The left side flank as source of information for animal behaviour and welfare in dairy cow.

The development of devices and methods able to spot changes of behaviours and physiological parameters from normality in a timely manner, is one of the aims of Precision Livestock Farming. With this vision, the aim of the present study was to develop a model to identify animals’ posture and predefined behaviours (moving, feeding, resting, ruminating and standing still) from data collected by a single tri-axial accelerometer located on the left side flank of dairy cows and evaluate its accuracy and precision. This spot was chosen because in ruminants, beyond behaviour, it potentially enables also the monitoring of rumen and turaco-abdominal contractions associated with breathing and involved in both urination and defecation. Twelve Italian Red-and-White lactating dairy cows were equipped with a tri-axial accelerometer located on the left side paralumbar fossa and were observed on average for 136 ± 29 min per cow by two trained observers who continuously recorded animals’ posture and behaviour as a reference. Acceleration data were grouped in time windows of 8s overlapping by 33%, for a total of 35133 rows. For each row, 32 different features were extracted and used by machine learning algorithms for the classification of posture and behaviour. To build up a predictive model, the dataset was split in training and testing datasets, characterized by 75 and 25% of the observations, respectively. Four algorithms were tested: Random Forest, K Nearest Neighbors, Extreme Boosting Algorithm (XGB) and Support Vector Machine. As regards behaviour classification a Convolutional Neural Network model (CNN, a Deep Learning Model), made of 8 layers, was also tested. Among machine learning models XGB showed the best accuracy (0.99) and Cohen’s kappa (0.99) in predicting posture, whereas Random Forest had the highest overall accuracy in predicting behaviours (0.76), showing a balanced accuracy from 0.96 for resting to 0.77 for moving. The higher rate of misclassification was found between feeding, moving and standing still. The Deep Learning model showed an overall accuracy in predicting behaviour of 0.92. Overall, the application of a single tri-axial accelerometer at the left side paralumbar fossa of mid-lactating dairy cows gave very accurate results regarding the prediction of posture and resting behaviour using machine learning models, whereas precision and accuracy for ruminating and feeding behaviours were greatly improved by the use of CNN.