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

Abstract

For each experiment, a qualitative analysis of the influence of complementary variables was conducted. I. EXPERIMENT 1A. Identification of type of jaw movement (B, C, CB) for each animal species separately (2 models, one per animal species)

Short-term feeding behaviour sound classification method for sheep using LSTM networks

“…Therefore, performance comparison was carried out between the developed classifier and the previous multi-class feeding behaviour classifiers, as shown in Table 5. Milone et al [13] MP IC, CB, IB 93.33% (cattle) Not provided Chelotti et al [1] MP IC, CB, IB 84.0% (cattle) Not provided Deniz et al [15] MP IC, CB, IB 76.4% (cattle) Not provided Giovanetti et al [40] Acc GR, RE, RU 93% (sheep) Not provided Zehner et al [42] NP, Acc EA, RU, DR, OB 94.5% # (cow) Not provided Chelotti et al [17] AR IC, CB, IB 90.74% (cattle) 92.39% # Decandia et al [41] Acc GR, RU, OB 89.7% (sheep) Not provided Galli et al [18] MP Generally speaking, the previous multi-class classifiers were established based on the sensor data acquired by electronic sensors such as pressure sensors, accelerometers, and so on, or acoustic signal obtained by microphones or audio recorders. Most of the sensor data-based classifier focused on recognizing long-term activities (rumination and grazing) rather than individual jaw movements [17] .…”

“…Classifiers established by Milone et al [13] and Galli et al [18] were both aimed for identifying short-term feeding behaviour sound for sheep, where hidden Markov models and linear discriminant analysis were respectively used as the foundation methods. The overall accuracy achieved by Milone et al [13] was higher than that of the classifier constructed by Galli et al [18] , both of which were lower than the overall accuracy achieved by C PW_MFCC in this study.…”

“…Short-term feeding behaviour sound classifiers [1,17,18] were established for cattle. For computational overhead reduction, only acoustic features in time domain were utilized to classify sound signal produced by jaw movement events (including bite, chew, and chew-bite) [1] .…”

“…For a better understanding of the performance of different classifiers, Equations (18) to (22) were carried out for each sound episode category to calculate the recall, specificity, precision, accuracy, and F1-score of each classifier. The results are presented in Table 3.…”

“…As a result, intake estimation for ruminants based on acoustic features in short-term feeding behaviour attracted great attention among researchers [2,12] . Identifying the sound produced by short-term feeding behaviour has previously been reported in the literatures [13][14][15][16][17][18] . Various methods, such as hidden Markov models [13] , support vector machine [16] , decision tree [17] , random forest [17] , radial basis function network [17] , linear discriminant analysis [18] , and so on, were employed as the foundation techniques to classify short-term feeding behaviour.…”