Using a neural network based vocalization detector for broiler welfare monitoring
The poultry industry in Flanders, Belgium, is characterized by highly efficient production systems.
To assure sustainable food production, where broiler quality of life is key, data collection systems and (automated) interpretation of broiler health and welfare are of capital importance.
This paper describes the realization and deployment of an acoustic detector for broiler vocalizations, as part of a larger set of behavior and welfare monitoring tools developed within the ICON-WISH project. The vocalization detector is based on a convolutional neural network.
For training, a labelled library with vocalizations is built (>2k samples), based on a large set of broiler audio recordings covering the full broiler life-span.
Four different types of vocalizations (pleasure notes, distress calls, short peeps and warbles) are identified in function of broiler age to account for spectral changes.
Based on this library, the neural network achieves a balanced accuracy of 87.9%.
To indicate its potential, the detector is applied in a real-life medium scale housing, in which several groups of broilers are exposed to different environmental conditions (heat stress a.o.).
The occurrence and type of vocalizations is analyzed, and the potential to identify broiler stress is investigated.