1,2 , Jingwei Qi 3, Yongfeng Li 1, Beibei Xu 1, Wenju Zhang 1,
Kaiwen Wang 1,4 and Wensheng Wang 1,*
1Agricultural Information Institute, Chinese Academy of Agriculture Sciences, Beijing 100086, China;
jinzhongming@caas.cn (Z.J.); hang.shu@doct.uliege.be (H.S.); yongfeng.li@student.uliege.be (Y.L.);
xuxiaobei224@163.com (B.X.); zhangwenju@caas.cn (W.Z.); kaiwen.wang@wur.nl (K.W.)
2AgroBioChem/TERRA, Precision Livestock and Nutrition Unit, Gembloux Agro-Bio Tech,
University of Liège, 5030 Gembloux, Belgium
3College of Animal Science, Inner Mongolia Agricultural University, Hohhot 010018, China;
qijingwei_66@126.com
4
Information Technology Group, Wageningen University and Research, 6708 PB Wageningen, The Netherlands
*Correspondence: guoleifeng@caas.cn (L.G.); wangwensheng@caas.cn (W.W.)
Simple Summary:
The monitoring and analysis of sheep behavior can reflect their welfare and health,
which is beneficial for grazing management. For automatic classification and the continuous monitoring
of grazing sheep behavior, wearable devices based on inertial measurement unit (IMU) sensors are
important. The accuracy of different machine learning algorithms was compared, and the best one was
used for the continuous monitoring and behavior classification of three grazing sheep on pasture with
three different sward surface heights. The results showed that the algorithm automatically monitored
the behavior of grazing sheep individuals and quantified the time of each behavior.
Abstract:
Behavior classification and recognition of sheep are useful for monitoring their health and
productivity. The automatic behavior classification of sheep by using wearable devices based on IMU
sensors is becoming more prevalent, but there is little consensus on data processing and classification
methods. Most classification accuracy tests are conducted on extracted behavior segments, with only
a few trained models applied to continuous behavior segments classification. The aim of this study
was to evaluate the performance of multiple combinations of algorithms (extreme learning machine
(ELM), AdaBoost, stacking), time windows (3, 5 and 11 s) and sensor data (three-axis accelerometer
(T-acc), three-axis gyroscope (T-gyr), and T-acc and T-gyr) for grazing sheep behavior classification on
continuous behavior segments. The optimal combination was a stacking model at the 3 s time window
using T-acc and T-gyr data, which had an accuracy of 87.8% and a Kappa value of 0.836. It was applied
to the behavior classification of three grazing sheep continuously for a total of 67.5 h on pasture with
three different sward surface heights (SSH). The results revealed that the three sheep had the longest
walking, grazing and resting times on the short, medium and tall SHH, respectively. These findings
can be used to support grazing sheep management and the evaluation of production performance.
Keywords:
behavior classification; grazing sheep; machine learning; sward surface heights;
behavior distribution
1. Introduction
Sheep provide a variety of products, and their environment and health are important
factors that affect production performance. As the temperature of the microenvironment
increases, heat stress significantly impairs the efficiency of meat and wool production [
1
].
Diseases not only affect sheep, but are also a major cause of economic loss for the sheep
industry [
2
]. The behavior of livestock can reflect their response to the environment and
Animals 2022,12, 1744. https://doi.org/10.3390/ani12141744 https://www.mdpi.com/journal/animals

Animals 2022,12, 1744 2 of 24
health [
3
]. Li et al. [
4
] pointed out that the lying time of Small Tail Han sheep increases
significantly when the ambient temperature rises, so forage intake is reduced to lower heat
production in sheep. When ruminal bloat happens, grazing and rumination slow down or
stop, and sheep keep getting up and lying down [
5
]. The duration of resting behavior can
reflect their social stress in animal husbandry [
6
]. Therefore, the continuous monitoring
and analysis of sheep behavior relays their welfare and health status in a timely manner,
which leads to the formulation of measures to improve their welfare, reduce economic
losses and help achieve efficient and sustainable development [7].
Traditional manual observation needs an observer to record livestock behavior, which
is both time- and labor-consuming and potentially has an impact on normal livestock behav-
ior [
8
,
9
]. Moreover, it makes continuous monitoring impossible, especially when there is a
large quantity and wide distribution of livestock [
3
]. Many studies [
10
–
13
] pointed out that
computer vision, sound analysis, motion sensing, satellite positioning and other technolo-
gies have been used to improve the ability of remote, large-scope and large-scale monitoring
of livestock behavior. The wearable motion sensor is more suitable for fine-scale monitoring
of the behavior of grazing sheep than computer vision and satellite positioning [
14
], and has
potential application prospects. Motion sensors have been widely used in cattle to identify
behavior, and many products—-for example, Lely [
15
], MooMonitors [
16
], IceTag3D
™
,
REDI, SCR/Allflex and CowManager Sensor systems [
17
]—are commercially available.
Due to the differences in physiology and behavior, the products of cattle monitoring cannot
be applied directly to sheep [
18
]. Moreover, research into the behavior classification of
grazing sheep is less than that of cattle [
19
]. As such, using motion sensors to monitor the
behavior of sheep is of great interest.
The behavior monitored for sheep generally includes grazing, lying, standing, walking,
ruminating, running, etc. The main factors affecting the classification accuracy of a motion
sensor are the wearing position, sensor type, data collection frequency, time window size,
feature construction and algorithm. However, the above factors chosen to achieve an
optimum classification accuracy of sheep behavior varied in previous studies (Table 1),
and very few of the trained models have been applied to continuous behavior segments
in actual situations. Therefore, the aim of this study was to assess the performance of a
range of machine learning (ML) algorithms (extreme learning machine (ELM), AdaBoost,
stacking) in classifying walking, standing, grazing, lying and running behavior of grazing
sheep at three different time windows (3, 5 and 11 s) using three different sensor data types
(three-axis accelerometer (T-acc), three-axis gyroscope (T-gyr), and T-acc and T-gyr). The
focus of this study was to evaluate multiple combinations of algorithms, time windows and
sensor data. The optimal combination was selected to analyze the behavior distribution of
three sheep grazing continuously on pasture with different sward surface heights (SSH).

Animals 2022,12, 1744 3 of 24
Table 1. Overview of algorithms and other parameters for sheep behavior classification to achieve the best classification accuracy.
Sensor Position Sensor Data Collection Frequency Time Window
The Number of
Features Finally Used
for Classification
Type of Behavior Algorithm Accuracy Source
Neck Tri-axial accelerometer 100 Hz 5.12 s 10
Lying
Standing
Walking
Running
Grazing
Quadratic Discriminant Analysis 89.7% Marais et al. [20]
Under-jaw Three-axis accelerometer 5, 10, 25 Hz 5 s 5
Grazing
Lying
Running
Standing
Walking
Decision Tree 85.5% Alvarenga et al. [8]
Neck Three-dimensional
accelerometer 100 Hz 5.3 s 27
Standing
Walking
Grazing
Running
Lying
Linear Discriminant Analysis 82.40% Le Roux et al. [21]
Ear Neck Tri-axial accelerometer 32 Hz 5, 7 s 44
Walking
Standing
Lying
Random Forest 95%F-score: 91–97% Walton et al. [22]
Neck Tri-axial accelerometer
and gyroscope 16 Hz 7 s 39 Grazing
Ruminating Random Forest 92% Mansbridge et al. [18]
Ear
Neck Leg Tri-axial accelerometer 12 Hz 10 s 14
Grazing
Standing
Walking
Quadratic Discriminant Analysis 94–99% Barwick et al. [9]
Under-jaw Three-axial accelerometer and
a force sensor 62.5 Hz 30 s 15
Grazing
Ruminating
Other activities
Discriminant Analysis 89.7% Decandia et al. [19]
Rear Accelerometers 40 Hz 3 s 30
Foraging
Walking
Running
Standing
Lying
Urinating
Random Forest 0.945(Kappa value) Lush et al. [23]
Ear Accelerometers 12.5 Hz 10 s 19
Grazing
Lying
Standing
Walking
Support Vector Machine 76.9% Fogarty et al. [3]
Neck Three-axial accelerometer and
ultrasound transducer 50 Hz 5 s 11
Infracting
Eating
Moving
Running
Standing
Invalid
Decision Tree 91.78% Nóbrega et al. [24]

Animals 2022,12, 1744 4 of 24
2. Materials and Methods
2.1. Experimental Site, Animals, and Instrumentation
This study was approved by the Animal Ethics Committee of the University of New
England, and followed the Code of Research Conduct for the University of New England,
to conform to the Australian Code for Care and Use of Animals (AEC17-006).
The experimental site was located in the ryegrass pasture (
−
30.5, 151.6) near the Uni-
versity of New England, New South Wales, Australia. To study the behavior distribution of
sheep on different SSH, three paddocks of 72 m
2
(48 m long
×
1.5 m wide) were set up, and
the SSH was cut to 2–3 cm (short), 5–6 cm (medium) and 8–10 cm (tall), as shown in Figure 1.
A water trough was provided at one end of each paddock for the sheep to drink from.
Animals 2022, 12, x FOR PEER REVIEW 4 of 24
B-Sheep, P-Sheep and R-Sheep were assigned. The IMU data of grazing sheep on pastures
with different SSH were collected for 5 consecutive days: at the short SSH on the 15th and
16th, at the medium SSH on the 17th and 18th, and at the tall SSH on the 19 May 2017. To
record the sheep behavior, a camera was fixed at one end of the paddock, while the cap-
tured videos were stored in the camera’s SD card. Sheep entered the paddock at 8:00 a.m.
on each test day and the camera began to collect data for 4 h.
Figure 1. Experimental paddock on pastures with different sward surface heights (SSH). Location
of inertial measurement unit (IMU) sensor and its orientation on sheep.
2.2. Sheep Behavior Definition and Labelling
Based on previous studies and the situation of this study, the behavior of grazing
sheep was classified into five categories: walking, standing, grazing, lying and running.
These behaviors are described in Table 2.
Table 2. Definition of the five behaviors for classification. Definition adapted from [3,9,17,22,24].
Behavior Description
Walking
The head moves forward/backward or sideways for at least two consecutive steps. From one place
to another, the legs on the diagonal of the sheep move at the same time. Slow movement during
grazing is excluded.
Standing Sheep in standing position. The limbs and head are still or slightly moved, including standing
chewing and ruminating.
Grazing Sheep graze with their heads down, chew and move slowly to find grass.
Lying Sheep in lying position. The head is down or up, and still or slightly moving. Chewing and rumi-
nating are included.
Running Sheep run faster to escape obstacles or catch up with other companions. In most cases, two
front/rear legs move at the same time, and there is no biting or chewing.
The video data of the three sheep collected on the 15th and 18th of May were labelled
using the behavior labelling software developed by the authors. A camera was placed at
one end of the paddock, but the sheep that were too far away or were too close together
were labelled as unknown because their specific behavior could not be observed. Each
video contained behavior records of the three sheep. To complete the behavior labelling
of the three sheep in the video, each video had to be labelled three times. Video labelling
Figure 1.
Experimental paddock on pastures with different sward surface heights (SSH). Location of
inertial measurement unit (IMU) sensor and its orientation on sheep.
Three 8-month-old Merino sheep of approximately 35 kg were used in this study. A
designed wearable device based on InvenSense MPU-9250 was worn on the neck of sheep
to collect inertial measurement unit (IMU) data. The collected data were stored in the
device’s secure-digital (SD) card. MPU-9250 provides a T-acc, a T-gyr and a three-axis
magnetometer. The x-, y-, and z-axis represent movement in vertical, horizontal, and lateral
directions, respectively (as shown in Figure 1). In more detail, the collection frequency of
IMU data was 20 Hz, the range of the T-acc value was
±
2 g (9.8 m/s
2
), and the range of
the T-gyr value was
±
2000 dps (
◦
/s). To distinguish each sheep clearly in the shooting
video, blue, purple, and red livestock-marking pigments were used as markers; hence, the
names B-Sheep, P-Sheep and R-Sheep were assigned. The IMU data of grazing sheep on
pastures with different SSH were collected for 5 consecutive days: at the short SSH on
the 15th and 16th, at the medium SSH on the 17th and 18th, and at the tall SSH on the
19 May 2017
. To record the sheep behavior, a camera was fixed at one end of the paddock,
while the captured videos were stored in the camera’s SD card. Sheep entered the paddock
at 8:00 a.m. on each test day and the camera began to collect data for 4 h.
2.2. Sheep Behavior Definition and Labelling
Based on previous studies and the situation of this study, the behavior of grazing
sheep was classified into five categories: walking, standing, grazing, lying and running.
These behaviors are described in Table 2.

