ArticlePDF Available

Machine learning based fog computing assisted data-driven approach for early lameness detection in dairy cattle

April 2020
Computers and Electronics in Agriculture 171:105286

DOI:10.1016/j.compag.2020.105286

License
CC BY-NC-ND 4.0

Authors:

Mohit Taneja

TSSG

John Byabazaire

University College Dublin

Nikita Jalodia

Walton Institute of Information and Communication Systems Science

Show all 6 authorsHide

Timely lameness detection is one of the major and costliest health problems in dairy cattle that farmers and practitioners haven't yet solved adequately. The primary reason behind this is the high initial setup costs, complex equipment and lack of multi-vendor interoperability in currently available solutions. On the other hand, human observation based solutions relying on visual inspections are prone to late detection with possible human error, and are not scalable. This poses a concern with increasing herd sizes, as prolonged or undetected lameness severely compromises cows' health and welfare, and ultimately affects the milk productivity of the farm. To tackle this, we have developed an end-to-end IoT application that leverages advanced machine learning and data analytics techniques to monitor the cattle in real-time and identify lame cattle at an early stage. The proposed approach has been validated on a real world smart dairy farm setup consisting of a dairy herd of 150 cows in Waterford, Ireland. Using long-range pedometers specifically designed for use in dairy cattle, we monitor the activity of each cow in the herd. The accelerometric data from these sensors is aggregated at the fog node to form a time series of behavioral activities, which are further analyzed in the cloud. Our hybrid clustering and classification model identifies each cow as either Active, Normal or Dormant, and further, Lame or Non-Lame. The detected lameness anomalies are further sent to farmer's mobile device by way of push notifications. The results indicate that we can detect lameness 3 days before it can be visually captured by the farmer with an overall accuracy of 87%. This means that the animal can either be isolated or treated immediately to avoid any further effects of lameness. Moreover, with fog based computational assistance in the setup, we see an 84% reduction in amount of data transferred to the cloud as compared to the conventional cloud based approach.

Mobile application developed specifically considering the needs of the farmer, including an offline first strategy. The figure presents data and notification flow in the developed mobile application.

…

Microservices based application flow for integration of services from different service providers.

…

Density distribution plot comparing the Normal and Lame profiles.

…

Comparing the distribution of cow 2463 against the normal and lame profile at three different stages.

…

Comparing the distribution of cow 1469 against the normal and lame profile at three different stages.

…

Figures - available via license: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International

Content may be subject to copyright.

Available via license: CC BY-NC-ND 4.0

Content may be subject to copyright.

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Machine learning based fog computing assisted data-driven approach for

early lameness detection in dairy cattle

Mohit Taneja

a,b,⁎

, John Byabazaire

a,b

, Nikita Jalodia

a,b

, Alan Davy

a,b

, Cristian Olariu

Paul Malone

Emerging Networks Laboratory, Telecommunications Software and Systems Group, Department of Computing and Mathematics, School of Science and Computing,

Waterford Institute of Technology, Waterford, Ireland

CONNECT- Centre for Future Networks and Communications, Dublin, Ireland

Innovation Exchange, Dublin, IBM, Ireland

ARTICLE INFO

Keywords:

Smart dairy farming

Fog computing

Internet of Things (IoT)

Cloud computing

Smart farm

Data analytics

Microservices

Machine learning

Clustering

Classiﬁcation

Data-driven

ABSTRACT

Timely lameness detection is one of the major and costliest health problems in dairy cattle that farmers and

practitioners haven't yet solved adequately. The primary reason behind this is the high initial setup costs,

complex equipment and lack of multi-vendor interoperability in currently available solutions. On the other hand,

human observation based solutions relying on visual inspections are prone to late detection with possible human

error, and are not scalable. This poses a concern with increasing herd sizes, as prolonged or undetected lameness

severely compromises cows' health and welfare, and ultimately aﬀects the milk productivity of the farm. To

tackle this, we have developed an end-to-end IoT application that leverages advanced machine learning and data

analytics techniques to monitor the cattle in real-time and identify lame cattle at an early stage.

The proposed approach has been validated on a real world smart dairy farm setup consisting of a dairy herd of

150 cows in Waterford, Ireland. Using long-range pedometers speciﬁcally designed for use in dairy cattle, we

monitor the activity of each cow in the herd. The accelerometric data from these sensors is aggregated at the fog

node to form a time series of behavioral activities, which are further analyzed in the cloud. Our hybrid clustering

and classiﬁcation model identiﬁes each cow as either Active, Normal or Dormant, and further, Lame or Non-

Lame. The detected lameness anomalies are further sent to farmer's mobile device by way of push notiﬁcations.

The results indicate that we can detect lameness 3 days before it can be visually captured by the farmer with an

overall accuracy of 87%. This means that the animal can either be isolated or treated immediately to avoid any

further eﬀects of lameness. Moreover, with fog based computational assistance in the setup, we see an 84%

reduction in amount of data transferred to the cloud as compared to the conventional cloud based approach.

1. Introduction

Internet of things (IoT), fog computing, cloud computing and data

driven techniques together oﬀer a great opportunity for verticals such

as the dairy industry to increase productivity by getting actionable in-

sights to improve farming practices, thereby increasing eﬃciency and

yield. There has been active initiation and movement in the agricultural

domain to move towards tech-enabled smart solutions to improve

farming practices, and increase productivity and yield. The concept of

Smart Dairy Farming is no longer just a futuristic concept, and has

started to materialize as diﬀerent ﬁelds such as machine learning have

found a practical applications in this domain.

Timely detection of lameness is a big problem in the dairy industry

which farmers are not yet able to adequately solve. It is one of the

factors for reduced performance on many dairy farms, at least through

reduced reproductive eﬃciency, milk production and increased culling

(Chapinal et al., 2009). Lameness is considered to be the third disease of

economic importance in dairy cows after reduced fertility and mastitis

(Van Nuﬀel et al., 2015). An all-encompassing deﬁnition of lameness

includes any abnormality which causes a cow to change the way that

she walks, and can be caused by a range of foot and leg conditions,

themselves caused by disease, management or environmental factors

(AHDB, 2016). Prevention, early detection and treatment of lameness is

therefore important to reduce these negative eﬀects of lameness in

https://doi.org/10.1016/j.compag.2020.105286

Received 16 September 2019; Received in revised form 6 February 2020; Accepted 8 February 2020

⁎

Corresponding author.

E-mail addresses: mtaneja@tssg.org (M. Taneja), byabazairej@gmail.com (J. Byabazaire), njalodia@tssg.org (N. Jalodia), adavy@tssg.org (A. Davy),

rioderaca@gmail.com (C. Olariu), pmalone@tssg.org (P. Malone).

Computers and Electronics in Agriculture 171 (2020) 105286

(http://creativecommons.org/licenses/BY-NC-ND/4.0/).

dairy cows (Alsaaod et al., 2015; Poursaberi et al., 2011). Early de-

tection of disease allows farmers to intervene earlier, leading to pre-

vention of antibiotic administration and improvement in the milk yield,

and savings on veterinary treatment for their herd.

With the increasing global demands of agricultural and dairy pro-

duce, the scale of farming and livestock management is only set to in-

crease. To increase the productivity on a dairy farm, farmers generally

look towards the following two:

1. Control welfare related issues like lameness

2. Increase the number of accurate estrus detection, so as to increase

the size of the herd for further proﬁtability.

The latter has been adequately addressed. In fact, there are major

industry partnerships like Fujitsu partnering with Microsoft to provide a

SaaS (Software as a Service) for accurate estrus detection powered by

Microsoft's Azure and Analytics (Meet the connected cow; Hadoop,

2015). So, the next major problem to be solved in smart dairy farming is

the ability to accurately and timely detection of lameness.

Our work is motivated by the following facts:

•Manual human observation based solutions for lameness detection

are susceptible to human error and are not scalable as the size of the

farm increases.

•Few of the existing solutions (discussed in the next section) require

the equipment in use for detection to be placed in a controlled po-

sition, and the cows need to be constrained to walk through them.

Guiding the cows to walk in a controlled manner and presence of a

human being leads to bias in the measurement because of stoic nature

of cows; as they will try to hide their weakness and pain compared to

measurements made during normal routine without the presence of a

human.

Although there are existing wearable sensor based cloud centric and

isolated oﬄine solutions, they suﬀer from the issue of multi-vendor

interoperability and vendor-lock in, which has been detailed in the next

section. Therefore, there is still a need for further automation and ad-

vanced machine learning based solution that:

•Monitors the animals everywhere they are – either in the ﬁelds

grazing, during milking, or lying down in the shed.

•Is Population Agnostic – takes into account individual animal be-

haviour.

•Is Environment Agnostic – takes into account variations in weather

and environment.

In this article, we present an end-to-end IoT application that le-

verages threshold based clustering and machine learning classiﬁcation

to detect lameness in dairy cattle. The application automatically mea-

sures and gathers activity data (lying time, step count and swaps per

hour) continuously, so that cows can be monitored daily. Furthermore,

the clustering technique employed ensures that the models dynamically

adjust depending on farm and weather conditions, and automatically

selects a custom learning model for that cluster.

Our contribution and core novelty of the work presented is sum-

marized as follows:

1. Although the existence of clusters in herds have been used before in

cattle behaviours (Stephenson and Bailey, 2017), this study is the

ﬁrst to use cluster speciﬁc model for lameness detection as opposed

to a one-size-ﬁts-all solution.

2. The technique used to form animal proﬁles eliminates the eﬀects of

external factors like weather, location and farm conditions. This

study is the ﬁrst to oﬀer an early lameness detection service that is

population and environment agnostic.

3. The study is the ﬁrst to propose a feedback based re-training based

on the inputs of the human in the loop, who could be an agricultural

expert or farmer.

4. The study is amongst one of the few to apply modern AI (Artiﬁcial

Intelligence) techniques to detect lameness in dairy cattle.

5. The methods are deployed in practice and real data has been col-

lected. The proposed approach has been validated in a real-world

IoT deployment in a Smart Dairy farm setup with a full herd of 150

dairy cows.

The paper has been further structured as follows: 2 presents the

literature review, background, state-of-the-art and motivation, 3 pre-

sents the experimental setup, system architecture and application de-

sign, 4 presents materials, methods and machine learning model de-

veloped, 5 presents discussion of the results, and ﬁnally 6 and 7 present

the conclusion, Ongoing and future work respectively.

2. Literature Review, background and motivation

2.1. Lameness in dairy cattle: geographical variance and associated costs

The prevalence of lameness has been reported diﬀerently in dif-

ferent regions and states. Ger reported that on an average Irish farm, 20

in every 100 cows will be aﬀected by lameness in a given year. In the

United States authors in (Cook, 2003) and (Espejo et al., 2006) reported

a mean lameness prevalence of 25%, whereas in California and the

northeastern United States, overall lameness prevalence was estimated

to be 34% and 63%, respectively (von Keyserlingk et al., 2012). British

and German studies reported a lameness prevalence of 37% and 48%

(Whay et al., 2003; Barker et al., 2010), whereas a prevalence of 16%

was reported in the Netherlands (Amory et al., 2006). In our experi-

mental deployment on a herd of 150 cows in Ireland, 26 cases of la-

meness were recorded during July to December 2017. Further up by the

end of the experiment in April 2018, a total of 32 cases were recorded

overall.

Lameness can be classiﬁed into three main categories: solar ulcers,

digital disease (white line abscess, foreign bodies in the sole, and

pricked or punctured sole), and inter digital disease (lesions of the skin

between claws and heel including foul in the foot, inter digital ﬁbroma

and dermatitis). More than 65% of cases of lameness are said to be

caused by diseases (Foditsch et al., 2016). Other causes include injuries

to the upper skeleton or major muscles, septic joints and injection site

lesions.

