Animal Environment and Welfare-2021ISAEWProceedings

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/369890676
Animal Environment and Welfare — Proceedings of International Symposium

2021
Book · October 2021
CITATIONS READS
0 294
3 authors:
Ji-Qin Ni Lingjuan Wang-Li

Purdue University North Carolina State University
191 PUBLICATIONS 4,227 CITATIONS 107 PUBLICATIONS 1,019 CITATIONS
SEE PROFILE SEE PROFILE
Chaoyuan Wang
China Agricultural University
84 PUBLICATIONS 927 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Heat and AQ stress in dairy cattle View project
Effects of crude protein reduction on gas emissions from an experimental swine building in Indiana View project
All content following this page was uploaded by Ji-Qin Ni on 08 April 2023.
The user has requested enhancement of the downloaded file.

Animal Environment and Welfare —Proceedings of International Symposium 2021
Animal
Environment
and Welfare
— Proceedings of International Symposium 2021
Edited by
Ji-Qin Ni, Lingjuan Wang-Li, Chaoyuan Wang
Chongqing Academy of Animal Sciences

Chongqing · China
Int. Symp. on Animal Environ. & Welfare Oct. 21–23, 2021, Chongqing, China
Animal Environment and Welfare

— Proceedings of International Symposium
October 21–23, 2021
Rongchang, Chongqing, China
动物环境和福利化养殖
国际研讨会论文集
2021 年 10 月 21–23 日
中国 ˑ 重庆 ˑ 荣昌
International Research Center for

Edited by
Ji-Qin Ni, Lingjuan Wang-Li, Chaoyuan Wang
Sponsored by
International Research Center for Animal Environment and Welfare (IRCAEW)
Chinese Society of Agricultural Engineering (CSAE)
China Agricultural University (CAU)
National Center of Technology Innovation for Pigs
Rongchang District Government, Chongqing
Hosted by
Chongqing Academy of Animal Sciences (CAAS)
Key Lab of Agricultural Engineering in Structure and Environment, MOA, China
Symposium Co-Chairs
Dr. Baoming Li, Professor, CAU, China
Dr. Hongwei Xin, Dean and Director, UT AgResearch, The University of Tennessee, USA
Proceedings Co-Chairs
Dr. Ji-Qin Ni, Professor, Purdue University, USA
Dr. Lingjuan Wang-Li, Professor, North Carolina State University, USA
Dr. Chaoyuan Wang, Professor, CAU, China
Program Co‐Chairs
Dr. Qiang Zhang, Professor, University of Manitoba, Canada
Dr. Weichao Zheng, Associate Professor, CAU, China
Organizing Committee
Dr. Yong Huang, Professor, CAAS, China (Chair)
Dr. Chaoyuan Wang, Professor, CAU, China
Dr. Feiyun Yang, Professor, CAAS, China
Ms. Wen Liu, Associate Professor, CAAS, China
Dr. Weichao Zheng, Associate Professor, CAU, China
Dr. Hao Li, Associate Professor, CAU, China
Dr. Qin Tong, Associate Professor, CAU, China
∙i∙
Acknowledgements
We wish to thank those who helped review part of the submitted full-length papers for the
symposium and worked with us to control and improve the quality of the papers. They were:
Prof. Jay Johnson USDA-ARS USA

Prof. Yansen Li Nanjing Agricultural University China
Prof. Longshen Liu Nanjing Agricultural University China
Prof. Allan P. Schinckel Purdue University USA
Prof. Qiang Zhang University of Manitoba Canada
The Editors
Disclaimer: It is the authors’ responsibility to obtain written permissions to use copyrighted

materials (e.g., figures, photos, tables) in their papers.
© 2021 of individual papers by the authors of the papers
∙ ii ∙
Welcome to ISAEW 2021

Dear ISAEW2021 Participants,
On behalf of the Board of Directors of the International Research Center for Animal Welfare
(IRCAW), we wish to extend a warm welcome to each of you, whether you are a returning veteran
or a first-timer participant of this international symposium!
Dating back to 2011 at the Chongqing Academy of Animal Sciences (CAAS) in Rongchang,
a group of researchers from seven institutions in USA, Canada, The Netherlands, and Australia
signed a historical Memorandum of Understanding to establish the IRCAW, headquartered at
CAAS, with a goal to promote international collaboration and advance the knowledge of animal
environment and welfare. This year marks the 10th anniversary of the Center’s establishment.
Today, 20 institutions representing 10 countries have joined the Center. During the past 10 years,
the Center has held six international symposia (including this one) and four center bi-annual
specialty workshops, plus the first Asian Conference on Precision Livestock Farming in 2020
(ACPLF2020). The symposia and workshops have attracted thousands of academic scientists,
graduate students, industry leaders and professionals, and government officials from more than
20 countries in six continents. Among the notable invited participants were Dr. Norm Scott of
Cornell University (member of the U.S. National Academy of Engineering), Dr. Jim Roth of Iowa
State University (member of the U.S. National Academy of Medicine), Dr. Robert Easter of
University of Illinois (former Chancellor of UIUC), Mr. Tom Hebert (former U.S. Undersecretary
of Agriculture), and Dr. Cathie Woteki of Iowa State University (member of the U.S. National
Academy of Medicine, former Chief Scientist of the White House, and former Dean of Iowa State
University), and Mr. Jeff Aiken (President of the Tennessee Farm Bureau Federation – the largest
state farm bureau in the United States with nearly 680,000 members), Prof. Maohua Wang of
China Agricultural University (member of the Chinese National Academy of Engineering), Prof.
Xiwen Luo of South China Agricultural University (member of the Chinese National Academy
of Engineering), and Prof. Defa Li of China Agricultural University (member of the Chinese
National Academy of Engineering). Over 70 technical sessions were conducted, with publication
of 268 research and review papers that composed six symposium proceedings.
These international symposia and specialty workshops have provided a convergence platform
for multi-disciplinary scientists from various parts of the world to jointly address global grand
challenges through exchange of recent research findings, identification of current challenges and
future research directions, and formation of synergistic focal groups to efficiently and effectively
conduct translational research; train next-generation scientists and leaders; and promote on-farm
adoption of technologies toward sustainably feeding the world. The characteristics of the
symposia in terms of global relevancy of the subject matters addressed, innovative and cutting-
edge nature of the research, quality of the presentations and publications, and unparalleled
hospitality of the hosts have established the symposium one of the must-attend academic
symposia, both nationally and internationally. The symposia have proven to be impactful in terms
of advancing scientific knowledge, developing technologies applicable for adoption by farmers
worldwide, training future leaders and workforce, and providing a roadmap for developing smart
or digital livestock farming that will be critical to sustainably feed the world for generations to
come.
The afore-mentioned achievements would not have been possible without the unwavering
support and partnership of CAAS under the outstanding leadership of President Zuohua Liu. The
tremendous collaborations and support by colleagues at China Agricultural University also have
been vital to the successes of the Center and are deeply appreciated.
∙ iii ∙
Although the COVID-19 situation is improving to various degrees globally, it remains

uncertain for overseas participants to travel to China and attend the symposium in person.
Therefore, the symposium will offer a hybrid format (in-person for the domestic participants and
online or video presentations for overseas attendees/contributors who are unable to attend in
person).
Wish ISAEW2021 a great success! Be well!
Hongwei Xin, Symposium Co-chair Baoming Li, Symposium Co-chair

Dean and Director Professor
The University of Tennessee, USA China Agricultural University, China
∙ iv ∙
Preface
International Symposium on Animal Environment and Welfare (ISAEW) 2021 is the sixth of
a series of biennial academic events. This proceedings book publishes scientific research and
review papers of the ISAEW 2021.
The ISAEW has a history of 11 years, starting from 2011. The first symposium, the
International Symposium on Health Environment and Animal Welfare (ISHEAW), was held from
October 19 to 22, 2011 in Rongchang, Chongqing City, China. The name of the symposium was
changed to International Symposium on Animal Environment and Welfare since 2013. All the
past five symposia were held in Rongchang in October (October 19–22, 2013; October 23–26,
2015; October 23–26, 2017; and October 21–24, 2019).
Planning of ISAEW 2021 began in October 2019 at the Board of Directors’ meeting of the
International Research Center for Animal Welfare (IRCAW). An Announcement and Call for
Abstracts of ISAEW 2021 was released in December 2020. The first abstract was submitted on
January 20, 2021. By May 30, 2021, 66 abstracts were received and 65 were accepted. Authors
of the accepted abstracts were invited for full-length paper submissions, which lasted for several
months. During this period, a couple of researchers, who had not submitted abstracts, were invited
to submit papers. From April 20 to September 30, 2021, 55 full papers were received. After format
correction, plagiarism checking, reviewing, editing, and requesting authors’ revisions, final
versions of 47 out of the 55 papers were accepted in this book. The 176 authors and co-authors of
these papers were from six countries (Belgium, Canada, China, Japan, The Netherlands, and
USA).
Papers in this book are divided into five Themes, which were outlined in the Announcement
and Call for Abstracts of ISAEW 2021 and are structured a little differently from the previous
ISAEW Proceedings. Authors selected the Themes for their papers at abstract submissions. A few
papers, of which the author did not indicate the Themes, were assigned to relevant Themes by the
Editors. The Themes of a couple of other papers were changed by the Editors, based on the
contents of the papers, from what the authors had selected.
Due to the worldwide COVID 19 pandemic and government-issued travel restrictions and
mandatory quarantine, the number of abstracts, especially those from out of China, submitted to
ISAEW 2021 was smaller compared with 2017 and 2019 (Table 1). However, this 2021
Proceedings book still publishes more papers than 2011 and 2015, and shows more participating
authors/coauthors than 2011, 2013, and 2015.
Table 1. Comparison of some statistics for the six ISAEW symposia.
Abstracts Papers Authors and
Symposium Countries, n b
received, n published, n co-authors, n a
ISHEAW 2011 NA 19 82 3
ISAEW 2013 NA 50 141 9
ISAEW 2015 62 44 121 11
ISAEW 2017 93 57 227 14
ISAEW 2019 97 51 183 13
ISAEW 2021 66 47 176 6
Note: NA = Not available. a An author or a coauthor of multiple papers in each proceedings
is only counted once; b Countries of the authors/co-authors, including China.
Although all authors were required to use the ISAEW 2021 paper template, not all
requirements were met in all the papers. Therefore, correction, typesetting, and typography were
performed on the papers to make the book consistent in style and format and look good on the
∙v∙
page with the best possible visual communication.

It was expected at the beginning of editorial work that the pandemic would greatly reduce the
in-person attendance at ISAEW 2021. For this reason, only an electronic version of this book will
be published so all in-person and on-line attendees can receive the same book on time. However,
the interior book design still considers the printed book. Therefore, some blank pages are included
to start each major book section, such as Table of Contents and all Themes, on a right page with
an odd page number. Although this book file in PDF format can be used to directly print a
hardcopy, a book cover with customized book spine is available upon request by contacting
jiqin@purdue.edu.
Like all the previous ISAEW Proceedings, this book is a valuable information source for
agricultural engineers, animal scientists, graduate students, and livestock and poultry producers.
It can also be a current reference for environmentalists, agricultural and environmental policy
makers, and animal production equipment manufacturers.
We would like to thank all the authors and co-authors who shared their research progresses,
findings, and experiences in this book. The patience and persistence of some authors were very
impressive as they had been requested to improve their papers in as many as five or six revisions.
Particular thanks must also go to IRCAEW; the symposium co-chairs, program co-chairs, and
organizing committee; and the five external paper reviewers. This book could not be completed
and published without the supports from all these organizations and individuals.
The Editors
October 10, 2021. Updated October 25, 2021
∙ vi ∙
Table of Contents
Acknowledgements .................................................................................................. ii
Welcome to ISAEW 2021 .......................................................................................iii
Preface ..................................................................................................................... v
Table of Contents ................................................................................................... vii
Theme I: Animal Environment Monitoring, Assessment, and Control .................. 1

Survival of Escherichia Coli in Airborne and Settled Poultry Litter Particles
Xuan Dung Nguyen, Yang Zhao, Jeffrey D. Evans, Jun Lin, Brynn Voy, .............. 3
Effects of Monochromatic Blue Light on Reducing the Adverse Impact of
Induced Cyclic Chronic Heat Stress During Thermal Manipulation of Broiler
Embryos
Li Zeng, Jinming Pan .......................................................................................... 10
Light Demand Characteristics, Production Performance and Changes of
Feeding Pattern of Broilers
Shouyi Wang, Jinming Pan ................................................................................. 18
A Comparison of Ventilation Systems in Three Different Types of Multi-
Floor Pig Buildings Using Numerical Simulation
Xiaoshuai Wang, Feiyue Hu, Tian Xie, Kaiying Wang ....................................... 26
Concentration and Emission of Particle Matters from an Intensive Dairy
Farm in Northern China
Yujian Lu, Lei E, Chaoyuan Wang, Zhenpeng Hu, Yu Liu .................................. 34
Comparative Assessment of the Breeding Egg Disinfection Using Slightly
Acid Electrolyzed Water and Benzalkonium Bromide
Chang Liu, Weichao Zheng, Ling Zhou, Yuxuan Sun, Shengqiang Han,
Zonggang Li ........................................................................................................ 42
Heating Time and Heating Load of Fattening Pig Houses in Different Climate
Regions in Winter
Fei Qi, Hao Li, Jinjun Huang, Zhengxiang Shi ................................................... 50
A Survey and Analysis of Antibiotic Resistance of Escherichia Coli in Feces
from Chinese Pig Farms
Mingyang Li, Yansen Li, Chunmei Li .................................................................. 58
Effect of Ventilation Fans and Types of Partition on the Physiological
Parameters of Dairy Pre-Weaned Individual-Housed Calves in a Calf Barn
Wanying Zhao, Zhengxiang Shi, Xinyi Du, Huiyuan Guan, Hao Li .................... 68
Assessing Air Quality in Three Types of Laying Hen Houses Using an
Internet of Things Technique
Zongyang Li, Shaojie Wang, Chaoyuan Wang, Boyu Ji, Yu Liu ......................... 76
∙ vii ∙
Disinfection Effect of Slightly Acidic Electrolyzed Water Applied in

Drinking Water Systems of Large-scale Dairy Farms
Xuedong Zhao, Chenxuan Zhu, Zhengxiang Shi ................................................. 84
Coupling Zirconium-based Metal-Organic Framework with Titanium
Dioxide for Photocatalytic Oxidation of Selected Livestock Odorant
Zun Man, Xiaorong Dai, Yuxuan He, Leiping Wang, Dezhao Liu ...................... 91
Investigation of Dielectric Barrier Discharge Reactor for Livestock Odorants
Reduction
Dezhao Liu, Xianwang Kong, Zhen Cai, Zhen Liu, Keping Yan ......................... 98
Air Resistance of Pigs in the Group Based on OpenFOAM: Influence of
Stocking Density and Live Weight
Xuefei Wu, Hao Li, Zhengxiang Shi .................................................................. 106
Impact of Carbon Dioxide on Hydrogen Sulfide Adsorption to Straw
Biochars
Dandan Huang, Mingshen Liang, Ning Wang, Qiyong Xu ............................... 114
Theme II: Animal Production Systems and Equipment ...................................... 121

A CFD Study on Ventilation Options for Cage-free Hen Houses in North
America
Long Chen, Eileen E Fabian (Wheeler) ............................................................ 123
Development of a Wide-Angle Egg Turning Incubator Capable of Enhancing
Gosling Embryonic Development, Hatchability, and Post-Hatch Growth
Binbin Guo, Zichun Dai, Aidong Sun, Huiman Ren, Zhendan Shi ..................... 131
Joint Data Envelopment Analysis and Life Cycle Assessment of the Canadian
Egg Industry, Differentiated by Housing Type
Ian Turner, Davoud Heidari, Nathan Pelletier ................................................. 138
Research Progress of Computer Vision-Based Intelligent Monitoring in Egg
Production
Zhonghao Chen, Jinming Pan, Hongjian Lin .................................................... 146
Performance Analysis and Validation of a Novel Dehumidification System
for Closed Livestock House in Cold Region
Ping Zheng, Jicheng Zhang, Jun Bao, Qiuju Xie .............................................. 155
Calibration of Discrete Element Method Parameters for Juvenile Manila
Clam (Ruditapes philippinarum) and Seeding Verification
Hangqi Li, Xiuchen Li, Hanbing Zhang, Guochen Zhang, Qian Zhang,
Wenbo Liu, Gang Mu ........................................................................................ 163
Development of a Test Strip for Detecting Sows’ Oestrus Based on Urine
Hormone
Kaidong Lei, Chao Zong, Yanbin Wen, Guanghui Teng, Hao Wang ................ 171
∙ viii ∙
Theme III: Animal Welfare Monitoring, Assessment, and Improvement .......... 177
Effects of L-leucine Administration on Thermoregulation in Chicks
Guofeng Han, Yangyang Cui, Mitsuhiro Furuse, Vishwajit S. Chowdhury,
Yansen Li, Chunmei Li ...................................................................................... 179
Physiological Responses in Dairy Cows to Increasing Temperatures at
Different Relative Humidity and Air Velocity Levels
Mengting Zhou, André J.A. Aarnink, Thuy T.T. Huynh, Peter W.G. Groot
Koerkamp .......................................................................................................... 186
Effects of Keel Bone Fracture on Behavior, Welfare, and Production
Performance of Laying Hens Housed Individually in Furnished Cages
Haidong Wei, Yanju Bi, Yanru Feng, Runxiang Zhang, Jun Bao ..................... 193
A Simulation Model for Detection of Sow Lameness and Estimation of
Ground Reaction Force
Xiaojie Yan, Qiang Zhang, Laurie Connor, Nicolas Devillers, Kristopher
Dick ................................................................................................................... 201
Broiler Gait Score in Relation to Bird Body Weight and Activity
Xiao Yang, Yang Zhao, Hao Gan, Shawn Hawkins, Liz Eckelkamp, Maria
Prado, Robert Burns, Tom Tabler ..................................................................... 209
Effects of Pre-Weaning Co-Mingling Piglets on Behavior and Growth
Performance of Pigs after Weaning
Tongshuai Liu, Xue Hui, Xiao Yang, Lei Xi, Pu Cheng, Zhenzhen Ji,
Xuanyang Li, Zhixiao Yang, Kunhua Zhu, Weidong Liu, Wei Ma, Zhifang
Shi, Jingfeng Zhang, Jingjing Kang .................................................................. 214
L-selenomethionine in Cardiotoxicity of Ammonia through Inhibiting
Autophagy Mediated via a Signaling Pathway
Zheng Cheng, Yutao Li, Yufu Shu, Xin Li, Xiaohong Zhang, Honggui Liu ....... 222
Impacts of Keel Bone Damage on Fear Behavior and Physiological
Responses of Laying Hens
Runxiang Zhang, Haidong Wei, Yanru Feng, Hanlin Yu, Susu Ding,
Haoyang Nian, Hengyi Zhang, Qian Zhao, Jianhong Li, Jun Bao ................... 230
Prediction of Sow Physiological Parameters Based on Support Vector
Machine and Extreme Gradient Boosting Models
Yanrong Zhuang, Mengbing Cao, Shulei Li, Yu Liu, Jianlong Zhang,
Guanghui Teng .................................................................................................. 238
Effect of High-Ammonia Environment on the Gene Networks Associated
with Oxidative Metabolism, Inflammation, and Apoptosis in “Gut-Brain”
Axis
Yutao Li, Jing Wang, Jianxing Wang, Jun Bao ................................................. 246
Effects of Three Ventilation Methods on the Environmental Conditions of
Pig Houses, Pig Production Performance and Abnormal Behavior during
Winter Fattening
Zhifang Shi, Xuanyang Li, Lei Xi ...................................................................... 252
∙ ix ∙
Relationship between Stereotypic Behaviors, Physiological Characteristics,

and Immune System Biomarkers of Confined Pregnant Sows
Lei Pan, Susu Ding, Haoyang Nian, Runxiang Zhang, Xiang Li, Jun Bao ....... 260
Theme IV: Digital and Smart Technologies for Managing and Improving Animal
Environment and Welfare .............................................................................. 267
Cross-modality Interaction Network for Equine Activity Recognition Using
Time-Series Motion Data
Axiu Mao, Endai Huang, Weitao Xu, Kai Liu ................................................... 269
Early Detection of Chicken Diseases Based on Clinical Symptom
Monitoring: A Review
Pengguang He, Jinming Pan, Hongjian Lin ..................................................... 277
A Key Frame Selection Method for Creating Deep Learning Training Set in
Animal Research Involving Time-Series Video Data
Endai Huang, Axiu Mao, Haiming Gan, Kai Liu .............................................. 287
Monitoring Environmental and Behavioral Aspects on Dairy Cattle Farms to
Reduce Heat Stress
Daniela Lovarelli, Marcella Guarino ............................................................... 295
Development of a Point Cloud Acquisition System and 3D Point Cloud
Reconstruction Method for Beef Cattle
Jiawei Li, Qifeng Li, Weihong Ma, Xianglong Xue, Luyu Ding, Ligen Yu,
Ronghua Gao .................................................................................................... 302
A Method for Pig Individual Identification and Its Body Parts Recognition
Mengru Wu, Qiuju Xie, Xin Li, Muyu Yang, Jun Bao, Honggui Liu ................. 310
Robust Audio Fingerprint Algorithm for Automatic Recognition of Laying
Hens’ Vocalizations
Tiantian Du, Ligen Yu, Tonghai Liu, Qingfei Xue, Rui Meng, Luyu Ding,
Weihong Ma, Ronghua Gao, Qifeng Li ............................................................. 318
A Deep Learning-Based Fusion Method of Infrared Thermography and
Visible Image for Pig Body Temperature Detection
Qiuju Xie, Mengru Wu, Muyu Yang,Jun Bao, Xin Li, Honggui Liu,
Haiming Yu, Ping Zheng ................................................................................... 326
Real-time Monitoring of Exhaust Fan Operation Status in a Livestock House
Using Image Analysis
Luyu Ding, Yang Lv, Qifeng Li, Ligen Yu, Ronghua Gao, Weihong Ma,
Qinyang Yu ........................................................................................................ 334
Feasibility Analysis of Ammonia Monitoring in Piggery Environmental
Monitoring System Based on Internet of Things
Jinrui Zhang, Hua Wang, Deyong She, Xiaoliu Xue, Zhonghong Wu, Jijun
Liu, Meizhi Wang .............................................................................................. 342
Automatic Recognition for Rumination Behavior of Dairy Cows Based on
Machine Learning
Liwei Wang, Qiuju Xie, Yidan Xu ..................................................................... 350
∙x∙
Why Precision Livestock Farming Can Generate a More Sustainable

Livestock Sector
Daniel Berckmans ............................................................................................. 358
Theme V: Manure and Mortality Management and Byproduct Development in

Animal Agriculture ......................................................................................... 369
Field Experience of Removing and Land Application of Dairy Lagoon Solids
Timothy Canter, Teng-Teeh Lim, Joseph Zulovich ........................................... 371
Indexes ...................................................................................................................... 379

Author Name Index .............................................................................................. 381
Author Affiliation Index ...................................................................................... 385
∙ xi ∙
∙ xii ∙
Theme I:
Animal Environment Monitoring,
Assessment, and Control
∙1∙
∙2∙
Survival of Escherichia Coli in Airborne and

Settled Poultry Litter Particles
Xuan Dung Nguyen a, Yang Zhao a,*, Jeffrey D. Evans b, Jun Lin a, Brynn Voy a,
Liesel Schneider a, Joseph L. Purswell b
a
Department of Animal Science, The University of Tennessee, Knoxville, TN 37996, USA
b
USDA, Agriculture Research Service, Poultry Research Unit, Mississippi State, MS 39762, USA
* Corresponding author email: yzhao@utk.edu
Abstract
Airborne Escherichia coli (E. coli) in the poultry environment can migrate inside and outside
houses through air movement. The airborne E. coli, after settling on surfaces, could be re-
aerosolized or picked up by vectors (e.g., caretakers, rodents, transport trucks) for further
transmission. To assess the impacts of E. coli transmission among poultry farms, understanding
the survivability of the bacteria is necessary. The objective of this study is to determine the
survivability of airborne and settled E. coli under laboratory environmental conditions. Poultry
litter inoculated with E. coli was aerosolized using a dust generator for 20 min in a biosecurity
cabinet. To determine the survivability of airborne E. coli, AGI-30 bioaerosol samplers were used
to collect the E. coli at 0 and 20 min after the aerosolization. The sampling at each time point
lasted for 10 minutes. The half-life time of airborne E. coli was then determined by comparing
the number of colony-forming units (CFU) of the two samplings. To determine the survivability
of settled E. coli, sterile Petri-dishes were placed on a cabinet floor right after the aerosolization
to collect settle E. coli. The total number of Petri-dishes were then divided into two groups, with
each group being quantified for culturable E. coli concentration and of dust particle weight at 24-
hr intervals. The survivability of settled E. coli was then determined by comparing the CFU of E.
coli per mg litter collected in the Petri dishes in the two groups. In addition, the survivability of
E. coli in the litter was also determined by the change in bacterial concentration in the inoculated
litter. The results of the study showed that the airborne E. coli and settled E. coli can survive a
long period of time, especially the settled one. The survivability test for airborne E. coli had shown
that the half-life time was 5.70 ±1.18 minutes, and 4.90 ±1.6 minutes for the bacteria in total dust
and the dust with the size larger than 2.5 µm, respectively. The survivability of settled E. coli was
much longer with the half-life time of settled E. coli was 9.63 ±1.64 hours.
Keywords: Airborne E. coli, settled E. coli, survivability, airborne transmission, poultry
1. Introduction
The U.S. poultry industry is among the World’s largest poultry production. Poultry products
originating in the United States primarily consist of meat from broilers and turkeys and eggs from
commercial layers. The combined value of production from these products in 2014 exceeded $48
billion. These products provide important and affordable sources of dietary protein to the domestic
population. In addition, approximately 18% of U.S. poultry products are exported and poultry
production in the U.S. has been estimated to provide over 1 million jobs. However, the outbreak
of infectious diseases is one of the biggest challenges for the poultry industry. For example, the
Highly Pathogenic Avian Influenza (HPAI) outbreaks in the U.S. in 2015 which resulted in losses
of over 50 million birds and 3.3 billion dollars (Torremorell et al., 2016).
Escherichia coli (E. coli) is a member of the Enterobacteriaceae family and is commonly
associated with the intestinal tract of warm-blooded animals and the environment in which these
animals reside. In poultry, E. coli primarily inhabits the lower gastrointestinal tract as an indicator
for the poultry environmental quality and exists there as an important commensal species.
Typically, E. coli are harmless, but some E. coli strains may be pathogenic in nature and their
virulence relates to the loss of the poultry industry. Pathogenic E. coli strains have been associated
with a wide range of diseases outside of the lower gastrointestinal tract in poultry and are
∙3∙
commonly referred to as APEC or avian pathogenic E. coli (Saif et al., 2008). Avian pathogenic
E. coli causes the systemic disease colibacillosis in broilers, which is commonly characterized by
the triad of lesions or early death. The severity of APEC disease depends on the health status of
the host, virulence characteristics of the E. coli strain, and other predisposing factors such as
stress. Approximately 30% of broiler flocks in the U.S are infected by subclinical colibacillosis
(Fancher et al., 2020).
The litter is a major reservoir of microorganisms in poultry environment (Carpenter, 1986).
The dry matter contents can be about 70–80% of litter mass and it can contain several biological
particles and compounds that can affect the quality of the poultry environment (Schulz et al.,
2016). With a large number of dust particles, and under bird activity, the poultry environment is
a highly dusty environment.
Air in the poultry houses not only contains odors and gases but also bacteria such as E. coli.
E. coli from manure first deposits into poultry litter and then may be aerosolized through bird
activities (Zhao et al., 2014). Ventilation systems can drive their migration across a poultry house
or even from barn to barn. Airborne E. coli have been shown to account for 2–6 % of the total
airborne bacteria in the poultry house (Zucker et al., 2000). With the high concentration of E. coli
and the possibility of barn-to-barn transmission, the airborne E. coli can harm the entire wide
range of environment outside the poultry house, and they can deposit on any surface near the
poultry houses. The modeling of barn-to-barn airborne transmission has been performed to assess
the probability of airborne transmission in a study conducted in 2019 (Zhao et al., 2019). The
probability of airborne infection is affected by several factors including farm type, flock size and
distance of transmission where the survivability of pathogen is known as a key factor of input
parameter to decide the accuracy of the modeling. Moreover, the survivability of E. coli on certain
surfaces (or settled E. coli) under room conditions had been reported to reach or exceed 28 days
(Wilks et al., 2005). With such a long stay on the surface, they can infect other chicken houses.
Moreover, through vectors, they can spread to larger areas. These all raise the question of how
long the airborne E. coli can survive in the air and on the physical surfaces.
The aim of this study was to investigate the survivability of airborne E. coli and settled E.
coli. The study involved assessing the survivability of airborne E. coli generated by a dust
aerosolization system and the survivability of settled E. coli on a flat surface overtime under the
representing room conditions.
2. Materials and Methods
To investigate the survivability of the airborne E. coli and the settled E. coli, the experiments
were run in a test chamber in a Biosafety Level 2 (BSL-2) laboratory.
2.1. Preparation of E. coli solution
The E. coli strain used in the present study was Escherichia coli GFP (ATCC® 25922GFP™).
E. coli strain was cultured at 37°C for 24 hours in ATCC® Medium 2855 (Tryptic Soy Agar
‘TSA’ and Tryptic Soy Broth ‘TSB’ with 100 mcg ml-1 Ampicillin). The bacterial concentrations
of E. coli in the solution after 24 hours were from 108 to 109 CFU ml-1.
2.2. Litter preparation
The litter from the commercial broiler farm was first collected and stored in a container placed
inside the test chamber. It was then brought back to the Biosafety Level 1 (BSL-1) laboratory to
analyze the dry matter content. After that, it was autoclaved and divided into identical-size
aluminum boxes with the amount of 6 kg per box. The boxes were sealed by aluminum foils and
covered by plastic caps to avoid contamination. They were stored in a 4oC fridge until being used.
It was important to prepare litter so that the bacteria were evenly distributed. To do that, 240
g of litter needed for this experiment was evenly distributed into 40 ceramic cups. The ability to
generate airborne dust has been tested in a previous experiment, and the results show that 240 g
of litter added to the mixer, would produce dust concentrations ranging from 0.9 to 1.1 mg m-3
∙4∙
which is within a typical range of dust concentration in commercial poultry farm. To prepare litter
inoculated with E. coli, a set of 40 ceramic cups with identical shapes was used to hold the litter.
In each cup, 6 g of litter was prepared and mixed with 6 ml of E. coli cultured solution. The 6 ml
bacteria solution was sprayed evenly onto the litter in each cup. In the meantime, an aluminum
spoon was used to gently mix the litter and E. coli solution. The mixtures then went through a
process of drying at 22oC and 52–67% relative humidity (RH) for 48 hours until the dry matter
content (DMC) of the mixture reached about 71.07 ± 5.42%. The bacteria concentration in each
cup was approximately 107 CFU g-1 litter after the drying process. The litter containing E. coli
was then transferred from 40 ceramic cups to a metal bowl of the mixer for aerosolization. In the
bowl, the litter was gently mixed up one more time before aerosolization.
2.3. Aerosolization system
The aerosolization system in this experiment includes a mixer for the airborne E. coli
experiment and two mixers, a compact stand mixer (model DCSM350GBRD02, New York City,
NY), and a professional blender (model BL610, Newton, MA), for settled E. coli. The mixer’s
dimension is 0.3 m L × 0.2 m W × 0.3 m H with a 3.3 l stainless steel bowl. The blender’s
dimension is 0.25 m L × 0.19 m W × 0.43 m H with a 2.1 l pitcher. They were set up at the highest
speed to ensure the bacteria concentration in the air high enough. A stir fan was also used to
distribute the airborne E. coli in the chamber evenly.
2.4. Test chamber
A test chamber used for assessing the three samplers was a laboratory 4-port glovebox. This
glovebox (2100 series, Cleatech, Orange, CA) was a non-vacuum unit with two internal access
doors with stainless steel frame, and a removable fully gasketed back wall. The dimension of the
test chamber is 1.5 m L × 0.6 m W × 0.6 m H. The chamber was well sealed to prevent dust-laden
particles from spilling out. It was also equipped with a temperature and RH sensor for
continuously monitoring the inside thermal environment.
In the settled E. coli experiment, the chamber was modified to meet the purpose of the
experiment. Initial results showed that the aerosolization space of the entire chamber was too large
which led to the low concentration of airborne E. coli and dust particles. Thus, the chamber had
been modified by halving the aerosolization space through the use of acrylic film that splits the
chamber. The aerosolization space after modification was 0.75 m L × 0.6 m W × 0.6 m H.
2.5. Air samplers
To evaluate and compare sampling performances, the AGI-30 impinger was used to collect
E. coli-laden dust particles in a test chamber in a BSL-2 laboratory at the Animal Science
Department, University of Tennessee. The AGI-30 operates at 12.5 l min-1. The airborne
compounds were sucked utilizing a vacuum pump through a fine nozzle in which the particles
were accelerated and then impact directly into the 20 ml TSB. The AGI-30 had been proven to
have the highest performance among three commonly used samplers (Anderson six-stage, AGI-
30, and BOBCAT ACD-200).
2.6. System setup and sample collection for airborne E. coli
Two hundred and forty grams (240 g) of litter, which contained ~104 log10 CFU [mg litter]-1
of E. coli, were prepared and placed in the mixer. The mixer was placed in the center of the
chamber to help evenly distribute the dust particles carrying E. coli. The mixer was fixed to the
chamber surface by means of suckers, preventing it from moving during the running process. The
stir fan was placed at the corner of the chamber to aid in distributing airborne particles. The AGI-
30 was placed near the steel bowl of the mixer 5 cm away.
The sampling process includes two-time points with a total running period of 50 min including
20 minutes of the running mixer, 10 minutes of 1st sampling, 10 minutes of resting time, and 10
minutes of 2nd sampling. After running the mixer with the fan on for 20 min, the mixer and the
fan had been stopped and started the AGI-30 and a dust monitor (Dusttrak DRX Aerosol Monitor
∙5∙
8533, TSI Inc., Shoreview, MN) for 10 min. Then, taking out the AGI-30 sampling tube
containing TSB-E. coli solution left the aerosolization system for 10 min (no run for 10 min) and
continued to run AGI-30 and the dust monitor for 10 min. The experiment had been repeated 7
times.
2.7. System setup and sample collection for settled E. coli
Two hundred and forty grams (240 g) of litter, which contained ~104 log10 CFU [mg litter]-1
of E. coli, were mixed gently and divided into two parts with 120 g for each mixer. The mixers (a
stand mixer and a blender) were placed side-by-side in the center of modified aerosolization space
to distribute airborne E. coli-dust particles evenly. The stir fan was placed at the corner to support
the aerosolization. Four empty Petri dishes were first weighed and placed inside the aerosolization
space. The Petri dishes were arranged evenly on both sides of the mixers. To avoid position
confounding effect, the Petri dishes were arranged randomly for each even of the total of 4
experiment evens.
The sampling process started with 15 min of generating dust-laden particles carrying E. coli
by turning on the aerosolization system. The dust concentration during the mixer running time
was also monitored by Dusttrak. Then, turned off the aerosolization system, and covered the four
Petri dishes with caps sealed by parafilm. Two Petri dishes were immediately analyzed to quantify
viable E. coli via traditional culture technique. The remaining 2 Petri dishes were left at room
temperature at 20oC, 62 % RH for 24 hours. After that, they were analyzed to quantify viable E.
coli by the same culture technique.
2.8. Statistical analysis
Means and standard deviations were calculated by using SAS 9.4. Total 7 replicates for
airborne E. coli experiment and 4 replicates for settled E. coli yields robust statical analysis for
calculating the half-life time.
2.9. Half-life time
The half-life time is the time interval needed for bacteria to decrease by half (Zhao et al.,
2011b). It was calculated by the following Eq. (1).
𝑙𝑜𝑔10 2 × 𝑇
𝑡1/2 = 𝐶 (1)
𝑙𝑜𝑔10 ( 𝑣𝑖𝑎𝑏𝑙𝑒 𝑏𝑎𝑐𝑡𝑒𝑟𝑖𝑎⁄𝐶 ′ )
𝑣𝑖𝑎𝑏𝑙𝑒 𝑏𝑎𝑐𝑡𝑒𝑟𝑖𝑎
where t1/2: half-life time (min); T= 20 (min) for airborne E. coli and 24 (hr) for settled E. coli test;
Cviable bacteria : Bacteria concentration for the first sampling event (in settled E. coli experiment
is viable bacteria concentration for 2 first picked up dishes) (CFU mg-1); C′ viable bacteria : Bacteria
concentration for the second sampling even (in settled E. coli experiment is viable bacteria
concentration for the 2 second picked up dishes) (CFU mg-1).
2.10. Determining airborne E. coli concentration
Each air sample collected by AGI-30 in liquid form (in TSB medium) was used to quantify
viable E. coli via traditional culture techniques. After vortexing for 5 s, a 0.1 ml subsample and
0.1 ml serially diluted (1:10) samples were be plated onto TSA agar plates. The plates were
aerobically incubated at 37oC for 24 h. The visible E. coli colonies formed on plates (30 to 300
colonies) were determined. Based on the culture results and the sampled air volume, airborne E.
coli concentrations were calculated in logarithm colony-forming units per cubic meter (log10 CFU
m-3) using Eq. (2).
𝑁 × 10𝑛 1
𝐶 = 𝑙𝑜𝑔10 ( × 𝑉𝑠 × ) (2)
𝑉𝑝 𝑉𝑎
where C = the airborne bacteria concentration, log10 CFU m-3;
N = the number of colonies on a
countable plate (30 to 300 colonies); n = serial dilution factor (n = 0 for undiluted sample, n = 1
for 10-fold diluted sample, etc.); VP = the sample volume plated, ml (VP = 0.1 ml in this study);
∙6∙
Vs = the total volume of the original liquid sample, ml; V a = the total air volume sampled using
the bioaerosol samplers, m3.
2.11. Determining settled E. coli concentration
Each settled sample on an empty petri dish was used to quantify viable settled E. coli. After
adding 10 ml of TSB medium (the culture medium) in each petri dish, the petri dish was gently
shaken to wash the petri dish surface and draw settled E. coli into TSB. Then, the viable E. coli
was determined as in the airborne E. coli concentration test with the Eq. (2).
3. Results and Discussion
3.1. Environmental conditions for airborne E. coli and settled E. coli survivability test
The environmental conditions applying for the airborne E. coli and the settled E. coli
experiments were shown in Table 1. The conditions applied in the airborne E. coli experiments
were stable with DMC for E. coli-litter mixture at 71%, E. coli concentration in the litter was
about 104 CFU mg-1, RH was 55% and the temperature was about 22oC. There was a difference
between the dust concentration of the airborne E. coli experiment and settled E. coli because of
the requirements of a different experiment.
In the settled E. coli test, the chamber had been modified and made the aerosolization space
smaller. Thus, the dust concentration was slightly higher compared to the airborne E. coli test.
There was a big fluctuation of temperature and RH during the aerosolization process. This could
be explained by the heat generated during the two mixers running in a smaller space compared to
one mixer in the airborne test. The bacteria concentration and DMC were quite stable and similar
to the airborne E. coli test.
Table 1. Environmental conditions (Mean±SD) for airborne E. coli and settled E. coli test.
Test Airborne E. coli Settled E. coli
DMC of litter (%) 71.07±5.42 72.13±1.12
Bacteria concentration in litter (CFU mg-1) 4.36±0.62 3.97±0.51
RH (%) 54.83±5.04 63.0±7.33
Temperature (oC) 22.1±1.41 27.7±5.1
Dust concentration (mg m-3) 1.0±0.1 1.16±0.4
3.2. Airborne E. coli survivability
Parameters collected from the experiment to calculate the half-life time were listed in the
following Table 2.
Table 2. Airborne E. coli parameters (Mean±SD).
Time 1st sampling 2nd sampling
-3
Total dust concentration (mg m ) 1.32±0.73 0.75±0.26
Dust particle >2.5 µm concentration (mg m-3) 0.03 ±0.02 0.016±0.012
Airborne E. coli concentration (log10 CFU m-3) 6.43±0.78 5.03±0.98
Airborne E. coli concentration divided by total dust
6.3±0.78 5.15±0.88
concentration (log10 CFU mg-1)
Airborne E. coli concentration divided by dust particle
7.95±0.66 6.82±0.46
>2.5 µm concentration (log10 CFU mg-1)
Based on the parameters from Table 2., the half-life time has been calculated. The data showed
that the half-life time of the E. coli strain based on total dust was 5.70 ± 1.18 minutes, and 4.90 ±
1.6 minutes for dust particles with a size larger than 2.5 µm. These number indicated that the half-
life time of the bacteria based on total dust was not significantly different compared to dust with
size larger than 2.5 µm. However, our initial hypothesis was that half-life time was longer for total
dust. This hypothesis assumes that survival of E. coli is better in large particles. With the large
particles, E. coli attached are provided a better protection from ambient environmental change.
∙7∙
The E. coli attached to the smaller particles are more sensitive under the environmental influences,
for instance in a study conducted by Zhao, E. coli show the vulnerability when they expose to
harsh environmental variation (Zhao et al., 2011a; Zhao et al., 2011b). The discrepancy between
the results and our hypothesis is because the environmental conditions applied in this study was
stable at room conditions which were quite favorable for the survival of bacteria. Furthermore,
the nature of large variation in bacteria culture may be another reason of not being able to detect
the difference. All the above mentioned warrant the investigation on the relationship of bacteria
survival with particle sizes in future research.
3.3. Settled E. coli survivability
The survivability of settled E. coli has been tested over 24 hours. After 24 hours, the
concentration of settled E. coli declined from 3.74 ± 0.07 to 2.97 ± 0.15 log10 CFU mg-1. The E.
coli concentration downed about 1 log in the first day, and the half-life time was 9.63 ± 1.64 hours.
The settled E. coli could be found everywhere near poultry barns where bacterial contamination
could deposit on any surfaces, e.g., poultry house walls, table, or farmers’ equipment. In addition,
with the longtime of survivals, settled E. coli can deposit on many types of vectors such as poultry
transport vehicles, and spread to a larger area. A study conducted in 2005 (Wilks et al., 2005) also
mentioned the survivability of E. coli at room temperature. The authors had indicated that the E.
coli strain used in their study (E. coli O157) can survive at room temperature on surfaces over 28
days. After first 180 min, the total number of viable E. coli fell by 1 log, and by 5 log after 2 days.
So, there is a difference between Wilks’s data and the data collected from this study. This
difference can be explained by the difference of Wilks’s study methodology and this study
methodology. This study tested the E. coli survivability after being aerosolized in the air and then
settled on surface. Wilks only tested the survivability of E. coli directly placed on some certain
surfaces without aerosolization process.
4. Conclusions
The study has demonstrated the survivability of airborne and settled E. coli under room
conditions. Half-life time of E. coli in total dust and in dust particles greater 2.5 µm has also been
reported. Based on the results, we conclude that 1) there was no statistically difference among
airborne E. coli survivability under two different categories of dust size (total dust, and dust with
size larger than 2.5 µm); 2) the settled E. coli can survive longer than airborne E. coli with half-
life time up to over 3 days. The results of this study will be tested in the future under the poultry
environment where the conditions are more fluctuated.
Acknowledgments
Financial support for the study was provided by the USDA National Program (Project
No.6064-13000-013-00D). I appreciate the cooperation and assistance of the broiler grower in
providing the broiler litter. This research was supported by the USDA National Institute of Food
and Agriculture multistate project under accession number 1026643.
References
Carpenter, G. (1986). Dust in livestock buildings—review of some aspects. Journal of Agricultural
Engineering Research, 33(4), 227–241. https://doi.org/10.1016/S0021-8634(86)80038-5
Fancher, C. A., Zhang, L., Kiess, A. S., Adhikari, P. A., Dinh, T. T., & Sukumaran, A. T. (2020). Avian
pathogenic Escherichia coli and Clostridium perfringens: Challenges in no antibiotics ever broiler production
and potential solutions. Microorganisms, 8(10), 1533. https://doi.org/10.3390/microorganisms8101533
Saif, Y., Barnes, H., Glisson, J., Fadly, A., McDougald, L., & Swayne, D. (2008). Diseases of poultry.
12. Ames, Iowa: Blackwell Pub Professional, 452–514.
Schulz, J., Ruddat, I., Hartung, J., Hamscher, G., Kemper, N., & Ewers, C. (2016). Antimicrobial-resistant
Escherichia coli survived in dust samples for more than 20 years. Frontiers in microbiology, 7, 866.
https://doi.org/10.3389/fmicb.2016.00866
Torremorell, M., Alonso, C., Davies, P. R., Raynor, P. C., Patnayak, D., Torchetti, M., & McCluskey, B.
(2016). Investigation into the airborne dissemination of H5N2 highly pathogenic avian influenza virus during
∙8∙
the 2015 spring outbreaks in the Midwestern United States. Avian diseases, 60(3), 637–643.
https://doi.org/10.1637/11395-021816-Reg.1
Wilks, S., Michels, H., & Keevil, C. (2005). The survival of Escherichia coli O157 on a range of metal
surfaces. International journal of food microbiology, 105(3), 445–454. https://doi.org/10.1016/j.ijfoodmicro.
2005.04.021
Zhao, Y., Aarnink, A., De Jong, M., & Groot Koerkamp, P. W. G. (2014). Airborne Microorganisms
From Livestock Production Systems and Their Relation to Dust. Critical Reviews in Environmental Science
and Technology, 44. https://doi.org/10.1080/10643389.2012.746064
Zhao, Y., Aarnink, A. J., Doornenbal, P., Huynh, T. T., Koerkamp, P. W. G., de Jong, M. C., & Landman,
W. J. (2011a). Investigation of the efficiencies of bioaerosol samplers for collecting aerosolized bacteria using
a fluorescent tracer. I: Effects of non-sampling processes on bacterial culturability. Aerosol Science and
Technology, 45(3), 423–431. https://doi.org/10.1080/02786826.2010.543196
Zhao, Y., Aarnink, A. J. A., Doornenbal, P., Huynh, T. T. T., Koerkamp, P. W. G. G., de Jong, M. C. M.,
& Landman, W. J. M. (2011b). Investigation of the efficiencies of bioaerosol samplers for collecting
aerosolized bacteria using a fluorescent tracer. II: sampling efficiency and half-life time. Aerosol Science and
Technology, 45(3), 432–442. https://doi.org/10.1080/02786826.2010.543197
Zhao, Y., Richardson, B., Takle, E., Chai, L., Schmitt, D., & Xin, H. (2019). Airborne transmission may
have played a role in the spread of 2015 highly pathogenic avian influenza outbreaks in the United States.
Scientific reports, 9(1), 1–10. https://doi.org/10.1038/s41598-019-47788-z
Zucker, B. A., Trojan, S., & Müller, W. (2000). Airborne gram-negative bacterial flora in animal houses.
J Vet Med B Infect Dis Vet Public Health, 47(1), 37–46. https://doi.org/10.1046/j.1439-0450.2000.00308.x
∙9∙
Effects of Monochromatic Blue Light on Reducing the

Adverse Impact of Induced Cyclic Chronic Heat Stress
During Thermal Manipulation of Broiler Embryos
Li Zeng, Jinming Pan *
College of Biosystems Engineering and Food Science, Zhejiang University,
Hangzhou, Zhejiang 310058, China
* Corresponding author. Email: panhouse@zju.edu.cn
Abstract
The aim of the present study was to investigate the effects of monochromatic blue light (BL)
on reducing the adverse effect of induced cyclic chronic heat stress in thermal manipulation of
broilers by subjecting embryos to heat stress during incubation development. Blue light
significantly lowered malondialdehyde and corticosterone concentrations in the embryonic liver.
Additionally, blue light reduced the tissue damage induced by heat stress in the embryonic liver.
Liver HSP70, HSP90, HSF1 and HSF3 gene expression in chicken liver were significantly
modulated by blue light, whereas the effects were different. Moreover, blue light significantly
modulated liver SOD and CAT antioxidant enzyme activity and their gene expression in
embryonic liver. However, blue light did not exert significant effects on body weight, late hatch
rectal temperature and tibia length of hatched chicks. Results suggested that monochromatic blue
light reduced the content of malondialdehyde and corticosterone in broilers under heat stress and
increased the relative expression of SOD and CAT genes. Moreover, the monochromatic blue
light may reduce the metabolic heat production of broilers during the embryonic stage, thus
reducing the damage of broilers due to heat stress during the embryonic heat acclimation stage.
Keywords: Thermal conditioning, thermotolerance, corticosterone, heat shock proteins,
antioxidant enzyme, liver histomorphology
1. Introduction
The effect of climatic variation has become a universal challenge in global poultry production
(Nawab et al., 2018). The economic loss caused by heat stress in the poultry industry alone is up
to $128 to $165 million per year, and the total annual economic loss of the United States livestock
industry is 16.9 to23.6 million dollars (Lara and Rostagno, 2013). Therefore, there is a crucial
need to develop various effective mitigation measures to reduce the detrimental effects of heat
stress in hot regions of the world. Heat treatment of broiler chickens during incubation by
exposing the embryo to heat stress has been shown to improve their ability to gain heat resistance
later in life (Yahav et al., 2004a; Yahav et al., 2004b). Broilers hatched after embryonic exposure
to intermittent TM of 39.5 ºC every 12h during days E7-E16 have lower calorific rates, larger
muscle growth, lower relative weight of abdominal fat pads and a thinner average fiber diameter
under heat stress (32 ºC for 12 h d-1) (Piestun et al., 2011). As well as the incubation period, from
the end of the TM till hatch, The metabolic rate and calorific value of TM treated embryos were
lower than those of the control (Piestun et al., 2015), and the main reason was lower plasma T3
concentration and the regulation of expression of DIO3 (Deiodinase, iodothyronine, type III) and
DIO2 (Deiodinase, iodothyronine, type II), respectively in muscles (Al-Zghoul et al., 2015).
However, heat exposure reduces animal welfare and growth performance (Loyau et al., 2015).
Additionally, heat stress triggers oxidative stress in liver, increases tissue damage (Abdo et al.,
2017). Elevated blood corticosterone concentration indicates that heat exposure increases the
stress response (Yahav et al., 2004b). Embryo growth and yolk consumption can be lowered by
embryo thermal treatment from E16 to E18 (Willemsen et al., 2010).
The severity of heat stress depends on the disturbance of the balance between the production
of oxygen reactive species (ROS) and the cellular antioxidant systems (Mahmoud and Edens,
2003; Halliwell and Whiteman, 2004; Feng et al., 2008; Akbarian et al., 2016). Cellular
∙ 10 ∙
antioxidant enzymes include superoxide dismutase (SOD) and catalase (CAT) which protect cells
from heat-stress-ROS-related effects. Heat stress conditions, species, and affected tissue will
impact the response of antioxidant enzyme (Akbarian et al., 2016). In addition, protein damage
and subsequent accumulation of unfolded proteins were caused by heat stress (McCormick et al.,
2003; Staib et al., 2007). The expression of chaperone proteins and shock proteins (HSPs) are
increased by heat stress, which leads to protein stability and heat resistance (Jaattela, 1999). The
HSPs consists of Hsp40, Hsp60, Hsp70, Hsp90, Hsp110, and the small HSPs. HSP70 and HSP90
work to protect the affected cell, preventing the aggregation of unfolding protein (Slimen et al.,
2016). Heat shock transcription factors (HSFs) are a family of DNA-binding proteins. The HSFs
can interact with a specific DNA sequence to modulate HSP expression (heat shock element
(HSE)) (Morimoto et al., 1997; Akerfelt et al., 2010; Gomez-Pastor et al., 2018).
Different mitigation methods have been adopted to reduce the destructive effects of heat
stress, including environmental strategies in shape of housing, ventilation and natural or artificial
shading, as well as feed additives in the diet and water (Nawab et al., 2018). Another approach
that can alleviate the heat stress deleterious effects is lighting management. Abdo et al. (2017)
described that susceptibility of broilers could be reduced by monochromatic blue light during heat
stress. Reducing the stress response and improving immune response in broilers has been proved
to be the effect of blue light (Zhang et al., 2014). In addition, blue light may alleviate stress
response by modulating expression of interleukin-1β (IL-1β) and ratio of heterophils to
lymphocytes (H/L) (Xie et al., 2008a; Mohamed et al., 2014). The objective of this present study
was to explore the effects of the monochromatic blue light on reducing the adverse impact of heat
stress of broilers in thermal manipulation by exposing embryos to periodic high temperature
during incubation development.
2.1. Incubation management
A total of 800 hatching eggs were selected by weight (60 ± 2 g) from Huangjiaoma broiler
breeders at a breeder company (Zhejiang Qunda Animal Husbandry Co., Jiaxing, China). The
eggs were incubated in separate environmentally controlled commercial incubators (NK-hatching,
Dezhou Nongke Incubation Equipment Co., Shandong, China) at the incubation room, Zhejiang
University, China. Air humidity of the incubator was kept at 60% during the experimental period.
The incubators were automatically maintained at 37.8 ± 0.1 ºC and 60 ± 1% relative humidity
(RH) during the incubation period. The eggs were turned from day 1 to day 18 while the time to
turn the eggs was 2h a day. The eggs were candled (Cool-Lite tester, GQF Manufacturing,
Savannah, Ga.) on day 18 during incubation, and nonviable eggs were removed.
2.2. Experimental design
Before incubation, the hatching eggs were disinfected (methanol plus potassium
permanganate fumigation) and divided into equal four groups (n=200), with four egg tray
replicates per group (n=50). Two groups (W+37.6 ºC and B+37.6 ºC) were incubated at normal
temperature with white and blue lighting, respectively. Other two groups (W+40 ºC and B+40 ºC)
were exposed to an experimental cyclic chronic heat stress. On this point, heat treatment was
carried out from the day14 day to the day18 in which eggs were exposed to 40 ± 1 ºC for 4 h a
day. After heat treatment the temperature was lowed to normal 37.6 ºC till hatch. The intensity of
blue light is 27 lux and the wavelength is 450 nm.
2.3. Sample collection
Twenty-four eggs from each group (six eggs from each replicate) were used for the sample
collection after heat acclimatization (after day 18 heat acclimatization). Cervical dislocation was
used to kill the embryos before Liver specimens of each bird were collected. 10% formalin was
used to preserve one third of liver specimens for histopathological examination; for RNA
extraction, one third of liver specimens were frozen in liquid nitrogen and then preserved at −80
∙ 11 ∙
ºC; for antioxidant enzymes activity analysis, one third of liver specimens were preserved at −20
ºC.
3. Results
3.1. Effect of blue light on MDA and corticosterone concentrations in embryonic liver
Figure 1(a) shows the MDA contents in chick embryos following heat stress. Light and
temperature effects on the level of MDA concentration were statistically significant (two-way
ANOVA, P<0.0001 for light; P<0.05 for temperature). Because MDA is an indicator of lipid
peroxidation, MDA concentration displayed a significant increase in the case of heat stress (W+40
ºC), and heat stress with blue light (B+40 ºC) compared to white (W+37.6 ºC) in chick embryos
(two-way ANOVA, for W+40 ºC P<0.001; and for B+40 ºC P<0.05). Interestingly, using blue
light and during heat stress (B+40 ºC) induced a lower concentration of MDA compared to white
light with heat stress (W+40 ºC) (two-way ANOVA, P<0.01).
Figure 1(b) shows the effect of light and temperature on corticosterone contents in the liver
following heat stress. Light effects on the level of corticosterone contents were statistically
significant (two-way ANOVA, P<0.0001 for light; P = 0.001 for temperature). A significant
increase of corticosterone contents was noticed in the case of heat stress (W+40 ºC) and heat stress
with blue light (B+40 ºC) compared to white (W+37.6 ºC) (two-way ANOVA, for W+40 ºC
P<0.0001; and for B+40 ºC P = 0.0001). Interestingly, using blue light during heat stress (B+40
ºC) induced a lower concentration of corticosterone compared to white light with heat stress
(W+40 ºC) (two-way ANOVA, P<0.01).
Figure 1. Blue light significantly lowers MDA and corticosterone concentration and increases
antioxidant enzyme activities in the liver of chick embryos following heat stress: (a) MDA
content. (b) corticosterone concentration. (c) SOD enzyme activity. (d) CAT enzyme activity.
Mean ± SEM is shown. *, **, *** and **** denote statistical significance (two-way ANOVA)
with P<0.05, P<0.01, P<0.001, P<0.0001, respectively.
3.2. Effect of blue light on the increase in liver interstitial cracks caused by heat stress
Figures 2(a) and 2(c) show that the cells of embryonic liver were closely arranged and
∙ 12 ∙
structurally intact in case of white light (W+37.6 ºC) and blue light alone (B+37.6 ºC). Figure 2(b)
shows that there were more cracks and loose arrangement in embryonic liver tissue in case of heat
stress (W+40 ºC), moreover, severe mononuclear cell infiltration and focal necrosis were found,
which was consistent with the discovery of Abdo et al. (2017) that heat stress at high temperature
may cause moderate to severe fatty changes and perivascular mononuclear cell infiltration and
focal necrosis. Figure 2(d) shows that there were fewer cracks and the liver tissue structure was
more compact in case of heat stress with blue light (B+40 ºC) compared to heat stress (W+40 ºC).
Figure 2. Liver histomorphology of Huangjiaoma chicken: (a) White light (W+37.6 ºC). (b)
White light during heat stress (W+40 ºC). (c) Blue light (B+37.6 ºC). (d) Blue light during heat
stress (B+40 ºC). Arrows point the interstitial cracks.
3.3. Effect of blue light on liver HSP70, HSP90, HSF1, and HSF3 gene expression in chicken
liver.
Figure 3 presents the relative gene expression profiles of HSPs (HSP70 and HSP90) as well
as HSF3 and HSF1 from 8 birds for each treatment compared to normal white light (W+37.6 ºC).
The analysis was performed in the liver of chick embryos following heat stress using qPCR. Blue
light modulated the expression level of HSP70 and HSP90 in chick embryos without temperature
differences (two-way ANOVA, P<0.0001 for blue light effect in the case of HSP70 and HSP90).
In the liver, normally, blue light (B+37.6 ºC) did not affect much the HSP70 gene expression
where only a slight increase of HSP70 gene expression were detected. However, heat stress caused
significant upregulation of HSP70 gene expression when used with white light (W+40 ºC) and
blue light (B+40 ºC) (two-way ANOVA, for W+40 ºC P<0.0001, P<0.0001 for B+40 ºC).
Nevertheless, when blue light was used during heat stress (B+40 ºC), compared to in the case of
heat stress (W+40 ºC), an interesting significant downregulation of HSP70 gene expression was
detected P =0.0094). For HSP90 gene (Figure 3), it showed a similar expression pattern to HSP70
gene. There was no temperature difference and only differences due to light treatment were
detected without light + temperature interaction (two-way ANOVA, P>0.05 for strain and
interaction; P<0.0001 for light treatment). In the liver, B+37 ºC revealed a nonsignificant
upregulation of HSP90 gene expression level. Additionally, white light and blue light during heat
∙ 13 ∙
stress stimulated more HSP90 gene expression compared to that under normal condition (W+37.6
ºC) (P<0.0001; P =0.0003, resp.). However, when blue light was used during heat stress (B+40
ºC), compared to heat stress only (W+40 ºC), which displayed a nonsignificant downregulation
of HSP70 gene expression level (P>0.05). These results demonstrate that blue light induces more
HSP70 and HSP90 gene expression. Most of the increases in HSP70 and HSP90 expression
occurred by using blue light during heat stress. Interestingly, using blue light and during heat
stress (B+40 ºC) induced a significant downregulation of HSP70 gene expression compared to
white light with heat stress (W+40 ºC).
In conclusion, using blue light during heat stress regulated the expression levels of HSP70,
HSP90, HSF1, and HSF3. A similar increase in the level of expression of HSPs (HSP70 and
HSP90) was found. However, in the case of HSF1 and HSF3, this resulted in an upregulation
expression level of HSF1 but a slight increase in the case of HSF3.
****
2.0 ****
** *** B
**
Fold change (log2)
1.5 **** H+W

****
1.0 H+B
* *
0.5 * **
0.0
-0.5
-1.0
D
T
70
90
3
SF
SF
A
SO
SP
SP
C
H
H
H
Figure 3. Liver HSP70, HSP90, HSF1, and HSF3 gene expression are significantly modulated
by blue light. The expression levels were presented as log2 fold change and shown in the figure
as mean ± SEM. *, **, *** and **** denote statistical significance (two-way ANOVA) with
P<0.05, P<0.01, P<0.001, P<0.0001, respectively.
3.4. Effect of blue light on liver SOD and CAT antioxidant enzyme activity and their gene
expression in embryonic liver
Figures 1(c) and 1(d) show the effect of temperature, light, and temperature + light interaction
on SOD and CAT enzyme activities in the liver after heat stress, respectively. Light effect on the
level of SOD activity was statistically significant (two-way ANOVA, P<0.0001 for light
treatment). SOD activity displayed a significant increase in the case of heat stress (W+40 ºC), and
heat stress with blue light (B+40 ºC) compared to white light (W+37.6 ºC) in chicks (two-way
ANOVA, for W+40 ºC P<0.01; and for B+40 ºC P<0.01). However, in the case of using blue light
alone (B+37.6 ºC), it induced a slight downregulation (P =0.1168). Besides, CAT enzyme showed
a similar response to SOD enzyme activity following heat stress. Light effect was statistically
significant (two-way ANOVA, P<0.0001). A significant increase of CAT enzyme activity was
noticed in the case of heat stress (W+40 ºC) and using blue light during heat stress (B+40 ºC)
compared to white (W+37.6 ºC) (two-way ANOVA, p =0.0007 for W+40 ºC and for B+40 ºC p
=0.0004).
To address how it modulates their gene expression levels, we measured SOD and CAT mRNA
expression levels. Figure 3 shows the temperature, light treatment, and temperature + light
interaction effects on SOD and CAT genes. For SOD, the light treatment caused significant
differences, while in the case of CAT, there were significant differences due to temperature effect
(two-way ANOVA, in the case of SOD for temperature and light + temperature interaction
P>0.05; for light treatment P =0.0112; in the case of CAT, temperature P =0.0214; light treatment
∙ 14 ∙
and strain + light interaction P>0.05). The SOD gene expression level was upregulated when only
blue light was used (B+37.6 ºC). This upregulation was significant when compared to heat stress
(W+40 ºC) which stimulated a nonsignificant downregulation (P =0.0543). Additionally, SOD
gene expression level was downregulated when blue light was used during heat stress (B+40 ºC).
The CAT gene reacted differently to SOD gene. Blue light resulted in a slight or significant
upregulation either normally (B+37.6 ºC) or when it was applied during heat stress (P =0.1298;
P<0.0239, resp.) compared to heat stress (H+W), which was characterized by a slight
downregulation. In conclusion, blue light regulated SOD and CAT gene expression levels and
their antioxidant enzyme activity during heat stress.
4. Discussion
Meat type chickens have limited ability to face the destructive impact of heat stress. Several
studies addressed heat challenge during embryogenesis enhance adaptive capacity of broilers to
cope with heat exposure. However, the heat challenge during embryonic development may trigger
oxidative stress, stress response and reduce growth performance. This study aims to investigate
monochromatic blue light exposure on alleviating the negative impact of heat stress during
chicken embryogenesis.
Figure 1 shows the effect of blue light on MDA and Corticosterone content following heat
stress. In this regard, for MDA (Figure 1(a)), exposure of eggs to blue light at the same time as
heat stress (B+40 ºC) significantly depressed its content compared to heat stress (W+40 ºC) which
demonstrated a significant enhance in MDA content. Higher MDA content is an indicator of lipid
peroxidation and therefore more oxidative damage (Ismail et al., 2013). The lowered MDA
concentration in the case of B+40C suggests a possible role of blue light in lowering the negative
effect of heat stress (Ke et al., 2011a). These findings are consistent with this of Abdo et al. (2017).
Similarly, the results with Corticosterone (Figure 1(b)) also showed that exposure of eggs to blue
light while heat stress (B+40 ºC) significantly lowered its content compared to heat stress (W+40
ºC) which demonstrated a significant enhance in Corticosterone content. The activity of the
neuroendocrine system of poultry could be altered by heat stress, which contribute to activation
of the hypothalamic–pituitary–adrenal axis. This disturbance causes elevation of corticosterone
levels and degradation of the health status of animals (Quinteiro-Filho et al., 2009; Costa-Pinto
and Palermo-Neto, 2010). Elevated corticosterone levels of meat type chickens have demonstrated
depressed lymphocyte number culminating in a higher H:L ratio (Siegel, 1995). The content of
corticosterone is a sensitive indicator of stress in broilers and it is commonly used as a tool to
assess the physiological stress (Selvam et al., 2018). The results of the present study indicated that
exposure of eggs to heat stress significantly enhanced MDA and corticosterone content in
embryonic liver, whereas supplementation of blue light in the environment was shown to restore
the MDA and corticosterone level toward normal.
Assessing those effects of exposure to monochromatic blue light on the histology of the liver
tissue were further examined in Figure 2. Heat stress (W+40 ºC) induced tissue injury, increasing
liver interstitial cracks. Incubation with blue light during heat stress (B+40 ºC) contributed to the
loss of tissue damage effect by heat stress compared to those incubation in white light. The blue
light could enhance broiler chickens’ resistance (Xie et al., 2008b).
In the present study, under normal ambient temperature, compared with white light group
(W+37.6 ºC), blue light (B+37.6 ºC) can significantly increase the expression of HSP70 and HSF3
genes, and these proteins acting as molecular chaperones to prevent protein aggregation can
improve cell heat tolerance to resist high temperature, thus regulating the balance of cell survival
and death. It is worth noting that under the condition of heat stress, compared with the white light
group (W+40 ºC), the gene expression of HSP70, Hsp90 and HSF3 in the blue light group (B+40
ºC) decreased slightly, which may be due to the fact that the way of improving the heat resistance
of Broilers by blue light is not single, and the need for heat shock protein in the embryo was
reduced due to the improvement of heat stress resistance under the monochromatic blue light.
∙ 15 ∙
Figure 1 shows that the activities of SOD and CAT in heat stress (40 ºC) group were
significantly higher than those in normal group (37.6 ºC), but blue light had no significant effect
on the activities of the two enzymes, which indicated that monochromatic blue light had no
regulatory effect on the activities of antioxidant enzymes. However, according to Figure 3,
monochromatic blue light could increase the relative expression of SOD and CAT genes. Under
normal temperature, the gene expression of SOD and CAT was increased by monochromatic blue
light. Under high temperature, the gene expressions of the two enzymes were significantly
decreased, but under this condition, compared with white light (W+40 ºC), monochromatic blue
light (B+40 ºC) also increased the gene expression of the two enzymes. This indicated that
monochromatic blue light can improve the heat tolerance of broilers by regulating the gene
expression of antioxidant enzymes rather than enzyme activity.
According to results, the weight of the body exposed to monochromatic blue light (B+37.6
ºC) was lower than that of broilers exposed to white light (W+37.6 ºC). Similarly, in the high
temperature group, the weight of the body exposed to monochromatic blue light (B+40 ºC) was
lower than that of broilers exposed to white light (W+40 ºC). In addition, at the same ambient
temperature, the weight percentage of egg yolk in blue light group was higher than that in white
light group, which indicated that monochromatic blue light might reduce the heat production of
broiler embryos by reducing the metabolic rate of broiler embryos, because egg yolk was the main
energy source of hatching embryos. However, after hatching, there was no difference in body
weight between blue light group and white light group, and there was no difference in tibia length
5. Conclusions
The experimental results suggested that monochromatic blue light can reduce the content of
MDA and corticosterone in broilers under heat stress and increase the relative expression of SOD
and CAT genes. Moreover, the monochromatic blue light may reduce the metabolic heat
production of broilers during the embryonic stage, thus reducing the damage of broilers due to
heat stress during the embryonic heat acclimation stage.
References
Abdo, S.E., S. El-Kassas, A.F. El-Nahas, S. Mahmoud, 2017. Modulatory Effect of Monochromatic Blue
Light on Heat Stress Response in Commercial Broilers. Oxidative Medicine and Cellular Longevity.
Akbarian, A., J. Michiels, J. Degroote, M. Majdeddin, A. Golian, S. De Smet, 2016. Association between
heat stress and oxidative stress in poultry; mitochondrial dysfunction and dietary interventions with
phytochemicals. Journal of Animal Science and Biotechnology. 7 (1), 37.
Akerfelt, M., R.I. Morimoto, L. Sistonen, 2010. Heat shock factors: integrators of cell stress, development
and lifespan. Nature Reviews Molecular Cell Biology. 11 (8), 545–555.
Al-Zghoul, M.B., S.M. El-Bahr, R.K. Al-Rukibat, A.S. Dalab, T.A. Althnaian, S.Y. Al-Ramadan, 2015.
Biochemical and molecular investigation of thermal manipulation protocols during broiler embryogenesis and
subsequent thermal challenge. BMC Veterinary Research. 11.
Costa-Pinto, F.A., J. Palermo-Neto, 2010. Neuroimmune interactions in stress. Neuroimmunomodulation.
17 (3), 196–199.
Feng, J., M. Zhang, S. Zheng, P. Xie, A. Ma, 2008. Effects of high temperature on multiple parameters
of broilers in vitro and in vivo. Poultry Science. 87 (10), 2133–2139.
Gomez-Pastor, R., E.T. Burchfiel, D.J. Thiele, 2018. Regulation of heat shock transcription factors and
their roles in physiology and disease. Nature Reviews Molecular Cell Biology. 19 (1), 4–19.
Halliwell, B., M. Whiteman, 2004. Measuring reactive species and oxidative damage in vivo and in cell
culture: how should you do it and what do the results mean? British Journal of Pharmacology. 142 (2), 231–
255.
Ismail, I., K. Al-Busadah, S. El-Bahr, 2013. Oxidative stress biomarkers and biochemical profile in
broilers chicken fed zinc bacitracin and ascorbic acid under hot climate. Am J Biochem Mol Biol. 3, 202–214.
Jaattela, M., 1999. Heat shock proteins as cellular lifeguards. Annals of Medicine. 31 (4), 261–271.
Ke, Y., W. Liu, Z. Wang, Y. Chen, 2011a. Effects of monochromatic light on quality properties and
antioxidation of meat in broilers. Poultry Science. 90 (11), 2632–2637.
Ke, Y.Y., W.J. Liu, Z.X. Wang, Y.X. Chen, 2011b. Effects of monochromatic light on quality properties
and antioxidation of meat in broilers. Poultry Science. 90 (11), 2632–2637.
∙ 16 ∙
Lara, L.J., M.H. Rostagno, 2013. Impact of heat stress on poultry production. Animals. 3 (2), 356–369.
Loyau, T., L. Bedrani, C. Berri, S. Metayer-Coustard, C. Praud, V. Coustham, S. Mignon-Grasteau, M.J.
Duclos, S. Tesseraud, N. Rideau, C. Hennequet-Antier, N. Everaert, S. Yahav, A. Collin, 2015. Cyclic
variations in incubation conditions induce adaptive responses to later heat exposure in chickens: a review.
Animal. 9 (1), 76–85.
Mahmoud, K.Z., F. Edens, 2003. Influence of selenium sources on age-related and mild heat stress-related
changes of blood and liver glutathione redox cycle in broiler chickens (Gallus domesticus). Comparative
Biochemistry and Physiology Part B: Biochemistry and Molecular Biology. 136 (4), 921–934.
McCormick, P., G. Chen, S. Tlerney, C. Kelly, D. Bouchier-Hayes, 2003. Clinically relevant thermal
preconditioning attenuates ischemia-reperfusion injury. Journal of Surgical Research. 109 (1), 24–30.
Mohamed, R., M. Eltholth, N. El-Saidy, 2014. Rearing broiler chickens under monochromatic blue light
improve performance and reduce fear and stress during pre-slaughter handling and transportation.
Biotechnology in Animal Husbandry. 30 (3), 457–471.
Morimoto, R.I., M.P. Kline, D.N. Bimston, J.J. Cotto, 1997. The heat-shock response: regulation and
function of heat-shock proteins and molecular chaperones. Essays in Biochemistry, Vol 32, 1997. 32, 17–29.
Nawab, A., F. Ibtisham, G.H. Li, B. Kieser, J. Wu, W.C. Liu, Y. Zhao, Y. Nawab, K.Q. Li, M. Xiao, L.L.
An, 2018. Heat stress in poultry production: Mitigation strategies to overcome the future challenges facing
the global poultry industry. Journal of Thermal Biology. 78, 131–139.
Piestun, Y., O. Halevy, D. Shinder, M. Ruzal, S. Druyan, S. Yahav, 2011. Thermal manipulations during
broiler embryogenesis improves post-hatch performance under hot conditions. Journal of Thermal Biology.
36 (7), 469–474.
Piestun, Y., I. Zimmerman, S. Yahav, 2015. Thermal manipulations of turkey embryos: The effect on
thermoregulation and development during embryogenesis. Poultry Science. 94 (2), 273–280.
Quinteiro-Filho, W., D.A. Righi, J. Palermo-Neto, 2009. Effect of cyhalothrin on Ehrlich tumor growth
and macrophage activity in mice. Brazilian Journal of Medical and Biological Research. 42 (10), 912–917.
Selvam, R., S. Suresh, M. Saravanakumar, C. Chandrasekaran, D.S. Prashanth, 2018. Alleviation of heat
stress by a polyherbal formulation, Phytocee™: impact on zootechnical parameters, cloacal temperature, and
stress markers. Pharmacognosy Research. 10 (1), 1.
Siegel, H., 1995. Stress, strains and resistance. British Poultry Science. 36, 003-22.
Slimen, I.B., T. Najar, A. Ghram, M. Abdrrabba, 2016. Heat stress effects on livestock: molecular, cellular
and metabolic aspects, a review. Journal of Animal Physiology and Animal Nutrition. 100 (3), 401–412.
Staib, J.L., J.C. Quindry, J.P. French, D.S. Criswell, S.K. Powers, 2007. Increased temperature, not
cardiac load, activates heat shock transcription factor 1 and heat shock protein 72 expression in the heart.
American Journal of Physiology-Regulatory, Integrative and Comparative Physiology. 292 (1), R432–R439.
Willemsen, H., B. Kamers, F. Dahlke, H. Han, Z. Song, Z. Ansari Pirsaraei, K. Tona, E. Decuypere, N.
Everaert, 2010. High-and low-temperature manipulation during late incubation: effects on embryonic
development, the hatching process, and metabolism in broilers. Poultry Science. 89 (12), 2678–2690.
Xie, D., Z. Wang, Y. Dong, J. Cao, J. Wang, J. Chen, Y. Chen, 2008a. Effects of monochromatic light on
immune response of broilers. Poultry Science. 87 (8), 1535–1539.
Xie, D., Z.X. Wang, Y.L. Dong, J. Cao, J.F. Wang, J.L. Chen, Y.X. Chen, 2008b. Effects of
monochromatic light on immune response of broilers. Poultry Science. 87 (8), 1535–1539.
Yahav, S., A. Collin, D. Shinder, M. Picard, 2004a. Thermal manipulations during broiler chick
embryogenesis: Effects of timing and temperature. Poultry Science. 83 (12), 1959–1963.
Yahav, S., R.S. Rath, D. Shinder, 2004b. The effect of thermal manipulations during embryogenesis of
broiler chicks (Gallus domesticus) on hatchability, body weight and thermoregulation after hatch. Journal of
Thermal Biology. 29 (4–5), 245–250.
Zhang, Z., J. Cao, Z. Wang, Y. Dong, Y. Chen, 2014. Effect of a combination of green and blue
monochromatic light on broiler immune response. Journal of Photochemistry and Photobiology B: Biology.
138, 118–123.
∙ 17 ∙
Light Demand Characteristics, Production Performance and

Changes of Feeding Pattern of Broilers
Shouyi Wang, Jinming Pan *
* Corresponding author. Email: panhouse@zju.edu.cn
Abstract
Poultry is a light-sensitive animal, and the light environment has an important influence on
the growth and development of poultry. Previous studies mainly focused on the effects of light
environment on various physiological indicators of poultry but have seldom explored the light
demand characteristics of broilers under active selection. This experiment mainly studies the light
demand characteristics of broilers under the yellow LED light environment, and the influence of
different breeding densities on the production performance and diet characteristics of broilers.
Studies have shown that the production performance indexes of low-density (2.5 chickens m-2)
are higher than those of the high-density group (7.5 chickens m-2). The feed and water
consumption in the bright area of the two experimental groups were significantly higher than those
in the dark area, which means the broilers showed a great preference for the bright area. However,
as the age of the broilers increased, the food and water consumption of the broilers had a decrease,
indicating that broilers had a lower preference for light in the middle and late stages of growth.
The statistical results of residence frequency distribution characteristics show that broiler chickens
have different light requirements at different growth stages under the condition of active selection:
1) Low-density breeding environment: young chickens 23.8L:0.2D, adult chickens 22.3L:1.7D;
2) High-density breeding environment: young chickens 22.6L:1.4D, adult chickens 15.0L:9.0D.
Keywords: Broiler, light-demand characteristics, production performance, feeding pattern.
1. Introduction
Poultry is a light-sensitive animal with a highly developed visual system and light stimulation
has an important impact on the growth, behavior and rhythm of poultry (Lewis and Morris, 2000;
Tang and Ma, 2004). Therefore, controllable artificial light sources have become the main source
of breeding light. Compared with traditional incandescent lamps and fluorescent lamps, light-
emitting diode lights (LEDs) have very different spectral characteristics. LEDs have a narrow
spectrum and can form a special spectrum by concentrating energy in a specific nano-wavelength
band. Due to the acute vision of poultry, the optimized lighting environment according to the light
demand characteristics of poultry can effectively improve the production performance of poultry
(Rozenboim et al., 2004; Pan et al., 2019).
The light environment includes three factors: light intensity, light color, and photoperiod.
Changes in various factors of the light environment will have different effects on the growth of
poultry. Different light intensity has no significant effect on the performance of broiler chickens,
but low light intensity often affects the welfare of poultry (Alvino et al., 2009a; Alvino et al.,
2009b; Blatchford et al., 2009; Deep et al., 2012). High light intensity will also cause behaviors
such as pecking in the breeding process, which will cause certain negative effects on poultry
production. The light color is mainly determined by the wavelength and relevant studies have
found that blue and green light has a significant advantage in promoting the performance of broiler
chickens (Ke et al., 2011). The photoperiod includes two aspects: the duration of light and the
light system. A recent study showed that night length can seriously affect the sleep of starlings
(van Hasselt et al., 2020) and previous experiments have proved that changes in the photoperiod
will affect the feed conversion rate of poultry and the incidence of chicken diseases like leg
disease, and continuous light will inhibit the growth of broilers to a certain extent (Mahmud et al.,
2011; Fidan et al., 2017). On the whole, previous studies mainly focused on the effects of light
∙ 18 ∙
environment on various physiological indexes of poultry and did not take into account the light
demand characteristics of poultry themselves.
This research chose the yellow LED, which is commonly used in actual production, as the
light source and maintained the ambient temperature at the room temperature (26℃). In this
research, we investigated the light demand characteristics of different growth stages and the
influence of different breeding densities on broiler weight and diet characteristics, and we hope
to provide references for further application to the broiler breeding lighting.
All procedures in this experiment were approved by the Laboratory Animal Center, Zhejiang
University (Hangzhou, Zhejiang, China).
2.1. Animals and Experimental Treatments
Sixty 1-day-old yellow-feathered chicks (Meihuang, Zhejiang Guangda Breeding Poultry Co.,
Ltd., Zhejiang, China) were used for the experiment. The chicks were raised intensively to adapt
to the environment in the first two days of the brooding period and then separately in three
replicate experimental groups. Each experimental group included two independent chicken cages
[2 m × 1 m × 1 m, length (L) × width (W) × height (H)] raised 5 and 15 chicks, respectively [low-
density: 2.5 chicks m-2 and high-density: 7.5 chicks m-2] (Figure 1). The cages were equally
divided into two connected areas, namely the LA (light area) and the DA (dark area), where the
broilers could choose to stay freely. Cameras were installed on the top of the LA to record the
activities of broilers in real time. The manure receiving boards were placed 0.5 m below the
chicken coop and cleaned once a day. Feeding and drinking devices of the cage were placed at
specific positions respectively where the chicks can eat and drink freely and the feed and drinking
water were changed every day to prevent feed spoilage or odors production which may affect the
experimental results. During the experiment, the Ta (ambient temperature) was controlled at 26–
28 ℃ and RH (relative humidity) was controlled at 56%–70%. The vaccination procedures for
broiler chickens were strictly in accordance with the vaccine standards for commercial broiler
chickens of Zhejiang Qinqin Breeding Co., Ltd (Zhejiang, China), and the epidemic prevention
work was performed by the company's professional technicians. The feed for different growth
stages of broilers is provided by the company, and the main ingredients are corn and soybean
meal.
Figure 1. Structure diagram and real shots of breeding area.

2.2. Lighting environment
Straight yellow LED tubes were selected as the light source in the experiment. The LED tubes
and light adjustment controllers used are provided by Hangzhou LightTalk Biotechnology Co.,
Ltd (Zhejiang, China). The main wavelength range of the LED tubes is 587.9–589.1 nm, and the
tube power is 4.32 W. Each LED tube contains 108 lamp beads, and the length of the tube is 60
cm. In the formal experiment, two tubes were installed at the top of the LA and the lighting system
∙ 19 ∙
of the LA is 24L:0D (24 h light: 0 h dark) while DA was covered with shading cloth. The light
intensity setting value of LA was 15 lx which was commonly used in the growing period in actual
production. We measured the five points in the trapezoidal area with illuminance meter (SMART
SENSOR AR823+; Hong Kong SMART SENSOR Co., Ltd, China) and get the average value.
We cleaned the surface of the lamp every week to reduce the influence of dust on the light
intensity, thereby reducing the influence of light intensity on the experimental results. The
windows and doors of the breeding area are sequentially covered by high white plastic nets and
silver shading cloth to avoid the influence of external light on the breeding area.
2.3. Production performance and behavior observation
Consumption of feed and water (remaining amount, added amount) in LA and DA was
measured daily to calculate the chick preference for eating behavior. To reduce the influence of
too frequent manual weighing on the accuracy of the experiment, each chick was weighed once a
week (4, 11, 18, 25, 32, 39, 46, 53 days old). After each weighing, the chicks were placed in the
area opposite to the last time, and the death of the chickens was recorded daily. At the end of the
experiment, all broiler chickens were killed and various slaughter indicators were measured.
Behavior observations were carried out on the day before weighing to avoid the influence of
manual weighing on the behavior of broilers. The video recording of the chick behavior is
completed by a CCD camera (DH-CA-D481/3; Hangzhou Dahua Technology Co., Ltd., China),
and the shooting range could cover the entire breeding surface of the LA. We took screenshots of
the videos at 5 min intervals and counted the number of chicks both in the LA and DA. It is proved
that the 15 min interval can cover 95% of the change interval in the total result (Kristensen et al.,
2007). When counting the numbers, the basis for judging the individual chick in a certain area is
the eyeballs location. If most of the body of the broiler is in the bright area but its eyeballs are in
the DA, the broiler is still counted as the DA when counting. To ensure the accuracy of the
experimental results, the videos at certain special moments are not counted, including operations
such as artificial addition of feed and water or dust removal.
2.4. Statistical Analysis
During the experiment, some production performance indicators including body weight,
relative growth rate (RGR), feed consumption, water consumption, feed conversion ratio (FCR),
and mortality rate were used. Relative growth rate is the percentage of poultry's weight gain in a
certain period. The formulas are as follows:
Wt1  Wt 0
RGR  (1)
Wt 0
where RGR represents relative growth rate (%), Wt0 represents the weight of broiler at the age of t0
(g), Wt1 represents the weight of broiler at the age of t1 (g).
Feed conversion ratio is an important index to characterize production performance and feed
to weight ratio is a kind of feed conversion rate. The formulas are as follows:
d
E i
FCRW  i 1
(2)
Wt1  Wt 0
where FCRW represents feed to weight ratio (%), Ei represents feed intake on day i, Wt0 represents
the weight of broiler at the age of t0 (g), Wt1 represents the weight of broiler at the age of t1 (g).
The percentage of staying frequency directly reflected the broilers' preference for a certain
area. The formulas are as follows:
Nl
Fsl   100% (3)
N
∙ 20 ∙
Fsd  1  Fsl (4)

where Fsl represents percentage of staying in the LA frequency (%), Nsl represents the number of
broilers staying in the LA at a certain moment, Fsd represents percentage of staying in the DA
frequency (%), Nsd represents the number of broilers staying in the DA at a certain moment, N
represents the number of all broilers. All data such as body weight, feed consumption, water
consumption and regional residence time are analyzed by Office Excel 2019.
3. Results
3.1. Production performance
In this experiment, different breeding densities had significant effects on broiler body weight,
weight gain rate and other production performance indicators, but had no significant effect on
feed-to-weight ratio (Table 1). The final body weight, weight gain rate and feed consumption of
low-density broilers were significantly higher than those of high-density broilers. Although there
was no significant difference in the body weight of broiler chickens between the two groups in
the early period of the experiment (4, 11, and 18 days of age), the final body weight of the broilers
in the low-density group was significantly higher than that in the high-density group from the
perspective of the later period and the whole period of the experiment (Figure 2).
Table 1. Effects of different rearing densities on production performances of broilers.
Index
Group 4-day-old 52-day-old Relative growth Feed Feed-to-weight
-1 1 1
weight g weight g- rate consumption g- ratio
Low-density 71.2±1.2 1766.3±156.3 2379.82±2.0 3914.2±552.1 2.3
High-density 68.9±1.3 1546.1±13.5 2145.27±0.6 3508.2±161.0 2.4
Note: The values in the Table are expressed as: Mean ± SD.
Figure 2. The average weight of broilers in different rearing densities.

3.2. Feeding and drinking characteristics
The preference of eating and drinking was measured by the value of feed and water
consumption of an average single broiler per day calculated at intervals of 7 days. The daily feed
intake and drinking water preference of broilers reflect the time the broilers stayed in the LA and
DA to a certain extent, so the law of light demand for broilers can be detected.
The feed and water consumption of all test groups in the LA are much larger than those in the
DA (Figure 3) and broilers with different densities have certain similarities. In the early stage of
broiler growth, it increased with the increase in age, and then began to decrease in the later stage
of growth. However, from a numerical point of view, the LA feed and water consumption of the
low-density group are much higher than that of the high-density group.
For low-density groups, the inflection points of feed and water consumption in the LA all
∙ 21 ∙
appear in the 5-week old, before when the consumption had been increasing with the increase of
the age of the broilers, and the consumption decreases in the later period. From the perspective of
the entire experiment cycle, the consumption in the DA maintained at a very low level, but still
improved to a certain extent in the later stage of growth.
For high-density groups, the inflection points of feed and water consumption in the LA appear
in the 6-week old, but the feed consumption declined and then increased in the third week. We
also find that the feed consumption in the LA increased slowly and decreased significantly at 7
weeks of age. Before the 4 weeks of age, the water consumption of broilers in the LA increased
slightly. Since then, the water consumption of broilers in the LA has greatly increased. The feed
and water consumption of broilers in the DA is low, but it is still increasing to a certain extent in
the late growth period.
Figure 3. The average daily feed and water consumption in different rearing densities per week.
Low-density-LA refers to the light area of the low-density experimental group, and so on.
Comparing the groups with different densities, the LA feed and water consumption of low-
density group are significantly higher than the high-density group, indicating that even in the same
light area, broilers in the low-density group have a stronger preference for the light area than the
high-density group.
3.3. Staying frequency distribution characteristics
We processed the videos of the broiler's activities in each area and obtained the time of the
broiler's stay in the LA and the DA. The residence frequency distribution of broilers in each zone
can more intuitively reflect the change in light preference of broilers during the whole growth
process than feed and water consumption, and we finally obtained the light demand of broilers
during the whole growth stage.
All groups of broilers have a significant difference in the staying frequency in the LA and the
DA (Figure 4). Starting from the 4 days of age, the broiler's staying frequency in the LA has
always occupied an absolute advantage and this advantage gradually weakens with the increase
of day age. For low-density broilers, the staying frequency in the LA occupies an absolute
advantage during the whole process. Before 25 days of age, broilers of low-density stayed in the
LA with a frequency of almost 100%, and then continued to decrease as the age increased,
eventually reaching the lowest level of 78%. The DA has the opposite law to the LA. Staying
frequency of broilers in the high-density group has similar laws, decreasing with the increase of
the age. Until the age of 53 days, the staying frequency of DA exceeded that of the LA. These
∙ 22 ∙
show that broiler preference for DA gradually increased, and the increasing is slower in the early
stage and faster in the middle and late stages.
Figure 4. The distributions of broilers in different rearing densities.

In addition, statistics on the percentage of broilers' staying frequency throughout the entire
process showed that the average staying percentage of broilers in the LA of the low-density group
was 94.82%, which was significantly higher than that of the high-density group of 79.51%. When
the light and temperature are the same, broilers under different breeding densities also have a
certain degree of difference in light demand, and the low-density group has a higher light demand.
We took two typical day-old (24-day-old and 45-day-old) broilers as the representatives of
the early and late growth stages and counted the changes in the staying frequency of the broilers
at different densities throughout the day (Figure 5). The daily occurrence frequency of broilers in
LA was divided into two time periods: the daytime (DT, 06:00–17:00) and the nighttime (NT,
18:00–5:00). The demarcation line of day and night was determined by referring to the average
sunrise time and sunset time of Hangzhou during the reference experiment, and the stay time of
the broiler in the day and night bright area was used for the calculation of the lighting system.
Figure 5. The daily presence frequency in the LA at two ages (25 d and 46 d).
Note: Low-24d-LA represents the light area of the low-density experimental group at the age of
25 days; others can be deduced by analogy, and the water and feed were updated daily at 9:00 a.m.
At 24-day and 45-day-old, the staying frequency was evenly distributed throughout the day.
However, 45-day-old broilers generally had a lower frequency in the LA than 24-day-old,
indicating that young broilers have more light demand. The frequency of high-density broilers in
the LA showed a certain regularity, that is, the activity pattern of broilers changed periodically
with the outside day and night changes. During the day, the frequency of broilers in the LA showed
a certain downward trend over time and the frequency of broilers in the DA showed a certain
upward trend over time at night. However, this pattern is not obvious and further investigation is
needed. In addition, the effect of broiler age on the staying time in the LA can be clearly observed
from the figure. The young broilers stayed in the LA longer than the grown.
After calculation, the light system under the active choice of broilers can be obtained. For
low-density groups, during the day and night, the 24-day-old broilers stayed in the LA for 11.9
and 11.9 hours respectively and the 46-day-old broilers stayed for 10.9 and 11.4 hours
respectively. For high-density groups, during the day and night, the 24-day-old broiler stayed in
the LA for 11.2 hours and 11.4 hours respectively and the 45-day-old broiler stayed for 7.0 hours
∙ 23 ∙
and 8.0 hours respectively. We calculated further and obtained two light systems at different
growth stages under the active selection of broilers: 1) Low-density breeding environment: young
broilers 23.8L:0.2D, adult broilers 22.3L:1.7D; 2) high-density breeding environment: young
broilers 22.6L:1.4D, adult broilers 15.0L:9D. As an experimental result, the lighting system has a
certain reference value for the lighting system setting of the chicken breed in actual production,
but it needs further exploration if it is to be applied in practice.
4. Discussion
There was no significant difference in the body weight of broiler chickens between the two
groups in the early stage of the experiment, but significant from the later stage of the experiment
and the whole process. This may be due to the increase in volume in the later stage of growth,
which squeezed the growth space, resulting in limited feeding and drinking water, slow growth of
broilers, and poor group uniformity. It is generally believed that an appropriate increase in the
stocking density can increase the slaughter rate of broilers, the weight of the glandular stomach
and the weight of abdominal fat (Yu et al., 2006). However, this experiment did not reach a
conclusion similar to the above results. The possible reason is that the breeding density used in
this experiment is generally small and the maximum breeding density is 7.5 broilers m-2, which is
not enough to make differences. Also, the light environment was different, for that the poultry
grew in the non-uniform LED light environment and could freely shuttle between the LA and the
DA in this experiment. Light duration and light cycle were determined by the broilers' independent
choice. Therefore, this conclusion needs further verification.
In the observation of the typical day-old broiler's single-day stay behavior, we found that the
broiler behavior did not have a significant correlation with the external time rhythm, which is
inconsistent with some studies. In previous study, Senaratna et al. (Senaratna et al., 2012) found
that under the combined effects of time phase (morning, evening, night), light color, and age, the
frequency of many behaviors (sleeping, lying down, feeding behavior) of broiler chickens is
significantly different, and other studies have similar conclusions (Kristensen et al., 2007). The
possible reason is that the experiment started from the broiler brooding period, which had already
destroyed the internal rhythm of the broilers themselves. The broilers formed a behavior pattern
based on their own comfort in a relatively constant light environment, which is the result of the
broilers' active choice and a response to many factors in the external environment.
At present, several commonly used lighting systems (12L:12D, 16L:8D, 20L:4D)
(Olanrewaju et al., 2006), 16L:8D are the most commonly used as the intermediate values
(Renden et al., 1996). This light regime has been shown to improve poultry welfare, reduce animal
psychological stress and enhance immune effect (Rozenboim et al., 1999). In our experiment,
broilers could choose their own light cycle according to their needs. The choice of broilers in the
bright and dark areas reflects their choice of light cycle. In the early stage of growth, the chicks
chose close to 24 hours of light might because the chicks just after hatching were not sensitive to
light and need to find food to survive. In the late growth period, broilers gradually adapted to the
environment and started to choose a light cycle that suits them. The broiler autonomous choice of
photoperiod is more in line with the laws of nature (Li et al., 2019), and whether this choice will
have an impact on physiological functions and production performance needs further research.
5. Conclusions
Under the yellow LED light environment, the production performance indexes of chicks in
the low-density group were better than those in the high-density group. The feed and water
consumption in the LA of the three experimental groups were significantly higher than those in
the DA. So, the chicks showed a great preference for the LA. However, the feed and water
consumption of the chicks in LA decreased in the middle and late stages indicating that broilers
had a lower preference for light in the middle and late stages of growth. Comparing the different
breeding density groups, the low-density group's LA feed and water consumption were
significantly higher than the high-density group, indicating that even in the same light area,
∙ 24 ∙
broilers in the low-density group have a stronger preference for the light area than the high-density
group. The statistical results of the residence frequency distribution characteristics show that
chicks have different light requirements at different growth stages under the condition of active
selection: 1) Low-density breeding environment: young chickens 23.8 L:0.2 D, adult chickens
22.3 L:1.7 D; 2) High-density breeding environment: 22.6 L:1.4 D for young chickens, 15.0 L:9.0
D for grown chickens. In addition, statistics on the frequency distribution of broilers' eating and
drinking behaviors in a day did not show a significant rule, and follow-up studies are needed.
References
Alvino, G.M., G.S. Archer, J.A. Mench, 2009a. Behavioural time budgets of broiler chickens reared in
varying light intensities. Applied Animal Behaviour Science. 118 (1–2), 54–61.
Alvino, G.M., R.A. Blatchford, G.S. Archer, J.A. Mench, 2009b. Light intensity during rearing affects
the behavioural synchrony and resting patterns of broiler chickens. British Poultry Science. 50 (3), 275–283.
Blatchford, R.A., K.C. Klasing, H.L. Shivaprasad, P.S. Wakenell, G.S. Archer, J.A. Mench, 2009. The
effect of light intensity on the behavior, eye and leg health, and immune function of broiler chickens. Poultry
Science. 88 (1), 20–28.
Deep, A., K. Schwean-Lardner, T.G. Crowe, B.I. Fancher, H.L. Classen, 2012. Effect of light intensity
on broiler behaviour and diurnal rhythms. Applied Animal Behaviour Science. 136 (1), 50–56.
Fidan, E.D., A. Nazligul, M.K. Turkyilmaz, S.U. Aypak, F.S. Kilimci, S. Karaarslan, M. Kaya, 2017.
Effect of photoperiod length and light intensity on some welfare criteria, carcass, and meat quality
characteristics in broilers. Revista Brasileira De Zootecnia-Brazilian Journal of Animal Science. 46 (3), 202–
210.
Ke, Y.Y., W.J. Liu, Z.X. Wang, Y.X. Chen, 2011. Effects of monochromatic light on quality properties
and antioxidation of meat in broilers. Poultry Science. 90 (11), 2632–2637.
Kristensen, H.H., N.B. Prescott, G.C. Perry, J. Ladewig, A.K. Ersboll, K.C. Overvad, C.M. Wathes, 2007.
The behaviour of broiler chickens in different light sources and illuminances. Applied Animal Behaviour
Science. 103 (1–2), 75–89.
Lewis, P.D., T.R. Morris, 2000. Poultry and coloured light. Worlds Poultry Science Journal. 56 (3), 189–
207.
Li, G., B. Li, Z. Shi, Y. Zhao, Q. Tong, Y. Liu, 2019. Diurnal rhythms of group-housed layer pullets with
free choices between light and dim environments. Canadian Journal of Animal Science. 100 (1), 37–46.
Mahmud, A., Saima, Rafiullah, I. Ali, 2011. Effect of Different Light Regimens on Performance of
Broilers. Journal of Animal and Plant Sciences. 21 (1), 104–106.
Olanrewaju, H., J. Thaxton, W. Dozier, J. Purswell, W. Roush, S. Branton, 2006. A review of lighting
programs for broiler production. International journal of poultry science. 5 (4), 301–308.
Pan, C., M. Lu, Y. Zhang, Y. Yu, Q. Lu, Y. Yang, J. Pan, 2019. Unevenly distributed LED light produces
distinct behavioral preferences and production performance of broilers. International Journal of Agricultural
and Biological Engineering. 12 (2), 49–53.
Renden, J., E. Moran Jr, S. Kincaid, 1996. Lighting programs for broilers that reduce leg problems without
loss of performance or yield. Poultry Science. 75 (11), 1345–1350.
Rozenboim, I., B. Robinzon, A. Rosenstrauch, 1999. Effect of light source and regimen on growing
broilers. British Poultry Science. 40 (4), 452–457.
Rozenboim, I., I. Biran, Y. Chaiseha, S. Yahav, A. Rosenstrauch, D. Sklan, O. Halevy, 2004. The effect
of a green and blue monochromatic light combination on broiler growth and development. Poultry Science.
83 (5), 842–845.
Senaratna, D., T. Samarakone, A. Madusanka, W. Gunawardane, 2012. Preference of broiler chicken for
different light colors in relation to age, session of the day and behavior. Tropical Agricultural Research. 23
(3), 193–203.
Tang, Z., Y. Ma, 2004. The tectofugal pathway in bird's visual system. Journal of Guangxi Normal
University. 22 (4), 79–82.
van Hasselt, S.J., M. Rusche, A.L. Vyssotski, S. Verhulst, N.C. Rattenborg, P. Meerlo, 2020. Sleep time
in the european starling is strongly affected by night length and moon phase. Current Biology. 30 (9), 1664–
1671.
Yu, Z., S. Zhengxiang, Z. Furong, 2006. Effects of music, stocking density on growth performance and
carcass quality of crossbreed broiler. Transactions of the CSAE. 22, 155–158.
∙ 25 ∙
A Comparison of Ventilation Systems in Three Different Types of

Multi-Floor Pig Buildings Using Numerical Simulation
Xiaoshuai Wang a, Feiyue Hu a, Tian Xie b, Kaiying Wang a,*
a
b
Qingdao Big Herdsman Machinery Co., Ltd, Qingdao, Shandong 266000, China
* Corresponding author. Email: zjuwky@zju.edu.cn
Abstract
An increasing number of large pig farms attempt to build multi-floor pig buildings (MFPBs)
instead of conventional one-floor pig buildings in China because MFPB has a great advantage in
saving land and increasing benefit margin. Currently, the structure of MFPB, especially the
ventilation system, varies greatly because MFPB is a fresh concept and lacks common standards.
As well known, the ventilation system is critical to indoor environmental control, and different
ventilation systems will result in different ventilation performance (i.e., ventilation rate and
airflow pattern) in MFPBs. However, no study has reported a comparison of the ventilation
performance among the different ventilation systems used in MFABs. This work aims to compare
the ventilation performance (ventilation rate and airflow pattern) of three popular MFPB types
using the Computational Fluid Dynamics (CFD) technique. To compare the ventilation
performance of the three MFPBs, CFD models for the three MFPBs were properly modelled and
validated using field-measured data. The CFD models were then extended to simulate the airflow
patterns and ventilation rates of the three MFPBs. The simulation results showed that the airflow
pattern in each MFPB was similar. Compared with the two buildings with fans installed on each
floor, the ventilation rate of building with openings in the end wall and fans installed in the top of
the yard was around 25% less. Besides, the ventilation rate of each floor increased from the first
to the top floor for both buildings with a yard, while no significant difference was observed in the
building without a yard. Increase the yard’s width could mitigate the variation in VR of each floor.
Therefore, from the ventilation perspective, the fans installed separately on each floor were
recommended rather than the ones that fans were installed together at the top of the yard.
Keywords: Airflow pattern, multi-floor pig building, numerical evaluation, ventilation rate,
yard’s width
1. Introduction
China has the world’s largest swine industry. A wave of African swine fever outbreaks these
years has wiped out many small-scale pig farms with conventional one-floor buildings because of
the outdated facilities and the poor control of bio-security since 2018. Instead, multi-floor pig
buildings (MFPB) has sprung out in many provinces in China, because MFPB 1) performs
potentially better in control of bio-safety, 2) requires much less land (the land-use cost continuedly
increase in China), and 3) increases the profit margin for the animal feeding operations (AFOs)
(Wang et al., 2021).
For a commercial and intensive pig building, the indoor environmental condition is important,
because it affects the performance, behavior, and health of pigs, as well as the quantity and quality
of the pork (Huynh et al., 2005; Spoolder et al., 2012; Baxter et al., 2015). The ventilation system
in the animal building plays an extremely important role in the control of the indoor environmental
condition because ventilation can remove the extra heat and moisture from the building and induce
fresh air to replace the polluted air (Ecim-Djuric and Topisirovic, 2010; Saha et al., 2011). Survey
shows that tunnel ventilation is mostly applied in the MFPB in the summertime, but the difference
in the tunnel ventilation system still exists. Therefore, it is necessary to evaluate the performance
of each tunnel ventilation system. This evaluation could provide useful information for the future
design of MFPB.
∙ 26 ∙
To date, few researchers have conducted systematic evaluations on the ventilation

performance in MFPB. Gao et al. (2018) reported differences in air temperature and relative
humidity among floors inside 4 seven-floor pig buildings and stated that the differences in
stocking density, pig size, and growth stage accounted for the differences in air temperature and
relative humidity. Besides, Wang et al. (2021) firstly assessed the effect of floor numbers (ranging
from 6-floor to 10-floor) and yard width (20m and 30m) on the ventilation performance for a
specific type of MFPG. According to the simulated results, they concluded that increasing the
floor number negatively affects the ventilation rates of MFPB and increasing the width of the yard
benefits to ventilation rates of MFPB. Despite these two studies, no more study has systematically
evaluated the performance of the current ventilation systems applied in MFPBs.
To perform such evaluation, researchers can conduct several field experiments with different
types of MFPBs. However, field experiment is costly, time-consuming, and labor-intensive, let
alone the difficulty in getting permission from AFOs under the ongoing ASF epidemic. Note that,
it is almost impossible to conduct a field experiment in commercial pig buildings in China
currently. Recently, the Computational Fluid Dynamics (CFD) simulation has been widely used
in the evaluation of the indoor environmental condition and the ventilation system in many animal
buildings (Qin et al., 2020; Yi et al., 2020; Wang et al., 2021). All the studies proved that CFD
simulation is a feasible, fast, and parameter-controllable way to conduct such evaluation. Thus,
this work was conducted by means of CFD simulation.
To sum up, this study aims to compare the performance of ventilation systems of the three
types of MFPBs. The ventilation rate, airflow pattern, and ventilation efficiency would be properly
evaluated by means of numerical simulations.
2.1. Assumption
This study was based on the following assumptions:
 The ambient environmental condition was stable during the field measurement. The
blockage effect of the steel posts and feeding troughs on the airflow has not been considered
due to the limited effect comparing with pigs.
 The evaluation of ventilation was based on isothermal simulations because the air
movement inside the building was mainly determined by exhausting fans, the effect of
thermal buoyance on the ventilation rate was limited in summer for a tunnel ventilated
building.
2.2. Description of the three MFABs
Generally, there are three popular types of MFPB, as shown in Figure 1. The first type of
MFPB (referred to as B1) simply stacks the conventional one-floor pig barn to a multi-floor pig
barn (Figure 1a). The second type of MFPB (referred to as B2) has a yard surrounding the
exhausting fans (Figure 1b), the outgoing air from each floor was released from the top of the
yard. In B1 and B2, the wet pads and exhausted fans were installed at the opposite endwalls at
each floor, serving as air inlets and outlets, respectively. The difference between B1 and B2 was
the existence of a yard. The third type of MFPB (referred to as B3) is similar to B2, expect the
location of fans installed (Figure 1c). In B3, the fans were installed together at the top part of a
higher yard (comparing with the one in B2) instead of installing fans at the endwall at each floor
as did in B1 and B2. There are four openings in the endwall. Both B2 and B3 had a yard, which
could collect polluted air and released the air from the top of the building before the pollutions
such as NH3, H2S, odors, and particulate matter, were well removed.
∙ 27 ∙
Figure 1. Sketches of the three different MFABs: (a) B1, (b) B2, and (c) B3.
2.3. CFD modeling
2.3.1. Computational domain and grid distribution
Considering the structural characteristics of MFPBs, only a unit of each MFPB was modeled
for the simulation. By doing so, the simulation cost (time and RAM) could be greatly reduced
without great sacrifice of the accuracy of the simulation. Figure 2a-c show the computational
domains for the three MFABs, which consist of a building unit and the ambient domain.
Solidworks (Dassault Systems S.A. France) was used for modeling the computational domain.
Geometric information for the computational domain was also illustrated in Figure 2a-c.
Figure 2. (a-c) Computational domain including the MFPBs (B1, B2, and B3, respectively) and
the corresponding expanded domain. (d) The mesh distribution of a virtual vertical plane of B2.
The structural (hexahedral) meshing strategy was applied to ensure the mesh of good quality,
in which the growth ratios of the grid were set to be less than 1.25. To ensure that the simulated
results were independent of the grid resolution, a grid independent test was conducted with four
levels of the grid resolution. Compared with the data from the finest one, the coarsest one with a
relative error of less than 1% was selected for the following simulations. Figure 2d shows the
selected mesh distribution strategy (using B2 as an example), and the mesh distribution for the
rest two buildings were similar. The meshing procedure was done using Ansys ICEM (ANSYS,
Inc. USA).
2.3.2. Governing equation and simulation scheme
The steady Reynolds-averaged Navier-Stokes (RANS) approach was applied because of its
effectivity and accuracy in air movement prediction (Rong et al., 2016). The general governing
equation for a steady case can be expressed using Eq (1) (Patankar, 1980):
𝑑𝑖𝑣(𝜌𝛷𝑢) = 𝑑𝑖𝑣(𝛤𝛷 𝑔𝑟𝑎𝑑𝛷) + 𝑆𝛷 (1 )
where ρ is density, kg m-3; u is the velocity, ms-1; Φ represents the common variables of interest,
i.e., velocity, m s-1, temperature, K, species (such as moisture), turbulent kinetic energy, m2 s-2,
∙ 28 ∙
and its dissipation rate, m2 s-3; ΓΦ is the transport coefficient dependent on Φ, and SΦ is the source
term dependent on Φ.
The governing equations were discretized by the SIMPLE (Semi-Implicit Method for
Pressure-Linked Equations) scheme. Second-order upwind discretization was selected for
momentum, turbulent kinetic energy, and specific dissipation rate to improve the accuracy of the
final solution. The Realizable k-ε turbulence model was chosen because it is deemed as a reliable
computational turbulence model in many similar studies related to airflow inside livestock
buildings (Norton et al., 2010; Bjerg et al., 2013). Enhanced wall treatment was enabled in the
simulation, and the area-weighted average values of Y+ in all the simulation cases were kept
within the required range (Y+ >30). Ansys Fluent (ANSYS, Inc. USA) was used to run all the
simulations in this study.
2.3.3. Simplifications of fans, wet-pad, slatted floor, and animal occupied zone
Exhaust fans produce pressure jump when operating, and the pressure jump is dependent on
the air velocity passing through the fan’s orifice. To properly model the fans on the endwalls, Fan,
a type of boundary condition in Fluent, was selected in the CFD model. The pressure jump in the
CFD model would automatically be modified according to the passing through air velocity,
following the Eq. (2), which was determined by the performance curve based on the final report
by Bess Lab, UIUC (Lab, 2017), with R2 of 0.99.
∆𝑃 = −0.2562𝑣 3 + 3.3931𝑣 2 − 19.88𝑣 + 101.96 (2)
where the ΔP denotes the pressure jump generated by a fan, Pa; v is the air velocity passing
through the fan.
Regarding the simplifications of wet-pads, slatted floors, and animal occupied zone (AOZ),
porous medium treatment was applied with specific viscous and inertial resistance coefficients for
all these three components. Detailed setups for all these three components could be found in
(Wang et al., 2021)’s study. All the used viscous and inertial resistance coefficients were listed in
Table 1.
Table 1. The viscous and inertial resistance coefficients for wet-pads, slatted floor, and AOZ.
X-direction1 Y-direction Z-direction
Item
R1, m-1 R2, m-2 R1, m-1 R2, m-2 R1, m-1 R2, m-2
AOZ 0.4 400 0.4 400 0.4 400
Slatted floors 80 15,000 80 15,000 - -
Wet pads 11.06 182,754 - - - -
1
Direction of X, Y, and Z were indicated in Figure 2b
2.4. Field experiment for CFD model validation
Field measurements were carried out in a five-floor pig building (Figure 3a) located in
Kaijiang, Sichuan province, China on May 27th, 2021. The pig building was new, in which the
first floor was used for finishing, the second and third floor were used for piglet nursery, and the
fourth and fifth floor were used for farrowing. The pig building is similar to the B3 (as illustrated
in Figure 1c). There was no pig raised in the building during the field measurement. In the field
experiments, the environmental parameters (i.e., air velocity (v), air temperature (Ta), and relative
humidity (RH) at each measuring location were simultaneously recorded every 5 seconds for 40
minutes using portable hot-wire anemometers (Figure 3b) (VelociCalc® Multi-Function
Ventilation Meter 9565-A, TSI, Shoreview, Minnesota, USA, the accuracy of ±1.5%, ±0.3°C, and
±3% for v, Ta, and RH, respectively). There were 5–6 measuring locations with a height of 1.5m
from the slatted floor at each floor. Figure 3c-e show the layout of the measurement locations on
different floors. The measurement was started from the first floor, and the measurement lasts for
7 hours from 8:30 am to 2:30 pm. During the measurement, the fans were all turned on.
CFD model was validated by comparing the simulated air velocities with the field measured
∙ 29 ∙
ones, because the airflow acted the most important role in this study. most of the simulated
velocities were close to the corresponding measured ones, and the variations of both the measured
and simulated air velocities were similar. Therefore, the CFD model developed in this study could
be deemed valid.
Figure 3. Photos of the (a) tested MFPB and (b) measurement sensor. Layout of the
measurement locations at (c) 1st floor (finishing room), (d) 2nd and 3rd floor (nursery room),
and (e) 4th and 5th floor (farrowing room) in the tested MFPB (marked as the blue dot).
2.5. Data processing and analysis
The ventilation rate, VR, was selected as the comparing parameter in this study to
quantitatively evaluated the difference among floors and MFPBs. To obtain the ventilation rate at
each floor under each type of MFPB, the averaged-velocity in the x-direction (vave_x) at the cross-
section of the fan cones for B1 and B2 and the openings for B3 (which were indicated in Figure
2c) was used, which was obtained from Ansys Fluent (ANSYS, Inc., Canonsburg, Pennsylvania,
USA). The equation VR=vave_x × A was used to calculate the ventilation rate of each floor, in
which the A denotes the cross-sectional area of either the fan cones or the openings. Besides,
Office Excel, CFD Post (ANSYS, Inc.), and Origin 2020 (OriginLab Co., Northampton,
Massachusetts, USA) were used for data analysis and plot.
3.1. Airflow pattern in MFPBs
Figure 4 shows the contours of air velocity and corresponding vector plot at the plane with a
height of 0.3m and 1.3m from the floor on the third floor of the three MFPBs. As can be seen, the
airflow patterns in all the three MFPBs were similar, which can be attributed to the similar
ventilation systems used in these three buildings. The air velocity at AOZ was much less than the
air velocity in the height of 1.3m from the floor, indicating that most of the fresh air did not go
through the AOZ. Besides, according to Figure 5a-c, there was a dark blue area at the pens close
to the inlet openings, indicating that the air movement there was very slow. Such a stagnant zone
occurred because the openings were set too high (0.9m from the floor, 0.3m higher than the animal
occupied zone). As well known, the heat, moisture, and pollutant gases would be accumulated
due to the lack of air movement in the stagnant zones, resulting in a poor living environment for
pigs there. To improve the living environment in the stagnant area, additional partial ventilation
systems (e.g., baffles or ceiling deflectors (Cheng et al., 2018; Zhou et al., 2019)) that can increase
the local air exchange rate and direct fresh air at headspace into the AOZ, were highly
recommended.
∙ 30 ∙
Figure 4. Contours and vector plots of the air velocity at a height of 0.3 m (a, b, c) and 1.3m (d,
e, f) of B1, B2, and B3, respectively.
3.2. Ventilation rates in MFPBs
Figure 5a shows that the ventilation rate at each floor of all three types of MFPB. For the
ventilation rates in B1, no significant difference was observed among each floor with an average
ventilation rate of 106.66 × 103 m3 h-1. This is because the ventilation system on each floor was
relatively independent in B1. The ventilation rates in B2 increased gradually from the first floor
(103.58 × 103 m3 h-1) to the fifth floor (107.81 × 103 m3 h-1). The relative difference between the
first floor and fifth floor was 3.9%. The variation tendency of the ventilation rates in B2 was
consistent with the previously published study (Wang et al., 2021), which can be attributed to the
existence of the yard in B2 and the different pressure in the yard.
Figure 5. Ventilation rate at each floor of the three five-floor pig buildings.
Similarly, the ventilation rate of each floor increased gradually from the first floor (77.62 ×
103 m3 h-1) to the fifth floor (79.85 × 103 m3 h-1) in B3, indicating that the pressure difference in
the yard also affected the ventilation rate of each floor. However, the ventilation rates were around
25% less than the ventilation rates of B1 and B2. The openings in the endwall (indicated in Figure
1c) accounted for the decrease in the ventilation rate. These openings could be deemed as orifices,
when the air passes through the openings, its pressure builds up slightly upstream of the openings.
Theoretically, enlarging the size of the openings could decrease the pressure drop when airflow
passing through the openings, resulting in an increase in the ventilation rates at each floor.
3.3. Effect of yard’s width on ventilation rates in MFPBs
To evaluate the yard’s width effect on the ventilation performance of B2 and B3, additional
simulations with three levels of the yard’s widths, i.e., 4m, 6m, and 8m, were simulated. The
buildings with a 6m-wide yard were deemed as practical cases. Figure 6a shows the ventilation
∙ 31 ∙
rate of each floor of B2 and B3 with different widths of the yard. As can be seen, the relative
difference in ventilation rate between the bottom floor and the top floor decreased as the width of
the yard increase for both types of buildings, indicating that prolonging the width of the yard could
narrow the difference in ventilation rates between the bottom and top floor. Similar results were
reported by Wang et al. (2021). This phenomenon could be attributed to the pressure difference
in the yard with a different width. Figure 6b illustrates the pressure profile inside the yard in B2
and B3 with three levels of the yard’s width. As it can be seen, the pressure increased along with
the height increase from the bottom to the top of the yard. Generally, given a similar ventilation
rate, the smaller sectional area of the yard (because of the smaller width of the yard) would cause
greater average air velocity, resulting in a greater pressure gradient field inside the yard. The
difference in the local pressure close to the fans among floors affected the fans’ ventilation
performance in B2 and the difference in the local pressure close to the openings among floors also
affect the ventilation rate that passing through the openings. and for B3, the pressure at the
resulting in the difference in ventilation rate among floors.
Figure 6. Ventilation rate at each floor (a) and contours of the pressure profile in B2 and B3(b)
with 4m, 6m, and 8m -wide yard.
4. Conclusions
The multi-floor pig building has been sprung out in China. Many types of MFPB have been
designed and constructed. In this study, we evaluated the performance of the ventilation system
in three popular types of MFPB. Based on the results, we concluded:
 The airflow patterns at each floor of all the three MFPBs were similar. The air movement
in the animal-occupied zone was limited with most of the fresh air passing through the
headspace.
 The openings at the endwall in B3 contribute to a ~25% decrease of ventilation rate at each
floor comparing with those of B1 and B2.
 Yard accounts for the difference in ventilation rate among floors in B2 and B3, while no
significant difference in the ventilation rate at each floor in B1 was observed. The pressure
gradient in the yard accounts for such a difference.
 The width of the yard positively affected the ventilation performance, and a wider yard
could narrow the relative difference of ventilation rate between the bottom floor and top
floor.
Acknowledgments
This study was supported by the Development Program of Zhejiang Province, China (Grant
No. 2018C02035). The authors appreciated the help from the engineers in the experimental pig
farm during field experiments.
∙ 32 ∙
References
Baxter, E.M., O.O. Adeleye, M.C. Jack, M. Farish, S.H. Ison, S.A. Edwards, 2015. Achieving optimum
performance in a loose-housed farrowing system for sows: The effects of space and temperature. Applied
Animal Behaviour Science. 169, 9–16.
Bjerg, B., G. Cascone, I.-B. Lee, T. Bartzanas, T. Norton, S.-W. Hong, I.-H. Seo, T. Banhazi, P. Liberati,
A. Marucci, G. Zhang, 2013. Modelling of ammonia emissions from naturally ventilated livestock buildings.
Part 3: CFD modelling. Biosystems Engineering. 116 (3), 259–275.
Cheng, Q., H. Li, L. Rong, X. Feng, G. Zhang, B. Li, 2018. Using CFD to assess the influence of ceiling
deflector design on airflow distribution in hen house with tunnel ventilation. Computers and Electronics in
Agriculture. 151, 165–174.
Ecim-Djuric, O., G. Topisirovic, 2010. Energy efficiency optimization of combined ventilation systems
in livestock buildings. Energy and Buildings. 42 (8), 1165–1171.
Gao, Y., Y. Diao, C. Lin, Y. Liu, C. Guo, M. Lei, Y. Tong, X. Li, 2018. Monitoring and analysis of
thermal environment and harmful gases in mechanically ventilated multistory pig buildings (in Chinese).
Transactions of the Chinese Society of Agricultural Engineering. 34 (4), 239–247.
Huynh, T.T.T., A.J.A. Aarnink, W.J.J. Gerrits, M.J.H. Heetkamp, T.T. Canh, H.A.M. Spoolder, B. Kemp,
M.W.A. Verstegen, 2005. Thermal behaviour of growing pigs in response to high temperature and humidity.
Applied Animal Behaviour Science. 91 (1), 1–16.
Lab, B., 2017. Final Report #17137.
Norton, T., J. Grant, R. Fallon, D.-W. Sun, 2010. Optimising the ventilation configuration of naturally
ventilated livestock buildings for improved indoor environmental homogeneity. Building and Environment.
45 (4), 983–995.
Patankar, S., 1980. Numerical heat transfer and fluid flow. Washington, DC, Hemisphere Publishing
Corp.: CRC press. 210 p.
Qin, C., X. Wang, G. Zhang, Q. Yi, Y. He, K. Wang, 2020. Effects of the slatted floor layout on flow
pattern in a manure pit and ammonia emission from pit — A CFD study. Computers and Electronics in
Agriculture. 177, 105677.
Rong, L., P.V. Nielsen, B. Bjerg, G. Zhang, 2016. Summary of best guidelines and validation of CFD
modeling in livestock buildings to ensure prediction quality. Computers and Electronics in Agriculture. 121,
180–190.
Saha, C.K., G. Zhang, J.-Q. Ni, Z. Ye, 2011. Similarity criteria for estimating gas emission from scale
models. Biosystems Engineering. 108 (3), 227–236.
Spoolder, H.A.M., A.A.J. Aarnink, H.M. Vermeer, J. van Riel, S.A. Edwards, 2012. Effect of increasing
temperature on space requirements of group housed finishing pigs. Applied Animal Behaviour Science. 138
(3), 229–239.
Wang, X., J. Wu, Q. Yi, G. Zhang, T. Amon, D. Janke, X. Li, B. Chen, Y. He, K. Wang, 2021. Numerical
evaluation on ventilation rates of a novel multi-floor pig building using computational fluid dynamics.
Computers and Electronics in Agriculture. 182, 106050.
Yi, Q., D. Janke, L. Thormann, G. Zhang, B. Amon, S. Hempel, Š. Nosek, E. Hartung, T. Amon, 2020.
Airflow Characteristics Downwind a Naturally Ventilated Pig Building with a Roofed Outdoor Exercise Yard
and Implications on Pollutant Distribution. Applied Sciences. 10 (14), 4931.
Zhou, B., X. Wang, M.R. Mondaca, L. Rong, C.Y. Choi, 2019. Assessment of optimal airflow baffle
locations and angles in mechanically-ventilated dairy houses using computational fluid dynamics. Computers
and Electronics in Agriculture. 165, 104930.
∙ 33 ∙
Concentration and Emission of Particle Matters from an

Intensive Dairy Farm in Northern China
Yujian Lu a,b, Lei E a,b, Chaoyuan Wang a,b,*, Zhenpeng Hu a,b, Yu Liu a,b
a
Department of Agricultural Structure and Bioenvironmental Engineering,
College of Water Resources and Civil Engineering, China Agricultural University, Beijing 100083, China
b
Key Laboratory of Agricultural Engineering in Structure and Environment,
Ministry of Agriculture and Rural Affairs, Beijing 100083, China
* Corresponding author. Email: gotowchy@cau.edu.cn
Abstract
Public concerns on air quality issues continuously grow in the intensification development of
animal industry in China. As an important aerosol pollutant, particulate matter (PM) production
and its emission from dairy farm have potential threat to the health of cows, workers and
neighbors, especially for the open and semi-open production systems. To characterize PM
concentration and emission from dairy farming in Northern China, a field measurement system
was developed with Internet of Things (IoT) technique for a long-term monitoring of PM2.5 and
total suspended particle (TSP), carbon dioxide (CO2) and ammonia (NH3) concentrations as well
as temperature (T), and relative humidity (RH) inside and outside the barns of an intensive dairy
farm. Ventilation rate of the barn was calculated following the CO2 mass balance method to
quantify the emissions. Results showed that the mean PM2.5, TSP, CO2 and NH3 concentrations
inside the barn in winter were 62.5 μg m-3, 104.5 μg m-3 and 1453.4 mg m-3, 0.35 mg m-3,
respectively. Indoor PM2.5 concentration were nearly two times higher than the ambient (62.5 μg
m-3 vs. 33.2 μg m-3), but not significantly different to the outdoor (62.5 μg m-3 vs. 57.7 μg m-3),
which illustrated the barn could be a main source of fine particle. Meanwhile, PM2.5 concentration
had significant correlations with NH3 concentration and RH, and peaks of TSP concentration were
observed during the feeding and milking period. Average emission rates of PM2.5, TSP and NH3
during the measurement were 8.6 mg h-1 hpu-1, 25.6 mg h-1 hpu-1 and 417.8 mg h-1 hpu-1,
respectively.
Keywords: Field measurement, long-period monitoring, IoT, emission rate, ventilation.
1. Introduction
Air quality issues originated from the intensification development of animal farming has
increasingly become a public concern. Because of the high stocking density, concentrated animal
feeding operations (CAFOs) are regarded as the main emission sources of particle matter (PM),
ammonia (NH3), greenhouse gas and other gaseous pollutants, having a considerable contribution
to the atmospheric environment among agricultural activities (Klimont et al., 2002). Among
which, PM has an adverse effect on animal warfare, production performance and efficiency (Al-
Homidan et al., 2003), as well as the health of farming workers and neighbors (Williams et al.,
2011; Mitchell et al., 2015; Wegesser et al., 2009). On the one hand, gaseous pollutants, bacteria,
and viruses carried by PM are associated with respiratory diseases, which cause increased
occurrence of chronic cough and/or phlegm, chronic bronchitis, allergic reactions and asthma-like
symptoms to livestock farmers (Cambra-Lopez et al., 2010). On the other hand, PM released from
CAFOs can transport harmful materials to the ambient environment, which may potentially
weaken the health of nearby people. Additionally, odorous molecules can be absorbed on PM and
transported to surrounding residents to decrease living condition and cause complaints.
Most PM studies on field measurement are focused on pigs and poultries (Cambra-Lopez et
al., 2010), while limited research is on dairy farming. Takai et al. (1998), Winkel et al. (2015) and
Joo et al (2013) investigated PM characteristics from dairies in northern Europe, the Netherlands
and America (Takai et al., 1998; Joo et al., 2013; Winkel et al., 2015). However, the results from
different studies varied considerably, and dynamic patterns of PM concentration and emission
∙ 34 ∙
from dairy farming are still not very clear. Furthermore, compared with pigs or poultries, dairies
are commonly using fully open or semi open barns with natural ventilation systems, resulting in
more challenges to quantify PM, and the PM monitoring instruments with high accuracy are
generally expensive, which also constrain the multi-points and long-period measurement to obtain
the field data directly. Generally, it is still lack of PM fundamental data of dairies with different
production systems, especially in Northern China, which is a dominating milk production area in
China. Currently available results from other countries or areas could not illustrate the real
conditions of dairy farming in Northern China neither.
Using self-developed environmental monitors with an open-source internet of things (IoT)
technique, this study conducted a long-term field measurement with the objectives to (1) quantify
indoor and outdoor PM, CO2 and NH3 concentrations in winter; and (2) determine PM2.5 and TSP
emissions from a typical naturally ventilated dairy barn in northern China.
2.1. Investigated dairy farm
A field measurement was conducted from January 1 to March 31, 2021 in a commercially
operated dairy farm (117°04'E, 39°25'N) in Beichen District, Tianjin city, Northern China. As
shown in Figure 1, the dairy farm, which was surrounded by crop fields, consisted of ten free-stall
barns with natural ventilation systems. All the barns had pitched-roof structures with the ridge
openings oriented from the west to east and were equipped with curtains rolling from top and
bottom on the sidewalls to adjust the ventilations in cold seasons.
The selected dairy barn for measurement (Barn A in Figure 1) located at the northwestern of
the farm, was 186 m long × 31 m wide × 4 m eave height. During the test, nearly 100 Holstein
heifers, 100 dry cows and 200 lactating cows were housed in four rows of head-to-head free-stall
cubicles inside Barn A. Generally, feed delivering (at 8:00, 15:00, and 18:00), manure removing
(at 8:00, and 15:00) and bedding litter renewing (every two days) were operated when the cows
left for milking at 8:00, 15:00, and 24:00 of each day.
2.2. Measurement
2.2.1 sampling locations
As shown in Figure 1, aerial pollutant concentrations (PM2.5, TSP, CO2, and NH3) and
environmental variables (T and RH) were monitored inside and outside Barn A. Indoor air
samplers were located at five locations. Four were installed in the center of four pens at an
approximately 2.5 m to 3 m height above the floor to monitor air quality of the animal occupation
zone. The other one was hung at ridge opening with a height of about 1 m to the roof to measure
outgoing gas concentrations, and the sampling point was close to the middle of the barn.
Four outdoor samplers were installed in the west (A1), north (A2), northeast (A3) and
southwest (A4) directions out of the farm boundary, which were approximately 198 m, 185 m,
370 m and 256 m away from the center of Barn A (Figure 1), respectively. The farm was in a
typical continental monsoon climate with prevailing winds from southeast toward northwest in
summer and from northwest to southeast in winter. The selection of outdoor positions was mainly
based on the wind direction and electricity availability.
Ambient PM concentration and meteorological data were obtained from the nearest stations
of China National Environmental Monitoring Centre (CNEMC, http://www.cnemc.cn/en/) and
China Meteorological Administration (http://www.cma.gov.cn/en), which were 11.4 km and 7.5
km away from the investigated dairy farm.
2.2.2 monitoring system and sensors
To achieve a real-time monitoring through multi-points for a seasonal period, an IoT
technique was applied in this study. The self-developed portable devices, including particulate
concentration monitoring unit (PCMU) and portable monitoring unit (PMU), were used to
∙ 35 ∙
continuously monitor the PM2.5, TSP concentrations as well as CO2 and NH3 concentrations, T,
and RH, respectively. Each PCMU and PMU were integrated by a microcontroller unit (MCU)
and several sensors to real time detect the gas concentrations and thermal environment parameters
(Li et al., 2020). Detail information on the sensors of PCMU and PMU are given in Table 1. After
being acquired by the portable units, the data were then sent to an IoT platform (Thingsboard-
3.1.1) through the 4G network, which was an open-source platform for device management, data
collection, processing and visualization. Time-series data uploaded from PMU and PCMU were
visualized on the platform interface and then saved into the Cassandra database.
Figure 1. The sampling locations inside and outside the Barn A.

The sampling interval of PM2.5 and TSP of PCMU was 1 min. The PM concentration of each
sampling point was averaged and uploaded at a 5 min duration after removing the maximum and
minimum values. The PMU had a 20 min sampling cycle. In each cycle, the PMU had a 5 min
purging duration to recover the accuracy of electrochemical NH 3 sensor, a 10 min sampling
duration to collect CO2, NH3, T and RH information of the air sucked into the test chamber, and
a 5 min sleeping duration to release the heat generated by the embedded pump and solenoids. The
data logging interval was 1 min.
Table 1. Detail specification of sensors integrated into PCMU and PMU.
Devices Variable Range Accuracy Sensor type
PM2.5 0–1 mg m-3 1 μg m-3 PM5003T
PCMU
TSP 0–20 mg m-3 1 μg m-3 SDS198
T -40–+80 °C ± 0.3 °C
AM2305
RH 0–99.9% ± 2% (10–90%)
PMU ± 50 ppm + 3.0%
CO2 0–9800 mg m-3 MH-Z14A
of the reading
NH3 0–38 mg m-3 0.1 mg m-3 ZEO3
2.3. Data processing and statistical analysis
2.3.1. ventilation and emission rate calculation
A CO2 mass-balance method was applied to estimate ventilation rate (VR) based on the total
∙ 36 ∙
CO2 production and its concentration difference between outgoing (CCO2, exhaust) and incoming
(CCO2, inlet) air of the barn by equation (1), where hpu is heat production unit of the cows and 1 hpu
equals 1000 W of total heat at 20 °C (Pedersen et al., 1998). The CCO2, exhaust and CCO2, inlet were
directly measured in the barn, and more detail of the total CO2 production determination is givens
by the CIGR report (CIGR, 2002). When calculating the ventilation rate, only the data with the
indoor and outdoor CO2 concentration difference being greater than 138 mg m-3 were used (Wang
et al., 2016). In this process, CO2 generated from manure was ignored.
Emission rates (ER) of the gaseous pollutants were calculated using equation (2), where
Cexhaust and Cinlet were NH3, PM2.5 and TSP concentrations in the exhaust and inlet air. To make
an easy comparison with other studies, emission rates were expressed in three different units of
production: mg h-1 hpu-1, mg h-1 LU-1(1LU=500 kg) and mg h-1 animal-1.
0.185×A
VR hpu = (1)
CCO2, exhaust -CCO2, inlet ×10-6
ER = VR×(Cexhaust -Cinlet ) (2)
2.3.2. data preprocessing
Original data were preprocessed, during which values of the obvious outliers such as over-
range data were firstly omitted and then a box plot was further applied to delete other outliers,
which can not directly reveal the real environment of the farm. Afterwards, all data were
resampled to hourly mean vaules as the basic data unit for analysis. Study showed that ambient
concentrations on smoggy days had a significant impact on indoor PM concentration. To reveal
the real PM dynamics of the dairy barn, data measured in those smoggy days with the ambient
AQI (air quality index) over 100 were excluded in the PM analysis in this study (Li et al., 2020).
Then, preprocessed data were used for correlation analysis by Spearman and Nonparametric
tests. Statistical significant difference was defined when the P-value was less than 0.05 (P<0.05).
The results were expressed as mean±standard deviation (SD). The software used included Excel
2016 (Microsoft, USA), Python 3 (version 3.0) and SPSS 22 (IBM, USA).
3.1. PM concentration
3.1.1. Statistical analysis
The indoor and outdoor hourly mean PM concentrations varied from 1st Jan to 30th March
were shown in Figure 2 and the descriptive statistics were presented in Table 2.
During the measurement, the hourly mean T and RH were 3.7 °C (-18.9 °C–23.1 °C) and
49.5% (8.4 %–51.3 %). The PM2.5 and TSP concentrations inside the cubicle dairy barn in winter
were 62.5 μg m-3 and 104.3 μg m-3, which were greater than the outside (57.7 μg m-3 and 94.1 μg
m-3 for PM2.5 and TSP). No significant difference of the indoor and outdoor PM2.5 concentrations
was observed (P>0.05). Meanwhile, the overall indoor PM2.5 concentration was nearly two times
higher than the ambient data (P<0.01) obtained from the nearest CNEMC stations, which
illustrates that the dairy barn could be a main source of fine particulates of dairy farming.
The PM2.5 concentration measured in our study was generally comparable to most results of
previous studies in winter (Takai et al., 1998; Kaasik et al., 2013). But it was higher than some
cases (Joo et al., 2015; Winkel et al., 2015), which might be due to bedding material difference of
the farms. Takai et al. (1998) found respirable dust concentrations of dairies with litter bedding in
England, the Netherlands, Denmark, and Germany were 170 μg m-3, 60 μg m-3, 50 μg m-3, and 30
μg m-3, respectively. Kaasik et al. (2013) measured nine uninsulated loose housing cowsheds in
Estonia and reported the mean PM10 concentration was 65 μg m-3. They also concluded that the
amounts of litter greatly contributed to the concentrations of respirable and inhalable particulates.
In our study, the recycled manure solids as bedding materials were renewed every two days.
As shown Figure 2, a significant fluctuation of average TSP concentrations with a wide
∙ 37 ∙
measuring range (5.6 μg m-3–624.5 μg m-3) was found in this study, and the concentration was
generally increased during feeding or milking operation and then settled down to the normal level
afterwards. TSP concentration levels in this study were also corresponded with the results of
Kaasik and Goodrich’s research on free-stall dairy barns with bedding materials (Kaasik et al.,
2013; Goodrich., 2006), but were relatively lower than other measurements (Joo et al., 2015;
Schmidt et al., 2002). This might be associated with the different sampling methods and sampling
heights. Most of the above studies measured PM concentrations at a 1 m height or at the cattle
respiratory level with intermittent sampling, while in our study the automatic devices were set at
2.5 m to 3 m heights without the cows’ interferes and the continuous measurement method was
adopted during the whole season.
Figure 2. Indoor and outdoor PM2.5 and TSP concentration variation during measurement.
Table 2. Statistic information of indoor, outdoor, and ambient PM concentrations.
Indoor Outdoor Ambient
PM
C* R* C R C R
PM2.5 62.5a* 1.5–225.3 57.7a 2.0–224.3 33.2b 1.0–74.0
TSP 104.3a 5.6–624.0 94.1b 5.2–611.7 NA* NA
* C and R represent concentration and range, respectively;
* Different superscript lowercase letters indicate significant difference (P<0.05);
* NA represent not available.
In general, the highest concentrations of PM2.5 and TSP mainly occurred around late evening
and early morning. The peak values of TSP appeared around 8:00, 15:00 and 18:00, and lasted
for 1 h, which agreed the farm management. During the periods, daily managements on milking,
feeding, and manure removing were operated, which caused more cow and vehicle activities to
consequently produce more coarse particulates to the air.
∙ 38 ∙
3.1.2. Diurnal patterns

The diurnal patterns of hourly mean PM concentrations and RH inside Barn A in winter were
shown in Figure 3. The PM2.5 concentrations were positively correlated with RH (r=0.35, p<0.05)
and NH3 concentration (r=0.63, p<0.05), which was also supported by Kaasik et al. (2013). Under
a moist condition, coarse fractions commonly assembled together into large particulates and easily
settled down to the floor, while fine fractions contained a large proportion of liquid components,
which may partially explain the relations. Furthermore, previous studies have shown a moist
environment with relatively high NH3 concentrations could accelerate the formation of PM2.5
(Hristov et al., 2011).
Figure 3. Diurnal patterns of ambient and indoor PM concentration and RH.

