Animal behavior analysis methods using deep learning: A survey

Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa Animal behavior analysis methods using deep learning: A survey Edoardo Fazzari a,b,c,∗, Donato Romano a,b, Fabrizio Falchi a,c, Cesare Stefanini a,b aThe BioRobotics Institute, Sant’Anna School of Advanced Studies, Viale Rinaldo Piaggio, Pontedera, 56025, Italy bDepartment of Excellence in Robotics and AI, Sant’Anna School of Advanced Studies, Piazza Martiri della Libertà, Pisa, 56127, Italy cInstitute of Information Science and Technologies, National Research Council of Italy, via G. Moruzzi, Pisa, 56124, Italy article i n f o Keywords: Animal behavior Deep learning Pose estimation Object detection Bio-acoustics Machine learning a b s t r a c t Animal behavior serves as a reliable indicator of the adaptation of organisms to their environment and their overall well-being. Through rigorous observation of animal actions and interactions, researchers and observers can glean valuable insights into diverse facets of their lives, encompassing health, social dynamics, ecological relationships, and neuroethological dimensions. Although state-of-the-art deep learning models have demonstrated remarkable accuracy in classifying various forms of animal data, their adoption in animal behavior studies remains limited. This survey article endeavors to comprehensively explore deep learning architectures and strategies applied to the identification of animal behavior, spanning auditory, visual, and audiovisual methodologies. The survey categorizes techniques into pose estimation-based and non-pose estimation-based methods, analyzing their applications, effectiveness, and limitations. Furthermore, the manuscript scrutinizes extant animal behavior datasets, offering a detailed examination of the principal challenges confronting this research domain. The article culminates in a comprehensive discussion of key research directions within deep learning that hold potential for advancing the field of animal behavior studies. 1. Introduction The study of animal behavior involves the observation, description, and comprehension of how animals engage with one another and their surroundings. Presently, the landscape of animal behavior research is undergoing rapid evolution, propelled by the continual introduction of innovative experimental methodologies and the advancement of sophisticated behavior detection systems (Wang, Du, Wang, Hu, & Zhao, 2021b). This progression holds particular significance in advancing our understanding of neuroethological aspects, exemplified by the utilization of mice in exploring diseases like Alzheimer’s (Pedersen et al., 2006), and in refining animal welfare practices within agriculture (Mishra & Sharma, 2023). The impetus behind this surge in progress is the integration of cutting-edge technologies, with deep learning standing out as a transformative force that reshapes the approaches researchers employ to investigate and interpret animal behaviors (Brown & de Bivort, 2017). Deep learning has emerged as a pivotal tool in the exploration of animal behavior. This advanced branch of artificial intelligence empowers computers to autonomously discern patterns and features from extensive datasets, improving upon earlier methods based on classical Machine ∗Corresponding author. E-mail addresses: [email protected], [email protected] (E. Fazzari), [email protected] (D. Romano), [email protected] (F. Falchi), [email protected] (C. Stefanini). Learning, as discussed in Valletta, Torney, Kings, Thornton, and Madden (2017). As researchers amass increasingly intricate datasets through state-of-the-art monitoring technologies, including high-resolution cameras, GPS tracking devices, and sensors, the capability of deep learning algorithms to extract meaningful insights becomes indispensable (Benaissa et al., 2023; Koger et al., 2023). This not only expedites the analysis process but also reveals nuanced aspects of animal behavior that were previously challenging to decipher. Furthermore, deep learning plays a crucial role in the development of sophisticated behavior detection systems. These systems can automatically recognize and classify various behaviors, allowing researchers to redirect their focus from laborious manual data annotation to the interpretation of results (Arablouei et al., 2023a). This acceleration in data processing and behavior recognition enhances the scalability and efficiency of animal behavior studies, ushering in a new era of discovery and understanding in this dynamic field. This survey article makes a threefold contribution to the current understanding of the study of animal behavior through deep learning: •We provide a thorough examination of existing technologies and algorithms employed in the analysis of animal behavior. This entails https://doi.org/10.1016/j.eswa.2025.128330 Received 11 December 2024; Received in revised form 30 April 2025; Accepted 23 May 2025 Expert Systems With Applications 289 (2025) 128330 Available online 26 May 2025 0957-4174/© 2025 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ ). E. Fazzari et al. a detailed exploration of methodologies and approaches currently prevalent in this domain, providing readers with a nuanced understanding of the technological landscape. •We compile and present a comprehensive list of publicly available datasets relevant to the research field. This compilation serves as a valuable resource for researchers and practitioners, facilitating access to essential data for furthering investigations into animal behaviors through data-driven methodologies. •We engage in a substantive discussion regarding potential directions for the evolution of the field. Emphasizing the integration of deep learning techniques, our discourse aims to enhance the quality of existing technologies, thereby advancing the understanding of animal behaviors. This forward-looking analysis provides insights into potential avenues for improvement and innovation. To the best of our knowledge, this survey is the only examination of the topic to date. The comprehensive overview, dataset compilation, and forward-looking discussions collectively contribute to a nuanced understanding of current advancements and lay the groundwork for future developments in the application of deep learning to the study of animal behavior. 2. Motivation and problem statement In this section, we expound upon the foundational motivations that underscore the study of animal behavior through deep learning, articulating the diverse advantages and objectives inherent to this specialized research domain. Concurrently, we meticulously scrutinize the principal limitations that characterize this field, recognizing the nuanced challenges that arise from variations in research setups. Our intention is to furnish a comprehensive guide for prospective researchers, endowing them with a thorough understanding of potential impediments prior to initiating investigations within this domain. Concurrently, we underscore the myriad opportunities inherent in the study of animal behavior, fostering an appreciation for the intricate dimensions of this research frontier. 2.1. Limitations in studying animal behaviors The exploration of animal behavior is confronted by a multitude of challenges that intricately shape the effectiveness and practicality of its applications. These challenges permeate the methodologies employed for data acquisition, the intricacies of data analysis (both in terms of location and computational demands), and the nuanced process of data annotation. An integral aspect of animal behavior studies is the utilization of sensors for data collection. However, the attachment of sensors to animals introduces a unique set of challenges. Notably, there is a risk of inducing stress responses and altering normal behaviors (Zhang, Li, Huang, & Chen, 2020). This necessitates a profound reflection on the authenticity of the collected data, urging researchers to question whether stress-induced behaviors accurately mirror natural patterns. Moreover, the task of differentiating genuine behavioral signals from background noise in sensor data adds complexity to interpretation, emphasizing the requirement for advanced algorithms capable of discerning meaningful patterns amidst the noise (Kavlak, Pastell, & Uimari, 2023). Beyond stress responses and noise challenges, sensor equipment grapples with limitations in battery life (Mekruksavanich, Jantawong, & Jitpattanakul, 2022), impacting the duration and scope of behavioral studies, especially in scenarios requiring continuous data collection over extended periods. Researchers are challenged to strike a balance between the need for comprehensive, continuous monitoring and the practical constraints imposed by limited battery capacities. Transitioning into the domain of mobile devices introduces another layer of complexity to the challenges encountered in deep learning applications. The implementation of models on these devices confronts the persistent issue of storage limitations (Cao, Zhao, Liu, & Sun, 2020). Balancing robust object detection with efficient data compression becomes a paramount concern, with techniques like Quantized-CNN (Wu, Leng, Wang, Hu, & Cheng, 2016) attempting to address this challenge. However, the ongoing quest is to achieve this balance without compromising precision, a crucial consideration for the reliability of behavioral analyses. Furthermore, light-weight processing enables deployment of trained models on-board, reducing reducing latency and bandwidth requirements in harsh farming environments (Temenos et al., 2024). A pivotal challenge arises in the realm of labeling and annotation, where economic and practical constraints hinder the tagging of large datasets for each animal (Bhattacharya & Shahnawaz, 2021). This bottleneck impedes the scalability of deep learning models, heavily reliant on labeled datasets for effective training. The impracticality of manual labeling raises fundamental concerns about the breadth and accuracy of behavioral datasets, impacting the reliability of subsequent analyses. Subjectivity compounds these challenges, influencing the accuracy and consistency of behavioral annotations. Visual inspection, often subjective, is limited in providing objective insights into complex animal behaviors (Bernardes, Lima, Guedes, da Silva, & Martins, 2021). Manual annotation, while traditional, is labor-intensive and susceptible to inter-annotator disagreements (Segalin et al., 2021; Zhou et al., 2022). The inherent subjectivity introduces variability, raising questions about the replicability and reliability of experiments (Dell et al., 2014; Hou et al., 2020). Moreover, these challenges extend to innovative techniques, such as multi-view recordings, which hold promise for providing richer insights into animal behaviors (Jiang et al., 2021). However, challenges arise in correlating social behaviors from different perspectives due to the lack of correspondence across data sources. Effectively coordinating information from multiple viewpoints demands inventive approaches to ensure accuracy and reliability, representing a frontier where deep learning methodologies can contribute significantly. A major challenge surfaces when comparing laboratory studies, where challenges often revolve around the subjectivity introduced by the controlled environment, with ethological studies conducted in the wild presenting a distinct set of challenges, notably in-field tracking (Marshall, Li, Wu, & Dunn, 2022). The diverse and unpredictable environments encountered in the wild introduce complexities not found in controlled laboratory settings. Bridging the gap between these two disciplines necessitates adaptable detection and tracking algorithms that seamlessly operate in both environments. Robust algorithms capable of handling varying animal sizes, changing appearances, clutter, occlusions, and unpredictable environments are vital for extracting meaningful insights (Haalck, Mangan, Webb, & Risse, 2020; Hou et al., 2020; Lauer et al., 2022). These challenges underscore the critical need for technological innovation that aligns with the demands of both controlled and wild settings. 