Deep metric learning for individual cattle identification using coat patterns: Proposal for a best practice

Effective individual cattle identification is crucial for Precision Livestock Farming (PLF), particularly in managing growth and improving animal welfare through remote identification in unconfined environments. This study proposes a best practice, a systematic strategy for training and hyperparameter optimisation to identify the optimal-performing model, for the identification of individual cattle using a straightforward Deep Metric Learning (DML) structure based on coat patterns. The proposed best practice was developed using the open-access Unmanned Aerial Vehicle (UAV)-RGB dataset and validated on the Ground-RGB dataset. Both datasets, contributed by the authors, were obtained from the same semi-free-range Simmental male beef cattle herd of 96 individuals on a commercial farm in China but were captured using different platforms: UAV and stationary frames, respectively. The study also evaluates the impact of the semi-hard mining technique on the performance of DML. Analysis of the optimal models derived by using the best practice for the two datasets indicates that cattle in standing postures yield slightly better performance than those with diversity in postures, including both standing and lying, with a 1.41% improvement in Cumulative Matching Characteristic at rank 1 (CMC@1). However, both optimal models demonstrate performance comparable to that in state-of-the-art studies. When the best practice was tested for generalisability in the publicly available OpenCows2020 (breed: Holstein Friesian) dataset, the optimal model achieved remarkable results, with the CMC@1 and CMC@5 both at 99.80% and the mean Average Precision (mAP) at 99.67%, surpassing previous studies using the same dataset. These results confirm the effectiveness of the proposed best practice in terms of accuracy and generalisability for individual cattle identification and offer valuable insights for similar biometric identification tasks, such as individual identification using facial images.