Internet of Things and Machine Learning techniques in poultry health and welfare management: A systematic literature review

The advent of digital technologies has brought substantial improvements in various domains. This article provides a comprehensive review of research emphasizing AI-enabled IoT applications in poultry health and welfare management. This study focused on poultry welfare since modern poultry management is confronted with issues relating to standardized parameters for welfare assessment and robust monitoring systems, particularly for broilers ’ health and disease outbreak prevention. Evidence has shown that modern digital technologies have high possibilities for intelligent automation of current and future poultry management operations to facilitate high-quality and low-cost poultry production. Therefore, this study presents a systematic review of the current state-of-the-art AI-enabled IoT systems and their recent advances in developing intelligent systems in this domain. Also, the study provides an overview of the critical applications of identified digital technologies in poultry welfare management. Lastly, the study discusses the challenges and opportunities of AI and IoT in poultry farming.


Introduction
The demand for poultry meat is progressively increasing because of its high protein, low energy, and low cholesterol (Lashari et al., 2018).However, poultry high production depends on the environmental condition, disease outbreaks, breeding process, and active management operations (Lashari et al., 2018).Therefore, efficient poultry health and welfare management is essential to prevent infectious diseases, boost production, and ensure healthy broilers.Nevertheless, traditional chicken poultry welfare management approaches are fraught with high labour costs and inefficient resources management, i.e., feed, water, and power consumption.In this context, the integration of Internet of Things (IoT) and Machine Learning (ML) has been considered promising technologies for delivering smart poultry farming, continuous data monitoring and prescriptive analytics in order to address the above identified challenges for efficient resource control and optimal decision-making (Fang et al., 2021;Ribeiro et al., 2019).Accordingly, Raj and Jayanthi (2018) posited that AI-enabled IoT systems could help poultry farm owners enhance production while substantially lowering costs.
IoT comprises many physical sensing devices connected to a Wide Area Network (WAN) to collect, share, and convey information for analysis purposes, while ML is a computational process of unearthing new insights and facts through analytics and a learning process (Michalski et al., 2013).Evidence abounds in the literature on how IoT technologies have been used to assess and control variables such as temperature, humidity, vibration, and air pollutants in poultry houses (Lashari et al., 2018;Lin et al., 2016;W. Pereira et al., 2020).Moreso, poultry feeding and watering systems can incorporate IoT for optimal disease control and management to increase production, enhance safety, and improve profit (Lashari et al., 2018;Ribeiro et al., 2019).When fed into ML, IoT data have been used to detect and classify diseases that are known to have devastating impacts on poultry production and human health, especially the zoonotic poultry pathogens (Cuan et al., 2020).Due to the recent advances in AI-enabled IoT systems for poultry welfare management, this study presents a systematic survey of the current state-of-the-art technologies regarding poultry health and welfare management.A growing body of literature on digital technology applications in the agro-industry has been published in recent years, however, available peer-reviewed articles on AI-enabled IoT systems in poultry health and welfare management are very scarce.Besides, most of the earlier studies are focused on specific aspects of poultry welfare after this procedure.
Hereafter, the quality of documents was evaluated through an exhaustive review to ensure they were relevant to the study's research questions.For example, concerning poultry health management, each article was checked to determine whether it satisfies the following: Explains the use of ML or deep learning in poultry health and welfare, Describes the use of IoT or sensors to enrich data collection in poultry health and welfare, and, Describes smart poultry management systems in the context of Precision Livestock Farming (PLF).
A Precision Livestock Farming (PLF) system employs data collection and monitoring, analysis, ML, control systems, and ICT.Consequently, a total of 93 articles for the period between 2010 and 2022 were left for further analysis after the selection and filtering procedures.These 93 articles provided a representative sampling of IoT and ML applications in poultry health and welfare management.Table 2 shows the number of selected articles by journals and conference proceedings.The current study found the top four journals outlets related to IoT and AI in poultry health and welfare management as Computers and Electronics in Agriculture (26 papers), Biosystems Engineering (14 papers), Information Processing in Agriculture (3 Papers), Poultry Science (3), and Sensors (3 papers).Also, IEEE organized conferences have the largest number of papers (9) related to IoT and ML when compared to other conferences.

Quantitative content analysis
(1) Publication's timeline-The 89 articles published from January 2010 to June 2022 revealed the rapid interest in AI-enabled IoT technologies for poultry welfare management.As shown in Fig. 1, IoT or ML applications for chicken poultry health and welfare management could be divided into three phases according to their publication year.The first phase, 2010 to 2013, had a relatively small number of articles due to the complexity and limited coverage of applying the latest digital technologies in poultry health management.Thus, they have not received much academic attention.The second phase, 2014 to 2018, showed the number of related articles increasing slightly, indicating a growth in the research interest in digital technologies for chicken poultry health and safety management.Finally, the third phase from, 2019 to 2022, revealed a sharp increase in IoT and AI, especially the deep learning research for poultry health and welfare management.A slight drop in publications is noticeable in 2021 and 2022, probably due to Covid-19 impacts on quality research outputs.However, it is anticipated that this number will continue to surge, considering the success of deep learning-based models over classical ML-based approaches in other application domains.
(2) Geographical distribution-Fig. 2 summarizes the country of origin of the selected papers.Every continent is represented by at least one research work as follows: Africa (4, 3.22 %), America (24, 26.88 %), Asia (33,35.48 %), Europe (31,33.33 %), and Oceania (1, 1.07 %).China has the largest number of peerreviewed papers, followed by the USA, Brazil, and Belgium.It is interesting to note that Chinese researchers have been actively involved in this research area because China is home to several artificial intelligence and sensor superpowers and are heavily investing in agriculture, including animal production.
(3) Word distribution-In this study, 86 articles were text mined for common words relating to IoT and ML in poultry management and constructed a word cloud (a holistic picture) of IoT and ML's main applications in this domain.Fig. 3 revealed common IoT and ML applications as environment monitoring (i.e., temperature and humidity), behavior monitoring (i.e., eating, walking, pecking), disease detection and classification, and control.Most frequently used ML techniques include Support Vector Machines (SVM), deep learning, Bayesian networks, linear regression, neural networks, and decision trees.Everyday IoT technology for data acquisition includes cameras, microphones, and sensors.The visualization in Fig. 3 will enable stakeholders (designers and farm managers) to be well informed of relevant IoT and ML knowledge in poultry welfare management.