2.2. Objectives in studying animal behaviors Studying animal behavior offers manifold advantages, enriching our comprehension of the natural world and presenting practical applications across diverse domains, including neuroscience, pharmacology, medicine, agriculture, ecology, and robotics. Six key advantages of studying animal behavior are identified: 1. Biodiversity conservation: Understanding animal behavior is crucial for the conservation of biodiversity (Ditria et al., 2020; Hou et al., 2020; Nilsson et al., 2020; Pillai, Gupta, Sharma, & Bansal, 2023; Wijeyakulasuriya, Eisenhauer, Shaby, & Hanks, 2020). Knowledge of behaviors such as migration patterns, feeding habits, and reproductive strategies is essential for designing effective conservation strategies and protecting endangered species. 2. Ecological understanding: Animal behavior provides insights into the ecological dynamics of ecosystems (Akçay et al., 2020; Gotanda, Expert Systems With Applications 289 (2025) 128330 2 E. Fazzari et al. Farine, Kratochwil, Laskowski, & Montiglio, 2019). Behavioral studies help researchers understand how animals interact with their environment, including their roles in nutrient cycling, seed dispersal, and predator-prey relationships (Chen et al., 2020; Nasiri, Amirivojdan, Zhao, & Gan, 2023; Yamada, Shawe-Taylor, & Fountas, 2020). 3. Human health and medicine: Studying animal behavior can have implications for human health and medicine (Gnanasekar, Yanushkevich, Van den Hoogen, & Trang, 2022; Hart, 2011; Manduca et al., 2023; Shaw & Lahrman, 2023). For example, research on animal models helps in understanding certain diseases and developing potential treatments. Behavioral studies on animals also contribute to our understanding of the neurobiology and psychology that underlie human behavior (Coria-Avila et al., 2022; Mathis & Mathis, 2020; Saleh, Ahmed, Zaafan, Farouk, & Atia, 2023). 4. Animal welfare and husbandry: Knowledge of animal behavior is essential for promoting the welfare of domesticated animals and optimizing their husbandry practices (Bao & Xie, 2022; Jiang et al., 2021). Understanding how animals express natural behaviors can inform the design of environments that support their physical and psychological well-being Jiang, Rao, Zhang, and Shen (2020); Manoharan (2020); Tassinari et al. (2021). 5. Pest control and agriculture: Although right now it is very limited to pest identification, understanding the behavior of pest species can aid in the development of effective pest control strategies in agriculture (Coulibaly, Kamsu-Foguem, Kamissoko, & Traore, 2022; Júnior & Rieder, 2020; Mendoza et al., 2023). This knowledge helps farmers manage crop damage and reduce the need for harmful pesticides (Mankin, Hagstrum, Guo, Eliopoulos, & Njoroge, 2021; Teixeira, Ribeiro, Morais, Sousa, & Cunha, 2023). 6. Understanding social dynamics: Observing social behaviors in animals can provide insights into the principles governing social structures and interactions (Papaspyros et al., 2023; Xiao et al., 2023). This knowledge can have applications in fields such as sociology and psychology, contributing to our understanding of social dynamics in general (Alameer et al., 2022; Landgraf et al., 2021; Perez & TolerFranklin, 2023). In this context, deep learning plays a pivotal role and emerges as a major technology in advancing the field, opening new opportunities. Addressing the limitations discussed in the previous section, we will now elaborate on the advantages that deep learning and computational technologies bring to the study of animal behavior. We previously emphasized the integral aspect of employing sensors for data collection in animal behavior studies, which may induce stress and high noise levels. The use of multiple and diverse sensors for data acquisition, coupled with advanced architectures incorporating fusion layers, has been shown to mitigate noise and enhance the precision of analysis (Mahmud, Zahid, Das, Muzammil, & Khan, 2021). However, attaching sensors directly to animals may introduce bias, prompting researchers in livestock health assessment and neuroscience to adopt computer vision. The ability of computer vision to provide real-time, noninvasive, and accurate animal-level information through the use of cameras has gained popularity (Oliveira, Pereira, Bresolin, Ferreira, & Dorea, 2021). Nevertheless, this approach is limited to setups within the camera frame, except for innovations like Haalck et al. (2020) moving camera that tracks animals, creating a dynamic map of their environment. In larger scenarios, such as meadows where cows graze, sensors remain preferable. Nevertheless, collecting and analyzing sensor data from mobile devices on animals proves challenging and time-consuming. To address this, Dang, Dang, Tran, and Chung (2022) introduced Long Range Area Network (LoRaWAN), where sensors attached to cows connect to gateways transmitting information to the cloud. This not only overcomes the limitations of computational power associated with mobile devices but also ensures continuous, real-time data analysis. The efficacy of deep learning is contingent on annotated data, especially for supervised approaches. While manual annotation remains Table 1 Summary of the limitations and objectives in studying animal behaviors. DL stands for Deep Learning. Limitations Objectives Sensor-induced stress Biodiversity conservation Battery life Biodiversity preservation Data noise Ecological insight Storage constraints Health impact Labeling economics Welfare optimization Subjective annotation Pest management Computational demands Social dynamics Ethical considerations Non-invasive (DL related) In-field tracking Real-time applications (DL related) Environmental unpredictability unavoidable, in tasks such as pose estimation and classification, labeling can be iterative. This involves annotating a small portion of the dataset, training a network, predicting on new images, correcting labels, and repeating this process multiple times. This iterative approach accelerates the labeling process, as demonstrated by Pereira et al. (2019). Another approach is to generate artificial labels (Li & Lee, 2023). Table 1 succinctly encapsulates a consolidated overview of both primary limitations and advantages, providing a discerning reference for researchers navigating the sophisticated landscape of animal behavior studies. 2.3. Research questions The survey aims to address the following research questions: RQ1 Which animal species are more considered and why? This research question seeks to investigate commonly explored and wellstudied animal species in the context of behavior analysis. To address this question, the survey will explore existing literature and research to identify trends and biases in the selection of animal subjects. The objective is to understand why certain species are more frequently examined and to gain insights into potential research trends, contributing to a deeper understanding of established knowledge and guiding future studies in the field. RQ2 What deep learning methods have been used in the literature for animal behavior analysis? This research question focuses on summarizing and categorizing the existing deep learning methods employed in the literature for animal behavior analysis. The survey will review a wide range of studies to identify and classify the various deep learning techniques applied to analyze animal behavior. The objective is to analyze existing methodologies to identify trends, strengths, and limitations of current approaches in the field. RQ3 What are the deep learning strategies that are suitable and could enhance this task, but are not yet exploited? This research question looks forward, aiming to identify untapped potential in the application of deep learning to animal behavior analysis. The survey will involve a comprehensive review of the current literature to identify gaps or areas where deep learning strategies have not been extensively explored. This involves proposing novel applications of existing techniques or suggesting modifications to adapt deep learning methods for more effective analysis of animal behavior. 3. Method for literature survey In this section, we clarify the methodology applied in our survey. Our approach involved a thorough systematic review to carefully select the pertinent studies considered in this article. Following this, we accurately analyzed the gathered information, employing statistical methods to derive meaningful insights. Expert Systems With Applications 289 (2025) 128330 3 E. Fazzari et al. Table 2 Features extracted from papers. Reference: Assigned identifier for each retrieved article. Year: Publication year of the article. Country: Geographical location where the authors are based, as required in Section 3.2. Species: The specific species under investigation in the article. Pose estimation: Indicates whether the methodology incorporates pose estimation (True or False). Behavior analysis: Indicates whether the analysis considers an association between extracted features and observed behaviors (True or False). Feature methodology: The approach employed to extract salient features in the study. Behavior methodology: The methodological framework used to correlate features with observed behaviors. Authors’ research field: Indicates the research fields the authors mainly work on. 3.1. Search and selection strategies This section delineates the methodologies employed for data collection and synthesis. Initially, data acquisition was conducted through systematic searches on academic repositories, including Google Scholar, IEEE Xplore, and the Springer Database. The formulated search queries were as follows: animal behavior AND deep learning (insect OR wild) AND behavior AND deep learning The decision to employ distinct queries for insects and wild animals was necessitated by the observed paucity of literature in these categories relative to studies involving farm animals and neuroethology, commonly focused on mice. Thus, the formulation of specific queries was imperative to encompass a broader spectrum of animal species. Notably, while our search explicitly included “behavior”, we ensured that both American and British spellings (“behavior” and “behaviour”) were effectively considered. Our observations confirmed that academic indexing systems retrieve papers regardless of spelling variations, mitigating the need for separate queries for each variant. Regarding deep learning, we opted to use “deep learning” as a general term rather than specifying particular architectures such as CNNs or RNNs. This choice was informed by the observation that papers discussing neural networks typically reference deep learning in some capacity. Indeed, several