Lameness has many negative eﬀects, including reduction in feed

intake, reduction in milk production (mainly due to withdrawal due to

antibiotics usage) and weight loss. It therefore has a drastic eﬀect on the

performance of a dairy farm. Lameness is mostly detected at advanced

stage and thus requires immediate and often costly treatment. Once an

animal becomes lame, it can take several weeks to recover. Lameness

thus represents a signiﬁcant cost to dairy farmers in terms of time, ﬁ-

nancial expenditure for veterinary calls, medication and treatment, and

also for loss in production. Table 1 (Ger) summarizes the costs estimates

for each type of lameness.

2.2. Existing approaches for lameness detection

2.2.1. Pressure plate/load cell

In these solutions, the main aim is to investigate how the weight is

Table 1

Costs associated with each type of lameness.

Type of lameness Digital Inter digital Solar ulcer Average case

Prevalence (%) 45 35 20

Total cost of a single case

(in €) 282.85 136.12 504.58 275.26

M. Taneja, et al. Computers and Electronics in Agriculture 171 (2020) 105286

distributed across the legs of the animal as it walks through a marked

area. Neveux et al. (2006) studied the use of a platform outside the

automatic milking system to measure the weight distribution of cows

while standing on diﬀerent surfaces. Chapinal et al. (2010) and Pastell

et al. (2010) later adjusted the experimental setup to measure lameness

and hoof lesions.

The drawbacks of such solutions may not be only the costs of new

and complex equipment but also other technical concerns. For example,

Pastell et al. (2010) suggested that a cow may suﬀer pain when

walking, which is not as obvious when the cow is standing still. In this

setup, the cow must be guided or must be standing in controlled posi-

tion. Because cows are stoic in nature, this will aﬀect the measurements

and alter the results.

2.2.2. Image processing techniques

This category studies the use of image processing techniques to

analyse the posture of the animal as it walks through a milking parlour.

Poursaberi et al. (2010) proposed a method based on detecting the arc

of back posture and ﬁtting a circle through selected points on the spine

line of a cow as it walks. Viazzi et al. (2013) further studied the idea

and an algorithm based on Body Movement Pattern was tested under

farm conditions. Further study on this method shows that there remain

challenges on real farm conditions. For example, changing lighting

conditions cause noise and shadows in the images that impede extrac-

tion of the back posture, or continuous background changes that in-

terfere with cow segmentation from the images. Some of these chal-

lenges were explored by Poursaberi et al. (2009), Van Hertem et al.

(2013) and Viazzi et al. (2014).

2.2.3. Activity based techniques

Here, techniques use accelerometers (2D and 3D) and pedometers to

record movement patterns of the animal. This data is then used to build

the daily activities of the cow, e.g. walking, lying down. Munksgaard

et al. (2006) proposed the use of sensors that measure acceleration in

diﬀerent dimensions to automatically monitor activity (standing and

lying behaviour) of cows. Their results indicate excellent accuracy be-

tween the sensor data attached to the legs of the cows and observations

for lying and standing (0.99), activity (0.89), and number of steps

(0.84). Chapinal et al. (2011) used ﬁve 3D accelerometers on cows, one

on each limb, and concluded a single device attached to one of the legs

appeared to be suﬃcient to measure the walking speed of cows, which

was associated with locomotion scores. In other studies accelerometers

were mounted on a hind leg of 348 cows in 401 lactations on four

commercial farms (Thorup et al., 2015). Since then, a vast number of

studies have used accelerometers to measure dairy cow activity and

behaviour (Alsaaod et al., 2012; Yunta et al., 2012; O'Driscoll et al.,

2008; Blackie et al., 2011).

2.3. IoT, fog computing and data analytics in agriculture domain

There have been proposed systems in industry (Ag, 2017; Ireland,

2017; Boumatic, 2017) as well in academia (Taylor et al., 2013; Chen

et al., 2014; Wark et al., 2009; Brewster et al., 2017) for animal health

management in dairy farms. Authors in (Al-Fuqaha et al., 2015) provide

a detailed survey of IoT enabling technologies that can oﬀer automa-

tion, data aggregation and protocol adaptation in the wide ﬁeld of IoT.

They also present the required integration of IoT with emerging tech-

nologies such as data analytics and fog computing. Another survey in

(Rutten et al., 2013) identiﬁes a serious lack of analytics and in-

telligence in existing smart dairy farming systems, thus leading to gaps

between the desired requirement of the system and proposed solutions.

It articulates the pressing requirement of intelligence to be present on

the premises, in the on-farm systems. As a consequence, attention is

being drawn towards designing systems with intelligence and data

analytics capability being present on premises (Shi et al., 2016), and

utilizing fog computing comes to shore with those objectives in mind.

Authors in (Muangprathub et al., 2019) present the need of data

driven movement in agriculture in order to improve crop yields, im-

prove quality, and reduce costs. Another recent survey by authors in

(Jukan et al., 2017) identiﬁes the lack of interoperability provided by

such systems, and the need of developing an integrated system com-

bining edge, fog and cloud to provide application and services. The

authors here also identify that technology solutions with no con-

sideration of interoperability results in vendor lock-in, which not only

hinders innovation, but also results in higher costs for the farmer/user.

Authors in (Caria et al., 2017) present the use of Raspberry Pis as

edge devices which are further connected with the cloud to demon-

strate a smart farm computing systems for animal welfare monitoring.

Authors demonstrated that a low-cost and open computing and sensing

system can eﬀectively monitor multiple parameters related to animal

welfare. While animal welfare remains a broad concept, their paper

shows that many parameters relevant to various stakeholders can be

measured, collected, evaluated and shared, opening up new possibilities

to improve animal welfare and foster high-tech innovations in this

sector.

The authors in (Hsu et al., 2018) propose the use of fog computing

for innovative service creations for existing cloud based agriculture

system. By means of simulation, the authors demonstrate that fog

computing presents a unique capability for a creative IoT platform

adoption in agriculture with existing cloud support. Quite recently,

authors in (Zamora-Izquierdo and Santa, 2019) describe the design,

development and evaluation of a system that covers extreme precision

agriculture requirements by using automation, IoT technologies, and

edge and cloud computing through virtualisation.

2.4. How is the proposed system novel as compared to the existing ones?

Although most farms are equipped with some kind of estrus detec-

tion system (Miiaková et al., 2018) which is based on accelerometers,

lameness detection systems based on the same have not been successful.

This is because of vendor lock-in. Each of the system would require its

own hardware. Worth mentioning is the insight in the review by Van De

Gucht et al. (2017) that farmers who already had an estrus detection

system were willing to have an add-on for lameness detection. All the

current systems lack this kind of integration. Another assumption made

by all the current solutions is that all the animals will get lame the same

way. To put this into context, consider a farm in Ireland – there are two

main seasons, summer and winter. During summer, the animals stay in

the ﬁeld and graze freely, while during the winter the animals are kept

in house. In both cases the activity levels of the animals are diﬀerent.

Interestingly, as the activity levels are low during winter considering

the animals stay indoors rather than having much outdoor activity, the

entire herd's activity pattern will be closer to that of a lame animal

during the summer. Therefore, a learning model should be able to

consider such external factors.

In a review by authors in (Van Nuﬀel et al., 2015) about automatic

lameness detection, some suggestions were made. One of these was the

need for automatic and continuous measurement of the parameters.

This is because most solutions available require the animals to be

guided one way or another. The other suggestion was the need for

custom solutions, systems that need less space or those that can be

included in the existing farm infrastructure (Van Nuﬀel et al., 2015).

The presented work diﬀers from the existing sensor based system

solutions by oﬀering following advantages:

•Sensor agnostic: The model is built to take in activity data from any

kind of sensor used to monitor activity of the animal. This among

other thing will reduce the initial installation costs if a farm already

has a system in place.

•Avoids vendor lock-in: Design, creation and development of services

following a microservices based application design principles to

tackle the problem of vendor lock-in and to support multi-vendor

M. Taneja, et al. Computers and Electronics in Agriculture 171 (2020) 105286

interoperability.

•Multiple end-users: Since our system is designed as a service, this

makes it easy to integrate with the existing systems. The end user

therefore could be a farmer with an existing system or even an agri-

tech service provider who wants to provide more services to their

clients.

One of the primary limitations of the previously proposed systems is

that they follow the technique to process and analyse previously col-

lected data and perform only cloud based analytics without leveraging

and eﬃciently utilizing the resources (Taneja and Davy, 2016) avail-

able on the farm along the things-to-cloud continuum (Taneja and

Davy, 2017), moreover such techniques are not always suitable for real-

time tracking and monitoring of dynamic entities such as dairy cows.

The gaps with the existing research is that either it has been developed

out of the agricultural context, or addresses the issue of analytics and

control in isolation; this has also been identiﬁed as key limitations by

authors in (Zheleva et al., 2017).

While there has been a movement towards data-driven agriculture

in recent times for sustainable and productive growth, there is still a

void when it comes to leveraging emerging paradigms such as fog

computing, and applying innovating machine learning models to solve

a speciﬁc problem in the dairy sector. Most of the articles in literature

present results based on simulated experiments, and those which come

from real world deployment are mostly agriculture based, and rarely

based on dairy farming. Further, only some of them have a machine

learning element to automate their approach. However, to the best of

our knowledge, no prior work focuses on providing an end-to-end IoT

solution integrating edge, fog and cloud intelligence speciﬁcally in case

of smart dairy farming IoT settings.

We position our work as an answer to the issues mentioned above,

thus bridging the gap, and providing an innovative way that integrates

edge, fog, cloud computing and machine learning to provide a solution

speciﬁcally in case of smart dairy farming in an IoT setup. The novelty

of the proposed model comes from the standpoint that it has been

speciﬁcally designed and developed to address a speciﬁc vertical of the

IoT ecosystem i.e., dairy farming, and within that to address a speciﬁc

problem related to animal welfare i.e., detecting lameness at an early

stage before the clinical signs of it appear, with a microservices oriented

design making it multi-vendor interoperable.

3. Experimental setup – smart dairy farm setup: real world test-

bed deployment

As part of the experiment, the trial

was conducted on a local dairy

farm with a full dairy herd of 150 cows in Waterford, Ireland. Amongst

the available options for the sensors/wearables available for livestock

monitoring, we used commercially available radio based Long Range

Pedometer (433 MHz, ISM band, LRP — ENGS Systems

©®

, Israel) in our

deployment. These pedometers were attached to the front leg of cows in

the herd, as shown in Fig. 1.

3.1. Architecture design and system overview

The overall architecture of the test-bed is shown in Fig. 2. As shown

in Fig. 2, the Receiver is the master unit which sends the received data

to the communication unit (RS485 to USB) through wired connection,

which in turn then sends it to the gateway (a PC form factor device in

our case, which acts as controller and fog node. The conﬁguration

used

is Intel®Core™ 3rd Generation i7-3540 M CPU @ 3.00 GHz, 16.0 GB

RAM, 500 GB Local Storage) through wired connection via USB inter-

face. The fog node consists of a local database which stores all the data

from the sensors before it is preprocessed. The collected raw data is

then preprocessed and aggregated at fog node to form behavioural

activities, and summed to form daily time series. In this study, we used

three behavioural activities (step count, lying time, swaps) for the

analysis with their description below:

1. Step count: This is the number of steps an animal makes.

Fig. 1. Long Range Pedometer (LRP) attached as a part of the experiment to the

front leg of the cows.

Fig. 2. Overall architecture of the test-bed and system overview.

The ethical approval for the experimentation was taken from Research

Ethics Committee of Waterford Institute of Technology, Ireland prior to the

deployment in July 2017.

Note that our results presented in Taneja et al. (2019a) suggest that the

system is fully capable to run with fog node of lower computational power

(footnote continued)

without compromising on quality of service. The results on resource utilization

at fog node and discussion on platform performance on using fog node with low

level computational power, and system resilience has been discussed in greater

detail in our work available at Taneja et al. (2019a).