Applications of IoT and Machine learning in poultry welfare monitoring
This subsection presents the Systemic Literature Review (SLR) results (Tables 3, 4, and 5) concerning answers to research questions.Consequently, the leading IoT and ML applications in poultry health and welfare management were identified in the literature as (1) Behavior/ Environment monitoring, (2) Disease Analytics, and (3) Control/Intervention. Also, Fig. 4 summarizes a holistic view of digital technology for poultry health and welfare applications, AI/ML techniques, tools, data monitored, data types, processing types, diseases, and health status.

Behavior/Environmental monitoring
In this category (Table 3), the selected papers numbering about 57 (61.29 %) focused on systems that remotely monitor chicken behavioral characteristics, i.e., feeding, resting, running, and environmental parameters (temperature, relative humidity).This intervention's main idea is to add value to poultry farmers by facilitating automated acquisition of relevant data using IoT technologies (i.e., sensors, cameras, microphones, mobile phones) and transmitting such to a server or cloud for instant processing and visualization.These automated monitoring systems facilitate the continuous measurements of broilers in a non-intrusive and non-invasive way.

Control intervention in poultry health management
Proper monitoring of environmental parameters, i.e., temperature, humidity, ventilation, and lighting in poultry houses, is essential to guarantee optimal rearing conditions.In addition, their simultaneous supervision and control will reduce energy consumption and increase productivity (Sitaram et al., 2018).Studies (12,12.90 %) in this category (Table 5) highlight the use of sensors to monitor and control environmental conditions by activating appropriate devices, i.e., ventilations, lightning, cooling, and heating systems (Choukidar & Dawande, 2017;Demmers et al., 2010;Gunawan et al., 2019;Lahlouh et al., 2020;Li et al., 2020a;Lorencena et al., 2020;Mirzaee-Ghaleh et al., 2015;So-In et al., 2014).For instance, Zhang and Chen (2020) developed an automatic detection system for sick chickens based on the ResNet residual network (accuracy of 93.70 %) to monitor broilers' behavioral physiology and production performance.
Also, Lorencena et al. (2020) proposed a framework to control and supervise temperature and humidity to aid optimal decision-making in poultry farming.In the same vein, Gunawan et al. (2019) developed a system to maintain the optimum environmental conditions, where ammonia and carbon dioxide levels were regulated using exhaust DC fans.Similarly, humidity and temperature were controlled by DC fans and heat lamps.Water quality management in poultry farms is essential for chicken growth and for controlling bacterial diseases.Thus, Choukidar and Dawande (2017) connected sensors to a Raspberry Pi to control the water level and other parameters, i.e., temperature, smoke, gas, and food dispensing in a poultry farm.Furthermore, So-In et al.
(2014) developed a low computational complexity system of 80.00 % accuracy to automatically adjust the poultry environmental behavior using temperature, humidity, light intensity, and population density.
In addition, feeding systems optimization is necessary for improved production efficiency and animal welfare to promote leg health in broiler chickens.Thus, control strategies to optimize feeding over the entire period of growth to reduce the feed intake and cost have been developed (Demmers et al., 2010(Demmers et al., , 2018;;Kakhki et al., 2019;Klotz et al., 2022).Finally, Lahlouh et al. (2020) developed a system with a 97.00 % accuracy to control hygro-thermal parameters (temperature and relative humidity) and contaminant gases to provide appropriate conditions suitable for efficient poultry production.

AI/ML techniques for poultry health and welfare management
Research in machine learning (ML) has been concerned with building computer programs to construct new knowledge or improve already possessed knowledge by using input information (Michalski et al., 2013).Thus, ML-based techniques generally consist of feature extractors that transform raw data (i.e., pixel values of images) into feature vectors and learning subsystems that regress or classify patterns in the extracted features.However, constructing feature extractors in conventional ML requires laborious adjustment and considerable technical expertise, which limits their applications (LeCun et al., 2015).In contrast, deep learning techniques, developed from conventional ML techniques, are representation-learning methods that can automatically discover features (or data representation) from raw data without extensive engineering knowledge on feature extraction (LeCun et al., 2015).Different deep learning architectures available are deep neural networks,   Most ML models use historical data as input to predict new output values.ML is classified as either supervised or unsupervised (Milosevic et al., 2019), with many learning models (i.e., classification, regression, clustering) and learning algorithms (i.e., ANN, SVR, random forest, CNN, GLM).Supervised learning algorithms use labelled data in their development for accurate classifications or predictions.Unsupervised learning, on the other hand, uses no pre-assigned labels in unravelling unique patterns in datasets.ML techniques, i.e., deep learning (22 %), SVM (13 %), linear regression (9 %), neural networks (8 %), random forest (4 %), decision trees (4 %), logistic (4 %), k-means (4 %), and fuzzy logic (3 %) have, in reality, seen increasing usage in the poultry welfare management domain.A brief discussion on studies that have applied ML techniques for poultry health and management follows.
Also, the heat stress of broilers in commercial broiler-houses negatively affects poultry farm productivity and profitability.ML techniques adopted in estimating the body temperature (heat stress) of broilers include Lasso regression (Bloch et al., 2020), Fuzzy-GA (Hernández-Julio et al., 2020), and generalized sequential pattern (Branco et al., 2020).Furthermore, monitoring poultry welfare and behavioral activities, i.e., eating, drinking, preening, and resting, are important as good welfare promotes healthy chicken growth and improves production.Consequently, ML techniques used in monitoring welfare and behavioral activities include linear regression (Aydin et al., 2015;Li et al., 2020b;Neves et al., 2015) and decision trees (Reboiro-Jato et al., 2011) for monitoring feed and water utilization.Also, Kakhki et al. (2019) employed neuro-fuzzy techniques to optimize feed consumption in poultry farms.
In addition, Banerjee et al. (2012) compared the predictive performance of decision trees, ANN, Naïve Bayes, and radial basis function for poultry activity monitoring, Pereira et al. (2013) (Dawkins et al., 2021;Fernández et al., 2018) and Bayesian regression (Roberts et al., 2012) were also used to monitor and evaluate the risk of welfare issues.In the density map estimation of crowded chicken, Cheng et al. (2019) used deep learning to estimate the density and counting of poultry in farms.ANN was used to generate action plans for broiler house management (Ribeiro et al., 2019) and k-means for evaluating resource usage (Sibanda et al., 2020).Also, Rico-Contreras et al. ( 2017) used ANN and fuzzy logic to predict litters' moisture content, while So-In et al. (2014) employed k and fuzzy-C means to manage poultry population density.
Furthermore, weight is an essential parameter for estimating poultry farms' growth and feed conversion efficiency.Consequently, ML techniques, i.e., linear regression (Fontana et al., 2015;Kashiha et al., 2013;Mollah et al., 2010), support vector regression (Amraei et al., 2017b), and Bayesian artificial neural network (Mortensen et al., 2016) have been used for broilers' growth estimation.Other ML techniques used are deep learning (Demmers et al., 2010;Huang et al., 2021;Johansen et al., 2019), ANN (Amraei et al., 2017a;Ma et al., 2020), and quantile regression forests (Diez-Olivan et al., 2019).Likewise, lameness is one of the causes of poor welfare in poultry and early detection of lameness will allow farmers and veterinarians to take timely management actions in time.ML techniques for detecting lameness in broilers include decision trees (de Alencar Nääs et al., 2020) and linear regression (Van Hertem et al., 2018).
The occurrence of poultry diseases affects poultry welfare and production, food safety, and zoonotic infections.Hence, ML techniques have been employed for the timely detection of these diseases.For instance, Raj and Jayanthi (2019) evaluated the predictive performance of KNN, SVM, logistic decision, linear and quadratic discriminant techniques in detecting avian influenza.Qiang and Kou (2019) also benchmarked support vector machines, Bayesian networks, and kNN to predict avian flu.Other studies using ML techniques to detect avian influenza are SVM (Aziz & Othman, 2017;Huang et al., 2019;Zhuang et al., 2018), random forest, and maximum entropy (Belkhiria et al., 2020), deep learning (Cuan et al., 2020(Cuan et al., , 2022a) and association rules and sequential pattern mining (Xu et al., 2017).

