A Temporal Boosted YOLO-Based Model for Birds Detection around Wind Farms

Area of Science:

• Computer Vision
• Artificial Intelligence
• Environmental Science

Background:

• Object detection in sky surveillance is difficult due to small objects and dynamic backgrounds.
• Detecting birds near wind turbines is crucial for preventing collisions and reducing avian mortality.

Purpose of the Study:

• To develop and evaluate a YOLOv4-based ensemble model for accurate bird detection in wind farm environments.
• To improve small bird detection using techniques like tiling and temporal stacking.

Main Methods:

• Utilized two novel datasets (Klim and Skagen) from Denmark for training and testing.
• Developed three YOLOv4-based models, progressively incorporating tiling and temporal stacking.
• Implemented an ensemble detector combining the best-performing models.

Main Results:

• The final ensemble model achieved high mean Average Precision (mAP) values: 90% on the Klim dataset and 92% on the Skagen dataset.
• Individual models showed significant improvements with advanced techniques, reaching up to 90% mAP on the Klim dataset.
• Model 1 achieved 82% mAP on Klim and 60% mAP on Skagen, while Model 3 reached 90% mAP on Klim and 92% on Skagen.

Conclusions:

• The proposed YOLOv4 ensemble model effectively enhances bird detection accuracy in challenging wind farm conditions.
• Improved bird detection can inform wind farm siting and turbine placement, thereby mitigating bird mortality.
• The system offers potential for enhancing collision avoidance systems in renewable energy facilities.