TSM-YOLO: A deep learning framework for emergence detection and localization of tomato tray seedlings in greenhouse nurseries

Accurate detection and counting of emergence-stage tray seedlings are essential for intelligent greenhouse nursery management. However, this task is challenging due to the tiny size, dense distribution, subtle features, and susceptibility to complex backgrounds of seedlings. In this study, an improved YOLOv11-based framework, called Tray Seedling Monitoring YOLO (TSM-YOLO), is introduced for precise greenhouse tray seedling detection and localization. TSM-YOLO incorporates a Spatial Reorganization Downsampling Convolution (SRDC) for preserving spatial detail during downsampling, a Multi-branch Perception Attention (MPA) block to strengthen local feature discrimination, and a Bidirectional Auxiliary Feature Pyramid Network (BAFPN) for effective multi-scale feature fusion. Furthermore, a composite regression loss blending WIoU v3 and Normalized Wasserstein Distance (NWD) is used to stabilize the localization of small, densely packed targets. Experiments demonstrate that TSM-YOLO achieves 97.82% precision and 96.13% mAP@0.5, surpassing classic models. The model maintains high robustness under varying lighting and dense vermiculite conditions. Deployed on Jetson Xavier NX, TSM-YOLO sustains 96.41% precision and strong counting consistency. These results confirm TSM-YOLO as an effective solution for automated emergence monitoring in greenhouse nurseries.