Plant Disease Object Detection on PlantDoc Using YOLO26n

Abstract

Plant disease recognition from images is often reported as classification, although field inspection needs more than a class label. A farmer or agricultural officer must also know where the suspected leaf or disease region appears. This paper examines that localization problem using YOLO26n on the PlantDoc object detection dataset. PlantDoc is not a clean laboratory leaf dataset. It contains outdoor images with background clutter, uneven illumination, different leaf poses, overlapping objects, and visible symptom variation. YOLO26n was trained for 50 epochs with 416 × 416 input size and batch size 16. On the test set, the model obtained 0.534 precision, 0.560 recall, 0.547 F1-score, 0.573 mAP@0.50, and 0.417 mAP@0.50:0.95. Compared with the original PlantDoc detection benchmark, mAP@0.50 increased from 0.389 to 0.573. This result shows that a recent lightweight YOLO detector can improve object-level localization on PlantDoc. At the same time, the lower mAP@0.50:0.95 shows that precise bounding-box placement is still difficult. Most errors appear in visually similar symptoms, overlapping leaves, cluttered backgrounds, and under-represented classes. Thus, YOLO26n is better positioned as an initial baseline reference than as a deployable diagnostic model. Keywords: Object detection; Plant disease; PlantDoc; YOLO26n; Deep learning.

Plant phenotyping relevance

植物画像から病変領域を検出・局在化する手法をYOLOで評価し、ベンチマーク比較と性能検証を行っているため、病害状態の画像ベース表現型計測が中心です。

Code and data availability

The paper trains YOLO26n on the PlantDoc dataset but provides no authors' public code, trained checkpoints, or dataset deposit. PlantDoc is cited prior work ([20]), and the Ultralytics documentation URL is a generic framework reference, not a paper-specific asset. No availability statement for the authors' own analysis

No evidence-backed public reproduction asset is currently recorded.