Animals 2022,12, 1744 5 of 24
Table 2. Definition of the five behaviors for classification. Definition adapted from [3,9,17,22,24].
Behavior Description
Walking The head moves forward/backward or sideways for at least two consecutive steps. From one place to another, the legs on the
diagonal of the sheep move at the same time. Slow movement during grazing is excluded.
Standing Sheep in standing position. The limbs and head are still or slightly moved, including standing chewing and ruminating.
Grazing Sheep graze with their heads down, chew and move slowly to find grass.
Lying Sheep in lying position. The head is down or up, and still or slightly moving. Chewing and ruminating are included.
Running
Sheep run faster to escape obstacles or catch up with other companions. In most cases, two front/rear legs move at the same time,
and there is no biting or chewing.
The video data of the three sheep collected on the 15th and 18th of May were labelled
using the behavior labelling software developed by the authors. A camera was placed at
one end of the paddock, but the sheep that were too far away or were too close together
were labelled as unknown because their specific behavior could not be observed. Each
video contained behavior records of the three sheep. To complete the behavior labelling
of the three sheep in the video, each video had to be labelled three times. Video labelling
data were linked with IMU data by the corresponding timestamp and used as the label
of behavior. When constructing the dataset, all unknown labels were discarded, leaving
behind only the behavior label that could be observed clearly (Table 3). Data on the running
behavior in the labelled dataset were rare because the experimental site was very safe and
there were no threats that would make the sheep run away. In this study, the sheep ran
mainly because they touched the electrified fence of the paddock. However, this happened
very infrequently, which led to a very unbalanced dataset of five behaviors.
Table 3.
The total time and data size of observations available for each behavior. A row comprises an
x-, y- and z-axis accelerometer and gyroscope raw data.
Behavior Number Behavior Total Time (s) Total Data Rows
1 Walking 4352 87,040
2 Standing 15,499 309,980
3 Grazing 24,374 487,480
4 Lying 9534 190,680
5 Running 97 1940
Guo et al. [
25
] found that it was robust to use the model trained on a specific SSH to
classify grazing and non-grazing behavior of grazing sheep on pastures with different SSH
(2–10 cm). Therefore, this study adopted the data training model of three sheep on the 15th
and 18th of May, and applied the trained model on the 16th, 17th and 19th of May to the
7.5 h continuous behavior segments classification of the three grazing sheep on pastures
with three different SSH. To evaluate the practical application performance of the model
for sheep behavior classification in continuous behavior segments, part of the data of the
five behaviors were randomly labelled by video as a continuous behavior segments test
dataset (CBS test dataset), which included three sheep on pasture with three SSHs walking
for 1250 s, standing for 1800 s, grazing for 1800 s, lying for 1800 s, and running for 37 s.
2.3. Wavelet Transform Denoising
It was necessary to reduce the noise of the collected data as the collected IMU data
were inevitably disturbed by noise. The various sheep behaviors were completed by
various specific movements, each represented by different T-acc and T-gyr data, so it
was particularly important to save the peak signals and changing data signals for the
behavior representation. Wavelet filtering effectively filtered out the noise while retaining
the peak and mutation values to the maximum extent. The wavelet denoising experiment
for collected IMU data was carried out using different thresholds and rules in MATLAB.
Discrete wavelet db6 was selected as the basis function, and the raw data were decomposed
by five layers of wavelet. According to heursure, quantization was carried out under

Animals 2022,12, 1744 6 of 24
a soft threshold [
26
]. In the end, wavelet reconstruction was carried out to complete
wavelet transform denoising for the collected IMU data. Wavelet transformation effectively
removed the high-frequency noise from the static behavioral data of sheep, while retaining
change data in the dynamic behavior (Figure 2).
Animals 2022, 12, x FOR PEER REVIEW 6 of 24
(a) (b)
Figure 2. Comparison of the signal before and after wavelet denoising. (a) Time series of x-axis ac-
celerometer signal from 20 Hz sampling rate for observed behaviors of 10 s lying; (b) time series of
x-axis accelerometer signal from 20 Hz sampling rate for observed behaviors of 10 s walking.
2.4. Time Window Size Selection
The sheep behavior was not instantaneous but consisted of specific movements
throughout a period. During the experiment, 20 records were collected every second, but
each record was obviously insufficient to represent a behavior. To solve this problem, pre-
vious studies usually used the windowing method to complete data classification, and the
data in each time window were used to represent a kind of behavior. Studies found that
the dynamic behavior of an animal’s daily activity changed periodically [26], so it was
more reasonable to determine the time window of the specific behavior based on the
movement period of the animals’ behavior. Since dynamic behavior has a stronger move-
ment periodicity than static behavior, we analyzed the period of three dynamic behaviors
(walking, running, and grazing) to determine the appropriate time window for behavior
classification.
It was found that the typical walking and running behaviors of sheep have strong
periodicity. The T-acc and T-gyr signals of a typical 10 s of walking behavior of sheep are
shown in Figure 3.
Figure 2.
Comparison of the signal before and after wavelet denoising. (
a
) Time series of x-axis
accelerometer signal from 20 Hz sampling rate for observed behaviors of 10 s lying; (
b
) time series of
x-axis accelerometer signal from 20 Hz sampling rate for observed behaviors of 10 s walking.
2.4. Time Window Size Selection
The sheep behavior was not instantaneous but consisted of specific movements through-
out a period. During the experiment, 20 records were collected every second, but each record
was obviously insufficient to represent a behavior. To solve this problem, previous studies
usually used the windowing method to complete data classification, and the data in each
time window were used to represent a kind of behavior. Studies found that the dynamic
behavior of an animal’s daily activity changed periodically [
26
], so it was more reasonable
to determine the time window of the specific behavior based on the movement period of
the animals’ behavior. Since dynamic behavior has a stronger movement periodicity than
static behavior, we analyzed the period of three dynamic behaviors (walking, running, and
grazing) to determine the appropriate time window for behavior classification.
It was found that the typical walking and running behaviors of sheep have strong
periodicity. The T-acc and T-gyr signals of a typical 10 s of walking behavior of sheep are
shown in Figure 3.
As shown in Figure 4, the x-, y- and z-axis signals of the T-acc and T-gyr in Figure 3
were subjected to fast Fourier transformation. At the same time, the dominant frequency
and period were calculated using the Formulas (1) and (2).
f=(n−1)×fs
N(fsis the sam pling f requency,which is 20Hz in this study)(1)
T=1
f(2)
The maximum and minimum periods for calculating the typical walking behavior
segment (Figure 3) of sheep were 0.91 and 0.29 s, respectively. The maximum and mini-
mum frequencies were 3.5 and 1.1 Hz, respectively. X-axis accelerometer (x-acc), y-axis
gyroscope (y-gyr) and z-axis accelerometer (z-acc) signals had similar periods, while y-axis
accelerometer (y-acc) and z-axis gyroscope (z-gyr) signals had the same periods. Observing
the walking behavior of sheep through the video found that (i) to (iv) in Figure 5was
considered a complete period of a sheep’s walking. It was considered that the walking
time was about 0.29 s from (i) to (ii), about 0.45 s from (i) to (iii), and about 0.91 s from
(i) to (iv) in the 10 s walking behavior segment shown in Figure 3. A total of 24 typical

Animals 2022,12, 1744 7 of 24
walking segments of sheep were observed by video, and the average period was 0.93 s and
the maximum period was 1.25 s.
Animals 2022, 12, x FOR PEER REVIEW 6 of 24
(a) (b)
Figure 2. Comparison of the signal before and after wavelet denoising. (a) Time series of x-axis ac-
celerometer signal from 20 Hz sampling rate for observed behaviors of 10 s lying; (b) time series of
x-axis accelerometer signal from 20 Hz sampling rate for observed behaviors of 10 s walking.
2.4. Time Window Size Selection
The sheep behavior was not instantaneous but consisted of specific movements
throughout a period. During the experiment, 20 records were collected every second, but
each record was obviously insufficient to represent a behavior. To solve this problem, pre-
vious studies usually used the windowing method to complete data classification, and the
data in each time window were used to represent a kind of behavior. Studies found that
the dynamic behavior of an animal’s daily activity changed periodically [26], so it was
more reasonable to determine the time window of the specific behavior based on the
movement period of the animals’ behavior. Since dynamic behavior has a stronger move-
ment periodicity than static behavior, we analyzed the period of three dynamic behaviors
(walking, running, and grazing) to determine the appropriate time window for behavior
classification.
It was found that the typical walking and running behaviors of sheep have strong
periodicity. The T-acc and T-gyr signals of a typical 10 s of walking behavior of sheep are
shown in Figure 3.
Figure 3.
Time series of three-axis accelerometer and three-axis gyroscope signals from a 20 Hz
sampling rate for observed behavior of 10 s walking. Acceleration is in g (9.8 m/s2) units.
The T-acc and T-gyr signals of a typical 5 s of running behavior of sheep are shown in
Figure 6. The x-, y- and z-axis signals of T-acc and T-gyr were, respectively, subjected to fast
Fourier transformation, as shown in Figure 7. At the same time, the dominant frequency
and period were calculated.
The x-acc, y-gyr and z-acc signals that were used to calculate the typical running
behavior segment of sheep in Figure 6had similar periods, while the x-axis gyroscope
(x-gyr), y-acc and z-gyr signals had the same periods. The maximum and minimum periods
were 0.45 and 0.21 s, respectively. The maximum and minimum frequencies were 4.8 and
2.3 Hz, respectively. The maximum period of the running behavior segment of the sheep
(Figure 6) was about half of that of the walking behavior segment (Figure 3).
Figure 8presents the T-acc and T-gyr signals of a typical 10 s of grazing behavior of
sheep. The x-, y- and z-axis signals were, respectively, subjected to fast Fourier transforma-
tion, and the dominant frequency and period were calculated at the same time.
Compared with typical walking and running behavior, grazing had no significant
periodicity. Moreover, through observation, it was found that the period of sheep grazing
behavior on pasture with different SSH was also different. Therefore, the period of sheep
grazing behavior was mainly determined by video observation. The process of observing a
sheep’s grazing behavior could be roughly divided into: (i) biting once or several times and
then swallowing; (ii) biting once or several times, then chewing and finally swallowing;
(iii) grazing, then biting once or several times and finally swallowing; (iv) grazing, then
biting once or several times and then chewing (or chewing while foraging), and finally
swallowing. Due to the biting movement of sheep, the video was easily observed, and the
time interval between two biting movements was taken to be the grazing period. A total of
41 grazing segments of sheep in the video were observed, and the duration of each one
divided by the number of biting movements was taken as the period of grazing behavior:
the maximum period was 2.15 s, which was longer than the maximum period of walking
by 1.25 s. Considering that the period of walking behavior was about twice that for running

Animals 2022,12, 1744 8 of 24
behavior, the maximum period for observing dynamic behavior was 2.15 s. Therefore, a
minimum time window of 3 s was enough to satisfy a behavioral movement period; time
windows of 3, 5 and 11 s were used for time window comparison, being the maximum
period of 2.15 s rounded up 1, 2, and 5 times.
Animals 2022, 12, x FOR PEER REVIEW 7 of 24
Figure 3. Time series of three-axis accelerometer and three-axis gyroscope signals from a 20 Hz
sampling rate for observed behavior of 10 s walking. Acceleration is in g (9.8 m/s2) units.
As shown in Figure 4, the x-, y- and z-axis signals of the T-acc and T-gyr in Figure 3
were subjected to fast Fourier transformation. At the same time, the dominant frequency
and period were calculated using the Formulas (1) and (2).
𝑓
=(𝑛−1)×

