Fecha de publicación: -- Fuente: PubMed "Tobacco production" Front Plant Sci. 2026 Oct 7;17:1930225. doi: 10.3389/fpls.2026.1930225. eCollection 2026.ABSTRACTAccurate stem point localization is critical for intelligent tobacco production but remains challenging in low-altitude nadir-view UAV imagery. This is because the visible upper stem after topping is small, weakly textured, and visually similar to surrounding canopy tissues. This study formulates stem point localization as a topology-guided geometric inference problem based on canopy and leaf vein structures. A lightweight instance segmentation model, YOLO-ODS, was developed on YOLOv11n-Seg to jointly segment tobacco canopies and leaf veins from UAV imagery. The model integrates an Orientation-Aware Convolution Module (OACM), a Dual Interactive Attention Module (DIAM), and Slide Loss to enhance directional feature representation and improve discrimination of fine-grained structures. Based on segmentation results, leaf vein skeletons are extracted and local trend lines are fitted to estimate stem points via geometric convergence under canopy constraints. Experimental results showed that YOLO-ODS achieved a Recall of 86.79%, a mask mAP@0.5 of 90.19%, and a mask mAP@0.5:0.95 of 67.20%, outperforming the YOLOv11n-Seg baseline by 2.20, 1.14, and 2.10 percentage points, respectively. For the leaf vein category, the mask mAP@0.5:0.95 reached 51.49%. With 2.86 M parameters and 10.19 GFLOPs, YOLO-ODS remained lightweight for field deployment. The proposed localization method achieved an MAE of 12.05 px and an RMSE of 12.99 px, reducing MAE and RMSE by 51.2% and 52.6%, respectively, compared with YOLOv11n-pose. These results suggest that integrating lightweight instance segmentation with leaf vein topology may provide a promising approach for improving stem point localization under complex field conditions.PMID:42909929 | PMC:PMC13685701 | DOI:10.3389/fpls.2026.1930225