The red dotted line in Figure 3 showed a regulated 24-hour PM2.5 concentration (35 μg m-3)
in the National Ambient Air Quality Standards (NAAQS) of American (USEPA, 2011). During
our test, the mean indoor and outdoor PM2.5 concentrations (62.5 μg m-3 and 57.7 μg m-3) were
generally above the standard level, which may present potential threats for the farm workers for a
long period exposure into the environment. Pope et al. (2002) assessed the relationship between
long-term exposure to the air pollution and some diseases and found each 10-μg m-3 elevation in
fine particulate air pollution resulted in approximately 4%, 6%, and 8% increased risk of all-cause,
cardiopulmonary, and lung cancer mortality, respectively (Pope et al., 2002).
3.2. Ventilation and emission rate
Table 3 revealed the PM2.5, TSP and NH3 emissions of Barn A during winter period, which
were expressed in three different units for comparison.
Table 3. Emission rates of NH3, PM2.5 and TSP.
Emission rates
Pollutants
mg h-1 hpu-1 mg h-1 LU-1 mg h-1 animal-1
PM2.5 8.6 9.8 12.2
TSP 25.6 28.9 36.1
NH3 417.8 469.4 586.8
During measurement, the hourly mean CO2 concentrations inside and outside the barn were
1453.4 and 1274.0 mg m-3, which satisfied the application of CO2 mass balance method. The
hourly mean ventilation rate was 995.1 m3 h-1 hpu-1, which was in accordance with Winkle’s
measurements for natural ventilated dairy barns in the Netherland (Winkel et al., 2015). Due to
more ventilation, CO2 and NH3 concentrations in daytime were smaller than nighttime. The hourly
mean PM2.5, TSP and NH3 emission rates were 8.6 mg h-1 hpu-1, 25.6 mg h-1 hpu-1 and 417.8 mg
h-1 hpu-1, respectively. As for PM2.5 and NH3, similar or relatively low level of emission rates can
∙ 39 ∙
also be found in other studies for dairy barns with natural ventilation (Takai et al., 1998; Winkel
et al., 2015). Williams et al. (2011) conducted an exposure assessment of NH3 and PM at homes
varied with distance to scaled dairy operations, and found airborne biological contaminants
associated with airborne PM were statistically dispersed at distances up to 3 miles (4.8 km) away
from the dairies. Thus, relatively high level of PM2.5 and NH3 emissions from selected dairies, as
a major part of the aerosol pollutants of dairy farming, may potentially present healthy and
ecological threats to the adjacent communities.
4. Conclusions
Concentrations and emissions of PM, CO2 and NH3 from a naturally ventilated dairy barn
during winter in Northern China were monitored with self-developed devices and IoT technique.
During the measurement, hourly mean PM2.5, TSP, CO2, and NH3 concentrations inside the dairy
barn were 62.5 μg m-3, 104.3 μg m-3, 1453.4 mg m-3, and 0.35 mg m-3, respectively. Indoor PM2.5
concentration was nearly two times higher than the ambient, while no significant difference was
observed compared with outdoor concentration, and dairy farming elevated the fine particle
environment in the vicinity area. PM2.5 had a significant and positive correlation with indoor NH3
concentrations and RH, while daily management including feeding, milking, manure removing
and bedding renewing, was found to have a clear influence on TSP variation. Hourly mean PM2.5,
TSP and NH3 emissions were 8.6 mg h-1 hpu-1, 25.6 mg h-1 hpu-1 and 417.8 mg h-1 hpu-1, which
were comparable to reported emissions from naturally ventilated dairy barns.
Acknowledgements
This work was supported by the National Natural Science Foundation of China (grant number
31972614).
References
Homidan, A.A., J.F. Robertson, A.M. Petchey, 2003. Review of the effect of ammonia and dust
concentrations on broiler performance. World's Poultry Science Journal. 59 (3), 340–349. http://doi.org/
10.1079/WPS20030021.
Cambra-Lopez, M., A.J. Aarnink, Y. Zhao, S. Calvet, A.G. Torres, 2010. Airborne particulate matter
from livestock production systems: a review of an air pollution problem. Environ Pollut. 158 (1), 1–17.
http://doi.org/10.1016/j.envpol.2009.07.011.
Pedersen, S. and K. Sällvik, 2002. 4th Report of Working Group on Climatization of animal houses. Heat
and moisture production at animal and house levels (eds.). International Commission of Agricultural
Engineering (CIGR), Section II.
Goodrich, L. B., 2006. A PM10 Emission Factor for Freestall Dairies. (Master), MS Thesi, Department
of Biological and Agricultural Engineering, Texas A&M University, TX, USA.
Hristov, A.N., 2011. Technical note: Contribution of ammonia emitted from livestock to atmospheric fine
particulate matter (PM2.5) in the United States. Journal of Dairy Science. 94 (6), 3130–3136. http://doi.org/
10.3168/jds.2010-3681.
Joo, H.S., P.M. Ndegwa, A.J. Heber, J.Q. Ni, B.W. Bogan, J.C. Ramirez-Dorronsoro, E.L. Cortus, 2013.
Particulate matter dynamics in naturally ventilated freestall dairy barns. Atmospheric Environment. 69, 182–
190. http://doi.org/10.1016/j.atmosenv.2012.12.006.
Klimont, Z., Amann, M, 2002. European control strategies for fine particles: the potential role of
agriculture. Landbauforschung Völkenrode, 235, 29–35.
Kaasik, A., M. Maasikmets, IAgrE, 2013. Concentrations of airborne particulate matter, ammonia and
carbon dioxide in large scale uninsulated loose housing cowsheds in Estonia. Biosystems Engineering. 114
(3), 223–231.
Li, Z., C. Wang, B. Li, Z. Shi, W. Zheng, G. Teng, 2020. Concentration and size distribution of particulate
matter in a new aviary system for laying hens in China. Journal of the Air & Waste Management Association.
70 (4), 379–392. http://doi.org/10.1080/10962247.2020.1722291.
Mitchell, D.C., T.L. Armitage, M.B. Schenker, D.H. Bennett, D.J. Tancredi, C.E. Langer, S.J. Reynolds,
G. Dooley, J. Mehaffy, F.M. Mitloehner, 2015. Particulate Matter, Endotoxin, and Worker Respiratory Health
on Large Californian Dairies. Journal of Occupational and Environmental Medicine. 57 (1), 79–87.
http://doi.org/10.1097/jom.0000000000000304.
∙ 40 ∙
McGinn, S.M., T.K. Flesch, D. Chen, B. Crenna, O.T. Denmead, T. Naylor, D. Rowell, 2010. Coarse
Particulate Matter Emissions from Cattle Feedlots in Australia. Journal of Environmental Quality. 39 (3), 791–
798. http://doi.org/10.2134/jeq2009.0240.
Pedersen, S., H. Takai, J.O. Johnsen, J.H.M. Metz, P.W.G. Groot Koerkamp, G.H. Uenk, V.R. Phillips,
M.R. Holden, R.W. Sneath, J.L. Short, R.P. White, J. Hartung, J. Seedorf, M. Schröder, K.H.H. Linkert, C.M.
Wathes, 1998. A Comparison of Three Balance Methods for Calculating Ventilation Rates in Livestock
Buildings. Journal of Agricultural Engineering Research. 70 (1), 25–37. https://doi.org/10.1006/
jaer.1997.0276.
Pope, C.A., 3rd, R.T. Burnett, M.J. Thun, E.E. Calle, D. Krewski, K. Ito, G.D. Thurston, 2002. Lung
cancer, cardiopulmonary mortality, and long-term exposure to fine particulate air pollution. JAMA. 287 (9),
1132–1141.
Takai, H., S. Pedersen, J.O. Johnsen, J.H.M. Metz, P.W.G. Groot Koerkamp, G.H. Uenk, V.R. Phillips,
M.R. Holden, R.W. Sneath, J.L. Short, R.P. White, J. Hartung, J. Seedorf, M. Schröder, K.H. Linkert, C.M.
Wathes, 1998. Concentrations and Emissions of Airborne Dust in Livestock Buildings in Northern Europe.
Journal of Agricultural Engineering Research. 70 (1), 59–77. http://doi.org/https://doi.org/10.1006/
jaer.1997.0280.
Schmidt, D., Jacobson, L., Janni, K, 2002. Continuous monitoring of ammonia, hydrogen sulfide and dust
emissions from swine, dairy and poultry barns. 2002 ASAE Annual International Meeting/CIGR XVth World
Congress, Chicago, lllinois, USA, July 28, July 31, 2002.
USEPA. 2011. National ambient air quality standards. USEPA, Washington, DC.
Williams, D.A.L., P.N. Breysse, M.C. McCormack, G.B. Diette, S. McKenzie, A.S. Geyh, 2011. Airborne
cow allergen, ammonia and particulate matter at homes vary with distance to industrial scale dairy operations:
an exposure assessment. Environmental Health. 10. http://doi.org/10.1186/1476-069x-10-72.
Winkel, A., J. Mosquera, P.W.G. Groot Koerkamp, N.W.M. Ogink, A.J.A. Aarnink, 2015. Emissions of
particulate matter from animal houses in the Netherlands. Atmospheric Environment. 111, 202–212.
http://doi.org/10.1016/j.atmosenv.2015.03.047.
Wegesser, T.C., J.A. Last, 2009. Mouse lung inflammation after instillation of particulate matter collected
from a working dairy barn. Toxicology and Applied Pharmacology. 236 (3), 348–357. http://doi.org/10.1016/
j.taap.2009.02.023.
Wang, X., P.M. Ndegwa, H. Joo, G.M. Neerackal, C.O. Stöckle, H. Liu, J.H. Harrison, 2016. Indirect
method versus direct method for measuring ventilation rates in naturally ventilated dairy houses. Biosystems
Engineering. 144, 13–25. http://doi.org/https://doi.org/10.1016/j.biosystemseng.2016.01.010.
Zhao, L.Y., M.F. Brugger, R.B. Manuzon, G. Arnold, E. Imerman, L. Zhao, 2007. Variations in air quality
of new Ohio dairy facilities with natural ventilation systems. Applied Engineering in Agriculture. 23 (3), 339–
346.
∙ 41 ∙
Comparative Assessment of the Breeding Egg Disinfection Using

Slightly Acid Electrolyzed Water and Benzalkonium Bromide
Chang Liu a,b, Weichao Zheng a,b,c, *, Ling Zhou a, Yuxuan Sun a, Shengqiang Han d,
Zonggang Li a,b
a
b
c
Beijing Engineering Research Center for Animal Healthy Environment, Beijing 100083, China
d
Jiangsu Lihua animal husbandry Co., Ltd, Changzhou, Jiangsu 213168, China
* Corresponding author. Email: weichaozheng@cau.edu.cn
Abstract
Conventional chemical disinfectants, used for egg disinfection, could result in toxic residue
and endanger hatchability, chick quality, and pullet growth performance. Slightly acidic
electrolyzed water (SAEW) is known as a novel disinfectant for egg sterilization due to its high
efficiency and no residue. In this study, a comprehensive assessment of spraying sterilization
efficiency on eggshell and knock-on effect on eggshell quality and breeding egg incubation was
conducted by using SAEW and benzalkonium bromide solution (BBS), at different
concentrations, spraying volumes, and sterilization durations. The results showed that the
sterilization efficiency of SAEW increased with the increase of available chlorine concentrations
(ACC), spraying volumes, and sterilization durations. Slightly acidic electrolyzed water with
ACC of 150 mg L-1 and 10 g L-1 benzalkonium bromide had the same sterilization rates of
approximately 86.2% at a spraying volume of 0.5 mL egg-1 and sterilization durations of 180 s.
Neither had significant effect on eggshell strength or thickness. The eggshell cuticle quality in the
benzalkonium bromide group was significantly higher than the control group (no disinfection)
and the 150 mg L-1 SAEW group. Embryo weight, relative embryo weight, and hatchability in the
SAEW group had no significant differences with those in the benzalkonium bromide group.
SAEW should be more popular because of its simple preparation, low cost and no residue. The
results indicated that SAEW was an alternative disinfectant for the sterilization of breeding eggs
instead of conventional chemical disinfectants such as benzalkonium bromide.
Keywords: Breeding egg disinfection, alternative disinfectant, eggshell quality, egg incubation
1. Introduction
Breeding eggs laid from hen may contact feces, dust, and other pollutants, and the surface will
be contaminated with pathogenic microorganisms (De Reu et al., 2006). The incubation
temperature is suitable for the growth of pathogenic bacteria enter into the embryo through
eggshells, resulting in embryo deformities and even death (Zeweil et al., 2015). Therefore, it is
necessary to sterilize breeding eggs to inactivate bacteria on the eggshell (Olsen et al., 2017).
Formaldehyde, one of conventional chemical disinfectants widely used in egg disinfection,
could result in toxic residue and endanger hatchability, lower chick quality and pullet growth
performance (Oliveira et al., 2020). Several studies have endeavored to find alternative
disinfectants. Peracetic acid, UV-C, O3, hydrogen peroxide + alcohol, essential oil, and
benzalkonium bromide were also applied to assess the disinfection and hatching effect (Wells et
al., 2010; Al-Ajeeli et al., 2016; Li et al., 2018; Huang et al., 2019; Motola et al., 2020; Oliveira
et al., 2020; Tebrun et al., 2020; Wlazlo et al., 2020). Besides, eggshell cuticle plays a protective
function in trans-shell contamination with bacteria (Sheng et al., 2021). In addition to disinfection
and incubation results, avoiding damage to the cuticle during disinfection is also necessary.
Slightly acidic electrolyzed water (SAEW) is known as a novel disinfectant for egg
sterilization due to its high efficiency and no residue (Rahman et al., 2010). The effective form of
chlorine compounds in SAEW mainly is the hypochlorous acid (HOCl) having strong
∙ 42 ∙
antimicrobial activity (Cao et al., 2009). Cao et al. (2009) reported SAEW with a pH value of 5.0–
6.5 may be a promising disinfectant agent for eggshell washing without environmental pollution.
Ni et al. (2014) demonstrated that the bactericidal activity of SAEW against E.coli O157: H7,
Staphylococcus aureus, and total aerobic bacteria on the eggshell was significantly higher than
that of chlorine dioxide and NaClO solutions at the same ACCs of 80 or 100 mg L-1. Zang et al.
(2019) found that a complete inactivation of S. Enteritidis and E. coli on the eggshell resulted
from treatment with SAEW immersed at an ACC of 26 mg L-1 for 3 and 4 min. SAEW has been
shown to be effective in disinfecting eggshell, however, there is a lack of research breeding eggs
disinfection parameters and effects. Other disinfectants have been shown to be harmful to embryos
(Wlazlo et al., 2020). So, effect of SAEW on embryonic development and hatching performance
need to be studied. Application of SAEW on egg by washing or immersing cause damage to
cuticle (Zang et al., 2019), however, the effect of spraying SAEW on eggshell quality still
unknown. In short, there is a lack of published research on the combined effects of SAEW on the
microbial count, eggshell quality, and hatchability concomitantly.
The objective of this research is to evaluate the efficacy of SAEW spraying on breeding eggs,
in comparison benzalkonium bromide solutions (BBS) used in hatchery, on the reduction of
eggshell microbial count, eggshell quality and hatchability performance.
2.1. Disinfection experiment
All eggs were disinfected in a self-made spray equipment (Figure 1). The spray systems were
made by (JDT-12, Fog Machine, Chaoshan, China) and SAX00 high-pressure nozzle with a spray
flow of 0.02 L min-1. The ventilation systems were made by a blower (CZR-40, Shanghai Xin
long, Shanghai, China), and the air volume is 1.07 m3 min-1.
HEPA High-pressure
nozzle
Hyperbaric A blower
spraying
A container
for disinfectant
Figure 1. Self-made spray equipment.

A total of 1008 eggs (Jinghong No.1) were randomly divided into 5 groups: Control (no
treatment), Tap water, SAEW (50, 100 and 150 mg L-1), and BBS (10,000 mg L-1). Spray time
was 90 s, spray volume was 0.5 mL egg-1 during disinfection process, after disinfection the space
keep closed for 90 s. Then the recommend disinfection concentration can be determined through
the sterilization effect. Then 672 eggs divided into different groups and disinfected as outlined in
Table 1 to determine the recommend SAEW disinfection technology including spray time, spray
volume and closed time.
The SAEW was generated by an electrolyzed water generator (JH-120, Shenzhen Jianghe
Biotechnology Co., Ltd, Shenzhen, China). The pH values were measured by an AZ8601pH
meter. ACC was measured by RC-3F (Kasahara Chemical Instruments Corp., Saitama, Japan).
BBS was prepared by dilution of 5% of the original solution (Shanghai Yunjia Huangpu
∙ 43 ∙
Pharmaceutical Co., Ltd, Shanghai, China).

After disinfection, 12 random eggs from each group were placed in a sterile sampling bag
with 50 mL normal saline, and 1 mL of sample solution was taken out after shaking. 1 mL of the
resultant rinse solution was diluted serially using normal saline. The samples were plated on plate
count agar that were incubated at 37 °C for 24 h prior to counting the resultant colonies.
Table 1. Disinfection experiment design for determine spray volume and sterilization duration.
Spray volume Closed time
Disinfectant Spray time (s)
(mL egg-1) (s)
ND (control group/No disinfection) / / /
60 0.33 120
SAEW (recommend ACC) 90 0.5 90
120 0.67 60
2.2. Eggshell quality experiment
In total of 504 breeding eggs was randomly divided into 3 groups and disinfected with the
ACC, spray volume, and sterilization duration described in Section 2.1. After disinfection, 10
random eggs of each group were sampled for eggshell strength, eggshell thickness, and opacity
of cuticle. The experiment was repeated three times.
Eggshell Strength: The eggshell strength was measured using Model-I, Robotmation, Japan.
Eggshell Thickness: The shell from the blunt end was opened and the inner shell membrane
was separated. Then the thickness of the blunt end, medium end and sharp end eggshell was
measured using a helical micrometer (0–25 mm/0.001 mm, Mitutoy, Japan). Their mean values
represented the eggshell thickness.
Opacity of cuticle: The color of eggshell before and after dyeing with MST glue protective
blue dye was measured by color spectrophotometer, and then the opacity of cuticle (α value) was
calculated (Chen et al., 2019).
2.3. Hatching Evaluation
A total of 400 eggs was randomly divided into 2 groups and disinfected with the ACC, spray
volume, and sterilization duration described in Section 2.1, and incubated. The experiment was
also repeated three times.
Embryo weight and relative embryo weight: On the 6th, 10th, 14th, and 18th day of
incubation, 10 random eggs were taken from each group. Recording egg and embryos weight, and
using the Eq. (1). calculate relative embryo weight.
Relative embryo weight= (embryo weight / egg weight) × 100% (1)
Hatchability: After hatching, the numbers of breeding eggs, unfertilized eggs, and chicks were
counted. The hatchability was calculated using Eq. (2).
Hatchability= [chicks number / (breeding eggs number – unfertilized eggs number)] × 100% (2)
Embryonic mortality: After the hatching, the rest of eggs in the hatching tray were beaten and
analyzed. Embryonic mortality was calculated by using the Eq. (3), Eq. (4), and Eq. (5).
Early dead = (1 to 7 d of incubation dead embryos number / fertile eggs number) × 100% (3)
Mid dead = (8 to 14 d of incubation dead embryos number / fertile eggs number) × 100% (4)
Late dead = (15 to 21 d of incubation dead embryos number / fertile eggs number) × 100% (5)
2.4 Data Analysis
The data of disinfection and eggshell quality were calculated by single-factor variance
analysis, and the hatching data were analyzed by t-test. SPSS was used to analyze the data and
Origin was used to draw the charts. Statistical significance for all tests was considered at P<0.05.
∙ 44 ∙

3.1. Optimization of disinfection parameters of slightly acid electrolyzed water and benzalkonium
bromide
The bactericidal rate of SAEW increased with the increasing ACC. No significant difference
between the total bacteria after 150 mg L-1 SAEW treatment and that of 10,000 mg L-1 BBS
treatment (P >0.05) indicate they had comparable bactericidal effect (Figure 2).
Figure 2. Effect of different disinfectant and disinfection technology on bactericidal effect.

a–b Lowercase alphabets mean differences between mean values are significant at P<0.05.
SAEW1: spray 60 s and closed for120 s; SAEW2: spray 90 s and closed for 90 s; SAEW3:
spray 120 s and closed for 60 s.
∙ 45 ∙
This result showed that the antibacterial efficacy of SAEW depends on ACC, which were in
agreement with Zhang et al. (2021) who reported that there were significant differences among
different concentrations of SAEW (P<0.05) The disinfection technology used in this experiment
was spray instead of immersion, which explained the difference in the bactericidal effect on the
surface of eggshells by Zang et al. (2019). Besides, organic matter such as feces and dust on the
surface of the eggshell may make low concentration disinfection ineffective. Ni et al. (2015)
reported that the bactericidal efficiency of SAEW decreased with increasing bovine serum
albumin concentration. Hao et al. (2013) reported that SAEW with an ACC of 250 mg L-1
inactivated 100% of bacteria and fungi in solid materials (dusts, feces, feather, and feed). This
result explained ACC of 50 mg L-1 and 100 mg L-1 SAEW had no significant differences, while a
higher concentration SAEW (150 mg L-1) had a significant difference (P<0.05) because organic
existed.
Benzalkonium bromide solutions, a kind of quaternary ammonium salt disinfection, took
bactericidal action by inducing intracellular proteins and electrolytes leakage. According to the
report (Huang et al., 2019), BBS could bind to an extracellular matrix with electrostatic interaction
and making biofilm removal easier, which explains it’s high efficient sterilization and the reason
for 150 mg L-1 SAEW solution and 10,000 mg L-1 BBS had the same bactericidal effect.
From Figure 2 (B), under the treatment of spraying for 90 s and closed for 90 s, the bactericide
rate (86.2%) was the highest, which indicates that less spray quantity and less contacting time are
not sufficient to achieve a good disinfection effect. Similar results were reported by Ni et al.
(2016), which showed that the bactericidal activity of SAEW increased with the increasing ACC
and spraying duration. However, Wang et al. (2018) reported that a microbial reduction of almost
1.0 log CFU cm-2 was observed in chicken carcasses after spraying with 30 mg L-1 SAEW for 15
s. Different results in these studies may be due to the droplet size (Zang et al., 2017).
3.2. Effect of slightly acid electrolyzed water and benzalkonium bromide disinfection on eggshell
quality
Neither of the treatments had effect on eggshell strength nor eggshell thickness (P >0.05,
Figure 3). This result is consistent with the result of Melo et al. (2019) who found hydrogen
peroxide spraying, per acetic acid spraying, and water spraying did not affect eggshell quality.
Figure 3. Effect of different disinfectant treatments on eggshell strength and thickness of

breeding eggs. a–b Lowercase alphabets mean differences between mean values are significant at
P<0.05.
∙ 46 ∙
The opacity of eggshell cuticle of Control group, SAEW and BBS was 26.07 ± 2.65%, 21.37
± 2.62% and 38.34 ± 2.31%, respectively. There was no significant difference in the opacity of
eggshell cuticle between the Control group and the SAEW treatment group (P >0.05), but both
groups were significantly smaller than the BBS treatment group (P >0.05). Zang et al. (2019)
used the same disinfection but identified that SAEW may damage the egg cuticle, thus favoring
trans-shell contamination with bacteria. This result occurred may be due to Zang et al. (2019)
immersed eggs in SAEW, while SAEW spraying was used in this experiment. Fasenko et al.
(2009) also reported that there was no effect on cuticle structure by spray EO water. Besides, the
morphology of eggshells were usually directly measured by electron microscopy (Melo et al.,
2019), while we used an dyeing method reported by Chen et al. (2019). The reason why the
opacity of cuticle in this group is higher than other groups may be that there is more foam in the
BBS solution, which forms a thin film on the eggshell and confuses the results. Besides, the cuticle
layer contains antibacterial proteins, which may be denatured by BBS. Therefore, it is necessary
to further investigate if eggshell cuticle could be affected by BBS by using electron microscopy.
3.3. Effect of slightly acid electrolyzed water and benzalkonium bromide on hatching performance
The weight of the embryo can reflect the development of the embryo. Table 2 shows that there
were significant differences in embryos in the SAEW group and the BBS group (P<0.05) on day
14 and 18, while no significant differences in the other groups. However, we could not draw a
conclusion that SAEW treatment was better than BBS treatment in embryo development. High
embryonic weight may be related to egg weight, and it is more accurate to use relative embryonic
weight as a developmental evaluation standard. No significant effects of different disinfection
treatments were found on the weight of 1-day-old chicks between SAEW and BBS (Table 2),
which also proves that those disinfectants didn’t make a difference in affecting embryo
development.
Table 2. Effects of different disinfection treatments on embryo weight, relative embryo weight
during incubation and 1-day-old chick weight.
The incubation Embryo weight/1-day-old Relative embryo
Treatments
time chick weight (g) weight (%)
SAEW 0.418±0.016 0.689±0.245
Day 6
BBS 0.419±0.007 0.699±0.017
SAEW 2.622±0.042 4.544±0.085
Day 10
BBS 2.647±0.083 4.632±0.191
SAEW 10.763±0.194a 18.659±0.334
Day 14
BBS 9.777±0.332b 17.341±0.669
SAEW 26.307±0.218a 48.456±0.649
Day 18
BBS 24.962±0.564b 46.362±1.081
SAEW 41.751±0.195 -
1-day-old chicks
BBS 42.161±0.250 -
From Table 3, hatchability, early dead, mid dead, and late dead had no significance between
the two groups (P >0.05). Former studies also explore disinfectants effect on chick quality. Chick
quality, as determined by visual assessment and broiler weight at the time of hatch, was also not
affected by spray EO water (Fasenko et al., 2009). However, reduced hatching and significantly
increased mortality occurred in O3 group (Wlazlo et al., 2020). Different disinfectants show
different results, for the chemical disinfectants may toxic residue and endanger embryo after
disinfection. Main content of SAEW is hypochlorous acid which dissolves after disinfection (Cao
et al., 2009). BBS has antibacterial and low toxic properties (Huang et al., 2019). Even if there
was residue, it may have little effect on the development of the embryo.
∙ 47 ∙
Table 3. Hatchability and embryonic mortality of fertilized eggs incubated under different
disinfection treatments.
Hatchability (%) Early dead (%) Mid dead (%) Late dead (%)
SAEW 90.49±0.97a 3.17±0.48a 1.10±0.37a 4.69±0.18a
a
BBS 90.65±0.38 4.00±0.11 a 0.70±0.40 a 4.26±0.84a
4. Conclusions
In conclusion, SAEW can be an alternative disinfectant for the sterilization of breeding eggs
compared with BBS. The recommend disinfection technology is using SAEW as disinfectant with
ACC 150 mg L-1, spray volume 0.5 mL egg-1, and disinfection for 180 s. Although there were no
differences in germicidal efficacy, eggshell quality and hatching performance under the
recommended disinfection procedures, SAEW should be more popular because of its simple
preparation, low cost and no residue.
Acknowledgements
The animal study proposal was approved by The Laboratory Animal Ethical Committee of
China Agricultural University. The research team would like to acknowledge the financial support
for this research project from National Natural Science Foundation of China (31802109) and open
topic of Animal Disease Prevention and Food Safety Key Laboratory of Sichuan Province.
References
Al-Ajeeli, M.N., T.M. Taylor, C.Z. Alvarado, C.D. Coufal, 2016. Comparison of eggshell surface
sanitization technologies and impacts on consumer acceptability. Poult Sci. 95 (5), 1191–1197. http://doi.org/
10.3382/ps/pew014.
Cao, W., Z.W. Zhu, Z.X. Shi, C.Y. Wang, B.M. Li, 2009. Efficiency of slightly acidic electrolyzed water
for inactivation of Salmonella enteritidis and its contaminated shell eggs. Int J Food Microbiol. 130 (2), 88–
93. http://doi.org/10.1016/j.ijfoodmicro.2008.12.021.
Chen, X., X. Li, Y. Guo, W. Li, J. Song, G. Xu, N. Yang, J. Zheng, 2019. Impact of cuticle quality and
eggshell thickness on egg antibacterial efficiency. Poult Sci. 98 (2), 940–948. http://doi.org/10.3382/ps/
pey369.
De Reu, K., K. Grijspeerdt, W. Messens, M. Heyndrickx, M. Uyttendaele, J. Debevere, L. Herman, 2006.
Eggshell factors influencing eggshell penetration and whole egg contamination by different bacteria, including
Salmonella enteritidis. Int J Food Microbiol. 112 (3), 253–260. http://doi.org/10.1016/j.ijfoodmicro.
2006.04.011.
Fasenko, G.M., E.E. O'Dea Christopher, L.M. McMullen, 2009. Spraying hatching eggs with electrolyzed
oxidizing water reduces eggshell microbial load without compromising broiler production parameters. Poult
Sci. 88 (5), 1121–1127. http://doi.org/10.3382/ps.2008-00359.
Hao, X.X., B.M. Li, C.Y. Wang, Q. Zhang, W. Cao, 2013. Application of slightly acidic electrolyzed
water for inactivating microbes in a layer breeding house. Poult Sci. 92 (10), 2560–2566. http://doi.org/
10.3382/ps.2013-03117.
Huang, Z., Y. Lin, F. Ren, S. Song, H. Guo, 2019. Benzalkonium bromide is effective in removing
Bacillus cereus biofilm on stainless steel when combined with cleaning-in-place. Food Control. 105, 13–20.
http://doi.org/10.1016/j.foodcont.2019.05.017.
Li, X., D. Anderson, B. Rathgeber, N. McLean, J. MacIsaac, 2018. Fumigating broiler hatching eggs with
lysozyme product (Inovapure) to reduce eggshell microbial load. Poult Sci. 97 (12), 4252–4261. http://doi.org/
10.3382/ps/pey288.
Melo, E.F., W.L.S. Climaco, M.V. Triginelli, D.P. Vaz, M.R. de Souza, N.C. Baiao, M.A. Pompeu, L.J.C.
Lara, 2019. An evaluation of alternative methods for sanitizing hatching eggs. Poult Sci. 98 (6), 2466–2473.
http://doi.org/10.3382/ps/pez022.
Motola, G., H.M. Hafez, S. Bruggemann-Schwarze, 2020. Efficacy of six disinfection methods against
extended-spectrum beta-lactamase (ESBL) producing E. coli on eggshells in vitro. PLoS One. 15 (9),
e0238860. http://doi.org/10.1371/journal.pone.0238860.
∙ 48 ∙
Ni, L., W. Cao, W.-C. Zheng, H. Chen, B.-M. Li, 2014. Efficacy of Slightly Acidic Electrolyzed Water
for Reduction of Foodborne Pathogens and Natural Microflora on Shell Eggs. Food Science and Technology
Research. 20 (1), 93–100. http://doi.org/10.3136/fstr.20.93.
Ni, L., W. Cao, W.C. Zheng, Q. Zhang, B.M. Li, 2015. Reduction of microbial contamination on the
surfaces of layer houses using slightly acidic electrolyzed water. Poult Sci. 94 (11), 2838–2848. http://doi.org/
10.3382/ps/pev261.
Ni, L., W. Zheng, Q. Zhang, W. Cao, B. Li, 2016. Application of slightly acidic electrolyzed water for
decontamination of stainless steel surfaces in animal transport vehicles. Prev Vet Med. 133, 42–51.
http://doi.org/10.1016/j.prevetmed.2016.09.010.
Oliveira, G.D.S., V.M. Dos Santos, S.T. Nascimento, J.C. Rodrigues, 2020. Alternative sanitizers to
paraformaldehyde for incubation of fertile eggs. Poult Sci. 99 (4), 2001–2006. http://doi.org/10.1016/
j.psj.2019.11.032.
Olsen, R., E. Kudirkiene, I. Thofner, S. Pors, P. Karlskov-Mortensen, L. Li, S. Papasolomontos, C.
Angastiniotou, J. Christensen, 2017. Impact of egg disinfection of hatching eggs on the eggshell microbiome
and bacterial load. Poult Sci. 96 (11), 3901–3911. http://doi.org/10.3382/ps/pex182.
Rahman, S.M., T. Ding, D.H. Oh, 2010. Effectiveness of low concentration electrolyzed water to
inactivate foodborne pathogens under different environmental conditions. Int J Food Microbiol. 139 (3), 147–
153. http://doi.org/10.1016/j.ijfoodmicro.2010.03.020.
Sheng, X., D. Shu, Y. Li, Z. Zhan, X. Yuan, S. Liu, H. Wu, S. Bing, Y. Zang, 2021. Combined approach
consisting of slightly acidic electrolyzed water and chitosan coating to improve the internal quality of eggs
during storage. J Sci Food Agric. 101 (6), 2355–2361. http://doi.org/10.1002/jsfa.10858.
Tebrun, W., G. Motola, M.H. Hafez, J. Bachmeier, V. Schmidt, K. Renfert, C. Reichelt, S. Bruggemann-
Schwarze, M. Pees, 2020. Preliminary study: Health and performance assessment in broiler chicks following
application of six different hatching egg disinfection protocols. PLoS One. 15 (5), e0232825. http://doi.org/
10.1371/journal.pone.0232825.
Wang, H., J. Qi, D. Duan, Y. Dong, X. Xu, G. Zhou, 2018. Combination of a novel designed spray cabinet
and electrolyzed water to reduce microorganisms on chicken carcasses. Food Control. 86, 200–206.
http://doi.org/10.1016/j.foodcont.2017.11.027.
Wells, J.B., C.D. Coufal, H.M. Parker, C.D. McDaniel, 2010. Disinfection of eggshells using ultraviolet
light and hydrogen peroxide independently and in combination. Poult Sci. 89 (11), 2499–2505. http://doi.org/
10.3382/ps.2009-00604.
Wlazlo, L., K. Drabik, K.I.A. Al-Shammari, J. Batkowska, B. Nowakowicz-Debek, M. Gryzinska, 2020.
Use of reactive oxygen species (ozone, hydrogen peroxide) for disinfection of hatching eggs. Poult Sci. 99
(5), 2478–2484. http://doi.org/10.1016/j.psj.2019.12.039.
Zang, Y., B. Li, W. Zheng, X. Sheng, D. Shu, 2017. Influence of droplet size and deposition on slightly
acidic electrolyzed water spraying disinfection effect on livestock environment. Transactions of the Chinese
Society of Agricultural Engineering. 33 (9), 224–229.
Zang, Y.T., S. Bing, Y.J. Li, D.Q. Shu, A.M. Huang, H.X. Wu, L.T. Lan, H.D. Wu, 2019. Efficacy of
slightly acidic electrolyzed water on the microbial safety and shelf life of shelled eggs. Poult Sci. 98 (11),
5932–5939. http://doi.org/10.3382/ps/pez373.
Zeweil, H.S., R.E. Rizk, G.M. Bekhet, M.R. Ahmed, 2015. Comparing the effectiveness of egg
disinfectants against bacteria and mitotic indices of developing chick embryos. J Basic & App. Zoology. 70,
1–15. http://doi.org/10.1016/j.jobaz.2014.12.005.
Zhang, J., J. Wang, D. Zhao, J. Hao, 2021. Efficacy of the two-step disinfection with slightly acidic
electrolyzed water for reduction of Listeria monocytogenes contamination on food raw materials. Lwt. 140,
110699. http://doi.org/10.1016/j.lwt.2020.110699.
∙ 49 ∙
Heating Time and Heating Load of Fattening Pig Houses

in Different Climate Regions in Winter
Fei Qi a, Hao Li a,b, Jinjun Huang a, Zhengxiang Shi a,b,*
a
b
Beijing Engineering Research Center for Animal Healthy Environment, Beijing 100083, China
* Corresponding author. Email: shizhx@cau.edu.cn
Abstract
To reduce the cost of pig production and realize precise control of the environment in
commercial pig houses in winter, it is of great significance to reasonably design the minimum
ventilation and heating load in the houses. Based on the environmental meteorological data in
different regions, combined with minimum ventilation model of pig houses in winter, considering
the influence of pig weights, outdoor temperatures, feeding densities and building envelope
thermal resistances on the heating load of fattening pigs, this paper introduces two heating load
estimation models for pigs in winter. The model was used to study the heating time and heating
load of a model pig house with length and width of 110 m × 15 m in typical cities in five major
climate regions (Changchun, Beijing, Nanning, Wuhan, Guiyang). The resultant heating load
models for nursery pigs and finishing pigs have the correlation coefficient R2 in 0.966 and 0.963,
respectively. With increase in the thermal resistance of the enclosure structure of the pig house,
the heating load and its variation range are gradually reduced. Heating capacity of commercial
pig houses in typical cities of different climatic regions under the thermal resistance of the same
building envelope in an order of Changchun > Beijing > Guiyang > Wuhan > Nanning. The
heating load models provide references for the calculation of the heating time and heating load of
commercial pig houses in different climatic regions.
Keywords: Different climate regions, fattening pig houses, ventilation, heating load models,
heating duration
1. Introduction
Low environmental temperatures under the thermal natural zone in animal housing cause a
series of physiological changes, including accelerated breathing, vasoconstriction, enhanced
endocrine activities and accelerated nutrient and energy metabolism to maintain a normal body
temperature (Ji et al., 2021; Scharf et al., 2019; Quiniou et al., 2001). When pigs are in cold stress,
their weight gain rate decreases, feed intake and feed-to-weight ratio increase, which seriously
affects the economic benefits of pig farms (Lopez et al., 1991; Parvu et al., 2011). The
recommended air temperatures in the pig house in China (GBT 17824.3-2008) are 20–25℃ for
nursery pig houses with a low critical value of 16℃, and 15–23℃ for growth and finishing houses
with a low critical value of 13℃.
Heating of pig houses in winter is essential. Pigs in heated pig houses grow faster and have
fewer diseases than pigs in unheated ones (Mun et al., 2020; Pandorfi et al., 2005). Therefore,
reasonable design of the minimum ventilation and heating load in fattening pig house in winter is
of great significance for saving pig production and energy consumption costs. Ambient
environmental conditions determine the length of the heating season and the heating load of
fattening pig houses in different climate regions in winter. Li et al. (2018a, 2018b) compared the
ventilation rate calculated based on the heat balance, humidity balance and CO2 balance, the
minimum ventilation rate in pig houses in Guangxi was determined by the humidity balance in
winter, and the minimum ventilation rate in pig houses in North China was determined by the CO2
in winter. Besides, the stocking density and the thermal resistance of the building envelopes also
had significant impacts on the heating load (Wang et al., 2014; Zhao, 2017; Wang and Li, 2017).
However, there is still a lack of comparative studies on the heating load of pig houses in different
regions.
∙ 50 ∙
China has a vast territory and diverse climate types, and the differences among regions are
also obvious (Meteorological Data Room, Meteorological Information Center, China
Meteorological Administration, 2005). Due to the large differences in winter climate, the heating
load and heating time of the year-round fattening pig houses vary significantly in different regions.
The current research on the winter heating load and heating time of different cities is mainly
concentrated in the field of residential buildings. For example, the annual heating time in Harbin
is about 182 days, Beijing is about 125 days, and Xi'an and Jinan are about 100 days (Hu et al.,
2008). There is a lack of research on the heating time and load of the fattening pig houses.
In this research, the building size of a pig farm was taken as an example, considering the effect
of pig weights, outdoor temperatures, feeding densities and building envelope thermal resistances
on the heating load of fattening pigs. We introduced two heating load estimation models for pigs
in winter, and studied the annual heating time and heating load differences in the fattening pig
houses in different regions. This research provides a reference for the calculation of minimum
ventilation and heating energy consumption in the fattening pig houses in winter.
2.1. Building thermal zoning standards and selection of the research cities
This research quotes the thermal zoning standard in the "Code for Thermal Design of Civil
Buildings", which divided the building into five major thermal zoning: severe cold climate zone,
cold climate zone, hot summer and cold winter climate zone, hot summer and warm winter climate
zone and mild climate zone. In order to study differences in thermal engineering and
environmental control parameters of fattening pig house in different climate regions, this research
selected a typical city in each climate region to study, including Changchun (severe cold), Beijing
(cold), Wuhan (hot summer and cold winter), Nanning (hot summer and warm winter), Guiyang
(mild), and used a weather database software to obtain all-year weather data with hourly nodes.
2.2. Construction of the model pig house
A fattening pig house at Hongming Pig Farm in Qiongshan, Haikou, Hainan Province was
used as a model. The building size and thermal parameters are shown in Tables 1 and 2.
Table 1. Size of the model pig house.
Canopy Ridge Number and Number and Number and area
Length Width
height height area of pen area of doors of windows
(m) (m)
(m) (m) (n, m2) (n, m2) (n, m2)
110 15 3 4.75 72, 3 × 6 4, 2.1 × 1 56, 0.9 × 1.5
In production, if the stocking density is too high, pigs will attack and fight each other. If the
stocking density is too low, it will increase the average heating load of each pig, which is not a
good practice to saving production costs (Fu et al., 2016). Considering the above factors and pig
feeding welfare requirements, the feeding density of fattening pigs should meet Table 3.
The feeding densities of the nursery and finishing stage were 0.35 m2 head-1 and 0.8 m2
head-1, respectively (3702 pigs and 1620 pigs raised in the model pig house for nursery and
finishing pigs, respectively).
2.3. Calculation of minimum ventilation in pig houses in winter
The calculation of ventilation through the heat balance can be written as following:
𝑄s + 𝑄m + 𝑄h = 𝑄w + 𝑄v + 𝑄e (1)
where 𝑄s is the sensible heat produced by pigs, W; 𝑄m is the heat dissipated by lighting, motors,
and equipment, W, which value often ignored; 𝑄w is the heat consumption of the building through
the envelope, W; 𝑄v is the sensible heat loss of the air, W; 𝑄e is the sensible heat due to water
evaporation, W. Because some of the sensible heat have already considered this factor, so it is
∙ 51 ∙
generally not calculated separately; 𝑄h is the supplementary heat of the heating radiator, W.
Therefore, the equation of ventilation heat consumption can be simplified as:
𝑄v + 𝑄s + 𝑄w = 0 (2)
The value is positive when the air inside the house gets heat, and the value is negative when
the air inside the house loses heat.
Table 2. Thermal parameters of the model pig house.
Heat transfer Heat transfer
Building
Basic parameters Towards Area (m2) coefficient correction
envelope
(W m2 ℃-1) factor
East and west 81.6 2.02 0.91
Wall 240mm thick brick wall South 292.2 2.02 0.83
North 292.2 2.02 0.95
28*0.9*1.5 single-layer South 37.8 2.6 0.83
Window plastic steel window
North 37.8 2.6 0.95
outer seal plastic film
Door 4*1*2.1 iron gate East and west 8.4 6.4 1
Color steel plate filled
Roof - 1694.32 0.47 0.98
with rock wool
Floor Slatted floor - 1650 0.23 0.38
Table 3. Fattening pigs feeding process parameters.
Type of pigs Weight (kg) Feeding age Daily gain (kg) Stocking density (m2 head-1)
Nursery 14–34 49–77 0.44 0.3–0.4
Finishing 34–100 78–180 0.97 0.6–1.2
The minimum ventilation rate in winter can be calculated by moisture balance or the limit of
CO2 concentration in the house, as shown in equation (3).
𝑉min = max {𝑉t , 𝑉CO2 , 𝑉h } (3)
where 𝑉min is the minimum ventilation rate in winter in the fattening pig house, m3 h-1; 𝑉t is the
ventilation volume determined by the heat balance without heating, m3 h-1; 𝑉CO2 is the ventilation
volume determined by the CO2 balance, m3 h-1; 𝑉h is the ventilation volume determined by the
humidity balance, m3 h-1.
The amount of winter ventilation determined by the heat balance is derived from equation (2)
and shown in equation (4).
−𝑄s ×𝑌+𝑄w
𝑉t = (4)
𝐶p 𝜌w ∆𝑡
where 𝐶p is the specific heat capacity of air at constant pressure, 𝐶p=0.28W∙h (kg∙℃)-1; 𝜌w is the
air density, 𝜌w = 353/(𝑡 + 273), kg m-3; ∆𝑡 is the temperature difference between inside and
outside the house, ℃.
By determining the low-limit temperature 𝑡imin inside the house and the ambient temperature
𝑡 outside the house, the temperature difference ∆𝑡 between the inside and outside can be
calculated. 𝑡imin satisfied the equation (5) (Huang, 2020) of the weight of fattening pigs and the
low-limit temperature inside the house.
−0.2299𝑚 + 26.794(𝑚 < 60kg, 𝑚 is the weight of the fattening pig）
𝑡imin = { (5)
13（𝑚 ≥ 60kg, 𝑚 is the weight of the fattening pig）
The ventilation rate determined by the moisture balance in winter is shown in equation (6).
𝐹V
𝑉h = 3.6 × 106 (6)
𝜌w (𝑑i −𝑑0 )
∙ 52 ∙
where 𝐹V is the amount of water vapor removed by the ventilation of the fattening house, g h-1; 𝑑i
is the moisture content of the air inside the house, kg kg-1; 𝑑0 is the moisture content of the air
outside the house, kg kg-1; 3.6 × 106 is the unit conversion coefficient, 1kg s-1=3.6 × 106g h-1.
The amount of winter ventilation determined by the CO2 balance is shown in equation (7).
0.185×𝑄t20
𝑉CO2 = ×𝑌 (7)
1000×(𝐶i −𝐶0 )
where 𝑄t20 is the total heat output at 20 ℃, W; 𝐶i is the CO2 concentration, and the upper limit
concentration inside the house is 3.0 × 10-3, m3 m-3 (Yan and Li, 2011); 𝐶0 is the CO2
concentration outside the house, which is 0.3 × 10-3, m3 m-3; 𝑌 is the number of pigs, head; 0.185
is 0.185 m3 hpu-1 (hpu, total heat output of 1 000 pigs at 20 ℃) of CO2 output of fattening pigs.
When 𝑉min = 𝑉t , the heating load 𝑄p is 0, and the relative humidity and CO2 concentration
are both lower than the limit. When 𝑉min = 𝑉CO2 or 𝑉h , the heating load 𝑄p is not 0, so the
corresponding heat needs to be supplemented in the house to maintain the heat balance.
2.4. Calculation of heating load in pig house
In winter, when the sensible heat production of pigs in the fattening pig house is less than the
heat consumption of the building envelope and the ventilation heat dissipation, additional heat
needs to be provided to maintain the heat balance. This additional heating load satisfies the
following model:
𝑄p = 𝑄w + 𝐶p 𝜌w ∆𝑡𝑉min − 𝑄s × 𝑌 (8)
The heating load of each pig is determined by the weight of the pig, the temperature outside
the house, the stocking density and the thermal resistance of the building envelope.
2.5. Data entry
In order to study the influence of four factors (the feeding area of fattening pigs, the
temperature outside the house, the weight of the fattening pig, the thermal resistance of the
enclosure) on the heating load of fattening pigs, different gradients of stocking density and
building thermal resistance are set as initial values and imported into the simulation computing
platform for multiple calculations. The temperature and humidity data of Changchun City was
taken as the initial input of outdoor environmental data. The specific input parameters are shown
in Table 4.
Table 4. Initial input parameters of heating load calculation program for fattening pigs.
Weight Outdoor Stocking density Thermal resistance of
Type
(kg) environmental data (m2 head-1) envelope (m2 ℃ W-1)
Nursery Environmental data
14–34 0.2, 0.3, 0.4, 0.5
pigs of Changchun
1, 2, 3, 4, 5, 6
Finishing Environmental data 0.6, 0.7, 0.8, 0.9,
34–100
pigs of Changchun 1.0, 1.1, 1.2
The heating load of each fattening pig in the house can be got through the VBA model. Then
use SPSS to do a multiple regression analysis and obtain a mathematical model of the heating
load of the fattening pigs and the four factors.
To know the differences in heating time and heating load of fattening pig houses in different
regions, the environmental data of different cities and the size parameters, thermal parameters,
stocking density and other parameters of nursery pig houses and finishing pig houses were input
into the VBA program. Based on the calculated temperature outside the house in different regions
in winter and the temperature and humidity requirements in the fattening pig house, Yan et al.
(2011) obtained the low-limit heat resistance proposed for the fattening pig house in different
regions based on the calculation equation of the low-limit thermal resistance of the enclosure
structure: Nanning is no low-limit thermal resistance R0, Changchun’s R0 is 0.767, Beijing’s R0 is
0.469, Wuhan’s R0 is 0.479, Guiyang’s R0 is 0.282.Therefore, for the convenience of calculation,
∙ 53 ∙
the thermal resistance gradients of the wall and roof are set up as respectively 0.5, 1, 2, 3, 4, 5, 6
m2 ℃ W-1.
3.1. Minimum ventilation model in winter for fattening pigs of different weights
Inputting the external temperature and humidity data of the five regions in meteonorm 7 and
the corresponding indoor environmental parameter thresholds into the equation (3) algorithm
through the VBA language implementation, the minimum ventilation rates in winter under
different weights were generated and shown in Figure 1.
In this research, a regression analysis of pig weights and the corresponding theoretical winter
minimum ventilation was conducted rate under a comprehensively consideration of the
temperature conditions, relative humidity conditions and CO2 concentration in the house, the
winter minimum ventilation rate was established as shown in equation (9).
𝑉min = 5.1667 ∙ ln(𝑚) − 8.4354 (9)
where 𝑚 is the weight of the fattening pig, kg. Under the minimum ventilation rate in winter, if
the pig house still cannot meet the suitable temperature requirements for pig production, additional
heating measures should be considered.
Figure 1. The minimum ventilation rate in winter for pigs with different weights.
3.2. Mathematical model of heating load for fattening pigs
For nursery pigs, the heating load model for each pig in the house is shown in equation (10).
16.37 𝑡0
𝑄psi = 31.13 × 𝑛 − 3.048 × 𝑡0 + 3.72 × 𝑚0.75 + − 0.61 × − 1.89 × 𝑚 − 15.27 (10)
𝑅 𝑅
For finishing pigs, the heating load model for each pig in the house is shown in equation (11).
31.52 𝑡0
𝑄psi = 31.09 × 𝑛 − 5.42 × 𝑡0 − 52.01 × 𝑚0.75 + − 1.08 × − 13.06 × 𝑚 + 226 (11)
𝑅 𝑅
where 𝑄psi is the heating load for fattening pigs, W; 𝑛 is the feeding area of fattening pigs, m2
head-1; 𝑡0 is the temperature outside the house, ℃; 𝑚 is the weight of the fattening pig, kg; 𝑅 is
the thermal resistance of the enclosure, m2 ℃ W-1.
The correlation coefficients R2 of equations (10) and (11) are 0.966 and 0.963, respectively.
Analysis of the stocking density shows that the coefficient before the stocking density in two
equations is about 31. It means that for every increase of 0.1 m2 head-1 in the housing density, the
heating load of each fattening pig will decrease by about 3.1 W. Analysis of the temperature
outside the house shows that, taking the thermal resistance of the enclosure as 2 m2 ℃ W-1 as an
example, for every 1℃ decreases outside the nursery pig house, the heating load of each nursery
pig increases by about 3.3 W. For every 1℃ decreases outside the finishing pig house, the heating
load of each finishing pig increases by about 5.9 W. Analysis of the relationship between the
thermal resistance of the enclosure and the heating load of each fattening pig shows that the
∙ 54 ∙
reciprocal of the thermal resistance of the enclosure, that is, the heat transfer coefficient has a
linear relationship with the heating load. Since the sensible heat production of the fattening pigs
has a nonlinear relationship with weight, the relationship between the heating capacity of the
fattening pigs and weights of the fattening pigs is a nonlinear calculation model.
3.3. Differences in heating time and heating load of fattening pig houses in different regions
The heating time and load of the nursery and finishing pig houses in different regions were
obtained and are showed in figures 2 and 3.
a. Nursery pig house b. Finishing pig house

Figure 2. Heating time of year in fattening pig house.
a. Nursery pig house b. Finishing pig house

Figure 3. Heating load of year in fattening pig house.
As shown in Figure 2, due to the large difference in temperature in the climatic regions of the
five cities, the heating time of a heating season varies significantly. When the thermal resistance
of the building envelope is 1 m2 ℃ W-1, Changchun is the city with the longest heating time
among 5 cities. Nanning is the city with the shortest heating time. With the increase of the thermal
resistance of the building envelope, the heating time of the nursery and finishing pig houses in
various places is gradually shortened. And the effect of shortening the heating time gradually
decreases with the increase of the thermal resistance. When the thermal resistance of the envelope
exceeds 3 m2 ℃ W-1, increasing the thermal resistance of the building envelope has no obvious
effect on shortening the heating time. For nursery pig houses in Nanning, when the thermal
∙ 55 ∙
resistance of the building envelope exceeds 1 m2 ℃ W-1, it is almost unnecessary to heat the
house. For finishing pig houses in Wuhan and Guiyang, when the thermal resistance of the
building envelope is higher than 1 m2 ℃ W-1, almost no heating is needed in the house.
Analysis of the heating load of each nursery and finishing pig in the house (Figure 3) indicates
that the heating load changes with the thermal resistance of the building envelope. With the
increase of the thermal resistance, the heating load of each fattening pig gradually decreases, and
the magnitude of the decrease decreases with the increase of the thermal resistance. The heating
load in the five regions under the same thermal resistance of the building envelope follows the
order of Changchun > Beijing > Guiyang > Wuhan > Nanning. For Changchun area, when the
thermal resistance of the building envelope is greater than 3 m2 ℃ W-1, increasing the thermal
resistance has no obvious effect on reducing the heating load of each nursery pig and finishing
pig in the house. Therefore, the building thermal resistance of nursery and finishing pig houses in
Changchun area should not be greater than 3 m2 ℃ W-1. For nursery and finishing pig houses in
Beijing, when the thermal resistance of the building envelope is greater than 2 m2 ℃ W-1, the
heating load in the house will decrease as the thermal resistance increases, and the reduction is
small. Therefore, when the pig house is designed, the thermal resistance of the house in Beijing
should not be greater than 2 m2 ℃ W-1. The thermal resistance of the enclosure of the nursery pig
houses in Wuhan and Guiyang should not be higher than 2 m2 ℃ W-1, and the finishing pig house
should not be higher than 1 m2 ℃ W-1. For the nursery and finishing pig houses in Nanning, the
thermal resistance of the building envelope can be 0.5 m2 ℃ W-1, which can basically meet the
winter insulation requirements.
4. Conclusions
This research reveals the following observations and findings:
(1) In order to meet the requirements of temperature, humidity and CO2 concentration in pig
houses in winter, the minimum ventilation rate for fattening pigs of different weights can be
expressed as 𝑉min = 5.1667 · ln(𝑚) − 8.4354 with 𝑚 in kg of weight of fattening pigs.
(2) the heating load model for nursery pigs may be expressed as 𝑄𝑝𝑠𝑖 = 31.13 × 𝑛 − 3.048 ×
16.37 𝑡0
𝑡0 + 3.72 × 𝑚0.75 + − 0.61 × − 1.89 × 𝑚 − 15.27 . The heating load model for
𝑅 𝑅
31.52
finishing pigs may be expressed as 𝑄𝑝𝑠𝑖 = 31.09 × 𝑛 − 5.42 × 𝑡0 − 52.01 × 𝑚0.75 + −
𝑅
𝑡
1.08 × 0 − 13.06 × 𝑚 + 226, the correlation coefficient R2 of the two equations are 0.966 and
𝑅
0.963, respectively.
(3) Heating capacity of commercial pig houses in typical cities of different climatic regions
under the thermal resistance of the same building envelope follows the trend in Changchun >
Beijing > Guiyang > Wuhan > Nanning.
Acknowledgements
This work was supported by National Key Research and Development Plan of China -
Integration and demonstration of production technology, environment and engineering
technology for commercial pigs (2018YFD0501103).
References
Ji, H., Z. Y. Shao, Y. Liu, X. Zhang, C. Y. Niu, J. R. Guo, B. Xu, X. L. Zhan, J. X. Liu, J. F. Wang, 2021.
Effects of different degrees of cold stress on FIAF expression in pigs. Kafkas Universitesi Veteriner Fakultesi
Dergisi. 27(2): 135–140.
Scharf I, K. Wertheimer, J. L. Xin, T. Gilad, I. Goldenberg, A. Subach, 2019. Context-dependent effects
of cold stress on behavioral, physiological, and life-history traits of the red flour beetle. Insect Science. 26 (1):
142–153. https://doi.org/10.1111/1744-7917.12497
Quiniou, N., Noblet, J., van Milgen, J., Dubois, S., 2001. Influence of low ambient temperatures on heat
production and energy balance of single-housed growing pigs fed ad libitum: a comparison with group-housed
pigs. Animal Research. 50(4): 325–333. https://doi.org/10.1051/animres:2001134
∙ 56 ∙
Lopez, J., G. W. Jesse, B. A. Becker, M. R. Ellersieck, 1991. Effects of temperature on the performance
of finishing swine: II. Effects of a cold, diurnal temperature on average daily gain, feed intake, and feed
efficiency. Journal of Animal Science. 69(5):1850–1855. https://doi.org/10.2527/1991.6951850x
Parvu, M., A. T. Bogdan, H. Grosu, V. Andronie, I. C. Andronie, 2011. Effects of cold temperatures on
productive parameters at Mangalica and Large White pigs. In Animal hygiene and sustainable livestock
production. Vienna, Austria, July 3–7. Kofer, J.; Schobesberger, H. 2:685–687.
GBT 17824.3-2008, 2008. Environmental parameters and environmental management of large-scale pig
farms.
Mun, H.S., M. A. Dilawar, M. G. Jeong, D. Rathnayake, J. S. Won, K. W. Park, S. R. Lee, S. B. Ryu, C.
J. Yang, 2020. Effect of a heating system using a ground source geothermal heat pump on production
performance, Energy-Saving and Housing Environment of Pigs. Animals (Basel). Nov 9,10(11):2075.
https://doi.org/10.3390/ani10112075
Pandorfi, H., I. J. O. da. Silva, D. J. de. Moura, K. B. Sevegnani, 2005. Microclimate of creep for piglets
submitted to different systems of heating during the winter period. Revista Brasileira de Engenharia Agrícola
e Ambiental. 9(1): 99–106.
Li, X.S., X. L. Feng, B. M. Li, S. B. Bai, Z. C. Li, Z. L. Li, 2018. Determination of ventilation rate for
pregnant sows in Guangxi in winter. Foreign Animal Husbandry (Pig and Poultry). 38(03): 46–48.
Li, X.S., Wang, Q.C., Yan, Z.C., Chen, F.J., Yang, X.C., Dai, Y.Q., Tan, L.L., 2018. Study on the winter
ventilation of the new pig farm nursery in North China. China Pig Industry. 13(03): 63–65.
Wang, M.Z., J. J. Liu, J. H. Tian, Q. Li, Z. H. Wu, 2014. Energy saving and insulation effect of renovation
for existing swine houses in Beijing. Transactions of the Chinese Society of Agricultural Engineering. 30(05):
148–154.
Zhao, W.Y., 2017. Study on thermal insulation performance of exterior walls and heating energy of
nursery pig buildings. MS Thesis, College of Engineering, China Agricultural University, Beijing, China.
Wang, Y., B. M. Li, 2017. Analysis and experiment on thermal insulation performance of outer building
envelope for closed layer house in winter. Transactions of the Chinese Society of Agricultural Engineering.
33(07): 190–196.
Meteorological Data Room, Meteorological Information Center, China Meteorological Administration,
2005. Special meteorological data set for China's building thermal environment analysis.
Hu, X.Q., Niu, Q.Z., Fang, X.M., Zhang, J.J., 2008. Analysis of energy-saving rate of heating heat
metering in typical cities in different climate zones. National HVAC Refrigeration 2008 Academic Annual
Conference Data Collection. China Architectural Society HVAC Branch, China Refrigeration Society Air
Conditioning Heat Pump Professional Committee: China Refrigeration Society, 2008:1.
Fu, L.L., H. Z. Li, T. T. Liang, B. Zhou, Q. P. Chu, A. P. Schinckel, X. J. Yang, R. Q. Zhao, P. H. Li, R.
H. Huang, 2016. Stocking density affects welfare indicators of growing pigs of different group sizes after
regrouping. Applied Animal Behaviour Science.2016:174. https://doi.org/10.1016/j.applanim.2015.10.002
Huang, J.J., 2020. Study on Ventilation Rate and Cooling and Heating Load of the Nursery and Fattening
Pig House Based on Energy Balance. MS Thesis, College of Engineering, China Agricultural University,
Beijing, China.
Yan, P.S., R. Z. Li, 2011. Domestic Animal and Environmental Hygiene (4th Edition). Beijing: Higher
Education Press. (In Chinese).
∙ 57 ∙
A Survey and Analysis of Antibiotic Resistance of Escherichia Coli

in Feces from Chinese Pig Farms
Mingyang Li, Yansen Li, Chunmei Li *
College of Animal Science and Technology, Nanjing Agricultural University,
Nanjing, Jiangsu 210095, China
* Corresponding author. Email: chunmeili@njau.edu.cn
Abstract
Fecal carriage of bacteria is a major source of antibiotic resistance genes (ARGs) and a public
health risk. Still, the resistance of Escherichia coli (E. coli) in Chinese pig farms remains a main
gap in the available literature. Our goal is to conduct a meta-analysis of studies reporting antibiotic
resistance of fecal carriage of E. coli from pig farms in China, calculating the combined pooled
resistance rates and summarizing factors associated with it. We searched PubMed and Web of
Science for studies published in English up to February 28, 2021. We chose five major types of
antibiotics (ciprofloxacin, gentamicin, tetracycline, ampicillin, and florfenicol) to evaluate E. coli
resistance rate comprehensively. We used a random-effects model and Freeman-Tukey double
arcsine transformation to calculate the resistance rate and 95% confidence interval. Of 120
retrieved manuscripts, 16 studies (1985 E. coli isolates) were deemed eligible for our analysis.
The combined resistance rate of E. coli from fecal was 58.8% (95% CI: 45.3–71.7%) to
ciprofloxacin; 54.3% (95% CI: 35.3–72.6%) to gentamicin; 91.0% (95% CI: 83.1–96.7%) to
tetracycline; 81.4% (95% CI: 62.0–95.1%) to ampicillin; and 65.4% (95% CI: 33.9–90.9%) to
florfenicol. In conclusion, fecal carriage of E. coli in Chinese pig farms show high resistance to
ciprofloxacin, gentamicin, tetracycline, ampicillin, and florfenicol. Future confirmation of
antibiotic resistance and other information in pig farms should be more precise and depends on
local surveys. More attention should be given to the resistance changes after the withdraw of
antibiotic growth promoters in Chinese pig farms.
Keywords: antibiotic resistance, feces, Escherichia coli, pigs, China, meta-analysis.
1. Introduction
Antibiotics have been widely used in animal husbandry as growth promoters due to the
prophylactic effects on the animal disease. Simultaneously, resulting from long-term excessive
use of antibiotics in the pig industry, antibiotic resistance of Gram-negative bacteria emerged and
rapidly rose. The growing risk of antibiotic resistance among gram-negative bacteria is a global
problem (Kaye and Pogue, 2015). Domestic pigs, therefore, have become an extensive reservoir
of antibiotic resistant bacterial strains and genes, which can seriously endanger public health.
Escherichia coli (E. coli) is a Gram-negative bacterium commonly found in the intestine of
animals and humans. The rapid increase of antibiotic resistant intestinal pathogenic E. coli, such
as extended-spectrum beta-lactamases (ESBL)-producing E. coli, has raised the scientific
community concern(Cameron-Veas et al., 2015).
In the face of the wide occurrence of ARGs and their literally adverse impacts on human
health, the voice of restricting the use of feed is getting higher. Many countries begin to focus and
impose prohibitions on in-feed antibiotic growth promoters (AGPs). Sweden is the first country
to ban the use of AGPs in 1986. In 2014, the FDA (Food and Drug Administration) of the U.S.
claimed that the use of preventive antibiotics in animal feed would be banned within three years
to avoid antibiotic resistance problems. As the largest producer and user of antibiotics globally,
China has announced that the use of AGPs will be banned from 2020 (Announcement No. 194,
issued by the Ministry of Agriculture of China). However, withdrawal of AGPs from husbandry
can result in more challenges such as compromised gut health and rising gut diseases (Casewell
et al., 2003). Inevitably, conventional antibiotics are still needed in animal farming as a treatment
method, more importantly, because of the lack of alternative methods and their high cost in recent
∙ 58 ∙
years (Czaplewski et al., 2016). Considering the inevitable application of antibiotics, it is essential
for veterinary treatment decisions to thoroughly investigate the bacterial antibiotic resistance of
pig farms by which we can slow down the rapid growth of resistance.
Previous surveys showed that five major types of antibiotics having the highest monitoring
frequency are beta-lactams, tetracyclines, aminoglycosides, fluoroquinolones, and phenicols,
consistent with the common antimicrobials currently used in the world (Sjolund et al., 2016). By
investigating the sensitivity test of these antibiotics, the resistance characteristics of local E. coli
can be quantified as a scale to antibiotic resistance in pig farms. Announcement No. 2292, issued
by the Ministry of Agriculture of China in 2015, was regarded as the beginning of Chinese
restrictions on AGPs. In the current study, we investigate the changes before and after the period
when China begins to restrict AGPs use. Simultaneously, the resistance of E. coli in pig farms is
varied because of farm size and pig health (Xu et al., 2018). The increasing prevalence of ESBL
mechanisms of resistance among E. coli also needs to be focused (Kraemer et al., 2017).
This meta-analysis strictly followed the guidelines of the PRISMA [Preferred Reporting Items
for Systematic Reviews and Meta-analyses].
2.1. Search strategy
Relevant studies published in PubMed and Web of Science databases from their inception to
February 28, 2021, were collected and only the English language was applied (Table 1). The
reports of antibiotic resistant E. coli in Chinese pig farms were included. The antibiotic resistance
was defined by the drug-susceptibility test results (disk diffusion test, MIC (minimum inhibitory
concentration) test, which followed the guidelines of organizations such as CLSI (The Clinical
and Laboratory Standards Institute)). MESH (Medical Subject Headings) that reflect key
domains: antibiotics, swine, feces, Escherichia coli, and China. We also extended our search to
the cited references of retrieved articles to identify any additional relevant studies.
2.2. Selection criteria
A published article was eligible for inclusion in our meta-analysis under the conditions that
1) examined bacteria must include E. coli; 2) examined bacteria must be isolated from pig feces
or rectal swab; 3) all of E. coli isolates were subjected to antimicrobial susceptibility tests by the
standard disk diffusion method according to the CLSI guidelines; 4) sampling location must be in
China. Excluded from this analysis were as follows: 1) the samples of pig feces mixed with feces
of other animals or humans or Liquid feces; 2) studies included E. coli isolates with numbers
below 20; 3) did not report the resistance of antimicrobial drugs commonly assessed in three or
more studies; 4) included antibiotic interventions in the experimental design. 5) did not offer
sufficient data to create a contingency table.
2.3. Data extraction
Two of the authors independently screened the full text of each eligible study and extracted
the information from each included study. The assessment measures were as follows: study name
(the first author name), year of publication, region, sampling period, initial phase, health
condition, sample size, type of E. coli isolates, CLSI criteria, and data of antimicrobial
susceptibility test. The authors’ two datasets were combined, and disagreements during
independent data extraction were resolved by consensus at last.
We choose the most frequently occurring antibiotics (ciprofloxacin, gentamicin, tetracycline,
ampicillin, and florfenicol) in each of the five major antibiotics types to evaluate E. coli resistance.
We searched all CLSI standards in the original literature. The criteria for MIC breakpoints for E.
coli to these five antibiotics did not change according to the CLSI guidelines for different years.
∙ 59 ∙
Table 1. Characteristics of the included studies.

Health No. of CLSI
Reference Sampling period Type of E. coli
condition* E. coli criteria
(Gao et al., ESBL producing
2014.05–2014.07 healthy 29 CLSI 2013
2015b) E. coli
(Gao et al., ESBL producing
2013.04–2013.06 healthy 119 CLSI 2011
2015a) E. coli
(Hu et al., ESBL producing
2012–2013 healthy 31 CLSI 2012
2013) E. coli
(Lee and ESBL producing
2015.01–2015.12 sick 54 CLSI 2011
Yeh, 2017) E. coli
(Li et al., ESBL producing
2014.05–2014.12 healthy 31 CLSI 2011
2015) E. coli
(Tian et al.,
2006.09–2009.01 healthy 167 E. coli CLSI 2010
2012)
(Tian et al., sick and
2002–2007 212 E. coli CLSI 2006
2009) healthy
(Tong et al., CLSI 2013
2016.08–2017.09 healthy 152 E. coli
2018) and 2017
(Wang et CLSI 2004
2003–2008 sick 102 E. coli
al., 2011) and 2008
(Wu et al., NCCLS
2014.02–2014.07 healthy 30 E. coli
2019) 2004
post-
(Xu et al.,
2010.05–2013.08 weaning 458 E. coli CLSI 2012
2015)
diarrhea
(Zhang et ESBL producing
2015.05–2015.07 healthy 34 CLSI 2012
al., 2016) E. coli
(Zhang et
al., 2017)
(Zhang et
al., 2020)
(Zhou et al., NCCLS
2009–2010 healthy 400 E. coli
2012) 2003
(Zou et al., ESBL producing
2015 healthy 22 CLSI 2016
2019) E. coli
* We sent messages to relevant authors if no health condition was mentioned in the study.
NCCLS= National Committee for Clinical Laboratory Standards attendees.
We used the random-effects model to aggregate individual effect sizes and calculated the
summary resistance rate with 95% CI to assess the individual study proportion and pooled effects.
We used the Freeman-Tukey double arcsine transformation to minimize the effects of studies with
extremely high or low resistance rate estimates on the overall pooled estimates. Cochran’s Q
statistics (significance level of P ≤0.1) and degree of inconsistency (I2) were calculated to examine
the heterogeneity in all studies. The percentage of I2 values around 25% (I2≤25%), 50% (I2~50%),
and 75% (I2≥75%) were interpreted as low, medium, and high heterogeneity, respectively. We
further investigated potential heterogeneity sources by meta-regression analysis and subgroup
analysis, which attempts to relate differences in effect sizes to study characteristics. The studies
were classified by sampling period (before 2015 vs. after 2015), pig health (healthy vs. sick), type
of E. coli (ESBL E. coli vs. non-ESBL E. coli), and sample size (small: less than 100 isolates vs.
large: greater than 100 isolates) according to potentially relevant characteristics. We also
∙ 60 ∙
performed sensitivity analyses to identify the study (or studies) which contributed as the primary
source of the heterogeneity and recalculated the pooled estimates and heterogeneity after excluded
the specified study.
Meta-analysis was carried out in Stata 12.0 (Stata Corp., USA).
3. Results
One hundred and twenty studies were selected in the initial searches. We included 16 studies
(with 1985 E. coli isolates) and extracted the data for our meta-analysis. We extracted studies that
reported over three antibiotics for antimicrobial susceptibility testing. The most frequent of five
major types of antibiotics are 15 studies reported on ciprofloxacin, 14 on gentamicin and
tetracycline, 12 on ampicillin, and 8 on florfenicol.
In individual variable meta-regression analysis, sampling period, health condition, sample
size, and type of E. coli had no significant relation (P>0.05) to the heterogeneity of antibiotics
resistance. In Table 2, for three antibiotics (ciprofloxacin, gentamicin, and ampicillin), older
studies (before 2015) had a higher resistance rate than did newer studies (after 2015). Sick pigs
were more likely to have antibiotic resistant E. coli isolates than healthy pigs. All E. coli isolates
from the small sample size were higher than those from the large sample size. The resistance rate
of ESBL-producing E. coli strains was taller than non-ESBL-producing E. col strains. Sensitivity
analysis showed that the resistance rates did not significantly differ after excluding the intense
study, indicating no risk of bias in original studies.
Figure 1 presents the overall resistance rates of E. coli isolates to five antibiotics. Specifically,
the highest antibiotic resistance of E. coli was to tetracycline 91.0% (95%CI: 83.1–96.7%),
heterogeneity was substantial (I2= 95.434%, P= 0.000); followed by resistance to ampicillin
81.4% (95%CI: 62.0–95.1%) with high heterogeneity (I2= 98.727%, P= 0.000); resistance to
florfenicol was 65.4% (95%CI: 33.9–90.9%) with substantial heterogeneity (I2= 99.168%, P=
0.000); resistance to ciprofloxacin was 58.8% (95%CI: 45.3–71.7%) with high heterogeneity (I2=
96.768%, P= 0.000); and resistance to gentamicin was 54.3% (95%CI: 35.3–72.6%) with
substantial heterogeneity (I2= 98.435%, P= 0.000).
4. Discussion
This study aims to quantitatively define the antibiotic resistance rate of E. coli from pig farms
in China. As shown in our results, E. coli isolates showed various resistance to five majorly used
antibiotics and had the highest resistance rate to tetracycline. These findings are in accordance
with antibiotic resistance of bacteria from livestock environment (Zhou et al., 2017). Despite the
substantial heterogeneity coming from physiological, environmental, and economic factors, we
can obtain a globally accurate and reliable conclusion by choosing a reasonable model and data
transformation method following the basic requirements of meta-analysis.
Our subgroup analysis exhibited that the resistance to antibiotics was closely related to the
quantity and health condition of pigs. The resistances of ciprofloxacin, gentamicin, and ampicillin
decreased after 2015. Currently, they were still relatively high compared with the FDA data during
the same period (https://www.fda.gov/animal-veterinary/national-antimicrobial-resistance-
monitoring-system/narms-now-integrated-data). This result may be explained by the fact that the
use of antibiotics by farmers in treating food-producing animals, which could indicate the trend
of antibiotics use (Wang et al., 2020). According to these data, we can infer that there will be a
lower trend after 2020 upon the withdraw of AGPs. The resistance rates were higher in E. coli
isolates from diseased pigs than from healthy pigs to all antibiotics. The high heterogeneity
detected in gentamicin (I2= 98.435%, Pheterogeneity= 0) was majorly attributed to the healthy
subgroup (I2= 97.877%, Pheterogeneity= 0) rather than the sick subgroup (I2= 0%, Pheterogeneity= 0.49).
Hence, it is conceivable that gentamicin is used more for treatment than for prevention in Chinese
pig farms. Furthermore, as more antibiotics were used to treat diseased pigs, improper treatments
(e.g., overuse of antibiotics) may accelerate the spread of antibiotic genes. It is necessary to
∙ 61 ∙
monitor these five types of antibiotics used for veterinary treatment in pig farms after the cessation
of AGPs.
Table 2. Characteristics of the included studies.
Antibiotics& Resistance estimate Heterogeneity test
No. of studies
subgroups (95% CI) PHeterogeneity* I2 (%)
CIP
Before 2015 9 0.658 (0.503, 0.797) 0 96.703
After 2015 6 0.481 (0.244, 0.721) 0 95.939
Healthy 12 0.537 (0.384, 0.686) 0 95.796
Sick 3 0.766 (0.583, 0.910) NA NA
Small sample size 9 0.629 (0.436, 0.803) 0 92.591
Large sample size 6 0.533 (0.327, 0.733) 0 98.458
ESBL 7 0.638 (0.402, 0.845) 0 93.528
non-ESBL 8 0.548 (0.369, 0.721) 0 97.932
GEN
Before 2015 9 0.615 (0.425, 0.790) 0 97.984
After 2015 5 0.410 (0.067, 0.811) 0 98.385
Healthy 10 0.473 (0.258, 0.692) 0 97.877
Sick 4 0.723 (0.691, 0.753) 0.49 0
ESBL 6 0.691 (0.521, 0.839) 0 86.155
non-ESBL 8 0.433 (0.194, 0.690) 0 99.067
TET
Before 2015 8 0.892 (0.780, 0.969) 0 96.379
After 2015 6 0.932 (0.794, 1.000) 0 93.851
Healthy 11 0.905 (0.804, 0.974) 0 94.89
Sick 3 0.926 (0.775, 0.999) NA NA
Small sample size 8 0.967 (0.927, 0.993) 0.033 54.025
ESBL 6 0.973 (0.919, 1.000) 0.012 65.858
non-ESBL 8 0.852 (0.732, 0.942) 0 96.933
AMP
Before 2015 7 0.901 (0.722, 0.996) 0 98.514
After 2015 5 0.661 (0.208, 0.983) 0 98.863
Healthy 9 0.728 (0.479, 0.919) 0 98.427
Sick 3 0.973 (0.953, 0.989) NA NA
ESBL 4 1.000 (0.991, 1.000) 0.983 0
non-ESBL 8 0.648 (0.388, 0.869) 0 99.038
FFC
Before 2015 4 0.585 (0.104, 0.978) 0 99.58
After 2015 4 0.719 (0.409, 0.947) 0 96.788
Healthy 5 0.566 (0.174, 0.915) 0 98.991
Sick 3 0.784 (0.597, 0.926) NA NA
Small sample size 3 0.889 (0.819, 0.946) NA NA
ESBL 3 0.889 (0.819, 0.946) NA NA
non-ESBL 5 0.504 (0.139, 0.867) 0 99.477
* Significance level PHeterogeneity <0.05; NA= not available; CI= confidence interval.
∙ 62 ∙
∙ 63 ∙
Figure 1. Meta-analysis of differences in resistance of E. coli. A: ciprofloxacin; B: gentamicin;

C: tetracycline; D: ampicillin; E: florfenicol. n: sample size; CI: confidence interval. The small
squares represent study-specific estimates (size of the square reflects the study-specific
statistical weight), and the horizontal lines represent 95% CI. The open diamonds represent
summary estimates with corresponding 95% CI.
The differences in E. coli resistance were also associated with sample size and E. coli types.
Resistance from a small sample size tended to be higher than from a large sample size, suggesting
a more extensive scale E. coli resistance monitor system to enable more robust, valid, and reliable
resistance rates. ESBL-producing E. coli isolates have reported a higher resistance level than non-
ESBL-producing E. coli isolates in five antibiotics, especially in the subgroup analysis of
ampicillin (I2= 0, Pheterogeneity= 0.983). These results were consistent with other studies conducted
(Adenipekun et al., 2015) and corroborated that most ESBL-producing E. coli isolates could be
classified as multidrug-resistant (Magiorakos et al., 2012). The ESBL genes are often located on
mobile genetic elements and associated with other resistance genes, suggesting that the
application of multiple types of antibiotics, not only beta-lactams, could co-select for ESBL genes
and lead to the persistence of ESBL-producing E. coli in the environment (Allen, 2014).
∙ 64 ∙
Our meta-regression analysis showed that the influential factors of E. coli resistance in pig
farms are more complex as shown by the heterogeneity. However, the downgraded level of
heterogeneity after subgroup analysis demonstrated that these subgroups could still explain the
heterogeneity among studies. Thus, our summary resistance rates and the similarity and
differences among the subgroups could reflect the actual situation in the probability of resistance
in the fecal carriage of E. coli on the pig farm. The sensitivity analyses showed no significant
publication bias in the analyzed studies, which meant that our meta-analysis was plausible even
though it was based on a relatively small number (n = 16) of precious studies. Despite the
limitations of those studies, a pooled mean resistance rate of fecal carriage of E. coli was attempted
based on excluding different experimental designs and data treatments in every study.
The main implication of our results is that the identification and antibiotic susceptibility tests
for fecal carriage of E. coli resistance should be integral to the biosecurity assessment and
veterinary prescribing reference for pig farms. Due to the tremendous livestock size, animal
husbandry is the most extensive application sector for antibiotics in China, accounting for more
than half of total antibiotic consumption, which could discharge antibiotics residues into the
nearby environment. The association between overuse or wrongful use of antibiotics on animal
husbandry and bacterial resistance in animals and farmers has been convinced. The high
antibiotics resistance reported in this meta-analysis should be concerned.
Heterogeneity among studies results from random factors, which are usually neglected in such
research. For example, growth phase, temperature, and farm size are likely to play a role in
resistance cultivation. The enrichment degree of enrichment of ARGs may differ in different
growth periods of pigs. Evidence from an experimental investigation found that the carriage
prevalence of CTX-M-producing E. coli during preweaning was higher than during postweaning
and finishing pig (Hansen et al., 2014). Meanwhile, the temperature is also one of the most
influential factors to bacterial growth rates and may accelerate horizontal gene transfer
(MacFadden et al., 2018). It would be interesting to look into possible variations by farm size
because the farm size has different hygiene requirements for safe animal production, affecting the
total use of antibiotics and the selection intensity (Olesen et al., 2018). The subgroup analysis did
not include location because of uneven studies (half of 16 studies were from Shandong province,
China). However, previous reports show differences in ARGs abundance among provincial
regions (Li et al., 2020). Further research should monitor more regions, especially the provinces
with many pig farms in China, and describe more detailed farm size information.
One limitation is that our meta-analysis is based on cross-sectional designs, and therefore
inferences about causality cannot be made. However, several included studies reported the use of
antibiotics on pig farms (Wu et al., 2019). We did not identify enough longitudinal studies, which
would be practically tricky but necessary for future research because they would provide
information on the development of antibiotics resistance, especially risk factors and mediators.
5. Conclusions
This meta-analysis has calculated the resistance rate for fecal carriage of E. coli was 58.8%
(95%CI: 45.3–71.7%) to ciprofloxacin; 54.3% (95%CI: 35.3–72.6%) to gentamicin; 91.0%
(95%CI: 83.1–96.7%) to tetracycline; 81.4% (95%CI: 62.0–95.1%) to ampicillin; and 65.4%
(95%CI: 33.9–90.9%) to florfenicol. Further research is needed to identify origins for the
heterogeneity in E. coli resistance rates. Comprehensive information about pig farms, antibiotic
usage, pig health status, and trends of resistance rates among pig fecal carriage of E. coli is an
essential component of veterinary treatment decisions. In this sense, our meta-analyses may help
inform biosecurity assessment of pig farms and indicate that full-time monitoring is needed in the
future to assess the subsequent effects after the withdrawal of antibiotic growth promoters on
bacterial resistance in food-animal production.
∙ 65 ∙
Acknowledgements
This study was supported by the National Key Research and Development Program of China
(2016YFD0500505) and the National Natural Science Foundation of China (No. 31772648).
References
Allen, H.K., 2014. Antibiotic resistance gene discovery in food-producing animals. Curr Opin Microbiol.
19, 25–29. http://doi.org/10.1016/j.mib.2014.06.001.
Cameron-Veas, K., M. Sola-Gines, M.A. Moreno, L. Fraile, L. Migura-Garcia, 2015. Impact of the use
of beta-lactam antimicrobials on the emergence of Escherichia coli Isolates resistant to cephalosporins under
standard pig-rearing conditions. Appl Environ Microb. 81 (5), 1773–1778. http://doi.org/10.1128/aem.03916-
14.
Casewell, M., C. Friis, E. Marco, P. McMullin, I. Phillips, 2003. The European ban on growth-promoting
antibiotics and emerging consequences for human and animal health. J Antimicrob Chemoth. 52 (2), 159–
161. http://doi.org/10.1093/jac/dkg313.
Czaplewski, L., R. Bax, M. Clokie, M. Dawson, H. Fairhead, V.A. Fischetti, S. Foster, B.F. Gilmore,
R.E.W. Hancock, D. Harper, I.R. Henderson, K. Hilpert, B.V. Jones, A. Kadioglu, D. Knowles, S. Olafsdottir,
D. Payne, S. Projan, S. Shaunak, J. Silverman, C.M. Thomas, T.J. Trust, P. Warn, J.H. Rex, 2016. Alternatives
to antibiotics-a pipeline portfolio review. Lancet Infect Dis. 16 (2), 239–251. http://doi.org/10.1016/s1473-
3099(15)00466-1.
Gao, L., Y. Tan, X. Zhang, J. Hu, Z. Miao, L. Wei, T. Chai, 2015a. Emissions of Escherichia coli carrying
extended-spectrum beta-lactamase resistance from pig farms to the surrounding environment. Int J Env Res
Pub He. 12 (4), 4203–4213. http://doi.org/10.3390/ijerph120404203.
Gao, L.L., J.Q. Hu, X.D. Zhang, L.M. Wei, S. Li, Z.M. Miao, T.J. Chai, 2015b. Application of swine
manure on agricultural fields contributes to extended-spectrum beta-lactamase-producing Escherichia coli
spread in Tai'an, China. Front Microbiol. 6. http://doi.org/10.3389/fmicb.7015.00313.
Hansen, K.H., V. Bortolaia, P. Damborg, L. Guardabassi, 2014. Strain diversity of CTX-M-producing
Enterobacteriaceae in individual pigs: insights into the dynamics of shedding during the production cycle.
Appl Environ Microbio. 80 (21), 6620–6626. http://doi.org/10.1128/aem.01730-14.
Hu, Y.Y., J.C. Cai, H.W. Zhou, D. Chi, X.F. Zhang, W.L. Chen, R. Zhang, G.X. Chen, 2013. Molecular
typing of CTX-M-producing Escherichia coli isolates from environmental water, swine feces, specimens from
healthy humans, and human patients. Appl Environ Microb. 79 (19), 5988–5996. http://doi.org/10.1128/
aem.01740-13.
Kaye, K.S., J.M. Pogue, 2015. Infections Caused by Resistant Gram-Negative Bacteria: Epidemiology
and Management. Pharmacotherapy. 35 (10), 949–962. http://doi.org/10.1002/phar.1636.
Kraemer, J.G., J. Pires, M. Kueffer, E. Semaani, A. Endimiani, M. Hilty, A. Oppliger, 2017. Prevalence
of extended-spectrum beta-lactamase-producing Enterobacteriaceae and methicillin-resistant Staphylococcus
aureus in pig farms in Switzerland. Sci Total Environ. 603, 401–405. http://doi.org/10.1016/j.scitotenv.
2017.06.110.
Lee, W.C., K.S. Yeh, 2017. Characteristics of extended-spectrum beta-lactamase-producing Escherichia
coli isolated from fecal samples of piglets with diarrhea in central and southern Taiwan in 2015. Bmc Vet Res.
13 (1). http://doi.org/10.1186/s12917-017-0986-7.
Li, S., W.G. Song, Y.F. Zhou, Y.J. Tang, Y.X. Gao, Z.M. Miao, 2015. Spread of extended-spectrum beta-
lactamase-producing Escherichia coli from a swine farm to the receiving river. Environ Sci Pollut R. 22 (17),
13033–13037. http://doi.org/10.1007/s11356-015-4575-7.
Li, Y., X.R. Yang, H.N. Zhang, H.Y. Jia, X.G. Liu, B. Yu, Y.C. Zeng, Y. Zhang, X.Y. Pei, D.J. Yang,
2020. Prevalence and antimicrobial susceptibility of Salmonella in the commercial eggs in China. Int J Food
Microbiol. 325. http://doi.org/10.1016/j.ijfoodmicro.2020.108623.
MacFadden, D.R., S.F. McGough, D. Fisman, M. Santillana, J.S. Brownstein, 2018. Antibiotic resistance
increases with local temperature. Nat Clim Change. 8 (6), 510–510. http://doi.org/10.1038/s41558-018-0161-
6.
Olesen, S.W., M.L. Barnett, D.R. MacFadden, J.S. Brownstein, S. Hernandez-Diaz, M. Lipsitch, Y.H.
Grad, 2018. The distribution of antibiotic use and its association with antibiotic resistance. Elife. 7.
http://doi.org/10.7554/eLife.39435.
Sjolund, M., M. Postma, L. Collineau, S. Loesken, A. Backhans, C. Belloc, U. Emanuelson, E.G. Beilage,
K. Stark, J. Dewulf, M. Consortium, 2016. Quantitative and qualitative antimicrobial usage patterns in farrow-
to-finish pig herds in Belgium, France, Germany and Sweden. Prev Vet Med. 130, 41–50. http://doi.org/
10.1016/j.prevetmed.2016.06.003.
∙ 66 ∙
Tian, G.B., H.N. Wang, L.K. Zou, J.N. Tang, Y.W. Zhao, M.Y. Ye, J.Y. Tang, Y. Zhang, A.Y. Zhang,
X. Yang, C.W. Xu, Y.J. Fu, 2009. Detection of CTX-M-15, CTX-M-22, and SHV-2 extended-spectrum beta-
lactamases (ESBLs) in Escherichia coli fecal-sample Isolates from pig farms in China. Foodborne Pathog Dis.
6 (3), 297–304. http://doi.org/10.1089/fpd.2008.0164.
Tian, G.B., H.N. Wang, A.Y. Zhang, Y. Zhang, W.Q. Fan, C.W. Xu, B. Zeng, Z.B. Guan, L.K. Zou,
2012. Detection of clinically important beta-lactamases in commensal Escherichia coli of human and swine
origin in western China. J Med Microbiol. 61 (2), 233–238. http://doi.org/10.1099/jmm.0.036806-0.
Tong, H.X., J.X. Liu, X.H. Yao, H.Y. Jia, J.C. Wei, D.H. Shao, K. Liu, Y.F. Qiu, Z.Y. Ma, B.B. Li, 2018.
High carriage rate of mcr-1 and antimicrobial resistance profiles of mcr-1-positive Escherichia coli isolates
in swine faecal samples collected from eighteen provinces in China. Vet Microbiol. 225, 53–57. http://doi.org/
10.1016/j.vetmic.2018.09.018.
Wang, X.M., X.P. Liao, S.G. Liu, W.J. Zhang, H.X. Jiang, M.J. Zhang, H.Q. Zhu, Y. Sun, J. Sun, A.X.
Li, Y.H. Liu, 2011. Serotypes, virulence genes, and antimicrobial susceptibility of Escherichia coli isolates
from pigs. Foodborne Pathog Dis. 8 (6), 687–692. http://doi.org/10.1089/fpd.2010.0739.
Wang, Y., C. Xu, R. Zhang, Y.Q. Chen, Y.B. Shen, F.P. Hu, D.J. Liu, J.Y. Lu, Y. Guo, X. Xia, J.Y. Jiang,
X.Y. Wang, Y.L. Fu, L. Yang, J.Y. Wang, J. Li, C. Cai, D.D. Yin, J. Che, R. Fan, Y.Q. Wang, Y. Qing, Y.
Li, K. Liao, H. Chen, M.X. Zou, L. Liang, J. Tang, Z.Q. Shen, S.L. Wang, X.R. Yang, C.M. Wu, S.X. Xu,
T.R. Walsh, J.Z. Shen, 2020. Changes in colistin resistance and mcr-1 abundance in Escherichia coli of animal
and human origins following the ban of colistin-positive additives in China: an epidemiological comparative
study. Lancet Infect Dis. 20 (10), 1161–1171. http://doi.org/10.1016/s1473-3099(20)30149-3.
Wu, B., Q. Qi, X.D. Zhang, Y.M. Cai, G.L. Yu, J. Lv, L.L. Gao, L.M. Wei, T.J. Chai, 2019. Dissemination
of Escherichia coli carrying plasmid-mediated quinolone resistance (PMQR) genes from swine farms to
surroundings. Sci Total Environ. 665, 33–40. http://doi.org/10.1016/j.scitotenv.2019.01.272.
Xu, G.F., W. An, H.D. Wang, X.Y. Zhang, 2015. Prevalence and characteristics of extended-spectrum
beta-lactamase genes in Escherichia coli isolated from piglets with post-weaning diarrhea in Heilongjiang
province, China. Front Microbiol. 6. http://doi.org/10.3389/fmicb.2015.01103.
Xu, Y.C., L.L. Zhong, S. Srinivas, J. Sun, M. Huang, D.L. Paterson, S. Lei, J.X. Lin, X. Li, Z.C. Tang,
S.Y. Feng, C. Shen, G.B. Tian, Y.J. Feng, 2018. Spread of MCR-3 colistin resistance in China: an
epidemiological, genomic and mechanistic study. Ebiomedicine. 34, 139–157. http://doi.org/10.1016/
j.ebiom.2018.07.027.
Zhang, H., Z. Zhai, Q. Li, L. Liu, S. Guo, Q. Li, L. Yang, C. Ye, W. Chang, J. Zhai, 2016. Characterization
of extended-spectrum beta-lactamase-producing Escherichia coli Isolates from pigs and farm Workers. J Food
Protect. 79 (9), 1630–1634. http://doi.org/10.4315/0362-028x.Jfp-16-093.
Zhang, H., X. Zhao, X. Wang, W. Chang, 2017. Prevalence and antimicrobial resistance profiles of
Escherichia coli isolated from free-range pigs. Journal of Infection in Developing Countries. 11 (8), 652–655.
http://doi.org/10.3855/jidc.9269.
Zhang, L., H. Li, J. Gao, J. Gao, D. Wei, Y. Qi, 2020. Identification of drug-resistant phenotypes and
resistance genes in Enterococcus faecalis isolates from animal feces originating in Xinjiang, People's Republic
of China. Canadian Journal of Animal Science. 100 (4), 674–682. http://doi.org/10.1139/cjas-2018-0161.
Zhou, H.C., L. Liu, Z.F. Tian, W.J. Zhang, 2012. Prevalence and patterns of antimicrobial resistant of
fecal Escherichia coli isolates among pigs on eight farrow-to-finish farms and their production environments
in Shandong province, China. Turk J Vet Anim Sci. 36 (6), 718–722. http://doi.org/10.3906/vet-1008-446.
Zhou, Y.T., L.L. Niu, S.Y. Zhu, H.J. Lu, W.P. Liu, 2017. Occurrence, abundance, and distribution of
sulfonamide and tetracycline resistance genes in agricultural soils across China. Sci Total Environ. 599, 1977–
1983. http://doi.org/10.1016/j.scitotenv.2017.05.152.
Zou, H.Y., B.W. Zheng, M.L. Sun, J. Ottoson, Y.B. Li, B. Berglund, X.H. Chi, X. Ji, X.W. Li, C.S.
Lundborg, L.E. Nilsson, 2019. Evaluating dissemination mechanisms of antibiotic-resistant bacteria in rural
environments in China by using CTX-M-producing Escherichia coli as an indicator. Microb Drug Resist. 25
(7), 975–984. http://doi.org/10.1089/mdr.2018.0431.
∙ 67 ∙
Effect of Ventilation Fans and Types of Partition on the

Physiological Parameters of Dairy Pre-Weaned
Individual-Housed Calves in a Calf Barn
Wanying Zhao a,b, Zhengxiang Shi a,b,*, Xinyi Du a,b, Huiyuan Guan a,b, Hao Li a,b
a
b
Abstract
Calves raised in barns are kept in individual stalls separated by solid or mesh partitions.
Ventilation fans blowing air perpendicular to these stalls only provide the optimal airflow to the
first few calves, especially for calves separated by solid partitions. In warm seasons, to explore
the effects of airflow speeds and types of partition on the health of pre-weaned calves, 24 Holstein
female pre-weaned calves were divided into four groups as follows: (1) six calves were separated
by solid partitions and received low-speed airflow (SP-LA); (2) six calves were separated by mesh
partitions and received low-speed airflow (MP-LA); (3) six calves were separated by solid
partitions and received high-speed airflow (SP-HA); (4) six calves were separated by mesh
partitions and received high-speed airflow (MP-HA). The results showed that the average airflow
speeds of four groups were 0.13–0.14 m s-1, 0.61–0.65 m s-1, 2.51–2.60 m s-1 and 2.22–2.80
m s-1, respectively. The rectal temperatures and respiratory rates of calves of SP-LA were
significantly higher than other groups. Calves of MP-LA and MP-HA registered significantly
higher IgA and IgG than calves of SP-LA and SP-HA (P<0.05). Calves of SP-LA registered
significantly lower IgM and higher TNF-α and IL-6 than other groups. We concluded that in warm
seasons, with the axial-flow fan turned on, mesh partitions did not have negative impact on the
health of the calves. The low-speed airflow (<0.2 m s-1) in the calves’ pens with two solid
partitions increases the risk of calf disease.
Keywords: Pre-weaned calves, mesh partitions, solid partitions, airflow speed
1. Introduction
High calf morbidity and mortality rates are serious issues (Urie et al., 2018). Apart from the
cases of failure of the insufficient maturation of the immune system in calves, functions of the
calf immune system can be, at least to some extent, modulated by environmental factors in barn
(Cobb et al., 2014). For calves, the reported upper critical thermoneutral zone ranges from 26 to
32 C (Spain and Spires, 1996) and the physiological range of calf body temperature lies between
38.1 to 39.2 C with diurnal variation (Piccione et al., 2003). Housing calves indoor can protect
the cows from direct sunlight in hot seasons (Bruoucek et al., 2008) and help to maintain the body
core temperature of the cows in cold seasons (Pempek et al., 2016). The calves reared in individual
pens could receive precise management separately based on their situations, and the solid
partitions between pens prevented the nose-to-nose contact, reducing the prevalence of infectious
diseases (Gulliksen et al., 2009). Logo et al. (2006) suggested that ideal pen had solid panels on
2 sides to separate each calf from the next, mesh panels in front and deep loose bedding during
months when temperatures fall below the thermoneutral zone of the calf. Therefore, some farmers
would adopt individual pens with 2 solid partitions to rear pre-weaned calves in every season.
Besides, in warm seasons, to against heat stress, fans increase the speed of airflow and
therefore can more quickly remove convective heat from the surface of a calf (Spiers et al., 2018).
However, if the fans are arranged to propel air in a direction perpendicular to the compartments
divided by solid partitions, the airflow will only benefit those calves in the first few pens
(Nordlund and Halbach, 2019). We have found that, in summer, when the calves (0.5-month-old
∙ 68 ∙
to 1.5-month-old) were located far from fans, <0.2 m s-1 airflow speed may impair the immune
functions of calves (Zhao et al., 2021). There is a trade-off between airflow speed in calves’
occupied zone and contact between individual-housed calves in warm seasons. Calves were reared
in pens with mesh partition on four sides, which with higher social contact events and airflow
speed between calves reared in pens with solid partition.
The objective of this study was to determine the effect of axial-flow fan and types of partition
on calves. We hypothesized that calves raised in pens with mesh partitions on four sides would
perform most optimally in terms of thermoregulation, health responses when compared with
calves raised in other pens.
2.1. Experiment house and calf management
Field experiments were carried out during the warm season on a commercial pre-weaned calf
house (3234’N, 12054’E) located in Yancheng, Jiangsu, China between May 23 and June 18,
2020. The calf house was 26 m wide, 240 m long, orientated in a north-south direction, and 8.7 m
high at its peak with approximately 500 pre-weaned calves. The calf house had a natural-
ventilation system augmented with axial fans (Shanghai Terrui Mechanical Equipment Co., Ltd.,
China). The fans (diameter = 0.97 m) are usually operated at full speed (25430 m3 h-1, 410 W, 720
rpm).
Each fan was installed on a column (located at 12 m intervals) at a height of 2.2 m and angled
downward from the vertical at about 38°. The fan system was manually controlled and switched
on when the ambient temperature inside the barn reached 20 °C. They were separated from their
dams and placed inside individual pens (1.2  1.2  1.5 m), and each pen was supplied with 5 cm
of sawdust to serve as bedding. Two detachable planks (solid penal or wire-mesh penal) formed
the walls that separated the pens, and to sustain each calf, two stainless steel buckets were used,
one to supply concentrated feed and the other to supply milk and water. All the calves were
provided milk manually at 07:00, 14:00, and 19:00 every day.
Twenty-four Holstein female calves (ages range from 3–11 days at the beginning of the test)
were selected and divided into four groups: (1) six calves were separated by two solid partitions
and received low-speed airflow (SP-LA); (2) six calves were separated by mesh partitions and
received low-speed airflow (MP-LA); (3) six calves were separated by two solid partitions and
received high-speed airflow (SP-HA); (4) six calves were separated by mesh partitions and
received high-speed airflow (MP-HA). The total serum protein concentrations of all born calves
were tested on day 3 and found to be over 5.2 g L-1, which meant successful passive immunization
from colostrum.
Eight digital data loggers ((HOBO U23 Pro v2; Onset Computer Co. Ltd., Bourne, MA, USA)
were installed inside the calf pens to measure the inside air temperature and humidity located
1.0m above the floor (Figure 1). The source of air temperature and humidity outside was China
Meteorological Data Service Center (http://data.cma.cn/en). The air velocities at each pen were
measured using a multifunctional heat-wire anemometer (9565-P with 964 probe, TSI Inc.,
Shoreview, MN, USA; 3% of reading, 0.01 m s-1 resolution), and each pen were detected five
times in five days. The airflow in center of the pens was monitored at heights of 0.4 m and 1.0 m
above the floor (1.0 m is the average height of a standing calf’s back and 0.4 m is the height of an
average calf’s back when lying down; both heights are typically monitored when measuring the
airflow inside in a calf pen).
2.3. Physical parameters and blood collection and measurement
Respiratory rate and rectal temperature were measured between 13:00–15:00 daily.
Respiratory rate was measured by counting flank movements in the 30s and multiplying by 2.
Rectal temperature was measured by a digital thermometer. The weight of calves was measured
∙ 69 ∙
at birth and at the end of the test, and average daily gain (ADG) was calculated. The morbidity
rate of bovine respiratory disease (BRD) and diarrhea of calves were recorded. Blood samples
from all study calves (10 mL in a vacuum blood collector without anticoagulant) were collected,
using jugular venipuncture (May 22, May 29, June 5, June 12, 2020). The samples were kept at
an incubator (35C) until the blood clotted, then the serum was separated by centrifugation at
4000  g for 5 min and frozen at -20C before the concentrations of immunity parameters
determination. The concentrations of IgA, IgG, IgM, TNF-, and IL-6 were detected using
commercially available Enzyme-Linked Immunosorbent Assay (ELISA) kits (Jianglai Biological
Technology Co. Ltd., Shanghai, China).
Figure 1. Schematic diagram of experimental design groups.

2.4 Statistical analysis
All analyzes were performed using SPSS software version 17.0 (IBM Corp., Armonk, NY,
USA). The descriptive statistics were expressed as mean  standard error. Differences between
treatments were deemed statistically significant if the associated P-value <0.05. The effect of IgA,
IgG, IgM, TNF-, and IL-6 content on the airflow speed and the age of calves were analyzed with
a linear mixed model. The statical model used was: Yijk = μ + ASi +WKj +Rk + ASi  WKj + ijK,
where Yijk = parameters investigated; μ = model constant; ASi = effect of airflow speed (i = 1 to
4); WKj = effect of age (week) (j = 1 to 5); Rk = replicate; ASi  WKj = effect produced by the
interaction of airflow speed and age; ijk = the residual error term.
3. Results
3.1. Environmental conditions and rectal temperature, respiratory rate, and ADG of tested calves
During the experimental period, the average daily temperature and relative humidity were
ranging from 18.57 to 28.37 C and 61.93% to 97.92% in calf house, meanwhile, the value of
average daily temperature and relative humidity fluctuated between 17.32 to 30.46 C and 48.46%
to 96.08% outside. Average daily THI of calf house was similar to those of outside.
According to the data, the airflow speeds did not differ significantly during the five days,
which meant each group was essentially subjected to a steady airflow speed of pens. And the
average data for airflow speed were shown in Table 1. The group with highest airflow speed was
∙ 70 ∙
MP-HA, followed by SP-HA, MP-LA, SP-LA (P<0.05). The calves of group SP-LA had highest
rectal temperature, and the calves of group SP-NF had lowest rectal temperature. The calves of
group SP-LA has highest respiratory rate (P<0.05). There was no significant difference in ADG
among these groups, but with the airflow speed increasing, the ADG showed an increasing trend.
Table 1. Airflow speeds, rectal temperature, respiratory rate, and ADG (MeanSEM) of groups.
Airflow speed (m s-1) Rectal temperature Respiratory rate ADG
Group
0.4 m 1.0 m (C) (breaths min-1) (kg d-1)
SP-LA 0.13±0.01d 0.14±0.04d 39.16±0.04a 59.89±1.57a 0.77±0.11
MP-LA 0.65±0.05c 0.61±0.04c 39.13±0.04ab 54.06±1.76b 0.83±0.07
SP-HA 2.51±0.50b 2.60±0.11b 38.99±0.04b 51.22±1.62b 0.84±0.67
MP-HA 2.22±0.10a 2.80±0.06a 39.14±0.04ab 53.64±1.71b 0.98±0.01
a–d
Note: Values having different letters within the same column are significantly different
(P<0.05).
3.2. Immunity parameters and health status of calves
The total number of illnesses and duration of treatment of calves with mesh partitions were
lower than calves with solid partitions (Table 2). Calves of MP-LA had the lowest number of
illnesses. The total treatment days of calves with mesh partitions were lower than that of calves
with solid partitions, and the total number of sicks and treatment days of calves of SP-LA were
higher than other groups.
Table 2. Disease and treatment records of calves during the experimental.
The age of sick Duration of each Total treatment
Group Type of illness
(d) treatment (d) (d)
26 Diarrhea 4
SP-LA 16 Diarrhea 3 18
13/18/26 BRD / Diarrhea / BRD 3/3/5
MP-LA 11 Diarrhea 4 4
16 BRD 3
SP-HA 14
9/20/26 Diarrhea / BRD / BRD 3/4/4
21 Diarrhea 5
MP-HA 21 BRD 4 13
15 Diarrhea 4
In general, the concentrations of IgA and IgG of calves of MP-HA and MP-LA were higher
than that of claves of SP-HA, and the concentrations of IgA and IgG of calves of SP-LA were
lowest (Table 3). The concentration of IgM of calves in SP-LA was significantly lower than the
other three groups (P<0.05), and the concentrations of TNF- and IL-6 were significantly higher
than the other 3 groups (P<0.05). The concentrations of IgA, IgM, IgG were higher at 1 and 4
weeks than the rest of age; the concentrations of IL-6 and TNF- were higher at 2 and 5 weeks
than the rest of age.
In week 3, the concentration of IgG of calves (MP-HA) was significantly lower than that of
calves (MP-LA) (P<0.05) (Figure 2). In week 4, the concentration of IgM of calves (MP-HA) was
significantly higher than that of calves (MP-LA) (P<0.05). In addition, there were no significant
differences in other immune indicators of calves with mesh partitions in weeks 1–4. The results
were consistent with the results in Table 2. In weeks 2–5, the concentrations of IgA, IgM, and IgG
of calves (SP-LA) were lower than that of calves (SP-HA) (P<0.05); the concentrations of TNF-
, and IL-6 of calves (SP-LA) were significantly higher than that of calves (SP-HA). The
concentrations of IgA, IgM, and IgG of calves reared with mash partitions were higher than that
∙ 71 ∙
of calves reared with solid partitions, and the concentrations of TNF- and IL-6 of calves raised
with mash partitions were lower than that of calves raised with solid partitions.
Figure 2. The concentrations of IgA (A), IgM (B), IgG (C), and TNF-α (D) and IL-6 (E) among
calves of groups (SP-LA, MP-LA, SP-HA, and MP-HA). Data represent mean ± SEM.
a–c Values having different letters within the same column are significantly different (P<0.05).
4. Discussion
All that is known is that calves in pens with different partitions receive air at different speeds
from fans. Still air was defined as air moving at a speed of less than 0.3 m s-1 (Wathes et al., 1983),
thus, the air in SP-LA pens was still air. Calves in SP-LA has highest respiratory rates and rectal
temperature, which was similar to a previous study (Zhao et al., 2021). In warm seasons, airflow
speed (0.5–2.0 m s-1) can reduce the calf’s respiration and rectal temperature.
∙ 72 ∙
Table 3. Immunity parameters of calves.

IgA IgG IgM TNF- IL-6
Item
(g L-1) (g L-1) (g L-1) (pg mL-1) (pg mL-1)
Group
SP-LA 4.50c 4.16c 1.92b 153.70a 148.54a
MP-LA 5.17a 6.01a 2.42a 108.27b 94.36b
b b a b
SP-HA 4.78 5.08 2.30 116.66 99.11b
MP-HA 5.07a 6.13a 2.41a 110.14 b 95.49b
SEM 0.06 0.15 0.04 3.46 4.18
Age (Weeks)
1 4.94ab 5.44ab 2.34ab 121.49bc 114.01ab
c b c ab
2 4.68 4.99 2.15 126.33 116.29a
3 4.91b 5.35b 2.27b 121.65b 99.76b
a a a c
4 5.06 5.73 2.35 113.99 104.29b
5 4.75bc 5.17b 2.23bc 132.11a 120.30a
SEM 0.07 0.17 0.04 4.10 4.97
P-value
Group <0.05 <0.05 <0.05 <0.05 <0.05
Age <0.05 <0.05 <0.05 <0.05 <0.05
Group  Age NS1 NS NS NS NS
a,b,c
Note: Values having different letters within the same column are significantly different
(P<0.05); 1NS = no significance.
During the experiment, calves of each group infected with BRD or diarrhea. The treatment
days of calves raised in pens with mesh partitions were shorter than that of the solid partitions.
The prevalence of calf respiratory disease increased with increasing bacterial counts in the pens,
and decreased with the presence of solid partitions between pens (Lago et al., 2006), of which the
results were consistent with this study. In a study, the incidence of respiratory disorders of calves
housed in individual pens with solid partitions was 38.5%, compared with 60.0% for group-
penned calves in winter (Hanekamp et al., 1994).
Apart from production and health variables, we also decided to check concentrations of IgG,
IgM, and IgA. When calves are in poor health, they tend to have lower immunoglobulins. In calves
in weeks 2–4 (highest incidence of diseases), the concentration of IgG1, IgG2, and IgA in nasal
secretions are the lowest level (Van Donkersgoed et al., 1993). In the present study, the
concentrations of IgA, IgG, and IgM of calves reared in pens with mesh partitions were higher
than that of calves reared in pens with solid partitions, and the concentrations of IgA, IgG, and
IgM of calves subjected to SP-LA were lowest. The results show that, in summer, combination of
mesh partition + high-speed airflow (>0.5 m s-1) has a positive effect on the immunoglobulin
concentrations of calves, while the combination of solid partition + low-speed airflow (<0.2
m s-1) has a negative impact on the immunoglobulin concentrations of calves. In the practical
level, the suitable airflow speed was related to the ambient temperature. In the average
temperature was about 14 C, the airflow speed >0.8 m s-1 was associated with lung consolidation
in group-housed calves (van Leenen et al., 2020).
Apart from immunoglobulin concentrations, we investigated that some markers and mediators
of inflammation were commonly used in the diagnosis of inflammation, such as TNF- and IL-6
(El-Ashker et al., 2014). Indoor environments that do not have appropriate ventilation can have
greater airborne noxious gases and dust, and bacterial growth in moist bedding, which can all have
negative effects on a calf’s health and well-being (Cobb et al., 2014). The concentrations of IL-6
and TNF- of calves in SP-LA were highest in each week. It may be because there were more
bacteria in the animal-occupied zone of calves when those pens do not have appropriate
ventilation. In agreement with a previous study, the levels of IL-6 and TNF-α increase in the
∙ 73 ∙
presence of Gram-negative pathogens and viral agents (Asgharpour et al., 2020). Except for the
calves in the SP-LA, the concentrations of TNF- and IL-6 of calves in the SP-HA, MP-HA, and
MP-LA showed a downward trend with increasing age (7-35 d).
5. Conclusions
There were three major findings of this study: 1) when the average daily temperature and
relative humidity were ranging from 18.57 to 28.37 C and 61.93% to 97.92%, respectively, in
the calf house, the calves of group SP-LA had higher rectal temperature and respiratory rate than
other groups. 2) The combination of mesh partition + high-speed airflow (>0.5 m s-1) has a positive
effect on the immunoglobulin concentrations of calves’ blood, while the combination of solid
partition + low airflow speed (<0.2 m s-1) has a negative impact on the immunoglobulin
concentrations of calves’ blood. 3) The concentrations of IL-6 and TNF- of calves reared in pens
with solid partitions and subjected to low-speed airflow were higher than other groups in each
week. Therefore, we concluded that in the warm season, with the axial-flow fan turned on, the
contact between calves caused by mesh partitions has no negative impact on the health of the
calves. At the same time, the low-speed airflow (<0.2 m s-1) in the calves’ occupied zone caused
by solid partitions increases the risk of calf disease.
Acknowledgements
The study was supported by China Agriculture Research System of MOF and MARA. The
authors gratefully acknowledge the help provided by Shenfeng dairy farm in Yancheng, China,
where the experiments were carried out.
References
Asgharpour, P., M. R. M. Dezfouli, M. G. Nadealian, Z. Eftekhari, G. R. N. Borojeni. 2020. Effects of 1,
25-dihydroxy vitamin D3 on clinical symptoms, pro-inflammatory and inflammatory cytokines in calves with
experimental pneumonia. Research in Veterinary Science. 132: 186–193. https://doi.org/10.1016/
j.rvsc.2020.04.018
Bruoucek, J., P. Kisac, M. Uhrincat, A. Hanus, F. Benc. 2008. Effect of high temperature on growth
performance of calves maintained in outdoor hutches. Journal of animal feed science. 17 (2):137–
146. https://doi.org/10.22358/jafs/66477/2008
Callan, R. J., Garry, B. F., 2002. Biosecurity and bovine respiratory disease. The veterinary clinics of
North America. 18(1):57–77. https://doi.org/10.1016/s0749-0720(02)00004-x
Cobb, C. J., B. S. Obeidat, M. D. Sellers, A. R. Pepper-Yowell, M. A. Ballou. 2014. Group housing of
Holstein calves in a poor indoor environment increases respiratory disease but does not influence performance
or leukocyte responses. Journal of Dairy Science. 97(5):3099–3109. https://doi.org/10.3168/jds.2013-7823
Costa, J. H., R. R. Daros, M. A. von Keyserlingk, D. M. Weary. 2014. Complex social housing reduces
food neophobia in dairy calves. Journal of Dairy Science. 97(12):7804–7810. https://doi.org/10.3168/
jds.2014-8392
Dado-Senn, B., V. Ouellet, G. E. Dahl, J. Laporta. 2020. Methods for assessing heat stress in preweaned
dairy calves exposed to chronic heat stress or continuous cooling. Journal of Dairy Science.103(9):8587–8600.
https://doi.org/10.3168/jds.2020-18381
Duve, L. R., M. B. Jensen. 2011. The level of social contact affects social behaviour in pre-weaned dairy
calves. Applied Animal Behaviour Science 135(1–2):34–43. https://doi.org/10.1016/j.applanim.2011.08.014
El-Ashker, M., M. Salama, A. Rizk, M. El-Boshy. 2014. The use of inflammatory markers as a prognostic
aid for traumatic reticuloperitonitis in water buffalo. Veterinární Medicína 59 (5):239–
246. https://doi.org/10.17221/7517-vetmed
Grell, S. N., T. Kirsten, L. E. Larsen, P. M. H. Heegaard. 2005. Marked induction of IL-6, haptoglobin
and IFNγ following experimental BRSV infection in young calves. Veterinary Immunology and
Immunopathology. 103(3–4):235–245. https://doi.org/10.1016/j.vetimm.2004.09.025
Gulliksen, S. M., K. I. Lie, T. Løken, O. Østerås. 2009. Calf mortality in Norwegian dairy herds. Journal
of Dairy Science. 92(6):2782–2795. https://doi.org/10.3168/jds.2008-1807
Hanekamp, W. J. A., A. C. Smits, H. K. Wierenga. 1994. Open versus closed barn and individual versus
group-housing for bull calves destined for beef production. Livestock Production Science. 37(3):261–
270. https://doi.org/10.1016/0301-6226(94)90121-x
∙ 74 ∙
Hill, T. M., H. G. Bateman, J. M. Aldrich, R. L. Schlotterbeck. 2011. Comparisons of housing, bedding,

and cooling options for dairy calves. Journal of Dairy Science. 94(4):2138–2146. https://doi.org/10.3168/
jds.2010-3841
Lago, A., S. M. McGuirk, T. B. Bennett, N. B. Cook, K. V. Nordlund. 2006. Calf respiratory disease and
pen microenvironments in naturally ventilated calf barns in winter. Journal of Dairy Science. 89(10):4014–
4025. https://doi.org/10.3168/jds.s0022-0302(06)72445-6
Mohammadi, G. R., K. Ghazvini, H. Abbas Panah. 2006. Antimicrobial susceptibility testing of
Mannheimia haemolytica and Pasteurella multocida isolated from calves with dairy calf pneumonia. Archives
of Razi Institute 61(2):91–96.
Nordlund, K. V., C. E. Halbach. 2019. Calf Barn Design to Optimize Health and Ease of Management.
Veterinary Clinics of North America: Food Animal Practice. 35(1):29–45. https://doi.org/10.1016/
j.cvfa.2018.10.002
Pempek, J. A., M. L. Eastridge, S. S. Swartzwelder, K. M. Daniels, T. T. Yohe. 2016. Housing system
may affect behavior and growth performance of Jersey heifer calves. Journal of Dairy Science. 99(1):569–
578. https://doi.org/10.3168/jds.2015-10088
Piccione, G., G. Caola, R. Refinetti. 2003. Daily and estrous rhythmicity of body temperature in domestic
cattle. BMC Physiology 3:7. https://doi.org/10.1186/1472-6793-3-7
Spain, J. N., D. E. Spires. 1996. Effects of supplemental shade on thermoregulatory response of calves to
heat challenge in a hutch environment. Journal of Dairy Science.79(4):639–646. https://doi.org/10.3168/
jds.s0022-0302(96)76409-3
Spiers, D. E., J. N. Spain, M. R. Ellersieck, M. C. Lucy. 2018. Strategic application of convective cooling
to maximize the thermal gradient and reduce heat stress response in dairy cows. Journal of Dairy Science.
101(9):8269–8283. https://doi.org/10.3168/jds.2017-14283
Urie, N. J., J. E. Lombard, C. B. Shivley, C. A. Kopral, A. E. Adams, T. J. Earleywine, J. D. Olson, F. B.
Garry. 2018. Preweaned heifer management on US dairy operations: Part V. Factors associated with morbidity
and mortality in preweaned dairy heifer calves. J Dairy Sci 101(10):9229–9244. https://doi.org/10.3168/
jds.2017-14019
Van Donkersgoed, J., C. S. Ribble, L. Boyer, and H. Townsend. 1993. Epidemiological study of enzootic
pneumonia in dairy calves in Saskatchewan. Canadian Journal Veterinary Research. 57 (4):247–254.
van Leenen, K., J. Jouret, P. Demeyer, L. Van Driessche, L. De Cremer, C. Masmeijer, F. Boyen, P.
Deprez, B. Pardon. 2020. Associations of barn air quality parameters with ultrasonographic lung lesions,
airway inflammation and infection in group-housed calves. Preventive Veterinary Medicine. 181:105056.
https://doi.org/10.1016/j.prevetmed.2020.105056
Wathes C. M., C. D. Jones, A. J. Webster. 1983. Ventilation, air hygiene and animal health. The
Veterinary Record. 113(24):554–559.
Zhao, W., C. Choi, D. Li, G. Yan, H. Li, Z. Shi. 2021. Effects of Airspeed on the Respiratory Rate, Rectal
Temperature, and Immunity Parameters of Dairy Calves Housed Individually in an Axial-Fan-Ventilated
Barn. Animals. 11(2):354. https://doi.org/10.3390/ani11020354
∙ 75 ∙
Assessing Air Quality in Three Types of Laying Hen Houses Using

an Internet of Things Technique
Zongyang Li a,b, Shaojie Wang a,b, Chaoyuan Wang a,b,*, Boyu Ji c, Yu Liu a,b
a
Department of Agricultural Structure and Bioenvironmental Engineering, College of Water Resources and
Civil Engineering, China Agricultural University, Beijing 100083, China
b
Key Laboratory of Agricultural Engineering in Structure and Environment, Ministry of Agriculture and
Rural Affairs, Beijing 100083, China
c
Pig Improvement Company, Shanghai 200030, China
* Corresponding author. Email: gotowchy@cau.edu.cn
Abstract
This paper presented an Internet of Things (IoT) system using the wireless sensor network
(WSN) technique to monitor indoor air quality in three typical laying hen houses of NM (natural
mating colony cage), CC (conventional cage), and AV (aviary) in China. Carbon dioxide (CO2),
ammonia (NH3), and particulate matter (PM) concentrations in the houses were continuously
measured year-round. The results suggested that the developed IoT system functioned reliably
and could provide information for the farmers to adjust environmental control strategy based on
the online measured data. In the NM house, average CO2, NH3, PM2.5, and total suspended
particulate (TSP) concentrations were 1518 ± 650 mg m-3, 1.72 ± 1.26 mg m-3, 0.041 ± 0.019
mg m-3, and 0.331 ± 0.200 mg m-3, respectively. The corresponding results were 3190 ± 1737
mg m-3, 1.11 ± 0.81 mg m-3, 0.046 ± 0.023 mg m-3, and 0.303 ± 0.269 mg m-3 in the CC house,
and 2390 ± 1502 mg m-3, 1.23 ± 0.86 mg m-3, 0.024 ± 0.014 mg m-3, and 0.623 ± 0.798 mg m-3 in
the AV house. The results showed that CO2 and NH3 concentrations were significantly affected
by daily management, including manure removal and ventilation operation, and the PM
environment was mainly influenced by birds’ activities. The NM and AV houses provided larger
space for laying hens to perform more natural behavior and resulted in higher coarse particle
concentrations. Generally, the air pollutants threatening the workers and chicken health, including
NH3 and PM2.5, remained at similar levels in the three investigated houses, suggesting both animal
friendly and conventional houses for laying hens could achieve desirable indoor environment
under appropriate management.
Keywords: Laying hen houses, carbon dioxide, ammonia, particulate matter, wireless sensor
network, big data
1. Introduction
Internet of Things (IoT) applied in precision agriculture has received increasing interest
during the past decade (Tzounis et al., 2017; Villa-Henriksen et al., 2020). The wireless sensor
network (WSN) was widely established as a part of IoT system for collecting information of
weather, crop, and soil, environment control of greenhouses, accurately measuring the fertilizer
and water requirement (Aqeel ur et al., 2014; Ojha et al., 2015). Some of the research also focused
on tracking animals and analyzing their behaviors using the WSN technology (Huircán et al., 2010;
Kwong et al., 2012). The growing scale of livestock farms makes it hard to manually control the
production system precisely. Introducing an intelligent management strategy based on IoT is
becoming more and more common on modern livestock farms.
Many researchers have applied the WSN technology on monitoring livestock building
environment, but few published studies have investigated the availability in real livestock houses
(Bustamante et al., 2012; Calvet et al., 2014; Li et al., 2015). The complex conditions in real
production system cannot be simply simulated in controlled chamber. For example, the hostile air
environment in poultry house contains large amount of corrosive gases and dust, which severely
influence the precision and lifespan of instruments and sensors. Many researchers found that the
routine maintenance of devices usually costs a lot of labor and time (Heber et al., 2006; Ni et al.,
∙ 76 ∙
2012; Zhao et al., 2015). Additionally, the complex topography or confined buildings equipped
with intensive cages bring high number of packet losses due to signal latency and attenuation
(Chen and Chen, 2019). Ensuring the velocity and veracity of big data transmission and analysis
is difficult but necessary (Kamilaris et al., 2017). It’s a big challenge to develop a robust and
feasible IoT system based on WSN technology for real-time monitoring the poultry house
environmental parameters stably for long periods.
Mature sensor technologies have decreased the cost of manufacturing and miniaturization,
allowing different sensors to be integrated into a portable instrument, which makes it easier to
monitor multi-variables in multi-points simultaneously. Ji et al. (2016) upgraded a Portable
Monitoring Unit (iPMU) for measuring temperature, pressure and concentration of carbon dioxide
(CO2), and ammonia (NH3) in livestock houses. Because of the different measurement principles
between gases and particulates, this version of PMU still lacked the ability to measure particulate
matter (PM) concentration. The Xbee module which iPMU used as wireless transmitter is
unsuitable for long-distance data transfer. In this study an IoT system was developed with sensor
nodes in the perception layer consisted of the optimized PMU (UPMU). UPMU added two sensors
based on the light scattering theory to measure concentrations of PM 2.5 and total suspended
particulate (TSP). The performance of these PM sensors was verified in laboratory and poultry
farms (Kelly et al., 2017; Li et al., 2020). LoRa and 4G technology were combined for establishing
the LAN-WAN structure, ensuring the stability of the network layer.
The objectives of this study were to (1) develop a reliable IoT system to continuously manage
environment data on laying hen farms; (2) provide a practical method for pre-processing the big
datasets; (3) verify the applicability of IoT system in three typical laying hen houses of NM
(natural mating colony cage), CC (conventional cage) and AV (aviary) system.
2.1. System frame design
The IoT system consisted of a perception layer, a network layer, and an application layer
(Figure 1). Several UPMUs distributed in laying hen houses constituted the perception layer. PM
cannot be sampled by vacuum pump sucking air through a tube because dust may clog the
sampling circuit. The PM sensor therefore must be separated from other sensors and installed at
the sampling point independently, which was usually in the cage. Every single PM sensor was a
LoRa node and formed the star-like structure LAN in case for the signal shielding. A pivot LoRa-
4G node was set outside the poultry house, collecting the PM concentration data transferred from
the PM sensor nodes, and uploaded the data to the OneNET Cloud IoT platform through the 4G
cellular network. Data of temperature, relative humidity (RH), and concentrations of CO2 and
NH3 were transmitted to the platform through 4G directly. Users can check the data remotely
through a personal computer (PC) or a mobile phone connected to the background system of
OneNET Cloud platform.
Figure 1. System frame for wireless monitoring chicken house environment.
∙ 77 ∙
2.2. Hardware Design

The UPMU consisted of a control module, a data acquisition module, and a transmission
module. The control module was a microcontroller board based on the Atmega2560 (Arduino,
www.arduino.cc). In this study, an AC-to-DC adapter supplied the power to the microcontroller.
The other modules were powered by the microcontroller board. The data acquisition module was
made up of four sensors and a time module (DS1502). Technical data of the sensors are shown in
Table 1. A GPRS module (YL-700 DTU, JX-YL, Shenzhen) had a maximum transmission rate
of 4.2 Mbps (4G) for transferring data wirelessly to the IoT platform.
Table 1. Technical data of the sensors in UPMU.
Technical data
Sensor Parameter Response Measuring Accuracy Maintenance
time (s) range (mg m-3) (mg m-3) period
MH-Z14 CO2 <450 0–9800 ±98
3 months
ME3-NH3 NH3 <250 0–38 ±1
PMS5003T PM2.5 <1 0–0.5 ±10
2 weeks
SDS198 TSP <6 0–20 ±30
2.3. Surveyed laying hen houses
Experiments were conducted for analyzing the robust performance of the IoT system in hostile
conditions and investigating the overall environment characteristics under the same monitoring
system. The general information of three studied laying hen houses is summarized in Table 2.
Table 2. Description of the three laying hen houses (Liu et al., 2020).
Items NM CC AV
Location 113.49° E, 36.41° N 115.06° E, 36.33° N 116.11° E, 40.08° N
Elevation (m) 583 39 50
Breed Hy-Line gray Hy-Line brown Jingfen NO. 6
Animal capacity 25,000 100,000 2,952
Cage size (m) 4.8 L×1.2 W×0.7 H 0.6 L×0.6 W×0.5 H 4.8 L×2.8 W×3.3 H
Stocking density
700 472 711
(cm2 hen-1)
Layout 4 rows, 4 tiers 5 rows, 8 tiers 2 rows
House size (m) 120 L×12 W×5.8 H 109 L×14 W×6.5 H 90 L×12 W×3.5 H
Mechanical, 6 side-
Mechanical Mechanical
Ventilation wall fans, 36 end-
16 end-wall fans 3 end-wall fans
wall fans
Manure removal Every 7 days Every 2 days Every 2 days
Jul. 1, 2019 – Dec. 15, 2019 – Nov. 1, 2019 –
Experiment time
Oct. 31, 2019 Dec. 1, 2020 Dec. 1, 2020
2.4. Big data processing
The IoT system produced large amount of data during the operation of UPMU. The common
method of processing data usually applies the three deltas theory of the normal distribution to find
out outliers. However, the daily data of environmental parameters varied temporally, the data size
was not sufficient to be considered a normal distribution. In this case, accidental errors introduced
from the interferences such as device vibration or production operation inconsistencies were
filtered out based on the theory of box plot. A box plot can display the first and third quartiles (Q1
and Q3) of a dataset. The difference value of Q1 and Q3 is the interquartile range (IQR). A
distance of 1.5 times the IQR from the lower and upper quartile distinguishes the normal data and
outliers. Pre-processing diurnal data by eliminating outliers of box plot was performed before
analyzing the whole dataset.
∙ 78 ∙
Besides, the influence of ambient (atmosphere) pollution should be taken into consideration
while processing the PM data. Research revealed that when the ambient air quality index (Chinese
AQI) exceeded 100, part of the indoor PM was derived from the ambient (Li et al., 2020). PM2.5
is one of the six pollutants for AQI evaluation and have significant impact on the indoor
environment. For investigating the real PM characteristics of the layer house precisely, the data
of PM2.5 and TSP were excluded in polluted days when individual air quality index (IAQI) of
PM2.5 was larger than 100.
For NH3 and PM concentrations, the minimum values were zero but no upper limit values
existed theoretically. Although the data size was big enough, the seasonal or annual NH3 and PM
concentration data showed as a normal distribution without the left part. The Kruskal-Wallis one-
way ANOVA non-parametric test was therefore performed in SPSS to compare the concentration
differences between seasons and farms.
2.5. Data processing and statistical analysis
The size of IoT dataset was so huge and complicated that cannot be processed by the common
method. Accidental errors generated from sensors needed to be eliminated based on the theory of
box plot. After excluding all the interferences, the values of six environmental parameters were
compared between seasons and different production systems. The data processing and statistical
analysis methods described above were performed using Python (version 3.0) and SPSS (version
25.0; IBM Crop., Armonk, NY, USA).
3.1. Performance of IoT system
To analyze the data transmission efficiency (the amount of data from the platform divided by
the amount of data from the devices) of the IoT system, the amount of data stored at the device
side and the amount of data stored at the platform side were counted. A total of more than 520,000
sets of data from the devices and nearly 370,000 sets of data at the platform were stored during
the experiment, and the overall data transmission efficiency was 71.2%. The data transmission
efficiency of the temperature, RH, and gases parameters which used 4G network to upload with
little network fluctuation was above 96%. This efficiency was much higher than the criteria set
by Ni et al. (2012) in that more than 18 h or 75% of valid data could be used to calculate the daily
means. On the other hand, the PM data efficiency was just above 40%. However, this size of valid
dataset at this efficiency was still enough to reflect the variation of PM concentration because the
measuring intervals of 5 min provided sufficient data to calculate the hourly means under the
tolerable range of packet loss rate. It meant that the establishment of LAN for monitoring PM
concentration was feasible. But the data transmission technology still needed further optimization
to promote the efficiency.
Data from different UPMUs often arrived at the gateway center simultaneously while
detecting the transmitted data, leading to dynamic memory overflowed in the gateway center and
failure in parsing the data. It illustrated two problems with the PM concentration monitoring
network:
(1). Data from 2–3 devices would arrive at the same time due to a long-time interval (2 s) for
the gateway center to detect and accept transmitted data. The phenomenon of “congestion” could
easily occur, causing the gateway center to fail in parsing the data.
(2) Although the new LoRa communication technology was adopted, the signal strength was
severely attenuated by the intensive production equipment in the house, resulting in a high packet
loss rate. Devices placed inside the cages were easily damaged by laying hens’ activities such as
chasing with each other, flapping wings or pecking. The robustness of UPMU needed to be
enhanced.
The congestion could be greatly alleviated when the interval of detecting data was set at 100
ms. The problem of serious signal attenuation was solved by replacing the glue stick antenna,
∙ 79 ∙
which sent out a weak signal strength, with a suction cup antenna or built-in patch antenna to
strengthen the signal and penetration ability.
3.2. Differences in environmental parameters between seasons and laying hen houses
The overall variations of temperature, RH, concentration of CO2, NH3, PM2.5, and TSP are
shown in Figure 2, which shows that monitoring four variables of temperature, RH, CO2, and NH3
performed much better than PM. It can correspond with higher maintenance frequency for PM
sensors than other devices during experiments. The stability of WSN for PM monitoring was
improved after the experiment in NM, few packet losses were found in the later period in CC and
AV. The data deficiencies in the summer of 2020 in CC and AV were attributed to the COVID-
19 pandemic. People were quarantined at home and broken devices in poultry farm cannot be
repaired timely.
Figure 2. Variations of different environmental parameters in three poultry production systems.

Six important environmental parameters, temperature, RH, concentrations of CO 2, NH3,
PM2.5, and TSP, were recorded during the experiments. The summary of descriptive statistics
(mean ± SD) is shown in Table 3.
∙ 80 ∙
Table 3. Descriptive statistic result (mean ± SD) of environmental parameters measured in three
laying hen houses and four seasons.
Parameter Season NM CC AV
Spring (Mar – May) NA 23.3 ± 2.4 Ac 20.2 ± 3.8 Bb
T Summer (Jun – Aug) 25.0 ± 1.9 Ba* 28.6 ± 2.0 Aa 24.7 ± 2.6 Ca
(°C) Autumn (Sept – Nov) 21.7 ± 2.4 Bb 24.0 ± 3.7 Ab 17.3 ± 3.6 Cc
Winter (Dec – Feb) NA 19.4 ± 2.7 Ad 16.9 ± 2.6 Bd
Spring (Mar – May) NA 47 ± 11 d 47 ± 18 d
RH Summer (Jun – Aug) 73 ± 21 Ca 88 ± 12 Aa 77 ± 17 Ba
(%) Autumn (Sept – Nov) 68 ± 17 Ab 54 ± 12 Bc 50 ± 16 Cc
Winter (Dec – Feb) NA 62 ± 10 Ab 54 ± 11 Bb
Spring (Mar – May) NA 2993 ± 970 Ab 1821 ± 1111 Bc
CO2 Summer (Jun – Aug) 1165 ± 236 Ab 1034 ± 511 Bc 708 ± 182 Cd
(mg m ) Autumn (Sept – Nov)
-3
2049 ± 712 Ca 3113 ± 946 Ab 2446 ± 1152 Bb
Winter (Dec – Feb) NA 5398 ± 901 Aa 3971 ± 1082 Ba
Spring (Mar – May) NA 1.21 ± 0.52 Bb 1.33 ± 0.80 Ab
NH3 Summer (Jun – Aug) 1.39 ± 0.87 Ab 1.20 ± 0.76 Bb 1.06 ± 0.76 Cc
(mg m ) Autumn (Sept – Nov)
-3
2.21 ± 1.56 Aa 0.45 ± 0.54 Cc 1.35 ± 1.06 Bb
Winter (Dec – Feb) NA 1.70 ± 0.79 Ba 2.45 ± 1.20 Aa
Spring (Mar – May) NA 0.033 ± 0.012 Ad 0.021 ± 0.013 Bc
PM2.5 Summer (Jun – Aug) 0.043 ± 0.020 Ba 0.044 ± 0.018 Ac 0.030 ± 0.012 Ca
(mg m-3) Autumn (Sept – Nov) 0.034 ± 0.011 Bb 0.053 ± 0.028 Aa 0.025 ± 0.015 Cb
Winter (Dec – Feb) NA 0.048 ± 0.019 Ab 0.022 ± 0.013 Bc
Spring (Mar – May) NA 0.317 ± 0.327 Ac 0.141 ± 0.166 Bc
TSP Summer (Jun – Aug) 0.330 ± 0.206 Ab 0.108 ± 0.070 Bd 0.105 ± 0.112 Cd
(mg m-3) Autumn (Sept – Nov) 0.333 ± 0.200 Ba 0.339 ± 0.196 Bb 1.049 ± 0.665 Ab
Winter (Dec – Feb) NA 0.469 ± 0.404 Ba 1.333 ± 0.923 Aa
* Different superscripts of uppercase letters within the same row indicate significant
difference (P<0.05) in the means between the different laying hen houses. Lowercase letters stand
for significant difference (P<0.05) between seasons.
3.2.1. Temperature
The stocking density and ambient temperature were two important factors influencing the
temperature inside the poultry houses. The temperature in CC was the highest among all laying
hen houses no matter in which season. It was mainly attributed to the highest stocking density of
472 cm2 hen-1 in CC. AV maintained the lowest indoor temperature with the largest area for laying
hens at 711 cm2 hen-1. The indoor temperature varied along with ambient temperature from the
hottest summer at 28.6 ± 2.0 °C in CC to the coldest winter at 16.9 ± 2.6 °C in AV. But the thermal
environments in all laying hen houses stayed within the desirable levels.
3.2.2. Relative humidity
The full-time operation of the evaporative pad introduced a large amount of moisture in
summer, leading to the highest RH in three poultry houses. Moreover, humidity increased with
growing latent heat produced by more laying hens, and thus the higher stocking density in CC led
to a higher indoor RH with 88 ± 12% in summer. Less ventilation and lower temperature in winter
made the moisture accumulate inside the houses, resulting in a higher level of RH than in spring
and autumn. The driest season was spring with 47% RH in both CC and AV.
3.2.3. Carbon dioxide
The ventilation rate decreased with the decreasing ambient temperature, resulting in the CO 2
accumulated inside the houses. The CO2 concentrations were therefore the highest in autumn
∙ 81 ∙
(NM) and winter (CC and AV). The stocking density also affected the CO2 concentration because
of the chicken respiration. The concentration of CO2 in CC was the highest of the three laying hen
houses. However, the CO2 concentration of NM was higher than that of CC in summer, which
could be attributed to the fact that the NM located in an area with cooler weather than CC. The
hottest period lasted at most 1–2 weeks in July and the ventilation rate declined during other time
in summer, leading to a higher CO2 concentration of NM than CC and AV.
3.2.4. Ammonia
Manure is the main source of NH3 in poultry houses. Timely manure removal inside the house
will lower the indoor NH3 concentration. A 7-day manure removing cycle was the reason why
NH3 concentration in NM was the highest at 1.39 ± 0.87 mg m-3 and 2.21 ± 1.56 mg m-3 in summer
and autumn, respectively. In addition, standardized management practices like washing floors and
sterilizing cages were helpful to keep the concentrations of hazardous gases at relatively low
levels. Despite cleaning manure belt every two days, the manure residue remained on the aviary
was hard to be removed thoroughly, producing more NH3 in AV. The ventilation rate was another
factor that affected the NH3 concentration. Similarly, NH3 accumulated easily when the
ventilation rate decreased. Ammonia concentrations were all higher in the colder season, but
maintained at a safe level much lower than the threshold of 15 mg m-3 in NY/T 388-1999.
3.2.5. PM2.5
The PM2.5 concentration probably related to the stocking density when there was no extra
PM2.5 source like litter on solid floors. CC had a higher concentration of PM 2.5 than NM, and the
lowest PM2.5 concentration was found in AV. Increasing ventilation rate did not seem to have any
effect on PM2.5 because the concentrations in summer and autumn were higher than in winter
when the ventilation rate was the lowest. The average concentrations in none of the housing
systems exceeded 0.075 mg m-3, which met the requirement proposed by GB 3095 (2012)
stipulating that 24-hour average concentration must be less than 0.075 mg m-3 in the rural area of
China. But bacteria attaching to PM affected the respiratory health of workers cannot be ignored.
3.2.6. TSP
Ventilation, living space, and the activities of laying hens were highly related to the
concentration of TSP. Birds were able to walk in the cage in NM, running and flying were even
allowed in AV. When the ventilation rate was low in autumn and winter, TSP concentrations in
AV were higher than those in CC. With the outdoor environment getting warmer in spring and
summer, more TSP was exhausted by ventilation in CC and AV. Similar to the ventilation impact
on CO2 in summer, the TSP concentration in NM was the highest among three laying hen houses
because of the coolest weather.
4. Conclusions
A system frame of portable environment monitoring equipment for poultry houses based on
IoT system was developed. The distributed wireless monitoring network for poultry barn
environment was established, which met the functional requirements of long-time continuous
monitoring of multiple environmental parameters. A LAN structure based on LoRa was feasible
for long-time monitoring in poultry facilities but needed further optimization.
Environmental parameters of the laying hen house in AV were generally comparable to those
of the NM and CC system due to the suitable system design and environment control strategy.
Unlike the usual AV systems in EU and the US, the surveyed AV laying hen house was equipped
with wire floor and no litter spreading on the floor, which greatly mitigated the generation of PM.
Larger space allowance per hen and relatively higher ventilation rates also provided a good air
quality, especially that the PM2.5 concentration was even the lowest in three poultry houses.
Acknowledgements
This work was supported by the National Key R&D Program Project of China (grant number
∙ 82 ∙
2017YFD0701602).
References
Aqeel ur, R., A.Z. Abbasi, N. Islam, Z.A. Shaikh, 2014. A review of wireless sensors and networks'
applications in agriculture. Computer Standards & Interfaces. 36 (2), 263–270. http://doi.org/https://doi.org/
10.1016/j.csi.2011.03.004.
Bustamante, E., E. Guijarro, F.-J. García-Diego, S. Balasch, A. Hospitaler, A.G. Torres, 2012.
Multisensor System for Isotemporal Measurements to Assess Indoor Climatic Conditions in Poultry Farms.
Sensors. 12 (5). http://doi.org/10.3390/s120505752.
Calvet, S., J.C. Campelo, F. Estellés, A. Perles, R. Mercado, J.J. Serrano, 2014. Suitability evaluation of
multipoint simultaneous CO2 sampling wireless sensors for livestock buildings. Sensors. 14 (6), 10479–
10496.
Chen, C.-S., W.-C. Chen, 2019. Research and Development of Automatic Monitoring System for
Livestock Farms. Applied Sciences. 9 (6). http://doi.org/10.3390/app9061132.
Heber, A.J., T.-T. Lim, J.-Q. Ni, P.-C. Tao, A.M. Schmidt, J.A. Koziel, S.J. Hoff, L.D. Jacobson, Y.
Zhang, G.B. Baughman, 2006. Quality-assured measurements of animal building emissions: Particulate matter
concentrations. J. Air Waste Manage. Assoc. 56 (12), 1642–1648. http://doi.org/10.1080/10473289.2006.
10464569.
Huircán, J.I., C. Muñoz, H. Young, L. Von Dossow, J. Bustos, G. Vivallo, M. Toneatti, 2010. ZigBee-
based wireless sensor network localization for cattle monitoring in grazing fields. Computers and Electronics
in Agriculture. 74 (2), 258–264. http://doi.org/https://doi.org/10.1016/j.compag.2010.08.014.
Ji, B., W. Zheng, R.S. Gates, A.R. Green, 2016. Design and performance evaluation of the upgraded
portable monitoring unit for air quality in animal housing. Computers and Electronics in Agriculture. 124,
132–140.
Kamilaris, A., A. Kartakoullis, F.X. Prenafeta-Boldú, 2017. A review on the practice of big data analysis
in agriculture. Computers and Electronics in Agriculture. 143, 23–37. http://doi.org/10.1016/j.compag.2017.
09.037.
Kelly, K.E., J. Whitaker, A. Petty, C. Widmer, A. Dybwad, D. Sleeth, R. Martin, A. Butterfield, 2017.
Ambient and laboratory evaluation of a low-cost particulate matter sensor. Environmental Pollution. 221, 491–
500.
Kwong, K.H., T.-T. Wu, H.G. Goh, K. Sasloglou, B. Stephen, I. Glover, C. Shen, W. Du, C. Michie, I.
Andonovic, 2012. Practical considerations for wireless sensor networks in cattle monitoring applications.
Computers and Electronics in Agriculture. 81, 33–44. http://doi.org/https://doi.org/10.1016/j.compag.2011.
10.013.
Li, H., H. Wang, W. Yin, Y. Li, Y. Qian, F. Hu, 2015. Development of a remote monitoring system for
henhouse environment based on IoT technology. Future Internet. 7 (3), 329–341.
Li, Z., C. Wang, B. Li, Z. Shi, W. Zheng, G. Teng, 2020. Concentration and size distribution of particulate
matter in a new aviary system for laying hens in China. Journal of the Air & Waste Management Association.
70 (4), 379–392. http://doi.org/10.1080/10962247.2020.1722291.
Liu, Y., G. Zhang, L. Rong, Z. Li, S. Wang, C. Wang, 2020. Concentrations and Emissions of Ammonia
from Different Laying Hen Production Systems of Conventional Cage, Aviary and Natural Mating Colony
Cage in North China Plain. Applied Sciences. 10 (19), 6820.
Ni, J.-Q., L. Chai, L. Chen, B.W. Bogan, K. Wang, E.L. Cortus, A.J. Heber, T.-T. Lim, C.A. Diehl, 2012.
Characteristics of ammonia, hydrogen sulfide, carbon dioxide, and particulate matter concentrations in high-
rise and manure-belt layer hen houses. Atmospheric Environment. 57, 165–174.
Ojha, T., S. Misra, N.S. Raghuwanshi, 2015. Wireless sensor networks for agriculture: The state-of-the-
art in practice and future challenges. Computers and Electronics in Agriculture. 118, 66–84. http://doi.org/
10.1016/j.compag.2015.08.011.
Tzounis, A., N. Katsoulas, T. Bartzanas, C. Kittas, 2017. Internet of Things in agriculture, recent advances
and future challenges. Biosystems Engineering. 164, 31–48. http://doi.org/10.1016/j.biosystemseng.2017.
09.007.
Villa-Henriksen, A., G.T. Edwards, L.A. Pesonen, O. Green, C.A.G. Sørensen, 2020. Internet of Things
in arable farming: Implementation, applications, challenges and potential. Biosystems Engineering. 191, 60–
84.
Zhao, Y., T.A. Shepherd, H. Li, H. Xin, 2015. Environmental assessment of three egg production
systems–Part I: Monitoring system and indoor air quality. Poultry Science. 94 (3), 518–533.
∙ 83 ∙
Disinfection Effect of Slightly Acidic Electrolyzed Water Applied in

Drinking Water Systems of Large-scale Dairy Farms
Xuedong Zhao a,b, Chenxuan Zhu c, Zhengxiang Shi a,b,*
a
b
Key Laboratory of Agricultural Engineering in Structure and Environment, Ministry of Agriculture and
c
Shanghai Academy of Agricultural Sciences, Shanghai, China
Abstract
Drinking water quality is an important factor on dairy cows' health and production
performance. Drinking water systems in large-scale dairy farms are difficult and inefficient to
clean, and the pipelines and drinking troughs are easily contaminated by biofilm coverage to affect
cows' health. This study investigated the effect of slightly acidic electrolyzed water (SAEW) as a
disinfectant on drinking water quality of the drinking water system, to provide a method for
designing an efficient, low-cost and healthy drinking water system for large-scale dairy farms.
Further, disinfection of SAEW with an available chlorine concentration (ACC) of 30 mg L-1 for
12 min can kill 90.4% of microorganism in drinking water. After cleaning the drinking water
system using SAEW with 30 mg L-1 ACC, the residual chlorine concentration of drinking water
was maintained above 0.05 mg L-1 for 2 h and no residual chlorine remained after 6 h. The working
time interval for cleaning the drinking water system with SAEW was 48 h according to turbidity.
In dairy farms, dairy performance and the average milk yield per day can be promoted and
increased by 9.3% after SAEW cleaning.
Keywords: Water quality, disinfection, slightly acidic electrolyzed water, drinking water system,
large-scale dairy farm
1. Introduction
Drinking water for dairy cows is extremely sensitive and important for their health, body
immunity and milk yield. Even if the difference in diet levels is not significant, it is possible that
milk yield may drop by 18% to 20% due to low water quality (Murphy, 1992; Cardot et al., 2008;
Golher et al., 2021). In dairy drinking water management, water-soluble additives are often used
through the drinking water system in order to ensure that the drinking water quality meets sanitary
standards (Plaizier et al., 2005; Vargas et al., 2020). Long-term accumulation of impurities and
microorganisms in the water can form biofilm on inner wall of drinking water pipe and drinking
troughs (Schutz et al., 2019), which causes pollution of the drinking water system of the dairy
farms and affects the health of the cows seriously. Therefore, seeking for an efficient, non-toxic
and environmentally friendly disinfectant is essential to ensure the safety of drinking water for
dairy cows.
Currently, the main methods of disinfection of drinking water are chemical methods such as
chlorine, and ozone disinfection, and physical methods such as UV disinfection. Chlorine is the
most widely used disinfectant due to its low cost, ease of use and ability to kill microorganisms,
but chlorine produces toxic byproducts such as trihalomethanes and haloacetic acids during
disinfection and is unsafe to use and store (Wang et al., 2021). Ozone has a broad spectrum and
good disinfection effect, and is reduced to oxygen after disinfection, but ozone is limited by its
half-life, high cost and has no lasting disinfection capability (Zhang et al., 2020). UV disinfection
is simple to use and does not produce harmful disinfection by-products, but dissolved substances
in the water can interfere with the disinfection effect of UV (Wang et al., 2016). The realization
of an efficient, low-cost, green drinking water cleaning system is a key issue that needs to be
addressed in large-scale dairy farms.
Slightly acidic electrolyzed water (SAEW) can be used to replace some of the traditional
∙ 84 ∙
disinfection chemicals because of its wide range and high efficiency of disinfection, low price, no
pollution, no residue, and no harm to human body (Chen et al., 2016; Liu et al., 2020). Ni et al.
(2016) studied available chlorine concentration (ACC) and disinfection time on disinfection
process with SAEW, and the disinfection effect of SAEW was significantly greater than that of
phenol solution when the ACC was higher than 50 mg L-1 (Ni et al., 2016). The SAEW has been
commonly used in various fields and has played a good role in the environmental disinfection of
livestock houses. In application of drinking water systems, Bodas et al. (2013) added SAEW at
3% into the drinking water of ewes. The addition of 3% SAEW effectively reduced the growth of
bacteria in the water tank and significantly improved water quality without affecting the health
performance of the animals themselves (Bodas et al., 2013). Although SAEW has been used in
numerous systems for environmental disinfection of livestock houses, its application in the
drinking water system of dairy cows has not been addressed.
In this paper, the effect of using SAEW as a disinfectant for dairy water system cleaning on
drinking water system was tested to investigate whether the use of SAEW cleaning can be
promoted as a more reasonable way to clean the dairy water system, thus providing a basis for an
efficient, low-cost and green drinking water system design process for large-scale dairy farms.
The study on the feasibility and effectiveness of applying SAEW to dairy farm drinking water
system cleaning can provide a theoretical basis for developing new efficient and green
disinfectants, and can further improve staff welfare and cow welfare.
2.1. Dairy farm
The trial site is located at the Beijing Dairy Center Yanqing base. Holstein lactating cows
were kept in loose pens in fully enclosed barns, and calves were kept in outdoor hutches. Fully
automatic intelligent milking robots were used to milk the cows in the barn all day. The barn was
cleaned with an automatic scraper and manure channel system. Drinking water was taken from
the groundwater, and the water tank was made of stainless steel, which was automatically filled
with water according to water level sensing all day. The water tank and drinking troughs were
cleaned manually every four days, and the cleaning time was mostly at night.
2.2. Materials and equipment
The SAEW preparation device used in this experiment was jointly developed by the Key
Laboratory of Agricultural Engineering in Structure and Environment of MOARA of China
Agricultural University and Rui'ande Environmental Protection Equipment (Beijing) Co. The
device has an external dimension of 500 × 400 × 650 mm and can generate SAEW in a volume
of 100–250 L per hour, with an oxidation-reduction potential (ORP) of 800–1150 mV, a pH of
5.0–6.5, and an ACC of 80–250 mg L-1. WZB-170 portable turbidity meter and PHBJ-260F
portable pH meter from Shanghai Yidian Scientific Instruments Co., and RC-3F digital available
chlorine tester from Kasasa Principle Chemical Co. were used to test water quality.
2.3. Preparation of slightly acidic electrolyzed water
The SAEW was prepared by electrolysis of NaCl solution in an electrolyzer with a diaphragm,
and acidic electrolytic water containing hypochlorous acid, hypochlorite and dilute hydrochloric
acid was obtained at the anode, and alkaline electrolytic water was obtained at the cathode. In this
test, 5% mass fraction of NaCl solution was prepared, and after cyclic electrolysis by device, ACC
in the prepared SAEW was measured by using available chlorine concentration meter, and
electrolytic water with ACC of 60, 50, 30, 10, and 5 mg L-1 was prepared by diluting with
deionized water. The pH of the obtained electrolyzed water was adjusted to 5.5–6.5 by using 1%
mass fraction of HCl solution, and the pH and actual ACC of the electrolytic water were measured
for use.
2.4. Operation of cleaning system
The total length of the single line of water supply pipe is about 200 m and inner diameter of
∙ 85 ∙
the pipe is 32 mm, and the volume of drinking troughs is 200 × 300 × 4000 mm. So the total
volume in the pipe and trough is 0.4 m3 (0.16+0.24 m3) according to the calculation of the size
respectively, which is the minimum one-time demand of SAEW used for cleaning the drinking
water system. Due to the maximum water supply flow (Q = 2 m3 h-1) of pipe booster pump, the
working time of 12 min to fill with SAEW can be determined.
When using SAEW to clean the cow drinking water system, firstly, closing the drinking water
supply valve and the cold water return valve of the cow drinking water system; secondly, opening
the inlet valve of electrolyzed water after the tank was empty; thirdly, closing the water inlet valve
and leaving it for a period of time after the tank was full of SAEW; finally, turning on the switch
of the tank to empty the SAEW, closing the tank, and opening the drinking water inlet valve of
the system after emptying. The cleaning system consisted of the prepared SAEW pumped into the
system through the water supply pipe, which was discharged from the water tank drain after
cleaning the inner wall of the water supply pipe and drinking troughs, and clean water was
discharged to the rainwater collection pipe through the dark pipe in the barn. The free residual
chlorine, pH and turbidity of the drinking water were tested after the cleaning was completed and
the water was restored.
2.5. Test of drinking water properties
Under laboratory conditions, disinfection effects of SAEW with different ACCs were
investigated in 12 min. In practice, after the sink was cleaned with SAEW and started to be filled
with water, the free residual chlorine concentration of the drinking water in the sink was measured
after 0, 0.5, 1, 2, 4, 6, 12, 24, and 48 h. To investigate the effect of cleaning water system drinking
water quality after using SAEW, the pH and turbidity of drinking water in drinking troughs were
measured after drinking troughs started to be filled with water for 0, 0.5, 1, 2, 4, 6, 12, 24, and 48
h after cleaning was completed using SAEW.
2.6. Cultivation of colonies
The 9 ml of normal saline or SAEW at 60, 50, 30, 10, and 5 mg L-1 of ACC was placed in a
centrifuge tube with a pipette, and 1 mL of water sample diluted 100 times with saline was added
into the centrifuge tube with saline or SAEW and shake thoroughly. After mixing well for 12 min,
100 μL of the mixture was coated onto the surface of culture plates, and then incubated in a
biochemical incubator at 37℃ for 48 h. After the culture was completed, culture plates with the
number of colonies between 30 and 300 was selected for whole bacteria counting. Three repeated
samples were set for each group of colonies above.
3.1. Disinfection effect of SAEW with different ACCs
According to the results of the total colony culture to calculate the logarithmic value of the
total bacterial concentration, disinfection effects of SAEW with different ACCs on drinking water
are shown in Table 1. Analysis of the disinfection effects of SAEW with different ACCs in
drinking water showed that the bactericidal effect of SAEW was significantly enhanced (P<0.05)
with the ACC increasing. This was consistent with the results obtained from other studies. Thorn
et al. (2012) treated mung beans and their sprouts with 20 mg L-1 and 80 mg L-1 of ACC
respectively for 20 min effectively reduced the amount of E. coli and Salmonella enteritidis
without affecting the activity of the seeds (Thorn et al., 2012).
The trend analysis of disinfection efficiency versus ACC is shown in Figure 1. When the ACC
of SAEW was 10 mg L-1, the contact with drinking water for 12 min to kill the bacteria in it was
68.9%, and the total colony concentration in drinking water was 1.87 ± 0.23 lg (CFU mL-1). When
the ACC of SAEW was 30 mg L-1, the contact with drinking water for 12 min could largely kill
the bacteria in water, which was selected as the optimal concentration in actual operation. When
the ACC of SAEW was 50 mg L-1 and above, contact with drinking water disinfection for 12 min
could completely kill the bacteria in the water.
∙ 86 ∙
Table 1. Disinfection effects of SAEW with different ACCs.