(
𝑓
 𝑖𝑠 𝑡ℎ𝑒 𝑠𝑎𝑚𝑝𝑙𝑖𝑛𝑔
𝑓
𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦, 𝑤ℎ𝑖𝑐ℎ 𝑖𝑠 20𝐻𝑧 𝑖𝑛 𝑡ℎ𝑖𝑠 𝑠𝑡𝑢𝑑𝑦) (1)
𝑇=1
𝑓
(2)
(a)
(b)
(c)
(d)
(e) (f)
Figure 4. Fourier transform performed on the three-axis accelerometer and three-axis gyroscope
signals of the 10 s walking behavior shown in Figure 3 to calculate the main frequency and period.
(a) x-axis accelerometer: n = 23; N = 200; f = 2.2 Hz; T ≈ 0.45 s; (b) x-axis gyroscope: n = 12; N = 200; f
= 1.1 Hz; T ≈ 0.91 s; (c) y-axis accelerometer: n = 36; N = 200; f = 3.5 Hz; T ≈ 0.29 s; (d) y-axis gyroscope:
n = 24; N = 200; f = 2.3 Hz; T ≈ 0.43 s; (e) z-axis accelerometer: n = 24; N = 200; f = 2.3 Hz; T ≈ 0.43 s; (f)
z-axis gyroscope: n = 36; N = 200; f = 3.5 Hz; T ≈ 0.29 s.
Figure 4.
Fourier transform performed on the three-axis accelerometer and three-axis gyroscope
signals of the 10 s walking behavior shown in Figure 3to calculate the main frequency and period.
(a)x-axis accelerometer: n= 23; N= 200; f= 2.2 Hz; T≈0.45 s; (b)x-axis gyroscope: n= 12; N= 200;
f= 1.1 Hz; T
≈
0.91 s; (
c
)y-axis accelerometer: n= 36; N= 200; f= 3.5 Hz; T
≈
0.29 s; (
d
)y-axis
gyroscope: n= 24; N= 200; f= 2.3 Hz; T
≈
0.43 s; (
e
)z-axis accelerometer: n= 24; N= 200; f= 2.3 Hz;
T≈0.43 s; (f)z-axis gyroscope: n= 36; N= 200; f= 3.5 Hz; T≈0.29 s.

Animals 2022,12, 1744 9 of 24
Animals 2022, 12, x FOR PEER REVIEW 8 of 24
The maximum and minimum periods for calculating the typical walking behavior
segment (Figure 3) of sheep were 0.91 and 0.29 s, respectively. The maximum and mini-
mum frequencies were 3.5 and 1.1 Hz, respectively. X-axis accelerometer (x-acc), y-axis
gyroscope (y-gyr) and z-axis accelerometer (z-acc) signals had similar periods, while y-
axis accelerometer (y-acc) and z-axis gyroscope (z-gyr) signals had the same periods. Ob-
serving the walking behavior of sheep through the video found that (i) to (iv) in Figure 5
was considered a complete period of a sheep’s walking. It was considered that the walking
time was about 0.29 s from (i) to (ii), about 0.45 s from (i) to (iii), and about 0.91 s from (i)
to (iv) in the 10 s walking behavior segment shown in Figure 3. A total of 24 typical walk-
ing segments of sheep were observed by video, and the average period was 0.93 s and the
maximum period was 1.25 s.
Figure 5. Decomposition movement of sheep’s typical walking behavior in one period. (i) The start-
ing state of walking behavior; (ii) the sheep lifts the left leg; (iii) the body moves forward with the
left leg; and (iv) the sheep lifts the right leg and moves forward with the right leg to the starting
state.
The T-acc and T-gyr signals of a typical 5 s of running behavior of sheep are shown
in Figure 6. The x-, y- and z-axis signals of T-acc and T-gyr were, respectively, subjected
to fast Fourier transformation, as shown in Figure 7. At the same time, the dominant fre-
quency and period were calculated.
Figure 5.
Decomposition movement of sheep’s typical walking behavior in one period. (
i
) The starting
state of walking behavior; (
ii
) the sheep lifts the left leg; (
iii
) the body moves forward with the left leg;
and (iv) the sheep lifts the right leg and moves forward with the right leg to the starting state.
Animals 2022, 12, x FOR PEER REVIEW 8 of 24
The maximum and minimum periods for calculating the typical walking behavior
segment (Figure 3) of sheep were 0.91 and 0.29 s, respectively. The maximum and mini-
mum frequencies were 3.5 and 1.1 Hz, respectively. X-axis accelerometer (x-acc), y-axis
gyroscope (y-gyr) and z-axis accelerometer (z-acc) signals had similar periods, while y-
axis accelerometer (y-acc) and z-axis gyroscope (z-gyr) signals had the same periods. Ob-
serving the walking behavior of sheep through the video found that (i) to (iv) in Figure 5
was considered a complete period of a sheep’s walking. It was considered that the walking
time was about 0.29 s from (i) to (ii), about 0.45 s from (i) to (iii), and about 0.91 s from (i)
to (iv) in the 10 s walking behavior segment shown in Figure 3. A total of 24 typical walk-
ing segments of sheep were observed by video, and the average period was 0.93 s and the
maximum period was 1.25 s.
Figure 5. Decomposition movement of sheep’s typical walking behavior in one period. (i) The start-
ing state of walking behavior; (ii) the sheep lifts the left leg; (iii) the body moves forward with the
left leg; and (iv) the sheep lifts the right leg and moves forward with the right leg to the starting
state.
The T-acc and T-gyr signals of a typical 5 s of running behavior of sheep are shown
in Figure 6. The x-, y- and z-axis signals of T-acc and T-gyr were, respectively, subjected
to fast Fourier transformation, as shown in Figure 7. At the same time, the dominant fre-
quency and period were calculated.
Animals 2022, 12, x FOR PEER REVIEW 9 of 24
Figure 6. Time series of three-axis accelerometer and three-axis gyroscope signals from a 20 Hz
sampling rate for observed behavior of 5 s of running. Acceleration was in g (9.8 m/s
2
) units.
(a) (b)
(c) (d)
(e) (f)
Figure 7. Fourier transform analysis of three-axis accelerometer and three-axis gyroscope signals for
5 s of running behavior, as shown in Figure 6, calculating the main frequency and period. (a) x-axis
accelerometer: n = 23; N = 100; f = 4.4 Hz; T ≈ 0.23 s; (b) x-axis gyroscope: n = 12; N = 100; f = 2.2 Hz;
T ≈ 0.45 s; (c) y-axis accelerometer: n = 12; N = 100; f = 2.2 Hz; T ≈ 0.45 s; (d) y-axis gyroscope: n = 23;
Figure 6.
Time series of three-axis accelerometer and three-axis gyroscope signals from a 20 Hz
sampling rate for observed behavior of 5 s of running. Acceleration was in g (9.8 m/s2) units.
2.5. Classification Feature Construction
Based on the labelled T-acc data and T-gyr data, the feature datasets were constructed
out of the time and frequency domains with time windows of 3, 5 and 11 s. Table 4shows
the selected time and frequency domain features.

Animals 2022,12, 1744 10 of 24
Animals 2022, 12, x FOR PEER REVIEW 9 of 24
Figure 6. Time series of three-axis accelerometer and three-axis gyroscope signals from a 20 Hz
sampling rate for observed behavior of 5 s of running. Acceleration was in g (9.8 m/s
2
) units.
(a) (b)
(c) (d)
(e) (f)
Figure 7. Fourier transform analysis of three-axis accelerometer and three-axis gyroscope signals for
5 s of running behavior, as shown in Figure 6, calculating the main frequency and period. (a) x-axis
accelerometer: n = 23; N = 100; f = 4.4 Hz; T ≈ 0.23 s; (b) x-axis gyroscope: n = 12; N = 100; f = 2.2 Hz;
T ≈ 0.45 s; (c) y-axis accelerometer: n = 12; N = 100; f = 2.2 Hz; T ≈ 0.45 s; (d) y-axis gyroscope: n = 23;
Figure 7.
Fourier transform analysis of three-axis accelerometer and three-axis gyroscope signals for
5 s of running behavior, as shown in Figure 6, calculating the main frequency and period. (
a
)x-axis
accelerometer: n= 23; N= 100; f= 4.4 Hz; T
≈
0.23 s; (
b
)x-axis gyroscope: n= 12; N= 100;
f= 2.2 Hz
;
T
≈
0.45 s; (
c
)y-axis accelerometer: n= 12; N= 100; f= 2.2 Hz; T
≈
0.45 s; (
d
)y-axis gyroscope:
n= 23
;
N= 100; f= 4.4 Hz; T
≈
0.23 s; (
e
)z-axis accelerometer: n= 25; N= 100; f= 4.8 Hz; T
≈
0.21 s; (
f
)z-axis
gyroscope: n= 12; N= 100; f= 2.2 Hz; T≈0.45 s.

Animals 2022,12, 1744 11 of 24
Animals 2022, 12, x FOR PEER REVIEW 10 of 24
N = 100; f = 4.4 Hz; T ≈ 0.23 s; (e) z-axis accelerometer: n = 25; N = 100; f = 4.8 Hz; T ≈ 0.21 s; (f) z-axis
gyroscope: n = 12; N = 100; f = 2.2 Hz; T ≈ 0.45 s.
The x-acc, y-gyr and z-acc signals that were used to calculate the typical running be-
havior segment of sheep in Figure 6 had similar periods, while the x-axis gyroscope (x-
gyr), y-acc and z-gyr signals had the same periods. The maximum and minimum periods
were 0.45 and 0.21 s, respectively. The maximum and minimum frequencies were 4.8 and
2.3 Hz, respectively. The maximum period of the running behavior segment of the sheep
(Figure 6) was about half of that of the walking behavior segment (Figure 3).
Figure 8 presents the T-acc and T-gyr signals of a typical 10 s of grazing behavior of
sheep. The x-, y- and z-axis signals were, respectively, subjected to fast Fourier transfor-
mation, and the dominant frequency and period were calculated at the same time.
Figure 8. Time series of three-axis accelerometer and three-axis gyroscope signals from a 20 Hz
sampling rate for observed behavior of 10 s of grazing. Acceleration was in g (9.8 m/s
2
) units.
Compared with typical walking and running behavior, grazing had no significant
periodicity. Moreover, through observation, it was found that the period of sheep grazing
behavior on pasture with different SSH was also different. Therefore, the period of sheep
grazing behavior was mainly determined by video observation. The process of observing
a sheep’s grazing behavior could be roughly divided into: (i) biting once or several times
and then swallowing; (ii) biting once or several times, then chewing and finally swallow-
ing; (iii) grazing, then biting once or several times and finally swallowing; (iv) grazing,
then biting once or several times and then chewing (or chewing while foraging), and fi-
nally swallowing. Due to the biting movement of sheep, the video was easily observed,
and the time interval between two biting movements was taken to be the grazing period.
A total of 41 grazing segments of sheep in the video were observed, and the duration of
each one divided by the number of biting movements was taken as the period of grazing
behavior: the maximum period was 2.15 s, which was longer than the maximum period
of walking by 1.25 s. Considering that the period of walking behavior was about twice
that for running behavior, the maximum period for observing dynamic behavior was 2.15
s. Therefore, a minimum time window of 3 s was enough to satisfy a behavioral movement
Figure 8.
Time series of three-axis accelerometer and three-axis gyroscope signals from a 20 Hz
sampling rate for observed behavior of 10 s of grazing. Acceleration was in g (9.8 m/s2) units.
Table 4.
Features calculated for each time window (3, 5 and 11 s) based on x-, y- and z-axis accelerom-
eter and gyroscope data. Equations adapted from [3,8,9,17,22,23].
Feature The Number of Features Equations/Description #
Minimum value 6 Minimum value of all window values
Maximum value 6 Maximum value of all window values
Median 6 Median value of all window values
Upper quartile 6 Upper quartile value of all window values
Lower quartile 6 Lower quartile value of all window values
Kurtosis 6 Kurtosis calculated from window values
Skewness 6 Skewness calculated from window values
Range 6 Max −Min
Mean 6 1
n∑n
i=1X(i)
Variance 6 1
n−1∑n
i=1Xi−X2
Standard deviation 6 q1
n−1∑n
i=1Xi−X2
Root mean square (RMS) 6 q1
n∑n
i=1Xi2
Signal magnitude area (SMA) 2 1
n∑n
i=1(|Xi|+|Yi|+|Zi|)
Energy 2 1
n∑n
i=1Xi2+Yi2+Zi22
Entropy 2 1
n∑n
i=11+(Xi+Yi+Zi)2ln1+(Xi+Yi+Zi)2
Dominant frequency 6 After applying Fourier transformation, this is the frequency at which
the signal has its highest power
Spectral energy 6 1
N∑N
k=1(F(k))2;N=n∗f
2
Spectral entropy 6 −1×∑N
k=1(F(k)
Z1Nlog2
F(k)
Z1N);Z1=1
N∑N
k=1F(k)
Vectorial dynamic body acceleration (VeDBA) 15 qax2+ay2+az2
Features of VeDBA
Minimum value, maximum value, median, upper quartile, lower
quartile, kurtosis, skewness, range, mean, variance, standard deviation,
RMS, dominant frequency, spectral energy, spectral entropy
#
Where nis the total number of all window values; where Nis
n∗f
2
, f is data collection frequency of the device (20 Hz).
A total of 63 features were constructed based on the T-acc data, 48 features were
constructed based on the T-gyr data, and 111 features were constructed based on the T-acc
and T-gyr data. In order to compare the accuracy of three kinds of sensor data for behavior
classification, AdaBoost was used to rank the feature importance of behavior classification
in T-acc data, and T-acc and T-gyr data. The top 48 important features were selected to
construct its feature dataset.

