Abstract
Monitoring deer behaviour continuously and objectively is essential for accurate welfare assessment in pastoral farming, yet conventional observation methods are labour-intensive and prone to subjective bias. We present an unsupervised machine-learning pipeline for classifying farmed red deer (Cervus elaphus) behaviour from wearable accelerometer data. Tri-axial accelerometers mounted on the head and ear of stags recorded movement at 50Hz over 48h. k-means clustering of the normalised signals yielded eight prototypical movements spanning three behavioural categories: inactive states (lying, resting with or without rumination/panting), grazing (stationary and walking), and rapid ear flicks. A first-order Markov chain then linked these prototypes into higher-order sequences we call ”super-prototypes”. These compound behaviours included ear-flick bouts and alternating grazing-stepping cycles. Within the single 24-h window, activity varied with time of day, ear flicks being markedly more frequent by day. Against independent video annotation, the eight categories agreed with the observer’s labelling at a mean of 0.96 across prototypes. Because classification reduces to a nearest-centroid operation, the algorithm runs on a 16MHz microcontroller at over 4×108 classifications per second, four orders of magnitude faster than the sensor sampling rate. The pipeline requires no hand-labelled training data and runs in real time on the animal, a feasible basis for welfare monitoring in extensive deer farming systems once validated under field conditions.