Progress of deep learning techniques in poultry health and welfare management
Deep learning (DL) is widely used in different fields.For example, it has proven to be an efficient method commonly used in various agriculture-related areas (Yao et al., 2020).The great advantage of using deep learning is the reduced need for feature engineering since deep neural networks are directly involved in the extraction of intrinsic attributes, i.e., color, shape, and texture information (LeCun et al., 2015).
The use of deep learning in poultry health and welfare management is on the rise, especially, DL-based approaches, including Faster R-CNN, You Only Look Once (YOLO), and Single Shot multibox Detector (SSD) have been applied for object detection in poultry in recent years.For instance, Cao et al. (2021) proposed an automated chicken counting method with Dense CNN.Jung et al. (2021) developed CNN models to automate the classification of vocalizations of laying hens and cattle.In addition, Zhuang and Zang (2019) implemented a deep learning variantimproved Single Shot MultiBox Detector, to automatically diagnose broilers' health status.The proposed algorithm achieved 99.70 % mean average precision.Similarly, Huang et al. (2021) used combined RNN, LSTM, and GTU to detect poultry eating behavior based on vocalization signals with an accuracy of 96.00 %, while Cheng et al. (2019) used fully convolutional networks for density estimation and poultry counting.Zhang and Chen (2020) also designed an automatic detection system for sick chickens using an improved residual Network (ResNet), which resulted in 93.70 % recognition accuracy.Also, You et al. (2021) developed a DNN model to predict oviposition events for individual broiler breeders for efficient bird management with the area under the receiver operating characteristic (ROC) curve at 0.94.Also, Yao et al. (2020) used deep neural networks (DNNs) to estimate the gender ratio of chickens, and the experimental results achieved an average accuracy of 96.90 %.Similarly, a performance comparison of CNN, LSTM, and GRU was performed to determine chicken gender, with CNN obtaining the highest accuracy of 91.25 % (Cuan et al., 2022a).Also, Geffen et al. (2020) used the Faster R-CNN method to detect and count laying hens in battery cages with 89.60 % accuracy of a validation dataset of 2000 images.These studies on deep learning confirm its reliability and efficiency in poultry welfare management.GANs application in poultry welfare is a gray area.However, GANs and autoencoders can be combined for robust image classification to provide an unsupervised data augmentation method for poultry-related computer vision problems.In addition, GANs can synthesize additional realistic images that resemble those in the training set.
From the papers reviewed, the applications of a few DL techniques in poultry welfare management are basically for object detection (Geffen et al., 2020; Zhang & Chen, 2020), growth prediction (Demmers et al., 2018), disease management (Zhuang & Zhang, 2019), acoustic analysis (Jung et al., 2021), tracking (Cao et al., 2021;Fang et al., 2021;Li et al., 2021;Pu et al., 2018), and habitat monitoring (Debauche et al., 2020).Deep Reinforcement Learning (DRL) is an advanced and active research field in ML that combines deep and reinforcement learning to solve impossible problems (Lei et al., 2020).Moreover, it promises to solve more complex problems in robotics, resource management, and other fields requiring decision-making capabilities.But unfortunately, deep reinforcement learning is yet to be applied in poultry welfare and health management.Huang et al. (2021) also opined that fully functional robots developed and implemented in the poultry production system are limited.However, in poultry welfare management, DRL would be an essential ingredient for interactive perception, mainly for object segmentation, object recognition, or categorization tasks.In addition to perception capability, DRL will facilitate planning and executing low-productivity tasks, reasoning, and communication capabilities to support optimal decision-making by production managers.Fig. 5 describes the current state of DRL applications in poultry health and welfare management.

Processing techniques
This subsection presents an overview of the poultry data processing types regarding computer vision and vocalization techniques.Consequently, classification, regression, and clustering algorithms in poultry health and welfare management were briefly discussed.

