Reproducible YOLO11m Baseline for Paddy Leaf Disease Object Detection: Repairing a Defective Dataset Split and Documenting Per-Class Performance

Published

11 September 2026

Venue

IEEE

DOI

—

Abstract

Automated detection of paddy (Oryza sativa) leaf diseases can support timely field intervention, but public datasets for this task are not always shipped in a directly usable form. This paper documents a reproducible object-detection pipeline built on a 480-image, 6-class Roboflow paddy-disease dataset whose raw export placed all annotated pairs under a single partition and provided only placeholder numeric class labels. We first validate every annotation, then construct a seeded (42), dominant-class-approximate-stratified 70/15/15 train/validation/test split with a post-split integrity check, and fine-tune a COCO-pretrained YOLO11m detector (80 epochs, patience 15, image size 640, AdamW) on the resulting train/validation partitions. Evaluated once on the held-out test split (72 images, 257 instances), the model reaches precision 0.570, recall 0.597, mAP50 0.585, and mAP50-95 0.269, at 46.65 ms/image on a Tesla T4 GPU. Per-class analysis shows AP50 ranging from 0.342 to 0.878 across the six classes, with the numerically dominant class exhibiting the lowest AP50 despite having the most training examples, an effect traced to a high rate of missed (background-classified) detections in the confusion matrix. We report this as a documented, reproducible baseline rather than a state-of-the-art claim, given open questions about class semantics and possible image-level leakage that the available metadata cannot rule out.