M. Taneja, et al. Computers and Electronics in Agriculture 171 (2020) 105286

2. Lying time: The number of hours an animal spends lying down.

3. Swaps: This is the number of times an animal moves from lying

down to standing up.

We used Message Queue Telemetry Transport (MQTT) (MQTT,

2017) as the connectivity protocol between fog node and cloud (service

instances running on IBM Cloud) in our deployment setting. MQTT is an

open-source protocol originally invented and developed by IBM

(Getting to know MQTT, 2017). It is a lightweight publish-subscriber

model based protocol designed on top of the TCP/IP stack. It is speci-

ﬁcally targeted for remote location connectivity with characteristically

unreliable network environments such as high delays and low band-

width (Lee and Kim, 2013), which is one of the issues in remote farm

based deployments such as ours. Hence, we chose MQTT as the con-

nectivity protocol in our deployment.

The MQTT architecture comprises of two functional components,

namely MQTT clients (such as publishers and subscribers) and MQTT

broker (for mediating messages between publishers and subscribers). In

our setup these components are as follows:

•MQTT Publisher: Script running on fog node (i.e., local PC at farm)

•MQTT Broker: IBM Watson IoT Platform (as a service on IBM

Cloud)

•MQTT Subscriber: Application designed and hosted on IBM Cloud

Thus, the data from fog node after pre-processing, aggregation and

classiﬁcation as described above and shown in Figs. 2 and 3is streamed

to IBM Waston IoT platform using MQTT, the IBM Watson IoT platform

receives all these messages, and the MQTT subscriber listening to the

events of this broker picks up all the data and stores it in Cloudant

NoSQL JSON Database at IBM Cloud.

3.2. Designing and developing an IoT based software system: objectives and

challenges

Building an IoT application is an intricate process involving end-to-

end components, each of which is adapted to the use case being ad-

dressed. Generally speaking, an end-to-end IoT solution towards a

smart scenario involves the following steps:

•Connecting the Unconnected: This step involves the installation of

sensors on physical entities such as objects (both static or in mo-

tion), remote infrastructure or living entities towards achieving a

speciﬁed objective such as monitoring.

•Data Acquisition: This involves attaining the sensor data and

transferring it to the data analytics platform(s) to achieve actionable

insights for better decision making. This becomes a critical problem

in scenarios such as ours, wherein a farm location has little or no

Internet connectivity.

•Architecting, Integrating and Management: This crucial phase

involves key decisions on the software architecture and design

principles to be used during development of the system. Once

ﬁnalized, the next step is to integrate, optimize and manage the

computing system thus built, which is usually an ongoing process.

•Data Analytics: Once the data is at the desired platform (be it fog or

cloud), this part involves ﬁguring out how to analyze the data to get

the desired information to achieve the speciﬁed objective, given the

constraints.

We also had the same objectives in mind while developing the ap-

plication in our scenario, the ﬁrst three of these objectives have been

explained above and further below in this section, and the data ana-

lytics objective has been explained in greater detail in the next section.

The end-to-end data and work ﬂow of the developed application has

been presented in Fig. 3. The primary challenge is to design an end to

end IoT solution to meet the speciﬁed objective given the highly vari-

able, harsh and resource constrained environment in a smart dairy

farming setting. This includes making the system resilient and fault

tolerant to cope up with the variable farm environments, including

weather-based network outages and connectivity issues because of re-

mote location of the farm. A detailed discussion on test-bed deployment

challenges, technical challenges faced during deployment and devel-

opment, and critical decisions made, was presented in Taneja et al.

(2019b).

The use of fog computing brings eﬃciency and sustainability to the

overall IoT solution being proposed. In most cases, farms are located in

remote locations and can suﬀer from phases of low or no Internet/

network connectivity. In such adverse connectivity scenarios it becomes

ideal to process the data locally as much as possible and send the ag-

gregated or partial outputs over the internet to the cloud for further

enhanced analytical results. In view of this we design our solution

utilizing fog computing which aims to bring computation capabilities

closer to the source of data. The fog computing based approach leads to

eﬀective utilization of limited available resources (Taneja and Davy,

2017) and also leads to signiﬁcant reduction in the amount of the data

being transferred to the cloud.

3.3. Oﬄine-ﬁrst model for mobile application design

Farms are usually located in geographically remote locations facing

constrained network connectivity. Most of the IoT deployments in such

settings are faced with limited cellular coverage. Existing solutions are

mostly cloud based or completely oﬄine. This limits the farmers' ability

to interact with the application anytime and anywhere. The system

developed in this study has an oﬄine enabled strategy via the mobile

application and cloud dashboard. Fig. 4 shows the data ﬂow of the

oﬄine enabled design approach.

Once the model produces notiﬁcations, these are sent to the farmer's

mobile device as push notiﬁcations. On board the application is a

PouchDB (Pouch, 2019) database which synchronizes with the cloudant

database in IBM cloud using a REST API whenever a connection is es-

tablished. The application in general helps to achieve the following

tasks:

Fig. 3. Work ﬂow and data ﬂow in the test-bed deployment.

M. Taneja, et al. Computers and Electronics in Agriculture 171 (2020) 105286

•Push notiﬁcations: Whether on WiFi or limited cellular network, or

whether the application is open or not, these will go through each

time the status of the farm changes.

•Data annotation: During the training process, this feature was used

by the human operator to annotate the data. In our case, this was

done weekly by an agricultural science student.

•Feedback to improve model learning: When a notiﬁcation is gen-

erated, the farmer has the option of conﬁrming if the identiﬁed cow

is actually lame, or tag it as a false alarm or even report a missed

alert. All this information is sent back to the model to improve its

accuracy.

3.4. Microservices based application ﬂow for multi-vendor interoperability

Unlike the existing systems that are based on a monolithic design

approach, the application designed in this study follows a microservices

(Balalaie et al., 2016) based approach for design, creation and de-

ployment. The aim is to make the developed system as Application/

Software as a Service', which can be used by the service providers to

integrate with their existing systems. For example, an agri-tech com-

pany could be a service provider for any other solution such as mastitis

detection, who wants to expand their system or integrate any of the

services such as lameness or heat detection into their system. A visual

representation of such a possible integration is presented in the Fig. 5.

Feature engineering layer as shown in the Fig. 5 ensures that data is

transformed to output only the required features and also reject those

that cannot be engineered to form the required features for a desired

service; for example Lameness Detection and Heat Detection Service

expects lying time, step count and swaps but a service provider might

have activity counter instead of step count and (Stand up + Liedown)

instead of swaps.

It is important to note that this layer will be diﬀerent for each

service provider since the underlying sensor technology might be dif-

ferent. This in turn makes the developed system sensor agnostic. The

output from the feature engineering layer is then passed to the access

layer, which includes both mobile and web components. This then goes

through a REST API which in turn calls the desired service.

Fig. 4. Mobile application developed speciﬁcally considering the needs of the farmer, including an oﬄine ﬁrst strategy. The ﬁgure presents data and notiﬁcation ﬂow

in the developed mobile application.

Fig. 5. Microservices based application ﬂow for integration of services from

diﬀerent service providers.

M. Taneja, et al. Computers and Electronics in Agriculture 171 (2020) 105286

4. Materials, methods and machine learning model description

4.1. Data

The data from the sensors is sent via the receiver to the fog node,

where it is pre-processed and aggregated into three behavioural activ-

ities—(1) Step count, (2) Lying time, and (3) Swaps. The choice of these

3 features is guided by literature study, which indicates that they are

among the best predictors of a lame cow, or one transitioning to la-

meness (Thorup et al., 2015). The data is then summed to form daily

time series. Out of 150 cows used in the trial, only 146 cows were used

in the analysis. Only data from July to December 2017 was included in

this analysis. During this period, 26 animals were reported as lame

(cows were checked for lameness by either the agricultural science

student or by the farmer). Because the number of lame animals was

small, splitting the data into training and testing folds was made in a

such a way that atleast 75% of the lame animals was put in the training

fold and the rest in the testing fold. This was a challenge as the dataset

was imbalanced, but because this was a live experiment, we hoped to

re-train the models after sometime. The performance on both the

training and testing are reported in a later section.

Fig. 6 gives a quick overview of the end-to-end pipeline of the de-

veloped solution illustrating: (1) data collection from sensors, (2) ob-

servation of the herd by an animal expert for locomotion scoring, (3)

translating the human observer's expertise into a machine learning

based system leading to early detection of lameness in dairy cattle.

Table 2 presents the locomotion scoring scale system used by the

agricultural science student during animal observation.

4.2. Machine learning model and data analytics

4.2.1. Cow proﬁling

In order to build robust proﬁles that are distinguishable by the

learning model, it is important to understand how each test proﬁle

(lame and non-lame) relates to the rest of the herd. The most common

approach would be to compare the activity level of lame and non-lame

Fig. 6. A diagrammatic representation of the end-to-end pipeline of the developed solution illustrating: (1) data collection from sensors, (2) observation of animals by

an animal expert to give locomotion score, (3) translating the human observer's expertise into a machine learning based system leading to early detection of lameness

in cattle.

Table 2

Locomotion scoring scale system used by the agricultural science student while

observing cows.

Fig. 7. Comparing the Mean and Median of the various Animal Activities.

Fig. 8. Relationship between herd mean and cow activity for cow 2346.

M. Taneja, et al. Computers and Electronics in Agriculture 171 (2020) 105286

animals and investigate how these deviate from the mean of the entire

herd. However, the mean can be aﬀected by a single value being too

high or low compared to the rest of the sample. This is why a median is

sometimes taken as a better measure. Fig. 7 compares the mean and

median of the herd. The results show that these almost trace out each

other for all the three activities; lying time, step count and swaps. This

is one of the features of a normal distribution, and therefore it would

not matter whether the mean or median is used. Thus, we decided to

use herd mean in our analysis.

Authors in (Stephenson and Bailey, 2017) have argued that animals

grazing within the same pasture can inﬂuence the movement, grazing

locations, and activities of other animals randomly, with attraction, or

with avoidance; therefore most of the animals will have their activity

levels equal to the herd mean. For this reason and the one discussed

above, the herd mean was used as the baseline and any deviation from

such behaviour due to lameness will be classiﬁed as an anomaly. It is

also important to note that this will eliminate the eﬀects of external

factors as these will be aﬀecting the whole herd and only leave the

individual eﬀects of lameness on the cow.

We further deﬁne the Lame Activity Region (LAR) and the Normal

Activity Region (NAR) as shown in Fig. 8. Once a cow is identiﬁed as

lame, we compare the herd mean for all the activities to that cow's

activities and deﬁne a region

>d D d

1 2

, where

is the day the ac-

tivity starts to deviate from the herd activity mean,

is the day that

cow is identiﬁed as lame. As lameness is a transition, we ascertain that

the cow will remain lame after that until it's out of its lameness cycle.

is the entire duration between

and the days after

until the cow is

out of its lameness cycle. It's the whole duration between

to the last

day when the cow was still lame. The values of

, and

will vary

for each cow as some may have longer lameness cycles than others, and

also depending on when the cow is identiﬁed as lame. This is motivated

by the fact that lameness is a transition from normal behaviour to la-

meness and back, it will probably start before it is seen and even con-

tinue after treatment until the cow becomes normal again. Once we

deﬁne the LAR, the rest of the graph is treated as the NAR.

Normal Proﬁle,

To form the normal proﬁle, we deﬁne a small window

within

NAR for each of the normal cows and calculate mean absolute deviation

Nmad

for a given period of time.

H C| |

mad

m i

(1)

Here

is the herd mean,

is the cow activity and

is the

window size of NAR.