Computer vision
Computer vision combines mathematics, computer science, and software programming to provide image-based automated process control (Okinda et al., 2020).Similarly, it comprises hardware and software (image processing and analysis algorithms), with the hardware consisting of computers, cameras, and lighting units (Abd Aziz et al., 2021;Nyalala et al., 2021) to track and monitor the behavior and health status of chickens.The advantages of computer vision systems lie in their non-invasive, non-invasive, and low-cost animal monitoring (Li et al., 2021).Fig. 6 depicts the components of computer vision for poultry health and welfare management.Cameras are one of the core components of computer vision systems, and they are classified into 2D and 3D cameras (Abd Aziz et al., 2021).Digital 2D cameras are common and cheap (Fernández et al., 2018;Van Hertem et al., 2018).They operate with a light-sensing chip (charge-coupled device) to detect the brightness of three filtered color channels (red, green, and blue) at each pixel.The 3D cameras, though more expensive and with fewer pixels per area, allow for broader application as they can capture additional information, i.e., depth (the vertical distance between target and camera), and are less prone to darkness and environmental influences (Okinda et al., 2020).
Depth cameras create a 3D image of the targeted scene or object and offer more discerning information to recover postures and recognize actions.Thus, they have been deployed to monitor poultry's behavior and health status.For instance, studies such as those from Aydin et al.The lighting units (i.e., LED and Halogen) illuminate the pen or poultry house to ensure that the illumination intensity is within an acceptable range (typically between 15 and 20 lx) (Wang et al., 2016), for improved image processing and analysis operations.From the software perspective, it is a set of programs or routines associated with computer system operations, and its development plays an essential role in computer vision (Abd Aziz et al., 2021).Software forms in computer vision include image pre-processing for image quality enhancement, image segmentation, and feature extraction, which extracts meaningful information from images.
Image segmentation partitions images into multiple segments and labels the segments with known classes to facilitate accurate classification (Abd Aziz et al., 2021;Zhuang et al., 2018).Though this field has a long research history, deep learning networks have delivered models with remarkable performance for segmentation and have thus become the new standard for image segmentation (Okinda et al., 2020).Techniques for image segmentation (i.e., extracting a chicken body from a background image) in poultry welfare management include the Watershed algorithm (Cao et al., 2021;Mortensen et al., 2016) and the Ellipse model (Kashiha et al., 2014;Zhuang et al., 2018).Others are Mean-shift clustering (Fang et al., 2020) and K-Means clustering (Zhuang et al., 2018).
Feature extraction algorithms transform raw data into a suitable internal representation to enhance ML models' predictive ability (LeCun et al., 2015).Thus, extracting features from images helps to correlate objects to a specific bio-response or bio-process under investigation (Okinda et al., 2020).Consequently, after the pre-processed and segmented operations, selected features of images are extracted for classification purposes.Common features used for conventional image analysis in poultry include age (1D), morphological features (i.e., length, breadth, area, perimeter), and 3D (volume and surface area) (Abd Aziz et al., 2021;Okinda et al., 2019Okinda et al., , 2020)).For instance, Mollah et al. (2010) used 1D and morphological features to examine the relationship between manual body weight and the number of surface-area pixels in the image.Features such as area are mostly used to estimate the broiler body size.The area was computed by summing pixels within a contour constituting the broiler image to analyse broiler behavior (Pereira et al., 2013).Aydin (2017) also used 1D features in detecting lameness in broilers.Morphological features have also been used to measure feed intakes (Aydin et al., 2015;Mehdizadeh et al., 2015;Neves et al., 2015) and strentching (Li et al., 2021) 2016) used age, morphological features, and 3D to determine optimal ways to control food, water supplies, and circadian rhythm of broilers for optimal growth patterns.
Finally, the gender ratio of free-range chickens is considered a major welfare problem in commercial broiler farming.Thus, systems to support free-range chicken producers have been developed for chicken counting (Cao et al., 2021) and identifying chicken gender for flock economic value (Yao et al., 2020).

Vocalization analysis
Analysis of broiler vocalizations can yield valuable insights into poultry welfare and how diseases manifest and progress over time.In addition, this technique can play an important role in detecting infection with pathogenic microorganisms, threat signals, information about feeding, activity monitoring, and population estimation.The key advantage of the technology is the continuous non-invasive audio measurements of the poultry environment at a relatively low cost (Jung et al., 2021).Thus, vocalization analysis applications in poultry welfare management abound.The following are some examples.
The frequency of rales produced by infected chickens was used to detect respiratory infection and initiate remedial actions to inhibit further infection or spread before clinical signs manifestation (Carroll et al., 2014;Cuan et al., 2020;Rizwan et al., 2016).Also, Carpentier et al. (2019) presented an algorithm with a precision of 88.40 to monitor the chicken sneezing sounds from multiple broilers' vocalizations in a noisy environment, while Banakar et al. (2016) proposed a system with a 91.20 % accuracy for detecting and diagnosing respiratory diseases.Du et al. (2018) developed a system to detect anomaly poultry status at night by monitoring the number of vocalizations and area distributions.The proposed approach (74.70 % accuracy) is a practical and feasible method for poultry behavior and welfare, especially in stress detection.Feeding behavior detection of broiler chickens is vital in distinguishing healthy from infected birds.Furthermore, pecking sounds have been used to monitor the feed intake of broilers by a real-time sound processing technology (Aydin et al., 2015;Aydin & Berckmans, 2016).Also, Fontana et al. (2015) proposed a method to automatically measure the growth rate of broiler chickens by sound analysis (R 2 = 0.98).Cuan et al. (2022a) obtained a classification accuracy of 91.25 % for chicken gender determination, while Jung et al. (2021) developed a strategy (75.80 % accuracy) for chicken gender identification for economic value estimation.
In summary, these studies have revealed potential applications of vocalization analysis due to its good performance and low computational cost to optimize conditions of the poultry environment and detect behavioral problems, i.e., feather pecking, feed intakes, infections, and stress.Thus, sound technology has real potential for practical commercial implementation to improve health and poultry welfare.Features commonly used in previous studies (i.e., Aydin et al., 2015;Aydin & Berckmans, 2016;Banakar et al., 2016;Fontana et al., 2015) for detecting and classifying poultry vocalization data include pitch, frequency, and time-frequency.Similarly, studies such as those from Carroll et al. ( 2014), Cuan et al. (2022b), Cuan et al. (2022a) and Jung et al. (2021) used the Mel frequency cepstrum coefficients (MFCCs) to represent the acoustic sound of hens to recognize behavioral meanings from those sounds.

