CNN and transfer learning-based classification model for automated cows feeding behaviour recognition from accelerometer data

Abstract

Due to technological developments, wearable sensors for monitoring farm animal behaviour have become cheaper, with longer life and more accessible for small farms and researchers. In this study, an acceleration measuring tag connected by BLE for monitoring behaviour of dairy cows was used. An optimal CNN-based model for the feeding behaviour classification was trained and the training process was analysed considering training dataset and the use of transfer learning. A classifier based on a neural network was trained by acceleration data collected in a research barn. Based on a dataset including 33.7 cow*days (21 cow recorded during 1-3 days) of labelled data and an additional free access dataset with similar acceleration data, a classifier with F1=93.9% was developed. The optimal classification window size was 90s. In addition, the influence of the training dataset size on the classifier accuracy was analysed for different neural networks using the transfer learning technique. During increasing of the training dataset size, the rate of the accuracy improvement decreased, and, starting from a specific point, the use of additional training data can be impractical. Relatively high accuracy was achieved with few training data when the classifier was trained using randomly initialised model weights, and higher accuracy was achieved when transfer learning was used. These findings can be used for estimation of the necessary dataset size for training neural network classifiers intended for other environments and conditions. HighlightsCNN cow feeding behaviour classifier was optimised for neck tags. Transfer learning technique significantly improves accuracy of CNN. Relation between the training dataset size and CNN accuracy was estimated. NN trained by one dataset can be inapplicable for others. BLE tags measuring acceleration transferred data in real time.