ENERGY-AWARE BENCHMARKING OF LIGHTWEIGHT YOLO MODELS FOR SUSTAINABLE BANANA QUALITY GRADING

Main Article Content

NGOC BAO HO
NGUYET THI MINH NGUYEN

Abstract

Post-harvest losses present a critical challenge in global agricultural supply chains, where inconsistent manual grading drives significant food waste and economic instability. While AI-based computer vision offers a promising automated solution, the environmental cost of deploying deep learning models is frequently ignored. Lightweight networks are often widely considered to be energy-efficient, yet this requires strict empirical validation. This study investigates the intersection of accuracy and sustainability by benchmarking four YOLO architectures on over 11,000 annotated banana images. Our empirical results challenge the assumption that legacy compact models guarantee environmental efficiency: YOLOv3-tiny consumed the highest energy (0.2036 kWh) while delivering the lowest accuracy (mAP50-95 = 0.6418). Conversely, modern nano-scale architectures achieve an improved balance. YOLOv8n minimized energy consumption to just 0.1711 kWh while maintaining strong accuracy (mAP50-95 = 0.7084), and YOLO11n maximized precision (mAP50-95 = 0.7112) with only a modest computational increase. Crucially, cost-benefit analysis reveals that the carbon emissions generated during the training of these optimized models are minimal compared to the considerable greenhouse gases avoided by drastically reducing post-harvest fruit waste. This evidence-based framework demonstrates that deploying Green AI in agriculture can secure high productivity while strictly ensuring long-term ecological preservation.

Article Details

Section
Chemical, Bio, Food, Environmental Technology