Lame Proﬁle

To form the lame proﬁle, we deﬁne a small window

within LAR

for each of the lame cows and calculate mean absolute deviation

Lmad

for a given period of time.

H C| |

mad

m i

(2)

Here

is the herd mean,

is the cow activity and

is the

window size of LAR.

The process is repeated for all the lame and non-lame cows. The

results of this are plotted in a density distribution plot as shown in

Fig. 9.

Relationship between individual cows, Normal (Non-Lame) and Lame

proﬁles

To test the viability of the proﬁles, randomly chosen cows that have

at least been identiﬁed as lame at some point during the experiment

were used (these were not used to form the proﬁles above). The goal

was to go back in time and see how these relate to the constructed

proﬁles before they get lame, when they are lame and when they

transition back to normal behaviour. Using Eq. (3), we deﬁne a window

slice

(the optimal number of days were chosen after a repetitive

process) starting at the beginning of the experiment when the cow is not

lame. We then slide through time as we calculate the Average Deviation

(AD) within each

for each of the cows, and for each of the activity.

=AD H C

m i

(3)

Here

is the herd mean within

and

is the cow activity.

This is repeated as we slide the window. The results of this are plotted

as density distribution and compared with Fig. 9.

The three graphs (a, b and c) for both Figs. 10 and 11 show periods

of transition from normal to lameness for the two cows 2463 and 1469.

Fig. 9. Density distribution plot comparing the Normal and Lame proﬁles.

Fig. 10. Comparing the distribution of cow 2463 against the normal and lame

proﬁle at three diﬀerent stages.

Fig. 11. Comparing the distribution of cow 1469 against the normal and lame

proﬁle at three diﬀerent stages.

M. Taneja, et al. Computers and Electronics in Agriculture 171 (2020) 105286

In Figs. 10(a) and 11(a), both animals are non-lame and the distribu-

tions relate to the Normal cows proﬁle distribution. In Figs. 10(b) and

11(b) the distribution starts to shift to the right. Fig. 10(b) has two

peaks. One relates more to the normal proﬁle and the other to the lame

proﬁle. Fig. 11(b) on the other hand has one peak and this is mid way

between both the normal and the lame proﬁle. This kind of behaviour is

justiﬁable because lameness is a transition. Perhaps this could be the

best stage for the system to identify early lameness. Finally in Fig. 10(c)

and 11(c), the distributions overlap with the lame proﬁle distribution. It

is important to note, even at this stage lameness is not yet visually

detectable by the farmer for both cows.

4.2.2. Clustering

From the above, it was discovered that not all animals behaved the

same way. For example, some animals had their activity levels (step

count, lying time and swaps) tracing out the herd mean, others with

activity levels always higher than the herd mean and, the other cate-

gory always lower than the herd mean. It's also important to note that

even when they became lame they had diﬀerent activity levels de-

pending on which category they belonged to. Therefore the clustering

model is based on this observation. We set thresholds, and based on this

we form three clusters.

To deﬁne a cluster, we deﬁne a window of size

days, and calculate

MAD (Mean Absolute Deviation) between the cow activity and the herd

mean for all the three activities.

H C

| |

MAD

m i

(4)

Here

is the herd mean within a deﬁned window,

is the cow

activity for activity

and

is the window size. We varied the values of

while testing the accuracy of the classiﬁcation model and concluded

that 14 days would be the optimal number of days to deﬁne a cluster.

Based on MAD, we deﬁned a threshold

. Now based on this threshold,

and the following criterion, we deﬁne three clusters. If any two of the

activity levels are below a certain threshold, then that animal is as-

signed into one of the below clusters:

Active

These are animals in the herd that have activity levels always higher

than the herd mean. These have the mean deviation of any two of the

activities is greater than threshold

Normal

These are animals in the herd that have activity levels always tra-

cing out the herd mean. These have the mean deviation of any two of

the activities is less than

but great or equal to zero.

Dormant

These are animals in the herd that have activity levels always lower

than the herd mean. These have the mean deviation of any two of the

activities less than zero.

The threshold was carefully chosen by a repetitive evaluation pro-

cess, and was set to 1.7. The results presented in next section have been

derived with

value equals to 1.7. It’s also important to note that these

clusters are dynamic, i.e., the animals keep changing the clusters they

belong to. This can be caused by many factors like age and weather. So

it is the role of the clustering model to keep regrouping the animals

before selecting the appropriate classiﬁcation model for that cluster

(the best amount of time to re-cluster was found to be two weeks i.e.,

14 days). Table 3 shows the distribution of the clusters as of writing of

this article. The total number used to build clusters was 146 as three of

the animals were eliminated due other health related issues and one

animal lost the tag during the experiment.

4.2.3. Classiﬁcation

Classiﬁcation algorithms are a family of machine learning algo-

rithms that output a discrete value. The output variables are sometimes

called labels or categories. These kind of problems always require the

examples be classiﬁed into two or more classes. Classiﬁcation problems

with two labels are called binary classiﬁcation problems while those

with more than two are called multi-class. We formulated our problem

as a binary class problem with Lame being the positive class and Non-

lame as the negative class. In general, to solve these kind of tasks, the

learning model is usually tasked to produce a function

f R n: {1, ..., }

where

is the number of labels. For example, let

X Y{ , }

denote the data

set (feature, label), and the parameters, where:

= =X

x x x

Lying Steps Swaps

· · ·

n n n nn n

11 12 13

1 2 3 12 3

= =Y

Lame

Non Lame

When

=y f x( )

, an input depicted by vector

will be assigned to a

class label identiﬁed by

. It is important to note that there are other

variants of functions

. For example

might output a probability dis-

tribution as opposed to a class label. At the time of writing, the feature

matrix

was made up 3 columns. Each of the columns is a feature.

These were chosen because most literature suggests that they more

representative of an animal transitioning to lameness or one that is

already lame. The vector

consists of labels

and

, where 0 indicates

non-lame and 1 otherwise. Fig. 12 shows the overall work ﬂow and data

ﬂow of training and testing of the designed model.

5. Results, evaluation and discussion

A brief demo-video of the developed system is available at Taneja

et al. (2018a). Our initial work on age-based clustering of cows com-

bined with data analytics to detect anomalies in their behaviour, and

microservices based application design for integration of diﬀerent ser-

vices has been presented in Taneja et al., 2019a; Byabazaire et al.,

2019; Taneja et al., 2018b respectively.

5.1. Rationale behind clustering

In a study about association patterns of visually observed cattle,

Stephenson et al. (2016) concluded that herds with 40 or less cows did

not exhibit preferential or avoidance associations. This means that they

lived together as a single group. In contrast, larger herd sizes (53–240

cows) tended to form associations with other cows stronger than what

you would expect by chance. Therefore, the clustering step is only re-

levant to large herd sizes. Needless to mention, automated lameness

solutions are meant for large herd sizes as it is assumed that for small

ones, the farmer can visually inspect and keep track of the cows' welfare

easily. We compared the results of a one-size-ﬁts-all model and a cluster

speciﬁc models. Overall, cluster speciﬁc models reduced the classiﬁ-

cation error by 8% as compared to a one-size-ﬁts-all model without

clustering. For example, Fig. 13 shows an animal that was conﬁrmed as

lame from 03/12/2017 to 15/12/2017. The activity clustering based

normal cluster model could correctly identify all the days the animal

was lame, which has been illustrated in Fig. 13 using the highlighted

red box. However, the one-size-ﬁts-all model could only pick up some

days as shown by the red points within the highlighted red box in

Fig. 13.

Table 3

Distribution of the clusters.

Active Normal Dormant

25 109 12

M. Taneja, et al. Computers and Electronics in Agriculture 171 (2020) 105286

5.2. Early lameness detection assessment

The problem was formulated as a binary classiﬁcation problem with

Lame as being the positive class and Non-lame as the negative class.

These are denoted as (

) for the negative class and (

) for the positive

class in the model diagram presented in Fig. 12. The training process

reported in this study is unique from the previous approaches because it

has a feedback loop added. After model validation, an agricultural ex-

pert or farmer re-annotates training data to improve the model accu-

racy.

Please note that once the activity based clustering (Section 4.2.2) is

done, and LAR and NAR have been deﬁned (cow proﬁling – Section

4.2.1), we then calculate the euclidean distance between all the points

on the herd-mean curve and cow activity curve within each of these

regions. This gives us two sets

and

, where

are the values from

NAR and

are the values from LAR, and these form the two classes that

are used in machine learning element of the developed system. These

two sets (

and

) are fed into a K-NN machine learning classiﬁcation

model.

We experimented on a number of sklearn (Pedregosa et al., 2011)

classiﬁcation algorithms ranging from Support Vector Machine (SVM),

Random Forest (RF), K-Nearest Neighbors (K-NN) and Decision Trees.

We selected K-NN classiﬁcation algorithm, as it was best balanced in

terms of accuracy and early lameness detection window as shown in

Tables 4 and 5.

It is also important to note that although a diﬀerent model was

trained and built for each of the three clusters (i.e., three classiﬁcation

models – one for each cluster), results reported (performance and ac-

curacy) in this study are only for the normal cluster. This is because it

was not possible to eﬃciently evaluate the other two clusters as testing

data in these was very small (i.e., imbalanced for a proper evaluation).

K-NN

This has a number of parameters that should be ﬁne-tuned in order

to achieve the desired results. Among these, we evaluated diﬀerent K-

values (2–5), which is the number of neighbours to consider while as-

signing the nearest class. We set the distance metric to minkowski. The

highest accuracy was obtained with

=k2

although this was over ﬁtting

the data. Table 5 presents the accuracy of detection with diﬀerent K-

values, and also the length of the early detection window.

Optimal results were obtained at

=k4

which gave an accuracy of

87% with 3 days before the visual signs could be seen. In all, the normal

cluster model had a sensitivity of 89.7% and speciﬁcity of 72.5%.

Fig. 12. Designed hybrid machine learning model and work ﬂow illustrating the steps in process of data collection, clustering, transformation, classiﬁcation, training,

evaluation and production mode.

Fig. 13. Animal conﬁrmed as lame between 03/12/2017 and 15/12/2017 but

could not be correctly identiﬁed by a one-size-ﬁts-all model. The highlighted

red box shows that using activity clustering based normal cluster in the de-

signed machine learning model, the system was able to detect animal as lame

on 03/12/2017 i.e., starting of the highlighted red box; whereas the one-size-

ﬁts-all system was able to pick only some days, shown as red points in the

ﬁgure. (For interpretation of the references to colour in this ﬁgure legend, the

reader is referred to the web version of this article.)

Table 4

Lameness detection accuracy of the developed system.

Classiﬁcation Model Accuracy (in %) Number of days before the visual

signs of lameness appears

Random Forest 91 1

K-Nearest Neighbors (K-

NN)

87 3

Table 5

Diﬀerent K-values, accuracy of the developed system and early detection

window size.

K-value Accuracy (in %) Number of days before the visual signs of lameness

appear

2 91 1

3 89 2

4 87 3

5 81 1

M. Taneja, et al. Computers and Electronics in Agriculture 171 (2020) 105286

Fig. 15 shows some of the correct detections. One particular cow was

conﬁrmed as lame between 16/10/2017 and 25/10/2017 and the

model could correctly classify all the days as shown by the red points in

Fig. 15.

Fig. 14 shows two cases where the model was able to detect a cow

being lame 3 days before its visual clues were available to the farmer.

The highlighted blue box shows the day when it was visually detected

by the farmer or animal expert, and the start of the red points shows

when the model detected the cow to be lame, and highlighted box

shows the number of days for which the visual sign didn't appear to be

seen by the farmer or animal expert.

5.3. Reduction in data transfer: fog-cloud data reduction

Among the downsides of the existing approaches is that they are

either fully cloud-centric in nature, i.e., all the data is sent to the cloud

Fig. 14. Early detection of lameness by the developed model and late observation by farmer.