Data processing
This subsection briefly describes data pre-processing techniques for poultry audio and image data.
Data processing involves collecting and manipulating data to produce meaningful information, and pre-processing of high-dimensional features is a general and robust method for improving the learning algorithm performance (Hastie et al., 2009).Data in the context of this study refers to text and multimedia data, i.e., images, audios, and videos employed in poultry welfare management.Consequently, data processing comprises techniques to improve data further to eliminate noise or unwanted frequencies and increase measurement precision and analysis reliability (Cuan et al., 2022a;Pereira et al., 2013;Huang et al., 2021).
For instance, in computer vision problems, the primary purpose of this process is to enhance the image quality for the segmentation step, especially in separating a digital image into distinct areas (Nyalala et al., 2021).Standard pre-processing techniques include filtering, normalization, approximation, enhancement, and cancellation of points (Okinda et al., 2020).In addition, many of these pre-processing methods used in poultry welfare management include dilation and erosion (Amraei et al., 2017a;Kashiha et al., 2014;Mollah et al., 2010;So-In et al., 2014), thresholding (Kashiha et al., 2013;Mehdizadeh et al., 2015;Neves et al., 2015;Okinda et al., 2019;Youssef et al., 2015), Otsu's method (Kashiha et al., 2014;Okinda et al., 2019), gaussian filter (Mortensen et al., 2016;Raj & Jayanthi, 2019), and binarization (Guo et al., 2020;Li et al., 2020b;Neves et al., 2015;Okinda et al., 2019;Pereira et al., 2013).Others include image cropping to emphasize a particular subset of a larger image (Li et al., 2021) and image size reduction (Li et al., 2020a), as large image sizes can decrease detection speed when input into CNN detectors.Also is data augmentation (Zhang & Chen, 2020), to increase the amount of data by adding slightly modified copies of already existing data or newly created synthetic data from existing data.
Also, in poultry vocalization analysis, studies such as those from Huang et al. (2021), Klotz et al. (2022), andPereira et al. (2013) used data normalization, a technique to convert data from spatial domain to frequency domain with either fast Fourier transform or Discrete cosine transform techniques (Cuan et al., 2022b;Huang et al., 2021;Mahdavian et al., 2021).Similarly, the uses of pre-emphasis and micro-analysis techniques are standard pre-processing techniques used on audio signals (Jung et al., 2021;Huang et al., 2019;Huang et al., 2021).

Classification
This subsection briefly describes the classification technique as one of the key processing techniques in poultry welfare management.
The classification technique categorizes data into classes to identify primarily the category under which new data or objects belong (Okinda et al., 2020), and it is the most used technique for delineating classes of output (usually categorical) based on some set of input features (Ajayi et al., 2020).Also, the classification task is essential for computer vision and vocalization analysis (Amraei et al., 2017b;Aydin, 2017;Aydin et al., 2015;Aydin & Berckmans, 2016) as it helps to identify objects and tasks performed by those objects.For example, given an image or sound, classification techniques can help discriminate between healthy and sick broilers or determine whether a chicken is laying eggs, walking, or drinking.Examples of classification techniques include CNN, SVM, ANN, DNN, random forests, logistic, and decision trees.However, the high CNN classification accuracy makes it a promising approach, especially for classifying animal sounds (Jung et al., 2021).As revealed in this study, typical classification applications in poultry welfare management are activity recognition and disease detection.A few examples are presented.
For activity recognition, Banerjee et al. ( 2012) compared decision trees, Naïve Bayes, and neural networks for broilers' activity recognition and reported that neural networks had the best overall accuracy of 82.10 %.Similarly, Pereira et al. (2013) employed the classification tree algorithm to identify hen white broiler breeder behavior and reported an overall success rate of 70.30 % on the validation set.Also, Rizwan et al. (2016) compared extreme learning machine and SVM classifiers to detect rales (a gurgling sound that is a symptom of respiratory diseases in poultry) and reported a classification accuracy of SVM (97.60 %) over the extreme learning machine technique.Pu et al. (2018) developed an automatic CNN-based method (accuracy of 99.17 %) to recognize the chicken behavior within a poultry architecture.Also, a CNN classification model, with a classification accuracy of 75.78 % on the validation dataset, was used to effectively recognize the sounds of laying hens (Jung et al., 2021).Geffen et al. (2020) used Faster R-CNN detection and tracking algorithms to detect hens in cages with an 89.60 % accuracy at cage level.Li et al. (2020a) used CNN to detect drinking behaviors of pullets in a lighting preference test system, and they reported a classification accuracy of 89.10 %.Also, Fang et al. (2021) used a Naive Bayesian model to classify and identify the poses of broiler chickens.They reported the test precision of behavior recognition at 0.75 (standing), 0.51(walking), 0.63 (running), 0.94 (eating), 0.96 (resting), and 0.93 (preening).Lastly, Li et al. (2021) used a faster R-CNN method to detect broiler stretching behaviors with an accuracy of 99.50 % on the testing dataset.Nasiri et al. (2022)  Similarly, timely disease detection is paramount in poultry production (Zhang & Chen, 2020).The following are a few examples of classification models in poultry disease detection and diagnosis.Huang et al. (2019) developed an audio analysis-based detection method to detect avian influenza in chickens using SVM (a binary classification) and reported an accuracy rate between 84.00 % and 90.00 %.Also, Cuan et al. (2020) proposed a CNN model to detect chickens with avian influenza from chicken sound extracts and reported the highest accuracy of 95.84 %.Also, Okinda et al. (2019) compared the performance of an SVM classifier with ANN and logit regression to establish a correlation of broiler feature variables with their health status.The authors reported that an SVM classifier outperformed all other models with an accuracy of 97.80 %.Zhang et al. ( 2018) developed a system to analyze broilers' postures and detect sick broilers with an SVM classifier.The authors reported an accuracy rate of 99.47 % on the test samples.Also, Zhuang and Zhang (2019) proposed a CNN-based recognition model with a 99.70 % mean average precision (mAP) for detecting the health status of broilers to support efficient flock management.Similarly, Zhang and Chen (2020) developed an automatic detection system for sick chickens based on ResNet residual (a CNN architecture) classifier.They reported accuracy of 93.70 % on the test set.
This methodology often used to describe relationships between independent and dependent variables, is a formula usually represented as a line to make predictions, including interpolations and extrapolations (Hastie et al., 2009).Multivariable regression uses more than one independent variable to predict an outcome (Küçüktopcu & Cemek, 2021).The two modes of tackling regression problems are linear and nonlinear regression methods.Linear regression models are simple and often provide an adequate and interpretable description of how the inputs affect the output (Hastie et al., 2009).The nonlinear models are complex and used for modeling the nonlinear pattern relationship between predictor and response variables.Regression methods mostly applied in poultry monitoring include linear, logistic, lasso, SVM, ANN, RNN, DNN, and tree model regression (Okinda et al., 2020).
Body weight is an essential indicator for determining poultry's growth and health status as it is closely related to production performance (Johansen et al., 2019).Consequently, with respect to poultry, regression-based models have been used to analyze poultry growth curves (Amraei, Abdanan, et al., 2017;Demmers et al., 2010Demmers et al., , 2018;;Fontana et al., 2015;Johansen et al., 2019;Mollah et al., 2010;Mortensen et al., 2016).For example, linear regression was applied to estimate chicken weight, with a relative error in weight estimation of chicken, expressed in terms of percent error of the residuals from surface area pixels between 0.04 and 16.47 (Mollah et al., 2010).Mortensen et al. (2016) applied Bayesian ANN regression for broiler weight estimation and reported a relative mean error of 7.8 % for the proposed model.Amraei et al. (2017a) used neural networks to estimate the live body weight of broilers and reported an R-Squared value of 0.98.Amraei et al. (2017b) used support vector regression (SVR) to estimate the weight of life broiler chickens with an R-Squared value of 0.98.Also, Fontana et al. (2015) proposed a method to automatically measure the growth rate of broiler chickens by sound analysis (R2 = 0.98).
Ammonia is a primary air pollutant in poultry farms, adversely affecting the ecosystem, environment, birds, and human health (Küçüktopcu & Cemek, 2021).Therefore, estimating NH3 concentration is essential for proper litter management and protecting environmental, human, and animal health (Debauche et al., 2020;Küçüktopcu & Cemek, 2021).Hence, Küçüktopcu and Cemek (2021) performed a performance comparison of four models (i.e., multilayer perceptron, integrated adaptive neuro-fuzzy inference systems with grid partitioning and subtractive clustering (ANFIS-GP and ANFIS-SC), and multiple linear regression analysis) to estimate ammonia concentration in poultry.They reported that ANFIS-SC was more accurate, with an R-Squared value of 0.86 on the validation set.
Also, Lin et al. (2016) developed an ammonia monitor for a poultry farm and reported a relative error (RE) of the monitor averaged 7 %.Debauch et al. ( 2020) also developed a system to monitor and predict the air quality in poultry using GRU and reported improvements in both the training and prediction speeds.Other applications of regressionbased models are for analyzing litter moisture content (Rico-Contreras et al., 2017), evaluating chicken gender ratio (Yao et al., 2020), and counting birds in poultry farms (Cao et al., 2021;Cheng et al., 2019;Geffen et al., 2020).

