Due to the long relevant
time span and the spatial conditions, the development of a sensor-based assessment
for cattle behaviour in the context of weaning under semi-natural conditions is
essential. Ten beef calves had been equipped with a rumination halter with
noseband and accelerometer at the cheek, an accelerometer at one hind leg and a
microphone at a collar. The sensor data were compared with observational data
(10 h/calf, direct continuous focal animal sampling). Using the leg-fitted
accelerometer lying was estimated very well. As storage and battery capacity are
high, it is a suitable tool for long-term assessment. The automatic
classification of the noseband data overestimated feeding and underestimated rumination.
Suckling was mainly classified as feeding. Even though sensitivity, specificity
and accuracy were above 70 %, which is acceptable. The classification
algorithm could be refined, however due to the high risk of injuries, halters should
not be used for long-term assessments. Machine learning was used to classify
sound data as feeding, ruminating, drinking and suckling. Behaviours were
classified correctly with an accuracy of 63 % which is not acceptable. In
conclusion, oral behaviours were more difficult to distinguish. Therefore, methods
need to be improved.