M. Taneja, et al. Computers and Electronics in Agriculture 171 (2020) 105286

for processing and analysis; or have just farm premises based system

which limits the accuracy and intelligence (Rutten et al., 2013) of such

systems as there are no dynamic and frequent updates.

In this work, we focused on the reduction of data exchanged be-

tween fog node and the cloud. We leveraged and utilized fog archi-

tecture in our work and were able to reduce data exchange between fog

and cloud node from 10.1 MB to 1.61 MB on daily basis. Fig. 16 shows

an 84% reduction in the amount of data that would otherwise have

streamed to the cloud throughout the day. This aspect of data reduction

becomes even more crucial while scaling up the farm and the herd, as

the amount of data collected and streamed would then rapidly increase.

6. Conclusion

Our results suggest that building custom models for small groups of

animals in the herd that share similar features within the herd improves

the accuracy of the lameness detection as opposed to a one-size ﬁts all

approach. This approach becomes more important and practically vi-

able with increase in size of the herd. Insights from our real world

deployment suggest that activity based cluster speciﬁc models reduce

the classiﬁcation error of lameness detection by 8% as opposed to a one-

size-ﬁts-all approach. Using these clusters to then identify the anoma-

lies in animal behaviour gives a better early detection. In our case,

feeding the resultant cluster in K-NN (K-Nearest Neighbours) based

classiﬁcation models gives an accuracy of 87% with an early detection

of 3 days window before any visual or clinical sign of lameness appears.

It is because of this carefully blended design of clustering and classiﬁ-

cation model that results in a hybrid model for early lameness detection

in dairy cattle.

Further, the fog-based computational assistance enables the in-

telligent processing of data closer to the source, thereby leading to an

84% reduction in the amount of data transfer to cloud. Another key

lesson learned is that any of the edge/fog/cloud resources of the overall

architecture if considered in isolation would not be able to manage the

developed IoT application, without compromising on functionalities or

performance. And thus a careful coordination of edge, fog and cloud

components is required as presented in this work.

7. Ongoing and future work

To further validate the proposed approach for early lameness de-

tection, we are expanding the work undertaken to date through the

execution of a use case in the IoF2020 project

named MELD

. The

MELD project is building and expanding upon this existing work, and

integrating it into the IoF2020 dairy farming technology trials with

deployments in Portugal, Israel and South Africa. It aims to leverage

sensor technologies from two diﬀerent vendors on a combined total of

approximately 1000 cattle, consisting of both beef and dairy.

In future work we also intend to investigate a more robust clustering

technique as the current one is only based on threshold. Also, we plan

to evaluate the other cluster models. Further, in context of eﬃcient in-

network resource utilization and increasing system resilience, one of

the possible directions of future work is to look into distributed learning

(Konecný et al.) and distributed data analytics (Taneja et al., 2019c;

Chang et al., 2017) based approaches in such real-world IoT based

deployments.

Once the developed technology has been validated on a number of

farms with diﬀerent geographical and environmental settings, the goal

is to roll out the technology to the vendors' customer base as an added

feature through licensing.

CRediT authorship contribution statement

Mohit Taneja: Investigation, Software, Methodology, Data cura-

tion, Writing - original draft, Formal analysis, Validation. John

Byabazaire: Investigation, Software, Methodology, Data curation,

Formal analysis, Validation. Nikita Jalodia: Methodology, Writing -

review & editing, Visualization. Alan Davy: Supervision, Funding ac-

quisition, Conceptualization. Cristian Olariu: Resources, Formal ana-

lysis, Supervision. Paul Malone: Writing - review & editing, Project

administration.

Declaration of Competing Interest

The authors declare that they have no known competing ﬁnancial

interests or personal relationships that could have appeared to inﬂu-

ence the work reported in this paper.

Acknowledgment

This work has emanated from research conducted with the ﬁnancial

support of Science Foundation Ireland (SFI) and is co-funded under the

European Regional Development Fund under Grant Number 13/RC/

2077. Mohit Taneja is also supported by CISCO Research Gift Fund.

Fig. 15. Red points indicating lameness anomalies identiﬁed by the normal

cluster model for cow ID 1988. (For interpretation of the references to colour in

this ﬁgure legend, the reader is referred to the web version of this article.)

Fig. 16. Daily reduction in the amount of data between the fog node and the

cloud.

Internet of Food & Farm 2020, https://www.iof2020.eu/https://www.

iof2020.eu/.

MELD stands for Machine Learning based Early Lameness Detection in Beef

and Dairy Cattle.

The collected data from the existing real-world deployment is available to

be shared with the academic research community upon request.

M. Taneja, et al. Computers and Electronics in Agriculture 171 (2020) 105286

The future work (MELD) is funded through the IoF2020 which has

received funding from the European Union's Horizon 2020 research and

innovation programme under grant agreement no. 731884.

References

“Connected Cows?” - Joseph Sirosh (Strata + Hadoop 2015) – YouTube. https://www.

youtube.com/watch?v=oY0mxwySaSo (accessed on 04/29/2019).

Ag, Q., Cattle health management, tags, and software | quantiﬁed ag. http://quantiﬁedag.

com/ (accessed on 12/07/2017).

Dairy Cow Mobility and Lameness - AHDB Dairy. Accessed on 1st Nov 2016. URL https://

dairy.ahdb.org.uk/technical-information/animal-health-welfare/lameness.

Al-Fuqaha, A., Guizani, M., Mohammadi, M., Aledhari, M., Ayyash, M., 2015. Internet of

things: A survey on enabling technologies, protocols, and applications. IEEE

Commun. Surv. Tutor. 17 (4), 2347–2376. https://doi.org/10.1109/COMST.2015.

2444095.

Alsaaod, M., Römer, C., Kleinmanns, J., Hendriksen, K., Rose-Meierhöfer, S., Plümer, L.,

Büscher, W., 2012. Electronic detection of lameness in dairy cows through measuring

pedometric activity and lying behavior. Appl. Animal Behav. Sci. 142 (3–4),

134–141. https://doi.org/10.1016/j.applanim.2012.10.001.

Alsaaod, M., Niederhauser, J., Beer, G., Zehner, N., Schuepbach-Regula, G., Steiner, A.,

2015. Development and validation of a novel pedometer algorithm to quantify ex-

tended characteristics of the locomotor behavior of dairy cows. J. Dairy Sci. 98 (9),

6236–6242. http://linkinghub.elsevier.com/retrieve/pii/

S0022030215004609https://doi.org/10.3168/jds.2015-9657.

Amory, J., Kloosterman, P., Barker, Z., Wright, J., Blowey, R., Green, L., 2006. Risk

factors for reduced locomotion in dairy cattle on nineteen farms in The Netherlands.

J. Dairy Sci. 89 (5), 1509–1515. http://linkinghub.elsevier.com/retrieve/pii/

S0022030206722184https://doi.org/10.3168/jds.S0022-0302(06)72218-4.

Balalaie, A., Heydarnoori, A., Jamshidi, P., 2016. Microservices Architecture Enables

DevOps: Migration to a Cloud-Native. Architecture. https://doi.org/10.1109/MS.

2016.64.

Barker, Z.E., Leach, K.A., Whay, H.R., Bell, N.J., Main, D.C.J., 2010. Assessment of la-

meness prevalence and associated risk factors in dairy herds in England and Wales. J.

Dairy Sci. 93 (3), 932–941. https://doi.org/10.3168/jds.2009-2309..https://doi.

org/10.3168/jds.2009-2309.

Blackie, N., Bleach, E., Amory, J., Scaife, J., 2011. Impact of lameness on gait character-

istics and lying behaviour of zero grazed dairy cattle in early lactation. Appl. Animal

Behav. Sci. 129 (2–4), 67–73. https://doi.org/10.1016/j.applanim.2010.10.006.

Boumatic. https://boumatic.com/us_en/ (accessed on 12/07/2017).

Brewster, C., Roussaki, I., Kalatzis, N., Doolin, K., Ellis, K., 2017. Iot in agriculture:

Designing a europe-wide large-scale pilot. IEEE Commun. Mag. 55 (9), 26–33.

https://doi.org/10.1109/MCOM.2017.1600528.

Byabazaire, J., Olariu, C., Taneja, M., Davy, A., 2019. Lameness detection as a service:

Application of machine learning to an internet of cattle. In: 2019 16th IEEE Annual

Consumer Communications Networking Conference (CCNC), pp. 1–6. https://doi.

org/10.1109/CCNC.2019.8651681.

Caria, M., Schudrowitz, J., Jukan, A., Kemper, N., 2017. Smart farm computing systems

for animal welfare monitoring. In: 2017 40th International Convention on

Information and Communication Technology, Electronics and Microelectronics

(MIPRO), pp. 152–157. https://doi.org/10.23919/MIPRO.2017.7973408.

Chang, T.-C., Zheng, L., Gorlatova, M., Gitau, C., Huang, C.-Y., Chiang, M., 2017.

Decomposing data analytics in fog networks. In: Proceedings of the 15th ACM

Conference on Embedded Network Sensor Systems, SenSys '17, ACM, New York, NY,

USA, pp. 35:1–35:2. https://doi.org/10.1145/3131672.3136962.

Chapinal, N., de Passillé, A., Weary, D., von Keyserlingk, M., Rushen, J., 2009. Using gait

score, walking speed, and lying behavior to detect hoof lesions in dairy cows. J. Dairy

Sci. 92 (9), 4365–4374. http://linkinghub.elsevier.com/retrieve/pii/

S002203020970760Xhttps://doi.org/10.3168/jds.2009-2115.

Chapinal, N., de Passillé, A., Rushen, J., Wagner, S., 2010. Automated methods for de-

tecting lameness and measuring analgesia in dairy cattle. J. Dairy Sci. 93 (5),

2007–2013. http://linkinghub.elsevier.com/retrieve/pii/

S002203021000189Xhttps://doi.org/10.3168/jds.2009-2803.

Chapinal, N., de Passillé, A., Pastell, M., Hänninen, L., Munksgaard, L., Rushen, J., 2011.

Measurement of acceleration while walking as an automated method for gait as-

sessment in dairy cattle. J. Dairy Sci. 94 (6), 2895–2901. http://linkinghub.elsevier.

com/retrieve/pii/S0022030211002773https://doi.org/10.3168/jds.2010-3882.

Chen, M.-C., Chen, C.-H., Siang, C.-Y., 2014. Design of information system for milking

dairy cattle and detection of mastitis. Math. Probl. Eng. 2014. https://doi.org/10.

1155/2014/759019.

Cook, N.B., 2003. Prevalence of lameness among dairy cattle in Wisconsin as a function of

housing type and stall surface. J. Am. Vet. Med. Assoc. 223 (9), 1324–1328. https://

doi.org/10.2460/javma.2003.223.1324.

Espejo, L., Endres, M., Salfer, J., 2006. Prevalence of lameness in high-producing holstein

cows housed in Freestall Barns in Minnesota. J. Dairy Sci. 89 (8), 3052–3058. http://

linkinghub.elsevier.com/retrieve/pii/S0022030206725796https://doi.org/10.3168/

jds.S0022-0302(06)72579-6.

Foditsch, C., Oikonomou, G., Machado, V.S., Bicalho, M.L., Ganda, E.K., Lima, S.F., Rossi,

R., Ribeiro, B.L., Kussler, A., Bicalho, R.C., 2016. Lameness prevalence and risk

factors in large dairy farms in Upstate New York. Model development for the pre-

diction of claw horn disruption lesions. PLoS ONE 11 (1), e0146718. https://doi.org/

10.1371/journal.pone.0146718.

C. C. V. Ger, C. of XLVets), Economic cost of lameness in Irish dairy herds, Forage Nutrit.

https://www.xlvets.ie/sites/xlvets.ie/ﬁles/press-article-ﬁles/XLVets%2520Article

%2520Forage%2520Guide%25202012.pdf.

