Detection of the number of herbage prehensions represents the key element for calculating total herbage intake in grazing ruminants. The reference is to the well-known definition of intake which arises from the product eating time x bite rate x bite mass. The study presents a deep learning-based approach for the automatic recognition of prehension in ewes using audio signals. Mono audio files, sampled at 16 kHz, were transformed into spectrograms and segmented into 500 ms time windows, which were subsequently labeled into two classes: prehension and non-prehension. A YOLOv11n model in classification configuration was trained on these time-frequency repre sentations. The dataset shows a slightly unbalanced distribution between classes, which is reflected in the metrics obtained on the test set: the overall accuracy reached approximately 0.83 with a macro F1-score of ~0.69. Per formance highlights a strong discriminative ability for the non-prehension class (precision ≈ 0.97, recall ≈ 0.84), while the prehension class shows a high recall (≈ 0.75) but a lower precision (≈ 0.36), suggesting the presence of false positives. The same study on cattle (data not published) yielded better results. This is because the ingestion behavior of sheep, which is very different from that of cattle, is difficult to recognize due to their tendency to penetrating vegetation in search of their favorite species, thus producing very loud noises that can interfere with the recognition of prehension sounds. Nonetheless, global metrics showed higher micro and weighted accuracy values (≈ 0.83 and ≈ 0.90), indicating that the model maintains good overall performance despite the difficulty in recognizing the minority class. The results demonstrate the potential of spectrogram-based audio analysis and lightweight architectures for automatic behavioral monitoring in precision livestock farming, while also high lighting potential directions for improvement in the trade-off between sensitivity and accuracy of grasping events.

Automatic recognition of herbage prehensions in grazing ewes

S. Bognanno;R. Avanzato;L. Beritelli;F. Beritelli;F. Gimmillaro;M. Avondo
2026-01-01

Abstract

Detection of the number of herbage prehensions represents the key element for calculating total herbage intake in grazing ruminants. The reference is to the well-known definition of intake which arises from the product eating time x bite rate x bite mass. The study presents a deep learning-based approach for the automatic recognition of prehension in ewes using audio signals. Mono audio files, sampled at 16 kHz, were transformed into spectrograms and segmented into 500 ms time windows, which were subsequently labeled into two classes: prehension and non-prehension. A YOLOv11n model in classification configuration was trained on these time-frequency repre sentations. The dataset shows a slightly unbalanced distribution between classes, which is reflected in the metrics obtained on the test set: the overall accuracy reached approximately 0.83 with a macro F1-score of ~0.69. Per formance highlights a strong discriminative ability for the non-prehension class (precision ≈ 0.97, recall ≈ 0.84), while the prehension class shows a high recall (≈ 0.75) but a lower precision (≈ 0.36), suggesting the presence of false positives. The same study on cattle (data not published) yielded better results. This is because the ingestion behavior of sheep, which is very different from that of cattle, is difficult to recognize due to their tendency to penetrating vegetation in search of their favorite species, thus producing very loud noises that can interfere with the recognition of prehension sounds. Nonetheless, global metrics showed higher micro and weighted accuracy values (≈ 0.83 and ≈ 0.90), indicating that the model maintains good overall performance despite the difficulty in recognizing the minority class. The results demonstrate the potential of spectrogram-based audio analysis and lightweight architectures for automatic behavioral monitoring in precision livestock farming, while also high lighting potential directions for improvement in the trade-off between sensitivity and accuracy of grasping events.
2026
precision grazing, audio tool, herbage prehensions, sheep
File in questo prodotto:
Non ci sono file associati a questo prodotto.

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.11769/740073
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact