Classification of sheep behaviour though sensor-based collars and machine learning

• Autor(a) principal: Fonseca, Luís Carlos Duarte
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Texto Completo: http://hdl.handle.net/10773/38704
• topic: Accelerometer; Animal behaviour; Collar; Decision tree; Gyroscope; Machine learning; Rumination; Sheep

Animal behaviour monitoring is a topic of increasing importance in the past few years, due to the needs to optimize the multiple fields of animal production to answer worldwide population growth. Considering that animal behaviour is an indicator of animal welfare, by automating and scaling this process, it is possible to constantly monitor entire flocks, allowing for the early detection of diseases which might otherwise incur into huge costs and losses. Simultaneously, the cost of microprocessors and small sensors is decreasing. This allows for the construction of low-powered devices capable of both retrieving and classifying animal behaviour, resorting to Machine Learning. Here, the study revolves around a collar equipped with 3 sensors: accelerometer, gyroscope, and thermometer, resulting in an initial set of 12 features. From the wide variety of algorithms, Decision Trees are particularly suitable for this task due to their many advantages, namely simplicity and readability. The latter allows for a conversion of the developed models into raw code, passive of being directly applied on the devices. Therefore, the goal of this dissertation is to evaluate the ability of Decision Trees to classify six main behaviours (Lying, Standing, Inclined, Eating, Walking and Ruminating), while simultaneously assessing which sensors and features are more valuable to optimize the production costs of said devices. Overall, it is possible to classify these behaviours reliably (with a balanced accuracy around 90%) and remove one of the sensors, without impacting model’s performance significantly. Notwithstanding, Temperature surprisingly arises as a potential important feature to predict postures. This thesis was realized with the support of the Instituto de Telecomunicações.