Keypoint detection-based localization of weed growth points for laser weeding in sugarcane fields
Smart Agricultural Technology · 2026-12 · 15 · 102686
DOI: 10.1016/j.atech.2026.102686 ↗Kurz gefasst
Fachartikel in der Zeitschrift Smart Agricultural Technology, 2026-12. Verfasst von Yiqi Huang und 5 weiteren Autor:innen. Die Arbeit nennt 42 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Shenzhen Science and Technology Innovation Commission, Science and Technology Planning Project of Shenzhen Municipality, National Natural Science Foundation of China.
Abstract (Originalsprache)
Efficient and precise weed management is essential for modern sugarcane production. To support precision laser weeding, this study proposes MEL-YOLO, a lightweight keypoint detection model based on YOLOv8s-Pose for simultaneous crop–weed detection and weed growth point localization during the sugarcane tillering stage. To address accuracy degradation caused by dense weed distribution, mutual occlusion, and illumination variation, as well as excessive computational complexity under limited computing resources, three improvements were introduced. First, a Mixed Local Channel Attention (MLCA) mechanism was embedded in the backbone to enhance weak-feature extraction and interference suppression. Second, an Efficient Multi-Scale Convolution with Position-aware Pooling (EMSCP) module was adopted in the neck to improve multi-scale feature fusion while maintaining a lightweight structure. Third, a lightweight LSCD-LQE detection head incorporating localization quality estimation was designed to improve the consistency between prediction confidence and localization quality. A custom annotated dataset containing sugarcane seedlings and five common weed species in Guangxi sugarcane fields was constructed to train and evaluate the model. Ablation and comparative experiments showed that MEL-YOLO achieved an mAPkpt@0.5 of 94.31% ± 0.09%, significantly outperforming YOLOv8s-Pose by 3.23 percentage points (p < 0.001). The model size was reduced to 17.60 MB, and the inference speed reached 186.73 ± 1.83 FPS on the test platform. Qualitative results showed stable performance under dense weed distribution, severe occlusion, and complex illumination. These results indicate that MEL-YOLO has potential to provide reliable visual perception support for precision laser weeding in sugarcane fields.
Förderung
- Shenzhen Science and Technology Innovation Commission
- Science and Technology Planning Project of Shenzhen Municipality
- National Natural Science Foundation of China
- The Agricultural Science and Technology Innovation Program
Zitieren
APA
Huang, Y., Tang, C., Zou, M., Chen, J., Qian, W., & Qiao, X. (2026). Keypoint detection-based localization of weed growth points for laser weeding in sugarcane fields. Smart Agricultural Technology, 15, 102686. https://doi.org/10.1016/j.atech.2026.102686
Vancouver
Huang Y, Tang C, Zou M, Chen J, Qian W, Qiao X. Keypoint detection-based localization of weed growth points for laser weeding in sugarcane fields. Smart Agricultural Technology. 2026;15:102686. doi:10.1016/j.atech.2026.102686
BibTeX
@article{Huang2026,
author = {Huang, Yiqi and Tang, Chao and Zou, Mingfeng and Chen, Jiahao and Qian, Wanqiang and Qiao, Xi},
title = {Keypoint detection-based localization of weed growth points for laser weeding in sugarcane fields},
journal = {Smart Agricultural Technology},
year = {2026},
volume = {15},
pages = {102686},
doi = {10.1016/j.atech.2026.102686}
}
EUVIMED