Clustering
This subsection presents a brief highlights of clustering techniques and their application in poultry welfare management.
The clustering technique divides data into a group of similar objects that are different from objects of other groups (Ismail et al., 2016).The clustering technique is often one of the first steps in data mining analysis and therefore supports the development of population segmentation models (Zhang et al., 2018).Examples of clustering techniques include K-means, Gaussian mixture, K-Medoids, and Fuzzy C-Means.However, K-means application is more common in poultry welfare management (Carroll et al. (2014;Feiyang et al., 2016;Küçüktopcu & Cemek, 2021).The K-means technique aggregates similar objects concerning their characteristics based on a distance measure, i.e., Euclidean distance (Hastie et al., 2009).
For instance, Carroll et al. (2014) used the k-means algorithm to cluster the vectors of MFCCs from chicken audio data into 60 clusters, to yield a single cluster index for each time slice when detecting that detecting chicken making rales sound.Feiyang et al. (2016) used the Kmeans method to classify poultry into sick, normal, and active to boost a classifier while analyzing chicken behavior characteristics, i.e., speed, ability to snatch food, and resting time.Thus, ensuring that chicken diseases are timely detected and the accurate growth states of chickens are immediately known.Clustering is a useful technique for discovering data distribution and patterns in the underlying data.Guo et al. (2021) developed a model to discover chicken distribution and compared a method integrating the GB (Green/Blue) color space and twodimensional Otsu processing with K-means and Fuzzy C-Means techniques concerning the processing time and target extraction.The Fuzzy C-Means method was reported to extract individual broiler from original images with a reasonable visualization efficiency (e.g., clearness and completeness of chicken areas).Ismail et al. (2016) evaluated the rapid centroid estimation (RCE), a lightweight swarm clustering algorithm with k-means to cluster the Newcastle disease dataset, and reported that RCE shows better external clustering quality measures than K-means.
Similarly, while estimating the ammonia concentration in poultry farms, Küçüktopcu and Cemek (2021) used a subtractive clustering technique to determine optimal input parameters for the regression model.Zhuang et al. (2018) also proposed a real-time poultry segmentation algorithm based on K-means clustering and the ellipse model for automated diagnoses of broilers' health status.

