Deep learning for Amur tiger re-identification in camera traps: A tool assisting population monitoring and spatio-temporal analysis

• Construct new datasets from camera traps in boreal forest ecosystem for Amur tiger segmentation and re-identification. • Compare 10 outstanding deep learning backbones based on improved CNN or Transformer. • Design a new two-stage re-identification pipeline with a feature fusion method to improve performance. • Support practical population monitoring and spatio-temporal analyses of Amur tigers: a case study in home range. Reliable and accurate individual data is crucial for the conservation and recovery of endangered Amur tigers. With the growing volume of data accumulated from camera traps in recent years, deep learning has emerged as an effective tool for wildlife detection and identification. While several studies have focused on Amur tiger re-identification, most rely on relatively ideal datasets and may not be suitable for practical monitoring. Less work has been done in the open-world scenario. To fill this gap, we focused on 16 wild Amur tigers in the Northeast Tiger and Leopard National Park, and (1) constructed two datasets more aligned with the wild environment, containing 1,121 and 34,691 images respectively; (2) designed a two-stage re-identification pipeline that includes the segmentation and classification steps; (3) proposed a fusion module based on representation and metric learning to improve performance. After comparing various deep learning backbones, we chose DDRNet-39 and ConvNeXt-small to test this pipeline, which yielded an accuracy of 95.49% on the test set. Using a case study on home range and dispersal behavior, we illustrated how our approach could assist in population monitoring and spatio-temporal analyses of Amur tigers, as well as other species with distinctive coat patterns. Future researches can proceed to enhance this approach by expanding datasets or integrating multi-modal data, further promoting its applications in the real world.