OPTIFARM: Benchmarking YOLO Architectures for Location-Robust Potato Quality Detection.

Abstract

Potato sorting in post-harvest processing relies heavily on manual visual inspection, which is physically demanding, subjective, and insufficiently scalable for modern packing lines. This study investigates the feasibility of a low-cost RGB-based optical inspection system for automated potato quality detection using deep learning-based object detection. A controlled imaging platform was constructed using commodity hardware, and a dataset of 19,805 manually annotated instances across 1361 images was collected from two geographically distinct farm locations in Slovenia. A systematic benchmark of 25 model configurations spanning five YOLO architecture families-YOLOv8, YOLOv9, YOLOv10, YOLOv11, and YOLO26-was conducted across three practical quality classes (Edible, Feed, Rotten) using a strict cross-location evaluation protocol in which models were trained on one location and tested on a completely unseen second location. All models achieved strong in-distribution performance (F1 ≥ 0.906), but showed considerable variation under cross-location conditions, with external F1 ranging from 0.792 to 0.918. The yolo26_l configuration achieved the best cross-location performance (F1 = 0.918, mAP@0.5:0.95 = 0.816, ΔF1 = 0.029), demonstrating that transferable representations are achievable under a standard supervised training protocol. Per-class analysis identified feed detection as the primary generalization bottleneck. The results confirm that affordable RGB-based sorting systems are technically feasible and highlight cross-location evaluation as an essential protocol for assessing real-world deployment readiness.

Plant phenotyping relevance

ジャガイモ塊茎の品質・腐敗状態をRGB画像と物体検出で推定する撮像システムを構築し、複数YOLOモデルの交差地域ベンチマークと外部検証を行っており、表現型取得法が中心である。

Code and data availability

The paper's core potato image dataset (1361 images, 19,805 annotated instances from two Slovenian locations) is only available from the corresponding author upon reasonable request, so the primary phenotyping asset is not publicly actionable. The MDPI supplementary materials (training configurations, per-epoch metrics,

No evidence-backed public reproduction asset is currently recorded.