Sensor technologies in poultry welfare
Recently, tremendous advances have been achieved in sensing technologies in terms of diversity, accuracy, and affordability.Sensors, if appropriately deployed, can provide a timely diagnosis of diseases in animals, eventually decreasing economic losses.Furthermore, the primary environmental conditions to control in the poultry buildings are the hygro-thermal parameters (temperature and relative humidity) and contaminant gases, i.e., ammonium and carbon dioxide (Lahlouh et al., 2020).Therefore, such devices are particularly useful for poultry health management (Carpentier et al., 2019).In this regard, a tool for data visualization, as shown in Fig. 7, was created to summarize the relevant and tested sensor-based applications and sensor types used in the studies reviewed.The innermost circle represents the unifying name for the devices, whereas the outer circle represents the sensor types, and the outermost circle represents specific sensors.
Environmental sensors are primarily used to monitor environmental conditions, i.e., temperature, humidity, and air quality, to provide the appropriate conditions suitable for the significant efficiency of animal production.Importantly, inadequate temperature, relative humidity, and the length of exposure have substantial impacts on broiler welfare, mortality, and performance (Debauche et al., 2020;Diez-Olivan et al., 2019).Furthermore, exposure to elevated levels of noxious gases like Fig. 7. Sensor types used in poultry welfare management.DHT11, DHT22, SHT75, and HX71-VI sensors can all be used to monitor temperature and humidity.carbon dioxide and ammonia can reduce weight, feed conversion, overall viability, and loss of profit in the poultry industry (Küçüktopcu & Cemek, 2021).Thus, efforts to monitor and control environmental conditions will directly impact bird welfare and permit the development of systems for precise control of the production environment.Several wireless-based sensors and use IoT tools to monitor environmental parameters have been proposed as a promising tool in PLF (Debauche et al., 2020;Gunawan et al., 2019;Lahlouh et al., 2020;Lashari et al., 2018;Lin et al., 2016;Lorencena et al., 2020;Mirzaee-Ghaleh et al., 2015;Pereira et al., 2020;So-In et al., 2014).
Acoustic sensors are used to measure the characteristics of sounds emitted by hens.Analysis of acoustic data can serve as a reliable stress indicator (Du et al., 2018), measure food intake (Aydin et al., 2015), and disease detection (Cuan et al., 2020;Huang et al., 2019).Avian influenza poses a potential health threat to both chickens and humans.Banakar et al. (2016) developed an avian influenza monitoring system to simulate the spread of highly pathogenic avian influenza viruses in chickens.Results showed the sensor's capability to detect the virus accurately.In addition, other respiratory diseases can be timely detected using appropriate sound technologies (Carpentier et al., 2019).Raj and Jayanthi (2019) used an acoustic sensor, SEN-14262, for the real-time identification of infected hens.
Infrared or thermal sensors (i.e., FLIR Lepton (FLIR Systems, Inc., Oregon, USA, Thermo GEAR-G120, Kinect cameras (Microsoft Corporation)) are non-invasive welfare devices to assess the body's superficial temperature distribution from the infrared radiation emitted by objects.For example, systems incorporating FLIR Lepton cameras have been proposed to measure broiler temperature (Bloch et al., 2020;Hernández-Julio et al., 2020).In addition, the movement of layers across perches and other housing equipment is a risk factor for bone breakage, a typical condition of poor welfare.Thus, Kinect sensors were used to monitor and study different aspects of the movement of broilers and layers (Aydin, 2017;Feiyang et al., 2016;Pu et al., 2018).Also, Banerjee et al. (2012) used the MTS510 accelerometer (Crossbow Technology, Inc., San Jose, California), placed in laying hens to monitor their movement.Other sensors, i.e., FSR402 weight sensors (Interlink Electronics, Camarillo, California), were used to assess the average broiler weight (Lahlouh et al., 2020), and Light Dependent Resistor (LDR) was used to monitor luminosity (Lahlouh et al., 2020).

Key challenges in poultry welfare management
Although there are several studies relating to poultry welfare management, researchers have also improved the method of measuring the pen house atmospheric conditions, health, behavior, weight, and growth of chickens.However, there are still several issues to be addressed.The section describes the issues and challenges in poultry management, including the quality of raw data, the precision of image segmentation, and the reliability of prediction or classification.
(i) New datasets to support further challenging tasks: -although researchers have collected several datasets recently for everyday welfare management tasks, there is a need for new large-scale datasets for more challenging tasks.These new tasks include responding to sudden disease outbreaks, handling new insects and vermin, efficient nutrient utilization, and optimizing feed for improved production.The increasing availability of large data sources and data sets, obtained through sensors will encourage more initiatives, projects and new ventures in the poultry health and welfare management.(ii) Raw data quality: -Challenges in ensuring raw data quality are related to physical actions affecting changes in postures, orientations, and the diversity of birds' body dimension measurement.
Besides, images could be poor due to dust bathing of hyperactive broilers when stretching out wings.Also, image dimension varies due to factors affecting chicken locations (i.e., below cameras, feather level, lighting, image threshold values, and the distance of chickens from cameras).Some researchers omitted the head and tail positions during the feature extraction phase to overcome this challenge (Mortensen et al., 2016).However, this strategy will lead to underestimating the broiler's body weight and behavior compared to the actual.(iii) Deep learning and adversarial perturbations: -in another direction, deep models are fragile to adversarial perturbation on inputs (Papernot et al., 2016).Those changes in data distribution can unpredictably trigger weak features.Thus, leading to a slight decline in performance and ultimately causing deep learning models to make wrong predictions with high confidence.For instance, imperceptible pixel differences in images can trick deep learning models.Such adversarial attacks are an important obstacle to the successful deployment of deep learning, especially in applications involving a proactive response to sudden poultry disease outbreaks.Though some early solutions have been proposed (i.e., distillation, feature aggregation, applying denoising autoencoders on data, using multi-scale networks,) a significant challenge is to develop effective defence mechanisms against these attacks.(iv) Interpretable of deep learning models: -While DL models have achieved promising performance in various challenging problems, their limitation regarding interpretability (" black-box" problem), which aims to explain their output, is a challenge.For example, why a model outperforms another model on one dataset but underperforms on other datasets?What exactly have DL models learned?What is a minimal architecture for achieving a definite accuracy on a given dataset?Although attention and selfattention mechanisms, widely used in many fields because of their ability to distinguish features, can provide some insight toward answering these questions, a detailed study of these models' underlying behavior and dynamics is still lacking.Nevertheless, a thorough understanding of their theoretical aspects can help develop enhanced models for various poultry health and welfare management analysis scenarios.

