Two-Stage Wildlife Event Classification for Edge Deployment.

Area of Science:

• Ecology
• Computer Science
• Artificial Intelligence

Background:

• Camera-based wildlife monitoring faces challenges with non-target triggers and slow manual review, hindering timely intervention in human-wildlife conflicts.
• Cloud-dependent inference limits real-time processing, especially in connectivity-limited environments.

Purpose of the Study:

• To develop a deployable, fully offline edge vision sensor for near-real-time, accurate wildlife event classification.
• To improve the efficiency and reliability of wildlife monitoring systems for human-wildlife conflict management.

Main Methods:

• A two-stage approach combining a You Only Look Once (YOLO)-family detector for empty-image suppression and an EfficientNet-based classifier for puma confirmation.
• Utilizing staged transfer learning and robust design for low-quality, nighttime monochrome imagery.
• Field deployment and ablation studies to evaluate performance and adaptability.

Main Results:

• Achieved near-real-time classification with end-to-end latency of approximately 4 seconds.
• Demonstrated high performance on a held-out test set: precision 0.983, recall 0.975, F1 0.979, accuracy 0.986.
• Substantially reduced false alarms compared to full-image classifiers while maintaining high recall; adaptable to other species like ringtails.

Conclusions:

• The proposed offline edge vision sensor effectively addresses limitations of current wildlife monitoring systems.
• The two-stage classification approach offers a robust, accurate, and efficient solution for high-stakes human-wildlife conflict intervention.
• The system's adaptability and real-time capabilities provide flexible actuation for various ecological and conservation applications.