Real-World Evaluation of YOLOv4–YOLOv8 for Agronomic Feature Detection in Cotton Using the CBD-750 Dataset

Abstract

This research presents a comprehensive object detection framework, PE-ACOD (Productivity Enhancement in Agriculture Crops by Object Detection), designed to optimize cotton crop productivity using real-time computer vision techniques. This work rigorously implements and evaluates multiple YOLO-based object detection models—including pre-YOLOv5 (YOLOv4), YOLOv5, and post-YOLOv5 models (YOLOv6, YOLOv7, YOLOv8)—on a real-world Dataset: Cotton Boll Dataset (CBD-750) collected from five cotton-growing regions in Telangana, India. Each model was trained, validated, and deployed to detect key agronomic indicators such as boll maturity, foliar stress, and weed intrusion. Deployment was validated using edge devices like Intel NCS2. Comparative results indicate that YOLOv8 achieved the highest detection performance (mAP@0.5 = 96.1%), while YOLOv5s offered the best balance between speed (35.4 FPS), size (14.2 MB), and deployment readiness. This implementation-focused study highlights how successive YOLO versions impact precision agriculture outcomes, offering insights for selecting optimal models in constrained farm environments.