Detected total bacterial Removed total bacterial
Samples
concentration, lg (CFU mL-1) concentration, lg (CFU mL-1)
Drinking water 6.02±0.22 ---
ACC 5 mg L-1 2.57±0.14 3.45±0.36
ACC 10 mg L-1 1.87±0.23 4.15±0.45
ACC 30 mg L-1 0.58±0.05 5.44±0.27
ACC 50 mg L-1 0 6.02±0.22
ACC 60 mg L-1 0 6.02±0.22
Figure 1. Disinfection efficiency of SAEW with different ACCs.

3.2. SAEW's continuous disinfection capability
As shown in Figure 2, the tank was filled with drinking water after 12 minutes, and the
concentration of residual chlorine decreased with time going by. The result showed 2 h residual
chlorine concentration to maintain the residual chlorine concentration of 0.05 mg L-1 or more, 4
h residual chlorine concentration of about 0.03 mg L-1, 6 h and after the residual chlorine
concentration in drinking water to 0. It indicated that after the use of SAEW cleaning, residual
chlorine in the system had a continuous disinfection capacity of 2–4 h for the drinking water in
the system, and no residual chlorine remained after 6 h.
3.3. Effect of SAEW on drinking water quality of dairy cows
In Figure 3a and 3b, according to the monitoring of the water intake of cows, the water
consumption of cows in the test group was about 180 ± 10 L h-1 per hour from 12:00 to 22:00.
Based on the pipe inner diameter of 32 mm, pipe length of 200 m, and the size of drinking trough
was 200 × 300 × 3500 mm, it was calculated that the turnover of drinking water in the whole
system was completed in about 2 hours. After using SAEW to clean the drinking water system,
the pH value of drinking water decreased in a short period of time due to the residue of SAEW in
the pipe and in the inner tank, and basically remained stable for 4 h.
Analysis of turbidity changes, turbidity in drinking water increased slightly for a short period
of time, which was due to the dissolution and stripping of biofilm on the inner wall of the pipes
to increase the content of turbid substances in the water. The turbidity of drinking water dropped
to the lowest level after 4 h, and the pH change pattern synchronized. After 6 h, the turbidity in
the water increased due to the feed and other impurities in the mouth and nose of cows or exposed
to air, but the turbidity range was controlled within 3 NTU within 48 h. Therefore, according to
∙ 87 ∙
the turbidity standard, it is recommended to use SAEW to clean the drinking water system at an
interval of 48 h. As shown in Figure 3 c, the SAEW effectively cleaned the biofilm on the inner
wall of the drinking water pipes of the cows. Measurements of drinking water quality in the
cleaned drinking troughs confirmed that SAEW can effectively clean the inner walls of drinking
water pipes and troughs and help maintain the quality of drinking water quality.
Figure 2. Residual chlorine concentration of drinking water after cleaning with SAEW.
3.4. Promotion of SAEW cleaning drinking water for the performance of dairy cows
To investigate the effect of the SAEW cleaning system on performance of dairy cows, 15
cows in good health were selected in the experimental group and 15 cows in the control group.
Drinking pipes and troughs of the experimental group were cleaned every 48 h with SAEW at an
ACC of 30 mg L-1, and other breeding environments and conditions were the same for both
groups. The average milk yield of cows in the test and control groups before the experiment was
equivalence, with an average daily milk yield of 39.7 kg and 38.9 kg, respectively. Then the
average daily milk yield of cows in both groups was collected for 26 consecutive days in October
as a production performance parameter.
The milk yield results of the two groups of cows are shown in Figure 4. Under the same
breeding conditions, the production performance of the two groups of cows was different. 39.9 ±
1.5 kg of daily milk yield was obtained from the control group and 43.6 ± 1.7 kg of daily milk
yield was obtained from the experimental group. The average daily milk yield of the experimental
group was 9.3% higher than that of the control group, indicating that the production performance
of the whole group of cows in the experimental group was better than that of the control group.
Analysis of the milk yield curves of cows in the control and test groups showed a significant
correlation (P<0.01), indicating the same trend of milk yield in both groups. It reflected the
environmental factors such as temperature had basically the same effect on the two groups of
cows, and the improved production performance of the experimental group was only attributed to
the improved water quality due to the cleaning effect of SAEW, which benefitted the health of the
cows and improved the milk yield.
∙ 88 ∙
Figure 3. a. pH of drinking water with time after cleaning with SAEW; b. Turbidity of drinking
water with time after cleaning with SAEW; c. The inner wall of drinking water pipe before and
after the cleaning of SAEW.
Figure 4. Milk yield of cows of the experimental and control groups.

4. Conclusions
A 12-min contact disinfection of SAEW with ACC of 30 mg L-1 could basically kill the
bacteria in drinking water. No residual chlorine remained 6 h after cleaning the dairy water system
∙ 89 ∙
with SAEW of 30 mg L-1 and the residual chlorine concentration of drinking water could be
maintained above 0.05 mg L-1 for 2 h. The drinking water cleaned with SAEW could improve the
production performance of cows, and the average daily milk yield of cows in the experimental
group was 9.3% higher than that of the control group. In practical application, SAEW should be
replenished every 48 h with the initial ACC value of 30 mg L-1.
Acknowledgements
The study was supported by China Agriculture Research System of MOF and MARA.
References
Bodas, R., D.J. Bartolome, M.J.T. De Paz, R. Posado, J.J. Garcia, L. Rodriguez, S. Olmedo, A.B. Martin-
Diana, 2013. Electrolyzed water as novel technology to improve hygiene of drinking water for dairy ewes.
Research in Veterinary Science. 95 (3), 1169–1170. http://doi.org/10.1016/j.rvsc.2013.07.009.
Cardot, V., Y. Le Roux, S. Jurjanz, 2008. Drinking behavior of lactating dairy cows and prediction of
their water intake. Journal of Dairy Science. 91 (6), 2257–2264. http://doi.org/10.3168/jds.2007-0204.
Chen, T.Y., S.H. Kuo, S.T. Chen, D.F. Hwang, 2016. Differential proteomics to explore the inhibitory
effects of acidic, slightly acidic electrolysed water and sodium hypochlorite solution on Vibrio
parahaemolyticus. Food Chemistry. 194, 529–537. http://doi.org/10.1016/j.foodchem.2015.08.019.
Golher, D.M., B.H.M. Patel, S.H. Bhoite, M.I. Syed, G.J. Panchbhai, P. Thirumurugan, 2021. Factors
influencing water intake in dairy cows: a review. International Journal of Biometeorology. 65 (4), 617–625.
http://doi.org/10.1007/s00484-020-02038-0.
Liu, Y., C.Y. Wang, Z.X. Shi, B.M. Li, 2020. Optimization and Modeling of Slightly Acidic Electrolyzed
Water for the Clean-in-Place Process in Milking Systems. Foods. 9 (11). http://doi.org/10.3390/
foods9111685.
Murphy, M.R., 1992. Symposium - Nutritional Factors Affecting Animal Water and Waste Quality -
Water Metabolism of Diary Cattle. Journal of Dairy Science. 75 (1), 326–333. http://doi.org/10.3168/
jds.S0022-0302(92)77768-6.
Ni, L., W.C. Zheng, Q. Zhang, W. Cao, B.M. Li, 2016. Application of slightly acidic electrolyzed water
for decontamination of stainless steel surfaces in animal transport vehicles. Preventive Veterinary Medicine.
133, 42–51. http://doi.org/10.1016/j.prevetmed.2016.09.010.
Plaizier, J., D. Fulawka, A. Nikkhah, A. Kennedy, 2005. Effects of feeding time and forage to concentrate
ratio on water intake and drinking behavior of dairy cows. Journal of Animal Science. 83, 188–188.
Schutz, K.E., F.J. Huddart, N.R. Cox, 2019. Manure contamination of drinking water influences dairy
cattle water intake and preference. Applied Animal Behaviour Science. 217, 16–20. http://doi.org/10.1016/
j.applanim.2019.05.005.
Thorn, R.M.S., S.W.H. Lee, G.M. Robinson, J. Greenman, D.M. Reynolds, 2012. Electrochemically
activated solutions: evidence for antimicrobial efficacy and applications in healthcare environments. European
Journal of Clinical Microbiology & Infectious Diseases. 31 (5), 641–653. http://doi.org/10.1007/s10096-011-
1369-9.
Vargas Bello Perez, E., R. Dhakal, M.O. Nielsen, L. Ahrne, H.H. Hansen, 2020. Short communication:
Effects of electrochemically activated drinking water on bovine milk production and composition, including
chlorate, perchlorate, and fatty acid profile. Journal of Dairy Science. 103 (2), 1208–1214. http://doi.org/
10.3168/jds.2019-17510.
Wang, X.F., X.L. Hu, X. Gong, Y.X. Ban, 2016. UV-TiO2 photocatalytic disinfection and
photoreactivation of pathogenic bacterium in municipal wastewater. Journal of Central South University. 23
(12), 3115–3121. http://doi.org/10.1007/s11771-016-3377-5.
Wang, Y.H., Y.H. Wu, L.W. Luo, Q. Wang, X. Tong, Y. Bai, X.Y. Ni, H.B. Wang, G.Q. Chen, I. Nozomu,
Z. Chen, H.Y. Hu, 2021. Metagenomics analysis of the key functional genes related to biofouling aggravation
of reverse osmosis membranes after chlorine disinfection. Journal of Hazardous Materials. 410. http://doi.org/
10.1016/j.jhazmat.2020.124602.
Zhang, L., Y. Luo, W. Wang, Y. Sun, J. Zhang, M. Fatima, X. Jia, H.-J. Qiu, 2020. Efficient inactivation
of African swine fever virus by ozonized water. Veterinary Microbiology. 247, 108796. http://doi.org/
10.1016/j.vetmic.2020.108796.
∙ 90 ∙
Coupling Zirconium-based Metal-Organic Framework with

Titanium Dioxide for Photocatalytic Oxidation of Selected
Livestock Odorant
Zun Man a, Xiaorong Dai b, Yuxuan He a, Leiping Wang b, Dezhao Liu a,*
a
Key Laboratory of Equipment and Informatization in Environment Controlled Agriculture,
Ministry of Agriculture and Rural Affairs, College of Biosystems Engineering and Food Science,
Zhejiang University, Hangzhou, Zhejiang 310058, China
b
Center for Excellence in Regional Atmospheric Environment & Key Laboratory of Urban Environment
and Health, Institute of Urban Environment, Chinese Academy of Sciences, Xiamen, Fujian 361021, China
* Corresponding author. Email: dezhaoliu@zju.edu.cn
Abstract
Photocatalytic technology is considered a promising method for removing odor from livestock
and poultry farms. However, the currently used titanium dioxide (TiO2) photocatalyst has low
light utilization efficiency and poor adsorption performance for pollutants such as sulfur
compounds including dimethyl sulfide. The combination of metal-organic framework (MOF) and
inorganic semiconductors can construct a charge transfer path, improve the utilization efficiency
of electrons and holes, thereby improving photocatalytic performance. In this study, zirconium-
based metal-organic framework (UiO-66) was prepared by hydrothermal method and coated with
different amounts of TiO2 layer on its surface to prepare a series of UiO-66@TiO2 catalysts, and
its photocatalytic degradation performance on odor compounds was investigated. Experimental
results show that the UiO-66@TiO2 catalysts performance was better than a single UiO-66 or
TiO2. The four repeated use experiments showed only slight inactivation, proving its stable
photocatalytic performance. X-ray diffraction (XRD) and other characterization methods were
used to investigate its physical and chemical properties, and the improvement mechanism of UiO-
66@TiO2 photocatalytic performance was explored based on the characterization results and
photocatalytic activity data. The results show that thanks to the super large specific surface area
of UiO-66, the prepared catalyst can effectively adsorb pollutants and intermediate products to
achieve the purpose of deep mineralization. In addition, the TiO 2 layer and UiO-66 form a
heterojunction, which can effectively broaden the light absorption range and improve the
utilization of photo generated electron-hole pairs. This study can provide reference for the design
of core-shell structured TiO2@MOF and photocatalytic purification of odor gas from livestock.
Keywords: UiO-66, TiO2, photocatalytic oxidation, desulfurization
1. Introduction
Dimethyl sulfide (DMS) is one of the first eight key monitoring malodorous substances in
China, mainly from biological anaerobic fermentation process in sewage treatment plants (Han et
al., 2021), landfills (Liu et al., 2020), paper mills (Fontaine et al., 2021), livestock farms (Dalby
et al., 2018), etc. DMS has poor chemical stability, easy volatility and low olfactory threshold
(Long et al., 2017). It is extremely destructive to human central nervous system and blood
circulation system (Huang et al., 2021). After entering the troposphere, DMS gas easily reacts
with ozone, hydroxyl and other free radicals to form SO2 and methanesulfonic acid, thus making
an important contribution to the natural acidity of rainwater. At the same time, DMS can be
oxidized or photolyzed to aerosols in the stratosphere, which intensifies the production of chlorine
atoms and ozone destruction (Long et al., 2017). Photocatalytic technologies with mild reaction
conditions and ease of use are expected to be very promising in treating DMS. However, the
commonly used TiO2 photocatalysts have low light utilization efficiency and poor adsorption of
pollutants, leading to unsatisfactory degradation efficiency, which limits the application of this
technology.
Development of novel catalysts is an important way to solve the above problems. Metal
∙ 91 ∙
organic framework materials (MOFs) with semiconductor properties, large specific surface area,
and simple synthesis methods are gaining increasing attention and have shown excellent
performance in gas adsorption and separation (Chen et al., 2020a), photocatalytic hydrogen
production (Chen et al., 2020b), CO2 reduction (Cui et al., 2020), chemical synthesis (Shen et al.,
2013), and NOx degradation (Wang et al., 2018), etc. The huge specific surface of MOFs
facilitates the adsorption of intermediates and is expected to achieve deep degradation of
pollutants. However, the direct use of MOFs materials for catalysis is not satisfactory, mainly due
to the problems of low utilization of photogenerated electrons and insufficient stability.
It was shown that the compounding of MOFs with inorganic semiconductors could build
charge transfer paths and improve the efficiency of electron and hole utilization. Wang et al.
prepared the catalyst HKUST-1@TiO2 for the degradation of isopropanol, and HKUST-1 could
selectively capture intermediates, thus improving the mineralization rate of isopropanol (Wang et
al., 2016). Li et al. (2018) grew amorphous titanium dioxide in situ within the pores of various
MOFs, which greatly facilitated the migration and separation of photogenerated carriers. titanium
dioxide, which greatly promoted the migration and separation of photogenerated carriers and
improved the photocatalytic performance. Based on the above studies, in this study, UiO-
66@TiO2 catalysts with excellent activity and stability were prepared by hydrothermal
compounding of UiO-66 with TiO2, and the enhancement mechanism of its photocatalytic
performance was investigated in combination with series characterization.
2.1 Experimental drugs
Zirconium tetrachloride (ZrCl4), N,N-dimethylformamide (DMF), ethanol and methanol
reagents were purchased from China National Pharmaceutical Group Chemical Reagent
Company, terephthalic acid reagent was purchased from Sigma-Aldrich, and TiO2 was purchased
from Nanjing Xianfeng Nano Reagent Company. The purity of the above reagents was
analytically pure, and no subsequent purification was performed.
2.2. Synthesis of UiO-66
Synthesis of UiO-66 (Qiu et al., 2017): 0.600 g of ZrCl4 and 0.456 g of terephthalic acid were
dissolved in 150 mL of DMF, 0.475 mL of deionized water was added dropwise, stirred at room
temperature for 10 min, and then transferred to a hydrothermal reactor with a capacity of 200 mL
of PTFE liner and heated at 120 °C for 24 h. The reaction product was filtered, washed with DMF
and methanol, and dried under vacuum at 150 °C for 6 h.
2.3. Synthesis of UiO-66@TiO2 composites
Synthesis of UiO-66@TiO2: 0.2 g of UiO-66 and 0.6 g of TiO2 were sonicated and dispersed
into 400 mL of methanol solution and stirred for 15 min at room temperature. The resulting
solution was dried at 85 °C.
2.4. Characterization
The microstructure, surface properties and optoelectronic properties of the composites were
characterized by X-ray diffraction (XRD), scanning electron microscopy (SEM), N2 adsorption-
desorption (BET) and electrochemical workstations.
2.5. Photocatalytic oxidation tests
The photocatalytic reactor was a plexiglass box with a volume of 50 L. The UV light was
provided by an 8 W ultraviolet lamp tube installed at the center of the reactor. The glass dish with
coated photocatalyst was placed under the UV lamp with a constant distance of 15 cm. The
experimental procedure was divided into two stages: the dark adsorption stage and the
photodegradation stage.
∙ 92 ∙

3.1. Characterization of photocatalysts
Figure 1 shows the XRD patterns of all the samples prepared in the experiment. The two
strong peaks located at 5°–10° are the diffraction peaks of UiO-66. This is consistent with the
literature report and indicates the successful preparation of UiO-66 (Sha et al., 2015). After
compounding with TiO2, the peak intensity of UiO-66 decreased significantly, and the
characteristic peaks of TiO2 could be observed in the compounded samples.
Figure 1. XRD patterns of the samples.

Figure 2 shows the SEM images of UiO-66 and UiO-66@TiO2. The particle size of UiO-66
is about 400 nm, and it has an angular octahedral morphology. The large particles of UiO-66 are
closely surrounded by small TiO2 particles, and the lattice is clearly visible at the edges, indicating
that TiO2 has been loaded on the surface of UiO-66 after hydrothermal method.
Figure 2. SEM images of (a) UiO-66, (b) UiO-66@TiO2.

The N2 adsorption-desorption curve of UiO-66 shows a typical type I adsorption isotherm,
indicating that it has a microporous structure (Figure 3). The specific surface area / pore volume
of TiO2, UiO-66 and UiO-66@TiO2 are 123.5 m2 g-1 / 0.172 cm3 g-1, 350.9 m2 g-1 / 0.327 cm3 g-1
and 892.3 m2 g-1 / 0.562 cm3 g-1, respectively. The decrease in specific surface area and pore
volume of the composite material is caused by the coverage of TiO 2 on the surface.
∙ 93 ∙
Figure 3. N2 adsorption-desorption isotherms (a) and pore size distribution curves (b) of UiO-66,
TiO2, and UiO-66@TiO2 composite material.
According to the UV-Vis DRS (Figure 4a), UiO-66 has a shoulder peak at 283 nm with strong
absorption in the UV spectral region, and the absorption edge is significantly blue-shifted after
compounding with TiO2. Based on ahν = A(hν-Eg) n/2, the band gaps of UiO-66, TiO2 were
calculated as 3.90 and 3.33 eV.
According to the Mott-Schottky curve (Figure 4c), the flat-band potential of UiO-66 is -0.71
eV vs. Ag / AgCl (i.e., -0.51 eV vs. NHE). The difference between the bottom of the conduction
band and the flat-band potential is chosen to be 0.1 eV, which gives the position of the bottom of
the conduction band of UiO-66 as -0.61 eV. Depending on the forbidden band width, the positions
of HOMO and LOMO of UiO-66 can be calculated as 3.29 and -0.61 eV, respectively (vs. NHE).
Similarly, the positions of the valence and conduction bands of TiO2 can be calculated as 3.01 and
-0.29 eV, respectively (vs. NHE).
Figure 4. (a) UV-vis spectroscopy for the composite material; (b) The converted Tauc plot of
(αhν)2 versus photo energy (hν); (c) The Mott-Schottky diagrams of UiO-66.
3.2. DMS photodegradation performance of synthesized photocatalysts
Figure 5a shows the comparison of DMS photocatalytic degradation activity of different
samples. UiO-66 showed the best adsorption capacity, decreasing the concentration of DMS by
17% in the dark reaction stage. The DMS adsorption performance of TiO2 was improved from
3% to 9% for UiO-66@TiO2 by the doping of UiO-66. The improved adsorption performance
could contribute to the subsequent photocatalysis. In the photocatalytic stage, the concentration
of DMS only decreased by 7% under the blank control condition without any photocatalyst. This
indicated that DMS was not easily photolyzed. The DMS degradation efficiency of pristine TiO 2
was 61%. The rapid recombination of photogenerated electron-hole pairs is the main reason for
its low efficiency. The DMS degradation efficiency of pure UiO-66 was 44%. This indicates that
UiO-66 is also involved in the photocatalytic reaction. Compared with TiO 2 and UiO-66 alone,
the UiO-66@TiO2 composite significantly improved the degradation efficiency of DMS. After
∙ 94 ∙
only 80 min of photolysis, 99% of DMS was decomposed with a degradation efficiency of 13.48
mg g-1 h-1. The heterojunction formed by UiO-66 and TiO2 accelerated the transfer of
photogenerated electrons and holes and improved the material's photocatalytic activity. Figure 5b
shows the fitting results of the pseudo-first-order kinetics of DMS degradation. The model agrees
well with the experimental data, where the degradation rates of UiO-66@TiO2 (k=0.0459) are
17.8 and 7.1 times higher than those of pristine UiO-66 (k=0.0025) and TiO2 (k=0.0064),
respectively.
The samples after the reaction were subjected to repeated test experiments under the same
conditions, and there was a slight decrease in activity after four repetitions, presumably because
the degradation by-products generated by incomplete DMS mineralization were adsorbed on the
sample surface and covered the reaction active site. The activity of the sample could be largely
recovered after ethanol washing, which shows that the catalyst has good photochemical stability.
From Figure 5d, the photocurrent of the composite sample was significantly increased compared
with that of UiO-66 alone, indicating that the hybrid structure of UiO-66 and TiO2 was beneficial
to the rapid separation of photogenerated carriers.
Figure 5. Photocatalytic degradation of DMS (a), and the fitting results of the pseudo-first-order
kinetics of DMS degradation (b), recycling test for DMS degradation over UiO-66@TiO2 (c),
and transient photocurrent response of the prepared photocatalytic material (d).
3.3. Mechanisms for photocatalytic degradation
Combined with the results of the energy band position calculation, this paper speculates that
the increased activity of the composite sample may be due to the formation of type-I
heterojunction between the two components (as shown in Figure 6). Compared with TiO2, the
formation of migration paths reduces the complexation rate of electrons and holes because the
conduction band position of UiO-66 is higher and the valence band position is lower, so the
photogenerated electrons are transferred to the TiO2 surface and the photogenerated holes are
∙ 95 ∙
enriched on the TiO2 surface. The conduction band position of TiO2 is lower than the reduction
potential of O2 / ∙O2- (-0.046 eV vs. NHE), so ∙O2- is produced on its conduction band. Dimethyl
sulfate is eventually degraded under the attack of strongly oxidizing free radicals.
Figure 6. Scheme of the proposed degradation mechanism of DMS over UiO-66@TiO2.

4. Summary and Conclusions
In this paper, DMS degradation experiments under UV light were carried out for UiO-66,
TiO2 and UiO-66@TiO2 composite photocatalysts. The results showed that UiO-66@TiO2 had
the best performance, and the DMS degradation efficiency could reach 99% in 80 min and was
17.8 times higher than that of UiO-66 alone. The combination of UiO-66 and TiO2 could promote
carrier migration and separation and improve the utilization of photogenerated charges. Combined
with the analysis of the energy band structure, in the UiO-66@TiO cm3 g-1 system, TiO2 and UiO-
66 formed a strong interfacial interaction, and the transfer of electrons and holes occurred under
photoexcitation. Electrons combined with O2 in the conduction band of TiO2 to generate ∙O2-,
while holes were better utilized to produce more -OH. The increase in the generation of reactive
radicals achieved the improvement of the DMS degradation.
References
Chen, R., Z. Yao, N. Han, X. Ma, L. Li, S. Liu, H. Sun, S. Wang, 2020a. Insights into the Adsorption of
VOCs on a Cobalt-Adeninate Metal-Organic Framework (Bio-MOF-11). Acs Omega. 5 (25), 15402–15408.
http://doi.org/10.1021/acsomega.0c01504.
Chen, X., Y. Cai, R. Liang, Y. Tao, W. Wang, J. Zhao, X. Chen, H. Li, D. Zhang, 2020b. NH2-UiO-
66(Zr) with fast electron transfer routes for breaking down nitric oxide via photocatalysis. Applied Catalysis
B-Environmental. 267. http://doi.org/10.1016/j.apcatb.2020.118687.
Cui, L., X. Zou, Y. Liu, X. Li, L. Jiang, C. Li, L. Yang, M. Yu, Y. Wang, 2020. Dramatic enhancement
of photocatalytic H-2 evolution over hydrolyzed MOF-5 coupled Zn0.2Cd0.8S heterojunction. Journal of
Colloid and Interface Science. 577, 233–241. http://doi.org/10.1016/j.jcis.2020.05.023.
Dalby, F.R., M.J. Hansen, A. Feilberg, 2018. Application of Proton-Transfer-Reaction Mass
Spectrometry (PTR-MS) and S-33 Isotope Labeling for Monitoring Sulfur Processes in Livestock Waste.
Environmental Science & Technology. 52 (4), 2100–2107. http://doi.org/10.1021/acs.est.7b04570.
Fontaine, D., J. Eriksen, P. Sorensen, 2021. Sulfur from biogas desulfurization: Fate of S during storage
in manure and after application to plants. Science of the Total Environment. 754. http://doi.org/10.1016/
j.scitotenv.2020.142180.
Han, Z., R. Li, H. Shen, F. Qi, B. Liu, X. Shen, L. Zhang, X. Wang, D. Sun, 2021. Emission characteristics
and assessment of odors from sludge anaerobic digestion with thermal hydrolysis pretreatment in a wastewater
treatment plant. Environmental Pollution. 274. http://doi.org/10.1016/j.envpol.2021.116516.
Huang, S., C. Peng, G. Cao, Z. Wang, 2021. Corrosion Behavior of Copper-Nickel Alloys Protected by
BTA in Simulated Urban Atmosphere. Acta Metallurgica Sinica. 57 (3), 317–326. http://doi.org/10.11900/
0412.1961.2020.00206.
∙ 96 ∙
Li, X., Y. Pi, Q. Hou, H. Yu, Z. Li, Y. Li, J. Xiao, 2018. Amorphous TiO2@NH2-MIL-125(Ti)
homologous MOF-encapsulated heterostructures with enhanced photocatalytic activity. Chemical
Communications. 54 (15), 1917–1920. http://doi.org/10.1039/c7cc09072b.
Liu, Y., H. Yang, W. Lu, 2020. VOCs released from municipal solid waste at the initial decomposition
stage: Emission characteristics and an odor impact assessment. Journal of Environmental Sciences. 98, 143-
150. http://doi.org/10.1016/j.jes.2020.05.009.
Long, Y., S. Zhang, Y. Fang, Y. Du, W. Liu, C. Fang, D. Shen, 2017. Dimethyl sulfide emission behavior
from landfill site with air and water control. Biodegradation. 28 (5–6), 327–335. http://doi.org/10.1007/
s10532-017-9799-4.
Qiu, J., Y. Feng, X. Zhang, M. Jia, J. Yao, 2017. Acid-promoted synthesis of UiO-66 for highly selective
adsorption of anionic dyes: Adsorption performance and mechanisms. Journal of Colloid and Interface
Science. 499, 151–158. http://doi.org/10.1016/j.jcis.2017.03.101.
Sha, Z., H.S.O. Chan, J. Wu, 2015. Ag2CO3/UiO-66(Zr) composite with enhanced visible-light promoted
photocatalytic activity for dye degradation. Journal of Hazardous Materials. 299, 132–140. http://doi.org/
10.1016/j.jhazmat.2015.06.016.
Shen, L., S. Liang, W. Wu, R. Liang, L. Wu, 2013. CdS-decorated UiO-66(NH2) nanocomposites
fabricated by a facile photodeposition process: an efficient and stable visible-light-driven photocatalyst for
selective oxidation of alcohols. Journal of Materials Chemistry A. 1 (37), 11473–11482. http://doi.org/
10.1039/c3ta12645e.
Wang, H., T. Yu, X. Tan, H. Zhang, P. Li, H. Liu, L. Shi, X. Li, J. Ye, 2016. Enhanced Photocatalytic
Oxidation of Isopropanol by HKUST-1@TiO2 Core-Shell Structure with Ultrathin Anatase Porous Shell:
Toxic Intermediate Control. Industrial & Engineering Chemistry Research. 55 (29), 8096–8103. http://doi.org/
10.1021/acs.iecr.6b01400.
Wang, X., X. Zhao, D. Zhang, G. Li, H. Li, 2018. Microwave irradiation induced UIO-66-NH2 anchored
on graphene with high activity for photocatalytic reduction of CO2. Applied Catalysis B-Environmental. 228,
47–53. http://doi.org/10.1016/j.apcatb.2018.01.066.
∙ 97 ∙
Investigation of Dielectric Barrier Discharge Reactor for

Livestock Odorants Reduction
Dezhao Liu a,*, Xianwang Kong a,b, Zhen Cai a, Zhen Liu c, Keping Yan c
a
b
School of Petrochemical and Environment, Zhejiang Ocean University,
Zhoushan, Zhejiang 316022, China
c
College of Chemical and Biological Engineering, Zhejiang University,
*
Corresponding author. Email: dezhaoliu@zju.edu.cn
Abstract
Non-thermal plasma technology has potential for efficiently treating odorants from pig houses.
This lab-scale study was performed to investigate the removal of typical odorants in a dielectric
barrier discharge (DBD) reactor and its by-products formation. Four specific energy densities
(SED, 57.3, 98.5 114.6 and 146.0 J L-1) were evaluated at gas residence time of ca. 0.2 s. The
tested odorants were ammonia, organic acids (acetic acid, propionic acid, butyric acid, n-valeric
acid), sulfur compounds (hydrogen sulfide, dimethyl sulfide, dimethyl disulfide, dimethyl
trisulfide), phenols (p-cresol, 4-ethylphenol) and indoles (indole, skatole). Preliminary results
showed that NH3 and H2S were sufficiently removed (>95%) at all SED. As for other odorants,
increased SED in general had positive effect on compounds removal. The sulfur compounds
achieved >95% removal at all SED, and similar result was found for p-cresol. The indole and
skatole had relatively lower removal being 68–86% and 68–72%, respectively. The removal
patterns varied among organic acids. The butyric acid had the highest removal efficiency
with >92% at all SED, while propionic acid and n-valeric acid had only 30% removal at low SED
which was dramatically increased to 80–98% when SED was elevated to 114.6 and 146.0 J L-1. A
production of 27–265% of acetic acid was detected and increased SED could largely reduce the
production. In summary, DBD reactor could efficiently remove most of the livestock odorants,
while the formation of by-products (O3, N2O) and its environmental effects need to be considered
when using this technology.
Keywords: Non-thermal plasma, wasted gas treatment, nitrous oxide, ozone, gas removal
1. Introduction
Odor emissions from livestock industries are regarded as serious air pollution worldwide and
proper control of such emissions still remains challenging even in developed countries (Wang et
al., 2021). As the fast development of commercial and large-scale livestock farms, developing
countries, such as China, have put great efforts in minimizing the air pollution associated with
livestock farms. For example, a number of livestock farms have been forced to shut down in recent
years since these farms were not able to meet the national standard for odor emissions.
Swine farms are major component of livestock industries in China, and the ventilated air of
pig houses are characterized by low concentration and high complexity in gas components (Van
der Heyden et al., 2015). Biological treatment based on biofilm degradation (i.e., biofilter,
biotrickling filter) are widely accepted solutions for removal of livestock odorants, such as NH 3,
H2S, organic acids, as well as odors (Dumont et al., 2020). On successful operation, removal
efficiencies of >80% for NH3, >90% for organic acids and phenols, 60% for odors could be
achieved (Wang et al., 2021). Currently in China most of the new-built swine farms have adopted
biological methods to clean the off-gas from pig houses. However, biological methods may be
negatively affected by cold environment and fluctuations in inlet loading, etc. Results from Liu et
al. (2013) showed large variations (22–90%) of H2S removal, attributed to differences in air
loading rates and changes in biological activities. In addition, regular maintenance of
∙ 98 ∙
biofilter/biotrickling filter is probably not a simple task to the farm owners.

Non-thermal plasma (NTP) technology stands as an alternative option to coupe with VOCs
pollution for decades (Chung et al., 2019). In plasma, high energetic electrons induce molecule
excitation, ionization and dissociation; moreover, free radicals are produced to initiate chemical
oxidation. Through plasma chemical processes pollutants can thus be removed from a particular
air stream containing low- or medium-concentration odorous gases. NTP has been applied in full-
scale for remediation treatment of wasted gases from food industry, composting facility and
tobacco factories, etc., but only tentative tests have been conducted for those from animal houses.
Zhang et al. (1996) applied ferroelectric packed-bed plasma reactor to treat odorous gas collected
from a nursery pig room. Up to 95% NH3 decomposition efficiencies were obtained under a
combination of test conditions in the lab (e.g., voltages, gas residence time). As for odors, panelist
could not smell the original odor therefore the plasma reactor was effective in destroying the
odorous compounds. In a pilot-scale investigation (Andersen et al., 2013), plasma reactor with
corona-discharge could remove 100% of 3-methyl-1H-indole and >80% of indole from wasted
gas of a pig house under negative discharge modes. However, averaged removal efficiencies for
other compounds were relatively low ranging from -2% (acetone) to 33% (methanethiol). The
authors attributed such differences to the differential reactivity towards OH radicals and possibly
surface reactions induced by electron capture. The feasibility of using NTP to eliminate odorants
(such as H2S, acetic acid, dimethyl sulfide) has also been shown in former studies (Karatum and
Deshusses, 2016; Zhu et al., 2018), although gas concentrations were distinct from those of pig
houses. Therefore, NTP may present as a suitable choice under some circumstances for alleviating
air pollution of pig houses. Besides, the capacity of NTP system in terms of removing livestock
odorants remains to be explored and more research are warranted.
The specific energy density (SED) is a vital parameter for NTP reactor and normally higher
discharge power result in better gas removal (Chung et al., 2019). However, the unexpected
production of ozone (O3) or NO/NO2 from NTP reactor may also raise as SED increase (Lee et
al., 2020). In addition, N2O can be generated due to reactions (R1–R3) initiated by excited N2
molecules or other N-containing substances (Tang et al., 2017). This phenomenon is of particular
concern because N2O is a potent greenhouse gas contributing, together with CO2 and CH4, to the
global warming. Studies demonstrated that N2O formation from NTP reactors had a positive
response to increased SED as well as concentrations of NO/NO2 (Zhang et al., 2016).
In this laboratory study, we investigated the removal efficiencies of typical odorants present
in ventilation of pig houses in a dielectric barrier discharge (DBD) reactor powered by a pulse
power supply. The effects of applied voltage and inlet loading on gas removal were assessed.
Moreover, the production of N2O at varied levels of NH3 concentrations were determined and
compared.
2.1. Laboratory DBD reactor
This study was carried out at lab-scale, and NTP system based on a DBD reactor was set up.
The discharge mode in DBD can be highly efficient in energy use which is beneficial to plasma
chemical processes and the pollutant removal. Figure 1 gives an overview of the DBD reactor.
The body of the reactor (reaction chamber) was a quartz tube (ID: 10 mm, wall thickness: 2 mm,
length: 600 mm) with a solid stainless-steel rod (OD: 6 mm, length: 80 cm) inserted right in the
middle. The rod was fitted through a customized pipe joint and served as high voltage electrode.
The entire reactor was gas-tight except the air inlet and outlet and the discharge gap was 2 mm.
∙ 99 ∙
The outer surface of the quartz tube was wrapped by copper foil connected to earth electrode. The
cover length was ca. 450 mm therefore resulting an effective reaction volume of 22.6 cm3. A pulse
power supply was applied to provide the high voltage under different frequencies. The DBD
reactor was operated at frequency of 1 kHz under four voltages: 8.6, 10.6, 12.5 and 13.5 kV.
Figure 1. Schematic overview of the different components of the DBD reactor and gas line.
2.2. Gas preparations
In total 13 compounds were selected for this experiment (Table 1) as representatives of typical
odorants emitting from pig houses. The tested odorants concentrations were close to those found
in real situations.
Table 1. Compounds included in this experiment.
Category Compound name Chemical formula
N-containing compound Ammonia NH3
Acetic acid CH3COOH
Propionic acid CH3CH2COOH
Organic acids
Butyric acid C4H8O2
n-Valeric acid C5H10O2
Hydrogen sulfide H2S
Dimethyl sulfide C2H6S
Sulfur compounds
Dimethyl disulfide C2H6S2
Dimethyl trisulfide C2H6S3
p-cresol C7H8O
Phenols
4-ethylphenol C8H10O
Indole C8H7N
Indoles
Skatole C9H9N
The NH3 and H2S were supplied by certified gas cylinder (Hangzhou Jingong Co., Ltd, China)
and an air pump provided the air for dilution. For gas removal, NH3 at concentration levels of 6.2,
∙ 100 ∙
13.8, 19.6, 29.7 and 41.2 ppm were tested. The other gas compounds were prepared as follows:
firstly, pure chemicals of GC grade were added into PTFE tubes (OD: 6.35 mm, wall thickness 1
mm, length: 350 mm) with both ends tightly capped. Secondly these PTFE tubes were placed
inside the inner cavity of a temperature-controlled (60 ℃) oven which has one air inlet and one
air outlet. Thirdly air at constant flow rate entered the oven through the air inlet and captured the
chemical molecules that were volatilized from various PTFE tubes. By this method, the gas
concentrations in oven outlet were relatively stable after ca. 2 h period of equilibrium.
The air flow rate was modulated and controlled by mass flow controllers (CSA200, Beijing
Sevenstar Flow Co., Ltd, China). The air flow of the whole gas line was also check at the outlet
of DBD reactor using Gilibrator 2 (Sensidyne, USA) to verify the air tightness of DBD reactor
and the oven. In this experiment, air flow rate entering the oven was not changed and gas residence
time was calculated as 0.20 s.
2.3. Gas detection and measurements
Concentration differences between inlet and outlet air were used to calculate gas removal
efficiencies (%). Samplings of outlet gas were performed ca. 5 min after initiation of each test
condition. The NH3 concentrations (ppm in volume) were monitored by a portable NH3 detection
instrument (Shenzhen Anpal Technology Co., Ltd, China). The H2S and SO2 concentrations (ppb
in volume) were determined using a H2S-SO2 analyzer (model 450i, Thermo Fisher Scientific
Inc., USA) that gave results every minute. The O3 concentrations (ppb in volume) were recorded
by UV-100 O3 detector (Eco Sensors, USA) and the N2O (ppm in volume) were analyzed using
gas chromatography (Trace 1300 E, Thermo Fisher Scientific Inc., USA). A thermal imaging
camera enabled the measurement of surface temperature of DBD reactor.
The VOCs were analyzed using thermal desorption-gas chromatography-mass spectrometry
(TD-GC-MS) method. The detection system is composed of a fully automatic thermal desorption
instrument (TD100-xr, Markes, UK) and a gas chromatograph-mass spectrometer (7890B/5977B,
Agilent, USA). The chromatographic column is Agilent HP-INNOWax: model 19091N-133. The
minimum temperature of cold trap was set to -10 ℃ and the desorption time 5 min with maximum
temperature of 320 ℃. For GC the temperature of column was programmed as follows: initial 60
℃ for 1 min, then increased to 230 ℃ at 10 ℃ min-1, and maintained for 1 minute. Gas samples
were taken by adsorption tubes (C2-AAXX-5032, Markes, UK) through a sampling pump (GSP-
300FT-2, GASTEC, Japan) at flow rate of 100 mL min -1 and for 3 min. Standard curves of each
compound were prepared using chemicals of GC grade and used for quantification. At each
sampling, triplicate gas samples were collected.
The operating voltage applied on reactor was measured using an oscilloscope (Tektronix Inc.,
USA). The measured voltage and charge values are plotted against each other in a V−Q cyclogram,
better known as a Lissajous figure. The discharge power of the DBD reactor was calculated from
the enclosed area of a V−Q Lissajous figure.
Due to the nature of NTP related experiments, all tests were finished with one DBD reactor
and statistical analyses were not performed.
Preliminary results of this ongoing experiment are presented in this section.
3.1. Performance of DBD reactor
The corresponding discharge power calculated from Lissajous figure was 6.4, 11.0, 12.8 and
16.3 W, respectively, under the four voltages of 8.6, 10.6, 12.5 and 13.5 kV. The measured outlet
flow rate was around 6.6–6.8 L min-1, which was 92–94% of that of inlet. The SED was then
estimated as an average of 57.3, 98.5, 114.6 and 146.0 J L-1 (energy per unit of air volume),
respectively.
Generally, no NH3 was detected at the reactor outlet for all tested inlet NH3 concentration
∙ 101 ∙
levels even under the lowest SED indicating 100% removal. These removal rates were comparable
to those from Zhang et al. (1996), but much higher compared to 15–28% from Andersen et al.
(2013). This may be due to the difference in discharge mode. Figure 2 shows the changes of H2S
and concomitant evolution of SO2 at original H2S concentrations of 930 and 1854 ppb under SED
of 57.3 J L-1. The instrument results indicated that concentrations of outlet H2S decreased to nearly
zero after 8 or 10 min, but considering gas residence time of only 0.2 s, the response of the
instrument may have been delayed because of the long sampling tube (4 m) that enlarged the
effects of gas adsorption onto the tube wall. SO2 as a byproduct was gradually formed and finally
accounted for 54%–63% of the initial H2S concentration. When the SED increased the proportion
of transformed SO2 slightly increased to 65–69%. The residual H2S was probably converted into
elementary sulfur. In conclusion this DBD reactor can be efficient in destroying NH 3 and H2S,
but would release SO2 as a byproduct.
Figure 2. Dynamics of H2S and SO2 under SED of 57.3 J L-1 at initial H2S concentrations of 930
ppb (a) and 1854 ppb (b).
Inlet concentrations of VOCs were close to those found in ventilation air of pig houses (p-
cresol: 41.5 μg m-3, 4-ethylphenol: 39.6 μg m-3, indole: 28.4 μg m-3, skatole: 10.4 μg m-3), while
sulfur compounds had relatively higher concentrations than normal conditions (Figure 3). In all,
sulfur compounds and phenols were better removed (>90%) in this DBD reactor compared to
organic acids and indoles. Skatole had the removal efficiencies between 68–72%. These results
were different from those of Andersen et al. (2013) where nearly 100% removal of indole and
skatole were reached, much higher than other substances. As for organic acids, the removal
efficiencies and responses to increased SED were not consistent. The butyric acid had removal
efficiency of >92% that were not influenced by levels of SED. For propionic acid and n-valeric
acid, the effects of increasing SED on gas removal were obvious, changing from ca. 30 to ca. 90%
when SED shifted from the lowest to the highest. Acetic acid was the only compound that showed
negative removal in this DBD reactor, i.e., larger concentrations in outlet relative to inlet. The
production of acetic acid may result from the degradation of long-chain acids (propionic acid,
∙ 102 ∙
butyric acid, n-valeric acid). When SED increased, the production of acetic acid decreased
indicating that SED larger than 114.6 J L-1 would be optimal for this DBD reactor to destroy the
acetic acid. Deposits sometimes occurred inside the DBD reactor wall (Karatum and Deshusses,
2016), but this phenomenon was not observed in the present study possibly due to the low
concentrations of VOCs tested. Overall, this DBD reactor achieved higher gas removal compared
to that in Andersen et al. (2013), suggesting the potential of using NTP systems to remediate
wasted gas from pig houses.
7.0
100
6.0
Initial concentration (mg m )
-3
Removal efficiency (%)

5.0 0
Concentration
4.0 SED 57.3 J L-1
SED 98.5 J L-1
-100
3.0 SED 114.6 J L-1
SED 146.0 J L-1
2.0
-200
1.0
0.0 -300
ol
ol
ole
le
cid
cid
cid
cid
de
de
lfid
en
res
ato
ulfi
ulfi
Ind
ca
ca
ic a
a
lph
isu
p-c
Sk
nic
yl s
ris
eti
tyri
ler
thy
yl d
yl t
pi o
eth
Ac
Bu
4-e
eth
eth
n-V
Pro
Di m
Di m
Di m
Figure 3. Inlet VOCs concentrations and removal efficiencies under different SED.
3.2. Assessment of byproducts from DBD reactor
Ozone is a regular byproduct escaping from NTP systems and excessive O 3 in atmosphere
may cause air pollution such as photo-chemical smog. Therefore, measurement of O3 formation
is an important step to evaluate the environmental impact of NTP systems. Since O 3
decomposition would be promptly accelerated when the ambient temperature increases, higher
SED may not necessarily end up with higher O3 production. The outlet O3 concentrations were
initially highest but then declined overtime until reaching steady levels except for those under
lowest SED (Figure 4). The SED of 98.5 J L-1 had the highest O3 concentration (630 ppm)
followed by the order of 508 ppm (114.6 J L-1), 420 ppm (57.3 J L-1), and 332 ppm (146.0 J L-1).
In terms of reactor surface temperature, it was intimately correlated with SED, i.e., higher SED
led to higher surface temperature ranging from 60–127 ℃. The O3 production in this experiment
were all much higher than the combined concentrations of tested odorants under different SED.
Another byproduct of NTP systems can be N2O that has not been frequently brought up. The
removal of odorants would deplete O or radicals that are involved in N2O production. Besides the
NH3 may be oxidized into NO or NO2 by NTP that would possibly affect the N2O formation. With
room air containing no odorants, the outlet N2O concentration was already ten times higher than
the background room air under the lowest SED (Figure 5). Besides a trend was found that elevated
SED resulted in larger production of N2O, in accordance with previous studies (Tang et al., 2017;
Zhang et al., 2016). It was also claimed that presence of O2 was favorable to N2O production. The
O2 content in off-gas of animal houses were always close to atmospheric air (ca. 21%), thus the
potential of N2O production from DBD reactors treating wasted gas of pig houses may be
relatively high. When VOCs were present, production of N2O were basically not changed. But
once 19.6 ppm NH3 were introduced, the N2O production under all SED were increased
suggesting that NH3 or its conversion products may play a role in N2O formation. Dependence of
N2O production on inlet NH3 concentrations were evaluated under SED of 114.6 J L-1. The
∙ 103 ∙
increase of N2O production were clearly not proportional to the increase of inlet NH 3
concentrations. This phenomenon needs further research in order to understand the effects of NH3
addition on N2O production.
1000
57.3 J L-1
98.5 J L-1
O3 concentrations (ppm)
800
114.6 J L-1
146.0 J L-1
600
400
200
0
50 100 150 200 250 300 350 400 450 500 550 600
Discharge time (s)

Figure 4. Outlet O3 concentrations of DBD reactor along with discharge time.
15 (a) (b)
Background
Room air
VOCs
N2O production (ppm)
12
NH3
0
Background 57.3 98.5 114.6 146 6.2 13.8 19.6 29.7 41.2
SED (J L-1) inlet NH3 concentration (ppm)
Figure 5. N2O production from DBD reactor with inlet gas containing room air, VOCs and NH 3
(a), respectively. N2O production at SED of 114.6 J L-1 with different inlet concentrations of
NH3 (b).
4. Conclusions
The DBD reactor in the present study was highly efficient in removing NH3 and H2S but
would release SO2 as a byproduct. Overall satisfactory removal (>90%) of odorants could be
achieved but relatively high SED were required for compounds such as propionic acid and n-
valeric acid, and probably even higher for compounds like acetic acid. The formation of N 2O from
DBD reactor is of great concern in the global context of carbon neutrality and controlling of
greenhouse gas emissions.
Acknowledgements
The research was financially supported by the National Natural Science Fund of China (No.
31672468) and Thousand Talents Program (Youth Project 2016).
∙ 104 ∙
References
Andersen, K.B., J.A. Beukes, A. Feilberg, 2013. Non-thermal plasma for odour reduction from pig
houses-A pilot scale investigation. Chemical Engineering Journal. 223, 638–646. http://doi.org/10.1016/
j.cej.2013.02.106
Chung, W.C., D.H. Mei, X. Tu, M.B. Chang, 2019. Removal of VOCs from gas streams via plasma and
catalysis. Catalysis Reviews-Science and Engineering. 61 (2), 270–331. https://doi.org/10.1080/01614940.
2018.1541814
Dumont, E., M. Poser, P. Peu, A. Couvert, 2020. Biotrickling filters for the removal of gaseous ammonia
emissions from livestock facilities. Theoretical prediction of removal efficiency and experimental validation.
Chemical Engineering Journal. 402, 126188. https://doi.org/10.1016/j.cej.2020.126188
Karatum, O., M. A. Deshusses, 2016. A comparative study of dilute VOCs treatment in a non-thermal
plasma reactor. Chemical Engineering Journal. 294, 308–315. http://doi.org/10.1016/j.cej.2016.03.002
Lee, B., D.W. Kim, D.W. Park, 2020. Decomposition of heptane by dielectric barrier discharge (DBD)
plasma reactor having the segmented electrode: comparison of decomposition mechanisms to toluene. Plasma
Chemistry and Plasma Processing. 40 (1), 61–77. https://doi.org/10.1007/s11090-019-10024-7
Liu, D., M.M. Lokke, A.L. Riis, K. Mortensen, A. Feilberg, 2014. Evaluation of clay aggregate
biotrickling filters for treatment of gaseous emissions from intensive pig production. Journal of Environmental
Management. 136, 1–8. https://doi.org/10.1016/j.jenvman.2014.01.023
Tang, X., J. Wang, H. Yi, S. Zhao, F. Gao, Y. Huang, R. Zhang, Z. Yang, 2017. N2O formation
characteristics in dielectric barrier discharge reactor for environmental application: Effect of operating
parameters. Energy & Fuels. 31 (12), 13901–13908. https://doi.org/10.1021/acs.energyfuels.7b02428
Van der Heyden, C., P. Demeyer, E.I.P. Volcke, 2015. Mitigating emissions from pig and poultry
housing facilities through air scrubbers and biofilters: State-of-the-art and perspectives. Biosystems
Engineering. 134, 74–93. https://doi.org/10.1016/j.biosystemseng.2015.04.002
Wang, Y. C., M.F. Han, T.P. Jia, X.R. Hu, H.Q. Zhu, Z. Tong, Y.T. Lin, C. Wang, D.Z. Liu, Y.Z. Peng,
G. Wang, J. Meng, Z.X. Zhai, Y. Zhang, J.G. Deng, H.C. Hsi, 2021. Emissions, measurement, and control of
odor in livestock farms: A review. Science of the Total Environment. 776, 145735.
https://doi.org/10.1016/j.scitotenv.2021.145735
Zhang, R., T. Yamamoto, D.S. Bundy, 1996. Control of ammonia and odors in animal houses by a
ferroelectric plasma reactor. IEEE Transactions on Industry Applications. 32 (1), 113–117.
Zhang, Y., X. Tang, H. Yi, Q. Yu, J. Wang, F. Gao, Y. Gao, D. Li, Y. Cao, 2016. The byproduct
generation analysis of the NOx conversion process in dielectric barrier discharge plasma. RSC Advances. 6
(68), 63946–63953. https://doi.org/10.1039/c6ra08488e
Zhu, T., R. Wang, W. Bian, Y. Chen, W, 2018. Jing. Advanced oxidation technology for H 2S odor gas
using non-thermal plasma. Plasma Science and Technology. 20 (5), SI. https://doi.org/10.1088/2058-
6272/aaae62
∙ 105 ∙
Air Resistance of Pigs in the Group Based on OpenFOAM:

Influence of Stocking Density and Live Weight
Xuefei Wu a,b, Hao Li a,b,*, Zhengxiang Shi a,b
a
b
Key Laboratory of Agricultural Engineering in Structure and Environment of Ministry of Agriculture and
* Corresponding author. Email: leehcn@hotmail.com
Abstract
Stocking density and live weight are the main factors determining the air resistance in animal
occupied zone (AOZ). To study the impact of live weight and stocking density on air resistance,
OpenFOAM was adopted. Firstly, tube bundle with three boundary layer treatments (10 boundary
layers with 0.04 m total thickness, 2 boundary layers with 0.04 m total thickness, and 10 boundary
layers with 0.02 m total thickness) and turbulence models (standard k-ϵ, realizable k-ϵ, RNG k-ϵ,
standard k-ω, and SST k-ω) were simulated to ensure the accuracy of OpenFOAM. Secondly, pig
groups of different live weights (30, 45, 60, 75, 90, 105, 120 and 135 kg) and stocking densities
(0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 2.0, and 2.6 pig m-2) were used to study the air resistance in the
horizontal and vertical directions of AOZ. The results indicated that 10 boundary layers with 0.04
m total thickness and realizable k-ε model showed the best performance. The air resistance
increased with the stocking density or live weight increasing and the fluctuation occurred in pig
groups of 45 kg and 105 kg in the horizontal direction. The proportion of impact on air resistance
was the live weight (about 70%), the interaction (15%), and the stocking density (10%) in the
vertical direction. In the horizontal direction, the stocking density influenced air resistance mostly
(about 60%) and the body weight influenced mostly (about 40%) in light (30–60 kg) and heavy
body weight (75–135 kg), respectively.
Keywords: CFD, pressure drop, stocking density, live weight
1. Introduction
Animals as an obstacle influence the airflow in the ventilation system (Bjerg et al., 2008; Iqbal
et al., 2021). However, the air resistance in the zone occupied by animals is uncertain. Meanwhile,
the animal occupied zone (AOZ) is determined by the animals, and the live weight and stocking
density are the main two keys for AOZ. The average body weights of finishing pigs are from 15
to 135 kg. On the other hand, pigs are often stocked at a high density for efficient management.
However, research concerning the stocking density and body weight of pig groups on airflow
resistance are scarce and needed to be studied further.
Field experiment, lab study, and computational fluid dynamics (CFD) are available to
calculate the pig group’s air resistance. And considering the drawbacks of field experiment (the
limited measurement points), and lab experiment (difficulty in representing the real phenomenon),
therefore, CFD was used in this study for its ability to analyze complex airflow patterns under
arbitrary conditions. Furthermore, CFD software is an important factor that influences the
accuracy. The widely used commercial CFD software provides convenient user interfaces but
cannot be customized easily (Hong et al., 2017). On the other hand, open-source software
packages have access to the source code which means the possibility of customizing it for produce
spin-off program applications (Seo et al., 2015). However, the open-source CFD has rarely been
used for studies on agriculture engineering. So, OpenFOAM was used in our study as a base for
further spin-off especially in agricultural engineering. Meanwhile computing performance should
be investigated for open-source packages. Determining the appropriate mesh size is always
challenging in OpenFOAM. Therefore, the first key step is to determine the mesh treatment of
OpenFOAM. On the other hand, there is no turbulence model suitable for different cases. So
different turbulence models were evaluated to establish the most appropriate model in this study.
∙ 106 ∙
The objectives of this study were to 1) investigate the boundary layer treatment of
OpenFOAM; 2) evaluate the performance of different turbulence models; 3) test the impact of
live weight and stocking density on air resistance; and 4) establish the weight-density-related
regressive equations for widely use.
2.1. Case studies based on tube bundles
Firstly, tube bundles geometrically-like pig groups were simulated to ensure the accuracy of
OpenFOAM. Then semi-experimental equations of tube bundle available in the literature as the
calculated pressure drop were compared with the simulated results.
2.1.1. Tube bundles for validation
The tube bundle with fifteen tube rows was in a staggering array, which was 16.55 m (L) ×
1.70 m (W) × 0.20 m (H) (Figure 1). The distance between the tubes was both 0.45 m in the
horizontal and vertical directions. The thickness of the tube bundle was 0.20 m and the
characteristic diameter was 0.40 m, which was close to the body diameter of 50 kg mass pigs.
2.1.2. Meshing strategy in OpenFOAM
To study the mesh treatment of boundary, three treatments of boundary layers were used: a)
10 boundary layers with 0.04 m total thickness; b) 2 boundary layers with 0.04 m total thickness;
c) 10 boundary layers with 0.02 m total thickness shown in Figure 2. A grid-sensitivity analysis
was conducted with k-ε model at 1.5 m s-1 and three mesh resolutions of the refinement region
were tested: level 1 (77,274 meshes), level 2 (220,418 meshes), and level 3 (1,596,726 meshes),
with 0.025, 0.0125, and 0.00625 m mesh sizes.
Figure 1. Geometry of tube bundle in the virtual wind tunnel.
(a) (b) (c)

Figure 2. Three boundary layer treatments: (a) 10 boundary layers (0.04 m); (b) 2 boundary
layers (0.02 m); (c) 10 boundary layers (0.02 m).
2.1.3. Turbulence modelling
Four common RANS turbulence models were compared, e.g., standard k-ε, RNG k-ε,
realizable k-ε, standard k-ω, and SST k-ω model. The four models were simulated with three
boundary layer treatments (boundary layer treatments of a, b, and c).
2.1.4. Boundary conditions and numerical method
Detailed boundary conditions are listed in Table 1. The simulations were solved in a steady-
state by the simpleFoam solver of OpenFOAM and the pressure field and the velocity field were
coupled by the SIMPLE algorithm. GAMG was used for solving the pressure equation.
2.1.5. Empirical correlation of tube bundles
The pressure drops from empirical correlations of tube bundle were calculated following the
empirical equations:
2𝑓′ 𝐺𝑚𝑎𝑥 2 𝑁 𝜇𝜔 0.14
∆p = ( ) (1)
𝜌 𝜇𝑏
∙ 107 ∙
where Gmax is the mass velocity at min flow area, kg m-2 s-1; ρ is density at free-stream conditions,
kg m-3; N is number of transverse rows; μb is average free-stream viscosity, N s m-2; μω is wall
viscosity, N s m-2.
The empirical friction factor f’ for staggered tube arrangements is:
0.08
f ′ = {0.25 + −0.16
}𝑅𝑒𝑚𝑎𝑥 (2)
[𝑆𝑛 −𝑑/𝑑]1.08
where Sn is the distance between the center of the two adjacent tubes, m; d is the diameter of the
tube bundles, m.
Table 1. Boundary and initial conditions for the simulation of tube bundle.
Boundaries Boundary conditions
airflow speeds 0.2, 0.5, 1.0, 1.5, 2.0, 2.5 and 3.0 m s -1，
Inlet
turbulence intensity of 5%
Outlet 0 Pa in pressure outlet
Tube bundles surface Smooth wall
Other boundaries Symmetry boundary
2.2. Studies based on pig models in the group
2.2.1. Live weight and stocking density of pig group
The live weights and stocking density in the level of 0.6–2.6 pig m-2 are listed in Table 2.
Table 2. Stocking density and live weight of pig groups.
Live weight, kg
30 45 60 75 90 105 120 135
0.6 0.6 0.6 0.6
0.8 0.8 0.8 0.8 0.8 0.8 0.8 0.8
1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0
Stocking density, 1.2 1.2 1.2 1.2 1.2 1.2 1.2
pig m-2 1.4 1.4 1.4 1.4 1.4
1.6 1.6 1.6 1.6 1.6
2.0 2.0 2.0
2.6 2.6 2.6
2.2.2. Geometry and computational domain
Pig groups were placed into a virtual wind tunnel. Each pig groups have three uniform
distributions, take the 90 kg pig group for example in Figure 3 left. The horizontal direction (x
and y directions) of AOZ was homogeneous and only the x and z directions (vertical direction) of
AOZ were simulated. The wind tunnel were 10.0 m long and 2.0 m × 0.8 m in cross-section (x-
direction) and 13.0 m long and 2.0 m × 2.5 m in cross-section (z-direction) in Figure 3 middle and
right.
(a) x-direction (b) z-direction

Figure 3. Left: Three distributions (90 kg and 1.4 pig m-2). Middle and right: Wind tunnel for
different directions of AOZ.
∙ 108 ∙
2.2.3. Mesh generation and Turbulence modeling

It was the same as that described in section 2.1.2.
The turbulence modeling was the best performance choose in section 2.1.4.
2.2.4. Boundary conditions and numerical method
The boundary conditions were generally the same as those in section 2.1.4. Numerical method
and convergence criteria were also kept unchanged.
Parameter definitions and calculations are shown in Eq. (3). Darcy friction factor f, simply
called friction factor, is the description of frictional losses and the formulae are:
2g×D×∆p
f= (3)
𝐿×𝜌×𝜇2
where, L is the length of the object, m; D is the characteristic diameter, m; ρ is the density of the
fluid, kg m-3; μ is the velocity of the fluid, m s-1; g is the gravity, m s-2.
3. Result and discussion
3.1. Modeling validation
In total, 168 cases were simulated based on tube bundle for modeling validation which was
dependent on the semi-experimental equations of tube bundle available in the literature.
3.1.1. Grid independence evaluation based on mesh size and boundary layer treatment
As can be observed, meshes in level 1 and level 2 with average error less than 11% and the
“a” wall treatment shows best results with relative error 2.6% and 3.6%, respectively, in mesh
level 1 and level 2 (Table 3). However, the relative error of Level3 increased about 10% compared
with the other two mesh levels. Extremely small meshes decreased the accuracy and refinement
meshes were not always better compared with coarse meshes. So, Level 2 was used for the next
simulation.
Table 3. Pressure drops with different boundary treatments with k-ε model at 1.5 m s-1.
Mesh level of refinement region Calculated pressure
Wall treatment
Level 1 Level 2 Level 3 drop, Pa m-1
a 2.630 2.603 2.130
b 2.465 2.412 1.961 2.701
c 2.494 2.472 1.963
*Wall treatment a means 10 boundary layers with 0.04 m total thickness; b means 2 boundary
layers with 0.04 m total thickness; c means 10 boundary layers with 0.02 m total thickness.
3.1.2. Influence of the turbulence models
The results of four turbulence models with three different boundary treatments were shown
in Figure 4. Both standard k-ε and realizable k-ε performed well (relative error <15%) and
realizable k-ε model performed best. This was because compared to k-ε model, realizable k-ε
model has improved the prediction for boundary layer under strong pressure gradients. Hence
realizable k-ε model with “a” boundary layer treatment (relative error 6.08%) was used for further
study. RNG k-ε and k-ω SST underestimated the results that the relative errors decreased from -
55.05% to -26.40%, from -51.05% to -34.31% at 0.2 and 3.0 m s-1, in “a” boundary treatment. it
is maybe the surface mesh density was not fine enough with denser mesh that SST k-ω model
required. As for k-ω model, the pressure drop ranged from -3.59% to 51.61% and performed better
at low air velocities (0.2 and 0.5 m s-1) with small relative errors (-3.59% and 12.29%) in “b”
boundary treatment. But they performed rather worse and significantly overestimated the pressure
drop with the air velocity increased.
3.1.3. Mesh strategy and turbulence on the Y plus
The y+ of the tube bundles was calculated in Table 4. The different turbulence has little
influence on y+ and y+ becomes bigger with the velocity increases. In b treatment, y+ is always
∙ 109 ∙
larger than the other two wall treatments due to the difference of the first grid size 0.852 × 10-3,
0.426 × 10-3, and 0.0174 m in a, b, and c treatments. The existing wall functions have been
modified to provide the accurate result, so wherever the position of the first cell center in
OpenFOAM, y+ was acceptable.
standard k-ε realisable k-ε RNG k-ε standard k-ω k-ω SST calculated
18 20
14 18
16
14 12 16
14
12 10
Δp, Pa m-1
12
Δp, Pa m-1
ΔP, pa m-1
10 8 10
8
6 8
6 6
4
4 4
2 2 2
0 0 0
0.0 1.0 2.0 3.0 0.0 1.0 2.0 3.0 0.0 1.0 2.0 3.0
Air velocity, m s-1 air velocity, m s-1 Air velocity, m s-1
(a) (b) (c)
Figure 4. Pressure drops in three boundary layer treatments.
Table 4. The y+ in different wall treatments.
Turbulence 0.2 (m s-1) 1.5 (m s-1) 3.0 (m s-1)
model a b c a b c a b c
standard k-ε 0.14 4.22 0.12 0.27 13.05 0.17 0.42 18.32 0.21
realizable k-ε 0.14 4.22 0.12 0.26 13.04 0.17 0.39 18.30 0.20
RNG k-ε 0.122 4.43 0.12 0.33 13.74 0.19 0.50 19.31 0.26
standard k-ω 0.12 4.26 0.12 0.38 13.21 0.22 0.50 18.84 0.30
k-ω SST 0.12 4.50 0.12 0.47 14.03 0.24 0.73 18.55 0.38
* note “a” means the total thickness 0.04 m with 10 boundary layers; “b” means the total
thickness 0.04 m with 2 boundary layers; “c” means the total thickness 0.02 m with 10 boundary
layers.
3.2. Air resistance of AOZ influenced by stocking density
The friction factors of three distributions were averaged and then the averaged friction factor
of 258 sets in total under different stocking densities were analyzed. Pig groups of different
stocking densities were simulated. In z-direction, f increased with the stocking density increasing.
This is because that AOZ is filled with pigs as obstacles and the remaining space is the porosity
through which the airflow flows. Taking pig group of 45 kg for example, stocking density
decreased from 2.6 to 1.0 animal m-2 and AOZ porosity increased from 75.85% to 90.7%, so
friction factor increased from 8.713 to 0.655. On the other hand, the range of friction factor was
large from less than 1.0 to above 8.0. The biggest friction factor gap of 45 kg pig group is from
0.655 to 8.713 (1.0 to 2.6 pig m-2). Meanwhile, in high stocking, friction factor increases faster
than that in low stocking density which can be seen from the growth trend in Figure 5. In x-
direction, friction factor has an increasing trend with the stocking density increasing, however,
fluctuations occurred in pig groups of the 45 and 105 kg. and under the joint influence of global
and local porosity, the friction factor fluctuation occurred.
∙ 110 ∙
4
30 kg
8 3
45 kg
3
Friction factor
Friction factor
60 kg 6
2
75 kg 4 2
90 kg 1
2
105 kg 1
120 kg 0 0
0.4 0.8 1.2 1.6 2.0 2.4 0.4 0.8 1.2 1.6 2.0 2.4
135 kg Stocking density, pig m-2 Stocking density, pig m-2
(a) z-direction (b) x-direction
Figure 5. Friction factor in two directions.

Table 5. Stocking density and friction factor regression.
Live weight Variety range, pig m-2 Z direction R2 X direction R2
30 1.0-2.6 f=0.3939x2.0414 0.981 f=0.4129x0.3294 0.969
45 1.0-2.6 f=0.5139x2.7198 0.961 F=0.5694x0.747 0.927
60 1.0-2.0 f=0.6657x2.7697 0.962 F=0.9422x0.7007 0.898
75 0.8-1.6 f=1.0623x2.5554 0.960 F=1.1507x1.2177 0.900
90 0.6-1.6 f=1.5169x2.5621 0.925 F=1.5002x1.1385 0.866
105 0.6-1.4 f=1.7031x2.1222 0.980 F=1.6339x0.4383 0.621
120 0.6-1.4 f=2.057x2.393 0.983 F=2.686x0.7021 0.937
135 0.6-1.2 f=2.4452x2.4939 0.988 F=2.3264x1.4295 0.931
3.3. Air resistance of AOZ influenced by live weight
Exponential fitting curves with stocking densities are shown in Table 7. In z-direction, friction
factor increased with the live weight in Figure 6. In small live weight, friction factor increased
slowly and in high live weight, increased obviously. For example, in 1.2 pig m-2, friction factor
increased from 0.442 to 0.655 (light weight from 30 to 60 kg) and 1.817 to 2.307 (heavy weight
from 105 to 135 kg). In x-direction, friction factor had an increasing trend with the weight
increasing in stocking density of 1.4, 1.6, 2.0, and 2.6 pig m-2. There were fluctuations in the low
stocking density 0.6, 0.8, 1.0, and 1.2 pig m-2. The fluctuation occurred in the body weight of 105,
120, and 135 kg. This is because in low stocking density, the heavy pigs with large volume were
places crowd in a relatively small area and their orientation is limited and when the certain wind
direction of the pig’s trunk axis caused less air disturbance around the body which means the less
air resistance occurred (Cao et al., 2021).
Table 6. Live weight and friction factor regression.
Live Variety range,
Z direction R2 X direction R2
weight pig m-2
30 1.0-2.6 f=0.3939x2.0414 0.981 f=0.4129x0.3294 0.969
45 1.0-2.6 f=0.5139x2.7198 0.961 F=0.5694x0.747 0.927
60 1.0-2.0 f=0.6657x2.7697 0.962 F=0.9422x0.7007 0.898
75 0.8-1.6 f=1.0623x2.5554 0.960 F=1.1507x1.2177 0.900
90 0.6-1.6 f=1.5169x2.5621 0.925 F=1.5002x1.1385 0.866
105 0.6-1.4 f=1.7031x2.1222 0.980 F=1.6339x0.4383 0.621
120 0.6-1.4 f=2.057x2.393 0.983 F=2.686x0.7021 0.937
135 0.6-1.2 f=2.4452x2.4939 0.988 F=2.3264x1.4295 0.931
∙ 111 ∙
10 4
0.6 pig m-2
3
0.8 pig m-2 8
Friction factor
Friction factor 3
1.0 pig m-2
6 2
1.2 pig m-2
4 2
1.4 pig m-2
1
1.6 pig m-2 2
1
2.0 pig m-2
0 0
2.6 pig m-2 15 30 45 60 75 90 105 120 135 15 30 45 60 75 90 105 120 135
Live weight, kg Live weight, kg
(a) z-direction (b) x-direction
Figure 6. Friction factor under different live weights in two directions.
3.4. Interaction of live weight and density on air resistance
The two-factor variance analysis was conducted to know the sensitivity of stocking density
and live weight in the given range in Table 7. In z-direction, as for stocking density, live weight,
and the stocking density and live weight interaction, F was large than Fcrit and the P-value was
smaller than the significance level 0.01, so it indicated that stocking density, live weight, and their
interaction have significant effects on the air resistance. By the comparison of η2 in the z-direction,
about 10% of air resistance was accounted for by stocking density, about 70% by live weight, and
about 15% by their interaction.
In x-direction, as for stocking density and live weight, F was large than Fcrit and the P-value
was smaller than the significance level 0.01, so stocking density, live weight has significant effects
on the air resistance. However, as for the stocking density and live weight interaction, the F was
smaller than Fcrit and the P-value was larger than the significance level 0.01 in the situations, so
it indicated that their interaction has no significant effect on the air resistance. By the comparison
of η2 in the z-direction, about 10% of air resistance was accounted for by stocking density, about
70% by live weight, about 15% by their interaction. In x-direction, in live weight of 30-60 kg, the
stocking density plays the most important role (the biggest η2 was 56.40%) and in live weight of
75-135 kg, the live weight did (the biggest η2 42.24%).
Table 7. Two-factor analysis variance between body weight and stocking density.
AOZ Weight Source of
F P-value Fcrit η2
Direction Levels, kg variance
Stocking density 51.24 2.94×10-11 3.26 10.86%
30–60 Live weight 129.83 5.48×10-22 2.48 68.82%
Interaction 15.57 2.83×10-10 2.11 16.50%
z-direction
75–135 Live weight 129.83 5.48×10-22 2.48 75.40%
Interaction 15.57 2.83×10-10 2.11 11.28%
30–60 Live weight 6.84 1.41×10-4 2.48 19.14%
Interaction 0.77 0.65 2.11 4.32%
x-direction
75–135 Live weight 30.96 1.98×10-15 2.37 42.24%
Interaction 1.10 0.38 1.75 5.98%
∙ 112 ∙
4. Conclusions
The following conclusions can be drawn:
1) OpenFOAM in 10 boundary layers with 0.04 m total thickness and realizable k-ε model
shows the best performance with an average relative error of 6.80%.
2) In both z-direction and x-direction, the air resistance increased with the stocking density
or live weight increasing. In high stocking, air resistance increases faster than that in low stocking
density. In z-direction, the proportion of impact on air resistance is the live weight (about 70%),
the stocking density (about 10%), and their interaction (about 15%).
3) In light body weight (30-kg), the stocking density influences air resistance mostly (about
60%), in heavy body weight, the body weight was influenced mostly (about 40%). And the
interaction has no significant influence on the air resistance in the x-direction.
Acknowledgements
The support of National Key Research and Development Plan of China-Integration and
demonstration of production technology, environment and engineering technology for
commercial pigs (2018YFD0501103) to the first author (Xuefei Wu) is very appreciated.
References
Bjerg, B., G. Zhang, P. Kai, 2008. CFD investigation of a partly pit ventilation system as method to
reduces ammonia emission from pig. Central theme, technology for all: sharing the knowledge for
development. Proceedings of the International Conference of Agricultural Engineering, XXXVII Brazilian
Congress of Agricultural Engineering, International Livestock Environment Symposium - ILES VIII, Iguassu
Falls City, Brazil, 31st August to 4th September, 2008, unpaginated-unpaginated.
Cao, M., C. Zong, Y. Zhuang, G. Teng, S. Zhou, T. Yang, 2021. Modeling of Heat Stress in Sows Part 2:
Comparison of Various Thermal Comfort Indices. Animals. 11 (6). http://doi.org/10.3390/ani11061498.
Hong, S., V. Exadaktylos, I. Lee, T. Amon, A. Youssef, T. Norton, D. Berckmans, 2017. Validation of
an open source CFD code to simulate natural ventilation for agricultural buildings. Computers and Electronics
in Agriculture. 138, 80–91. http://doi.org/10.1016/j.compag.2017.03.022.
Iqbal, A., K.R. Gautam, G. Zhang, L. Rong, 2021. Modelling of animal occupied zones in CFD.
Biosystems Engineering. 204, 181–197. http://doi.org/10.1016/j.biosystemseng.2021.01.022.
Seo, I.-h., I.-b. Lee, S.-w. Hong, H.-s. Noh, J.-h. Park, 2015. Web-based forecasting system for the
airborne spread of livestock infectious disease using computational fluid dynamics. Biosystems Engineering.
129, 169–184. http://doi.org/10.1016/j.biosystemseng.2014.10.004.
∙ 113 ∙
Impact of Carbon Dioxide on Hydrogen Sulfide Adsorption to

Straw Biochars
Dandan Huang a,b, Mingshen Liang b, Ning Wang b, Qiyong Xu b,*
a
School of Ecology, Shenzhen Campus of Sun Yat-sen University,
Shenzhen, Guangdong 518107, China
b
Shenzhen Engineering Laboratory for Eco-efficient Recycled Materials,
School of Environment and Energy, Peking University Shenzhen Graduate School,
Shenzhen, Guangdong 518055, China
* Corresponding author. Email: qiyongxu@pkusz.edu.cn
Abstract
Biochar is a promising H2S adsorbent in biogas desulfuration or purification. To clarify the
impact of carbon dioxide (CO2) on hydrogen sulfide (H2S) removal by biochars, this study
conducted H2S adsorption tests using rice straw biochar (RS), wheat straw biochar (WS), and
maize straw biochar (MS). Hydrogen sulfide (100 ppm) with different CO2 contents (10%, 20%,
and 50%) in the syngas were simulated. Humidity demonstrated to efficiently improve the H2S
removal on all biochars compared to that in dry condition. Carbon dioxide proved to be negative
for the H2S adsorption on dry straw biochars due to the competition between CO2 and H2S for
adsorption sites. However, humidity successfully alleviated this negative impact of CO2 for RS
and MS and even reversed the CO2 effect from negative to positive for WS. The alleviation effect
of humidity was ascribed to a favorable carbon surface environment created by the dissociation
of bicarbonates and carbonates from the reactions of CaO and especially MgO with CO 2.
Consequently, the CO2 competition was avoided, and H2S dissociation on this carbon surface was
also accelerated. This may have facilitated the subsequent catalytic reactions towards H 2S
oxidation.
Keywords: Carbon, odor, biogas, H2S removal, gas purification, desulfuration
1. Introduction
Hydrogen sulfide (H2S) is one common contaminant in biogas produced during anaerobic
digestion of organic wastes including livestock manure (Gwenzi et al., 2021). It is well known
that H2S is toxic and lethal at high concentrations (Sitthikhankaew et al., 2014). Long-term
exposure to low levels of H2S can also result in detrimental effects on human health such as
headache, dizziness, and throat and respiratory irritation (Lewis and Copley, 2015). Because of
its very low olfactory threshold (0.7 μg m-3), H2S is also regarded as a main contributor to odor
nuisance (Elsayed et al., 2009). Apart from that, H2S will pollute the metal catalysts used for
biogas upgrading, and produce sulfur dioxide during biogas combustion, which is the main cause
of acid rain (Woolcock and Brown, 2013). Therefore, different national standards have been
implemented for restricting H2S emissions to the atmosphere (Velasco et al., 2019).
Adsorption is considered to be the most promising method for the removal of H2S because of
its high removal efficiency, simple operation, low cost, and easy regeneration (Ma et al., 2021).
Using biochars as absorbents in gas effluent treatment can be a sustainable and cost-effective
approach especially if the biomass source for producing the biochars are waste materials (Bamdad
et al., 2018). Biochars from a wide range of wastes have been examined as promising materials
for H2S adsorption, including wood waste, crop waste, livestock manure, food waste, sludge, and
others (Martin et al., 2019; Sethupathi et al., 2017; Pelaez-Samaniego et al., 2018; Shang et al.,
2016; Hervy et al., 2018).
Carbon dioxide (CO2) is another common component in biogas from anaerobic digestion of
livestock manure. It is also an acid gas owning potential impact on the H2S adsorption to carbons
including biochars (Sitthikhankaew et al., 2014; Bamdad et al., 2018). So far, the knowledge
regarding the impact of CO2 on H2S adsorption to carbons is still very limited, especially very few
∙ 114 ∙
for biochars, and no consistent conclusion can be drawn. Sitthikhankaew et al. (2014) reported a
competitive adsorption between CO2 and H2S on the surface of activated carbons. In contrast,
Hervy et al. (2018) observed no obvious effect of CO2 on activated carbons from food waste and
sludge under dry condition. Gonçalves et al. (2018) concluded a cooperative effect of CO 2
molecules that favored H2S adsorption on commercial activated carbons in the 8.9 Å pore when
the CO2 concentration in the syngas was greater than 5%. Specifically for biochar, Sethupathi et
al. (2017) reported that simultaneous adsorption of CH4 and H2S resulted in a competition for
sorption sites. The effect of CO2 on the H2S adsorption performance of biochars worths being
further explored given the considerable CO2 (i.e., up to 20–47%) in actual biogas (Hervy et al.,
2018; Sitthikhankaew et al., 2014).
Therefore, we selected three typical straw waste derived biochars to conduct H2S adsorption
tests in this study. Different humidity of biochar (dry and 20%) and CO2 contents (10%, 20%, and
50%) in the syngas were considered to simulate various adsorption scenarios. The major purpose
was to assess the impact of CO2 on the H2S adsorption to the three straw biochars and the effect
of humidity in affecting the CO2 impact on their H2S adsorption.
2.1. Biochar preparation and characterization
Three straw wastes derived biochars, including rice straw, wheat straw, and maize straw
biochars, were obtained from Nanjing Zhirong Technology Co., LTD, China, labeled as RS, WS,
and MS, respectively. For producing the biochars, pyrolysis of these straw wastes was conducted
with a heating rate of 8.5℃ min-1 to a maximum temperature of 500℃. The final temperature was
maintained until no gas effluent from the outlet was detected. Then, the heating program was
terminated, and the furnace was cooled to the room temperature. The produced biochars were then
sieved through No. 18 and 40 sieves so that biochars with particle sizes of 0.425–1 mm were
obtained for the following adsorption tests.
Table 1. Physicochemical properties of the three biochars.
Biochar type
Properties
Rice straw Wheat straw Maize straw
VS (%) 63.42±0.33 70.65±1.21 67.42±1.27
TS (%) 98.85±0.31 98.58±0.39 87.04±1.27
Fixed carbon (%) 31.84±0.92 24.33±1.43 15.25±0.31
Ash content (%) 3.59±0.32 3.60±0.70 4.38±0.29
pH 9.82±0.02 9.69±0.03 9.60±0.07
BET surface area (m2 g-1) 33.38 46.77 43.30
Average pore diameter (nm) 0.5999 0.5912 0.5915
Micropore volume (cm3 g-1) 0.0263 0.0356 0.0335
Organic element (%) C 54±1.25 66±0.57 56±0.68
H 2.74±0.37 2.30±0.04 2.41±0.00
N 2.12±0.45 1.55±0.40 2.57±0.07
S 0.48±0.15 0.27±0.02 0.27±0.03
O 41±0.28 30±1.03 39±0.58
Major metal oxides (%) CaO 21.41 26.81 26.12
MgO 17.82 19.42 17.35
Al2O3 8.94 8.86 10.28
Na2O 18.48 15.42 -
K2O 12.40 - 20.04
Table 1 shows the physicochemical properties of the three biochars. Biochar pH was
determined using a pH meter (FiveEasy Plus, Mettler Toledo, Switzerland). The Brunauer-
∙ 115 ∙
Emmett-Teller (BET) method was applied to estimate the specific surface area through CO 2
adsorption at 298.15 K and the Horvath-Kawazoe model was used for determining the micropore
volume and average pore diameter. The contents of fixed carbon, ash, and volatiles were obtained
by the weight difference method. A scanning electron microscope (ZEISS SUPRA® 55, Carl
Zeiss, German) was used to observe the microstructure. The C, H, and N content was measured
using an elemental analyzer (Perkin Elmer 2400 Ⅱ, USA), and then the O content was derived.
The contents of metal oxides were determined by X-ray fluorescence spectrometer (XRF, ZSX
Primus Ⅱ, Rigaku) for original biochar. Besides, X-ray photoelectron spectroscopy (XPS) analysis
was conducted for chemical states of sulfur before and after H 2S adsorption using a PHI 5500
multi-technique system (Physical Electronics).
The H2S adsorption tests were conducted using a plexiglass reactor (#1), with an inner
diameter of 1.5 cm and height of 26 cm, at the room temperature (Figure 1). A mass of 1 g biochar
was packed in the reactor for all the tests. A H2S concentration of 100 ppm was selected in
simulating the syngas with nitrogen (99.999%) as the balance gas (Hervy et al., 2018). For
obtaining the adsorption curves, an infrared biogas analyzer (Gasboard-3200plus, Hubei Cubic-
Ruiyi Instrument Co., Ltd, China) was used to monitor the outlet H2S concentrations, with a
measurement range of 200 ppm and an accuracy of ± 2% full scale. The analyzer was calibrated
before each adsorption test. Moreover, as shown in Figure 1, a back-up reactor (#2) was used to
demonstrate the repeatability and stability of the adsorption system. The adsorption tests were
performed simultaneously for reactors #1 and #2 under dry biochar and pure H2S. Results showed
good agreements between the two reactors but were not given in this paper. For the below results
and discussion, only the results from the reactor #1 were used.
To investigate the impact of biochar humidity on the H2S adsorption, two humidity levels (0%
and 20%) were designed for the adsorption tests. Biochars with zero humidity were obtained
through oven drying of original biochars for 24 hrs. Biochars with 20% humidity level were
achieved by adding the calculated mass of water to the pre-dried biochars.
Syngas
Reactor Reactor
#1 #2
Mass
flowmeter
Figure 1. A picture of the experimental set-up.

To investigate the effect of CO2 on the H2S adsorption to the three biochars, three CO2 contents
(10%, 20%, and 50%) were designed with 100 ppm H2S in the syngas. Then the adsorption tests
were performed for the biochars under dry and 20% humidity conditions, respectively, all at an
inlet H2S flow rate of 100 mL min-1.
∙ 116 ∙
2.3. Equation
The H2S adsorption capacity was calculated by Eq. (1):
Q   H 2 S  Ci t Ct 
q=
m  1  C dt
0 i
(1)
where q is adsorption capacity, in mg g-1; Q is the inlet flow rate, in mL min-1; ρH₂S is the density
of H2S, in g cm-3; t is the time, in min; Ci and Ct are the inlet and outlet concentrations, respectively;
and m is the mass of biochar, in g.
3.1. Effect of CO2 on H2S removal
As displayed in Figure 2, under dry condition, the adsorption capacities were 0.33, 0.41, and
0.43 mg g-1, respectively, for RS, WS, and MS, without obvious differences. When 20% humidity
was present, the H2S removal capacity was greatly improved to 1.28 mg g-1 for RS, 0.99 mg g-1
for WS, and 1.13 mg g-1 for MS. The benefits of humidity in improving H2S adsorption to biochars
included the enhanced H2S dissolution and ionization in water film (Xu et al., 2014;
Sitthikhankaew et al., 2014). Additionally, metal oxides can perform catalytic oxidation of H2S;
their reactivity with H2S are very low in dry conditions but improved in humid conditions
(Bagreev and Bandosz, 2005).
Figure 2 also shows the H2S removal capacities under different CO2 contents. Under dry
condition, the presence of CO2 significantly inhibited the adsorption of H2S to all three biochars
due to the competition for adsorption sites (Sethupathi et al., 2017), regardless of the CO2
contents. The inhibition was most obvious for RS (reduced by 70–75% from 0.33 to 0.08–0.10
mg g-1) followed by MS (reduced by 67–75% from 0.43 to 0.11–0.14 mg g-1) and WS (reduced
by 39–63% from 0.41 to 0.15–0.25 mg g-1). The less inhibition of CO2 in the dry condition for
WS could be ascribed to its higher BET surface area and pore volume (Table 1), providing more
adsorption sites and space than that of RS and MS (Shang et al., 2016).
Dry 20% humidity
0.6 1.8
Adsorption capacity (mg g-1)
Adsorption capacity (mg g-1)
RS WS MS RS WS MS
0.5 1.5
0.4 1.2
0.3 0.9
0.2 0.6
0.1 0.3
0 0
0% 10% 20% 50% 0% 10% 20% 50%
CO2 content CO2 content
Figure 2. H2S adsorption capacities for the rice straw (RS), wheat straw (WS), and maize straw
(MS) biochars under dry + different CO2 contents and 20% humidity + different CO2 contents.
In the presence of 20% humidity, the inhibition of CO2 was alleviated for RS and MS with
their adsorption capacities rising back to 0.71–0.86 mg g-1 and 0.84–0.94 mg g-1, respectively. For
WS, 20% humidity even reversed the impact of 10% and 20% CO2 from negative to positive as
the adsorption capacity were slightly increased in comparison with that under dry condition. The
negative impact of 50% CO2 was also alleviated for WS when 20% humidity was present. The
adsorption capacities were 0.15 mg g-1 for the dry + 50% CO2 condition but 0.94 mg g-1 for the
humid + 50% CO2 condition.
3.2. Mechanism of humidity in alleviating CO2 inhibition on H2S removal
Given the relatively superior performance of WS in the presence of CO2 + H2O, the XPS
analysis was conducted for only WS to quantify the proportions of different sulfur products.
Figure 3 shows the sulfur products for WS after H2S adsorption tests in the presence of 20%
∙ 117 ∙
humidity and 20% CO2 + 20% humidity.

Sulfate Sulfite H₂S(ads) Metal sulfide Elemental sulfur
1.74 1.22
100%
20.26 21.4
80% 32.46
27.4
Proportion
60%
25.03
40% 79.74
20.73
49.98
20%
20.04
0%
Original WS WS + 20% humidity WS+20% humidity+20% CO₂
Groups
Figure 3. Sulfur products characterized by XPS analysis for the original WS and WS after H2S
adsorption under 20% humidity and 20% humidity + 20% CO2 in the syngas.
The original WS only presented sulfur products of sulfate (SO42-) (79.74%) and metal sulfides
(MSx) (20.26%). With 20% humidity, the relative proportion of sulfur products after H2S
adsorption increased most in MSx (from 20.26% to 32.46%), followed by H2S in adsorbed state
(H2Sads) (from 0 to 25.03%), sulfite (from 0 to 20.73%), and elemental sulfur (from 0 to 1.74%).
When 20% CO2 was present with 20% humidity, the peak for H2S(ads) disappeared and the
content of MSx also decreased from 32.46% to 21.40%; in contrast, the SO42- content rose from
20.04% to 49.98%.
Table 2. Final pH values of the rice, wheat, and maize straw biochars
Final pH
Adsorption scenarios Rice straw Wheat straw Maize straw
biochar biochar biochar
Dry 8.59 8.10 8.22
20% humidity 8.17 7.55 7.74
Dry + 10% CO2 7.71 8.47 7.59
Dry + 20% CO2 7.58 8.94 7.74
Dry + 50% CO2 7.84 7.50 7.86
20% humidity + 10% CO2 7.67 8.35 7.71
20% humidity + 20% CO2 7.72 8.68 7.75
20% humidity + 50% CO2 7.93 7.30 7.77
As mentioned above, the catalytic H2S oxidation by metal oxides could be improved in the
presence of humidity. Table 1 shows that CaO and MgO are the two most abundant metal oxides
commonly in the three biochars. CaO and specifically MgO barely dissolve in water but can react
with CO2 to form bicarbonates and carbonates. Due to the reaction, CaO and MgO are present as
a separate solid phase (bicarbonates and carbonates) covered with a thin film of hydroxides or
bicarbonates, the dissociation of which can provide a high pH value and thus a favorable carbon
surface environment for H2S dissociation (Bagreev and Bandosz, 2005). The disappeared
H2S(ads) for WS after H2S adsorption under 20% humidity + 20% CO2 compared to that under
20% humidity supported the accelerated H2S dissociation. Considering the CO2 inhibition on H2S
adsorption capacity was alleviated most for WS and WS had the highest MgO content, MgO
content was expected to be more beneficial than CaO in creating the above mentioned carbon
∙ 118 ∙
surface environment. In line with this, a less reduction in pH was also observed for WS than RS
and MS (Table 2). However, the buffer against CO2 provided by CaO and MgO could be
exhausted at a high CO2 content (i.e., 50%). Thus, the competition of CO2 for the adsorption sites
with H2S (Sethupathi et al., 2017) explained the decreased adsorption capacity and pH when CO2
content further increased from 10% and 20% to 50% for WS.
4. Conclusions
The influence of CO2 on the H2S adsorption was demonstrated to be negative for the straw
biochars in dry condition. Humidity successfully alleviated the negative impact of ≤ 20% CO2 for
RS and WS. A shift of the CO2 effect from negative to positive was particularly noted for WS,
ascribed to a favorable carbon surface environment created by the dissociation of bicarbonates
and carbonates from the reactions of CaO and especially MgO with CO2. However, this buffer
against CO2 could be exhausted at a high CO2 content (e.g., 50%). In achieving an optimum H2S
removal by straw biochars in actual biogas purification or desulfuration, humidity is vital in
activating the catalytic H2S oxidation or alleviating and even inducing a favorable change of the
negative CO2 impact.
Acknowledgements
This research was funded by the National Key R&D Program of China (2018YFC1902903).
References
Bagreev, A., T.J. Bandosz, 2005. On the mechanism of hydrogen sulfide removal from moist air on
catalytic carbonaceous adsorbents. Industrial & Engineering Chemistry Research. 44 (3), 530–538.
https://doi.org/10.1021/ie049277o.
Bamdad, H., K. Hawboldt, S. MacQuarrie, 2018. A review on common adsorbents for acid gases removal:
Focus on biochar. Renewable and Sustainable Energy Reviews. 81 (Part 2), 1705–1720. https://doi.org/
10.1016/j.rser.2017.05.261.
Elsayed, Y., M. Seredych, A. Dallas, T.J. Bandosz, 2009. Desulfurization of air at high and low H2S
concentrations. Chemical Engineering Journal. 155 (3), 594–602. https://doi.org/10.1016/j.cej.2009.08.010.
Gonçalves, D.V., M.A.G. Paiva, J.C.A. Oliveira, M. Bastos-Neto, S.M.P. Lucena, 2018. Prediction of the
monocomponent adsorption of H2S and mixtures with CO2 and CH4 on activated carbons. Colloids and
Surfaces A. 559, 342–350. https://doi.org/10.1016/j.colsurfa.2018.09.082.
Gwenzi, W., N. Chaukura, T. Wenga, M. Mtisi, 2021. Biochars as media for air pollution control systems:
Contaminant removal, applications and future research directions. Science of the Total Environment, 753,
142249. https://doi.org/10.1016/j.scitotenv.2020.142249.
Hervy, M., D.P. Minh, C. Gérente, E. Weiss-Hortala, A. Nzihou, A. Villot, L.L. Coq, 2018. H2S removal
from syngas using wastes pyrolysis chars. Chemical Engineering Journal. 334, 2179–2189. https://doi.org/
10.1016/j.cej.2017.11.162.
Lewis, R.J., G.B. Copley, 2015. Chronic low-level hydrogen sulfide exposure and potential effects on
human health: A review of the epidemiological evidence. Critical Reviews in Toxicology. 45 (2), 93–123.
https://doi.org/10.3109/10408444.2014.971943.
Ma, Q., W. Chen, Z. Jin, L. Chen, Q. Zhou, X. Jiang, 2021. One-step synthesis of microporous nitrogen-
doped biochar for efficient removal of CO2 and H2S. Fuel. 289, 119932. https://doi.org/10.1016/
j.fuel.2020.119932.
Martin, L.M., G. Hernandez, Z. Montoya, J. Koranten-Amoako, L. Derry, D. Valles-Rosales, C.E.
Brewer, 2019. Adsorption of hydrogen sulfide on biochars from pallet wood waste. In ASABE 2019 Annual
International Meeting. Boston, USA, July 7–10. St. Joseph, Mich.: ASABE. https://doi.org/10.13031/
aim.201900359.
Pelaez-Samaniego, M.R., M.W. Smith, Q. Zhao, T. Garcia-Perez, C. Frear, M. Garcia-Perez, 2018.
Charcoal from anaerobically digested dairy fiber for removal of hydrogen sulfide within biogas. Waste
Management. 76, 374–382. https://doi.org/10.1016/j.wasman.2018.03.011.
Sethupathi, S., M. Zhang, A.U. Rajapaksha, S.R. Lee, N.M. Nor, A.R. Mohamed, M. Al-Wabel, S.S. Lee,
Y.S. Ok, 2017. Biochars as potential adsorbers of CH4, CO2 and H2S. Sustainability. 9, 121. https://doi.org/
10.3390/su9010121.
∙ 119 ∙
Shang, G., Q. Li, L. Liang, P. Chen, X. Huang, 2016. Adsorption of hydrogen sulfide by biochars derived
from pyrolysis of different agricultural/forestry wastes. Journal of the Air & Waste Management Association.
66 (1), 8–16. https://doi.org/10.1080/10962247.2015.1094429.
Sitthikhankaew, R., D. Chadwick, S. Assabumrungrat, N. Laosiripojana, 2014. Effects of humidity, O2,
and CO2 on H2S adsorption onto upgraded and KOH impregnated activated carbons. Fuel Processing
Technology. 124, 249–257. https://doi.org/10.1016/j.fuproc.2014.03.010.
Velasco, A., J.M. Morgan-Sagastume, A. González-Sánchez, 2019. Evaluation of a hybrid
physicochemical/biological technology to remove toxic H2S from air with elemental sulfur recovery.
Chemosphere. 222, 732–741. https://doi.org/10.1016/j.chemosphere.2019.02.005.
Woolcock, P.J., R.C. Brown, 2013. A review of cleaning technologies for biomass-derived syngas.
Biomass & Bioenergy. 52, 54–84. https://doi.org/10.1016/j.biombioe.2013.02.036.
Xu, X., X. Cao, L. Zhao, T. Sun, 2014. Comparison of sewage sludge and pig manure derived biochars
for hydrogen sulfide removal. Chemosphere. 111, 296–303. https://doi.org/10.1016/j.chemosphere.
2014.04.014.
∙ 120 ∙
Theme II:
Animal Production Systems and
Equipment
∙ 121 ∙
∙ 122 ∙
A CFD Study on Ventilation Options for Cage-free Hen Houses in

North America
Long Chen a,*, Eileen E Fabian (Wheeler) b
a
Tianjin Academy of Agricultural Sciences, Tianjin 300384, China
b
Department of Agricultural and Biological Engineering, Pennsylvania State University,
State College, PA 16802, USA
* Corresponding author. Email: lchen6316@gmail.com
Abstract
A shift to cage-free eggs is underway in North America. This contributes to the most
significant evolution that poultry facilities have been facing in decades. This study aims to provide
practical recommendations to refining ventilation system design in cage-free hen houses with the
goal of assuring bird welfare through comfortable conditions and improving the capacity of
containing contaminants, such as ammonia. Four three-dimensional CFD models were developed
based on a full-scale floor-raised hen house, and three alternatives. In addition, 2,365 birds were
individually modeled with simplified shapes to represent one-eighth of a real commercial hen
house. Performance of the standard ventilation configuration and alternative designs were
analyzed and evaluated by comparing their simulation outputs at locations represented by cross-
sectional planes in the model. Critical environment parameters such as temperature, pressure, air
speed, and ammonia concentration were captured and exported for further analyses. Additionally,
both indoor and outdoor contours of those key factors were created for data visualization and
evaluation. Five animal zones were also created to represent the Animal Occupation Zone (AOZ)
for specialized data analysis. Considerable simulation results and subsequent analyses
substantially demonstrated these alternative models had the comparable capacity to create
satisfactory indoor conditions for the cage-free hen house. In addition, statistical analyses were
conducted to confirm the significance of vital factors and verify significant differences for
particular comparisons.
Keywords: Floor-raised hen housing, computational fluid dynamics, full-scale simulation, indoor
air conditions, animal welfare
1. Introduction
Poultry facilities are going through some significant transition because of demands for a wide
range of building configurations to accommodate cage-free, enriched cages, and organic
standards. Common cage-free housing systems include aviary systems, convertible cages and
floor housing systems (Chen et al., 2020). Each housing configuration has its advantages and
drawbacks with regards to stocking densities, mortality rates, disease control, and nuisance of
eggs laid outside of nest-boxes (floor eggs). The lack of unified guidelines and industry conflict
regarding what “cage-free” means have left egg producers without clear understanding about how
to proceed the transition to cage-free housing, or to which particular system to switch. In addition,
both egg producers and poultry house contractors are confronting some difficulty in finding the
most effective management of cage-free facilities to maintain optimal production (Mongeau and
Risser, 2018). Thereby, current ventilation systems have not necessarily kept pace with the new
poultry buildings due to a lack of performance-based systematic design procedures.
Computational fluid dynamics (CFD) has been used as a powerful tool to model fluid
movement in diverse applications. Hence, the agricultural engineering community has embraced
CFD to predict the performance of ventilation systems under different conditions for the sake of
comprehensive design (Mistriotis and Jong, 1997). CFD modelling allows full control of the
influencing factors, provides universal data in the computational domain, and is relatively low
cost in terms of time and computational expense. Numerous studies have demonstrated that CFD
models can reliably predict airflow, heat, and mass transfer in animal housing systems (Li, Rong,
∙ 123 ∙
and Zhang, 2016; Li, Rong, and Zhang, 2017; Osorio-Saraz et al., 2011; Rong et al. 2015).
This study applied CFD simulations in characterizing three alternative ventilation schemes
applied to a floor-raised hen house, in comparison with the standard ventilation scheme (Chen et
al., 2020) in terms of evaluating the indoor conditions within each design. Simulation results of
critical environmental parameters for bird comfort, air speed and temperature, along with the
driving force for ventilation air exchange (static pressure difference), were analyzed visually and
quantitatively at the whole house and bird levels. The goal was to assess the performance of these
ventilation options for cage-free poultry houses to maintain a desired and comfortable indoor
microclimate that satisfies the needs of production demand and animal welfare concerns.
Conditions during cold weather were modelled as the most challenging for maintaining
comfortable, healthy conditions due to low ventilation air exchange and fresh air distribution in
the hen house. Due to the limitation of this article’s length, contents of temperature and pressure
analyses were not presented and discussed.
Four CFD models and corresponding simulations were conducted using the commercial
software package FLUENT v19.1 (ANSYS, 2009). The standard k-ε turbulence model (Cengel
and Cimbala, 2013; Launder and Spalding, 1974) with enhanced wall functions was adopted for
the development of CFD models, based on the previous investigations (Harral and Boon, 1997;
Osorio Saraz et al., 2016; Rong et al., 2015; Worley and Manbeck, 1995). Simulations were
conducted on the Pennsylvania State University’s Institute for Computational and Data Sciences’
Roar supercomputer.
2.1. Development of the CFD model
The computational domain of each model included the barn itself and ambient air to properly
simulate airflows inside and outside the building. All models shared the same size computational
domain where the height was 24.4 m (80 ft) and the width 128.2 m (420.6 ft). An outdoor wind
speed of 2.0 m s-1 (4.5 mph) was assumed perpendicular to the barn sidewalls, blowing from left
to right constantly across the computational domain. Note that the distance between the left end
of the domain and the house was much shorter than that from the right end to the house because
an extended domain far from the target layer house at the downwind side minimized reverse flows
at domain boundaries (Pawar et al., 2007).
Four ventilation schemes were modelled and investigated. The standard TISE ventilation
scheme was modelled with ten inlets above each sidewall (symmetrical about the building
centerline), spaced evenly along the building eaves, and half of one of the four sidewall exhaust
fans located at the right-sidewall (Figure 1). Ventilation inlets at the top of both sidewalls were
2.5 m (8.3 ft) above the ground.
Three alternative ventilation systems were designed, including “mid-wall inlet ceiling
exhaust” [MICE], “mid-wall inlet ridge exhaust” [MIRE], and “mid-wall inlet attic exhaust”
[MIAE] (Figure 1) (Chen, 2019). Inlets of MICE were positioned with the base 1.5 m (60 in.)
above the floor and had a wall-plate at the top (Figure 1) to direct incoming air horizontally from
the inlet opening baffle. The exhaust fan of MICE was positioned at the middle of the ceiling
cross-section. Note that the fan chute was modelled as an attached duct whose length was 2.9 m
(112.5 in.). Inlets of MIRE and MIAE were modelled identically with those of MICE. However,
the major differences between three alternative designs were the positions of exhaust fans. The
MICE configuration resembled some ventilation designs more common in European construction,
while no ceiling was included in MIRE and its exhaust fan was placed at the middle of the roof,
3.3 m (131 in.) above the nest-boxes, with a short duct length of 1.6 m (64.5 in.). MIAE had a
partial ceiling, forming an “attic space” with a 2.3 m (90 in.) opening along the central length of
the layer house. The exhaust fan of MIAE with an attached duct length of 2.5 m (100 in.) was
positioned at the middle of the roof ridge at a distance of 2.4 m (95.5 in.) above the nest-boxes.
∙ 124 ∙
Figure 1. Geometry of the study layer house with the standard ventilation scheme (TISE) and
three alternative ventilation designs (MICE, MIRE, and MIAE). Arrows indicate locations of
inlets and the exhaust fan.
A significant aspect of the study was to determine conditions at bird-level for welfare comfort
conditions in addition to overall environment patterns in the building air space. For this reason,
hen models were included in the simulation. In total 2,365 individual hen models were included
in each model to represent approximately 1/8 of the total hens housed in the 1/8 house section.
An estimated distance between hens was calculated based on the stocking density by assuming all
the hens were evenly distributed according to the actual scenario observed in the study barn.
2.2. Boundary conditions
Six types of boundary conditions or cell zones were adopted in the CFD simulation inside and
outside the layer house:
• The ground, ceiling, roof, slatted floor, nesting area, litter area, inlet baffles, sidewalls,
animal surfaces, and the top surface of the computational domain were defined as “walls”.
Note that all “walls” were defined as non-slip walls except for the top surface of the
computational domain, which was defined as a zero-shear stress wall with no resistance
along the surface.
• The front and back surfaces of the computational domain along the z-axis and both near and
far ends of the house were defined as “symmetry” boundary conditions, whereas those
surfaces represented internal faces that accounted for 1/8 of the actual scenario.
• Two faces of each inlet perpendicular to the wind direction were assigned boundary
conditions of “interior” to represent an interior portion of the computational domain through
which air could flow. Additionally, the exhaust fan had two faces defined as interior.
• Each hen model was defined as a solid volume, whose surface was defined as solid wall
with a constant typical hen body temperature of 42°C (107.6°F). A heat generation rate of
4,467 W m-3 was assigned to the individual hen models (Pawar et al., 2007).
• The left surface of the computational domain was defined as a “velocity inlet” with a
specified wind magnitude of 2.0 m s-1 (393.7 ft min-1).
∙ 125 ∙
• The right surface of the computational domain was assigned a boundary condition of
“pressure outlet” through which flow exits to atmospheric pressure
• The volume of the exhaust fan was defined as a “3D fan zone” where the entire fan volume
was considered a fluid cell zone, which simulated the effect of an axial fan by applying a
distributed momentum source. In addition, the fan was defined with a hub radius of 5 cm (2
in.), a tip radius of 46 cm (18 in.), and a thickness of 5 cm (2 in.). Rotational speed was
specified as 60 rad s (573 rpm). A constant pressure jump of 18 Pa (0.072 in. of water) was
applied across all the cells in the fan zone to achieve the desired hen ventilation rate.
Four ventilation schemes were analyzed and compared for airflow patterns. Simulation results
of air velocity vectors were visualized at three reference planes with contour plots.
Contours of air velocity magnitude and velocity vectors were created to visualize airflow
patterns inside the layer house with each ventilation scheme. In general, fast incoming air jets
were observed from both inlets in all four ventilation schemes, and the air speed decreased
gradually when approaching the center of house (Figure 2). Although the air speed from the
downwind inlet did not travel as far as the upwind inlet, expected air circulation patterns were
observed throughout the house for all the models.
Figure 2. Isometric view of the study layer house showing the location of three reference planes
as indicated in (1), and overall airflow patterns and air velocity magnitudes within four
ventilation schemes: (1) TISE, (2) MICE, (3) MIRE, (4) MIAE.
The patterns of indoor airflow varied with the type of ventilation schemes. The standard TISE
ventilation system possessed the strongest incoming air jets compared to three alternative designs,
because TISE model had inlets at the top of the sidewalls along a flat ceiling (Figure 3). Without
the ceiling along which the air jets could move, the fresh air trajectories in the other three models
more quickly dropped towards the hen occupied area after traveling fairly horizontal paths for a
short distance. Another observation was that the incoming air from the downwind (right-side)
inlet had relatively smaller size except for the MIAE model (Figure 3).
∙ 126 ∙
Inlet Inlet
“TISE”
“MICE”
Inlet Inlet
“MIRE”
Inlet Inlet
“MIAE”
Inlet Inlet
Figure 3. Indoor air velocity vector-contours at Plane I of four models.

Air movement patterns varied dramatically at Plane N that contained no ventilation features.
Strong air movement was observed throughout the cross-section for all models (Figure 4).
Airflows around 1.5 m s-1 (295 ft min-1) moved from the upper right to the lower left of the house
in TISE.
In the MICE model, strong air movements were observed at the right portion of the house
close to the sidewall, moving towards the central nest-box area. There was also a small portion of
fast airflow at the upper left in MICE, which is close to an obvious swirl nearby. The MIRE model
had strong airflows starting from the left portion of the house, where the position and the trajectory
of the airflows were coincident to the upwind incoming air jet. Even vigorous airflows with similar
trajectories were observed in the MIAE model from both sides. Because Plane N represented the
majority (69%) of cross-sectional locations in the layer house, air movements in this Plane would
be more representative of overall house conditions.
The air movement patterns at Plane F exhibited the influence of the exhaust fan on
performance of each model (Figure 5). The TISE model presented different patterns compared to
the other three models due to the distinct sidewall location of the fan. Air moved from the middle
of the house towards the fan in TISE within the right side and its speed increased gradually as it
approached the fan influence area. Nonetheless, a strong horizontal air movement was also
observed towards the left sidewall within the left portion of the house, with even higher air speeds
than that of the right side.
However, some slanted airflows were observed in the MICE model moving from the nest-box
area to the upper left corner. In the MIAE model, most of the attic space was filled with airflows
moving upwards, excluding some airflows close to the duct and fan areas.
∙ 127 ∙
“TISE”
“MICE”
“MIRE”
“MIAE”
Figure 4. Indoor air velocity vector-contours at Plane N of four models.
“TISE”
Fan
Fan
“MICE”
Fan
“MIRE”
Fan
“MIAE”
Figure 5. Indoor air velocity vector-contours at Plane F of four models.
∙ 128 ∙
4. Conclusions
One standard and three alternative ventilation schemes were modelled in a real commercial
floor-raised cage-free hen house with the goal of evaluating the indoor air conditions within each
design. The simulated ventilation rate of four models (TISE, MICE, MIRE, MIAE) for the same
hen house was 1.97 m3 s-1 (4174 ft3 min-1), 1.93 m3 s-1 (4089 ft3 min-1), 1.96 m3 s-1 (4153 ft3
min-1), and 1.91 m3 s-1 (4047 ft3 min-1), respectively. These ventilation rates were on the higher
end of recommended air exchange range for hens during cold weather, which was appropriate for
the evaluation at 0°C. The egg-laying flock was modelled as heated, hen-shaped individuals
evenly distributed throughout the building. By observing air velocity vector-contours, MICE,
MIRE, and MIAE were able to provide strong incoming air jets that can reach the central region
of the house almost as well as the TISE model. Furthermore, the alternative designs provided
indoor air movements similar to the standard model, even better at the hen level (Plane I). Air
speeds on average were maintained at 0.35 m s-1 (69 ft min-1) at hen level in the majority of the
house for TISE, MIRE, and MIAE. Airflow pattern visualization showed two large circular air
eddies that included the incoming air jets in TISE whereas the alternative models had these large
eddies along with more numerous, smaller circulation patterns.
This study recognizes CFD modelling as a powerful tool to analyze ventilation performance
and indoor microclimate. Using realistic dimensions, including modelling animals, improved the
usefulness of CFD simulation. The four models can be fine-tuned to assess other existing
ventilation schemes or evaluate proposed ventilation options for various types of poultry houses.
In summary, making full use of CFD modelling enables investigators to tackle practical problems
and explore sophisticated solutions related to animal housing.
Acknowledgements
The authors would like to thank the Egg Industry Center (EIC) who funded this project.
Thanks to several Pennsylvania cage-free hen egg producers, hen house construction companies,
and advisors for their participation in the project efforts. Acknowledgement to the USDA National
Institute of Food and Agriculture Federal Appropriations under Project PEN04614 and Accession
number 1011207. Thanks to Dr. Zhanxiong Xu for his professional statistic consultancy. Thanks
also goes to the Institute for Computational and Data Sciences at The Pennsylvania State
University for providing high performance computing (HPC) resources and technical support.
References
ANSYS. 2009. User guide. Release 12.0. Lebanon, NH, USA.
Cengel, Y. A., and J. M. Cimbala. 2013. Fluid mechanic: fundamental and applications. 3rd ed. McGraw-
Hill, New York, N. Y. 880–930 pp.
Chen, L. (2019). Cage-Free-Hen Housing Ventilation Options to Improve Air Quality and Bird Welfare.
Ph.D. Dissertation, Department of Agricultural and Biological Engineering, The Pennsylvania State
University. State College, PA, USA.
Chen, L., Fabian-Wheeler, E. E., Cimbala, J. M., Hofstetter, D., and Patterson, P. (2020). Computational
fluid dynamics modelling of ventilation and hen environment in cage-free egg facility. Animals, 10(6), 1–17.
https://doi.org/10.3390/ani10061067
Chen, L., Fabian-Wheeler, E. E., and Hofstetter, D. (2018). Ventilation options modelling for laying hens
in cage-free environment: three-dimensional case. ASABE 2018 Annual International Meeting. Detroit, MI,
July 29–August 1. Paper # 1801613. https://doi.org/10.13031/aim.201801613
Harral, B. B., and C. R. Boon. 1997. Comparison of predicted and measured airflow patterns in a
mechanically ventilated livestock building without animals. Journal of Agricultural Engineering Research
66:221–228. doi:10.1006/jaer.1996.0140
Launder, B. E., and Spalding, D. B. (1974). The numerical computation of turbulent flows. Computer
Methods in Applied Mechanics and Engineering, 3(2), 269–289. https://doi.org/10.1016/0045-
7825(74)90029-2
Li, H., Rong, L., and Zhang, G. (2016). Study on convective heat transfer from pig models by CFD in a
virtual wind tunnel. Computers and Electronics in Agriculture, 123, 203–210. https://doi.org/10.1016/
j.compag.2016.02.027
∙ 129 ∙
Li, H., Rong, L., and Zhang, G. (2017). Reliability of turbulence models and mesh types for CFD
simulations of a mechanically ventilated pig house containing animals. Biosystems Engineering, 161, 37–52.
https://doi.org/10.1016/j.biosystemseng.2017.06.012
Mistriotis, A, and Jong, T. De. (1997). Computational Fluid Dynamics (CFD) as a tool for the analysis of
ventilation and indoor microclimate in agricultural buildings. In Netherlands Journal of Agricultural Science
(Vol. 45, pp. 81–96). https://library.wur.nl/ojs/index.php/njas/article/view/527
Mongeau, E., and Risser, K. (2018). A look at cage-free housing solutions and practices for a changing
market.
Osorio-Saraz, J. A., Olivera Rocha, K. S., Ferreira Tinôco, I. de F., Gates, R. S., Barreto Mendes, L.,
Zapata, O. L., and Damasceno, F. A. (2011). Use of CFD modelling for determination of ammonia emission
in non-insulated poultry houses with natural ventilation. ASABE, Louisville, Kentucky, August 7–10, 2011.
Paper # 1110781, 2006, 2–11. https://doi.org/10.13031/2013.37308
Osorio Saraz, J. A., Ferreira Tinôco, I. de F., Olivera Rocha, K. S., Barreto Mendes, L., and Norton, T.
(2016). A CFD based approach for determination of ammonia concentration profile and flux from poultry
houses with natural ventilation. Revista Facultad Nacional de Agronomia Medellin, 69(1), 7825–7834.
https://doi.org/10.15446/rfna.v69n1.54750
Pawar, S. R., Cimbala, J. M., Wheeler, E. F., and Lindberg, D. V. (2007). Analysis of Poultry House
Ventilation Using Computational Fluid Dynamics. Transactions of the ASABE, 50(4), 1373–1382.
Rong, L., Bjerg, B., and Zhang, G. (2015). Assessment of modelling slatted floor as porous medium for
prediction of ammonia emissions – Scaled pig barns. Computers and Electronics in Agriculture, 117, 234–
244. https://doi.org/10.1016/j.compag.2015.08.007
Worley, M. S., and Manbeck, H. B. (1995). Modelling particle transport and air flow in ceiling inlet
ventilation systems. Transactions of the ASABE, 38(1), 231–239. https://doi.org/10.13031/2013.27835
∙ 130 ∙
Development of a Wide-Angle Egg Turning Incubator Capable of

Enhancing Gosling Embryonic Development, Hatchability,
and Post-Hatch Growth
Binbin Guo a, Zichun Dai a, Aidong Sun b, Huiman Ren c, Zhendan Shi a,*
a
Key Laboratory of Protected Agriculture Engineering in the Middle and Lower Reaches of Yangtze River,
Ministry of Agriculture and Rural Affairs; Laboratory of Animal Improvement and Reproduction, Institute
of Animal Science, Jiangsu Academy of Agricultural Sciences, Nanjing, Jiangsu 210014, China
b
Jiangsu Key Laboratory for Food Quality and Safety-State Key Laboratory Cultivation Base of Ministry of
Science and Technology, Nanjing, Jiangsu 210014, China
c
Foshan Renshi Machinery Technology Co., Ltd., Foshan 511000, Guangdong, China
* Corresponding: zdshi@jaas.ac.cn
Abstract
A new type of incubator with egg turning angle adjusted larger to 60° and 70° was designed
to improve goose egg hatching performance. We studied the impact on embryonic and gosling
growth performances, as well as the endocrine and molecular regulation mechanisms of wider egg
(WA) turning angle. Results showed that WA egg turning during incubation improved goose egg
hatchability. The improved hatchability was associated with improvements in embryonic and
muscular developments from embryonic days 22 to 29 through upregulation of gene expression,
such as hypothalamic growth hormone-releasing hormone, pituitary growth hormone (GH), and
liver insulin-like growth factor 1(IGF-1). The growth enhancing effect by the WA egg turning
during incubation continued through the gosling post-hatch rearing period with higher body
weight and relative leg and muscle weights. WA goslings also had higher plasma concentrations
of growth-related hormones, such as GH, IGF-1 and triiodothyronine (T3). These results
demonstrated that the WA egg turning during incubation could upregulate growth axis hormone
gene expression and secretions to enhance gosling embryo development, hatchability and growth
performance.
Keywords: Egg turning, embryonic development, growth performance, goose, hatching
performance, somatotropic axis
1. Introduction
Egg turning is important during incubation of domestic fowl eggs. Most of the early studies
have been concerned with its effect on hatchability, embryo positioning and gas exchange
(Deeming, 2009; Elibol and Brake, 2006; Tona et al., 2003a). Recent studies demonstrated that
turning during incubation, such as turning angle, frequency and turning duration, influence the
embryonic growth and the quality of newly hatched chicks and goslings (Cutchin et al., 2009; Dai
et al, 2017; Tona et al., 2003b). Better quality chicks or goslings are usually characterized by
higher body mass, dry and clean feather, bright eyes, healthy legs and closed navel (Tona et al.,
2003a).
Since chicken is the mostly produced domestic fowl and the chicken egg incubators and
incubation methods have become the standard of hatchery practices. With regard to egg turning,
the angle of 45° (either side from the vertical) is best appropriate for incubation of chicken eggs
because it produces the highest fertile egg hatchability (Funk and Forward, 1960; Wilson, 1991;
Deeming, 2002). However, when this 45° angle egg turning incubator is used for incubation of
much larger and heavier, approximately150 g–200 g, goose eggs, hatchabilities are much lower
and more variable. Apart from the higher metabolic heat production of the developing goose
embryo that could render ventilation more difficult, inappropriate egg turning may be another
important factor. For example, enlarging or widening egg turning angle could provide sufficient
relative movement of the embryo within the egg shell, thus to better position the allantois and
yolk sac membrane so to improve nutrients and gas exchanges to the embryo (Baggott et al., 2002;
∙ 131 ∙
Cutchin et al., 2009). In order to improve goose egg hatchability and also gosling quality, this
study aimed to design a new type of incubator with egg turning angle adjusted larger to 60° and
70°, and studied its impact on embryonic and gosling growth performances, as well as the
endocrine and molecular regulation mechanisms.
2. Design of wide-angle egg turning incubator and incubation performance tests
2.1. Design of wide-angle egg turning incubator
The conventional 45° egg turning incubator works by tilting the tray supports at 45° angle
from the vertical either side, so the egg trays sway up and down within a total angle of 90° (Figure
1). As enlarging egg turning angle through tilting the egg trays requires expanding the space or
distance between each layer of tray, doing so could reduce incubator space usage and number of
trays and therefore number of eggs incubated. The strategy of realizing WA egg turning while
still maintain the incubation capacity is to fix the distance between tray layers and to sway the
whole rack that hold the trays up and down at 70° angle from the vertical either side (Figure 1).
The racks can be fixed in the incubator or on wheeled trolleys (Figure 2) which facilitates loading
and unloading of the eggs into incubator. Four trolleys each containing 3200 goose eggs are fitted
into one incubator (Figure 3). Turing of the rack is realized by an easy connection to a turning
motor through shafts and gear wheels. Running of motor back and forth is controlled by limit
switches and relays till the rack sways to a certain angle or position. Other parts of the incubator,
including temperature, humidity and ventilation controlling systems, are adopted from the
conventional type incubators.
Figure 1. Diagram of conversion of the conventional 45° to 70° WA egg turning incubator,
which also shows changing of the egg turning mode and direction.
Figure 2. The trolley with the rack for holding egg trays on each layer. 1, rack; 2, carrier; 3,
wheel; 4, egg turning transmission mechanism.
∙ 132 ∙
Figure 3. Overall structural diagram of the WA egg turning incubator. 1, outer wall; 2,
ventilation fan; 3, egg rack; 4, humidifier; 5, turning gear box.
2.2. Wider angle egg turning to improve goose egg hatchability
Two experiments were carried out to test the incubation performance of the WA egg turning
incubator in comparison with the conventional type 45 turning incubator. In experiment 1,
medium sized Yangzhou goose eggs of 140–160 g of weight were used, and in experiment 2 used
large sized Shitou goose eggs with weight of 190–230 g. Three tests were conducted in experiment
1 each with 432 Yangzhou geese eggs set, while 6 tests were performed that each set 196 eggs, in
either of one WA or conventional incubator. Incubation followed the standard practices to control
temperature, humidity and ventilation. On days 7 and 18, eggs were candled to pick out the
unfertile eggs and dead embryos. The hatching performance data are listed in Table 1. In general,
WA incubation achieved better performances than did the conventional egg turning incubation.
Both the embryo livability on E7 and E18 were significantly higher in WA incubation, so were
set egg hatchability (Table 1). WA incubation increased the set egg hatchability by 3.55
percentage for Yangzhou goose eggs, and 6.11 percentage for Shitou goose eggs (Table 1). Such
results demonstrate that WA egg turning improves goose egg hatchability more efficiently for
large sized eggs.
Table1. Effects of egg turning angle on Yangzhou and Shitou geese egg incubation
performances.
Experiment 1 Experiment 2
Items (Yangzhou goose) (Shitou goose)
WA Conventional WA Conventional
Incubation batches 3 3 6 6
Number of eggs in each batch 432 432 196 196
E7 embryo livability (%) 93.67±0.20A 90.97±0.27B 70.57±1.61 70.22±1.79
E15 embryo livability (%) 92.94±0.21a 89.63±0.40b 68.32±1.38 68.03±2.05
Set egg hatchability (%) 86.42±0.34a 82.87±0.35b 61.72±1.29a 55.61±1.64b
Note: Values with different superscript letters in each row within the same type of geese are
significantly different (a–b: P<0.05, A–B: P<0.01).
3. WA improved embryo and muscle growth and gosling post-hatch growth performance
3.1. WA improved embryo and muscle growth
This study was conducted to compare different angle, 50°, 60° and 70°, egg turning effects on
embryo development and embryonic muscle growth, which were associated with hatchability and
gosling quality. Three tests were conducted that each used 1728 eggs, 576 eggs for each of the 3
turning angles, that were selected with an average weight of 145.0 g (range from 135.0 g to 155.0
∙ 133 ∙
g) to eliminate errors that might arise from egg weight variations. During the 28 days incubation
period, the eggs were turned once every 2 hours. On the 8th and 18th day of incubation, eggs were
candled as described above to pick out the unfertile eggs and dead embryos. After 28 days, the
eggs were then transferred into the hatcher with no further egg turning. The goslings were hatched
on 29th–30th days. The cumulative embryo livabilities to different stage of incubation, fertile egg
hatchability, hatch gosling weight and quality were summarized in Table 2. It can seen that, in
comparison with 50° egg turning, WA turning at 60° or 70°maintained embryo livability or
reduced mortality between E18 to E30, though the beneficial effects by 70° started to show before
E18. This contributed to improving the total fertile egg hatchability. Most importantly, the 70°
egg turning produced goslings on average weighed more than 6 grams heavier than did other angle
goslings, which also contributed to improve their quality.
Table 2. Effects of egg turning angles during incubation on embryo livability, hatchability,
gosling weight and quality.
Egg turning angle
Items
50° 60° 70°
Embryo livability on E8 (%) 95.46±0.56 95.16±0.39 95.66±0.55
Embryo livability on E18 (%) 92.69±0.43 92.45±0.77 93.46±0.43
Embryo livability on E30 (%) 82.58±1.29a 86.68±0.97b 85.96±0.98b
Fertile egg hatchability (%) 83.41±0.56b 87.23±0.51a 86.53±0.89a
b b
Gosling weight (g, n = 50) 98.31±8.72 99.05±9.81 105.52±8.65a
High quality goslings (%) 80.95±0.98b 91.28±0.57a 89.38±0.65a
Note: Values are presented as means ± SEM. Different letters in each line indicates significant
difference (P<0.05).
In this study, embryo weight and muscle development were also monitored on E15, E22 and
E29 or prior to hatch. The results were showed in Figure 4. Wider angle egg turning enhanced
embryo and muscle growth at the later stage of incubation, from E22 to E29, so that on E29 70°
turned embryos had the highest weight and leg muscle weight. This suggested that proper turning
or movement of the increasing heavier and bigger embryos could enhance nutrient usage that is
pertinent to growth and development.
Figure 4. Relative embryonic body weight (A) and leg muscle weight (B) of eggs of 3 turning
angles. Data were presented as means ± SEM (n = 8). Different letters above each bar indicate
differences being significant (P<0.05) at a given embryonic age. Leg muscle weight on E15
included the leg bones.
Figure 5 shows mRNA gene expression levels of the somatotropic axis genes in E29 embryo
of the 3 different egg turning angles. Egg turning by 70° angle significantly upregulated the
∙ 134 ∙
expression levels of the hypothalamic GHRH, pituitary GH and liver IGF-I than did the other 2
angle turnings. These results indicated that large or 70° angle egg turning enhanced embryo and
muscle development through augmentation of the somatotropic axis hormone gene expression
and secretion.
Figure 5. Relative mRNA expression levels of the somatotropic axis genes on E29 embryos of
50°, 60° and 70° turned eggs, respectively. Data are shown as mean values ± SEM (n=8).
Different letters above the bars denote significant (a–b, b–c: P<0.05, a–c: P<0.01) differences.
3.2. WA improved gosling post-hatch growth performance
This study tested the post-hatch growth performance of goslings from 70° egg turning versus
that of 50°. One hundred and twenty newly hatched Yangzhou goslings from each group of egg
turning angle were selected and equally separated into 6 pens of 20 birds each. All the goslings
were fed the same creep feed for the first 3 days, then the starter feed till day 40 of age, then
growing and fattening feed till day 70 or marketing age.
The goslings of 70° turning group started with a hatch weight of 106.2 ± 1.5 g in contrast to
the 100.5 ± 1.7 g of the 50° turning group (Figure 6). In the entire 70 day growing period, the
former group goslings continued to show the weight advantage over the latter group, with the
difference being significant till day 50. The growth promoting effect by 70° egg turning also was
demonstrated in leg muscle which exhibited earlier from hatch to day 30 of growth (Figure 6B),
and later from day 30 to 50 in breast muscle (Figure 6C).
Figure 6. Gosling body weight (A), relative leg (B) and breast muscle (C) weights of 50° and
70°egg turning groups during the rearing period. Asterisk on the bars indicate significant
difference between groups (*: P<0.05; **: P<0.01).
We also assayed blood concentrations of GH, IGF-I and T3 that were important to growth of
fowls. In the post-hatch period, all three hormones showed an increase in concentrations to the
peak level that was earlier by day 30 of age for GH, but later by day 50 for IGF-I and T3.
Concentrations of all 3 hormones decreased to lower levels by day 70 of age (Figure 7). Paralleled
∙ 135 ∙
to the faster growth in early stages of the 70° turning group goslings were the markedly higher
concentrations of all 3 hormones than in the 50° group goslings (Figure 7).
Figure 7. Plasma concentrations of GH (A), IGF-I (B) and T3 (C) in female goslings that were
hatched from 70° or 50° egg turning in incubation and reared to day 70 of age. Values are
presented as mean ± SEM. Asterisk indicate significant differences (*: P<0.05, **: P<0.01).
4. Conclusions
A new type of incubator has been developed through changes in design of egg trolley and egg
turning mode and direction. It allows turning eggs at wider angles for incubation of large sized
goose eggs. Incubation tests showed WA during incubation improved egg hatchability which was
more effective for the large sized goose eggs. The improved hatchability was associated with
improvements in embryonic and muscular developments through upregulation of gene expression
and secretion of somatotropic axis hormones. The growth enhancing effect by the WA during
incubation continued through the gosling post-hatch rearing period, which is again supported by
higher secretions of growth-related hormones. These results imply that the new type of incubator
can provide the goose industry a new tool for improving production efficiency of both goose
breeding and rearing.
Acknowledgements
This study was supported by China Agriculture Research System of MOF and MARA （
CARS-40-20), the Open Grant of Key Laboratory of Protected Agriculture Engineering in the
Middle and Lower Reaches of Yangtze River, Ministry of Agriculture and Rural Affairs (KF2018,
KF2019) and Jiangsu Province Postdoctoral Science Foundation Grant (2020Z396).
References
Baggott, G.K.; Deeming, D.C. and Latter, G.V. 2002. Electrolyte and Water Balance of the Early Avian
Embryo: Effects of Egg Turning. Avian and Poultry Biology Reviews, 13(2), 105–119.
Cutchin, H. R., M. J. Wineland, V. L. Christensen, S. Davis, and K. M. Mann. 2009. Embryonic
development when eggs are turned different angles during incubation. Journal of Applied Poultry Research
18(3):447–451.
Deeming, D.C., 2002. Patterns and significance of egg turning. pages 161–178 in Avian Incubation:
Behavior, Environment and Evolution. Deeming, D.C. Ed. Oxford University Press, Oxfordshire, UK.
Deeming, D. C., 2009. The role of egg turning during incubation. Avian Biology Research. 2(1):67–71.
Dai, Z., J. Yao, Y. Ren, X. Shao, Y. Huang, and Z. Shi. 2017. Development of large egg turning angle
hatching machine and its application in goose egg hatching. China Poultry. 39(6):63–66. (In Chinese)
Elibol, O., and J. Brake. 2006. Effect of egg turning angle and frequency during incubation on hatchability
and incidence of unhatched broiler embryos with head in the small end of the egg. Poultry. 85(8):1433–1437.
Funk, E. M., and J. Forward. 1960. The relation of angle of turning and position of the egg to hatchability
of chicken eggs. Poultry Science. 39(3):784–785.
Tona, K., O. Onagbesan, B. D. Ketelaere, E. Decuypere, and V. Bruggeman. 2003a. Effects of turning
duration during incubation on corticosterone and thyroid hormone levels, gas pressures in air cell, chick
quality, and juvenile growth. Poultry Science. 82(12):1974–1979.
∙ 136 ∙
Tona, K., F. Bamelis, B. D. Ketelaere, V. Bruggeman, V. M. Moraes, J. Buyse, O. Onagbesan, and E.

