Cotton Leaf Disease Detection Dataset (6 classes, 4,173 images)

Dataset accompanying the paper "SCD-YOLO: A Lightweight Model for Cotton Leaf Disease Detection and an Adaptive Localization Loss for Training". CONTENTS4,173 images with 6,012 YOLO-format bounding-box annotations across six classes: blight, curl, grey_mildew, healthy, leaf_spot, wilt. Split 8:1:1 into 3,338 / 417 / 418 images (train / val / test). The split is random at the image level and is provided as-is so that the reported results can be reproduced exactly. Per-class instance counts (train / val / test): blight 2327/282/249, curl 725/80/110, grey_mildew 149/13/27, healthy 379/42/48, leaf_spot 442/57/64, wilt 784/117/117. The class order in data.yaml is load-bearing: it indexes the per-class upper bound of the Cotton-IoU loss described in the paper. SOURCESThe dataset merges two subsets, distributed here as a single collection. 1. 3,746 images (89.8%) curated from a publicly available cotton disease dataset released on Roboflow Universe under CC BY 4.0. The original release contained 5,400 images; 1,654 low-quality samples (near-duplicates, out-of-focus images, abnormal exposure) were removed. Screening was based solely on image quality and redundancy, not on model outputs or annotation content. 2. 427 images (10.2%) collected and annotated by the authors at the cotton experimental base in Taonan City, Jilin Province, China. Images were captured under natural light from the flowering and boll-forming stage to the boll-opening stage with the main camera of a 48-megapixel smartphone. Annotation was carried out by the authors under the guidance of provincial-level agricultural experts, with class definitions consistent with subset 1. LICENCEReleased under CC BY 4.0, matching the licence of subset 1. Please cite both this record and the original Roboflow Universe dataset. CODEhttps://github.com/yichenwang-jluct/SCD-YOLOhttps://doi.org/10.5281/zenodo.22057957