DeepFaune New England: A Species Classification Model for Trail Camera Images in Northeastern North America

Area of Science:

• Ecology
• Computer Science
• Wildlife Biology

Background:

• Automated wildlife identification from camera trap images is crucial for ecological monitoring.
• Existing models often lack regional specificity, requiring adaptation for new geographic areas.

Purpose of the Study:

• To develop and validate an automated wildlife classification model for northeastern North America.
• To assess the effectiveness of transfer learning for adapting a European wildlife model to a new continent.

Main Methods:

• Adaptation of the existing DeepFaune model (originally for European wildlife) for North American species.
• Training and validation using trail camera images from northeastern North America.
• Leveraging open-source software and openly licensed data for accessibility.

Main Results:

• The DeepFaune New England model achieved high accuracy (97%) in identifying 24 wildlife taxa.
• Demonstrated successful transfer learning across continents, adapting a European model for North American wildlife.
• The model and its software are open source, facilitating integration into existing camera trapping workflows.

Conclusions:

• The DeepFaune New England model provides an accurate and accessible tool for wildlife monitoring in North America.
• Transfer learning is a practical approach for developing specialized wildlife identification models efficiently.
• Open-source availability promotes wider adoption and integration in ecological research.