Animals 2022,12, 1744 12 of 24
Nine behavioral feature datasets were constructed for three different time windows
and three different kinds of sensor data combinations: in the constructing the datasets,
the larger the time window, the fewer the rows of feature data (Figure 9). There were
50,302 rows of feature data with a 3 s time window; 46,101 rows of feature data with a 5 s
time window; and 23,015 rows of feature data with an 11 s time window in the labelled
behavioral data on the 15th and 18th of May. The duration of the running segments did
not exceed 11 s, and only a few of them exceeded 5 s. As a result, this sheep behavior
could not be constructed within an 11 s time window in the labelled data, and only a few
running behavior features could be constructed within a 5 s time window. Therefore, it was
not included in the feature datasets with an 11 s or 5 s time window. The nine behavioral
feature datasets were standardized: 80% of each was used as the training and 20% as the
test dataset.
Animals 2022, 12, x FOR PEER REVIEW 12 of 24
behavioral data on the 15th and 18th of May. The duration of the running segments did
not exceed 11 s, and only a few of them exceeded 5 s. As a result, this sheep behavior could
not be constructed within an 11 s time window in the labelled data, and only a few run-
ning behavior features could be constructed within a 5 s time window. Therefore, it was
not included in the feature datasets with an 11 s or 5 s time window. The nine behavioral
feature datasets were standardized: 80% of each was used as the training and 20% as the
test dataset.
Figure 9. The features calculation process of three time windows (3, 5 and 11 s) on the 33 s observed
behavior segments.
2.6. ML Classification Algorithms
Ensemble learning is a technique for improving prediction performance by construct-
ing and integrating multiple machine learners [27]. According to different integration
strategies for different machine learners, ensemble learning can be divided into boosting
[28], bagging [29] and stacking [30]. Boosting and bagging usually integrate homogeneous
learners: boosting adopts sequence integration, and bagging adopts parallel integration.
Stacking integrated heterogeneous learners is a hierarchical structure, and the outputs of
multiple heterogeneous learners in the first layer are used as learner inputs in the second
layer of the training model. The integration of multiple learners can reduce the possible
deviation of a single classifier when dealing with unbalanced data and prevent over-fit-
ting, resulting in a better performance than a single algorithm. Therefore, more studies
apply ensemble learning to the classification learning of unbalanced data [31,32].
The ELM method has a ML training speed thousands of times faster than that of a
traditional back-propagation neural network. It is based on a generalized, single-hidden
layer, feedforward neural network and has good generalization performance [33–35].
ELM, AdaBoost (a concrete implementation of the Boosting algorithm) [36] and
stacking were used to classify sheep behavior in this study. The basic learner of the stack-
ing algorithm adopted AdaBoost, random forest (RF, an improvement on bagging), and
support vector machine (SVM), which has been applied well in previous research on
sheep behavior classification. The secondary learner of the stacking algorithm adopted
ELM. Trained ELM, AdaBoost, and stacking were compared for accuracy and practical
application in sheep behavior classification.
2.7. Performance of the Classification
The accuracy of the trained models was evaluated on the test dataset and the CBS
test dataset, and the evaluation indexes were accuracy and Kappa value, which is an index
used to test whether the model prediction is consistent with the actual values. Accuracy
was calculated using Formula (3): accurac
y
=
 (3)
True positive (TP) indicated that both the actual category and the model prediction
category were positive. True negative (TN) indicated that both were negative. False posi-
tive (FP) indicated a positive model prediction category, but a negative actual category.
False negative (FN) indicated a negative model prediction category, but a positive actual
Figure 9.
The features calculation process of three time windows (3, 5 and 11 s) on the 33 s observed
behavior segments.
2.6. ML Classification Algorithms
Ensemble learning is a technique for improving prediction performance by construct-
ing and integrating multiple machine learners [
27
]. According to different integration strate-
gies for different machine learners, ensemble learning can be divided into boosting [
28
],
bagging [
29
] and stacking [
30
]. Boosting and bagging usually integrate homogeneous
learners: boosting adopts sequence integration, and bagging adopts parallel integration.
Stacking integrated heterogeneous learners is a hierarchical structure, and the outputs of
multiple heterogeneous learners in the first layer are used as learner inputs in the second
layer of the training model. The integration of multiple learners can reduce the possible
deviation of a single classifier when dealing with unbalanced data and prevent over-fitting,
resulting in a better performance than a single algorithm. Therefore, more studies apply
ensemble learning to the classification learning of unbalanced data [31,32].
The ELM method has a ML training speed thousands of times faster than that of a
traditional back-propagation neural network. It is based on a generalized, single-hidden
layer, feedforward neural network and has good generalization performance [33–35].
ELM, AdaBoost (a concrete implementation of the Boosting algorithm) [
36
] and stack-
ing were used to classify sheep behavior in this study. The basic learner of the stacking
algorithm adopted AdaBoost, random forest (RF, an improvement on bagging), and support
vector machine (SVM), which has been applied well in previous research on sheep behavior
classification. The secondary learner of the stacking algorithm adopted ELM. Trained ELM,
AdaBoost, and stacking were compared for accuracy and practical application in sheep
behavior classification.
2.7. Performance of the Classification
The accuracy of the trained models was evaluated on the test dataset and the CBS test
dataset, and the evaluation indexes were accuracy and Kappa value, which is an index
used to test whether the model prediction is consistent with the actual values. Accuracy
was calculated using Formula (3):
accuracy =TP +TN
TP +FP +FN +TN (3)

Animals 2022,12, 1744 13 of 24
True positive (TP) indicated that both the actual category and the model prediction
category were positive. True negative (TN) indicated that both were negative. False positive
(FP) indicated a positive model prediction category, but a negative actual category. False
negative (FN) indicated a negative model prediction category, but a positive actual category.
The Kappa value was calculated based on a confusion matrix, and the calculated result was
between
−
1 and 1, but usually between 0 and 1. The larger the value, the higher the model
classification accuracy. The Kappa value is very suitable for evaluating the performance of
the model for classifying the unbalanced quantity of samples in various categories [37].
The classification performance of each behavior was evaluated for precision, recall
and F-score, and calculated by using Formulas (4)–(6).
Precision =TP
TP +FP (4)
Recall =TP
TP +FN (5)
F−score =2×Precision ×Recall
Precision +Recall (6)
3. Results
3.1. Model Training and Test Results
The training dataset was used to train the model, during which a 5-fold cross validation
was conducted to select the optimal hyperparameters; the test dataset was used to evaluate
the performance of the trained model. The performance of the trained models on the test
dataset is shown in Table 5.
Table 5.
Summary of accuracy and kappa values for different ML predictions of walking, standing,
grazing, lying and running at 3, 5 and 11 s time windows on the test dataset. Bold indicates the
highest accuracy and Kappa values combination.
Time
Window Sensor
ELM AdaBoost Stacking Number of
Classified Behaviors
Accuracy (%) Kappa Value Accuracy (%) Kappa Value Accuracy (%) Kappa Value
3 s
Accelerometer 91.5 0.871 94.7 0.921 94.6 0.918 5
Gyroscope 90.1 0.850 92.4 0.885 93.1 0.895 5
Accelerometer
and Gyroscope 92.7 0.890 97.1 0.956 97.2 0.959 5
5 s
Accelerometer 93.3 0.899 97.5 0.963 97.6 0.964 4
Gyroscope 93.0 0.894 95.0 0.923 94.9 0.922 4
Accelerometer
and Gyroscope 94.7 0.920 98.9 0.983 98.9 0.983 4
11 s
Accelerometer 98.26 0.983 99.3 0.989 99.3 0.988 4
Gyroscope 97.0 0.951 96.6 0.945 96.2 0.939 4
Accelerometer
and Gyroscope 98.5 0.976 99.7 0.995 99.7 0.995 4
It was found that the accuracy of the three models in the three time windows from
three types of sensor data were all above 90%, and the Kappa values were above 0.85. The
larger the time window, the higher the classification accuracy. Data classification accuracy
using T-acc and T-gyr sensors was higher than using them separately, and accuracy using
T-acc was higher than using T-gyr. Model accuracy was stacking > AdaBoost > ELM.
3.2. Practical Application of the Trained Models
The 27 trained models were applied to the behavior classification of three grazing
sheep on pasture with three different SSH from 9:00 a.m. to 4:30 p.m. on the 16th, 17th
and 19th of May. The moving mode of the time window during classification comprised
jumping and sliding. For example, the sheep touched the electrified fence during grazing,
then ran for 4 s, walked for 6 s, and finally stood up. Assuming that the model classified
every behavior feature with 100% accuracy, behavior classification with 3, 5 and 11 s time

Animals 2022,12, 1744 14 of 24
window jump-moving is shown in Figure 10, and classification with slide-moving is shown
in Figure 11.
Animals 2022, 12, x FOR PEER REVIEW 14 of 24
window jump-moving is shown in Figure 10, and classification with slide-moving is
shown in Figure 11.
Figure 10. The process of 33 s of continuous behavior segment classification using 3, 5 and 11 s time
windows with jump-moving. Assuming that the model accurately classified each behavioral fea-
ture, the classification accuracy at 3, 5 and 11 s was 97.0, 93.9 and 84.8%, respectively.
Theoretically, the slide-moving window classification accuracy should be higher, so
the slide-moving window was used when using the 27 trained models for continuous be-
havior segments classification. The final classification result at each moment was deter-
mined by the behavior with the highest prediction score among all behaviors classified by
the trained model for all time windows containing that moment, as shown in Figure 11.
The results of 27 trained models on the CBS test dataset are shown in Table 6.
Table 6. Summary of the accuracy and Kappa values for different ML predictions of walking, stand-
ing, grazing, lying and running at 3, 5 and 11 s time windows on the CBS test dataset. Bold indicates
the highest accuracy and Kappa values combination.
Time Window Sensor
ELM AdaBoost Stacking
Number of Clas-
sified Behaviors
Accuracy
(%) Kappa Value Accuracy
(%) Kappa Value Accuracy
(%) Kappa Value
3 s
Accelerometer 84.8 0.796 80.2 0.735 81.2 0.749 5
Gyroscope 82.8 0.770 78.9 0.716 77.6 0.700 5
Accelerometer and
Gyroscope 85.2 0.801 85.3 0.803 87.8 0.836 5
5 s
Accelerometer 83.2 0.774 78.9 0.717 83.8 0.782 4
Gyroscope 82.7 0.767 78.6 0.711 81.4 0.750 4
Accelerometer and
Gyroscope 87.4 0.830 86.2 0.814 86.2 0.815 4
11 s Accelerometer 72.7 0.631 76.9 0.689 74.4 0.656 4
Gyroscope 66.3 0.542 63.9 0.510 64.8 0.524 4
Figure 10.
The process of 33 s of continuous behavior segment classification using 3, 5 and 11 s time
windows with jump-moving. Assuming that the model accurately classified each behavioral feature,
the classification accuracy at 3, 5 and 11 s was 97.0, 93.9 and 84.8%, respectively.
Theoretically, the slide-moving window classification accuracy should be higher, so the
slide-moving window was used when using the 27 trained models for continuous behavior
segments classification. The final classification result at each moment was determined by
the behavior with the highest prediction score among all behaviors classified by the trained
model for all time windows containing that moment, as shown in Figure 11. The results of
27 trained models on the CBS test dataset are shown in Table 6.
Table 6.
Summary of the accuracy and Kappa values for different ML predictions of walking, standing,
grazing, lying and running at 3, 5 and 11 s time windows on the CBS test dataset. Bold indicates the
highest accuracy and Kappa values combination.
Time
Window Sensor
ELM AdaBoost Stacking Number of
Classified Behaviors
Accuracy (%) Kappa Value Accuracy (%) Kappa Value Accuracy (%) Kappa Value
3 s
Accelerometer 84.8 0.796 80.2 0.735 81.2 0.749 5
Gyroscope 82.8 0.770 78.9 0.716 77.6 0.700 5
Accelerometer
and Gyroscope 85.2 0.801 85.3 0.803 87.8 0.836 5
5 s
Accelerometer 83.2 0.774 78.9 0.717 83.8 0.782 4
Gyroscope 82.7 0.767 78.6 0.711 81.4 0.750 4
Accelerometer
and Gyroscope 87.4 0.830 86.2 0.814 86.2 0.815 4
11 s
Accelerometer 72.7 0.631 76.9 0.689 74.4 0.656 4
Gyroscope 66.3 0.542 63.9 0.510 64.8 0.524 4
Accelerometer
and Gyroscope 78.0 0.702 67.8 0.565 71.5 0.616 4

