Classification of Damaged Potatoes Using YOLO11

Abstract

Post-harvest sorting of potatoes is a key step in ensuring quality, preventing degradation of stored stocks and reducing subse-quent losses. Traditional manual sorting methods are laborious, subjec-tive and inconsistent, especially in large operations. Automatic detection of potato multi-type defects remains a challenge because of the diversifi-cation in defect size and visual similarity among multi-type defects. This study proposes a comprehensive potato sorting framework based on the YOLO11 object detection family, explicitly targeting real-world conveyor sorting scenarios. A novel mixed-condition dataset was constructed, com-bining both cleaned (washed) and uncleaned (harvest-state) potatoes captured under continuous conveyor motion and fixed industrial light-ing. The dataset includes four practically relevant classes: healthy pota-toes, defective potatoes, stones with soil residues, and plant debris. Unlike previous studies that focus on a single model configuration, all YOLO11 variants were systematically evaluated to analyze the trade-offs between detection accuracy, model complexity, and deployment feasibil-ity. The proposed approach achieves consistently high performance across both clean and unclean test sets, with macro F1-scores exceeding 0.96. The YOLO11-medium model demonstrated the best balance between accuracy and computational e fficiency, reaching F1-scores of 0.9908 on cleaned samples and 0.9752 on unwashed samples. The main contribu-tions of this work are threefold: 1) the creation of a realistic, mixed-condition post-harvest potato dataset reflecting industrial conveyor-line constraints; 2) a comparative analysis of the entire YOLO11 model fam-ily for agricultural object detection under heterogeneous visual conditions; and 3) the demonstration of a deployment ready solution that combines high accuracy with moderate computational requirements. These findings highlight the suitability of YOLO11 based systems for robust, real-time postharvest sorting and contribute to advancing practical deeplearning applications in precision agriculture.