YOLOv8-ECFS: A lightweight model for weed species detection in soybean fields

Abstract

Precision agriculture technology has become a crucial means of improving the quality of crop production. As an emerging technology in farmland management, intelligent weeding robots utilize intelligent spraying systems to effectively manage weeds, adjusting the types and dosages of herbicides in a timely manner. The accuracy and real-time performance of weed identification algorithms are the keys to intelligent weeding. This study established a proprietary dataset comprising 6690 images of soybean seedlings and weeds and proposed an improved lightweight algorithm, YOLOv8-ECFS. Based on YOLOv8s, this model introduces the EfficientNet network to improve feature extraction capability and accelerate the inference speed, replacing the CIoU loss with Focal_SIoU to optimize the regression accuracy of the bounding boxes. Furthermore, the coordinate attention module is introduced into the neck to enable the model to precisely capture textural and color differences between various weeds and soybean crops, thereby ensuring precise identification of multiple weed species. The results demonstrate that YOLOv8-ECFS achieves precision, mAP, and F1 values of 92.2%, 95.0%, and 90.9%, representing an increase of 2.5%, 1.3%, and 1.6%, respectively, compared to YOLOv8s. Simultaneously, the model's GFLOPs and model size have been reduced by 11.1G and 9.1 MB, respectively, ensuring both recognition accuracy and lightweight performance. The test set results show that YOLOv8-ECFS accurately identifies densely growing and mutually occluding weeds, reducing cases of false positives and missed detections. Compared to other mainstream YOLO algorithms, YOLOv8-ECFS demonstrates the best overall performance, thus providing support for intelligent weeding robots in farmland management and unmanned farms.