Intelligent transplantation is often required to accurately detect the missing seedlings in the densely arranged tomato plug seedlings in the greenhouse. In this study, an improved YOLOv8n-based model (named YOLO-SFEP) was proposed to detect the missing-seedlings in greenhouse environments. Small-target detection was also improved by the illumination robustness. Computational complexity was also reduced to optimize the training efficiency. Four innovations were integrated into the framework: feature extraction refinement, attention-driven interference suppression, network lightweighting, and adaptive optimization of the loss function. Extensive experiments were carried out to validate the practical deployment. Firstly, the P3 detection layer of YOLOv8n was reconstructed to integrate the SPDConv (space-to-depth convolution) into an Omni-Kernel module. Dual enhancement strengthened the feature extraction for small targets, effectively improving the detection accuracy in dense seedling arrangements. Secondly, an EMA (efficient multi-scale attention) mechanism was embedded into the neck network, in order to amplify the critical features of missing seedlings, while suppressing interference from uneven lighting conditions in greenhouses. Thirdly, the lightweight deployment was achieved to replace the Bottleneck modules in C2f (cross-stage partial bottleneck with two convolutions) layers with FasterNet Block modules from FasterNet. Redundant parameters were then reduced without compromising recognition accuracy, indicating compatibility with the resource-constrained devices. Finally, the CIoU (complete intersection over union) loss function was replaced by PIoU (powerful-IoU). The adaptive penalty factors were introduced into the target scales. Thereby, the convergence stability was optimized under the varying seedling sizes. Experimental results revealed that the YOLO-SFEP model attained a precision of 96.7%, a recall of 98.8%, and a mean average precision at an intersection over union threshold of 0.5 (mAP50) of 97.9%, in order to detect the missing seedlings in tomato plug trays. There was an improvement over the baseline YOLOv8n model, where the precision was enhanced by 2.4 percentage points. The parameter count and computational load of the proposed model are 2.62 million(M) and 7.6 Gigaflops(G), respectively, representing a reduction of 0.39 M and 0.5 G compared with YOLOv8n. The model size was compacted to 5.5 megabytes(MB), decreasing by 0.8 MB relative to the baseline. Importantly, the model showed high robustness across a spectrum of natural light intensities (low, medium, and high), attaining mAP50 values of 96.7%, 98.9%, and 98.1%, respectively, which underscores its superior performance in challenging illumination environments. The improved model was also deployed on an edge computing device, the Raspberry Pi 5. A detection speed of 13.8 frames per second was achieved, equal to an inference time of 72 milliseconds(ms) per image. This speed was 12.2 ms faster than that of the baseline YOLOv8n model on the same platform. These outcomes highlighted the model’s aptitude for real-time greenhouse applications, enabling it to deliver precise and prompt positional data on missing seedlings to support automated replenishment by transplantation equipment. In conclusion, the YOLO-SFEP model markedly improves the detection of missing seedlings in tomato plug trays within greenhouse environments. By integrating advanced modules for small target detection, attention mechanisms for feature enhancement, and efficient components for model lightweighting, it achieves elevated detection accuracy alongside reduced computational complexity. Its successful deployment on an edge device underscores its viability for on-site use. This research offers a solid technical foundation for intelligent missing seedling detection and automated seedling replenishment in greenhouse tomato cultivation. For further research, techniques such as knowledge distillation, quantization and pruning can be considered in the future to lighten the YOLO-SFEP model, and at the same time, analyze the resource consumption of the model in actual deployment, including memory occupation and power consumption, in order to comprehensively evaluate its applicability on edge devices, and eventually build an inspection system to realize the integration of image acquisition and processing by using edge devices.

