Discriminative power of acoustic features for jaw movement classification in cattle and sheep

Measuring foraging behaviour and pasture intake of ruminants is useful but difficult. Acoustic monitoring is one of the most promising methods for this task. In this work, we test its potential to classify jaw movements (JM) according to type (bite, chew, chew-bite), activity (grazing or rumination), and forage species being consumed. Experiments with cows and sheep grazing and ruminating several forages were conducted. First, each JM was manually identified, classified and described by two sets of sound features: i) one containing energy bands of the spectrum ($EB$), and ii) one containing four complementary (or global) variables ($C_V$). Two models were evaluated, one with $EB$ alone and the other one combining $EB$ and $C_V$. Jaw movements were correctly classified by type with 73.0% and 78.5% average accuracy. Accuracy was better for cows than sheep (85 vs 66%). Simultaneous identification of the type of JM and plant species was about 78%. Classification accuracy of activity based on chews averaged 68.5% and 77.0% for rumination and grazing. Models including global variables performed better than using only the spectrum. Acoustic monitoring is a very promising method for further development, particularly to study diet selection.