Getting to know MQTT. https://www.ibm.com/developerworks/library/iot-mqtt-why-

good-for-iot/index.html (last accessed on August 03, 2017).

Hsu, T.-C., Yang, H., Chung, Y.-C., Hsu, C.-H., 2018. A creative iot agriculture platform for

cloud fog computing. Sustain. Comput. Inform. Syst. http://www.sciencedirect.com/

science/article/pii/S2210537918303275https://doi.org/10.1016/j.suscom.2018.10.

006.

Ireland, D., Milking parlours & machines, heat detection, scrapers, feeders, milk cooling -

dairymaster Ireland. http://www.dairymaster.ie/ (accessed on 12/07/2017).

Jukan, A., Masip-Bruin, X., Amla, N., 2017. Smart computing and sensing technologies for

animal welfare: A systematic review. ACM Comput. Surv. 50 (1), 10:1–10:27. http://

doi.acm.org/10.1145/3041960https://doi.org/10.1145/3041960.

Konecný, J., McMahan, H.B., Ramage, D., Richtárik, P., Federated optimization:

Distributed machine learning for on-device intelligence, CoRR abs/1610.02527.

arXiv:1610.02527. URL http://arxiv.org/abs/1610.02527.

Lee, S., Kim, H., Hong, D.K., Ju, H., 2013. Correlation analysis of mqtt loss and delay

according to qos level, in: The International Conference on Information Networking

2013 (ICOIN), pp. 714–717. https://doi.org/10.1109/ICOIN.2013.6496715.

Meet the “connected cow” | Financial Times. https://www.ft.com/content/2db7e742-

7204-11e7-93ﬀ-99f383b09ﬀ9 (accessed on 04/29/2019).

Miiaková, M., Strapák, P., Szencziová, I., Strapáková, E., Hanušovský, O., 2018. Several

methods of Estrus detection in cattle dams: a review. Acta Universitatis Agriculturae

et Silviculturae Mendelianae Brunensis 66 (2), 619–625. https://www.researchgate.

net/publication/324902783_Several_Methods_of_Estrus_Detection_in_Cattle_Dams_A_

Review https://acta.mendelu.cz/66/2/0619/https://doi.org/10.11118/

actaun201866020619.

MQTT. http://mqtt.org/ (last accessed on August 03, 2017).

Muangprathub, J., Boonnam, N., Kajornkasirat, S., Lekbangpong, N., Wanichsombat, A.,

Nillaor, P., 2019. Iot and agriculture data analysis for smart farm. Comput. Electron.

Agric. 156, 467–474. http://www.sciencedirect.com/science/article/pii/

S0168169918308913https://doi.org/10.1016/j.compag.2018.12.011.

Munksgaard, L., van Reenen, C.G., Boyce, R., 2006. Automatic monitoring of lying,

standing and walking behavior in dairy cattle. Anim. Sci. 84 (suppl), 304.

Neveux, S., Weary, D., Rushen, J., von Keyserlingk, M., de Passillé, A., 2006. Hoof dis-

comfort changes how dairy cattle distribute their body weight. J. Dairy Sci. 89 (7),

2503–2509. http://linkinghub.elsevier.com/retrieve/pii/

S0022030206723256https://doi.org/10.3168/jds.S0022-0302(06)72325-6.

O'Driscoll, K., Boyle, L., Hanlon, A., 2008. A brief note on the validation of a system for

recording lying behaviour in dairy cows. Appl. Animal Behav. Sci. 111 (1–2),

195–200. https://doi.org/10.1016/j.applanim.2007.05.014.

Pastell, M., Hänninen, L., de Passillé, A., Rushen, J., 2010. Measures of weight distribu-

tion of dairy cows to detect lameness and the presence of hoof lesions. J. Dairy Sci. 93

(3), 954–960. http://linkinghub.elsevier.com/retrieve/pii/

S0022030210000627https://doi.org/10.3168/jds.2009-2385.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M.,

Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D.,

Brucher, M., Perrot, M., Duchesnay, E., 2011. Scikit-learn: Machine learning in

Python. J. Mach. Learn. Res. 12, 2825–2830.

PouchDB. The JavaScript Database that Syncs! https://pouchdb.com/ (accessed on 05/

02/2019).

Poursaberi, A., Bahr, C., Pluk, A., Van Nuﬀel, A., Berckmans, D., 2010. Real-time auto-

matic lameness detection based on back posture extraction in dairy cattle: Shape

analysis of cow with image processing techniques. Comput. Electron. Agric. 74 (1),

110–119. https://doi.org/10.1016/j.compag.2010.07.004.

Poursaberi, A., Pluk, A., Bahr, C., Martens, W., Veermäe, I., Kokin, E., Praks, J.,

Poikalainen, V., Pastell, M., Ahokas, J., Van Nuﬀel, A., Vangeyte, J., Sonck, B.,

Berckmans, D., 2009. Image based separation of dairy cows for automatic lameness

detection with a real time vision system. In: American Society of Agricultural and

Biological Engineers Annual International Meeting 2009, ASABE 2009, vol. 7, pp.

4763–4773. https://doi.org/10.13031/2013.27258. URL http://www.scopus.com/

inward/record.url?eid=2-s2.0-77649156361 partnerID=tZOtx3y1.

Poursaberi, A., Bahr, C., Pluk, A., Berckmans, D., VeermÃďe, I., Kokin, E., Pokalainen, V.,

2011. Online lameness detection in dairy cattle using body movement pattern (bmp).

In: 2011 11th International Conference on Intelligent Systems Design and

Applications, pp. 732–736. https://doi.org/10.1109/ISDA.2011.6121743.

Rutten, C., Velthuis, A., Steeneveld, W., Hogeveen, H., 2013. Invited review: Sensors to

support health management on dairy farms. J. Dairy Sci. 96 (4), 1928–1952. http://

www.sciencedirect.com/science/article/pii/S0022030213001409https://doi.org/

10.3168/jds.2012-6107.

Shi, W., Cao, J., Zhang, Q., Li, Y., Xu, L., 2016. Edge computing: Vision and challenges.

IEEE Internet Things J. 3 (5), 637–646. https://doi.org/10.1109/JIOT.2016.

2579198.

Stephenson, M.B., Bailey, D.W., 2017. movement patterns of gps-tracked cattle on ex-

tensive rangelands suggest independence among individuals? Agriculture 7 (7).

http://www.mdpi.com/2077-0472/7/7/58https://doi.org/10.3390/

agriculture7070058.

M. Taneja, et al. Computers and Electronics in Agriculture 171 (2020) 105286

Stephenson, M.B., Bailey, D.W., Jensen, D., 2016. Association patterns of visually-ob-

served cattle on Montana, USA foothill rangelands. Appl. Animal Behav. Sci. 178,

7–15. http://www.sciencedirect.com/science/article/pii/

S0168159116300442https://doi.org/10.1016/j.applanim.2016.02.007.

Taneja, M., Davy, A., 2017. Resource aware placement of iot application modules in fog-

cloud computing paradigm. In: 2017 IFIP/IEEE Symposium on Integrated Network

and Service Management (IM), pp. 1222–1228. https://doi.org/10.23919/INM.2017.

7987464.

Taneja, M., Davy, A., 2016. Resource aware placement of data analytics platform in fog

computing. Proc. Comput. Sci. 97, 153–156. http://www.sciencedirect.com/science/

article/pii/S1877050916321111https://doi.org/10.1016/j.procs.2016.08.295.

Taneja, M., Jalodia, N., Byabazaire, J., Davy, A., Olariu, C., Smartherd-connected_cows_

demo.mp4 - google drive. https://drive.google.com/ﬁle/d/

1QIrKSp8SkAZRRAFQDQHdPXu1TVYaVyv3/view (accessed on 10/04/2018).

Taneja, M., Byabazaire, J., Davy, A., Olariu, C., 2018. Fog assisted application support for

animal behaviour analysis and health monitoring in dairy farming. In: 2018 IEEE 4th

World Forum on Internet of Things (WF-IoT), pp. 819–824. https://doi.org/10.1109/

WF-IoT.2018.8355141.

Taneja, M., Jalodia, N., Malone, P., Byabazaire, J., Davy, A., Olariu, C., 2019b. Connected

cows: Utilizing fog and cloud analytics toward data-driven decisions for smart dairy

farming. IEEE Internet of Things Mag. 2 (4), 32–37. https://doi.org/10.1109/IOTM.

0001.1900045.

Taneja, M., Jalodia, N., Byabazaire, J., Davy, A., Olariu, C., 2019a. SmartHerd manage-

ment: A microservices-based fog computing–assisted IoT platform towards data-

driven smart dairy farming. Software: Pract. Exp. 49 (7), 1055–1078. https://doi.

org/10.1002/spe.2704..https://onlinelibrary.wiley.com/doi/pdf/10.1002/spe.

2704. In this issue.

Taneja, M., Jalodia, N., Davy, A., 2019c. Distributed decomposed data analytics in fog

enabled iot deployments. IEEE Access 7, 40969–40981. https://doi.org/10.1109/

ACCESS.2019.2907808.

Taylor, K., Griﬃth, C., Lefort, L., Gaire, R., Compton, M., Wark, T., Lamb, D., Falzon, G.,

Trotter, M., 2013. Farming the web of things. IEEE Intell. Syst. 28 (6), 12–19. https://

doi.org/10.1109/MIS.2013.102.

Thorup, V.M., Munksgaard, L., Robert, P.E., Erhard, H.W., Thomsen, P.T., Friggens, N.C.,

2015. Lameness detection via leg-mounted accelerometers on dairy cows on four

commercial farms. Animal 9 (10), 1704–1712. https://doi.org/10.1017/

S1751731115000890.

Van De Gucht, T., Saeys, W., Van Nuﬀel, A., Pluym, L., Piccart, K., Lauwers, L., Vangeyte,

J., Van Weyenberg, S., 2017. Farmers' preferences for automatic lameness-detection

systems in dairy cattle. J. Dairy Sci. 100 (7), 5746–5757. http://linkinghub.elsevier.

com/retrieve/pii/S0022030217305027https://doi.org/10.3168/jds.2016-12285.

Van Hertem, T., Maltz, E., Antler, A., Romanini, C., Viazzi, S., Bahr, C., Schlageter-Tello,

A., Lokhorst, C., Berckmans, D., Halachmi, I., 2013. Lameness detection based on

multivariate continuous sensing of milk yield, rumination, and neck activity. J. Dairy

Sci. 96 (7), 4286–4298. http://linkinghub.elsevier.com/retrieve/pii/

S0022030213003561https://doi.org/10.3168/jds.2012-6188.

Van Nuﬀel, A., Zwertvaegher, I., Pluym, L., Van Weyenberg, S., Thorup, V.M., Pastell, M.,

Sonck, B., Saeys, W., 2015. Lameness detection in dairy cows: Part 1. How to dis-

tinguish between non-lame and lame cows based on diﬀerences in locomotion or

behavior. Animals 5 (3), 838–860. https://doi.org/10.3390/ani5030387.

Van Nuﬀel, A., Zwertvaegher, I., Van Weyenberg, S., Pastell, M., Thorup, V.M., Bahr, C.,

Sonck, B., Saeys, W., 2015. Lameness detection in dairy cows: Part 2. Use of sensors

to automatically register changes in locomotion or behavior. Animals. https://doi.

org/10.3390/ani5030388.

Viazzi, S., Bahr, C., Schlageter-Tello, A., Van Hertem, T., Romanini, C., Pluk, A.,

Halachmi, I., Lokhorst, C., Berckmans, D., 2013. Analysis of individual classiﬁcation

of lameness using automatic measurement of back posture in dairy cattle. J. Dairy

Sci. 96 (1), 257–266. https://linkinghub.elsevier.com/retrieve/pii/

S0022030212008466https://doi.org/10.3168/jds.2012-5806.