Decuypere. 2003b. Effects of egg storage time on spread of hatch, chick quality, and chick juvenile growth.
Poultry Science. 82(5):736–741.
Wilson, H. R. 1991. Physiological requirements of the developing embryo: Temperature and turning.
Chapter 9. Pages 145–156 in: Avian Incubation. S. G. Tullett, Ed. Butterworth-Heinemann, London, UK.
∙ 137 ∙
Joint Data Envelopment Analysis and Life Cycle Assessment of the

Canadian Egg Industry, Differentiated by Housing Type
Ian Turner *, Davoud Heidari, Nathan Pelletier
University of British Columbia, Kelowna, British Columbia, Canada
* Corresponding author. Email: ian.turner@ubc.ca
Abstract
The Canadian egg industry is currently undergoing a transition away from conventional cage
egg production systems in favour of alternative (i.e., enriched cage, single- and multi-tier free run,
free range, and organic) production systems. These systems may be characterized by differences
in environmental impacts, driven by differences in resource-use, and other management factors.
To ensure net-positive sustainability outcomes during this transition, it is important to understand
levels of resource-use efficiency currently realized in the Canadian egg industry, key drivers of
efficiency, and the environmental impact mitigation potential of industry-wide efficiency gains.
In this research, joint data envelopment analysis (DEA) and life cycle assessment (LCA) is applied
to these ends.
Using data collected from ~200 farms across Canada in 2019, DEA results indicate that
Canadian egg farms are operating at high levels of efficiency relative to one another. Variability
in calculated efficiency scores indicates that further efficiency gains are possible; however, key
drivers of efficiency were unable to be identified using the least selection and shrinkage operator
(LASSO). Environmental impact mitigation potentials of industry-wide efficiency gains to peak
levels currently realized with respect to feed and pullet inputs indicate that environmental impacts
may decrease by up to 17%, depending on the impact and housing system. Future research should
continue to focus on identifying key drivers of resource use efficiency to provide farmers with
priority target for improving environmental sustainability outcomes.
Keywords: Efficiency, egg production, impact mitigation, LASSO, LCA, poultry
1. Introduction
Livestock production systems are key contributors to many environmental impacts (Marques
et al., 2020). These impacts may be decreased through implementation of novel technologies,
increases in efficiency, or other means. Egg production is the fastest growing industrial livestock
sector in the world (UN FAO, 2020). The majority of global egg production takes place in
conventional cage housing systems; these systems, however, are being phased out due to
increasing animal welfare concerns (Ochs et al., 2018). Rather, egg production is transitioning to
alternative (i.e. enriched cage, single- and multi-tier free run, free range, and organic) production
systems on a global scale (Balsiger, 2016; Egg Farmers of Canada, 2020) These systems provide
hens with a greater space allowance, as well as access to different furnishings and enrichments,
such as perches, nests, and others.
In addition to altering animal welfare outcomes, this transition may also have significant
effects on the environmental footprint of egg production. Trade-offs in environmental impacts
between housing systems are driven by differences in farm-level resource use (particularly feed
conversion ratio), and other factors (Pelletier, 2017). These trade-offs may be investigated using
life cycle assessment (LCA), a holistic assessment framework that uses a systems-level approach
to assess the cumulative resource demands and environmental burdens of product supply chains.
LCA is a multi-criteria assessment framework, allowing for identification of impact trade-offs
and hotspots throughout product supply chains, and avoidance of burden shifting between impacts
or supply chain steps when selecting sustainability interventions (Guinée et al., 2002).
Previously, LCA studies of the Canadian egg industry, differentiated by housing system type,
indicated significant variability in levels of on-farm resource use and emissions intensities per
tonne of eggs produced (Pelletier, 2017; Turner et al., 2021). This variability was particularly
∙ 138 ∙
pronounced in alternative production systems, indicating the potential for reductions in

environmental impacts through the identification of management strategies for increasing farm-
level resource use efficiency. Identification of environmental impact mitigation potential therefore
requires assessment of levels of resource-use efficiency currently realized, and quantification of
the impact mitigation potential of industry-wide efficiency gains.
Joint application of the efficiency benchmarking technique data envelopment analysis (DEA)
and LCA is a promising avenue for performing such an assessment. DEA is a non-parametric
frontier analysis method used to explore the relationship between input and output levels. The
analysis is performed on groups of units where individuals are called decision making units
(DMUs). These may be farms, factories, etc. Efficiency levels realized by individual DMUs are
compared and an efficiency frontier is identified at the maximum realized efficiency, enveloping
all inefficient DMUs. Deviations from the efficiency frontier represent potential efficiency gains
(Cooper et al., 2007).
In this research, joint DEA-LCA is applied to the Canadian egg industry to investigate current
levels of resource-use efficiency in different housing systems, key housing system and
management variables that are associated with changes in efficiency scores, and the
environmental impact mitigation potential of industry-wide efficiency gains to the peak levels
currently realized by the most efficient farms. This research will provide important information
to the Canadian egg industry to ensure net-positive outcomes in sustainability decision making by
identifying priority strategies for sustainable intensification as the industry navigates the transition
away from conventional cage production systems.
A separate manuscript (Turner et al., 2021) describes data collection and treatment, as well as
all methodological choices and assumptions made during development of baseline Canadian egg
industry LCA models, differentiated by housing system type. Briefly, the data used in model
development was collected from ~200 egg farms, distributed across different housing system
types and provinces in Canada in 2019. The baseline LCA models are used to quantify
environmental impacts against which efficient LCA models are compared to investigate
environmental impact mitigation potentials. The baseline life cycle inventory (LCI) models (with
proprietary, third party LCI providers disconnected) used to generate life cycle impact assessment
(LCIA) results may be found on the Open Science Framework at https://osf.io/4qetu/?show=view.
The LCA performed using the efficiency targets calculated by DEA follows the same LCA
modeling principles described in Turner et al., (2021), except for quantitative uncertainty
assessment, which was not performed here.
Vázquez-Rowe et al., (2010) describe a five-step method for joint DEA-LCA. The five steps
are described as: Creation of separate LCIs for each individual DMU
1) Performance of LCIA calculations for each DMU
2) Performance of DEA using the LCIs generated in step 1 to generate efficiency targets for
each DMU
3) LCIA calculations for each DMU using the efficiency targets generated in step 3
4) Comparison of impacts calculated in steps 2 and 4
Since the goal of this analysis was to investigate environmental impact mitigation potentials
of industry-wide efficiency gains, rather than specific farm-level gains, a modified five-step
method was used here. Instead of performing LCIA calculations for each individual farm as in
steps 2 and 4, a single LCIA calculation was performed for each housing system using production
weighted average inputs per tonne of eggs produced prior to application of DEA, followed by a
second LCIA calculation using production weighted averages based on the efficiency targets
calculated for each farm.
Non-radial, slacks-based measures of efficiency were used in the DEA, and an input
∙ 139 ∙
orientation was chosen to reduce excesses of inputs while holding outputs constant. The
production-possibility set (PPS) to be used in the DEA also needed to be defined. The two most
commonly used PPS’ are the constant-returns-to-scale (CRS) and the variable-returns-to-scale
(VRS) models. These models differ in the property of scalability, which only holds in CRS
models. In CRS models, a change in inputs/outputs causes a proportional change in inputs/outputs;
this scalability is not observed in VRS models (Cooper et al., 2007). To inform the choice of PPS
used, regression analyses were performed for each housing system to investigate scalability.
Dozens of eggs produced per hen and total dozens produced on farm were each regressed on total
number of hens housed. The results of these regressions informed PPS choice.
Rules regarding the number of DMUs to be included in DEA are underdeveloped and lacking
in consensus (Wilson, 2018). Similarly, little guidance is available on choosing specific inputs
and outputs for inclusion. Vázquez-Rowe et al., (2010) suggest inclusion of the most relevant
inputs and outputs. Inputs and outputs were therefore chosen based on relative contributions to
total environmental impacts as defined by the baseline LCA models (Turner et al., 2021). Across
the majority of environmental impacts, feed, pullets, and manure management were the greatest
contributors. While manure management processes make large contributions to many
environmental impacts, it is more efficacious to include feed inputs in DEA, as reductions in feed
inputs will reduce both impacts associated with feed, impacts from manure management due to
lower amounts of manure being excreted. Based on relative contributions to impacts, therefore,
feed and pullet inputs were included as inputs in the DEA. Eggs produced was the only output
included in the analysis. DMUs were chosen for inclusion on the basis of availability of data for
these three variables. No input or output weighting was considered, and targets were used to create
efficient LCIs for each housing system. All DEA modeling was done in PIM-DEA (Emrouznejad
and Thanassoulis, 2011). LCIA calculations were performed as described in Turner et al., (2021),
and environmental impacts of the industry-wide efficiency scenario were compared to those of
the baseline models.
Finally, the relationship between efficiency scores and possible explanatory variables
representing different housing system characteristics and management practices were
investigated. Possible predictor variables were chosen on the basis of data availability, and
included length of lay cycle, mortality rate, space allowance, whether or not the pullet rearing
system matched the laying system, the feeding system used (i.e. chain and trough, feed cart, or
feed pan), number of feedings per day, type of feed (i.e. crumble, mash, pellets, other) hours of
light received per day, average barn temperature, and whether the farm used white or brown birds.
The cross-validated least shrinkage and selection operator (LASSO) was used for investigation of
drivers of efficiency. The LASSO is a regularization method that shrinks coefficient estimates
relative to the number of coefficients in the fitted model. The LASSO includes a penalty term that
forces some coefficients to zero, removing them from the model and increasing model
interpretability. This method is suitable for fitting of models in which the number of predictor
variables is greater than sample size.
Since the goal of this LASSO was not to generate models for prediction of efficiency, the data
sets were not split into training and testing subsets for model validation. Due to data availability
constraints, LASSO models could not be generated for single-tier free run, or organic production
systems. Predictor variables which exhibited no variability within housing systems were omitted
during model generation. No weights were imposed on predictor variables during model fit. All
LASSO models were generated using the glmnet package in R studio (Friedman et al. 2010;
RStudio Team 2019).
A detailed breakdown of the LCIA results for the baseline egg production models is available
in Turner et al., (2021), section 3.1. Across all housing systems, total eggs produced exhibited a
strong linear relationship with number of hens housed, while dozens of eggs per hen exhibited
∙ 140 ∙
little to no relationship with number of hens housed (Table 1). Based on these results, the property
of scalability holds for the Canadian egg industry, indicating a CRS model is the most appropriate
PPS for use in DEA.
Table 1. Coefficients and R2 values for the regression analyses performed between total dozens
produced and number of hens housed, and dozens per hen and number of hens housed, per
housing system.
Total dozens vs. Number of Dozens/hen vs. Number of hens
hens housed housed
Coefficient R2 value Coefficient R2 value
Conventional 26.79 0.96 0.000002 0.0012
Enriched 26.85 0.99 -0.000007 0.0084
Single-tier free run 26.47 0.99 0.00001 0.0039
Multi-tier free run 25.83 0.99 -0.00005 0.27
Free range 26.26 0.99 -0.0002 0.23
Organic 29.98 0.97 0.0002 0.26
3.1. Determination of efficiency targets for LCA and impact mitigation potentials
In total, 86 conventional, 41 enriched, 6 single-tier free run, 15 multi-tier free run, 9 free range,
and 10 certified organic DMUS were included in the DEA models. Generally, high levels of
efficiency and low variability were seen within all housing systems (Figure 1).
Figure 1. Levels of efficiency exhibited across each housing system.

Organic production systems exhibited the highest homogeneity in calculated efficiency
values, including the highest average percent efficiency with respect to feed and pullet inputs, and
the lowest variability in efficiency scores. Free range systems exhibited the lowest average percent
efficiency, and the highest level of variability in calculated efficiency scores. With the exception
of the low variability in organic systems and the high variability in free range systems, variability
in efficiency scores were comparable across all other housing systems (Table 2). Three farms were
identified as operating at 100% efficiency in conventional, enriched, and multi-tier free run
∙ 141 ∙
systems, two farms were identified as operating efficiently in free range systems, and one farm
was identified as operating efficiently in single-tier free run and organic systems.
Table 2. Average efficiency scores, and variability and range in calculated efficiency scores for
each housing system.
Average efficiency Variance Range
Conventional 96.43128 3.775988 9.21
Enriched 96.86561 3.520255 6.92
Single-tier free run 96.65833 4.868537 6.64
Multi-tier free run 97.62933 5.547021 8.25
Free range 96.07 15.29829 11.14
Organic 98.217 1.448046 3
Calculated efficiency targets for feed and pullet inputs indicate the potential for large
reductions in these inputs under industry-wide efficient conditions (Table 3). Under 100%
efficient conditions, feed inputs were reduced to under 2 tonnes of feed per tonne of eggs produced
in all housing systems except free range and organic, though these systems also saw large
decreases in feed inputs. Decreases in pullet inputs per tonne of eggs produced were also observed
under 100% efficient conditions for all housing systems, include a reduction to fewer than 50
pullets in conventional, enriched, and both free-run systems. Overall, under 100% efficient
conditions, decreases in pullet and feed inputs for each housing system ranged from 3.55% to
13.22%.
Table 3. Production weighted average feed and pullet inputs per tonne of eggs produced under
baseline and 100% efficient conditions, per housing system.
Feed (tonnes) Pullets (items)
2019 DEA % 2019 DEA %
Baseline efficient reduction Baseline efficient reduction
Conventional 1.92 1.79 6.77 50.74 47.32 6.74
Enriched 1.95 1.79 8.21 51.73 48.23 6.77
Single-tier free run 2.06 1.88 8.74 52.15 46.86 10.14
Multi-tier free run 2.06 1.96 4.85 51.97 49.4 4.95
Free range 2.42 2.10 13.22 54.72 51.06 6.69
Certified organic 2.25 2.10 6.67 54.34 52.41 3.55
3.2. Environmental impact mitigation potential of efficiency increases
The observed decreases in feed and pullet inputs under efficient conditions translated to
industry-wide decreases in the environmental impacts of egg production across all impacts and
housing systems compared to baseline levels reported by Turner et al., (2021). Under efficient
conditions, impact reductions ranged from 7.24% to 10.76% across all housing systems. The
largest average decrease was observed for acidifying emissions, while the lowest was observed
for freshwater aquatic ecotoxicity. Free range systems showed the highest reduction in all
environmental impacts (average: 14%) driven by the high variability in efficiency scores
observed, while organic systems showed the lowest or second lowest mitigation potential across
all impact categories (average: 5.84%). Of the remaining housing systems, average impact
reductions under efficient conditions were 8.84% for conventional, 9.02% for enriched, 8.41% for
single-tier free run, and 6.51% for multi-tier free run systems. Figure 2 presents the decreases in
greenhouse gas emissions under industry-wide efficient conditions.
∙ 142 ∙
3500
GHG Emissions (Kg CO2eq)
3000
2500
2000
1500
1000 Baseline
500
DEA-Efficient
0
Figure 2. Greenhouse gas (GHG) emissions per tonne of eggs produced in each housing system
under baseline and efficient conditions.
Across the LASSO models generated for conventional, aviary, and free range production
systems, coefficients for all predictor variables were shrunk to 0, indicating none of the included
predictor variables were significant drivers of efficient scores. As such, from this analysis it is not
possible to conclude any of the predictor variables used are key drivers of efficiency in these
housing systems. In contrast, the LASSO model generated for enriched systems shrank all
coefficients to 0 except those associated with length of lay cycle, and the use of brown-feathered
birds. Specifically, in enriched housing systems an increase of 1 day in lay cycle length was
associated with a 0.001% increase in efficiency, while the use of brown-feather birds was
associated with a 0.95% decrease in efficiency, with respect to pullet and feed inputs.
4. Discussion
The Canadian egg industry is currently navigating a transition away from conventional cage
production systems in favour of alternatives. These systems may be characterized by significant
trade-offs in environmental impacts (Turner et al., 2021). To ensure this transition results in net-
positive sustainability outcomes, it is important to understand levels of, and homogeneity in
resource-use efficiency in different housing systems used in the Canadian egg industry. This
knowledge will help determine priority strategies for improving environmental sustainability
outcomes of Canadian egg production systems.
Joint DEA-LCA of the Canadian egg industry indicates that farms are operating at high levels
of feed- and pullet-use efficiency, with relatively low levels of variability in efficiency scores.
This result is unexpected considering the housing system transition only began in Canada in
earnest in 2017, and producer experience has been repeatedly shown to play a significant role in
determining efficiency levels. This phenomenon has been observed in cotton production
(Karimov, 2014), aquaculture systems, (Murshed-E-Jahan and Pemsi, 2011) and commercial
fisheries (Vázquez-Rowe and Tyedmers, 2013). It is promising, therefore, that observed
efficiency levels are already high in alternative systems, and will likely continue to increase as
farmers continue gaining experience with these new systems.
Industry-wide increases in feed- and pullet-use efficiency also translated to reductions in
environmental impacts associated with egg production across all housing systems and impacts
investigated. Decreases in environmental impacts of up to 17% per tonne of eggs produced may
be realized through increases in feed- and pullet-use efficiency. Additionally, increased efficiency
with respect to these resources may also represent significant cost savings to farmers, as feed
represents the largest share of production costs.
Notably, under efficient conditions, feed inputs were reduced to under 2 tonnes and pullet
∙ 143 ∙
inputs under 50 birds per tonne of eggs produced for all housing systems except for free range
and organic. These results suggest that peak achievable levels of efficiency may be higher in these
systems than more intensive free range or organic production systems. This conclusion is
promising considering current trends in the housing system transition in Canada, as the proportion
of hens housed in enriched and free-run systems is growing faster than that in free range and
organic production (Egg Farmers of Canada, 2020).
As the distribution of egg farms operating different housing systems in Canada continues to
change, it is important to consider the long-term environmental sustainability implications of
different distributions. Assuming trends continue, farmers will predominantly transition to
enriched, or free-run housing systems. Feed and pullet inputs per tonne of eggs produced are lower
in the baseline enriched LCI than in either single- or multi-tier free run, translating to lower
environmental impacts in enriched production systems. Further, though percent reductions in
inputs to enriched systems were in between those of single- and multi-tier free run systems in
magnitude, under efficient conditions feed inputs to enriched systems are lower than those to free-
run systems per tonne of eggs produced. Assuming resource-use efficiency gains continue as
farmer experience grows, it seems likely that feed inputs per tonne of eggs produced will continue
to be lower than those in free-run systems. This may translate to large reductions in cumulative
environmental impacts of Canadian egg production if farmers disproportionately transition to
enriched cage systems, as feed inputs are the greatest contributor to many environmental impacts
(Turner et al., 2021). To ensure net positive sustainability outcomes, however, it is also necessary
to consider other impacts and trade-offs between enriched and free-run systems. These include
animal welfare outcomes (Ochs et al., 2018), economic outcomes related to consumer preferences
for cage-free eggs (Balsiger, 2016), and other social impacts (Pelletier, 2018).
One important drawback of DEA as an efficiency assessment tool is that it is a relative
assessment method. As such, efficiency analysis using DEA assumes that current realized levels
of efficiency are truly representative of the maximal achievable efficiency levels. This, however,
may not be the case. Additional DEA methods have been developed to explore this possibility,
such as super-efficiency metrics (Cooper et al., 2007). Further investigation of efficiency may
also be done using other methods from operations research (Heidari et al., 2021). In particular,
predictive machine learning models may be used to predict levels of efficiency beyond those
currently realized. The LASSO would be one such technique capable of this task, with some
modifications from how it was applied here. Specifically, use of the LASSO as a model to predict
efficiency scores would require splitting of available data into training and testing data sets. Once
the optimal model is found, it would be possible to predict efficiency scores based on new values
of the predictor variables included in the model.
Doing so, however, would require further investigation into identification of predictor
variables that are key drivers of farm level resource-use efficiency. Unfortunately, few concrete
results were obtained here in this regard. In the future, larger data sets and additional predictor
variables not included in the analysis would be useful for finding key drivers of efficiency. This
includes levels of producer experience, data which was not collected here. As farmers continue to
transition to alternative housing systems, ongoing efforts to continue to collect data is vital to
increase sample sizes such that key drivers of efficiency can be determined.
The work presented here represents the first report application of joint DEA-LCA to egg
production systems. The results presented here indicate that, while high levels of resource-use
efficiency are being realized on Canadian egg farms, further efficiency gains are possible. This
may translate to large decreases in environmental impacts. Further research to investigate drivers
of efficiency in each housing system is imperative in the future.
∙ 144 ∙
References
Balsiger P., 2016. Moral struggles in markets: The fight against battery cages and the ride of cage-free
eggs in Switzerland. Eur J Sociol 57:419–450.
Cooper W.W., Seiford L., Tone K., 2007. Data envelopment analysis. A comprehensive text with models,
applications, references, and DEA-Solver Software, Second. New York, NY: Springer. 490 p.
Egg Farmers of Canada, 2020. Annual Report 2020. https://www.eggfarmers.ca/wp-content/uploads/
2021/03/2020_Egg-Farmers-of-Canada_Annual-Report.pdf. Accessed June 15, 2021.
Emrouznejad A., Thanassoulis E., 2011. Performance Improvement Management Software PIM-DEA
Soft-V3.0.
Friedman J., Hastie T., Tibshirani R., 2010. Regularization paths for generalized linear models via
coordinate descent. J Stat Softw 33:1–22.
Guinée J., Gorree M., Heijungs R., Huppes G., Kleijn R., de Koning A., van Oers L., Sleeswijk A., Suh
S., Udo de Haes H., de Bruijn H., van Duin R., 2002. Handbook on Life Cycle Assessment. An operational
guide to the ISO standards. Dordrecht: Kluwer Academic Publishers. 692 p.
Heidari D., Turner I., Ardestani-Jaafari A., Pelletier N., 2021. Operations research for environmental
assessment of crop-livestock production systems. Agric Syst. In press.
Karimov A.A., 2014. Factors affecting efficieny of cotton producers in rural Khorezm, Uzbekistan: Re-
examining the role of knowledge indicators in technical efficiency improvement. Agric Food Econ 2:1–16.
Marques G.M., Teixeira C.M.G.L., Sousa T., Morais T.G., Teixeira, R.F.M., Domingos, T., 2020.
Minimizing direct greenhouse gas emissions in livestock production: The need for a metabolic theory. Ecol
Modell 434: 109259.
Murshed-E-Jahan K., Pemsi D.E., 2011. The imapct of integrated aquaculture-agriculture on small-scale
farm sustainability and farmers’ livelihoods: Experience from Bangladesh. Agric Syst 104:392–402.
Ochs D.S., Wolf C.A., Widmar N.J.O., Bir C., 2018 Consumer perceptions of egg-laying hen housing
systems. Poult Sci 97:3390–3396.
Pelletier N., 2017. Life cycle assessment of Canadian egg products, with differentiation by hen housing
system type. J Clean Prod 152:167–180.
Pelletier N., 2018. Social Sustainability Assessment of Canadian Egg Production Facilities: Methods,
Analysis, and Recommendations. Sustainability 10:1601.
RStudio Team, 2019. RStudio: Integrated Development for R
Turner I., Heidari D., Pelletier N., 2021. Life cycle assessment of Canadian egg production systems,
differentiated by housing type. Submitted to Res Conserv Recyl. Submission number: RECYCL-D-21-
01527R1.
UN FAO, 2020. The state of food security and nutrition in the world. http://www.fao.org/3/ca9692en/
ca9692en.pdf. Accessed June 12, 2021.
Vázquez-Rowe I., Iribarren D., Moreira M.T., Feijoo G., 2010. Combined application of life cycle
assessment and data envelopment analysis as a methodological approach for the assessment of fisheries. Int J
Life Cycle Assess 15:272–283.
Vázquez-Rowe I., Tyedmers P., 2013. Identifying the importance of the “skipper effect” within sources
of measured inefficiency in fisheries through data envelopment analysis (DEA). Mar Policy 38:387–396.
Wilson P.W., 2018. Dimension reduction in nonparametric models of production. Eur J Oper Res
267:349–367.
∙ 145 ∙
Research Progress of Computer Vision-Based Intelligent

Monitoring in Egg Production
Zhonghao Chen, Jinming Pan, Hongjian Lin *
* Corresponding author. Email address: linhongjian@zju.edu.cn
Abstract
Poultry egg production provides an important protein source. Egg quality, including volume
and weight, dirt and defects, must be strictly monitored in production and processing. Computer
vision-based monitoring technology is an efficient method for quality detection and is widely used
in egg industry. Recent advances such as better camera resolution, improved lighting system and
configuration, and data processing algorithms further improved detection accuracy, and also lead
the technology to a more intelligent direction. In addition, collecting egg production and quality
data throughout the laying period in each cage and egg conveyor helps implement the best
management practices and optimize the performance of laying hens. This review summarized the
latest studies on computer vision-based intelligent monitoring systems in egg production and
processing and emphasized on its application in external and internal egg quality inspection in
real time. It then discussed existing challenges and trends in its study and application.
Keywords: Computer vision technology, egg inspection, deep learning, intelligent monitoring
1. Introduction
Poultry eggs are rich in nutrients and are an important source of dietary protein, essential
amino acids (EAA), total amino acids (TAA), and lipid intake (Sun et al., 2019). Recent studies
found that chicken egg yolks are rich in complex structural lipids for lipid homeostasis in the
central neural system. As an important source of protein and lipid with various benefits to health,
poultry egg production and consumption is continuously expanding According to FAO statistics
(Figure 1), the global egg production has annually increased over the last half century. Eggs
production alone in China exceeded 31 million tons in 2016.
Figure 1. Annual production statistics in major egg producing countries.
∙ 146 ∙
In traditional poultry farms, poultry farmers usually inspect external qualities of eggs by
visually checking eggs, remove cracked eggs, and manually sort eggs by sizes. However, the
manual inspection is often time-consuming and inaccurate. For large-scale poultry enterprise,
researchers proposed several methods for egg quality inspection according to acoustic, optical and
electrical characteristics for the determination of egg volume, weight, cracks and metamorphism.
Computer vision has been widely used in the detection of poultry egg quality, including egg size
grading, cracked egg detection, dirty and blood spotty egg detection, egg freshness estimation,
egg fertilization detection and so on. Figure 2 illustrates the egg intelligent monitoring system
comprehensively. Computer vision system is mainly composed of hardware and software part.
The hardware part can be narrowed down into camera and the lighting unit, which are replaceable
for different monitoring purposes, meanwhile the software part is further classified into image
procession and data analysis (Figure 2). In recent years, computer vision technology with higher
efficiency and accuracy was developed, and the technology has become a promising tool for real-
time and nondestructive automation quality inspection, not only for size grading, but also for
internal and external quality detection (Nyalala et al., 2021). Computer vision technology
combined with deep learning algorithms has rapidly evolved in recent years and its applications
have shown a great promise (Smith et al., 2021). This review summarized the recent advances,
and further discussed about the possibility of intelligent monitoring in the context of Precision
Livestock Farming (PLF) that aims to improve animal health and welfare, and finally analyzed
the challenges and future trends.
Figure 2. Egg intelligent monitoring system based on computer vision technology.

2. Lighting unit and camera in computer vision system
Computer vision is an interdisciplinary technology that deals with automatically collecting
and applying digital images or videos for pattern recognition of objects (Choi et al., 2018). The
technology has been widely used in various processes of egg production system, including weight
prediction (Ab Nasir et al., 2018), volume estimation (Chan et al., 2018), stain identification
(Yang et al., 2018), crack detection (Guanjun et al., 2019) and fertilized eggs assessment (Dong
et al., 2021). Realization of computer vision technology follows the steps of image acquisition,
∙ 147 ∙
image processing, and pattern recognition. After image is acquired, raw images are partitioned
into multiple segments, features are then extracted, and then regression models are applied. Linear
regression, SVM classifier, Convolutional Neural Network (CNN), Artificial Neural Network
(ANN) (Soltani and Omid, 2015), KNN, Principal Component Analysis (PLS-DA) (Chen et al.,
2015), Bayesian Network (BN) classifier (Soltani and Omid, 2015) are among the most commonly
used methods for modeling.
Table 1. Hardware and software summary of computer vision for poultry egg products.
Software/
Hardware
Algorithms
Image type Parameters Achievement References
Portable Linear Volume and Use different
RGB – (Widiasri et
webcam & regression & mass angles for
Grayscale al., 2019)
LED lighting ANOVA measurement measurement
HD Webcam Reduce the (Siswantor
RGB – Volume
c270 h & LED ANN
Grayscale prediction
processing o et al.,
lamps time 2017)
Overcome the
Kinect v2 Regression bad ambient
depth camera models Depth and Volume light (Okinda et
& Natural (SVM, GPR, color images estimation conditions and al., 2020)
light ANN) egg occlusion
problems
LOG
Monochrome Overcome the
algorithm &
industrial Monochrome complicated (Guanjun et
Hysteresis Cracks identify
digital camera images egg surface al., 2019)
thresholding
& Cold light conditions
algorithm
Industrial
SVM RGB – High detection (Haoran et
camera & Crack detection
classifier Grayscale accuracy al., 2020)
Candle light
Sequenced
Extract other
High-speed wave signal
RGB – characteristic (Sun et al.,
digital camera extraction and Crack detection
Grayscale signal of 2017b)
& LED lamps identification
cracked eggs
algorithm
A non- (Soltani
CCD camera ANN, SVM,
Freshness destructive
& LED BN & ____
inspection detection of
and Omid,
lighting unit Decision tree 2015)
egg freshness
Visible
PLS-DA,
NIR absorbance
KNN and Blood spots (Chen et
spectrometer ____ spectroscopy
BLR detection al., 2015)
camera combined with
algorithm
BLR model
Identify
Digital CCD Hatching
LDA, NB & RGB – infertile duck (Dong et
camera & characteristics
SVM Grayscale egg using al., 2021)
LED lamps (infertile)
machine vision
The choice of computer vision configuration primarily depends on egg parameters to be
detected, as shown in Table 1. The hardware mainly consists of camera and the lighting unit, and
the software is developed for the purpose of data acquisition and analysis. To ensure the good
quality of images, a suitable lighting unit and camera are necessary, which combined with robust
algorithms can perform well. Light source illuminate objects and is reflected to camera. Light
directly affects image quality of texture and color. At present, natural light (Okinda et al., 2020)
and light-emitting diode (LED) (Siswantoro et al., 2017) are two main sources that are in use.
∙ 148 ∙
Camera unit is another key component in the computer vision system. Cameras used in egg
farms are mainly charge coupled device (CCD) cameras. Kinect V2 Depth Camera can obtain
depth information of egg and is used for precise estimation of the egg volume (Okinda et al.,
2020). Monochrome Digital Camera was used to capture images of egg surface for cracks
detection (Guanjun et al., 2019). High-speed digital camera was used for online identification of
cracked eggs with high accuracy and processing speed (Sun et al., 2017b). To detect the internal
information of eggs parameters, such as blood spots inside eggs (Zhu et al., 2017), egg freshness
(Soltani and Omid, 2015), egg fertilization (Dong et al., 2019), and fake egg detection (Joshi et
al., 2020), researchers often use NIR camera (Chen et al., 2015).
3. Applications of computer vision system in egg industry
There are multiple applications of computer vision system in egg farms. The egg size
classification depends on weight and volume estimation, the external defects detection focuses on
stains and cracks detection, and the internal quality mainly includes egg freshness, blood spots
and fertilization of eggs. The next section elaborates the research advances in those three aspects.
In precision livestock farming, many egg quality parameters are necessary for monitoring so that
managers can adjust the poultry farming strategy in time. However, researchers mainly focus on
the egg quality inspection of egg products, including weight and size, volume and mass, cracks
and dirt, freshness and egg yolk status (Table 2). The applications of the poultry egg’s external
and internal quality inspection are summarized in Table 2, which shows that egg industry need a
more comprehensive monitoring with higher accuracy.
3.1. Egg grading
Egg size grading depends on accurately estimating the egg weight, volume or other geometric
dimension. Experimental studies showed the egg weight and its geometric parameters are related,
so the egg weight can be estimated through image. A study reported that linear correlation
coefficient between egg weight and area is 0.9891. Models based on this linear relationship
achieved classification errors of 2.5 % for training set and 12.5 % for testing set (Alikhanov et al.,
2017). Therefore, egg weight prediction model was not sufficiently accurate. Some studies
designed a Fuzzy Inference System (FIS) for better egg size automatic sorting (Omid et al., 2013).
A depth-image based egg volume estimation system was designed (Okinda et al., 2020) to
evaluate single-egg (no occlusion) and multi-eggs (partially occluded, i.e., simple and complex)
situations. Contour curvature analysis and K-closest M-circle center algorithms were used to
segment egg images, achieving RMSE of 1.175 cm3 and R2 of 0.984 for single egg image and
RMSE of 1.080 cm3 and 1.294 cm3 for multi-egg image with simple and complex occlusion,
respectively. Depth camera or 3D Camera can be used to obtain the depth information or 3D
surface image of eggs. Combined with the point cloud processing algorithm, the accurate
estimation of egg volume and egg shape index was achieved (Chan et al., 2018). Zhang et al.
obtained the 3D point cloud of eggs by using multi-angle images, adopted refinement algorithm
based on incremental Delaunay triangulation. The study reached a coefficient of determination of
99% (Zhang et al., 2016).
Researchers have made different optimization in software to improve the accuracy in real-
time detection. For example, the CVS for predicting the volume of eggs using ANN structure
showed a good accuracy and less computing time (Soltani et al., 2015; Siswantoro et al., 2017).
2-D imaging was assisted with geometrical transformation of an egg into one of the known ovoid
to determine its volume and surface area (Narushin et al., 2020a; Narushin et al., 2020b). Deep
Convolutional Neural Network (CNN) or K-nearest Neighbor Network (K-NN) were used to
perform egg volume classification tasks (Ab Nasir et al., 2018; Nasiri et al., 2020).
3.2. External quality inspection
Eggshell defects and dirt may allow bacteria crack into the egg interior, causing egg to rot
(Wang et al., 2020) or deteriorating eggshell appearance and the sale value of commodity eggs
∙ 149 ∙
(Wang et al., 2019). The process of detecting eggshell cracks and stains by using image processing
technology includes several steps. First, egg images are collected, and then cracks or dirt region
are identified as features in training samples (Yang et al., 2018). Next, the training samples will
be input into the neural network for model training. The last step is to use the trained neural
network model to test recognition using testing samples. Following a similar procedure, the vision
system can be adapted for detection of defects such as corrugated, wrinkled, pimpled, odd-shaped
and misshapen eggs.
Table 2. Applications of the poultry egg’s external and internal quality inspection.
Advantages
Category Parameters Algorithm Accuracy References
/Limitations
Fuzzy logic Sample size is (Omid et al.,
Size grading 95%
classifier small 2013)
Weight and Poor
Size Weight Regression performance of (Alikhanov et
97.5%
estimation analysis the regression al., 2017)
model
The The disc
Volume & Regression
Volume and significant calculating (Widiasri et al.,
mass analysis
Mass value is method was 2019)
measurement
0.982 applied
LOG
Use negative
algorithm &
Cracks LOG algorithm (Guanjun et al.,
Hysteresis 92.5%
identify to enhance the 2019)
thresholding
cracks
algorithm
Achieve a
Cracks, blood higher
BiLSTM (Turkoglu,
and dirt 99.17% accuracy using
Cracks and Dirt network 2021)
detection BiLSTM
network
94.3%
(white Low accuracy
Dirt stains egg); cannot meet (Yang et al.,
FCM cluster
detection 90.0% the practical 2018)
(brown application
egg)
88.75% to Investigate
LIBSVM,
100.00% different
Freshness DCA, LDA, (Dai et al.,
(different incident angles
evaluation KNN, RF & 2020)
incident influencing the
NB
Freshness and angle) egg’s freshness
Egg yolk CARS, IPLS
PLSDA, and SPA were
Blood spot (Zhu et al.,
SVM and 95% used to
detection 2017)
SBDA model simplify the
wavebands
The accuracy
Infertile duck LDA, NB
could be (Dong et al.,
egg and SVM 92.06%
Hatching further 2021)
identification algorithm
characteristics improved
(fertile and The VIS/NIR
Unfertilized
infertile) SNV and NB transmittance (Dong et al.,
duck eggs 94.54%
model spectroscopy 2019)
identification
was applied
∙ 150 ∙
A negative LOG operator was proposed to effectively enhance the cracks in the egg image,
and an improved LFI (Local Fitting Image) index was used to distinguish the crack region from
the mislabeled region. The experimental results showed a cracked egg recognition rate of 92.5%
(Guanjun et al., 2019). A method consists of a sequenced wave signal extraction algorithm was
tested to extract signals from the translucent images of duck eggs, and then the characteristic
parameters of the signals were calculated and analyzed. The accuracy rate was 93.1% (Sun et al.,
2017a). Muammer Turkoglu proposed a novel Sequential Multiple Image-based Convolutional
Neural Network-BiLSTM (SMI-CNN-BiLSTM) model (Turkoglu, 2021), which obtained deep
features from egg images based on pretrained residual network (DenseNet201) model. The
obtained features were then fed into the Bidirectional Long-Short-Term-Memory (BiLSTM). The
experimental results showed the proposed model achieved a superior accuracy score of 99.17%,
better than SVM method (Haoran et al., 2020).
A vision system was improved for defect detection by scanning a laser pattern of structured
light and imaging, highlighting the changes in geometry as a result of deformation of the laser
transitions generated by scanning the egg surface. The obtained information was used to classify
the defective samples, and achieved a 97.5% efficiency (Mota-Grajales et al., 2019). Similarly,
another vision system based on a modified pressure chamber and a continuously rotating egg was
designed and implemented for egg crack detection. The experiments were validated using 750 egg
surface images with 94% accuracy (Priyadumkol et al., 2017).
3.3. Internal quality inspection
Eggs are highly perishable and quality deterioration occurs during storage, distribution, sale,
and processing. The main indicators of egg internal quality include the freshness, blood spots,
yolk morphology and fertilization (Dong et al., 2019). Non-destructive techniques are of
paramount importance in internal quality inspection (Loffredi et al., 2021). Computer vision
system is usually difficult to obtain satisfactory information for internal quality inspection if
directly using CCD camera (Sun et al., 2014; Soltani and Omid, 2015; Dong et al., 2021).
Therefore, spectroscopy cameras, i.e., VISNIR, NIR, Raman, microwaves, hyperspectral imaging,
pulsed IR thermography systems, are adopted for non-invasively detection.
Researchers studied the influence of incident angles of light source on the accuracy. A variety
of classifiers based on spectra after preprocessing and feature wavelength extraction obtained
three classifiers with the highest accuracy. The three classifiers were used as metamodels of
stacking ensemble learning and improved the highest accuracy from 96.25% to 100% (Dai et al.,
2020). In another study, the scattering parameters of the eggs were acquired in the range of 0.9–
1.7 GHz of microwave spectra. PLS and ANN regression methods were implemented to predict
the egg quality indices and SIMCA and ANN classification methods were applied to classify the
eggs in terms of storage time. The ANN resulted in a total accuracy of 100% (Akbarzadeh et al.,
2019).
Blood spots in eggs are another important internal quality parameter; however, the absorption
feature of pigment in brown eggshell is similar to that of blood spots. Chen et al. explored the
optimal discrimination method based on visible absorbance spectroscopy. The spectra of 96
brown-shell normal eggs and 98 brown-shell artificial bloodspot eggs were collected in the
spectral range of 200–1100 nm by a prototype egg internal quality detection system with a
conveyor speed of 4 eggs per second. Three discrimination methods, i.e., partial least squares
discriminant analysis (PLS-DA), k-nearest neighbor (KNN) and binary logistic regression (BLR)
were compared. The study found the BLR method was better than PLS-DA and KNN, and the
best discrimination rates for the training set and prediction set were 95.4% and 96.9% (Chen et
al., 2015). In related studies, researchers simplified the whole band by using Competitive
Adaptive Reweighed Sampling (CARS), Interval Partial Least Squares (IPLS) and Successive
Projections Algorithm (SPA). Partial Least Square Discriminant Analysis (PLSDA) and Support
Vector Machine (SVM) models were established then by using spectral characteristic variables
∙ 151 ∙
selected by the three methods. Also, other models like the stepwise Bayesian discriminant analysis
model can reach a 95% accuracy (Zhu et al., 2017).This model greatly reduces the number of
characteristic spectral variables, which can avoid the influence of excessive spectral variables on
the accuracy in real-time online detection of blood spot eggs.
The viability evaluation of embryos of hatching eggs is important for both the food processing
and poultry farming industries. A hyperspectral visible/near infrared line-scan imaging system
was used to determine the viability of fertilized eggs. The acquired image of each egg (88 eggs in
total) was analyzed using Principal Component Analysis (PCA) with a texture co-occurrence
method for image processing. The generated PCA-core images exhibited significant differences
between the blood vessels of the fertilized viable and nonviable eggs. The results indicated that
by utilizing a single-band (560 nm) image and simple image processing methods, the viability of
fertilized eggs can be determined with an accuracy of approximately 99% (Park et al., 2019). In
further research, a new classification model based on Fully Convolutional Networks (FCNs) and
a Gated Recurrent Unit (GRU) was proposed. The model detected an embryo viability by
determining the indicators of fetal heartbeat signals. Experimental results based on the data set
showed that the model is the most accurate and capable of classifying 83 embryos per second
(Geng et al., 2019).
4. Conclusion and future perspectives
This review summarized computer vision system for application in egg’s external and internal
quality inspection. Some recent studies have focused on improving detection accuracy by
combining deep learning methods or improving visual systems, while others have focused on
optimizing algorithmic frameworks to improve the detection speed of online real-time detection.
In addition to ensuring grading and quality of egg products, a real-time monitoring system may
also optimize feeding and breeding management through timely feedback of egg production
information. The integration of data from computer vision monitoring system and management
practices showed a great promise in future precision livestock farming.
References
Ab Nasir, A.F., S.S. Sabarudin, A.P.P. Abdul Majeed, A.S. Abdul Ghani, 2018. Automated egg grading
system using computer vision: Investigation on weight measure versus shape parameters. IOP Conference
Series: Materials Science and Engineering. 342, 012003. http://doi.org/10.1088/1757-899x/342/1/012003.
Akbarzadeh, N., S.A. Mireei, G. Askari, A.H. Mahdavi, 2019. Microwave spectroscopy based on the
waveguide technique for the nondestructive freshness evaluation of egg. Food Chem. 277, 558–565.
http://doi.org/10.1016/j.foodchem.2018.10.143.
Alikhanov, J., S.M. Penchev, T.D. Georgieva, A. Moldazhanov, Z. Shynybay, P.I. Daskalov, 2017. An
indirect approach for egg weight sorting using image processing. Journal of Food Measurement and
Characterization. 12 (1), 87–93. http://doi.org/10.1007/s11694-017-9619-x.
Chan, T.O., D.D. Lichti, A. Jahraus, H. Esfandiari, H. Lahamy, J. Steward, M. Glanzer, 2018. An Egg
Volume Measurement System Based on the Microsoft Kinect. Sensors (Basel). 18 (8), 2454. http://doi.org/
10.3390/s18082454.
Chen, M., L. Zhang, H. Xu, 2015. On-line detection of blood spot introduced into brown-shell eggs using
visible absorbance spectroscopy. Biosystems Engineering. 131, 95–101. http://doi.org/10.1016/
j.biosystemseng.2015.01.006.
Choi, I., J. Kim, J. Jang, 2018. Development of Marker-Free Night-Vision Displacement Sensor System
by Using Image Convex Hull Optimization. Sensors (Basel). 18 (12), 4151. http://doi.org/10.3390/s18124151.
Dai, D., T. Jiang, W. Lu, X. Shen, R. Xiu, J. Zhang, 2020. Nondestructive Detection for Egg Freshness
Based on Hyperspectral Scattering Image Combined with Ensemble Learning. Sensors (Basel). 20 (19), 5484.
http://doi.org/10.3390/s20195484.
Dong, J., X. Dong, Y. Li, Y. Peng, K. Chao, C. Gao, X. Tang, 2019. Identification of unfertilized duck
eggs before hatching using visible/near infrared transmittance spectroscopy. Computers and Electronics in
Agriculture. 157, 471–478. http://doi.org/10.1016/j.compag.2019.01.021.
Dong, J., B. Lu, K. He, B. Li, B. Zhao, X. Tang, 2021. Assessment of hatching properties for identifying
multiple duck eggs on the hatching tray using machine vision technique. Computers and Electronics in
Agriculture. 184, 106076. http://doi.org/10.1016/j.compag.2021.106076.
∙ 152 ∙
Geng, L., H. Wang, Z. Xiao, F. Zhang, J. Wu, Y. Liu, 2019. Fully Convolutional Network With Gated
Recurrent Unit for Hatching Egg Activity Classification. IEEE Access. 7, 92378–92387. http://doi.org/
10.1109/access.2019.2925508.
Guanjun, B., J. Mimi, X. Yi, C. Shibo, Y. Qinghua, 2019. Cracked egg recognition based on machine
vision. Computers and Electronics in Agriculture. 158, 159–166. http://doi.org/10.1016/j.compag.
2019.01.005.
Haoran, C., H.E. Chuchu, J. Minlan, L.I.U. Xiaoxiao, 2020. Egg crack detection based on support vector
machine. 2020 International Conference on Intelligent Computing and Human-Computer Interaction (ICHCI),
80-83 p. http://doi.org/10.1109/ichci51889.2020.00025.
Joshi, R., S. Lohumi, R. Joshi, M.S. Kim, J. Qin, I. Baek, B.-K. Cho, 2020. Raman spectral analysis for
non-invasive detection of external and internal parameters of fake eggs. Sensors and Actuators B: Chemical.
303, 127243. http://doi.org/10.1016/j.snb.2019.127243.
Loffredi, E., S. Grassi, C. Alamprese, 2021. Spectroscopic approaches for non-destructive shell egg
quality and freshness evaluation: Opportunities and challenges. Food Control. 129, 108255. http://doi.org/
10.1016/j.foodcont.2021.108255.
Mota-Grajales, R., J.C. Torres-Peña, J.L. Camas-Anzueto, M. Pérez-Patricio, R. Grajales Coutiño, F.R.
López-Estrada, E.N. Escobar-Gómez, H. Guerra-Crespo, 2019. Defect detection in eggshell using a vision
system to ensure the incubation in poultry production. Measurement. 135, 39–46. http://doi.org/10.1016/
j.measurement.2018.09.059.
Narushin, V.G., G. Lu, J. Cugley, M.N. Romanov, D.K. Griffin, 2020a. A 2-D imaging-assisted
geometrical transformation method for non-destructive evaluation of the volume and surface area of avian
eggs. Food Control. 112, 107112. http://doi.org/10.1016/j.foodcont.2020.107112.
Narushin, V.G., M.N. Romanov, G. Lu, J. Cugley, D.K. Griffin, 2020b. Digital imaging assisted geometry
of chicken eggs using Hügelschäffer's model. Biosystems Engineering. 197, 45–55. http://doi.org/10.1016/
j.biosystemseng.2020.06.008.
Nasiri, A., M. Omid, A. Taheri-Garavand, 2020. An automatic sorting system for unwashed eggs using
deep learning. Journal of Food Engineering. 283, 110036. http://doi.org/10.1016/j.jfoodeng.2020.110036.
Nyalala, I., C. Okinda, C. Kunjie, T. Korohou, L. Nyalala, Q. Chao, 2021. Weight and volume estimation
of poultry and products based on computer vision systems: a review. Poult Sci. 100 (5), 101072. http://doi.org/
10.1016/j.psj.2021.101072.
Okinda, C., Y. Sun, I. Nyalala, T. Korohou, S. Opiyo, J. Wang, M. Shen, 2020. Egg volume estimation
based on image processing and computer vision. Journal of Food Engineering. 283, 110041. http://doi.org/
10.1016/j.jfoodeng.2020.110041.
Omid, M., M. Soltani, M.H. Dehrouyeh, S.S. Mohtasebi, H. Ahmadi, 2013. An expert egg grading system
based on machine vision and artificial intelligence techniques. Journal of Food Engineering. 118 (1), 70–77.
http://doi.org/10.1016/j.jfoodeng.2013.03.019.
Park, E., S. Lohumi, B.-K. Cho, 2019. Line-scan imaging analysis for rapid viability evaluation of white-
fertilized-egg embryos. Sensors and Actuators B: Chemical. 281, 204–211. http://doi.org/10.1016/
j.snb.2018.10.109.
Priyadumkol, J., C. Kittichaikarn, S. Thainimit, 2017. Crack detection on unwashed eggs using image
processing. Journal of Food Engineering. 209, 76–82. http://doi.org/10.1016/j.jfoodeng.2017.04.015.
Siswantoro, J., M.Y. Hilman, M. Widiasri, 2017. Computer vision system for egg volume prediction using
backpropagation neural network. IOP Conference Series: Materials Science and Engineering. 273, 012002.
http://doi.org/10.1088/1757-899x/245/1/012002.
Smith, M.L., L.N. Smith, M.F. Hansen, 2021. The quiet revolution in machine vision – A state-of-the-art
survey paper, including historical review, perspectives, and future directions. Computers in Industry. 130,
103472. http://doi.org/10.1016/j.compind.2021.103472.
Soltani, M., M. Omid, 2015. Detection of poultry egg freshness by dielectric spectroscopy and machine
learning techniques. LWT - Food Science and Technology. 62 (2), 1034–1042. http://doi.org/10.1016/
j.lwt.2015.02.019.
Soltani, M., M. Omid, R. Alimardani, 2015. Egg volume prediction using machine vision technique based
on pappus theorem and artificial neural network. J Food Sci Technol. 52 (5), 3065–3071. http://doi.org/
10.1007/s13197-014-1350-6.
Sun, C., J. Liu, N. Yang, G. Xu, 2019. Egg quality and egg albumen property of domestic chicken, duck,
goose, turkey, quail, and pigeon. Poult Sci. 98 (10), 4516–4521. http://doi.org/10.3382/ps/pez259.
Sun, K., L. Ma, L. Pan, K. Tu, 2017a. Sequenced wave signal extraction and classification algorithm for
duck egg crack on-line detection. Computers and Electronics in Agriculture. 142, 429–439. http://doi.org/
10.1016/j.compag.2017.09.034.
∙ 153 ∙
Sun, K., W. Zhang, L. Pan, K. Tu, 2017b. Recognition of a Cracked Hen Egg Image Using a Sequenced
Wave Signal Extraction and Identification Algorithm. Food Analytical Methods. 11 (4), 1223–1233.
http://doi.org/10.1007/s12161-017-1105-x.
Sun, L., L.-m. Yuan, J.-r. Cai, H. Lin, J.-w. Zhao, 2014. Egg Freshness on-Line Estimation Using Machine
Vision and Dynamic Weighing. Food Analytical Methods. 8 (4), 922–928. http://doi.org/10.1007/s12161-
014-9944-1.
Turkoglu, M., 2021. Defective egg detection based on deep features and Bidirectional Long-Short-Term-
Memory. Computers and Electronics in Agriculture. 185, 106152. http://doi.org/10.1016/j.compag.2021.
106152.
Wang, C., J. Zhou, H. Wu, J. Li, Z. Chunjiang, R. Liu, 2020. Research on the Evaluation Method of
Eggshell Dark Spots Based on Machine Vision. IEEE Access. 8, 160116–160125. http://doi.org/
10.1109/access.2020.3020260.
Wang, D.H., H. Chen, R.Y. Zhou, C.X. Huang, H.X. Gao, B.L. Fan, G.J. Liu, Z.H. Ning, 2019. Study of
measurement methods on phenotype of translucent eggs. Poult Sci. 98 (12), 6677–6683. http://doi.org/
10.3382/ps/pez539.
Widiasri, M., L.P. Santoso, J. Siswantoro, 2019. Computer vision system in measurement of the volume
and mass of egg using the disc method. IOP Conference Series: Materials Science and Engineering. 703,
012050. http://doi.org/10.1088/1757-899x/703/1/012050.
Yang, Q., M. Jia, Y. Xun, G. Bao, 2018. Detection of egg stains based on local texture feature clustering.
International Journal of Agricultural and Biological Engineering. 11 (1), 199–205. http://doi.org/10.25165/
j.ijabe.20181101.2592.
Zhang, W., X. Wu, Z. Qiu, Y. He, 2016. A novel method for measuring the volume and surface area of
egg. Journal of Food Engineering. 170, 160–169. http://doi.org/10.1016/j.jfoodeng.2015.08.025.
Zhu, Z.-H., W.-Q. Li, Q.-H. Wang, Y. Tang, F.-L. Cao, R. Ma, 2017. Online Discriminant Model of
Blood Spot Eggs Based on Spectroscopy. Journal of Food Process Engineering. 40 (3), e12435. http://doi.org/
10.1111/jfpe.12435.
∙ 154 ∙
Performance Analysis and Validation of a Novel Dehumidification

System for Closed Livestock House in Cold Region
Ping Zheng a, Jicheng Zhang a, Jun Bao b, Qiuju Xie a,*
a
College of Electrical and Information, Northeast Agricultural University,
Harbin, Heilongjiang 150030, China
b
College of Animal Science and Technology, Northeast Agricultural University,
* Corresponding author. Email: xqj197610@163.com
Abstract
To solve the common contradiction between ventilation dehumidification and heat
preservation in closed livestock house in cold region in winter, a novel dehumidification system
was studied based on the principle of condensation dehumidification using the climate
characteristics of big Temperature Difference Between Indoor and Outdoor (TDBIO). Natural-
cooling refrigerant was applied circularly in this system to reduce indoor water vapor mass, which
could avoid sending indoor air to outdoor directly to decrease massive heat loss. An experiment
was carried out in a closed chamber to test and verify the performance of this dehumidification
system in January 2021. Dehumidification rate (DR) was calculated with the theory of wet air,
and the mean total heat transfer coefficient (HTC) was deduced by the effectiveness-number of
transfer units (ε- NTU) in this study. The results showed that there were two important factors
influencing on dehumidification, i.e., TDBIO and fan air flow rate. When TDBIO was 28.9 °C
and fan air flow rate was 2.5 m s-1, the averaged and maximum DR were 0.9 kg h-1 and 3.63
kg h-1, respectively. When TDBIO was about 30 ℃, the mean total HTCs were 17.72 W m-2
°C-1, 14.79 W m-2 °C-1 and 10.2 W m-2 °C-1 at different air flow rate of 2.5 m s-1, 2.0 m s-1 and 1.5
m s-1, respectively. The mean total HTCs could be used for DR calculation when TDBIO was
given. The results also showed that this new developed dehumidification system saved about half
heat loss compared to positive-pressure ventilation system theoretically to achieve the same
requirement of dehumidification.
Keywords: Closed livestock building, dehumidifier, effectiveness-number of transfer units (ε-
NTU), wet air theory, energy saving.
1. Introduction
At present, there is a contradiction between ventilation and heat preservation in livestock and
poultry production in cold winter (Barber et al., 1989; Harmon et al., 2010). In order to keep an
appropriate temperature for animals in winter in the closed house, the environment control
strategies of less or non-ventilation are always used to keep warm, which could result in an
extremely high humid indoor environment condition. This high humid living environment may
increase the opportunity of microorganisms’ growth, and the possibility of animal respiratory
diseases (Berckmans, 2014; Stinn and Xin, 2014). Although ventilation could decrease the
humidity and exchange the fresh air, the indoor temperature will be decreased rapidly at the same
time. It is reported that the heat lost by livestock buildings through ventilation during winter
accounts for 70–90% (Spengler and Stombaugh, 1983), and it can be reduced by using a
recuperative air-to-air heat exchanger (Allen and Payne, 1987; Rösmann and Büscher, 2010;
Abbaspour-Fard et al., 2011).
Currently, there are some measures for indoor environment control, i.e., dehumidifiers, air
conditioners and ground source heat pumps (Badescu, 2007; Licharz et al., 2020) in some
commercial or scientific research farms. But the costs and energy consumption for these
environmental device and equipment increase extra economy burden to livestock farms (Boris,
2008). Up to now, there are few reports on special dehumidification system with low energy
consumption designed for livestock and poultry houses in cold region. (Krommweh et al., 2014).
∙ 155 ∙
Dehumidifier of Condensation Dehumidification is an effective and common way to reduce

the humidity in civil house (Campen and Bot, 2002). It is generally based on the principle of
condensation and water separation. When the hot-humidity air pass through the cold medium of
dehumidifier, the air temperature is reduced to lower than some dew point temperature under a
certain relative humidity (RH), and leads to some water vapor be separated out. This method
avoids indoor air exhaust to the outside directly and maximizes the heat retention through the air
indoor circulation. However, the high cost of equipment and power consumption of compressor
limit the wide use of dehumidifiers in animal farming (Tikhomirov et al., 2020).
The objective of this study is to design a new dehumidification system that takes full
advantages of the characteristics of natural low temperature in cold winter and follows the
principle of condensation and water separation to achieve dehumidification effect with low indoor
heat loss and low equipment cost.
2.1. Description of dehumidification system
The overall design of this dehumidification system is mainly consisted of indoor and outdoor
fin-tube heat exchangers, indoor and outdoor fans, a refrigerant tank, a refrigerant circulating
pump and pipes (Figure 1). The indoor and outdoor heat exchangers, circulating pump and
refrigerant tank constitute a set of refrigerant circulation system through pipes.
1. Humidifier; 2. Indoor fan; 3. Indoor heat exchanger; 4. Outdoor heat exchange; 5. Refrigerant
bank; 6. Pump; 7. Tubes; 8. Temperature and humidity sensors; 9. Anemometer
Figure 1. Systemic structure of the dehumidification system.
Refrigerant tank is located outdoor in low temperature environment (Figure 1, a). When this
system works, liquid refrigerant being cooled naturally is circulated in the refrigerant circulation
system by a refrigerant pump. Low-temperature refrigerant passing through the tubes of indoor
heat exchanger (Figure 1, b) makes indoor heat exchanger to be local low temperature (Figure 1,
c). It makes the temperature of indoor moisture air imported by the fan drop below the dew point
of the saturated humidity and precipitates condensate water (Figure 1, d). Meanwhile, the
refrigerant in the tubes of indoor heat exchanger is heated up. Then, the refrigerant consequently
flows back to the outdoor refrigerant bank for cooling. To quickly reduce the temperature of
flowback refrigerant and improve dehumidification efficiency, a heat exchanger with fan can be
installed outdoors to accelerate refrigerant releasing heat.
2.2. Experiment description
To evaluate the main impact factors of dehumidification and cooling performance of this
humidification system, an experiment was carried out in the laboratory of Northeast Agricultural
University, Harbin, China from January 2 to 22, 2021. An indoor closed chamber 5.0 m × 4.0 m
∙ 156 ∙
× 3.0 m (L × W × H) was built (Figure 2) in the laboratory. The experiments were carried out
under the conditions of indoor fan air flow rate of 1.0 m s-1, 1.5 m s-1 and 2.0 m s-1 and refrigerant
flow rate of 8.0 L min-1. Harbin is located between 125°41’ to 130°13’N, and 44°04’ to 46°04’E.
The average outdoor temperature in January was about -19 °C, and the RH was about 72%
(http://www.tianqihoubao.com/lishi/haerbin/month/202101.html). The dimension of inside and
outside fin-tube heat exchanger was 0.53 m × 0.18 m × 0.53 m (L × W × H) (Figure 2). The heat
exchange area and cross-sectional area of heat exchanger was about 18 m2 and 0.16 m2,
respectively. The diameter of variable rate fan fixed on heat exchanger was 0.4 m. Power of the
fan and refrigerant pump were 190 W and 46 W, respectively. Three temperature and humidity
sensors (model SHT20, Renzhi Measurement and Control Technology Co., Ltd., Jinan, China)
were equipped in front and back of the heat exchanger, and in the middle of this chamber. An
anemometer (model AS836, Jingxun Electronic Technology Co., Ltd., Jinan, China) with 0.2
m s-1 accuracy was installed at the inlet of the indoor heat exchanger. The control system collected
data of indoor and outdoor temperature and humidity to control the operation on fans and
refrigerant pump.
A B
Figure 2. The picture of dehumidification system. A: indoor; B: outdoor.

To simulate the high-humidity environment of animal house in cold winter, the indoor relative
humidity of the closed chamber was controlled higher than 90% using a humidifier. Then the
dehumidification system was operated, i.e., indoor and outdoor fans, the refrigerant circulation
pump were turned on, at the same time. The air temperature and humidity of inlet and outlet of
heat exchanger were monitored continuously until the temperature and RH of indoor closed
chamber were stable.
2.3. Numerical model and physical properties
2.3.1. Calculation of dehumidification rate based on wet air theory
The dehumidification efficiency of this new dehumidification system is simulated by using
the theory of wet air. To simplify the calculation, the following assumptions are made: (1) wet air
has a constant physical property, which is composed of dry air and water vapor. (2) The
condensation process is based on the central nucleation hypothesis, ignoring the direct
condensation caused by the water vapor in the wet air cooling at the non-fluid solid interface; (3)
The contact thermal resistance of fin-tube is ignored, and the influence of environmental thermal
radiation is ignored; (4) The fouling factor is ignored.
pw
W  0.622
p  pw (1)
mvapor  mdryW
(2)
∙ 157 ∙
mvapor
mcond .  (3)
t
where W is moisture content, kg kg-1(dry air); Pw is the partial vapor pressure, kPa; Pws is the
partial saturation vapor pressure, kPa; P is atmosphere pressure, kPa; mdry is dry air mass in closed
chamber, kg; mvapor is water vapor mass in air, kg; Δmvapor is water vapor mass difference between
two time points, kg; Δt is time difference, min; mcond . is DR, kg h-1. The physical properties of
indoor air are obtained in saturation vapor pressure table and air density table (Incropera and
Dewitt, 2011).
The temperature and RH in the test chamber are collected by the temperature and humidity
sensors. The air moisture content is calculated using Eq. (1). Combined with the air density table
and the chamber volume, water vapor mass is calculated using Eq. (2). The DR of water vapor in
any length of time can be calculated using Eq. (3).
2.3.2 Heat calculated by temperature difference
The heat loss of dehumidification can be calculated by the air temperature difference.
Q  c p , h mT (4)
)  r 2vt )
TindoorV  (Tindoor  Tout（
Tindoor '  (5)
V
Qdehum.
  100% (6)
Qvent .
where Q is heat loss, kJ; cp,h is specific heat capacity at constant pressure of air kJ kg-1 °C-1; m is
air mass, kg; ΔT is temperature difference, °C ; Tindoor and Tout are the indoor and outdoor
temperature, respectively, °C ; Tindoor’ is new indoor temperature after ventilation, °C ; V is the
volume of the chamber, m3; r, v and t are the radius of fan (m), air flow rate (m s-1) and duration
time (s); Qdehum. and Qvent. are the heat loss of dehumidification system and positive-pressure
ventilation, respectively, kJ; δ is the ratio of heat loss of dehumidification system and positive-
pressure ventilation.
2.3.3 Effectiveness-number of transfer units (ε- NTU method) calculation
In order to simplify the calculation, it is assumed that: (1) the physical properties are constant;
(2) The heat resistance and fouling of heat exchanger can be ignored; (3) The heat transfer
coefficient of the whole fin-tube heat exchanger is the same.
During dehumidification period, the temperature of refrigerant increases in the tube of indoor
heat exchanger, so the refrigerant absorbs energy. While air temperature decreases when flowing
through indoor heat exchanger, so the indoor air releases energy. The mass flow rates of air and
refrigerant are
mh  hvAfan (7)
mc =c H (8)
where mh is mass flow rate of air, kg s-1; h is air density, kg m-3; v is air flow rate, m s-1; Afan is
the area of heat exchanger, m2; mc is mass flow rate of the refrigerant, kg s-1; c is refrigerant
density, kg m-3; H is flow rate of the refrigerant e, m3 s-1.
Cmin  mhc p , h (9)
Cmax  mc c p ,c (10)
∙ 158 ∙
where Cmin is heat-capacity flow rate of air, W °C-1; cp,h is average specific heat of constant
pressure of air, J kg-1 °C-1; Cmax is heat-capacity flow rate of the refrigerant, W °C-1; cp,c is average
specific heat of constant pressure of the refrigerant, J kg-1 °C-1.
The heat transfer of water vapor caused by temperature change and phase transition are
calculated using Eqs. (11) and (12), respectively.
q1  Cmin (Th,i  Th, o ) (11)
q2  mcond .h fg (12)
where q1 and q2 are heat transfer rate of water vapor in temperature change stage and phase
transition stage, respectively, W; Th,i and Th,o are the temperatures of air at the inlet and outlet of
heat exchanger, °C ; hfg is the enthalpy of water vapor transition, J kg-1.
The maximum possible heat transfer rate is calculated in Eq. (13).
qmax  Cmin (Th,i  Tc ,i ) (13)
where qmax is the maximum possible heat transfer rate, W; Cmin equals to the smaller of the air and
refrigerant heat flow rates, W °C-1; Th,i and Tc,i are the inlet temperature of air and refrigerant
respectively, °C.
The ε- NTU method is one of the methods used to verify heat exchanger (Incropera and Dewitt,
2011). The efficiency  is the ratio of the actual heat transfer rate to the maximum possible heat
transfer rate, which is calculated in Eq. (14).
  (q1  q2 ) / qmax (14)
When phase change occurs on the air side, NTU can be calculated in Eq. (15).
NTU   ln(1   ) (15)
U  Cmin NTU / A (16)
where NTU is a dimensionless parameter widely used in heat exchanger analysis; U is the total
heat transfer coefficient (HTC), W m-2 °C-1; A is the effective heat transfer area of fin-tube
dehumidifier, m2 (Song et al., 2017; Lee et al., 2020).
3. Results and analysis
3.1 Performance analysis
3.1.1. Dehumidification rate
From saturation vapor pressure table and air density table (Incropera and Dewitt, 2011), air
density is about 1.293 kg m-3, the specific heat capacity is 1007 J kg-1 °C-1, and the latent enthalpy
is 2.534 × 106 J kg-1 on 1 atm. The density of refrigerant is about 1.14 × 103 kg m-3, and the
specific heat capacity is 2300 J kg-1 °C-1 from the specification of this refrigerant.
The influence factors, i.e., TDBIO and fan air flow rate, were designed in 4 different
experimental cases. During the experiment, the air RH and water vapor mass decreased gradually,
then finally to a stable state. at the same time, the air temperature decreased by condensation
dehumidification. The DR was calculated using Eq. (1)–(3). The variation of temperature,
humidity, and water vapor mass of indoor air during the test were presented in 4 experimental
cases (Figure 3).
The air flow rate and TDBIO of case 2 were bigger than other cases (Table 1), so the averaged
and maximum DR were higher, meanwhile its temperature dropped quickly. On the contrary, the
condition of case 1 was detrimental to dehumidification. The averaged DR and the maximum DR
of the system increased significantly as TDBIO increased. The TDBIOs in the last three cases
(Table1) were about 30 °C, the higher the fan air flow rate was, the higher the averaged and
maximum DR of the system was.
∙ 159 ∙
3.1.2. Heat loss

According to Eq. (4), heat losses and heat loss rates of the four cases in Table 1 were calculated
(Table 2). In case 2 and case 3, the indoor initial temperature and TDBIO were almost the same,
and fan air flow was 2.5 m s-1 and 2.0 m s-1, respectively. The fan air flow affected the length of
dehumidification time and heat loss rate in these two cases. In case 1 and case 3, there were the
same length of dehumidification time, TDBIO influenced on dehumidification water mass and
heat loss rate.
140 1.2 140 1.2
Temperature Temperature
B
Temperature(oc) RH(%)
120 RH
120 A RH
Water vapor mass(kg)

1

1
Water vapor mass Water vapor mass
100 100
0.8 0.8
80 80
0.6 0.6
60 60
0.4 0.4
40 40
20 0.2 20 0.2
0 0 0 0
1 5 10 15 20 25 30 35 40 45 1 4 7 12 17 22 27 32 37
Time(min) Time(min)
140 1.2 140 1.2

Temperature Temperature
C D
120 RH 1 120 RH

1
Water vapor mass Water vapor mass
100 100
0.8 0.8
80 80
0.6 0.6
60 60
0.4 0.4
40 40
20 0.2 20 0.2
0 0 0 0
0 5 10 15 20 25 30 35 40 45 0 10 20 30 40 50 60
Time(min) Time(min)
Figure 3. The variation of temperature, humidity and water vapor mass. A: TDBIO was 20.2 °C,
air flow rate was 2.0 m s-1; B: TDBIO was 28.9 °C, air flow rate was 2.5 m s-1; C: TDBIO was
31.1 °C, air flow rate was 2.0 m s-1;D: TDBIO was 29.7 °C, air flow rate was 1.5 m s-1.
Table 1. DRs in different cases.
Mean TDBIO, Air flow rate, Averaged DR, kg Maximum DR,
Case
°C m s-1 h-1 kg h-1
1 20.2 2.0 0.01 0.81
2 28.9 2.5 0.90 3.63
3 31.1 2.0 0.78 2.25
4 29.7 1.5 0.58 2.22
Table 2. Heat loss in different dehumidification (DH) cases.
Indoor Indoor Mean DH Heat loss
Time, Heat
Case initial T, final T, TDBIO, water rate,
min loss, kJ
°C °C °C mass, kg kJ min-1
1 15.2 11.4 3.8 0.29 45 295.98 6.58
2 17.1 9.0 8.1 0.47 37 630.91 17.05
3 17.6 9.2 8.4 0.58 45 654.28 14.54
4 14.1 7.8 6.3 0.35 50 490.71 9.81
3.1.3. Total heat transfer coefficient
In the dehumidification system, both air temperature change and phase transition were
exothermic processes. The entropy change was negative. Given phase transition, the maximum
∙ 160 ∙
possible heat transfer rate can be calculated by the air heat flowrate. The heat transfer coefficients
were calculated by the Eqs. (14) – (16) using ε- NTU. The mean total HTCs of the last three
cases were 17.72 W m-2 °C -1, 14.79 W m-2 °C -1 and 10.20 W m-2 °C -1 when the averaged TDBIO
was about 30 °C and the air flow rate was 2.5 m s-1, 2.0 m s-1 and 1.5 m s-1 respectively, as shown
in Table 3.
Table 3. Total HTC in different air flow rates.
2.5 m s-1 2.0 m s-1 1.5 m s-1
Total Total Total
Inlet T, Inlet T, Inlet T,
HTC, HTC, HTC,
°C °C °C
W m-2 k-1 W m-2 k-1 W m-2 k-1
16.1 22.00 16.8 15.16 14.1 11
13.1 18.20 13.9 19.15 13.4 9.78
11.9 18.56 12.3 13.57 12.2 11
10.8 17.04 11.5 12.23 10.9 11
10.4 15.50 11 12.91 10.2 10.58
9.2 15.02 10.4 15.77 9.6 9
17.72 14.79 10.20
Based on mean total HTC of this dehumidification system in Table 3, when TDBIO is known,
the air flow rate can be adjusted and DR can be obtained by Eqs. (9) - (16). Given the averaged
TDBIO was maintained at 30 °C, DR could be calculated by air flow rate like this equation
mcond .  0.0102  v 2  0.0058  v  0.55 . DR increased along with fan air flow increases, which
can be used as the basis for the controller design.
3.2 Comparison theoretically with positive-pressure ventilation
In positive-pressure ventilation system, outdoor cold and dry air was sent into by ventilation
fan, while indoor warm air with high humid comes out under the same volume of air exchange.
In order to verify and compare the energy consumption of dehumidification system with positive-
pressure ventilation system when removing almost the same mass of air water vapor, a theoretical
temperature calculation of positive-pressure ventilation was employed as shown in Eq. (5). Given
the radius of fan for positive pressure and air flow rate were 0.2 m and 2.0 m s-1, respectively,
TDBIO were consistent with those of above humidification system. The temperature difference
was calculated by using Eq. (5), and energy losses were 599.76 kJ, 1518.86 kJ, 1846.00 kJ,
1433.18 kJ, respectively by Eq. (4). While energy losses by humidification system were 295.98
kJ, 630.91 kJ, 654.28 kJ, 490.71 kJ, respectively (Table 2). Therefore, the heat losses of
dehumidification system in these four cases were 49.35%, 41.54%, 35.44%, 34.24% of those with
positive-pressure ventilation calculated by Eq. (6). It is notable that, the time for dehumidification
by the positive-pressure ventilation was shorter than that by the new developed dehumidification
system, so there is a significant heat loss rate (Table 4) that could lead to a sharp temperature drop.
Table 4. Heat loss of positive-pressure ventilation.
Dehumidi- Heat loss
Heat loss,
Case Initial T, °C Final T, °C TDBIO, °C fication water Time, s rate,
kJ
mass, kg kJ min-1
1 15.2 7.5 7.7 0.23 100 599.76 359.86
2 17.1 -2.4 19.5 0.53 190 1518.86 479.64
3 17.6 -6.1 23.7 0.58 240 1846.00 461.50
4 14.1 -4.3 18.4 0.34 180 1433.18 477.73
4. Conclusions
(1) The new dehumidification system provided an innovative and effective method for
∙ 161 ∙
dehumidification of livestock and poultry houses in cold winter. It made full use of the natural
cooling refrigerant to reduce the indoor air humidity in cold region.
(2) When TDBIO was about 30 °C, the mean total HTCs calculated by ε- NTU were 17.72 W
m-2 °C-1, 14.79 W m-2 °C-1 and 10.2 W m-2 °C-1 at different air flow rate of 2.5 m s-1, 2.0 m s-1 and
1.5 m s-1, respectively. The mean total HTC could be used for DR calculation when TDBIO was
given.
(3) The heat losses of dehumidification system were 49.35%, 41.54%, 35.44%, 34.24% of
those of positive-pressure ventilation, which achieved a significant energy saving.
Acknowledgements
This work was supported by the project of National Natural Science Foundation of China
(NSFC) (32072787); the project of scholar plan at Northeast Agriculture University (19YJXG02),
China; the Earmarked Fund for China Agriculture Research System (CARS-35), China; the Key
Laboratory of Swine Facilities Engineering, Ministry of Agriculture, P.R. China. Also, the authors
acknowledge the support by the researchers and managers, and the help by all others for this study.
References
Abbaspour-Fard, M.H., A. Gholami, M. Khojastehpour, 2011. Evaluation of an earth-to-air heat
exchanger for the northeast of Iran with semi-arid climate. International Journal of Green Energy. 8, 499e510.
Allen, W.H., F.A. Payne, 1987. Designing animal ventilation schedules with counterflow heat
exchangers. Transactions of the ASAE. 30 (3), 782–788.
Badescu, V., 2007. Economic aspects of using ground thermal energy for passive house heating.
Renewable Energy. 32, 895e903.
Barber, E.M., H.L. Classen, P.A. Thacker, 1989. Energy use in the production and housing of poultry and
swine-an overview. Canadian Journal of Animal Science. 69 (1), 7–21.
Berckmans, D., 2014. Precision livestock farming technologies for welfare management in intensive
livestock systems. Revue Scientifique et Technique (International Office of Epizootics). 33 (1), 189–196.
Boris, R.E., 2008. Energy savings with heat pads and lighting. Advances in Pork Production: Banff Pork
Seminar.
Campen, J.B., G.P.A. Bot, 2002. Dehumidification in greenhouses by condensation on finned pipes.
Biosystems Engineering. 82 (2), 177–185.
Harmon, J.D., H.M. Hanna, D. Petersen, 2010. Farm Energy: Sizing minimum ventilation to save heating
energy in swine housing. Agriculture and Environment Extension Publications. 28.
Incropera, F.P., D.P. Dewitt, 2011. Fundamentals of heat and mass transfer. Staff General Research
Papers. 27 (1–2), 139–162.
Krommweh, M., P. Rosmann, W. Buscher, 2014. Investigation of heating and cooling potential of a
modular housing system for fattening pigs with integrated geothermal heat exchanger. Biosystems
Engineering. 121 (1), 118–129.
Lee, J.H., Y.W. Son, S.M. Chang, 2020. Numerical analysis on natural convection heat transfer in a single
circular fin-tube heat exchanger (Part 2): correlations for limiting cases. Entropy. 22 (3), 358.
Licharz, H., P. Rsmann, M. Krommweh, E. Mostafa, W. Buscher, 2020. Energy efficiency of a heat pump
system: case study in two pig houses. Energies. 13 (3), 662.
Rösmann, P., W. Büscher, 2010. Rating of an air-to-air heat exchanger in practice. Landtechnik. 65 (6),
418–420.
Song, K.W., Z.P. Xi, M. Su, L.C. Wang, L.B. Wang, 2017. Effect of geometric size of curved delta
winglet vortex generators and tube pitch on heat transfer characteristics of fin-tube heat exchanger.
Experimental Thermal & Fluid Science. 82, 8–18.
Spengler, R.W., D.P. Stombaugh, 1983. Optimization of earth-tube heat exchangers for winter ventilation
of swine housing. Transactions of the ASAE. 26 (4), 1186–1193.
Stinn, J., H. Xin, 2014. Heat lamp vs. heat mat as localized heat source in swine farrowing crate. Animal
Industry Report.
Tikhomirov, D.A., S.S. Trunov, A.V. Kuzmichev, S.A. Rastimeshin, O.V. Shepovalova, 2020. Energy-
efficient thermoelectric unit for microclimate control on cattlebreeding premises. Energy Reports. 6, 293–305.
∙ 162 ∙
Calibration of Discrete Element Method Parameters for Juvenile

Manila Clam (Ruditapes philippinarum) and Seeding Verification
Hangqi Li a, Xiuchen Li a,b, Hanbing Zhang a,b, Guochen Zhang a,b, Qian Zhang a,b,
Wenbo Liu a, Gang Mu a,b,*
a
Dalian Ocean University, Dalian, Liaoning 116023, China
b
R&D Center of Fisheries Equipment and Engineering of Liaoning Province,
Dalian, Liaoning 116023, China
* Corresponding author. Email: mugang@dlou.edu.cn
Abstract
The manila clam is an important economic shellfish in China. To promote the design and
optimization of juvenile manila clam seeding equipment. The juvenile clam discrete element
model was established based on 3D scanning and EDEM software. The contact parameters of
clam were calibrated by combining the result of measurement and simulation test (slope sliding
and dropping) of clam contacting with acrylic and stainless steel. The clam seeding discrete
element method (DEM) simulation and realistic verification test in different seeding wheel’s
rotational speeds were carried out with calibrated clam. The results showed that: (a) the static
friction coefficient of clam-acrylic and clam-stainless steel were 0.31 and 0.23, respectively. (b)
the restitution coefficient of clam-clam, clam-acrylic, and clam-stainless steel were 0.32, 0.48,
0.32, respectively. The results of the angle of repose from the response surface test showed when
the contact wall was acrylic, the coefficient rolling friction and static friction of clam-clam were
0.17, 1.12, respectively, the coefficient rolling friction of clam-acrylic was 0.20. When the contact
wall was stainless steel, the coefficient rolling friction and static friction of clam-clam were 0.33,
1.25, respectively. The coefficient rolling friction of clam-stainless steel was 0.20. The results of
the verification test showed that the average error between the realistic and simulation value was
less than 5.0%, the clam DEM model could be applied in clam seeding equipment simulation and
development.
Keywords: Shellfish, mechanical seeding, DEM model, simulation
1. Introduction
Manila clam (Ruditapes philippinarum) is one of the most important farmed shellfish in China,
which production of China has exceeded 4 × 106 t in 2019. The quality of clam seeding is an
important factor affecting the harvest yield. At present, artificial seeding is the main clam seeding
method which has the problem with uneven seeding and serious damage (Yuan et al., 2014), so
mechanized seeding is imperative. It is difficult to analyze the complex dynamic behavior of
juvenile manila clams (clam) in the mechanized seeding process accurately by traditional
measurement methods. The discrete element method (DEM) can better reflect the motion state of
particles in complex motion and has been widely used in the design and research of agricultural
equipment (Zeng et al., 2021). The simulation contact parameters are the key factors that affect
the accuracy of the simulation results. Because of the geometry difference between clam and its
DEM model, there are some contact parameters errors. Therefore, the simulation contact
parameters need to be calibrated.
There are mainly two ways: direct measuring approach and DEM simulation calibration
approach used to obtain simulation contact parameters at present (Coetzee, 2017). The DEM
simulation calibration approach is to make use of a procedure where the experiment is performed
to measure a specific material property. The experiment is then simulated by following the
experimental procedures as closely as possible. The DEM parameter values are then changed
iteratively until the predicted response matches the measured result. DEM simulation calibration
approach with the advantage of a smaller error and higher accuracy has been widely used in
agricultural engineering and agricultural granular materials calibrating. The calibration objects
∙ 163 ∙
include dried cherry fruit (Santos et al.), wheat (Liu et al., 2016), soybean (Ghodki et al., 2018),
and other granular. However, there are few studies on simulation contact parameters calibration
of heterogeneous granular material such as shellfish. To provide reference to the design and
improvement of clam seeding equipment by DEM, the clam was taken as the research object in
this paper. The clam DEM model was established and the simulation contact parameters between
clam and different materials were calibrated.
2.1. Materials
In this paper, the test object was clam (shell length: 15–20mm) used for seeding in the
shellfish farm sea of Dandong City, Liaoning Province, China. According to the design
requirements, 5 mm thick acrylic and stainless steel plate were selected as the contact wall
materials (Table 1).
Table 1. The initial set of material parameters for AC (acrylic), SS (stainless steel), and clam.
Material parameters Symbol Value
AC Poisson’s ratio vac 0.36
AC shear modulus (Pa) Gac 2.3 × 107
AC density (g mm-3) ρac 1.06
SS Poisson’s ratio vss 0.3
SS shear modulus (Pa) Gss 7 × 1010
SS density (g mm-3) ρss 7.80
Clam Poisson’s ratio vc 0.33 (Mu, 2019)
Clam shear modulus (Pa) Gc 5.56 × 107
Clam density (g mm-3) ρc 1.58
2.2. Clam DEM model
Because the shape of the clam was oval and irregular (Figure 1). The clam contour model
(Figure 1) was established by reverse engineering with the three-dimensional scanner (Einscan-
S, Shining3D, CHN) and Geomagic Studio 2016 (Geomagic Studio 2016, Geomagic, US). Then
the clam DEM contour model (Figure 1) referring to the clam contour model was filled with
different size particles based on the multi-sphere method (MSM) (Santos et al.) and Hertz–
Mindlin no-slip contact model in EDEM software (EDEM 2018, DEM Solutions Ltd, UK). The
simulation contact parameters were calibrated by the DEM simulation calibration approach.
Length
Height
Figure 1. Clam (left), Clam contour model (middle), and Clam DEM contour model (right).
2.3. Measurement
2.3.1. The coefficient of static friction (μs)
The particle–wall coefficient of static friction (μs-pw) was determined by following the inclined
plate method. When the clam began to slide down, the value of inclination angle (β) was recorded
by digital inclinometer (DMI 410, Rion, CHN) and calculated the value of μs-pw (Figure 2(a)). The
value of the particle-particle coefficient of static friction (μs-pw) was determined in Section 2.4.2.
2.3.2. The coefficient of restitution (e)
The determination of the coefficient of restitution (e) was based on drop tests (González-
Montellano et al., 2012). Stainless steel, acrylic, and clam (clams had been adhered to stainless
steel plate closely) were used as the bottom plate, and the scale paper (5 × 5mm) was selected as
∙ 164 ∙
the background. The clam fell freely from the height (H0) of 150mm and rebounded to a maximum
height (H1). The video of the process of drop tests was recorded with a high-speed camera
(Fastcamsa5, Photron, JPN) running at 1024 × 1024 pixels@1000 fps (Figure 3(a)). The video
was converted into images by Matlab to measure H1. The value of e was calculated by Eq. (1).
2 gH1 H1
e  (1)
2 gH 0 H0
2.3.3. The static angle of repose (θ)

The angle of repose of clam was measured by lifting cylinder test. The cylinder was uplifted
at a speed of 0.1m s-1 when the clams were poured into a hollow cylinder (R=50mm, H=350mm)
which was placed on the bottom plate. When clams were deposited by gravity in a stable state,
the angle of repose of four directions: X+, X−, Y+, Y− were measured with a digital inclinometer
(Figure 4(a)).
2.4. DEM simulation calibration
2.4.1. The determined contact parameters calibration
The DEM simulation calibration test of the simulation contact parameters single-factor
experiment that has been repeated in EDEM (Figure 2(b), 3(b), 4(b)) to make the DEM simulation
calibration test the same as realistic test. The range of simulation contact parameters was predicted
and divided into 5 groups on average (Grima and Wypych, 2011). The DEM simulation test data
were fitted by curve fitting, and the quadratic polynomial fitting curve and equation were obtained.
The accuracy of the quadratic polynomial fitting equation was judged. Then, the directly measured
corresponding parameters (β and H1) were brought into the quadratic polynomial fitting equation
as the target values to calculate the simulation contact parameters. The calibrated simulation
contact parameters were re-entered into the EDEM software to repeat the DEM simulation test
for obtaining the simulation corresponding parameters. The relative error (δ) between the
corresponding parameters results of the DEM simulation test and direct measurement was
calculated to verify the accuracy of the simulation contact parameters.
2.4.2. Response surface simulation test
The other clam contact parameters such as μr-pw, μs-pp, μr-pp were difficult to get by direct
measurement and DEM simulation calibration, so the response surface simulation test was used
to get them. The DEM calibrated simulation contact parameters (eg: μ’s-pw, e’pp, e’pw) were input
into EDEM software for a pre-test of clam stacking simulation test, and the value range of the
other uncalibrated simulation parameters (μ’r-pp-ss, μ’s-pp-ss, μ’r-pw-ss, μ’r-pp-ac, μ’s-pp-ac, μ’r-pw-ac) was
predicted in a range. The response surface test with 3 factors and 3 levels was designed (Santos
et al., 2014), and the significance was set at 3 levels of high (+1), medium (0), and low (−1).
Moreover, taking the measured repose angle of clams as the target value, the simulation repose
angle of clam-stainless steel (θ’ss) and clam-acrylic (θ’ac) were evaluated as the response index in
the response surface simulation test. In addition, the angle of repose error (δa) between the
simulation and the measurement was used as the evaluation index. Based on the analysis of
variance (ANOVA) of quadratic polynomial model of response surface test, the multiple
regression equation without the non-significant factors was obtained, and the optimal
combinations of uncalibrated simulation contact parameters under different contact parameters
were calculated.
2.5. Clam seeding verification test
The accuracy of calibrated clam DEM model was judged in the clam seeding verification test
by comparing the average error of the coefficient of variation between the realistic and simulation
seeding tests. Taking the rotation speed of the seeding wheel as the variable, the clam seeding test
of 5 groups was designed. The weight of clam discharged from 5 continuous silos of seeding tray
was measured in each test. The clam seeding equipment was mainly composed of a blanking
∙ 165 ∙
hopper, the seeding tray and seeding wheel were designed and trial produced in realistic clam
seeding test. The seeding tray and seeding wheel were manufactured through 3D printing, and the
inner wall of the seeding wheel adhered to acrylic plates (Figure 5(a)). The 3D clam seeding
equipment, clam DEM model, and acrylic were imported into the EDEM software for the clam
seeding simulation test (Figure 5(b)). The coefficient of variation was calculated according to Eq.
(2)−(4). The test design of realistic seeding was the same as the simulation seeding test.
1 n
S  (M i  M 0 )2
n i 1
(2)
1 n
M0   Mi
n i 1
(3)
S (4)
CV   100%
M0
where CV was the coefficient of variation in circumferential distribution, %; S was the standard
deviation, g; n was the grid number; M0 was the average feed quality was collected by the grid in
the collection domain, g; Mi was the weight of feed particles was collected in i computing grid, g.
1 Wall 2 Clam 3 Inclinometer 4 Lifting platform

Figure 2. Calibration experiment of static friction coefficient: (a) direct measurement (b) DEM
simulation measurement.
(a) (b)
Before impact Impact After impact Before impact Impact After impact
1 Wall 2 Clam 3 Coordinate paper 4 Testbed
Figure 3. Calibration experiment of restitution coefficient: (a) direct measurement (b) DEM
simulation measurement.
The data of response surface test results were calculated by Design-Expert software (Design
expert 10, Stat-Ease, US), and the significant impact level was set at 95% (P<0.05), and the
extremely significant impact level was set at 99% (P<0.01).
Figure 4. Clam stacking test: (a) direct measurement (b) DEM simulation measurement.
∙ 166 ∙

3.1. The coefficient of static friction (μs)
The results of β and μs-pw between clam and the stainless steel as well as acrylic were measured
directly (Table 2). Based on the direct measurement results, the particle–wall simulation
coefficient of static friction (μ’s-pw) was predicted to be in the range of 0.20–0.40. The simulated
inclination angle (β’) was measured by the protractor function in EDEM software. The simulation
results of β' were substituted into the quadratic polynomial fitting regression equation ys-ss, ys-ac
(Figure 6(a)) to calculate the μ’s-pw (Table 2).
(a) (b)
1 Clam 2 Blanking hopper 3 Seeding tray

4 Seeding wheel 5 Motor 6 Conveyor
Figure 5. Clam seeding verification test: (a) direct measurement (b) DEM simulation
measurement.
Table 2. Particle–wall coefficient of static friction (μs-pw) for SS and AC.
Symbol Value (°)
Approach Parameters
SS AC SS AC
Direct Inclination angle β βss βac 14.40 18.78
measurement Coefficient of static friction μs μs-pw-ss μs-pw-ac 0.26 0.34
DEM Inclination angle β’ β’ss β’ac 14.97 19.12
simulation Coefficient of static friction μ’s μ’s-pw-ss μ’s-pw-ac 0.22 0.31
The μ’s-pw-ss, μ’s-pw-ac were substituted into the static friction coefficient simulation verification
test, and the value of β’ss, β’ac was measured by the protractor function in EDEM, respectively.
The value of inclination angle relative error (δa) between β’ and β was: δa-ss = 4.0%, δa-ss = 1.8%,
respectively. The simulation inclination angle was consistent with the direct measurements.
The value of μ’s-pw was significantly smaller than μs-pw (Table 2). The reason may be that the
clam DEM model was composed of smooth particles, but the realistic clam had a growth line on
its surface, with the increase of clam size, the surface lines of clam gradually deepened, so the
surface roughness and the static friction coefficient also gradually increased. This agreed with the
results reported (Liu et al., 2020).
3.2. The coefficient of restitution (e)
The results of H1 and e obtained for the clam with the stainless steel, clam, and acrylic wall
material were measured directly (Table 3). Based on the measured results, the simulation
coefficient of restitution (e’) was predicted to be in the range of 0.25–0.55. The simulation
rebound height (H’1) was measured by the ruler function of the EDEM. The regression equations
yh-ss, yh-ac, and yh-pp (Figure 6(b)) were obtained by quadratic polynomial fitting of results of H’1.
The results of H1 were substituted into the equation yH-ss, yH-ac, yH-pp to calculate the value of e’pw-
ss, e’pw-ac, e’pp (Table 3).
The e’pw-ss, e’pp, e’pw-ac were substituted into the restitution coefficient simulation verification
test, and the value of H’1-pw-ss, H’1-pp, H’1-pw-ac was measured, respectively. The rebound height
relative error (δH1) between H’1 and H1 were: 1.7%, 1.7%，2.1%, respectively. DEM simulation
was close to the direct measurement, which could effectively replace the realistic drop test.
∙ 167 ∙
The simulation restitution coefficient was quite different from direct measurements. The
relative error (δpw) between e’pw-ac and epw-ac was 17.1%. The reason may be that the shape of
clams was irregular and the center of gravity was also different, so when each clam impacted the
bottom plate, the impacting position of clam and height was different. However, the discrete
element model of clam in the drop DEM simulation was filled with solid spherical particles, the
center of gravity position was relatively stable, and the impact position was relatively fixed, so
the value of δac was big.
Table 3. The coefficient of restitution (e) for SS, AC, and clam.
Symbol Value (mm)
Approach Parameters
SS Clam AC SS Clam AC
Direct Rebound height H1 H1-pw-ss H1-pp H1-pw-ac 13.18 6.79 25.20
measurement coefficient of restitution e μs-pw-ss epp epw-ac 0.30 0.21 0.41
DEM Rebound height H’1 H’1-pw-ss H’1-pp H’1-pw-ac 12.96 6.93 25.63
simulation coefficient of restitution e e’pw-ss e’pp e’pw-ac 0.38 0.32 0.48
(a) (b)
Figure 6. The fitting curve of relation (a) static friction coefficient and inclination angle (b)
coefficient and rebound height in DEM simulation calibration.
3.3. Response surface simulation test
The results of directly measured repose angle of clam-stainless steel (θss) and clam-acrylic
(θac) were θss = 31.75°, θac = 38.07° (Table 4).
When the bottom plate material was stainless steel, the simulation contact parameters: μ'r-pp,
μ's-pp, μ'r-pw, μ'r-ppμ's-pp, μ'r-ppμ'r-pw, μ's-pp2 and μ'r-pp 2 had high significant influence, whereas μ's-ppμ'r-
pw and μ'r-pw had insignificant influence (Table 5). When the bottom plate material was acrylic,
2
the simulation contact parameters μ'r-pp, μ'r-pw2 had high significant influence and μ's-pp, μ's-ppμ'r-pw,
μ's-pp2 had significant influence, whereas μ'r-ppμ'r-pw and μ'r-pw2 had insignificant influence. In both
stainless steel and acrylic models, the parameters (the P value of lack of fit, the coefficient of
variation (CV), determination coefficient (Rs2), correction determination coefficient (Adj−Rs2), the
Adeq−Precision indicated the accuracy of the fitted equation was high, and the simulation repose
angle regression equations were obtained (Ep. (5), (6)).
ss  41  1.67R  PP  2.05S  PP  0.94R  PW  1.94R  PP S  PP  1.46R  PP R  PW  1.76R  PP 2  2.52S  PP 2 (5)
ac  31.86  2.95R  PP  0.66S  PP  0.86S  PP R  PW  0.91S  PP 2  1.21R  PW 2 (6)
The results of some simulation contact parameters were obtained: μ’r-pp-ss = 0.33, μ’r-pp-ac =
0.20, μ’s-pp-ss = 1.25, μ’s-pp-ac = 1.12, μ’r-pw-ss = 0.34, μ’r-pw-ac = 0.17. The simulation repose angles
of clams: θ’ss = 31.55°, θ’ac = 37.90°, and the repose angles relative error: δa-ss = 0.04%, δa-ac =
0.06%. Because there was no obvious difference between the simulation and the direct
∙ 168 ∙
measurement results, the accuracy of clam simulation contact parameters was high. Therefore, the
clam discrete element model could be used for EDEM simulation for clam seedling.
Table 4. Test design and results of repose angle response surface.
Number μ'r-pp μ's-pp μ'r-pw 𝜃’ss (°) 𝜃’ac (°)
1 -1 -1 0 38.25 29.48
2 1 -1 0 37.22 33.63
3 -1 1 0 31.08 25.95
4 1 1 0 37.79 32.76
5 -1 0 1 38.71 29.40
6 1 0 1 39.64 35.99
7 -1 0 1 33.90 29.42
8 1 0 1 40.68 35.49
9 0 -1 1 41.00 33.81
10 0 1 1 35.84 31.65
11 0 -1 1 38.83 30.97
12 0 1 1 34.23 32.24
13 0 0 0 40.49 31.77
14 0 0 0 40.00 31.25
15 0 0 0 40.89 32.37
16 0 0 0 40.25 31.54
17 0 0 0 40.17 32.38
Table 5. ANOVA of the quadratic polynomial model of response surface test.
SS AC
Source
Sum of squares df P value Sum of squares df P value
Model 129.92 9 <0.0001** 89.13 9 <0.0003**
μ'r-pp 22.41 1 <0.0001** 69.74 1 <0.0001**
μ's-pp 33.46 1 <0.0001** 3.50 1 0.0297*
μ'r-pw 7.13 1 0.0022** 0.93 1 0.2030
μ'r-ppμ's-pp 14.98 1 0.0002** 1.77 1 0.0942
μ'r-ppμ'r-pw 8.56 1 0.0013** 0.07 1 0.7164
μ's-ppμ'r-pw 0.08 1 0.6357 2.94 1 0.0413*
μ'r-pp2 13.02 1 0.0004** 1.05 1 0.1793
μ's-pp2 26.66 1 <0.0001** 3.47 1 0.0302*
μ'r-pw2 0.57 1 0.2228 6.19 1 0.0085**
Residual 2.24 7 3.31 7
Lack of Fit 1.76 3 0.0784 2.29 3 0.1569
Pure Error 0.48 4 1.01 4
Cor Total 132.16 16 92.43 16
CV 1.47% 2.16%
Rs2 0.9642 0.9831
Adj−Rs2 0.9183 0.9613
Adeq−Precision 23.674 18.636
The repose angle measured in the stacking test was θss < θac by comparing the test result of
acrylic and stainless steel. The main reason may be that the larger surface roughness of acrylic
than that of stainless steel, so the larger static friction coefficient (μs-pw-ac) led to the larger repose
angle on the acrylic bottom plate. The μ’r-pp and μ’s-pp had significant effects on the repose angle
of clams (Table 4). It could be inferred that the clam contact parameters: μs-pw, μr-pp, μs-pp had a
significant effect on the application of the clam DEM model (Liu et al., 2016).
∙ 169 ∙
3.4. Clam seeding verification test

The results of the clam seedling verification test (Table 7) showed the average value of δt
between the verification test and the realistic seeding test by calculating was 4.98 % when the
feeding speed of the dropping port of the seeding equipment was constant and the rotation of the
seeding wheel was different. The realistic test and simulation test were the same and the accuracy
of the clam DEM model was high. Therefore, the clam seeding simulation could effectively
simulate the realistic situation of clam seeding. The optimal working range of the rotation speed
selected of the clam seeding equipment was 23−38 r min-1 (Table 5).
Table 7. Clam seeding verification test results.
Seeding wheel’s speed (r min-1)
Variation coefficient
15 23 30 38 45
Realistic value (%) 17.96 11.28 7.16 12.35 20.56
Simulation value (%) 9.79 7.77 5.56 9.84 10.91
Error δt (%) 8.17 3.51 1.60 2.51 9.65
In this study, the clam DEM model was established in EDEM software. The simulation
contact parameters of clam between different contact materials: stainless steel and acrylic were
determined by the DEM simulation calibration approach. In addition, the simulation contact
parameters (μ’s-pp, μ’r-pp, μ’s-pw) had significant effects on the simulation repose angle of clam. The
accuracy of the discrete element model of clam was high and was verified in the clam seeding
verification test. Therefore, the calibrated clam discrete element model could be directly used in
the simulation of clam seeding in EDEM software in the future and could effectively understand
the motion state of clams and provide technical reference to the design and improvement in clam
seeding devices.
References
Coetzee, C.J., 2017. Review: Calibration of the discrete element method. Powder Technology. 310, 104-
142. http://doi.org/10.1016/j.powtec.2017.01.015.
Ghodki, B.M., M. Patel, R. Namdeo, G. Carpenter, 2018. Calibration of discrete element model
parameters: soybeans. Computational Particle Mechanics. 6 (1), 3–10.
González-Montellano, C., J.M. Fuentes, E. Ayuga-Téllez, F. Ayuga, 2012. Determination of the
mechanical properties of maize grains and olives required for use in DEM simulations. Journal of Food
Engineering. 111 (4), 553–562. http://doi.org/10.1016/j.jfoodeng.2012.03.017.
Grima, A.P., P.W. Wypych, 2011. Investigation into calibration of discrete element model parameters for
scale-up and validation of particle–structure interactions under impact conditions. Powder Technology. 212
(1), 198–209. http://doi.org/10.1016/j.powtec.2011.05.017.
Liu, F., J. Zhang, B. Li, JunChen, 2016. Calibration of parameters of wheat required in discrete element
method simulation based on repose angle of particle heap. Transactions of the Chinese Society of Agricultural
Engineering. http://doi.org/10.11975/j.issn.1002-6819.2016.12.035.
Liu, W., H. Zhang, X. Li, G. Zhang, Q. Zhang, S. Qu, J. Gao, B. Fan, G. Mu, 2020. Analysis of
biomechanical properties of juvenile Manila clam for mechanization sowing. Journal Of Dalian Ocean
University. 35 (03), 455–461.
Mu, G., 2019. Study on Key Technology and Design of Chain-tine Dredger for Ruditapes Philippinarum.
Doctor, Dalian University of Technology.
Santos, K.G., A.V.P. Campos, O.S. Oliveira, L.V. Ferreira, M.C. Francisquetti, M.A.S. Barrozo, 2014.
DEM simulations of dynamic angle of repose of acerola residue: a parametric study using a response surface
technique.
Yuan, H., H. Shen, L. Wang, L. Shi, 2014. Ecological Culture Technology of Ruditapes Philippines
Beach. Fisheries Science and Technology Information. 41 (01), 54–56.
Zeng, Z., X. Ma, X. Cao, Z. Li, X. Wang, 2021. Critical Review of Applications of Discrete Element
Method in Agricultural Engineering. Transactions of the Chinese Society for Agricultural Machinery.
∙ 170 ∙
Development of a Test Strip for Detecting Sows’ Oestrus Based on

Urine Hormone
Kaidong Lei a, Chao Zong a, Yanbin Wen a, Guanghui Teng a,*, Hao Wang b
a
b
Chongqing Academy of Animals Sciences, Rongchang, Chongqing 402460, China
* Corresponding author. Email: futong@cau.edu.cn
Abstract
To overcome the limitation of the time and labor consuming in traditional oestrus testing
methods, a test strip method based on urine pheromone estradiol (E2) and luteinizing hormone
(LH) hormone was developed to detect the oestrus state of post-weaning sows. The results showed
that: 1) The accuracy of the paper strip assisted testing method based on urine pheromone was
87.50%; 2) After mating, 7 out of 8 sows were successfully pregnant and transferred to gestation
period. The rate of successful mating reached to 85.71%. The method of the test strip for detecting
sows’ oestrus based on urine hormone was proved useful, and it can be applied to detect the
oestrus state of sows in large scale in future studies.
Keywords: Swine, PLF, test paper, test method, welfare, weaning
1. Introduction
The precision livestock farming (PLF) has been developed in recent years (Berckmans, 2008),
and a large number of studies have been carried out and are tend to be non-contact, stress-free,
and animal welfare improved (Turner et al., 2018; Pietrosemoli and Tang, 2020). The number of
weaned piglets per sow per year (PSY) is the main indicator to measure the production efficiency
of sows. Accurate detection of sows’ oestrus is desired and is useful for making production plans,
and helpful in avoid wasting non-productive days in the pig farm.
Traditionally, manual observation methods and boar test methods have been widely used in
oestrus detection. However, the traditional methods are not only time-consuming and labor-
intensive, but also dependent on the staff’s relevant experience in oestrus testing. This can largely
affect the accuracy of oestrus detection. The experienced staff always use some phenomenon to
help them on oestrus detection, for example, during the period of oestrus in sows, there are
decrease in the feed intake but increases in the sensitivity to environmental changes and the
activities. Besides, the changes in the vulva and the reaction of the sow to a back pressure test
have been always indications of a sow in oestrus state.
At present, many research have shown the automatic identification of the process of sow
oestrus. 1) Through the relationship between the temperature change of the vulva and buttocks of
sows before and after oestrus. Sykes et al. (2012) found that the maximum and average
temperatures of the vulva during the oestrus period of sows were higher than those during the
oestrus cycle period, but there was no difference between the minimum values. Scolari et al.
(2011) explored the changes in the temperatures of the vulvas and buttocks of sows before and
after oestrus and found that the temperature of each sow’s vulva increased significantly at the
beginning of oestrus and decreased significantly before ovulation, while the temperature of the
buttock surface did not change significantly. 2) The accelerometer was used to detect oestrus.
Altmann (1941) detected the activity level of the sow. Bressers et al. (1993) set the activity
threshold to detect the oestrus of sows. 3) Infrared sensors were used to detect oestrus, Freson et
al. (1998) used infrared sensors to detect oestrus with 86% accuracy. 4) Radio frequency sensors
were used to identify and record sows’ oestrus. Houwers (1988) used sensors to automatically
record the frequency of sows visiting boars, and the results showed that when the sows had never
been in oestrus and entered the oestrus period, the frequency of boar visits gradually increased.
Korthals (1999) counted the average length of sow visits. The statistical results showed that the
length of sow visits obeyed the 92 s/d Poisson distribution. The sensitivity of heat detection was
∙ 171 ∙
76.4%.
The above studies have achieved good results in physiological monitoring through image
technology and sensor technology, and it also makes intelligent oestrus detection a hot spot.
However, a large amount of research based on advanced image sensor technology requires
expensive electronic equipment and a lot of training professionals. Therefore, a quick test strip
method has been proposed to assist in detecting the sows’ oestrus state.
In this study, urine hormone indicator test papers were developed and carried out on post-
weaning sows. The method was on the basis of the cyclic variation of the urinary pheromone
estradiol (E2) and luteinising hormone (LH) and the color changes of the test papers, which
assisted to detect the oestrus state in sows.
2.1. Animals
The experiment was carried out in Chongqing Rongchang National Core Pig Farm from
October 15, 2020 to January 15, 2021. The sows were Yorkshire. In this experiment, eight sows
with similar body conditions, 200 ± 4.5 kg weight, were used.
2.2. Experimental method
During the experiment, the non-destructive sampling method was used to continuously track
and collect the urine samples from individual sows. At the same time, the results of manual
inspection of sows and the number of litters were recorded.
Generally, the collected urine needed be immediately frozen below 0 °C to avoid causing
instability in hormone measurement values. In the pen, 50 mL of morning urine was collected
continuously from the first day of weaning until the end of mating period. The collected urine was
marked with date and the ear number of the sow and stored at -20 °C. For the acquisition of
hormone data, a radioimmunoassay kit (Tianjin Depu Biotechnology and Medical Products Co.,
Ltd.) was used to determine the content of estradiol and luteinizing hormones in urine. Origin and
MATLAB software were used to fit the hormone content value and the curve of weaning days, it
concluded that the hormone content in the urine of sows changed regularly during the oestrus
cycle, and a quick test paper was designed for detection. In the test paper design, the test paper is
coated with antigen and antibodies. If the value of the detected substance was higher than the
threshold value, a color reaction would occur.
2.3. Analysis of data
The data standardization (normalization) processing has been introduced to intuitively
compare the two hormones under the same dimension. Different evaluation indicators have
different dimensions, and it will affect the results of data analysis. To reduce the unit influence
between each indicator, the data of estradiol hormone and luteinizing hormone were processed
for data standardization. After that, the two indicators were at the same magnitude and were
compared and analyzed under the same dimension. The MATLAB software and the Z-score
standardization method were used for conversion with Eq. (1).
X* = (x − µ) / σ (2)
where X* is the normalized value; x is each value in the original data; μ is the mean value of the
hormone sample data; σ is the standard deviation of the hormone sample data.
The standard deviation was standardized in Eq. (1). The mean value of the processed data was
0, and the standard deviation was 1. To analyze the changing trend of estradiol and luteinizing
hormones better during the oestrus cycle, the values of the two hormones were normalized by
MATLAB.
∙ 172 ∙

3.1. Estradiol hormone changes after weaning
In this paper, the results showed that the increase of estradiol hormone is related to oestrus.
Eight weaned sows were randomly selected on site for testing, as shown in Figure 1. the estradiol
hormones rose and fell regularly from the first day after weaning. The conducted daily manual
checkups (including manual back pressure, vulvar observation, etc.) showed that the estradiol
hormone content reached a peak when sows determined to be in oestrus by manual back pressure
test.
600
Sow-1 Sow-2 Sow-3 Sow-4
Estradiol hormone, pg mL-1
500
400
300
200
100
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
Day
Figure 1. Changes of estradiol hormone in sows after weaning with weaning days. Sow-1 to
Sow-8 are sow numbers that were tracked in the study; Sow-6 returned to oestrus; Sow-7 had
poor body condition after weaning.
3.2. Luteinizing hormone changes after weaning
By fitting the curve of luteinizing hormone and weaning time, the results of luteinizing
hormone changes after weaning is shown in Figure 2. It has shown that the luteinizing hormone
data fluctuated greatly on the first day after weaning. The sow-6 had a small peak on the sixth
day, and it was found this sow was in oestrus by manual back pressure; the sow-2 had a small
peak on the 18th day, and the sow was also found in oestrus by manual back pressure on that day.
40
Luteinizing hormone, miu mL-1
35
30
25
20
15
10
5
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
Day
Figure 2. Changes of Luteinizing Hormone-Releasing hormone in sows after weaning with

weaning days.
3.3. Comparison of luteinizing hormone and estrogen after normalization
As shown in Figure 3, due to the large individual differences, although the sows were weaned
∙ 173 ∙
at the same time, there were still huge fluctuations among the sows. It was found that the
difference was related to the body cycle between individual sows and postpartum recovery.
However, during the period from weaning to oestrus, the levels of luteinizing hormone and
estradiol hormone were higher than those of non-oestrus hormones.
0.35
0.30 E2 LH
0.25
Overall desirability
0.20
0.15
0.10
0.05
0.00
Sow-1 Sow-2 Sow-3 Sow-4 Sow-5 Sow-6 Sow-7 Sow-8
Sow
Figure 3. Normalized estradiol and luteinizing hormone.
3.4. The identification of oestrus in sows using hormone changes
To realize identification of oestrus detection in sows using sows’ urine, a rapid urine test paper
based on hormone response was designed. The responses of the estradiol and luteinizing hormone
to the antigen and antibody on the test paper, each of the two hormone responses accounted for
50%, were used for the identification.
3.5. Discussion
The sow's urine data can be used for detecting the sow's oestrus state. Some scholars have
studied that sows were usually in oestrus in 1–5 days after weaning, and they mated (Ostersen et
al., 2010). This paper aimed at the continuous tracking of weaned sows for 1–7 days, and the
results showed that hormone test paper was used to auxiliary detection of sows’ oestrus status
during the oestrus cycle. To verify the actual effect of the test paper on the state of the sow's
oestrus detection, the sow's condition check and back pressure test were carried out on a daily
base, and the determination of oestrus status is to observe the sow's response to back pressure
(BPT) (Turner et al., 2018). To explain the hormone curve changes better, the hormone changes
were placed on the same dimension for comparison. Within 1–7 days after weaning, if the average
content of estrogen was lower than 280.53 pg ml-1, and the content of luteinizing hormone was
lower than 15.17 miu ml-1, the sow was not in oestrus.
Compared with the results of manual oestrus checking, the accuracy rate of test strip method
was 87.5%. The results have shown that one sow returned to oestrus and the remaining seven
sows entered the gestation period. The mating success rate reached 85.71%.
With the aid of test strip proposed in this study, the staff used the hormone test paper to assist
in judging the oestrus status. The staff could finish the work without much experience and
professional theoretical knowledge. During the period from weaning to oestrus, the urine hormone
analysis technology could be used to obtain the physiological and behavioral changes of sows. In
this study, by analyzing and comparing the response of the two hormones to the sow’s behavior
during the estrous cycle, it was found that the detection accuracy was higher than that of Freson
et al. (1998), who used infrared sensors to detect oestrus of sows, and the cost of test paper was
much less than sensors. The test paper method in this study obtained the same results compared
with pressure back test method. It was found that poor body condition or overly obese body
condition could directly affect the oestrus of sows.
According to the verification data on the pig farm, it was feasible to assist the judgment of
∙ 174 ∙
oestrus on the urine hormone of the sow. Compared with other studies adopting RFID and other
sensors to monitor the visit frequency of sows to boars to detect oestrus (Bressers et al., 1991),
the method proposed in this study was faster and more economical and it did not need much
technical training for staff. Compared with the method based on the real boars to detect the oestrus
state, the proposed method had advantages in time-saving and biological safety. Moreover, the
most important thing was that the urine test paper was recyclable and sustainable.
The results above proved that an assisting method based on urine hormone test paper to detect
the oestrus state of sows was useful. However, due to the limited sample data, this method still
needs more tests before large scale applications. In the future, it is necessary to obtain a large
number of hormones from sows and increase the variety of sows to obtain better results, providing
technical support to staff.
4. Conclusions
This study proposed a method for detecting oestrus of sows based on the urine hormone of
sows and drew the following conclusions:
1) Based on the estradiol (E2) and luteinizing hormone (LH) in the urine of sows, the oestrus
state was detected. Twenty non-pregnant sows were examined and judged successively, and the
accuracy rate reached 85.71% by analyzing the color of hormone test paper.
2) It was found that 7 out of 8 paper test detected oestrous sows successfully entered into the
gestation period.
3) Using this method to guide the optimal conception time of sows, the success rate of the
pregnancy in this experiment reached 85.71%
Acknowledgments
National Key Research and Development Program of China: 2016YFD0700204, Chongqing
Technology Innovation, and Application Development Project: cstc2019jscx-gksbX0093.
References
Altmann, M., 1941. Interrelations of the sex cycle and the behavior of the sow. Journal of Comparative
Psychology. 31, 481–498. http://doi.org/10.1037/h0060248.
Berckmans, D., 2008. Precision livestock farming (PLF). Computers and Electronics in Agriculture. 62,
1–1. http://doi.org/10.1016/j.compag.2007.09.002.
Bressers, H.P.M., J. Brake, J. Noordhuizen, 1991. Estrus detection in group-housed sows by analysis of
data on visits to the boar. Applied Animal Behaviour Science. 31 (3–4), 183–193. http://doi.org/10.1016/0168-
1591(91)90004-h.
Bressers, H.P.M., J.H.A. Tebrake, B. Engel, J. Noordhuizen, 1993. Feeding order of sows at an individual
electronic feed station in a dynamic group-housing system. Applied Animal Behaviour Science. 36 (2–3),
123–134. http://doi.org/10.1016/0168-1591(93)90004-9.
Freson, L., S. Godrie, N. Bos, J. Jourquin, R. Geers, 1998. Validation of an infra-red sensor for oestrus
detection of individually housed sows. Computers and Electronics in Agriculture. 20, 21–29. http://doi.org/
10.1016/S0168-1699(98)00005-2.
Houwers, H., 1988. Locality registration as a way of oestrus detection in an integrated group-housing for
sows. Proceedings of the International Congress on Applied Ethology in Farm Animals, 44–50.
Korthals, R.L., 1999. The effectiveness of using electronic identification for the identification of estrus
in swine. ASAE/CSAE-SCGR Annual International Meeting, Toronto, Ontario, Canada, 18–21 July. 10 pp.
Ostersen, T., C. Cornou, A.R. Kristensen, 2010. Detecting oestrus by monitoring sows’ visits to a boar.
Computers and Electronics in Agriculture. 74 (1), 51–58. http://doi.org/10.1016/j.compag.2010.06.003.
Pietrosemoli, S., C. Tang, 2020. Animal welfare and production challenges associated with pasture pig
systems: A review. Agriculture. 10 (6). http://doi.org/10.3390/agriculture10060223.
Scolari, S.C., S.G. Clark, R.V. Knox, M.A. Tamassia, 2011. Vulvar skin temperature changes
significantly during estrus in swine as determined by digital infrared thermography. Journal of Swine Health
and Production. 19, 151–155.
∙ 175 ∙
Sykes, D.J., J.S. Couvillion, A. Cromiak, S. Bowers, E. Schenck, M. Crenshaw, P.L. Ryan, 2012. The use
of digital infrared thermal imaging to detect estrus in gilts. Theriogenology. 78 (1), 147–152. http://doi.org/
10.1016/j.theriogenology.2012.01.030.
Turner, S., I. Camerlink, E. Baxter, R. D'Eath, S. Desire, R. Roehe, 2018. Breeding for pig welfare:
Opportunities and challenges. In: Advances in Pig Welfare. 399–414. Woodhead Publishing. Cambridge, MA,
United States.
∙ 176 ∙
Theme III:
Animal Welfare Monitoring,
Assessment, and Improvement
∙ 177 ∙
∙ 178 ∙
Effects of L-leucine Administration on Thermoregulation in Chicks

Guofeng Han a, Yangyang Cui a, Mitsuhiro Furuse b, Vishwajit S. Chowdhury b,
Yansen Li a, Chunmei Li a,*
a
Research Center for Livestock Environmental Control and Smart Production,
College of Animal Science and Technology, Nanjing Agricultural University,
Nanjing, Jiangsu 210095, China
b
Graduate School of Bioresource and Bioenvironmental Science, Kyushu University,
Fukuoka 819-0395, Japan
* Corresponding author. Email: chunmeili@njau.edu.cn
Abstract
Climate change and summer heat waves are inducing heat stress, which is as an increasing
threat for poultry production. Our recent studies found that leucine (Leu) concentrations of
embryonic brain and liver were declined by higher incubation temperature. Then L-Leu was in
ovo injected into eggs and caused hypothermia at hatching. This study aimed to investigate the
effects of L-Leu injection in ovo, oral or central on body temperature changes in chicks under
acute heat stress. In Experiment 1, fertilized eggs were incubated under control temperature and
in ovo injected with L-Leu solution on embryonic day 7. After hatching, male chicks were exposed
to thermal challenge. In Experiment 2, chicks were orally injected with L-Leu and exposed to
thermal stress. In Experiment 3, chicks were subjected to central injection of L-Leu. The results
indicated that L-Leu in ovo injection enhanced thermotolerance in male chicks, as the body
temperature was significantly lower in L-Leu in ovo injected chicks as compared with the control
group, and the mRNA expressions of heat shock proteins, a biomarker of thermal stress, were
attenuated by L-Leu under acute heat stress. In addition, daily oral injection of L-Leu
(continuously for 6 days) did not affect the body temperature changes in neonatal chicks and also
had no effects on thermotolerance under heat stress. Moreover, the body temperature was also not
affected by central injection with L-Leu in neonatal chicks. In conclusion, L-Leu in ovo
administration could be a novel approach to develop thermotolerant broilers.
Keywords: Heat stress, amino acid, in ovo injection, body temperature, heat shock protein.
1. Introduction
Summer high ambient temperature (HT) causes thermal stress on animal health, as well as
negative effects on animal production (Luber and McGeehin, 2008). The high temperature
scenarios haves been intensified due to the global warming. Poultry meat and eggs provide high-
quality protein and low levels of fat in human diets (Hargis et al., 1993). However, chickens are
very sensitive to heat stress as they have a relatively high metabolic rate and lack of sweat gland.
Many studies have been conducted to propose suitable and effective strategies for reducing
negative effects of heat stress on poultry production (Yahav, 2009; Renaudeau et al., 2012).
Amino acids not only serve as building blocks of protein, but also play important roles in growth
(Wu, 2009), regulation of body temperature (Chowdhury et al. 2020), food intake (Furuse et al.
2006) and behavior (Tran et al. 2015). It was suggested that amino acids may have some
thermoregulation-related functions under heat stress and potential to improve thermotolerance in
chicks (Chowdhury et al., 2020).
Branched chain amino acids (BCAAs), which includes leucine (Leu), isoleucine and valine,
are essential amino acids that are involved in different physiological functions, including insulin
resistance, metabolism regulation and recovery of heat-injured cells (Hsu-Ming et al., 2012; Shin
et al., 2014). Leu, which an important substrate for protein synthesis of skeletal muscle, activates
signaling factor of mammalian target of rapamycin (mTOR) to promote protein synthesis in
skeletal muscle and in adipose tissue, and involves in central nervous system regulation of food
intake and energy balance (Cota et al. 2006; Li et al., 2011). In our previous study (Han et al.,
2017), the hepatic and diencephalic Leu concentration in embryos was significantly reduced by
∙ 179 ∙
thermal manipulation, which is a technique of raising incubation temperature to enhance

thermotolerance in post-hatched chicks. Then, L-Leu was in ovo administrated to embryos and
caused hypothermia at hatching (Han et al., 2017). In this study, we conducted in ovo, oral and
central administration of L-Leu and examined its effects on body temperature response under
acute heat stress. The objectives of the study were to clarify whether or not L-Leu administration
could affect thermoregulation and improve thermotolerance in chicks.
All experimental protocols were approved under the guidelines for animal experiments in the
Faculty of Agriculture, Kyushu University and complied with Law No. 105 and Notification No.
6 of the Japanese government.
2.1. Experiment design
In Experiment 1, 84 eggs (Chunky strain) were purchased from a local hatchery (Mori
Hatchery, Fukuoka, Japan) in 2016, and placed into incubator under an incubation temperature of
37.6°C with 58–68% relative humidity and auto-turning every hour. The eggs were divided into
control and L-Leu groups (n = 42/group). The control group was in ovo fed sterile water (500
µl/egg) and the L-Leu group was in ovo fed 34.5 µmol L-Leu dissolved in 500 µl sterile water/egg
as described elsewhere (Han et al., 2017). Unfertilized eggs were detected by candling and were
discarded before the in ovo feeding on embryonic day (ED) 7. After disinfection and placement
of a small hole at the larger end of each egg, sterile water or L-Leu solution was injected to a
depth of 25 mm with a 1 ml disposable syringe with a 25-gauge needle. After completion of the
injection, the small holes were immediately sealed with scotch tape, and the eggs were returned
to the incubator. At ED 18, the eggs were moved to hatching trays. After hatching, chicks were
housed in groups in metal cages under a thermal neutral temperature (TN). At 2 days old, males
were separated by feather identification.
At 9 days old, the male chicks from each group were exposed to HT (35 ± 1°C) or CT (28 ±
1°C) for 180 min. All the birds were provided with free access to food and water. Rectal
temperature (RT) was measured at 0, 60, 120 and 180 min after the bird was placed in the chamber,
using a digital thermometer with an accuracy of ± 0.1°C (Thermalert TH-5, Physitemp
Instruments Inc., USA), by inserting the thermistor probe into the rectum (colon) through the
cloaca to a depth of 2 cm as reported previously (Han et al., 2017). Food intake was measured on
the basis of the amount of food that had disappeared over 180 min.
In Experiment 2, twenty 1-d-old layer chicks (Julia, male) were purchased from a local
hatchery (Murata Hatchery, Fukuoka, Japan) in 2017. Chicks were housed at the constant
temperature of 30 ± 1°C and constant light with free access to commercial starter diet and water.
Chicks were divided into two groups (Control and L-Leu; n = 10 group-1) based on body weight
on 1-d-old and were orally injected with sterile water (1 ml per 100 g body weight) or L-Leu
solution (112.5 mg / ml / 100 g body weight) from 2 to 7 day old, as the first week of life is critical
for epigenetic thermotolerance improvement in chicks. On 14-d old, chicks from two groups were
put into chambers under 35 ± 1°C for 180 min. The RTs were recorded at 0, 30, 60, 120, 180 min
after putting into the chambers. Food intake was also measured at 0 and 180 min.
In Experiment 3, fifty 1-d-old layer chicks (Julia, male) were purchased from a local hatchery
(Murata Hatchery, Fukuoka, Japan) in 2017. The rear condition was the same as for Experiment
2. Chicks were isolated into individual small cages and be assigned to three groups (n = 12 per
group) based on body weight on 3-d-old. Treatments were as follows: 1. Saline (0 µg per 10 µl);
2. L-Leu low dose (100 µg per 10 µl); 3. L-Leu high dose (200 µg per 10 µl) based on a previous
report (Izumi et al., 2004; Wang et al., 2012) where the highest dose stimulated food intake. On
4-d-old, chicks were intracerebroventricular (ICV) injected randomly with one of the 3 solutions
(10 µl bird-1) at three min intervals. Chicks in each treatment group were fasted after injection
because the food intake could be stimulated by L-Leu (Izumi et al., 2004; Wang et al., 2012)
which may affect the rectal temperature. Rectal temperature was measured at the time of 0-, 1-,
∙ 180 ∙
2- and 3- h after ICV injection. At the end of the experiment, chicks were decapitated following
the anesthesia with isoflurane. The brains were removed, and the location of the Evan Blue dye
was confirmed. Data from chick without dye in the later ventricle were excluded from the study.
2.2. Statistical analyses
Data concerning rectal temperature, food intake and gene mRNA expression were statistically
analyzed by two-way or three-way analysis of variance (ANOVA) followed by the Holm-Sidak’s
multiple comparisons test as a post-hoc analysis when a significant interaction was detected. Body
weight was statistically analyzed by student’s test. Statistical analysis was conducted using
GraphPad Prism 6 (GraphPad Software, Inc., San Diego, CA, USA). Results are expressed as
mean ± standard error of mean (SEM).
3.1. L-Leu in ovo administration enhances thermotolerance
In this study, we conducted in ovo administration of L-Leu once on ED 7 and examining its
effects on post-hatch chicks. It was found that in ovo administration of L-Leu caused a significant
(P<0.05) suppression of rectal temperature in chicks compared with the control group (Han et al.,
2007). Moreover, the rectal temperature and food intake were significantly suppressed in L-Leu
treated group in comparison with control chicks under acute heat stress (HT: 35°C vs CT: 28°C)
in male chicks (Han et al., 2018) (Figure 1). Body temperature is a critical parameter of
thermotolerance acquisition in mammals and avian (Yahav, 2015). Therefore, a lower body
temperature under heat stress indicates that the thermotolerance was improved by L-Leu in ovo
administration. Lower body temperature might have resulted from lower heat production or from
higher heat loss, or from both (Aggarwal and Upadhyay, 2012). Reduced food intake might have
contributed to reducing heat production in male chicks under heat stress by in ovo administration
of L-Leu, which is beneficial in alleviating heat stress in chickens (Etches et al., 2008).
Figure 1. The effects of in ovo administration of L-Leu on rectal temperature (A) and food
intake (B) in 9-day-old male broiler chicks exposed to either thermal neutral temperature (TN; 28
± 1°C) or a high ambient temperature (HT; 35 ± 1°C) for 180 min. The number of chicks in each
group was as follows: Control = 7, L-Leu (CT = 6, HT = 7). Different superscripts indicate
significant (P<0.05) differences. * indicates significant (P<0.05) differences between Control and
L-Leu groups under HT. Values are mean SEM. L-Leu, L-leucine. (J. Therm. Biol., 71:74–82,
with the permission from Elsevier as the authors’ right).
The diencephalon is the neural center of thermoregulation (Yahav, 2015), and is very sensitive
to thermal stress. Acute heat stress significantly increased diencephalic heat shock protein (HSP)
gene expression in chicks, which was reported in many studies (Yun et al., 2012; Sheikh et al.,
∙ 181 ∙
2016; Yang et al., 2016). However, in ovo administration of L-Leu caused not to alter the
diencephalic gene expression of HSP-70 and -90 under heat stress (Figure 2). HSP family genes
are activated in cells under exposure to different stress conditions, including thermal stress
(Horowitz, 2002; Khalil et al., 2011). HSP-70 and -90 have been extensively studied among HSP
families and their gene expression serves as a biomarker of thermal stress or oxidative stress in
organisms (Tedeschi et al., 2015). HSP-70 expression is also a useful marker of cellular injury
and neurotoxicity (Murakami, 2014). Therefore, it could be hypothesized that cellular injury and
the anti-oxidative status would be improved by the in ovo administration of L-Leu, since it was
reported that pine bark extract (flavangenol), a strong anti-oxidant, dose-dependently reduced the
HSP mRNA expression in the brain (Yang et al., 2016). In addition, supplements of other anti-
oxidative substances, such as zinc and vitamin C, also reduced expression of HSPs and attenuated
thermal stress (Yun et al., 2012; Sheikh et al., 2016). Therefore, it could be postulated that in ovo
administration of L-Leu may enhance the cellular protective function under heat stress and
attenuate the over-expression of HSP family genes.
Figure 2. The effects of in ovo administration of L-Leu on HSP-70 (A) and HSP-90 (B) mRNA
expressions in the brain of 9-day-old male broiler chicks exposed to either thermal neutral
temperature (TN; 28 ± 1°C) or high ambient temperature (HT; 35 ± 1°C) for 180 min. The number
of chicks in each group was as follows: Control (n = 7), L-Leu (CT = 6, HT = 7). Different
superscripts indicate significant (P<0.05) differences. Values are mean SEM. L-Leu, L-leucine;
HSP, heat shock protein. (Poult. Sci., 98:1243–1253, with the permission from Elsevier as the
authors’ right)
3.2. Body temperature changes after Oral or central administration of L-Leu in chicks
In last experiment, it was confirmed that L-Leu in ovo injection influenced thermoregulation
and afforded thermotolerance functions in chicks. Experiment 2 and Experiment 3 investigated
the thermoregulatory matter of chicks due to the oral or central administration of L-Leu in post
hatch period. It was found that the rectal temperature was significantly increased (Figure 3A),
which is a common physiological response (Chowdhury et al., 2020), under HT after L-Leu oral
injection. However, the body weight was also not affected by six-days oral administration of L-
Leu (Figure 3B). Recently, it was found that the free BCAAs concentrations in brain were
significantly declined under HT (40°C for 5 hours; Ito et al., 2015). Central administration of L-
Leu stimulated food intake in chicks (Izumi et al., 2004; Wang et al., 2012). However, the body
temperature was not affected by central administration of L-Leu in current study (Figure 4). Thus,
it was confirmed that central/oral administration of L-Leu did not affect body temperature
regulation in neonatal chicks.
∙ 182 ∙
Figure 3. The rectal temperature changes under heat challenge (35 ± 1°C; 180 min) in 14-d
old layer male chicks, which were daily orally injected with L-Leu or control solutions from 2- to
7-d old (A); The body weight after heat challenge in 14-d old layer male chicks (B). Values are
mean ± SEM with n=6-7/group. L-Leu, L-leucine.
Figure 4. The rectal temperature changes after central administration of saline or L-Leu
solution in 4-d old male chicks. Values are mean ± SEM with n=9–12/group. L-Leu, L-leucine.
In order to clarify the different effects of L-Leu in ovo and central/oral administration on
thermoregulation in chicks, the amino acid changes following L-Leu in ovo injection in embryos
and chicks were investigated. It was found that hepatic and plasma Leu concentration was not
affected at ED 14, which suggested that the injected L-Leu was totally metabolized before ED 14
(Han et al., 2021). Moreover, the Leu concentrations in plasma, liver and brain were also not
affected by L-Leu in ovo injection under CT or HT conditions in chicks (Han et al., 2019). These
facts suggested that L-Leu may be a trigger rather than a long-term regulator for L-Leu-mediated
thermotolerance in broiler chicks.
4. Conclusions
It was confirmed that L-Leu acted as an agent to enhance thermotolerance in male chicks. It
is the first time to identify that L-Leu has thermoregulatory functions and affords thermotolerance
in birds. Finding of this research suggested that L-Leu in ovo administration improved
thermotolerance under acute heat stress and improved growth performance under chronic heat
∙ 183 ∙
stress in adult broiler chickens. So, the new biotechnology, L-Leu in ovo administration could be
a potent to be a suitable technology for application in poultry production in future.
Acknowledgements
This study was partly supported by JSPS KAKENHI Grant (JP15K07694 and JP18K19271)
to VSC.
References
Aggarwal, A., R. Upadhyay, 2012. Heat Stress and Animal Productivity. Springer, India.
Chowdhury, V.S., G. Han, H.M. Eltahan, S. Haraguchi, E.R. Gilbert, M.A. Cline, J.F. Cockrem, T.
Bungo, M. Furuse, 2020. Potential role of amino acids in the adaptation of chicks and market-age broilers to
heat stress. Front Vet Sci, 7.
Cota, D., K. Proulx, K.A.B. Smith, S.C. Kozma, G. Thomas, S.C. Woods, R.J. Seeley, 2006.
Hypothalamic mTOR signaling regulates food intake. Science. 312(5775), 927–930.
Etches, R.J., T.M. John, A.M. Verrider Gibbins, 2008. Behavioural, physiological, neuroendocrine and
molecular responses to heat stress, in: Daghir N. J. (Eds.), Poultry Production in Hot Climates. CABI,
Wallingford, UK, 48–79 p.
Furuse, M., T. Takagi, T. Bungo, D.M. Denbow, 2006. Functions of L-pipecolic acid as an L-lysine
metabolite in the brain. In: Chen FJ (Eds), Trends in Brain Mapping Research. Hauppauge, New York, pp.
145–165.
Han, G., Y. Ren, D. Shen, S. Li, V.S. Chowdhury, Y. Li, M. Furuse, C. Li, 2020. L-Leucine in ovo
administration causes growth retardation and modifies specific amino acid metabolism in broiler embryos. J
Poult Sci, In press.
Han, G., H. Yang, Y. Wang, R. Zhang, K. Tashiro, T. Bungo, M. Furuse, V.S. Chowdhury, 2019. Effects
of in ovo feeding of L-leucine on amino acids metabolism and heat-shock protein-70, and -90 mRNA
expression in heat-exposed chicks. Poult Sci, 98: 1243–1253.
Han, G., H. Yang, T. Bungo, H. Ikeda, Y. Wang, L.T.T. Nguyen, H.M. Eltahan, M. Furuse, V.S.
Chowdhury, 2018. In ovo L-leucine administration stimulates lipid metabolisms in heat-exposed male, but not
female, chicks to afford thermotolerance. J Therm Biol, 71, 74–82.
Han, G., H. Yang, M.A. Bahry, P.V. Tran, P.H. Do, H. Ikeda, M. Furuse, V.S. Chowdhury, 2017. L-
Leucine acts as a potential agent in reducing body temperature at hatching and affords thermotolerance in
broiler chicks. Comp Biochem Physiol A Mol Integr Physiol, 204, 48–56.
Hargis, P.S., M.E. Van Elswyk, 1993. Manipulating the fatty acid composition of poultry meat and eggs
for the health-conscious consumer. Worlds Poult Sci J, 49(3), 251–264.
Horowitz, M., 2002. From molecular and cellular to integrative heat defense during exposure to chronic
heat. Comp Biochem Physiol A Mol Integr Physiol, 131(3), 475–483.
Hsu-Ming, W., K. Naito, Y. Kinoshita, H. Kobayashi, K.I. Honjoh, K. Tashiro, T. Miyamoto, 2012.
Changes in transcription during recovery from heat injury in Salmonella typhimurium and effects of BCAA
on recovery. Amino Acids, 42(6), 2059–2066.
Ito, K., M.A. Bahry, Y. Hui, M. Furuse, V.S. Chowdhury, 2015. Acute heat stress up-regulates
neuropeptide Y precursor mRNA expression and alters brain and plasma concentrations of free amino acids
in chicks. Comp Biochem Physiol A Mol Integr Physiol, 187, 13–19.
Izumi, T., K. Kawamura, H. Ueda, T. Bungo, 2004. Central administration of leucine, but not isoleucine
and valine, stimulates feeding behavior in neonatal chicks. Neurosci Lett, 354(2), 166–168.
Khalil, A.A., N.F. Kabapy, S.F. Deraz, C. Smith, 2011. Heat shock proteins in oncology: Diagnostic
biomarkers or therapeutic targets? BBA. 1816(2), 89–104.
Li, F., Y. Yin, B. Tan, X. Kong, G. Wu, 2011. Leucine nutrition in animals and humans: mTOR signaling
and beyond. Amino Acids, 41(5), 1185.
Luber, G., M. McGeehin, 2008. Climate change and extreme heat events. Am J Prev Med, 35(5), 429–
435.
Murakami, A., 2014. Dose-dependent functionality and toxicity of green tea polyphenols in experimental
rodents. Arch Biochem Biophys, 557, 3–10.
Renaudeau, D., A. Collin, S. Yahav, V. de Basilio, J.L. Gourdine, R.J. Collier, 2012. Adaptation to hot
climate and strategies to alleviate heat stress in livestock production. Animal, 6, 707–728.
Sheikh, A.A., A. Aggarwal, O. Aarif, 2016. Effect of in vitro zinc supplementation on HSPs expression
and Interleukin 10 production in heat treated peripheral blood mononuclear cells of transition Sahiwal and
Karan Fries cows. J Therm Biol, 56, 68–76.
∙ 184 ∙
Shin, A.C., M. Fasshauer, N. Filatova, L.A. Grundell, E. Zielinski, J.Y. Zhou, T. Scherer, C. Lindtner,
P.J. White, A.L. Lapworth, O. Ilkayeva, 2014. Brain insulin lowers circulating BCAA levels by inducing
hepatic BCAA catabolism. Cell Metab, 20(5), 898–909.
Tedeschi, J.N., W.J. Kennington, O. Berry, S. Whiting, M. Meekan, N.J. Mitchell, 2015. Increased
expression of Hsp70 and Hsp90 mRNA as biomarkers of thermal stress in loggerhead turtle embryos (Caretta
caretta). J Therm Biol, 47, 42–50.
Tran, P.V., V.S. Chowdhury, M. Nagasawa, M. Furuse, 2015. Changes in free amino acid and monoamine
concentrations in the chick brain associated with feeding behavior. Springer Plus. 4(1), 1–9.
Wang, S., P. Khondowe, S. Chen, J. Yu, G. Shu, X. Zhu, L. Wang, P. Gao, Q. Xi, Y. Zhang, Q. Jiang,
2012. Effects of "Bioactive" amino acids leucine, glutamate, arginine and tryptophan on feed intake and
mRNA expression of relative neuropeptides in broiler chicks. J Anim Sci Biotechnol, 3(1), 1–8.
Wu, G., 2009. Amino acids: metabolism, functions, and nutrition. Amino Acids. 37(1), 1–17.
Yahav, S., 2015. Regulation of body temperature: strategies and mechanisms, in: Scanes CG. (Eds),
Sturkie's Avian Physiology, 6th edition, Academic Press, London, UK, 869–905 p.
Yahav, S., 2009. Alleviating heat stress in domestic fowl: different strategies. World's Poultry Science
Journal, 65, 719–732.
Yang, H., V.S. Chowdhury, M.A. Bahry, P.V. Tran, P.H. Do, G. Han, R. Zhang, H. Tagashira, M.
Tsubata, M. Furuse, 2016. Chronic oral administration of pine bark extract (flavangenol) attenuates brain and
liver mRNA expressions of HSPs in heat-exposed chicks. J Therm Biol, 60, 140–148.
Yun, S.H., Y.S. Moon, S.H. Sohn, I.S. Jang, 2012. Effects of cyclic heat stress or vitamin C
supplementation during cyclic heat stress on HSP70, inflammatory cytokines, and the antioxidant defense
system in Sprague Dawley rats. Exp Anim, 61(5), 543–553.
∙ 185 ∙
Physiological Responses in Dairy Cows to Increasing Temperatures

at Different Relative Humidity and Air Velocity Levels
Mengting Zhou, André J.A. Aarnink *, Thuy T.T. Huynh, Peter W.G. Groot Koerkamp
Department of Biosystems Engineering, Wageningen University and Research,
Wageningen, The Netherlands
* Corresponding author. Email: andre.aarnink@wur.nl
Abstract
The effects of increasing ambient temperature (T) at different relative humidity (RH) and air
velocity (AV) levels on physiological responses were studied using 20 cows inside climate
respiration chambers. The climate inside the chamber was programmed to follow a daily pattern
of lower night and higher day temperatures with 9°C difference. The T was gradually increased
from 7 to 21°C during nighttime (12 hrs) and 16 to 30 °C at daytime (12 hrs) within 8 days (step
size of 2°C per day for both night and day temperatures). The experimental treatments comprised
three RH levels and three AV levels. The RH levels varied also during day-night: RH_l: 30-50%;
RH_m: 45-70%; RH_h: 60-90%. Three AV levels were created: AV_l: fan off (AV 0.1m/s);
AV_m: fan speed level 1 (AV 1.0m/s); AV_h: fan speed level 2 (AV 1.5m/s). The AV_m and
AV_h were only combined with RH_m. The animals had free access to feed and water.
Respiration rate (RR), rectal temperature (RT) and skin temperature (ST) were measured and
analyzed. The inflection point temperatures (IPt’s) at which a certain physiological variable
started to change were determined for the different RH and AV levels. Results showed that IPT’s
for RR (first indicator) varied between 18.9 and 25.5 °C, for RT (a delayed indicator) between
20.1 and 25.9 °C. The IPt’s decreased with higher RH levels for both RR and RT while IPt’s
increased with higher AV for RR and had a relatively minor change with AV for RT. The ST was
positively correlated (P<0.01) with ambient temperatures. ST was not affected by RH levels but
significantly affected by AV levels.
Keywords: Dairy cow, temperature, relative humidity, air velocity, heat stress, physiological
response
1. Introduction
With the expected increase of the global temperature (Gauly et al., 2013), the problem of heat
stress in dairy cows in summer has become increasingly prominent, with negative effects on health
(de Andrade Ferrazza et al., 2017; Kadzere et al., 2002), production (Hill and Wall, 2015) and
reproduction (García-Ispierto et al., 2007; Schüller et al., 2014) and increased mortality risk
(Vitali et al., 2009). Dairy cows are particularly sensitive to higher ambient temperatures (Kadzere
et al., 2002). Generally, cows can maintain their body heat balance at a constant level (Mount,
1979) at given ambient thermal condition. When ambient temperatures rise above the so-called
thermal comfort conditions, the cows must recruit extra physiological responses in order to get
rid of the produced heat. The physiological responses could include increased respiration rate,
increased blood flow from the core body to the skin surface, increased sweating rate, increased
water consumption, reduced rumination activities, increased heart rate, reduced heat production,
and increased deep body temperature (Amamou et al., 2019; Burfeind et al., 2012; Galán et al.,
2018; Hill and Wall, 2015).
With regards to the response of increasing respiration rate, Kadzere et al. (2002) reported that
about 15% of accumulated metabolic heat is used by the respiratory tract for evaporative heat loss.
Dikmen and Hansen (2009) defined heat stress as the sum of external forces acting on an animal
that causes an increase in body temperature and evokes a physiological response. Li et al. (2020)
determined that the rectal temperature of a high producing cow already increased above 20.4°C.
The temperature threshold was modelled as 23°C for metabolic heat production by McArthur
(1987), above which, metabolic rate began to decline in response to elevated body temperature.
∙ 186 ∙
Recognizing the importance of ambient air humidity, temperature-humidity index (THI) is

frequently used to assess heat stress magnitude in dairy cows. For example, Pinto et al. (2020)
determined the THI threshold on respiration rate, heart rate and rectal temperature of high
producing cows and suggested actions could be taken when THI raises above 65.
Still little is known about the exact temperatures above which cows start to adapt to increasing
heat loads. These temperatures will depend on humidity and air velocity levels. Therefore, the
objectives of this study were (1) to quantify the effects of increasing temperatures on physiological
parameters of Holstein cows; (2) to determine the inflection point temperatures for activation of
adaptive mechanisms at different relative humidity and air velocity levels.
The experiment was conducted in 2021, at the animal research facilities “Carus” of
Wageningen University and Research (WUR), Wageningen, the Netherlands. The experimental
procedures were approved by the Institutional Animal Care and Use Committee of Wageningen
University and was conducted under the Dutch Law on Animal Experiments (Project No. 2019.D-
0032).
Twenty Holstein cows were used and all cows underwent a pre-trial period of conditioning to
reduce the reactivity and improve the adjustment to the new environment and intensive
management. For this purpose, the cows were maintained with halters and were managed and tied
with ropes in the pens in an area close to the climate-controlled respiration chambers for 7 d.
Each cow was subjected to a 11-d experimental period in climate-controlled respiration
chambers, including 3 d for adaptation and 8 d for the experiment. The experiment consisted of
five treatments and for each treatment, four replicates were measured. For all dairy cows, the
inside air temperatures followed a daily pattern of lower night and higher day temperatures with
9°C difference and were set for d 1 to 3 as 7°C during the night and 16°C during the day.
Subsequently, inside air temperature (T) was gradually increased from 7 to 21°C at night and 16
to 30°C at daytime within 8 days (step size of 2°C per day for both night and day temperatures).
The experimental treatments comprised three relative humidity (RH) levels and three air
velocity (AV) levels. RH levels varied also during day-night: RH_l: 30–50%; RH_m: 45–70%;
RH_h: 60–90%. Three AV levels were created: AV_l: fan off (AV 0.1 m s-1); AV_m: fan speed
level 1 (AV 1.0 m s-1); AV_h: fan speed level 2 (AV 2.0 m s-1). For AV_m and AV_h the T range
was increased with 2°C. The AV_m and AV_h were only combined with RH_m. The daily
patterns during the 8 experimental days are shown in Figure 1. Between 0700 to 1000 h, T and
RH changed gradually into day levels and stayed constant until 1900 h, while fans were kept on
from 0900 to 2100 h for two high AV circumstances. Between 1900 to 2200 h, T and RH level
gradually changed into night levels and stayed constant again until next day 0700 h, while fans
were kept off from 2100 h to next day 0900 h during the 11-d experimental period for each
treatment. Because of the data collection procedures carried out by the researchers, there was a
1-h time lag between chamber 1 and 2 in T, RH and AV change, switching on or off of the lights,
feeding and milking.
All cows were subjected to the same feed management. Feed and water were available ad
libitum. Feed was given twice a day (0500 and 1530). The diet was formulated to meet or exceed
the nutritional requirements of lactating Holstein cows. The cows were milked twice a day at 0500
and 1530.
2.2. Data collection
The T and RH of the cells were continuously recorded at 30-sec intervals automatically. AV
was measured by an anemometer (Testo 5-412-983, Testo SE & Co. KGaA, Germany) at 5 spots:
close to the neck, middle trunk and rump above the cow, and lateral sides of the cow. All
physiological parameters on animals were measured at 1800 in the afternoon after 8 hours
∙ 187 ∙
exposure to warm conditions. Respiration rate (RR) was measured by recording the time used for
counting 10 flank movements by means of a stopwatch. Skin temperature (ST) was measured with
a thermometer probe (Testo 0602 0393, Testo SE & Co. KGaA, Germany) and a handy data logger
(Testo 435-4, Testo SE & Co. KGaA, Germany) in close touch with four skin surface areas under
hair coat including heart, back, belly and rump and was calculated based on the average value
from these four values. The rectal temperature (RT) was obtained by inserting a digital
thermometer (VT 1831, Microlife, Switzerland) to the depth of 3 cm into the rectum of cows and
maintaining until beep.
100 100 40 80
90 90 36 70
80 80 32
60
Temperature, °C
28
Temperature, C
70 70
24 50
60 60
RH, %
RH, %
50 50 20 40
40 40 16 30
30 30 12
20
20 20 8
10 10 4 10
0 0 0 0
0 1 2 3 4 5 6 7 8 4:00 7:00 10:00 13:00 16:00 19:00 22:00
Day number Time
Air temperature 30 - 50% RH 45 - 70% RH 60 - 90% RH Air temperature 45 - 70% RH
Figure 1. Schematic temperature and relative humidity patterns during the 8 experimental days.
Between 0700 to 1000 h temperature and relative humidity changed gradually into day levels
and stayed constant until 1900 h. Between 1900 to 2200 h, temperature and relative humidity
gradually decreased into night levels and stayed constant again until next day 0700 h.
One cow (receiving treatment of RH_m and AV_m) was excluded from the experiment
because of mastitis infection. A cow was considered as the experimental unit for all parameters.
All statistical analyses were performed in SAS 9.4 (SAS Institute Inc., Cary, NC) using the
MIXED procedure. A nonlinear mixed model was used to investigate the influence of increasing
T for each treatment. The model can be written as
𝑦𝑖𝑗𝑘 = 𝐶 + 𝑇𝑟𝑡𝑖 +𝑎𝑖 ∙ 𝑧 + 𝑐𝑜𝑤𝑖𝑗𝑘 + 𝜀𝑖𝑗𝑘 (1)
where, 𝑦𝑖𝑗𝑘𝑙 is the observed response variables; C is the general mean; 𝑇𝑟𝑡𝑖 is the fixed effect of
the i-th treatment (five treatments); 𝑎𝑖 is the regression coefficient of treatment i; 𝑧 is the
structural part to create a broken-line regression of the daily T with an inflection point (IPt);
𝑐𝑜𝑤𝑖𝑗𝑘 is the random effect of j-th cow for the k-th observation; and 𝜀𝑖𝑗𝑘 is the residual error.
𝑧 = (𝑇 > 𝐼𝑃𝑡) ∗ (𝑇 − 𝐼𝑃𝑡), 𝑤ℎ𝑒𝑟𝑒 (𝑇 − 𝐼𝑃𝑡) 𝑖𝑠 𝑑𝑒𝑓𝑖𝑛𝑒𝑑 𝑎𝑠 𝑧𝑒𝑟𝑜 𝑖𝑓 (𝑇 ≤ 𝐼𝑃𝑡) (2)
The broken-line regression model was fit and IPt was determined for RR and RT by meeting
fit statistics criterion, using SAS NLMixed including cow random effects. The 𝜒 2 test was used
to test the significance of the model, and t-test was applied to pairwise compare the differences
between treatments.
If the model failed to converge, a linear regression model was used. A mixed model was used
to investigate the influence of increasing T for each treatment. Model assumptions were evaluated
by examining the distribution of residuals (homogeneity of variance and normality). The linear
regression model was as follows:
𝑦𝑖𝑗𝑘 = C + 𝑇𝑟𝑡𝑖 + 𝑏𝑖 ∙ (T × 𝑇𝑟𝑡𝑖 ) + 𝑎𝑖 ∙ 𝑇 + 𝑐𝑜𝑤𝑖𝑗𝑘 + 𝜀𝑖𝑗𝑘 (3)
where, 𝑦𝑖𝑗𝑘 is the observed response variables; 𝐶 is the intercept; 𝑇𝑟𝑡𝑖 is the fixed effect of the i-
th treatment; 𝑎𝑖 , 𝑏𝑖 are regression coefficients for 𝑇 and (T × 𝑇𝑟𝑡𝑖 ), respectively; (T × 𝑇𝑟𝑡𝑖 ) is
interaction between ambient temperatures and treatments; 𝑐𝑜𝑤𝑖𝑗𝑘 is the random effect of j-th cow
for the k-th observation; and 𝜀𝑖𝑗𝑘 is the random residual. The 𝜒 2 test was used to test the
∙ 188 ∙
significance of the model, and t-test was applied using PDIFF statement to pairwise compare the
differences between treatments. Significance was declared when P ≤ 0.05 unless otherwise
indicated.
Data in Table 1 shows that the broken-line model fits for RR and RT and a linear model fits
for ST. Treatment effects on respiration rate were presented. With increasing T, RR remained at
stable level until reaching to the IPt, after which it increased linearly. Relative humidity affected
the IPt for RR significantly (25.45, 20.98 and 18.88°C for RH_l, RH_m and RH_h, respectively;
P<0.05). Interestingly, the effect of higher AV did delay the IPt, but only with high level AV
rather than medium level. Air velocity with high fan speed level affected the IPt for RR (20.98,
22.84 for AV_l and AV_h, respectively; P<0.05) while the IPt of the AV_m treatment did not
differ from IPt of the AV_l treatment. Similarly, the IPt of the AV_m treatment was significantly
lower than that of AV_h treatment.
Table 1. Estimation of different dependent variables on temperatures for broken-line or linear
models.
Dependent Respiration rate, Rectal Skin
variables breath min-1 temperature, °C temperature, °C
Adjusted R2 0.725 0.581 0.811
Model IPt1 IPt Intercept a
components Constanta Constant a
c
Treatments I 25.45±0.42 25.34±0.92a 30.12±0.76a
RH_l AV_l 36.51±3.23c 38.61±0.13a 0.21±0.03a
c bc
9.46±1.34 0.08±0.03
II 20.98±0.90a 25.89±0.45a 30.36±0.83a
RH_m AV_l 30.06±3.70a 38.66±0.07a 0.19±0.04a
ab a
5.13±0.88 0.14±0.05
III 18.88±1.04d 20.14±1.2b 31.30±0.83a
ab c
RH_h AV_l 28.41±6.93 38.38±0.08 0.18±0.04a
5.31±0.63b 0.10±0.02ab
IV 20.98±0.97a 20.95±1.19b 25.07±0.92b
b b
RH_m AV_m 24.81±2.88 38.09±0.15 0.36±0.04b
4.17±0.54a 0.07±0.02c
V 24.20±0.84b 25.25±1.32a 25.26±0.66b
RH_m AV_h 32.70±2.35a 38.43±0.10c 0.35±0.03b
5.34±1.23ab 0.07±0.02c
a,b,c,d values within a column with different superscripts differ, P<0.05.
1IPt=inflection point temperature; a=slope.
Rectal temperature was affected by increasing temperature (Table 1). Until the T reached
23.51°C (on average), the RT of cows was constant at an average of 38.43°C. Above the IPt, it
increased 0.09°C/°C. The RH affected the IPt (25.34, 25.89 and 20.14°C for RH_l, RH_m and
RH_h, P<0.05). AV did not have a clear effect on RT, and for AV_m condition, the IPt even
appeared at lower T than AV_l (P<0.05).
Skin temperature increased linearly with increasing T (P<0.001). On average, ST increased
by 0.19 °C with fan off and 0.36 °C with fan on for every Celsius degree increase in T,
respectively. There was no RH effect on ST, while average ST was lower in two AV_m and AV_h
treatment groups than that in three no fan treatment groups (P<0.05). There was an interaction
effect between AV and T (P<0.05, Figure 2).
The development of new strategies for heat mitigation that consider the physiological
adaptations of the animal are needed. In the Netherlands, within a site-specific on farm location,
∙ 189 ∙
humidity does not vary much during a hot period (André et al., 2011). Using only temperature has
the advantage that temperature of the weather forecasts can be directly compared with the
inflection point temperatures to indicate the risk for heat stress in the near future.
(a1) (a2)
(b1) (b2)
Figure 2. Relationships between different treatments and (a1) respiration rate and (b1) skin
temperature (treatment I, II and III); (a2) respiration rate and (b2) skin temperature (treatment II,
IV and V).
This study evaluated the effect of increasing temperatures under different relative humidity
and air velocity levels on thermoregulatory responses to elucidate the adaptive mechanisms of
heat dissipation in cattle. The increase in respiration rate was the first reaction of cows under
warm conditions attempting to maintain thermoregulation by increasing evaporative heat loss
from the respiratory-tract surface (Silanikove, 2000). Berman et al. (1985) noted that RR in the
dairy cow started to rise when ambient temperature surpassed 25°C. They suggested that
respiratory evaporative heat loss is extremely important in large cattle for maintenance of thermal
stability due to their large body size. In our study, RR started to increase when ambient T was
above 19°C for 60% RH with low AV conditions and the IPt’s increased with lower RH levels
which was comparable to the results from other studies (Pinto et al., 2020). The medium AV had
no effect on the IPt while high AV increased IPt with 2°C. According to Spiers et al. (2018), the
benefit of fan cooling was highly dependent on the air T and because of the small difference
between air T and ST, the medium AV could not help a lot to dissipate the heat by convection,
which resulted in a little difference in RR between two treatments (RH_m*AV_l and
RH_m*AV_m).
When the cow failed to dissipate all the heat loss from the body, the RT increased. The
relationship between RT and T has been studied by many researchers and a recent study by Li et
al. (2020) showed the RT started to rise at ambient T of 20.4 without taking into account RH and
∙ 190 ∙
AV, which was comparable to our IPt’s receiving treatment of RH_h*AV_l. With lower RHs, the
IPt’s for RT were raised by about 5°C in our study. Pinto et al. (2020) reported that the critical
THI thresholds for RT was 70, which was consistent with our result from treatments without fans
when recalculating the air T and RH to THI. Our results indicated that IPt for RT with medium
AV was lower than that with low AV, which seems not logical in reality since the most common
cooling system is mechanical fans for increasing air velocity around cows in the Netherlands
during summer period (André et al., 2011). The possible reason could be that the specific cows
under this treatment were, by random chance, less heat adapted or produced more heat than the
cows in the other treatments. Some works confirm that the individual cow factors can influence
the physiological parameters and susceptibility of dairy cows to heat stress (Berman, 2005;
Gaughan et al., 2000), however, those cow factors were not included in our analysis. The lowest
and highest IPt values were found at the treatments with high RH with low AV and medium RH
with low AV, respectively. This indicated that we should pay more attention to keep RH inside
the barn at a relatively low level during the hot periods of warm days to avoid heat stress effects.
Our results showed that skin temperature increased at higher ambient temperatures. In this
way the cow enables easier dissipation of heat to the environment and maintains a gap between
ST and T. At low temperature conditions, ST with high AV was much lower than that with low
AV, which could be explained by the heat transfer theory: the amount of convective heat loss was
increased by high AV and the cows had to decrease vasodilation for reducing transfer of heat from
blood to the periphery (Silanikove, 2000). According to (Gebremedhin and Wu, 2001), above
35°C of skin surface temperature, all routes of heat exchange become very active and cows slowly
began to store heat and thus rectal temperature increased. When recalculating our results at the
same skin temperature (35°C), the corresponding ambient T’s were comparable to IPt’s for RT
for treatments without fans but for two treatments with high AV, the calculated ambient T’s were
higher than IPt’s for RT. Differences between different studies can be due to the lack of AV effects
in other studies (Yadav et al., 2017).
From this study it can be concluded that for high productive dairy cows physiological signs of
heat stress already occur at moderate temperatures of higher than 20°C (start to increase rectal
temperature). The modern western dairy cow has a high metabolic activity because of bigger body
size and higher milk production and thereby a high heat production. However, dairy cows in
moderate climate regions (e.g., the Netherlands) possibly are less heat adapted. Dairy cow farmers
should take this into account during summer periods to avoid negative consequences.
4. Conclusions
From this study it was concluded:
1. Above inflection point temperatures, thermal physiological parameters of dairy cows
substantially changed. Respiration rate was the first indicator showing that the cow is
reacting to high ambient temperatures.
2. Rectal temperature increased above an average ambient temperature of 23.5°C. Increase
in RT is an indicator that inside ambient temperature is above the upper limit of thermal
neutral zone.
3. Profound effects of relative humidity on physiological parameters were found. The IPt’s
of RR and RT decreased with higher relative humidity levels.
4. Generally, the effects of air velocity were small on RR and RT but significantly different
in ST. The difference of ST between low AV and higher AV got smaller with increasing
T.
Reference
Amamou, H., Beckers, Y., Mahouachi, M., Hammami, H., 2019. Thermotolerance indicators related to
production and physiological responses to heat stress of holstein cows. Journal of Thermal Biology 82, 90–
98.
∙ 191 ∙
André, G., Engel, B., Berentsen, P., Vellinga, T.V., Lansink, A.O., 2011. Quantifying the effect of heat
stress on daily milk yield and monitoring dynamic changes using an adaptive dynamic model. Journal of Dairy
Science 94, 4502–4513.
Berman, A., 2005. Estimates of heat stress relief needs for Holstein dairy cows 1. Journal of Animal
Science 83, 1377–1384.
Berman, A., Folman, Y., Kaim, M., Mamen, M., Herz, Z., Wolfenson, D., Arieli, A., Graber, Y., 1985.
Upper Critical Temperatures and Forced Ventilation Effects for High-Yielding Dairy Cows in a Subtropical
Climate. Journal of Dairy Science 68, 1488–1495.
Burfeind, O., Suthar, V.S., Heuwieser, W., 2012. Effect of heat stress on body temperature in healthy
early postpartum dairy cows. Theriogenology 78, 2031–2038.
de Andrade Ferrazza, R., Mogollón Garcia, H.D., Vallejo Aristizábal, V.H., de Souza Nogueira, C.,
Veríssimo, C.J., Sartori, J.R., Sartori, R., Pinheiro Ferreira, J.C., 2017. Thermoregulatory responses of
Holstein cows exposed to experimentally induced heat stress. Journal of Thermal Biology 66, 68–80.
Dikmen, S., Hansen, P.J., 2009. Is the temperature-humidity index the best indicator of heat stress in
lactating dairy cows in a subtropical environment? Journal of Dairy Science 92, 109–116.
Galán, E., Llonch, P., Villagrá, A., Levit, H., Pinto, S., del Prado, A., 2018. A systematic review of non-
productivity-related animal-based indicators of heat stress resilience in dairy cattle. PloS one 13, e0206520.
García-Ispierto, I., López-Gatius, F., Bech-Sabat, G., Santolaria, P., Yániz, J.L., Nogareda, C., De Rensis,
F., López-Béjar, M., 2007. Climate factors affecting conception rate of high producing dairy cows in
northeastern Spain. Theriogenology 67, 1379–1385.
Gaughan, J., Holt, S., Hahn, G., Mader, T., Eigenberg, R., 2000. Respiration rate: Is it a good measure of
heat stress in cattle? Asian-Australasian Journal of Animal Sciences 13, 329–332.
Gauly, M., Bollwein, H., Breves, G., Brügemann, K., Dänicke, S., Daş, G., Demeler, J., Hansen, H.,
Isselstein, J., König, S., Lohölter, M., Martinsohn, M., Meyer, U., Potthoff, M., Sanker, C., Schröder, B.,
Wrage, N., Meibaum, B., von Samson-Himmelstjerna, G., Stinshoff, H., Wrenzycki, C., 2013. Future
consequences and challenges for dairy cow production systems arising from climate change in Central Europe
– a review. Animal 7, 843–859.
Gebremedhin, K.G., Wu, B., 2001. A model of evaporative cooling of wet skin surface and fur layer.
Journal of Thermal Biology 26, 537–545.
Hill, D.L., Wall, E., 2015. Dairy cattle in a temperate climate: the effects of weather on milk yield and
composition depend on management. Animal 9, 138–149.
Kadzere, C., Murphy, M., Silanikove, N., Maltz, E., 2002. Heat stress in lactating dairy cows: a review.
Livestock Science 77, 59–91.
Li, G., Chen, S., Chen, J., Peng, D., Gu, X., 2020. Predicting rectal temperature and respiration rate
responses in lactating dairy cows exposed to heat stress. Journal of Dairy Science 103, 5466–5484.
McArthur, A.J., 1987. Thermal interaction between animal and microclimate: a comprehensive model.
Journal of Theoretical Biology 126, 203–238.
Mount, L.E., 1979. Adaptation to thermal environment. Man and his productive animals. Edward Arnold
(Publishers) Ltd., London.
Pinto, S., Hoffmann, G., Ammon, C., Amon, T., 2020. Critical THI thresholds based on the physiological
parameters of lactating dairy cows. Journal of Thermal Biology 88, 102523.
Schüller, L.K., Burfeind, O., Heuwieser, W., 2014. Impact of heat stress on conception rate of dairy cows
in the moderate climate considering different temperature–humidity index thresholds, periods relative to
breeding, and heat load indices. Theriogenology 81, 1050–1057.
Silanikove, N., 2000. Effects of heat stress on the welfare of extensively managed domestic ruminants.
Livestock Production Science 67, 1–18.
Spiers, D.E., Spain, J.N., Ellersieck, M.R., Lucy, M.C., 2018. Strategic application of convective cooling
to maximize the thermal gradient and reduce heat stress response in dairy cows. Journal of Dairy Science 101,
8269–8283.
Vitali, A., Segnalini, M., Bertocchi, L., Bernabucci, U., Nardone, A., Lacetera, N., 2009. Seasonal pattern
of mortality and relationships between mortality and temperature-humidity index in dairy cows. Journal of
Dairy Science 92, 3781–3790.
Yadav, B., Singh, G., Wankar, A., 2017. The use of infrared skin temperature measurements for
monitoring heat stress and welfare of crossbred cattle. Indian Journal of Dairy Science 70, 127–131.
∙ 192 ∙
Effects of Keel Bone Fracture on Behavior, Welfare, and