Animals 2022,12, 1744 15 of 24
Animals 2022, 12, x FOR PEER REVIEW 15 of 24
Accelerometer and
Gyroscope 78.0 0.702 67.8 0.565 71.5 0.616 4
Figure 11. The process of 33 s of continuous behavior segment classification using 3, 5 and 11 s time
windows with slide-moving. Assuming that the model accurately classified each behavioral feature,
the classification accuracy at 3, 5 and 11 s was 100, 100 and 87.9%, respectively.
When applying the trained models to continuous behavior segments, the accuracy
on the CBS test dataset obviously decreased. Based on T-acc and T-gyr data, the 3 s time
window and stacking model had the highest sheep behavior classification accuracy of
87.8% and a Kappa value of 0.836. The larger the time window, the lower the classification
accuracy, which was contrary to the results in the test dataset. The accuracy of classifica-
tion by using T-acc data was higher than using T-gyr data, which was the same as the
results in the test dataset. In most instances, the classification accuracy of using the two
kinds of sensor data was higher than that by using them separately. Stacking and ELM
models performed better on the CBS test dataset.
The classification accuracy (Table 7) of the optimal model for each behavior across
the three time windows was calculated, and the main reason for the decline in perfor-
mance was the confusion of standing and lying behavior. Although the training samples
of running behavior were very few, the F-score of classification still reached 82.4% in prac-
tical application due to its very special features.
Figure 11.
The process of 33 s of continuous behavior segment classification using 3, 5 and 11 s time
windows with slide-moving. Assuming that the model accurately classified each behavioral feature,
the classification accuracy at 3, 5 and 11 s was 100, 100 and 87.9%, respectively.
When applying the trained models to continuous behavior segments, the accuracy
on the CBS test dataset obviously decreased. Based on T-acc and T-gyr data, the 3 s time
window and stacking model had the highest sheep behavior classification accuracy of
87.8% and a Kappa value of 0.836. The larger the time window, the lower the classification
accuracy, which was contrary to the results in the test dataset. The accuracy of classification
by using T-acc data was higher than using T-gyr data, which was the same as the results
in the test dataset. In most instances, the classification accuracy of using the two kinds
of sensor data was higher than that by using them separately. Stacking and ELM models
performed better on the CBS test dataset.
The classification accuracy (Table 7) of the optimal model for each behavior across the
three time windows was calculated, and the main reason for the decline in performance
was the confusion of standing and lying behavior. Although the training samples of
running behavior were very few, the F-score of classification still reached 82.4% in practical
application due to its very special features.

Animals 2022,12, 1744 16 of 24
Table 7. Performance statistics of three optimal ML models (with the highest classification accuracy
at the 3, 5 and 11 s time window on the CBS test dataset) used for classifying walking, standing,
grazing, lying and running on the CBS test dataset.
Time Window Model Performance Walking Standing Grazing Lying Running
3 s Stacking Precision (%) 99.3 81.9 98.5 75.8 90.3
Recall (%) 95.4 72.8 100.0 85.4 75.7
F-score (%) 97.3 77.1 99.3 80.3 82.4
5 s ELM Precision (%) 99.6 73.2 95.4 89.9 –
Recall (%) 88.6 91.9 100.0 69.3 –
F-score (%) 93.8 81.4 97.6 78.3 –
11 s ELM Precision (%) 97.9 63.3 82.4 92.5 –
Recall (%) 59.1 94.6 100.0 52.6 –
F-score (%) 73.7 75.9 90.4 67.1 –
The accuracy of the trained models on the CBS test dataset drops obviously compared
to the classification accuracy on the test set. The models that performed the best classifica-
tion at the 3, 5 and 11 s time windows were tested to see if they were over-fitted because
of having too many features. Using the top 12 important features from T-acc and T-gyr
data to retrain these models and test the accuracy of the trained models on the test dataset
and the CBS test dataset, the results are presented in Table 8. It was found that when the
number of features was reduced from 48 to 12, the performance of the trained model on the
test dataset was not significantly affected, but the accuracy dropped obviously on the CBS
test dataset. Compared with 12 features, training the model with 48 improved the practical
performance of sheep behavior classification in continuous behavior segments.
Table 8.
Summary of accuracy and the Kappa values for three ML models (with the highest classifica-
tion accuracy at the 3, 5 and 11 s time windows on the CBS test dataset, trained by 12 features from
three-axis accelerometer and three-axis gyroscope data) at the 3, 5 and 11 s time windows on the test
dataset and the CBS test dataset separately.
Time Window Model Accuracy (%) Kappa Value
Test dataset 3 s Stacking 95.1 0.927
5 s ELM 94.4 0.915
11 s ELM 98.4 0.975
CBS test dataset 3 s Stacking 79.0 0.719
5 s ELM 78.2 0.707
11 s ELM 63.3 0.501
3.3. Behavior Classification of Three Grazing Sheep on Pasture with Three Different SSH
We selected the combination (trained stacking model, 3 s time window, T-acc and
T-gyr data) with the best classification performance on the CTB test dataset to classify the
behavior of three grazing sheep for 7.5 h (from 9:00 a.m. to 4:30 p.m.) each on pasture with
three different SSH. The classification results are shown in Figure 12.

