UC Davis

Animal detection and counting via image classification is an emerging area of research for the poultry industry as it offers a non-invasive approach to monitoring poultry. A similar approach can be applied to relevant websites and social media in order to detect non-commercial poultry activities. These types of analysis can enable a deeper understanding of the spatiotemporal patterns of non-commercial poultry sales (e.g. backyard chickens and game fowl), for various purposes including disease modeling and targeted university agricultural extension activities. In this research, five image classification models were used to evaluate their accuracy in distinguishing chicken from non-chicken species in diverse settings ranging from commercial to non-commercial settings. Results showed an overall accuracy ranging from 35% to 99%. When comparing the models, the convolutional neural network ConvNeXt model demonstrated superior performance in terms of its accuracy, model training time, and prediction time. The ability to use image classification in concert with social media and other online platforms offers a novel approach toward surveying backyard chickens and game fowl flocks.