Production Performance of Laying Hens Housed Individually in
Furnished Cages
Haidong Wei a, Yanju Bi a, Yanru Feng a, Runxiang Zhang a, Jun Bao a,*
a
b
College of Life Science, Northeast Agricultural University, Harbin, Heilongjiang 150030, China
* Corresponding author. Email: jbao@neau.edu.cn
Abstract
Keel fracture has adverse effects on welfare and behavior of laying hens. To investigate this,
90 Lohmann white laying hens at 17 weeks of age (WOA) were housed individually in furnished
cages. Hens were marked with fractured keel (FK) or normal keel (NK) based on the keel bone
status through palpation at 22, 27, 32, 37, and 42 WOA. After each palpation, the behavior was
observed, and the egg production, dirty eggs, broken eggs and feed intake of FK and NK hens
were recorded at 27–32, 32–37, and 37–42 WOA, respectively. Following each behavioral
observation, 10 fresh FK hens and 10 NK hens were selected to determinate the welfare and egg
quality. The results showed that the incidences of keel fracture increased with the age. Compared
with NK hens, the sitting and standing behaviors increased (P<0.05), while feeding, walking,
perching and jumping behaviors decreased (P<0.05) in FK hens. During the experiment period,
the egg production rate, body weight, daily feed intake, and eggshell strength, thickness and
weight decreased (P<0.05), and duration of tonic immobility increased (P<0.05) in FK hens
compared with NK hens. At 27–32 WOA, FK hens had elevated broken-egg rate (P<0.05). There
were no differences in the dirty-egg rate, protein height, Haugh unit, feather quality, and toe and
foot pad health (P >0.05). Therefore, keel fracture in hens altered behavior and reduced the
welfare, production performance, feed intake and eggshell quality.
Keywords: Keel bone damage, behavior, animal welfare, egg quality, laying hen
1. Introduction
In modern intensive production systems for laying hens, furnished cages and non-cage
systems have become common due to the increasing attention given to hen welfare and with the
ban imposed by the European Union on the traditional cage-raising system from 2012. Non-cage
systems allow more space for hens to perform natural behaviors (Stratmann et al., 2015), however,
it evidentially increases the production and management cost, and increases the incidences of
mortality and keel fracture (Wilkins et al., 2004).
Furnished cages with perches, nests, litter and other facilities allow the native behavior, and
improve the welfare of hens than conventional cages. However, their usage can lead to some
welfare and health problems such as keel damage (including keel fracture and deviation). Some
studies have shown that keel fractures cause pain and stress response (Nasr et al., 2012a; Wei et
al., 2019). The incidences of keel fractures increase with the hen’s age, with a peak record during
laying period, that gradually decreases after 42 weeks of age (WOA) (Stratmann et al., 2015;
Gebhardt-Henrich and Fröhlich, 2015). Heerkens et al. (2015) found that incidences of keel
fractures were positively correlated with foot pad disorders. Additionally, keel fractures
influenced the performance and egg quality in laying hens. Nasr et al. (2012b) and Candelotto et
al. (2017) found that in laying hens with keel fractures the egg production, its weight and eggshell
surface area was lower than that in hens without keel fractures.
The keel is an important structural bone in birds that helps flying and breathing. Keel fracture
changes behavior and resting locations, reduces motility and affects the utilization of facilities for
laying hens (Nasr et al., 2012a, 2012b). Casey-Trott and Widowski (2016) studied the changes in
behavior of laying hens with and without keel fractures in small and large furnished cages. They
∙ 193 ∙
found that fractured keel (FK) hens spent more time resting on the perch than standing on floor
compared to the hens without FK. However, Nasr et al. (2012a) found that FK hens spent more
time sleeping on the floor, and rarely used the perch that was 100 cm above the ground.
Additionally, nesting time of FK hens was longer than normal keel (NK) laying hens (Gebhardt-
Henrich and Fröhlich, 2015). However, most of the behaviors, such as walking, jumping, sitting,
preening and comforting, have not been systematically studied. Therefore, the objective of this
study was to investigate the effects of keel fractures on behavior, welfare, production performance
and egg quality in laying hens housed individually in furnished cages.
2.1. Experimental animals and management
Total 90 healthy Lohmann white laying hens of 17 WOA were individually housed in
furnished cages, each of size 50 cm length × 70 cm width × 70 cm height, having two wooden
square perches (20 cm and 40 cm above the wire-mesh floor, and 45 cm and 25 cm away from
front wire-mesh sidewall, respectively), one closed nest box (35 cm length × 20 cm width × 25
cm height, installed on the left and rear of the sidewalls), one nipple drinker and one rectangular
feeder (installed outside of the front sidewall). The laying hens house was semi-enclosed with a
combination scheme of natural light and artificial light. Artificial light was programed for 16
hours (5:00–21:00 h), followed by 8 hours of dark, and light intensity was 18–22 lux. The house
had natural ventilation. During the entire experiment period (17–42 WOA), all laying hens had
free access to feed and water. The temperature of laying hens house was 20–28°C and the relative
humidity was 45–70% throughout the experiment. Laying hens were provided with a standard
commercial layer diet of 2800 kcal kg-1 ME and 16.08% CP.
2.2. Assessment of keel fracture
Assessment of keel damages was done by one worker according to the palpated method
(Casey-Trott et al., 2015). Palpations were performed by running thumb and index fingers down
the spine/ventral, and lateral surface edge of keel bone. Keel status can be classified as normal
keel (NK), deviated keel (DK), and fractured keel (FK). DK with the deviation of the keel bone
from a two-dimensional straight plane in either the transverse or median sagittal planes, including
bending, S-shaped, twisted, or curved keels (Casey-Trott et al., 2015). The FK with the presences
of sharp bends, shearing, fragmented sections, callus materials and/or clear bumps on the ventral
and lateral surfaces of keel bone (Casey-Trott et al., 2015). If a hen had both a fractured and
deviated keel, then it was categorized as FK. In this experiment, keel status was assessed by
palpation starting at 17 WOA, and then palpated every 5 weeks thereafter. The FK and NK hens
were marked at each palpating examination according to their keel bone status, and the incidence
of keel damage at each time-point was calculated. At the end of the experiment (42 WOA), all
hens were euthanized and dissected to validate the accuracy of palpation.
2.3. Behavioral observation
After palpation of hens at 22, 27, 32, 37 and 42 WOA, behavioral observations were
conducted for 2 consecutive days using a video recording system (Hikvision, Hangzhou, China).
On each behavioral observation day, the behaviors of hens with NK and fresh FK (n = 6 each)
were observed for 7 hours. This included 2 hours each, in morning (08:30–10:30 h), afternoon
(12:30–14:30 h), evening (16:00–18:00 h), and 30 min before and after the artificial lighting was
turned off (20:30–21:30 h). The behaviors were classified as ‘state’ and ‘event’ behaviors (Table
1). The state behaviors were collected using focal animal sampling method at 5 s interval over 7
hours observation period, and each behavior was counted as the percentage of the total observation
time. For event behaviors, continuous recording and One-Zero sampling method was used to
sample each behavior over 7 hours observation period, and each event behavior was expressed as
the total numbers sampled.
∙ 194 ∙
Table 1. Behavioral categories and definitions

Behaviors Definitions
State behaviors
Standing Laying hen standing on feet, legs extended, no movement of the body but
with eyes open. Head is erect or relaxed posture
Feeding Laying hen directing its head in feed trough and carrying out pecking,
ingesting or eating feed, once or repeatedly. Including standing breaks of
≤ 5s followed by resumption of behavior
Walking Moving more than three paces in one direction, and head erect
Sitting Laying hen sitting with abdomen touching on the ground, wings closed,
eyes opened and without any extra activity
Perching Laying hen with two feet on a perch for more than 3s, including standing,
sitting, and walking
Nesting Including standing, sitting, laying, preening in the nest box
Event behaviors
Drinking Laying hen directing its head to nipple drinker, pecking and swallowing
water. Including standing breaks of ≤ 5s followed by resumption of
behavior
Preening Laying hen using its beak to gently peck, nibble, comb, preen feathers, or
using claw to scratch its wings or head
Comforting Including body shaking, tail shaking, wing raising or stretching, leg
stretching, simultaneous wing and leg stretching, and sham dustbathing
Cage-pecking Pecking or scratching perch, nest box, or floor of cage
Jumping Laying hen jumping up or down from the perch
2.4. Production performance and egg quality
To compare the performance of the NK and FK hens, the hens from three different age groups
(27–32, 32–37, and 37–42 WOA) were used as the focal birds. Firstly, the number of eggs laid,
dirty eggs, broken eggs, and feed intake from all hens (except FK hens) were noted at 27 WOA.
These laying hens were then palpated at 32 WOA to determine NK or FK. The performance data
and feed intake of the NK and FK hen groups at 27–32 WOA was measured based on the
assessment of keel status at 32 WOA. Namely, if a hen had NK at 32 WOA, the performance data
and feed intake of this bird at 27–32 WOA period were belonged to the NK group. At 32 WOA,
if a hen had FK, its corresponding data was categorized into the FK group. Similarly, hens had
NK or FK at 37 and 42 WOA, their performance data and feed intake at 32–37 and 37–42 WOA,
were also classified into the FK or NK group, respectively. Following the day after behavioral
observation at each time-point, 10 eggs were collected from other NK and FK hens (n = 10, each)
to analyze its quality in terms of egg weight (g), eggshell weight (g), egg-shape index, yolk weight
(g), eggshell thickness (mm), eggshell strength (kg cm-3), albumen height (mm) and Haugh unit.
2.5. Assessment of welfare
After each palpation at 27, 32, 37 and 42 WOA, new hens from NK and FK groups (n = 10
except in KF group at 42 WOA, where n = 6) were randomly selected to assess welfare indexes.
2.5.1. Tonic immobility test (TI)
The selected NK and FK hens were brought into a quiet room and placed lying on their backs
in a U-shaped wooden groove. To induce TI, a tester gently pressed the hen’s head for 15 s with
one hand and their chest with another. To ensure the success of TI, the testing hens had to stay
still for at least 10 seconds. If each trial was successful, the TI duration of the bird kept in lying
until it turned over and stood up was recorded. Each time the bird was induced 3 times. If the hens
were stationary for 20 minutes, a TI duration of 20 minutes was recorded.
∙ 195 ∙
2.5.2. Feather cover

Feather cover score was taken on eight parts of the body: head, neck, back, tail, wings, chest,
cloaca and legs. The conditions of feather cover were scored from 0 to 3; 0: best feather condition
with no damage, 1: having slight damage but skin well covered, 2: feather with moderate damage
and area skin of <1 cm × 1 cm exposed, 3: feather having severe damage and area skin of >1 cm
× 1cm exposed. Total feather cover scores for all parts were calculated for each bird, the lowest
(best feather) score was 0 and the highest (worst feather) score was 24.
2.5.3. Foot pad and toe health
Foot pad condition and toe health was assessed for hyperkeratosis, dermatitis, and bumble
foot. Hyperkeratosis of both foot pad and toe was scored as 0 (absent) or 1 (present). Foot pad and
toe dermatitis were scored as 0–2, namely, 0: perfect foot pad and toe, with no damage, 1: slight
dermatitis on foot pad and toe with small epithelium lesion (<0.2 cm), 2: severe dermatitis on foot
pad and toe with a large epithelium lesion (>0.2 cm). Bumble foot was scored as absent (0) or
present (1). The mean score was obtained from all hens for the data analysis.
Statistical analysis was performed using SPSS 22. Data were tested for normal distribution
using the Kolmogorov-Smirnov test. The differences in behaviors, egg quality and welfare of NK
and FK hens at each time-point were analyzed using one-way ANONA with Duncan’s multiple
comparison. Production performance and feed intake data were analyzed by repeated measures
analysis with multivariate ANOVA under general linear model to analyze the difference using the
tests of between-subjects factor (Group) and with-subjects factor (Time). The results were
expressed as mean ± SEM, and differences of P ≤ 0.05 was considered statistically significant.
3.1. Assessment of keel bone fracture
As shown in Figure 1, the keel bones of all hens were normal at the initiation of the study.
However, at 22 WOA, 26.7% of hens had DK bones, but no FK. At 27 to 42 WOA, DK remained
generally stable at 21.1-24.4%, while FK consistently increased from 16.7% to 54.4%, and NK
consistently decreased from 61.1% to 21.2%. Therefore, with increase of age, the incidence of
keel fracture also increased from 16.7% at 27 WOA up to 54.4% at 42 WOA.
Figure 1. Percentages of laying hens with normal keel (NK), deviated keel (DK) and fractured
keel (FK) bones at six time-points.
Some researchers also suggest that the incidence of keel fracture increase with age of hens
(Rufener et al., 2018). Consistent with these results, it was found that keel fracture did not occur
at the onset of the study (17–22 WOA), however, the incidences increased to over 50% at 42
WOA. This may indicate that hens lose the calcium in their bones for eggshell mineralization at
∙ 196 ∙
the high peak of lying (Whitehead, 2004). As the laying period continues, the reduction of calcium
content in bone increases bone fragility resulting in a higher risk of fractures in hens.
3.2. Behavioral observation
As shown in Table 2, compared with the NK hens, the standing and sitting increased (P<0.05),
while feeding, walking, perching, and jumping decreased in FK hens (P<0.05). Overall, it was
found that keel fracture impairs the mobility of laying hens. Standing and sitting behavior
increased and the feeding, walking, perching and jumping behavior decreased in FK hens. This
finding was consistent with the report of Nasr et al. (2012b), indicating that FK hens spent more
time sitting on the floor and rarely used perches. Furthermore, keel fracture decreased the odds of
perch or nest access and increased latencies to jump from perches in laying hens housed in non-
cage systems (Nasr et al., 2012a). These results suggested that keel fracture causes pain and
impairs the mobility, thus may result in increased sitting and decreased perching.
Table 2. Mean (±SEM) values of behaviors of normal keel (KN) and fractured keel (FK) hens.
Item NK FK
State behaviors
Standing 28.60b (0.67) 34.22a (1.02)
Feeding 35.21a (0.83) 31.03b (1.31)
Walking 8.38a (0.45) 6.05b (0.29)
Sitting 7.63b (0.23) 10.53a (1.14)
Perching 13.07a (0.99) 10.03b (0.46)
Nesting 7.03 (0.60) 8.07 (0.54)
Event behaviors
Drinking 177.25 (9.92) 169.75 (16.94)
Preening 124.13 (9.85) 119.88 (11.20)
Comforting 42.63 (3.83) 32.50 (3.07)
Cage-pecking 59.00 (3.52) 54.63 (3.97)
Jumping 23.88a (2.91) 13.13b (1.19)
a, b Mean values with superscripts within the same row differ significantly, P ≤ 0.05.
Table 3. Mean (±SEM) values of performance and feed intake between normal keel (NK) and
fractured keel (FK) laying hens at three periods.
27-32 WOA 32-37 WOA 37-42 WOA
Item (unit)
NK FK NK FK NK FK
Number of birds 42 29 28 40 19 49
Number of eggs 1388 931 949 1318 651 1645
94.44a 91.72b 96.84a 94.14b 97.90a 95.92b
Egg production (%)
(0.75) (0.88) (1.03) (0.97) (0.95) (0.86)
2.32 3.69 2.12 2.94 2.00 2.99
Dirty-egg rate (%)
(0.53) (0.64) (0.69) (0.58) (0.77) (0.48)
0.97b 2.42a 1.19 1.90 0.93 2.31
Broken-egg rate (%)
(0.31) (0.38) (0.54) (0.45) (0.62) (0.39)
116.24a 112.21b 121.76a 119.45b 126.11a 123.67b
Daily feed intake (g)
(1.41) (1.14) (0.88) (0.72) (0.92) (0.74)
3.3. Measurement of production performance

The production performance of NK and FK laying hens at three experimental periods (27–32,
32–37, and 37–43 WOA) were showed in Table 3. During these periods, the mean egg production
and daily feed intake of FK hens significantly decreased compared with NK hens (P<0.05). The
percentage of broken egg from FK hens was significantly increased at 27–32 WOA (P<0.05);
∙ 197 ∙
however, there was no significant difference observed at 32–37 and 37–42 WOA (P >0.05). There
was no significant difference in dirty-egg rate between NK and FK hens at each experimental
period (P >0.05). Overall, keel fracture decreased the egg production and daily feed intake during
experimental period. Thiruvenkadan et al. (2010) indicated that keel fracture causes pain and
stress, and disrupts the hormones required for ovulation and decreases the responses of granulosa
cells to luteinizing hormone, which could reduce egg production. Additionally, the fact that FK
hens had lower egg production was found by Rufener et al. (2018), who speculated that it might
be due to the enhancement of bone metabolism and the redistribution of available resources to
fracture healing. Likewise, a major reason for the decrease in egg production of FK hens may be
the reduction of feed behavior and intake as seen in this study.
3.4. Measurement of egg quality
As showed in Table 4, at 27 WOA, the egg weight, eggshell thickness, eggshell strength, and
egg surface area of FK hens were less than those of NK hens (P<0.05). At 27 WOA, the eggshell
weight of FK hens tended to be lower than NK hens (P = 0.051), but the difference between the
groups was not significant. Haugh unit from FK hens was lower than that of NK hens at 37 WOA
(P<0.05). Compared to NK hens, FK hens had lower eggshell thickness, strength, and weight at
32, 37 and 42 WOA (P<0.05).
Table 4. Mean (±SEM) values of egg quality from normal keel (NK) and fractured keel (FK)
laying hens at four time-points.
27 WOA 32 WOA 37 WOA 42 WOA
Item (unit)
NK FK NK FK NK FK NK FK
56.89a 54.88b 58.94 57.98 52.42 51.41 62.12 61.27
EW (g)
(0.66) (0.59) (1.81) (0.57) (0.62) (1.63) (1.18) (2.01)
77.00 76.75 77.46 77.21 77.64 78.05 77.38 76.25
ESI (%)
(0.07) (0.09) (0.50) (1.02) (0.61) (0.61) (1.06) (2.25)
0.39a 0.35b 0.41a 0.37b 0.39a 0.37b 0.41a 0.37b
EsT (mm)
(0.02) (0.02) (0.04) (0.01) (0.08) (0.05) (0.01) (0.02)
7.40 7.02 8.10a 7.64b 7.31a 6.78b 8.70a 8.14b
EsW (g)
(0.11) (0.15) (0.14) (0.07) (0.17) (0.11) (0.07) (0.08)
4.63a 4.11b 4.59a 4.01b 4.38a 4.00b 4.86a 4.29b
EsS (kg cm-3)
(0.19) (0.11) (0.07) (0.18) (0.15) (0.09) (0.18) (0.13)
68.86a 67.13b 69.79 70.56 65.00 64.07 73.26 72.52
ESA (cm2)
(0.56) (0.51) (0.48) (1.52) (0.55) (1.44) (0.98) (1.68)
97.03 94.65 92.55 89.87 92.26a 90.13b 91.94 91.01
HU
(1.25) (1.94) (1.83) (0.85) (0.83) (0.54) (1.18) (1.46)
14.56 14.17 15.38 14.61 14.58 15.33 16.72 16.16
YW (g)
(0.49) (0.36) (0.25) (0.43) (0.42) (0.23) (0.34) (0.40)
9.35 8.78 7.97 8.55 8.08 7.81 8.56 8.35
AH (mm)
(0.26) (0.41) (0.17) (0.41) (0.16) (0.10) (0.27) (0.32)
Abbreviation: EW = egg weight; ESI = egg-shape index; EsT = eggshell thickness; EsW =
eggshell weight; EsS = eggshell strength; ESA = egg surface area; HU = haugh unit; YW = yolk
weight; AH = albumen height; WOA = weeks of age.
Overall, keel fractures mainly affected the external, but not the internal egg quality. Eggshell
quality like thickness and strength was strongly associated with its calcification. Candelotto et al.
(2017) found that an elevated susceptibility to keel fractures was corelated to thinner eggshells
and lower egg breaking strength in laying hens. Nasr et al. (2012b) found FK hens had reduced
eggshell weight and had a tendency to lay lighter eggs. Similarly, it was found in this study that
eggshell weight, thickness, and strength of FK hens was lower than NK hens. Generally, a hen
∙ 198 ∙
needs to absorb approximately 3 g of calcium per day to synthesize normal eggshell. However,
after keel fracture, this calcium may be used for fracture healing and calluses formation instead
of eggshell formation. Therefore, eggshell thickness and strength of FK hens were decreased.
Furthermore, the lower daily feed intake and feeding behavior observed in FK hens could have
directly contributed to the decrease in calcium resources for eggshell formation.
3.5. Assessment of welfare
As shown in Table 5, FK hens had a higher TI value at all time-points (P<0.05), and a
decreased body weight at 27 and 42 WOA (P<0.05). However, there was no significant difference
in feather cover score, foot pad and toe hyperkeratosis scores, and toe dermatitis score between
NK and FK hens during the experimental periods (P>0.05), except in FK hens at 37 WOA
showing a greater score in foot pad dermatitis (P<0.05). Furthermore, the incidence of bumble
foot was not found in NK and FK hens. The high incidences of keel fracture were positively
related to foot injuries in barn and organic production systems (Riber et al., 2016). Gebhardt-
Henrich and Fröhlich (2015) showed that hens with bumblefoot on both feet had a keel fracture
in aviary system. Thus, foot pad and toe health of hens may somehow be related to the occurrences
of keel fracture in group-housing systems or may be related to reduced perch use in hens with keel
damage. There were no differences in hyperkeratosis and dermatitis on both feet and toes in this
study. This may be related to the hens being housed individually in furnished cages. In non-cage
and organic production systems, FK hens had severe claw injuries associated with litter quality
and housing resources (Hocking and Wu, 2013). The present study did not provide litter for hens
in furnished cages, and there was no difference observed in the foot pad and toe injuries between
NK and FK hens. Thus, the litter could be a major cause for the foot pad and toe injuries.
Assessment of feather condition was also one of the important indicators of welfare in laying hens,
and a greater feather quality represents a higher welfare state, and vice versa. However, there was
no difference in feather quality observed in NK and FK hens in this study. This perhaps suggests
that keel fracture and feather quality were not related in this study of hens housed individually in
furnished cages.
Table 5. Mean (±SEM) values of welfare assessment between normal keel (NK) and fractured
keel (FK) laying hens at four time-points.
27 WOA 32 WOA 37 WOA 42 WOA
Item (unit)
NK FK NK FK NK FK NK FK
1.56a 1.46b 1.62 1.55 1.65 1.57 1.69a 1.58b
BW (kg)
(0.05) (0.10) (0.02) (0.03) (0.03) (0.08) (0.04) (0.04)
5.64b 7.08a 6.35b 7.83a 5.30b 7.52a 7.30b 9.82a
TI (min)
(0.28) (0.50) (0.47) (0.41) (0.26) (0.27) (0.83) (0.80)
0.75 1.38 1.00 1.39 1.38 1.63 1.01 1.63
FC (score)
(0.25) (0.26) (0.27) (0.32) (0.26) (0.18) (0.27) (0.38)
0.25 0.25 0.25 0.38 0.38 0.38 0.38 0.50
TH (score)
(0.16) (0.16) (0.16) (0.18) (0.18) (0.18) (0.18) (0.19)
0.25 0.38 0.38 0.50 0.38 0.50 0.50 0.88
TD (score)
(0.25) (0.18) (0.21) (0.18) (0.26) (0.19) (0.27) (0.29)
FPH 0.13 0.13 0.25 0.25 0.38 0.25 0.25 0.38
(score) (0.13) (0.13) (0.16) (0.16) (0.18) (0.16) (0.16) (0.18)
FPD 0.50 0.88 0.88 1.25 0.75b 1.63a 1.25 1.75
(score) (0.19) (0.35) (0.23) (0.25) (0.31) (0.26) (0.31) (0.25)
Abbreviation: BW = body weight; TI = tonic immobility; FC = feather cover; TH = toe

hyperkeratosis; TD = toe dermatitis; FPH = foot pad hyperkeratosis; FPD = foot pad dermatitis;
WOA = weeks of age.
∙ 199 ∙
4. Conclusions
In conclusion, the results demonstrated that keel fracture caused changes in state behaviors
(i.e., standing, feeding, walking, sitting, and perching), but not event behaviors (i.e., drinking,
preening, comforting, and cage-pocking). Keel fracture led to a longer duration of TI in laying
hens, and decreased their daily feed intake, egg production, broken-egg rate, and eggshell quality.
Therefore, keel fracture can change parts of behavior and reduce welfare, performance and
external egg quality of laying hens.
Acknowledgements
This work was supported by the National Natural Science Foundation of China (grant numbers
31672466 and 31972608).
References
Candelotto, L., A. Stratmann, S.G. Gebhardt-Henrich, C. Rufener, T. van de Braak, M.J. Toscano, 2017.
Susceptibility to keel bone fractures in laying hens and the role of genetic variation. Poultry Science. 96,
3517–3528.
Casey-Trott, T.M., J.L.T. Heerkens, M. Petrik, P. Regmi, L. Schrader, M.J. Toscano, T. Widowski, 2015.
Methods for assessment of keel bone damage in poultry. Poultry Science. 94, 2339–2350.
Casey-Trott, T.M., T.M. Widowski, 2016. Behavioral differences of laying hens with fractured keel bones
within furnished cages. Frontiers in Veterinary Science. 3, 42.
Gebhardt-Henrich, S.G., E.K. Fröhlich, 2015. Early onset of laying and bumblefoot favor keel bone
fractures. Animals. 5, 1192–1206.
Heerkens, J.L.T., E. Delezie, T.B. Rodenburg, I. Kempen, J. Zoons, B. Ampe, F.A.M. Tuyttens, 2015.
Risk factors associated with keel bone and foot pad disorders in laying hens housed in aviary systems. Poultry
Science. 95, 482–488.
Hocking, P.M., K. Wu, 2013. Traditional and commercial turkeys show similar susceptibility to foot pad
dermatitis and behavioural evidence of pain. British Poultry Science. 54, 281–288.
Nasr, M.A., C.J. Nicol, J.C. Murrell, 2012a. Do laying hens with keel bone fractures experience pain?
PloS One. 7, e42420.
Nasr, M.A., J. Murrell, L.J. Wilkins, C.J. Nicol, 2012b. The effect of keel fractures on egg-production
parameters, mobility and behaviour in individual laying hens. Animal Welfare. 21, 127–135.
Riber, A.B., L.K. Hinrichsen, 2016. Keel-bone damage and foot injuries in commercial laying hens in
Denmark. Animal Welfare. 25, 179–184.
Rufener, C., S. Baur, A. Stratmann, M.J. Toscano, 2018. Keel bone fractures affect egg laying
performance but not egg quality in laying hens housed in a commercial aviary system. Poultry Science. 98,
1589–1600.
Stratmann, A., E.K. Fröhlich, S.G. Gebhardt-Henrich, A. Harlander-Matauschek, H. Würbel, M.J.
Toscano, 2015. Modification of aviary design reduces incidence of falls, collisions and keel bone damage in
laying hens. Applied Animal Behaviour Science. 165, 112–123.
Thiruvenkadan, A.K., S. Panneerselvam, R. Prabakaran, 2010. Layer breeding strategies: an overview.
Worlds Poultry Science Journal. 66, 477–502.
Wei, H.D., C. Li, H.W. Xin, S. Li, Y.J. Bi, X. Li, J.H. Li, R.X. Zhang, J. Bao, 2019. Keel Fracture Causes
Stress and Inflammatory Responses and Inhibits the Expression of the Orexin System in Laying Hens.
Animals. 9, 804.
Whitehead, C.C., 2004. Overview of bone biology in the egg-laying hen. Poultry Science. 83, 193–199.
Wilkins, L.J., S.N. Brown, P.H. Zimmerman, C. Leeb, C.J. Nicol, 2004. Investigation of palpation as a
method for determining the prevalence of keel and furculum damage in laying hens. Veterinary Record. 155,
547–549.
∙ 200 ∙
A Simulation Model for Detection of Sow Lameness and Estimation

of Ground Reaction Force
Xiaojie Yan a, Qiang Zhang a,*, Laurie Connor b, Nicolas Devillers b, Kristopher Dick a
a
Department of Biosystems Engineering, University of Manitoba, Winnipeg, MB R3T 2N2, Canada
b
Department of Animal Science, University of Manitoba, Winnipeg, MB R3T 2N2, Canada
c
Agriculture and Agri-Food Canada, Sherbrooke Research and Development Centre,
Sherbrooke, Québec, J1M 0C8, Canada
* Corresponding author. Email: Qiang.Zhang@umanitoba.ca
Abstract
Leg problems of sows, such as lameness, are common concerns in the swine industry for
animal welfare and productivity. This paper presents a two-dimensional stick model to simulate
sow walking on concrete floors and detect lameness. Data were extracted from recorded video
images of sows walking on a concrete floor to establish motion equations in the simulation model.
A step of walking was modelled as three consecutive phases of motion: acceleration of foot
leaving the ground; constant speed of foot swinging; and deceleration of foot touching the ground.
One gait cycle was modelled in a time sequence to match the movement phases of all joints in the
stick model. The simulated joint movements were used to determine a selected set of gait
parameters, including the stride length and stance time for fore- and hindlimbs, back angle,
diagonality and walking speed. These gait parameters were then used to detect sow lameness. To
determine the GRF (ground reaction force) on sow feet, the stick model was simplified as a rigid
trunk connected by two joints (mass) on massless legs, and the acceleration and time-dependent
mass were then used to calculate the vertical GRF. The results showed that an accuracy greater
than 92% for lameness detection was achieved. The calculated GRF’s were in good agreement
with the double-hump GRF profile reported in the literature.
Keywords: Simulation model, sow gait, lameness detection, ground reaction force
1. Introduction
Leg problems (e.g., lameness) are major welfare issues in the swine industry (Kirk et al.,
2005). Understanding the biomechanics of pig walking is critical for the lameness assessment,
thus improving animal welfare and productivity. Unsymmetrical gait is usually considered as an
indicator of lameness for pigs and other quadrupeds (Griffin et al., 2004). Therefore, gait analysis
is commonly used in lameness assessment.
Visual locomotion scoring is a common method of gait assessment in commercial swine barns
but is subjective and not very reliable for detecting subtle lameness (Channon et al., 2009).
Kinematics analysis with camera systems and kinetics using force plates are two objective
techniques to analyze pig gait and lameness (Clayton and Schamhardt, 2001; Conte et al., 2014).
However, most studies of lameness assessment on pigs have been focused on the abnormal gait
using the measurement techniques, and few have attempted to simulate the gait with consideration
of the inter-limb coordination.
The stick model is a sophisticated method to simulate human motion (Aggarwal and Cai,
1999; Mikić et al., 2003; Chan et al., 2016), which has been adopted by several researchers to
simulate gait of quadrupeds (Lacquaniti et al., 1999). Catavitello et al. (2015) used a stick model
to represent the segments and markers of dogs to study the kinematic coordination of different
gait (walk, trot, gallop, and swim). The inter-segmental coordination patterns of the dogs were
studied and compared during the four forms of quadrupedal locomotion. Specific studies about
gait simulation models of pigs and other farm animals are still scarce. Yan et al., (2019) proposed
a 2D stick model to simulate sow walking on concrete floors and detect lameness. In their model,
the key kinematical variables (acceleration and speed) of pig motion of walk were simulated, but
forces were not considered.
∙ 201 ∙
Understanding the biomechanics of pig walking with consideration of the forces involved is
critical to minimize slip injury and foot abrasions of animals (von Wachenfelt et al., 2009a;
Applegate et al., 1988). A key force involved is the ground reaction force (GRF), which is exerted
by the ground on the animal body in contact with it (Scalley, 2012). Force plates are commonly
used to measure GRF for both human and animals. von Wachenfelt et al. (2009a) used a force
plate to record the horizontal and vertical ground reaction forces, as well as the time of peak
vertical forces when pig walked on the clean and fouled concrete floors. They found that the pigs
reduced the horizontal GRF more than the vertical GRF to adapt the gait to the fouled concrete
floors. Another method of GRF measurement is motion capture which can be used to determine
the acceleration, from which the GRF is estimated using Newton's 2nd Law. A challenge of using
the motion capture method is how to define/quantify the mass when calculating the force from
acceleration because the body mass of an animal is generally unevenly distributed among different
feet and the distribution varies during walking. Clauser et al. (1969) discussed segmental masses
and positions of segmental mass centers of a human body, however, there are few comprehensive
studies about pigs. Bobbert et al. (1991) calculated the vertical GRF from positional data for the
landing phase in running of human using mass property data from Clauser et al. (1969). Some of
the human GRF studies (Corazza et al., 2010; Scalley, 2012) used laser scan or markerless motion
capture system to get the positions and volume of human body, which was then used to estimate
the body mass distribution by assuming a constant density of human body.
The aims of this study were: (1) apply and optimize a stick model of sow walking simulation
for lameness detection; (2) use the model to estimate ground reaction forces.
2.1. Two-dimensional stick model
A two-dimensional stick model proposed by Yan et al. (2019) was used in this study. The
model was based on the video images of the markers on sow’s body (Figure 1a) reported by
Devillers et al (2019). A sow was represented by 13 joints and 12 sticks in the stick model and
the motions of the sow body were simulated by coordinated movements of these joints and sticks
during walking (Figure 1b).
Figure 1. Sow’s body skeleton and approximation by joints and sticks. (a) Illustration of
positions of the nine reflective markers on the sow in recorded video images by Devillers et al.
(2019); (b) A stick model of sow body with 13 joints and 12 sticks.
Each step of walk was modelled as three consecutive phases of movements: acceleration of
foot leaving the ground; constant speed of foot swinging; and deceleration of foot touching the
ground. A gait cycle for each limb was defined as the time interval between two successive foot
lift, as indicated by the movements of joints 1, 5, 6 and 10. A gait cycle consisted of several time
∙ 202 ∙
steps based on the movement characteristics and the movements of each joint at a time step in x-
and y- directions were calculated as time functions (Yan et al., 2019).
x  x0  f (t )
(3)
y  y0  g (t )
where, x and y are the x- and y- coordinates of a joint, respectively; x0 and y0 are the initial
coordinates defining the start location of the joint at each step; 𝑓(𝑡) and 𝑔(𝑡) represent the
displacement in x and y direction, respectively, as a function of time.
A phase of movement could be mathematically described by a 2 nd order polynomial for
constant acceleration and deceleration or a linear function for constant speed to fit the data. The
acceleration and deceleration values were then determined by differentiating the 2 nd order
polynomial function twice with respect to time. The constant speed was determined by
differentiating the linear function once.
The joint coordinates predicted by equation 1 at each time step were then used to calculate a
selected set of characteristic gait parameters, including the stride length and stance time for fore-
and hindlimbs, diagonality, back angle and walking speed. Based on the comparisons of the
predicted gait parameters between the lame and healthy sows, the criteria were established for
using the predicted gait parameters to detect sow lameness (Yan et al., 2019).
2.2. Video data analysis
The model parameters were determined from the measured movements of sows by digitizing
the recorded video images. Three sets of video data were used respectively for: (1) model
development, including 81 videos of non-lame sows; (2) model validation, including 15 videos of
non-lame sows; and (3) model evaluation for lameness detection, including 75 videos of lame
sows. Still images were extracted from the videos at 30 Hz using GIMP (Version 2.10.10, the
GIMP Development Team). The x- and y-coordinates of each reflective marker in the images
were determined using MATLAB (Version 7.1.0.183 (R14), the MathWorks, Inc., Natick, MA,
USA), following a calibration procedure proposed by Clayton and Schamhardt (2001). The
coordinate data of the left and right limbs obtained from the videos were pooled together due to
symmetrical gait. Once the x- and y-coordinates were extracted from the images, they were plotted
against time to quantify the movements of each marker as functions of time (Yan et al., 2019).
2.3. Estimation of ground reaction force
The ground reaction force acting on a limb consisted of a vertical and a horizontal component
and the fore- and hindlimbs were considered as single rigid segments during the estimation of
GRF (Figure 2). To estimate mass distribution, the stick model in Figure 1b was simplified as a
rigid trunk connected by two point-masses and each point-mass connected to two massless limbs.
Based on the method developed by van Gurp et al. (1987), the mass carried by a limb was assumed
to increase linearly to a maximum mass ms, then stabilized at ms during the single support of the
corresponding biped, and decrease to zero linearly at the end of the stance phase (Figure 3). Past
research of GRF using force plates for pigs showed that the forelimbs carry more weight (60%)
than the hindlimbs (40%) (Merkens et al., 1986; Devillers et al., 2020; von Wachenfelt et al.,
2009a). Therefore, the maximum mass ms was determined as 0.6 and 0.4 times the pig’s total body
mass for the forelimbs (mF) and hindlimbs (mH), respectively. The mass variation on a limb during
the stance phase was be calculated as follows:
t
m(t )  ms   when 0≤t˂t1
 t1 
m(t )  ms when t1≤t˂t2 (2)
∙ 203 ∙
 t t 
m(t )  ms 1  2  when t2≤t≤t1+ t2
 t1 
where ms is the maximum mass carried by a limb during the stance phase: mF and mH for the
forelimbs and hindlimbs, respectively; t1 is the time from a foot landing on ground to single
support of the corresponding biped; t2 is the time that a foot is in contact of the ground before
lifting off.
Figure 2. Forces acting on a limb (m(t) is the mass carried by a limb; g is the gravitational
constant; Fv is the vertical ground reaction force; Fh is the horizontal ground reaction force).
Figure 3 Variation of mass carried by a limb during walk of a quadruped (ms is the maximum
mass; t1 is the time before single support during the stance phase; t2 is the time before liftoff
during the stance phase).
The limb movements in the model were still considered as two dimensional in sagittal plane.
The vertical GRF could be determined using Newton's 2nd Law:
Fv  m(t )  ( g  a) (3)
where Fv is the vertical GRF on a limb; m(t) is the time-dependent mass carried by the limb; g is
the gravitational constant; and a is the instantaneous acceleration of the limb in the vertical
direction. The instantaneous acceleration data were calculated by differentiating the y-coordinates
extracted from the kinematics videos with respect to time at interval of 33.3 ms, and then were
digitally smoothed using a moving-average filter.
3.1. Lameness detection
Sixty lame sows were tested to verify the proposed simulation model for lameness detection
of sows. The selected gait parameters with corresponding deviation thresholds (Table 1) were
used to detect lameness. If the percentage difference between predicted value (for non-lame sows)
and measured value exceeded the deviation threshold of the corresponding gait parameter, the
sow was considered as lame.
∙ 204 ∙
Table 1. Percentage deviation thresholds of gait parameters.

Stride Stance Stride Stride
Gait Walking Back
Diagonality length time length length
parameters speed angle
forelimb forelimb hindlimb hindlimb
Threshold (%) 8 3 10 9 10 7 10
The accuracies using the seven gait parameters listed in Table 1 to detect lameness
individually are shown in Table 2. The highest accuracy of 93% was obtained using the forelimb
stance time parameter and the lowest accuracy 75% was from the back angle parameter. The
reason for variable accuracies of different parameters was that lameness was not necessarily the
same for all sows, and the way gait was affected could vary depending on the cause of lameness,
and which legs and the number of legs were affected. Therefore, combinations of multiple
parameters might provide a more accurate indicator to detect lameness for whatever the cause or
the severity.
Table 2. Lameness detection accuracy of the selected gait parameters in the model.
Gait parameters Accuracy Average weight
Walking speed 90% 10±16
Back angle 75% 2±2
Diagonality 85% 7±5
Forelimb
Stride length 82% 9±16
Stance time 93% 5±4
Hindlimb
Stride length 77% 10±10
Stance time 90% 6±4
A weighting method was then used to optimize the combination of the gait parameters for
lameness detection. The weight was calculated as the difference between the deviation and the
threshold in Table 1. For example, if a deviation of diagonality was 12% while the threshold was
10%, the weight of diagonality would be 2 (12%–10%). It meant a weight could reflect its
sensitivity to lameness. However, the calculated weight in Table 1 showed the large weight did
not necessarily result in high accuracy, while a low weight could (e.g., 5 for the hindlimb stance
limb associated with the high accuracy of 93%). This meant that a single parameter probably was
not the best model indicator for lameness detection. Therefore, a two-parameter combination
approach was used to improve the reliability of the model in terms of lameness detection. The
weight index for each combination was calculated as the sum of any two parameter weights
divided by the sum of seven parameter weights. All 21 parameter combinations were then used to
detect lameness of 60 sows and the accuracies are shown in Figure 4.
The combination of the walking speed and the forelimb stance time had the highest weight
index (60%) and the highest accuracy (98%). The combination of the diagonality and the forelimb
stride length showed lowest accuracy of 92% with a low weight index of 3%. It was apparent that
the two-parameter combinations improved the accuracy of lameness detection. According to the
high accuracy and high weight index, the WS+STF combination was considered as the best
indicator to detect lameness (Figure 4). The accuracies achieved in this study were close to the
model of Poursaberi et al. (2010). They developed an automatic system for lameness assessment
of dairy cattle with an accuracy of 96% by analyzing the back curvature based on 2D side view
images. Zhu and Zhang (2010) built a stick model of pig limbs to detect pig lameness by assessing
the swing angle of carpal and tarsal joints and the model achieved an accuracy of 90%. Although
the model proposed in the current study has not been tested on large datasets, it showed a
promising ability to detect sow lameness.
∙ 205 ∙
Figure 4. Accuracy of lameness detection using different two-parameter combinations (WS:

walking speed; BA: back angle; D: diagonality; SLF: stride length – forelimb; STF: stance time
– forelimb; SLH: stride length – hindlimb; STH: stance time - hindlimb)
3.2. Estimation of GRF
Acceleration obtained from the video data showed that a peak vertical acceleration occurred
at the moment the sow foot landed the floor. The average peak acceleration during the touchdown
phase of 20 healthy sows were summarized in Table 3 and compared with the data of 20 lame
sows with visual scores of 1–3. The forelimb had peak acceleration of 24 m s-2 and 17 m s-2 for
non-lame and lame sows, respectively; the hindlimb showed acceleration of 20 m s-2 and 14 m s-
2 for non-lame and lame sows, respectively. The peak vertical accelerations of non-lame sows
were significantly greater than the lame sows. Several studies showed the lame pigs tended to
avoid putting weight on the lame leg (Thorup et al., 2007; von Wachenfelt et al., 2009b). It was
also observed that the ratio of the forelimb to hindlimb acceleration was about 1.2 for both non-
lame and lame sows, which was lower than the weight distribution (60% vs. 40% or 1.5) between
the fore- and hindlimbs typically reported for non-lame sows. The GRF of non-lame and lame
sows were not compared due to unknown weight-shifting of lame sows. The lame sows could
have various lameness problems and the weight distribution on limbs would vary depending on
the cause of lameness and which limbs and the number of limbs were affected. Therefore, the 60%
vs. 40% weight distribution for non-lame sows could not be used for the approximation of weight
distribution for lame sows.
Table 3. The average peak vertical acceleration determined from the video data (n=20 for non-
lame and lame sows each).
Average peal vertical acceleration (standard deviation) (m s-2)
Limb
Non-lame sows Lame sows
Forelimb 24 (3) 17 (4)
Hindlimb 20 (3) 14 (4)
The predicted vertical GRF for non-lame sows followed a double-hump pattern as other
quadrupeds, such as dogs (Polet and Bertram, 2019) and horses (van Gurp et al., 1987; Clayton
and Schamhardt, 2001) (Figure 5). The first peak (8.1 N kg-1 and 7.0 N kg-1 for forelimb and
hindlimb, respectively) in the vertical GRF curve reflected the force of touchdown phase during
the gait cycle and the second peak (7.5 N kg-1 and 5.5 N kg-1 for forelimb and hindlimb,
respectively) represented the force when the limb lifted off from the ground.
The vertical GRF remained relatively constant when the sow foot was fully in contact with
the floor. Generally, the vertical GRF of the hindlimb showed smaller values than the forelimb,
∙ 206 ∙
which may explain why leg problems occur more frequently in the forelimbs than in the hindlimbs
of pigs (Jørgense, 2000).
Figure 5. A typical example of vertical GRF exerted by the forelimb and hindlimb of a sow
walking on concrete floor.
The vertical GRF values in this study were in good agreement with the experimental results
of von Wachenfelt et al. (2009a). They used a force plate to record the force during the walking
of pigs and found that the peak vertical GRF of forelimb and hindlimb were 8.44 N kg-1 and 6.06
N kg-1, repectively. Thorup et al. (2007) measured GRF using a force plate when pigs walked on
concrete floors and reported the average peak vertical GRF of 5.63 N kg-1 and 4.43 N kg-1 for
forelimb and hindlimb, respectively, smaller than this study. The differences in GRF might be due
to the animal size (weight), walking speed and floor conditions (von Wachenfelt et al., 2009a).
4. Conclusions
A 2D stick model consisting of 13 joints and 12 sticks was used successfully for lameness
detection of sows. The sow was considered lame if the deviation of the measured value(s) from
the predicted value(s) for non-lame sows for a specific gait parameter(s) exceeded the
threshold(s). The model detection of lameness had an accuracy greater than 92%. A simplified
model consisting of a rigid trunk connected by two point-masses on massless limbs was adequate
in estimating the vertical GRF (ground reactions force). The estimated vertical GRF showed a
double-hump profile, which was in agreement with the results reported in the literature.
Acknowledgements
The authors would like to acknowledge the funding support by Swine Innovation Proc of
Canada and thank Don Chaput and Archimedes Isit, the staff of the Glenlea Research Station,
University of Manitoba for help in the technical support. Authors also wish to thank Lindsey
Lippens and Marjolaine St-Louis for recording the kinematics videos during this study.
References
Aggarwal, J., Q. Cai, 1999. Human motion analysis: A review. Computer Vision and Image
Understanding. 73, 428–440.
Applegate, A., S, Curtis, J. Groppel, J. McFarlane, and T. Widowski, 1988. Footing and gait of pigs on
different concrete surfaces. Journal of Animal Science. 66 (2), 334–341.
Bobbert, M., H. Schamhardt, and B. Nigg, 1991. Calculation of vertical ground reaction force estimates
during running from positional data. Journal of Biomechanics. 24 (12), 1095–1105.
Catavitello, G., I. Yuri, I. and F. Lacquaniti, 2015. Planar covariation of hindlimb and forelimb elevation
angles during terrestrial and aquatic locomotion of dogs. Plos One. https://doi.org/10.1371/journal.
pone.0133936.
∙ 207 ∙
Chan, C., W. Loh, W. and I. Abd Rahim, 2016. Human motion classification using 2D stick-model
matching regression coefficients. Applied Mathematics and Computation. 283 (20), 70–89.
Channon, A., A. Walker, T. Pfau, I. Sheldon, and A. Wilson, 2009. Variability of Manson and Leaver
locomotion scores assigned to dairy cows by different observers. Veterinary Record. 164, 388–392.
Clayton, H. M., H.C. Schamhardt, 2001. Chapter two: Measurement techniques for analysis. Equine
Locomotion. 55–76.
Conte, S., R. Bergeron, H. Gonyou, J. Brown, F.C. Rioja-Lang, L. Connor, and N. Devillers, 2014.
Measure and characterization of lameness in gestating sows using force plate, kinematic and accelerometer
methods. Journal of Animal Science. 92 (12), 5693–5703.
Corazza, S., T. Andriacchi, E. Gambaretto, and L. Mundermann, 2010. Automatic generation of a subject-
specific model for accurate markerless motion capture and biomechanical applications. IEEE Transactions on
Biomedical Engineering. 57, 806–812.
Devillers, N., F. Janvier, F. Delijani, S. Methot, K. Dick, Q. Zhang, and L. Conno, 2019. Effect of slat
and gap width of slatted concrete flooring on sow gait using kinematics analysis. Animals. 9, 206–221.
Devillers, N., X. Yan, K. Dick, Q. Zhang, and L. Connor, 2020. Determining an effective slat and gap
width of flooring for group sow housing, considering both sow comfort and ease of manure management.
Livestock Science. 242, 104275.
Griffin, T. M., R.P. Main, and C.T. Farley, 2004. Biomechanics of quadrupedal walking: how do four-
legged animals achieve inverted pendulum-like movements? The Journal of Experimental Biology. 207,
3545–3558.
Jørgensen, B., 2000. Osteochondrosis/osteoarthrosis and claw disorders in sows, associated with leg
weakness. Acta Vet. Scand. 41, 123–138.
Kirk, R., B. Svensmark, L. Ellegaard, and H. Jensen, 2005. Locomotive disorders associated with sow
mortality in Danish pig herds. Journal of Veterinary Medicine. A, Physiology, Pathology, Clinical Medicine.
52, 423–428.
Lacquaniti, F., R. Grassco, R. and M. Zago, 1999. Motor patterns in walking. News in Physiological
Sciences. 14, 168–174.
Merkens, H., H. Schamhardt, W. Hartmans, and A. Kerjes, 1986. The ground reaction force pattern of the
Dutch Warmblood horse at normal walk. Equine Veterinary Journal. 18, 207–214.
Polet, D., J. Bertram, 2019. An inelastic quadrupedal model discovers four-beat walking, two-beat
running, and pseudo-elastic actuation as energetically optimal. PLOS Computational Biology. https://doi.org/
10.1371/journal.pcbi.1007444
Scalley, T., 2012. Vertical ground reaction force estimation using position data measured from a
markerless motion capture system. MSc Thesis. Department of Mechanical Engineering, The Ohio State
University, Ohio, USA.
Thorup, V. M., F.A. Tøgersen, B. Jørgensen, and B.R. Jensen, 2007. Biomechanical gait analysis of pigs
walking on solid concrete floor. Animal. 1, 708–715.
van Gurp, M., H. Schamhardt, and A. Crowe, 1987. The ground reaction force pattern from the hindlimb
of the horse simulated by a spring model. Acta Anaesthesiologica Scandinavica. 129, 31–33.
von Wachenfelt, H., S. Pinzke, and C. Nilsson, 2009a. Gait and force analysis of provoked pig gait on
clean concrete surfaces. Biosystems Engineering. 104, 534–544.
von Wachenfelt, H., S. Pinzke, C. Nilsson, O. Olsson and C. Ehlorsson, 2009b. Force analysis of
unprovoked pig gait on clean and fouled concretesurfaces. Biosystems Engineering. 104, 250–257.
Yan, X., Q. Zhang, L. Connor, N. Devillers, and K. Dick, 2019. A 2D stick model for simulation of sow
walking. In: ASABE Annual International Meeting. Boston, USA, July 7–10.
∙ 208 ∙
Broiler Gait Score in Relation to Bird Body Weight and Activity

Xiao Yang a, Yang Zhao a,*, Hao Gan b, Shawn Hawkins b, Liz Eckelkamp a, Maria Prado a,
Robert Burns b, Tom Tabler c
a
Department of Animal Science, The University of Tennessee, Knoxville, TN 37996, USA
b
Department of Biosystems Engineering and Soil Science, The University of Tennessee,
Knoxville, TN 37996, USA
c
Department of Poultry Science, Mississippi State University, Mississippi State, MS 39762, USA
* Corresponding author. Email: yzhao@utk.edu
Abstract
Lameness in broilers may be associated with pain and is considered as a major broiler welfare
concern. Manual gait score assessment in commercial broiler houses is discrete, time consuming,
and laborious. As such, the development of automated methods for broiler gait score assessment
would be beneficial to the broiler industry. The objective of this study was to identify the
correlation of broiler gait score with several production and behavioral metrics that could
potentially be used for automatic gait score estimations. Variables included bird body weight and
activity. Sixteen pens were used to rear Cobb 500 and Ross 708 broilers for eight and nine weeks,
respectively (8 pens strain-1, 12 birds pen-1). Gait score of all birds was assessed weekly by trained
assessors following a six-point (0-5) scoring protocol from the third week. The overall body
weight of birds in each pen was measured weekly. Top-view cameras were installed to
continuously record videos of broilers in all 16 pens. Images were extracted from video clips (10
min hour-1) during light period (16L:8D) and imported to MATLAB to determine activity index
through image processing. Daily mean activity indices of the first three days after each gait score
assessment were used for analyses. The results show that gait score of both bird strains increased
with bird age and body weight, indicating worse walking ability as broilers get older and heavier.
Correlations between broiler gait score and body weight fitted well in third-order polynomial
regressions with high coefficients of determination (R2 = 0.97 for both Cobb 500 and Ross 708).
Strong correlation was also observed between broiler gait score and activity in an exponential
regression. The coefficients of determination were 0.76 for Cobb 500 and 0.83 for Ross 708. The
findings of this study demonstrated the potential of estimating broiler gait score using bird body
weight and activity index. This information will help to develop automated gait score assessment
systems in the future.
Keywords: Poultry, lameness, body weight, activity, image processing
1. Introduction
Lameness in broiler chickens is a major issue that may cause welfare concerns and economic
loss (Gocsik et al., 2017). It was reported that 14%-50% of broilers are moderately to severely
lame in commercial farms at the late phase of production cycle (Bassler et al., 2013; De Jong et
al., 2016; Kestin et al., 1992; Sanotra et al., 2003). Lame birds may suffer from pain (Caplen et
al., 2014; Hothersall et al., 2016), and have difficulty in getting access to feed and water resources
(Bessei, 2006; Butterworth et al., 2002). Furthermore, birds with leg abnormalities may cause
economic loss due to the higher mortality (Verma, 2007; Wideman Jr et al., 2012) and slower
body weight gain (Gocsik et al., 2014).
Gait score (GS) is a common metric for lameness assessment of broilers. One of the most
widely used scoring systems is the six-point scoring protocol (0-5, a higher value indicates worse
walking ability) developed by Kestin et al. (1992). However, implementing animal-based
measures at farm level by human assessors and manual methods is discrete, laborious and time-
consuming.
Advances in computational power and artificial intelligence have made image processing a
viable solution for automatic animal behavior monitoring and welfare assessment (Pluk et al.,
∙ 209 ∙
2012; Stavrakakis et al., 2014; Van Hertem et al., 2018). Aydin (2017) developed a machine
vision-based early detection system of broiler lameness, and found high correlations (R2 = 0.988)
between feature variables and gait scores. Broilers were required to walk through a corridor during
the measurement however, which may limit the application of the method in commercial farms.
Flock activity has been identified as a potential indicator of predicting broiler gait score in a
previous study (Aydin et al., 2010). Van Hertem et al. (2018) developed a prediction model for
broiler gait score based on automated measurements of flock distribution, bird activity and body
weight. The study was conducted in Europe where flock management and housing environments
are different from those practiced in the US. Modification of the model may be required for farms
in US.
The objective of this study was to identify the broiler gait score as affected by bird body
weight and activity for Cobb 500 and Ross 708 broilers. Both body weight and activity could be
potentially used for automatic gait score assessment.
2.1. Housing, animals and management
The study was conducted in an experimental facility at the Johnson Animal Research and
Teaching Unit at The University of Tennessee, Knoxville, TN, USA. Sixteen pens, each
measuring 1.5 × 1.5 m (L × W), were used to raise Cobb 500 and Ross 708 broilers (8 pens
strain-1) for eight and nine weeks, respectively. Day-old chicks were procured from a commercial
hatchery and randomly assigned to different pens (12 birds pen-1). A tube feeder and a bell drinker
were installed in each pen. Feed and water were provided ad libitum. Used litter (wood shavings
and manure) was collected from a commercial broiler house, mixed well, and placed onto the
floor at around 5 cm in depth. Temperature was maintained at approximately 32°C at the start of
experiment and was gradually reduced as birds grew with a final temperature set point of 22°C on
38d and thereafter.
2.2. Body weight (BW)
Overall body weight of all birds in each pen was measured weekly using a bench scale
(OHAUS, OHAUS corporation, Parsippany, NJ, USA; capacity: 200 kg; resolution: 0.05 kg;) to
estimate the correlation between body weight and gait score of broilers.
2.3 Broiler activity
Eight top-view cameras (IP5M-B1186EW-28MM, Amcrest Technologies, Huston, TX) were
installed to continuously monitor broilers in the 16 pens. Each camera covered two pens. Videos
were saved in an external hard drive and exported for analysis after the experiment. Images were
extracted from the video clips (first 10 min of each hour) during light period (16L:8D) at the frame
rate of 5 frames sec-1. The pen areas in the images were cropped out and imported to MATLAB
(2018b, The MathWorks, Inc. Natick, MA) for activity index (AI) determination. The AI was
determined by calculating the ratio of bird moving pixel to overall bird-representative pixel in
consecutive images. A detailed description of the method used is provided in Yang et al. (2020).
2.4. Gait score (GS)
Gait score of all birds were determined weekly by two trained assessors beginning at the third
week of bird age. The assessment procedures were adapted from the Welfare Quality Assessment
Protocol (Welfare Quality, 2009) and the scoring system proposed by Garner et al. (2002). Each
bird was manually assigned a score from 0 to 5 based on their walking ability, with 0 being normal
and 5 being unable to walk.
3.1. Correlation of gait score (GS) with body weight (BW)
Figure 1 shows the correlation between GS and BW for Cobb 500 and Ross 708 broilers. In
general, GS continuously increased as bird grew. Correlations of GS and BW for Cobb and Ross
∙ 210 ∙
broilers were fitted polynomial regressions (degree = 3) with high coefficients of determination
(0.97 for both). Positive correlation between BW and GS was also reported in other studies
(Kristensen et al., 2006; Van Hertem et al., 2018). It indicated that BW was an important
determinant of broiler GS.
(a) (b)
Figure 1. Correlations of gait score and body weight for Cobb 500 (a) and Ross 708 (b). Greater
gait score means worse walking ability.
3.2. Correlation of gait score (GS) with activity index (AI)
Figure 2 shows the correlation between GS and AI of Cobb 500 and Ross 708 broilers. In
general, broilers with lower GS showed higher AI. The correlation between AI and GS for Cobb
and Ross well fitted exponential regressions. The coefficients of determination were 0.76 for Cobb
and 0.83 for Ross. In a previous study by Aydin et al. (2010), the maximum activity level was
observed for broilers with scores 2 and 3 as compared to groups scored 0, 1, 4 and 5. The
difference in outcome is probably due to all birds were grouped by GS in their study. However,
broilers with different GS were mixed together for AI determination in current study. Therefore,
the effects of GS on AI may be reduced.
(a) (b)
Figure 2. Correlations of gait score and activity for Cobb 500 (a) and Ross 708 (b). Greater gait
score means worse walking ability, and greater activity index means more active birds.
4. Conclusions
In this study, correlations of broiler gait score with body weight and activity for Cobb 500 and
Ross 708 broilers were investigated. We concluded that gait score of Cobb and Ross broilers were
positively correlated with body weight (R2 = 0.97 for Cobb and Ross), while negatively correlated
with activity level (R2 = 0.76 for Cobb and R2 = 0.83 for Ross). As such, a prediction model based
on body weight and activity index could be developed to determine broiler gait score at flock
∙ 211 ∙
level. These findings provide insight into the development of automatic broiler gait score
assessment system that can be used for commercial farms.
Acknowledgements
Financial support was provided by Foundation for Food & Agriculture Research (FFAR)
SMART Broiler Initiative and the USDA National Institute of Food and Agriculture multistate
project under accession number 1026643. We thank the assistance provided by staff in JARTU,
bird care takers and helpers for score assessment and litter removal.
References
Aydin, A., 2017. Development of an early detection system for lameness of broilers using computer
vision. Computers and Electronics in Agriculture 136, 140–146. https://doi.org/10.1016/j.compag.
2017.02.019.
Aydin, A., Cangar, O., Ozcan, S.E., Bahr, C., Berckmans, D., 2010. Application of a fully automatic
analysis tool to assess the activity of broiler chickens with different gait scores. Computers and Electronics in
Agriculture 73, 194–199. https://doi.org/10.1016/j.compag.2010.05.004.
Bassler, A., Arnould, C., Butterworth, A., Colin, L., De Jong, I., Ferrante, V., Ferrari, P., Haslam, S.,
Wemelsfelder, F., Blokhuis, H., 2013. Potential risk factors associated with contact dermatitis, lameness,
negative emotional state, and fear of humans in broiler chicken flocks. Poultry Science 92, 2811–2826.
https://doi.org/10.3382/ps.2013-03208.
Bessei, W., 2006. Welfare of broilers: a review. World's poultry science journal 62, 455–466.
https://doi.org/10.1079/wps2005108.
Butterworth, A., Weeks, C., Crea, P., Kestin, S., 2002. Dehydration and lameness in a broiler flock.
Animal Welfare 11, 89–94.
Caplen, G., Hothersall, B., Nicol, C., Parker, R., Waterman-Pearson, A., Weeks, C., Murrell, J., 2014.
Lameness is consistently better at predicting broiler chicken performance in mobility tests than other broiler
characteristics. Animal Welfare 23, 179–187. https://doi.org/10.7120/09627286.23.2.179.
De Jong, I., Hindle, V., Butterworth, A., Engel, B., Ferrari, P., Gunnink, H., Moya, T.P., Tuyttens, F.,
Van Reenen, C., 2016. Simplifying the Welfare Quality® assessment protocol for broiler chicken welfare.
Animal 10, 117–127. https://doi.org/10.1017/s1751731115001706.
Garner, J., Falcone, C., Wakenell, P., Martin, M., Mench, J., 2002. Reliability and validity of a modified
gait scoring system and its use in assessing tibial dyschondroplasia in broilers. British poultry science 43,
355–363. https://doi.org/10.1080/00071660120103620.
Gocsik, E., Kortes, H., Lansink, A.O., Saatkamp, H., 2014. Effects of different broiler production systems
on health care costs in the Netherlands. Poultry Science 93, 1301–1317. https://doi.org/10.3382/ps.2013-
03614.
Gocsik, É., Silvera, A., Hansson, H., Saatkamp, H., Blokhuis, H., 2017. Exploring the economic potential
of reducing broiler lameness. British poultry science 58, 337–347. https://doi.org/10.1080/00071668.2017.
1304530.
Hothersall, B., Caplen, G., Parker, R., Nicol, C., Waterman-Pearson, A., Weeks, C., Murrell, J., 2016.
Effects of carprofen, meloxicam and butorphanol on broiler chickens' performance in mobility tests. Animal
Welfare 25, 55–67. https://doi.org/10.7120/09627286.25.1.055.
Kestin, S., Knowles, T., Tinch, A., Gregory, N., 1992. Prevalence of leg weakness in broiler chickens and
its relationship with genotype. Veterinary Record 131, 190–194. https://doi.org/10.1136/vr.131.9.190.
Kristensen, H.H., Perry, G., Prescott, N., Ladewig, J., Ersbøll, A.K., Wathes, C., 2006. Leg health and
performance of broiler chickens reared in different light environments. British poultry science 47, 257–263.
https://doi.org/10.1080/00071660600753557.
Pluk, A., Bahr, C., Poursaberi, A., Maertens, W., Van Nuffel, A., Berckmans, D., 2012. Automatic
measurement of touch and release angles of the fetlock joint for lameness detection in dairy cattle using vision
techniques. Journal of Dairy Science 95, 1738–1748.
Sanotra, G.S., Berg, C., Lund, J.D., 2003. A comparison between leg problems in Danish and Swedish
broiler production. Animal Welfare 12, 677–683.
Stavrakakis, S., Guy, J., Warlow, O., Johnson, G., Edwards, S., 2014. Walking kinematics of growing
pigs associated with differences in musculoskeletal conformation, subjective gait score and osteochondrosis.
Livestock Science 165, 104–113.
∙ 212 ∙
Van Hertem, T., Norton, T., Berckmans, D., Vranken, E., 2018. Predicting broiler gait scores from activity
monitoring and flock data. Biosystems Engineering 173, 93–102. https://doi.org/10.1016/j.biosystemseng.
2018.07.002.
Verma, D.S., 2007. Effects of mycotoxins on the bone structure and leg weakness in poultry: Estimation
of the economic loss by leg weakness.
Welfare Quality, 2009. The Welfare Quality® assessment protocol for poultry (broilers, laying hens). The
Welfare Quality® Consortium, Lelystad, The Netherlands.
Wideman Jr, R., Hamal, K., Stark, J., Blankenship, J., Lester, H., Mitchell, K., Lorenzoni, G., Pevzner,
I., 2012. A wire-flooring model for inducing lameness in broilers: evaluation of probiotics as a prophylactic
treatment. Poultry Science 91, 870–883. https://doi.org/10.3382/ps.2011-01907.
Yang, X., Huo, X., Li, G., Purswell, J.L., Tabler, G.T., Chesser, G.D., Magee, C.L., Zhao, Y., 2020.
Effects of Elevated Platform and Robotic Vehicle on Broiler Production, Welfare, and Housing Environment.
Transactions of the ASABE 63, 1981–1990. https://doi.org/10.13031/trans.14115.
∙ 213 ∙
Effects of Pre-Weaning Co-Mingling Piglets on Behavior and

Growth Performance of Pigs after Weaning
Tongshuai Liu a,b, Xue Hui b,c, Xiao Yang d, Lei Xi a,b,*, Pu Cheng a,b, Zhenzhen Ji b,c,
Xuanyang Li a,b, Zhixiao Yang b,c, Kunhua Zhu b,c, Weidong Liu a,b, Wei Ma a,b,
Zhifang Shi a,b, Jingfeng Zhang a,b, Jingjing Kang e
a
College of Animal Science and Technology, Henan University of Animal Husbandry and Economy,
Zhengzhou, Henan 450046, China
b
Henan Engineering Research Center for Animal Healthy Environment and Intelligent Equipment,
c
School of Energy and Intelligence Engineering, Henan University of Animal Husbandry and Economy,
d
Department of Animal Science, The University of Tennessee, Knoxville, Tennessee 37996, USA
e
College of Veterinary Medicine, Henan University of Animal Husbandry and Economy, Zhengzhou,
Henan 450046, China
* Corresponding author. Email: xileihn@163.com
Abstract
Co-mingling piglets before weaning can effectively alleviate stress when piglets are
transferred from farrowing pens to nursery houses and mixed with other litters. Previous research
indicates that co-mingling piglets could reduce aggression with no impact on pig performance. It
is hypothesized that regrouping with familiar pigs after weaning may improve the effect of pre-
weaning co-mingling on growth performance of pigs. Based on the hypothesis, 20 lactating sows
(Large White × Landrace) and their 229 litters were used in this study. The litters were randomly
divided into two groups (10 litters group-1), pre-weaning co-mingling group (C) and no pre-
weaning co-mingling group (N). Litters were co-mingled one week before weaning in group C,
whereas litters were kept individually in farrowing pens before weaning in group N. Piglets stayed
in the farrowing pens for one more week after weaning at 28 days of age. All piglets were then
transferred from farrowing pens to nursery houses at day 35. Piglets in group C were mixed with
familiar pigs while every two litters were randomly mixed in group N in nursery house. Co-
mingling piglets before weaning reduced aggression after weaning. The reduction was observed
three days after mixing (P=0.01). In addition, pre-weaning co-mingling also increased the average
proportion of eating behavior over 2 days after mixing in nursery pig house (P<0.05). Notably,
piglets that were co-mingled before weaning and mixed with familiar pigs after weaning had
higher body weight at 97 days of age as compared with the control group (P<0.0001). The result
of this research would be helpful to improve the welfare and efficiency of pig production.
Keywords: Pre-weaning co-mingling, behavioral expression, growth performance, welfare, pig
1. Introduction
With transferring from farrowing pens to nursery houses and mixing with unfamiliar litters,
weaned pigs may suffer from stress (Merlot et al., 2004), such as adapting to a new environment
and diet, and forming new social hierarchy (Hayne et al., 2003; Weary et al., 2008). This process
may cause aggression between piglets and make them nervous, wounded and loss of appetite,
which have negative effects on growth rate and feed conversion efficiency. A number of strategies
have been employed to alleviate the stress, such as pre-weaning socialization (D’Eath, 2005; Ko
et al., 2020), nutritional supplementation and use of appeasing pheromones (Peden et al., 2018),
group farrowing (Hillmann et al., 2003), environmental enrichment (Martin et al., 2015),
imprinted and guided weaners (Mesarec et al., 2017), using mixed weight pens, night-time mixing,
using specific mixing pens, odor masking agents, sufficient space and so on. Ison et al. (2018)
conducted an online survey regarding techniques that can be used to minimize aggression at
mixing for nursery to finishing pigs. It is shown that co-mingling piglets before weaning was rated
∙ 214 ∙
as most useful. Co-mingling piglets before weaning, also called early life socialization, could
improve social hierarchy formation when pigs were mixed after weaning (D’Eath, 2005), thereby
reduce aggressive behavior and skin lesions (Camerlink et al., 2018). The early social environment
may affect the development and regulation of aggressive behavior in pigs (Verdon et al., 2016).
However, the effects of co-mingling of piglets before weaning on growth performance were not
consistent in previous research. Turpin et al. (2017b) reported intermittent suckling technique with
co-mingling before weaning (18 days of age) could improve performance after weaning. Pluske
and Williams (1996) found performance of pigs in the first 2 weeks after weaning was not
improved by co-mingling during lactation. Parratt et al. (2006) reported that co-mingling piglets
5 days prior to weaning (day 0) could reduce fighting in the immediate post-weaning period, while
had no effect on growth performance after weaning (days 0 to 7). Turpin et al. (2017a) found that
litters co-mingled one weeks before weaning (15 days of age) had no effect on performance after
weaning (36 days of age). Additionally, Salazar et al. (2018) reported that growth rate was lower
in socialized than in non-socialized piglets, which was probably due to socialized piglets were
more active than non-socialized pigs. Theoretically, less aggression after weaning could improve
welfare of piglets and may enhance the growth performance. Most piglets were regrouped with
unfamiliar ones in research that reported that pre-weaning mixing had no significant influence on
growth performance in post-weaning period (D’Eath, 2005; Turpin et al., 2017a; Salazar et al.,
2018). we hypothesized that regrouping with unfamiliar pigs after weaning may compromise the
effect of pre-weaning co-mingling on growth performance of pigs after mixing, thus the effect of
co-mingling before weaning on the later stage of growing pigs was not fully shown up.
The objective of this research was to investigate the effects of pre-weaning co-mingling on
behavior and growth performance of pigs that were mixed with familiar pigs after weaning. The
results of this research would be helpful to optimize the process of pig management and improve
the efficiency of commercial pig breeding.
2.1. Animals and housing
This research was conducted at a commercial pig farm in Hebi (Henan province, China)
during October 2019 and February 2020. Twenty multiparous sows (Large White × Landrace)
mated to Duroc boars and confined in farrowing crates (2.20 × 0.63 m2) were selected. All sows
were randomly distributed to a room measured 30.0 × 9.0 m2. Farrowing pen dimensioned 2.20
× 1.75 m2 with double row layout. Sows were fed with pelleted lactation diet twice per day over
the period of parturition to weaning. Each farrowing pen had an infrared heating lamp and a warm-
box. The temperature under infrared heating lamps was up to 30 °C. The average air temperature
and relative humidity in farrowing house were 21.06 ± 1.88 °C and 71.59 ± 10.32%, respectively.
Creep feed was provided ad libitum from week 2. Sows and piglets had unlimited access to water.
On the day of weaning, sows were removed from farrowing pen and piglets were kept in the
pens until 35 days of age. The nursery pig house (29.0 × 8.0 m2) had 18 pens (3.3 × 3.1 m2) and
each pen had two litters. The temperature of the nursery house was controlled by infrared heating
lamps and floor heating system. Average temperatures of floor and air in nursery house were
23.95 ± 2.19 °C and 16.59 ± 1.20 °C, respectively. Average relative humidity in nursery house
was 73.78% ± 3.00%. Piglets were fed ad libitum with a commercial balanced dry feed mixture
for weaned piglets and had free access to water. Piglets were kept in the nursery house until 97
days of age. The concentration of total volatile organic compounds in the air of farrowing and
nursery pig houses were 0.667 ± 0.204 ppm and 0.609 ± 0.042 ppm, respectively.
Sows and their litters were randomly divided into two groups, pre-weaning co-mingling group
(C) and no pre-weaning co-mingling group (N). Each group had ten sows and litters. Every litter
had 7–14 piglets, which was 229 piglets in total. Average litter sizes at birth for groups C and N
∙ 215 ∙
were 11.2 ± 2.6 and 11.7 ± 2.6, respectively. Group C allowed piglets from two adjacent farrowing
pens to co-mingle by removing parts of the pens 7 days prior to weaning (Figure 1). Group N was
the control and their piglets were kept in the farrowing pens individually. Piglets in both groups
were weaned at 28 days of age, thereafter they were transported to nursery house at 35 days of
age. In nursery pig house, piglets from group C were mixed with familiar ones, and piglets from
group N were mixed randomly with unfamiliar ones. From 1 to 35 days of age, 22 weaners were
lost to follow-up and 9 died. From 36 to 97 days of age, 26 piglets were lost to follow-up and 10
died.
Figure 1. Photos of farrowing pig house (left) and nursery pig house (right).
2.3. Piglet performance
Piglets born alive were weighed 1 day after birth, 35 days of age, and 97 days of age,
respectively.
2.4. Behavioral observations during lactation
A camera was installed in each nursery pig pen at the height of 2 m above the floor to
continuously monitor behaviors of piglets. Parratt et al. (2006) reported that the number of fights
in pig pens increased dramatically within 3 days after mixing and returned to baseline levels by 4
days post-mixing. Therefore, the behaviors of piglets (aggressive, social, resting and eating) after
mixing in the nursery house were observed for three days. The descriptions of behaviors were
shown in Table 1. Daily observations were carried out in morning (from 9:00 to 11:00 h) and
afternoon (from 14:00 to 16:00 h). During observation, behavior of all piglets was assessed every
15 s, and the occurrence of each behavior was then divided by the total number of piglets in each
pen to obtain the proportions of different behavior in each pen.
Table 1. Behavior categories and a detailed description for each category.
Category Description
Resting Whole length of body on the floor or on other pigs. Hind
quarters on the floor, front legs supporting body.
Eating Pigs were deemed as eating if they were interacting with
feed.
Aggressive Head knocks, bites, pushing, fighting and bullying.
Social Nosing body of other pig or nose-nose contact.
The number of piglets, body weight and proportion of each behavior were analyzed using
Proc NPAR1WAY in SAS version 9.2 (SAS Institute Inc., Cary, NC, USA). As the data were not
normally distributed, untransformed data was used in a non-parametric Wilcoxon Rank-Sum test
to compare the differences of two groups. Statistical significance was accepted when P<0.05.
Results were reported as means with standard error (±SE).
∙ 216 ∙
3. Results
3.1. Behavior of piglets
Average proportions of different behavior categories were shown in Figure 2. Average
proportion of aggression in group C was numerically lower than that in group N during two days
after piglets were transferred to nursery house (P=0.29 for the first day & P=0.20 for the second
day). The variable coefficients of aggression proportion in the first two days after mixing were
0.34 and 0.29 for group C, while 0.20 and 0.13 for group N.
Figure 2. Average proportion of different behavior over 3 days after mixing in nursery pig
house. Note: N, no co-mingling before weaning; C, co-mingling before weaning; The vertical
error bars represent standard error of proportion of behavior; Different superscripts on the bar
chart indicate significant difference (P<0.05).
∙ 217 ∙
Fluctuation of aggression proportion for group C was higher than that of group N during 48 h
after mixing. On the third day, average proportion of aggressive behavior in group C (10.3 ± 2.2%)
was lower than that in group N (13.2 ± 3.5) (P=0.01). In addition, the proportion of aggressive
behavior in group C decreased with time. There was no significant difference on social behavior
between two groups over the 3 days mixing in nursery pig house (P>0.05). The proportion of
social behavior for both groups fluctuated dramatically, with the variable coefficients ranged from
0.30 to 0.39 during the 3 days after mixing.
The proportion of eating behavior in both groups fluctuated dramatically during the 3 days
after mixing. Variable coefficients of eating behavior proportion ranged from 0.58 to 0.76 for both
groups. Average proportion of eating in group C was significantly higher than that in group N
during the first two days after mixing (P=0.03 for the first day & P= 0.0009 for the second day),
while was surpassed by group N in the third day (P=0.0001). There was a trend that the average
proportion of eating decreased along with time in each group.
On the first day after mixing, average proportion of resting behavior of group N was
significantly higher than that of group C (P<0.0001). On the second day, average proportions of
resting behavior of groups N and C were 60.5 ± 33.0% and 75.9 ± 26.8%, respectively (P<0.0001).
However, no significant difference on proportions of resting behavior between two groups was
observed on the third day (P=0.07).
3.2. Body of weight
Average body weights of piglets at 1 day of age, 35 days of age, and 97 days of age were
shown in Table 2. There was no significant difference for body weight at 1 day of age (P=0.37)
and 35 days of age (P=0.53). Nevertheless, the average body weight of group C (35.6 ± 3.3 kg)
was higher than that of group N (33.1 ± 3.0 kg) at 97 days of age (P<0.0001).
Table 2. Average body weights (mean ± SE) of pigs at 1 day of age, 35 days of age, and 97 days
of age.
Average body weight kg-1
Treatment
1 day of age 35 days of age 97 days of age
No co-mingling pre-weaning 1.4±0.2 (117) 9.7±2.2 (100) 33.1±3.0 a (90)
Co-mingling pre-weaning 1.5±0.2 (112) 8.6±1.9 (98) 35.6±3.3 b (91)
Note: The numbers in brackets are count of pigs weighed; Values with a different superscript
correspond to a significant difference (P<0.05).
4. Discussion
4.1. Effects of pre-weaning co-mingling on aggression of piglets after weaning
Socializing piglets before weaning could improve social hierarchy formation and reduce
fighting (D’Eath, 2005). Though a few researchers have reported that pre-weaning socialization
could reduce aggression (Parratt et al., 2006; Camerlink et al., 2018), the significant reduction
was only observed on the third day after mixing in nursery house in this study. Similar results
have been reported by Morgan et al. (2014) that there was no significant difference in aggression
between socialized and control pigs during 21–24 h after mixing. We speculated that piglets may
spend more time adapting to the new environment instead of fighting when they were transferred
to a nursery house. In addition, greater fluctuation of aggression proportion for group C was
probably due to piglets co-mingled before weaning were more confident (Camerlink et al., 2018)
about adapting to new environment.
It is noteworthy that the age when piglets is co-mingled during lactation may have some
influence on aggression. However, the effect of mixing pigs at different ages on aggression is
uncertain. Pitts et al. (2000) reported that pigs mixed at younger ages (5 days of age) before
weaning had shorter time of fighting, whereas Salazar et al. (2018) indicated that co-mingling
piglets at 14 days of age had less aggression during lactation than at 7 days of age (Salazar et al.,
∙ 218 ∙
2018).
4.2. Effects of pre-weaning co-mingling on social behavior of piglets after weaning
It may take piglets more than 72 h to establish the social hierarchy after mixing in nursery
house according to previous research (Parratt et al., 2006). As a result, the proportion of social
behavior fluctuated dramatically in the first three days after mixing. No significant difference on
the amount of social behavior was observed for both groups. The result was consistent with
previous research reported by Salazar et al. (2018) that the proportion of social behavior did not
differ between the co-mingling piglet litters and the control group during lactation.
4.3. Effects of pre-weaning co-mingling on eating behavior of piglets after weaning
Compared with the control group, average proportion of eating behavior for piglets socialized
in lactation was higher during the first two days after regrouping in nursery house. Consistent with
this, Turpin et al. (2017b) reported that intermittent suckling techniques with co-mingling before
weaning could improve feed intake 2 to 8 d after weaning. Over time, piglets may adapt to the
new environment and improve the efficiency of feed intake, thus reduce the time spent on eating
and decrease the proportion of eating behavior.
4.4. Effects of pre-weaning co-mingling on resting behavior of piglets after weaning
Average proportions of resting behavior in group C was significantly lower than that in group
N on the first day after mixing. This finding was different from previous studies reporting that
piglets socialized pre-weaning had higher proportion of resting than that of non-socialized piglets
at initial stage after weaning (Parratt et al., 2006; Turpin et al., 2017a). The difference may be
related to different observing periods of behavior and methods of co-mingling before weaning.
Parratt et al. (2006) recorded lying number of piglets during 3 h following weaning and they co-
mingled piglets 5 days prior to weaning. Turpin et al. (2017a) observed post-weaning behaviors
of piglets for 2-hour periods (morning 09:00 h and afternoon 14:00 h). Ko et al. (2020) found
significant increase of salivary cortisol and chromogranin A for piglets that did not co-mingle pre-
weaning after weaning, while the increase was not significant for piglets co-mingled pre-weaning
with environmental enrichment. The level of salivary cortisol and chromogranin A were two
parameters positively correlated with stress, which may indicate that pre-weaning socialization
could alleviate stress for piglets after weaning. Higher proportion of resting behavior in group N
during the initial 24 h after mixing was probably because piglets were too nervous about the new
environment and partners to do other things. Conversely, the piglets co-mingled pre-weaning
might have relative stable hierarchy and could adapt to the new environment in a short time.
Therefore, they performed more exploration behaviors and spent less time resting after they were
transferred to the nursery house. Along with time, socialized piglets gradually got used to the
environment and had fewer exploration, which resulted in an increase of resting behavior in group
C on the second day after mixing.
4.5. Effects of pre-weaning co-mingling on body weight of piglets at later stage of growing period
No significant difference on average body weight between socialized and non-socialized
piglets before weaning was observed by 35 days of age. Similar results have been reported by
previous studies (Morgan et al., 2014; Camerlink et al., 2018; Kanaan et al., 2008; Turpin et al.,
2017a). However, Salazar et al. (2018) found socialized piglets were lighter than the control group
during lactation. The inconsistency was probably due to the birth weight difference of treatments
for Salazar et al. (2018) found that average body weight of socialized piglets had been heavier
than that of the control at 7 days of age when pre-weaning co-mingling was conducted. Morise et
al. (2008) reported that low birth weight was associated with reduced growth rate throughout the
growing period.
In this study, it was found that pre-weaning socialization could improve growth performance
of pigs. The improvement of body weight may be the result of reduced aggression and increased
feed intake at the initial stage after mixing in nursery house. Similar to this, Camerlink et al.
∙ 219 ∙
(2018) also found that there was a trend that body weight of piglets socialized during lactation
(SOC) was heavier than the control (CON) between 6 weeks and 11 weeks of age, while the
difference of body weight between the SOC and CON was not significant. In line with this,
Morgan et al. (2014) also found this phenomenon on 7 days after weaning. As Morgan et al.
(2014) and Camerlink et al. (2018) did not record body weight of piglets at later stage of growth,
the effect of pre-weaning socialization on body weight of pigs at later stage could not be
determined. In contrast to this study, D’Eath (2005) reported that socializing piglets before
weaning did not improve weight gain of pigs (at 60 days of age). Camerlink et al. (2018) also
demonstrated that there was no significant difference of body weight between socialized and
control piglets at 11 weeks of age. The reason of this contradiction is probably due to D’Eath
(2005) and Camerlink et al. (2018) regrouped pigs with unfamiliar ones in nursery house. Piglets
were usually more active and aggressive in the groups with foreign pigs (Hwang et al., 2016).
However, in practice, it would be better for producers to avoid regrouping pigs with unfamiliar
ones at weaning (Ison et al., 2018; Camerlink et al., 2021).
5. Conclusions
Co-mingling piglets before weaning not only reduced aggression, but also significantly
increased average proportion of eating behavior in the first two days after mixing in nursery house.
Additionally, pre-weaning co-mingling and regrouping with familiar piglets in nursery house
significantly improved body weight of pigs at later stage of growing period compared with the
control group. Effects of mixing pigs at different ages on aggression should be furtherly explored
to improve the welfare of piglets and alleviate stress of mixing piglets.
Acknowledgements
We thank all our colleagues for their support during this study and staffs of the pig farm for
their assistance in this research. This study was supported by the National Key Research and
Development Program of China (2018YFD0501103) and Doctoral research start-up funding of
Henan University of Animal Husbandry and Economy (2019HNUAHEDF037).
References
Camerlink, I., C. Proegger, D. Kubala, K. Galunder, J. Rault, 2021. Keeping littermates together instead
of social mixing benefits pig social behaviour and growth post-weaning. Applied Animal Behaviour Science.
235, 105230. https://doi.org/10.1016/j.applanim.2021.105230.
Camerlink, I., M. Farish, R.B. D’Eath, G. Arnott, Turner, S.P., 2018. Long Term Benefits on Social
Behaviour after Early Life Socialization of Piglets. Animal. 8,192. https://doi.org/10.3390/ani8110192.
Camerlink, I., S.P. Turner, 2016. Farmers’ perception of aggression between growing pigs. Applied
Animal Behaviour Science. 192, 42–47. http://doi.org/10.1016/j.applanim.2016.11.009.
D’Eath, R.B., 2005. Socialising piglets before weaning improves social hierarchy formation when pigs
are mixed post-weaning. Applied Animal Behaviour Science. 93, 199–211. http://doi.org/10.1016/j.applanim.
2004.11.019.
Hayne, S.M., H.W. Gonyou, 2003. Effects of regrouping on the individual behavioural characteristics of
pigs. Applied Animal Behaviour Science. 82, 267–278. http://doi.org/10.1016/S0168-1591(03)00084-4.
Hillmann, E., F.V. Hollen, B. Bünger, D. Todt, L. Schrader, 2003. Farrowing conditions affect the
reactions of piglets towards novel environment and social confrontation at weaning. Applied Animal
Behaviour Science. 81, 99–109. http://doi.org/10.1016/S0168-1591(02)00254-X.
Hwang, H.S., J.K. Lee, T.K. Eom, S.H. Son, J.K. Hong, K.H. Kim, S.J. Rhim, 2016. Behavioral
Characteristics of Weaned Piglets Mixed in Different Groups. Asian-Australasian Journal of Animal Sciences.
29,1060–1064. https://doi.org/10.5713/ajas.15.0734.
Ison, S.H., R.O. Bates, C.W. Ernst, J.P. Steibel, J.M. Siegford, 2018. Housing, ease of handling and
minimising inter-pig aggression at mixing for nursery to finishing pigs as reported in a survey of North
American pork producers. Applied Animal Behaviour Science. 205, 159–166. https://doi.org/10.1016/
j.applanim.2018.05.004.
Kanaan, V.T., E.A. Pajor, D.C. Lay Jr., B.T. Richert, J.P. Garner, 2008. A note on the effects of co-
mingling piglet litters on pre-weaning growth, injuries and responses to behavioural tests. Applied Animal
Behaviour Science. 110, 386–391. http://doi.org/10.1016/j.applanim.2007.05.002.
∙ 220 ∙
Ko, H.L., Q. Chong, D. Escribano, I. Camerlink, X. Manteca, P. Llonch, 2020. Pre-weaning socialization
and environmental enrichment affect life-long response to regrouping in commercially-reared pigs. Applied
Animal Behaviour Science. 229, 105044. https://doi.org/10.1016/j.applanim.2020.105044.
Martin, J.E., S.H. Ison, E.M. Baxter, 2015. The influence of neonatal environment on piglet play
behaviour and post-weaning social and cognitive development. Applied Animal Behaviour Science. 163, 69–
79. http://doi.org/10.1016/j.applanim.2014.11.022.
Merlot, E., M.C. Meunier-Salaün, A. Prunier, 2004. Behavioural, endocrine and immune consequences
of mixing in weaned piglets. Applied Animal Behaviour Science. 85, 247–257. https://doi.org/10.1016/
j.applanim.2003.11.002.
Mesarec, N., M.P. Povše, D. Škorjanc, J. Skok, 2017. Gangs of piglets: Welfare and growth of imprinted
and guided weaners. Applied Animal Behaviour Science. 195, 44–49. http://doi.org/doi:10.1016/
j.applanim.2017.05.018.
Morgan, T., J. Pluske, D. Miller, T. Collins, A.L. Barnes, F. Wemelsfelder, P.A. Fleming, 2014.
Socialising piglets in lactation positively affects their post-weaning behaviour. Applied Animal Behaviour
Science. 158, 23–33. http://doi.org/10.1016/j.applanim.2014.06.001.
Morise, A., I. Louveau, I.L. Huërou-Luron, 2008. Growth and development of adipose tissue and gut and
related endocrine status during early growth in the pig: impact of low birth weight. Animal. 2, 73–83.
https://doi.org/10.1017/S175173110700095X.
Parratt, C.A., K.J. Chapman, C. Turner, P.H. Jones, M.T. Mendl, B.G. Miller, 2006. The fighting
behaviour of piglets mixed before and after weaning in the presence or absence of a sow. Applied Animal
Behaviour Science. 101, 54–67. http://doi.org/10.1016/j.applanim.2006.01.009.
Peden, R.S.E., S.P. Turner, L.A. Boyle, I. Camerlink, 2018. The translation of animal welfare research
into practice: the case of mixing aggression between pigs. Applied Animal Behaviour Science. 204, 1–9.
https://doi.org/10.1016/j.applanim.2018.03.003.
Pitts, A.D., D.M. Weary, E.A. Pajor, D. Fraser, 2000. Mixing at young ages reduces fighting in
unacquainted domestic pigs. Applied Animal Behaviour Science. 68, 191–197. https://doi.org/10.1016/
S0168-1591(00)00104-0.
Pluske, J.R., I.H. Williams, 1996. Reducing stress in piglets as a means of increasing production after
weaning: administration of amperozide or co-mingling of piglets during lactation? Animal Science. 62, 121–
130. http://doi.org/10.1017/S1357729800014405.
Salazar, L.C., H.L. Ko, C.H. Yang, L. Llonch, X. Manteca, I. Camerlink, P. Llonch, 2018. Early
socialisation as a strategy to increase piglets’ social skills in intensive farming conditions. Applied Animal
Behaviour Science. 206, 25–31. https://doi.org/10.1016/j.applanim.2018.05.033.
Turpin, D.L., P. Langendijk, C. Sharp, J.R. Pluske, 2017a. Improving welfare and production in the peri-
weaning period: Effects of co-mingling and intermittent suckling on the stress response, performance,
behaviour, and gastrointestinal tract carbohydrate absorption in young pigs. Livestock Science. 203, 82–91.
http://doi.org/10.1016/j.livsci.2017.07.006.
Turpin, D.L., P. Langendijk, P. Kate, J.R. Pluske, 2017b. Intermittent suckling with or without co-
mingling of non-littermate piglets before weaning improves piglet performance in the immediate post-
weaning period when compared with conventional weaning. Journal of Animal Science and Biotechnology.
8: 14. http://doi.org/10.1186/s40104-017-0144-x.
Verdon, M., R.S. Morrison, P.H. Hemsworth, 2016. Rearing piglets in multi-litter group lactation
systems: Effects on piglet aggression and injuries post-weaning. Applied Animal Behaviour Science. 183, 35–
41. http://doi.org/10.1016/j.applanim.2016.05.008.
Weary, D.M., J. Jasper, M.J. Hötzel, 2008. Understanding weaning distress. Applied Animal Behaviour
Science. 110, 24–41. http://doi.org/10.1016/j.applanim.2007.03.025.
∙ 221 ∙
L-selenomethionine in Cardiotoxicity of Ammonia through

Inhibiting Autophagy Mediated via a Signaling Pathway
Zheng Cheng, Yutao Li, Yufu Shu, Xin Li, Xiaohong Zhang, Honggui Liu *
* Corresponding author. E-mail address: liuhonggui1312@163.com
Abstract
Ammonia is a major harmful gas in livestock houses, and excessive ammonia inhalation has
adverse effects to pig hearts. However, the mechanism of ammonia-induced cardiac toxicity in
pigs has not been reported. This study aims to enrich the toxicological mechanism of ammonia
and provide valuable reference for future intervention of ammonia toxicity. Twenty-four 125-day-
old fattening pigs were randomly divided into 4 groups, namely C (control) group, A (ammonia)
group (with the concentration of 88.2 mg m-3–90.4 mg m-3), Se (Selenomethionine) group (with
a content of 0.5 mg kg-1) and A(ammonia)+Se group, to investigate the mechanism of
mitochondrial autophagy in ammonia inhalation-induced cardiac injury in pigs and the protective
effect of L-Selenomethionine. The results showed that a large number of myocardial fiber edema,
cytoplasmic bleakness and mitochondrial autophagy was observed in the A group; ATP content
and ATPase activities decreased significantly; Endoplasmic reticulum stress (ERS) markers
(GRP78, IRE1α, ATF4, ATF6, and CHOP) were significantly induced at mRNA and protein
levels; PI3K/Akt/mTOR signaling pathway was activated; and autophagy key genes and proteins
(Beclin 1, LC3，ATG3 and ATG5) were significantly up-regulated. Compared with the A group,
various indexes in the A+Se group were significantly alleviated by adding L-Selenomethionine
in the feed. The results indicated that excessive ammonia inhalation caused cardiac injury in pigs,
induced ERS, activated PI3K/Akt/mTOR signaling pathway, and led to mitochondrial autophagy.
Keywords: Ammonia exposure, cardiac damage, autophagy, PI3K/Akt/mTOR signaling
pathway, endoplasmic reticulum stress
1. Introduction
Ammonia is a colorless but has a strong pungent smell of harmful gas, easily soluble in water
(Xing et al., 2016). With the intensification and large-scale development of the breeding industry,
the density of livestock and poultry is getting higher and higher, and the harmful gases are also
increasing, such as ammonia, carbon dioxide, hydrogen sulfide and methane (Ni et al., 2017).
Ammonia in the environment comes from a variety of sources, such as Power plants (Kang et al.,
2020a), agriculture (Beltran et al., 2021), Domestic waste incineration (Kang et al., 2020b) and
animal husbandry (Ignacio et al., 2021), etc. Ammonia cannot be underestimated for its damage
to animals while destroying the atmospheric balance. In recent years, the toxicological hazards of
ammonia to chickens (An et al., 2019), rats (Zielińska et al., 2016) and other animals are
constantly updated.
It has been found that oxidative stress is one of the important mechanisms of ammonia
exposure-induced organism damage (Wang et al., 2020). When ammonia exposure induces
oxidative stress in the body, it is bound to cause the change of antioxidant enzymes and the
increase of reactive oxygen species (ROS) (Gao et al.,2021), and then lead to the occurrence of
endoplasmic reticulum stress (ERS) (Liang et al., 2016). ERS is a kind of suborganelle
pathological state, such as the disorder of physiological function of ER, dysregulation of calcium
homeostasis, and excessive accumulation of unfolded and misfolded proteins in ER lumen due to
some reasons (Bilekova et al., 2020). Autophagy is a lysosomal self degradative metabolic process
that occurs under the control of autophagy related genes, which helps maintain the synthesis of
cellular proteins, the degradation and turnover recycling of organelles, and the homeostatic
balance of the intracellular environment (Levine and Kroemer et al., 2019). PI3K/Akt/mTOR acts
∙ 222 ∙
as a classical autophagy signaling pathway to regulate a variety of biological functions, including

proliferation, survival, escape from apoptosis and protein synthesis (Elsayed et al., 2021). Studies
have found that ammonia exposure can cause oxidative stress, which further leads to endoplasmic
reticulum stress and autophagy (Ali et al., 2021). However, the role of oxidative stress,
endoplasmic reticulum stress, and autophagy in the mechanisms of ammonia-induced cardiac
injury remains unclear.
Selenium (Se) is an important trace element in animal body, which plays an important
biological function in anti-oxidative stress, improving body immunity, antagonism of heavy metal
toxicity and so on (Sordillo and López-Alonso, 2013). The main form of Se function in the body
is selenoprotein, especially selenium-enzyme. It has been widely used in pig industry as feed
additives within the past four decades (Surai and Fisinin, 2016). To our knowledge, the mitigating
effect of Se on ammonia toxicity has not been reported. Therefore, in the present study, we
evaluated the toxic effects of ammonia exposure on pig heart from the perspectives of oxidative
stress, energy metabolism, and autophagy, and for the first time examined the mitigating effects
of L-selenomethionine on ammonia-induced cardiotoxicity. This study provides a theoretical
basis for improving the toxicological mechanism of ammonia, and also provides reference for
comparative medicine.
2.1. Experimental animals
All procedures in our experiment were approved by the Animal Care and Use Committee of
Northeast Agricultural University. Twenty-four 125-day-old healthy fattening sows with similar
body weight were randomly assigned to four groups (six pigs in each group): control group (C
group), ammonia group (A group), Se group, and ammonia+Se group (A+Se group). The
ammonia concentration in the C group and the Se group was lower than 5 mg m-3. The ammonia
concentration in the A group and the A+Se group was in the range of 88.2-90.4 mg m-3. The pigs
in the C group and the A group were feed basal diets. The Se content in the basic diet was 0.218
± 0.021 mg kg-1. The Se content in the diet in the Se group and the A + Se group was 0.5 mg
kg-1 by adding L-Selenomethionine. The formal experiment lasted 30 days. At the end of the
experiment, all pigs were euthanized, and heart tissue was taken.
2.2. Microscopic examination
The heart tissues were fixed in 10% neutral formalin solution for fixation for at least 24 h,
dehydrated in ethanol, embedded in paraffin. Then tissue sections were stained with hematoxylin
and eosin (H&E).
2.3. Ultrastructural examination
The heart tissue sections (1.0 mm3) were fixed in 2.5% glutaraldehyde phosphate-buffered
solution, and then in 1% osmium tetroxide. The samples were dehydrated with a graded series of
ethanol, then transferred to acetone and embedded in EPON 812. The ultrathin sections (50 nm)
were cut, mounted on a copper grid, and stained with uranyl acetate and lead citrate.
Ultrastructural observation was performed using a transmission electron microscopy (TEM)
(HITACHI, Japan).
2.4. Determination of ROS
Reactive oxygen species was detected by fluorescence staining. The slides were observed
under a fluorescence microscope and the images were collected. The nuclei stained by DAPI are
blue under UV excitation, and the positive expression is red light. The cumulative optical Density
(IOD) and the corresponding tissue pixel Area (Area) of the three positive visual fields in each
section were analyzed by using Image-Pro Plus 6.0 analysis software, and the Areal Density
(IOD/Area) was calculated.
∙ 223 ∙
2.5. Quantitative real-time PCR (qRT-PCR)

Total RNA in the heart of pigs was extracted. β-actin was used as a reference gene. Autophagy
related genes (Beclin1, LC3-I, LC3-II, ATG3, and ATG5), endoplasmic reticulum stress related
genes (GRP78, IRE1α, PERK, ATF4, ATF6, and CHOP) and β-actin were measured. The qRT-
PCR was performed on Light Cycler®480 System (Roche, Basel, Switzerland). The change in
mRNA level was analyzed using the 2−△△Ct method.
2.6. Western blot
Total protein was extracted from tissues and electrophoresis was performed on 12% SDS-
polyacrylamide gel, then transferred to nitrocellulose membrane in 20% methanol triglycine
buffer. Sealed with 5% skim milk at 37°C for 2h and incubated overnight with diluted primary
antibodies, then HRP-bound secondary antibodies against rabbit IgG were incubated at 37°C for
1h (1:2000, Santa Cruz, USA).
3.1. Effect of excess ammonia and/or Se on histological structure
In the C group and the Se group (Figure 1A and 1B), bundle cells were large and clear, and
regularly arranged. Myocardial fibers were closely arranged and intermuscular blood vessels were
abundant. No obvious abnormalities and inflammation were observed in the interstitium. In the
A+Se group (Figure 1C), a small amount of myocardial fiber edema was observed, and the
cytoplasm was loose and pale (black arrow). In the A group (Figure 1D), there was a lot of
myocardial fiber edema, and the cytoplasm was loose and pale (black arrow).
Figure 1. Effects of ammonia or/and Se on microstructure.
∙ 224 ∙
In the C group and the Se group (Figure 2A and 2B), the nuclei were normal, the nucleoli
were elliptic, the nuclear membrane was clear and complete. In the A group (Figure 2D), a large
number of mitochondrial ridge breaks (yellow arrows) and mitochondrial vacuolation (blue
arrows), as well as a large number of autophagosomes (red arrows) and autophagosolysosomes
(green arrows) were observed. In the A + Se group (Figure 2C), the number of mitochondrial
ridge breaks (yellow arrow) and mitochondrial vacuolation (blue arrow) as well as
autophagosomes (red arrow) and autophagosomes (green arrow) were significantly lower than
those in ammonia treatment group to some extent.
Figure 2. Effects of ammonia or/and Se on ultrastructure of the heart.

3.2. Effect of excess ammonia and/or Se on ROS content
Compared with the C group, ROS content in the Se group was no significant change, while
ROS content in the A +Se group and A group increased significantly (P<0.05). Compared with A
+ Se group, ROS content in A group was increased significantly (P<0.05, Figure 3).
Figure 3. Effects of ammonia or/and Se on ROS level.

3.3. Effect of excess ammonia and/or Se on the expression level of ERS-related genes
ERS-related genes were detected by qRT-PCR and Western blot and were shown in Figure 4.
Compared to the C group, the mRNA levels of ERS-related genes (GRP78, ATF6, IRE1α, ATF4,
and CHOP) displayed a significant increase (P ˂ 0.05) in the A group. Those changes were
remarkably alleviated in the A + Se group. There was no significant difference (P >0.05) in ERS-
related genes between the C group and the Se group. The protein expressions of GRP78, IRE1α,
ATF4, ATF6 and CHOP showed a similar trend with their mRNA levels.
∙ 225 ∙
Figure 4. Effects of ammonia or/and Se on the mRNA level and protein expression levels of
ERS-related genes.
3.4. Effect of excess ammonia and/or Se on PI3K/Akt/mTOR signaling pathway

Effects of excess ammonia exposure and/or Se on PI3K/Akt/mTOR signaling pathway are
shown in Figure 5. Compared with the C group, the p-PI3K/PI3K, p-AKT/AKT and p-
mTOR/mTOR in the protein level in the A+Se and A group were significantly reduced (P<0.05).
There was no significant difference in the protein level of the p-PI3K/PI3K, p-AKT/AKT, p-
mTOR/mTOR between the C group and the Se group (P >0.05). These results indicated that
excessive ammonia exposure activated the PI3K/Akt/mTOR signaling pathway, while the
addition of L-selenomethionine has an inhibitory effect on the PI3K/Akt/mTOR signaling
pathway activated by ammonia exposure.
3.5. Effect of excess ammonia and/or Se on the expression level of autophagy-related genes
Effects of excess ammonia exposure and/or Se on the mRNA level and the protein expression
of autophagy-related genes in pig hearts are shown in Figure 6. The mRNA levels of autophagy-
related genes (LC3-I, LC3-II, ATG3, ATG5 and Beclin-1) in the A+Se and the A groups were
significantly increased (P<0.05) compared with the C group.
In addition, the mRNA levels and the protein expression of autophagy-related genes (LC3-I,
LC3-II, ATG3, ATG5 and Beclin-1) in the A group were significantly increased (P<0.05)
compared with the A+Se group. These results indicated that excess ammonia exposure could
activate autophagy in pig hearts, thus causing heart damage, and the addition of L-
∙ 226 ∙
selenomethionine could inhibit the excessive autophagy caused by excessive ammonia exposure
to some extent.
Figure 5. Effects of ammonia or/and Se on the expression levels of PI3K/Akt/mTOR pathway-

related proteins.
4. Conclusion
In conclusion, the results indicated that excessive ammonia inhalation caused cardiac injury
in pigs, induced ERS, activated PI3K/Akt/mTOR signaling pathway, and led to mitochondrial
autophagy. The L-Selenomethionine supplementation could alleviate the cardiac injury caused by
excessive ammonia inhalation to a certain extent by inhibiting ERS and P13K/AKT/mTOR
signaling pathway. The findings provided a new perspective for investigating the effect of
ammonia on heart injury and provided a reference for comparative medicine.
Acknowledgements
The study was supported by the Earmarked Fund for China Agriculture Research System
(Project No. CARS-35).
∙ 227 ∙
Figure 6. Effects of ammonia or/and Se on the mRNA and protein levels of autophagy-related
genes.
References
Ali, S. S. W., Zhang, S., Ishfaq, M., Tang, Y., Teng, X., 2021. PTEN/AKT/mTOR pathway involvement
in autophagy, mediated by miR-99a-3p and energy metabolism in ammonia-exposed chicken bursal
lymphocytes. Poultry Science. 100(2), 553–564. https://doi.org/10.1016/j.psj.2020.11.015.
An, Y., Xing, H., Zhang, Y., Jia, P., Gu, X., Teng, X., 2019. The evaluation of potential immunotoxicity
induced by environmental pollutant ammonia in broilers. Poultry Science. 98, 3165–3175. https://doi.org/
10.3382/ps/pez135.
Beltran, I., Weerden, T. J., Alfaro, M. A., Amon, B., Klein, C. A. M., Grace, P., Hafner, S., Hassouna,
M., Hutchings, N., Krol, D. J., Leytem, A. B., Noble, A., Salazar, F., Thorman, R., Velthof, G. L., 2021.
DATAMAN: A global database of nitrous oxide and ammonia emission factors for excreta deposited by
livestock and land‐applied manure. Journal of Environmental Quality. 50(2), 513–527. https://doi.org/
10.1002/jeq2.20186.
Bilekova, S., Sachs, S., Lickert, H., 2020. Pharmacological Targeting of Endoplasmic Reticulum Stress
in Pancreatic Beta Cells. Trends Pharmacol. Sci. 42(2), 85–95. https://doi.org/10.1016/j.tips.2020.11.011.
Elsayed, A. M., Bayraktar, E., Amero, P., Salama, S. A., Abdelaziz, A. H., Ismail, R. S., Zhang, X., Ivan,
C., Sood, A. K., LopezBerestein, G., RodriguezAguayo, C., 2021. PRKAR1B-AS2 Long Noncoding RNA
Promotes Tumorigenesis, Survival, and Chemoresistance via the PI3K/AKT/mTOR Pathway. International
Journal of Molecular Sciences. 22(4), 1882. https://doi.org/10.3390/ijms22041882.
Gao, X.-Q., Fei, F., Huang, B., Meng, X. S., Zhang, T., Zhao, K.-F., Chen, H.-B., Xing, R., Liu, B.-L.,
2021. Alterations in hematological and biochemical parameters, oxidative stress, and immune response in
Takifugu rubripes under acute ammonia exposure. Comparative Biochemistry and Physiology, Part C. 243,
108978. https://doi.org/10.1016/j.cbpc.2021.108978.
∙ 228 ∙
Ignacio, B., J., W. T., A., A. M., Barbara, A., M., K. C. A., Peter, G., Sasha, H., Mélynda, H., Nicholas,
H., J., K. D., B., L. A., Alasdair, N., Francisco, S., E., T. R., L., V. G., 2021. DATAMAN: A global database
of nitrous oxide and ammonia emission factors for excreta deposited by livestock and land‐applied manure.
Journal of Environmental Quality. 50(2), 513–527. https://doi.org/10.1002/jeq2.20186.
Kang, S., Kim, S.-D., Jeon, E.-C., 2020a. Ammonia Emission Sources Characteristics and Emission
Factor Uncertainty at Liquefied Natural Gas Power Plants. Int. J. Env. Res. Public Health. 17(11),3758.
https://doi.org/10.3390/ijerph17113758.
Kang, S., Woo, J., Jeon, E., 2020b. Ammonia Emission Characteristics and Emission Factor of Municipal
Solid Waste Incineration Plant. Sustainability. 12(18), 7309. https://doi.org/10.3390/su12187309.
Levine, B., Kroemer, G., 2019. Biological Functions of Autophagy Genes: A Disease Perspective. Cell.
176, 11–42. https://doi.org/10.1016/j.cell.2018.09.048.
Liang, Z., Liu, R., Zhao, D., Wang, L., Sun, M., Wang, M., Song, L., 2016. Ammonia exposure induces
oxidative stress, endoplasmic reticulum stress and apoptosis in hepatopancreas of pacific white shrimp
(Litopenaeus vannamei). Fish and Shellfish Immunology. 54, 523–528. https://doi.org/10.1016/j.fsi.2016.
05.009.
Ni, J.-Q., Liu, S., Diehl, C. A., Lim, T.-T., Bogan, B. W., Chen, L., Chai, L., Wang, K., Heber, A. J.,
2017. Emission factors and characteristics of ammonia, hydrogen sulfide, carbon dioxide, and particulate
matter at two high-rise layer hen houses. Atmos. Environ. 154, 260–273. https://doi.org/10.1016/j.atmosenv.
2017.01.050.
Sordillo, L. M., López-Alonso, M., 2013. Selenium-Dependent Regulation of Oxidative Stress and
Immunity in Periparturient Dairy Cattle. Veterinary Medicine International. 2013, 154045. https://doi.org/
10.1155/2013/154045.
Surai, P. F., Fisinin, V. I., 2016. Selenium in sow nutrition. Anim. Feed Sci. Technol. 211, 18–30.
https://doi.org/10.1016/j.anifeedsci.2015.11.006.
Wang, H., Zhang, Y., Han, Q., Xu, Y., Hu, G., Xing, H., 2020. The inflammatory injury of heart caused
by ammonia is realized by oxidative stress and abnormal energy metabolism activating inflammatory pathway.
Sci. Total Environ. 742, 140532. https://doi.org/10.1016/j.scitotenv.2020.140532.
Xing, H., Luan, S., Sun, Y., Sa, R., Zhang, H., 2016. Effects of ammonia exposure on carcass traits and
fatty acid composition of broiler meat. Animal Nutrition. 2, 282–287. https://doi.org/10.1016/j.aninu.
2016.07.006.
Zielińska, M., Dąbrowska, K., Hadera, M. G., Sonnewald, U., Albrecht, J., 2016. System N transporters
are critical for glutamine release and modulate metabolic fluxes of glucose and acetate in cultured cortical
astrocytes: changes induced by ammonia. J. Neurochem. 136(2), 329–338. https://doi.org/10.1111/jnc.13376.
∙ 229 ∙
Impacts of Keel Bone Damage on Fear Behavior and

Physiological Responses of Laying Hens
Runxiang Zhang a,b, Haidong Wei a, Yanru Feng a, Hanlin Yu a, Susu Ding a,
Haoyang Nian a, Hengyi Zhang a, Qian Zhao a, Jianhong Li c, Jun Bao a,b,*
a
b
Key Laboratory of Chicken Genetics and Breeding, Ministry of Agriculture and Rural Affairs,
c
Abstract
Keel bone damage (KBD) is prevalent in all laying hen housing systems and its incidence
differs between strains. To investigate the incidence of KBD and its effect on emotional welfare
in commercial laying hens and native breed of Lindian chickens, a 2 (Strain) × 3 (Group) factorial
study was designed. Total 100 Hy-line Brown and 100 Lindian chickens (10 birds per furnished
cage) at 25-week-old were studied for 7 weeks. At the peak of laying (at 32-week-old), birds were
palpated and classified as normal (NK), deviated (DK), and fractured (FK) groups according to
the keel bone status. Birds in each group of strains were determined the fear behaviors and fear-
related physiological indicators. Results showed that the incidence of KBD (especially fractures)
in Lindian chickens was higher than Hy-line Brown hens. Meanwhile, Lindian chickens had
significantly high whole blood 5-HT content, ISO-vocalization, NOT-latency, HAT-score, INV-
flapping number and duration (p<0.01), and lower serum IL-6 content compared to Hy-line Brown
hens. Additionally, FK hens had significantly elevated the contents of whole blood CORT and
serum IL-1β and IL-6, TI-duration, and NOT-latency (p<0.05), as well as reduced 5-HT content
and INV-flapping wings number (p<0.05) in comparison to NK and DK hens, while these
indicators were not different between NK and DK hens (p >0.05). Therefore, KBD negatively
affects fear related physiological and behavioral responses of laying hens, and this negative
welfare impact is mainly reflected by keel fracture.
Keywords: Laying hen, fear, stress, animal welfare, keel bone fracture, strain
1. Introduction
With the transform of laying hens from conventional cage to alternative systems like furnished
cages, cage-free and aviary systems, the occurrence of keel bone damage (KBD, including
deviations and fractures) has become a serious health and welfare problem. Previous studies have
reported that KBD (especially fractures) not only change behavior (Nasr et al., 2012; Casey-Trott
and Widowski, 2016), reduce egg production and egg quality (Rufener et al., 2019; Wei et al.,
2020), but also cause pain (Nasr et al., 2012) and physiological stress (Wei et al., 2019), as well
as induce a depressive-like state (Armstrong et al., 2020) in laying hens.
It was found that the incidence of KBD varies by chicken strains. In commercial laying hens,
some studies reported that brown strains have more sever and frequent keel bone deviations
(Habig and Distl, 2013) and more keel bone fractures (Fawcett et al., 2020) when compared with
white strains; however, other studies reported that white laying hens have more and serve keel
deviations and fractures in comparison with brown hens (Candelotto et al., 2017; Ali et al., 2020).
Besides, there was strain differences in occurrence of KBD in native chickens from different
countries or regions (Kittelsen et al., 2020).
Chickens are similar to mammals and other birds that have behavioral and physiological
responses to various environmental stimulations and stressors (Cockrem, 2013). Fear and stress
responses of chickens to stressor or stimuli are common. Tonic immobility (TI) is a well-validated
and widely-used test for the measurement of fear response in chickens, and TI duration is a valid
∙ 230 ∙
indicator to assess fear. A longer TI duration indicates a stronger fear. Novel object test (NOT) is
another popular method to measure fear in poultry, and the latency of avoidance or approach of
birds to novel object can directly reflect the fear levels (Hocking et al., 2001). The approach of
birds to novel object usually suggests a lower fear and a positive emotion state; whereas the
avoidance of birds to novel object suggests a higher fear and a negative emotion state. In addition,
isolation test (IT), human approach test (HAT), and inversion test are also applied to assess fear
of chickens in many studies (Nelson et al., 2020). Generally, fear and stress can reflect emotion
state, it was proved that corticosterone, serotonin (5-HT) and cytokines interleukin 1β (IL-1β) and
IL-6 are associated with positive and negative emotion in human and animals (Korte, 2001).
Lindian chicken, a native strain from Lindian County, Heilongjiang province, China, is
famous for unique texture and flavor of its meat and eggs, and stronger cold resistance and crude
feed resistance traits (Qiu et al., 2015). However, the percentage of keel bone damage and
measurement of behavioral and physiological responses of Lindian chickens with different keel
bone status has not been assessed. Therefore, the present study was conducted to shed some light
as to how keel bone fractures influence the fear-related responses of commercial and native laying
hens.
2.1. Animals and management
A total of one hundred Hy-line Brown commercial laying hens and 100 China-Lindian native
chickens with normal keel bone, both at 25 weeks of age (WOA), was used in this study. Birds
were housed in furnished cages with a size of 150 cm length × 70 cm width × 70 cm height, and
cage was enriched with two quadrate wood perches (20 cm and 40 cm above the wire-mesh floor,
and 45 cm and 25 cm away from front wire-mesh sidewall, respectively), an elevated enclosed
red nest box with a size of 40 cm length × 70 cm width × 25 cm height (installed right wire-mesh
sidewall and 45 cm above the wire-mesh floor), a rectangular feed trough and a waterline with
four nipple drinkers. Each strain laying hens was randomly assigned to ten replicated cages with
10 birds each. Furnished cages for per strain were allocated at a semi-enclosed hen house with
natural ventilation. Artificial light was provided for 16 h from 5:00 to 21:00, and light intensity
was 20-24 lux. The ambient temperature and relative humidity of hen house were 18–21 °C and
50–70%, respectively. All laying hens were fed a commercial corn-soybean meal layer diet with
16.08% crude protein and 2800 kcal kg-1 metabolic energy, and they can free access to feed and
drink water throughout the entire experimental period from 25 to 32 WOA.
2.2. Palpation of keel bone status
Keel bone status of all lying hens was assessed through palpation method according to the
description of Casey-Trott et al. (2015). Keel bone status can be classified as three main
categories: normal keel (NK), deviated keel (DK) and fractured keel (FK) bones based on the
presence or absence of deviations and fractures. In the present study, if a laying hen had both DK
and FK bone, the keel bone was recognized as FK.
2.3. Focal animal selection and sample collection
After the assessment of KBD for all laying hens at the 1st day of 32 WOA, 10 birds were
randomly selected for each group of NK, DK and FK per strain and marked with different
numerical leg-tags. In this study, these selected laying hens were used as focal animals to weight
individually, collect blood, and measure fear responses. Total 3 ml blood of each focal bird was
collected, and 1 ml blood was collected in 5-mL EDTA tubes for the determination of
corticosterone (CORT) and serotonin (5-HT). The remaining 2 ml blood was then centrifugated
at 3000 rpm for 15 min at 4°C, thus serum samples were obtained for the determination of
emotion-related cytokines.
∙ 231 ∙
2.4. Determination of physiological indicators in blood

Concentrations of blood CORT and 5-HT were measured by using commercially available
ELISA kits following the manufacturer’s instructions (Shanghai Jinma Laboratory Equipment
Corporation Ltd, Shanghai, China). Levels of IL-1β and IL-6 in serum of each group chicken were
measured by using commercially available ELISA kits (Shanghai Xinle Biotechnology Co., Ltd,
Shanghai, China) in accordance with the manufacturer’s instructions.
2.5. Determination of fear response
Measurement of fear response in two chicken strains with different keel bone status was
performed at 32 WOA. Fear measurement mainly includes isolation (ISO) test, novel object test
(NOT), human approach test (HAT), tonic immobility (TI), and inversion (INV) test. All
behavioral tests followed conventional testing procedure.
Statistical analysis was performed by SPSS 22 (SPSS Inc., Chicago, IL, USA). Data of serum
fear-related physiological indicators and fear behavior responses were firstly tested for normality
with Shapiro-Wilk test. Data on INV-flapping wings duration was transformed by the log method
and data on ISO-vocalizations was transformed by the square root method for normal distribution.
The ANOVA of 2 × 3 (Strain × Group) factorial arrangement of treatments with Duncan’s
multiple range test was then applied to compare the differences in each indicator between
treatments. Each bird in cage was recognized as an experimental unit. The results were expressed
as mean ± SEM, the difference was considered as statistically significant when p ≤ 0.05.
3.1. Percentage of keel bone damage
The strain difference in KBD was found in the present study (Figure 1), The percentage of
NK, DK, and FK bones in Hy-line Brown laying hens was 43.3%, 30.0%, and 26.7%, and that in
Lindian chickens was 40%, 23.3%, and 36.7%, respectively. It was confirmed that the percentage
of keel bone damage (especially FK) in Lindian chickens was higher than that in Hy-line Brown
laying hens at 32 WOA. Previous studies found that the prevalence of KBD was different between
white and brown commercial laying hens, as well as within noncommercial chickens (Kittelsen
et al., 2020).
Figure 1. Percentage of normal keel (NK), deviated keel (DK), and fractured keel (FK) in two
chicken strains.
3.2. Measurement of physiological indicators
The levels of physiological indicators are shown in Table 1. The interaction of chicken’s strain
(Strain) and keel bone status (Group) had a significant effect on the concentration of IL-1β (F
(2,54) = 4.36, p = 0.02), but no effects on the concentrations of CORT (F (2,54) = 0.64, p = 0.53),
5-HT (F (2,54) = 0.99, p = 0.38), and IL-6 (F (2,54) = 0.54, p = 0.59). Studies on laying hens
∙ 232 ∙
found that keel bone fractures induced pain (Nasr et al., 2012), stress and inflammation (Wei et
al., 2019). The physiological indicators might confirm the negatively affect induced by keel bone
fractures.
The concentration of whole blood CORT was affected by Group (F (2,57) = 10.23, p<0.01)
but not affected by Strain (F (1,58) = 4.36, p = 0.93), and FK birds had evidently elevated CORT
concentration compared to NK and DK birds. At the physiological level, an elevated level of
blood CORT usually represents the organism experiences stress response (Post et al., 2003).
Previous studies showed that keel bone fractures increased the concentration of serum CORT in
caged laying hens (Wei et al., 2019). Similarly, this study found that KBD had significantly impact
on whole blood CORT content, and FK hens had elevated blood CORT content compared to NK
and DK hens, indicating that keel bone fractures lead to stress.
The concentration of whole blood 5-HT was affected by Strain (F (1,58) = 11.22, p<0.01) and
Group (F (2, 57) = 5.16, p = 0.02). There was a significantly increased 5-HT concentration in
Lindian chickens compared to Hy-line Brown hens (F (1,58) = 11.22, p<0.01), and the 5-HT
concentration in NK birds was significantly higher than FK birds (F (2,57) = 5.16, p = 0.02). It
was showed that blood 5-HT content was related to the degree of fearfulness and stress in laying
hens and broilers (Uitdehaag et al., 2011). The results showed that FK hens had decreased blood
5-HT content compared to NK hens, suggesting that FK hens experienced more fear than NK
hens.
Table 1. Levels of stress and fear-related indicators in blood of two chicken strains with
different keel bone status.
Strain Group CORT (ng L-1) 5-HT (ng L-1) IL-1β (ng L-1) IL-6 (ng L-1)
NK 456.64±13.72 469.61±8.72 171.89±2.20 45.88±1.11
Hy-line
DK 471.39±14.55 456.53±9.74 170.61±3.16 48.14±0.82
Brown
FK 509.95±7.93 452.27±10.27 177.77±2.97 50.86±1.06
NK 472.83±6.77 497.17±4.06 166.60±2.38 44.28±0.64
Lindian DK 483.76±3.05 487.48±7.67 171.74±4.38 44.96±0.81
FK 505.05±6.01 461.59±7.64 189.85±2.13 49.25±0.73
Main effect
Hy-line
479.66 459.47b 173.42 48.29a
Brown
Strain
Lindian 487.21 482.08a 176.06 46.16b
SEM 5.54 4.77 1.72 0.51
NK 465.24b 483.39a 169.25b 45.08b
b ab b
DK 477.58 472.01 171.18 46.55b
Group
FK 507.50a 456.93b 183.81a 50.06a
SEM 6.79 5.85 2.10 0.62
Strain 0.340 0.001 0.281 0.014
p-value Group <0.001 0.017 <0.001 <0.001
Strain×Group 0.531 0.378 0.018 0.587
Strain 0.93 11.22 1.18 8.87
F-value Group 10.23 5.16 14.17 17.06
Strain×Group 0.64 0.99 4.36 0.54
Data are expressed as mean and SEM, n = 10 per group; a, b, c Means in the same column with
different superscripts are significantly different, p ≤ 0.05; Note: NK = normal keel bone; DK =
deviated keel bone; FK = fractured keel bone; CORT = corticosterone; 5-TH = serotonin; IL =
interleukin.
The concentration of serum IL-1β was affected by Group (F (2,57) = 14.17, p<0.01) but not
affected by Strain (F (1, 58) = 1.18, p = 0.28), and FK birds had significantly increased IL-1β
∙ 233 ∙
concentration compared to NK and DK birds (F (2,57) = 14.17, p<0.01). Additionally, the

concentration of serum IL-6 was affected by Strain (F (1,58) = 8.87.17, p = 0.01) and Group (F
(2,57) = 17.06, p<0.01). The IL-6 concentration in Hy-line Brown hens was evidently higher than
Lindian chickens (F (1,58) = 8.87.17, p = 0.01), and that in FK birds was evidently higher than
NK and DK birds (F (2,57) = 17.06, p<0.01). It has been confirmed that proinflammatory
cytokines like IL-1β and IL-6 are involved in the reaction of emotion such as anxiety and fear in
human and animals, and positive emotion is related to low levels of proinflammatory cytokines
(Song, 2011). The results of this study showed that the concentrations of cytokines IL-1β and IL-
6 in serum of FK hens was significantly higher than NK and DK hens, indicating that keel bone
fractures could induce a negative emotion (especially fear) in laying hens.
3.3. Measurement of fear response
The present study showed that fear response of chickens to TI, NOT, and INV tests was
affected by bird’s strain and keel bone status (Figure 2). The ISO-vocalization was affected by
Strain (F (1,58) = 1082.49, p<0.01) but not affected by Group (F (2,57) = 1.66, p = 0.20), and the
number of vocalizations in Lindian chicken was significantly higher than Hy-line Brown hens in
ISO test (F (1,58) = 1082.49, p<0.01, Figure 2a). The latency of NOT test was affected by Strain
(F (1,58) = 884.33, p<0.01) and Group (F (2,57) = 7.26, p<0.01), and Lindian chickens had
evidently higher NOT-latency than Hy-line Brown hens (F (1,58) = 884.33, p<0.01); the latency
of NOT test in FK chickens was elevated when compared to NK and DK chickens (F (2,57) =
7.26, p<0.01, Figure 2b). The score of HAT test was affected by Strain (F (1,58) = 28.79, p<0.01)
but not affected by Group (F (2,57) = 2.63, p = 0.08), and Lindian chickens had a significantly
high score of HAT test compared to Hy-line Brown hens (F (1,58) = 28.79, p<0.01, Figure 2c).
For the INV test, the number and duration of flapping wings was significantly affected by
Strain (number: F(1,58) = 30.20, p<0.01; duration: F(1,58) = 13.41, p<0.01), and the number of
flapping wings was also affected by Group (F(2,57) = 4.62, p = 0.01), but the duration was not
influenced by Group (F(2,57) = 2.17, p = 0.12); there were significant increase in the number
(F(1,58) = 30.20, p<0.01) and duration (F(1,58) = 13.41, p<0.01) of flapping wings in Lindian
chickens in comparison to Hy-line Brown hens (Figure 2d and e), and the number of flapping
wings in FK birds was evidently lower that NK birds (F(1,58) = 4.62, p = 0.01, Figure 2e). The
intensity of flapping wings was not significantly affected by Strain (F (1,58) = 3.86, p = 0.06) and
Group (F (2,57) = 1.25, p = 0.29) in INV test (Figure 2f). Additionally, the TI duration was
evidently affected by Strain (F (1,58) = 35.03, p<0.01) and Group (F (2,57) = 16.49, p<0.01). The
duration of TI in Hy-line Brown hens was significantly longer than Lindian chickens (F (1,58) =
35.03, p<0.01), and TI duration in FK birds was significantly increased compared with NK and
DK birds (F (2,57) = 16.49, p<0.01, Figure 2g).
From the above, we found that Hy-line Brown laying hens had evidently increased TI-duration
and decreased ISO-vocalizations, NOT-latency, HAT-score, INV-flapping wing number and
duration in comparison to Lindian chickens, indicating that the difference of chickens to fear
response depending on the strain. This result was similar with the findings of Kozak et al., (2019)
and Nelson et al., (2020), who reported that behavioral response of chickens to TI, ISO, NOT and
HAT as fear test differed between white and brown strains, or commercial and native strains. The
results from TI test indicated that Hy-line Brown hens with a longer TI duration had a higher
fearfulness than Lindian chickens, but from the results of other fear tests such as ISO, HAT, NOT
and INV showed that Lindian chickens had a higher fearfulness than Hy-line Brown hens. The
behavioral performance of birds to TI test is passive fear response, while behavioral performance
to HAT, NOT, ISO and INV tests is active fear response, which may indicate that Lindian
chickens are willing to perform active behavioral response to fearfulness, while commercial Hy-
line Brown hens tend to perform passive fear behaviors. Overall, the results indicated that
behavioral response of chickens to fear test varied by strain. On the other hand, the result indicated
that keel bone fractures caused fear response that was reflected by TI, NOT and INV tests.
∙ 234 ∙
Figure 2. Behavioral responses of Lindian chickens and Hy-line Brown hens with normal keel
(NK), deviated keel (DK) and fractured keel (FK) bones to fear tests. Data are expressed as
mean ± SEM, n = 10 each group per strain.
Previous studies found that keel bone fractures caused pain and inhibited mobility in laying
hens (Nasr et al., 2012; Rentsch et al., 2019). It was approved that TI is a passive behavioral
response that is reflected by immobility; while NOT and INV are active response to fear reflected
by avoidance and struggle (Forkman et al., 2007). Keel bone fractures impaired hen’s mobility,
which may induce a longer duration of TI and latency of NOT. Because keel bone fracture is one
of the factors caused pain to hens, which may impair the behavior motivation of flapping wings
in INV test, thus resulting in a reduction in the number of flapping wings. Therefore, it might be
speculated that fear response reflected by TI, NOT and INV tests in FK hens was associated with
∙ 235 ∙
the pain and reduced mobility induced by keel bone fractures.