Animals 2022,12, 1744 17 of 24
Animals 2022, 12, x FOR PEER REVIEW 17 of 24
lying behaviors were combined into resting behavior. As shown in Table 9, the behavioral
distribution of the three sheep had some common characteristics: the longest walking time
was on short SSH, the longest grazing time was on medium SSH, and the longest resting
time was on tall SSH. Individual differences were found in which the R-Sheep had the
shortest grazing and the longest walking times, while the P-sheep had the longest grazing
and the shortest resting times.
Figure 12. Behavior distribution of three grazing sheep on pastures with three different sward sur-
face heights (SSH) classified by the trained stacking model.
Table 9. Percentage of behavior time of the three grazing sheep for 7.5 h on pasture with three dif-
ferent sward surface heights (SSH).
SSH B-Sheep P-Sheep R-Sheep Average Proportion
Short
Walking 9.91%
Grazing 50.79%
Resting 39.18%
Running 0.12%
Medium
Walking 6.50%
Grazing 54.83%
Resting 38.63%
Running 0.04%
Tall
Walking 6.72%
Grazing 50.70%
Resting 42.52%
Running 0.06%
Figure 12.
Behavior distribution of three grazing sheep on pastures with three different sward surface
heights (SSH) classified by the trained stacking model.
The behavioral distribution of the three sheep on pasture with same SSH was similar,
but grazing behavior on short SSH was relatively scattered with a short duration for each
bout, and more continuous long-term grazing was found on medium and tall SSH. Walking
behavior was usually mixed with grazing behavior. Sheep would stand or lie down after
grazing for a period of time, accompanied by rumination, and then start grazing repeatedly.
To quantitatively analyze the behavior classification of the three sheep on pasture
with three different SSH, the average time each sheep spent grazing in 7.5 h was counted.
Because of the misclassification between standing and lying behaviors, the standing and
lying behaviors were combined into resting behavior. As shown in Table 9, the behavioral
distribution of the three sheep had some common characteristics: the longest walking time
was on short SSH, the longest grazing time was on medium SSH, and the longest resting
time was on tall SSH. Individual differences were found in which the R-Sheep had the
shortest grazing and the longest walking times, while the P-sheep had the longest grazing
and the shortest resting times.
Table 9.
Percentage of behavior time of the three grazing sheep for 7.5 h on pasture with three
different sward surface heights (SSH).
SSH B-Sheep P-Sheep R-Sheep Average Proportion
Short
Animals 2022, 12, x FOR PEER REVIEW 17 of 24
lying behaviors were combined into resting behavior. As shown in Table 9, the behavioral
distribution of the three sheep had some common characteristics: the longest walking time
was on short SSH, the longest grazing time was on medium SSH, and the longest resting
time was on tall SSH. Individual differences were found in which the R-Sheep had the
shortest grazing and the longest walking times, while the P-sheep had the longest grazing
and the shortest resting times.
Figure 12. Behavior distribution of three grazing sheep on pastures with three different sward sur-
face heights (SSH) classified by the trained stacking model.
Table 9. Percentage of behavior time of the three grazing sheep for 7.5 h on pasture with three dif-
ferent sward surface heights (SSH).
SSH B-Sheep P-Sheep R-Sheep Average Proportion
Short
Walking 9.91%
Grazing 50.79%
Resting 39.18%
Running 0.12%
Medium
Walking 6.50%
Grazing 54.83%
Resting 38.63%
Running 0.04%
Tall
Walking 6.72%
Grazing 50.70%
Resting 42.52%
Running 0.06%
Animals 2022, 12, x FOR PEER REVIEW 17 of 24
lying behaviors were combined into resting behavior. As shown in Table 9, the behavioral
distribution of the three sheep had some common characteristics: the longest walking time
was on short SSH, the longest grazing time was on medium SSH, and the longest resting
time was on tall SSH. Individual differences were found in which the R-Sheep had the
shortest grazing and the longest walking times, while the P-sheep had the longest grazing
and the shortest resting times.
Figure 12. Behavior distribution of three grazing sheep on pastures with three different sward sur-
face heights (SSH) classified by the trained stacking model.
Table 9. Percentage of behavior time of the three grazing sheep for 7.5 h on pasture with three dif-
ferent sward surface heights (SSH).
SSH B-Sheep P-Sheep R-Sheep Average Proportion
Short
Walking 9.91%
Grazing 50.79%
Resting 39.18%
Running 0.12%
Medium
Walking 6.50%
Grazing 54.83%
Resting 38.63%
Running 0.04%
Tall
Walking 6.72%
Grazing 50.70%
Resting 42.52%
Running 0.06%
Animals 2022, 12, x FOR PEER REVIEW 17 of 24
lying behaviors were combined into resting behavior. As shown in Table 9, the behavioral
distribution of the three sheep had some common characteristics: the longest walking time
was on short SSH, the longest grazing time was on medium SSH, and the longest resting
time was on tall SSH. Individual differences were found in which the R-Sheep had the
shortest grazing and the longest walking times, while the P-sheep had the longest grazing
and the shortest resting times.
Figure 12. Behavior distribution of three grazing sheep on pastures with three different sward sur-
face heights (SSH) classified by the trained stacking model.
Table 9. Percentage of behavior time of the three grazing sheep for 7.5 h on pasture with three dif-
ferent sward surface heights (SSH).
SSH B-Sheep P-Sheep R-Sheep Average Proportion
Short
Walking 9.91%
Grazing 50.79%
Resting 39.18%
Running 0.12%
Medium
Walking 6.50%
Grazing 54.83%
Resting 38.63%
Running 0.04%
Tall
Walking 6.72%
Grazing 50.70%
Resting 42.52%
Running 0.06%
Walking 9.91%
Grazing 50.79%
Resting 39.18%
Running 0.12%
Medium
Animals 2022, 12, x FOR PEER REVIEW 17 of 24
lying behaviors were combined into resting behavior. As shown in Table 9, the behavioral
distribution of the three sheep had some common characteristics: the longest walking time
was on short SSH, the longest grazing time was on medium SSH, and the longest resting
time was on tall SSH. Individual differences were found in which the R-Sheep had the
shortest grazing and the longest walking times, while the P-sheep had the longest grazing
and the shortest resting times.
Figure 12. Behavior distribution of three grazing sheep on pastures with three different sward sur-
face heights (SSH) classified by the trained stacking model.
Table 9. Percentage of behavior time of the three grazing sheep for 7.5 h on pasture with three dif-
ferent sward surface heights (SSH).
SSH B-Sheep P-Sheep R-Sheep Average Proportion
Short
Walking 9.91%
Grazing 50.79%
Resting 39.18%
Running 0.12%
Medium
Walking 6.50%
Grazing 54.83%
Resting 38.63%
Running 0.04%
Tall
Walking 6.72%
Grazing 50.70%
Resting 42.52%
Running 0.06%
Animals 2022, 12, x FOR PEER REVIEW 17 of 24
lying behaviors were combined into resting behavior. As shown in Table 9, the behavioral
distribution of the three sheep had some common characteristics: the longest walking time
was on short SSH, the longest grazing time was on medium SSH, and the longest resting
time was on tall SSH. Individual differences were found in which the R-Sheep had the
shortest grazing and the longest walking times, while the P-sheep had the longest grazing
and the shortest resting times.
Figure 12. Behavior distribution of three grazing sheep on pastures with three different sward sur-
face heights (SSH) classified by the trained stacking model.
Table 9. Percentage of behavior time of the three grazing sheep for 7.5 h on pasture with three dif-
ferent sward surface heights (SSH).
SSH B-Sheep P-Sheep R-Sheep Average Proportion
Short
Walking 9.91%
Grazing 50.79%
Resting 39.18%
Running 0.12%
Medium
Walking 6.50%
Grazing 54.83%
Resting 38.63%
Running 0.04%
Tall
Walking 6.72%
Grazing 50.70%
Resting 42.52%
Running 0.06%
Animals 2022, 12, x FOR PEER REVIEW 17 of 24
lying behaviors were combined into resting behavior. As shown in Table 9, the behavioral
distribution of the three sheep had some common characteristics: the longest walking time
was on short SSH, the longest grazing time was on medium SSH, and the longest resting
time was on tall SSH. Individual differences were found in which the R-Sheep had the
shortest grazing and the longest walking times, while the P-sheep had the longest grazing
and the shortest resting times.
Figure 12. Behavior distribution of three grazing sheep on pastures with three different sward sur-
face heights (SSH) classified by the trained stacking model.
Table 9. Percentage of behavior time of the three grazing sheep for 7.5 h on pasture with three dif-
ferent sward surface heights (SSH).
SSH B-Sheep P-Sheep R-Sheep Average Proportion
Short
Walking 9.91%
Grazing 50.79%
Resting 39.18%
Running 0.12%
Medium
Walking 6.50%
Grazing 54.83%
Resting 38.63%
Running 0.04%
Tall
Walking 6.72%
Grazing 50.70%
Resting 42.52%
Running 0.06%
Walking 6.50%
Grazing 54.83%
Resting 38.63%
Running 0.04%
Tall
Animals 2022, 12, x FOR PEER REVIEW 17 of 24
lying behaviors were combined into resting behavior. As shown in Table 9, the behavioral
distribution of the three sheep had some common characteristics: the longest walking time
was on short SSH, the longest grazing time was on medium SSH, and the longest resting
time was on tall SSH. Individual differences were found in which the R-Sheep had the
shortest grazing and the longest walking times, while the P-sheep had the longest grazing
and the shortest resting times.
Figure 12. Behavior distribution of three grazing sheep on pastures with three different sward sur-
face heights (SSH) classified by the trained stacking model.
Table 9. Percentage of behavior time of the three grazing sheep for 7.5 h on pasture with three dif-
ferent sward surface heights (SSH).
SSH B-Sheep P-Sheep R-Sheep Average Proportion
Short
Walking 9.91%
Grazing 50.79%
Resting 39.18%
Running 0.12%
Medium
Walking 6.50%
Grazing 54.83%
Resting 38.63%
Running 0.04%
Tall
Walking 6.72%
Grazing 50.70%
Resting 42.52%
Running 0.06%
Animals 2022, 12, x FOR PEER REVIEW 17 of 24
lying behaviors were combined into resting behavior. As shown in Table 9, the behavioral
distribution of the three sheep had some common characteristics: the longest walking time
was on short SSH, the longest grazing time was on medium SSH, and the longest resting
time was on tall SSH. Individual differences were found in which the R-Sheep had the
shortest grazing and the longest walking times, while the P-sheep had the longest grazing
and the shortest resting times.
Figure 12. Behavior distribution of three grazing sheep on pastures with three different sward sur-
face heights (SSH) classified by the trained stacking model.
Table 9. Percentage of behavior time of the three grazing sheep for 7.5 h on pasture with three dif-
ferent sward surface heights (SSH).
SSH B-Sheep P-Sheep R-Sheep Average Proportion
Short
Walking 9.91%
Grazing 50.79%
Resting 39.18%
Running 0.12%
Medium
Walking 6.50%
Grazing 54.83%
Resting 38.63%
Running 0.04%
Tall
Walking 6.72%
Grazing 50.70%
Resting 42.52%
Running 0.06%
Animals 2022, 12, x FOR PEER REVIEW 17 of 24
lying behaviors were combined into resting behavior. As shown in Table 9, the behavioral
distribution of the three sheep had some common characteristics: the longest walking time
was on short SSH, the longest grazing time was on medium SSH, and the longest resting
time was on tall SSH. Individual differences were found in which the R-Sheep had the
shortest grazing and the longest walking times, while the P-sheep had the longest grazing
and the shortest resting times.
Figure 12. Behavior distribution of three grazing sheep on pastures with three different sward sur-
face heights (SSH) classified by the trained stacking model.
Table 9. Percentage of behavior time of the three grazing sheep for 7.5 h on pasture with three dif-
ferent sward surface heights (SSH).
SSH B-Sheep P-Sheep R-Sheep Average Proportion
Short
Walking 9.91%
Grazing 50.79%
Resting 39.18%
Running 0.12%
Medium
Walking 6.50%
Grazing 54.83%
Resting 38.63%
Running 0.04%
Tall
Walking 6.72%
Grazing 50.70%
Resting 42.52%
Running 0.06%
Walking 6.72%
Grazing 50.70%
Resting 42.52%
Running 0.06%
4. Discussion
4.1. Time Window Size and Sensor Type
In this study, several kinds of time windows, sensor data and ML algorithms were
used to classify five behaviors of grazing sheep. Depending on the maximum period of the

Animals 2022,12, 1744 18 of 24
dynamic behavior, we used 3, 5 and 11 s time windows. The results demonstrated that the
larger the time window, the higher the behavior classification accuracy on the test dataset.
This was consistent with the related studies. For example, Fogarty et al. [
3
] compared the 5
and 10 s time window and found that the 10 s one classified grazing, lying, standing and
walking behaviors with higher accuracy. Walton et al. [
22
] compared the time windows of
3, 5 and 7 s, and found that the 7 s time window was more accurate for classifying walking,
standing and lying behavior. This was because single-behavior windows [
38
] were usually
used during model training and testing, and the behavioral features of the large window
were more typical, which could increase the discrimination between each time window
segment. However, the large time window resulted in less data available for training and
validation because the larger the time window, the more it spanned several behaviors, and
it was usually removed from the dataset (Figure 9). The time window usually affected the
accuracy of behavioral classification depending on the frequency of data collection [
22
].
According to the different data collection frequencies and time windows used to achieve
the best classification performance in previous studies, it was found that the rows of raw
data contained in each time window were usually more than 100 [10,18].
When the trained models were applied to the CBS test dataset, the smaller the time
window, the higher the accuracy (3 > 5 > 11 s). This was completely opposite to the
performance of the trained models on the test dataset, which was a new finding of this study.
The duration of each continuous behavior in the actual grazing process could not all be
greater than or equal to 11 s, which led to the inclusion of two or more behaviors in a single
11 s time window [
39
]. This inevitably resulted in the misclassification of behaviors with
durations less than 11 s (as shown in Figure 11), degrading the classification performance
of trained models on the CBS test dataset. Therefore, when classifying sheep behaviors by
IMU sensors, attention should be paid to the balance between the time window with the
highest model training accuracy and the shortest duration of each behavior. To precisely
classify the sheep behavior at every moment, the time window should be long enough to
accommodate the maximum movement period of each behavior, but not too large, because
the smaller the time window, the more sensitive it is to behavior classifications, especially
for grazing sheep that have frequent movement changes. If the time windows were too
large, some short-duration behaviors could not be precisely classified. We should pay more
attention to the behavior classification performance in the CBS test set, which is more in
line with the actual application and more conducive to commercialization [10].
We compared three kinds of sensor data (T-acc, T-gyr, T-acc and T-gyr) for sheep
behavioral performance in this study. The result showed that the highest accuracy using
T-acc, T-gyr, T-acc and T-gyr data was 84.8, 82.8 and 87.8%, respectively. The accuracy of
sheep behavior classification using T-acc data was higher than that for T-gyr data. Many
studies have used T-acc data for sheep behavior classification with good accuracy [
8
–
10
,
20
].
Using two types of sensor data at the same time was beneficial for improving classification
accuracy in most cases. Similarly, Mansbridge et al. [
18
] and Walton et al. [
22
] have argued
that using both a gyroscope and accelerometer should improve the classification accuracy
of some behaviors, such as lying and eating. Nevertheless, some studies have reported that
gyroscopes increase power consumption and sometimes do not deliver major classification
performance improvements [
40
,
41
]. This indicates that we need to determine the type of
sensor to use based on the specific behavior we want to classify.
4.2. Sheep Behavior Classification Algorithm
The results from this study showed that, compared with the classification accuracy
in the test dataset, the accuracy of the trained models on the CBS test dataset decreased
obviously, but not because the trained models were over-fitted. In regards to selecting
model hyperparameters by five-fold cross-validation on the training data dataset, the
learning curves of optimum hyperparameters of a stacking model 3 s time window and
an ELM model 5 s and 11 s time window are shown in Figure 13. The accuracy of the
trained models on the test dataset was similar to that on the validation dataset and had not

