Smart IoT-Enabled Sensor Network for Real-Time Livestock Welfare Monitoring

Thesis Abstract

The livestock sector faces increasing welfare and productivity challenges due to limited real-time monitoring capabilities, variability in behavior, and environmental stressors that can lead to disease outbreaks and economic losses. This study proposes and empirically evaluates a Smart IoT-Enabled Sensor Network for Real-Time Livestock Welfare Monitoring to enhance early detection of welfare deviations, improve timely decision-making, and reduce intervention costs. The aim is to develop an integrated, scalable sensing and analytics framework that infers welfare states from multi-modal data streams and provides actionable insights to farm managers. Specific objectives are to (i) design and deploy a low-power IoT sensor network (accelerometers, gyroscopes, temperature-humidity, heart-rate patches, and GPS) on a representative cohort of 320 cattle across three commercial farms, (ii) develop a data fusion architecture and feature extraction pipeline to derive welfare indicators such as activity patterns, thermal comfort, respiration rate proxies, and social proximity metrics, (iii) implement machine learning models for real-time anomaly detection and welfare classification, and (iv) evaluate system performance against ground-truth veterinary assessments and production metrics over a 12-month period. Methodologically, a mixed-methods approach is adopted within a pragmatic research design. The population comprises lactating dairy cattle managed under conventional intensive farming systems in the southeastern region. A stratified random sample of 320 animals is selected, with equal representation across parity and lactation stages. Data collection integrates continuous sensor streams transmitted via a LoRaWAN-based network to a cloud platform, complemented by weekly veterinary welfare scoring using a standardized checklist and biweekly milk yield records. Instrument validity and reliability are ensured through calibration protocols for physiological sensors, pilot testing of the welfare scoring system, and cross-validation with existing farm automation records. Data analysis employs time-series preprocessing, feature engineering (movement entropy, accelerometer variance, skin-temperature proxies, and proximity graphs), and supervised learning using Random Forest and Gradient Boosting for welfare classification, with Recurrent Neural Networks for sequence modeling of welfare trajectories. Model performance is evaluated using cross-validated metrics (accuracy, F1-score, AUC-ROC) and compared against baseline rule-based thresholds. An unsupervised clustering approach (K-means, DBSCAN) is used to identify welfare states and transition patterns, while SHAP values facilitate interpretability of feature importance. The theoretical grounding draws on the Theory of Planned Behavior to understand farmer adoption of IoT-based welfare monitoring and the Wildlife Welfare Indicators framework to structure welfare metrics; these inform the design of user-centered dashboards and alerting mechanisms. Key expected findings include (i) high-resolution welfare indicators enabling early detection of lameness, heat stress, and social disruption with lead times exceeding 24–48 hours prior to clinical manifestation, (ii) robust multimodal models achieving validation accuracy above 82% and AUC-ROC above 0.88 for welfare classification, (iii) demonstrable reductions in treatment costs and non-productive days by at least 12% in participating farms, and (iv) positive shifts in farmer decision-making efficiency driven by actionable alerts and transparent model explanations. The study contributes to knowledge by integrating IoT sensing with advanced analytics to operationalize welfare indicators at individual and group levels, offering a scalable blueprint for real-time livestock welfare management and revealing practical barriers to adoption through theory-driven insights. The main conclusion anticipated is that a well-designed IoT sensor network, underpinned by transparent analytics and farmer-centered interfaces, can significantly improve real-time welfare monitoring and decision support in commercial cattle farming. Recommendations include scaling the platform to diverse species and production systems, incorporating adaptive alert thresholds to mitigate alarm fatigue, integrating cloud-based privacy and security controls, and conducting longitudinal investigations into long-term welfare and productivity outcomes.