Quantification of grass-severing bites performed by grazing cattle using halter-mounted accelerometers and machine learning

Grasslands represent a key element of agroecosystems for sustainable food systems. A better understanding of the grazing behaviour of domestic herbivores is essential to support innovations for grassland management and define grazing practices that support rather than enter into conflict with biodiversity. A key component of the grazing process is the grass-severing bite by which the herbivore collects forage from a pasture. How often, where, and when such bites are performed are relevant indicators of the grazing behaviour of cattle and could be used as indicators to guide farmers in pasture management. In this work, we developed a methodology to create a Machine Learning (ML) model for identifying grass-severing bite events from the Inertial Measurement Unit (IMU) signals of a sensor placed on the neck of cows. The two-phase process consisted of classifying every period of behaviour of cattle into two mutually exclusive behaviours: “ingestion” and “other” (phase 1), and then counting the number of bites taken during each period classified as “ingestion” (phase 2). Seven dry red-pied Holstein cattle and two Blonde d'Aquitaine x Belgian White and Blue cross-breds were observed. A total of 39 h and 25 min of video were recorded and tagged for the different behaviours to train several ML algorithms. During phase 1, four different window segmentations and two different splits of the data were used to train and test four ML classification algorithms: Bagged Tree, Medium k-NN, Fine tree and linear SVM. The results show that Bagged Tree algorithms with 30 s windows and 90 % overlap gave the best results during the first phase, with an accuracy of 97.83 % for split 1 and 98.07 % for split 2. During phase 2, the same four window segmentations as for phase 1 were used, to test regression algorithms to quantify the number of bites taken during each time-window. Two machine learning algorithms were tested: Bagged Tree and Medium NN, on 5 sessions of 30 min. The sessions ranged between 0 % and 94 % of ingestion time. Phase 2 results showed that Bagged Tree regression algorithms with 10 s windows and 90 % overlap performed the best, with an average RMSE of 1.83 for the tested value and an error percentage of -1.93 % and 0 % for the session with 94 % or 0 % of ingestion time, and between +15.06 % and +26.97 % of error for sessions where the animal alternates frequently between both behaviours. The data and code used in this study are openly available on a public depository