Continuous monitoring of cattle foraging behavior is a major requirement for precision livestock farming applications. Several strategies have been proposed for this task but monitoring of free-ranging cattle for a long period of time has not been fully achieved yet. In this study, an algorithm is proposed for long-term analysis of foraging behavior that uses the regularity of this behavior to recognize grazing and rumination bouts. Acoustic signals are analyzed offline in two main stages: segmentation and classification. In segmentation, a complete recording is analyzed to detect regular masticatory events and to define the time boundaries of foraging activity blocks. This stage also defines blocks that correspond to no foraging activity (resting bouts). The detection of event regularity is based on the autocorrelation of the sound envelope. For classification, the energy of sound signals within a block is analyzed to detect pauses and to characterize their regularity. Rumination blocks present regular pauses, whereas grazing blocks do not. The evaluation of the proposed algorithm showed very good results for the segmentation task and activity classification. Both tasks were extensively analyzed with a new set of multidimensional metrics. Frame-based F1-score was up to 0.962, 0.891 and 0.935 for segmentation, rumination classification, and grazing classification, respectively. The average time estimation error was below 0.5 min for classification of rumination and grazing on recordings of several hours in length. In addition, a comparison for rumination time estimation was done between the proposed system and a commercial one (Hi-Tag; SCR Engineers Ltd., Netanya, Israel). The proposed algorithm showed a narrower error distribution, with a median of −2.56 min compared to −13.55 min in the commercial system. These results suggest that the proposed system can be used in practical applications.Web demo available at: http://sinc.unl.edu.ar/web-demo/rafar/.
Read full abstract