Deep learning based classification of sheep behaviour from accelerometer data with imbalance

Classification of sheep behaviour from a sequence tri-axial accelerometer data has the potential to enhance management. Sheep is inherently imbalanced (e.g., more ruminating than walking) resulting in underperforming classification for minority activities which hold importance. Existing works have not addressed class imbalance and use traditional machine learning techniques, e.g., Random Forest (RF). We investigated Deep Learning (DL) models, namely, Long Short Term Memory (LSTM) Bidirectional LSTM (BLSTM), appropriate sequential data, data. Two sets were collected normal grazing conditions using jaw-mounted ear-mounted sensors. Novel this study, alongside typical single classes, walking, depending on behaviours, samples labelled with compound walking_grazing. The number steps performed observed 10 s time window was also recorded incorporated models. designed several multi-class studies being synthetic DL models achieved superior performance ML especially augmented 4-Class + Steps: 88.0%, RF 82.5%). methods showed generalisability unseen (i.e., F1-score: BLSTM 0.84, 0.83, 0.65). LSTM, sub-millisecond average inference time, making them suitable real-time applications. results demonstrate effectiveness conditions. techniques can generalise across different sheep. study presents strong foundation development such animal monitoring.