NUST Institutions Library Catalogue

Cotton sustains the textile economies of 80+ countries with annual yield of 113 million
bales, yet 30–60 % yield losses are still attributed to viral, bacterial and fungal diseases
that remain undetected until manual field scouting. To address this critical issue, the
use of Unmanned-aerial-vehicle (UAV) imagery combined with deep learning offers a
scalable early-warning system, but existing models suffer from low early-stage
accuracy, poor transferability and high computational cost. In this paper, we propose
Cotton-Vision, a fully convolutional single-stage detector that synergizes three lightweight architectural innovations. The architectural changes include ghost bottleneck
modules for 20 % parameter reduction, efficient channel-attention (ECA) for adaptive
feature recalibration, and cross-stage partial (CSP) fusion for enriched gradient flow.
The network is trained on local datasets of cotton collected from 7 different fields and
comprised 2700+ drone images from local cotton farms in Punjab Province of Pakistan.
Extensive augmentation (rotation, shear, exposure, saturation) yields 11704 training
samples. On a held-out test set, Cotton-vision achieved a precision of 97.7%, a recall
of 77.7%, a mAP@50 of 85.6%, and a mAP@50-95 of 76.6%. These results represent
10-20% improvement in precision, recall and mAP over the leading models, including
YOLOv5, YOLOv8, YOLOv11, YOLOv12, RT-DETR and RF-DTR. Through a
comprehensive ablation study, we demonstrate that each of our proposed architectural
enhancements contributes to the model's superior performance with 12%-16%
improvement from the base model. Our findings confirm that this synergy of
lightweight and attention-based modules provides a robust and computationally viable
solution for early cotton disease detection.