Supplementary code and dataset for individual identification of wild raptors using a deep learning approach: A case study of the white-tailed eagle

 Explanation of the Supplementary Code and Dataset

 (1) Background
Recognizing individual animals plays a vital role in ecological researches. The evolution of computing power has paved the way for deep learning techniques to be applied in wildlife identification. Nevertheless, distinguishing individual animals in the unpredictable natural environment continues to pose a significant hurdle. To tackle this issue, a novel deep learning framework, termed the object tracking–face extraction–sampling–recognition (OFSR) framework, has been introduced. This approach employs advanced deep learning to isolate facial characteristics and integrates a multitask system that shares information across tasks to incorporate additional data, boosting the precision of individual recognition.

 (2) Code and Dataset
The study offers complete, executable code for the individual identification process, enabling others to replicate and expand upon the work. The dataset was gathered using the DS-2TD91LY-AFG model PTZ camera, which can detect subjects up to 6.7 km away, making it well-suited for monitoring white-tailed eagles across wide regions. The camera provides a resolution of 1920×1080 pixels and supports multiple encoding options such as H.264, H.265, Smart264, Smart265, and MJPEG, ensuring high-quality and efficient video recording. In December 2022, a total of 131 video clips were captured, featuring 47 distinct white-tailed eagles, forming a comprehensive dataset for testing and refining the identification model.

(3) Notification in Model test

Rather than directly dividing all individual images into training and test sets, we implemented a video-level partitioning approach. Specifically, individuals appearing across multiple videos were designated as test subjects, with 20% of their corresponding videos allocated to the test set. Videos featuring individuals that appeared only once were incorporated into the training set to enhance task complexity, while ensuring these videos were excluded from the test set to avoid overlap between the training and test datasets.

Article link: https://doi.org/10.1016/j.ecoinf.2025.103379