Smartphone-based behavior monitoring for real-time livestock welfare assessment

Thesis Abstract

The rapid expansion of smartphone technology and machine vision offers transformative potential for real-time livestock welfare assessment, addressing gaps in continuous monitoring that traditional farm visits and manual scoring cannot satisfy. A central problem is the lack of scalable, non-invasive, accessible tools that integrate behavioral indicators with welfare outcomes to enable timely decision-making on commercial farms. The study aims to develop and validate a smartphone-based framework that automatically detects and interprets critical behavioral patterns to inform welfare status, with a focus on cattle in intensive grazing and feedlot systems. Specific objectives are to (i) identify a core set of behavioral indicators strongly associated with welfare states (e.g., restlessness, foraging patterns, social interactions, vocalization frequency), (ii) design a mobile sensing and analytics pipeline incorporating computer vision, inertial measurement unit (IMU) data, and shallow machine learning models to classify welfare states, (iii) evaluate the framework’s accuracy against expert welfare assessments and standardized scoring systems, and (iv) assess practicality, acceptance, and reliability under commercial farm conditions. The methodology adopts a sequential explanatory mixed-methods design conducted over 18 months in three commercial livestock operations. The population comprises 1,200 dairy and beef cattle across farms in temperate and subtropical climates. A stratified random sample of 300 animals will be used for quantitative analysis, with a subsample of 60 animals subjected to in-depth qualitative observation. Data collection instruments include a smartphone app-enabled video capture module for behavioral detection, IMU sensors attached temporarily to body-mounted collars, standardized welfare scoring protocols (e.g., the Welfare Quality indicators), and a farmer-completed usability questionnaire. The smartphone app integrates computer vision techniques for gait, posture, and social proximity estimation, and time-series feature extraction from IMU data. Data validity will be enhanced by parallel human observer annotations and calibration sessions with 15 welfare experts. Reliability will be addressed using test-retest procedures, inter-rater reliability (Cohen’s kappa) for observational data, and intraclass correlation for sensor-derived metrics. Analytical approaches comprise - descriptive statistics to characterize baseline welfare indicators across farms and seasons; - supervised machine learning (random forest, gradient boosting, and support vector machines) to classify welfare states from multimodal features, with performance evaluated via 10-fold cross-validation and area under the ROC curve; - regression analyses (multivariate linear and logistic) to quantify associations between smartphone-derived metrics and established welfare scores; - time-series analysis (ARIMA and LSTM-based models) to forecast welfare trajectories using continuous behavioral data; - thematic analysis of interview transcripts from farmers to elucidate perceived barriers, usability issues, and workflow integration. The study will use a theoretical lens grounded in the Welfare Science framework and the Technology Acceptance Model (TAM) to interpret adoption determinants and interpretive validity of smartphone-based assessments. A conceptual model will link smartphone-derived behavioral metrics to welfare outcomes through mediating factors such as environmental conditions, feeding regime, and social dynamics. Ethical considerations address animal welfare during any sensor attachment, data privacy, and informed consent from farm owners. Expected findings include (i) identification of a robust feature set from video and IMU data that predicts welfare status with an AUC exceeding 0.85 in cross-species validation; (ii) demonstration of high agreement (kappa > 0.70) between smartphone-derived welfare classifications and expert human scores; (iii) evidence that real-time feedback reduces latency in welfare interventions by at least 24 hours; and (iv) insights into factors affecting adoption, including perceived ease of use, perceived usefulness, and compatibility with farm workflows. The contribution to knowledge lies in integrating mobile sensing, computer vision, and machine learning within a practical, scalable framework for real-time welfare assessment, thereby advancing non-invasive livestock monitoring and providing a decision-support tool for farmers and veterinarians. The study will culminate in actionable recommendations for the design and deployment of smartphone-based welfare monitoring systems, guidelines for sensor integration on commercial farms, and policy considerations for data privacy and animal welfare standards. It may also identify avenues for future research, such as cross-species generalization, integration with environmental sensors, and adaptive alerting mechanisms to optimize welfare outcomes.