retrieved papers had titles mentioning CNNs or other architectures without explicitly stating “deep learning,” yet they addressed deep learning concepts in their introductions. Since referencing the broader field is common in academic writing, this approach ensured comprehensive coverage of relevant studies. Lastly, the inclusion of “animal” in our search strategy required careful consideration. As indicated by our need for a distinct query for insects, many papers on insect studies did not explicitly mention the term “animal”. To address this, we incorporated (insect OR wild) into our search queries. The term “wild” was particularly valuable, as it serves as the root of “wildlife”, a term frequently associated with key research themes such as wildlife conservation, biodiversity, and ecological studies. This inclusion allowed us to capture a broader range of studies focusing on animals in non-domesticated settings, ensuring a more comprehensive literature retrieval. Subsequently, the acquired data underwent systematic tabulation based on features explicated in Table 2. These features were derived from discerned patterns identified during a comprehensive analysis of the extant literature. Synthesizing the outcomes of this analysis, Sections 4 and 5 encapsulate the aggregated findings, summarizing the respective papers that expound upon solution methodologies grounded in pose estimation and those that do not. Finally, a judicious filtering operation was executed to extract only those articles germane to the objectives of this survey, resulting in 161 articles. Each article within this subset underwent thorough examination, and pertinent references therein were scrutinized and subsequently incorporated into our survey to enrich its content. The overall process is described via the PRIMA flowchart in Fig. 1. 3.2. Comprehensive science mapping analysis 3.2.1. Annual scientific production In the process of retrieving articles, our attention was exclusively directed towards research publications spanning the temporal spectrum from 2020 to 2023. Fig. 2(a) elucidates this distribution through a histogram, illustrating the quantitative representation of papers across each respective year. 3.2.2. Scientific production based on animal considered Fig. 2(b) illustrates the distribution of percentages pertaining to the various animal species under consideration in the selected articles. Evidently, a predominant emphasis is placed on research concerning livestock, notably focusing on cows and pigs, as well as studies involving mice, related mostly to neuroscience. 3.2.3. Research field of authors Given that animal behavior analysis is an interdisciplinary and multidisciplinary field, it becomes imperative to comprehend the research background of individuals engaged in this domain. Despite the predominant focus on articles related to deep learning technologies, it is noteworthy that a considerable number of non-artificial intelligence practitioners are actively entering this field, as illustrated in Fig. 2(c). Interestingly, when combining "computer science" (encompassing computer engineering) and "artificial intelligence," they constitute only 18% of the scholarly contributions. In contrast, bio-related fields, including biology, animal science, agriculture, veterinary, and ecology, collectively contribute 30 % to the research landscape. Noteworthy is the active participation of various engineering fields, even those with a mechanicalelectrical background, in the exploration of animal behavior. Additionally, a compelling correlation is observed in the fields of neuroscience and psychology, where the majority of articles are dedicated to the study of mice. 4. Pose estimation-based methods A potential first step in measuring behavior is to identify meaningful keypoints on the animal’s body to track the movements of specific body parts and quantify behavioral patterns. Pose estimation, the process of identifying and locating the position and orientation of objects, is a fundamental technique widely used in the examination of animal behaviors alongside object detection, as discussed in Section 5.3. Originating from Human Pose Estimation (HPE), the evolution into Animal Pose Estimation (APE) was spearheaded by Mathis et al. (2018) through DeepLabCut and Pereira et al. (2019) via LEAP, subsequently evolving into SLEAP (Pereira et al., 2022). This section delves into an in-depth analysis of these two methodologies juxtaposed with emerging trends within the field of research. Given the primary focus of our survey on animal behavior analysis, subsequent to the introduction of these predominant approaches, we elucidate the utilization of pose estimation outputs for behavior analysis and classification. For a more comprehensive understanding of animal pose estimation, we recommend perusing the survey conducted by Jiang, Lee, Teotia, and Ostadabbas (2022a). Expert Systems With Applications 289 (2025) 128330 4 E. Fazzari et al. Fig. 1. PRIMA flowchart regarding our literature search. Fig. 2. (a) illustrates a histogram depicting the distribution of research articles per year, focusing exclusively on papers obtained and cataloged during the initial scavenging phase. (b) presents a pie chart detailing the variety of animals utilized in behavioral studies leveraging deep learning techniques. (c) displays another pie chart showcasing the diverse research fields of the authors. Expert Systems With Applications 289 (2025) 128330 5 E. Fazzari et al. Fig. 3. (a) shows the architecture of LEAP Pereira et al. (2019); (b) the one exploited in T-LEAP Russello et al. (2022). LEAP (Pereira et al., 2019) is a single-animal pose estimation model employing convolutional layers culminating in confidence maps that delineate the probability distribution for each distinct body keypoint. This architectural design, depicted in Fig. 3, is characterized by its simplicity, featuring three sets of convolutional layers. The initial two sets are terminated by max pooling to alleviate computational complexity. Subsequently, transposed convolution is applied to restore the original dimensions of the images, yet with a depth corresponding to the number of keypoints, thereby generating a confidence map for each. Despite its simplicity, the LEAP model encounters challenges in non-laboratory settings due to issues such as occlusion, prompting the introduction of T-LEAP (Russello, van der Tol, & Kootstra, 2022). T-LEAP preserves the architecture of LEAP but diverges in its use of 3D convolution instead of 2D convolution. The input to T-LEAP comprises four consecutive frames extracted from videos, enhancing the model’s robustness. Notably, T-LEAP maintains a focus on single-animal pose estimation, as elucidated in Fig. 3. Subsequently, the author of LEAP introduced a refined version known as Social LEAP (SLEAP) (Pereira et al., 2022), designed to proficiently address the challenges associated with multianimal pose estimation through the integration of both bottom-up (Papandreou et al., 2018) and top-down strategies (Nguyen & Kresovic, 2022). In the top-down strategy, SLEAP first identifies individuals and subsequently detects their respective body parts. Unlike LEAP, SLEAP seamlessly incorporates this approach without the need for an additional object detection architecture. On the other hand, the bottom-up strategy in SLEAP involves detecting individual body parts and subsequently grouping them into individuals based on their connectivity. A key advantage of this dual-strategy framework is its efficiency, requiring only a single pass through the neural network. The output of this strategy produces multi-part confidence maps and part affinity fields (PAFs) (Cao, Simon, Wei, & Sheikh, 2017), constituting vector fields that intricately represent spatial relationships between pairs of body parts. Additionally, SLEAP undergoes a structural enhancement by transitioning from LEAP’s backbone to a more intricate U-Net architecture (Ronneberger, Fischer, & Brox, 2015), thereby significantly improving accuracy in the realm of multi-animal pose estimation scenarios. Similarly, DeepLabCut (DLC) has evolved significantly over time. Initially designed as a single-animal pose estimation method, it utilizes a pretrained ResNet-50 (He, Zhang, Ren, & Sun, 2016) backbone with subsequent deconvolutional layers to generate confidence maps for keypoints. This approach, taking advantage of Imagenet pretrained weights, allowed DLC to effectively estimate skeletons with minimal data. The model’s capabilities were later expanded to include 3D pose estimation through the use of multiple cameras (Nath et al., 2019). Each camera view was trained independently, and sophisticated camera calibration techniques were employed to derive 3D locations. A subsequent milestone in DLC’s development involved addressing the challenges of multianimal pose estimation (Lauer et al., 2022). This evolution introduced DLCRNet, a structural modification that replaced the ResNet backbone. DLCRNet employs a bottom-up multi-animal pose estimation approach, featuring a multi-fusion architecture and a multi-stage decoder. The decoder utilizes multiple stages of score maps and PAFs (Cao et al., 2017) to predict keypoints for each animal. Further innovation is exemplified by SuperAnimal (Ye et al., 2022), which introduced transformer layers into the model architecture. While DLC and SLEAP currently stand as the predominant pose estimation methodologies in behavior analysis for animal behavior classification, it is imperative to acknowledge recent advancements in animal pose estimation architectures. Several notable methodologies have been introduced: •OptiFlex (Liu et al., 2020b) is a video-based animal pose estimation method that, given a skip ratio 𝑠 and a frame range 𝑓, assembles a sequence of 2𝑓+ 1 images with indices ranging from 𝑡−𝑠×𝑓 to 𝑡+𝑠×𝑓. This sequence is input to a model based on residual blocks with intermediate supervision, generating predictions for each image and producing a sequence of heatmap tensors. These tensors are then fed into an OpticalFlow model, ultimately yielding the final heatmap prediction for index 𝑡. OptiFlex has demonstrated superior accuracy compared to DeepLabCut, LEAP, and DeepPoseKit (Graving et al., 2019). •SemiMultiPose (Blau, Gebhardt, Bendesky, Paninski, & Wu, 2022) introduces a semi-supervised multi-animal pose estimation approach, building upon DeepGraphPose (Wu et al., 2020) and DirectPose (Tian, Chen, & Shen, 2019). Taking both labeled and unlabeled frames as input, the method processes them using a ResNet backbone, generating a compact representation fed into three branches: one for detecting keypoint heatmaps (B1), one for bounding box heatmaps (B2), and a third for keypoint detection (B3). SemiMultiPose aims to generate pseudo keypoint coordinates from B2 and B3 for the self-supervised branch, contributing to B1. The network has shown improved accuracy compared to SLEAP. However, the authors note that in cases of abundant labeled data, their method may not significantly outperform others, and for single-animal pose estimation with unlabeled frames from a sequential video, DeepGraphPose might outperform SemiMultiPose, benefiting from the consideration of spatial and temporal information. •Lightning pose (Biderman et al., 2023) exploits spatiotemporal statistics of unlabeled videos in two ways. Firstly, it introduces unsupervised training objectives penalizing the network for predictions violating the