Data Science Approaches for the Analysis of Animal Behaviours

Abstract

Animals’ welfare can be categorized and predicted by observing their daily routine and behaviour and consequently, conclusions about their health status and their ecology can be made. Yet, observing the animals’ activity is time-consuming and labour intensive, therefore there is a need to develop automated behavioural monitoring systems for more efficient and effective computerised agriculture. In this study, accelerometer and gyroscope measurements were collected from seven Hebridean ewes located in Cheshire, UK. Once the activities of the animals were labelled as grazing, resting, walking, browsing, scratching, and standing, data analysis was conducted. The performance of the random forest was evaluated as we have previously suggested that this algorithm can provide advantages and has been proven to adequately classify the behaviours of the animals. When using features from both the accelerometer and gyroscope, the algorithm obtained the best results having accuracy and kappa value of 96.43%, and 95.02%, respectively. However, using data from the accelerometer exclusively, only decreased the accuracy by 0.40% and kappa value by 0.56%. Therefore, in future work, we will consider only the use of accelerometer sensor and we will test the performance of the same features and random forest for real-time activity recognition using larger cohort of animals as well as mixed breed animals.