Animals 2022,12, 1744 19 of 24
decreased, which indicated that the trained models were not over-fitted. In the practical
application, the classification accuracy of the trained models on the CBS test dataset
decreased, which was mainly caused by the confusion of standing and lying behaviors as
the movement of standing and lying had very similar neck movements (Figure 14). Even
though the same three sheep were used during the entire experiment, the device they wore
had to be taken off every day to output data and put on again the next day, which led to
different wearing positions for each sheep every day. The device position has a significant
influence on standing and lying behaviors, which are usually difficult to classify with ear
tags and collar-mounted sensors [
3
,
42
]. As can be seen in Table 7, as the time window
became larger, the precision of lying behavior increased from 75.8 to 92.5%, while the recall
decreased from 85.4 to 52.6%, and the precision of standing behavior decreased from 81.9
to 63.3%, indicating that more lying behaviors were misclassified as standing behaviors.
The F-score of standing behavior decreased from 77.1 to 75.9%, and the F-score of lying
behavior decreased from 80.3 to 67.1%. This contrasts with Walton et al. [
22
], where the
recall of standing and lying increased as the time window became larger, and the F-score
range of standing and lying was between 94% and 97%. At the same time, it was also found
that the recall of walking behavior decreased obviously in the 11 s time window, and the
precision of grazing behavior also decreased, indicating that some walking behavior was
misclassified as grazing behavior. This was because the misclassification mainly existed
between dynamic and static behaviors. The classification performance of grazing was the
best of the five behaviors, which was consistent with the findings of Fogarty et al. [
3
]. In
order to achieve robust and high-performing classification models, more balanced data
for each behavior from more sheep need to be collected to train and validate the behavior
classification models [8,18].
As for the three classification algorithms, the accuracy of stacking and AdaBoost on the
test dataset were high, as was the classification accuracy of stacking and ELM on the CBS
test dataset, indicating that stacking and ELM showed better robustness for sheep behavior
classification. These results provided a new reference for the algorithm selection of sheep
behavior classification since these two algorithms were hardly reported in previous studies
on sheep behavior classification [10].
Animals 2022, 12, x FOR PEER REVIEW 19 of 24
4.2. Sheep Behavior Classification Algorithm
The results from this study showed that, compared with the classification accuracy
in the test dataset, the accuracy of the trained models on the CBS test dataset decreased
obviously, but not because the trained models were over-fitted. In regards to selecting
model hyperparameters by five-fold cross-validation on the training data dataset, the
learning curves of optimum hyperparameters of a stacking model 3 s time window and
an ELM model 5 s and 11 s time window are shown in Figure 13. The accuracy of the
trained models on the test dataset was similar to that on the validation dataset and had
not decreased, which indicated that the trained models were not over-fitted. In the prac-
tical application, the classification accuracy of the trained models on the CBS test dataset
decreased, which was mainly caused by the confusion of standing and lying behaviors as
the movement of standing and lying had very similar neck movements (Figure 14). Even
though the same three sheep were used during the entire experiment, the device they
wore had to be taken off every day to output data and put on again the next day, which
led to different wearing positions for each sheep every day. The device position has a
significant influence on standing and lying behaviors, which are usually difficult to clas-
sify with ear tags and collar-mounted sensors [3,42]. As can be seen in Table 7, as the time
window became larger, the precision of lying behavior increased from 75.8 to 92.5%, while
the recall decreased from 85.4 to 52.6%, and the precision of standing behavior decreased
from 81.9 to 63.3%, indicating that more lying behaviors were misclassified as standing
behaviors. The F-score of standing behavior decreased from 77.1 to 75.9%, and the F-score
of lying behavior decreased from 80.3 to 67.1%. This contrasts with Walton et al. [22],
where the recall of standing and lying increased as the time window became larger, and
the F-score range of standing and lying was between 94% and 97%. At the same time, it
was also found that the recall of walking behavior decreased obviously in the 11 s time
window, and the precision of grazing behavior also decreased, indicating that some walk-
ing behavior was misclassified as grazing behavior. This was because the misclassification
mainly existed between dynamic and static behaviors. The classification performance of
grazing was the best of the five behaviors, which was consistent with the findings of
Fogarty et al. [3]. In order to achieve robust and high-performing classification models,
more balanced data for each behavior from more sheep need to be collected to train and
validate the behavior classification models [8,18].
(a)
Figure 13. Cont.

Animals 2022,12, 1744 20 of 24
Animals 2022, 12, x FOR PEER REVIEW 20 of 24
(b)
(c)
Figure 13. Learning curves for 5-fold cross-validation of three models (with the best performance
on three different time windows) on the three-axis acceleration and gyroscope data training set. (a)
Learning curve of stacking by 5-fold cross-validation on 3 s time window; (b) learning curve of ELM
by 5-fold cross-validation on 5 s time window; (c) learning curve of ELM by 5-fold cross-validation
on 11 s time window.
Figure 14. Time series of three-axis accelerometer signals from a 20 Hz sampling rate for 10 s of
observed behaviors of B-Sheep standing and P-Sheep lying. Acceleration was in g (9.8 m/s
2
) units.
As for the three classification algorithms, the accuracy of stacking and AdaBoost on
the test dataset were high, as was the classification accuracy of stacking and ELM on the
CBS test dataset, indicating that stacking and ELM showed better robustness for sheep
Animals 2022, 12, x FOR PEER REVIEW 20 of 24
(b)
(c)
Figure 13. Learning curves for 5-fold cross-validation of three models (with the best performance
on three different time windows) on the three-axis acceleration and gyroscope data training set. (a)
Learning curve of stacking by 5-fold cross-validation on 3 s time window; (b) learning curve of ELM
by 5-fold cross-validation on 5 s time window; (c) learning curve of ELM by 5-fold cross-validation
on 11 s time window.
Figure 14. Time series of three-axis accelerometer signals from a 20 Hz sampling rate for 10 s of
observed behaviors of B-Sheep standing and P-Sheep lying. Acceleration was in g (9.8 m/s
2
) units.
As for the three classification algorithms, the accuracy of stacking and AdaBoost on
the test dataset were high, as was the classification accuracy of stacking and ELM on the
CBS test dataset, indicating that stacking and ELM showed better robustness for sheep
Figure 13.
Learning curves for 5-fold cross-validation of three models (with the best performance
on three different time windows) on the three-axis acceleration and gyroscope data training set.
(
a
) Learning curve of stacking by 5-fold cross-validation on 3 s time window; (
b
) learning curve
of ELM by 5-fold cross-validation on 5 s time window; (
c
) learning curve of ELM by 5-fold cross-
validation on 11 s time window.
Animals 2022, 12, x FOR PEER REVIEW 20 of 24
(b)
(c)
Figure 13. Learning curves for 5-fold cross-validation of three models (with the best performance
on three different time windows) on the three-axis acceleration and gyroscope data training set. (a)
Learning curve of stacking by 5-fold cross-validation on 3 s time window; (b) learning curve of ELM
by 5-fold cross-validation on 5 s time window; (c) learning curve of ELM by 5-fold cross-validation
on 11 s time window.
Figure 14. Time series of three-axis accelerometer signals from a 20 Hz sampling rate for 10 s of
observed behaviors of B-Sheep standing and P-Sheep lying. Acceleration was in g (9.8 m/s
2
) units.
As for the three classification algorithms, the accuracy of stacking and AdaBoost on
the test dataset were high, as was the classification accuracy of stacking and ELM on the
CBS test dataset, indicating that stacking and ELM showed better robustness for sheep
Figure 14.
Time series of three-axis accelerometer signals from a 20 Hz sampling rate for 10 s of
observed behaviors of B-Sheep standing and P-Sheep lying. Acceleration was in g (9.8 m/s2) units.
4.3. Behavior Classification of Grazing Sheep on Pasture with Different SSH
The three merino sheep had a walking behavior while grazing (as shown in Figure 12).
By observing the video, it was found that they avoided grazing facing the sun. They always

Animals 2022,12, 1744 21 of 24
grazed along the long side of the paddock with their backs to the sun, and returned when
they reached the short side of the paddock. When they returned, they did not graze much
since they now faced the sun. They walked for a distance to the other short side and then
turned around to graze continually with their backs facing the sun. It may be the case
that they were changing the incident direction and area of solar radiation by adjusting
their posture, which is an effective way for animals to adjust the amount of environmental
radiant heat to maintain a constant body temperature [4,43].
Compared with medium and tall SSH, the grazing behavior on the short SSH was
relatively dispersed and continuous grazing time was shorter, possibly because insufficient
grass forced them to search for new grass more frequently, which resulted in the highest
proportion of walking behavior time. Animut et al. [
44
] have reported that decreasing
herbage allowance [
45
] increases the number of sheep’s steps. Moreover, the grazing
time on short SSH was less than on medium SSH, indicating that sheep might not eat
enough grass on short SSH. This was supported by the increase in running on short SSH,
as a result of trying to eat the grass outside the paddock and being electrocuted. The
sheep ran mainly because they touched the electrified paddock. The grazing time for
sheep on tall SSH was less than that on the medium SSH, but the resting time was the
longest, indicating that sheep took less time to consume enough grass. This was also the
conclusion of Wang et al. [
46
], who found that the relationship between grazing time and
SSH (
13 ≥SSH ≥5 cm
) was parabolic, opening upward. However, why the three sheep in
this study had similar average grazing times on both tall and short SSH requires further
investigation. Since only three sheep were used for analysis, statistical tests could not be
performed. To overcome this limitation, behavioral data collected from more individual
sheep are expected in future studies.
It was found that under the same sward conditions, the forage intake of grazing
livestock correlated positively with grazing time and speed, and individual forage in-
take [
47
,
48
]. Given that the three sheep were similar in age and weight, and assuming that
the individual grazing speed and grass intake were the same, the grass intake of three sheep
could be inferred from the predicted grazing time of sheep: P-Sheep > B-Sheep > R-Sheep.
R-Sheep always had the shortest grazing and longest walking times, which prompted
us to study the reasons for this phenomenon further. Our results demonstrated a good
potential for detecting individual differences in behaviors, and will facilitate the monitoring
of grazing sheep health, support farm decision-making and improve production efficiency.
5. Conclusions
In the training process of the sheep behavior classification model, testing the trained
model on continuous behavior segments was very important for evaluating the general-
ization ability and practical application performance of the trained model. Sensor type,
time window size, time window moving mode and algorithms all affected the accuracy
of continuous behavior segments classification. The accuracy of behavior classification
using T-acc data was higher than that for T-gyr data, and still higher when both data were
used simultaneously. The time window should be larger than the movement period of the
behavior. The 3 s time window showed higher accuracy than the 5 or 11 s time windows
when classifying the behavior of each second in continuous time. Stacking and ELM showed
stronger robustness on the CBS test dataset. The approach followed in this study can be
used to study individual behavior of sheep. In follow-up research, it will be necessary to
collect more data on individual sheep, to optimize the unbalance of training data datasets,
and to explore the method of judging the health of sheep through behavior time.
Author Contributions:
Conceptualization, Z.J. and W.W.; methodology, Z.J. and H.S.; software, Z.J.
and L.G.; validation, Y.L., J.Q. and W.Z.; formal analysis, W.W. and H.S.; investigation, W.Z. and Z.J.;
resources, W.W. and L.G.; data curation, L.G., K.W. and Y.L.; writing—original draft preparation,
Z.J.; writing—review and editing, H.S., Z.J. and B.X.; visualization, Z.J.; supervision, W.W.; project
administration, L.G. and J.Q.; funding acquisition, W.W. and L.G. All authors have read and agreed
to the published version of the manuscript.