smoothness of physical motion, multiple-view geometry, or departing from a low-dimensional subspace of plausible body configurations. Secondly, it proposes a novel network architecture predicting poses for a given frame using temporal context from surrounding unlabeled frames. The resulting pose estimation networks exhibit superior performance with fewer labels, generalize effectively to unseen videos, and provide smoother and more reliable pose trajectories for downstream analysis (e.g., neural decoding analyses) compared to previously mentioned approaches. Expert Systems With Applications 289 (2025) 128330 6 E. Fazzari et al. •Bhattacharya and Shahnawaz (2021) introduced a novel model for recognizing the pose of multiple animals from unlabeled data. The approach involves the removal of background information from each image and the application of an edge detection algorithm to the body of the animal. Subsequently, the motion of the edge pixels is tracked, and agglomerative clustering is performed to segment body parts. In a departure from previous methods, the end result is not specific keypoints but rather the segmentation of body parts. To achieve this, the authors utilized contrastive learning to discourage the grouping of distant body parts together. After obtaining the skeletal representation of each animal in every frame, whether from videos or images, the subsequent step involves processing the data to discern specific behaviors. The trajectories derived from pose estimation can be effectively analyzed through statistical methods. Weber, Mulders, Kaiser, Tackenberg, and Rust (2022) utilized DeepLabCut predictions (Mathis et al., 2018) and ANOVA (Kaufmann & Schering, 2014) to conduct behavioral profiling of rodents, with a focus on studying stroke recovery. In a similar vein, Lee et al. (2021), employing DLC, investigated the behavior of non-tethered fruit flies. Their study involved predicting locomotion patterns and identifying the centroid of the animals’ legs. Machine learning for analyzing pose estimation trajectories becomes crucial when classifying postures and relating them to specific behaviors. One of the simplest approaches is to use a Nearest-Neighbor classifier. Saleh et al. (2023) tested this method to classify mouse behaviors such as crossing and rearing in an open-field experiment, achieving 97% accuracy. Other machine learning approaches were employed by Fang, Zhang, Zheng, Huang, and Cuan (2021) and Nilsson et al. (2020). The former used a naive Bayesian classifier to identify eating, preening, resting, walking, standing, and running behaviors for poultry analysis, providing a disease warning system. The latter introduced the SimBa toolkit, importing DeepLabCut or DeepPoseKit projects to create classifiers using RandomForest Breiman (2001) and extracting features like velocities and total movements. Another application of Random Forest was employed by Higaki et al. (2024) for classifying on a scoring system basis dairy cows mobility. McKenzie-Smith, Wolf, Ayroles, and Shaevitz (2023) used trajectories obtained with SLEAP to identify stereotyped behaviors such as grooming, proboscis extension, and locomotion in Drosophila melanogaster, using resulting ethograms provided by MotionMapper (Berman, Choi, Bialek, & Shaevitz, 2014) to explore how flies’ behavior varies across time of day and days, finding distinct circadian patterns in all stereotyped behaviors. Other authors opted for recurrent and convolutional neural networks, with simple approaches such as using Long Short-Term Memory (LSTM) (Hochreiter & Schmidhuber, 1997) and 1D convolutional neural networks to process trajectories for drawing behavioral conclusions. Examples include detecting lameness in horses (Alagele & Yildirim, 2022) and determining chemical interactions experienced by crickets (Fazzari, Carrara, Falchi, Stefanini, & Romano, 2024). More complex approaches include Wittek, Wittek, Keibel, and Güntürkün (2023)’s use of InceptionTime (Ismail Fawaz et al., 2020), an ensemble of deep convolutional neural network models, to classify seven distinct behaviors in birds. Some authors simplified the classification process by introducing a nonlinear clustering phase to improve the feature space, followed by classification using Multilayer Perceptrons (MLP), demonstrating advantages in classification (Schneider, Lee, & Mathis, 2023; Ye et al., 2022). A recent emerging trend involves the utilization of unsupervised learning techniques in the analysis of animal behavior. Luxem et al. (2022) have innovatively proposed a methodology for processing trajectories derived from DeepLabCut by employing a Variational AutoEncoder (VAE) (Kingma & Welling, 2013). Subsequently, they apply a Hidden Markov Model (HMM) (Rabiner & Juang, 1986) to the new representation of trajectories to discern underlying motifs. Following a comprehensive analysis of motif usage, the authors iteratively employ HMM, limiting the number of motifs to those surpassing a 1 % usage threshold in the previous analysis. The refined motifs were attributed to specific behavior exhibited by the mice, such as exploration, rearing, grooming, pausing, or walking. Notably, this methodological approach outperforms conventional techniques, such as Auto-Regressive HMM (AR-HMM) or MotionMapper (Berman et al., 2014), when applied directly to the motion sequences. Motion trajectories extend their utility beyond predicting the behavior of individual animals; in multi-animal scenarios, they can also be applied to unravel the intricate web of social interactions among them. Segalin et al. (2021) introduced the Mouse Action Recognition System (MARS), a sophisticated automated pipeline tailored for pose estimation and behavior quantification in pairs of freely interacting mice. MARS adeptly discerns three specific social behaviors: close investigation, mounting, and attack. On a different note, Zhou et al. (2022) proposed the Cross-Skeleton Interaction Graph Aggregation Network (CSIGANet), a groundbreaking framework designed to capture the diverse dynamics of freely interacting mice. CS-IGANet successfully identifies a spectrum of behaviors, including approaching, attacking, chasing, copulation, walking away from another mouse, sniffing, and many others. Trajectories not only serve as a means to identify specific behaviors but are also instrumental in anomaly detection. For instance, Fujimori, Ishikawa, and Watanabe (2020) employed OneClassSVM (Boser, Guyon, & Vapnik, 1992) and IsolationForest (Liu, Ting, & Zhou, 2008) to detect outlier behaviors in domestic cats. Similarly, Gnanasekar et al. (2022) utilized pose estimation data to predict abnormal behavior in mice undergoing opioid withdrawal, employing pretrained convolutional neural networks for the classification of shaking behaviors. For a comprehensive summary of the various methods discussed in this section, please refer to Table 3, which outlines each method along with its advantages, limitations, and potential applications. 5. Non pose estimation-based methods In this section, we expound upon methodologies employed in the investigation of animal behaviors without recourse to pose estimation techniques. To enhance clarity and systematic presentation, we have delineated subsections corresponding to each methodology. 5.1. Sensor based approaches Sensor-generated data, typically originating from accelerometers or gyroscopes, has been extensively explored in the literature, as comprehensively in Kleanthous et al. (2022b); Neethirajan (2020) surveys. These surveys delve into the application of classical machine learning methods in modern animal farming and the study of animal behavior. More recently, a shift towards leveraging deep learning approaches has been observed. Arablouei et al. (2023a) utilized a wearable collar tag equipped with an accelerometer to collect data from grazing beef cattle. They applied a Multi-Layer Perceptron to classify behaviors such as grazing, walking, ruminating, resting, and drinking. Similarly, Eerdekens et al. (2020) employed tri-axial accelerometers on horses, strategically positioned at the two front legs’ lateral side. They proposed a Convolutional Neural Network to detect behaviors like standing, walking, trotting, cantering, rolling, pawing, and flank-watching based on the acquired data. Mekruksavanich et al. (2022), instead, segmented accelerometer data into 2-second windows and exploited a pre-trained ResNet model to perform sheep activity recognition. Dang et al. (2022) introduced the integration of multiple sensors, collecting environmental data (e.g., temperature, humidity) alongside cow behavior information obtained from accelerometers and gyroscopes. They preprocessed this information using a 1D-convolutional neural network and LSTM networks for classifying walking, feeding, lying, and standing. In a recent study, Pan, Chen, Zhong, Wang, and Zheng (2023) introduced four novel Convolutional Neural Network architectures tailored for Animal Action Recognition (AAR). These architectures, namely one-channel temporal Expert Systems With Applications 289 (2025) 128330 7 E. Fazzari et al. Table 3 Pose estimation techniques and associated processing methods applied to the keypoint location data obtained from the mentioned pose estimation model. Method Advantages Limitations Applications LEAP (Pereira et al., 2019) Simple and efficient model Single-animal pose estimation, problems in non-laboratory settings due to occlusions Single-animal keypoint estimation T-LEAP (Russello et al., 2022) Considers previous and subsequent frames for improving keypoint estimation Single-animal pose estimation, requires images from a video to work Single-animal keypoint estimation SLEAP (Pereira et al., 2022) Multi-animal pose estimation, both bottom-up and top-down approaches No significant limitations Single/Multi-animal keypoint estimation DeepLabCut (Mathis et al., 2018) Multi-animal pose estimation (DLCRNet), strong community, 3D location capabilities, various backbones (including Transformers, SuperAnimal) Limited to bottom-up multi-animal pose estimation Single/Multi-animal keypoint estimation OptiFlex (Liu et al., 2020b) Integrates optical flow to improve performance Requires images from a video to create optical flow Single-animal keypoint estimation SemiMultiPose (Blau et al., 2022) Uses semi-supervised learning In cases of abundant labeled data, may not significantly outperform other single-animal pose estimation methods Single-animal keypoint estimation Lightning Pose (Biderman et al., 2023) Uses spatiotemporal statistics to improve results, useful for unlabeled videos Limited to videos Single-animal keypoint estimation Bhattacharya and Shahnawaz (2021) Works with unlabeled data Segments body parts, does not individuate specific keypoints Body-part segmentation SBeA (Han et al., 2024) Requires fewer annotated samples for multi-animal pose estimation compared to DLC and SLEAP, enables 3D pose reconstruction Requires a heavy data augmentation process Single/Multi-animal keypoint estimation ANORA Robust for detecting unusual movements in animal behaviors May struggle with highly variable behaviors Behavioral profiling Nearest-Neighbor Classifier Simple, interpretable, works well with small datasets Computationally expensive for large datasets, sensitive to noise Animal behavior classification Naive Bayesian Classifier Fast, interpretable, works well with limited data Assumes feature independence, limiting accuracy Animal behavior classification (with probabilistic modeling) Random Forest Explainable May not be optimal compared to deep learning approaches when the number of samples is huge Feature extraction, classification LSTM Good for sequential data, captures temporal dependencies Requires large datasets, computationally expensive Animal behavior classification 1D Convolutional Neural Network Efficient for time-series pose data Less effective than LSTMs for long-range dependencies Animal behavior classification InceptionTime (Ismail Fawaz et al., 2020) Proven to be better than LSTM and 1D CNN Requires large training data and is more complex than LSTM and 1D CNN, since it uses ensembling and requires non-linear clustering and MLP training Animal behavior classification VAME (Luxem et al., 2022) Researchers do not need to state how many behaviors are present a priori Requires careful tuning and probabilistic interpretation Unsupervised behavior discovery MARS (Segalin et al., 2021) Extracts precise movement features (speed, acceleration, etc.), useful for unsupervised analysis Limited to predefined movement features Quantitative analysis of animal motion CS-IGANet (Zhou et al., 2022) Captures social behaviors between two mice Very complex architecture, requiring a lot of training data. Proven to work only on mice Identification of social behaviors OneClassSVM (Boser et al., 1992) Good for outlier detection in pose-based behaviors Sensitive to hyperparameters, may not generalize well Detection of abnormal movements IsolationForest (Liu et al., 2008) Efficient anomaly detection May not capture complex sequential dependencies Identification of unusual movement patterns (OCT), one-channel spatial (OCS), OCT and spatial (OCTS), and twochannel temporal and spatial (TCTS) networks, leverage data from 3D accelerometers and 3D gyroscopes. The core objective of their research was to scrupulously identify behaviors such as movement, drinking, eating, nursing, sleeping, and lying in lactating sows. More advanced techniques were employed by Otsuka et al. (2024) to classify wild animal behaviors using animal-borne accelerometers, leveraging Transformer architectures. Their attention mechanism allows for a more effective learning of temporal dependencies, enhancing classification performance. In addition to accelerometer and gyroscope data, GNSS (Global Navigation Satellite System) data emerges as a valuable tool for understanding animal behavior. Arablouei, Wang, Bishop-Hurley, and Liu (2023b) explored this avenue by employing GNSS to extract pertinent information about cattle behavior, including metrics like distance from water points, median speed, and median estimated horizontal position error. Integrating this GNSS data with accelerometry information, the researchers pursued two distinct approaches. The first involved concatenating features from both sensor datasets into a comprehensive feature vector, subsequently fed into a MLP classifier. Alternatively, the second approach centered on fusing the posterior probabilities predicted by two separate MLP classifiers. These methodologies enabled the accurate detection of behaviors such as grazing, walking, resting, and drinking. In conclusion, sensor-only data-based approaches are limited to unior multi-dimensional sequence analyses, depending on the type of sensor used. This makes such data comparable to pose estimation methods. In fact, the techniques discussed here share similarities with those used for processing pose estimation data, such as LSTMs, 1D CNNs, and Transformers. However, unlike pose estimation, which relies on capturing images or videos–a nearly cost-free process today due to the widespread availability of high-resolution smartphone cameras–sensorbased approaches introduce an additional expense related to sensor acquisition. Conducting an experiment requires purchasing multiple sensors, and if multiple animals need to be monitored simultaneously, a Expert Systems With Applications 289 (2025) 128330 8 E. Fazzari et al. separate set of sensors must be obtained for each animal, further increasing costs. The emergence of low-cost miniature sensors has helped mitigate these expenses (Jin et al., 2024), and researchers have explored methods to minimize the number of sensors needed by quantifying each sensor’s contribution to accurate behavior identification. Addressing this concern, Li, Yang, Su, Li, and Wang (2023) introduced a novel method aimed at optimizing sensor selection. Their approach involves assessing the self-information brought by the jth sensor concerning the occurrence of a specific activity 𝐴𝑖 and multiplying it by the universality of the same sensor j during instances of the same activity 𝐴𝑖. This innovative strategy has proven highly effective, leading to improved recognition rates and reduced computational time by eliminating redundant and noisy data, thereby lowering overall costs. 5.2. Bioacoustics While bioacoustics offers a captivating glimpse into animal behavior (Stowell, 2022) and their ecosystem (Oestreich, Oliver, Chapman, Go, & McKenna, 2024), given the integral role of sound in animal activities such as communication, mating, navigation, and territorial defense (Chalmers, Fergus, Wich, & Longmore, 2021), the current landscape of published articles predominantly emphasizes animal identification (Bravo Sanchez, Hossain, English, & Moore, 2021; Varma, Bateshwar, Rathi, & Singh, 2021; Xu, Zhang, Yao, Xue, & Wei, 2020) and sound event detection (Moummad, Serizel, & Farrugia, 2023; Nolasco et al., 2023). Notably, the existing literature reveals a scarcity of research endeavors combining acoustics and deep learning for the identification of animal behaviors. Wang, Wu, Cui, Xuan, and Su (2021a) stand out as pioneers in this domain, as they endeavored to classify sheep behaviors, including chewing, biting, chewing-biting, and ruminating sounds. This was accomplished using a recording device positioned proximal to the animal’s face, with a placebo class designated as noise. The acquired wavelet data were leveraged for classification tasks through both a feed-forward neural network and a recurrent neural network. Additionally, the information was further processed by transforming it into a log-scaled Mel-spectrogram, serving as input for a convolutional neural network. The findings underscore that while the recurrent neural network exhibited superior performance, the convolutional neural network outperformed the feed-forward approach, attributing its success to the enhanced signal representation offered by the Mel-spectrogram. CNNs have gained increasing attention in animal tasks, particularly due to the benefits of transfer learning, which enhances their performance on new datasets. Notable examples include the work of Manriquez P, Kotz, Ravignani, and De Boer (2024), who used CNNs to classify mammals vocalizations, and Schall, Kaya, Debusschere, Devos, and Parcerisas (2024), who applied them to detect baleen whales. Therefore, we believe that transfer learning could also improve behavior identification through sound, making it a valuable and effective approach to consider. 5.3. Object detection In conjunction with pose estimation techniques, object detection stands out as a widely employed deep learning methodology for analyzing animal behavior. Its prevalence may be attributed to its established utility in animal recognition and detection (Banerjee, Khan, & Sharma, 2023; Chen, Zhu, & Norton, 2021; Teixeira et al., 2023), prompting researchers to redirect their focus toward studying animal welfare and activity. Among the leading architectures for animal behavior identification, Faster R-CNN (Ren, He, Girshick, & Sun, 2015) and particularly YOLO (Redmon, Divvala, Girshick, & Farhadi, 2016) are frequently employed. Faster R-CNN follows a two-stage approach, first generating region proposals and then refining predictions, which provides high detection accuracy but comes at the cost of increased computational complexity, making it less suitable for real-time applications. In contrast, YOLO adopts a single-stage approach, directly predicting object classes and bounding boxes in one pass through the network. This design significantly enhances processing speed, making YOLO more favorable for real-time animal behavior analysis, albeit sometimes at the expense of detection precision. Alternative architectures have also been proposed. For instance, Samsudin, Harizan, Ibrahim, Karim, and Ibrahim (2022) utilized SSD MobileNetv2 (Sandler, Howard, Zhu, Zhmoginov, & Chen, 2018), a lightweight and efficient model, to detect abnormal and normal zebrafish larvae behaviors for examining the effects of neurotoxins. However, SSD MobileNetv2 typically trades off some accuracy for efficiency. To address spatiotemporal dependencies in behavioral analysis, McIntosh, Marques, Albu, Rountree, and De Leo (2020) introduced TempNet, which incorporates an encoder bridge and residual blocks with a two-stage spatial-temporal encoder. This architecture enhances the detection of dynamic behaviors, such as startle responses in fish, by capturing motion patterns more effectively than frame-based detectors like YOLO and Faster R-CNN. Object detection serves a dual role, encompassing instantaneous behavior detection through image or video frame analysis, as well as the quantification and tracking of specific behaviors. The accurate analysis of single frames, counting, and frame-by-frame examination enable researchers to quantify both the duration and frequency of distinct actions. For instance, the application of YOLO in the study by Alameer et al. (2022) facilitated the quantification of contact frequency among pigs, allowing the identification of peculiar behaviors such as rear snorting and tail-biting. In the context of cows and pigs, a crucial aspect involves quantifying movement and aggressive behavior (Alameer et al., 2022; Odo, Muns, Boyle, & Kyriazakis, 2023). Furthermore, efforts to discern rank relationships based on fighting behavior in animals like cows are of crucial importance (Uchino & Ohwada, 2021). Importantly, for tasks demanding prolonged animal identification, tracking is conventionally executed using DeepSort (Evangelista, Concepcion, Palconit, Bandala, & Dadios, 2022; Wojke, Bewley, & Paulus, 2017). Efficient instant detection can be accomplished by conducting a single analysis on the animal and directly classifying its behavior through a single image. In this context, deep learning object detection models prove instrumental in directly identifying behaviors such as positional activities (e.g., mating, standing, feeding, spreading, fighting, drinking) for the comprehensive analysis of animal health and stress behaviors (Manoharan, 2020; Riekert, Klein, Adrion, Hoffmann, & Gallmann, 2020; Wang, Wang, Li, & Ren, 2020). These models also find application in disease identification, such as the detection of wryneck (Elbarrany, Mohialdin, & Atia, 2023), and in studying behavioral adaptations to new environments (Li et al., 2019a). Furthermore, object detection models can be extended to operate with thermal and infrared images, which are getting more popular (Korelidou, Simitzis, Massouras, & Gelasakis, 2024). For example, Xudong, Xi, Ningning, and Gang (2020) utilized thermal images for the automatic recognition of dairy cow mastitis, introducing the EFMYOLOv3 model. Similarly, Lei et al. (2022) employed infrared images to discern feeding, resting, moving, and socializing behaviors in slow animals. Beyond these applications, notable approaches utilizing object detection include Fuentes, Yoon, Park, and Park (2020), who integrated YOLO and Optical Flow to detect actions in cows. Additionally, some researchers employ object detection solely for localizing the animal within the image or video. They subsequently crop that region and use it in other models, leveraging 2D pretrained networks or introducing 3D convolutional neural networks for video analysis (Feighelstein et al., 2023; Thanh & Netramai, 2022). 