Image-based individual identification of wild Black-necked Cranes (Grus nigricollis) using deep learning

Individual identification of wildlife is a cornerstone of behavioral and ecological research. In recent years, developments in deep learning for computer vision have broadened the applicability of image-based individual identification in animals. However, because high quality field images of birds are difficult to collect and curate, image-based individual identification in birds remains rare. In this study, we assessed the feasibility of individual identification in wild birds using field images of the Black-necked Crane (Grus nigricollis) as a focal species. A total of 45 individual cranes were photographed in the wild, yielding a dataset of 23,830 images. We processed these images using a three-stage workflow comprising object detection (YOLO11n), instance segmentation (Co-DINO), and individual classification (YOLO11s-cls). The model achieved a Top-1 accuracy of 90.5% on the original validation set, and this increased to 93.5% on a clarity-filtered validation set created by removing the 10% most blurred images from the original validation set. Finally, we evaluated its potential to identify new individuals. On the test set, when the model correctly identified 93.9% of images of new individuals, 18.8% of images of known individuals were falsely rejected. These results demonstrate the applicability of this approach for identifying individual Black-necked Cranes in the wild, and lay the groundwork by providing a methodological basis and practical reference for image-based identification in long-term field studies across large spatial scales.