Animals 2022,12, 1744 22 of 24
Funding:
This research was funded by the Major Science and Technology Program of Inner Mongolia
Autonomous Region, grant number 2020ZD0004; the National Key Research and Development
Program of China, grant number 2021YFD1300500; the Key Research and Development Program
of Ningxia Autonomous Region, grant number 2022BBF02021; and the Science and Technology
Innovation Project of the Chinese Academy of Agricultural Sciences, grant number CAAS-ASTIP-
2016-AII.
Institutional Review Board Statement:
This study was approved by the University of New England
Animal Ethics Committee, and followed the University of New England code of conduct for research
to meet the Australian Code of Practice for the Care and Use of animals (AEC17-006).
Data Availability Statement:
The data presented in this study are available upon request from the
corresponding author.
Acknowledgments:
We wish to thank Derek Schneider and Mitchell Welch (University of New
England) for their assistance, and the Precision Agriculture Research Group for the use of vehicles
and equipment for this research.
Conflicts of Interest: The authors declare no conflict of interest.
References
1.
Van Wettere, W.H.E.J.; Kind, K.L.; Gatford, K.L.; Swinbourne, A.M.; Leu, S.T.; Hayman, P.T.; Kelly, J.M.; Weaver, A.C.; Kleemann,
D.O.; Walker, S.K. Review of the impact of heat stress on reproductive performance of sheep. J. Anim. Sci. Biotechnol.
2021
,12,
859–876. [CrossRef]
2.
Roger, P.A. The impact of disease and disease prevention on welfare in sheep. In The Welfare of Sheep; Cathy, M.D., Ed.; Springer:
Dordrecht, The Netherlands, 2008; Volume 6, pp. 159–212. [CrossRef]
3.
Fogarty, E.S.; Swain, D.L.; Cronin, G.M.; Moraes, L.E.; Trotter, M. Behaviour classification of extensively grazed sheep using
machine learning. Comput. Electron. Agric. 2020,169, 105175. [CrossRef]
4.
Li, F.; Yang, Y.; Wang, H.; Lv, S.; Wei, W. Effect of heat stress on the behavior and physiology of sheep. J. China Agric. Univ.
2018
,
23, 43–49. (In Chinese)
5.
Guo, H. Prevention and treatment measures of common clinical diseases in sheep. Chin. Livest. Poult. Breed.
2020
,16, 155–156.
(In Chinese)
6.
Ren, K.; Karlsson, J.; Liuska, M.; Hartikainen, M.; Hansen, I.; Jørgensen, G.H.M. A sensor-fusion-system for tracking sheep
location and behaviour. Int. J. Distrib. Sens. Netw. 2020,16, 1550147720921776. [CrossRef]
7.
Yan, L.; Shao, Q.; Xi, G. Research progress of livestock behaviour intelligent monitoring system. J. Domest. Anim. Ecol.
2014
,35,
6–10. (In Chinese)
8.
Alvarenga, F.A.P.; Borges, I.; Palkoviˇc, L.; Rodina, J.; Oddy, V.H.; Dobos, R.C. Using a three-axis accelerometer to identify and
classify sheep behaviour at pasture. Appl. Anim. Behav. Sci. 2016,181, 91–99. [CrossRef]
9.
Barwick, J.; Lamb, D.W.; Dobos, R.; Welch, M.; Trotter, M. Categorising sheep activity using a tri-axial accelerometer. Comput.
Electron. Agric. 2018,145, 289–297. [CrossRef]
10.
Riaboff, L.; Shalloo, L.; Smeaton, A.F.; Couvreur, S.; Madouasse, A.; Keane, M.T. Predicting livestock behavior using accelerometers:
A systematic review of processing techniques for ruminant behaviour prediction from raw accelerometer data. Comput. Electron.
Agric. 2022,192, 106610. [CrossRef]
11.
Augustine, D.J.; Derner, J.D. Assessing herbivore foraging behavior with GPS collars in a semiarid grassland. Sensors
2013
,13,
3711–3723. [CrossRef]
12.
Sheng, H.; Zhang, S.; Zuo, L.; Duan, G.; Zhang, H.; Okinda, C.; Shen, M.; Chen, K.; Lu, M.; Norton, T. Construction of sheep
forage intake estimation models based on sound analysis. Biosyst. Eng. 2020,192, 144–158. [CrossRef]
13.
Fuentes, A.; Yoon, S.; Park, J.; Park, D.S. Deep learning-based hierarchical cattle behavior recognition with spatio-temporal
information. Comput. Electron. Agric. 2020,177, 105627. [CrossRef]
14.
Bailey, D.W.; Trotter, M.G.; Knight, C.W.; Thomas, M.G. Use of GPS tracking collars and accelerometers for rangeland livestock
production research. Trans. Anim. Sci. 2018,2, 81–88. [CrossRef]
15.
Bar, D.; Solomon, R.; Service, E. Rumination collars: What can they tell us. In Proceedings of the First North American Conference
on Precision Dairy Management, Toronto, Canada, 2–5 March 2010; Volume 2, pp. 214–215.
16.
Verdon, M.; Rawnsley, R.; Raedts, P.; Freeman, M. The behaviour and productivity of mid-lactation dairy cows provided daily
pasture allowance over 2 or 7 intensively grazed strips. Animals 2018,8, 115. [CrossRef]
17.
Barwick, J.; Lamb, D.W.; Dobos, R.; Welch, M.; Schneider, D.; Trotter, M. Identifying sheep activity from tri-axial acceleration
signals using a moving window classification model. Remote Sens. 2020,12, 646. [CrossRef]
18.
Mansbridge, N.; Mitsch, J.; Bollard, N.; Ellis, K.; Miguel-Pacheco, G.G.; Dottorini, T.; Kaler, J. Feature selection and comparison of
machine learning algorithms in classification of grazing and rumination behavior in sheep. Sensors 2018,18, 3532. [CrossRef]

Animals 2022,12, 1744 23 of 24
19.
Decandia, M.; Giovanetti, V.; Molle, G.; Acciaro, M.; Mameli, M.; Cabiddu, A.; Cossu, R.; Serra, M.G.; Manca, C.; Rassu, S.P.G.;
et al. The effect of different time epoch data settings on the classification of sheep behavior using tri-axial accelerometry. Comput.
Electron. Agric. 2018,154, 112–119. [CrossRef]
20.
Marais, J.; Le Roux, S.P.; Wolhuter, R.; Niesler, T. Automatic classification of sheep behaviour using 3-axis accelerometer data. In
Proceedings of the Twenty-Fifth Annual Symposium of the Pattern Recognition Association of South Africa (PRASA), RobMech
and AfLaT International Joint Symposium, Cape Town, South Africa, 27–28 November 2014; pp. 97–102.
21.
Le Roux, S.P.; Marias, J.; Wolhuter, R.; Niesler, T.; le Roux, S.P. Animal-borne behaviour classification for sheep (Dohne Merino)
and Rhinoceros (Ceratotherium simum and Diceros bicornis). Anim. Biotelem. 2017,5, 25. [CrossRef]
22.
Walton, E.; Casey, C.; Mitsch, J.; Vázquez-Diosdado, J.A.; Yan, J.; Dottorini, T.; Ellis, K.A.; Winterlich, A.; Kaler, J. Evaluation
of sampling frequency, window size and sensor position for classification of sheep behavior. R. Soc. Open Sci.
2018
,5, 171442.
[CrossRef]
23.
Lush, L.; Wilson, R.P.; Holton, M.D.; Hopkins, P.; Marsden, K.A.; Chadwick, D.R.; King, A.J. Classification of sheep urination
events using accelerometers to aid improved measurements of livestock contributions to nitrous oxide emissions. Comput.
Electron. Agric. 2018,150, 170–177. [CrossRef]
24.
Nóbrega, L.; Gonçalves, P.; Antunes, M.; Corujo, D. Assessing sheep behavior through low-power microcontrollers in smart
agriculture scenarios. Comput. Electron. Agric. 2020,173, 105444. [CrossRef]
25.
Guo, L.; Welch, M.; Dobos, R.; Kwan, P.; Wang, W. Comparison of grazing behaviour of sheep on pasture with different sward
surface heights using an inertial measurement unit sensor. Comput. Electron. Agric. 2018,150, 394–401. [CrossRef]
26.
Liu, Y. Research of Dairy Goat Behaviour Classification Based on Multi-Sensor Data. Master’s Thesis, Northwest A&F University,
Xianyang, China, 2020. (In Chinese).
27. Rokach, L. Ensemble-based classifiers. Artif. Intell. Rev. 2010,33, 1–39. [CrossRef]
28. Schapire, R.E. The strength of weak learnability. Mach. Learn. 1990,5, 197–227. [CrossRef]
29. Breiman, L. Bagging predictors. Mach. Learn. 1996,24, 123–140. [CrossRef]
30. Polikar, R. Ensemble based systems in decision making. IEEE Circuits Syst. Mag. 2006,6, 21–45. [CrossRef]
31.
Sun, Y.M.; Kamel, M.S.; Wong, A.K.C.; Wang, Y. Cost-sensitive boosting for classification of imbalanced data. Pattern Recognit.
2007,40, 3358–3378. [CrossRef]
32.
Galar, M.; Fernández, A.; Barrenechea, E.; Herrera, F. EUSBoost: Enhancing ensembles for highly imbalanced data-datasets by
evolutionary undersampling. Pattern Recognit. 2013,46, 3460–3471. [CrossRef]
33.
Huang, G.B.; Zhu, Q.Y.; Siew, C.K. Extreme learning machine: A new learning scheme of feedforward neural networks. In 2004
IEEE International Joint Conference on Neural Networks (IEEE Cat. No.04CH37541); IEEE: Budapest, Hungary, 25–29 July 2004;
Volume 2, pp. 985–990. [CrossRef]
34.
Huang, G.B.; Zhu, Q.Y.; Siew, C.K. Extreme learning machine: Theory and applications. Neurocomputing
2006
,70, 489–501.
[CrossRef]
35.
Huang, G.; Huang, G.B.; Song, S.J.; You, K.Y. Trends in extreme learning machines: A review. Neural Netw.
2015
,61, 32–48.
[CrossRef]
36.
Freund, Y.; Schapire, R.E. A decision-theoretic generalization of on-line learning and an application to boosting. J. Comput. Syst.
Sci. 1997,55, 119–139. [CrossRef]
37.
Santegoeds, O.J. Predicting Dairy Cow Parturition Using Real-Time Behaviour Data from Accelerometers. Master’s Thesis, Delft
University of Technology, Delft, The Netherlands, 2016.
38.
Tamura, T.; Okubo, Y.; Deguchi, Y.; Koshikawa, S.; Takahashi, M.; Chida, Y.; Okada, K. Dairy cattle behavior classifications based
on decision tree learning using 3-axis neck-mounted accelerometers. Anim. Sci. J. 2019,90, 589–596. [CrossRef] [PubMed]
39.
Andriamandroso, A.L.H.; Lebeau, F.; Beckers, Y.; Froidmont, E.; Dufrasne, I.; Heinesch, B.; Dumortier, P.; Blanchy, G.; Blaise, Y.;
Bindelle, J. Development of an open-source algorithm based on inertial measurement units (IMU) of a smartphone to detect cattle
grass intake and ruminating behaviors. Comput. Electron. Agric. 2017,139, 126–137. [CrossRef]
40.
Kleanthous, N.; Hussain, A.; Mason, A.; Sneddon, J. Data science approaches for the analysis of animal behaviours. In Intelligent
Computing Methodologies, Lecture Notes in Computer Science; Huang, D.S., Huang, Z.K., Hussain, A., Eds.; Springer International
Publishing: Cham, Switzerland, 24 July 2019; pp. 411–422. [CrossRef]
41.
Sakai, K.; Oishi, K.; Miwa, M.; Kumagai, H.; Hirooka, H. Behavior classification of goats using 9-axis multi sensors: The effect of
imbalanced datasets on classification performance. Comput. Electron. Agric. 2019,166, 105027. [CrossRef]
42.
Hamäläinen, W.; Jarvinen, M.; Martiskainen, P.; Mononen, J. Jerk-based feature extraction for robust activity recognition from
acceleration data. In Proceedings of the 2011 11th International Conference on Intelligent Systems Design and Applications
(ISDA), Córdoba, Spain, 22–24 November 2011; IEEE: Cordoba, Spain, 2011; pp. 831–836. [CrossRef]
43.
Hetem, R.S.; Strauss, W.M.; Heusinkveld, B.G.; de Bie, S.; Prins, H.H.T.; van Wieren, S.E. Energy advantage of orientation to solar
radiation in three African ruminants. J. Therm. Biol. 2011,36, 452–460. [CrossRef]
44. Animut, G.; Goetsch, A.L.; Aiken, G.E.; Puchala, R.; Detweiler, G.; Krehbiel, C.R.; Merkel, R.C.; Sahlu, T.; Dawson, L.J.; Johnson,
Z.B.; et al. Grazing behavior and energy expenditure by sheep and goats co-grazing grass/forb pastures at three stocking rates.
Small Rumin. Res. 2005,59, 191–201. [CrossRef]
45.
Lin, L.; Dickhoefer, U.; Müller, K.; Wurina; Susenbeth, A. Grazing behavior of sheep at different stocking rates in the Inner
Mongolian steppe, China. Appl. Anim. Behav. Sci. 2011,129, 36–42. [CrossRef]

Animals 2022,12, 1744 24 of 24
46.
Wang, S.P.; Li, Y.H. Behavioral ecology of grazing sheep V. Relationship between grazing behavior parameters and grassland
conditions. Acta Pratacult. Sin. 1997,4, 32–39. (In Chinese)
47.
Zhang, X.Q.; Hou, X.Y.; Zhang, Y.J. A comprehensive review of the estimation technology of feed intake and diet composition in
grazing livestock. Pratacult. Sci. 2012,29, 291–300. (In Chinese)
48.
Huang, R.Z. Study on Herbage Intake and Analysis of the Nutrient Surplus and Defict of Grazing Sheep in Seasonal Change.
Master’s Thesis, Shihezi University, Shihezi, China, 2016. (In Chinese)