4. Conclusions
In conclusion, the study found that the incidence of keel bone fractures in Lindian chickens
were higher than Hy-line Brown hens. Meanwhile, strain had a significant effect on fear-related
physiological and behavioral responses of hens by Lindian chickens which had increased whole
blood 5-HT content, ISO-vocalization, NOT-latency, HAT-score, INV-flapping number and
duration, and decreased serum IL-6 content and TI-duration compared to Hy-line Brown hens.
Additionally, FK hens had significantly elevated whole blood CORT, serum IL-1β and IL-6
contents, TI-duration and NOT-latency, as well as reduced whole blood 5-HT content and INV-
flapping wings number in comparison to NK and DK hens, which suggested that keel bone
fractures can lead to stress and fear and finally rise up to damage the emotion and welfare state of
laying hens.
Acknowledgements
This study was supported by National Nature Science Foundation of China (grant number
31972608).
References
Ali, A., Campbell, D., Siegford, J.M., 2020. A risk assessment of health, production, and resource
occupancy for 4 laying hen strains across the lay cycle in acommercial-style aviary system. Poult. Sci. 99,
4672–8464. https://doi.org/10.1016/J.PSJ.2020.05.057.
Armstrong, E.A., Rufener, C., Toscano, M.J., Eastham, J.E., Guy, J.H., Sandilands, V., Boswell, T.,
Smulders, T., 2020. Keel bone fractures induce a depressive-like state in laying hens. Sci. Rep. 10, 3007.
https://doi.org/10.1038/S41598-020-59940-1.
Candelotto, L., Stratmann, A., Gebhardt-Henrich, S.G., Rufener, C., van de Braak, T., Toscano, M.J.,
2017. Susceptibility to keel bone fractures in laying hens and the role of genetic variation. Poult. Sci. 96,
3517–3528. https://doi.org/10.3382/PS/PEX146.
Casey-Trott, T., Heerkens, J.L.T., Petrik, M., Regmi, P., Schrader, L., Toscano, M.J., Widowski, T., 2015.
Methods for assessment of keel bone damage in poultry. Poult. Sci. 94, 2339–2350. https://doi.org/10.3382/
PS/PEV223.
Casey-Trott, T.M., Widowski, T.M., 2016. Behavioral differences of laying hens with fractured keel
bones within furnished cages. Front. Vet. Sci. 3, 42. https://doi.org/10.3389/FVETS.2016.00042.
Forkman, B., Boissy, A., Meunier-Salaün, M.C., Canali, E., Jones, R.B., 2007. A critical review of fear
tests used on cattle, pigs, sheep, poultry and horses. Physiol. Behav. 92, 340–374. https://doi.org/10.1016/
J.PHYSBEH.2007.03.016.
Habig, C., Distl, O., 2013. Evaluation of bone strength, keel bone status, plumage condition and egg
quality of two layer lines kept in small group housing systems. Br. Poult. Sci. 54, 413–424. https://doi.org/
10.1080/00071668.2013.792405.
Hocking, P.M., Channing, C.E., Waddington, D., Jones, R.B., 2001. Age-related changes in fear, sociality
and pecking behaviours in two strains of laying hen. Br. Poult. Sci. 42, 414–423. https://doi.org/10.1080/
00071660120070686.
Kittelsen, K.E., Moe, R.O., Hansen, T.B., Toftaker, I., Vasdal, G., 2020. A descriptive study of keel bone
fractures in hens and roosters from four non-commercial laying breeds housed in furnished cages. Animals
10, 2192. https://doi.org/10.3390/ANI10112192.
Korte, S.M., 2001. Corticosteroids in relation to fear, anxiety and psychopathology. Neurosci. Biobehav.
Rev. 25, 117-142. https://doi.org/10.1016/S0149-7634(01)00002-1.
Kozak, A., Kasperek, K., Zięba, G., Rozempolska-Rucińska, I., 2019. Variability of laying hen behaviour
depending on the breed. Asian-Australas. J. Anim. Sci. 32, 1062–1068. https://doi.org/10.5713/AJAS.
18.0645.
Nasr, M.A.F., Murrell, J., Wilkins, L.J., Nicol, C.J. 2012., The effect of keel fractures on egg-production
parameters, mobility and behaviour in individual laying hens. Anim. Welf. 21, 127–135. https://doi.org/
10.7120/096272812799129376.
Nelson, J.R., Settar, P., Berger, E., Wolc, A., O’Sullivan, N., Archer, G.S., 2020. Brown and white egg-
layer strain differences in fearfulness and stress measures. Appl. Anim. Behav. Sci. 231, 105087.
https://doi.org/10.1016/J.APPLANIM.2020.105087.
∙ 236 ∙
Qiu, J.W., Wang, Z.P., Zhang, Y.L., Leng, L., Li, Y.M., Luan, P., Li, H., 2015. Genetic analysis on the
rose-comb,fishy taint and dwarf traits in lindian chicken population. China Poultry 37, 7–12. https://doi.org/
10.16372/j.issn.1004-6364.2015.14.002.
Rentsch, A.K., Rufener, C.B., Spadavecchia, C., Stratmann, A., Toscano, M.J., 2019. Laying hen’s
mobility is impaired by keel bone fractures and does not improve with paracetamol treatment. Appl. Anim.
Behav. Sci. 216, 19–25. https://doi.org/10.1016/j.applanim.2019.04.015.
Rufener, C., Baur, S., Stratmann, A., Toscano, M.J., 2019. Keel bone fractures affect egg laying
performance but not egg quality in laying hens housed in a commercial aviary system, Poult. Sci. 98, 1589–
1600. https://doi.org/10.3382/PS/PEY544.
Song, C., Wang, H., 2011. Cytokines mediated inflammation and decreased neurogenesis in animal
models of depression. Prog. Neuropsychopharmacol. Biol. Psychiatry 35, 760–768. https://doi.org/10.1016/
J.PNPBP.2010.06.020.
Wei, H.D., Bi, Y.J., Xin, H.W., Pan, L., Liu, R.Z., Li, X., Li, J.H., Zhang, R.X., Bao, J., 2020. Keel
fracture changed the behavior and reduced the welfare, production performance, and egg quality in laying
hens housed individually in furnished cages. Poult. Sci. 99, 3334–3342. https://doi.org/10.1016/
J.PSJ.2020.04.001.
Wei, H.D., Li, C., Xin, H.W., Li, S., Bi, Y.J., Li, X., Li, J.H., Zhang, R.X., Bao, J., 2019. Keel fracture
causes stress and inflammatory responses and inhibits the expression of the resin system in laying hens.
Animals 9, 804. https://doi.org/10.3390/ANI9100804.
∙ 237 ∙
Prediction of Sow Physiological Parameters Based on Support

Vector Machine and Extreme Gradient Boosting Models
Yanrong Zhuang a,b, Mengbing Cao a,b, Shulei Li a,b, Yu Liu a,b, Jianlong Zhang a,b,
Guanghui Teng a,b,*
a
b
* Corresponding author. Email: futong@cau.edu.cn
Abstract
The heat stress can affect the production performance of pig, especially sow. Multiple
indicators have been used to evaluate heat stress, but most of the models are built for people, cow
or young pig, not for sow. The equivalent temperature index for sows (ETIS) model, which
combines with environment paraments, can well evaluate the heat stress of sows, but the result in
test dataset had a gap with that in train dataset making the general applicability of ETIS necessary
to be improved. It may cause by the small dataset and no cross validation. This study built and
compared the performance of Support Vector Regression (SVR) and Extreme Gradient Boosting
(XGBoost) models to predict sow physiological parameters including skin temperature, rectal
temperature and respiration rate. One thousand and twenty-nine datasets were collected, and each
dataset included the parameters of temperature, relative humidity (RH) and wind speed as input
variables, skin temperature, rectal temperature and respiration rate as output variables. All of the
datasets were divided into train set (80%) and test set (20%) randomly. The results showed SVR
model and XGBoost models could well predict the physiological parameters of sow, especially
the skin temperature. The determination coefficients (R2) of SVR model and XGBoost model
were basically consistent with that between ETIS model and skin temperature. However, the
performances of SVR and XGBoost models in predicting sow rectal temperature and respiration
rate were bad. The bigger quantity, wider range, and more factors of data should be considered to
develop and improve the sow rectal temperature and respiration rate prediction model.
Keywords: Sow, heats stress, physiological parameters, support vector regression, extreme
gradient boosting
1. Introduction
At present, pig breeding is mainly procced in closed house, which makes it easier for pigs to
suffer heat stress (Mikovits et al., 2019) because of the nonfunctional sweat glands pig and its
extremely sensibility to heat (Dallaire et al., 1996). Heat stress can seriously affect the production
performance of sows and cause huge economic losses to producers (Baumgard and Rhoads, 2013;
Seibert et al., 2018). Heat stress would reduce the feed intake, body weight (Christon et al., 1999)
and milk yield of sows (Black et al., 1993; Christon et al., 1999) to decrease piglet weight gain
(Christon et al., 1999), and even cause death (Dallaire et al., 1996). In addition, sows affected by
heat stress might make the endocrine imbalance (Barb et al., 1991) to lead an inability to estrus
on schedule (Almond and Bilkei, 2005). Heat stress will also affect the skin temperature (Cao, et
al., 2021a, b), rectal temperature and respiration rate (Marple et al., 1974; Patience et al., 2005).
A series of negative effects affected by heat stress on sows seriously influence economic benefit
of farmers to make it urgent to well know the degree of heat stress of sow. Therefore, it is essential
to build a prediction model to evaluate the influence of heat stress to sows vividly.
Many models were established to assess thermal environment to reflect the influence of heat
stress, such as temperature-humidity index (THI) model (Thom, 1959), enthalpy (H) (Rodrigues,
et al., 2011), black globe-humidity index (BGHI) (Buffington, et al., 1981; de Oliveira Junior, et
al., 2011), and equivalent temperature index (ET) (Bjerg, et al., 2018; Bjerg and Kai, 2019).
∙ 238 ∙
However, these models mainly developed for young pigs or cows, and could not well reflect sows’
characteristics. Cao et al. (2021a; 2021b) used temperature, relative humidity (RH) and wind
speed as input variables, skin temperature as output variables to established equivalent
temperature index of the sows (ETIS) model, which could work better than the four models
mentioned before and show a good relationship with skin temperature of sows (R2 of train dataset
was 0.63, R2 of test dataset was 0.73). In addition, the correlations between THI and skin
temperature (R2 = 0.62), rectal temperature (R2 = 0.10), and respiratory rate (R2 = 0.14) were
calculated (Cao, et al., 2021a, b). However, the study of ETIS did not use the method of cross
validation and parameter adjustment to deduce the impact of small datasets, resulting in poor
generalization ability, and the result in test dataset had a gap with the result in train dataset that
meant the general applicability of ETIS needed to be improved. With the development of
computer technology, machine learning models, which support cross validation and parameter
adjustment, have been used in pig farm and got high accuracy in prediction of piglets’ stress
(Gorczyca, et al., 2018; Da Fonseca, et al., 2020), analysis of pig farm growth performance (Lee,
et al., 2019), and prediction of indoor air temperature and RH (Arulmozhi, et al., 2021), which
could help to develop high quality heat stress assessment model for sow.
This study collected sows’ physiological data, indoor environment data from a sow farm, and
applied two machine learning algorithms (Support Vector Regression (SVR), Extreme Gradient
Boosting (XGBOOST)) to develop model. The main objectives were: (1) to build models to
predict sow physiological parameters; (2) to compare the performance of different models in
predicting physiological parameters of sow.
2.1. Animals and Sow barn
This study was conducted from June to August of 2018 at the Ministry of Agriculture Feed
Industry Centre Fengning Animal Test Base (Fengning Hebei, China). Sows were crossbreeds of
Large White and Landrace, and raised in sow barn.
The barn could house 30 non-pregnant multiparous sows, which were waiting for breeding.
Sows were kept in crates with concrete solid floors (Figure 1 a, b). The size of every crate was
2.2 m L × 0.64 m W × 1 m H and each crate installed one feeder and one drinker. The ventilation
system consisted of 2 air inlets and a fan (YH900, Yinghe Company, Shenzhen, China) with the
capacity of 28,500 m3 h-1. The ventilation rate was controlled by workers based on inside air
temperature.
2.2. Data acquisition
In this study, 1029 datasets were collected, including temperature, RH, wind speed, skin
temperature, rectal temperature and respiration rate. Environment sensors were installed at 4
sample points at the height of 0.6 m above the floor (Figure 1b). Temperature and RH were
collected using wireless air temperature and RH sensors (testo 605i, Schwarzwald, Germany.
Ranges of -40 °C–60 °C, and accuracies of ± 0.5 °C for temperature. Ranges of 0–100%, and
accuracies of ± (1.8%RH+2.5%) for RH). Wind speed was monitored by hot wire anemometers
(testo 405i, Schwarzwald, Germany. Ranges of 0 m s-1 to 30 m s-1, and accuracies of ± (0.1
m s-1+5%)). 4 batches sows with 10 non-pregnancy sows randomly selected at air outlet, air inlet
and middle area of the barn in each batch were measured to get skin temperature, rectal
temperature and respiration rate.
2.3. Machine learning algorithms
The computer used in this study was equipped with Intel Core i7-9700 processor with 16GB
running memory and Microsoft Windows 10 (64 bit). All the processes of model development,
statistical analysis, and data mining were based on Python 3.6.5 and scikit-learn toolkit 0.19.1.
In this study, environmental data were used as input variables and physiology data of sows
were used as output factor. The ratio of the train set and the test set was 4:1, and the
∙ 239 ∙
hyperparameters were tuned with train set. The 5-fold cross validation (CV) was used to train the
data, and the test set data was used to test the reliability of the model. The dataset was trained by
using SVR and XGBoost models, and the R2 of the two models on the estimation of skin
temperature, rectal temperature, and respiration rate were compared. In addition, to test the
general applicability of the model, 6 models were tested and compared by using test dataset.
Figure 1. Barn schematic: (a) Side view; (b) Top view.

2.3.1. SVR model
Support Vector Machine (SVM) is a useful technology in statistical learning theory (Vapnik,
1999). The SVR is a powerful artificial intelligence tool to deal with regression problems based
on SVM (Vapnik, 1999; Yeh, et al., 2011).
Before training, the hyperparameters of the SVR model were adjusted. From the relevant
research of ETIS, it could be seen that the linearities of the evaluation model based on temperature,
RH and wind speed were high (Cao, et al., 2021a, b), so linear was selected as the kernel function
in this study. The hyperparameter C and ε are the two key hyperparameters of linear kernel. The
hyperparameter C is the penalty parameter of the error term, bad C may cause overfitting and
underfitting of the output. The hyperparameter ε is a unique hyperparameter of linear kernel, it
specifies the epsilon-tube within which no penalty is associated in the training loss function with
points predicted within a distance epsilon from the actual value. Grid search scheme was used to
choose good hyperparameters to build the SVR model, and the determination coefficient (R2) was
used to evaluate the optimized hyperparameters (Table 1).
Table 1. Interval of hyperparameter adjustments and optimized values in SVR models.
SVR hyperparameters C ε
Optimal values one 100 0.2
Optimal values two 100 0.4
Optimal values three 100 4.6
Note: Optimal values one, two, three represented optimal values of skin temperature model,
rectal temperature model, and respiration rate model, respectively.
2.3.2. XGBoost model
The XGBoost is a tree boosting (Friedman, 2001) machine learning system, which has used
∙ 240 ∙
in lots of data mining challenges and machine learning studies (Chen and Guestrin, 2016).
XGBoost is the model with high computation efficiency and good capability to solve overfitting
problems (Chen, et al., 2015).
Before training, hyperparameters of the XGBoost model were adjusted. Compared with the
SVR model, the XGBoost model had more hyperparameters to adjust (Table 2). The same
hyperparameter tuning method mentioned in 2.3.1 was used. The interval of adjusted
hyperparameters and optimized hyperparameters were shown in Table 2.
Table 2. Interval of hyperparameter adjustments and optimized values in XGBoost models.
XGBoost Lower Upper Optimal Optimal Optimal
hyperparameters limit limit values one values two values three
N_estimators 10 200 70 70 110
Max_depth 1 15 4 3 4
Min_child_weight 1 15 5 2 7
Learning_rate 0.0001 1 0.1 0.1 0.1
Reg_alpha 1 15 1 1 6
Reg_lambda 1 15 1 2 1
Note: Optimal values one, two, three represented optimal values of skin temperature model,
rectal temperature model, and respiration rate model, respectively.
3.1. Descriptive statistic
The dataset included 1029 sets of air temperature, RH, wind speed, skin temperature, rectal
temperature and respiration rate (Table 3).
Table 3. Statistic of the studied datasets.
Item Unit Max Mean Min SD
Temperature °C 34.0 28.7 21.9 2.6
RH % 89.8 65.8 40.4 10.0
Wind speed m s-1 0.29 0.07 0.00 0.07
Skin temperature °C 37.8 34.9 28.6 1.4
rectal temperature °C 41.27 38.44 37.21 0.46
respiration rate Times min-1 168 49 12 28
3.2. Model training
Compared the results of train dataset in different models (Table 4), it showed that: (1) The
mean R2 of SVR model with skin temperature train dataset was 0.65 and that of XGBoost model
was 0.71; (2) The mean R2 of SVR model with core temperature train dataset was 0.12 and that
of XGBoost model was 0.12; (3) The mean R2 of SVR model with respiration rate train dataset
was 0.17 and that of XGBoost model was 0.32.
Table 4. R2 of SVR and XGBoost model in predict skin temperature, rectal temperature and
respiration rate by using train dataset.
Item Fold 1 Fold 2 Fold 3 Fold 4 Fold 5 Mean
SVR-skin 0.71 0.68 0.69 0.60 0.58 0.65
XGBoost-skin 0.72 0.74 0.71 0.69 0.68 0.71
SVR-rectal temperature 0.17 0.11 0.12 0.08 0.12 0.12
XGBoost-rectal temperature 0.15 0.10 0.17 0.08 0.14 0.12
SVR-respiration rate 0.15 0.19 0.23 0.10 0.17 0.17
XGBoost-respiration rate 0.26 0.32 0.35 0.30 0.39 0.32
Note: Fold1 1, 2, 3, 4, 5 represented the 5 separate folds of 5-fold cross validation.
∙ 241 ∙
3.3. Model testing and comparison

The skin temperatures predicted by SVR and XGBoost models with test dataset were shown
in Figure 2. The R2 of SVR and XGBoost models were 0.64 and 0.71; The rectal temperatures
predicted by SVR and XGBoost models with test dataset were shown in Figure 3. R2 of SVR and
XGBoost models were 0.09 and 0.13; The respiration rate predicted by SVR and XGBoost models
with test dataset were shown in Figure 4. R2 of SVR and XGBoost models were 0.15 and 0.32.
The prediction accuracies of SVR and XGBoost models for skin temperature in train dataset
were higher than that of ETIS, of which the R2 were 0.65, 0.71, and 0.63 (Cao, et al., 2021a, b),
respectively. While for the test dataset, the R2 of SVR and XGBoost to predict skin temperature
were 0.64 and 0.71, which were lower than that of ETIS (R2 = 0.73) (Cao, et al., 2021a, b). But
the R2 of SVR and XGBoost models with train dataset and test dataset did not have gap that
illustrated the two models had better general applicability compared with ETIS. The results
showed machine learning models could be well used in pig farms to estimate the heat stress of
sow because of the good general applicability, especially the XGBoost model.
Both SVR and XGBoost models did not get good prediction result for rectal temperature. The
R2 of SVR and XGBoost models with train dataset were 0.12 and 0.12, and R2 with test dataset
were 0.09 and 0.13, respectively. It might because of the fact that sows were thermostatic animals,
and the environment of the experiment did not reach the lethal range of pigs. In order to solve this
problem, a wider range of data should be obtained in the future studies.
The prediction accuracies of SVR and XGBoost models for respiration rate were higher than
those for rectal temperature, but still low. The R2 of SVR and XGBoost models with train dataset
were 0.17 and 0.32, and R2 with test dataset were 0.15 and 0.32, respectively. The low prediction
accuracy for respiration rate might cause by the case that the respiration rate of sow was not only
affected by temperature, but also by the posture and breathing pattern of sow (Cabezon et al.,
2017; Li et al., 2018). However, the pearson correlation (R) between the prediction results of
XGBoost model and real respiration rate was 0.57 showed that XGBoost got well result, which
showed the performance of XGBoost was stronger than SVR in training complex parameters.
(a) (b)
Figure 2. Skin temperature prediction results using SVR model in test dataset(a), XGBoost
model (b).
∙ 242 ∙
(a) (b)
Figure 3. Rectal temperature prediction results using SVR model in test dataset (a), XGBoost
model (b).
(a) (b)
Figure 4. Respiration rate prediction results using SVR model in test dataset (a), XGBoost
model (b).
In this study, SVR and XGBoost algorithms were used to establish the sows’ physiological
parameters prediction model with temperature, humidity, and wind speed as input variables and
skin temperature, rectal temperature, and respiratory rate of sow as output factors. The conclusions
were as follows:
(1) The performances of SVR and XGBoost models in predicting sow skin temperature were
better than ETIS model, and the 5-fold cross validation used in SVR and XGBoost models
improved their general applicability.
(2) The performances of SVR and XGBoost models in predicting sow rectal temperature and
respiration rate were bad. The bigger quantity, wider range, and more factors of data
should be considered to develop and improve the sow rectal temperature and respiration
rate prediction model.
(3) The XGBoost model was more suitable to deal with the data of complex condition than
SVR.
∙ 243 ∙
Acknowledgements
This work was supported by the National Key R&D Program Project of China (grant number
2016YFD0700200).
Reference
Almond, P. K., Bilkei G., 2005. Seasonal infertility in large pig production units in an Eastern-European
climate. Australian Veterinary Journal. 83 (6), 344–346. https://doi.org/10.1111/j.1751-0813.2005.tb15627.x.
Arulmozhi, E., Basak, J.K., Sihalath, T., Park, J., Kim, H.T., Moon, B.E., 2021. Machine learning-based
microclimate model for indoor air temperature and relative humidity prediction in a swine building. Animals.
11 (1), 222. https://doi.org/10.3390/ani11010222.
Barb, C.R., Estienne, M.J., Kraeling, R.R., Marple, D.N., Rampacek, G.B., Rahe, C.H., Sartin, J.L., 1991.
Endocrine changes in sows exposed to elevated ambient-temperature during lactation. Domestic Animal
Endocrinology. 8 (1), 117–127. https://doi.org/10.1016/0739-7240(91)90046-M.
Baumgard, L. H., R. J. Rhoads, 2013. Effects of heat stress on postabsorptive metabolism and energetics.
Annual Review of Animal Biosciences. 1, 311-37. https://doi.org/10.1146/annurev-animal-031412-103644.
Bjerg, B.S., Rong, L., Zhang, G., 2018. Computational prediction of the effective temperature in the lying
area of pig pens. Computers and Electronics in Agriculture. 149 (SI), 71–79. https://doi.org/10.1016/j.compag.
2017.09.016.
Bjerg, B. S., P. Kai, 2019. CFD prediction of heat transfer in heated or cooled concrete floors in laying
areas for pig. In 2019 ASABE Annual International Meeting. Boston, MA, USA, July 7–10. https://doi.org/
10.13031/aim.201900735.
Black, J.L., Mullan, B.P., Lorschy, M.L., Giles, L.R., 1993. Lactation in the sow during heat-stress.
Livestock Production Science. 35 (1–2), 153–170. https://doi.org/10.1016/0301-6226(93)90188-N.
Buffington, D.E., Collazoarocho, A., Canton, G.H., Pitt, D., Thatcher, W.W., Collier, R.J., 1981. Black
globe-humidity index (BGHI) as comfort equation for dairy-cows. Transactions of the ASAE. 24 (3), 711–
714. https://doi.org/10.13031/2013.34325.
Cabezon, F.A., Schinckel, A.P., Marchant-Forde, J.N., Johnson, J.S., Stwalley, R.M., 2017.Effect of floor
cooling on late lactation sows under acute heat stress. Livestock Science. 206, 113–120. https://doi.org/
10.1016/j.livsci.2017.10.017.
Cao, M., Zong, C., Zhuang, Y., Teng, G., Zhou, S., Yang, T., 2021. Modeling of Heat Stress in Sows Part
2: Comparison of Various Thermal Comfort Indices. Animals. 11 (6), 1498. https://doi.org/10.3390/
ani11061498.
Cao, M., Zong, C., Wang, X., Teng, G., Zhuang, Y., Lei, K., 2021. Modeling of Heat Stress in Sows—
Part 1: Establishment of the Prediction Model for the Equivalent Temperature Index of the Sows. Animals.
11 (5), 1472. https://doi.org/10.3390/ani11061498.
Chen, J., Li, G., Xiao, B., Wen, Z., Lv, M., Chen, C., Jiang, Y., Wang, X., Wu, S., 2015. Assessing the
transferability of support vector machine model for estimation of global solar radiation from air temperature.
Energy Conversion and Management. 89, 318–329. https://doi.org/10.1016/j.enconman.2014.10.004.
Chen, T. Q., C. Guestrin, 2016. XGBoost: A Scalable Tree Boosting System. In Proceedings of the 22nd
ACM SIGKDD International Conference on knowledge discovery and data mining.San Francisco, CA,USA,
AUG 13–17. https://doi.org/10.1145/2939672.2939785.
Christon, R., Saminadin, G., Lionet, H., Racon, B., 1999. Dietary fat and climate alter food intake,
performance of lactating sows and their litters and fatty acid composition of milk. Animal Science. 69 (2),
353–365. https://doi.org/10.1017/S135772980005092X.
Da Fonseca, F.N., Abe, J.M., Naas, I.D.A., Da Silva Cordeiro, A.F., Do Amaral, F.V., Ungaro, H.C.,
2020. Automatic prediction of stress in piglets (Sus Scrofa) using infrared skin temperature. Computers and
Electronics in Agriculture. 168, 105148. https://doi.org/10.1016/j.compag.2019.105148.
DAllaire, S., R. Drolet, Brodeur, D., 1996. Sow mortality associated with high ambient temperatures.
Canadian Veterinary Journal-revue Veterinaire Canadienne. 37 (4), 237–239.
de Oliveira Junior, G.M., Ferreira, A.S., Oliveira, R.F.M., Silva, B.A.N., de Figueiredo, E.M., Santos,
M., 2011. Behaviour and performance of lactating sows housed in different types of farrowing rooms during
summer. Livestock Science. 141 (2–3), 194–201. https://doi.org/10.1016/j.livsci.2011.06.001.
Friedman, J. H., 2001. Greedy function approximation: A gradient boosting machine. Annals of Statistics.
29 (5), 1189–1232. https://doi.org/10.1214/aos/1013203451.
Gorczyca, M.T., Milan, H.F.M., Campos Maia, A.S., Gebremedhin, K.G., 2018. Machine learning
algorithms to predict core, skin, and hair-coat temperatures of piglets. Computers and Electronics in
Agriculture 151, 286–294. https://doi.org/10.1016/j.compag.2018.06.028.
∙ 244 ∙
Lee, W., Ham, Y., Ban, T., Jo, O., 2019. Analysis of Growth Performance in Swine Based on Machine
Learning. IEEE Access. 7, 161716–161724. https://doi.org/10.1109/ACCESS.2019.2951522.
Li, W., Yang, X.D., Dai, A.N., Chen, K., 2018. Sleep and Wake Classification Based on Heart Rate and
Respiration Rate. In 3rd International Conference on Automation, Control and Robotics Engineering.
Chengdu, CHINA, July 19–22. https://doi.org/10.110910.1088/1757-899X/428/1/012017.
Marple, D.N., Jones, D.J., Alliston, C.W., Forrest, J.C., 1974. Physiological and endocrinological changes
in response to terminal heat stress in swine. Journal of Animal Science. 39 (1), 79–82.
Mikovits, C., Zollitsch, W., Hoertenhuber, S.J., Baumgartner, J., Niebuhr, K., Piringer, M., Anders, I.,
Andre, K., Hennig-Pauka, I., Schoenhart, M., Schauberger, G., 2019. Impacts of global warming on confined
livestock systems for growing-fattening pigs: simulation of heat stress for 1981 to 2017 in Central Europe.
International Journal of Biometeorology 63 (2), 221–230. https://doi.org/10.1007/s00484-018-01655-0.
Patience, J.F., Umboh, J.F., Chaplin, R.K., Nyachoti, C.M., 2005. Nutritional and physiological responses
of growing pigs exposed to a diurnal pattern of heat stress. Livestock Production Science. 96(2–3), 205–214.
https://doi.org/10.1016/j.livprodsci.2005.01.012.
Rodrigues, V.C., Oliveira Da Silva, I.J., Correa Vieira, F.M., Nascimento, S.T., 2011. A correct enthalpy
relationship as thermal comfort index for livestock. International Journal of Biometeorology. 55 (3), 455–459.
https://doi.org/10.1007/s00484-010-0344-y.
Seibert, J., Graves, K., Hale, B., Keating, A., Baumgard, L., Ross, J., 2018. Characterizing the acute heat
stress response in gilts: I. Thermoregulatory and production variables. Journal of Animal Science. 96 (3), 941–
949. https://doi.org/10.1093/jas/skx036/4904292.
Thom, E. C., 1959. The discomfort index. Weatherwise. 12, 57–61.
Vapnik, V. N., 1999. An overview of statistical learning theory. IEEE Transactions on Neural Networks.
10 (5), 988–999. https://doi.org/10.1109/72.788640.
Yeh, C., Huang, C., Lee, S., 2011. A multiple-kernel support vector regression approach for stock market
price forecasting. Expert Systems with Applications. 38 (3), 2177–2186. https://doi.org/10.1016/j.eswa.
2010.08.004.
∙ 245 ∙
Effect of High-Ammonia Environment on the Gene Networks

Associated with Oxidative Metabolism, Inflammation, and
Apoptosis in “Gut-Brain” Axis
Yutao Li, Jing Wang, Jianxing Wang, Jun Bao *
Abstract
Ammonia (NH3) is an acknowledged environment pollutant in atmosphere with irritating
smell. Previous studies have shown that excessive NH3 has toxic effects on farm animals and
humans. However, the detail toxicity mechanism of NH3 to pigs is still unknown so far. To clarify
the mechanism of NH3 toxicity, we established a porcine exogenous NH3 poisoning model and
assessed the effect of NH3 on the gut-brain axis by transcriptome sequencing, qRT-PCR
validation, histological observation and chemical analysis. Our results showed that after 30 d of
NH3 exposure, 578 differentially expressed genes (DEGs) and 407 DEGs were obtained in the
hypothalamus and jejunum, respectively. These DEGs were enriched into Gene Ontology (GO)
terms associated with inflammation, oxidative metabolism, and apoptosis, and the highly
expressed genes among these DEGs were verified by qRT-PCR. The content of GSH and the
activities of GSH-Px and SOD were significantly decreased, while MDA content was increased
after NH3 exposure. CRF, SP, 5-HT and ghrelin content in serum elevated significantly.
Furthermore, pathologic observation in the NH3 group revealed infiltration of lymphocytes in the
hypothalamus and significant decrease of jejunal epithelial cells. Our results indicated that NH3
exposure could cause changes in transcriptional profiles, pathological damage, oxidative stress
and brain-gut peptide of the pig jejunum and hypothalamus, and lead to the imbalance of the brain-
gut axis through the “oxidative stress-inflammation-apoptosis” interaction network. Our study not
only provides a new perspective for the toxicity assessment of NH3, but also enhances our
understanding ofthe toxicology mechanism of NH3.
Keywords: Ammonia, pig, gut-brain axis, brain-gut peptide, toxicity assessment
1. Introduction
Ammonia (NH3) is a toxic environmental contaminant with irritating smell. It is one of the
representative harmful gases in livestock production, which usually originated from the
decomposition of carbamide in manure and urine (Sutton et al., 2008). Livestock production is
the main way of increasing environmental NH3 around the world. Some reports (An et al., 2019;
Zheng et al., 2019) have confirmed that high-concentration NH3 can cause sterile inflammation
on multiple organs. Poultry workers usually have cough or throat irritation and suffer from
paranasal sinus diseases (Kearney et al., 2014). Ammonia produced by H. pylori in stomach also
mediated gastric mucosal ulcer and cell necrosis and small intestines inflammation. Ammonia
caused ophthalmia and raised the apoptosis rate of neurons cells (Yang et al., 2003) and spleen
(An et al., 2019). They demonstrated that NH3 has a detrimental effect on digestive, nervous
systems. Recent research also found excessive NH3 in pigpens could induce abnormal
physiological responses in pigs (Wang et al., 2020a).
Oxidative stress is an expression form of imbalance between lesion and repair in organisms.
When environmental toxic pollutants caused oxidative stress (Zhao et al., 2020), the changes in
antioxidant indexes could be referred to as a direct manifestation. Ammonia can induce oxidative
stress via triggering the dysfunctions of redox homeostasis. It has been reported excessive NH 3
inhalation could increase MDA contents, decrease the activities of glutathione (GSH) and GSH-
Px in the bursa of fabricius of broiler (Shah et al., 2020). The toxicity of NH3 is related to its
oxidative stress, and oxidative damage may be one of the important ways that NH3 inhalation
causes organ damage. It has been demonstrated that oxidative stress can lead to histopathology
∙ 246 ∙
changes (Xing et al., 2019). According to the previous research, NH3 in the chicken house caused
inflammatory cells infiltration and congestion to appear and lymphocytes to swell in the spleen of
broilers. When chickens were exposed in high NH3 environment, nuclear atrophy and nuclear
membrane overlap in the liver were observed (Xu et al., 2020). In some cells, nucleoli were out
of shape, even disappeared; chromatin concentrated onto the nuclear membrane. However, the
pathological evidence is insufficient for a better assessment of the toxicity of NH3.
Transcriptomics profiling technique is a new high-tech tool to locate precisely differentially
expressed gene (Chen et al., 2019). Transcriptome analysis has a ubiquitous application in toxic
identification of environment pollutants (González-Velasco et al., 2020). Transcriptomics
profiling provides biologic information on mechanisms of action of oxidative stress induced by
toxicant at the genetic level (Garg et al., 2016). Pig is an important kind of livestock to agricultural
development. Thus, providing a harmless living environment is extremely necessary for animal
welfare and healthy farming. In medical science, pigs are ideal animal models in the field on the
mechanism of occupational poisoning. So far, in the field of environmental toxicology, the effect
of NH3 toxicity on the gut-brain axis of pigs has not been reported. Therefore, in this study, we
evaluated the toxicity effect of NH3 inhalation on the hypothalamus and jejunum of fattening pigs.
Our results provide novel views into the specific mechanism of NH3 toxicity and provide reference
for animal production management.
2.1. Replication of the exogenous NH3-poisoning model
Twelve 125-days-old three-way cross female pigs were randomly divided into two groups on
average, called control group (NH3 concentration <6.6 ppm) and NH3 group (117.6 ppm < NH3
concentration <121.9 ppm), respectively. Exogenous NH3 in the NH3 group was provided to the
environmental control chambers for up to 8 h per day. During the whole experiment, all pigs were
given enough feed and drinking water ad libitum. All processes were approved by Institutional
Animal Care and Use Committee of the Northeast Agricultural University (Number SRM-06).
The formal experiment lasted for 30 d. On the 30th d, all the pigs were euthanized after blood
collection from the heart, and the hypothalamus and jejunum was immediately separated. The
tissues were then homogenized under ice conditions, fixed at 10% formalin, and stored at -80 °C
after quick freezing, respectively.
2.2. Measurement of brain-gut peptide contents
The contents of 5-hydroxytryptamine (5-HT), corticotropin releasing factor (CRF), substance
P (SP) and ghrelin in serum were measured using commercial ELISA assay kits. Briefly, each
sample was mixed with specific solutions, and then was incubated at 37 °C. All procedures above
was consistent with the manufacturer's instructions. The OD values were determined. All brain-
gut peptide contents were calculated based on the corresponding standard curves.
2.3. Transcriptomics profiling technique
mRNA was decomposed into small fragment. Random primer and related reverse
transcriptase were used to synthesize the first strand cDNA. Subsequently, the synthesis of the
second strand cDNA was actualized. High-speed Illumina HiseqTM Sequencer was used to filter
and sort out the sequencing data. We applied DESeq2 to analyze those genes expressed
differentially when the | log2 (Fold Change) | greater than 1.5 and P value less than 0.05. The
differentially expressed genes (DEGs) were expressed by the heat maps with different colors. All
GO terms were divided into biological process, cellular component and molecular function.The
number of DEGs in jejunum and hypothalamus were calculated based on the R package. All
sequencing data was uploaded to NCBI database (BioProject Accession: PRJNA741846).
2.4. qRT-PCR verification of DEGs
Total RNA was extracted and the relative concentration and purity were measured. The
reverse transcription of mRNA was performed as directed by the manufacturer. The β-actin was
∙ 247 ∙
the internal reference gene. All the primers were designed by Primer Premier Software 5.0. The
qRT-PCR was performed by ABI QuantStudio 3 (Applied Biosystems Co. Ltd., USA). The
△△
relative expression level of mRNA was calculated by the2− Ct method.
2.5. Data-processing and statistical analysis
The data was analyzed by the SPSS 21.0. All results showed a normal distribution and were
described as the Mean ± Standard deviation. In our study, t-test was conducted to examine the
significance. Bar charts with the “*” symbol represented there was statistically significant
difference between control group and NH3 group. Statistical significance was defined as P<0.05.
3.1. Effects of NH3 exposure on brain-gut peptide contents
The contents of brain-gut peptide (5-HT, CRF, SP and ghrelin) in serum were measured to
examine the detail effect of NH3 on gut-brain axis. Compared with the control group, the relative
contents of 5-HT, CRF, SP and ghrelin in serum increased significantly (P<0.05) after 30 d of
NH3 exposure (Table 1).
Table 1. Effect of NH3 exposure on brain-gut peptide contents in serum.
Substance Group Content
5-HT (ng L-1) Control 63.560±1.386 *
NH3 78.948±2.046 **
CRF (ng L-1) Control 50.500±1.249 **
NH3 170.500±9.350 *
SP (ng L-1) Control 54.450±2.076 *
NH3 74.650±2.133 *
Ghrelin (pg mL-1) Control 352.400±12.954 **
NH3 670.000±9.083 *
3.2. DEGs analysis
A total of 578 DEGs and 407 DEGs was found in hypothalamus and jejunum, respectively.
For hypothalamus, 413 up-regulated genes and 165 down-regulated genes were annotated.
Similarly, in jejunum, 176 up-regulated genes and 231 down-regulated genes were annotated.
3.3. Gene ontology (GO) enrichment of DEGs
In the hypothalamus, top 5 GO terms for biological process were as follows (Table 2): Positive
regulation of transcription by RNA polymerase II; Regulation of transcription, DNA-templated;
Signal transduction; Positive regulation of transcription, DNA-templated; Protein
phosphorylation. Meanwhile, top 5 GO terms for cellular component were as follows: Membrane,
Integral component of membrane, Nucleus, Plasma membrane, and Cytoplasm. Top 5 GO terms
for molecular function were as follows: Protein binding; Metal ion binding; DNA binding;
Identical protein binding; Nucleotide binding. In the jejunum (Table 3), positive regulation of
transcription by RNA polymerase II was also the most typical biological process, membrane was
the most major category for cellular component, protein binding was the most representative GO
term for molecular function. Among the top 20 GO terms, in the hypothalamus, extracellular space
had the most gene number, external side of plasma membrane had the most significantly
difference while alpha-beta T cell receptor complex had the highest rich factor (Table 2).
Similarly, in the jejunum, the genes mapped into protein binding had maximum number, cell wall
disruption in other organism, ethanolaminephosphotransferase activity and thyroxine 5'-
deiodinase activity all had the highest rich factor, in parallel, Z disc was most significantly
differential among the top 20 GO terms (Table 3).
∙ 248 ∙
Table 2. Gene ontology (GO) enrichment statistics for hypothalamus (P<0.001).

Gene Rich
GO term
number factor
External side of plasma membrane 24 0.102
Alpha-beta T cell receptor complex 4 0.800
T cell differentiation 8 0.242
Synaptic vesicle 12 0.146
Synaptic vesicle membrane 8 0.235
RNA polymerase II activating transcription factor binding 8 0.222
T cell receptor complex 5 0.454
DNA-binding transcription activator activity, RNA polymerase II-
25 0.074
specific
Transmembrane signaling receptor activity 16 0.099
Skeletal muscle cell differentiation 8 0.186
Myelination 8 0.186
Paranode region of axon 4 0.400
Neuron projection 17 0.080
Axon guidance 12 0.102
Learning 7 0.166
Nervous system development 12 0.100
Regulation of T cell activation 4 0.363
Extracellular space 46 0.050
Immunological synapse 6 0.188
Cytoskeletal anchoring at nuclear membrane 4 0.333
Table 3. Gene ontology (GO) enrichment statistics for jejunum (P<0.001).
Gene Rich
GO term
number factor
Z disc 10 0.149
Defense response to virus 13 0.100
Extracellular region 34 0.046
Response to virus 8 0.137
Extracellular space 37 0.040
Calcium ion binding 28 0.045
Regulation of store-operated calcium entry 4 0.363
Integrin binding 9 0.105
Long-chain fatty acid biosynthetic process 3 0.600
Protein binding 87 0.029
Negative regulation of viral genome replication 5 0.192
ISG15-protein conjugation 3 0.500
Angiogenesis 10 0.079
Negative regulation of angiogenesis 7 0111
Postsynaptic density 10 0.076
Regulation of cytosolic calcium ion concentration 5 0.166
Cell wall disruption in other organism 2 1.000
Ethanolaminephosphotransferase activity 2 1.000
Thyroxine 5'-deiodinase activity 2 1.000
Sarcomere organization 5 0156
∙ 249 ∙
3.4. Validation of transcriptomics sequencing using qRT-PCR

As expected, relative mRNA levels of OASL, ISG15, S100A12, DIO1, CYP1B1, and
EPHA2 in the jejunum and relative mRNA levels of CD27, CD3G, E2F2, ACP5, IL2RB, BLK,
and CYP3A39 in the hypothalamus were up-regulated in the NH3 group. Compared to the control
group, relative mRNA level of ERO1A, EPHA7, TREM2 in the jejunum and relative mRNA level
of PFKFB1 and LIPM in the hypothalamus were down-regulated in the NH3 group (Figure 1).
These results above were consistent with the transcriptomics profiling results, indicating that our
transcriptomics analysis had a better credibility.
Figure 1. Change trend of DEGs expression level in hypothalamus (top) and

jejunum (bottom).
4. Conclusions
In conclusion, excessive NH3 exposure could induce oxidative stress in the jejunum and
hypothalamus of fattening pigs, and cause histology injury and changes in brain-gut peptide,
leading to the imbalance of the gut-brain axis. Furthermore, we established for the first time
transcriptome analysis database of hypothalamus and jejunum of pigs exposed to high NH3
environment, and demonstrated that chronic NH3 poisoning could induce the imbalance of the
gut-brain axis through the “oxidative stress-inflammation-apoptosis” interaction network. The
results not only provided a new perspective for the toxicity assessment of NH3, but also enrich the
toxicology mechanism of NH3.
∙ 250 ∙
Acknowledgements
This work was supported by the Earmarked Fund for China Agriculture Research System
(Project No. CARS-35-05B).
References
An, Y., Xing, H., Zhang, Y., Jia, P., Gu, X., Teng, X., 2019. The evaluation of potential immunotoxicity
induced by environmental pollutant ammonia in broilers. Poult. Sci. 98, 3165–3175.
Chen, J., Xu,Y., Han, Q., Yao, Y., Xing, H., Teng, X., 2019. Immunosuppression, oxidative stress, and
glycometabolism disorder caused by cadmium in common carp (Cyprinus carpio L.): Application of
transcriptome analysis in risk assessment of environmental contaminant cadmium. J. Hazard. Mater. 366, 386–
394.
González-Velasco, O., Papy-García, D., Le Douaron, G., Sánchez-Santos, J.M., De Las Rivas, J., 2020.
Transcriptomic landscape, gene signatures and regulatory profile of aging in the human brain. Biochim
Biophys Acta Gene Regul Mech. 1863(6), 194491.
Garg, R., Shankar, R., Thakkar, B., Kudapa, H., Krishnamurthy, L., Mantri, N.,Varshney, R.K., Bhatia,
S., Jain, M., 2016. Transcriptome analyses reveal genotype-and developmental stage-specific molecular
responses to drought and salinity stresses in chickpea. Sci. Rep. 6, 19228.
Kearney, G.D.,Shaw, R., Prentice, M., Tutor-Marcom, R., 2014. Evaluation of respiratory symptoms and
respiratory protection behavior among poultry workers in small farming operations. J. Agromed. 19, 162-170.
Shah, S.W.A., Chen, J., Han, Q., Xu, Y., Ishfaq, M., Teng, X., 2020. Ammonia inhalation impaired
immune function and mitochondrial integrity in the broiler bursa of fabricius: Implication of oxidative stress
and apoptosis. Ecotoxicol. Environ. Saf. 190, 110078.
Sutton, M.A., Erisman, J.W., Dentener, F., Möller, D., 2008. Ammonia in the environment:from ancient
times to the present. Environ. Pollut. 156, 583–604.
Wang, X.,Wang, M., Chen, S., Wei, B., Gao, Y., Huang, L., Liu, C., Huang, T., Yu, M., Zhao, S., Li, X.,
2020a. Ammonia exposure causes lung injuries and disturbs pulmonary circadian clock gene network in a pig
study. Ecotoxicol. Environ. Saf. 205, 111050.
Xu, Y., Li, Z., Zhang S., Zhang, H., Teng, X., 2020. miR-187-5p/apaf-1 axis was involved in oxidative
stress-mediated apoptosis caused by ammonia via mitochondrial pathway in chicken livers. Toxicol. Appl.
Pharmacol. 388, 114869.
Yang, L., Omori, K., Omori, K., Otani, H., Suzukawa, J., Inagaki, C., 2003. GABAC receptor agonist
suppressed ammonia-induced apoptosis in cultured rat hippocampal neurons by restoring phosphorylated
BAD level. J. Neurochem. 87, 791–800.
Zhao, H., Wang, Y., Guo, M., Liu, Y., Yu, H., Xing, M., 2020. Environmentally relevant concentration
of cypermethrin or/and sulfamethoxazole induce neurotoxicity of grass carp: Involvement of blood-brain
barrier, oxidative stress and apoptosis. Sci. Total Environ. 143054.
Zheng, S., Jin, X., Chen, M., Shi, Q., Zhang, H., Xu, S., 2019. Hydrogen sulfifide exposure induces jejunum
injury via CYP450s/ROS pathway in broilers. Chemosphere 214, 25–34.
Xing, H., Peng, M., Li, Z., Chen, J., Zhang, H., Teng, X., 2019. Ammonia inhalation-mediated mir-202-
5p leads to cardiac autophagy through PTEN/AKT/mTOR pathway. Chemosphere 235, 858–866.
∙ 251 ∙
Effects of Three Ventilation Methods on the Environmental

Conditions of Pig Houses, Pig Production Performance and
Abnormal Behavior during Winter Fattening
Zhifang Shi, Xuanyang Li, Lei Xi *
College of Animal Science and Technology, Henan University of Animal Husbandry and Economy,
* Corresponding author. Email: 80332@hnuahe.edu.cn
Abstract
A reasonable ventilation design is important to ensure the environmental quality and animal
health in livestock and poultry houses. In this study, 315 90-day-old “Duroc × Landrace ×
Yorkshire (DLY)” three-way crossbred healthy finishing pigs with similar body conditions and
similar body weight (39.5 ± 1.3 kg) were selected and evenly distributed to three experimental
pig houses. The environmental parameters in the house, behavior of pigs, and production
performance were monitored for 28 consecutive days. The temperature and humidity of pig houses
showed the same daily changes using the three ventilation methods, but the average temperature
and relative humidity of the tunnel-ventilated pig house were 19.22°C and 60.79%, respectively,
which were significantly higher than those of the naturally and mechanically ventilated pig houses
(P<0.05). No significant difference was found in the surface temperature of pigs and the wind
speed in the house using the three ventilation methods (P>0.05), but the NH3 and CO2
concentrations in the tunnel-ventilated pig house significantly reduced (P<0.05). Also, no
significant difference in feed intake was observed using the three methods (P>0.05). However,
the average daily gain of pigs in the tunnel-ventilated pig house increased by 14.88% (P<0.05)
compared with that in the naturally ventilated pig house. The feed-to-gain ratio (F/G) was reduced
by 11.44% in the tunnel-ventilated pig house compared with the naturally ventilated pig house
(P<0.05). The frequencies of the five abnormal behaviors of pigs, including gnawing, abdomen
rubbing, tail biting, hoof biting, and ear biting, in the tunnel-ventilated pigsty decreased
significantly (P<0.05). The tunnel ventilation system is conducive to improving the pig house
environment, reducing abnormal behavior of pigs, and improving pig production performance.
Keywords: Ventilation of pig house, tunnel ventilation, pig house environment, pig behavior,
production performance
1. Introduction
Ventilation is an important measure to ensure the environmental quality and animal health in
livestock and poultry houses. In the cold winter, the ventilation system of the livestock and poultry
house should not only remove the dirty air in time and maintain a good environmental quality but
also maintain the temperature in the house without affecting the health and growth of the animals.
Therefore, a reasonable ventilation design is particularly important in winter. At large-scale pig
farms, pigs are generally kept in closed houses with high density. The entire production
management process and pig activities are performed in the house. The environment in the house
is polluted, and the ventilation system plays an important role in the entire production process. At
present, the common ventilation methods of pig farms are natural and mechanical ventilation
(Wang et al., 2017; Gao et al., 2018; Zhu, 2007). Natural ventilation relies on wind pressure and/or
temperature difference between indoor and outdoor to create flow, which exchanges air inside and
outside the house. Regarding the power source of airflow, ventilation can be divided into wind
pressure ventilation and hot pressure ventilation (Zhao et al., 2009). The full and reasonable
application of natural ventilation is an economical and energy-saving measure, but its
disadvantage is that it is controlled by natural conditions and is not suitable to control the
microenvironment in the house. Regarding different driving principles, mechanical ventilation
∙ 252 ∙
can be divided into three types: negative-pressure ventilation, positive-pressure ventilation, and
equal-pressure ventilation. At present, negative-pressure ventilation is most widely used in farms
in China. This kind of ventilation system is suitable for pig houses with a span of less than 20 m,
but it is not suitable for windy and severely cold areas because it cannot control the temperature
and humidity of the air entering the house, thereby requiring a large economic investment.
Ventilation and heat exchange via tunnels take advantage of the higher temperature of the bottom
soil compared with the outdoor air temperature in winter. These tunnels use the bottom soil as a
heat source to increase the heat of the outdoor air flowing in the tunnel, thereby increasing the
indoor temperature and achieving the effect of ventilation and heat exchange. Tunnel ventilation
technology is widely used in public facilities and buildings due to its advantages of energy saving
and environmental protection (Wang et al., 2011; Hu, 2012). However, only three reports are
available on its application in the farming industry. Liu (1987) monitored the summer
environmental parameters of Shenzhen tunnel-ventilated pig houses and confirmed the role of
tunnel ventilation in the summer cooling of the farming industry. Wang et al. (2002) designed
tunnel air conditioning to improve the thermal environment of pig houses. Ma (1997) discussed
the problem of using layered tunnel wind to cool the livestock and poultry houses in the summer
of southern China and conducted a simulated calculation. However, the ventilation effect of tunnel
ventilation in winter pig houses and its influences on the microenvironment of pig houses, pig
production performance, and abnormal behavior of pigs have not been reported yet. This study
aimed to find a ventilation system that could improve the environment of piggery, reduce the
abnormal behavior of pigs, and improve the production performance of pigs. This study compared
natural ventilation, mechanical ventilation, and tunnel ventilation, and explored the effects of
three ventilation methods on the pig house microenvironment, behavior of pigs, and production
performance.
2.1. Basic overview of pig houses
The experiment was performed at a large-scale pig farm in Henan Province (Figure 1). The
experimental season was winter, and the experimental period was 28 days. During the test, the
average temperature outside the house was 4.5 ± 1.7 °C, and the average humidity outside the
house was 48.5 ± 9.7%. In addition, the average wind speed outside the house was 2.6 ± 1.1
m s-1. The three experimental pig houses had the same structure (Figure 1), and all pig houses
were oriented east–west with a length × width × eave height of 28 × 6 × 3 m3. The pig house wall
within 1.0 m from the ground was a brick-concrete structure, and the upper wall consisted of 10-
cm-thick, colored steel polystyrene sandwich panels. The roof was made of light steel brackets
and 25-cm-thick, colored steel polystyrene sandwich panels. Seven sliding windows (1.2 × 0.8
m2) were present on the south wall of pig houses, and seven sliding windows (0.6 × 0.6 m2) on
the north wall. Seven pens were present in each house, which were arranged in a single row, and
each pen was 5 × 4 m2 (length × width). The first pig house adopted natural ventilation. In the
first house, the windows on the north and south walls were opened for convection ventilation
(ventilation for 15 min at 9:00 in the morning and for 30 min at 13:00), and the windows were
closed during the rest of the day. The second pig house used negative-pressure longitudinal
mechanical ventilation. In the second house, the ventilation machine was turned on at 9:00 for 0.5
h and turned on again after an interval of 2.5 h; the cycle continued until the machine was turned
off at 20:00 in the evening. The third pig house was the tunnel-ventilated pig house, in which a
tunnel and a fan were used for ventilation. The tunnel was made of a cylindrical cement pipeline
(1 m in diameter and 30 m in length) and buried 5 m underground. Fans were installed in the
tunnel and run all day long. The wind entered the house through the tunnel for ventilation.
∙ 253 ∙
a. vent; b. Window; c. Fan; d. Mange; e. Water fountain; f. Leaky floor; g. Aisle

Figure 1. Floor plan of structural of pig house.
2.2. Pig herd and feeding management
In the experiment, 315 90-day-old "Duroc × Landrace × Yorkshire (DLY)” three-way
crossbred healthy finishing pigs with similar body conditions and body weight (39.5 ± 1.3 kg)
were randomly assigned to 3 experimental pig houses. Each house had 105 pigs, with 15 pigs in
each pen. During the experiment, the pigs were free to eat and drink, and they were fed and
managed according to routine procedures. The daily ration formula and nutritional level are shown
in Table 1.
Table 1. The composition (mass fraction) and nutrient levels of the basal diets.
Ingredient Composition Nutrient level Composition
Corn 65.0 DE (MJ Kg-1) 13.19
Wheat bran 8.0 CP (％) 17.23
Soybean 24.0 Ca (％) 0.80
Salt 0.3 P (％) 0.36
Premix 2.7 K (％) 0.53
Sum 100.0 Na (％) 0.14
Cl (％) 0.61
Lys (％) 0.79
Met + Cys (％) 0.59
2.3. Measurement method
The five-point method was used to evenly distribute the monitoring points in pig houses. The
monitoring points were 0.5 m above the ground. The monitoring indicators included temperature,
humidity, wind speed, NH3 concentration, and CO2 concentration in the house. An automatic
temperature and a humidity recorder (LGR-WSD20, Hangzhou Luge Technology Co., Ltd.,
China) were used to automatically record a set of temperature and humidity data in the house
every 5 min. A thermal breeze meter (Testo 404-v1, Germany) was used to measure the wind
speed in the house. An NH3 detection tube (CZY50, Hebi Yonghua Gas Detection Co., Ltd.,
China) and portable CO2 analyzer (JSA8, Shenzhen, China) were used to measure the changes in
NH3 and CO2 concentrations in the house, respectively. An infrared thermometer (TYR300,
Shanghai YourGood Company, China) was used to measure the surface temperature of pigs. The
measurement time of wind speed, NH3 concentration, CO2 concentration, and pig body surface
temperature in the house was three times a day at 8:00, 12:00, and 18:00 with continuous
monitoring for 28 days.
For determining pig production performance and abnormal behavior, automatic feeding
equipment and a 9SC-05 pig feed selection and automatic measurement system (Anhui Bodhi
Fruit Husbandry Technology Co., Ltd., China) were used to measure and calculate the starting
weight, final weight, total gain, average daily feed intake, average daily gain (ADG), and feed-to-
∙ 254 ∙
gain ratio (F/G) during the test period. A digital behavior image recorder (Noldus Information
Technology Company, Netherlands) was used to record and analyze the frequency of abnormal
behaviors, such as gnawing, abdomen rubbing, tail biting, hoof biting, and ear biting.
2.3. Data processing
The test data were analyzed by one-way analysis of variance and generalized linear model
using statistical software SPSS23.0. Multiple comparisons were performed by the least significant
difference method with a significance level of P<0.05. The test results were expressed as mean ±
standard deviation. GraphPad Prism 6.0 software was used for plotting.
3.1. Effect of ventilation methods on pig house environment
The effects of the three ventilation methods on pig house temperatures are shown in Figure
2a. The pig house temperatures using the three methods presented the same diurnal variation. The
highest temperatures appeared between 14:00 and 16:00, and the lowest temperatures appeared
between 6:00 and 10:00. The average temperatures of pig houses for the three ventilation methods
were all significantly higher than the temperature outside the houses during the study (6.8 ± 3.2°C)
(P<0.05), indicating that the three ventilation methods resisted changes in the outside
temperature. During the study, the average temperature of tunnel-ventilated, ventilated pig houses
was 19.22 ± 0.11°C, 17.26 ± 0.51°C, and 17.74 ± 0.29°C, respectively. The average temperature
of the tunnel-ventilated pig house was significantly higher than that of the naturally and
mechanically ventilated pig houses (P<0.05). The effect of the three ventilation methods on the
surface temperature of pigs is shown in Figure 2b.
The three ventilation methods had no significant effect on the surface temperature (P >0.05),
and all ventilation methods ensured stable surface temperature of the pigs. The relative humidity
changes in the house using the three ventilation methods are shown in Figure 2c. The three
ventilation methods had a significant impact on the relative humidity in pig houses. The relative
humidity in the tunnel-ventilated pig house was the highest, reaching 60.79%, which was
significantly higher than that of the naturally and mechanically ventilated pig houses (P<0.05).
The influence of the three ventilation methods on the wind speed in pig houses is shown in Figure
2d. The wind speed in the tunnel-ventilated pig house was 0.14 m s-1, which was not significantly
different from the wind speed in the naturally and mechanically ventilated pig houses (P >0.05).
The winter speeds in the three conditions were all less than the maximum wind speed limit of 0.15
m s-1 in the winter fattening pig house.
The effect of the three ventilation methods on the NH3 content in pig houses is shown in
Figure 2e. The average concentration of NH3 in pig houses using the natural, mechanical, and
tunnel ventilation methods was 18.52 ± 4.80 mg m-3, 24.63 ± 5.79 mg m-3, and 23.05 ± 5.07 mg
m-3, respectively. The average content of NH3 in the tunnel-ventilated pig house was the lowest,
which was significantly lower than that pig houses using the other two ventilation methods
(P<0.05), while the average contents of NH3 in the naturally and mechanically ventilated pig
houses were not significantly different from each other (P>0.05). The impact of the three
ventilation methods on the CO2 content in pig houses is shown in Figure 2f. Similar to the effects
on NH3, the CO2 content of the tunnel-ventilated pig house was the lowest at 1390 ± 241 mg m-3,
which was significantly lower than that of pig houses using the other two ventilation methods
(P<0.05).
The optimal temperature for winter fattening pig houses is 19 ℃–22 ℃ (Si et al., 1993), and
the optimal humidity range is 50%–80% (Yuan et al., 2014). In this study, the temperature and
humidity of pig houses using the three ventilation methods were at a level suitable for the growth
of finishing pigs. However, the temperature (19.22 ± 0.11 ℃) and relative humidity (60.79 ±
2.82%) of the tunnel-ventilated pig houses were significantly higher than those using the other
two ventilation methods. This might be because the outside air passed through the 30-m-long
∙ 255 ∙
ventilation tunnel to enter the pig house; besides absorbing the heat in the ventilation tunnel, the
outside air removed the water vapor in the ventilation tunnel, thus increasing the humidity of the
air in the pig house.
a. Indoor temperature b. Surface temperature c. Relative humidity d. Wind speed e. NH 3

concentration f. CO2 concentration
Figure 2. Environmental characteristics of the experimental piggery.
The average concentration of NH3 in the tunnel-ventilated pig house was 18.52 mg m-3, which
was lower than the upper limit of the daily average value of NH3 concentrations specified in the
environmental quality standards for livestock and poultry farms promulgated by the Ministry of
Agriculture of 25 mg m-3 (NY/T 388-1999, 1999) and close to the average range of NH3
concentration in winter pig houses of 13.89 mg m-3measured by Duchaine et al. (2006). In
contrast, the average NH3 concentration in the naturally ventilated and mechanically ventilated
pig houses reached 24.63 mg m-3 and 23.05 mg m-3, respectively, which were close to the upper
limit. Studies have shown that the CO2 concentration in pig houses should not exceed 1500 mg
m-3 (Ma et al., 2005); otherwise, it affects the production efficiency of pigs. The CO2 concentration
of the tunnel-ventilated pig house was 1390 ± 241 mg m-3, which was significantly lower than the
CO2 content in pig houses using the other two ventilation methods (P<0.05). These findings
∙ 256 ∙
indicated that the use of tunnel ventilation was beneficial to improve the ventilation efficiency
and air environment quality of pig houses.
3.2. Effect of ventilation methods on pig production performance
The effects of ventilation methods on pig performance are shown in Table 2. Before the
experiment, the initial weights of pigs in the three pig houses were not significantly different.
After the 28-day test period, the average pig weight in the tunnel-ventilated pig house reached
59.80 ± 1.54 kg, and the average daily weight gain reached 0.72 ± 0.03 kg, which was significantly
higher than that of the naturally and mechanically ventilated pig houses (P<0.05). During the
experiment, no significant difference was found between the total feed quality and the average
daily feed quality of pigs in the three experimental pig houses (P>0.05). The F/G of pigs in the
tunnel-ventilated pig house was 2.31 ± 0.14, which was significantly lower than that of the
naturally and mechanically ventilated pig house (P<0.05).
Table 2. Influence of ventilation on production performance of pigs.
Ventilation way Tunnel ventilation Tunnel ventilation Tunnel ventilation
Initial weight (kg) 39.75±1.20a 39.79±0.82a 39.60±1.41a
Final weight (kg) 59.80±1.54a 57.43±1.13b 58.35±1.20ab
Daily weight (kg) 0.72±0.03 a 0.63±0.05 b 0.67±0.04 ab
Total feed intake (kg) 48.71±2.43a 46.25±2.33a 47.84±2.08a
Daily feed quality (kg) 1.74±0.11 a 1.65±0.08 a 1.71±0.13 a
Feed conversion ratio 2.31±0.14 a 2.62±0.16 b 2.55±0.14ab
Note: the test period was 28 days. Different letters are marked in the same column of Numbers,
indicating significant differences between the two groups of data (P<0.05). The same number in
the same column is marked with the same letter, indicating no significant difference between the
two groups of data (P >0.05).
The feed intake of pigs did not show a significant difference using the three ventilation
methods, but the average daily weight gain of pigs in the tunnel-ventilated pig house was
significantly higher than that in the naturally ventilated pig house with an increase of 14.88%
(P<0.05), demonstrating that the pigs in the tunnel-ventilated house had a better feed conversion
rate. The reason for this difference might be that the tunnel ventilation provided suitable
environmental conditions for the pigs. A higher wind speed indicated better ventilation in pig
houses, and the level of ventilation in pig houses directly determined the various harmful gases in
the houses, such as NH3 concentration, as well as dust particles. When the concentration of NH3
in pig houses was too high, it reduced the function of the mucosal removal system and caused the
animal to be infected with respiratory diseases (Albert et al., 2015), thereby increasing the disease
incidence of animals and reducing production performance.
3.3. Effect of ventilation on abnormal behavior of pigs
The effect of the tunnel ventilation method on the behavior of pigs is shown in Figure 3. The
frequency of the five abnormal behavior of pigs, including gnawing, abdomen rubbing, tail biting,
hoof biting, and ear biting, in the tunnel-ventilated pigsty was the lowest (P<0.05). The frequency
of gnawing behavior in the tunnel-ventilated pigsty was 56.8% lower than that in the naturally
ventilated pig house (P<0.05) and 29.6% lower than that in the mechanically ventilated pig house
(P<0.05). The frequency of abdomen-rubbing and tail-biting behaviors in the tunnel-ventilated
pigsty was 19.9% and 15.7% lower than that in the mechanically ventilated pig house (P<0.05),
respectively. The frequency of hoof-biting behavior in the tunnel-ventilated pigsty was lower than
that in the naturally ventilated pig house and mechanically ventilated pig house by 24.9% and
25.5%, respectively (P<0.05). Moreover, the frequency of ear-biting behavior in the tunnel-
ventilated pigsty was 44.5% lower than that in the mechanically ventilated pig house (P<0.05)
and was not significantly different from that in the naturally ventilated pig house (P >0.05). No
difference was found in the frequency of abnormal behaviors of pigs in the naturally and
∙ 257 ∙
mechanically ventilated pig houses (P >0.05).
Figure 3. Influence of ventilation on abnormal behavior of pigs.

Studies have found that the environmental temperature is closely related to the various
behaviors of pigs. When the environmental temperature changes, the behaviors of animals, such
as feeding and lying, also change. At the same time, the change in external humidity promotes or
inhibits the influence of temperature on the behavior of pigs (Heetkamp et al., 2005). Under
suitable temperature conditions, the daytime lying time of finishing pigs is approximately 85%
(Ekkel et al., 2003). Every time the temperature of pig houses increases by 1°C, the side-lying
rate increases by 0.4% (Huynh et al., 2004), while the contact rate between pigs decreases by
3.7% (Aarnink et al., 2001). In the present study, the abnormal behavior of pigs in the tunnel-
ventilated pig house was significantly less than that in the naturally and mechanically ventilated
pig houses (P<0.05). This might be because the temperature of the tunnel-ventilated pig house
was significantly higher than that of naturally and mechanically ventilated pig houses. Thus,
tunnel ventilation improves the comfort of the pigs by improving the environment in the house,
thus avoiding the frequency of abnormal behaviors of pigs and improving the welfare of pigs.
4. Conclusions
All three ventilation methods ensured a reasonable pig house environment during the test
period and normal production activities of the pigs. Among the ventilation methods, the tunnel-
ventilated pig house had higher temperature, higher humidity, suitable ventilation speed, lower
NH3 concentration, and lower CO2 concentration, which provided a better growth environment
for pigs. The incidence of abnormal behavior of pigs living in the tunnel-ventilated pig house was
low. With tunnel ventilation, the pigs showed a higher ADG and final body weight, and the F/G
of pigs was lower than that of the naturally ventilated pig house. This finding indicated that the
tunnel ventilation system was beneficial for improving the pig house environment and the
production performance of pigs.
Acknowledgements
This work was supported by the National Key Research and Development Program of China
(2018YFD0501103) and the Key Scientific Research Project of Higher Education of Henan
Province (20B230005).
References
Aarnink, A. J. A., J. W.Schrama, E. J. E. Verheijen, J. Stefanowska, 2001. Pen fouling in pig houses
affected by temperature. Michigan USA, 2001,180–186.
Albert, W., M. Julio, W. G. Peter, K. Groot, W. M. O. Nico, and J. A. A. André, 2015. Emission of
particulate matter from animal houses in the Netherlands. Atmospheric Environment. 2015(111), 202–212.
∙ 258 ∙
Duchaine, C., Y. Grimard, Y. Cormier, 2000. Influence of building maintenance, environmental factors
and seasons on airborne contaminants of swine confinement buildings. A Journal for the Science of
Occupational and Environmental Health and Safety. 61(1), 56–63.
Ekkel, E. D., H. A. M. Spoolder, I. Hulsegge, and H. Hopster, 2003. Lying characteristics as determinants
for space requirements in pigs. Applied Animal Behaviour Science. 80(1), 19–30.
Gao, Y., Y. P. Diao, C. G. Lin, Y. X. Liu, C. M. Guo, M. G. Lei, Y. Tong, and X. Li, 2018. Monitoring
and analysis of thermal environment and harmful gases in mechanically ventilated multistory pig buildings.
Transactions of the Chinese Society of Agricultural Engineering. 34(4), 239–247.
Hu, S., 2012. Study on the Effect of Tunnel Wind on Natural Ventilation of Office Building. Ph.D Thesis,
Tianjin University, Tianjin, China.
Huynh, T., A. Aarnink, W. Gerrits, M. Heetkamp, T. Canh, H. Spoolder, B. Kemp, and M. Verstegen,
2005. Thermal behaviour of growing pigs in response to high temperature and humidity. Applied Animal
Behaviour Science. 91(1–2), 1–16.
Huynh, T., A. Aarnink, H. Spoolder, M. Verstegen, and B. Kemp, 2004. Effects of floor cooling during
high ambient temperatures on the lying behavior and productivity of growing finishing pigs. Transactions of
the ASABE. 47(5), 1773–1782.
Liu, Z., 1987. Research on the use of underground natural cold sources in closed pig houses to prevent
and cool heat. Journal of Agricultural Machinery. 18(3), 63–69.
Ma, C. W., 1997. Study on Earth Cooling of Ventilation Air for Animal House in South China in Summer.
Transactions of the Chinese Society of Agricultural Engineering. 13(5), 173–176.
Ma, C. W., X. W. Miao, 2005. Agricultural bioenvironmental engineering. Beijing: China Agriculture
Press.
NY/T 388-1999. Environmental quality standards for livestock and poultry farms. Beijing: Quality
Standards Office of Ministry of Agriculture of the People's Republic of China.
Si, J. L., Y. L. Wang, and H. F. Li, 1993. Current Farm-house Ventilation Technology. Transactions of
the Chinese Society of Agricultural Engineering. 9(5), 65–71.
Wang, K. T., 2011. Research on the Effect of the Air Cooling from Underground Tunnels’ Shape and
Ventilation Time. Changsha: Hunan University.
Wang, M. Z., J. J. Liu, W. Y. Zhao, Z. H. Chen, K. Ren, X. D. Cui, G. Wang, and Z. H. Wu, 2017. CO2
and NH3 emissions and water consumption at mechanically ventilated finishing pig house in Beijing equipped
with pull-plug manure removal system. Transactions of the Chinese Society of Agricultural Engineering,
33(17), 152–160.
Wang, Z. Y., B. Gu, and G. Zheng, 2002. A preliminary study on improving the thermal environment of
pig house with tunnel air conditioning. Underground Space. 2002(2), 165–170.
Yuan, Y. M., L. L. Sun, S. Q. Pan, X. W. Li, H. Z. Liu, and C. Y. Wang, 2014. Impact of tunnel ventilation
on thermal environment in solar heated swine housing. Transactions of the Chinese Society of Agricultural
Engineering. 30(16), 213–220.
Zhao, Y. H., 2009. Livestock and poultry environmental health and pasture design. Zhengzhou: Henan
Science and Technology Press.
Zhu, Z. P., 2007. Study on measurement of natural ventilation and harmful gas emissions in pig houses
by tracer method. Ph.D. Thesis, China Agricultural University, Beijing, China.
∙ 259 ∙
Relationship between Stereotypic Behaviors, Physiological

Characteristics, and Immune System Biomarkers of Confined
Pregnant Sows
Lei Pan, Susu Ding, Haoyang Nian, Runxiang Zhang, Xiang Li, Jun Bao *
Abstract
In order to clarify the responses of stereotypic behaviors and the levels of serum physiology,
immunity of pregnant sows housed in stalls, fifty pregnant sows (Large White × Landrace) were
selected to observe their behaviors in early, middle, and late gestation (27th, 62nd, and 91st day),
and then measured the concentrations of serum physiological parameters (cortisol, Pig-MAP, and
CRP) and immune parameters (IgA, IgG, IL-6, IL-10, TNF-α, and IFN-γ). The results showed
that with the progress of gestation, lateral lying behavior increased significantly (p<0.05), while
standing and ventral lying behavior decreased (p<0.05). Sham-chewing, bar-biting, trough-biting,
rooting behavior and frothy saliva score was significantly reduced (p<0.05). Furthermore, there
was no significant difference in physiological and immune levels in different gestational periods
(p >0.05). Sham-chewing was significantly positively correlated with serum cortisol (r = 0.314,
p<0.001), IL-6 (r = 0.244, p = 0.003), IL-10 (r = 0.169, p = 0.038) and negatively correlated with
serum IFN-γ (r = -0.243, p = 0.003). Trough-biting was significantly positively correlated with
serum cortisol (r = 0.187, p = 0.022), IgG (r = 0.164, p = 0.045) and TNF-α (r = 0.220, p = 0.007).
Frothy saliva score was significantly positively correlated with serum IL-10 (r = 0.196, p = 0.016)
and negatively correlated with serum IFN-γ (r = -0.201, p = 0.014). In conclusion, one of the
takeaways is that frothy saliva can be utilized as a reliable indicator for evaluating stereotypies of
sows. The sows with high incidence of stereotypic behaviors tried to improve stress and humoral
immunity level to cope with the confined environment, and long-term confined sows might be in
a chronic stress state.
Keywords: Confinement; pregnant sows; stereotypic behaviors; physiology; immunity
1. Introduction
Stereotypic behavior is one of the most concerned welfare problems in confined sows, it refers
to the constant, repetitive, and non-functional movement pattern of animals (Mason, 1991). The
stereotypic behaviors of sows principally include sham-chewing, bar-biting, trough-biting, and
rooting behavior. The latest study of Friedrich et al. (2020) indicated that the frothy saliva can be
used as a potential indicator to assess stereotypies of sows, but its reliability needs further
confirmation.
When sows are housed in a relatively barren environment, not only some behavioral
performances are inhibited, but also psychological and physiological circumstances are
dynamically altered (McGlone, 2013). For animals, physical discomfort can be assessed with
behavioral, physiological and biochemical indicators (Duncan, 2005), but psychological
frustrations are hard to find and evaluate. Some studies have suggested that stereotypic behavior
is an external manifestation of suffering or depression in farm animals (Cocchi et al., 2009), but
this view is not entirely reliable due to the lack of data and theoretical support. If stress is assumed
to play a crucial role in this mental state, pigs exposed to chronic stress can be considered as
animal models of depression (Willner, 1997). Behavior, physiology, immunity, and emotion of
animals are inseparable (Mauss et al., 2005). Sow is also an emotional animal, which is prone to
express emotional obstacles and behavioral abnormalities in adversity. When sows are stimulated
by multiple events such as barren environment, feed restriction and gestation, the performance of
stereotypic behaviors may be related to physiology, immunity, and negative emotions, which has
∙ 260 ∙
not been fully elucidated in previous studies.

Therefore, the stress-related physiological parameters (cortisol, Pig-MAP, and CRP) and
immune-related parameters (IgA, IgG, IL-6, IL-10, IFN-γ, and TNF-α) were selected in this study.
At the same time, frothy saliva was used as a new indicator to evaluate stereotypies of sows. The
purpose of our study was to investigate the relationship between behavior and physiological,
immune levels of pregnant sows housed in confinement environment.
2.1. Animals and animal management
Fifty healthy 3rd parity sows (Large White × Landrace) were housed individually in gestation
stalls (210 cm × 57.5 cm × 97.5 cm), which in the front part provided with food and water trough.
Sows were fed once a day at 6:30, and water was provided at 6:45, 9:30, 12:30, 14:30, and 17:00.
All sows were provided with commercial feed containing nutrient contents as follows: NE 9.5 MJ
kg-1, crude protein ≥ 15.0%, crude fiber ≤ 10.0%, calcium 0.5-1.3%, total phosphorus ≥ 0.5%,
lysine ≥ 0.5%. Each sow was fed 2-2.5 kg per day during the gestation of days 0-30 and was
increased to 2.5-3 kg per day during the gestation of days 31-95. Mechanical ventilation, natural
lighting, and temperature-controlled equipment (Big Dutchman, Vechta, Germany) were adopted
and the average daily temperature was 21 ± 1.1℃ with a relative humidity of 60 ± 4.5%.
2.2. Behavior observation and sampling
Behaviors of sows were recorded with the video camera (DS-IT5, Hikvision, China) in early
gestation (27th day), middle gestation (62nd day), and late gestation (91st day). Behavior
categories (Welfare Quality®, 2009) and definitions are listed in Table 1. The behavior of each
sow was continuously recorded at 9:00-11:00 and 14:00-16:00 on each behavior observation day.
Using continuous recording to sample general behaviors (standing, sitting, lateral lying, and
ventral lying behavior) of all sows, expressed as a percentage of the total observation time. Zero-
one sampling method was used to record grooming and stereotypic behaviors of all sows at a 1
min interval to count numbers during observation and recorded the behaviors once for each
occurrence. Meanwhile, the appearance of frothy saliva around the muzzle of sows was observed
on each observation day at 9:00-11:00. Frothy saliva scoring criteria of sows is shown in Table 2.
Table 1. Behavioral categories and definitions.
Behaviors Definitions
General behaviors
Standing Four feet stand upright to support the body
Sitting Hips touch the floor, hind legs curl up, and forelegs stand upright to
support the body
Lateral lying Lying with head, ears, shoulders and contacting the floor
Ventral lying Lying with chest and abdomen contacting the floor and head
straight
Grooming Scrub body with railings, hind limbs or quickly shake ears
Stereotypic behaviors
Sham-chewing Continuous chewing when there is no food in mouth
Bar-biting licking or biting bar with muzzle
Trough-biting licking or biting trough with muzzle
Rooting rooting or licking floor with muzzle
2.3. Physiological and immune measurements
To avoid the influence of the blood collection process on behavioral observation, the blood
collection of anterior vena cava was taken for all sows from 8:00 to 11:00 on days 28, 63, and 92
of gestation. The blood samples were centrifuged at 3000 r min-1 for 10 min, and the prepared
serum was stored at −80℃ before ELISA analysis. The concentrations of serum physiological
∙ 261 ∙
parameters (cortisol, Pig-MAP and CRP) and immune parameters (IgA, IgG, IL-6, IL-10, TNF-
α, and IFN-γ) were measured by ELISA kit (NO. JB158-Pg, JB448-Pg, JB383-Pg, JB176-Pg,
JB174-Pg, JB347-Pg, JB359-Pg, JB015-Pg and JB363-Pg, Jinma, Shanghai, China), and the
measurement procedure of the kit instructions was precisely followed. The intra-assay coefficient
of variation was less than 10%, and the inter-assay coefficient of variation was below 12%.
Table 2. Frothy saliva scoring criteria of sows.
Scoring rules Score
No frothy saliva appears on the lower muzzle area of sows 0
Frothy saliva covers less than 30% of the lower muzzle area of sows 1
Frothy saliva covers 30%-70% of the lower muzzle area of sows 2
Frothy saliva covers 70%-100% the lower muzzle area of sows and drips onto 3
ground
2.4. Statistical Analysis
Statistical analyses were carried out using SPSS 21.0 (SPSS Inc., Chicago, IL, USA). Before
statistical analysis, Shapiro-Wilk test was used to assess the normal distribution of all data. The
data of standing and sitting behavior conform to normal distribution after LN conversion, and the
data of sham-chewing behavior conform to normal distribution after SQRT (square root)
conversion. The homogeneity of variance was assessed by Levene test. One-way ANOVA under
the comparative mean was used to analyze the data on behaviors, physiological, and immune
parameters in different gestational periods. Duncan’s multiple comparison was used to compare
the differences in each parameter. The non-parametric Kruskal-Wallis test was used to analyze
the differences of bar-biting, trough-biting, rooting behavior, and frothy saliva score in different
gestation periods, pairwise comparisons were made using Wilcoxon rank-sum test. Spearman
correlation analysis was used to test the correlation between different stereotypic behaviors and
serum physiological, immune levels of sows during gestation. Results were expressed as mean ±
standard error of mean (SEM), p<0.05 indicates a significant difference.
3.1. General behaviors of sows during gestation
The statistical results of general behaviors (standing, sitting, lateral lying, ventral lying, and
grooming behavior) in early, middle, and late gestational periods are shown in Figure 1. With the
progress of gestation, lateral lying behavior of sows increased significantly (p<0.05), while
standing and ventral lying decreased significantly (p<0.05). Standing behavior in early gestation
was significantly higher than that in middle gestation (p<0.001) and late gestation (p<0.001).
Lateral lying behavior in early gestation was significantly lower than that in middle gestation (p
= 0.004) and late gestation (p<0.001), and that in middle gestation was significantly lower than
that in late gestation (p<0.001). Ventral lying in early and middle gestation was significantly
higher than that in late gestation (p<0.001). There were no significant changes in sitting and
grooming behavior in different gestational periods (p >0.05). The results of lying and standing
was similar to the previous research results (Peet-Schwering et al., 2003). It may suggest that
compared with early gestation, sows gradually adapt to the confined environment with the
progress of gestation and their negative emotions receded. It may also be due to the raised weight
of fetuses leading to decrease of sows’ activity, which makes sows tend to lie down. Behavior
observation found that the lateral lying and ventral lying of sows were basically complementary,
since ventral lying is easier to stand up than lateral lying, it can be speculated that sows may be
alert when ventral lying is compared with lateral lying.
∙ 262 ∙
Figure 1. Standing (a), sitting (b), lateral lying (c), ventral lying (d), and grooming (e) of sows in
early, middle, and late gestation. * p<0.05, ** p<0.01, *** p<0.001.
3.2. Stereotypic behaviors of sows during gestation
The statistical results of stereotypic behaviors (sham-chewing, bar-biting, trough-biting,
rooting behavior, and frothy saliva score) in early, middle, and late gestation are shown in Figure
2. With the progress of gestation of sows, sham-chewing, bar-biting, trough-biting, rooting
behavior, and frothy saliva score were significantly reduced (p<0.05). Sham-chewing in early
gestation was significantly higher than that in middle gestation (p = 0.035), and that in middle
gestation was significantly higher than that in late gestation (p<0.001). Bar-biting behavior in
early gestation was significantly higher than that in middle gestation (p = 0.017) and late gestation
(p = 0.017). Trough-biting in early gestation was significantly higher than that in middle gestation
(p = 0.029), and significantly higher than that in late gestation (p<0.001), there was still a
decreasing trend although the difference between middle and late gestation was not significant (p
= 0.057). Rooting behavior in early gestation was significantly higher than that in middle gestation
(p = 0.040), significantly higher than that in late gestation (p<0.001), and that in middle gestation
was significantly higher than that in late gestation (p = 0.008). Frothy saliva score in early
gestation was significantly higher than that in late gestation (p = 0.005). Obviously, the trends in
frothy saliva were consistent with other stereotypical behaviors during gestation. In this study, the
feed quantity of sows in middle and late gestation was increased compared with that in early
gestation. The gradual increase in feed quantity and fetal volume squeezed the gastrointestinal
tract of sows, and the combined effect of the two elevated the sense of satiation and inhibited the
feeding motivation of sows, which might result in the gradual decline of sham-chewing. The
“coping hypothesis” of stereotypic behaviors argues that the occurrence of stereotypic behaviors
in animals promotes the release of opioid neurotransmitters in the central nervous system, thus
producing a “calming effect” that enables animals to cope with suboptimal environments (Cronin
et al., 1986). Combined with the results of our study, it can be inferred that in the early gestational
stage of sows transfer to stalls, the sudden barrenness of the environment and the restricted space
make sows bored or depressed, and they need to respond by performing a lot of stereotypic
behaviors.
∙ 263 ∙
Figure 2. Sham-chewing (a), bar-biting (b), trough-biting (c), rooting (d), and frothy saliva score
(e) of sows in early, middle, and late gestation. * p<0.05, ** p<0.01, *** p<0.001.
3.3. Physiological and immune levels of sows during gestation
As listed in Table 3, there was no significant difference found in physiological and immune
levels of sows in early, middle, and late gestation periods (p>0.05). The results of this study are
not consistent with Zhang’s previous work, which found that sows performed higher stereotypic
behavior and increased serum cortisol, IgA and IL-6 concentrations as gestation went on (Zhang
et al., 2017). We speculate that this may be caused by breed effects. Compared with the Large
White sows used in Zhang’s study, the crossbred sows (Large White × Landrace) used in this
study have stronger adaptability and resistance to adversity.
Table 3. Physiological and immune levels of sows during gestation.
Gestational periods p-
Parameters
Early Middle Late value
Physiology
Cortisol (μg L-1) 126.776 (3.576) 123.148 (3.300) 122.89 (2.839) 0.641
Pig-MAP (μg L-1) 207.997 (6.550) 198.579 (4.947) 195.411 (5.295) 0.263
CRP (μg L-1) 2519.227 (89.758) 2453.64 (65.333) 2471.35 (61.804) 0.808
Immunity
IgA (μg mL-1) 44.872 (2.246) 40.319 (1.068) 39.393 (1.111) 0.402
IgG (μg mL-1) 84.091 (2.507) 85.934 (2.945) 84.429 (2.748) 0.880
IL-6 (pg mL-1) 1054.477 (35.528) 1008.250 (40.503) 986.064 (38.684) 0.438
IL-10 (pg mL-1) 196.997 (6.080) 195.362 (5.713) 198.540 (6.604) 0.935
IFN-γ (pg mL-1) 2337.673 (73.614) 2270.443 (63.773) 2286.529 (58.817) 0.752
TNF-α (pg mL-1) 397.288 (11.519) 403.111 (11.577) 395.316 (11.628) 0.885
3.4. Relationship between stereotypic behaviors and physiological, immune levels
Spearman correlation analysis between stereotypic behaviors and physiological, immune
∙ 264 ∙
levels of sows in early, middle, and late gestation is shown in Table 4. Sham-chewing was
significantly positively correlated with serum cortisol (r = 0.314, p<0.001), IL-6 (r = 0.244, p =
0.003), IL-10 (r = 0.169, p = 0.038) and negatively correlated with serum IFN-γ (r = -0.243, p =
0.003). Trough-biting was significantly positively correlated with serum cortisol (r = 0.187, p =
0.022), IgG (r = 0.164, p = 0.045) and TNF-α (r = 0.220, p = 0.007). Frothy saliva score was
significantly positively correlated with serum IL-10 (r = 0.196, p = 0.016) and negatively
correlated with serum IFN-γ (r = -0.201, p = 0.014). At present, there are still few studies on the
relationship between stereotypic behaviors and cortisol of pregnant sows, but previous studies
have shown that the stereotypic behaviors of sows in stalls are much more than that of group-
housed sows (Chapinal et al., 2017), and the blood cortisol level of group-housed sows is
significantly lower than that of sows in stalls (Spoolder et al., 2009). Liu et al. (2006) found a
positive correlation between occurrence of stereotypic behaviors and fecal cortisol level in giant
pandas. Therefore, we can speculate that the occurrence of stereotypic behaviors in sows may
indicate that sows were in a state of chronic stress. The results of serum IL-6 and IL-10 levels
may indicate that the occurrence of certain stereotypic behaviors will contribute to the
inflammatory response of sows, and the sows might be under psychological stress. Studies have
also suggested that long-term restraint stress increased serum IL-10 levels in mice with chronic
restraint stress depression model (Liang et al., 2015). IFN-γ levels were significantly reduced in
mice exposed to restraint stress, which may be due to stress inducing the tilt of Th1/Th2 balance
to Th2 dominant humoral immunity (Iwakabe et al., 1998). Therefore, the sows with high
incidence of stereotypic behaviors such as sham-chewing and trough-biting try to improve their
humoral immunity level to cope with and adapt to the confined environment in stalls. In addition,
the correlation between frothy saliva score and serum immune parameters IL-10 and IFN-γ levels
was the same as that between sham-chewing and serum IL-10 and IFN-γ levels, it is reasonable
to speculate that frothy saliva is a novel and reliable indicator for assessing stereotypies of sows
and has some practical value for evaluating welfare status of pregnant sows.
Table 4. Correlation between stereotypic behaviors and physiological, immune levels of sows.
Behaviors Frothy saliva
Parameters
Sham-chewing Bar-biting Trough-biting Rooting score
Physiology
r = 0.314 r = 0.023 r = 0.187 r = 0.065 r = -0.027
Cortisol
p<0.001 p = 0.778 p = 0.022 p = 0.426 p = 0.746
r = 0.081 r = 0.076 r = -0.021 r = 0.204 r = -0.008
Pig-MAP
p = 0.327 p = 0.354 p = 0.803 p = 0.012 p = 0.921
r = 0.009 r = 0.090 r = 0.057 r = -0.025 r = 0.101
CRP
p = 0.915 p = 0.271 p = 0.489 p = 0.764 p = 0.220
Immunity
r = 0.138 r = 0.059 r = 0.061 r = 0.168 r = -0.030
IgA
p = 0.091 p = 0.470 p = 0.462 p = 0.040 p = 0.712
r = 0.029 r = -0.026 r = 0.164 r = 0.231 r = -0.027
IgG
p = 0.722 p = 0.752 p = 0.045 p = 0.004 p = 0.741
r = 0.244 r = -0.814 r = -0.085 r = 0.109 r = -0.012
IL-6
p = 0.003 p = 0.869 p = 0.053 p = 0.184 p = 0.885
r = 0.169 r = 0.044 r = -0.002 r = 0.178 r = 0.196
IL-10
p = 0.038 p = 0.593 p = 0.982 p = 0.030 p = 0.016
r = -0.243 r = -0.119 r = -0.128 r = 0.011 r = -0.201
IFN-γ
p = 0.003 p = 0.148 p = 0.119 p = 0.890 p = 0.014
r = -0.084 r = -0.029 r = 0.220 r = 0.059 r = -0.156
TNF-α
p = 0.307 p = 0.721 p = 0.007 p = 0.472 p = 0.057
∙ 265 ∙
4. Conclusions
As gestation progresses, sows showed increased lateral lying behavior, while standing, ventral
lying, and stereotypic behaviors decreased. Besides, there was a weakly positive correlation
between sham-chewing, trough-biting behavior and the concentration of serum cortisol, a weakly
positive correlation between stereotypic behaviors and the Th2 humoral immune cytokines of IL-
6 and IL-10, and a weakly negative correlation between stereotypic behaviors and Th1 cellular
immune cytokines of IFN-γ. Thereby, we speculate that sows in stalls tried to cope with the
confined environment by performed stereotypic behaviors such as sham-chewing, trough-biting
behavior, as well as improving stress and humoral immunity level, and long-term restricted sows
may be in a chronic stress state. The results of our study also indicate that frothy saliva could be
utilized as a reliable indicator to evaluate stereotypies of sows. Further research needs to clarify
the molecular mechanism of the occurrence of stereotypic behaviors from the perspective of genes
or signaling pathways.
Acknowledgements
This study was funded by the National Natural Science Foundation of China (Grant No.
31472131) and China Agriculture Research System of MOF and MARA.
References
Chapinal, N., J.R. de la Torre, A. Cerisuelo, J. Gasa, M. Baucells, J. Coma, A. Vidal, X. Manteca, 2010.
Evaluation of welfare and productivity in pregnant sows kept in stalls or in 2 different group housing systems,
Journal of Veterinary Behavior. 5, 82–93.
Cocchi, M., L. Sardi, L. Tonello, G. Martelli, 2009. Do mood disorders play a role in pig welfare? Italian
Journal of Animal Science. 8, 691–704.
Cronin, G., P. Wiepkema, J.M. Van Ree, 1986. Andorphins implicated in stereotypies of tethred sows,
Experientia. 42, 198–199.
Duncan, I.J., 2005. Science-based assessment of animal welfare: Farm animals, Revue scientifique et
technique-Office international des epizooties. 24, 483–492.
Friedrich, L., J. Krieter, N. Kemper, I. Czycholl, 2020. Frothy saliva—a novel indicator to assess
stereotypies in sows? Applied Animal Behaviour Science. 222, 104897.
Iwakabe, K. M., Shimada, A. Ohta, T. Yahata, Y. Ohmi, S. Habu, T. Nishimura, 1998. The restraint stress
drives a shift in Th1/Th2 balance toward Th2-dominant immunity in mice, Immunology Letters. 62, 39–43.
Liang, S., T. Wang, X. Hu, J. Luo, W. Li, X. Wu, Y. Duan, F. Jin, 2015. Administration of Lactobacillus
helveticus NS8 improves behavioral, cognitive, and biochemical aberrations caused by chronic restraint stress,
Neuroscience. 310, 561–577.
Liu, J., Y. Chen, L. Guo, B. Gu, H. Liu, A. Hou, X. Liu, L. Sun, D. Liu, 2006. Stereotypic behavior and
fecal cortisol level in captive giant pandas in relation to environmental enrichment, Zoo Biology. 25, 445–
459.
Mason, G.J. 1991. Stereotypies and suffering, Behavioural Processes. 25, 103–115.
Mauss, I.B., R.W. Levenson, L. McCarter, F.H. Wilhelm, J.J. Gross, 2005. The tie that binds? Coherence
among emotion experience, behavior, and physiology, Emotion. 5, 175–190.
McGlone, J.J., 2013. Updated scientific evidence on the welfare of gestating sows kept in different
housing systems, The Professional Animal Science. 29, 189–198.
Spoolder, H., M. Geudeke, C. Van der Peet-Schwering, N. Soede, 2009. Group housing of sows in early
pregnancy: A review of success and risk factors, Livestock Science. 125, 1–14.
Van der Peet-Schwering, C., H. Spoolder, B. Kemp, G. Binnendijk, L. Den Hartog, M. Verstegen, 2003.
Development of stereotypic behaviour in sows fed a starch diet or a non-starch polysaccharide diet during
gestation and lactation over two parities, Applied Animal Behavioural Science. 83, 81–97.
Welfare Quality®, 2009. Welfare Quality®: Assessment Protocol for Pigs (Sows and Piglets, Growing
and Finishing Pigs). Welfare Quality® Consortium, Lelystad, The Netherlands.
Willner, P., 1997. Validity, reliability and utility of the chronic mild stress model of depression: a 10-year
review and evaluation, Psychopharmacology. 134, 319–329.
Zhang, M., X. Li, X. Zhang, H. Liu, J. Li, J. Bao, 2017. Effects of confinement duration and parity on
stereotypic behavioral and physiological responses of pregnant sows, Physiology & Behavior. 179, 369–376.
∙ 266 ∙
Theme IV:
Digital and Smart Technologies for
Managing and Improving Animal
Environment and Welfare
∙ 267 ∙
∙ 268 ∙
Cross-modality Interaction Network for Equine Activity

Recognition Using Time-Series Motion Data
Axiu Mao a, Endai Huang b, Weitao Xu b, Kai Liu a,*
a
Department of Infectious Diseases and Public Health, Jockey Club College of Veterinary Medicine and
Life Sciences, City University of Hong Kong, Hong Kong SAR, China
b
Department of Computer Science, City University of Hong Kong, Hong Kong SAR, China
* Corresponding author. Email: kailiu@cityu.edu.hk
Abstract
Automated equine activity recognition enables caretakers to monitor equine behavioral
variations continuously and remotely, provides rich insights into equine health and welfare, and
contributes to improved equine management and reduced workloads and costs in veterinary
clinics. Over the past decade, animal activity recognition has been studied increasingly with the
aid of various sensors and the use of machine learning and more recent deep learning models.
However, the existing deep learning methods tend to extract modality-specific features simply
using neural networks and fuse these features using concatenation or addition, ignoring deep inter-
modality interaction and resulting in limited performance. To tackle this issue, a cross-modality
interaction network (CMI-Net) involving a dual convolution neural network architecture and a
cross-modality interaction module (CMIM) was proposed to improve classification performance
for equine activities including eating, standing, trotting, galloping, and walking with and without
a rider. Specifically, the CMIM leveraged the information from two modalities (i.e., accelerometer
and gyroscope) to adaptively recalibrate the temporal and axis-wise features in each modality.
The CMI-Net was trained using time series motion data acquired from six neck-attached inertial
measurement units and verified with leave-one-out cross-validation. The results demonstrated that
our approach outperformed the existing algorithms by achieving the highest precision (82.66%),
recall (79.24%), f1-score (78.64%), and accuracy (93.00%). Furthermore, experiments conducted
on different variants of CMI-Net indicated that applying the CMIM in upper layer could obtain
better performance. To the best knowledge, the modality interaction mechanism is employed for
the first time in animal activity recognition with multiple sensors.
Keywords: Animal behavior, deep learning, cross-modality, interaction, wearable sensor
1. Introduction
Behaviors of equine provide rich insights into their physical status and circumstance and are
one of the most important indicators for their health and welfare conditions (Martiskainen et al.,
2009). However, behavioral monitoring for animals, to date, largely relies on manual
observations, which are labor-intensive, time-consuming, and prone to subjective judgments of
individuals (Frost et al., 2003). For instance, although the lameness of horse can be detected based
on changes in normal gaits, studies have reported that veterinarians may spend up to 40% of their
working time assessing lameness in horse, one of the most expensive health issues in the equine
industry (Weeren et al., 2017; Bosch et al., 2018). Therefore, it is of significant importance to
investigate and develop an automatic, objective, and quantifiable measurement system for equine
behaviors concerning health and welfare.
Over the past decades, automated animal activity recognition has been studied widely with
the aid of various sensors (e.g., accelerometers, gyroscopes, and magnetometers) and the use of
machine learning models. For instance, a Naïve Bayes classifier was applied to recognize horse
activities (e.g., eating and trotting) using tri-axial acceleration and obtained 90% classification
accuracy (Kamminga et al., 2019). Random Forest algorithm was utilized to detect cattle
behaviors (e.g., grazing and ruminating) based on tri-axial acceleration and the impact of sensor
placement (i.e., collar, halter, and ear) on classification accuracy was also evaluated (Rahman et
al., 2018). Quadratic Discriminant Analysis was used to categorize sheep activities using tri-axial
∙ 269 ∙
acceleration and gained high classification accuracy with 94%, 96%, and 99% for grazing,
standing, and walking events, respectively (Barwick et al., 2018). However, to classify animal
behaviors accurately using machine learning methods, feature extraction and method selection are
conducted manually and separately, which requires expert domain knowledge and often causes
feature engineering issues (Nweke et al., 2018). Moreover, hand-crafted features often fail to
capture general and complex features, resulting in low generalization ability, i.e., these extracted
features perform well in recognizing activities of some subjects, but badly for others.
To tackle the issues of machine learning, various deep learning approaches are being
increasingly and successfully adopted in recent years for automatic feature learning of wearable
sensors in animal activity recognition. For example, feedforward deep neural network (DNN) and
long short-term memory (LSTM) model were applied to automatically recognize cattle behaviors
(e.g., feeding and ruminating) using inertial measurement units (IMUs) (Peng et al., 2019;
Noorbin et al., 2020). Convolutional neural networks (CNNs), which well capture local temporal
dependency and scale invariance in signals, were developed in automated equine activity
classification based on tri-axial accelerometer and gyroscope data (Bocaj et al., 2020; Eerdekens
et al., 2020). However, in these studies, despite multiple sensors were used and multi-modal data
with different characteristics were yielded, these data were often processed simply by using fusion
strategies, such as early fusion, result fusion, and feature fusion. Early fusion used in these
methods, i.e., extract the same features without distinction of modalities, might cause
interferences between multi-modal information (Ha and Choi, 2016), while result fusion scheme
tends to ignore the correlations of multi-modal data and leads to information loss. Specifically,
feature fusion strategies are normally limited to linear fusion (e.g., concatenation and addition),
failing to achieve complicated multi-modal interaction and complementary-redundant
information combination between multiple modalities (Liu et al., 2020).
To improve the automated recognition performance for equine activities, we developed a
cross-modality interaction network (CMI-Net) involving a dual CNN trunk architecture and a
joint cross-modality interaction module (CMIM). The dual CNN architecture was used to extract
modality-specific features and aggregate information of multiple modalities. The CMIM
leveraged the knowledge of multi-modal data (i.e., data from accelerometer and gyroscope) to
adaptively recalibrate temporal and axis-wise features in each modality. The CMI-Net was trained
using an extensively labeled dataset from (Kamminga et al., 2019), to automatically recognize
equine activities including eating, standing, trotting, galloping, and walking with and without a
rider. The leave-one-out cross-validation (LOOCV) method was applied to test the generalization
ability of our method, and the results were further compared to the existing algorithms.
2.1. Data Description
The dataset under this study is a public dataset created by (Kamminga et al., 2019). In this
dataset, more than 1.2 million 2-s data samples were collected from 18 individual equines using
neck-attached IMUs. The sampling rate was set to 100 Hz for both tri-axial accelerometer and
gyroscope and 12 Hz for tri-axial magnetometer. Amongst, the majority of the samples were
unlabeled, whereas data from 6 equines and 6 activities (i.e., eating, standing, trotting, galloping,
and walking with and without a rider) were labeled extensively (87,621 2-s samples in total) and
were used to classify equine activities in previous studies (Kamminga et al., 2019; 2020). In this
study, data from tri-axial accelerometer and gyroscope among the 87,621 samples were exploited
separately, forming up to two tensors with a size of 1 × 3 × 200 for each sample. Specifically, the
activities of eating, standing, trotting, galloping, and walking with and without a rider occupied
18.32%, 5.84%, 28.62%, 4.50%, 3.80%, and 38.94% of the total sample number, respectively. In
addition, the samples were processed using Z-Score before fed into the network.
∙ 270 ∙
2.2. Proposed Model

Our proposed CMI-Net, where accelerometer and gyroscope data were fed into two CNN
branches (represented by CNNacc and CNNgyr) separately, is shown in Figure 1(a). Amongst, the
dual CNN enabled to extract modality-specific features and concatenate these features before the
final dense layer. To achieve deep interaction between the two-modality data and capture the
complementary information and suppress unrelated information (e.g., noise, redundant signals,
and potentially confusing signals) from them, a joint cross-modality interaction module (CMIM)
was designed and inserted in the upper layer. The details are described below.
Figure 1. The architecture of our proposed model. (a) Our proposed CMI-Net. The size of the
feature maps is marked after every Res-LCB layer. Here, “GAP” and “FC” mean global
average-pooling layer and fully connected layer, respectively. (b) Residual-like convolution
block (Res-LCB). (c) Cross-modality interaction module (CMIM).
2.2.1. Dual CNN trunk architecture
CNNacc and CNNgyr contained four convolution blocks, three max-pooling layers, one global
average-pooling layer, and one fully connected layer, respectively, followed by concatenation and
one joint fully connected layer. Inspired by the residual unit in the deep residual network that
behaves like ensembles and has smaller magnitudes of responses (He et al., 2016), to promote
representation ability and robustness of the model, we designed a residual-like convolution block
(Res-LCB) as demonstrated in Figure 1(b). The definition is described below.
X l 1  RELU (Conv11( X l )  Conv13 ( X l )) , (1)
where, X l and X l 1 denoted feature maps in l and l  1 layer, respectively. Conv11 () and
Conv13 () represented 1 × 1 and 1 × 3 convolution operations, respectively.  denoted the
element-wise addition, and RELU () denoted Rectified Linear Unit activation function.
2.2.2. Cross-modality interaction module (CMIM)
Inspired by the multimodal transfer module (MMTM) that recalibrates channel-wise features
of each modality based on multi-modal information (Woo et al., 2018), and convolutional block
attention module (CBAM) that focuses on the spatial information of the feature maps (Joze et al.,
∙ 271 ∙
2020), we proposed a CMIM based on attention mechanism to adaptively recalibrate temporal

and axis-wise features in each modality by utilizing multi-modal information. The detailed CMIM
is illustrated in Figure 1(c).
Let A  RC  H W and G  RCH W represented the features at a given layer of CNNacc and
CNNgyr, respectively. Here, C , H , and W denoted the channel number and spatial dimensions
of features. Specifically, the H and W corresponded to the axial and temporal signals, respectively.
The CMIM received A and G as input features. We first applied average-pooling operations
along channels of the input features, generating two 2D maps which were further concatenated
and mapped into a joint representation Z  RC H W . The operation is shown as follow:
'
Z  RELU (Conv13 ([ Avgpool ( A), Avgpool (G )])) , (2)

'
where C denoted the channel number of feature Z , Avgpool()
denoted the average-pooling
operation, and   denoted concatenation operation. Further, two spatial attention maps
AA  R1H W and AG  R1H W were generated through two independent convolution layers with
sigmoid function (  () ) using the joint representation Z :

13 13
AA   (Conv (Z )) , AG   (Conv (Z )) . (3)
AA
and AG
were then used to recalibrate the input features, generating two final refined
features, i.e., A'  RCH W and G '  RCH W :
A'  A  AA  A , G'  G  AG  G , (4)
where  denoted the element-wise multiplication. Specifically, each convolution operation
under this study was followed by a batch normalization operation. The increases in channel
numbers and decreases in spatial dimensions were implemented through Res-LCB and Max-
pooling operations, respectively.
2.3. Implementation Details
To attain subject-dependent results, the LOOCV method was used, in which four subjects
were chosen for training, one for validation, and one for test each time and rotated in a circular.
Precision, recall, F1-score, and accuracy were used as metrics to indicate the classification
performance (Peng et al, 2019). A cross-entropy function added by an L2 regularization term with
weight decay of 0.1 was adopted as our loss function. During training, an Adam optimizer with
an initial learning rate of 1e-4 was employed, and the learning rate decreased by 0.1 times every
20 epochs. The number of epochs and batch size were set to 100 and 256, respectively. The best
model with the highest validation accuracy was saved and verified using test data. To further
explore the performance of our CMIM, we ran the network without CMIM and with it inserted
after the first, second, and third max-pooling layer, respectively, to get four different variants. All
experiments were executed using Pytorch framework on a NVIDIA Tesla V100 GPU.
Overall, experiments conducted on the public dataset demonstrated that our proposed CMI-
Net outperformed the existing algorithms. Ablation studies were further carried out to verify that
applying the CMIM in upper layer of CMI-Net could obtain better performance. Moreover,
visualization analysis was presented to help us probe into the effectiveness of our method. The
details are described as follows.
3.1. Comparison with the Existing Algorithms
To verify the effectiveness of our proposed method, the performance of CMI-Net was
compared with the existing algorithms including Naïve Bayes (Kamminga et al., 2019) and
∙ 272 ∙
ConvNets (Bocaj et al., 2020), as illustrated in Table 1. These existing algorithms were also
conducted on the same dataset. The results revealed that our method outperformed Naïve Bayes
by obtaining an F1-score of 78.64% and an accuracy of 93.00%. The reason for this superior
performance was the convolution operation in CNN, which could achieve automated feature
learning and extract more complex patterns without any domain knowledge (Zeng et al., 2015).
Our model also performed better than the ConvNets with higher precision, recall, F1-score, and
accuracy of 82.66%, 79.24%, 78.64%, and 93.00%, respectively. This was attributed to the ability
of our architecture to effectively capture the complementary information and inhibit unrelated
information of multi-modal data through deep multi-modality interaction.
In Figure 2, we show the confusion matrix aggregating the classification results under 6-fold
cross-validation. All activities had more than 90% accuracy (i.e., 95.15% for eating, 91.98% for
galloping, 93.94% for standing, 97.82% for trotting, and 98.42% for walking-rider) except for the
walking-natural activity, which only obtained a low accuracy. This low classification accuracy
occurred due to two main reasons. The first reason was class imbalance. Walking-natural as the
minority class in the dataset only occupies 3.8% which is much less than the occupation of 38.94%
for walking-rider as the majority class. The second reason was the confusion with other activities.
As shown in Figure 2, the walking-natural activity was mostly confused with eating and walking-
rider. This was because during eating the equine was slowly walking so that some samples of
eating might contain walking activity (Kamminga et al., 2019). In addition, the movement patterns
of walking-natural and walking-rider were very similar, which interfered with the learning ability
of the network for these two behavioral characteristics (Figure 3). To further improve the
classification performance, especially for the walking-natural activity, we would investigate some
feasible strategies (Liu et al., 2016; Hermans et al., 2017) to enlarge inter-class distance and reduce
the intra-class distance between similar activities in future work.
Table 1. Comparison of our method with the existing algorithms.
Methods Precision (%) Recall (%) F1-score (%) Accuracy (%)
Kamminga et al. (2019) - - 72.60* 78.46*
Bocaj et al. (2020) 78.84 78.02 77.57 91.31
Our CMI-Net 82.66 79.24 78.64 93.00
* Denotes the values cited from the reported results directly.
Table 2. Comparison of our method with its variants.
Methods Precision (%) Recall (%) F1-score (%) Accuracy (%)
Variant0 79.02 77.09 76.88 91.76
Variant1 79.15 76.46 76.27 91.54
Variant2 77.57 78.43 77.50 92.34
Variant3 80.42 77.27 76.47 91.98
Our CMI-Net 82.66 79.24 78.64 93.00
3.2. Ablation study
To evaluate the effectiveness of CMIM and the impact of its position in network on
classification performance, the results corresponding to four different variants are shown in Table
2. Our proposed CMI-Net showed superior performance to Variant0 (i.e., network without
CMIM), indicating the effective performance of our interaction module. Whereas Variant1,
Variant2, and Variant3 (i.e., networks with CMIM inserted after each max-pooling layer
respectively) did not perform better than Variant0. This might be explained by the reason that
modality-specific features learned in the shallow layer were simple and contain noise, which
interfered the process that CMIM learned complex inter-modality correlations, leading to poor
predictions (Wei et al., 2021). In addition, our architecture obtained the best performance since it
∙ 273 ∙
applied the CMIM after a deeper layer, which enabled the network to discover more
discriminative patterns and suppress irrelevant variations more effectively (Lecun et al., 2015).
Figure 2. Confusion matrix for classification with six activities.
Figure 3. Example of accelerometer and gyroscope data for walking-natural and walking-rider.
3.3. Visualization
The results above have proved that the inclusion of an interaction module in the network
provided quantifiable improvements in identification performance. This was also reflected in the
qualitative visualization of the embeddings and the corresponding clusters in Figure 4, with the
help of t-distributed Stochastic Neighbour Embedding (t-SNE), a technique for visualizing high-
dimensional data by giving each datapoint a location in a two or three-dimensional map (Maaten
and Hinton, 2008). Figure 4 visualizes the two-dimensional embedded features from part test
dataset after the fully connected layers of both CNN branches under the network without and with
CMIM, respectively, by using the t-SNE technique with init of ‘pca’ and perplexity of 30.
Comparing the left and right columns in Figure 4, it can be observed that more compact clusters
were generated under the network with CMIM by reducing intra-class distance and enlarging
inter-class distance. The core technical point was that the joint interaction module enabled to
adaptively amplify salient features and suppress unrelated features based on information of two-
modality data. Therefore, both quantitative and qualitative findings reinforced the suitability of
our proposed CMI-Net to the tasks using two-modality data.
4. Conclusions
To improve equine activity classification performance of the deep learning-based methods,
we successfully employed our CMI-Net involving a dual CNN trunk architecture and a CMIM.
The CMI-Net could capture complementary information and suppress unrelated information of
multi-sensor data through deep inter-modality interaction. Specifically, the dual CNN architecture
∙ 274 ∙
extracted modality-specific features, and the CMIM recalibrated temporal and axis-wise features
in each modality by utilizing multi-modal knowledge. We also concluded that CMIM inserted in
deeper layer performed better than that in shallow layer since high-level features contained more
general patterns. Regarding the class imbalance problem in this study, we would try to apply some
useful balancing techniques (e.g., re-sampling and cost-sensitive re-weighting) in future work.
We also suggest further study on reasonable loss functions like triplet loss and L-softmax loss to
better distinguish activities with similar patterns.
(a) Accelerometer (without CMIM) (b) Accelerometer (with CMIM)
(c) Gyroscope (without CMIM) (d) Gyroscope (with CMIM)
Figure 4. Typical embeddings visualization of the features extracted from tri-axial accelerometer
and gyroscope data under network without and with CMIM, respectively.
Acknowledgements
We would like to thank Jacob W. Kamminga et al. at the pervasive systems group, University
of Twente for providing the public dataset. Funding for conducting this study was provided by
the new research initiatives at City University of Hong Kong.
References
Barwick, J., D.W. Lamb, R. Dobos, M. Welch, M. Trotter, 2018. Categorising sheep activity using a tri-
axial accelerometer. Computers and Electronics in Agriculture. 145, 289–297.
Bocaj, E., D. Uzunidis, P. Kasnesis, and C.Z. Patrikakis, 2020. On the benefits of deep convolutional
neural networks on animal activity recognition. In 2020 International Conference on Smart Systems and
Technologies. October. 83–88.
∙ 275 ∙
Bosch, S., F. Serra Bragança, M. Marin-Perianu, R. Marin-Perianu, B.J. Van der Zwaag, J. Voskamp, W.
Back, R.V. Weeren, P. Havinga, 2018. EquiMoves: a wireless networked inertial measurement system for
objective examination of horse gait. Sensors. 18(3), 850.
Frost, A.R., D.J. Parsons, K.F. Stacey, A.P. Robertson, S.K. Welch, D. Filmer, A. Fothergill, 2003.
Progress towards the development of an integrated management system for broiler chicken
production. Computers and Electronics in Agriculture. 39(3), 227–240.
Ha, S., and S. Choi, 2016. Convolutional neural networks for human activity recognition using multiple
accelerometer and gyroscope sensors. In 2016 International Joint Conference on Neural Networks. July. 381–
388.
He, K., X. Zhang, S. Ren, and J. Sun, 2016. Deep residual learning for image recognition. In Proceedings
of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. June. 770–778.
Hermans, A., L. Beyer, and B. Leibe, 2017. In defense of the triplet loss for person re-identification. arXiv
preprint. arXiv:1703.07737.
Joze, H.R.V., A. Shaban, M.L. Iuzzolino, K. Koishida, 2020. MMTM: multimodal transfer module for
CNN fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. June.
13289–13299.
Kamminga, J.W., L.M. Janßen, N. Meratnia, P.J. Havinga, 2019. Horsing around—a dataset comprising
horse movement. Data. 4(4), 131.
Kamminga, J.W., N. Meratnia, D.V. Le, and P.J. Havinga, 2020. Towards deep unsupervised
representation learning from accelerometer time series for animal activity recognition. In 6th Workshop on
Mining and Learning from Time Series, MiLeTS 2020. August. 8.
Kamminga, J.W., N. Meratnia, and P.J. Havinga, 2019. Dataset: horse movement data and analysis of its
potential for activity recognition. In Proceedings of the 2nd Workshop on Data Acquisition To Analysis.
November. 22–25.
LeCun, Y., Y. Bengio, G. Hinton, 2015. Deep learning. Nature. 521(7553), 436–444.
Liu, N., N. Zhang, and J. Han, 2020. Learning selective self-mutual attention for RGB-D saliency
detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. June.
13756–13765.
Liu, W., Y. Wen, Z. Yu, and M. Yang, 2016. Large-margin softmax loss for convolutional neural
networks. In International Conference on Machine Learning. June. 2(3), 7.
Van der Maaten, L., G. Hinton, 2008. Visualizing data using t-SNE. Journal of Machine Learning
Research. 9(11), 2579–2605.
Martiskainen, P., M. Järvinen, J.P. Skön, J. Tiirikainen, M. Kolehmainen, J. Mononen, 2009. Cow
behaviour pattern recognition using a three-dimensional accelerometer and support vector machines. Applied
Animal Behaviour Science. 119(1–2), 32–38.
Noorbin, S.F.H., S. Layeghy, B. Kusy, R. Jurdak, G. Bishop-hurley, and M. Portmann, 2020. Deep
learning-based cattle activity classification using joint time-frequency data representation. arXiv preprint.
arXiv:2011.03381.
Nweke, H.F., Y.W. Teh, M.A. Al-Garadi, U.R. Alo, 2018. Deep learning algorithms for human activity
recognition using mobile and wearable sensor networks: State of the art and research challenges. Expert
Systems with Applications. 105, 233–261.
Peng, Y., N. Kondo, T. Fujiura, T. Suzuki, H. Yoshioka, E. Itoyama, 2019. Classification of multiple
cattle behavior patterns using a recurrent neural network with long short-term memory and inertial
measurement units. Computers and Electronics in Agriculture. 157, 247–253.
Rahman, A., D.V. Smith, B. Little, A.B. Ingham, P.L. Greenwood, G.J. Bishop-Hurley, 2018. Cattle
behaviour classification from collar, halter, and ear tag sensors. Information Processing in Agriculture. 5(1),
124–133.
Van Weeren, P.R., T. Pfau, M. Rhodin, L. Roepstorff, F. Serra Bragança, M.A. Weishaupt, 2017. Do we
have to redefine lameness in the era of quantitative gait analysis? Equine Veterinary Journal. 49(5), 567–569.
Wei, J., Q. Wang, Z. Li, S. Wang, S. K. Zhou, and S. Cui, 2021. Shallow feature matters for weakly
supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition. June. 5993–6001.
Woo, S., J. Park, J. Y. Lee, and I. S. Kweon, 2018. Cbam: Convolutional block attention module.
In Proceedings of the European conference on computer vision. September. 3–19.
Zeng, M., L.T. Nguyen, B. Yu, O.J. Mengshoel, J. Zhu, P. Wu, and J. Zhang, 2014. Convolutional neural
networks for human activity recognition using mobile sensors. In 6th International Conference on Mobile
Computing, Applications and Services. November. 197–205.
∙ 276 ∙
Early Detection of Chicken Diseases Based on

Clinical Symptom Monitoring: A Review
Pengguang He, Jinming Pan, Hongjian Lin *
*Corresponding author. Email address: linhongjian@zju.edu.cn
Abstract
Global demand for protein has been steadily increasing as a result of population growth and
economic development. Poultry industry provides a large portion of meat and eggs for human
consumption; however, infectious diseases in poultry farms occasionally lead to a declined
productivity and even widespread chicken mortality. Early detection and warning of poultry
infectious diseases play a critical role in breeding and production system, improving animal
welfare and reducing losses. The health status of poultry is often reflected by its physiological and
behavioral clinical symptoms, such as higher body temperature resulted from a fever, abnormal
vocalization caused by respiratory diseases, and abnormal behaviors due to pathogenic infection.
The use of technologies for symptom detection can monitor the health status of broilers and laying
hens in a continuous, noninvasive and automated way. This review summarized recent literature
on poultry disease detection and highlighted clinical symptom monitoring technologies of sick
poultry. The review concluded that some recently developed technologies are demonstrated to be
superior to manual inspection in small scales, improve farm productivity and reduce loss, but the
clinical symptom-based monitoring systems have not been fully utilized for farm level
management.
Keywords: Clinical symptoms, disease warning, noninvasive detection, physiological
characteristics, behavioral characteristics
1. Introduction
The growth of population and life quality increases the global demand for protein products
such as meat and eggs (Cao and Li, 2013; Daghir et al., 2021). Poultry industry, providing a large
number of protein products for human beings (Gerber et al., 2007), is the most rapidly growing
agricultural sub-sector and plays a critical role in guaranteeing food safety (Mottet and Tempio,
2017). There are more than 68 billion chickens farmed in 2018, representing a third of global meat
production in addition to 1.38 trillion eggs (Blake et al., 2020).
Disease control is among the challenges in maintaining and further developing the industry.
Inadequate disease prevention leads to continuous occurrence of disease outbreaks. The outbreaks
not only interfere poultry production and cause huge economic losses, but also endangers human
health and life safety in some circumstances like pathogenic avian influenza (Kalthoff et al.,
2010). Conventional disease detection methods in farms mainly rely on sophisticated
veterinarians for manual inspection or biochemical examination with relevant instruments.
Manual inspection is time-consuming and labor-intensive and may produce inaccurate results.
Although biochemical examination is sensitive and more accurate, it cannot be performed in real-
time and tends to be expensive. With the continuous development of large-scale and mechanized
breeding, the traditional manual-based methods are unable to meet new requirements. Therefore,
it will make a significant contribution to the poultry industry if early disease detection can be
automatically done, and the corresponding solutions can be proposed in time.
Poultry sickness can be reflected from multiple dimensions, such as vocalization, body
temperature, and daily behaviors. When abnormal changes in chicken are observed as clinical
symptoms, health status may be determined. Figure 1 shows a simplified relationship between
clinical symptoms and possible reasons. Some of the existing reviews summarized the relevant
studies of poultry disease detection from the technical point of view (Ben Sassi et al., 2016; Vidic
∙ 277 ∙
et al., 2017; Astill et al., 2018; Li et al., 2020a). This review focuses on the characteristics that
can be used for early disease detection in poultry farms and expounds these physiological and
behavioral characteristics in detail.
Figure 1. A simplified relationship between clinical symptoms and possible reasons.

