Machine Learning Classification of Free-Ranging Livestock Behaviour Using Wearable Sensors

Abstract

Introduction: Effective animal conservation and management need periodic data on the behaviour of free-ranging livestock and the environment influencing it. The traditional method is precise but sporadic and laborious for large and/or inaccessible landscapes. Materials and methods: The current paper introduces the artificial intelligence-based methodology for assessing livestock movement using accelerometer, gyroscope, GPS, ambient temperature, pasture state and direct observations of behaviour. The experimental data set comprised 21,600 labelled windows measured on 36 free-ranging cows for 30 observation days. Data were filtered and aligned, as well as split into training, validation and testing sets. Results and discussion: The suggested methodology allowed obtaining 96.2% accuracy, 95.8% precision, 95.4% recall and 95.6% F1-score. Significant connections between animals and environment were linked to pasture availability, heat load, travelling distance, water presence and land slope. The algorithm helped to increase recognition performance under different environmental conditions and produce explainable alerts for fieldwork verification. Conclusion: The results prove the possibility of expanding the range of non-invasive monitoring in cases where outcomes still stay related to environmental context and expert observations