Proposed poultry welfare framework
Having reviewed existing interventions by other researchers in poultry welfare management, a scalable, resilient, extensible, and secured framework is presented to support the precision livestock farming concept for smart poultry health and welfare in addressing some of the challenges in the existing systems.The critical components of the framework (Fig. 8) include the deep learning (DL) module, Digital Twin module, cloud edge computing (cloud-fog-based) module, communication, security, and user-interface module.blockchain, edge, sensors, and users) is wireless and mobile, as shown in Fig. 8. (iv) Cloud-Edge computing-The successful deployment of a smart poultry welfare solution depends on this technology.The cloudfog-based module stores DL models, databases, and datasets due to better data management.The execution of DL models is carried out in fog computing nodes for increased network performance due to reduced latency.(v) Security module: -The recurrent news about security breaches, private data leaks, and inappropriate use of data has recently made the security of the IoT platforms necessary.Thus, the blockchain techniques will be deployed to secure the poultry welfare network platform due to the multitude of devices, sensors, and services involved in data collection and transmission.(vi) Digital-twins module will support real-time data processing and simulations of complex dynamic processes in poultry production with environmental and behavioral-based parameters (i.e., preening, chicken sound, temperature, humidity, ammonia) by advanced deep learning techniques.This module will facilitate better business decisions, improve poultry health and welfarebeing, and maximize the return from agricultural resources.(vii) The user interface module allows users to interact with the system.Through this module, users can receive alarms, visualize data in real-time, perform prospective analysis, and validate actions carried out by the RBL agent in controlling actuators.

Implications for practitioners and academics
The need for transparent, efficient, and sustainable poultry production systems is driving the digital transformation in poultry welfare both for ethical and economic reasons.The stakeholders, therefore, need to reassess their present position concerning the emerging technologies disrupting the agro-industry.Furthermore, with the increasing demand for poultry meat and eggs, and the incessant poultry disease outbreak, which has continued to be a threat and economic burden to the poultry industry, there is a need for a massive deployment of digital technologies to improve poultry disease management and productivity.The review reveals enormous benefits of IoT and ML to poultry welfare.Furthermore, it suggests how contemporary AI and IoT techniques can be harnessed to confront current challenges in poultry welfare to increase efficiency.
Based on the discussions in this study, the following areas need further investigation from researchers.First, more studies on how AI and IoT algorithms can facilitate the optimal use and utilization of resources in poultry welfare are required.Similarly, studies on identifying the relationships between the various barriers to implementing AI-based IoT systems for commercial poultry farming are required.The identification of the driving and dependence barriers will help in expediting the AIbased IoT implementation.

Conclusion
In this article, a comprehensive and systematic review of the applications of AI and IoT in poultry health and welfare management, especially for poultry production, has been provided.Also presented were the latest applications of AI-enabled IoT using various representative studies highlighting processing techniques, data, hardware, and software parts used in poultry welfare systems.In addition, this study illustrated the significant divisions of IoT/AI interventions in poultry and presented a reliable, robust, and extendable framework for poultry welfare, specifically in realizing robust poultry disease outbreak prevention.This study contributes to knowledge by helping stakeholders understand and better harness advanced digital technologies and critically analyse the limitations of poultry farms in recognizing the possible applications and patterns of technological advances in the domain.Also, information regarding technologies for poultry welfare management and optimization of its production process will help facilitate reaching a high quality, short process time, and low-cost production in poultry.Thus, the review will stimulate new lines of reasoning that will improve productivity and profitability in the poultry farming industry.

Fig. 3 .
Fig. 3. Word Cloud of key IoT and AI applications in poultry.

Fig. 4 .
Fig. 4. Overview of IoT and ML in poultry health management.
of broilers.Amraei et al. (2017b) also used 2D (morphological) features for the broiler weight estimation.Debauche et al. (2020) also used morphological features for poultry monitoring.Mortensen et al. ( developed a CNN-LSTM-based model for skeleton-based lameness recognition in broilers, and the model achieved a per-class classification accuracy of 97.5 %, while Cuan et al. (2022a) obtained a classification accuracy of 91.25 % for chicken gender determination.
(v)  The conventional fuzzy-logic controllers deployed in studies such as those fromLahlouh et al. (2020) andMirzaee-Ghaleh et al. (2015) are heavily reliant on the user's knowledge of the system and complicated rules.Besides, conventional controllers (i.e., proportional integral derivative (PID), fuzzy logic, predictive controller optimized by genetic algorithm, particle swarm optimization) deployed in studies such as those from Hernández-Julio et al. (2020),Kakhki et al. (2019),Küçüktopcu and Cemek (2021),Lahlouh et al. (2020) andMirzaee-Ghaleh et al. (2015) suffer from the inability to learn in real-time.However, to overcome this shortcoming, recent technologies such as reinforcement learning (RL), a hotspot in artificial intelligence, and ML techniques can be deployed to design intelligent systems to optimize a policy for complex tasks.(vi) Robust and accurate processing techniques: -The next problem is image processing techniques, i.e., segmentation and feature extraction used in convectional ML techniques.These techniques are confronted with problems (i.e., noise, contrast issues) that affect classification models' accuracy.In addition, the strong dependence on domain knowledge for designing features makes these methods difficult to generalize to new tasks.Finally, these models cannot take advantage of large amounts of training data because the features are pre-defined.However, for DL, the tasks of segmentation, feature extraction, and feature selection are eliminated using CNNs.Similarly, variational autoencoders, based on their ability to learn model parameters through the encoder-decoder path automatically, are commonly used to create compact data representations for efficient decisionmaking.As a result, DL techniques have achieved much success in animal monitoring systems.

Table 2
Selected articles and conference proceedings.

Table 3
IoT and ML for poultry behavior and environment monitoring.
R-Squared 0.93 Historical data ANN, FL Density, temperature, days, feeder type, drinker type Predict litters moisture content (continued on next page) R.O.Ojo et al. recurrent neural networks, deep belief networks, convolutional neural networks, autoencoders, generative adversarial networks, and deep reinforcement learning.Critical factors for deep learning networks' success include deeper architectures to capture invariant properties of data, regularisation techniques to support robust optimization for enhanced performance, massive datasets availability, and efficient computing hardware to solve complex problems (Oyedele et al., 2021).Convolutional neural network is one of the most representative deep learning algorithms in digital image processing (Zhuang & Zhang, 2019).

Table 4
IoT and ML for poultry disease analytics.
R.O.Ojo et al.
also used decision trees, Fang et al. (2020) utilized deep learning techniques, while Zhang et al. (2016) employed k-means clustering for poultry activity monitoring.Linear regression technique

Table 5
IoT and ML for poultry control and intervention.