YOLOWeeds: A novel benchmark of YOLO object detectors for multi-class weed detection in cotton production systems

Abstract

Weeds are among the major threats to cotton production. Overreliance on herbicides for weed control has accelerated the evolution of herbicide-resistance in weeds and caused increasing concerns about environments, food safety and human health. Machine vision systems for automated/robotic weeding have received growing interest towards the realization of integrated, sustainable weed management. However, in the presence of unstructured field environments and significant biological variability of weeds, it remains a serious challenge to develop reliable weed identification and detection systems. A promising solution to address this challenge are the development of arge-scale, annotated image datasets of weeds specific to cropping systems and data-driven AI (artificial intelligence) models for weed detection. Among various deep learning architectures, a diversity of YOLO (You Only Look Once) detectors is well-suited for real-time application and has enjoyed great popularity for generic object detection. This study presents a new dataset (CottoWeedDet12) of weeds important to cotton production in the southern United States (U.S.); it consists of 5648 images of 12 weed classes with a total of 9370 bounding box annotations, collected under natural light conditions and at varied weed growth stages in cotton fields. A novel, comprehensive benchmark of 25 state-of-the-art YOLO object detectors of seven versions including YOLOv3, YOLOv4, Scaled-YOLOv4, YOLOR and YOLOv5, YOLOv6 and YOLOv7, has been established for weed detection on the dataset. Evaluated through the Monte-Caro cross validation with 5 replications, the detection accuracy in terms of mAP@0.5 ranged from 88.14 % by YOLOv3-tiny to 95.22 % by YOLOv4, and the accuracy in terms of mAP@[0.5:0.95] ranged from 68.18 % by YOLOv3-tiny to 89.72 % by Scaled-YOLOv4. All the YOLO models especially YOLOv5n and YOLOv5s have shown great potential for real-time weed detection, and data augmentation could increase weed detection accuracy. Both the weed detection dataset22https://doi.org/10.5281/zenodo.7535814 and software program codes for model benchmarking in this study are publicly available33https://github.com/DongChen06/DCW, which will be to be valuable resources for promoting future research on big data and AI-empowered weed detection and control for cotton and potentially other crops.