Viazzi, S., Bahr, C., Van Hertem, T., Schlageter-Tello, A., Romanini, C.E.B., Halachmi, I.,

Lokhorst, C., Berckmans, D., 2014. Comparison of a three-dimensional and two-di-

mensional camera system for automated measurement of back posture in dairy cows.

Comput. Electron. Agric. 100, 139–147. https://doi.org/10.1016/j.compag.2013.11.

005.

von Keyserlingk, M., Barrientos, A., Ito, K., Galo, E., Weary, D., 2012. Benchmarking cow

comfort on North American freestall dairies: Lameness, leg injuries, lying time, fa-

cility design, and management for high-producing Holstein dairy cows. J. Dairy Sci.

95 (12), 7399–7408. http://linkinghub.elsevier.com/retrieve/pii/

S0022030212007606https://doi.org/10.3168/jds.2012-5807.

Wark, T., Swain, D., Crossman, C., Valencia, P., Bishop-Hurley, G., Handcock, R., 2009.

Sensor and actuator networks: Protecting environmentally sensitive areas. IEEE

Pervas. Comput. 8 (1), 30–36. https://doi.org/10.1109/MPRV.2009.15.

Whay, H.R., Main, D.C.J., Green, L.E., Webster, A.J.F., 2003. Assessment of the welfare of

dairy caftle using animal-based measurements: direct observations and investigation

of farm records. Vet. Rec. 153 (7), 197–202. http://veterinaryrecord.bmj.com/cgi/

doi/10.1136/vr.153.7.197https://doi.org/10.1136/vr.153.7.197.

Yunta, C., Guasch, I., Bach, A., 2012. Short communication: Lying behavior of lactating

dairy cows is inﬂuenced by lameness especially around feeding time. J. Dairy Sci. 95

(11), 6546–6549. http://linkinghub.elsevier.com/retrieve/pii/

S0022030212006261https://doi.org/10.3168/jds.2012-5670.

Zamora-Izquierdo, M.A., Santa, J., Martínez, J.A., Martínez, V., Skarmeta, A.F., 2019.

Smart farming iot platform based on edge and cloud computing. Biosyst. Eng. 177,

4–17. http://www.sciencedirect.com/science/article/pii/

S1537511018301211https://doi.org/10.1016/j.biosystemseng.2018.10.014.

Zheleva, M., Bogdanov, P., Zois, D.-S., Xiong, W., Chandra, R., Kimball, M., 2017.

Smallholder agriculture in the information age: Limits and opportunities. In:

Proceedings of the 2017 Workshop on Computing Within Limits, LIMITS '17, ACM,

New York, NY, USA, pp. 59–70. https://doi.org/10.1145/3080556.3080563. URL

http://doi.acm.org/10.1145/3080556.3080563.

Mohit Taneja is currently working as a software research

engineer, and also as a PhD researcher with Emerging

Networks Laboratory at Telecommunications Software and

Systems Group of Waterford Institute of Technology,

Ireland. He joined them in 2015 as a Masters Student as a

part of the Science Foundation Ireland (SFI) funded

CONNECT Research Centre. He was a part of the SFI-

CONNECT-IBM project termed SmartHerd, which has now

been extended with follow-on IoF2020 project termed

MELD, which is a project to detect early stage lameness in

cattle with the help of technology. His research focuses on

fog computing support for Internet of Things applications.

His current research interests include Fog and Cloud

Computing, Internet of Things (IoT), Distributed Systems,

and Distributed Data Analytics. He received his Bachelor's Degree in Computer Science

and Engineering from The LNM Institute of Information Technology, Jaipur, India in

2015.

John Byabazaire is currently a PhD student in school of

computer science, University College Dublin working on

IoT systems for data collection in precision agriculture.

Before that, John pursed a MSc in Computer Science from

Waterford Institute of Technology and due to graduate. He

received his BSc in Computer Science from Gulu University

in 2013. Alongside his current PhD study, John continues to

research e-learning for low bandwidth environments,

Software Deﬁned Networking, Network Function

Virtualization, and Remote Sensing, IoT (Internet of Things)

and Fog Analytics.

Nikita Jalodia is a PhD Researcher in the Department of

Computing and Mathematics at the Emerging Networks Lab

Research Unit in Telecommunications Software and

Systems Group, Waterford Institute of Technology, Ireland.

She is working as a part of the Science Foundation Ireland

funded CONNECT Research Centre for Future Networks and

Communications, and her research is based in Deep

Learning and Neural Networks, NFV, Fog Computing and

IoT. She received her Bachelor's Degree in Computer

Science and Engineering from The LNM Institute of

Information Technology, Jaipur, India in 2017, with an

additional diploma specialization in Big Data and Analytics

with IBM.

Alan Davy received the B.Sc. (with Hons.) degree in ap-

plied computing and the Ph.D. degree from the Waterford

Institute of Technology, Waterford, Ireland, in 2002 and

2008, respectively. Since 2002, he has been with the

Telecommunications Software and Systems Group, origin-

ally as a student and then, since 2008, as a Post-Doctoral

Researcher. In 2010, he was with IIT Madras, India, as an

Assistant Professor, lecturing in network management sys-

tems. He was a recipient of the Marie Curie International

Mobility Fellowship in 2010, which brought him to work at

the Universitat Politecnica de Catalunya for two years. He is

currently Head of Department of Computing and

Mathematics in School of Science and Computing at

Waterford Institute of Technology (WIT), Waterford,

Ireland. Previously, he was Research Unit Manager of the Emerging Networks Laboratory

with the Telecommunications Software Systems Group of WIT. He is coordinator of a

number of national and EU projects such as TERAPOD. His current research interests

include Virtualised Telecom Networks, Fog and Cloud Computing, Molecular

Communications and TeraHertz Communication.

M. Taneja, et al.

Computers and Electronics in Agriculture 171 (2020) 105286

Cristian Olariu received his B.Eng. degree in 2008 from

the Faculty of Electronics and Telecommunications,

Politehnica University of Timisoara, Romania, and his

Ph.D. degree in 2013 WIT in VoIP over wireless and cellular

networks. He was a research engineer with the Innovation

Exchange, IBM Ireland, and is currently a research scientist

with the Huawei Research Centre, Ireland. His interests are

in computer networks optimisation, applied predictive

modelling, anomaly detection, root cause analysis, deep

neural nets, machine learning and AI.

Paul Malone graduated from Waterford Institute of

Technology with a 1st class honours degree in 1998 and

completed a research MSc in 2001. Paul has worked on

researching technologies and techniques related to IT se-

curity in the areas of distributed trust and reputation

management and privacy and data protection controls. Paul

is currently coordinating an Agri-Tech use case sub-project

of the H2020 IoF2020.eu project applying machine learning

technologies in detecting early stage lameness in cattle.

M. Taneja, et al.

Computers and Electronics in Agriculture 171 (2020) 105286

Supply Chain Risks in the Dairy Industry

Article

Full-text available

Dec 2024
Benchmark Int J

Purpose Supply chain risk (SCR) has been extensively explored in various sectors, yet there is a notable scarcity of SCR studies in the dairy industry. This study aims to identify the primary and distinctive risks in the dairy supply chain (DSC), propose a typological model for SCR, highlight challenges specific to the DSC and offer mitigation strategies. Design/methodology/approach We employ a systematic literature review to collect and review relevant research articles published between 2010 and 2019 to identify the main risks and mitigation strategies associated with the DSC, enabling the construction of a typological model of DSC risks. Findings Results of the systematic review of the SCR literature show that the main DSC risks include on-farm risk (e.g. risks originating from the farming system), off-farm risk (e.g. supply risk, demand risk and manufacturing risk) and inherent SCR (e.g. logistics risk, information risk and financial risk). Notably, we find that the farming system plays a key role in today’s agricultural supply chain operations, indicating the importance of considering on-farm risk in the entire DSC. Additionally, mitigation strategies are located in response to the identified DSC risks by the typology of DSC risks. Originality/value This paper is the first attempt to develop a typological model of SCR for the dairy industry by a systematic literature review. The findings contribute to providing a comprehensive understanding of DSC risks by bridging the gap of ignoring the on-farm risks of the DSC in the existing literature. The typology may serve as a guide in practice to develop mitigation strategies in response to DSC risks.

Machine Learning en la detección y predicción de enfermedades del ganado: una visión general

Article

Full-text available

Dec 2024

Marco Vieto-Vega

La detección temprana y la predicción de enfermedades en el ganado son esenciales para garantizar la salud y el bienestar de los animales, mejorar la productividad y reducir las pérdidas económicas. En este contexto, el Machine Learning (ML), un avance prominente dentro de la inteligencia artificial emerge como una herramienta revolucionaria para transformar el proceso de identificación y manejo de enfermedades en los animales. Esta tecnología permite desarrollar algoritmos complejos capaces de analizar grandes volúmenes de datos clínicos y ambientales, identificando patrones de alerta temprana en síntomas y comportamientos asociados a enfermedades. A través de modelos predictivos, el ML evalúa factores de riesgo y estima la probabilidad de aparición de enfermedades, lo que mejora significativamente la precisión diagnóstica y la efectividad de los tratamientos. Este artículo revisa de manera exhaustiva el uso de ML en la producción ganadera, abordando aplicaciones, modelos y técnicas de vanguardia para la detección y manejo sanitario del ganado, y plantea oportunidades para una gestión ganadera más eficiente y ética, considerando además los desafíos éticos y de privacidad inherentes a la implementación de estas tecnologías

Redefınıng Water Resources: Innovatıve Water Harvestıng Strategıes In Agrıcultural Drought Management

Chapter

Full-text available

Dec 2024

Drought management in agriculture is of critical importance not only for the conservation of water resources but also for ensuring the continuity of sustainable agricultural production. Traditional water management methods are proving inadequate in the face of rapid climate change and increasing drought events, which highlights the growing need for innovative water management strategies (Mishra and Singh, 2010). In this context, water harvesting emerges as a significant strategy for evaluating alternative water resources and reducing the dependency of agricultural production on water

Justification of the component composition of the drug “DB” for dry hoof baths

Article

Full-text available

Nov 2024

This paper presents the results of substantiation of the component composition of the medicinal product «DB» for dry hoof baths for the treatment and prevention of infectious and non-infectious diseases of the distal extremities of cattle. The selection of the most optimal composition was carried out based on the results of studying the adsorption activity of 13 experimental samples and determining the minimum bactericidal (fungicidal) concentration of 10 experimental samples. The maximum adsorption and antimicrobial activity of the sample, which includes bentonite, copper sulfate, zinc sulfate, polyvinyl alcohol, perlite, camphor oil and sodium benzenesulfochloramide (called «DB») is shown. The tests showed that the composition of «DB» is the most optimal combination of the substances used, which complemented each other, leading to increased adsorption, bactericidal and anti-inflammatory properties. The developed means for dry hoof baths installed in places where animals are driven will be most effective in the initial (maceration, microtrauma, fresh wounds, etc.) and middle stages of infectious diseases of the distal extremities, which will significantly reduce the incidence and development of severe infectious forms of diseases. In this regard, the introduction of this means into the technology of cattle breeding will reduce labor costs for the prevention and treatment of cattle, will contribute to an increase in the period of economic use of animals, a decrease in the development of antibiotic resistance of pathogenic microorganisms and an increase in the profitability of cattle breeding as a whole.