2. Early disease detection through physiological characteristics
Physiological characteristics of animals contain health information, which usually can be used
to reflect the comfort of environment; emotional state, social class and health status of animals.
In poultry research area, some physiological characteristics of a single bird or a flock, including
vocalization, feces, and body temperature, etc., were studied for disease detection and early
warning as these factors were confirmed to have causal relationship with disease (Aziz and Bin
Othman, 2017; Carpentier et al., 2018; Li et al., 2020b). This section focuses on the forementioned
common physiological characteristics.
2.1. Abnormal vocalization
As an important physiological characteristic, vocalization of poultry is considered as an
indicator of health and welfare (Manteuffel et al., 2004; Du et al., 2020). Vocalization generated
with specific organs, is an expression of the physiological signal caused by internal or external
factors, which is closely related to the health of the respiratory system. The change in vocalization
of chickens may be to indicate a poor environment or disease infection. It has been successfully
used in the prediction of poultry weight and the analysis of pecking activity (Fontana et al., 2015;
Aydin and Berckmans, 2016; Gonzalez et al., 2020; Nasirahmadi et al., 2020).
Rale is a common symptom of respiratory diseases in poultry. Carroll et al. (2014) infected 6
chicks of 15-day-old with bronchitis virus and recorded the sound of healthy and sick chickens
with microphones. The Mel Frequency Cepstrum Coefficients (MFCCs) of the audio signals was
calculated as feature, and the classification accuracy of the decision tree was 73.4%. Then a
dictionary learning and sparse coding method was applied to this task to classify audio signals,
and the accuracy was improved to 97.85% (Whitaker et al., 2014).
The above research were conducted by breeding the experimental subjects in a closed
environment without environmental noise interference, but this is not consistent with the actual
situation of commercial breeding farms. Carpentier et al. (2019) randomly assigned 51 7-day-old
among 308 broilers to two cages in an open environment and vaccinated the subjects with
Newcastle vaccine after two days. The sound of the chickens was recorded for 972 h with a
microphone. The acquired sound signals were converted into spectrograms and eight features
were extracted for the classification of sneezing sound, and a result with a sensitivity of 66.7%
and an accuracy of 88.4% was achieved by the algorithm proposed in their study. In a recent study,
Newcastle disease was also used as a target disease in the experiment. Mahdavian et al. (2021)
extracted five acoustic features (MFCCs, energy of mel sub bands, wavelet energy, wavelet
entropy and spectral flatness) from healthy and sick chickens for sound classification, and they
∙ 278 ∙
found that wavelet energy and MFCCs were able to detect Newcastle disease on the fourth day
after inoculation with an accuracy of 80% and 78%, respectively. Although the two features
exhibited a similar accuracy, wavelet energy had a better performance on detecting healthy
chickens, while MFCCs was better at detecting challenged birds.
On the one hand, these studies proved that it’s feasible to extract sound features for the early
warning of poultry disease; on the other hand it also showed that the manually extracted features
may have certain bias for one particular disease, i.e., the acoustic features presented by chickens
with different diseases cannot be described by the same set of features, which leads to a low
generalization ability of the models and restricts the popularization and application of these
methods. To address this problem, deep-learning RNN (Recurrent Neural Network) and CNN
(Convolutional Neural Network) models, which can extract features from input data
automatically, were used to classify the sounds of healthy chickens and those infected with avian
influenza (AI) for disease detection and early warning (Cuan et al., 2020). The results of this study
showed that the highest recognition accuracies of CNN were 93.01%, 95.05% and 97.43% on
Day-2, Day-4, and Day-6 after injection of H9N2 virus, respectively, which were much higher
than the recognition accuracies of RNN (49.97%, 55.86%, and 57.10%, respectively). This 2-
day’s early warning may alert farmers to the AI infection, which will greatly reduce economic
losses if necessary measures were taken before an outbreak.
What the current studies have in common is that sound samples from both healthy and infected
chickens were manually screened and processed using sophisticated pretreatment methods, and
then the sound was further classified after obtaining high-quality data. However, the sound data
collected in real-time is continuous, and the abnormal sound from infected chickens is usually just
a moment and its occurrence may not have any pattern. Therefore, how to process and analyze
the collected real-time data is a difficult problem. In addition, environmental noise is an important
factor affecting the application of sound for early detection of poultry diseases in commercial
farms. If the abnormal sound can be located based on real-time analysis (Du et al., 2018), it will
further improve the potential of its application in modern poultry management.
2.2. Abnormal body temperature
Chicken is a homeothermic animal. They produce and dissipate heat to maintain a relatively
constant body temperature, including body core and surface temperature. However, under
pathological or stress conditions, the body temperature will change as a result of infection or
emergency response. So, body temperature can be utilized as an important physiological index in
early warning and clinical diagnosis of chicken diseases.
Footpad dermatitis (FPD) is a condition that causes lesions on the plantar surface of the
footpads in growing chicken, which is considered as an important animal welfare issue and will
cause huge economic losses (Krautwald-Junghanns et al., 2011). The prevalence and the severity
of FPD in broiler breeders increase over time (Kaukonen et al., 2016), so FPD is considered as a
threat to poultry production and need to be early diagnosed. The dielectric constant (DC) has been
shown to be useful in the diagnosis of FPD. The values of the DC decreased monotonically with
increasing FPD score and the negative correlation between DC and FPD score was significant
(P<0.0001) (Hoffmann et al., 2013). FPD infection is also usually accompanied by a change in
the temperature of plantar surface of the poultry foot. Plantar temperature measurements were
carried out on 80 turkeys of 10-week-old by using thermal imaging. The results showed that foot
pad temperatures were significantly lower than foot temperatures under normal condition, and
severity of mild foot pad dermatitis as scored visually was associated with the temperatures of the
plantar surface of the foot and foot pad (Moe et al., 2018). Furthermore, the birds were classified
into three categories of negative, suspect, and positive according to the temperature change from
the metatarsal footpad to the digit extremities (Wilcox et al., 2009). By analyzing the temperature
of plantar surface of the feet of 150 randomly selected hens, it was found that the Pearson’s
correlation between thermal images classified as positive and the visual score of clinical for FPD
∙ 279 ∙
was 86.7%, whereas the correlation between the thermal images classified as suspect and the
visual score of mild was only 26.7%. The result indicates that thermal imaging therefore
represents a novel tool for the reliable and noninvasive early detection of foot pathologies but is
a more sensitive indicator of subclinical infection than visual appraisal. In addition, Liu et al.
(2017) demonstrated that thermal imaging can be used to obtain the body surface temperature of
poultry and to identify healthy and sick chickens based on the temperature difference between
their heads and legs.
For AI diagnosis, researchers have focused on developing portable AI detection devices based
on PCR and biosensor technology in recent years (Wu et al., 2017; Astill et al., 2018; Zhang et
al., 2018; Zhang et al., 2019); however, these techniques are relatively expensive and require
sophisticated operating protocols. Thus, a prototype wearable wireless node with a thermistor and
an accelerometer was developed to detect chickens infected with the highly pathogenic avian
influenza (Okada et al., 2009). The result showed that weakness and fever of the infected chickens
were detectable before chickens showed clinical signs and gross lesions, and the death of infected
chickens was noticed by a sudden decrease in body temperature. There are also studies which
applied thermal imaging technology to AI detection. Blas et al. (2014) successfully detected dead
chickens in poultry production by high resolution thermal imaging technology, which improved
the welfare level and reduced the risk of disease transmission. The maximum surface temperature
change of chickens was monitored using thermal imaging camera devices. Further analysis found
that the maximum surface temperature increased at least 24 h before the manifestation of clinical
signs of AI infection, depending on the severity of the infection, which means the surface
temperature of chicken could be considered as an early indicator of AI (Noh et al., 2021).
At present, the disease detection based on abnormal body temperature in poultry is mostly
focused on leg diseases and avian influenza, while there are few studies on the relationship
between other diseases and body temperature in poultry. The environment of poultry farms is
complex, and the results of thermal infrared temperature measurement are easily affected by
environmental factors (Shang et al., 2017). Therefore, more efforts should be put into the research
on how to obtain high-quality thermal infrared images in a varying environment. Furthermore,
when the object of detection is expected to be a certain chicken rather than the whole flock, the
overlap of chickens will seriously interfere with the detection results. Therefore, the segmentation
of chickens in temperature measurement will be a key research direction in the future. Wearable
temperature sensors are another area that deserves more attention to develop more portable and
lightweight device with a better stability and longer transmission distance.
2.3. Abnormal feces
Digestive system diseases are widespread in chicken farm, seriously affecting productivity
and animal welfare (Ducatelle et al., 2018). When the gastrointestinal tract of chicken is infected,
different degrees of lesions will occur depending on the severity of infection, and abnormal fecal
characteristics will be observed as an early clinical symptom. Table 1 lists the common diseases
causing diarrhea in chickens and the corresponding feces external characteristics. Feces can be
used for early disease diagnosis in poultry. The traditional abnormal feces identification highly
relies on experienced veterinarians for manual inspection, which is feasible to some extent, but is
labor-intensive and cannot diagnose diseases at an early stage. Researchers try to use machine
learning approaches to address this task.
Aziz et al. (2017) were the first to combine the external characteristics of chicken feces with
machine learning for disease detection. They used the gray level co-occurrence matrix to extract
19 feces texture features. The feature set was used as input to support vector machine to classify
healthy and sick chickens, and the study achieved the highest accuracy of 93.75%. This result
proved the feasibility of detecting sick chickens by feces characteristics but was not representative
as only 20 feces samples were used. The development of deep learning technology makes it
possible to deal with larger sample set with high accuracy. Wang et al. (2019) collected 10,000
∙ 280 ∙
feces images of 25 to 35-day-old Ross broilers in commercial chicken farms and manually marked
them into five categories including normal, abnormal shape, abnormal color, abnormal water
content, abnormal shape and water content. In their study, two target detection models based on
CNN were proposed to detect and classify feces. The results showed that Faster R-CNN had better
performance on test dataset with a recall of 99.1% and an average mean precision of 93.3%. This
study proved the possibility of identifying diseases based on feces images in commercial scale.
Table 1. Common diseases causing diarrhea in chickens and the corresponding feces
characteristics
Name of disease Vulnerable time Degree of hazard Characteristics of feces
Avian influenza All age  Yellow-brown and thin
Newcastle disease All age  Yellow-green and thin
Infectious bursa 3-6 weeks  Lime watery
Avian cholera Adult  Grass-green and thin
pullorum disease 0-2 weeks  White mushy
Coccidiosis 4-6 weeks  Brown red thin or blood
Tapeworm disease All age  Thin or with blood like mucus
The counting and species identification of eggs in feces is also an important part in chicken
intestinal disease detection, which can be achieved with high accuracy today (Zhang et al., 2014;
Li et al., 2019; Thevenoux et al., 2021). Although this method can accurately identify a specific
disease, it usually has a complex operation process and is not suitable for disease early warning.
Conversely, it is difficult to diagnose specific diseases through the feces external characteristics,
but it has pratical significance for farmers to start disease prevention and control measures at an
early stage. More research have to be conducted to distinguish the features of the feces external
characteristics of different diseases at the early, middle and late stages of infection, so as to
achieve the purpose of disease early warning. In addition, digestive system diseases can often be
diagnosed by specific feces markers. As a technology that can identify certain substances
qualitatively and quantitatively, spectral technology has been used to detect feces on the surface
of meat products (Liu et al., 2003; Yoon et al., 2011; Chowdhury and Morey, 2020). But research
on applying the technology to directly detecting feces markers has not been reported.
3. Early disease detection through behavioral characteristics
As an important animal welfare and health indicator, behavior, which is expresses by activity
and posture, is a simple and common index that reflects the health status of animals and poultry
(de Jong et al., 2016; Li et al., 2020a). Therefore, the real-time, automatic and nondestructive
monitoring of poultry behavior plays a critical role in welfare improvement and disease early
detection of sick chickens.
3.1. Abnormal activity
Lameness is a common leg disease and approximately 30% of chickens were seriously lame
(Knowles et al., 2008). When the activity deviates from a normal threshold, it often indicates the
presence of disease. Some recent work suggests that automatic monitoring of flock activity can
provide disease early warning in poultry.
Kristen et al. (2011) used cameras to record and describe the undisturbed and abnormal
activity level of broiler chickens at 1 to 3 weeks of age. The results showed that the detection
system could notify producers when the activity deviated from an expected level at a given age.
Gait score is the most commonly used conventional method for lameness detection, which highly
relies on the visual observation. In order to realize automatic monitoring of the activity level of
broilers, Aydin (2017b) utilized a 3D vision camera with a depth sensor to record the behaviors
of broilers, and an image processing algorithm was employed to detect the number of lying and
latency to lie down events. The result showed that there were significant correlations between the
two events and gait score (R2 = 0.934 and 0.949, respectively). Later on, Aydin (2017a) found
∙ 281 ∙
that the speed, step frequency, step length and lateral body oscillation of broilers were correlated
with gait score (R2 = 0.831, 0.882, 0.844 and 0.861, respectively). Furthermore, the amplitude
(difference between highest activity peak and baseline level) was proved to be significantly related
to gait score as well, which will decrease with declining walking ability (Silvera et al., 2017).
Optical flow analysis, a simple imaging technique that measures changes in brightness in a
series of images, was used to detect abnormal flock activity at an early stage (Dawkins et al.,
2017). Colles et al. (2016) used optical flow method to detect diseased chickens infected by
Campylobacter. The result showed that Campylobacter-free chicken flocks have higher mean and
lower kurtosis of optical flow than those Campylobacter-positive. The subclinical symptoms can
be identified within the first 7 to 10 days through this way, which is much earlier than that of the
traditional microbiological methods. Flock activity monitoring has also been shown to be useful
for early detection of abnormal conditions associated with increased mortality and keel bone
fractures in laying hens (Malladi et al., 2011; Kozak et al., 2016). Furthermore, the flock activity
can be transformed into an animal distribution index, which has the potential to indicate some
underlying animal health problems (Kashiha et al., 2013; Peña Fernández et al., 2018; Guo et al.,
2020).
The above results suggest that automatic monitoring of poultry activity can detect disease at
an early stage. However, most of the current related studies focus on flocks, and how to precisely
identify the suspicious single chicken with abnormal activity from the group is another challenge.
3.2. Abnormal posture
Different from activity, posture represents a static characteristic. It is understood that chickens
will have not enough energy to maintain a normal posture when the body function deviates from
homeostasis.
Zhuang et al. (2018) have successfully identified healthy and avian influenza infected
chickens by observing the posture change in broilers. In their study, the segmentation algorithms
were used to separate chicken from background to extract the contour and skeleton information.
Six eigenvalues including concavity, skeleton attitude angle, shape features, area-linear rate,
elongation, and circularity were calculated manually to constitute the eigenvector, and finally
machine learning algorithms were used to evaluate the health status of broilers. The results
showed that Support Vector Machine model got the highest accuracy of 99.469%. The algorithms
proposed in their study can automatically detect sick chickens, but the procedures are complicated.
Before the health status assessment, the head area of chickens needs to be extracted to determine
whether they are in the foraging state, i.e., the algorithms cannot work when chickens are not
completely facing the camera. In order to analyze the postural differences between healthy and
sick chickens infected with Newcastle disease virus, Okinda et al. (2019) designed and calculated
another set of posture shape features including circle variance, elongation, convexity, complexity,
and eccentricity, and walk speed was also included in the eigenvector in their study. The results
suggested that early warning can be achieved before a major outbreak, and the earliest time that
Newcastle disease can be detected was on the 4th day based on circle variance and elongation,
and the 6th day based on eccentricity and walk speed. Similarly, the Support Vector Machine
model achieved the highest accuracy of 98.8%. The improvement is that this method only needs
to obtain the top view of the chicken to carry out subsequent recognition, reducing the dependence
on the orientation of chickens when taking images.
Artificial feature extraction algorithm is mainly based on morphological description.
Although it is feasible to some extent, the recognition accuracy is highly dependent on image
quality, which limits its use and promotion in complex background. Therefore, the posture
recognition methods based on deep learning was developed. To automatically detect sick broilers
in a flock, Zhuang et al. (2019) proposed an optimized model named Improved Feature Fusion
Single Shot MultiBox Detector based on the original Single Shot MultiBox Detector model. The
model can detect and identify the health status of broilers simultaneously, and achieved a mean
∙ 282 ∙
average precision of 99.7% (intersection-over-union >0.5). If automatic and accurate

classification of different chicken postures like standing, sleeping, water-taking, food-taking
feather pecking and fighting can be achieved, it will provide with richer information and further
promote efficient flock management (Thenmozhi et al., 2020).
4. Conclusions
Disease control is a critical challenge for poultry farmers. Traditional manual-based disease
inspection methods cannot meet the needs of large-scale operations. Clinical symptoms are the
most direct manifestation of poultry disease. Automatic disease detection and early warning
technologies based on clinical symptoms has great potential to monitor poultry health status in a
real-time and non-destructive way. Most of the relevant studies were carried out in experimental
environment, which can potentially be extended to commercial conditions. Table 2 lists the
comparison of five clinical symptoms used for disease early warning.
Table 2. Comparison of five clinical symptoms used for disease warning.
Clinical Collection Monitoring
Advantages Disadvantages
symptoms devices level
Microphone and  Low cost  High interference from
Vocalization Group
camera  Easy installation environmental noise
 Insufficient endurance
Temperature  Relatively mature
Individual  Short transmission
sensor  Widely used
distance
Temperature  Remote temperature
measurement
Infrared thermal  Easily affected by the
Group  High visualization
imager environment
degree of temperature
distribution
 Bad environment may
 Low cost
result in poor equipment
 Can monitor the
Camera Group life
chickens in the whole
Feces  Strict light conditions are
coop
required
 Can identify specific  No relevant studies have
Spectroscope Group
biomarkers been found
Acceleration  Low cost
Individual  Easy to fall off
sensor  High sensitivity
Activity
 Can monitor multiple  Severely affected by light
Camera Group
objects simultaneously conditions
 Algorithms for solving the
 Little effect on chicken
Posture Camera Individual chicken overlap problem
 High intelligence
are not well developed
Although the importance of clinical symptoms in early warning of poultry diseases has been
recognized, it has not been fully utilized. Cockscomb abnormalities (abnormal color, swollen,
scabby, shrunken), ocular anomalies (congestion, lacrimation) and the change of feathers (fluffy,
matte, and fall off) of chickens are important clinical symptoms as well. Multimodal technology
is a means to integrate multi-source data for analysis and decision-making, and has been
successfully applied to many fields such as medicine and emotion recognition. In poultry industry,
the application of this technology to comprehensively analyze and understand the information of
multiple clinical symptoms will further improve the accuracy of disease diagnosis and facilitate
the establishment of disease early warning system.
∙ 283 ∙
References
Astill, J., R.A. Dara, E.D.G. Fraser, S. Sharif, 2018. Detecting and Predicting Emerging Disease in Poultry
With the Implementation of New Technologies and Big Data: A Focus on Avian Influenza Virus. Front Vet
Sci. 5, 263.
Aydin, A., D. Berckmans, 2016. Using sound technology to automatically detect the short-term feeding
behaviours of broiler chickens. Computers and Electronics in Agriculture. 121, 25–31.
Aydin, A., 2017a. Development of an early detection system for lameness of broilers using computer
vision. Computers and Electronics in Agriculture. 136, 140–146.
Aydin, A., 2017b. Using 3D vision camera system to automatically assess the level of inactivity in broiler
chickens. Computers and Electronics in Agriculture. 135, 4–10.
Aziz, N.A., M.F. Bin Othman, 2017. Binary Classification using SVM for Sick and Healthy Chicken
based on Chicken's Excrement Image. Pertanika Journal of Science and Technology. 25, 315–324.
Ben Sassi, N., X. Averos, I. Estevez, 2016. Technology and Poultry Welfare. Animals (Basel). 6 (10), 62.
Blake, D.P., J. Knox, B. Dehaeck, B. Huntington, T. Rathinam, V. Ravipati, S. Ayoade, W. Gilbert, A.O.
Adebambo, I.D. Jatau, M. Raman, D. Parker, J. Rushton, F.M. Tomley, 2020. Re-calculating the cost of
coccidiosis in chickens. Vet Res. 51 (1), 115.
Blas, A., B. Diezma, A. Moya, C. Gomez, 2014. Early detection of mortality in poultry production using
high resolution thermography. 1499–1504 p.
Cao, Y., D. Li, 2013. Impact of increased demand for animal protein products in Asian countries:
Implications on global food security. Animal Frontiers. 3 (3), 48–55.
Carpentier, L., D. Berckmans, A. Youssef, D. Berckmans, T. van Waterschoot, D. Johnston, N. Ferguson,
B. Earley, I. Fontana, E. Tullo, M. Guarino, E. Vranken, T. Norton, 2018. Automatic cough detection for
bovine respiratory disease in a calf house. Biosystems Engineering. 173, 45–56.
Carpentier, L., E. Vranken, D. Berckmans, J. Paeshuyse, T. Norton, 2019. Development of sound-based
poultry health monitoring tool for automated sneeze detection. Computers and Electronics in Agriculture. 162,
573–581.
Carroll, B.T., D.V. Anderson, W. Daley, S. Harbert, D.F. Britton, M.W. Jackwood, Ieee, 2014. Detecting
Symptoms of Diseases in Poultry through Audio Signal Processing. 1132–1135 p.
Chowdhury, E.U., A. Morey, 2020. Application of optical technologies in the US poultry slaughter
facilities for the detection of poultry carcase condemnation. Br Poult Sci. 61 (6), 646–652.
Colles, F.M., R.J. Cain, T. Nickson, A.L. Smith, S.J. Roberts, M.C.J. Maiden, D. Lunn, M.S. Dawkins,
2016. Monitoring chicken flock behaviour provides early warning of infection by human pathogen
Campylobacter. Proceedings of the Royal Society B: Biological Sciences. 283 (1822), 20152323.
Cuan, K.X., T.M. Zhang, J.D. Huang, C. Fang, Y. Guan, 2020. Detection of avian influenza-infected
chickens based on a chicken sound convolutional neural network. Computers and Electronics in Agriculture.
178, 105688.
Daghir, N., M. Diab-El-Harake, S. Kharroubi, 2021. Poultry production and its effects on food security
in the Middle Eastern and North African region. Journal of Applied Poultry Research. 30 (1), 100110.
Dawkins, M.S., S.J. Roberts, R.J. Cain, T. Nickson, C.A. Donnelly, 2017. Early warning of footpad
dermatitis and hockburn in broiler chicken flocks using optical flow, bodyweight and water consumption. Vet
Rec. 180 (20), 499.
de Jong, I.C., V.A. Hindle, A. Butterworth, B. Engel, P. Ferrari, H. Gunnink, T.P. Moya, F.A.M. Tuyttens,
C.G. van Reenen, 2016. Simplifying the Welfare Quality((R)) assessment protocol for broiler chicken welfare.
Animal. 10 (1), 117–127.
Du, X.D., F.D. Lao, G.H. Teng, 2018. A Sound Source Localisation Analytical Method for Monitoring
the Abnormal Night Vocalisations of Poultry. Sensors. 18 (9), 2906.
Du, X.D., L. Carpentier, G.H. Teng, M.L. Liu, C.Y. Wang, T. Norton, 2020. Assessment of Laying Hens'
Thermal Comfort Using Sound Technology. Sensors. 20 (2), 473.
Ducatelle, R., E. Goossens, F. De Meyer, V. Eeckhaut, G. Antonissen, F. Haesebrouck, F. Van Immerseel,
2018. Biomarkers for monitoring intestinal health in poultry: present status and future perspectives. Veterinary
Research. 49.
Fontana, I., E. Tullo, A. Butterworth, M. Guarino, 2015. An innovative approach to predict the growth in
intensive poultry farming. Computers and Electronics in Agriculture. 119, 178–183.
Gerber, P., C. Opio, H.J.A.p. Steinfeld, F. health division, V.d.T.d.C. Agriculture Organization of the
United Nations, 2007. Poultry production and the environment—a review. 153.
Gonzalez, J.J., A. Nasirahmadi, U. Knierim, 2020. Automatically Detected Pecking Activity in Group-
Housed Turkeys. Animals (Basel). 10 (11), 2034.
∙ 284 ∙
Guo, Y.Y., L.L. Chai, S.E. Aggrey, A. Oladeinde, J. Johnson, G. Zock, 2020. A Machine Vision-Based
Method for Monitoring Broiler Chicken Floor Distribution. Sensors. 20 (11), 3179.
Hoffmann, G., C. Ammon, L. Volkamer, C. Suerie, D. Radko, 2013. Sensor-based monitoring of the
prevalence and severity of foot pad dermatitis in broiler chickens. British Poultry Science. 54 (5), 553–561.
Kalthoff, D., A. Globig, M. Beer, 2010. (Highly pathogenic) avian influenza as a zoonotic agent. Vet
Microbiol. 140 (3–4), 237–245.
Kashiha, M., A. Pluk, C. Bahr, E. Vranken, D. Berckmans, 2013. Development of an early warning
system for a broiler house using computer vision. Biosystems Engineering. 116 (1), 36–45.
Kaukonen, E., M. Norring, A. Valros, 2016. Effect of litter quality on foot pad dermatitis, hock burns and
breast blisters in broiler breeders during the production period. Avian Pathology. 45 (6), 667–673.
Knowles, T.G., S.C. Kestin, S.M. Haslam, S.N. Brown, L.E. Green, A. Butterworth, S.J. Pope, D. Pfeiffer,
C.J. Nicol, 2008. Leg Disorders in Broiler Chickens: Prevalence, Risk Factors and Prevention. Plos One. 3
(2), e1545.
Kozak, M., B. Tobalske, D. Springthorpe, B. Szkotnicki, A. Harlander-Matauschek, 2016. Development
of physical activity levels in laying hens in three-dimensional aviaries. Applied Animal Behaviour Science.
185, 66–72.
Krautwald-Junghanns, M.E., R. Ellerich, H. Mitterer-Istyagin, M. Ludewig, K. Fehlhaber, E. Schuster, J.
Berk, S. Petermann, T. Bartels, 2011. Examinations on the prevalence of footpad lesions and breast skin
lesions in British United Turkeys Big 6 fattening turkeys in Germany. Part I: prevalence of footpad lesions.
Poult Sci. 90 (3), 555–560.
Kristensen, H.H., C. Cornou, 2011. Automatic detection of deviations in activity levels in groups of
broiler chickens – A pilot study. Biosystems Engineering. 109 (4), 369–376.
Li, N., Z. Ren, D. Li, L. Zeng, 2020a. Review: Automated techniques for monitoring the behaviour and
welfare of broilers and laying hens: towards the goal of precision livestock farming. Animal. 14 (3), 617–625.
Li, X., C. Nie, Y. Liu, Y. Chen, X. Lv, L. Wang, J. Zhang, W. Yang, K. Li, C. Zheng, Y. Jia, Z. Ning, L.
Qu, 2020b. The Genetic Architecture of Early Body Temperature and Its Correlation With Salmonella
Pullorum Resistance in Three Chicken Breeds. Frontiers in Genetics. 10, 1287.
Li, Y., R. Zheng, Y. Wu, K. Chu, Q. Xu, M. Sun, Z.J. Smith, 2019. A low-cost, automated parasite
diagnostic system via a portable, robotic microscope and deep learning. J Biophotonics. 12 (9), e201800410.
Liu, X., F. Wang, Y. Liu, Y. Wu, H. Lu, H.J.C.P. Yan, 2017. Comparative study on surface temperature
between diseased and healthy layers. China Poultry. 39 (2), 53–56.
Liu, Y.L., W.R. Windham, K.C. Lawrence, B. Park, 2003. Simple algorithms for the classification of
visible/near-infrared and hyperspectral imaging spectra of chicken skins, feces, and fecal contaminated skins.
Applied Spectroscopy. 57 (12), 1609–1612.
Mahdavian, A., S. Minaei, P.M. Marchetto, F. Almasganj, S. Rahimi, C. Yang, 2021. Acoustic features
of vocalization signal in poultry health monitoring. Applied Acoustics. 175, 107756.
Malladi, S., J.T. Weaver, T.L. Clouse, K.E. Bjork, D.W. Trampel, 2011. Moving-Average Trigger for
Early Detection of Rapidly Increasing Mortality in Caged Table-Egg Layers. Avian Diseases. 55 (4), 603–
610.
Manteuffel, G., B. Puppe, P.C. Schön, 2004. Vocalization of farm animals as a measure of welfare.
Applied Animal Behaviour Science. 88 (1–2), 163–182.
Moe, R.O., J. Bohlin, A. Flo, G. Vasdal, H. Erlandsen, E. Guneriussen, E.C. Sjokvist, S.M. Stubsjoen,
2018. Effects of subclinical footpad dermatitis and emotional arousal on surface foot temperature recorded
with infrared thermography in turkey toms (Meleagris gallopavo). Poult Sci. 97 (7), 2249–2257.
Mottet, A., G. Tempio, 2017. Global poultry production: current state and future outlook and challenges.
Worlds Poultry Science Journal. 73 (2), 245–256.
Nasirahmadi, A., J. Gonzalez, B. Sturm, O. Hensel, U. Knierim, 2020. Pecking activity detection in
group–housed turkeys using acoustic data and a deep learning technique. Biosystems Engineering. 194, 40-
48.
Noh, J.Y., K.J. Kim, S.H. Lee, J.B. Kim, D.H. Kim, S. Youk, C.S. Song, S.S. Nahm, 2021. Thermal
Image Scanning for the Early Detection of Fever Induced by Highly Pathogenic Avian Influenza Virus
Infection in Chickens and Ducks and Its Application in Farms. Frontiers in Veterinary Science. 8, 616755.
Okada, H., T. Itoh, K. Suzuki, K. Tsukamoto, 2009. Wireless sensor system for detection of avian
influenza outbreak farms at an early stage. 2009 Ieee Sensors, Vols 1–3, 1374–+.
Okinda, C., M. Lu, L. Liu, I. Nyalala, C. Muneri, J. Wang, H. Zhang, M. Shen, 2019. A machine vision
system for early detection and prediction of sick birds: A broiler chicken model. Biosystems Engineering.
188, 229–242.
∙ 285 ∙
Peña Fernández, A., T. Norton, E. Tullo, T. van Hertem, A. Youssef, V. Exadaktylos, E. Vranken, M.
Guarino, D. Berckmans, 2018. Real-time monitoring of broiler flock's welfare status using camera-based
technology. Biosystems Engineering. 173, 103–114.
Shang, Y.H., D.X. Tao, Y.Q. Hou, J. Wang, Y.F. Pei, 2017. Accurate Measuring Temperature by Infrared
Thermal Imaging system in Vacuum and Cryogenic Environment. 2017 International Conference on Optical
Instruments and Technology – Optoelectronic Measurement Technology and Systems. 10621, UNSP
106210L.
Silvera, A.M., T.G. Knowles, A. Butterworth, D. Berckmans, E. Vranken, H.J. Blokhuis, 2017. Lameness
assessment with automatic monitoring of activity in commercial broiler flocks. Poult Sci. 96 (7), 2013–2017.
Thenmozhi, M., M. Saravanan, K.P.M. Kumar, S. Suseela, S. Deepan, 2020. Improving the prediction
rate of unusual behaviors of animal in a poultry using deep learning technique. Soft Computing. 24 (19),
14491–14502.
Thevenoux, R., V.L. Le, H. Villessèche, A. Buisson, M. Beurton-Aimar, E. Grenier, L. Folcher, N.
Parisey, 2021. Image based species identification of Globodera quarantine nematodes using computer vision
and deep learning. Computers and Electronics in Agriculture. 186.
Vidic, J., M. Manzano, C.M. Chang, N. Jaffrezic-Renault, 2017. Advanced biosensors for detection of
pathogens related to livestock and poultry. Vet Res. 48 (1), 11.
Wang, J., M. Shen, L. Liu, Y. Xu, C. Okinda, 2019. Recognition and Classification of Broiler Droppings
Based on Deep Convolutional Neural Network. Journal of Sensors. 2019, 1–10.
Whitaker, B.M., B.T. Carroll, W. Daley, D.V. Anderson, Ieee, 2014. Sparse Decomposition of Audio
Spectrograms for Automated Disease Detection in Chickens. 1122–1126 p.
Wilcox, C.S., J. Patterson, H.W. Cheng, 2009. Use of thermography to screen for subclinical bumblefoot
in poultry. Poultry Science. 88 (6), 1176–1180.
Wu, K., T. Klein, V.D. Krishna, D.Q. Su, A.M. Perez, J.P. Wang, 2017. Portable GMR Handheld Platform
for the Detection of Influenza A Virus. Acs Sensors. 2 (11), 1594–1601.
Yoon, S.C., B. Park, K.C. Lawrence, W.R. Windham, G.W. Heitschmidt, 2011. Line-scan hyperspectral
imaging system for real-time inspection of poultry carcasses with fecal material and ingesta. Computers and
Electronics in Agriculture. 79 (2), 159–168.
Zhang, J., Y. Lin, Y. Liu, Z. Li, Z. Li, S. Hu, Z. Liu, D. Lin, Z. Wu, 2014. Cascaded-Automatic
Segmentation for Schistosoma japonicum eggs in images of fecal samples. Comput Biol Med. 52, 18–27.
Zhang, R.S., D. Yao, J.F. Chen, W. Ye, X.H. Ou, T.M. Chen, B.C. Sun, 2018. Development and
evaluation of a real-time RT-PCR assay for detection of a novel avian influenza A (H5N6) virus. Journal of
Virological Methods. 257, 79–84.
Zhang, Z.J., D. Liu, J. Hu, W.Q. Sun, K.T. Liu, J. Li, H.X. Xu, J. Liu, L.H. He, D.X. Jiang, M. Gu, S.L.
Hu, X.Q. Wang, X.W. Liu, X. Liu, 2019. Multiplex one-step real-time PCR assay for rapid simultaneous
detection of velogenic and mesogenic Newcastle disease virus and H5-subtype avian influenza virus. Archives
of Virology. 164 (4), 1111–1119.
Zhuang, X., M. Bi, J. Guo, S. Wu, T. Zhang, 2018. Development of an early warning algorithm to detect
sick broilers. Computers and Electronics in Agriculture. 144, 102–113.
Zhuang, X., T. Zhang, 2019. Detection of sick broilers by digital image processing and deep learning.
Biosystems Engineering. 179, 106–116.
∙ 286 ∙
A Key Frame Selection Method for Creating Deep Learning

Training Set in Animal Research Involving Time-Series Video Data
Endai Huang a, Axiu Mao b, Haiming Gan b,c, Kai Liu b,*
a
Department of Computer Science, City University of Hong Kong, Hong Kong SAR, China
b
Department of Infectious Diseases and Public Health, Jockey Club College of Veterinary Medicine and
Life Sciences, City University of Hong Kong, Hong Kong SAR, China
c
Colleges of Electronic Engineering and Artificial Intelligence, South China Agricultural University,
Guangzhou, Guangdong, China
* Corresponding author. Email: kailiu@cityu.edu.hk
Abstract
Computer vision has been widely used in animal behavior and health monitoring, which
enables us a huge amount of image and video data. To date, various deep learning methods have
been proved to be powerful in analyzing these data. However, constructing training datasets with
a large number of well-labeled images/videos are needed as an essential step, and data labeling is
a labor-intensive and time-consuming task. In general, in animal studies involving time-series
video data, one video usually contains hundreds of thousands of frames, and many frames are
highly similar especially when they are consecutive with minimal animal movement. These
similar frames make minimal marginal contribution to the training but cost equal efforts when
labeling them. To solve this problem, we developed a global and local similarity index (GLSI)
and a key frame selection method, i.e., uniform selection based on GLSI (GLSI-US), which
selected images uniformly after being sorted by GLSI. To evaluate the training set generated by
GLSI-US, the object detection was used as a case deep learning task. The GLSI-US was compared
with random selection (RS) and the selection using the top k highest and the lowest GLSI (GLSI-
SH and GLSI-SL), respectively. The results revealed that the GLSI-US outperformed the GLSI-
SH, the GLSI-SL, and the RS method (up to 6.4%, 3.4%, and 1.6% in mean average precision,
respectively). The results of the three GLSI-based strategies demonstrated that training set
selection could largely affect the training result and there was an exclusiveness between image
uniqueness and image frequency.
Keywords: Image similarity, representative dataset, perceptual hash algorithm, precision
livestock farming, computer vision.
1. Introduction
Computer vision has been widely used in animal behavior and health monitoring, which
enables us a huge amount of image and video data (Fernandes et al., 2020). For example, a public
penguin dataset with over 500,000 penguin images was constructed for penguin counting (Arteta
et al., 2016). The Caltech-UCSD Birds-200-2011 dataset has about 12,000 images of over 200
bird species (Wah et al., 2011). In total, 3436 natural bird scenes of more than 38 bird species
were collected from 21 different sites in Turkey for counting the birds (Akçay, et al., 2021).
Continuous (24 h per day with 6s intervals) 3D images were collected during the late gestation
and farrowing period of sows for sow behaviors analysis (Lao et al., 2016). Dataset of farrowing
pens with 377,572 frames in 1,009 episodes was constructed and labeled for sow nursing behavior
detection (Yang et al., 2018). With the huge amount of data generated, how to efficiently make
use of them became a challenge.
To analyze the image and video data, various deep learning methods have been proposed and
proved to be powerful. For instance, different object detection methods in deep learning were
compared to count the bird in UAV captured photos in Korea (Hong et al., 2019). Faster R-CNN
(Ren et al., 2015) and Neural Architecture Search (NAS) based network were combined for pig
position and posture detection (Riekert et al., 2020). Pig videos were fed to a convolutional neural
network (CNN) and then to the Long Short-term Memory (LSTM) to recognize the feeding
∙ 287 ∙
behavior and their feeding time (Chen et al., 2020). An adopted counting network was proposed
to count the pig in actual piggery farming images and achieved a mean absolute error (MAE) of
1.67 per image (Tian et al., 2019).
However, when applying the deep learning method to animal industry, a large number of well-
labeled images or videos are needed to construct a training dataset for training the deep learning
model. Data labeling is indeed a labor-intensive and time-consuming task, which limits the
application of deep learning method. As a common practice, animal videos were first recorded
and then the animal image dataset was generated by extracting frames uniformly from these videos
(e.g., one frame per second or per 30 frames). After being labeled by human, these images were
used to train a deep learning model, and then the trained model were used to analyze the video or
entire image dataset. In general, one animal video contains hundreds of thousands of frames and
many frames are highly similar, especially when they are consecutive and with no or minimal
animal movement. These frames with high similarity make no or minimal marginal contribution
to the model training but cost equal efforts when labeling them. Therefore, it is of significant
importance to investigate how to efficiently select key frames for constructing training datasets
for deep learning methods.
The objective of this study was to develop a key frame selection method to efficiently choose
representative frames to construct a training set from a huge number of consecutive frames.
Specifically, a global and local similarity index (GLSI) was developed, which considered both the
global similarity of an individual frame with all other frames within a video and the local similarity
with its neighbor frames. Then a uniform selection based on GLSI (GLSI-US) was proposed,
which selected images uniformly after they were sorted by GLSI. The similarity was calculated
based on the perceptual hash algorithm (Liu et al., 2013) and Hamming distance (Hamming,
1950). Our method aimed to optimize the training dataset construction by retaining representative
frames and enhance the training results when training set size is limited (i.e., by removing
redundant frames while keeping unique frames). Furthermore, our method is a user-adjustable
method and can be extended and applied in other scenarios where a training set need to be
constructed.
2.1. Animals and videos
Videos of pigs in farrowing pens collected at the Swine Teaching and Research Center at the
University of Pennsylvania School of Veterinary Medicine were used in this study. The farrowing
pen scenario was chosen because the sow and piglets were still for most of the time (resting or
sleeping periodically throughout the daytime), which enlarged the similarity between consecutive
video frames. During data acquisition, a camera (IPX DDK-1700D Infrared IP Dome Camera,
Farmingdale, New Jersey, USA) with a resolution of 1024 × 768 pixels was positioned 2 m above
each farrowing pen to record lactating sows (Line 241; DNA Genetics, Columbus, NE) and their
piglets (Figure 1). The detailed information about the pigs, farrowing pens, and the video
collection procedure has been described in a separate study (Ceballos et al., 2020). After data
collection, images were extracted every 50 frames from a video during 8:40 and 15:20 when there
was sufficient light, and when the sow and piglets were more active. There were one sow and nine
piglets in this video and extracted images. In total, 3,130 images were collected and two objects,
i.e., piglets and sows, were labeled using bounding boxes via software LabelImg (Tzutalin, 2015).
2.2. Perceptual hash algorithm
The perceptual hash algorithm is a class of hashing algorithms that transform an image to a
fingerprint (e.g., a 64-bit binary string) (Figure 2). Firstly, the input image was resized to 32 × 32
pixels (Figure 2b) and then transformed to a 32 × 32 pixels gray image (Figure 2c). Then the gray
image was transformed to a 32 × 32 matrix by Discrete Cosine Transform (DCT) (Figure 2d). The
upper-left 8 × 8 matrix (denoted as Mul) was reserved and binarized to a binary matrix MB (Figure
∙ 288 ∙
2f) using the average value that,

(𝑖,𝑗)
= {1, 𝑖𝑓 𝑀𝑢𝑙 ≥ 𝑎𝑣𝑒𝑟𝑎𝑔𝑒(𝑀𝑢𝑙 )
(𝑖,𝑗)
𝑀𝐵 (1)
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
and this 8 × 8 binary matrix MB was flattened to a 64-bit binary fingerprint as the algorithm output
(Figure 2e).
Figure 1. Experimental setting of the farrowing pen. (a) Farrowing pen layout and (b) an image
example captured by the camera.
Figure 2. The flow of the perceptual hash algorithm. DCT stands for Discrete Cosine Transform.
If two images were similar, the perceptual hash algorithm would output two similar
fingerprints, and the difference between the two fingerprints could be estimated by Hamming
distance. Specifically, the XOR operation was performed on the two fingerprints and then the total
number of ‘1’s in the results was the Hamming distance (an example was showed in Figure 3).
Thereafter, the Hamming distance was rescaled to [0,1] to represent the similarity that,
𝑠𝑖𝑚(𝑓1 , 𝑓2 ) = 1 − ℎ𝑎𝑚(𝑓1 , 𝑓2 )/64 (2)
where ‘1’ meant high similarity and ‘0’ meant low similarity, and f1, and f2 were fingerprints of
two images after the perceptual hash algorithm.
∙ 289 ∙
Figure 3. An example of calculating Hamming distance between two fingerprints. Two 8-bit
fingerprints are used for demonstration.
2.3. Global and local similarity index (GLSI)
Similarities between images need to be defined and calculated. We denoted our image
fingerprint set as F = {f1, f2, ..., fn}, where n was the total number of images (n = 3,130 in this
experiment) and the set was sorted by their frame order. Two kinds of similarity were defined:
global similarity (designated as Sglobal) and local similarity (designated as Slocal).
The global similarity Sglobal was a positive symmetric matrix which indicated the similarity
between an image and all rest images within a set as
(𝑖,𝑗)
𝑆𝑔𝑙𝑜𝑏𝑎𝑙 = 𝑠𝑖𝑚(𝑓𝑖 , 𝑓𝑗 ), for 𝑖, 𝑗 = 1,2, . . . , 𝑛. (3)
The local similarity Slocal was represented as
1
𝑆𝑙𝑜𝑐𝑎𝑙 = (𝑊 − 𝐼) ⊙ 𝑆𝑔𝑙𝑜𝑏𝑎𝑙 (4)
𝜎√2𝜋
where W was Gaussian weight for neighbor images that
(𝑗−𝑖)2
1
𝑊 (𝑖,𝑗) = 𝑒− 2𝜎2 , for 𝑖, 𝑗 = 1,2, . . . , 𝑛 (5)
𝜎√2𝜋
and I was an identity matrix. ⊙ was Hadamard product (element-wise product) for the two
matrices, and sigma was the standard deviation for Gaussian distribution. Herein the sigma
controlled the Gaussian weight curve to be sharp or flat, and thus controlled the range of the
neighbor images that the local similarity concerned.
The meaning of the local similarity matrix was that when two images were from closer frames
(e.g., i was near j), Gaussian weight W(i,j) became larger and therefore the similarities of closer
images had a higher weight. A term with the identity matrix (Eq. (4)) was deducted to eliminate
the similarity of an image with itself (when i = j).
Finally, the global similarity matrix (Eq. (3)) and the local similarity matrix (Eq. (4)) were
combined and summed up to the global and local similarity index (GLSI) with a weight lambda
that,
1
𝐺𝐿𝑆𝐼 = (𝑆 − 𝐼)𝟏 + 𝜆𝑆𝑙𝑜𝑐𝑎𝑙 𝟏 (6)
𝑛−1 𝑔𝑙𝑜𝑏𝑎𝑙
where 1 was a column vector with all elements equal to 1 and multiplying this vector was to sum
up the row values of a matrix. 𝐺𝐿𝑆𝐼 ∈ 𝑅𝑛 was a column vector, where each element corresponded
the similarity of each image fingerprint in F. There were two hyperparameters for GLSI, one was
sigma (Eq. (5)) controlling the range of neighbor images the local similarity concerned, and the
other was a sum weight lambda (Eq. (6)) determining how important the local similarity was.
These two hyperparameters were user-adjustable parameters and could be adjusted according to
the user’s dataset and the purpose of selection.
2.4. Uniform selection based on GLSI (GLSI-US)
There was an exclusiveness between the image uniqueness and image frequency that an image
with high similarity with other images meant this image appeared in the video with high frequency
(less uniqueness), and vice versa. Therefore, the training set could not be formed directly by
choosing images with minimum GLSI. To include both images with high variation with others
∙ 290 ∙
and the images appeared with high frequency in the training set, all images were sorted by their
GLSI from low to high and then the training dataset was generated by selecting certain number
(e.g., k images) of images uniformly in the sorted image list. Herein k could be determined by
users according to the characteristics of the dataset and the time available in data labeling and
training. This uniform image selection according to their sorted GLSI was what we called the
uniform selection based on GLSI (GLSI-US).
2.5. Method evaluation and comparison
To assess whether the GLSI could detect and rank the similarities among images, three GLSI-
based selection strategies were compared, including the selection using the top k images with
highest GLSI (GLSI-SH), the selection using the top k images with lowest GLSI (GLSI-SL), and
the GLSI-US. The GLSI-SH and GLSI-SL were two extreme strategies that the GLSI-SH would
choose images with the highest similarity and thus decrease the generalization of deep learning
model, while the GLSI-SL would choose images with the highest variation but omit the images
with high frequency. In comparison, the GLSI-US, to a large extend, could potentially avoid these
two dilemmas as it selected images uniformly according to the sorted GLSI list. Therefore, the
accuracy enhancement may be expected for the GLSI-US as compared to the GLSI-SH and GLSI-
SL, if the GLSI could rank the images according to their similarities effectively and correctly.
To evaluate whether the GLSI-US can select the more representative images from consecutive
video frames than other selection methods, the GLSI-US was compared with a widely used
selection strategy, namely, random selection (RS). The RS was performed twice with different
random seeds. Here the ‘representative images’ meant a small image subset selected from a
complete image dataset and deep learning models trained on this subset could produce a
comparable performance as the training on the entire image dataset. ‘More representative images’
meant an image subset with a fixed train set size that enabled the deep learning model to have a
better test results on the entire image dataset than other subsets. With using the GLSI-US, a total
of 10 image subsets, with k = 100, 200, ..., 1000, respectively, were extracted from a total of 3,130
images and formed 10 different training sets as case studies.
A state-of-the-art deep learning object detection method Yolov4 (Bochkovskiy et al., 2020)
was selected as a case deep learning model. The Yolov4 detected both piglets and sows in the
images and output bounding boxes for each individual animal. The common metric of Yolov4
was mAP (when Intersection over Union (IoU) threshold = 0.5), and this metric was used to
evaluate the model performance.
2.6. Experimental settings
All the training epochs were 6,000 iterations and all the trained models were tested on the
entire image set (3,130 images). For the two user-adjustable hyperparameters, i.e., sigma and
lambda, sigma (Eq. (5)) was set to 2 so that the Gaussian weight W affected about 10 nearby
frames with significant impacts, and weight lambda (Eq. (6)) was set to 3. These values were set
according to the experiment and the characters of our dataset. All experiments were conducted on
a NVIDIA GeForce RTX 2080 Ti GPU and a NVIDIA GeForce RTX 3090 GPU.
To assess whether the GLSI can detect and rank the similarities among images, the GLSI-US
was compared with the GLSI-SH and the GLSI-SL (Figure 4). The results showed that the GLSI-
US achieved the best performance (mAP ranged from 96.6% to 99.2%) in each of the 10 cases
with different training set size, while the GLSI-SH showed the worst performance when the
training set was small and less than 600 images (mAP ranged from 90.4% to 95.2%). In addition,
the GLSI-SH and GLSI-SL had comparable performance when the training set size ranged from
600 to 1,000 images (mAP ranged from 96.1% to 97.2%). Besides, the enhancement from the
GLSI-US was more significant when training set size was small. Specifically, in the best case
(i.e., when k = 100), the mAP of the GLSI-US was 6.4% and 3.4% higher than the mAP of the
∙ 291 ∙
GLSI-SH and GLSI-SL, respectively. These results indicated that the GLSI ranked the images
effectively by their similarity.
To evaluate whether more representative images were selected from consecutive video
frames, the GLSI-US was compared with the RS (Figure 4). The GLSI-US outperformed the RS
in all cases. Specifically, the mAP of the GLSI-US was 0.2-1.6% higher than the mAP of the RS
in each of the 10 cases with different training set size. The two RS experiments showed
comparable results, and the differences between the two results disappeared when training set size
became larger. These results indicated that the GLSI-US could construct a better training set with
more representative images and thus enhance the training results than the common RS method.
Figure 4. Test results on whole images using different selection methods and different training
set sizes. GLSI, global and local similarity index; GLSI-US, uniform selection based on GLSI;
GLSI-SH, selection method using the highest GLSI; GLSI-SL, selection method using the
lowest GLSI; IoU, Intersection over Union; mAP, mean Average Precision; RS, random
selection.
To further illustrate that the GLSI could detect and rank the similarities among consecutive
images, the top eight images with the lowest GLSI and the highest GLSI were visualized in Figure
5 and Figure 6, respectively. The theoretical range of the GLSI was 0 (if all images were extremely
different) to 3.402 (if all images were the same) in this experiment. The top eight lowest GLSI
varied from 1.694 to 2.346, while the top eight highest GLSI were similar and ranged from 3.186
to 3.190. In theory, frames at the beginning and the end of the video tended to have lower GLSI
as these frames had fewer neighbor frames comparing with others. This assumption was supported
by the findings in our study. Specifically, images with the first (GLSI=1.694), second
(GLSI=1.935), and the third (GLSI=2.138) lowest GLSI (Figure 5a-c) were two beginning frames
and one end frame in the video, respectively. As for the rest five images displayed (Figure 5d-g),
there were obvious animal movements observed, e.g., the sow was turning over from lying
suddenly (Figure 5h). In terms of the top eight images with the highest GLSI (Figure 6), all of
them were images in which the sow was lying with no or minimum movement, and the piglets
were lying or sucking with tiny movement. This indicated that the lying was indeed a high-
frequency posture for sows, which was consistent with the findings in other research that sows
spent 84.0% of time lying down (Lao et al., 2016).
Comparing the three GLSI-based select strategies, the training set selection largely affected
the training results and there existed an exclusiveness between the image uniqueness and image
frequency when constructing training set using the GLSI. The GLSI-SH showed the worst
performance in this study especially when the training set had limited number of images.
Specifically, when the training set had less than 600 images, the GLSI-SH only chose the top
∙ 292 ∙
images with high similarity, and thus the deep learning model could only learn a limited number
of situations. Taking an extreme case as an example, if only top eight images with highest GLSI
(Figure 6) were used to train the deep learning model, the model indeed equivalently learned only
three situations (i.e., Figure 6a, 6d, and 6e as situation 1; Figure 6b and c as situation 2; Figure 6f-
h as situation 3), which was a waste of time in both data labeling and training. In contrast, the
images selected by the GLSI-SL varied a lot (Figure 5), which increased the generalization of the
deep learning model trained. However, as the GLSI-SL eliminated the images with high GLSI
(high frequent images), to a large extend, the capacity of the GLSI-SL was also restricted.
Consequently, despite the deep learning model trained using the eight images from Figure 5 could
learn from different situations, the GLSI-US, which selected images uniformly according to their
sorted GLSI, outperformed other two GLSI-based methods in this study.
Figure 5. Top eight images with lowest GLSI.
Figure 6. Top eight images with highest GLSI.

4. Conclusions
To select representative frames from videos and save time in data labeling, we developed a
global and local similarity index (GLSI) and a key frame selection method, namely, uniform
selection based on GLSI (GLSI-US). In GLSI-US, the similarity of video frames was determined
using the perceptual hash algorithm. In this study with images extracted from videos of farrowing
pens, the GLSI-US outperformed the GLSI-SH, the GLSI-SL, and random selection, i.e., up to
6.4%, 3.4%, 1.6% in mAP, respectively. The visualization showed that our method could detect
and rank the similarities among consecutive images. Comparing the three GLSI-based methods,
the training set selection indeed could largely affect the training results and there was an
∙ 293 ∙
exclusiveness between image uniqueness and image frequency.

Acknowledgements
Thanks to the staff at the swine teaching and research center, School of Veterinary Medicine,
University of Pennsylvania for animal care and data collection. Funding for conducting this study
was provided in part by the new research initiatives at City University of Hong Kong.
References
Akçay, H. G., Kabasakal, B., Aksu, D., Demir, N., Öz, M., & Erdo\ugan, A, 2020. Automated bird
counting with deep learning for regional bird distribution mapping. Animals, 10(7), 1207.
Arteta, C., Lempitsky, V., & Zisserman, A, 2016. Counting in the wild. European Conference on
Computer Vision, 483–498.
Bochkovskiy, A., Wang, C.-Y., & Liao, H.-Y. M, 2020. Yolov4: Optimal speed and accuracy of object
detection. ArXiv Preprint ArXiv:2004.10934.
Ceballos, M. C., Góis, K. C. R., & Parsons, T. D, 2020. The opening of a hinged farrowing crate improves
lactating sows’ welfare. Applied Animal Behaviour Science, 230(November 2019), 105068. https://doi.org/
10.1016/j.applanim.2020.105068
Chen, C., Zhu, W., Steibel, J., Siegford, J., Han, J., & Norton, T, 2020. Recognition of feeding behaviour
of pigs and determination of feeding time of each pig by a video-based deep learning method. Computers and
Electronics in Agriculture, 176(June), 105642. https://doi.org/10.1016/j.compag.2020.105642
Fernandes, A. F. A., Dórea, J. R. R., & Rosa, G. J. de M, 2020. Image Analysis and Computer Vision
Applications in Animal Sciences: An Overview. Frontiers in Veterinary Science, 7(October), 1–18.
https://doi.org/10.3389/fvets.2020.551269
Hamming, R. W., 1950. Error detecting and error correcting codes. The Bell System Technical Journal,
29(2), 147–160.
Hong, S. J., Han, Y., Kim, S. Y., Lee, A. Y., & Kim, G, 2019. Application of deep-learning methods to
bird detection using unmanned aerial vehicle imagery. Sensors (Switzerland), 19(7), 1–16. https://doi.org/
10.3390/s19071651
Johnson, A. K., & Marchant-Forde, J. N, 2008. Welfare of Pigs in the Farrowing Environment. In The
Welfare of Pigs (pp. 141–188). Springer Netherlands. https://doi.org/10.1007/978-1-4020-8909-15
Lao, F., Brown-Brandl, T., Stinn, J. P., Liu, K., Teng, G., & Xin, H, 2016. Automatic recognition of
lactating sow behaviors through depth image processing. Computers and Electronics in Agriculture, 125, 56–
62. https://doi.org/10.1016/j.compag.2016.04.026
Liu, X., Zhang, Q., Luan, R., & Yu, F, 2013. Applications of perceptual hash algorithm in agriculture
images. Proceedings of the 2013 6th International Congress on Image and Signal Processing, CISP 2013,
2(Cisp), 698–702. https://doi.org/10.1109/CISP.2013.6745255
Moustsen, V. A., Hales, J., Lahrmann, H. P., Weber, P. M., & Hansen, C. F, 2013. Confinement of
lactating sows in crates for 4 days after farrowing reduces piglet mortality. Animal, 7(4), 648–654.
https://doi.org/10.1017/S1751731112002170
Ren, S., He, K., Girshick, R., & Sun, J, 2015. Faster R-CNN: Towards Real-Time Object Detection with
Region Proposal Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6), 1137–
1149. https://doi.org/10.1109/TPAMI.2016.2577031
Riekert, M., Klein, A., Adrion, F., Hoffmann, C., & Gallmann, E, 2020. Automatically detecting pig
position and posture by 2D camera imaging and deep learning. Computers and Electronics in Agriculture,
174, 105391.
Tian, M., Guo, H., Chen, H., Wang, Q., Long, C., & Ma, Y, 2019. Automated pig counting using deep
learning. Computers and Electronics in Agriculture, 163(September 2018), 104840. https://doi.org/10.1016/
j.compag.2019.05.049
Tzutalin, 2015. LabelImg. Git code. https://github.com/tzutalin/labelImg
Wah, C., Branson, S., Welinder, P., Perona, P., & Belongie, S, 2011. The caltech-ucsd birds-200-2011
dataset.
Yang, A., Huang, H., Zhu, X., Yang, X., Chen, P., Li, S., & Xue, Y, 2018. Automatic recognition of sow
nursing behaviour using deep learning-based segmentation and spatial and temporal features. Biosystems
Engineering, 175, 133–145. https://doi.org/10.1016/j.biosystemseng.2018.09.011
∙ 294 ∙
Monitoring Environmental and Behavioral Aspects on

Dairy Cattle Farms to Reduce Heat Stress
Daniela Lovarelli *, Marcella Guarino
Department of Environmental Science and Policy, University of Milan, via Celoria 2, 20133, Milan, Italy
* Corresponding author. Email: daniela.lovarelli@unimi.it
Abstract
For dairy cows, heat stress has been recognized as a big issue for decades, and recently this
problem is rising even more due to global warming and the hit of summer heat waves. Heat stress
strongly affects multiple aspects of cows’ lives, among which milk production, fertility, behavior
and welfare. In this study, 8 dairy cattle farms located in Northern Italy were monitored
considering the barn building, the barn microenvironment, external local weather and dairy cattle
behavior. The monitoring lasted for 1 year, by collecting continuously data for 1 week during the
three periods of thermoneutral, hot and cold seasons, in each farm. These data include
temperature, relative humidity and illuminance through environmental sensors; daily lying and
standing time, number and duration of daily lying bouts through accelerometers; and the
respiration rate through eye monitoring. The temperature-humidity index (THI) was calculated
both inside and outside of the barn. The external data were obtained by downloading temperature
and relative humidity of the local ground-based control units. Results show that cows reduced the
lying time and increased the standing time, the lying bouts and respiration rate as the increasing
THI in the barn. Therefore, both the environmental conditions were not optimal and the barn
structure and/or forced ventilation were not suitable to respond efficaciously to animals’ welfare.
From this survey, it was very useful to monitor continuously the barn conditions and animal
behavior with sensors in early detecting undesired environmental conditions.
Keywords: Animal behavior, dairy cattle, temperature-humidity index, lying time, barn structure.
1. Introduction
Heat stress is putting more and more pressure on dairy cattle living conditions. Hot weather
can be a problem at the latitudes, and heat strongly affects cows in terms of productivity, fertility,
health and welfare (Arcidiacono et al., 2017; Gernand et al., 2019; Tucker et al., 2008).
Furthermore, this is especially valid when high air temperature and relative humidity last long for
days (Armstrong, 1994). Therefore, it is very important to manage properly a livestock farm, by
paying attention to the barn building and to the solutions for protecting animals from heat, such
as shading systems and natural and forced ventilation (Firfiris et al., 2019; Schütz et al., 2009).
A direct evidence of their sufferance from heat can be responsed by focusing on the cows’
behavior, since the obvious effects such as a longer standing time and a lower feed intake were
observed (Lovarelli et al., 2020a; Mattachini et al., 2019; Porto et al., 2017). Consequently, to the
behavioral changes and the negative effects on fertility and health, system inefficiencies can
affect the overall livestock sustainability, including both the economic and environmental issues.
Thus, it is fundamental for the future of livestock to invest in solutions able to monitor
continuously animals and help farmers to make proper decisions (Berckmans and Guarino, 2017).
In this context, a first aspect to which to pay attention is the definition of suitable features to
build new barn structures or the structural interventions necessary to improve already existing
buildings. In fact, the barn is the first “fixed” rearing feature that needs to be adequate for animals
(CIGR, 2014).
In temperate climates, cows are reared in open barns, where air exchanges are essential.
Among the main options, natural ventilation supported by the thermal buoyancy and the lateral
openings are the most important solutions to adopt, for which the barn design and building are
fundamental aspects (Berman, 2019). When this is not sufficient, the forced ventilation is one of
the most adapt interventions to introduce (CIGR, 2014). In addition, shades introduced with
∙ 295 ∙
shading curtains or trees offer a great protection against solar radiation and heat stressing
conditions; another solution is the possibility of introducing air cooling systems or systems
watering the animals’ skin such as sprinklers, which are able to lower down the body temperature.
In Italy, most of the livestock farms locate in the Northern part of the country. About 70% of
the Italian dairy and meat cattle and 87% of the Italian swine can be found in this part of Italy
(ISTAT, 2020). Moreover, the composition of livestock farms is still quite variable, since both
big and innovative farms as well as small and traditional farms are present. Commonly, the biggest
and innovative farms, with a higher willingness to invest are located in plain areas, where,
however, important heat waves hit the country during summer.
To avoid or at least reduce the undesired negative effects caused by heat stress, farmers need
to invest in this direction and policy makers must support them. Technology, wireless
connectivity, sensors and monitoring tools, known as Precision Livestock Farming, are giving the
opportunity to farmers to get managerial advantages from this animal-by-animal control, enabling
the system to be effectively sustainable respect to the environmental, economic and welfare
aspects (Lovarelli et al., 2020b).
In this context, the aim of this study is to monitor a sample of dairy cattle farms in Northern
Italy to analyze their behavioral response to micro-environmental conditions, and to identify the
possible specific strategies to improve animal welfare, production and reproduction.
2.1. Monitored farms
The studied farms were located in the Lombardy region of Italy, where most of the livestock
farms are present. Eight farms were investigated, firstly observing the building structures, the
ventilation systems and the animals’ housing systems; secondly, animals were selected randomly
for the monitoring of their productivity and behavioral responses to the different living conditions.
The structural aspects were observed on each farm on the first day of monitoring, while the
animals’ parameters (e.g., milk production, behavior) were measured and processed. In particular,
10 randomly selected cows per farm and per period were analyzed for their behavior with the
accelerometers, therefore in every farm 30 random cows were monitored in the year. Regarding
milk production, it was an average data obtained from the monthly analyses that are commonly
performed by the local regional association of livestock producers. All these monitored cows were
Holstein Friesian that were housed in free-stall pens and loose-housing layout. Not only the
building and ventilation differed among farms, but also the herd management, cubicles, feeding
places and bedding materials. These data are summarized in Table 1. All farms had no external
paddock and no automatic milking system.
The monitoring lasted for 1 year, during which three weeks of monitoring were performed to
take into account the effect of seasonality. In particular, the studied periods were:
 Thermoneutral: 1 week of monitoring during a thermally neutral season (e.g.,
spring/autumn);
 Hot: 1 week of monitoring during the warm season (e.g., summer);
 Cold: 1 week of monitoring during the cold season (e.g., winter).
During these weeks, data were continuously collected (24 h–7 d) both for the environmental
monitoring and for the behavioral aspects of animals.
2.2. Environmental monitoring
For the environmental monitoring inside each barn, two sensors (HOBO U12, Onset
Computer Corporation, Bourne, MA, USA) were installed in 2 different positions at a 2-m height,
respectively. The methodology for the installation was carried out according to Mattachini et al.
(2013).
∙ 296 ∙
Table 1. Main farm characteristics.

Ridge Roof Cooling Cooling Milk
Barn Ridge Bedding
Farm height slope system: system: rest production
orientation opening -1 material
(m) (%) feed area area (kg d )
A EW 7 23 N.A. fan + fog fan 31 Straw
B NW-SE 12 28 Yes fan + fog N.A. 32 Straw
C NE-SW 8 15 N.A. N.A. fan 32 Mattress
fan +
D EW 7 11 Yes N.A. 41 Mattress
sprinkler
E NW-SE 7 13 Yes sprinkler fan 29 Mattress
F NW-SE 13 33 Yes N.A. fan 32 Straw
fan +
G NW-SE 5 13 Yes sprinkler 32 Sand
sprinkler
fan +
H EW 8 10 Insuff. fan 36 Mattress
sprinkler
Note: EW=East-West; NW-SE=North West – South East; NE-SW= North East – South West;
N.A.=Not Available; Insuff.=Insufficient.
These sensors were installed on day 1 and uninstalled on day 7 and then reinstalled on the day
1 of the following monitoring period. They were installed inside the barn to collect micro-
environmental data, including temperature, relative humidity and illuminance in order to make
possible the evaluations on the living conditions of dairy cows. Sensors recorded data every 30
min and then they were downloaded for the data processing. In particular, hourly data were
obtained for both temperature and relative humidity. With these data, the THI was calculated
using Eq.1 by ASABE (2006):
THI= T+0.36 ×T_dp+41.2 (1)
where, T dry bulb temperature (°C), Tdp dew point temperature (°C), THI Temperature-humidity
index.
In order to evaluate the environmental conditions outside the barn and analyze the differences
between inside and outside the barn, the climatic conditions external to the barn were studied by
downloading the data of the weather stations closest to every farm from the website of the
Regional Agency for the Protection of the Environment (ARPA Lombardia). These variables were
the same of the internal ones: air temperature (T; °C) and relative humidity (RH; %). Similarly to
the internal micro-environmental conditions, they were averaged on a hourly basis and used to
calculate the THI outside the barn, using the equation 1.
2.3. Animals’ behavior monitoring
In order to monitor the behavior of the selected animals, the accelerometers HOBO Pendant
G Data Loggers (Onset Computer Corporation, Pocasset, MA, US) were installed on the hind leg
of 10 dairy cows using tape and plastic tough leg bands to avoid undesired changes of position.
Similarly, these sensors were positioned on day 1 and uninstalled on day 7 during each monitoring
period. The sensors recorded continuously cows’ activities by processing the data on the leg
orientation, which resulted from the 3 axis measurements of the accelerometers. The focus was
paid on the lying (or standing) activity, including standing and lying time per day (h d-1),
frequency of lying bouts (n bout d-1) and duration of the lying bouts (min bout-1) per cow and per
day. The devices recorded data at 1-min intervals and then were processed to get the daily average.
In each barn, 10 lying-down cows were randomly selected per farm to measure the respiration
rate (r.r; n. breaths min-1) during every monitoring period. This assessment was carried out
through eye-monitoring, by measuring the number of breaths for 1 min; this operation was done
twice per monitoring period (i.e., on day 1 and day 7).
∙ 297 ∙

Regarding the collected data and subsequent analysis, the raw dataset included about 65000
data on environmental parameters and 8600 data on behavioral ones. From these, a second dataset
was calculated with the daily averages for environmental and behavioral data, which included 146
data per parameter. The analyses were carried out using MS Office Excel 2016 and MS Office
Access 2016.
3.1 Environmental aspects
Table 2 shows the average values of the environmental parameters monitored with the
environmental sensors in the barn (microenvironment) and the data collected from the
meteorological stations (outside the barn). The collected hourly data were then averaged for the
full monitoring period.
Table 2. Mean and standard deviation values of temperature (T; °C), relative humidity (RH; %)
and THI, in the barn and outside the barn of the 8 farms.
Microenvironment Meteorological station
Period
T (°C) RH (%) THI T (°C) RH (%) THI
T 19.0 (3.1) 66.7 (13) 64.6 (4.0) 17.8 (4.2) 76.8 (20) 62.5 (5.6)
H 26.3 (3.0) 68.9 (10) 75.5 (3.4) 25.9 (4.2) 69.4 (17) 74.5 (4.5)
C 6.7 (3.8) 69.8 (13) 46.1 (6.1) 4.1 (4.0) 81.6 (21) 40.8 (7.8)
Note: T = thermoneutral period; H = hot period; C = cold period.
Although the farms are geographically close to each other, there were some differences in the
environmental parameters among them. Focusing on the meteorological data outside the barns,
differences among the 3 periods are quite similar, with RH having the widest variability,
especially in thermoneutral and cold periods. For what regards the micro-environmental aspects,
RH highlights some differences, and in particular the THI in the barn, which shows an average
value close to the welfare threshold of 72.
The temperature is always higher in the barn, while RH is higher outside. Except for the cold
period, THI is higher inside the barn than outside during another 2 periods (thermoneutral and
hot).
Since the THI is the most important parameter for cows’ welfare inside the barn, some classes
were built to analyze in detail its trend. Since THI starts presenting undesired conditions from a
value equal to about 69 and presents a threshold for undesired condition from the value of 72, the
first class included all data below the value of 69, and then smaller THI classes were considered..
These classes include: (a) THI<=69 (comfort zone), (b) THI between 69 and 72 (low comfort
zone), (c) THI between 72 and 75 (heat stress alert zone), (d) THI between 75 and 78 (heat stress
zone) and (e) THI>78 (severe heat stress zone). The share of data in each of these classes per farm
and per period is reported in Figure 1.
From this figure emerges clearly that the hot period has long periods with THI values much
higher than 72, and that these periods should be avoided, at the expenses of animals’ health,
welfare and productivity. Farms A and B highlight as predominant class the most unwanted one,
the one of THI>78, followed by farms G and H.
3.2 Behavioral aspects
Table 3 reports mean and standard deviation values of the monitored cows’ behaviors,
including the lying and standing time as well as the number of lying daily bouts per day and their
duration. These results are achieved from the processing of accelerometer data, from which first
daily averages and then full monitored periods averages were achieved.
From these values, it emerges that the average lying time of the 8 farms is affected by
seasonality. The lowest average lying time is observed in the hot period (10 h d-1), while the
longest in the cold one (12.0 h d-1). The highest number of bouts was observed in the thermoneutral
∙ 298 ∙
period, while the lowest number, with the longest duration, was observed in the cold period. In
the hot period, the average duration of bouts was the shortest (65.4 min).
THI<=69 69<THI<=72 72<THI<=75 75<THI<=78 THI>78
100%
90%
80%
70%
Share per THI class
60%
50%
40%
30%
20%
10%
0%
A B C D E F G H A B C D E F G H A B C D E F G H
Cold Hot Thermoneutral
Figure 1. Share of the different THI classes per farm and per monitoring period.
Table 3. Mean and standard deviation values of the 8 farms for the behavioral results of lying
and standing time (h d-1) and of the number (n. d-1) and duration (min bout-1) of lying bouts.
Behavior
Period
Standing (h d-1) Lying (h d-1) N. bouts (n. d-1) Duration bouts (min)
T 13.4 (1.5) 10.6 (1.5) 10.7 (3.3) 70.0 (22.1)
H 14.0 (1.4) 10.0 (1.4) 10.4 (2.7) 65.4 (13.0)
C 12.0 (1.1) 12.0 (1.1) 10.1 (1.8) 81.4 (14.2)
As expected, with hot air temperature and high relative humidity in hot period, cows stand
more than during temperate-cold periods, indicating environmental conditions in the cold period
are more adapt to their welfare. This aspect emerges also from the analysis of lying bouts. Hence,
animals rest more during the cold season. Figure 2 shows the reduction in lying time with the
increasing THI. Figure 3 shows the average daily distinction in standing and lying time per farm.
14.0
12.0
Lying time, h d-1
10.0
8.0
6.0
< 48 48-60 60-66 66-72 72-78 >78
THI class
Figure 2. Average daily lying time at the different classes of THI internal to the barn.
∙ 299 ∙
T - Lying time T - Standing time H - Lying time

H - Standing time C - Lying time C - Standing time
100%
90%
Share of average daily time (%)
80%
70%
60%
50%
40%
30%
20%
10%
0%
A B C D E F G H
Figure 3. Share of the average daily lying and standing time (%) per farm in the three monitored
periods. T=thermoneutral period; H=hot period; C = cold period.
Concerning the respiration rate, Table 4 shows the respiration rate (r.r.) and the internal THI
in the barn, highlighting an increased number of breaths per minute when hotter conditions are
present.
Table 4. Mean and standard deviation values of the respiration rate (r.r.; n. breaths min-1) and
THI in the barn per period.
T H C
Farm
r.r. THI r.r. THI r.r. THI
A 54.6 (5.9) 68.9 79.7 (5.5) 76.4 30 (0.3) 46.1
B 49 (0.6) 65 75.4 (1.1) 77.3 30.6 (2.8) 40.7
C 51.2 (8.2) 63.1 56.5 (2.4) 66.5 33.7 (2.7) 52.2
D 59 (2.8) 67.5 50.6 (2) 64.4 31.5 (3.5) 48.5
E 48 (0.1) 61.8 51.5 (4.9) 66.5 34.1 (0.1) 50.4
F 53.8 (3.3) 66.4 56.3 (3.3) 69.5 37.1 (2.1) 55.7
G 51.5 (5.5) 62.3 54.5 (1.8) 77 32.6 (1.4) 54.2
H 42.1 (2.7) 55.4 62.1 (5.2) 73.1 34.9 (0.6) 53.6
4. Conclusions
In this study, all farms showed difficulties in maintaining adequate environmental conditions
inside the barn to meet the welfare requirements of dairy cows, most of all during the hot seasons.
From this analysis emerged that 2 of the farms (i.e., A and B) showed the most undesired
microenvironments, and here and in a third farm (i.e., E), the behavior of cows was influenced.
Both Farms A and B have structural issues that need further study (i.e., lack of ridge opening,
insufficient roof height or inclination and insufficient forced ventilation).
The problem of heat stressing conditions is becoming very important under the climate crisis
that has been hitting at mid latitudes. It is very useful to monitor micro-environmental conditions
and the animal behavior to assess the living conditions of dairy cattle. This study highlights that
the wide differences can exhibit among livestock farms, even in a reduced geographic area. In
addition, the relationship between THI and the environmental and behavioral characteristics is
closely related, therefore if barns do not guarantee appropriate environmental conditions,
immediate effects emerge on the behavior of reared animals.
∙ 300 ∙
References
ASABE, 2006. Design of Ventilation Systems for Poultry and Livestock Shelters. In Standards 2006
American Society of Agricultural and Biological Engineers 53rd Edition St. Joseph, Michigan pp. 652-670.
Armstrong, D.V, 1994. Heat Stress Interaction with Shade and Cooling. J Dairy Sci. 77, 2044–2050.
https://doi.org/10.3168/jds.S0022-0302(94)77149-6
Arcidiacono, C., Porto, S.M.C., Mancino, M., Cascone, G., 2017. Development of a treshold-based
classifier for real-time recognition of cow feeding and standing behavioural activities from accelerometer data.
Computers and Electronics in Agriculture, 134, 124–134. http://doi.org/10.1016/j.compag.2017.01.021
ARPA Lombardia, Regional Agency for the Protection of the Environment. Available online:
https://www.arpalombardia.it/Pages/ARPA_Home_Page.aspx. (Accessed: 18 May 2021)
Berckmans, D., Guarino, M., 2017. From the Editors: Precision livestock farming for the global livestock
sector. Animal Frontiers, 7, 4–5. https://doi.org/10.2527/af.2017.0101
Berman, A., 2019. An overview of heat stress relief with global warming in perspective. International
Journal of Biometeorology, 63(4), 493–498. https://doi.org/10.1007/s00484-019-01680-7
CIGR, 2014. The design of dairy cow and replacement heifer housing. Report of the CIGR Section II
Working Group n. 14 – cattle housing. ISBN 978-2-9552352-0-1.
Firfiris, V.K., Martzopoulou, A.G., Kotsopoulos, T.A., 2019. Passive cooling systems in livestock
buildings towards energy saving: A critical review. Energ Buildings. 202, 109368. https://doi.org/10.1016/
j.enbuild.2019.109368
Gernand, E., König, S., Kipp, C., 2019. Influence of on-farm measurements for heat stress indicators on
dairy cow productivity, female fertility, and health. J Dairy Sci. 102(7), 6660–6671. https://doi.org/10.3168/
jds.2018-16011
ISTAT. National Institute of Statistics. Available online: http://dati.istat.it/ (Accessed: 18 May 2021)
Lovarelli, D., Finzi, A., Mattachini, G., Riva, E., 2020a. A Survey of Dairy Cattle Behavior in Different
Barns in Northern Italy. Animals, 10, 713.
Lovarelli, D., Bacenetti, J., Guarino, M., 2020b. A review on dairy cattle farming: Is precision livestock
farming the compromise for an environmental, economic and social sustainable production? J Clean Prod.
262, 121409. https://doi.org/10.1016/j.jclepro.2020.121409
Mattachini, G., Pompe, J., Finzi, A., Tullo, E., Riva, E., Provolo, G., 2019. Effects of feeding frequency
on the lying behavior of dairy cows in a loose housing with automatic feeding and milking system. Animals,
9(4), 121. https://doi.org/10.3390/ani9040121
Porto, S. M. C., D’Emilio, A., Cascone, G., 2017. On the influence of the alternation of two different
cooling systems on dairy cow daily activities. J Agr Eng. 48(1), 21–27. https://doi.org/10.4081/jae.2017.577
Schütz, K.E., Rogers, A.R., Cox, N.R., Tucker, C.B., 2009. Dairy cows prefer shade that offers greater
protection against solar radiation in summer: Shade use, behaviour, and body temperature. Appl Anim Behav
Sci. 116(1), 28–34. https://doi.org/10.1016/j.applanim.2008.07.00 5
Tucker, C.B., Rogers, A.R., Schütz, K.E., 2008. Effect of solar radiation on dairy cattle behaviour, use of
shade and body temperature in a pasture-based system. Appl Anim Behav Sci. 109(2–4), 141–154.
https://doi.org/10.1016/j.applanim.2007.03.015
∙ 301 ∙
Development of a Point Cloud Acquisition System and 3D Point

Cloud Reconstruction Method for Beef Cattle
Jiawei Li a,b, Qifeng Li a, Weihong Ma a,*, Xianglong Xue a, Luyu Ding a, Ligen Yu a,
Ronghua Gao a
a
Beijing Research Center for Information Technology in Agriculture, Beijing, China
b
College of Information and Electrical Engineering, China Agricultural University, Beijing, China
* Corresponding author. Email: mawh@nercita.org.cn
Abstract
Body size, weight, and body condition score are key indicators for monitoring cattle growth;
these parameters can also be utilized to predict the beef cattle yield and evaluate the economic
traits. However, it puts intense stress on cattle when the livestock body measurements are carried
out manually, giving rise to negative effects on their feeding and weight-gain. To resolve this
problem, point cloud preprocessing, registration and 3D reconstruction algorithms were
developed based on an automatically triggered five-angle beef cattle point cloud instantaneous
acquisition system. The algorithm realizes the point cloud filtering, registration, segmentation,
down-sampling and the point cloud 3D reconstruction of the global point cloud of cattle and target
recognition. The uncertainty of the point cloud reconstruction in this work is 20 mm, and the
acquisition time is within 0.08 s. As a result, non-constrained, non-contact, and stress-free real-
time collection of the cattle 3D point cloud is clearly demonstrated. This system provides technical
support for automatic extraction of key features during livestock body measurements.
Keywords: Unconstrained-collection, registration, 3D-reconstruction, down-sample, super-4pcs
1. Introduction
As one of the main protein sources for human-beings, beef supply is in high demand
(Emanuela et al., 2019), which stimulates the demand for precise breeding in turn (Charlotte et
al., 2020). In fact, the precise breeding relies on the successful detection of livestock body
parameters, including cattle's body shape, weight, and body condition scores (Yongliang et al.,
2021). The most commonly used method to collect body information is manual measurement
(Dingwell et al., 2006). However, tremendous pressure will be put on the cattle if they are driven
and confined to a closed space to do such measurements; this will also hampers the weight gain
process (Augspurger and Ellis, 2002). Recent studies show that cattle point cloud data and depth
images can be used for segmentation and extraction of body parts (Nan et al., 2021), Cominotte
developed an automatic algorithm to extract measurement points such as the back and buttocks
for beef cattle (Cominotte et al., 2020).
Local point cloud cannot restore the precise shape characteristics of animals, 3D
reconstruction for animals is a better solution. Cozler designed a scanning system that scans the
cattle body through laser stripes to achieve high-precision 3D reconstruction. The system has high
accuracy, but requires cattle to keep completely still within 5 s of the capturing process (Cozler et
al., 2019). Alexey designed a non-rigid algorithm to generate accurate 3D models of cattle based
on the point cloud collected by the Kinect v1(Alexey et al., 2020). Different from the 3D point
cloud reconstruction for plants, the point cloud collection needs to be conducted synchronously.
It is apparent that large body scale of cattle and the complex environment in farm create challenges
to data collection. To cope with these issues, a system for cattle point cloud instantaneous
collection with a gantry structure was designed and a registration method for cattle was developed
to achieve non-contact alignment and stress-free 3D model reconstruction.
2.1. Design of the cattle point cloud collection system
A gantry-type cattle point cloud collection device was designed based on the cattle transfer
∙ 302 ∙
corridor in a farm. The device consists of a computer, an RFID identifier, a set of through-beam
grating sensors, and five Kinect DK cameras (denoted as cameras A-E in Figure 1). The device
shows a height of 2.30 m and a width of 2.32 m. Four depth cameras are evenly deployed on both
sides and the last one is installed at the middle of the top beam. The RFID reader is set at the
midpoint of the gantry beam. The grating trigger sensor is set on both sides of the gantry, and all
devices are connected to an industrial integrated computer through USB3.0 cable or 485 data
cables, as shown in Figure 1.
Figure 1. Cattle point cloud collection device.

The computer runs with a Windows10 system version Core-I5 processor and 4GB RAM.
Kinect DK sensor collects depth images based on Laser Time of Flying (TOF). Kinect DK
equipped with a depth camera with 120°*120° width can collect depth image data in the range of
0.25-2.88 m; uncertainty of the measured distance is 17 mm, and the single-frame exposure time
is 12.8 ms. The infrared grating sensor is composed of four groups of beam sensing points. When
three groups or more beam points are blocked, pulse information is triggered. The response speed
is within 10 ms. The device can accurately identify the moment when the cattle enters.
2.2. Experimental environment and acquisition logic
The experiment was carried out in the transfer corridor using the natural light as the lighting
source in Zhuochen Ranch, Beijing in March, 2021. Five cattle with RFID tags were selected,
with the weight range of 312-534 kg. The floor of the transfer corridor is concrete, with a slope
of about 15°, and cannot exclude the interference of feces and dust. The same trial on the 1:2 beef
cattle model was also conducted in the laboratory.
When cattle pass directly under the device, the through-beam grating sensor and the RFID
reader are driven to read the cattle ear tag number. Once the required information is obtained, five
Kinect DK cameras take turns to collect cattle depth image data, followed by a computational
conversion into the point cloud data through the camera internal parameters based on formula (1).
1  (1)
f0 0 
x   x   x 
 y  D  0 1  
0  y 
   fy
 z   1
0  
0 1
 
 

where x , y are point cloud coordinates, x , y are depth image coordinates, D is the value of depth,
f x , f y are internal parameters of DK cameras.
2.3. Point cloud filtering and preprocessing
As the accuracy of the registration can be influenced by dust and debris, these interference
factors were filtered using radius filtering based on the Guess local distance distribution. The 3D
∙ 303 ∙
Gaussian distribution formula can be expressed as the formula (2).

(2)
1 1  1 
N(x∣u, A)  exp   (x  u )T A1 (x  u ) 
 2 
3 1
(2 ) | A |
2 2
where X is 3D vector, u is the mean of the vector, and A is the covariance matrix of all vectors.
Each point in the cattle point cloud was recorded as a 3D vector, the neighborhood of each
point is analyzed, and the average distance of all neighbor points was calculated. By calculating
the Gaussian distribution, points outside the sensitive neighborhood, regarded as outliers,were
filtered. The filtering process is shown in Figure 2.
Figure 2. Comparison before and after point cloud filtering.

2.4. Cattle and ground point cloud extraction
An environmental filtering algorithm was built based on the regional growth Kdtree
neighborhood clustering (Meijer et al., 2020) to reduce other interferences in the beef cattle
passage environment (such as passage railings, ground manure, sensor brackets, etc.). The
processing process is shown in Figure 3. The blue part in the figure is the interference.
Figure 3. Environmental point cloud filtering.

Due to the close contact with cattle, the ground cannot be removed by the neighborhood
clustering method. In this research, Random Sample Consensus (RANSAC) was selected for
plane extraction and filtering (Fatemeh, 2019). Extraction results are shown in Figure 4. In
particular, the vector that fits the ground is thereafter denoted as the vector M.
Figure 4. Ground point cloud extraction.

2.5. Multi-angle cattle point cloud registration
The point cloud data collected by DK cameras with different angles has a small repetitive part
and cannot be directly completed by conventional registration methods such as ICP. In order to
realize the 3D reconstruction of cattle based on five different perspectives point clouds (collected
by cameras (A, B, C, D, and E)), in this research, an algorithm based on the super-4PCS (Martín
et al., 2018)was developed. Cattle point cloud registration is divided into two steps:
(1) Point cloud registration and fusion on the same side
The point cloud collected by the depth camera on the same side of cattle usually showed a
∙ 304 ∙
large overlap. The registration was carried out by the Super-4PCS algorithm (Martín et al., 2018).
The affine invariance of the point pair was used to achieve registration. Here, the point cloud
collected by camera E is the source point cloud (recorded as S), and the target point cloud collected
by camera A is recorded as W. 4 points are randomly selected from S based on the principle of
maximizing distance to ensure that the distance between points is large but does not exceed the
threshold; these four points were denoted as:
B  Sa , Sb , Sc , Sd (3)
Assuming the line of Sa , Sb and Sc , Sd intersect at point e, then the define two independent
ratios and the distance between two points are defined as:
sa  se sc  se
r1  r2  (4)
sa  sb sc  sd
d1  sa  sb d2  sc  sd (5)
Then the point set P and Q were selected in the target point cloud W
P t , t ∣t , t
i j i j  W, si  s j [ d1   , d1   }
(6)

Q   ti , t j ∣
 ti , t j  W, si  s j [ d2  , d2   } (7)
If each point ti in W was taken as the center of the sphere, drawing the sphere with d 2 and
d1 as the radius respectively, ti and the points falling in the range of ( d1 -μ, d1 +μ) satisfy the
P set condition, and ti falling in ( d 2 -μ) , d 2 +μ) points in the range satisfy the Q set condition.
For each pair of points ( ti , t j ) in P, the e1 from r1 can be obtained based on formula (8)
e1  ti  r1  t j  ti  (8)
For each pair of points ( ti , t j ) in Q, the e2 from r2 can be obtained based on formula (9)
e2  ti  r2  t j  ti  (9)
When e1 and e 2 are approximately equal to point pairs and the angles are approximately
equivalent, transformation matrix T is calculated between Sa , Sb , Sc , Sd and ta , tb , tc , td .
T is used to transform the point cloud S into R, and calculate the registration evaluation coefficient
Pr (the proportion of the points whose distance between the transformed point cloud R and the
target point cloud W is less than the threshold value).
The above process was repeated several times, until the registration coefficient was greater
than 0.7, and then the registration was stopped and the matrix T was saved. The logic pseudo code
is shown in Table 1.
(2) Registration of the point cloud on both sides with the vertical point cloud
The overlap between the vertical-point cloud and side-point cloud, and the registration
uncertainty of direct measurement by 4-PCS was relatively large. The ground-plane point cloud
was treated as the z-axis origin. The normal vector M was in z-axis positive direction. The cutting
plane perpendicular to the normal vector at z = 0.6n was denoted as U. All the point clouds
between the cut surface of the point cloud and the ground were filtered, and the remaining part of
the point cloud was used as the basis for registration. Finally, super 4-PCS in the remaining area
was utilized to realize the registration.
∙ 305 ∙
Table 1. Pseudo code of trigger logic for point cloud collection equipment.
Algorithm 1 Cattle point cloud registration logic
Definition
S: Source point cloud W: Target point cloud
Pr：The ratio of the distance between transformed and target point cloud <12
θ: Angle between two sets of points in H ξ: Angle between two sets of points in D
T: Transformed matrix
Input: S, W
Process
1 set xi∈S，xj∈W，Pr=0
2 while (Pr<0.7) do
3 select any 4 non-collinear xi as H, i= (0,1,2,3)
4 calculate R, r2, d1, d2 base on formula (4), (5)
5 for xj in W:
6 generate group P\Q from n based on formula (6), (7)
7 for (ti, tj) in P:
8 calculate e1 based on formula (8)
9 for ta, tb in Q:
10 calculate e2 based on formula (9)
11 if (e1-e2)<μ
12 find （ti, tj ,ta, tb） as D
13 if ((θ-ξ) <0.04(θ+ξ))
14 calculate transform matrix T between (ti, tj ,ta, tb) and (x0, x1,x2,x3)
15 Transform all the points in S to R with T
16 Calculate Pr between R and W
17 Return Pr
Output: T
Based on the methods above, 6 rounds of point cloud reconstruction experiments were
conducted on the actual breeding environment and model cattle respectively, with the aim of
testing the accuracy and speed of the 3D reconstruction.
3.1. Point cloud reconstruction
One result was selected as an example to show the result of point cloud reconstruction. The
point cloud collected by the two cameras on the same side of the gantry were registered as shown
in Figure 5. In each figure, the two pictures on the left are point clouds collected by cameras from
different angles, and the middle and right are the point clouds after registration.
Figure 5. Point cloud registration on the left.

After the point cloud registration on the same side was completed, we registered the point
cloud collected by the camera on the top of the gantry with the point cloud after registration on
both sides. The final registration result is shown in Figure 6.
∙ 306 ∙
Figure 6. Point cloud registration result.

In order to ensure uniform point cloud density and facilitate subsequent processing, octree
was used to realize the point cloud downsampling uniformly, and the result is shown in figure 7.
Figure 7. Point cloud downsampling results.

3.2. Acquisition speed test
The device was used in the laboratory and a breeding house to test the speed of point cloud
collection and synthesis. Each column in Table 2 represents the time spent in each stage of an
independent collection experiment.
Table 2. System point cloud collection time.
T1/s T2/s T3/s T4/s T5/s T6/s T7/s
Laboratory experiment
1 0.09 4.23 4.09 82.53 4.93 95.87 13.34
2 0.08 3.72 3.87 85.95 5.90 99.52 13.57
3 0.08 3.72 3.54 94.71 4.03 106.08 11.37
4 0.08 3.01 5.85 83.38 5.83 98.15 14.77
5 0.08 3.15 4.60 93.96 4.02 105.81 11.85
6 0.07 3.43 3.13 81.96 5.28 93.87 11.91
Ranch experiment (a breeding house)
1 0.08 2.03 5.36 92.10 5.98 105.55 13.45
2 0.08 2.54 4.83 88.09 6.07 101.61 13.52
3 0.07 2.68 6.42 103.63 6.59 119.39 15.77
4 0.07 2.62 6.46 94.83 7.04 111.02 16.19
5 0.08 2.48 5.97 87.54 5.13 101.20 13.67
6 0.08 3.38 4.59 119.86 7.55 135.47 15.61
T1: Depth image acquisition time; T2: Point cloud generation and storage time; T3: point cloud
pre-processing time; T4: point cloud registration time; T5: 3D reconstruction time; T6: Total time;
T7: Total time except registration.
The collection time revealed in Table 2 represents the efficiency when collecting the cow
point cloud in motion. In the processing, the most time-consuming part was the point cloud
registration process, which accounted for 86.99 % of the total time on average. In order to improve
the efficiency of point cloud collection, the registration algorithm was only executed when the
device was successfully installed for the first time, and then the obtained registration matrix was
directly used for calculations. On this condition, point cloud reconstruction time took about 11 -
16 s each time. And T3-T5 were put into the background to run through multi-threading, and the
point cloud collection thread only retained the T1/T2 part, which took about 2 -4 s.
∙ 307 ∙
3.3. Point cloud accuracy test

The reconstructed cattle point cloud was manually marked to calculate the cattle's body height,
body width, and the calculated data were compared with the corresponding measured values. In
particular, there are inevitable errors in the results when measuring live cattle with a tape or a
measuring stick. In order to evaluate the theoretical error effectively, this project only selected the
beef cattle model in the laboratory as the object.
Table 3. Point cloud accuracy evaluation results.
Body height Body height Body width Body width average
(mm) uncertainty (mm) uncertainty uncertainty
1 912 0 469 0 0
2 898 0.015 476 0.015 0.015
3 915 0.003 472 0.006 0.0045
4 928 0.018 481 0.026 0.022
5 916 0.004 470 0.002 0.003
6 902 0.01 454 0.032 0.021
7 906 0.007 467 0.004 0.0055
* 1 for manual measurement values; 2-7 for calculated values in different rounds.
As shown in table 3, the present error is no more than 3.2 % in the results of measurement by
point cloud data, compared with manual measurement with a tape, and the absolute error is within
20 mm, which can meet the actual application requirements.
3.4. Disturbance caused by sunlight
The results of the collection speed test (Figure 8) shows that the time to collect the point cloud
of cattle in the laboratory was slightly longer than that in the ranch. This unexpected result was
carefully checked and it was found that the density and quality of the point cloud collected in
ranch was not as good as in the laboratory. Therefore, the same beef cattle model was placed in
the different environments (indoors and outdoors) for point cloud collection. By counting the
number of points, the impact of light on point cloud collection can be evaluated.
Figure 8. Light interference experiment results.

Point cloud a-h are collected in the environment as follows:
a: indoors without sunlight b: outdoor under the shelter of the building
c: outdoor under the cover of board outdoor d: outdoor under the cover of thick cardboard
e: outdoor direct sunlight only use cardboard to cover the breasts
f: outdoor direct sunlight only use cardboard to cover the belly
g\h: outdoor exposed in sunlight without obstructed.
From the result above, it is known that when exposed to strong light, the reflection of light
from the leather is messy. Due to the limitation of the principle of the TOF camera, the point cloud
appears incomplete. When used outdoors in strong sunlight, an effective shading shed must be
installed to collect effective beef cattle point clouds. However, it is inconvenient to deploy the
canopy on the scene and extra stress may be put on the cattle, additional efforts are still required.
∙ 308 ∙
4. Conclusions
We established a point cloud instantaneous collection gantry system, enabling livestock body
measurements for active cattle in the ranch. This system can realize cattle entry perception, ear
tag reading, and instantaneous collection of multi-view point clouds without stress and
interference on cattle. A set of beef cattle point cloud collection, preprocessing, registration,
thinning and 3D reconstruction algorithms are developed based on the gantry systems.
The algorithm can collect point cloud data from five different perspectives in 0.09 s. The
process of point cloud reconstruction of cattle 3D model needs another 10–15s. The uncertainty
limit of cattle point cloud data does not exceed 0.02 m, and the overall error is between 0.5–2 %,
which shows that the reconstructed cattle point cloud has restored the true shape of the cattle. 3D
cattle point cloud can provide important support for weight evaluation, body size calculation, body
condition score calculation, and body feature analysis. The system provides a possibility to
comprehensively and conveniently monitor growth and improve animal welfare.
Acknowledgements
This thesis was supported by Key Technologies Research and Development Program (CN),
funding number, 2018YFE0108500; Beijing Science and Technology Planning, funding number
Z191100004019007.
References
Emanuela, T., F. Alberto, G. Marcella, 2019. Review: Environmental impact of livestock farming and
Precision Livestock Farming as a mitigation strategy. Science of The Total Environment. 650, 2751–2760.
https://doi.org/10.1016/j.scitotenv.2018.10.018.
Charlotte, G., B. Ludovic, D. Jean-Yves, 2020. Improvement of feed and nutrient efficiency in pig
production through precision feeding. Animal Feed Science and Technology. 268, 114611. https://doi.org/
10.1016/j.anifeedsci.2020.114611.
Yongliang, Q., K. He, C. Cameron, L. Sabrina, S. Daobilige, E. Stuart, S. Salah, 2021. Intelligent
perception for cattle monitoring: A review for cattle identification, body condition score evaluation, and
weight estimation. Computers and Electronics in Agriculture. 185, 106143.
Dingwell, R.T., M.M. Wallace, C.J. McLaren, C.F. Leslie, K.E. Leslie, 2006. An Evaluation of Two
Indirect Methods of Estimating Body Weight in Holstein Calves and Heifers. Journal of Dairy Science. 89
(10), 3992–3998.
Augspurger, N.R., M. Ellis, 2002. Weighing affects short-term feeding patterns of growing-finishing pigs.
Canadian Journal of Animal Science. 82 (3), 445–448.
Nan, J., K. Gert, K. Peter Groot, S. Zhengxiang, D. Songhuai, 2021. Segmentation of body parts of cows
in RGB-depth images based on template matching. Computers and Electronics in Agriculture. 180, 105897.
https://doi.org/10.1016/j.compag.2020.105897.
Cominotte, A., A.F.A. Fernandes, J.R.R. Dorea, G.J.M. Rosa, M.M. Ladeira, E.H.C.B, G.L. Pereira, W.A.
Baldassini, O.R, 2020. Automated computer vision system to predict body weight and average daily gain in
beef cattle during growing and finishing phases. Livestock Science. 232, 103904. http://doi.org/10.1016/
j.livsci.2019.103904.
Y, C. Allain, A. Caillot, J.M. Delouard, L. Delattre, T. Luginbuhl, P. Faverdin, 2019. High-precision
scanning system for complete 3D cow body shape imaging and analysis of morphological traits. Computers
and Electronics in Agriculture. 157, 447–453. https://doi.org/10.1016/j.compag.2019.01.019.
Alexey, R., K. Vitaly, D. Konstantin, K. Vladimir, M. Sergei, 2020. Accurate body measurement of live
cattle using three depth cameras and non-rigid 3-D shape recovery. Computers and Electronics in Agriculture.
179, 105821.
Meijer, C., M.W. Grootes, Z. Koma, Y. Dzigan, R. Gonçalves, B. Andela, G, E. Ranguelova, N. Renaud,
W.D. Kissling, 2020. Laserchicken—A tool for distributed feature calculation from massive LiDAR point
cloud datasets. SoftwareX. 12, 100626.
Fatemeh, H.-L., 2019. Structural-based point cloud segmentation of highly ornate building façades for
computational modelling. Automation in Construction. 108, 102892.
Martín, B., B. Frédéric, G.-J. Higinio, M.-S. Joaquín, A. Pedro, 2018. 4-Plane congruent sets for
automatic registration of as-is 3D point clouds with 3D BIM models. Automation in Construction. 89, 120–
134. https://doi.org/10.1016/j.autcon.2018.01.014.
∙ 309 ∙
A Method for Pig Individual Identification and

Its Body Parts Recognition
Mengru Wu a, Qiuju Xie a,*, Xin Li b, Muyu Yang a, Jun Bao b, Honggui Liu b
a
b
*Corresponding author. Email: xqj197610@163.com
Abstract
Ear-tag and the rectal temperature measurement are most commonly used for pig individual
identification and pig body temperature measurement, they are invasive methods that have some
shortcomings of being easy to drop off, being easy to contamination and exerting stress on pigs.
To solve these problems, a non-invasive method based on machine vision for pig identification
and body temperature measurement was proposed in this paper. The method was developed for
pig face and body parts recognitions based on YOLOv4 to achieve and assist pig individual
identification and body surface temperature measurement. Data were collected in a pig farm with
front and top view videos from 6 pigs. A total of 5928 images was extracted with an interval of
40 frame from the videos for data training (4742 images), verifying (593 images) and testing (593
images). The results showed that, compared with other models, the model developed on YOLOv4
had optimal performances. The averaged precision and recall rate of the model for pig individual
identification was 98.13% and 99.57%, respectively; and the mean average precision (mAP) for
body parts recognition was 99.16%, when the threshold of the intersection over union (IOU) was
set 0.9. Therefore, the YOLOv4 model is an optimal model for pig individual identification and
body parts recognition. It also provides a basis for automatic non-invasive methods for pig
identification and body temperature detection.
Keywords: Individual identification, body parts, deep learning, YOLOv4, target detection.
1. Introduction
In intensive pig farming, automatic pig individual identification and body temperature
detection are critical to realize precision breeding and disease prediction (Gao, 2010; Menzel et
al., 2014). Traditional pig identification methods mainly include wearing electrical ear tags, pig
body color markings, etc. (Wang, 2020). Also, rectal temperature is most commonly used for body
temperature measurement (Alsaaod et al., 2014). However, these invasive measurements always
cause some stress on pigs.
With the development of machine vision in recent years, people began to explore low-cost,
non-contact technologies for pig individual identification (Marsot et al., 2020) and body surface
temperature measurement based on images or videos (Amezcua et al., 2014; Lu et al., 2018;
Zhang, 2019). Guo (2016) used the Isomap algorithm to fuse the extracted individual color,
texture and shape features of the pig, and used a vector machine to complete the classification,
but the individual recognition rate was low. Wada et al. (2013) used the feature space to crop the
pig face image and segment the important parts manually, the accuracy of the individual
identification of 6 pigs reached 97.9%. However, it is difficult to achieve an automatic pig face
feature extraction. Marsot et al. (2020) proposed an adaptive pig face recognition method and
established the class activation heat map (gradCAM). It proved that pig classification using neural
network was mainly achieved by its facial feature extraction. It has a good research prospect and
promotion space to realize pig individual recognition by extracting pig facial features.
Infrared thermal imaging technology has some advantages of non-contact, effective, etc. It is
gradually used for detections on body temperature and physiological diseases of livestock and
poultry (Bagavathiappan et al., 2013; Soerensen and Pedersen, 2015). Studies showed that the
∙ 310 ∙
rectal temperature of pigs has certain close relationship to the temperature of the ear root region
in the thermal infrared images (Siewert et al., 2014; Lu et al., 2018). It caused much attention on
most scholars and researchers in this field. However, there are still some problems need to be
solved. For example, the SVM algorithm was used to locate pig ears and perform the temperature
extraction in the infrared image (Lu et al., 2018), however, some deviations for pig heads and ear
positioning always exist. An improved Ostu algorithm was used for the ear root region detection
in the thermal infrared image (Zhou, 2016), but it could not extract ear root features completely
resulting an incorrect detection.
The objective of this study is to develop an accurate method based on YOLOv4 for pig
individual identification and body key part recognition using visible images, and to provide basis
for non-contact and rapid individual identification and body temperature detection.
The data used in this study were collected from a grouped pig house of a pig farm located in
Yabuli, Heilongjiang Province, China. There were 100 fattening pigs in the pig house. Each pig
was 270 days old and weighed about 200 kg. Six finisher pigs (Duroc × (Landrace × Large White))
were randomly selected for video data collection. Videos of pig face and ear root were captured
from the front and top perspectives, respectively. From the front view, data of pig face and
forehead could be obtained. From the top view, data of the ear root could be gotten. The video
data were captured from the original status of pig face without cleaning. The length of the video
recorded for each pig from front and top view was 4 minutes, with a video sampling rate of 30
Frames Per Second (FPS) and a resolution of 1920 × 1080 pixels. OpenCV was used for data
extraction from the video stream, and the images were extracted from the videos with an interval
of 40 frames. A total of180 images was obtained for each pig from the front and top views.
2.2. Data preprocessing
2.2.1. Data selection
In order to remove some unusable image data of blurring and targets border outing, the data
used in study were selected manually. The structural similarity (SSIM) was used to measure the
similarity between two images. In this paper, the threshold of SSIM was set as 0.5, when the value
of SSIM was higher than 0.5, the two pictures were considered to be too similar and be deleted.
After two data filtering methods, the number of images corresponding to front and top view for
each pig was shown in Table 1.
Table 1. Image Data selection for each pig.
View Pig 1 Pig 2 Pig 3 Pig 4 Pig 5 Pig 6
Front 137 121 146 113 99 120
Top 139 145 124 120 108 110
2.2.2. Data enhancement
Due to the complexity of the real commercial pig farm environment, some challenges of
different facial angles, different brightness of light may exert the influence on individual’s
identification and body key parts recognition. Therefore, to improve the generalization ability of
the model, an offline method for data enhancement was applied for image transformations of
randomly adjusting brightness, adding random occlusion, increasing Gaussian noise, and rotating.
2.2.3. Data set construction
After data enhancement, the original colored visible images collected from 6 pigs was
expanded from 1482 to 5928. Among them, there were 2350 and 2392 images from front and top
view, respectively. The dataset was established with these images in a disorder way, and be
divided into three sets of training, validation, and testing as a ratio of 8:1:1 (Table 2).
∙ 311 ∙
Table 2. Dataset division.

Dataset Forehead Ear root Pig 1 Pig 2 Pig 3 Pig 4 Pig 5 Pig 6
Training 2350 2392 430 396 474 352 318 380
Validation 297 296 59 44 55 50 39 50
Testing 297 296 59 44 55 50 39 50
In this paper, Labelimg, an open-source image annotation tool, was used to annotate the
enhanced pig faces and ear roots in the top view image dataset. The annotation format was
PASCAL VOC. The pig face, forehead and the ear root were marked as pig x (x is the number of
the pig), forehead, and ear root, respectively. The 5928 visible images with RGB three-channel
were annotated to generate the corresponding xml tag files to construct the dataset.
3. Model development for pig individual identification and body key parts recognition
3.1. Model structure
The model consists of image data input (Input), backbone feature extraction network
(Backbone), feature enhancement network (Neck) and prediction network (YOLO head). The
overall structure of the model is shown in Figure 1.
3.1.1. Input
The size of the model input image is fixed, it was divided into two sizes of 416 × 416 and 608
× 608. Considering the video memory and detection speed of computers, the size of 416 × 416
was selected as the image input mode in this research. Therefore, the pig image with resolution of
1920 × 1080 pixels need to be scaled to 416 × 416 pixels before input into the model, and the
blank were filled with grey bars to uniform the input for all data.
3.1.2. The backbone feature extraction network
The backbone feature extraction network CSPDarknet-53 is mainly composed of convolution
block (DarknetConv2D-BN-Mish) and multiple residual convolution blocks (resblock-body). The
DarknetConv2D-BN-Mish is composed of a single convolution, normalization and activation
function (Mish), which is shown in Eq. (1). The YOLOv4 uses the CSPnet structure (Wang et al.,
2019) in the residual convolution block, and splits the stack of the original residual block into two
parts. The main part continues to stack the original residual blocks, and the other part is like a
residual edge, directly connected to the last after a small amount of processing. This structure can
not only improve the model training accuracy, but also speed up the convergence rate. Three
effective feature layers (Figure 1), feat1 (52 × 52 × 256), feat2 (26 × 26 × 512), feat3 (13 × 13 ×
1024), could be obtained from the input pig image data through the backbone feature extraction
network.
  
Mish  x  tan h ln 1  e x (1)
3.1.3. Feature-enhanced network

The feature enhancement network uses two structures of Spatial Pyramid Pooling
(SPP)module (He et al., 2015) and Path Aggregation network (PANet) (Liu et al., 2018) to
enhance the feature extraction work. SPP uses four different scale pooling cores of 13 × 13, 9 ×
9, 5 × 5, and 1 × 1 to extract and re-aggregate the feature layer from different angles, which can
greatly increase the receptive field and separate the most significant contextual features. The three
effective feature layers were sampled in the PANet from top to bottom and from bottom to top,
so as to achieve repeated feature extractions and output feature maps for pig identification at three
scales.
3.1.4. Prediction Network
The prediction network (YOLO head) detects the three effective feature maps output by the
feature enhancement network. Take the detection of pig's forehead as an example. The three
∙ 312 ∙
effective feature maps were divided into grids of 76 × 76, 38 × 38 and 19 × 19, and each grid was
responsible for the detection of a region. The best detection frame for pig forehead recognition
was obtained by scoring the prediction frame and non-maximum suppression screening, as shown
in Figure 2. Each grid unit of YOLOv4 consists of 3 boxes (anchor), each box contains five
parameters (x, y, w, h, c), these parameters represent the horizontal coordinate of the central point,
the vertical coordinate of the central point, the length of the prediction box, the width of the
prediction box, and the confidence level. In the prediction process, the recognition model
continuously adjusts these parameters to realize detection for the target.
Input (416,416,3)
Feature enhancement network Prediction
Backbone feature
network
extraction network
PANet
CSPDarknet53 Conv YOLO
Concat + Conv ×5
DarknetConv2D_BN_Mish(416,416,32) Head
Conv + UpSampling DownSampling
YOLO
Concat + Conv ×5 Concat + Conv ×5
Resblock_body(208,208,64)×1 Head
Conv + UpSampling DownSampling
Resblock_body(104,104,128)×2
YOLO
Concat + Conv ×5 Head
feat1
Conv
Resblock_body(26,26,512)×8 SPP
feat2 5 Concat + Conv ×3
9
feat3
13
Conv ×3
Figure 1. Model structure diagram.
Figure 2. Prediction for pig forehead region.

3.2. Model training parameter setting and environment
The training process of the model was divided into two stages, and the model parameters were
set differently in each stage. In the first stage, the pre-trained weights of YOLOv4 on the coco
data set (YOLOv4_weights.h5) was loaded, and the backbone network was frozen to adjust the
weights to speed up the training. In the first stage, the epochs were 50, the initial learning rate was
set to 0.001, and the batch size was 4. In the second stage, the network was unfrozen, the epochs
were 150, the initial learning rate was 0.0001, and the batch size was 2.
The entire training process was optimized using the adaptive moment estimation (Adam)
method. To prevent overfitting, the label smoothing was 0.01, and the learning rate was updated
with cosine annealing decay. At the same time, the early stopping strategy was adopted. If the
model had no change in the loss value of the validation set within 10 epochs, the training will be
ended.
∙ 313 ∙
A desktop computer with hardware configured of Intel i7-9700 3.0 GHz CPU, GTX1660
GPU, 6GB independent video memory and 32GB memory. An operating system of Microsoft
windows, development tools of OpenCV and Tensorflow deep learning framework.
3.3. Evaluation Index
Using the loss, precision rate (P), the recall rate (R), the comprehensive evaluation index F1
score (F1) of the precision rate and the recall rate, average precision (AP) and the mean average
precision (mAP) as the model performance evaluation index, shown in Eqs. (2)–(5). In target
detection, each category can draw a P-R curve based on P and R. AP was the area between the P-
R curve and the coordinate axis, and mAP was the average of all categories of AP.
loss  lossciou  lossconfidence  lossclass (2)
TP (3)
P  100%
Tp  FP
TP (4)
R  100%
Tp  FN
2PR
F1  (5)
PR
where, lossciou is the loss of predicted bounding boxes, lossconfidence is the loss of predicted
confidence losses, and lossclass is the loss of predicted categories, TP is the number of samples
correctly identified, FP is the number of samples incorrectly identified, FN is the number of
unrecognized samples.
4.1. Training results analysis
The loss curves in the training set and validation set were similar (Figure 3), they all showed
a downward trend as the number of epochs increase, and finally oscillated within a certain range
and reached a stable state. There was a rapid drop on the training and validating losses in the first
25 epochs. From the 25th epoch to the 60th epoch, the losses started slowly. After 70th epoch of
training epochs, the loss of the training set and the validation set oscillated around 1.0 and 1.1,
respectively.
Figure 3. Loss curve with epochs.

4.2. Recognition results analysis
The trained YOLOv4 model was tested on the test set with 593 images. The partial recognition
results were shown in Figure 4. Figure 4 A and Figure 4 B were the recognition results of pig
individuals, foreheads and ear roots in the original image data under the front and top view
respectively, while Figure 4 C and Figure 4 D were the recognition results of Figure 4 A and
∙ 314 ∙
Figure 4 B enhanced by brightness and random rotation. By comparing the four images in pairs
in the longitudinal direction, it can be found that the YOLOv4 model could accurately detect pigs,
foreheads and ear roots with almost the same recognition accuracy, whether in the original image
or after the data has been enhanced.
A B
C D
Figure 4. Partial test results. The green box was the pig face recognition, the pink box was the
forehead recognition, and the blue box represented the recognition result of the ear root.
4.3. Comparison of different models
To verify the performance of the YOLOv4-based recognition model established in this paper,
the models of YOLOv3, EfficientNet, and RFB were compared using the same experimental
environment and dataset.
4.3.1. Training loss comparison
During the training process of the four models, the variations of loss were very different
(Figure 5). The overall trend of the loss curves of the four models was that as the number of epochs
increased, the loss value continued to decrease and gradually reached convergence. Compared
with the other three groups of models, the loss curve of YOLOv4 dropped the fastest and reached
the state of convergence in the shortest epochs. Although EfficientNet had the smallest training
loss value from the beginning to the end, the process of model convergence was too slow. It
converged after 120 epochs, twice as those of YOLOv4.
A B
Figure 5. Loss curve of the four models. A: training losses; B: validation losses.
4.3.2. Performance comparison
Four groups of trained models were tested and compared on the same test set. It can be seen
from Table 3 that the F1 score of YOLOv4 model for individual identification of six pigs was
2.07% higher than that of YOLOv3 of the same series, 0.11% and 24.48% higher than that of
EfficientNet and RFB. The performances of body key parts of the forehead and ear root
recognition on the test set were compared (Table 4). The mAP of YOLOv3, EfficientNet and RFB
∙ 315 ∙
were 98.48%, 98.48%, 58.84% respectively. They were 0.68%, 0.68% and 40.32% lesser than the
YOLOv4 model, respectively. It can be seen from Table 5 that the number of parameters in the
EfficientNet model were 3886253，it was less than one-tenth of those in other models. When
identifying individual pigs and body key parts on the test set, the values of F1 and mAP were
second only to YOLOv4. The lightweight feature of EfficientNet model increases the opportunity
to be employed on terminal devices with a smaller storage capacity and lower configuration to
develop a model for real-time pig recognition.
4.3.3. Performance comparison
Four groups of trained models were tested and compared on the same test set. It can be seen
from table 3 that the F1 score of YOLOv4 model for individual identification of six pigs was
2.07% higher than that of YOLOv3 of the same series, 0.11% and 24.48% higher than that of
EfficientNet and RFB. The performances of body key parts of the forehead and ear root
recognition on the test set were compared (Table 4). The mAP of YOLOv3, EfficientNet and RFB
were 98.48%, 98.48%, 58.84 respectively. They were 0.68%, 0.68% and 40.32% lesser than the
YOLOv4 model, respectively. The parameters of YOLOv4, YOLOv3, EfficientNet and RFB
models were 6404685, 61614037, 3886253 and 37978464, respectively. However, the parameters
of EfficientNet were less than one-tenth of other models. When identifying individual pigs and
body key parts on the test set, the values of F1 and mAP were second only to YOLOv4. The
lightweight feature of EfficientNet model increases the opportunity to be employed on terminal
devices with a smaller storage capacity and lower configuration to develop a model for real-time
pig recognition.
Table 3. Effects of different models on individual identification of pigs in the test set.
Models P R F1
YOLOv4 98.13% 99.57% 98.84%
YOLOv3 94.50% 99.15% 96.77%
EfficientNet 97.80% 99.67% 98.73%
RFB 59.34% 99.57% 74.36%
Table 4. The performances of body key parts recognition on the test set using different models.
AP
Models mAP
Ear root Forehead
YOLOv4 99.66% 98.65% 99.16%
YOLOv3 98.53% 98.42% 98.48%
EfficientNet 98.91% 98.55% 98.48%
RFB 52.03% 65.65% 58.84%
5. Conclusions
(1) The identification model developed based on YOLOv4 in this study could achieve an
accurate recognitions of pig individual and body key parts both in the front and top view. The
averaged precision and recall rate of pig face recognition were 98.13% and 99.57%, respectively.
The mAP for the identification on the body key parts of the ear root and forehead was 99.16%.
(2) Compared with the models of YOLOv3, EfficientNet, and RBF using the same dataset,
the F1 of YOLOv4 for individual identification of pigs were 2.07%, 0.11% and 24.48% higher
than theirs, respectively, and the mAP for identifying key parts were increased by 0.68%, 0.68%
and 40.32%, respectively.
(3) Among the four models, the EfficientNet model had the least number of parameters, and
the second performance compared to the YOLOv4 model, which may give an opportunity for a
real-time recognition achieved on the onsite devices.
∙ 316 ∙
Acknowledgements
Laboratory of Swine Facilities Engineering, Ministry of Agriculture, P.R. China. Also, the authors
References
Alsaaod, M., C. Syring, J. Dietrich, M.G. Doherr, T. Gujan, A. Steiner, 2014. A field trial of infrared
thermography as a non-invasive diagnostic tool for early detection of digital dermatitis in dairy cows. The
Veterinary Journal. 199 (2).
Amezcua, R., S. Walsh, P.H. Luimes, R.M. Friendship, 2014. Infrared thermography to evaluate lameness
in pregnant sows. Canadian Veterinary Journal La Revue Vétérinaire Canadienne. 55 (3), 268–272.
Bagavathiappan, S., B.B. Lahiri, T. Saravanan, J. Philip, T. Jayakumar, 2013. Infrared thermography for
condition monitoring - A review. Infrared Physics & Technology. 60, 35–55.
Gao, L., Duan, G., Yin, G., Zou, F., Yang, L., Yang, G., 2010. Trial Results of Infrared Thermometer in
the Application of Ante-mortem Body Temperature Screening in the Pig Slaughterhouse. China Animal
Husbandry & Veterinary Medicine. 37 (09), 235–237 (in Chinese with English abstract).
Guo, Y., Zhu, W., Ma, C., Chen, C., 2016. Top-view recognition of individual group-housed pig based
on Isomap and SVM. Transactions of the Chinese Society of Agricultural Engineering. 000 (003), 182–187
(in Chinese with English abstract)
He, K., X. Zhang, S. Ren, J. Sun, 2015. Spatial Pyramid Pooling in Deep Convolutional Networks for
Visual Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence.
Liu, S., L. Qi, H. Qin, J. Shi, J. Jia, 2018. Path Aggregation Network for Instance Segmentation. In: IEEE.
Lu, M., J. He, C. Chen, C. Okinda, M. Shen, L. Liu, W. Yao, T. Norton, D. Berckmans, 2018. An
automatic ear base temperature extraction method for top view piglet thermal image. Computers and
Electronics in Agriculture. 155, 339–347.
Marsot, M., J. Mei, X. Shan, L. Ye, P. Feng, X. Yan, C. Li, Y. Zhao, 2020. An adaptive pig face
recognition approach using convolutional neural networks. Computers and Electronics in Agriculture. 173,
105386.
Menzel, A., M. Beyerbach, C. Siewert, M. Gundlach, D. Hoeltig, R. Graage, H. Seifert, K. Waldmann, J.
Verspohl, I. Hennig-Pauka, 2014. Actinobacillus pleuropneumoniae challenge in swine: diagnostic of lung
alterations by infrared thermography. BMC veterinary research. 10.
Siewert, C., S. Dänicke, S. Kersten, B. Brosig, D. Rohweder, M. Beyerbach, H. Seifert, 2014. Difference
method for analysing infrared images in pigs with elevated body temperatures. Zeitschrift Fur Medizinische
Physik. 24 (1).
Soerensen, D.D., L.J. Pedersen, 2015. Infrared skin temperature measurements for monitoring health in
pigs: A review. Acta veterinaria Scandinavica. 57 (1), 5.
Wada, N., M. Shinya, M. Shiraishi, 2013. [Short Paper] Pig Face Recognition Using Eigenspace Method.
ITE Transactions on Media Technology and Applications. 1 (4).
Wang, C.Y., H. Liao, I.H. Yeh, Y.H. Wu, P.Y. Chen, J.W. Hsieh, 2019. CSPNet: A New Backbone that
can Enhance Learning Capability of CNN.
Wang, R., Shi, Z., Gao, R., Li, Q., 2020. Individual Identification of Pigs based on Multi-scale
Convolutional Network in a Variable Environment. Acta Agriculturae Universitatis Jiangxiensis. 042 (002),
391–400 (in Chinese with English abstract).
Zhang, Z., Zhang, H., Liu, T., 2019. Study on body temperature detection of pig based on infrared
technology: A review. Artificial Intelligence in Agriculture. 1.
Zhou, L., Chen, Z., Chen, D., Yuan, Y., Li, Y., Zheng, J., 2016. Pig Ear Root Detection Based on Adapted
Otsu. Transactions of The Chinese Society of Agricultural Machinery (4), 228–232 (in Chinese with English
abstract).
∙ 317 ∙
Robust Audio Fingerprint Algorithm for Automatic Recognition of

Laying Hens’ Vocalizations
Tiantian Du a,b,c, Ligen Yu a,b, Tonghai Liu c, Qingfei Xue a,b,c, Rui Meng a,b, Luyu Ding a,b,
Weihong Ma a,b, Ronghua Gao a,b, Qifeng Li a,b,*
a
National Engineering Research Center for Information Technology in Agriculture, Beijing 100097, China
b
Beijing Research Center for Information Technology in Agriculture, Beijing 100097, China
c
College of Computer and Information Engineering, Tianjin Agricultural University, Tianjin 300384, China
* Corresponding author. Email: liqf@nercita.org.cn
Abstract
Vocalizations were recorded non-invasively, and their dependence on inner states makes them
useful indicators of animal welfare states, emotional states and behavior contents. Therefore,
research on methods for classification and recognition of farm animal vocalizations based on the
time domain and frequency domain features are becoming more and more interested in the field
of precision livestock farming. In this paper, an audio fingerprint algorithm based on the local
energy features was proposed for automatic recognition of laying hens’ vocalizations. Sounds of
the laying hens were acquired from an experimental arena with perch system and four categories
of vocalizations including the Gakel, Alarm, Squawk and Laying calls were collected. The
frequency bands of each frame of laying hens’ vocalizations were divided into four groups of sub-
regions, and the energy of each sub-region was calculated simultaneously. Then the audio
fingerprint was encoded by comparing the adjacent sub-regions energy, whereas the acoustic
features could be described by a series of binary strings to classify laying hens’ vocalizations. An
adaptive audio fingerprint matching method was proposed for automatic recognition. Results
showed that the recall rates of the audio fingerprint algorithm used for classification of the Gakel,
Alarm, Squawk and Laying calls were 91.7%, 88.3%, 93.3% and 88.3%, respectively; the
precision rates were 88.7%, 91.4%, 90.3% and 91.4%, respectively; the F1-scores were 0.902,
0.898, 0.918 and 0.898, respectively. This work shows that the audio fingerprint algorithm can
achieve quick and automatic recognition for laying hens’ vocalizations.
Keywords: Laying hens, vocalizations, audio fingerprint, binary strings, automatic recognition
1. Introduction
With the development of digital audio technologies, the vocalizations of farm animals can be
analyzed, characterized and identified, which helped us to know the body condition and emotional
status non-invasively (Norton et al., 2019). It is of great significance to judge the health states of
animals and to improve the digitalization, intelligence and automation of livestock management
(McLoughlin et al., 2019). At least in theory, the sounds can be automatically identified from data
provided by the autonomous recorders. Compared with traditional methods by detecting
physiological and biochemical performances, sound monitoring will not cause additional stress to
the animals (Cuan et al., 2020). To date, many recognition methods have been applied to the study
of livestock vocalizations. The convolutional neural networks (CNNs) were used to classify cattle
vocals, and the accuracy was 81.96% (Jung et al., 2021). The support vector machines (SVM)
was determined to identify the sounds of chickens infected with avian influenza, and the
recognition rate in the test was 84% to 90% (Huang et al., 2019). Feed intake estimation using
acoustic analysis has been widely studied (Aydin and Berckmans, 2016).
Numerous acoustic features were proposed to use for sound identification and classification
in precision livestock farming, such as the relative sound intensity (Moura et al., 2008), power
spectra density (PSD) (Wang et al., 2019), peak frequency (Abdel-Kafy et al., 2020), energy
envelope (Exadaktylos et al., 2014) and Mel Frequency Cepstral Coefficients (MFCCs)
(Ntalampiras et al., 2019). In particular, MFCCs are commonly adopted as the acoustic features
(Liu et al., 2020). Along with more and more research on audio, which put forward a higher
∙ 318 ∙
requirement for the efficiency of acoustic features, quantities of digital audio technologies have
appeared and gave rise to the audio fingerprint algorithm.
Audio fingerprint represents the digital signature of the important acoustic features of the
audio which can be seen as a brief summary of the audio signals (Cano et al., 2005). Compared
with traditional feature extraction methods based on the time domain and frequency domains,
audio fingerprint has relatively small and precise characterization for classification and
identification. It is widely used in music identification, environmental sound identification,
common carotid artery signals recognition and other application scenarios (Pires et al., 2018). One
of the most classic algorithms is Philips’s fingerprint algorithm (Haitsma et al., 2003), which
computes audio fingerprints by comparing the energies in adjacent bands. When the audio signal
is interfered by noise, the algorithm retrieval performance can be maintained; but Philips
fingerprints algorithm cannot resist the time stretch of the audio signal. In the paper, the bits of
the signal depend on the comparing the adjacent sub-regions energy. Another issue that needs to
be considered is how to match audio fingerprint similarity reasonably to identify the unknown
laying hens’ vocalizations. Currently, more methods can be used to compare the bits error rates
(BER) of the two audio fingerprints. That is, the lower the bits error rate of the audio fingerprint
and the audio template, the more likely the sound type is the same to the audio template.
Different from previous studies on recognition of animal vocalizations, this study attempts to
precisely characterize and identify the laying hens’ vocalizations. The objectives of this research
were: (1) to develop an audio fingerprint algorithm for characterizing the typical calls of laying
hens in perch system; (2) to provide an adaptive matching method for recognition of laying hens’
vocalizations non-invasive.
2.1. Animals and Housing
The experiments started on January 13, 2021 and were carried out on Shangzhuang
Experimental Station of China Agricultural University (Haidian District, Beijing, China), and
were lasted for three consecutive weeks. The laying hens of Jing Tint 6 Variety (a group of 2700
hens) were raised on the perch system of a 28.0 m × 9.0 m × 4.3 m (L × D × H) building over a
69–71 weeks’ period. All experimental procedures were conducted in accordance with the
management guidelines of Jing Tint 6 Variety and treatments for the care and use of laboratory
birds, naturally intake and drinking, and the lighting time of the breeding house was from 6:00
AM to 10:00 PM. All recording procedures were non-invasive and did not cause any disturbance
to the birds in their regular daily activity.
2.2. Data Collection
The sounds were recorded by BK 2270-S-C (a sound monitoring analyzer) with a 4189 free-
field microphone (audio sample rate of 44,100 Hz, 16 bits resolution, single-channel and WAVE
format). The microphone was installed approximately 2.2 m above the building floor in the center
of the breeding house with perch system.
2.3. Sound Signal Processing and Labelling
The software of Adobe Audition CC 2018 (Adobe Systems Incorporated, San Jose, State of
California, USA) was used to visualize, identify, replay and label the sounds manually from the
whole recordings. In order to reduce the data amount of the audio, the audio signal was down-
sampled to 3,000 Hz. The selected sound segments were classified into four categories with the
Gakel, Alarm, Squawk, and Laying calls. The description of the laying hens’ vocalizations with
four categories in the perch system (Table 1).
2.4. Audio Fingerprint Construction
2.4.1. Sub-region Division Based on Frequency Spectrum
The length of vocalizations was selected with 5 frames, and the Fourier transform was
∙ 319 ∙
performed on each frame to obtain the frequency spectrum, and the frequency spectrum was
divided into 19 frequency bands. The sub-regions were composed of several adjacent frequency
bands and were allowed to overlap with the adjacent frequencies. The dimension and location of
each sub-region and the total number of sub-regions depend on the number of total bits of the
audio fingerprint we need. The local energy features were composed of sub-regions, its energy
was represented by the total energy of several frequency bands.
Table 1. Description of the laying hens’ vocalizations with four categories, each with 120
samples, in the perch system.
Sound type Description
Short repetitive notes generally had a wide frequency range, the base
Gakel frequency was often concentrated in the low frequency part, and the notes
had a significant and clear harmonic structure.
High pitch sounds had a significant harmonic structure, and the sound
Alarm
intensity was moderate.
Short notes had steep starting and ending wave forms, with wide frequency
Squawk
bands and moderate sound intensity.
The sounds made by the laying hens during the laying process had a series
Laying
of shorter notes and a unique harmonic structure.
The local energy features were constructed shown in three sub-images for better visualization
(Figure 1). They were defined as follows: the first group encoded the horizontal evolution of the
energy, consisting of the frequency bands contained in the 1a–1h sub-region; the second group
encoded the vertical evolution of the energy, consisting of frequency bands contained in the 2a–
2d sub-region; the third group encoded the most immediate region around the center of the energy,
consisting of frequency bands contained in the 3a–3d sub-region; the fourth group encoded
marginal evolution of the energy, consisting of frequency bands contained in the 4a–4h sub-
region. In total, 22 sub-regions were defined. More sub-regions can be easily obtained by defining
other sub-regions.
3a 4e 4f
Frames
1a 1c
3c
Bands
(a) Sub-image 1
2a
2b
Frames
1b 1d 1e 1g
2c
2d
Bands
(b) Sub-image 2
1b 1b 3b
Frames
3d 1b 1b
Bands
(c) Sub-image 3
Figure 1. The division of the sub-regions of 5 frames in every 19 bands, divided into 3 sub-
images to better visualize the overlapping regions.
2.4.2. Encoding Rules Based on Local Energy Features
The energy values vary along the time and frequency domain. When the energy differences
∙ 320 ∙
between the two parts of the compared local energy feature were positive, the corresponding bits
code was set to 1; when the differences were negative, the corresponding bits code was set to 0.
To a certain extent, the generated audio fingerprints represented the topological structure
information of the energy changes of the laying hens’ vocalizations in different frequency bands,
which was a brief description of the energy information in the frequency bands. The encoding
rules based on the local energy features were constructed (Table 2).
Table 2. The encoding rules based on local energy features.
Bits number Group type Comparisons
1-7 Horizontal 1a-1b, 1b-1c, 1c-1d, 1d-1e, 1e-1f, 1f-1g, 1g-1h
8-10 Vertical 2a-2b, 2b-2c, 2c-2d
11-14 Central 3a-3b, 3d-3c, 3a-3d, 3b-3c
15-22 Marginal 4a-4b, 4c-4d, 4e-4f, 4g-4h, 4a+4b-4c-4d, 4e+4f-4g-4h,
4c+4d-4e-4f, 4a+4b-4g-4h
The generated audio fingerprints have a clear frequency band energy information: 1–7 bits
represented band energy change information between different frequency bands in the same
frames; 8–10 bits represented band energy change information of the same frequency band in
different frames; 11–14 bits represented band energy change information at the center position;
15–22 bits represented band energy change information at the marginal position.
2.4.3. Binary audio fingerprint generation
For the generated binary audio fingerprint, the probability of each audio fingerprint 60 clips
of each bit was counted, where 1 or 0 accounted for more than 50% of the binary probability, then
22 bits audio fingerprint template were generated according to the calculation of the binary
probabilities.
2.5. Audio Fingerprint Matching
The classical similarity measurement used in audio fingerprint matching were the bits error
rate and strong or weak bits. Due to the limitations of the two classical methods, A new audio
fingerprint matching method based on the variance of each audio fingerprint as an adaptive
weight matching template was proposed and compared with the bits error rate matching and the
strong or weak bits matching methods.
(1) Bits Error Rate Matching
During the similarity matching process, two sound signals were declared a match if the bits
error rate between the template matching audio fingerprint and audio fingerprint template was
below a certain threshold. The classic of bits error rate between the two audio fingerprints F(n,m)
and F’(n,m) were calculated and shown here in Eq. (1).
N M
  F (n, m)  F '(n, m)
BER  n 1 m 1
PQ (1)
where F(n,m) was the audio fingerprint template database, F’(n,m) was the template matching
audio fingerprint,⨂was XOR logic operation, n was the time frame, m was the bits of frame.
(2) Strong or Weak Bits Matching
The average value of relative energy deviation between sub-region of the same vocalizations
categories was calculated.The formula for calculating the relative deviation of defined energy was
shown here in Eq. (2) and Eq. (3).
1 N
a b
P
N

i 1 ab
100%
(2)
∙ 321 ∙
N M
 pF (n, m)  F '(n, m)
BER  n 1 m 1
PQ (3)
where a represented the previous sub-region energy value and b represented the latter sub-
region energy value, which encoded weak bits by p=0 and strong bits by p=1.
(3) Adaptive Weighted Matching
The calculation of the adaptive weighted matching was shown here in Eq. (4).
N M
 (1   ) pF (n, m)  F '(n, m)
BER  n 1 m 1
PQ (4)
where σ represented the energy variances for bits.
2.6. Evaluating Algorithms
The precision rate, recall rate and F1-scores were calculated for evaluating the recognition of
laying hens’ vocalizations. The performance of similarity matching was evaluated by calculating
the recall and precision. The F1-scores was evaluated by calculating mean of precision and recall.
The calculation methods for the precision rate, recall rate, and F1-scores were shown here in Eq.
(5), Eq. (6) and Eq. (7).
TP (5)
precision  100%
TP  TF
TP (6)
recall  100%
TP  FN
2TP (7)
F1  scores 
2TP  FN  FP
where TP signified the number of query shots correctly cases as positive, TF referred to the
number of query shots incorrectly cases as positive, FN indicated the number of query shots
incorrectly cases as negative, and FP was the number of negative cases classified as positive cases.
2.7. Audio Fingerprint Robust
To test the robustness of audio fingerprint algorithm based on the local energy features, 5 dB
and 10 dB white noise were added to the test sample data, and the influences of the noises on the
precision rate, recall rate and F1-scores were collected and compared.
3.1. Audio Fingerprint Template of the Laying Hens’ Four Categories Vocalizations
The audio fingerprint template of four categories of the laying hens’ vocalizations
was shown in Figure 2 (1 is indicated by black).
Gakel
Alarm
Squawk
Laying
Figure 2. The audio fingerprint template of four categories of the laying hens’ vocalizations.
3.2. Matching Algorithms for Recognition of Laying Hens’ Vocalizations
The precision rate, recall rate and F1-scores of the matching algorithms for recognition of
laying hens’ vocalizations were shown in Figure 3. The adaptive weight matching algorithm had
∙ 322 ∙
the best recognition performance, while the F1-scores of the Gakel, Alarm, Squawk and Laying
calls were 0.902, 0.898, 0.918, and 0.898, respectively. The F1-scores of the same four categories
of the laying hens’ vocalizations using the strong or weak bits matching algorithm were 0.830,
0.807, 0.841, and 0.803, respectively. Based on the bits error rate matching algorithm, the
recognition effect was the worst. The F1-scores of the same four sounds were 0.756, 0.739, 0.787,
and 0.783, respectively.
(a) the recall rate (b) the precision rate
(c) the F1-scores

Figure 3. (a) The performances of the matching algorithms for recognition of laying hens’
vocalizations with (a) the recall rate, (b) the precision rate and (c) the F1-scores.
The acoustic characteristics of the same category vocalizations produced by the laying hens
have certain differences and stability. While the band energy differences of the sub-regions with
the frequency spectrum was close to zero and were vulnerable to background noises. During the
generation of the audio fingerprint template, the probability of each bit appearing 1 (or 0) was not
exactly the same, and some bits may appear 1 (or 0). The variance was very high, and the
possibility of 1 (or 0) in some bits were very small.
The classic bits error rate matching could not consider the differences among the energy in
the process of forming the fingerprint. This sub-region of the frequency spectrum was recognized
as weak bits, when the compared sub-region energy was less than the threshold. Similarly, strong
bits were extracted from the sub-regions when the compared sub-region energy was more than
the threshold. This method retained the temporal and spatial sequential characteristics of the raw
audio sub-region energy difference, as well as the internal structure and logical correlation of the
data, and retained the statistical value of the raw energy distribution. After that a strong or weak
bits audio fingerprint matching algorithm was proposed.
In the audio fingerprint matching process, the energy variance of each bit is not exactly the
same. The greater the energy variance, the higher the degree of randomness in the process of
generating fingerprints, the worse the certainty in the matching process will be. It is the first time
∙ 323 ∙
to propose the adaptive weight matching by using the variance of each bit.
The acoustic parameters of the vocalizations produced by the laying hens varied non-linearly
within a certain range, which would lead to a different probability of 1 (or 0) in each bit within
the fingerprint generation process. The weight of classic bits error rate audio fingerprint matching
algorithm of each bit was the same, and the difference in the probability of each bit was not used,
resulting in the worst recognition effect. The strong or weak bits matching algorithm was taken
into account that the bits were unstable due to the small energy difference, and the bits with the
small energy difference were assigned a weight of 0, which was ineffective in the fingerprint
matching process. The adaptive weighted audio fingerprint matching algorithm used the
probability of 1 (or 0) occurrence each bit as the matching weight, and the matching effect have
been significantly improved.
3.3. The Result of Audio Fingerprint Robust Test
The noise of audio fingerprints means that the audio fingerprint remains stable under different
noise resistance, and it can accurately identify the original audio from the audio database. The
precision rate, recall rate and F1-scores on the Signal Noise Ratio (SNR) were acquired (Table 3).
With the increase of Alarm added noise ratio, the analysis reason is the small adjacent local
energy difference and the poor robustness. The analysis found that the added noise is less, so the
impact is not very much on the F1-scores, because the less noise has a smaller impact on the
comparative local energy.
Table 3. The precision rate, recall rate and F1-scores on the Signal Noise Ratio (SNR).
SNR Type Recall (%) Precision (%) F1-scores
Gakel 91.7 88.7 0.902
Alarm 88.3 91.4 0.898
Clean Squawk 93.3 90.3 0.918
Laying 88.3 91.4 0.898
Gakel 85.0 87.9 0.864
Alarm 86.6 83.9 0.852
5 dB Squawk 91.6 87.3 0.894
Laying 86.6 91.2 0.888
Gakel 81.7 87.5 0.845
Alarm 83.3 79.4 0.813
Squawk 90.0 88.5 0.892
10 dB
Laying 85.0 85.0 0.850
Summary and Conclusions
This paper proposes an innovative audio fingerprint algorithm for classification and
recognition of laying hens’ vocalizations. The calls were characterized by binary strings and were
identified by an adaptive matching method. The F1-scores of the Gakel, Alarm, Squawk and
Laying calls reached to 0.902, 0.898, 0.918 and 0.898, respectively. The precision rate and recall
rate of the experiments indicated that the algorithm worked well. This method has the advantages
of precise characteristic parameters and matching method. It provides a new approach for the
sound identification of livestock and poultry in the future.
Acknowledgments
The authors would like to acknowledge the financial support for this research provided by the
Youth Funds of Beijing Academy of Agriculture and Forestry Sciences (GNJJ201913), the
Innovation Capacity Building Project of Beijing Academy of Agriculture and Forestry Sciences
(KJC202111007) and the National Natural Science Foundation of China Youth Funds
(31402113). In addition, the authors wish to express their appreciation to Prof. Baoming Li and
Prof. Guanghui Teng for the support of experimental sites and technical guidance.
∙ 324 ∙
References
Abdel-Kafy, E.M., S.E. Ibraheim, A. Finzi, S.F. Youssef, F.M. Behiry, G. Provolo, 2020. Sound analysis
to predict the growth of turkeys. Animals (Basel). 10 (5). http://doi.org/10.3390/ani10050866.
Aydin, A., D. Berckmans, 2016. Using sound technology to automatically detect the short-term feeding
behaviours of broiler chickens. Computers and Electronics in Agriculture. 121, 25–31. http://doi.org/10.1016/
j.compag.2015.11.010.
Cano, P., E. Batlle, T. Kalker, J. Haitsma, 2005. A review of audio fingerprinting. Journal of VLSI Signal
Processing Systems for Signal, Image and Video Technology. 41 (3), 271–284. http://doi.org/10.1007/s11265-
005-4151-3.
Cuan, K., T. Zhang, J. Huang, C. Fang, Y. Guan, 2020. Detection of avian influenza-infected chickens
based on a chicken sound convolutional neural network. Computers and Electronics in Agriculture. 178.
http://doi.org/10.1016/j.compag.2020.105688.
Exadaktylos, V., M. Silva, D. Berckmans, 2014. Automatic identification and interpretation of animal
sounds, application to livestock production optimization. In: Soundscape Semiotics - Localization and
Categorisation.
Haitsma, J., T. Kalker, 2003. A highly robust audio fingerprinting system with an efficient search strategy.
Journal of New Music Research. 32 (2), 211–221. http://doi.org/10.1076/jnmr.32.2.211.16746.
Huang, J.D., W.Q. Wang, T.M. Zhang, 2019. Method for detecting avian influenza disease of chickens
based on sound analysis. Biosystems Engineering. 180, 16–24. http://doi.org/10.1016/j.biosystemseng.
2019.01.015.
Jung, D.H., N.Y. Kim, S.H. Moon, C. Jhin, H.J. Kim, J.S. Yang, H.S. Kim, T.S. Lee, J.Y. Lee, S.H. Park,
2021. Deep learning-based cattle vocal classification model and real-time livestock monitoring system with
noise filtering. Animals (Basel). 11 (2). http://doi.org/10.3390/ani11020357.
Liu, L., B. Li, R. Zhao, W. Yao, M. Shen, J. Yang, 2020. A novel method for broiler abnormal sound
detection using WMFCC and HMM. Journal of Sensors. 2020, 2985478. http://doi.org/10.1155/
2020/2985478.
McLoughlin, M.P., R. Stewart, A.G. McElligott, 2019. Automated bioacoustics: methods in ecology and
conservation and their potential for animal welfare monitoring. J R Soc Interface. 16 (155), 20190225.
http://doi.org/10.1098/rsif.2019.0225.
Moura, D.J., W.T. Silva, I.A. Naas, Y.A. Tolon, K.A.O. Lima, M.M. Vale, 2008. Real time computer
stress monitoring of piglets using vocalization analysis. Computers and Electronics in Agriculture. 64 (1), 11–
18. http://doi.org/10.1016/j.compag.2008.05.008.
Norton, T., C. Chen, M.L.V. Larsen, D. Berckmans, 2019. Review: Precision livestock farming: building
'digital representations' to bring the animals closer to the farmer. Animal. 13 (12), 3009–3017. http://doi.org/
10.1017/S175173111900199X.
Ntalampiras, S.A., L.A. Ludovico, G. Presti, E.P. Prato Previde, M. Battini, S. Cannas, C. Palestrini, S.
Mattiello, 2019. Automatic classification of cat vocalizations emitted in different contexts. Animals (Basel).
9 (8). http://doi.org/10.3390/ani9080543.
Pires, I.M., R. Santos, N. Pombo, N.M. Garcia, F. Florez-Revuelta, S. Spinsante, R. Goleva, E.
Zdravevski, 2018. Recognition of activities of daily living based on environmental analyses using audio
fingerprinting techniques: a systematic review. Sensors (Basel). 18 (1). http://doi.org/10.3390/s18010160.
Wang, X., X. Zhao, Y. He, K. Wang, 2019. Cough sound analysis to assess air quality in commercial
weaner barns. Computers and Electronics in Agriculture. 160, 8–13. http://doi.org/10.1016/j.compag.
2019.03.001.
∙ 325 ∙
A Deep Learning-Based Fusion Method of Infrared Thermography

and Visible Image for Pig Body Temperature Detection
Qiuju Xie a, Mengru Wu a, Muyu Yang a,Jun Bao b,*, Xin Li b, Honggui Liu b,
Haiming Yu a, Ping Zheng a
a
b
Abstract
Body temperature is a critical indicator for pig health. Traditional rectum temperature
measurement is inconvenient and brings some stress on pigs. It is necessary to explore a non-
invasive method for body temperature monitoring for health condition monitoring and early-stage
disease prevention on grouped pigs. In this study, an Infrared Thermography (IRT) technology is
employed for pig skin temperature monitoring. Temperatures were measured from seven different
parts on pig body skin from 16 group-housed pigs in different life cycle (i.e., gilts, pregnant sows,
lactating sows, and suckling piglets) to determine pig body surface for automatic temperature
measurement. A deep learning-based fusion method of IRT and visible image (VI) were
developed for different parts recognition and body skin temperature measurement. It was shown
that temperatures on pig nose was the lowest among all the measured body skin parts with a mean
of 33.6 ℃. Higher temperatures were found on the body surface of Tail Root (TR), Ear Root (ER),
anal (AN) and forehead (FR) with means of 36.2 ℃, 37.0 ℃, 37.1 ℃, and 36.0 ℃, respectively.
For lactating sows, a higher mean temperature of 37.8 ℃ on nipples was also found. The averaged
recognition accuracy of for VI on different body parts of forehead and ear root was 94.2%, the the
mean Pixels Accuracy (mPA) for IRT and VI fusion was 96.81%. Therefore, the proposed fusion
method could be used as a feasible way to monitor abnormal body temperatures and provide some
references for indoor environmental control and early-stage disease warning.
Keywords: Pig body temperature, deep learning, infrared thermography, fusion, indoor
temperature dependence.
1. Introduction
Pig body temperature and its changes are important indicators that reflect the health and
physiological response of pigs (Arfuso et al., 2016). The clinical measurement for pig body
temperature is rectal temperature with a digital contact thermometer (Smith et al., 2015; Giannetto
et al., 2020). Due to the inconvenience and time-lags of the invasive rectal temperature
measurements, some studies have been conducted to explore the non-invasive methods for body
surface temperature detection to reflect on pig body temperature and the health status, e.g., (Jiao
et al., 2016; Ricci et al., 2019; da Fonseca et al., 2020). Infrared thermography (IRT) is a non-
invasive method that uses specialized imaging cameras to capture heat radiation emitted from pig
body surface. It transforms the visualization and quantification of the temperature distribution into
a thermography. In the animal and welfare science fields, IRT has been getting more and more
attention for its convenience and efficiency for animal body temperature measurement to assess
the health status (Caldara et al., 2014; Foster and Ijichi, 2017), agonistic behavior (Boileau et al.,
2019), and stress (Ricci et al., 2019; da Fonseca et al., 2020). It is necessary to select a suitable
body surface area as an indicator to characterize the variations of pig body temperature for a non-
invasive automatic temperature measurement.
However, most IRT related studies mentioned above were focused on some points of
relationship between core temperature and skin temperature (Brown-Brandl et al., 2013; Cugmas
et al., 2020), body parts selection for measurement (Lu et al., 2018), skin temperature rhythm
∙ 326 ∙
(Giannetto et al., 2020). In intensive pig farms, it is a tough challenge on developing a smart and
instant detection method for body surface temperature using IRT as a tool to automatically
monitor the abnormal pig body temperature in group-housed pigs.
The objective of this study was to develop a fusion method for IRT and VI to assist automatic
pig body surface temperature detection for early-stage disease warning. On this purpose, firstly,
the regions of pig skin for IRT measurement were compared and determined; then, a deep learning
based automatic detection method for body parts was developed using VI; lastly, the extraction
and fusion algorithms based on matching of IRT and VI, and semantic segmentation were
proposed for pig body surface temperature extraction.
The data in this study were collected from two different feeding process (i.e., restricted fence
and group-housed) at Yabuli pig farm, Heilongjiang Province, China, using an infra thermal
camera (Model 287-L20, Fotric, Texsa, USA). The infra thermal camera captured the IRT and VI
at the same shot. The resolution of visible image and infrared thermography are 1600 × 1200 (W
× H) pixels, 512 × 384 (W × H) pixels, respectively.
Data collected from pigs raised in the restricted fences were used for exploring the key part
of body surface temperature measurement. They were collected from 16 pigs in different life cycle
(i.e., gilts, pregnant sows, lactating sows, and suckling piglets) on seven different pig body
surfaces area in restrict fences, twice a day around 8:00 am and 15:00 pm, on August 5 to 11,
2020. A total of 3322 images (1661 visible and 1661 thermography) was collected for
determination of body surface temperature measurement in this study (Figure 1).
A B C D E F G
Figure 1. Partial visible image and IRT in seven different body parts. A: Ear Root (ER); B:
Forehead (FR); C: Nipple (NP); D: Nose (NS); E: Anus (AN); F: Back (BK); G: Tail Root (TR).
There were totally 912 images (456 images for IRT, another 456 images for VI) were collected
from 100 group-housed pigs used for automatic detection and fusion of IRT and VI. Each pig
aged of 270 days and weighing about 200 kg. Among the 456 VI images, there were 299 images
for ear root, 227 images for forehead.
2.2. IRT and VI fusion method
Through the infrared thermal camera, the skin temperatures could be obtained directly.
However, the infrared thermography is not as clear as the visible picture to be recognized for
automatic identification of body parts. Therefore, a fusion method was developed to provide an
automatic skin temperature detection on body key parts using the deep learning-based algorithm,
which includes three steps: 1) body key parts recognition, 2) temperature extraction from IRT,
and 3) visible image and IRT fusion.
2.2.1. Automatic detection of pig body key parts
Automatic detection of pig body key parts was developed base on YOLOv4 model to achieve
an optimal and accuracy real-time object detection (Bochkovskiy et al., 2020) in the complex
environment of pig farm. The model consisted of four parts, i.e., Input, Backbones, Neck, and
Heads (Figure 2). The detailed recognition method can be found in the conference paper titled “A
method for pig individual identification and its body parts recognition” in this proceeding.
∙ 327 ∙
In the Input part, the visible images of forehead and ear root were divided into the size of 416
× 416 to input into the model.
The Backbone part was used to extract the features from the input images. It was constructed
by six blocks, i.e., one DarknetConv2D_BN_Mish convolution block and multiple resblock body
residual convolution blocks. From the backbone feature extraction, three effective features in size
of 52 × 52 × 256, 26 × 26 × 512, and 13 × 13 × 1024 were obtained for object detection.
In the Neck part, Spatial Pyramid Pooling (SPP) and Path Aggregation network (PANet) was
used to perform the feature enhancement. The SSP was used for the feature extraction and re-
aggregation to achieve a wide receptive field and significant contextual features in three pooling
scales of 13 × 13, 9 × 9, 5 × 5, and 1 × 1. The PANet achieved repeated feature extractions and
output feature maps at three scales using up-sampling and down-sampling for body key parts
prediction.
The Head was responsible for the prediction that detects three different scaled feature maps
output by the Neck. Three different scaled feature maps could be divided into grid matrixes of
{(76,76), (38,38), (19,19)}. Each grid represented a region that expressed by three boxes. Each
box had five parameters (x, y, w, h, c) to represent the horizontal coordinate of the central point,
the vertical coordinate of the central point, the length of the prediction box, the width of the
prediction box, and the confidence level, respectively. These parameters were continually adjusted
for an optimal object detection.
Input Backbone Neck Head
PANet
CSPDarknet53
Darknet(416*416) Predictions
Concat+Conv*5
Forehead
Resblock(208*208) Conv+UpSampling
Concat+Conv*5
Resblock(104*104
416*416
) Conv+UpSampling
Forehead
Resblock(52*52) DownloadSampling
SPP Concat+Conv*5
Resblock(26*26)
Ear root
DownloadSampling
Resblock(13*13) Concat+Conv*5
Ear root Concat+Conv*3

Conv*3
Figure 2. The structure of the body key parts detection.

2.2.2. Temperature extraction from IRT
The regions of forehead and ear root were detected in the visible images based on the
automatic detection model developed in Section 2.2.1. Temperatures on the forehead and ear root
were extracted from the infrared thermography (Figure 3). Due to different resolution and scene
parallax, there was an object position error in the same shot of the visible and thermal images.
Therefore, in order to get the same region in the infrared thermography corresponding to that in
the visible image, the two images of infrared thermal and visible must be matched before the
temperature extraction from the infrared thermography. To match the two images of VRT and VI,
the VI needed to be transformed by taking the coordinate position of (800, 600) pixels as a central
point in the VI, separating out an image in size of 640 × 480 pixels, scaling in proportion of 0.8
to get the image in the same viewpoint as the IRT in size of 521 × 384.
According to the trained detection model developed based on YOLOv4, the box-shaped
detection region in the visible image was obtained. At the same time, the four pairs of coordinates
of the box on the top, bottom, left and right were mapped to the infra thermography. The highest,
lowest and mean temperatures were acquired from the temperature matrix within this box.
∙ 328 ∙
Temperature
Detection matrix
model
▫▫▫▫▫▫ Max()
Images ▫▫▫▫▫▫ Min()
match ▫▫▫▫▫▫ Mean()
▫▫▫▫▫▫
The same region in
infrared thermograph
Figure 3. The flow chart for pig body temperature extraction.
2.2.3. Fusion
To achieve the fusion of IRT and VI, two steps were taken: 1) segmentation from the VI; 2)
IRT fusion with the background of the segmented VI.
Semantic segmentation model based on DeepLabv3+ was developed for body key parts
segmentation. It consisted of four parts: Input, Encoder, Decoder, and Predict. The Input was
responsible for image data input. However, because the size of matched VI is 521 × 384 pixels,
the edge needed to be filled with the grey bar to get an unfirmed images in size of 512 × 512
pixels. The Encoder was used for feature extraction from the input images. The Mobilenetv2 was
used for the main network to extract some effective features. After five concurrent for effective
feature layers output from DCCN, one stack was made. Then the channel number adjusted was
completed by 1 × 1 convolution. The Decoder was responsible for feature extraction on lower
sematic information. The Predict was used for pixel levelled classification between the
background and pigs.
After the sematic segmentation, the pixels in the VI were classified into two categories of
background and pigs. To achieve the fusion on IRT and VI, the RGB channels representing for
pigs in the VI were replaced by those in the IRT to generate a new image. Then a pixel levelled
fusion were performed with TIF algorithm using the new image and the VI as the input to complete
the IRT and VI fusion (Figure 4).
Source-images
Final
Base layers
base layers
Two-scale image Two-scale image

Image2 reconstruction
decomposition
Final detail
Detail layers
layers
Image1
Figure 4. The flow chart of the IRT and VI fusion.

2.2.4. Dataset division and parameters setting
The total 456 images of IRT and VI are divided into datasets of training, validation and testing
in a proportion of 8:1:1. During the whole training process of sematic segmentation, Adam was
used for the model optimization, and early stop strategy will be start if there is no variation in the
training loss within ten epochs. There were two stages on data training of sematic segmentation.
In the first stage, the weights were adjusted as the main network was frozen to speed up the
training, epochs were 50, initial learning rate was 0.0001, batch size was 8; in the second stage,
∙ 329 ∙
unfrozen the network for training, epochs were 150, initial learning rate was 0.00001, batch size
was 4.
3.1. Pig body surface temperature variation and its key part selection
The statistical data of temperatures on seven different body surface parts, i.e., TR, BK, NS,
ER, AN, FR, NP, which were collected using infra thermal camera, are shown in Table 1.
The temperatures of different body surface parts were very different. However, their variation
trends were consistent (Table 1). The maximum body surface temperature appeared in the AN
area (40.3 ℃), and the lowest value appears in the NS area (24.7 ℃). The averaged body surface
temperatures of AN, ER and FR were higher and the variation were smaller than those on other
parts. The temperatures on pig body surface in the morning were slightly lower than those on the
corresponding parts in the afternoon.
Table 1. Statistics of pig skin temperature on seven body parts.
Minimum Maximum
Body parts Mean±Std, °C
temperature, °C temperature, °C
TR 27.3 39.0 34.2±2.3
BK 27.7 38.1 34.2±2.1
NS 24.7 35.7 31.2±2.7
ER 31.5 37.7 35.4±1.3
AN 29.5 40.3 36.2±1.7
FR 30.2 36.9 33.9±1.7
NP 29.4 38.3 34.4±2.2
Based on the body surface temperature monitored at actual pig farms, this paper considers the
temperature and relative humidity in the house as influencing factors and the correlations (Pearson
correlation analysis at P<0.05) among temperatures on seven body surface parts were shown in
Table 2. Strong positive correlations between pig body surface temperatures and indoor air
temperatures existed. The highest correlation coefficient of 0.7472 was found in temperatures
between NP and indoor air; the weakest correlation existed between NS and indoor air
temperature, with a correlation coefficient of 0.4954. Also, strong negative correlations between
indoor air relative humidity and pig body surface temperatures existed. The indoor air relative
humidity had a greater influence on the temperature of FR, and the correlation coefficient was -
0.7275.
Table 2. Correlation coefficients analysis for pig skin temperature on different body parts, and
indoor air temperature (T) and relative humidity (RH).
Indoor T, Indoor RH, TR, BK, NS, ER, AN, FR, NP,
℃ % ℃ ℃ ℃ ℃ ℃ ℃ ℃
Indoor T, ℃ 1
Indoor RH, % -0.8211 1
TR, ℃ 0.7142 -0.6925 1
BK, ℃ 0.6406 -0.5065 0.5569 1
NS, ℃ 0.4954 -0.5003 0.5892 0.2542 1
ER, ℃ 0.5901 -0.5863 0.6851 0.6036 0.6821 1
AN, ℃ 0.6008 -0.6062 0.7482 0.3434 0.6240 0.5044 1
FR, ℃ 0.7115 -0.7275 0.5940 0.6541 0.7225 0.7610 0.4876 1
NP, ℃ 0.7472 -0.6028 0.6264 0.5252 0.5897 0.5316 0.5966 0.6813 1
In addition, it can be seen from Table 2 that temperatures on different parts of pig body surface
exhibits strong positive correlations. The correlation coefficient between FR and ER is the largest
of 0.7610, and the smallest correlation coefficient of 0.2542 is between BK and NS. Temperatures
∙ 330 ∙
of ER, FR, and AN have a strong positive correlation with other parts of the pig body surface, and
the averaged correlation coefficients are 0.6280, 0.6501 and 0.6333, respectively.
The temperature changes in ER, FR, and AN can well reflect the variations in pig body surface
temperature of the other parts. Therefore, the three areas of ER, FR and AN can be used as the
key temperature measurement areas of the pig body surface. At the same time, considering the
convenience of automatic data collection of pig body surface temperature in the intensive pig
farming environment, the two areas of ER and FR was selected in this paper as the automatic data
collection areas for pig body surface temperatures.
3.2. Visible and IRT fusion
3.2.1. Performances of the automatic recognition on pig body key parts
The YOLOv4 model for pig forehead and ear root detection exhibit optimal speed and
accuracy (Figure 5). Performances of the model were evaluated by indicators of precision (P),
precision and recall comprehensive evaluation index F1 value, average detection accuracy (AP)
and mean average accuracy (mAP) (Table 3). On the testing dataset, the optimal performances of
P, AP and F1 were achieved with higher values of 95.24%and 92.68%, 96.17% and 92.23%, 0.89
and 0.90, respectively, for forehead and ear root detection; the mAP of forehead and ear root
detection were 94.20%. Therefore, the model could provide an excellent basis for automatic pig
body parts detection to assist real-time body temperature monitoring.
Figure 5. Partial results of automatic detection for pig forehead and ear root.
Table 3. Performances of the automatic detection model.
Pig body parts P, % F1 AP, % mAP, %
Forehead 95.24 0.89 96.17
94.20
Ear root 92.68 0.90 92.23
3.2.2. Temperature extraction from the infrared thermography
The temperatures were extracted from the same region in the infrared thermograph as those
in the visible images. The maximum, minimum and mean temperatures were obtained using the
automatic detection and extraction. Compared with the temperatures that manually extracted from
the infrared thermography on pig forehead and ear root (Figure 6), the errors of maximum and
mean temperatures between the automatic and manual acquisition from the infrared thermography
were very small both on pig forehead and ear root. However, there were large errors of minimum
temperatures between the automatic and manual temperature acquisition, especially for the pig
ear root temperature (Figure 6 B). It was mainly caused by the region deviation of temperature
detecting and labeling, and some unexpected regions out of the pig body part were included.
Notably, the maximum temperature was creditable for pig body temperature detection using
infrared thermal camera.
3.2.3. Fusion analysis
Using two indicators of the mean Pixels Accuracy (mPA) and the mean Intersection of Union
(mIoU) for the fusion results evaluation, high fusion performances of mPA and mIoU were
achieved for both background and pig (Table 4). The maximum mPA and mIoU of 97.87% and
∙ 331 ∙
94.89 were obtained for the pig fusion; the averaged mIoU and mPA for the whole images were
93.98% and 96.81%, respectively. Therefore, the fusion method based on sematic segmentation
model could achieve an optimal fusion for IRT and VI.
45 50
Max-auto Max-manual Max-auto Max-manual
A Min-auto Min-manual 46 B Min-auto Min-manual
40 Mean-auto Mean-manual Mean-auto Mean-manual
42
Temperature (℃)
Temperature (℃)
35 38
34
30 30
26
25
22
20 18
1 3 5 7 9 11 13 15 17 19 21 23 25 1 5 9 13 17 21 25 29 33 37 41
Number Number
Figure 6. Temperatures comparison between the automatic and manual extraction on pig
forehead and ear root. A: forehead; B: ear root.
Table 4. Evaluation indicators of the fusion.
mIoU mPA Average mIoU Average mPA
Background, % 93.07 95.75
93.98 96.81
Pig, % 94.89 97.87
4. Conclusions
The following conclusions were drawn from this study:
(1) Pig body parts of NP and NS had the strongest and weakest correlation with the indoor air
temperature of 0.7472 and 0.4954, respectively. Moreover, temperatures on body surface were a
little bit lower in the morning than those in the afternoon.
(2) Pig body parts of forehead and ear root could be used as temperature detection region and
represent for body temperature variations.
(3) The maximum temperature was creditable for pig body temperature detection using the
infrared thermal camera.
(4) The integration methods of DeepLabv3+ and TIF could achieve an optimal fusion for IRT
and VI in the case of small dataset.
Acknowledgements
Laboratory of Swine Facilities Engineering, Ministry of Agriculture, China. Also, the authors
References
Arfuso, F., M. Rizzo, C. Giannetto, E. Giudice, F. Fazio, G. Piccione, 2016. Age-related changes of serum
mitochondrial uncoupling 1, rumen and rectal temperature in goats. Journal of Thermal Biology. 59, 47–51.
http://doi.org/10.1016/j.jtherbio.2016.05.002.
Bochkovskiy, A., C.-Y. Wang, H.-Y.M. Liao, 2020. Yolov4 Optimal speed and accuracy of object
detection. arXiv.org https://arxiv.org/abs/2004.10934.
Boileau, A., M. Farish, S.P. Turner, I. Camerlink, 2019. Infrared thermography of agonistic behaviour in
pigs. Physiol Behav. 210, 112637. http://doi.org/10.1016/j.physbeh.2019.112637.
Brown-Brandl, T.M., R.A. Eigenberg, J.L. Purswell, 2013. Using thermal imaging as a method of
investigating thermal thresholds in finishing pigs. Biosystems Engineering. 114 (3), 327–333. http://doi.org/
10.1016/j.biosystemseng.2012.11.015.
Caldara, F.R., I.D.A. Ns, R.G. Garcia, 2014. Infrared Thermal Image for Assessing Animal Health and
Welfare. Journal of Animal Behaviour and Biometeorology. 2 (3), 66–72.
∙ 332 ∙
Cugmas, B., P. Susteric, N.R. Gorenjec, T. Plavec, 2020. Comparison between rectal and body surface
temperature in dogs by the calibrated infrared thermometer. Vet Anim Sci. 9, 100120. http://doi.org/10.1016/
j.vas.2020.100120.
da Fonseca, F.N., J.M. Abe, I. de Alencar Nääs, A.F. da Silva Cordeiro, F.V. do Amaral, H.C. Ungaro,
2020. Automatic prediction of stress in piglets (Sus Scrofa) using infrared skin temperature. Computers and
Electronics in Agriculture. 168, 105148. http://doi.org/10.1016/j.compag.2019.105148.
Foster, S., C. Ijichi, 2017. The association between infrared thermal imagery of core eye temperature,
personality, age and housing in cats. Applied Animal Behaviour Science. 189, 79–84. http://doi.org/10.1016/
j.applanim.2017.01.004.
Giannetto, C., F. Arfuso, E. Giudice, M. Gianesella, F. Fazio, M. Panzera, G. Piccione, 2020. Infrared
methodologies for the assessment of skin temperature daily rhythm in two domestic mammalian species.
Journal of Thermal Biology. 92, 102677. http://doi.org/10.1016/j.jtherbio.2020.102677.
Jiao, L., D. Dong, X. Zhao, P. Han, 2016. Compensation method for the influence of angle of view on
animal temperature measurement using thermal imaging camera combined with depth image. Journal of
Thermal Biology. 62 (Pt A), 15–19. http://doi.org/10.1016/j.jtherbio.2016.07.021.
Lu, M., J. He, C. Chen, C. Okinda, M. Shen, L. Liu, W. Yao, T. Norton, D. Berckmans, 2018. An
automatic ear base temperature extraction method for top view piglet thermal image. Computers and
Electronics in Agriculture. 155, 339–347. http://doi.org/10.1016/j.compag.2018.10.030.
Ricci, G.D., K.O.d. Silva-Miranda, C.G. Titto, 2019. Infrared thermography as a non-invasive method for
the evaluation of heat stress in pigs kept in pens free of cages in the maternity. Computers and Electronics in
Agriculture. 157, 403–409. http://doi.org/10.1016/j.compag.2019.01.017.
Smith, V.A., V. Lamb, A.R. Mcbrearty, 2015. Comparison of axillary, tympanic membrane and rectal
temperature measurement in cats. Journal of Feline Medicine and Surgery. 17 (12), 1098612X14567550.
∙ 333 ∙
Real-time Monitoring of Exhaust Fan Operation Status in a

Livestock House Using Image Analysis
Luyu Ding a,b, Yang Lv a,b, Qifeng Li a,b,*, Ligen Yu a,b, Ronghua Gao a,b,
Weihong Ma a,b, Qinyang Yu a,b
a
National Engineering Research Center for Information Technology in Agriculture,
Beijing 100097, China
b
Beijing Academy of Agriculture and Forestry Sciences, Beijing 100097, China
* Corresponding author. Email: liqf@nercita.org.cn
Abstract
Real-time monitoring of fan operation status is helpful to supervise and regulate the
ventilation rate in a mechanical ventilated livestock house. This study tried to monitor the
operation state including fan position and ventilation levels of frequency adjustable fans in a pig
house with image processing and mathematical modelling. Hough transform was used to locate
the position of the fan, the motion of pixels obtained from the image of running fans were
calculated with dense optical flow. The dense optical flow data in the Cartesian coordinate system
was converted in polar coordinates, and the fan operation status was judged by the relationship
between the threshold value and the polar diameter. The numerical characteristics of dense optical
flow data were extracted and examined for curve fitting to obtain the airflow rate of exhaust fan
at different operating levels (15, 25, 35, and 50 Hz). Ventilation rates were effectively estimated
when fans ran under 70% of the maximum power (50 Hz; 18,000 m³ h-1). At this condition, the
average error of ventilation rate monitored using the method in this study was 1.84% (115.30 m³
h-1). The detection accuracy of the fan running status reached 99%, and all fans in sight of the
camera could be monitored at the same time. This study provided a new approach to monitor the
operation status of exhaust fans using video and image analysis.
Keywords: Variable frequency fan, ventilation rate, running state, image processing, dense
optical flow, Hough transform.
1. Introduction
Microclimate is complicated in a livestock or poultry house, and is affected by many factors
such as outside environment, animals raised in the house, the building envelope, and the operating
status of the environmental control equipment (Kim et al., 2019). Ventilation is an important
means to regulate the microclimate of livestock and poultry houses. Insufficient ventilation has a
major harmful effect on the growth of livestock and poultry including overheating and heat stress
for animals, the accumulation of water vapor and harmful gases in the house. This will slow down
the growth of animals, reduce the feed conversion rate and even cause death, decline the indoor
air quality and increase the risk of disease (Janni et al., 2018; Jeong et al., 2020). Under the large-
scale closed housing condition, large number of fans are needed in a single livestock or poultry
house to meet the ventilation needs of different seasons and to ensure that animals are in a suitable
environment (Ni et al., 2017). The variable frequency fans are increasingly used in livestock and
poultry houses because of their high energy efficiency ratio and adjustable air volume of a single
fan. Thus, it is important to monitor the operation status and ventilation rate of fans to ensure the
most basic ventilation status and the ventilation rate of the livestock house is within the suitable
range.
At present, the operating status of a fan is usually monitored by a sensor to detect the
perceptible signal generated when the fan is running to infer the operating status (Chen et al.,
2010; Abid et al., 2012). The operation status of the fan with signal sensing is high in investment
and maintenance cost during actual use (Ni et al., 2016). For example, it is necessary to modify
the original fan in the livestock or poultry house, considering whether the device can be
successfully installed in the original livestock and poultry house, there are certain requirements
∙ 334 ∙
for the design of the livestock or poultry house. For various mechanical detection devices, the
core component of detection is the internal sensor. The service life and maintenance cost of the
sensor have become the main factors affecting the use effect. Therefore, the accuracy, durability,
after-sales and other aspects of the equipment are relatively high. For the installed equipment,
safety must also be considered. When the fan is working, the fan blades need to have a very high
speed. As the speed increases, a strong airflow is generated in the fan duct (Curi et al., 2017),
sometimes mixed with some large particles such as feed scraps. The corrosion of the exposed
testing equipment may cause great safety risks. So does it when the parts of the testing equipment
fall off and become involved in the blades of the fan. Therefore, it is particularly important to
develop a safe, convenient, and low-cost monitoring method for fan operating status and variable
frequency fan ventilation.
This study developed a method to detect the operation status and ventilation levels of variable
frequency fans by image processing. With this method, videos can be record by a camera or
webcam away from fans and no sensors are needed to install in fans. All fans in the sight of camera
or webcam can be detected. This would lower the cost and safety risks in monitoring fan operation
and provide new ideas and new methods for the monitoring of wind turbines.
1.1. Equipment and measurement
Video and ventilation rates of fans were record or measured in a pig farm of Zhejiang Huateng
Animal Husbandry Co., Ltd. Fiber reinforced plastics (FRP) variable frequency fans were used
for ventilation and the parameters of the were as follows:
Table 1. Fan parameters.
Parameter Numerical value
Rated voltage 380 V
Rated frequency (maximum) 50 Hz
Propeller diameter 0.946 m
Rated speed 710 r min-1
Maximum ventilation 18,000 m³ h-1
Figure 1. Fan outline.

