Green AI for Sustainable Agriculture: Benchmarking Energy Efficiency and Accuracy Trade-Offs of Lightweight YOLO Models in Banana Quality Grading

Post-harvest quality grading is critical for reducing losses in primary banana production. However, the adoption of artificial intelligence at the farm level is often constrained by limited energy and computational infrastructure. This study evaluates whether reliable banana quality grading can be achieved under such constraints by systematically com-paring lightweight object detection models. Four YOLO-based architectures were bench-marked using a curated dataset of over 11,000 annotated images across six commercial quality classes. Energy consumption and carbon emissions during model training and inference were quantified using CodeCarbon, while detection errors were diagnosed using the TIDE framework to assess practical visual grading limitations. The results reveal that legacy compact models do not inherently guarantee energy efficiency. Instead, modern lightweight architectures achieve a superior balance between spatial accuracy and energy use. Error analysis indicates that grading reliability is primarily challenged by localiza-tion errors in deteriorated fruit rather than class confusion. Ultimately, the environmental cost of model training is marginal compared to the potential reduction in post-harvest waste. These findings highlight that energy-aware model selection is essential for deploy-ing sustainable computer vision solutions in resource-constrained agricultural systems.