Individual identification of brown bears using pose-aware metric learning

Area of Science:

• Wildlife ecology
• Computer vision
• Conservation biology

Background:

• Individual identification is crucial for ecological research but traditional methods are invasive and costly.
• Distinguishing individuals in unmarked species using computer vision remains a significant challenge.
• Alaskan coastal brown bears (Ursus arctos) present a difficult case for visual identification due to natural variations.

Purpose of the Study:

• To develop and validate a novel AI model for non-invasive individual re-identification (ReID) of Alaskan brown bears.
• To create a comprehensive dataset of visually identified brown bears for training and testing AI models.
• To assess the potential of AI-driven ReID for wildlife research and conservation.

Main Methods:

• Curated a novel dataset of 72,940 high-resolution images of 109 individual Alaskan brown bears.
• Developed a pose-aware metric-learning-based AI model (PoseSwin) leveraging biometric information.
• Tested the model's accuracy in re-identifying individuals across seasons and varying conditions.

Main Results:

• The AI model achieved promising accuracy in re-identifying individual brown bears across multiple years.
• The model successfully detected unknown individuals in a real-world, open dataset.
• Demonstrated the feasibility of non-invasive ReID for challenging unmarked species.

Conclusions:

• PoseSwin offers a viable, non-invasive approach for individual brown bear re-identification.
• This AI-driven method can significantly advance wildlife research and conservation strategies.
• The approach shows potential for application to other unmarked species, expanding ecological research possibilities.