Detection and gradation of sweet potato storage roots by machine vision and deep learning

Real-time detection and quality grading of sweet potato storage roots remain persistent challenges during and after harvest, often relying on labor‑intensive manual inspection and resulting in substantial inefficiencies. To address these limitations, this study presents two complementary AI-driven solutions tailored for smart agricultural applications. First, we propose Ghost‑YOLOv5, a lightweight detection model optimized for real-time sweet potato identification under field harvesting conditions. The model achieves high detection accuracy (92.6% mAP) while significantly reducing computational demands, requiring only 1.28 million parameters, 2.6 GFLOPs, and 3.0 MB of memory—substantially outperforming conventional architectures in resource‑constrained environments. Second, we develop a knowledge‑distilled grading framework (YOLOv5n‑m) that integrates the representational strengths of YOLOv5m with the compactness of YOLOv5n. The distilled model achieves an overall grading accuracy of 0.922, surpassing the baseline YOLOv5n accuracy (0.897) while maintaining the same parameter count (1.76 million). Notably, the model demonstrates excellent performance in sweet potato quality grading, achieving 0.984 accuracy for Ideal-grade and 0.941 for Normal-grade roots. These results highlight the effectiveness of knowledge distillation in transferring high-level feature representations from larger teacher networks to lightweight student models without increasing computational cost. Together, the proposed detection and grading frameworks offer practical, deployable AI solutions that bridge the gap between laboratory model development and real-world agricultural implementation. By enhancing operational efficiency while minimizing computational requirements, this work contributes to sustainable smart agriculture and provides scalable benefits for sweet potato production and other root crop management systems.