Proposal for Computationally Efficient Fog Computing System for Coffee Berry Borer Detection via Optimized YOLOv26

Area of Science:

• Agricultural Science
• Computer Vision
• Pest Management

Background:

• The Coffee Berry Borer is a major threat to global Coffea arabica production.
• Early detection of infestation is difficult due to small symptom size and complex field conditions.

Purpose of the Study:

• To evaluate optimized object detection architectures for improved accuracy and computational efficiency in identifying Coffee Berry Borer damage.
• To find the best balance between detection accuracy and computational cost for pest identification.

Main Methods:

• Established baseline models: YOLOv8n (M0), YOLOv11n (M1), and YOLOv26n (M2).
• Developed seven variants (M3-M9) by integrating FasterNet, SimSPPF, and EMA.
• Utilized Pareto analysis to determine the optimal model configuration.

Main Results:

• Model M0 showed the highest detection accuracy (mAP@0.5 = 0.9534).
• Model M6 (FasterNet + SimSPPF) achieved the best accuracy-efficiency trade-off (mAP@0.5 = 0.9446, 5.12 GFLOPs).
• In situ validation demonstrated a mean F1-score of 0.7255 for detecting infected berries, even with shadows.

Conclusions:

• Model M6 is the optimal configuration for detecting Coffee Berry Borer infestation.
• The developed models are suitable for real-time agricultural deployment in pest management.
• Optimized object detection significantly enhances the ability to manage coffee crop pests.