5.4. Others Several research endeavors have employed unique deep learning methodologies, distinct from those discussed in the preceding section. Due to the relative scarcity of deep learning strategies for evaluating Expert Systems With Applications 289 (2025) 128330 9 E. Fazzari et al. Fang, C., Zhang, T., Zheng, H., Huang, J., & Cuan, K. (2021). Pose estimation and behavior classification of broiler chickens based on deep neural networks. Computers and Electronics in Agriculture, 180, 105863. Fazzari, E., Carrara, F., Falchi, F., Stefanini, C., & Romano, D., et al. (2023). A workflow for developing biohybrid intelligent sensing systems. In Proceedings of the italia intelligenza artificiale - thematic workshops co-located with the 3rd CINI national lab AIIS conference on artificial intelligence (ital IA 2023) (p. 555–560). CEUR Workshop Proceedings. Fazzari, E., Carrara, F., Falchi, F., Stefanini, C., & Romano, D. (2024). Using ai to decode the behavioral responses of an insect to chemical stimuli: Towards machine-animal computational technologies. International Journal of Machine Learning and Cybernetics, 15, 1985–1994. Fazzari, E., Romano, D., Falchi, F., & Stefanini, C. (2025). Selective state models are what you need for animal action recognition. Ecological Informatics, 85, 102955. Feichtenhofer, C., Fan, H., Malik, J., & He, K. (2019). Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision (pp. 6202–6211). Feighelstein, M., Ehrlich, Y., Naftaly, L., Alpin, M., Nadir, S., Shimshoni, I., Pinho, R. H., Luna, S. P., & Zamansky, A. (2023). Deep learning for video-based automated pain recognition in rabbits. Scientific Reports, 13, 14679. Fuentes, A., Yoon, S., Park, J., & Park, D. S. (2020). Deep learning-based hierarchical cattle behavior recognition with spatio-temporal information. Computers and Electronics in Agriculture, 177, 105627. Fujimori, S., Ishikawa, T., & Watanabe, H. (2020). Animal behavior classification using deeplabcut. IEEE. In 2020 IEEE 9th global conference on consumer electronics (GCCE) , pp. 254–257. Gnanasekar, S. T., Yanushkevich, S., Van den Hoogen, N. J., & Trang, T. (2022). Rodent tracking and abnormal behavior classification in live video using deep neural networks. IEEE. In 2022IEEE symposium series on computational intelligence (SSCI), pp. 830–837 Gore, S. V., Kakodkar, R., Del Rosario Hernández, T., Edmister, S. T., & Creton, R. (2023). Zebrafish larvae position tracker (z-lap tracker): A high-throughput deep-learning behavioral approach for the identification of calcineurin pathway-modulating drugs using zebrafish larvae. Scientific Reports, 13, 3174. Gotanda, K. M., Farine, D. R., Kratochwil, C. F., Laskowski, K. L., & Montiglio, P. O. (2019). Animal behavior facilitates eco-evolutionary dynamics. arXiv preprint arXiv:1912.09505. Graving, J. M., Chae, D., Naik, H., Li, L., Koger, B., Costelloe, B. R., & Couzin, I. D. (2019). Deepposekit, a software toolkit for fast and robust animal pose estimation using deep learning. ELife, 8, e47994. Gu, A., & Dao, T. (2023). Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752. Haalck, L., Mangan, M., Webb, B., & Risse, B. (2020). Towards image-based animal tracking in natural environments using a freely moving camera. Journal of Neuroscience Methods, 330, 108455. Han, F., Zhu, J., Liu, B., Zhang, B., & Xie, F. (2020). Fish shoals behavior detection based on convolutional neural network and spatiotemporal information. IEEE Access, 8, 126907–126926. Han, Y., Chen, K., Wang, Y., Liu, W., Wang, Z., Wang, X., Han, C., Liao, J., Huang, K., & Cai, S. (2024). Multi-animal 3d social pose estimation, identification and behaviour embedding with a few-shot learning framework. Nature Machine Intelligence, 6(1), 48–61. Nature Publishing Group UK London Hara, K., Kataoka, H., & Satoh, Y. (2018). Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 6546–6555). Hart, B. L. (2011). Behavioural defences in animals against pathogens and parasites: Parallels with the pillars of medicine in humans. Philosophical Transactions of the Royal Society B: Biological Sciences, 366, 3406–3417. He, K., Gkioxari, G., Dollár, P., & Girshick, R. (2017). Mask r-cnn. In Proceedings of the IEEE international conference on computer vision (pp. 2961–2969). He, K., Zhang, X., Ren, S., & Sun, J. (2016). Identity mappings in deep residual networks. Springer. In Computer vision–ECCV 2016: 14th European conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part IV 14, pp. 630–645. Higaki, S., Matsui, Y., Miwa, M., Yamamura, T., Hojo, T., Yoshioka, K., Vang, A., Negreiro, A., & Dórea, J. R. R. (2024). Leveraging computer vision-based pose estimation technique in dairy cows for objective mobility analysis and scoring system. Computers and Electronics in Agriculture, 217, 108573. Elsevier Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9, 1735–1780. Hou, J., He, Y., Yang, H., Connor, T., Gao, J., Wang, Y., Zeng, Y., Zhang, J., Huang, J., & Zheng, B., et al. (2020). Identification of animal individuals using deep learning: A case study of giant panda. Biological Conservation, 242, 108414. Ismail Fawaz, H., Lucas, B., Forestier, G., Pelletier, C., Schmidt, D. F., Weber, J., Webb, G. I., Idoumghar, L., Muller, P. A., & Petitjean, F. (2020). Inceptiontime: Finding alexnet for time series classification. Data Mining and Knowledge Discovery, 34, 1936–1962. Jaafra, Y., Laurent, J. L., Deruyver, A., & Naceur, M. S. (2019). Reinforcement learning for neural architecture search: A review. Image and Vision Computing, 89, 57–66. Elsevier Jia, Y., Li, S., Guo, X., Lei, B., Hu, J., Xu, X. H., & Zhang, W. (2022). Selfee, self-supervised features extraction of animal behaviors. ELife, 11, e76218. Jiang, L., Lee, C., Teotia, D., & Ostadabbas, S. (2022a). Animal pose estimation: A closer look at the state-of-the-art, existing gaps and opportunities. Computer Vision and Image Understanding, (p. 103483). Jiang, M., Rao, Y., Zhang, J., & Shen, Y. (2020). Automatic behavior recognition of group-housed goats using deep learning. Computers and Electronics in Agriculture, 177, 105706. Jiang, Z., Liu, Z., Chen, L., Tong, L., Zhang, X., Lan, X., Crookes, D., Yang, M. H., & Zhou, H. (2022b). Detecting and tracking of multiple mice using part proposal networks. IEEE Transactions on Neural Networks and Learning Systems. Jiang, Z., Zhou, F., Zhao, A., Li, X., Li, L., Tao, D., Li, X., & Zhou, H. (2021). Multi-view mouse social behaviour recognition with deep graphic model. IEEE Transactions on Image Processing, 30, 5490–5504. Jin, Z., Shu, H., Hu, T., Jiang, C., Yan, R., Qi, J., Wang, W., & Guo, L. (2024). Behavior classification and spatiotemporal analysis of grazing sheep using deep learning. Computers and Electronics in Agriculture, 220, 108894. Elsevier Joo, K. H., Duan, S., Weimer, S. L., & Teli, M. N. (2022). Birds’ eye view: Measuring behavior and posture of chickens as a metric for their well-being. arXiv preprint arXiv:2205.00069. Júnior, T. D. C., & Rieder, R. (2020). Automatic identification of insects from digital images: A survey. Computers and Electronics in Agriculture, 178, 105784. Kamminga, J. W., Le, D. V., Meijers, J. P., Bisby, H., Meratnia, N., & Havinga, P. J. (2018). Robust sensor-orientation-independent feature selection for animal activity recognition on collar tags. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 2, 1–27. Kamminga, J. W., Meratnia, N., & Havinga, P. J. (2019). Dataset: Horse movement data and analysis of its potential for activity recognition. In Proceedings of the 2nd workshop on data acquisition to analysis (pp. 22–25). Kasnesis, P., Doulgerakis, V., Uzunidis, D., Kogias, D. G., Funcia, S. I., González, M. B., Giannousis, C., & Patrikakis, C. Z. (2022). Deep learning empowered wearable-based behavior recognition for search and rescue dogs. Sensors, 22, 993. Kaufmann, J., & Schering, A. (2014). Analysis of variance ANOVA. John Wiley & Sons, Ltd. https://doi.org/10.1002/9781118445112.stat06938, Kavlak, A., Pastell, M., & Uimari, P. (2023). Disease detection in pigs based on feeding behaviour traits using machine learning. Biosystems Engineering, 226, 132–143. Kingma, D. P., & Welling, M. (2013). Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114. Kleanthous, N., Hussain, A., Khan, W., Sneddon, J., & Liatsis, P. (2022a). Deep transfer learning in sheep activity recognition using accelerometer data. Expert Systems with Applications, 207, 117925. Kleanthous, N., Hussain, A. J., Khan, W., Sneddon, J., Al-Shamma’a, A., & Liatsis, P. (2022b). A survey of machine learning approaches in animal behaviour. Neurocomputing, 491, 442–463. Knight, E. C., & Bayne, E. M. (2019). Classification threshold and training data affect the quality and utility of focal species data processed with automated audio-recognition software. Bioacoustics, 28, 539–554. Koch, G., Zemel, R., & Salakhutdinov, R., et al. (2015). Siamese neural networks for oneshot image recognition. Lille. In ICML deep learning workshop, Koger, B., Deshpande, A., Kerby, J. T., Graving, J. M., Costelloe, B. R., & Couzin, I. D. (2023). Quantifying the movement, behaviour and environmental context of groupliving animals using drones and computer vision. Journal of Animal Ecology. Korelidou, V., Simitzis, P., Massouras, T., & Gelasakis, A. I. (2024). Infrared thermography as a diagnostic tool for the assessment of mastitis in dairy ruminants. Animals, 14(18), 2691. MDPI Labuguen, R., Matsumoto, J., Negrete, S. B., Nishimaru, H., Nishijo, H., Takada, M., Go, Y., Inoue, K. i., & Shibata, T. (2021). Macaquepose: A novel “in the wild” macaque monkey pose dataset for markerless motion capture. Frontiers in Behavioral Neuroscience, 14, 581154. Landgraf, T., Gebhardt, G. H., Bierbach, D., Romanczuk, P., Musiolek, L., Hafner, V. V., & Krause, J. (2021). Animal-in-the-loop: Using interactive robotic conspecifics to study social behavior in animal groups. Annual Review of Control, Robotics, and Autonomous Systems, 4, 487–507. Lauer, J., Zhou, M., Ye, S., Menegas, W., Schneider, S., Nath, T., Rahman, M. M., Di Santo, V., Soberanes, D., & Feng, G., et al. (2022). Multi-animal pose estimation, identification and tracking with deeplabcut. Nature Methods, 19, 496–504. Lea, C., Vidal, R., Reiter, A., & Hager, G. D. (2016). Temporal convolutional networks: A unified approach to action segmentation. Springer. In Computer vision–ECCV 2016 workshops: Amsterdam, The Netherlands, October 8–10 and 15–16, 2016, Proceedings, Part III 14, pp. 47–54, Lecomte, C. G., Audet, J., Harnie, J., & Frigon, A. (2021). A validation of supervised deep learning for gait analysis in the cat. Frontiers in Neuroinformatics, 15, 712623. Lee, S., Waugh, B., O’Dell, G., Zhao, X., Yoo, W. S., & Kim, D. H. (2021). Predicting fruit fly behaviour using tolc device and deeplabcut. IEEE. in 2021 IEEE 21st international conference on bioinformatics and bioengineering (BIBE), pp. 1–6 Lei, Y., Dong, P., Guan, Y., Xiang, Y., Xie, M., Mu, J., Wang, Y., & Ni, Q. (2022). Postural behavior recognition of captive nocturnal animals based on deep learning: A case study of bengal slow loris. Scientific Reports, 12, 7738. Li, C., & Lee, G. H. (2023). Scarcenet: Animal pose estimation with scarce annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 17174–17183). Li, D., Zhang, K., Li, Z., & Chen, Y. (2020). A spatiotemporal convolutional network for multi-behavior recognition of pigs. Sensors, 20, 2381. Li, J., Xu, C., Jiang, L., Xiao, Y., Deng, L., & Han, Z. (2019a). Detection and analysis of behavior trajectory for sea cucumbers based on deep learning. IEEE Access, 8, 18832–18840. Li, S., Li, J., Tang, H., Qian, R., & Lin, W. (2019b). Atrw: a benchmark for amur tiger re-identification in the wild. arXiv preprint arXiv:1906.05586. Li, Y., Yang, G., Su, Z., Li, S., & Wang, Y. (2023). Human activity recognition based on multienvironment sensor data. Information Fusion, 91, 47–63. Lin, T. Y., Dollár, P., Girshick, R., He, K., Hariharan, B., & Belongie, S. (2017). Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 2117–2125). Expert Systems With Applications 289 (2025) 128330 16 E. Fazzari et al. Liu, D., Hou, J., Huang, S., Liu, J., He, Y., Zheng, B., Ning, J., & Zhang, J. (2023). LoTE-animal: A long time-span dataset for endangered animal behavior understanding. In Proceedings of the IEEE/CVF international conference on computer vision (pp. 20064–20075). Liu, F. T., Ting, K. M., & Zhou, Z. H. (2008). Isolation forest. IEEE. In 2008 eighth IEEE international conference on data mining, pp. 413–422, Liu, R., Xu, G., Jia, C., Ma, W., Wang, L., & Vosoughi, S. (2020a). Data boost: Text data augmentation through reinforcement learning guided conditional generation. arXiv preprint arXiv:2012.02952. Liu, X., Yu, S. y., Flierman, N., Loyola, S., Kamermans, M., Hoogland, T. M., & De Zeeuw, C. I. (2020b). Optiflex: Video-based animal pose estimation using deep learning enhanced by optical flow. BioRxiv, 2020–04. Liu, Y., Xu, H., Liu, D., & Wang, L. (2022). A digital twin-based sim-to-real transfer for deep reinforcement learning-enabled industrial robot grasping. Robotics and ComputerIntegrated Manufacturing, 78, 102365. Lostanlen, V., Palmer, K., Knight, E., Clark, C., Klinck, H., Farnsworth, A., Wong, T., Cramer, J., & Bello, J. P. (2019). Long-distance detection of bioacoustic events with per-channel energy normalization. arXiv preprint arXiv:1911.00417. Lu, W., Zhao, Y., Wang, J., Zheng, Z., Feng, L., & Tang, J. (2023). Mammalclub: An annotated wild mammal dataset for species recognition, individual identification, and behavior recognition. Electronics, 12(21), 4506. MDPI Luxem, K., Mocellin, P., Fuhrmann, F., Kürsch, J., Miller, S. R., Palop, J. J., Remy, S., & Bauer, P. (2022). Identifying behavioral structure from deep variational embeddings of animal motion. Communications Biology, 5, 1267. Mahmud, M. S., Zahid, A., Das, A. K., Muzammil, M., & Khan, M. U. (2021). A systematic literature review on deep learning applications for precision cattle farming. Computers and Electronics in Agriculture, 187, 106313. Manduca, G., Zeni, V., Moccia, S., Milano, B. A., Canale, A., Benelli, G., Stefanini, C., & Romano, D. (2023). Learning algorithms estimate pose and detect motor anomalies in flies exposed to minimal doses of a toxicant. Iscience, 26. Mankin, R., Hagstrum, D., Guo, M., Eliopoulos, P., & Njoroge, A. (2021). Automated applications of acoustics for stored product insect detection, monitoring, and management. Insects, 12, 259. Manoharan, D. S. (2020). Embedded imaging system based behavior analysis of dairy cow. Journal of Electronics and Informatics, 2, 148–154. Manriquez P, R., Kotz, S. A., Ravignani, A., & De Boer, B. (2024). Bioacoustic classification of a small dataset of mammalian vocalisations using deep learning. Bioacoustics, 33(4), 354–371. Taylor & Francis. Marks, M., Jin, Q., Sturman, O., von Ziegler, L., Kollmorgen, S., von der Behrens, W., Mante, V., Bohacek, J., & Yanik, M. F. (2022). Deep-learning-based identification, tracking, pose estimation and behaviour classification of interacting primates and mice in complex environments. Nature Machine Intelligence, 4, 331–340. Marshall, J. D., Li, T., Wu, J. H., & Dunn, T. W. (2022). Leaving flatland: Advances in 3d behavioral measurement. Current Opinion in Neurobiology, 73, 102522. Martin-Abadal, M., Ruiz-Frau, A., Hinz, H., & Gonzalez-Cid, Y. (2020). Jellytoring: Real-time jellyfish monitoring based on deep learning object detection. Sensors, 20, 1708. Mathis, A., Biasi, T., Schneider, S., Yuksekgonul, M., Rogers, B., Bethge, M., & Mathis, M. W. (2021). Pretraining boosts out-of-domain robustness for pose estimation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision (pp. 1859–1868). Mathis, A., Mamidanna, P., Cury, K. M., Abe, T., Murthy, V. N., Mathis, M. W., & Bethge, M. (2018). Deeplabcut: Markerless pose estimation of user-defined body parts with deep learning. Nature Neuroscience, 21, 1281–1289. Mathis, M. W., & Mathis, A. (2020). Deep learning tools for the measurement of animal behavior in neuroscience. Current Opinion in Neurobiology, 60, 1–11. McIntosh, D., Marques, T. P., Albu, A. B., Rountree, R., & De Leo, F. (2020). Movement tracks for the automatic detection of fish behavior in videos. arXiv preprint arXiv:2011.14070. McKenzie-Smith, G. C., Wolf, S. W., Ayroles, J. F., & Shaevitz, J. W. (2023). Capturing continuous, long timescale behavioral changes in drosophila melanogaster postural data. arXiv preprint arXiv:2309.04044. Mekruksavanich, S., Jantawong, P., & Jitpattanakul, A. (2022). Resnet-based deep neural network using transfer learning for animal activity recognition. IEEE. In 2022 6th international conference on information technology (InCIT), pp. 445–449. Mendoza, Q. A., Pordesimo, L., Neilsen, M., Armstrong, P., Campbell, J., & Mendoza, P. T. (2023). Application of machine learning for insect monitoring in grain facilities. AI, 4, 348–360. Mishra, S., & Sharma, S. K. (2023). Advanced contribution of iot in agricultural production for the development of smart livestock environments. Internet of Things, 22, 100724. Modlmeier, A. P., Colman, E., Hanks, E. M., Bringenberg, R., Bansal, S., & Hughes, D. P. (2019). Ant colonies maintain social homeostasis in the face of decreased density. ELife, 8, e38473. Mohialdin, A. M., Elbarrany, A. M., & Atia, A. (2023). Chicken behavior analysis for surveillance in poultry farms. Mondal, A., Nag, S., Prada, J. M., Zhu, X., & Dutta, A. (2023). Actor-agnostic multi-label action recognition with multi-modal query. In Proceedings of the IEEE/CVF international conference on computer vision (pp. 784–794). Morfi, V., Bas, Y., Pamuła, H., Glotin, H., & Stowell, D. (2019). Nips4bplus: A richly annotated birdsong audio dataset. PeerJ Computer Science, 5, e223. Mori, K., Yamauchi, N., Wang, H., Sato, K., Toyoshima, Y., & Iino, Y. (2022). Probabilistic generative modeling and reinforcement learning extract the intrinsic features of animal behavior. Neural Networks, 145, 107–120. Moummad, I., Serizel, R., & Farrugia, N. (2023). Regularized contrastive pre-training for few-shot bioacoustic sound detection. arXiv preprint arXiv:2309.08971. Nasiri, A., Amirivojdan, A., Zhao, Y., & Gan, H. (2023). Estimating the feeding time of individual broilers via convolutional neural network and image processing. Animals, 13, 2428. Nath, T., Mathis, A., Chen, A. C., Patel, A., Bethge, M., & Mathis, M. W. (2019). Using deeplabcut for 3d markerless pose estimation across species and behaviors. Nature Protocols, 14, 2152–2176. Neethirajan, S. (2020). The role of sensors, big data and machine learning in modern animal farming. Sensing and Bio-Sensing Research, 29, 100367. Ng, X. L., Ong, K. E., Zheng, Q., Ni, Y., Yeo, S. Y., & Liu, J. (2022). Animal kingdom: A large and diverse dataset for animal behavior understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 19023–19034). Nguyen, T. D., & Kresovic, M. (2022). A survey of top-down approaches for human pose estimation. arXiv preprint arXiv:2202.02656. Nilsson, S. R., Goodwin, N. L., Choong, J. J., Hwang, S., Wright, H. R., Norville, Z. C., Tong, X., Lin, D., Bentzley, B. S., & Eshel, N., et al. (2020). Simple behavioral analysis (simba)–an open source toolkit for computer classification of complex social behaviors in experimental animals. BioRxiv, 2020–04. Nolasco, I., Singh, S., Morfi, V., Lostanlen, V., Strandburg-Peshkin, A., Vidaña-Vila, E., Gill, L., Pamuła, H., Whitehead, H., & Kiskin, I., et al. (2023). Learning to detect an animal sound from five examples. Ecological Informatics, 77, 102258. Odo, A., Muns, R., Boyle, L., & Kyriazakis, I. (2023). Video analysis using deep learning for automatic quantification of ear biting in pigs. IEEE Access. Oestreich, W. K., Oliver, R. Y., Chapman, M. S., Go, M., & McKenna, M. F. (2024). Listening to animal behavior to understand changing ecosystems. Trends in Ecology & Evolution. Elsevier Oliveira, D. A. B., Pereira, L. G. R., Bresolin, T., Ferreira, R. E. P., & Dorea, J. R. R. (2021). A review of deep learning algorithms for computer vision systems in livestock. Livestock Science, 253, 104700. Otsuka, R., Yoshimura, N., Tanigaki, K., Koyama, S., Mizutani, Y., Yoda, K., & Maekawa, T. (2024). Exploring deep learning techniques for wild animal behaviour classification using animal-borne accelerometers. Methods in Ecology and Evolution, 15(4), 716–731. Wiley Online Library Özda¸s, M. B., Uysal, F., & Hardalaç, F. (2023). Classification of retinal diseases in optical coherence tomography images using artificial intelligence and firefly algorithm. Diagnostics, 13, 433. Pan, Z., Chen, H., Zhong, W., Wang, A., & Zheng, C. (2023). A cnn-based animal behavior recognition algorithm for wearable devices. IEEE Sensors Journal, 23, 5156–5164. Papandreou, G., Zhu, T., Chen, L. C., Gidaris, S., Tompson, J., & Murphy, K. (2018). Personlab: Person pose estimation and instance segmentation with a bottom-up, part-based, geometric embedding model. In Proceedings of the european conference on computer vision (ECCV). Papaspyros, V., Escobedo, R., Alahi, A., Theraulaz, G., Sire, C., & Mondada, F. (2023). Predicting long-term collective animal behavior with deep learning. bioRxiv, 2023–02. Park, J. S., O’Brien, J., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023b). Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual ACM symposium on user interface software and technology (pp. 1–22). Pedersen, W. A., McMillan, P. J., Kulstad, J. J., Leverenz, J. B., Craft, S., & Haynatzki, G. R. (2006). Rosiglitazone attenuates learning and memory deficits in tg2576 alzheimer mice. Experimental Neurology, 199, 265–273. Peng, X. B., Coumans, E., Zhang, T., Lee, T. W., Tan, J., & Levine, S. (2020). Learning agile robotic locomotion skills by imitating animals. arXiv preprint arXiv:2004.00784. Pereira, T. D., Aldarondo, D. E., Willmore, L., Kislin, M., Wang, S. S. H., Murthy, M., & Shaevitz, J. W. (2019). Fast animal pose estimation using deep neural networks. Nature Methods, 16, 117–125. Pereira, T. D., Tabris, N., Matsliah, A., Turner, D. M., Li, J., Ravindranath, S., Papadoyannis, E. S., Normand, E., Deutsch, D. S., & Wang, Z. Y., et al. (2022). Sleap: A deep learning system for multi-animal pose tracking. Nature Methods, 19, 486–495. Perez, M., & Toler-Franklin, C. (2023). Cnn-based action recognition and pose estimation for classifying animal behavior from videos: A survey. arXiv preprint arXiv:2301.06187. Pham, T. D. (2022). Classification of caenorhabditis elegans locomotion behaviors with eigenfeature-enhanced long short-term memory networks. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 20, 206–216. Piergiovanni, A., & Ryoo, M. (2019). Temporal gaussian mixture layer for videos. PMLR. In International Conference on Machine learning, pp. 5152–5161, Pillai, R., Gupta, R., Sharma, N., & Bansal, R. K. (2023). A deep learning approach for detection and classification of ten species of monkeys. IEEE. In 2023 International conference on smart systems for applications in electrical sciences (ICSSES), pp. 1–6, Rabiner, L., & Juang, B. (1986). An introduction to hidden markov models. IEEE ASSP Magazine, 3, 4–16. Rahman, S. A., Song, I., Leung, M., Lee, I., & Lee, K. (2014). Fast action recognition using negative space features. Expert Systems with Applications, 41, 574–587. https://www.sciencedirect.com/science/article/pii/S0957417413005812, https://doi.org/10.1016/j.eswa.2013.07.082. Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 779–788). Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28. Riekert, M., Klein, A., Adrion, F., Hoffmann, C., & Gallmann, E. (2020). Automatically detecting pig position and posture by 2d camera imaging and deep learning. Computers and Electronics in Agriculture, 174, 105391. https://doi.org/10.1016/j.compag.2020. 105391. Expert Systems With Applications 289 (2025) 128330 17 E. Fazzari et al. Rijsdijk, J., Wu, L., Perin, G., & Picek, S. (2021). Reinforcement learning for hyperparameter tuning in deep learning-based side-channel analysis. IACR Transactions on Cryptographic Hardware and Embedded Systems, (pp. 677–707). Robinson, T. P., Wint, G. W., Conchedda, G., Van Boeckel, T. P., Ercoli, V., Palamara, E., Cinardi, G., D’Aietti, L., Hay, S. I., & Gilbert, M. (2014). Mapping the global distribution of livestock. PloS ONE, 9, e96084. Romano, D., & Stefanini, C. (2021). Unveiling social distancing mechanisms via a fishrobot hybrid interaction. Biological Cybernetics, 115, 565–573. Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional networks for biomedical image segmentation. Springer. In Medical image computing and computerassisted intervention–MICCAI 2015: 18th International conference, Munich, Germany, October 5–9, 2015, Proceedings, Part III 18, pp. 234–241, Russello, H., van der Tol, R., & Kootstra, G. (2022). T-Leap: Occlusion-robust pose estimation of walking cows using temporal information. Computers and Electronics in Agriculture, 192, 106559. Saleh, A., Sheaves, M., Jerry, D., & Azghadi, M. R. (2022). Adaptive uncertainty distribution in deep learning for unsupervised underwater image enhancement. arXiv preprint arXiv:2212.08983. Saleh, D., Ahmed, M., Zaafan, M., Farouk, Y., & Atia, A. (2023). A pharmacology toolkit for animal pose estimation, tracking and analysis. IEEE. In 2023 International mobile, intelligent, and ubiquitous computing conference (MIUCC), pp. 1–7. Samsudin, W., Harizan, M., Ibrahim, M., Karim, R., & Ibrahim, W. (2022). Zebrafish larvae locomotor activity detection using convolutional neural network (cnn). Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., & Chen, L. C. (2018). Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4510–4520). Schall, E., Kaya, I. I., Debusschere, E., Devos, P., & Parcerisas, C. (2024). Deep learning in marine bioacoustics: A benchmark for baleen whale detection. Remote Sensing in Ecology and Conservation, 10(5), 642–654. Wiley Online Library Schneider, S., Lee, J. H., & Mathis, M. W. (2023). Learnable latent embeddings for joint behavioural and neural analysis. Nature, (pp. 1–9). Segalin, C., Williams, J., Karigo, T., Hui, M., Zelikowsky, M., Sun, J. J., Perona, P., Anderson, D. J., & Kennedy, A. (2021). The mouse action recognition system (mars) software pipeline for automated analysis of social behaviors in mice. ELife, 10, e63720. Selvaraju, R. R., Das, A., Vedantam, R., Cogswell, M., Parikh, D., & Batra, D. (2016). Gradcam: Why did you say that? arXiv preprint arXiv:1611.07450. Shamoun-Baranes, J., Burant, J. B., van Loon, E. E., Bouten, W., & Camphuysen, C. (2017). Short distance migrants travel as far as long distance migrants in lesser black-backed gulls larus fuscus. Journal of Avian Biology, 48, 49–57. Shaw, J. K., & Lahrman, S. (2023). The human–animal bond–a brief look at its richness and complexities. Canine and Feline Behavior for Veterinary Technicians and Nurses, (pp. 88–105). Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556. Stowell, D. (2022). Computational bioacoustics with deep learning: A review and roadmap. PeerJ, 10, e13152. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., & Rabinovich, A. (2015). Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 1–9). Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016). Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 2818–2826). Tassinari, P., Bovo, M., Benni, S., Franzoni, S., Poggi, M., Mammi, L. M. E., Mattoccia, S., Di Stefano, L., Bonora, F., & Barbaresi, A., et al. (2021). A computer vision approach based on deep learning for the detection of dairy cows in free stall barn. Computers and Electronics in Agriculture, 182, 106030. Teixeira, A. C., Ribeiro, J., Morais, R., Sousa, J. J., & Cunha, A. (2023). A systematic review on automatic insect detection using deep learning. Agriculture, 13, 713. Temenos, A., Voulodimos, A., Korelidou, V., Gelasakis, A., Kalogeras, D., Doulamis, A., & Doulamis, N. (2024). Goat-CNN: A lightweight convolutional neural network for poseindependent body condition score estimation in goats. Journal of Agriculture and Food Research, 16, 101174. Elsevier Thanh, V. Q., & Netramai, C. (2022). Deep learning-based monitoring system for distress on mice using behavior analysis. IEEE. In 2022 International electrical engineering congress (iEECON), pp. 1–4, Tian, Z., Chen, H., & Shen, C. (2019). Directpose: Direct end-to-end multi-person pose estimation. arXiv preprint arXiv:1911.07451. Tilak, O., Mukhopadhyay, S., Tuceryan, M., & Raje, R. (2010). A novel reinforcement learning framework for sensor subset selection. IEEE. In 2010 International conference on networking, sensing and control (ICNSC), pp. 95–100. Tucker Edmister, S., Del Rosario Hernández, T., Ibrahim, R., Brown, C. A., Gore, S. V., Kakodkar, R., Kreiling, J. A., & Creton, R. (2022). Novel use of fda-approved drugs identified by cluster analysis of behavioral profiles. Scientific Reports, 12, 6120. Uchino, T., & Ohwada, H. (2021). Individual identification model and method for estimating social rank among herd of dairy cows using yolov5. IEEE. In 2021 IEEE 20th international conference on cognitive informatics & cognitive computing (ICCI* CC), pp. 235–241, Ullah, N., Khan, J. A., Alharbi, L. A., Raza, A., Khan, W., & Ahmad, I. (2022). An efficient approach for crops pests recognition and classification based on novel deeppestnet deep learning model. IEEE Access, 10, 73019–73032. Valletta, J. J., Torney, C., Kings, M., Thornton, A., & Madden, J. (2017). Applications of machine learning in animal behaviour studies. Animal Behaviour, 124, 203–220. Elsevier Varma, A. L. S., Bateshwar, V., Rathi, A., & Singh, A. (2021). Acoustic classification of insects using signal processing and deep learning approaches. IEEE. In 2021 8th international conference on signal processing and integrated networks (SPIN), pp. 1048– 1052, Wang, J., Wang, N., Li, L., & Ren, Z. (2020). Real-time behavior detection and judgment of egg breeders based on yolo v3. Neural Computing and Applications, 32, 5471– 5481. Wang, K., Wu, P., Cui, H., Xuan, C., & Su, H. (2021a). Identification and classification for sheep foraging behavior based on acoustic signal and deep learning. Computers and Electronics in Agriculture, 187, 106275. Wang, X., Du, C., Wang, Y., Hu, S., & Zhao, Y. (2021b). Behavioral recognition of mice based on a deep network. IEEE. In 2021 IEEE 11th annual international conference on CYBER technology in automation, control, and intelligent systems (CYBER), pp. 840– 844. Weber, R. Z., Mulders, G., Kaiser, J., Tackenberg, C., & Rust, R. (2022). Deep learningbased behavioral profiling of rodent stroke recovery. BMC Biology, 20, 1–19. Wijeyakulasuriya, D. A., Eisenhauer, E. W., Shaby, B. A., & Hanks, E. M. (2020). Machine learning for modeling animal movement. PloS ONE, 15, e0235750. Wittek, N., Wittek, K., Keibel, C., & Güntürkün, O. (2023). Supervised machine learning aided behavior classification in pigeons. Behavior Research Methods, 55, 1624–1640. Wojke, N., Bewley, A., & Paulus, D. (2017). Simple online and realtime tracking with a deep association metric. IEEE. In 2017IEEE international conference on image processing (ICIP), pp. 3645–3649. Wu, A., Buchanan, E. K., Whiteway, M., Schartner, M., Meijer, G., Noel, J. P., Rodriguez, E., Everett, C., Norovich, A., & Schaffer, E., et al. (2020). Deep graph pose: A semisupervised deep graphical model for improved animal pose tracking. Advances in Neural Information Processing Systems, 33, 6040–6052. Wu, J., Leng, C., Wang, Y., Hu, Q., & Cheng, J. (2016). Quantized convolutional neural networks for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4820–4828). Xiao, S., Wang, Y., Perkes, A., Pfrommer, B., Schmidt, M., Daniilidis, K., & Badger, M. (2023). Multi-view tracking, re-id, and social network analysis of a flock of visually similar birds in an outdoor aviary. International Journal of Computer Vision, 131, 1532–1549. Xu, W., Zhang, X., Yao, L., Xue, W., & Wei, B. (2020). A multi-view cnn-based acoustic classification system for automatic animal species identification. Ad Hoc Networks, 102, 102115. Xudong, Z., Xi, K., Ningning, F., & Gang, L. (2020). Automatic recognition of dairy cow mastitis from thermal images by a deep learning detector. Computers and Electronics in Agriculture, 178, 105754. Yamada, J., Shawe-Taylor, J., & Fountas, Z. (2020). Evolution of a complex predatorprey ecosystem on large-scale multi-agent deep reinforcement learning. IEEE. In 2020 International joint conference on neural networks (IJCNN), pp. 1–8, Yamaguchi, S., Naoki, H., Ikeda, M., Tsukada, Y., Nakano, S., Mori, I., & Ishii, S. (2018). Identification of animal behavioral strategies by inverse reinforcement learning. PLoS Computational Biology, 14, e1006122. Yang, Y., Yang, J., Xu, Y., Zhang, J., Lan, L., & Tao, D. (2022). Apt-36k: A large-scale benchmark for animal pose estimation and tracking. Advances in Neural Information Processing Systems, 35, 17301–17313. Ye, S., Filippova, A., Lauer, J., Vidal, M., Schneider, S., Qiu, T., Mathis, A., & Mathis, M. W. (2022). Superanimal models pretrained for plug-and-play analysis of animal behavior. arXiv preprint arXiv:2203.07436. Yu, H., Xu, Y., Zhang, J., Zhao, W., Guan, Z., & Tao, D. (2021). Ap-10k: A benchmark for animal pose estimation in the wild. arXiv preprint arXiv:2108.12617. Zhang, K., Li, D., Huang, J., & Chen, Y. (2020). Automated video behavior recognition of pigs using two-stream convolutional networks. Sensors, 20, 1085. Zhou, F., Yang, X., Chen, F., Chen, L., Jiang, Z., Zhu, H., Heckel, R., Wang, H., Fei, M., & Zhou, H. (2022). Cross-skeleton interaction graph aggregation network for representation learning of mouse social behaviour. arXiv preprint arXiv:2208.03819. Zhu, Y., Lan, Z., Newsam, S., & Hauptmann, A. (2019). Hidden two-stream convolutional networks for action recognition. In Computer vision–ACCV 2018: 14th Asian conference on computer vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, Springer. pp. 363–378, Expert Systems With Applications 289 (2025) 128330 18