The video of a fan running at different ventilation levels of 15hz, 25hz, 35hz, and 45hz were
collected using a digital camera (ixus285HS, Canon, Japan) with a resolution of 1080 × 1920 (60
fps) at about 15 m away from the side wall where fans were installed. The elevation angle θ
between the camera and the fan was kept between 0–20° to prevent severe deformation of fans in
collected images and affecting subsequent processing. Two fans could be seen in the view of
camera and the acquisition period was 15 minutes for each ventilation level. Only one fan in the
view of camera was used to study the detecting method of ventilation rate with image analysis
while two fans in the view were used for simultaneous monitoring of multiple targets. The actual
ventilation rate of fan at different ventilation levels were sampled using an airflow rate detection
system (VTFS, Handan Changke Agricultural Machinery Co. LTD, China) mounted on the fans.
After collection, videos were processed with scientific computing libraries such as Scipy, Pandas,
Numpy, and Opencv of Python with image analysis technique.
∙ 335 ∙
1.2. Fan location and target area cut

The first thing to do was to locate the fan in the image. Considering the subsequent use of
dense optical flow algorithm, which can record the movement of objects in the image in two
consecutive frames, the position of the fan needed to be identified in the image to eliminate the
influence of other moving objects. The feature of the fan in appearance was its regular shape, so
that the position of the fan could be identified by using the Hough transform to recognizes the
circle of airduct. The Hough transform can find characters that can be represented by
mathematical expressions and detect possible defects in the characters. The distance between the
center of two fans in the image were calculated and other circles whose central distance less than
that was excluded in order to find the circle corresponding to the fan instead of other unwanted
circles. Combining with the size of the fan in image, the maximum and minimum radius of the
circle were defined to easily obtain the unique center coordinates and the radius of the circle
representing the fan in the image.
Afterwards, the circumscribed rectangle of the circle was cropped based on the center
coordinates and radius of the fan found by the Hough transform. This would eliminate the
interference outside the target area of fans and enhance the anti-interference and accuracy of the
subsequent calculation results.
1.3. Dense optical flow
Fan blades are difficult to track when rotating at a high speed. While a blade was always
rotating, the fan background normally remained static. Thus, the dense optical flow algorithm was
used in this paper to track the rotating of fan blades. The dense optical flow is an algorithm in
image analysis to estimate the movement of all pixels in two consecutive frames and to convert
the data form, which is shown as below:
𝑝11 ⋯ 𝑝1𝑛
P=[ ⋮ ⋱ ⋮ ] (1)
𝑝𝑚1 ⋯ 𝑝𝑚𝑛
where P is the matrix calculated by dense optical flow; 𝑝𝑚𝑛 is the pixel moving position of m
rows and n columns in the image; m is the maximum height of the image, n is the maximum width
of the image, all in pixels. The 𝑝𝑚𝑛 in the above formula is:
𝑝𝑚𝑛 = (𝑥𝑚𝑛 , 𝑦𝑚𝑛 ) (2)
where 𝑥𝑚𝑛 is the moving distance of the pixel in m rows and n columns in the x direction, and
𝑦𝑚𝑛 is the moving distance of the pixel in m rows and n columns in the y direction.
1.4. Identifying the operating status of the fan
The result of dense optical flow was divided into two parts, one part was the movement
distance in the x direction, and the other part was the movement distance in the y direction. The
two parts mentioned above can be understood as two characteristic dimensions, and the
recognition of the operating state of the fan was divided into two states: "yes" and "no". Under
the two feature dimensions, it was difficult to use simple methods to judge "yes" or "no" and data
simplifying was considered. The result of optical flow was converted into polar coordinates
(𝜌, 𝜃), where ρ represented the polar diameter and can be understood as the sum vector of the
displacement vectors in the two directions. θ represented the polar angle, which can be understood
as the angle of pixel movement. In the scenario of fan running status monitoring, fan blades always
move in a circle around the center of the circle, and the direction is less important than the angular
displacement. Therefore, the importance of polar angle can be ignored in the polar coordinate.
In the selected rectangular area, the data points whose polar diameter greater than the
threshold (𝜌∗ ) were determined and then calculating the percentage P of these data points in the
rectangular area:
𝐼(𝜌𝑖 >𝜌∗ )
𝑃= (𝑖 = 1,2,3, … , 𝑛 ∗ 𝑚) (3)
𝐼
∙ 336 ∙
where 𝜌𝑖 represents the polar diameter of the i-th data; 𝜌∗ represents the threshold of the polar
diameter; I represents the number of data points in the rectangular selection area.
The constituency of 𝜌∗ needs to be combined with the statistics of a large amount of empirical
data. 𝜌∗ can be understood as the boundary value that distinguishes whether the point was in
motion. If it is greater than 𝜌∗ , the point is a moving point; if it is less than 𝜌∗ , the point is not in
motion.
In order to prevent large fluctuations in forecasting, a time buffer window was established to
stabilize the forecast results. The result of P mentioned above was the ratio of the motion points
in the fan area at the current moment (𝑃𝑡 ), where t represents the time period at this moment. The
size of the time window is selected as T, and the final stable prediction result is:
1
𝑃̅ = ∑𝑇𝑖=1 𝑃𝑡−𝑖 (4)
𝑇
Let 𝑃0 be the threshold for judging whether the fan is running or not. If 𝑃̅ is greater than 𝑃0 ,
the fan was in the running state, otherwise not.
In this study, the accuracy was calculated as below to evaluate the index of efficiency of this
developed method to monitor the operating state of fans.
judge the correct number
Accuracy = (5)
total
1.5. Identify the ventilation levels of the variable frequency fan
The displacement vector matrix P of each frame pixel was obtained after applying the dense
optical flow algorithm. Numerical characteristics of the distribution of matrix P was calculated in
the x and y directions, respectively. A function fitting was performed to obtain the functional
relationship between ventilation rate and the numerical characteristics. Then, the ventilation rate
at different ventilation levels can be identified by this functional relationship.
𝑄𝑥 = 𝑓(𝐹𝑥 ) (6)
𝑄𝑦 = 𝑔(𝐹𝑦 ) (7)
𝑄 = 𝜆𝑄𝑥 + 𝜇𝑄𝑦 (8)
Among them, 𝐹𝑥 and 𝐹𝑦 are the numerical characteristics of the data distribution of matrix P
in the x and y directions; 𝑄𝑥 is the fitting function of 𝐹𝑥 with respect to the ventilation; 𝑄𝑦 is the
fitting function of 𝐹𝑦 with respect to the ventilation; 𝑄 is the reconciled ventilation, 𝜆 And 𝜇 are
the harmonic coefficients.
Establish a time window to stabilize forecast results. The result of Q mentioned above is the
ventilation rate at a certain moment (𝑄𝑡 ), where t represents the time period at the certain moment.
The window size of the time was selected as T, and the final stable prediction result is:
1
𝑄̅ = ∑𝑇𝑖=1 𝑄𝑡−𝑖 (9)
𝑇
Average absolute percentage error and residual value were adopted to evaluate the efficiency
of the developed method on monitoring the ventilation rate of variable frequency fan at different
ventilation levels.
𝑅esidual = |𝑦̂𝑖 − 𝑦𝑖 | (10)
𝑛
100% 𝑦̂𝑖 −𝑦𝑖
𝑀ean absolute percentage error = ∑ | | (11)
𝑛 𝑖=1 𝑦𝑖
The predicted value is 𝑦̂ = {𝑦̂1 , 𝑦̂2 , … , 𝑦̂𝑛 }, the true value is 𝑦 = {𝑦1 , 𝑦2 , … , 𝑦𝑛 }, the average
absolute percentage error range is [0,+∞), 0 % Means a perfect model, and more than 100% means
a poor quality model.
∙ 337 ∙

3.1. Monitoring of fan running status
Figure 2 demonstrates the visualization results in polar coordinate when the fan was running
at different ventilation levels from 15 hz to 45 hz. The black part of the image represents the
stationary background, and the gray part represents the motion of a fan. Results show that the
accuracy was 99% using the developed method in this study to monitor the fan running status.
Figure 2. From left to right are the recognized fan at 15 hz, 25 hz, 35 hz and 45 hz, respectively.
3.2. Identification of ventilation levels
It was found that the variance of the numerical characteristics of the dense optical flow data
showed an obvious stratification at different ventilation levels, and this stratification existed in
both x and y dimensions.
Figure 3. Variance scatter plot of data in x (left) and y (right) direction after dense optical flow
calculation for each frame of image under different ventilation levels.
According to the results found in Figure 3, the data were further simplified and their variances
under the same ventilation level was averaged for curve fitting to identify the ventilation rates at
different ventilation levels.
Table 2. Mean of variance of optical flow data of the fan operated at different ventilation level.
Rated Power Variance Variance Ventilation rate
frequency (hz) (w) (X-axis) (Y-axis) (m³ h-1)
15 250 0.8655 0.5079 2595
25 330 1.6086 0.8272 7454
35 430 2.3294 1.3991 13028
45 630 2.3364 1.3948 16556
Linear fitting shows a good performance when the fan was operated below 35 hz (P<0.05).
Figure 4 shows the results of linear curve fitting between ventilation rate and the averaged
∙ 338 ∙
variance of dense optical flow data calculated from each frame of image. It can be seen from the
fitting results that there is an obvious linear relationship between the ventilation rate and the
variance below 35 hz which was 70% of the maximum power. This suggests that the detection
capability of the method developed in this study reaches its upper detection limit when the fan ran
at the level greater than 70% of the maximum power. The monitored ventilation rate kept stable
when the ventilation level was greater than its upper limit of detection.
Figure 4. Variances in x (left) and y (right) direction after dense optical flow calculation for each
frame of image under different powers.
Figure 5 shows the scatter plot of the prediction results of ventilation rate identified by
developed method at different ventilation levels. The predicted ventilation rate was relatively
stable and the prediction showed a very good performance. The average residual error between
predicted ventilation rate and the measured ventilation rate was 115.30 m³ h-1, and the average
absolute percentage error was 1.84%.
Figure 5. Comparisons between predicted and measured ventilation rate.

3.3. Discussion of influencing factors
The main factors that affect the monitoring method developed in this study are as follows: the
distance between the image capture device and the fan, and the angle between the image capture
device and the fan. After further analysis, the external factors can be attributed to the pixels of the
image (resolution) and the degree of image deformation.
3.3.1. The impact of resolution
To study the effect of resolution on monitoring fan operation, the resolution of the same video
was changed and images with three different resolutions were obtained. The dense optical flow
algorithm was performance to these three kinds of data and their variance were calculated,
∙ 339 ∙
respectively. The calculation results of each frame of image were recorded and drawn into scatter
plots (Figure 6 – 8).
Figure 6. From left to right are the variance calculated from dense optical flow at different
ventilation levels in the x direction and y direction at a resolution of 640 × 480.
ventilation levels in the x direction and y direction at a resolution of 1280 × 720.
ventilation levels in the x direction and y direction at 1920 ×1080 resolution.
The stratification of variance at the resolution of the 640 × 480 was not good. There are many
intersecting parts, and it had a large degree of dispersion. This would result to a large error when
fitting curves to calculate the ventilation rate. However, this did not affect the determination of
whether the fan was running or not. A relatively good stratification of variance was observed at
∙ 340 ∙
the resolution of 1280 × 720 or 1920 × 1080. There was no significant difference between the
resolution of 1280 × 720 and 1920 × 1080 when using the fitted curve to calculate the ventilation
rate (P>0.05). Hence, an image acquisition device with a relatively high resolution should be
selected for identifying the fan ventilation levels and the lowest resolution of 1280 × 720 was
recommended.
3.2.2. The influence of the image deformation of the fan
The placement angle of the image acquisition device affects the deformation of the fan in the
image. Image deformation affected the circular positioning in locating the fan and cutting the
target area in the very beginning. This could be solved by adjusting the threshold of the
accumulator at the center of the Hough transform detection stage. A small threshold can detect a
more severely deformed circle.
4. Conclusions
A high accuracy can be obtained to monitor the operation status and ventilation levels of a
variable frequency fan based on image processing method developed in this study. After
performing the dense optical flow algorithm on the images of a running fan, the stratification of
variances was found for each frame of the dense optical flow data at different ventilation levels.
A good linear curve fitting was found between the variance of and the corresponding ventilation
rate and the detection for ventilation levels can be realized by using this function. However, this
method will reach the limit of monitoring when the fan ran at the ventilation levels over 70% of
the maximum power.
Acknowledgements
This research was funded by the Tianjin Science and Technology Planning Project (funding
20YFZCSN00220), Beijing Science and Technology Planning Project (Z191100004019007).
References
Abid, M., H. Trabelsi, M. Taktak, J. Antoni, J.M. Ville, T. Fakhfakh, M. Haddar, 2012. Tonal prediction
of a faulty axial fan. Applied Acoustics. 73 (10), 1022–1028.
Chen, Y., J.Q. Ni, C.A. Diehl, A.J. Heber, B.W. Bogan, L.L. Chai, 2010. Large Scale Application of
Vibration Sensors for Fan Monitoring at Commercial Layer Hen Houses. Sensors. 10 (12), 11590–11604.
http://doi.org/10.3390/s101211590.
Curi, T., D.J. de Moura, J.M. Massari, M. Mesquita, D.F. Pereira, 2017. Computational fluid dynamics
(CFD) application for ventilation studies in broiler houses. Engenharia Agricola. 37 (1), 1–12.
Fournel, S., A.N. Rousseau, B. Laberge, 2017. Rethinking environment control strategy of confined
animal housing systems through precision livestock farming. Biosystems Engineering. 155, 96–123.
http://doi.org/10.1016/j.biosystemseng.2016.12.005.
Janni, K.A., M. Torremorell, L.D. Jacobson, C. Alonso, B.P. Hetchler, 2018. Modeling airborne virus
concentrations in filtered swine barns with negative-pressure ventilation systems. Transactions of the Asabe.
61 (3), 1089–1099. http://doi.org/10.13031/trans.12561.
Jeong, M.G., D. Rathnayake, H.S. Mun, M.A. Dilawar, K.W. Park, S.R. Lee, C.J. Yang, 2020. Effect of
a Sustainable Air Heat Pump System on Energy Efficiency, Housing Environment, and Productivity Traits in
a Pig Farm. Sustainability. 12 (22). http://doi.org/10.3390/su12229772.
Kim, R.W., J.G. Kim, I.B. Lee, U.H. Yeo, S.Y. Lee, 2019. Development of a VR simulator for educating
CFD-computed internal environment of piglet house. Biosystems Engineering. 188, 243–264. http://doi.org/
10.1016/j.biosystemseng.2019.10.024.
Ni, J.Q., D. Kaelin, I.M. Lopes, S.L. Liu, C.A. Diehl, C. Zong, 2016. Design and performance of a direct
and continuous ventilation measurement system for variable-speed pit fans in a pig building. Biosystems
Engineering. 147, 151–161.
Ni, J.Q., S.L. Liu, I.M. Lopes, Q.J. Xie, P. Zheng, C.A. Diehl, 2017. Monitoring, modeling, and
characterizing single-speed ventilation fans for an animal building. Building and Environment. 118, 225–233.
http://doi.org/10.1016/j.buildenv.2017.03.036.
Rojano, F., P.E. Bournet, M. Hassouna, P. Robin, M. Kacira, C.Y. Choi, 2019. Modelling the impact of
air discharges caused by natural ventilation in a poultry house. Biosystems Engineering. 180, 168–181.
http://doi.org/10.1016/j.biosystemseng.2019.02.001.
∙ 341 ∙
Feasibility Analysis of Ammonia Monitoring in Piggery

Environmental Monitoring System Based on Internet of Things
Jinrui Zhang a, Hua Wang a, Deyong She b, Xiaoliu Xue a,c, Zhonghong Wu a,
Jijun Liu a, Meizhi Wang a,*
a
College of Animal Science and Technology, China Agricultural University, Beijing 100193, China
b
College of Biological and Pharmaceutical Engineering, West Anhui University,
Luan, Anhui 237012, China
c
* Corresponding author. Email: meizhiwang@cau.edu.cn
Abstract
With the rapid development of mobile Internet, big data and artificial intelligence, piggery
environmental monitoring system based on Internet of Things has become an important research
field. Nowadays, studies on the system mainly focus on the construction and optimization of
system software, neglecting accuracy and necessity analysis of long-term monitoring of basic
environmental data. In this test, the numerical variation characteristics of two-month operation of
electrochemical ammonia sensors (EAS) in piggeries in typical seasons (summer and winter) were
analyzed. The standard method HJ533-2009 was used to monitor the ammonia to evaluate the
accuracy of two brands of EAS based on a remote environmental monitoring system. The results
showed that during the operation of 17 EAS, 16 EAS showed obviously downward trend, and 6
EAS showed zero value frequently and continuously within 24 hours. On day 21 of the operation,
73% of the sensors had decay rates >65%. The average relative errors of the two brands of EAS
were -71% and -73% in the system, indicating low accuracy of EAS in the test. According to the
available research reports, the ranges of ammonia concentrations in large-scale pig farms are 0.6–
14.3 mg m-3 in summer and 2.0–20.6 mg m-3 in winter and are lower than the upper limit of various
standards. Moreover, the actual cost of EAS is high and can be 3.074 million yuan per year for a
basic 2600-sow farm. Therefore, it is not necessary to use EAS to monitor ammonia in large-scale
commercial pig farms. However, to meet the needs of modern intelligent pig farms for multi-
index environmental monitoring, it is urgent to develop ammonia sensors with low cost and high
reliability.
Keywords: Remote monitoring, IoT, ammonia concentration, electrochemical sensor, pig
building
1. Introduction
With the rapid development of new generation information technology such as mobile
Internet, big data, cloud computing and artificial intelligence, precision feeding with "Internet of
things" has become an important development trend of animal husbandry in the future (Xiong et
al., 2015). At present, the researches of remote environmental monitoring mainly focus on the
construction and optimization of environmental control system, such as the construction of
environmental control model (Christian et al., 2017), algorithm optimization (Yang et al., 2019),
data analysis method (Duan et al., 2017), data storage (Ji et al., 2016) and improvement of data
transmission form (Wang et al., 2020), etc., and the running time of the system is short, mostly 3-
7 days, thus the accuracy and necessity of long-term monitoring of some basic environmental data
are ignored. At present, the common methods for monitoring ammonia concentration are
electrochemical method, photoacoustic infrared method and spectroscopy. From the perspective
of cost, electrochemical ammonia sensors (EAS) are mainly used in large-scale commercial pig
farms (Li et al., 2019). Compared with sensors based on photoacoustic infrared and spectral
principles, the short-term measurement data of EAS may be accurate, but it is easy to be affected
by other gases in long-term use (Gao et al., 2018). Although Zhuang et al. (2019) and Tan et al.
(2020) have developed a real-time online monitoring device for ammonia concentration in pig
∙ 342 ∙
house based on tunable absorption spectroscopy technology, it has not been put into operation.
In the test, based on two-month operation effect of electrochemical ammonia sensors (EAS)
in pig houses, the numerical variation characteristics of EAS in typical seasons (summer and
winter) were analyzed. Then, based on the piggery environmental monitoring system, the national
standard HJ 533-2009 method was used at the same time to monitor the ammonia concentration
in the pig house, and evaluate the accuracy of two kinds of EAS values in the remote
environmental control system. Finally, based on the research of ammonia concentration in
commercial scale piggery in recent five years and the total investment cost of EAS, the feasibility
of ammonia monitoring in remote environmental monitoring system of piggery at present is
comprehensively analyzed, which provides reference for the optimization of remote
environmental monitoring system of piggery.
2.1. Long-term operation effect of EAS in typical seasons
2.1.1. Pig houses
From August 24, 2017 to September 22, 2017, the summer experiment was conducted in a
farrowing house in Jilin Province. The size of building was 28.8 × 10.3 × 2.5 m (L × W × H). The
wet curtain cooling system was used for ventilation and cooling. Sows were fed automatically for
three times a day. The manure was stored in deep-pit.
From December 21, 2017 to January 25, 2018, the winter experiment was carried out in a
farrowing house in Beijing. The pig house was north-south oriented, with the length and width of
15.5 m and 7.6 m, respectively. The house was well organized by mechanical ventilation and
artificial feeding. The manure was cleaned by labors every morning
2.1.2. Environmental data measurement
In summer, five measuring points (numbered A1-A5) were evenly selected for EAS (model
GT-903, Korno, Shenzhen, China). These EAS had measuring range of 0–37 mg m-3, and with
accuracy ≤ ±3%, working temperature and humidity -30–60 ℃ and ≤ 95% RH, respectively, and
error ≤ ±1%). The EAS installation height was 1.5 m from the ground. The EAS measurement
data was recorded every 30 min. To ensure all facilities of the pig house run well during the
experiment and no extreme situation occurred to interfere with the life of the sensor, temperature,
humidity and CO2 concentration in the house were monitored at the same time. Four measuring
points were evenly selected for the instruments, and the height was 1.5 m from the ground.
Temperature and humidity were measured using automatic recorders (Model 179A-TH, Apresys,
America). The measurement range was -40–100 ℃ and 0%–100% RH, respectively, with
accuracy of ±0.2 ℃ and ±1.8% RH. The concentration of carbon dioxide was measured using
automatic recorders, with the measurement range of 0–10,000 mg m-3 and the accuracy of ±150
mg m-3 (Model EZY-1S, Tianjianhuayi, Beijing, China).
In winter, six measuring points were selected for EAS in the house, and the height was 0.7m
from the ground, numbered B1-B6, all of which were new sensors. At the same time, three
measuring points were selected in the house to monitor temperature, humidity and CO 2
concentration. All of the instruments were the same as the summer experiment.
2.2. Numerical verification of ammonia sensor in remote environmental monitoring system
2.2.1. Pig houses
From January 19 to January 28, 2021, a fattening house in Beijing was selected, with the
building size of 24.6 × 10.0 × 2.5 m. The airtightness of the house was poor; therefore, the fans
were combined with the doors and windows for ventilation. The machine was used to clean the
manure every day. The house was equipped with remote environmental monitoring system to
ensure the real-time online work of the system.
∙ 343 ∙
2.2.2. Remote environmental monitoring system

In the experiment, the system included modules such as environmental information
acquisition, environmental control, upper computer control and wireless communication. After
the environmental information was collected by the environmental sensors, the environmental
information was converted and processed by fusion, and the monitored environmental data was
transmitted to the controller through the wireless or wired data receiving / transmitting control
module to control the equipment in the house automatically (Xie et al., 2017). On the other hand,
the environmental data was sent to the upper computer, and the dynamic display and storage of
environmental data were realized through the monitoring software, so as to realize the automation
of environmental remote control (Gao et al., 2019). The overall framework of the system is shown
in Figure 1a, and the structural diagram of ammonia instrument is shown in Figure 1b.
a. System framework b. Ammonia instrument
Figure 1. Schematic diagram of remote environmental control system.

In this experiment, the wired networking mode of RS-485 bus was adopted, and the GPRS
wireless remote transmission mode was adopted in the remote transmission layer to realize
convenient remote transmission of piggery environmental information.
2.2.3. Environmental data measurement
There are two brands of EAS (6 sets) in the system, numbered C1-C3 and D1-D3. Calibration
test would be conducted before the experiment. Meanwhile, the concentration of ammonia at the
same measuring points was determined by the national standard (HJ533-2009, 2010). Based on
the daily feeding time, the time of cleaning the manure and indoor temperature distribution,
sampling and analysis were conducted three times every day at 9:00, 15:00 and 20:00, for 2 hours
each time.
The temperature and humidity sensors were used in the above-mentioned system. The
concentration of carbon dioxide was measured by carbon dioxide instruments (Model WEZY-1,
Tianjianhuayi, Beijing, China), with the measurement range of 0–982 1 mg m-3 and the accuracy
of ±98 mg m-3.
2.3. Data analysis
The data of 17 EAS were analyzed by Statistical Product and Service Solutions (SPSS 25.0).
The single factor variance analysis was used to determine the difference significance of P<0.05
and P<0.01. P<0.05 was significant difference, P<0.01 was extremely significant difference.
3.1. Basic environmental conditions of experimental pig houses
Table 1 shows the temperature, humidity and carbon dioxide concentration in the three houses.
Temperature in the farrowing house should be 18–22 ℃, the maximum should not exceed 27 ℃,
the relative humidity should be 60%–70%, and the carbon dioxide concentration should be less
than 5893 mg m-3. The temperature in the fattening house should be 15–23 ℃, the minimum
should not be lower than 13 ℃, the relative humidity should be 65%–75%, and the CO2
concentration should be less than 5893 mg m-3 (Wang et al., 2016). The results showed that the
∙ 344 ∙
average temperature, relative humidity and CO2 concentration in the houses met the standard,
indicating that the environmental control equipment in the houses worked well.
Table 1. Temperature, relative humidity, and CO2 concentration in test pig houses.
Temperature (℃) Humidity (%) CO2 (ppm)
Average Range Average Range Average Range
2017-Summer 23.0 19.0-27.4 60.7 26.3-84.0 943.5 423.4-3350.5
2017-Winter 20.1 14.6-24.7 66.7 50.3-81.2 2232.7 1197.3-3995.3
2021-Winter 15.02 11.7-17.8 69.0 47.5-80.4 2246.3 1114.5-3051.5
3.2. Long-term operation effect of EAS in typical seasons
The new EAS with the same model and different production batches were selected uniformly
in two test houses from August 24 to September 22, 2017 and December 20, 2017 to January 26,
2018 respectively. The results are shown in Figure 2.
Figure 2. Long term effect of the electrochemical ammonia sensors. Top left and top right: Test
in summer. Bottom: Test in winter.
Results showed that the values of the five sensors in the summer test fluctuated greatly from
August 24th to September 1st. Except for A3, the other four sensors showed zero values for many
times, and the difference of the five sensors was extremely significant (P<0.01). It can be seen
from Figure. 2b that the values of the six sensors had an obvious abnormal downward trend over
time, and the values of B5, B2 and B3 tended to zero at 1:00 on January 11st, 10:00 on January
19th and 12:00 on January 27, 2018 respectively, indicating poor numerical stability of the
sensors. And the difference of 6 sensors was very significant (P<0.01).
∙ 345 ∙
3.3. Numerical verification of ammonia sensor in remote environmental monitoring system

The changes of EAS values of the two brands with time are shown in Figure 3. Figures 3a and
3b show that the values of the sensors showed zero values many times, and had a certain
attenuation trend over time, which was consistent with the above results.
a Sensor C b Sensor D
Figure 3. Numerical variation of EAS in remote environmental control system.
To further explore the accuracy of EAS, the national standard method (HJ 533-2009) was
used to detect the ammonia concentrations in the same period, and the concentrations measured
by the national standard method were taken as the true value to calculate the relative error of the
sensors. As shown in Figure 4 and Table 2 below.
30 Senseor C
25
EAS (mg m-3)
20
15
10
0
0 10 20 30
Standard method (mg m-3)
Figure 4. Accuracy analysis of the EAS.

Table 2. Relative errors of the EAS.
C1 C2 C3 D1 D2 D3
Ave ± SD (%) -81.6±16.9 -83.7±11.0 -91.2±9.1 -71.2±18.9 -47.2±17.8 -55.5±15.0
Max (%) -100.0 -100.0 -100.0 -80.3 -63.7 -77.9
Min (%) -58.3 -67.2 -75.9 -53.0 -26.4 -39.8
Figure 4 shows that the ammonia concentrations measured by the EAS were seriously
deviated from the values measured with the national standard method. It can be seen from Table
2 that the relative errors of the two EAS were very large and were all less than -40%; The average
relative error was -85.5% and -58.0%, respectively. The accuracy of model C was lower.
Table 3 below shows the performance analysis results of four models of EAS (17 sets) in this
study.
∙ 346 ∙
Table 3. Analysis of application effect of the EAS.

Reduction Reduction Reduction
First time Relative
Test rate after 7 rate after 14 rate after 21
Number zero is error
period days of use days of use days of use
displayed (%)
(%) (%) (%)
A1 21 d 9.85 d 2.5 100 100
A2 29 d 10 h -131.3 35.4 100
A3 21 d none -11.8 -35.3 -59.4
A4 30 d 11 h -62.5 3.6 40.6
A5 30 d 14 h -31.4 -28.8 100
B1 36 d 31 d 77.7 99.3 99.9 —
B2 37 d 36 d 20.5 39.1 67.7
B3 24 d none 7.0 24.6 43.3
B4 24 d none 50.5 69.4 79.2
B5 38 d 17 d 18.3 79.8 98.7
B6 24 d 19 d 50.8 89.3 98.0
C1 9d 18 h 80.3 -81.6
C2 9d 16 h 75.1 -83.7
C3 9d 16 h 68.7 -91.2
—
D1 10 d 8d 89.1 -71.2
D2 10 d none 56.8 -47.2
D3 10 d none 44.1 -55.5
Note: 1. The time when zero value appears for the first time refers to the time when the sensor
displays zero value for at least 5 hours; 2. The reduction rate of sensors are based on the average
values of the first day of operation, and the negative values indicate increases.
In the three groups of experiments, the number of pigs, the manure treatment modes,
ventilation and feed composition were basically constant, and there was no significant difference
in temperature, relative humidity and carbon dioxide concentration over time. However, it can be
seen from the table that during the operation of 17 sensors in pig houses, 12 sensors showed zero
value for many times, and 6 of them showed zero value continuously within 24 hours after use.
In addition, the values of 16 sensors showed abnormal attenuation, and 13 of them showed the
law of continuous attenuation, which showed that the value reduction rate continued to increase
over days. On day 21 of the operation, 73% of the sensors had decay rates >65%. In the system,
the average relative errors of the two sensors were -71% and -73%, respectively, indicating the
accuracy of EAS was low in the test.
The liquid electrolyte is mostly in common use for electrochemical sensors, and the
evaporation or pollution of electrolyte can also lead to sensor signal degradation (Pu et al., 2019).
There is a certain concentration of ammonia in the piggery. When the ammonia is continuously
monitored by the EAS, the working electrode of the electrochemical sensor will continue to carry
out chemical reaction, resulting in the rapid consumption of electrolyte. In addition, there are high
concentrations of dust, carbon dioxide and other interfering gases in the piggery, which will
further damage the life of EAS. Therefore, when the EAS is continuously used in the piggery, the
data will gradually decline. However, the current research on the performance optimization of gas
sensors, such as response time, measurement accuracy and the elimination and supplement of gas
interference substances, are mostly concentrated in the chemical and automotive fields, and less
specially developed for the livestock environment (Jie et al., 2015).
3.4. Necessity analysis of ammonia monitoring in remote environmental monitoring system
3.4.1. Ammonia concentration in pig houses
Ammonia is one of the important pollutants in pig houses. The growth performance of pigs
∙ 347 ∙
will be affected by long term stimulation of high concentration ammonia. According to the reports
in recent five years, the ammonia concentration range in the summer of large-scale pig farm is
0.6-14.3 mg m-3 (Jo et al., 2020; Jeppsson et al., 2021), and the winter range is 2.0-20.6 mg m-3
(Xie et al., 2017; Ehab et al., 2016). Ni et al. (2010) used INNOVA to monitor the environment
of three large-scale pig houses for two years. The average ammonia concentration in the house
was (10.7 ± 5.0) mg m-3. The concentration ranges in summer and winter were 0.9 – 15.2 mg m-3
and 5.9–29.8 mg m-3, respectively. Most of the ammonia concentration in the house is lower than
that which affects the performance of pigs (18.9 mg m-3) (Philippe et al., 2011), and lower than
the national standard (NH3 concentration in farrowing and nursery pig house is 20 mg m-3, the
concentration of NH3 in gestation and fattening pigs is 25 mg m-3) and the upper limit of CIGR
concentration (15 mg m-3). But even in summer, the concentration of ammonia in pig house is
higher than the optimal limit of ammonia concentration (8 mg m-3) (Wang et al., 2016). Therefore,
it is necessary to monitor ammonia concentration in pig house and feed back to control system in
real time, which means that we need to develop new sensors.
3.4.2. Cost analysis of EAS
Considering the cost of the system, EAS is used in the research on ammonia monitoring of
remote environmental control system in pig houses. If the system is used in a large-scale pig farm
for a long time, because of the poor stability and sensitivity of the long-term operation of EAS in
the house, it is necessary to calibrate it with standard gas many times and replace the sensor in
time. In addition, the internal core components of EAS on the market are mostly imported from
abroad, and the price range is 800-5000 yuan / piece. If there are at least two ammonia measuring
points in the pig house, the sensors are replaced every 15 days, and calibrated every 7 days, a pig
house unit costs about 53,000 yuan a year (assuming that the EAS is 1000 yuan / piece, and the
calibration cost is 100 yuan / time). For 2600 basic sow farms (30 units in fattening house, 14
units in nursery house, 12 units in farrowing house), The investment of EAS in one year should
be at least 53,000 × 58 = 3.074 million yuan. From the perspective of cost, the investment cost of
EAS is too high, and the cost performance of using EAS is low.
4. Conclusions
 During the operation of 17 EAS, 16 EAS showed obviously downward trend, and 6 EAS
showed zero value frequently and continuously within 24 hours. On day 21 of the system
operation, 72.7% of the sensors had decay rates >65%. In the evaluation of numerical accuracy
of ammonia sensor, the average relative errors of the two kinds of sensors in the first three
days were -85.5% and -58.0%, respectively, indicating that the numerical accuracy, sensitivity
and service life of EAS were all low in this test.
 The range of ammonia concentration in large-scale pig farm is 0.6–14.3 mg m-3 in summer
and 2.0–20.6 mg m-3 in winter, which is basically lower than the upper limit of various
standards, but higher than the optimal limit of ammonia concentration in pig house
(8 mg m-3). It is necessary to monitor the ammonia concentration in pig house and feed it back
to the control system in real time.
 At present, the technology of ammonia sensors is not mature, and the investment cost is too
high due to frequent replacement. It is not recommended to use EAS to monitor ammonia in
the remote environmental system of commercial pig farms. However, to meet the needs of
modern intelligent pig farms for multi-index environmental monitoring, it is urgent to develop
low-cost and high reliability ammonia sensors.
References
Christian, M., V. Ronja, and S. Gunther, 2017. Simulation of indoor climate of Livestock Building to
Assess Adaptive Measures for Reducing Heat Stress Due to Climate Change. Proceeding of International
Symposium of Animal Environment and Welfare. 53(3): 322–328.
∙ 348 ∙
Duan, Q.-L., X.-Y. Xiao, Y.-R. Liu, L. Zhuang, and K. Wang, 2017. Data fusion method of livestock and
poultry breeding internet of things based on improved support function. Transactions of the CSAE. 33(S1):
239–245.
Ehab, M., H. Richard, D. Bernd, A.E. Ghaly, and B. Wolfgang, 2016. Evaluation of two indoor air
pollution abatement techniques in forced-ventilation fattening pig barns. Atmospheric Pollution Research.
8(3).
Gao, Y., Y.-P. Diao, C.-G. Lin, Y.-X. Liu, C.-M. Guo, M.-G. Lei, Y. Tong, and X. Li, 2018. Monitoring
and analysis of thermal environment and harmful gases in mechanically ventilated multistory pig buildings.
Transactions of the CSAE. 34(4): 239–247.
Gao, Y, Z.-H. Chen, Y. Wang, X.-P. Li, and J.-L. Guo, 2019. Design for pig breeding chamber under
multiple environment variable control and analysis of internal flow field. Transactions of the CSAE. 35(2):
203–212.
HJ533-2009 Air and exhaust gas-determination of ammonia-Nesslers reagent spectrophotometry.
Beijing: China Environmental Press, 2010 (in Chinese).
Jeppsson, K. H., A.C. Olsson, A. Nasirahmadi, 2021. Cooling growing/finishing pigs with showers in the
slatted area: Effect on animal occupation area, pen fouling and ammonia emission. Livestock Science. 243:
104377.
Ji, B., W. Zheng, R.S. Gates, and A.R. Green, 2016. Design and performance evaluation of the upgraded
portable monitoring unit for air quality in animal housing. Computers and Electronics in Agriculture. 124(12):
132–140.
Jie, D.-F, J.-M. Pan, and Y.-B. Ying, 2015. Advances in methods and instruments for determining
concentration of gaseous air pollutants in large-scaled livestock farms. Transactions of the CSAE. 31(1): 236–
246.
Jo, G., T. Ha, Y.N. Jang, O. Hwang, S. Seo, S.E. Woo, S. Lee, D. Kim, and M. Jung, 2020. Ammonia
emission characteristics of a mechanically ventilated swine finishing facility in Korea. Atmosphere.
11(10) :1088.
Li, H.-L,2018. Reasearch on environmental monitoring methods of livestock and poultry facilities. Hefei:
College of Control Science and Engineering, University of Science and Technology of China.
Ni, J.-Q., A.J. Heber,.2010. An on-site computer system for comprehensive agricultural air quality
research. Computers and Electronics in Agriculture. 71(1): 38–49.
Philippe, F.X., M. Laitat, J. Wavreille, N. Bartiaux-Thill, B. Nicks, and J.F. Cabaraux, 2011. Ammonia
and greenhouse gas emission from group-housed gestating sows depends on floor type. Agriculture,
Ecosystems and Environment. 140(3):498–505.
Pu, S.-H., T.-T. Li, H. Wang, Y. Jian, Z.-H. Liu, D.-B. Long, and R.-D Ruan, 2019. Monitoring
technology of air pollution generated by animal feeding operations: A review. Journal of Agro-Environment
Science. 38(11):2439–2448.
Tan, H.-Q., X.-A. Li, and Z.-M. Ai, 2020. Detection of ammonia concentration in livestock poultry houses
based ontunable diode laser absorption spectroscopy. Transactions of the CSAE. 36(13):186–194.
Wang, M.-Z., Z.-H. Wu, J.-J. Liu, and W.-Y. Zhao, 2016. Review on environmental parameters of
harmful gases and particles in pig house. Swine Industry Science. 33(04):94-97.
Wang, S.-J, 2020. Research and Application of Wireless Monitoring System for Animal housing
Environment based on IoT. Beijing: College of Water Resources and Civil Engineer, China Agricultural
University.
Xie, Q.-J, Z.-B. Su, J.-Q., Ni, and P. Zheng, 2017. Control system design and control strategy of multiple
environmental factors in confined swine building. Transactions of the CSAE. 33(6): 163–170.
Xiong, B.-H., Z.-G. Yang, L. Yang, and X.-H. Pan, 2015. Review on application of Internet of Things
technology in animal husbandry in China. Transactions of the CSAE. 237-246.
Yang, L, C.-H. Liu, Y.-C. Guo, H. Deng, D.-L. Li, and Q.-L. Duan, 2019. Prediction of ammonia
concentration in fattening piggery based on EMD-LSTM. Transactions of the Chinese Society for Agricultural
Machinery. 50(S1):353–360.
Zhuang, S.-J., V.O. Philippe, V. Jürgen, S. Bart, and D. Peter, 2019. Evaluation of a cost-effective
ammonia monitoring system for continuous real-time concentration measurements in a fattening pig barn.
Sensors. 19(17): 3669.
∙ 349 ∙
Automatic Recognition for Rumination Behavior of Dairy Cows

Based on Machine Learning
Liwei Wang a, Qiuju Xie a,*, Yidan Xu b
a
College of Computer Science and Software, College of Big Data, Zhaoqing University,
Zhaoqing, Guangdong 526061, China
b
College of Mechanical and Automotive Engineering, Zhaoqing University,
Zhaoqing, Guangdong 526061, China
* Corresponding author. E-mail: xqj197610@163.com
Abstract
Rumination is a critical physical digestion process of dairy cows that reflects the health status
and estrus and is most concerned by the breeding production. To improve the automatic
monitoring of dairy cows’ rumination behavior and achieve the modern dairy cow production, it
is necessary to realize the intelligent recognition of rumination behavior. Method for sound and
activity monitoring were employed for rumination behavior recognition in this study. A novel
multi-source parameter portable neck-wearing monitoring device was developed to achieve a
hardware support and automatic algorithms of K-means and Support Vector Machine (SVM) were
used for intelligent classification of rumination behavior. Data of 36 representative days from six
Chinese Holstein cows were selected. It was shown that the frequency range of rumination sound
was between 1.2 kHz and 3.2 kHz; the axis parameters of 3D acceleration sensor for rumination
activity were from 0.15 g to 0.93 g for x-axis, from -0.14 g to 0.29 g for y-axis, and from 0.13 g
to 1.04 g for z-axis. The recognition rate of rumination behavior based on K-means combined
with SVM identification model and SVM was between 83.3% and 93.2%. The results showed a
good recognition and a feasible reference for automatic rumination monitoring in the complex
environment of large-scale dairy cows breeding farm.
Keywords: Sound detection, rumination recognition, SVM, K-means, MFCC.
1. Introduction
Rumination is one of the important physical digestion processes of dairy cows. A complete
rumination includes four stages: regurgitation, mastication, mixing saliva, and swallowing
(Beauchemin et al., 1994). The rumination behavior of dairy cows is affected by some feeding
conditions of time, quantities, and type. The mastication number after rumination is related with
the fiber content of the feed (Reiter et al., 2018), so the ration formula and nutrition problem of
dairy cow can be known by monitoring rumination behavior. The rumination time that reflects on
individual health status of dairy cows was significantly reduced when the cows are affected by
some diseases related to uterine, and metabolic and subclinical ketosis (Kaufman et al., 2016;
Stangaferro et al., 2016a; Stangaferro et al., 2016b). Also, the rumination time is an important
indicator to measure the estrus. About 94% of dairy cows in estrus had a rumination time reduction
(Reith and Hoy, 2012), up to about 17% (Reith et al., 2014).
Traditional methods for the rumination measurement are direct observation and manual
counting. In recent years, some intelligent monitoring methods have emerged, especially the
wearable devices and automatic classification and recognition algorithms for livestock behavior
monitoring.
The wearable hardware devices are the basis of intelligent monitoring. Some devices were
developed and worn around the neck, forehead, or jaw of a dairy cow for monitoring rumination,
feed intake and behavior (Chelotti et al., 2016; Shen et al., 2019). Examples of such devices
included the intelligent wearable monitoring device designed for animal welfare and the economic
efficiency of livestock farm (Borchers and Bewley, 2015), the HR-LD neck strap and ear tags of
SCR in Israel (Anonymous, 2020), and the common wearable devices for monitoring rumination
behavior using sensors of nose-band (Ruuska et al., 2016; Shen et al., 2020), sound (Milone et al.,
∙ 350 ∙
2012; Navon et al., 2013a; b), and acceleration (Tani et al., 2013; Giovanetti et al., 2016).
Classification and recognition algorithms are necessary for automatic livestock behavior
monitoring. In the past decade, some research on animal behavior recognition and classification
algorithms were conducted, they included Neural Net (NN), K-nearest neighbor (KNN), K-means
(KM), Support Vector Machine (SVM) and Hidden Markov Machine (HMM) (Milone et al.,
2012b; Nadimi et al., 2012). Among these recognition algorithms, the acoustical signal and
activity pattern classification and recognition method including Decision Trees (DS), SVM and
HMM (Chung et al., 2013; Poblete and Muñoz, 2016; Barker et al., 2018) were most popular
approaches for animal behavior classification and recognition. However, the rumination behavior
recognition for dairy cows using sound signal combined with activity information has not been
reported in the literature up to date.
The objective of this paper is to recognize rumination behavior through a cluster analysis
combined with classification methods based on a multi-source-parameter wearable monitoring
device to provide theoretical support for judging the health status and estrus of individual dairy
cows to improve the management and reproduction efficiency of the dairy farm.
2.1. Data monitoring
The experiment was carried out from July 7, 2017 to July 7, 2018 for more than 10 months at
Yinlang Dairy Farm, Daqing, China (Figure 1). Due to the comparisons, data completeness, health
status of the cows, 36 days of representative data for six adult Holstein cows were selected for the
comparison experiment. The cows had an average age of 4.5 ± 0.5 years and average weight of
about 480 kg. Two of the cows were in a state of estrus and four others were in un-estrus states
for six days. All six cows were kept in the same barn and feed with Leymus chinensis (fodder)
every morning and evening in the barn.
Figure 1. The Yinlang Dairy Farm.

The rumination monitoring wearable devices for the dairy cows consisted of five modules: a
single chip microcomputer (CPU), a sound and activity monitoring module, a storage module (a
SD card) and a wireless transmission module, to perform tasks of rumination information
management, collection, storage, and transmission, respectively. To get clear signals of sound and
activity, the monitoring device was worn on the neck of the cow with an adjustable necklace
(Figure 2).
Figure 2. The multi-source monitoring device on a cow neck.
∙ 351 ∙
The ruminations of two dairy cows (1# and 6#) in estrus period and four cows (2#, 3#, 4#, and
5#) un-estrus were monitored in the experiments. Dairy cows 1# and 2# were monitored for pre-
estrus, main-estrus, and the end-estrus. Each dairy cow was individually monitored for six days.
2.2. Automatic recognition Algorithm
After data monitoring, the Mel Frequency Cepstrum Coefficient (MFCC) extracted sound
characteristic parameters based on the particularity of sound signals. The rumination activity
characteristics were not extracted twice. The KM classified similar samples into the same category
based on similarity measure, but the disadvantage of the KM was that it could not automatically
determine the clustering number. Therefore, based on data analysis, this paper improved the KM
algorithm for the sample marking stage of classification. The Support Vector Machine (SVM)
was used for further classification and recognition of rumination signals. The automatic
recognition process of rumination behavior data is shown in Figure 3.
Initialize rumination data
Use MFCC to complete

feature extraction
Previous classification Improved classification

experiment experiment
The K-means algorithm is

Build SVM module
applied to get the cluster in
the training set
The classification of
rumination behavior was Run the SVM classification
done by SVM model algorithm
End of the classification
Figure 3. The automatic recognition process in the study.

2.2.1. K-means Algorithm
The KM algorithm is a clustering method used to divide the dataset into different categories
through iterative process (Dunnachie et al., 2010). The KM has the characteristics of fast
clustering speed and can cluster raw data with high dimensions without being given sample type
in advance. Sum of the squared errors (SSE) criterion function was used to evaluate clustering
performance in the algorithm, shown in Eqs. (1) and (2), respectively.
c
E 
2
pn  mi cn (1)
i =1 pn  i
In Eq. (1), E represents error in clustering, pn represents individual sample, i represents a

cluster, mi represents mass, c represents categorical variable, i represents sample example, n
represents the number of individual sample。
The sample value is:
pn   xi  xi 1  i  1,2, , n
2
(2)
After rumination feature extracting, the KM clustering algorithm was used. The characteristics
of the rumination activity was divided into two clusters. The larger cluster was selected from the
∙ 352 ∙
results obtained. The KM continued the splitting operation until K clusters, which represented the
characteristics of rumination features, were obtained and the algorithm terminated. The specific
steps are given as follows:
1) The rumination and feeding behavior of dairy cows was classified into regurgitation,
swallowing, and chewing according to the characteristic analyses, and three centers of mass
mi 
k
 k  1,2, , n  were set;
2) Each pn was assigned to the cluster that was closest to the center of mass mi  k  ;
 k 1
3) The new center of mass mi  k 1 and the error value E were calculated after
distribution;
4) Repeat steps 2) and 3) until k reached the maximum iterations;
5) The cluster number of k was obtained.
2.2.2. SVM Classification Algorithm
To realize a better recognition of rumination from the similarity between the activities of
swallowing and regurgitation, SVM was used to perform a successive classification, which had
been clustered by the KM clustering algorithm. Each class was divided into two subclasses, which
could be repeatedly performed until all subclasses were separated.
The classification of dairy cow activity characteristics by SVM is shown in Figure 4. Because
there could be some swallowing in the process of feeding and rumination during the process of
rumination monitoring that could affect the detection of rumination signals, it was found necessary
to classify activity characteristics of the regurgitation and the swallowing of feeding and
rumination for dairy cows to reduce the complexity of the rumination activity identification. The
rumination classification was completed using SVM 1 and SVM 2. The sample size of total
experiment data was 2724. The ratio of training data and testing data sets were 2:1.
Behavior
characteristics
SVM 1
Swallowing Regurgitation
SVM 2
Feeding Rumination
swallowing swallowing
Figure 4. Activity characteristics classification of dairy cows.

The multidimensional data were classified using rumination classification algorithm of SVM
until all classes could not be subdivided and the final classification results were obtained. The
model training and testing were executed using the function of svmtrain with MATLAB 2014a
(Martiskainen et al., 2009).
2.3. Extraction and analysis of sound feature
The Mel Frequency Cepstrum Coefficient (MFCC) is mostly common used for voice
recognition. The characteristic parameters of rumination sound were extracted using MFCC
(Boujelben and Bahoura, 2018). It was critical for acoustic recognition to perform an endpoint
∙ 353 ∙
detection after the sound preprocessing. So, the zero-crossing rate and short time energy were
used as dual-threshold to remove the silent segment of rumination sound after noise reduction.
The process of determining the starting point of rumination sound was a double threshold endpoint
detection of the chewing and swallowing sound after the signal preloading, windowing, and
framing.
Windowing sound signal：
S  n   s  n     n  (5)
In Eq. (5),   n  represents window function, weighted by Hamming function.
Short time energy：
En   m  0  X n  m 
N 1 2
(6)
 
The specific relationship between Mel frequency and actual frequency：
Mel  f   2595lg 1  f 700  (7)
In Eq. (7), f represents frequency, Hz。

2.4. Analysis of sound and activity data
After the extraction and analysis of sound feature of dairy cows, the SVM was used to classify
the sound and activity of the swallowing, rumination and chewing behavior of dairy cows. The
sound recognition process is shown in Figure 5.
Endpoint MFCC feature

Audio Preprocessin
detection selection and extraction
input g
Prognostic classification Determine classification

models
Figure 5. Sound recognition process.
The recognition of rumination activity of dairy cows was completed using algorithms after
the preliminary selection. Two recognition methods of SVM and KM combined with SVM (KM
+ SVM) were compared.
As the number of cluster centers (K) increased, more details of sample divisions became
available. The aggregation of each cluster increased gradually and the sum of the squared errors
(SSE) decreased. The initial cluster centers (K) in this experiment were set from 1 to 5, and the
corresponding SSE were 1167.6, 725.4, 416.7, 334.3 and 262.5, respectively.
When the initial cluster center K was 3, the three features, i.e., swallowing, rumination and
chewing, were clustered and marked. The three features were then divided into five features after
KM repeated clustering. The five features, i.e., rumination-swallowing, feeding-swallowing,
regurgitation, swallowing and chewing, were finally classified.
The classification effect was better than SVM, not only the selection of support vectors
becoming clearer, but also the categories was more obvious (Figure 6). After the KM was added
and the clustering and classification were repeated, a more accurate division of samples was
achieved, and the recognition rate was increased. A higher rumination recognition rate of more
than 90% was also achieved. The comparison of accuracy rate between KM + SVM method and
SVM are listed in Table 1.
∙ 354 ∙
Swallowing Regurditation Support Vectors Swallowing(classified) Regurgitation(classified)

4.5 4.5
4 4
3.5 3.5
Point
Point
3
3
2.5
2.5
2
2
4 5 6 7 8 4 5 6 7 8
Point Point
Figure 6. Classification results of using KM + SVM. Left: Training; Right: Classification.
Table 1. Comparison of training classification accuracy.
Accuracy (%)
Activity Feature Training data Testing data
KM + SVM SVM
Rumination-swallowing 352 176 93.2 75.8
Feeding-swallowing 238 118 85.6 73.6
Regurgitation 806 404 92.1 90.1
Swallowing 590 294 87.3 82.5
Chewing 420 210 83.3 81.3
The minimum and maximum accuracies of SVM were 75.8% and 90.1%, and those of KM +
SVM were 83.3% and 93.2%, respectively. Obviously, the accuracies of KM + SVM were higher
than those of SVM. Especially, in the classification of rumination-swallowing, the maximum
accuracy of KM + SVM was 93.2%, which was 17.4% higher than that of SVM.
The total rumination time of six dairy cows in each day is shown in Figure 7. The normal
daily total rumination times for un-estrus cows (2#, 3#, 4#, and 5#) were 449 ± 25 min. However,
from day 2 the rumination times for the estrus cows (1# and 6#) decreased dramatically, and on
day 4 they were declined into the lowest value of 384 min and 364 min, respectively. They then
gradually increased after day 4 and reached a normal value of 434 min from day 5. The results
showed that the rumination time of dairy cow during estrus was significantly lower than un-estrus
period. This coincided with the law of estrus for dairy cows (Reith and Hoy, 2012).
Figure 7. The daily rumination time for six dairy cows.
∙ 355 ∙
4. Conclusions
(1) The multisource device developed with sound and activity sensing was a feasible way for
rumination monitoring.
(2) The MFCC and SVM were useful methods for rumination sound and activity feature
extraction and pre-classification.
(3) The combination of KM algorithm and SVM algorithm was an effective method to solve
the classification problem in high dimensional data sets and achieved the classification accuracy
rate between 83.3% and 93.2% for rumination behavior of dairy cow and provided an automatic
reference for dairy cow production to judge the health and estrus status.
Acknowledgements
This work was supported by the Plan of Rural Revitalization Project of Guangdong Province,
China (2020ZDZX1041), Research on Key Technologies of Rumination Behavior Perception and
Intelligent Discrimination of Dairy Cows Based on Machine Learning (QN202130). The authors
acknowledge the support by the workers and managers at the Yinlang Dairy Farm, Daqing, China
during the rumination measurement, and the help by all others for this study.
References
Anonymous, 2020. Scrdairy, http://www.scrdairy.com/cow-intelligence/sensetime.html.
Barker, Z.E., Vázquez, J.D., Codling, E.A., Bell, N.J., Hodges, H.R., Croft, D.P., Amory, J.R., 2018. Use
of novel sensors combining local positioning and acceleration to measure feeding behavior differences
associated with lameness in dairy cattle. Journal of Dairy Science 101(7), 6310–6321.
Beauchemin, K.A., Mcallister, T.A., Dong, Y., Farr, B.I., Cheng, K.J., 1994. Effects of mastication on
digestion of whole cereal grains by cattle. Journal of Animal Science 72, 236–246.
Borchers, M.R., Bewley, J.M., 2015. An assessment of producer precision dairy farming technology use,
prepurchase considerations, and usefulness. Journal of Dairy Science 98, 4198–4205.
Boujelben, O., Bahoura, M., 2018. Efficient FPGA-based Architecture of an Automatic Wheeze Detector
using a Combination of MFCC and SVM Algorithms. Journal of Systems Architecture 88, 54–64.
Chelotti, J.O., Vanrell, S.R., Milone, D.H., Utsumi, S.A., Galli, J.R., Rufiner, H.L., Giovanini, L.L., 2016.
A real-time algorithm for acoustic monitoring of ingestive behavior of grazing cattle. Computers and
Electronics in Agriculture 127, 64–75.
Chung, Y., Lee, J., Oh, S., Park, D., Chang, H.H., Kim, S., 2013. Automatic Detection of Cow’s Oestrus
in Audio Surveillance System. Asian-australasian Journal of Animal Sciences 26, 1030–1037.
Dunnachie, M.E., Shields, P.W., Crawford, D.H., Davies, M., 2010. Filler models for Automatic Speech
Recognition created from Hidden Markov Models using the K-Means algorithm, Signal Processing
Conference, 2009 European, pp. 544–548.
Gáspárdy, A., Efrat, G., Bajcsy, A.C., Fekete, S.G., 2015. Electronic monitoring of rumination activity as
an indicator of health status and production traits in high-yielding dairy cows. Acta Veterinaria Hungarica 62,
452–462.
Giovanetti, V., Decandia, M., Molle, G., Acciaro, M., Mameli, M., Cabiddu, A., Cossu, R., Serra, M.G.,
Manca, C., Rassu, S.P.G., 2016. Automatic classification system for grazing, ruminating and resting behaviour
of dairy sheep using a tri-axial accelerometer. Livestock Science 196, 42–48.
Kaufman, E.I., Leblanc, S.J., Mcbride, B.W., Duffield, T.F., Devries, T.J., 2016. Association of
rumination time with subclinical ketosis in transition dairy cows. Journal of Dairy Science 99, 5604–5618.
Martiskainen, P., Järvinen, M., Skön, J.-P., Tiirikainen, J., Kolehmainen, M., Mononen, J., 2009. Cow
behaviour pattern recognition using a three-dimensional accelerometer and support vector machines. Applied
Animal Behaviour Science 119, 32–38.
Milone, D.H., Galli, J.R., Cangiano, C.A., Rufiner, H.L., Laca, E.A., 2012. Automatic recognition of
ingestive sounds of cattle based on hidden Markov models. Computers and Electronics in Agriculture 87, 51–
55.
Nadimi, E.S., Blanes-Vidal, V., Christensen, S., 2012. Monitoring and classifying animal behavior using
ZigBee-based mobile ad hoc wireless sensor networks and artificial neural networks. Computers & Electronics
in Agriculture 82, 44–54.
Navon, S., Mizrach, A., Hetzroni, A., Ungar, E.D., 2013a. Automatic recognition of jaw movements in
free-ranging cattle, goats and sheep, using acoustic monitoring. Biosystems Engineering 114, 474–483.
∙ 356 ∙
Navon, S., Mizrach, A., Hetzroni, A., Ungar, E.D., 2013b. Automatic recognition of jaw movements in
free-ranging cattle, goats and sheep, using acoustic monitoring ☆. Biosystems Engineering 114, 474–483.
Poblete, C.M., Muñoz, N.A.R., 2016. Detector of walking state for cows through support vector machine
based on accelerometers and gyroscopes installed on ear-tags and collars. IEEE International Conference on
Automatica, October 19–21, 2016. University of Talca, Curicó, Chile. Page: 1–6.
Reiter, S., Sattlecker, G., Lidauer, L., Kickinger, F., M, Ã.h., Auer, W., Schweinzer, V., Klein-Jöbstl, D.,
Drillich, M., Iwersen, M., 2018. Evaluation of an ear-tag-based accelerometer for monitoring rumination in
dairy cows. Journal of Dairy Science 101, 3398–3411.
Reith, S., Brandt, H., Hoy, S., 2014. Simultaneous analysis of activity and rumination time, based on
collar-mounted sensor technology, of dairy cows over the peri-estrus period. Livestock Science 170, 219–227.
Reith, S., Hoy, S., 2012. Relationship between daily rumination time and estrus of dairy cows. Journal of
Dairy Science 95, 6416–6420.
Ruuska, S., Kajava, S., Mughal, M., Zehner, N., Mononen, J., 2016. Validation of a pressure sensor-based
system for measuring eating, rumination and drinking behaviour of dairy cattle. Applied Animal Behaviour
Science 174, 19–23.
Shen, W., Cheng, F., Zhang, Y., Wei, X., Fu, Q., Zhang, Y., 2019. Automatic recognition of ingestive-
related behaviors of dairy cows based on triaxial acceleration. Information Processing in Agriculture 7 (3),
427–443.
Shen, W., Zhang, A., Zhang, Y., Wei, X., Sun, J., 2020. Rumination recognition method of dairy cows
based on the change of noseband pressure. Information Processing in Agriculture 7 (4), 479–490.
Stangaferro, M.L., Wijma, R., Caixeta, L.S., Al-Abri, M.A., Giordano, J.O., 2016a. Use of rumination
and activity monitoring for the identification of dairy cows with health disorders: Part I. Metabolic and
digestive disorders. Journal of Dairy Science 99, 7395–7410.
Stangaferro, M.L., Wijma, R., Caixeta, L.S., Alabri, M.A., Giordano, J.O., 2016b. Use of rumination and
activity monitoring for the identification of dairy cows with health disorders: Part II. Mastitis. Journal of Dairy
Science 99, 7411–7421.
Tani, Y., Yokota, Y., Yayota, M., Ohtani, S., 2013. Automatic recognition and classification of cattle
chewing activity by an acoustic monitoring method with a single-axis acceleration sensor. Computers and
Electronics in Agriculture 92, 54–65.
Yin, L., Liu, C.X., Hong, T.S., Zhou, H.E., Kaehsiang, K., 2010. Design of system for monitoring dairy
cattle's behavioral features based on wireless sensor networks. Transactions of the Chinese Society of
Agricultural Engineering 26, 203–208.
Zhang, X., Du, J., Fan, C., Liu, D., Fang, J., Wang, L., 2017. A wireless sensor monitoring node based
on automatic tracking solar-powered panel for paddy field environment. IEEE Internet of Things Journal 4
(5), 1304–1311.
∙ 357 ∙
Why Precision Livestock Farming Can Generate a

More Sustainable Livestock Sector
Daniel Berckmans a,b
a
Department of Biosystems, Catholic University Leuven, Belgium
b
Department of Biosystems Engineering and Soil Science, University of Tennessee, USA
Email: daniel.berckmans@kuleuven.be
Abstract
On 25 September 2015, the 193 Member States of the United Nations adopted the 2030 UN
Agenda for Sustainable Development. The agenda has generated 17 aspirational objectives, the
Sustainable Development Goals (SDGs). About half of them are related to livestock.
Livestock is positioned at the interface of the world’s human and natural systems, which has
been at the basis of the Global Agenda’s understanding of sustainability. In agriculture we use
natural resources (land, water, biodiversity, forests, fish, nutrients, and energy) and environmental
services and transform them into agricultural products (food, feed, fiber, fuel). Agriculture does
not only serve immediate needs, but also provides economic and social services (food security,
economic growth, poverty reduction, health, and social and cultural value). Because of the
increasing population growth and welfare, the conventional meat industry cannot follow
consumer-increasing demands worldwide. Expectations show an increase of over 60 % by 2050
in the worldwide demand for animal products. In Europe we see a trend that makes choices and
realization of European policy in relation to livestock, animal welfare and environmental impact
even more challenging.
This paper aims to think about how the livestock sector can act to fulfill the worldwide
increasing demand of animal products and at the same time become more sustainable.
Keywords: Precision livestock farming, sustainability, metabolic energy balance.
1. Introduction
On 25 September 2015, the 193 Member States of the United Nations adopted the 2030 UN
Agenda for Sustainable Development. The agenda had 17 aspirational objectives, the Sustainable
Development Goals (SDGs), which will serve governments, international organizations, the
private sector, and civil society to shape the path of human advancement over the next 15 years
(United Nations, 2015). Five out of 17 SDGs received significantly higher priority from all
perspectives: SDG 1 (no poverty), SDG 2 (zero hunger), SDG 13 (climate action), SDG 15 (life
on land) and SDG 17 (partnership for the goals). When asked as individuals, UN partners also
prioritized SDG 12 (responsible consumption and production). Livestock is positioned at the
interface of the world’s human and natural systems, which has been at the basis of the Global
Agenda’s understanding of sustainability. Five out of the 17 SDGs are strongly connected to
livestock production and received significantly higher priority from all perspectives in relation to
livestock: SDG 1 (no poverty), SDG 2 (zero hunger), SDG 13 (climate action), SDG 15 (life on
land) and SDG 17 (partnership for the goals) (United Nations, 2015; Keeling et al., 2019).
Due to the increasing population growth and welfare, the conventional meat industry cannot
follow consumer-increasing demands worldwide, without wondering how to become more
sustainable and to take measures in relation to climate change, animal welfare and animal health
in relation to human health.
The objective of this paper is to show how a large-scale field implementing of an appropriate
technology can contribute to an answer for the worldwide increasing demand for animal products
in a more sustainable way. We believe that, although still a lot of work needs to be done, the
science is available behind a very useful technology: “Precision Livestock Farming”.
∙ 358 ∙

Based upon scientific literature, our experience and logic deduction we try to explain the
reason why we believe that the technology of Precision Livestock Farming (PLF) can have a high
impact on sustainability. The main reason is that PLF can help to optimize the most essential part
and process of the livestock sector: the animal that turns feed energy into animal product.
2.1. What is Precision Livestock Farming?
Precision livestock farming is a tool for farmers to help them managing their animals based
upon continuous automatic real-time monitoring and control of production/reproduction, animal
health and welfare, and the environmental impact of livestock production. The PLF brings
automated monitoring by using cameras, microphones, sensors etc. The PLF assumes that
continuous monitoring 24/7 of animals will enable farmers to detect and control the health and
welfare status of their animals at any given time during the production process. Ultimately, an
animal in good health and enjoying high welfare might provide a better guarantee of product
quality in the long term with higher efficiency. Nowadays, the farmer is using modern
technologies to measure different parameters on the farm, such as ventilation rate, feed supply
and heating/cooling inputs, but few of the tools used at large scale up to now have directly focused
on monitoring the most important part: the animal. That is what PLF is about (Berckmans, 2014).
2.2. Problems to be solved
Domestication of animals by human started over 13,000 years ago and can only be considered
as a revolutionary step in the evolvement of mankind (McHugo et al., 2019; Childe, 1928). Today,
the position of livestock within the worldwide food production system is challenged. Concerns
are expressed on several issues, notably: Lack of increasing efficiency in animal production,
especially beef; criticism on the guarantee of animal welfare; environmental pollution by intensive
animal production, absence of and so far no successful identification of appropriate technologies
to improve this; risk for disease transfer from livestock to humans. More questions are raised by
the society such as: Are lab-meat and plant-based protein a better technology and possibly a threat
for livestock producers? Can we reduce food loss and create food waste recovery as animal feed?
Do livestock of the rich eat the grain of the poor?
Excessive consumption of animal protein is considered to exert pressure on the global food
system. As a result, in Europe consumption trends indicate increase in alternative plant-based
diets, which could change the future balance of protein consumption. A gradual shift towards
alternatives, including novel plant-based meat alternative products, and in the future lab-grown
meat, could have a significant impact on agricultural production in the EU, over the next 10–20
years (European Commission, 2019). The worldwide COVID19 pandemic does not help to
improve people’s perception regarding the contribution of the livestock sector in the food chain
due to the risk for pandemics and disease transfer. Health issues and diseases can decrease
production efficiency of livestock by up to 33% (McHugo et al., 2019). The trend towards more
intensification of livestock farming systems increases productivity but can also have adverse
effects on animal health and welfare and might increase the risk of rapid and far-reaching disease
outbreaks. We will not be able to reduce the demand for animal products in countries and regions
that finally have the economic position to ask for more animal products. Bu it is also not realistic
to think that the increasing demand for animal products of around 60 %, can be solved by getting
more animals. When considering the efficiency of producing animal protein in relation to
environmental impact, then namely beef is under criticism for producing methane, carbon dioxide,
and ammonia.
Another claim is that animals are considered to consume food that could potentially be eaten
by people. The FAO has published that 86% of livestock feed is not suitable for human
consumption. If not consumed by livestock, crop residues and by-products could quickly become
an environmental burden as the human populations grows and consumes more and more processed
food. Grains account for 13% of the global livestock dry matter intake.
∙ 359 ∙
It is obvious that the livestock sector and related stakeholders must come up with solutions
and answers to create a more sustainable protein production by livestock (Norton et al., 2019).
We believe that, although yet a lot of work needs to be done, the science is available behind a
very useful technology: “Precision Livestock Farming”.
3. Results: How can the use of PLF create a more sustainable livestock production?
Most of the increase in animal products in the last decades has come from genetic selection
and an increase in animal numbers rather than from an improved management to increase
individual-animal productivity. When considering what we are doing for many decades in the
livestock sector, the whole biological process in the animals is about turning feed energy into
animal product. This can be rationally summarized in its essential components of a metabolic
energy balance as follows:
Total Energy intake = Energy for basal metabolism (including Energy for immune system) +
Energy for movement + Energy for thermal + Energy for mental +
Energy for production + losses. (1)
It is all about managing and supporting the animal to help it transfer the energy, delivered
from feed intake, in the most efficient way in the Energy production term (meat, eggs, milk and
fiber). It is not a matter of focusing on a specific single term in this equation [Eq. (1)]. It is not a
matter of optimizing OR this term OR another one in the energy equation. It is about working on
AND this term AND all other terms to maximize the animal product with minimal Total Energy
intake and minimal losses that end up in the environment. In other words, we need “more with
less”: more animal products from less feed intake and consequently with less manure production,
less emissions, less infections, and less losses.
The PLF has already developed several techniques and offers many more possibilities to work
on each of the terms of the energy equation [Eq. (1)].
3.1. Total Energy intake
There are several opportunities to improve feed management. There is a difference between
the amount of feed fed to animals and the real feed intake. In research stations feed delivery or
feed intake are measured since many years, but the used technology is too expensive for large
scale field implementations in livestock houses and for sure not affordable for small family farms.
PLF has shown techniques for very accurate measurements of feed intake by broilers using sound
analysis with a simple microphone integrated in the feeder pan. The algorithm detects the number
of peckings when feed is taken in. The average feed intake per pecking was found as 0.025 g. The
amount of feed intake and the pecking frequency were highly correlated (R 2= 0.985) (Aydin et
al., 2014). The amount of feed consumed was measured with mean absolute errors and mean
square errors (MAE and MSE) of 0.127 kg, and 0.034 kg, respectively (Ran Bezena et al., 2020).
For grazing cattle an initial pasture intake algorithm was established for time spent grazing:
pasture dry matter intake: DMI (kg day (-1)) = -4.13 + 2.325 x, where x stands for the hours spent
grazing (P = 0.010, R2 = 0.53, Relative Standard Deviation = 1.65 kg DM day (Greenwood et al.,
2017). There are several unused opportunities to improve feed management such as feeding in
meal packages, adapt the composition of the feed, appetite regulation, post-ruminal nutrient
absorption, and cellular energetics and metabolism to the efficiency of feed utilization in cattle
(Kenny et al., 2018).
3.2. Energy for basal metabolism
The basal metabolism is the absolute minimal amount of energy that an individual body needs
to keep all organs functioning. This is a totally individual characteristic depending on species,
age, weight, health condition, production phase, etc. So far, the individual feeding strategy is
depending on the energy production term in milking cow, but the real estimation of the basal
metabolism term is not yet realized, although feasible to do. Genetic selection has accomplished
huge advantage in working on this energy basal metabolism and has consequently played a huge
∙ 360 ∙
role in producing more animal product. More efficiency in terms of feed conversion, growth rate,
less infections will result in less manure, less emissions.
Today, PLF technologies can offer the advantage of collecting real-time data from animals to
study and to manage the efficiency of phenotypes in the field at very large scale. If someone is
for example interested in analyzing the aggressive behavior of a specific species and want to
collect data from e.g., 500,000 animals, PLF can make it happen. Knowing this, new breeding
opportunities can be defined for the potential of livestock species to acquire plasticity for
adaptation to for example current climate changing conditions or improved emission results.
3.3. Continuous welfare and health monitoring by image systems
Animal health is of course crucial in realizing a more efficient energy equation [Eq. (1)]. PLF
offers many possibilities for real-time health monitoring from which most of them are not yet
implemented in operational field systems. It is now possible to monitor animals using normal
cameras with an image speed of up to 25 images per second or more. Moreover, we can have
many different monitoring algorithms that are easy to implement using top-view cameras placed
in operating livestock houses. In broiler chickens, for example, the eYeNamic system has been
developed for continuous automatic monitoring of the behavior of housed birds (Aydin et al.,
2010). The EU specifies that a broiler farmer must carry out a visual inspection of the birds at
least twice a day, the eYeNamic system does this continuously in a fully automated way.
Lameness problems in dairy often are related with inflammation and infections. Viazzi et al
(2013) give an example of PLF technology by using an automated camera analysis to detect
lameness in milking cow. Lameness is a major health and welfare problem in modern dairy cows,
where up to 25% may be badly affected. An abundance on lameness literature exist that shows
the different risk factors related with lameness for milking cow and other species. In literature
over 200 possible causes are described. It is a matter of detecting as soon as possible a first sign
of lameness to start treatment immediately. This is only possible by continuous monitoring. A
camera is filming the individual cow each time she leaves the milking robot. An algorithm does a
gait analysis based upon the video and a gait score is given. By carrying out image analysis and
calculating model parameters from the image information, an algorithm was developed for
automated detection of lameness problems in dairy cows. Such techniques provide frequent and
fully automatic monitoring of each individual cow, a process that the farmer no longer can carry
out easily. As soon as the calculated individual gait parameters change, a warning alerts the
farmer. The objective is to detect the first sign of upcoming lameness problems by focusing on
the variation in gate analysis when it transfers from the individually sound gait to a first score of
lameness in a scale of 5 as scored by experts. It was shown that the individualizing of the algorithm
is an important asset to make the monitor up to 10 % more accurate in detecting animals with
problems (Viazzi et al., 2013). To make the principle applicable in the field, a simplified prototype
of the lameness monitor was developed based upon only a top view image in which the main
feature variable is the back arch of the cow. When a cow has pain in one of the feet or legs, she
will use the back muscles to reduce the weight on that leg which is seen in the back arch of the
animal. The prototype tested in field conditions measures the back arch by using a top view image.
But so far also this technology did not turn yet into a commercial product for large scale
implementation.
3.4. Continuous health monitoring by real-time sound analysis
Respiratory pathologies are widespread in intensive livestock farms like pig farms (Ferrari et
al., 2010); their incidence and prevalence are high, and their principal clinical sign is coughing.
The importance of these diseases must be viewed from an economic as well as a hygiene
perspective; veterinary intervention can be expensive, and farmers can experience substantial
profit losses due to high mortality rates in growing/fattening pigs (which can be as high as 15%)
(Islam et al., 2013) or to a drop in production because of reduced feed conversion and a lower
growth rate. It is unlikely that a pig will reach the slaughter weight without having respiratory
∙ 361 ∙
problem (Bauman et al., 2002). With the Covid-19 pandemic, the importance of continuous health
monitoring is demonstrated once more. Over 60 % of the diseases that humans get are zoonoses:
transferred from animal to human. It is also known that detecting illness in individual animals
and providing individual care or group-by-group mass treatment in response to illness are not very
effective and are costly. It is therefore beneficial to investigate animal sounds with the aim of both
understanding respiratory diseases and using bioacoustics for continuous real-time monitoring
(Carpentier et al., 2018). The importance of coughing as a predictor of respiratory disease applies
to animals as well as to humans. It has been shown that pig vocalization is directly related to pain
and classification of these sounds has been attempted (Marx et al., 2003; Chedad et al., 2001). It
is also common practice among veterinarians to observe cough sounds in livestock houses for
diagnostic purposes. In this regard, there have been attempts to identify the characteristics of
coughing in animals like bovine and pigs and to automate the identification of cough sounds from
field recordings for e.g., detection of bovine respiratory disease (Figure 4 in Vandermeulen et al.,
2016). The detection of infection by sound analysis and related production loss due to reduced
feed intake has been shown (European Commission, 2016). Also, for calves it is possible
discriminate cough sounds from other sounds and to use cough sound as a non-invasively
diagnostic tool for respiratory diseases in youngstock groups (Aydin et al., 2010).
3.5. Energy for movement and physical activities
All physical or mental performances of animals or humans take metabolic energy. So far, we
are not yet considering the effect of the movement of animals in the management of the energy
equation. Many solutions are described in the scientific literature to monitor movement
continuously mainly for cows, pigs, and bigger animals by using 3D accelerometers and
gyroscope technology. The wearable technology is making fast progress in terms of accuracy,
dimensions, weight, price, and battery energy use. For lameness detection several solutions have
been proposed for several species and position and gait analysis become standard techniques for
milking cow and other species (Aydin et al, 2010; Wurtz et al., 2019).
The potential of technology for active management of the Energy movement component
however is not yet explored. When combining such technology with heart rate monitoring,
interesting opportunities become available for active management of this component. Do we get
more happy animals when they move more like has been shown for humans when doing sports?
What is the effect on body composition, meat quality, feed conversion, etc.? For large animals
like cow, beef cattle, horses, etc. the metabolic energy for moving the body is not neglectable
which does not mean that active movement management cannot be a good option for health
management or animal welfare?
Another example of using movement information is the automated continuous monitoring of
cows to detect problems during delivery. When farmers with cow and sheep are in the period of
deliveries, the follow up takes a lot of man hours. In collaboration with the Teagasc research
institute in Ireland we developed a system that reduces the image of the complex animal to only
ten dots. The real-time monitoring of the position of the simplified model allows to detect when
the delivery does not proceed as can be expected and the algorithm can give an alert to the farmer
that human help is needed.
3.6. Energy for the thermal component
What we do for many decades is to work on the energy thermal term by housing livestock in
structures to protect them from the varying outside weather conditions. With climate change there
are new problems to be solved. With genetic engineering many possibilities are still unexploited
in relation to phenotypes of livestock where PLF can collect many data. There are several
candidate genes that are associated with adaptation of ruminants, monogastric and poultry to heat
stress (Rovelli et al., 2020). Also, the use of new technologies and materials in climate control of
livestock houses has many unexploited opportunities when combining with the concept of letting
animals decide more for themselves by using PLF technologies with camera’s, microphones,
∙ 362 ∙
sensors etc. These technologies can monitor what the animals aims for and supports it in doing
so. The objective on this term is to make more body energy available for the basal metabolism
and for production term by reducing less energy spent in the thermal component (Figure 1).
Figure 1. Reduce energy in the thermal component to add metabolic energy in the basal
metabolism and the production term.
3.7. Energy for the mental component (animal welfare)
The term for mental energy includes mainly animal welfare as a central element for several
reasons. It is important to note that animal welfare is not only required for ethical reasons but also
for reasons of efficiency of the animal production process. As said higher it is a challenge to work
on each term of the energy equation 1 [Eq. (1)]. Animal welfare or the energy for mental term is
the term that is related to most other terms in Eq. (1). Animal welfare is also the part that is related
to many SDG’s (Sustainable Development Goals) (United Nations, 2015).
A first way to improve the efficiency in animal production is to reduce the number of health
issues since these always generate production losses. Animal welfare is closely connected to
animal health since stressed animals are using the feed energy for mental stress instead of bringing
this amount of energy into the production term (meat, milk, eggs, fiber). Moreover, stressed
animals depress the immune system with increasing risk for infections. How many of the more
than 70 billion livestock animals, slaughtered this year for the worldwide demand, are not
stressed? All the energy, used for the mental component when stressed, is not available for the
basal component, neither for the immune system, the thermal or the production term in the
equation.
What we expect to become a real disruptive technology for the livestock sector is the
continuous real-time monitoring of the mental energy term or animal welfare based upon
physiological variables. The technology is available for humans and will become available for
animals as soon as the appropriate sensor is realized (Joosen et al., 2019; Luwei et al., 2020; Piette,
2020).
In the past, Darwin (1872) and Porges (1995) have already shown that there is a dynamic
relationship between the central nervous system and the expression of emotions and that
physiological variables offer potential for monitoring stress. When an animal produces metabolic
energy within the aerobic zone, the inhaled air is brought into the blood in the lungs. Then the
heart is pumping the blood to the cell level where the metabolic energy is produced. This means
that the level of heart rate is a measure for the possible total aerobic production of metabolic
energy. This means that Eq. (1) can be written in the form of a heart rate Eq. (2).
HRTOTAL = HRBASAL + HRMOV + HRTHERM + HRMENTAL + HRPROD + ξ (2)
∙ 363 ∙
The decomposition of heart rate components in mental and physical components remains a
challenge on moving subjects, which leads to the consequence that most methods for stress
monitoring based on heart rate are limited to non-moving subjects, like heart rate variability.
The mental component in the energy equation can be monitored based upon physiological
measurements (heart rate and movement) in combination with individualized algorithms that
adapt to each individual animal. It has been shown on pigs that it is possible to monitor the
response of animals in real-time by measuring heart rate and movement in real-time in a
synchronized way. We stressed pigs 6 times with a sound-signal to induce a negative mental
response, and we gave them 6 times a toy to induce a positive mental response. The algorithm
picked up both types of responses in real-time. We could also show the agreement between the
output of the real-time algorithm with the blood hormone response of noradrenaline (Joosen et
al., 2019).
To introduce this technology in the field, we need a sensor that can accurately measure heart
rate and movement for example in an ear tag. A candidate technology to monitor heart rate in
livestock is the meanwhile well know ppg technology, standing for photoplethysmography
(Luwei et al., 2020). It can be expected that such an ear tag, monitoring heart rate and movement,
will come soon and it seems obvious that the first species to use it will be the more expensive
individual animal such as milking cow, beef cattle, racehorses etc. Such objective and continuous
monitoring of animal welfare based upon objective physiological measurements will be a huge
step in creating a more efficient production process. It finally will create the possibility to give
animals in the production system a life worth living (Wathes, 2010; Yeates, 2011).
The PLF solutions are mostly built upon the combination of one or more sensors (e.g., 3D
accelerometer, temperature sensor,…) or sensing systems (sound, image, …) and algorithms to
calculate the so-called target variable aimed for, e.g., lameness detection.
In the higher mentioned Eq. (1), the minimizing of the amount of feed energy going into the
terms Energy Immune system, Energy thermal and Energy mental, heavily effects the process
efficiency and the amount of energy that finally can go into the production term (Figure 2).
Figure 2. PLF aims to put more metabolic energy from the feed into the production term (meat,
milk, eggs, and fibers) with minimizing manure and emissions ξ.
This process efficiency in the transformation of feed energy into animal products determines
how much energy is wasted in manure, emissions and in polluting the environment or term ξ in
Eq. (1). So far, not enough attention is given to this very important focus since reducing the
environmental impact of the livestock sector has become a top priority. There is enough potential
in PLF technologies to work on this point. Tullo et al. (2019), give an extensive overview of the
potential of PLF technologies to mitigate environmental impact.
4. Conclusions
While the worldwide demand for animal products is increasing, the livestock sector is
challenged to answers several questions from the society. For the world and mainly for developing
countries, livestock is of crucial social-economic importance. It is not possible to reduce the world
wide increasing demand in countries that finally get the economic possibility to get more animal
∙ 364 ∙
protein.
To come up with solutions, an important step is to improve the efficiency in animal production
rather than once again increasing the number of animals to fulfil the demand. From scientific
literature and the so far rather limited commercial examples in the field, we can see that Precision
Livestock Farming (PLF) has a high potential to play a role in the more efficient management of
producing animal products. Today there is an inadequate demonstration of how livestock can play
a key role in the development of sustainable agriculture in different agroecosystems, and so far
we fail to transfer this appropriate technology to large scale field implementations.
Considering the huge increasing demand for animal products, other solutions should as well
be considered: Lab-meat, food losses and grain use by livestock. Europe aims for eating less meat
and dairy products and looks for alternative protein solutions. This might make livestock
producers nervous. The possibilities of lab-meat and plant-based protein should not be considered
as a treat since the worldwide demand for animal protein is expected to increase with over 60 %
by 2050 (Wehberg et al., 2017; Gerhardt et al., 2019). Less feed must generate more animal
products. Alternative solutions are welcome to contribute to the increasing demand. The use of
insects as food component for humans and animals is an alternative for the increasing demand for
animal meat and has environmental and social advantages over intensive production of livestock.
However, an issue that might compromise the success of insect rearing is the outbreak of insect
diseases and virus transmission. More understanding is required of the different factors that
interact in insect mass rearing.
At the same time, we struggle with serious food loss and food waste. We must consider
whether we can use food wastes as animal feed to collaborate with waste management processes
and food security challenges.
The more fundamental PLF research has started in the 90’s with experimental laboratory work
(Wouters et al., 1990) but so far, thirty years later not many systems have been successfully
implemented in the field at large scale. There is a lack of successful identifications of appropriate
technologies that have proven to work in the field at large scale. It is now time to start
implementing PLF technologies, developed at lab-scale to bring them to commercial livestock
applications. Although the high number of animals per farm, we need to bring the animals closer
to the farmer by using this technology (Leman, 1992).
Knowing that in most countries, where the demand for animal products is increasing, most of
the farmers are small farms. PLF technology will be affordable for them when the big producers
of animal products buy PLF technology in high numbers which makes the technology cheap. PLF
technology must be adapted to small farms and family farms to create the worldwide impact that
it can have in creating a more sustainable animal production.
Instead of re-inventing the wheel, researchers and industry should focus on large scale
implementation of developed PLF technology in commercial farms. To come up with real
solutions, a collaboration between different research disciplines (animal scientists, veterinarians,
engineers, etc.) is needed as well as a strong collaboration between researchers and industry.
References
Aydin, A., Bahr, C., Viazzi, S., Exadaktylos, V., Buyse, J., Berckmans, D. 2014. A novel method to
automatically measure the feed intake of broiler chickens by sound technology. Computers and Electronics in
Agriculture. 101 (February), 17–23.
Aydin, A., Cangar, O., Ozcan, S.E., Bahr, C., Berckmans, D. 2010. Application of a fully automatic
analysis tool to assess the activity of broiler chickens with different gait scores. Computers and Electronics in
Agriculture. 73 (2), 194–199. https://doi.org/10.1016/j.compag.2010.05.004
Baumann, B., Bilkei, G. 2002. Emergency culling and mortality in growing/fattening pigs in a large
‘farrow-to-finish’ production unit. Dtsch. tierärztl. Wochenschr. 109, 26–33.
∙ 365 ∙
Berckmans, D., 2014. Precision livestock farming technologies for welfare management in intensive
livestock systems. Revue Scientifique et Technique de l'OIE. 33 (1), 189–196. https://doi.org/10.20506/
rst.33.1.2273
Carpentier, L., Berckmans, D., Youssef, A., Berckmans, D jr., van Waterschoot, T., Johnston, D.,
Ferguson, N., Earley, B., Fontana, I., Tullo, E., Guarino, M., Vranken, E., Norton, T., 2018. Automatic cough
detection for bovine respiratory disease in a calf house. Biosystems Engineering. 173, 46–56.
Chedad, A., Moshou, D., Aerts, J.-M., Van Hirtum, A., Ramon, H., Berckmans, D. 2001. Recognition
system for pig cough based on probabilistic neural networks. Journal of Agricultural Engineering Research.
79 (4), 449–457.
Childe, V. G. 1928. The Most Ancient East: the Oriental prelude to European prehistory. London: Kegan,
Paul, Trubner Ltd;
Darwin, C. 1872. The Expression of the Emotions in Man and Animals. London, John Murray; Reprinted
Chicago, IL, University of Chicago Press, 1965.
European Commission. 2019. EU Agricultural Outlooks for 2019. For markets and Income 2019–2030.
European Commission, DG Agriculture and Rural Development, Brussels, pp. 92.
European Commission. 2016. Report of the EU-PLF Project title: Bright Farm by Precision Livestock
Farming. Nov 2012–2016. pp. 45. https://cordis.europa.eu/docs/results/311/311825/final1-20170707v5-
finalreport-gano311825.pdf
Ferrari, S., Piccinini, R., Silva, M., Exadaktylos, V., Berckmans, D., Guarino, M. 2010. Cough sound
description in relation to respiratory diseases in dairy calves. Preventive Veterinary Medicine. 96 (3–4), 276–
280. https://doi.org/10.1016/j.prevetmed.2010.06.013
Gerhardt, C., G. Suhlmann, F. Ziemßen, D. Donnan, M. Warschun, H-J. Kühnle, 2019. How will cultured
meat and meat alternatives disrupt the agricultural and food industry? Industrial Biotechnology. 16 (5).
https://doi.org/10.1089/ind.2020.29227.cge.
Greenwood, P.L., Paull, D. R., McNally, J., Kalinowski, T., Ebert, D., Little, B., Smith, D. V., Rahman,
A.,Valencia, P., Ingham, A. B., Bishop-Hurley, G. J. 2017. Use of sensor-determined behaviours to develop
algorithms for pasture intake by individual grazing cattle. Crop & Pasture Science. 68 (12), 1091–1099.
Islam, Z.U., Bishop, S.C., Savill, N.J., Rowland, R.R.R., Lunney, J.K., Trible, B., Doeschl-Wilson, A.B.
2013. Quantitative analysis of porcine reproductive and respiratory syndrome (PRRS) viremia profiles
experimental infection: a statistical modelling approach. PLoS ONE. 8 (12).
https://doi.org/10.1371/journal.pone. 0083567
Joosen, P., Norton, T., Marchant-Ford, J., Berckmans, D. 2019. Animal welfare monitoring by real-time
physiological signals. Proceedings of the ECPLF2019, the European Conference of Precision Livestock
Farming, Cork Ireland. 337–344.
Keeling, L., Håkan Tunón, H. , Gabriela Olmos Antillón, G., Berg, C., Jones, M., Stuardo, M., Swanson,
J., Wallenbeck, A., Winckler, C., Blokhuis, H. 2019. Animal welfare and the United Nations sustainable
development goals. Frontiers in Veterinary Science. 6, 336. https://doi.org/10.3389/fvets.2019.00336
Kenny, D. A., Fitzsimonsa, C., Waters, S. M., McGee, M. 2018. Invited review: Improving feed efficiency
of beef cattle - the current state of the art and future challenges. Animal. 12, 1815-1826, https://doi.org/
10.1017/S1751731118000976
Leman, A.D., Straw, B., Glock, R., Mengeling, W.L., D’Allaire, S., Taylor, D.J. 1992. Diseases of Swine,
7th Ed. Iowa State University Press, Ames.
Luwei, N., Berckmans, D., Wang, C., Li, B. 2020. Is continuous heart rate monitoring of livestock a
dream or is it realistic? A review. Sensors. 20, 2291; pp. 27. https://doi.org/10.3390/s200822
Marx, G., Horn, T., Thielebein, J., Knubel, B., von Borell, E. 2003. Analysis of pain-related vocalisation
in young pigs. Journal of Sound Vibration, 266, 687–698.
McHugo, G.P., Dover, M.J., MacHugh ,D.E. 2019. Unlocking the origins and biology of domestic
animals using ancient DNA and paleogenomics. BMC Biology. 17, 98. https://doi.org/10.1186/s12915-019-
0724-7
Norton, T. , Chen, C., Larsen, M.L.V., Berckmans, D., 2019. Review: Precision livestock farming:
building ‘digital representations’ to bring the animals closer to the farmer. Animal. 13 (12), 3009–3017.
Piette, D. 2020. Depression and burnout: a different perspective. PhD to obtain the degree of Philosophy
Degree in Biosience engineering at the Catholic University of Leuven, pp. 138.
Porges, S.W. 1995. Cardiac vagal tone: a physiological indexx of stress. Neuroscience and BioBehavioral
Reviews. 19 (2), 225–233.
Ran Bezena, R., Edanb, Y., Halachmi, I. 2020. Computer vision system for measuring individual cow
feed intake using RGB-D camera and deep learning algorithms. Computers and Electronics in Agriculture.
172 (May), 105345.
∙ 366 ∙
Rovelli, G., Ceccobelli, S., Perini, F., Demir, E., Mastrangelo, S., Conte, G., Abeni, F., Marletta, D.,
Ciampolini, R., Cassandro, M., Bernabuci U., Lasagna E. 2020. The genetics of phenotypic plasticity in
livestock in the era of climate change: a review. Italian Journal of Animal Science. 19 (1), 997–1014.
Tullo, E., Finzi, A., Guarino, M. 2019. Review: Environmental impact of livestock farming and Precision
Livestock Farming as a mitigation strategy. Science of the Total Environment. 650, 2751–2760.
United Nations. 2015. http://www.un.org/sustainabledevelopment.
Vandermeulen, J., Bahr, C., Johnston, D., Earley, B., Tullo, E., Fontana, I., Guarino, M., Exadaktylos,
V., Berckmans, D. 2016. Early recognition of bovine respiratory disease in calves using automated continuous
monitoring of cough sounds. Computer and Electronics in Agriculture. 129, 15–29.
Viazzi, S., Bahr, C., Schlageter-Tello, A., Van Hertem, T., Romanini, C.E.B., Pluk, A., Halachmi, I.,
Lokhorst, C., Berckmans, D. 2013. Analysis of individual classification of lameness using automatic
measurement of back posture in dairy cattle. Journal of Dairy Science. 96 (1), 257–266.
Wathes, C. 2010. Lives worth living? The Veterinary Record. London 166 (15), 468–472.
Wehberg, G., W. Vaessen, F. Nijland, T. Berger. 2017. Smart Livestock Farming: Potential of
Digitalization for Global Meat Supply. Discussion Paper, Deloitte. Issue 11/2017.
Wouters, P., Geers, R., Parduyns, G., Goossens, K., Truyen, B., Goedseels, V., Van der Stuyft, E. 1990.
Image-analysis parameters as inputs for automatic environmental temperature control in piglet houses.
Computers and Electronics in Agriculture. 5 (1990), 233–246, 233.
Wurtz, K., Camerlink, I., D'Eath, R.B., Fernandez, A.P., Norton, T., Steibel, J., Siegford, J. 2019.
Recording behaviour of indoor-housed farm animals automatically using machine vision technology: A
systematic review. Plos One. 14, 12. https://doi.org/10.1371/journal.pone.0226669
Yeates, J.W. 2011. Is ‘a life worth living’ a concept worth having? Animal Welfare. 20, 397–406.
∙ 367 ∙
∙ 368 ∙
Theme V:
Manure and Mortality Management and
Byproduct Development
in Animal Agriculture
∙ 369 ∙
∙ 370 ∙
Field Experience of Removing and Land Application of

Dairy Lagoon Solids
Timothy Canter, Teng-Teeh Lim *, Joseph Zulovich
Agricultural Systems Technology, University of Missouri, Columbia, MO 65211, USA
* Corresponding author. Email: limt@missouri.edu
Abstract
Manure lagoon systems remain popular in many livestock farms to manage water inventory
and as a means of treatment. Excessive solids build-up in such lagoons have been reported for
older systems and lagoons that are under-sized, and especially if the lagoon receives high amounts
of bedding and undigested fibrous materials. Accumulation of too many solids can result in
significant odor emissions, environmental degradation and decrease operational flexibility of the
farm. This paper documents procedures to prepare for and complete land application of manure
solids from a flush dairy lagoon in central Missouri, USA. The dairy farm uses mattress bedding
with supplemental cedar shavings in a freestall barn housing 140-160 lactating cows. Efforts to
estimate cost and plan nutrient application included measuring the lagoon sludge depth and lab
analysis of sludge characteristics were completed prior to awarding the project to a contractor
through a bidding process. The contractor utilized specialized equipment to dilute, agitate, pump
and land apply approximately 8 million gallons of diluted lagoon solids in just a few days. The
slurry being pumped out of the lagoon was sampled throughout the process to monitor the mass
of nutrients applied to specific plots of land. The purpose of this paper is to document the process
of effective lagoon solids removal for land application, considering the preparation, specialty
equipment and trained professional, timing of the crop fields, and adequate field working days.
Keywords: Manure management, flush dairy, lagoon agitation, sludge removal, nutrient
management.
1. Introduction
Lagoons can be an effective means of treating and storing flushed manure and may also serve
as a holding basin for water that is to be reused (e.g., manure flushing). Years, if not decades, can
go by without any significant maintenance requirements to the system outside of water
management, pumps and valves if the lagoon is properly designed and built (Pfost and Fulhage,
1992). Annual pump-out, in this context, consists of mixing and pumping from the lagoon to be
spread on nearby fields as supplemental crop fertilizer. Draw down (pumping) of the lagoon may
be needed one or more times per year, depending on lagoon size and loading, and rainfall, to
manage the water inventory.
Below the liquid layer is a blanket of settled solids. Some of these solids are digested by
microbes over time but some are not degradable. The USDA provides general generation rates
and characterizations animal wastes and bedding types (NRCS, 2008). Build-up of solids is
expected over time, and designers of lagoons will allocate additional lagoon volume for solids
storage. It may take five, ten, or twenty years, but the water holding capacity of the lagoon will
become diminished to a point where normal operation is affected.
The operator of the manure system can continue operating the lagoon with diminished
capacity for some time by drawing down the liquid supernatant level more frequently. However,
the longer this continues the more likely the draw down will have to occur during times that are
not ideal, possibly even illegal, such as when the ground is frozen or saturated. Lagoon water
topping the lagoon is another possibility, which can, in turn, lead to berm failure (Miller and
Major, 2013).
Pumping the solids out of the lagoon can be resource-intensive; requiring specialized
∙ 371 ∙
equipment with trained operators, seasonal planning, and suitable land to which solids can be
applied. Sometimes finding suitable land can be the most difficult aspect to overcome. Manure
slurry is heavy and expensive to transport, so land must be as close to the lagoon as possible. Crop
cover and rotation, crop yield and nutrient uptake, weather and current nutrient concentrations in
the soil are other aspects to consider (Fulhage, Charles D, 1994).
The purpose of this case study is to highlight some of the aspects involved with a recent solids
removal effort at a flush dairy lagoon in central Missouri. A contractor who specializes in lagoon
agitation, pumping, and land application was hired to complete the project. The effort was so great
that the dairy’s managers decided to install a coarse-solids removal system to reduce the amount
of non- and slowly degradable solids that enter the lagoon.
2.1. Background
FM Dairy is in Boone County, Missouri and is home to a herd of 140-160 lactating cows. The
dairy uses mattress bedding with supplemental cedar shavings (2.5 m3 d-1) in a freestall barn, the
layout of which is shown in Figure 1. Potable water is used to flush the milking parlor and holding
pen, while the freestall barn is flushed using recycled lagoon water.
Water Towers
x x x x
Milking
x 12” Butterfly Valve Parlor
Bedding Area
Throughway
Slope Screen
(Abandoned)
Storage Lagoon
Figure 1. Layout of FM Dairy Freestall Barn, Milking Parlor, and Manure Management System.
The 30.5 m by 61 m freestall barn has four flush alleys. The alleys are served by two flush
water towers that are each approximately 53 m3. The barn is flushed two to three times per day,
with three times per day being typical. Greater detail provided by the operators revealed the dairy
may flush as much as four times per day; morning flush, then scrape, then flush again, followed
by flushing around noon and again in the evening. The milking parlor and holding pen are also
flushed to the lagoon, but the water contribution is only 2% of the total daily flush water usage.
There was, at one time, an elevated screen that helped remove the large solids from the flush,
but the screen system fell into disrepair several years ago and was abandoned. It had been
approximately twenty years as of the time of this writing since solids have been effectively
removed from the lagoon at FM dairy. Solids accumulation had become so great that, by 2020,
islands could be seen in the middle of the lagoon. The operational capacity of the lagoon had been
decreased so much that the manure flushing system was on the verge of being inoperable.
∙ 372 ∙
2.2. Solids Removal and Land Application

The solids removal and application process should begin with an estimate of the volume of
solids and mass of nutrients that are needed to be removed and land applied. The mass of nutrients
must be calculated to understand the amount of land needed for application and if hauling a portion
of the slurry away will be required. Lastly, contracting with a firm may be needed due to
specialized equipment requirements.
2.2.1. Estimating the mass of solids and nutrients
Several companies with specialized equipment and training were contacted for estimates
removing the lagoon solids. The driving expense factors for solids removal include (a) the number
of gallons pumped and (b) the distance between the lagoon and the location where the solids are
deposited. Often there is a small window of opportunity for the land application of solids in the
spring, prior to crops being planted, and in the fall, between post-harvest and winter. Delays can
occur if rain or snow melt causes the ground to be saturated. The first step was to gather
information on the lagoon volume and its solid contents so that specialized contractors can provide
accurate proposals.
The lagoon at FM Dairy is oblong with curved ends, which means the exact dimensions are
nearly impossible to determine without extensive land surveying. There were no “as built” plans
available that would indicate precise length, width, side slope, or depth. The manager believed
the depth of the lagoon to be 3–3.7 m. It is common to have a slide slope of approximately
3-to-1, which is often the maximum slope allowed to maintain berm integrity while minimizing
footprint. Dimensions could be estimated from an online satellite view even though they were not
exact. These basic characteristics are used by contractors to estimate the volume of the lagoon
(FM Dairy lagoon was estimated to be 28,000 m3). However, the cost of the project is determined
by the total volume of effluent pumped or dredged, including dilution water if needed, which
means sludge inventory is important for an accurate estimate.
A sludge survey is needed to determine solids depth, volume, and nutrient content. These
factors will play an important role in both cost of removal and land requirements for application.
There are numerous methods to accomplish a sludge survey and a farm manager may wish to
review different options to determine what will work best for their operation (Sharara, 2020).
2.2.2. Lagoon agitation and land application
Solids in the lagoon at FM Dairy were agitated and pumped out from May 21, 2020 through
June 8, 2020. Precision Pumping (Quincy, IL) was contracted to apply a total of 30,000 m3 over
130 hectares, averaging approximately 231 m3 ha-1. Equipment names and details are provided
for clarity and are not an endorsement or recommendation of equipment.
The process began by agitating from the shore using two 20.3-cm (8-inch) lagoon agitators
(NUHN, Sebringville, ON, Canada) (Figure 2) connected to two tractors (John Deere, Moline,
IL). including an 8100 119 PTO kW using 34 L of diesel per hour and an 8295R 180 PTO kW
using 49 L of diesel per hour.
Figure 2. PTO-drive lagoon agitators and agitation boat in operation.
∙ 373 ∙
The agitators require a minimum of 71 kW to produce a maximum flow of 30 m3 min-1. As

seen in Figure 2, these agitators began operation by breaking up the surface solids using high-
powered jet, before they were moved to different locations around the lagoon and ultimately set
up on opposite corners to create a circular flow.
An agitation boat (2067C, Puck Enterprises, Manning, IA) was used once the two agitators
had sheared off enough of the upper solids blanket to allow for the boat’s draft (Figure 2). The
boat is powered by a 6.8 L, 205 kW John Deere engine that consumes up to 25 L of diesel per
hour. It operates a Cornell pump with a 43 cm impeller that feeds 3 guns with a total of 15 m3
min-1 at 0.2 MPa. These guns provide jet propulsion for the boat and break-up the solids blanket
below the boat. The boat is remote controlled with optional GPS setting to allow for line-of-sight
operation and automated hands-free operation if chosen.
A Zoske (Iowa Falls, IA) pump trailer was used (Figure 3) to deliver the agitated lagoon
effluent to the injection sites through a 20 cm, 1.4 MPa hose. The pump trailer consists of a Cornell
centrifugal pump powered by a John Deere motor. An Atlas Copco (Rock Hill, SC) compressor
is set adjacent to the pump trailer to provide compressed air to unclog the discharge line if needed.
Figure 3. Primary lagoon pump (trailer mounted, left) and dilution pump (right).
A dilution pump, consisting of a Thompson (Port Orange, FL) centrifugal pump mounted on
a 45 kW KHD Deutz (Deutz, Norcross, GA) motor, was used to pump fresh water from two nearby
ponds into the lagoon (Figure 3). The exact amount of dilution water used is unknown. The
contractor reported optimal pumping conditions occur when the slurry is approximately 5 to 8
percent solids. The initial slurry had a solids content of 10–13%, so a significant amount of
dilution water was needed to dilute the solids content to the target range. The contractor had to
continually balance the flow of dilution water with the speed of the agitation boat to hit the target
solids content. Too high of a solids content could clog hoses and injectors. Too low a solids
content results in inefficient operation, and both affect the amount of effluent to be land applied
and can be constrained by equipment or soil moisture condition.
A booster pump, consisting of a Smart Turner (Brantford, Ontario, Canada) centrifugal pump
paired with a John Deere engine (Figure 4), was used to maintain a flowrate over 6.8 m3 min-1
through the drag hose, which was reduced from 20 cm to 15 cm immediately after the booster
pump.
A toolbar (Hydro Engineering, Norwood Young America, MN) with 13 Dietrich shanks on
61 cm centers was used for manure injection. The toolbar was mounted to a John Deere tractor
(8320R, 196 PTO kW using 57 L of diesel per hour), as seen in Figure 4.
Figure 4. Booster pump (left), application toolbar/tractor (center), and hose cart/humper (right).
∙ 374 ∙
The injection equipment was supported by a John Deere tractor (8285R with 178 PTO kW
using 57 L of diesel per hour fuel use) that used a Zoske Open Spool Hose Humper to feed drag
line from a Zoske Raptor 32 Hose Cart (Figure 4). The hose cart has a capacity of either 16 sections
of 20 cm hose in 3.2 km lengths or 32 sections of 15 cm hose in 6.4 km lengths and is capable of
layout speeds up to 16.1 km (10 miles) per hour when the spool is under power.
3.1. Sludge Survey
The sludge survey at FM Dairy was unable to be fully executed. Surveyors discovered the
sludge blanket in the lagoon was so high that the boat would get stuck as soon as it was placed in
the water. This made estimation of the solids volume relatively easy insomuch as it was
approximately equal to the volume of the lagoon. However, that was not equivalent to the volume
of effluent that needed to be pumped from the lagoon due to the solids concentration. Three
samples were taken; one from a mound of visible sludge in the middle of the lagoon, one using a
conventional sludge judge between sludge mound in the middle of the lagoon, and one using a
custom sludge judge form the shore. Figure 5 shows the concentrations of nutrient for the three
samples as well as the average among the samples. The moisture contents of the samples were
86%, 94% and 88% for the sludge mound, lagoon center, and shore samples, respectively, with
an average of 89%. This solids concentration is close to Sharara’s (2020) estimate of 14%. The
contractor indicated that they liked to pump at solids concentration of 5–8 % and would use water
from a nearby pond to dilute the lagoon solids. A lagoon volume of 28,000 m3 full of sludge with
a 13% solids content results in an estimated overall solids mass of 370,000 kg. The contractor
pumped approximately 210,000 kg of solids, leaving roughly 43% of the original mass of solids
in the lagoon.
Nitrogen Ammonia-N P2O5 Phosphorous K2O Potassium

Concentration (g L-1)
Sludge Mound Sludge Judge

Custom SJ (Shore) Average
(Boat) (Boat)
Nitrogen 3.02 2.52 2.39 2.64
Ammonia-N 0.49 0.43 0.51 0.47
P2O5 3.63 1.71 2.61 2.65
Phosphorous 1.56 0.73 1.12 1.14
K2O 1.21 1.11 1.19 1.17
Potassium 1.00 0.92 0.98 0.97
Figure 5. Nutrient analysis of lagoon solids.

As seen in Figure 5, nutrient varied significantly between sample locations and methods. It
may be prudent to take multiple samples and consider applying concentrations at the 90th to 95th
percentile among the samples when planning land application. Samples can, and should, be taken
during the land application process to track the mass of nutrients being applied as well as the
solids content in the slurry.
3.2. Lagoon Effluent Characteristics
The process of agitation, dilution, pumping, and land application began on May 21, 2020.
∙ 375 ∙
Samples were taken from the pump trailer discharge (Figure 3) and analyzed for nitrogen,
ammonia-nitrogen, P2O5, K2O and moisture, as seen in Figure 6. It should be noted that
application did not occur on Saturday or Sunday (May 24–25 and May 30–31), as well as on May
28 and May 29 due to inclement weather.
3.5 98
Nitrogen Ammonia-N P2O5 K2O Moisture
3.0 97
96
Concentration (g L-1)
2.5
Moisture (%)
95
2.0
94
1.5
93
1.0
92
0.5 91
0.0 90
5/21 5/22 5/23 5/24 5/25 5/26 5/27 5/28 5/29 5/30 5/31 6/1 6/2 6/3 6/4
Date in 2020, m/dd
Figure 6. Nutrient contents of lagoon effluent samples.
Figure 6 displays a general pattern of decreasing concentration as moisture content increased
from May 22 through June 1. Nitrogen and phosphorus increased after June 1 as the moisture
content decreased (i.e., solids concentration becomes less dilute). A lower nutrient concentration
between days with similar moisture contents (e.g., May 27 vs. June 2) could be explained by solids
coming from a lower stratum in the solids blanket on the later date, meaning the solids were older
and more thoroughly digested than the solids in the upper stratum, or simply as noise. It is not
clear why potassium concentration trends were diametrically opposed to nitrogen and
phosphorous in the June samples. Multiple samples were taken on June 1 (11:00 and 15:00) and
June 2 (8:00, 11:00 and 15:00).
Variability seen in Figure 6 can be evaluated by comparing the sample difference to the mean
value, or simply taking the maximum difference between samples and dividing by the average
value of the samples. This method provided the percentage change relative to the sample average,
presented in Figure 7. The comparison suggests that potassium has the most variability, nitrogen
has the least variability, and phosphorous is inconclusive. However, the variability among the
nutrients is, in general, higher than the moisture content (i.e., solids concentration). Therefore,
one may conclude that the nutrient concentrations in the sludge vary based on the sludge and as a
function of moisture content. This poses a challenge for uniform distribution of nutrients from
manure lagoons. Perhaps the most reliable solution is real-time nutrient analysis and load
determination electronics that some land-application vendors are offering as part of their
equipment package.
3.3. Land Application
Solids were applied to 360 hectares, with the furthest being 2.5 km from the lagoon. The
contractor provided data that included GPS descriptions and field maps of where application
occurred, number of hectares of each field, a list of days that solids were applied to a given field,
overall average application rate (m3 ha-1) for each field, and total volume applied to each field.
Data provided via communication with the contractor was inconsistent that, as a result, nutrient
and solids concentrations were extrapolated from May 22 and applied to May 21 and May 23, as
∙ 376 ∙
well as from June 3 and applied to June 8, to estimate the mass of nutrients applied per hectare as
shown in Figure 8. After the lagoon was pumped down, the FM dairy installed a pull-plug
sediment basin (PPSB) system (Canter et al., 2020) to help remove some of the solids in the
manure flush. The PPSB is relatively low-maintenance and easy to operate solid removal system,
that would allow the farm improve managing solids inventory in the lagoon.
35
Percent Change Relative to Average
Nitrogen Ammonia-N P2O5 K2O Moisture

30
25
20
15
10
0
6/1/2020 6/2/2020
Nitrogen 2.7 0.90
Ammonia-N 8.8 11
P2O5 11 0.91
K2O 27 26
Moisture 0.42 0.85
Figure 7. Percent change of same day samples relative to the same day sample average.
700
Mass Application (kg ha-1)
Nitrogen Ammonia-N P2O5 Phosphorous K2O Potassium

600
500
400
300
200
100
0
5/21 5/22 5/23 5/24 5/25 5/26 5/27 5/28 5/29 5/30 5/31 6/1 6/2 6/3 6/4 6/5 6/6 6/7 6/8
5/21 5/22 5/23 5/26 5/27 6/1 6/2 6/3 6/8
Nitrogen 554 535 603 386 322 302 313 347 347
Ammonia-N 124 119 134 84 85 69 65 70 70
P2O5 292 282 318 339 269 203 216 261 261
Phosphorous 126 121 137 146 116 87 93 112 112
K2O 191 185 208 130 105 168 168 134 134
Potassium 159 153 173 108 88 139 139 111 111
Figure 8. Mass of nutrients applied by date.

4. Conclusions
Manure management can be a burden for all animal feeding operations; one that has the
potential to become a significant threat to the finance and management of farm if not proactively
managed. Sludge surveys of the lagoon prior to solids pumping can help reduce costly budget
overruns by providing the owner and the contractor with an estimate of the volume of solids and
nutrients that must be removed from the lagoon and the area of land needed for the application.
Farm managers would be well-advised to survey their lagoon every two to three years to track
sludge accumulation and plan for proper pump-down as a means of reducing sludge buildup.
Proper solids removal from the lagoon, particularly if regular and effective manure solids removal
has been neglected, requires specialized equipment and trained personnel for complete agitation,
∙ 377 ∙
that is required to reduce liquid supernatant on an annual or semiannual basis. Land application is
the most effective method for recycling the manure nutrients if proper nutrient management plan
is planned and followed (Pfost et al., 2000). Considerations and training should be taken to prepare
for potential safety hazards during agitation and land application (LPELC Admin, 2019).
No amount of planning can help avoid the variability seen in nutrient concentrations
throughout the day during lagoon pumping. Real-time nutrient analysis equipment coupled with
automated application rate can be very helpful in supplying the operator with real-time
information and ensure even distribution of nutrients. Such systems may not be practical for small
operators but efficient for custom manure applicators and/or larger farms. This can be another
reason for animal farms to consider professional custom applicator for the entire suite of manure
removal, from soil and manure sampling, nutrient management planning, to proper agitating and
land application of the manure. It should be noted that, due to heavy solids build-up and the
dilution needed, complete agitation and homogeneity of the lagoon was very difficult to achieve.
Regular and effective solids removal of the manure influent would ensure a greater portion of
lagoon volume to be utilized for recycling, which may also help reduce nutrient variability in the
lagoon effluent. Supplying consistent nutrients to the crop fields that received the lagoon effluent
can ensure and improve crop yields, making lagoon effluent valuable nutrient source.
Acknowledgements
The authors acknowledge the assistance of the FM dairy farm manager and staff, and Mr. Tim
Reinbott, Field Operations and Agricultural Experiment Station.
References
Canter, T., T-T. Lim, T. Chockley, 2020. Considerations of pull-plug sedimentation basin for dairy
manure management. University of Missouri Extension. https://extension.missouri.edu/publications/
eq302. Accessed July 6, 2021.
Fulhage, C.D., 1994. Solids removal from livestock manure lagoons. University of Missouri Extension.
https://extension.missouri.edu/publications/wq324. Accessed April 3, 2021.
LPELC Admin, 2019. Manure storage safety. Livestock and Poultry Environmental Learning
Community. https://lpelc.org/manure-storage-safety/. Accessed July 6, 2021.
Miller, R., J. Major, 2013. Lagoon startup and maintenance for optimal livestock waste treatment. Utah
State University Cooperative Extension. https://extension.usu.edu/agwastemanagement/ou-files/pdfs/Lagoon
_Startup_and_Maintenance_2013.pdf. Accessed April 24, 2021.
NRCS, 2008. Agricultural Waste Management Field Handbook; Chapter 4 - Agricultural Waste
Characteristics. United States Department of Agriculture. https://directives.sc.egov.usda.gov/
OpenNonWebContent.aspx?content=31475.wba. Accessed March 18, 2021.
Pfost, D., C. Fulhage, 1992. Lagoons for storage and treatment of dairy waste. University of Missouri
Extension. https://extension.missouri.edu/publications/wq304. Accessed April 3, 2021.
Pfost, D., C. Fulhage, D. Rastorfer, 2000. Anaerobic lagoons for storage/treatment of livestock manure.
University of Missouri Extension. https://extension.missouri.edu/publications/eq387. Accessed April 10,
2021.
Sharara, M., 2020. Sludge Survey methods for anaerobic lagoons. NC State Extension.
https://content.ces.ncsu.edu/sludge-survey-methods-for-anaerobic-lagoons. Accessed May 2, 2021.
∙ 378 ∙
Indexes
∙ 379 ∙
∙ 380 ∙
Author Name Index

Author name Page
Aarnink, André J.A. .................................................................................................................. 186

Bao, Jun ..................................................................................... 155, 193, 230, 246, 260, 310, 326
Berckmans, Daniel.................................................................................................................... 358
Bi, Yanju .................................................................................................................................. 193
Burns, Robert............................................................................................................................ 209
Cai, Zhen .................................................................................................................................... 98
Canter, Timothy ........................................................................................................................ 371
Cao, Mengbing ......................................................................................................................... 238
Chen, Long ............................................................................................................................... 123
Chen, Zhonghao ....................................................................................................................... 146
Cheng, Pu ................................................................................................................................. 214
Cheng, Zheng ........................................................................................................................... 222
Chowdhury, Vishwajit S. .......................................................................................................... 179
Connor, Laurie .......................................................................................................................... 201
Cui, Yangyang .......................................................................................................................... 179
Dai, Xiaorong ............................................................................................................................. 91
Dai, Zichun ............................................................................................................................... 131
Devillers, Nicolas ..................................................................................................................... 201
Dick, Kristopher ....................................................................................................................... 201
Ding, Luyu ................................................................................................................ 302, 318, 334
Ding, Susu ........................................................................................................................ 230, 260
Du, Tiantian .............................................................................................................................. 318
Du, Xinyi .................................................................................................................................... 68
E, Lei .......................................................................................................................................... 34
Eckelkamp, Liz ......................................................................................................................... 209
Evans, Jeffrey D. .......................................................................................................................... 3
Fabian (Wheeler), Eileen E....................................................................................................... 123
Feng, Yanru ...................................................................................................................... 193, 230
Furuse, Mitsuhiro ..................................................................................................................... 179
Gan, Haiming ........................................................................................................................... 287
Gan, Hao ................................................................................................................................... 209
Gao, Ronghua ........................................................................................................... 302, 318, 334
Groot Koerkamp, Peter W.G. ................................................................................................... 186
Guan, Huiyuan ............................................................................................................................ 68
Guarino, Marcella ..................................................................................................................... 295
Guo, Binbin .............................................................................................................................. 131
Han, Guofeng ........................................................................................................................... 179
Han, Shengqiang......................................................................................................................... 42
Hawkins, Shawn ....................................................................................................................... 209
He, Pengguang .......................................................................................................................... 277
He, Yuxuan ................................................................................................................................. 91
Heidari, Davoud ....................................................................................................................... 138
Hu, Feiyue .................................................................................................................................. 26
Hu, Zhenpeng ............................................................................................................................. 34
Huang, Dandan ......................................................................................................................... 114
∙ 381 ∙
Huang, Endai .................................................................................................................... 269, 287

Huang, Jinjun.............................................................................................................................. 50
Hui, Xue ................................................................................................................................... 214
Huynh, Thuy T.T. ..................................................................................................................... 186
Ji, Boyu ....................................................................................................................................... 76
Ji, Zhenzhen.............................................................................................................................. 214
Kang, Jingjing........................................................................................................................... 214
Kong, Xianwang ......................................................................................................................... 98
Lei, Kaidong ............................................................................................................................. 171
Li, Chunmei ........................................................................................................................ 58, 179
Li, Hangqi ................................................................................................................................. 163
Li, Hao .......................................................................................................................... 50, 68, 106
Li, Jianhong .............................................................................................................................. 230
Li, Jiawei .................................................................................................................................. 302
Li, Mingyang .............................................................................................................................. 58
Li, Qifeng ................................................................................................................. 302, 318, 334
Li, Shulei .................................................................................................................................. 238
Li, Xiang................................................................................................................................... 260
Li, Xin ...................................................................................................................... 222, 310, 326
Li, Xiuchen ............................................................................................................................... 163
Li, Xuanyang .................................................................................................................... 214, 252
Li, Yansen .......................................................................................................................... 58, 179
Li, Yutao ........................................................................................................................... 222, 246
Li, Zonggang .............................................................................................................................. 42
Li, Zongyang .............................................................................................................................. 76
Liang, Mingshen ....................................................................................................................... 114
Lim, Teng-Teeh ........................................................................................................................ 371
Lin, Hongjian.................................................................................................................... 146, 277
Lin, Jun ......................................................................................................................................... 3
Liu, Chang .................................................................................................................................. 42
Liu, Dezhao .......................................................................................................................... 91, 98
Liu, Honggui............................................................................................................. 222, 310, 326
Liu, Jijun................................................................................................................................... 342
Liu, Kai ............................................................................................................................. 269, 287
Liu, Tonghai ............................................................................................................................. 318
Liu, Tongshuai .......................................................................................................................... 214
Liu, Weidong ............................................................................................................................ 214
Liu, Wenbo ............................................................................................................................... 163
Liu, Yu.......................................................................................................................... 34, 76, 238
Liu, Zhen .................................................................................................................................... 98
Lovarelli, Daniela ..................................................................................................................... 295
Lu, Yujian ................................................................................................................................... 34
Lv, Yang ................................................................................................................................... 334
Ma, Wei .................................................................................................................................... 214
Ma, Weihong ............................................................................................................ 302, 318, 334
Man, Zun .................................................................................................................................... 91
Mao, Axiu ......................................................................................................................... 269, 287
Meng, Rui ................................................................................................................................. 318
Mu, Gang .................................................................................................................................. 163
Nguyen, Xuan Dung ..................................................................................................................... 3
Nian, Haoyang .................................................................................................................. 230, 260
∙ 382 ∙
Pan, Jinming ......................................................................................................... 10, 18, 146, 277

Pan, Lei ..................................................................................................................................... 260
Pelletier, Nathan ....................................................................................................................... 138
Prado, Maria ............................................................................................................................. 209
Purswell, Joseph L. ....................................................................................................................... 3
Qi, Fei ......................................................................................................................................... 50
Ren, Huiman ............................................................................................................................. 131
Schneider, Liesel .......................................................................................................................... 3
She, Deyong ............................................................................................................................. 342
Shi, Zhendan ............................................................................................................................. 131
Shi, Zhengxiang ...................................................................................................... 50, 68, 84, 106
Shi, Zhifang ...................................................................................................................... 214, 252
Shu, Yufu.................................................................................................................................. 222
Sun, Aidong .............................................................................................................................. 131
Sun, Yuxuan ............................................................................................................................... 42
Tabler, Tom .............................................................................................................................. 209
Teng, Guanghui ................................................................................................................ 171, 238
Turner, Ian ................................................................................................................................ 138
Voy, Brynn ................................................................................................................................... 3
Wang, Chaoyuan .................................................................................................................. 34, 76
Wang, Hao ................................................................................................................................ 171
Wang, Hua ................................................................................................................................ 342
Wang, Jianxing ......................................................................................................................... 246
Wang, Jing ................................................................................................................................ 246
Wang, Kaiying ............................................................................................................................ 26
Wang, Leiping ............................................................................................................................ 91
Wang, Liwei ............................................................................................................................. 350
Wang, Meizhi ........................................................................................................................... 342
Wang, Ning .............................................................................................................................. 114
Wang, Shaojie............................................................................................................................. 76
Wang, Shouyi ............................................................................................................................. 18
Wang, Xiaoshuai ........................................................................................................................ 26
Wei, Haidong.................................................................................................................... 193, 230
Wen, Yanbin ............................................................................................................................. 171
Wu, Mengru...................................................................................................................... 310, 326
Wu, Xuefei ............................................................................................................................... 106
Wu, Zhonghong ........................................................................................................................ 342
Xi, Lei............................................................................................................................... 214, 252
Xie, Qiuju ......................................................................................................... 155, 310, 326, 350
Xie, Tian ..................................................................................................................................... 26
Xu, Qiyong ............................................................................................................................... 114
Xu, Weitao................................................................................................................................ 269
Xu, Yidan ................................................................................................................................. 350
Xue, Qingfei ............................................................................................................................. 318
Xue, Xianglong......................................................................................................................... 302
Xue, Xiaoliu ............................................................................................................................. 342
Yan, Keping................................................................................................................................ 98
Yan, Xiaojie.............................................................................................................................. 201
Yang, Muyu ...................................................................................................................... 310, 326
Yang, Xiao........................................................................................................................ 209, 214
Yang, Zhixiao ........................................................................................................................... 214
∙ 383 ∙
Yu, Haiming ............................................................................................................................. 326

Yu, Hanlin ................................................................................................................................ 230
Yu, Ligen .................................................................................................................. 302, 318, 334
Yu, Qinyang ............................................................................................................................. 334
Zeng, Li ...................................................................................................................................... 10
Zhang, Guochen ....................................................................................................................... 163
Zhang, Hanbing ........................................................................................................................ 163
Zhang, Hengyi .......................................................................................................................... 230
Zhang, Jianlong ........................................................................................................................ 238
Zhang, Jicheng.......................................................................................................................... 155
Zhang, Jingfeng ........................................................................................................................ 214
Zhang, Jinrui ............................................................................................................................. 342
Zhang, Qian .............................................................................................................................. 163
Zhang, Qiang ............................................................................................................................ 201
Zhang, Runxiang ...................................................................................................... 193, 230, 260
Zhang, Xiaohong ...................................................................................................................... 222
Zhao, Qian ................................................................................................................................ 230
Zhao, Wanying ........................................................................................................................... 68
Zhao, Xuedong ........................................................................................................................... 84
Zhao, Yang ........................................................................................................................... 3, 209
Zheng, Ping ...................................................................................................................... 155, 326
Zheng, Weichao .......................................................................................................................... 42
Zhou, Ling .................................................................................................................................. 42
Zhou, Mengting ........................................................................................................................ 186
Zhu, Chenxuan ........................................................................................................................... 84
Zhu, Kunhua ............................................................................................................................. 214
Zhuang, Yanrong ...................................................................................................................... 238
Zong, Chao ............................................................................................................................... 171
Zulovich, Joseph ....................................................................................................................... 371
∙ 384 ∙
Author Affiliation Index

Author Affiliation Page
Agriculture and Agri-Food Canada, Sherbrooke Research and Development

Centre, Sherbrooke, Canada ................................................................................................. 201
Beijing Academy of Agriculture and Forestry Sciences, Beijing, China .................................. 334
Beijing Engineering Research Center for Animal Healthy Environment, Beijing,
China ................................................................................................................................ 42, 50
Beijing Research Center for Information Technology in Agriculture, Beijing,
China ............................................................................................................................ 302, 318
Catholic University Leuven, Department of Biosystems, Belgium .......................................... 358
China Agricultural University, College of Animal Science and Technology,
Beijing, China ....................................................................................................................... 342
China Agricultural University, College of Information and Electrical
Engineering, Beijing, China.................................................................................................. 302
China Agricultural University, College of Water Resources and Civil
Engineering, Beijing, China....................................... 34, 42, 50, 68, 76, 84, 106, 171, 238, 342
Chinese Academy of Sciences, Center for Excellence in Regional Atmospheric
Environment & Key Laboratory of Urban Environment and Health, Institute
of Urban Environment, Xiamen, China .................................................................................. 91
Chongqing Academy of Animals Sciences, Rongchang, China ............................................... 171
City University of Hong Kong, Department of Computer Science, Hong Kong,
China ............................................................................................................................ 269, 287
City University of Hong Kong, Department of Infectious Diseases and Public
Health, Hong Kong, China ........................................................................................... 269, 287
Dalian Ocean University, Dalian, China................................................................................... 163
Foshan Renshi Machinery Technology Co., Ltd., Foshan, China ............................................. 131
Henan Engineering Research Center for Animal Healthy Environment and
Intelligent Equipment, Zhengzhou, China ............................................................................ 214
Henan University of Animal Husbandry and Economy, College of Animal
Science and Technology, Zhengzhou, China ................................................................ 214, 252
Henan University of Animal Husbandry and Economy, School of Energy and
Intelligence Engineering, Zhengzhou, China ........................................................................ 214
Jiangsu Academy of Agricultural Sciences, Laboratory of Animal Improvement
and Reproduction, Institute of Animal Science, Nanjing, China .......................................... 131
Jiangsu Key Laboratory for Food Quality and Safety-State Key Laboratory
Cultivation Base of Ministry of Science and Technology, Nanjing, China .......................... 131
Jiangsu Lihua animal husbandry Co., Ltd, Changzhou, China ................................................... 42
Ministry of Agriculture and Rural Affairs, Beijing, China ................................................... 238
Kyushu University, Graduate School of Bioresource and Bioenvironmental
Science, Fukuoka, Japan ....................................................................................................... 179
Ministry of Agriculture and Rural Affairs, Key Laboratory of Agricultural
Engineering in Structure and Environment, Beijing, China .................... 34, 42, 68, 76, 84, 106
Ministry of Agriculture and Rural Affairs, Key Laboratory of Chicken Genetics
and Breeding, Harbin, China ................................................................................................ 230
Ministry of Agriculture and Rural Affairs, Key Laboratory of Equipment and
Informatization in Environment Controlled Agriculture, Hangzhou, China ........................... 91
∙ 385 ∙
Ministry of Agriculture and Rural Affairs, Key Laboratory of Protected

Agriculture Engineering in the Middle and Lower Reaches of Yangtze River,
Nanjing, China ...................................................................................................................... 131
Mississippi State University, Department of Poultry Science, Mississippi State,
USA ...................................................................................................................................... 209
Nanjing Agricultural University, College of Animal Science and Technology,
Nanjing, China ................................................................................................................ 58, 179
National Engineering Research Center for Information Technology in
Agriculture, Beijing, China ........................................................................................... 318, 334
Northeast Agricultural University, College of Animal Science and Technology,
Harbin, China ........................................................................ 155, 193, 222, 230, 260, 310, 326
Northeast Agricultural University, College of Electrical and Information,
Harbin, China ....................................................................................................... 155, 310, 326
Northeast Agricultural University, College of Life Science, Harbin, China ............. 193, 230, 246
Peking University Shenzhen Graduate School, Shenzhen Engineering
Laboratory for Eco-efficient Recycled Materials, Shenzhen, China ..................................... 114
Pennsylvania State University, Department of Agricultural and Biological
Engineering, State College, USA.......................................................................................... 123
Pig Improvement Company, Shanghai, China ............................................................................ 76
Qingdao Big Herdsman Machinery Co., Ltd, Qingdao, China ................................................... 26
R&D Center of Fisheries Equipment and Engineering of Liaoning Province,
Dalian, China ........................................................................................................................ 163
Shanghai Academy of Agricultural Sciences, Shanghai, China.................................................. 84
South China Agricultural University, Colleges of Electronic Engineering and
Artificial Intelligence, Guangzhou, China ............................................................................ 287
Sun Yat-sen University, School of Ecology, Shenzhen, China................................................. 114
The University of Tennessee, Department of Animal Science, Knoxville, USA .......... 3, 209, 214
The University of Tennessee, Department of Biosystems Engineering and Soil
Science, Knoxville, USA ...................................................................................................... 209
Tianjin Academy of Agricultural Sciences, Tianjin, China ...................................................... 123
Tianjin Agricultural University, College of Computer and Information
Engineering, Tianjin, China .................................................................................................. 318
University of British Columbia, Kelowna, Canada................................................................... 138
University of Manitoba, Department of Animal Science, Winnipeg, Canada .......................... 201
University of Manitoba, Department of Biosystems Engineering, Winnipeg,
Canada .................................................................................................................................. 201
University of Milan, Department of Environmental Science and Policy, Milan,
Italy....................................................................................................................................... 295
University of Missouri, Agricultural Systems Technology, Columbia, Missouri,
USA ...................................................................................................................................... 371
University of Tennessee, Department of Biosystems Engineering and Soil
Science, USA ........................................................................................................................ 358
USDA, Agriculture Research Service, Poultry Research Unit, Mississippi State,
USA .......................................................................................................................................... 3
Wageningen University and Research, Department of Biosystems Engineering,
Wageningen, The Netherlands .............................................................................................. 186
West Anhui University, College of Biological and Pharmaceutical Engineering,
Luan, China .......................................................................................................................... 342
Zhaoqing University, College of Computer Science and Software, College of
Big Data, Zhaoqing, China ................................................................................................... 350
∙ 386 ∙
Zhaoqing University, College of Mechanical and Automotive Engineering,

Zhaoqing, China ................................................................................................................... 350
Zhejiang Ocean University, School of Petrochemical and Environment,
Zhoushan, China ..................................................................................................................... 98
Zhejiang University, College of Biosystems Engineering and Food Science,
Hangzhou, China .............................................................................10, 18, 26, 91, 98, 146, 277
Zhejiang University, College of Chemical and Biological Engineering,
Hangzhou, China .................................................................................................................... 98
∙ 387 ∙
∙ 388 ∙
20 mm
Animal Environment and Welfare —Proceedings of International Symposium 2021

http://isaew2021.aconf.org
Cover design by Ji-Qin Ni

Cover photographs by Ji-Qin Ni
View publication stats

Animal Environment and Welfare-2021ISAEWProceedings

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Animal Environment and Welfare-2021ISAEWProceedings

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

Animal Environment and Welfare — Proceedings of International Symposium

Book · October 2021

Ji-Qin Ni Lingjuan Wang-Li

SEE PROFILE SEE PROFILE

Heat and AQ stress in dairy cattle View project

The user has requested enhancement of the downloaded file.

Chongqing Academy of Animal Sciences

Animal Environment and Welfare

International Research Center for

Prof. Jay Johnson USDA-ARS USA

Disclaimer: It is the authors’ responsibility to obtain written permissions to use copyrighted

© 2021 of individual papers by the authors of the papers

Welcome to ISAEW 2021

Although the COVID-19 situation is improving to various degrees globally, it remains

Hongwei Xin, Symposium Co-chair Baoming Li, Symposium Co-chair

page with the best possible visual communication.

Theme I: Animal Environment Monitoring, Assessment, and Control .................. 1

Disinfection Effect of Slightly Acidic Electrolyzed Water Applied in

Theme II: Animal Production Systems and Equipment ...................................... 121

Relationship between Stereotypic Behaviors, Physiological Characteristics,

Why Precision Livestock Farming Can Generate a More Sustainable

Theme V: Manure and Mortality Management and Byproduct Development in

Indexes ...................................................................................................................... 379

Animal Environment Monitoring,

Assessment, and Control

Survival of Escherichia Coli in Airborne and

Effects of Monochromatic Blue Light on Reducing the

1.5 **** H+W

Light Demand Characteristics, Production Performance and

Figure 1. Structure diagram and real shots of breeding area.

Fsd  1  Fsl (4)

Figure 2. The average weight of broilers in different rearing densities.

Figure 4. The distributions of broilers in different rearing densities.

A Comparison of Ventilation Systems in Three Different Types of

To date, few researchers have conducted systematic evaluations on the ventilation

Concentration and Emission of Particle Matters from an

Figure 1. The sampling locations inside and outside the Barn A.

3.1.2. Diurnal patterns

Figure 3. Diurnal patterns of ambient and indoor PM concentration and RH.

Comparative Assessment of the Breeding Egg Disinfection Using

Figure 1. Self-made spray equipment.

Pharmaceutical Co., Ltd, Shanghai, China).

3. Results and Discussion

Figure 2. Effect of different disinfectant and disinfection technology on bactericidal effect.

Figure 3. Effect of different disinfectant treatments on eggshell strength and thickness of

Heating Time and Heating Load of Fattening Pig Houses

a. Nursery pig house b. Finishing pig house

a. Nursery pig house b. Finishing pig house

A Survey and Analysis of Antibiotic Resistance of Escherichia Coli

Table 1. Characteristics of the included studies.

Figure 1. Meta-analysis of differences in resistance of E. coli. A: ciprofloxacin; B: gentamicin;

Effect of Ventilation Fans and Types of Partition on the

Figure 1. Schematic diagram of experimental design groups.

Table 3. Immunity parameters of calves.

Hill, T. M., H. G. Bateman, J. M. Aldrich, R. L. Schlotterbeck. 2011. Comparisons of housing, bedding,

Assessing Air Quality in Three Types of Laying Hen Houses Using

Figure 1. System frame for wireless monitoring chicken house environment.

2.2. Hardware Design

Figure 2. Variations of different environmental parameters in three poultry production systems.

Disinfection Effect of Slightly Acidic Electrolyzed Water Applied in

Table 1. Disinfection effects of SAEW with different ACCs.

Figure 1. Disinfection efficiency of SAEW with different ACCs.

Figure 4. Milk yield of cows of the experimental and control groups.