A machine learning approach to developing behaviour classification models of grazing cattle using Global Navigation Satellite System (GNSS) data

Monitoring the behaviour of free-ranging cattle is critical for effective management in extensive grazing systems. This study investigated whether a Global Navigation Satellite System (GNSS) tracking dataset could serve as a single data source for (1) classifying predominant grazing cattle behaviours (Grazing, Resting, and Walking) using Machine Learning (ML) models, and (2) generating diurnal activity profiles. Ten 17 – 22 months Angus steers weighing 363 ± 8.3 kg (mean ± SD) were grazed in a 29.6-ha paddock for 14 days. Each steer was fitted with a UNETracker II GNSS collar programmed to record five consecutive GNSS positions at 10-second intervals (Burst), followed by a 5-minute interval between Bursts, and was monitored via a high-definition video camera for 6 days. Grazing, Resting (Standing and Lying Down), and Walking behaviours were annotated and further synchronised with the GNSS data to generate a behaviour-annotated dataset. This dataset was segmented into three subsets: Burst 3 (B3), Burst 4 (B4), and Burst 5 (B5), containing sequences of 3, 4, and 5 GNSS points, respectively. To develop the ML models, 11 speed-based features, including acceleration, mean, standard deviation, coefficient of variation, and zero-speed count, were calculated from the annotated data for each Burst. Due to an imbalance in behavioural classes, the SMOTE oversampling technique was applied to the training data to improve model performance. Four ML algorithms (RF, XGBoost, k-NN, SVM) were used to develop behaviour classification models. Results showed that XGBoost at B3 outperformed other ML models in distinguishing between Grazing, Resting and Walking behaviours, and the oversampling improved the overall accuracy from 0.69 to 0.79. The Precision, Recall, F1 Score, Cohen’s kappa coefficient, and Matthew’s correlation coefficient (MCC) of the XGBoost model before vs. after oversampling were 0.68 vs 0.80, 0.62 vs 0.79, 0.64 vs 0.79, 0.41 vs 0.68 and 0.43 vs 0.69, respectively. These findings highlighted the potential of GNSS-derived speed features, with XGBoost emerging as a robust approach for classifying beef cattle behaviour at a large-scale pasture landscape.