Relationship of Phase Angle with Sarcopenia Components and Comprehensive Geriatric Assessment in Physically Independent Older Adults

Article

Dec 2024

ML-Based Anomaly Detection in 6G Networks: A Survey on the Current Status, Challenges, and Future Directions

Conference Paper

Oct 2024

AI-based prediction and detection of early-onset of digital dermatitis in dairy cows using infrared thermography

Article

Full-text available

Dec 2024

Digital dermatitis (DD) is a common foot disease that can cause lameness, decreased milk production and fertility decline in cows. The prediction and early detection of DD can positively impact animal welfare and profitability of the dairy industry. This study applies deep learning-based computer vision techniques for early onset detection and prediction of DD using infrared thermography (IRT) data. We investigated the role of various inputs for these tasks, including thermal images of cow feet, statistical color features extracted from IRT images, and manually registered temperature values. Our models achieved performances of above 81% accuracy on DD detection on ‘day 0’ (first appearance of clinical signs), and above 70% accuracy prediction of DD two days prior to the first appearance of clinical signs. Moreover, current findings indicate that the use of IRT images in conjunction with AI based predictors show real potential for developing future real-time automated tools to monitoring DD in dairy cows.

Energy-Aware Feature and Classifier for Behaviour Recognition of Laying Hens in an Aviary System

Article

Nov 2024

Securing smart dairy farms: A cybersecurity analysis of IoT-based cow health monitoring systems

Conference Paper

Jan 2024

A multimodal classification method: Cow behavior pattern classification with improved EdgeNeXt using an inertial measurement unit

Article

Nov 2024
COMPUT ELECTRON AGR

Distributed Decomposed Data Analytics in Fog Enabled IoT Deployments

Article

Full-text available

Mar 2019

The edge of the network plays a vital role in an IoT system, serving as an optimal site to perform operation on data before transmitting it over the network. We present the fog specific decomposition of multivariate linear regression as the predictive analytic model in our work using Statistical Query Model and Summation Form. The decomposition method used is not the contribution, but applying the decomposition method to the analytics model to run in a distributed manner in fog enabled IoT deployments is the contribution. What is novel is the decomposition made on a fog based distributed setting. To test the performance, our proposed approach has been applied to a real-world dataset and evaluated using a fog computing testbed. The proposed method avoids sending raw data to the cloud, and offers balanced computation in the infrastructure. The results show an 80% reduction in amount of data transferred to the cloud using the proposed fog based distributed data analytics approach as compared to the conventional cloud based approach. Furthermore, by adopting the proposed distributed approach, we observed a 98% drop in the time taken to arrive to the final result as compared to the cloud centric approach. We also present the results on quality of analytics solution obtained in both approaches, and they suggest that fog based distributed analytics approach can serve as equally as the traditional cloud centric approach.

Several Methods of Estrus Detection in Cattle Dams: A Review

Article

Full-text available

May 2018

The oestrus is defined as a complex of physiological signs and changes of behavior occurring immediately before the ovulation. It is the period of sexual receptivity of the cow and this period characterized by, typical cow behaviour which is standing when mounted by a bull or companion cow. Cow indicates this willingness by a slight arching of the back and immobility when approached. The bovine estrus cycle averages 20 days in heifers and 21 days in cows, but the fluctuations in the cycle length is laying between 18 and 25 days. As a result of this progress often only 50 % of the estrus cycles are recognized, and it is extremely hard to find all cows on heat. We state a few methods for estrus detection such as visual observation, locomotion activity and rumination time measurement and other. © 2008 Mendel University of Agriculture and Forestry Brno. All rights reserved.

Decomposing Data Analytics in Fog Networks

Conference Paper

Full-text available

Nov 2017

Fog computing, the distribution of computing resources closer to the end devices along the cloud-to-things continuum, is recently emerging as an architecture for scaling of the Internet of Things (IoT) sensor networking applications. Fog computing requires novel computing program decompositions for heterogeneous hierarchical settings. To evaluate these new decompositions, we designed, developed, and instrumented a fog computing testbed that includes cloud computing and computing gateway execution points collaborating to finish complex data analytics operations. In this interactive demonstration we present one fog-specific algorithmic decomposition we recently examined and adapted for fog computing: a multi-execution point linear regression decomposition that jointly optimizes operation latency, quality, and costs. The demonstration highlights the role fog computing can play in future sensor networking architectures, and highlights some of the challenges of creating computing program decompositions for these architectures. An annotated video of the demonstration is available at [5].

Connected Cows: Utilizing Fog and Cloud Analytics toward Data-Driven Decisions for Smart Dairy Farming

Article

Dec 2019

The Internet of Things (IoT) is about connecting people, processes, data, and things, and is changing the way we monitor and interact with things. An active incorporation of information and communication technology coupled with sophisticated data analytics approaches has the potential to transform some of the oldest industries in the world, including dairy farming. It presents a great opportunity for verticals such as the dairy industry to increase productivity by getting actionable insights to improve farming practices, thereby increasing efficiency and yield. Dairy farms have all the constraints of a modern business -- they have a fixed production capacity, a herd to manage, expensive farm labor, and other varied farm-related processes to take care of. In this technology-driven era farmers look for assistance from smart solutions to increase profitability and to help manage their farms well. We present an end-to-end IoT application system with fog assistance and cloud support that analyzes data generated from wearables on cows' feet to detect anomalies in animal behavior that relate to illness such as lameness. The solution leverages behavioral analytics to generate early alerts toward the animals' well being, thus assisting the farmer in livestock monitoring. This in turn also helps in increasing productivity and milk yield by identifying potential diseases early on. The project specializes in detecting lameness in dairy cattle at an early stage, before visible signs appear to the farmer or an animal expert. Our trial results in a real-world smart dairy farm setup, consisting of a dairy herd of 150 cows in Ireland, demonstrate that the designed system delivers a lameness detection alert up to three days in advance of manual observation.

IoT and agriculture data analysis for smart farm

Article

Jan 2019
COMPUT ELECTRON AGR

In this paper, we propose developing a system optimally watering agricultural crops based on a wireless sensor network. This work aimed to design and develop a control system using node sensors in the crop field with data management via smartphone and a web application. The three components are hardware, web application, and mobile application. The first component was designed and implemented in control box hardware connected to collect data on the crops. Soil moisture sensors are used to monitor the field, connecting to the control box. The second component is a web-based application that was designed and implemented to manipulate the details of crop data and field information. This component applied data mining to analyze the data for predicting suitable temperature, humidity, and soil moisture for optimal future management of crops growth. The final component is mainly used to control crop watering through a mobile application in a smartphone. This allows either automatic or manual control by the user. The automatic control uses data from soil moisture sensors for watering. However, the user can opt for manual control of watering the crops in the functional control mode. The system can send notifications through LINE API for the LINE application. The system was implemented and tested in Makhamtia District, Suratthani Province, Thailand. The results showed the implementation to be useful in agriculture. The moisture content of the soil was maintained appropriately for vegetable growth, reducing costs and increasing agricultural productivity. Moreover, this work represents driving agriculture through digital innovation.

Lameness Detection as a Service: Application of Machine Learning to an Internet of Cattle

Conference Paper

Jan 2019

Lameness is a big problem in the dairy industry, farmers are not yet able to adequately solve it because of the high initial setup costs and complex equipment in currently available solutions, and as a result, we propose an end-to-end IoT application that leverages advanced machine learning and data analytics techniques to identify lame dairy cattle. As part of a real world trial in Waterford, Ireland, 150 dairy cows were each fitted with a long range pedometer. The mobility data from the sensors attached to the front leg of each cow is aggregated at the fog node to form time series of behavioral activities (e.g. step count, lying time and swaps per hour). These are analyzed in the cloud and lameness anomalies are sent to farmer's mobile device using push notifications. The application and model automatically measure and can gather data continuously such that cows can be monitored daily. This means there is no need for herding the cows, furthermore the clustering technique employed proposes a new approach of having a different model for subsets of animals with similar activity levels as opposed to a one size fits all approach. It also ensures that the custom models dynamically adjust as weather and farm condition change as the application scales. The initial results indicate that we can predict lameness 3 days before it can be visually captured by the farmer with an overall accuracy of 87%. This means that the animal can either be isolated or treated immediately to avoid any further effects of lameness.

Smart farming IoT platform based on edge and cloud computing

Article

Jan 2019

Precision Agriculture (PA), as the integration of information, communication and control technologies in agriculture, is growing day by day. The Internet of Things (IoT) and cloud computing paradigms offer advances to enhance PA connectivity. Nevertheless, their usage in this field is usually limited to specific scenarios of high cost, and they are not adapted to semi-arid conditions, or do not cover all PA management in an efficient way. For this reason, we propose a flexible platform able to cope with soilless culture needs in full recirculation greenhouses using moderately saline water. It is based on exchangeable low-cost hardware and supported by a three-tier open source software platform at local, edge and cloud planes. At the local plane, Cyber-Physical Systems (CPS) interact with crop devices to gather data and perform real-time atomic control actions. The edge plane of the platform is in charge of monitoring and managing main PA tasks near the access network to increase system reliability against network access failures. Finally, the cloud platform collects current and past records and hosts data analytics modules in a FIWARE deployment. IoT protocols like Message Queue Telemetry Transport (MQTT) or Constrained Application Protocol (CoAP) are used to communicate with CPS, while Next Generation Service Interface (NGSI) is employed for southbound and northbound access to the cloud. The system has been completely instantiated in a real prototype in frames of the EU DrainUse project, allowing the control of a real hydroponic closed system through managing software for final farmers connected to the platform.

A Creative IoT Agriculture Platform for Cloud Fog Computing

Article

Nov 2018

The innovative service process is a process that uses newly developed technologies to improve the current service models. The study proposes a creative service process based on the cloud computing platform of the Internet of Things and it can be used to improve the integration of the current cloud-to-physical networking and to improve the computing speed of the Internet of Things. In the past, most of the computing technologies focused on high-speed computing in cloud computing or remote operations of a single object. If a service requires cloud or fog resources that can make device use of high-speed computing in the cloud and a single point of operation integration on the object side, it will be able to quickly increase in the process of collaboration, the required data will be moved back and forth between Cloud and Fog, speeding up the cloud computing integration schedule. This research uses innovative platform technology to be applied to the cloud agriculture platform. Through cloud integration, it can be applied to large-area data collection and analysis, allowing farmland with limited network information resources to be integrated and automated, including agricultural monitoring automation, pest management image analysis and monitoring, which can be used to solve the predicament of large-area automation construction.

Fog assisted application support for animal behaviour analysis and health monitoring in dairy farming

Conference Paper

Feb 2018

IoT in Agriculture: Designing a Europe-Wide Large-Scale Pilot

Article

Jan 2017

The technologies associated with the Internet of Things have great potential for application in the domain of food and agriculture, especially in view of the societal and environmental challenges faced by this sector. From farm to fork, IoT technologies could transform the sector, contributing to food safety, and the reduction of agricultural inputs and food waste. A major step toward greater uptake of these technologies will be the execution of IoT-based large-scale pilots (LSPs) in the entire supply chain. This article outlines the challenges and constraints that an LSP deployment of IoT in this domain must consider. Sectoral and technological challenges are described in order to identify a set of technological and agrifood requirements. An architecture based on a system of systems approach is briefly presented, the importance of addressing the interoperability challenges faced by this sector is highlighted, and we elaborate on requirements for new business models, security, privacy, and data governance. A description of the technologies and solutions involved in designing pilots for four agrifood domains (dairy, fruit, arable, meat and vegetable supply chain) is eventually provided. In conclusion, it is noted that for IoT to be successful in this domain, a significant change of culture is needed.

Machine learning based fog computing assisted data-driven approach for early lameness detection in dairy cattle

Abstract and Figures

Recommended publications

Lameness Detection as a Service: Application of Machine Learning to an Internet of Cattle

SmartHerd management: A microservices‐based fog computing–assisted IoT platform towards data‐driven...

Connected Cows: Utilizing Fog and Cloud Analytics toward Data-Driven Decisions for Smart Dairy Farmi...

Fog assisted application support for animal behaviour analysis and health monitoring in dairy farmin...