Winter wheat weed detection based on deep learning models

Li，Zhuolin1; Wang，Dashuai1; Yan，Qing2; Zhao，Minghu1; Wu，Xiaohu1; Liu，Xiaoguang1

2024-12-01

10.1016/j.compag.2024.109448

Computers and Electronics in Agriculture   影响因子和分区

0168-1699

Weeds seriously threaten winter wheat yields. Timely and effective weed control is crucial in modern precision agriculture. However, the diversity and morphological differences of weeds, and visual similarities between symbiotic weeds and crops pose significant challenges to weed detection. With the development of deep learning, crop detection has accomplished remarkable achievements in precision agriculture. Introducing the state-of-the-art (SOTA) deep learning algorithms into the winter wheat weed detection task is expected to provide a meaningful decision-making basis for precision weed management. However, current deep learning-based winter wheat weed detection methods still suffer severe challenges in terms of accuracy and speed. This can be mainly attributed to two aspects: the lack of high-quality and large-capacity datasets and the insufficient adaptability of detection models to specific datasets. This study is dedicated to filling the gap in winter wheat weed detection research from both the aspects of datasets and deep learning algorithms. Firstly, we present a new high-quality winter wheat weed dataset (3W dataset), which contains 9,278 manually annotated RGB images of 8 common winter wheat weed species. Then, we verify the feasibility of the 3W dataset by reproducing some SOTA models, including R-CNN, YOLO, and Transformer series. Based on the comparison results, we used the YOLOv8 model with faster inference speed and the DINO model with the best detection accuracy as the baselines and introduced some mainstream improving strategies such as spatial attention mechanism (SAM), Non-local block (NLB) and deformable convolution network v2 (DCNv2) to further improve their performance. The results of ablation experiments show that NLB can effectively improve the performance of YOLOv8 and DINO by enhancing their global modeling capabilities. Compared with the original YOLOv8, the mAP of the YOLOv8_NLB model is increased from 85.8 % to 88.3 %, and the inference speed is 84.6 FPS. The mAP of our improved DINO_NLB increased from 88.6 % to 89.4 %, occupying the highest detection accuracy. However, its detection speed is only about one-third of YOLOv8_NLB. To sum up, for weed detection tasks with high-speed requirements, it is recommended to use the YOLOv8_NLB model, and for scenarios that pursue high accuracy, the DINO_NLB model is an ideal choice. Finally, we deployed the faster YOLOv8_NLB model on the TensorRT format and quantized it in Jetson Xavier NX. The TensorRT_FP16 model achieves a better balance between speed and accuracy, with an inference speed of 40.5 FPS and an mAP of 86.7 %, which is more suitable for real-world weeding tasks.

Deep learning DINO Non-local block (NLB) Winter wheat weed YOLOv8

英语

第一 ; 通讯

2-s2.0-85205426157

Scopus

1.School of Microelectronics,Southern University of Science and Technology,Shenzhen,518055,China
2.School of Mechanical & Automotive Engineering,Liaocheng University,Liaocheng,252000,China

Li，Zhuolin,Wang，Dashuai,Yan，Qing,et al. Winter wheat weed detection based on deep learning models[J]. Computers and Electronics in Agriculture,2024,227.

Li，Zhuolin,Wang，Dashuai,Yan，Qing,Zhao，Minghu,Wu，Xiaohu,&Liu，Xiaoguang.(2024).Winter wheat weed detection based on deep learning models.Computers and Electronics in Agriculture,227.

Li，Zhuolin,et al."Winter wheat weed detection based on deep learning models".Computers and Electronics in Agriculture 227(2024).