Xinjiang, China's largest tomato production region, extensively employs plug seedling transplantation technology in modern agricultural practices. Grading and screening tomato plug seedlings to obtain qualified plants significantly enhances transplantation survival rates and crop yield consistency, making accurate pre-transplantation seedling grading critical for transplantation quality, operational efficiency, and economic viability. Traditional manual grading methods exhibit high labor intensity, time consumption, and substantial subjectivity, resulting in inconsistent results and reduced production efficiency. With growing demand for agricultural automation and precision agriculture technologies, advanced automated solutions capable of high-precision, high-reliability real-time quality assessment are urgently needed. This study proposes an innovative FAN-DETR (feature-adaptive network detection transformer) real-time grading detection model based on the advanced RT-DETR (real-time detection transformer) architecture to improve accuracy and speed of real-time tomato plug seedling grading detection. By combining complementary advantages of HGNet and CSPNet architectures, a novel lightweight hybrid backbone network with dynamic feature enhancement capabilities, designated as DFE-LHBNet (dynamic feature-enhanced lightweight hybrid backbone network), was meticulously designed to address computational redundancy issues. The HGStem module enables intelligent feature diversion and distribution, integrated with CSPNet's lightweight modular components to substantially reduce parameter count without compromising feature extraction capability. A hierarchical gating mechanism through HGBlock intelligently filters and prioritizes high-value features while suppressing redundant information. DWConv depthwise separable convolutions were integrated throughout the network to significantly reduce computational complexity while maintaining feature representation quality. An innovative DHAM (dual hierarchical attention mechanism) module was implemented for efficient feature fusion, balancing local detail preservation and global context understanding while reducing computational overhead. The model incorporates a state-of-the-art DTAB (dilated transformer attention block) attention mechanism to expand effective receptive field and capture comprehensive global contextual information crucial for precise seedling classification. A hierarchical processing pipeline effectively handles complex cross-channel and spatial information interactions through three parallel branches operating complementarily to extract diverse feature representations and enhance robustness. To optimize training efficiency and convergence characteristics, the PIoU (powerful IoU) loss function was adopted as the primary optimization criterion to prevent anchor box expansion issues and improve model training convergence speed and stability. Comprehensive experimental validation demonstrates that the proposed FAN-DETR model exhibits superior performance metrics. The system achieved an impressive mean average precision of 92.62% at IoU thresholds greater than 50%, while attaining excellent detection speed of 96.67 frames per second (FPS). Compared to the baseline RT-DETR model, these results represent significant improvements of 2.20 percentage points in accuracy and 22.82% in processing speed. The optimized architecture achieved substantial reductions of 29.59% in parameter count and 44.39% in computational complexity. Comparative analysis with mainstream YOLO series models (YOLOv5-m, YOLOv8-m, YOLOv10-m, and YOLOv12-m) revealed consistently superior performance across multiple evaluation metrics. Compared to YOLO models, FAN-DETR achieved average precision improvements of 6.77 percentage points, recall enhancements of 3.31 percentage points, and mean average precision increases of 1.50 percentage points at IoU thresholds greater than 50%. The 96.67 frames per second performance represents an average improvement of 34.73 frames per second compared to competing models, while maintaining an extremely compact model size of only 27.70 MB, demonstrating exceptional parameter compression and deployment efficiency suitable for resource-constrained agricultural environments. The FAN-DETR model exhibits exceptional performance characteristics in real-time tomato plug seedling grading detection applications, providing a highly reliable and accurate automated grading solution suitable for industrial deployment. This research provides compelling evidence for the feasibility and effectiveness of real-time plug seedling grading detection systems, contributing to advancing agricultural automation technology development and supporting intelligent agricultural solutions that improve production efficiency, reduce labor costs, and enhance crop quality consistency in tomato production operations.

Grading is one of the most important steps in the tomato hole-pan seedlings. In this study, an improved YOLOv5s target model was proposed to optimize the recognition accuracy of tomato hole-pan diseased seedlings using transfer learning. The backbone part of the lightweight network (EfficientNetv2) was used as the feature extraction. The spatial pyramid pooling fusion (SPPF) module in YOLOv5s was retained to compress the number of model parameters, in order to reduce the amount of computation; The lightweight up-sampling module of CARAFE was used to introduce a small number of parameters in the Neck part of the model; And the PANet was replaced with BiFPN to optimize the accuracy of tomato hole-pan seedling identification using transfer learning. The feature weight information was introduced to enhance the fusion of features at different scales; The efficient multi-scale attention mechanism (EMA) was introduced to increase the attention for the tomato hole-pan seedlings, and reduce the background interference; The CIoU loss function was replaced by the SIoU loss function to improve the model accuracy. Direction matching between the real and predicted frame was considered to improve the convergence of the model. The mean average accuracy of the improved YOLOv5s target model reached 95.6% after training by transfer learning, which was 0.7 percentage points higher than before; The improved YOLOv5s model was 53.1% of the number of parameters in the original model, and the computation was 20.0% of the original with only 3.20G, while the weights were 53.6% of the original and the weight size was only 7.35MB, and the mean average accuracy was improved by 2.6 percentage points, compared with the original; The visualization show that the stronger feature extraction was achieved in the improved model. The feature weights were centrally distributed in the center of the burrow and the edge of the burrow holes. Meanwhile, the heat map drawn with the GradCam showed that the improved YOLOv5s model was focused mainly on the burrow seedling itself and the edge of the burrow hole, which reduced the interference of the substrate; The ablation test verified that the improved YOLOv5s model was performed the best in the detection accuracy and the frame rate, compared with the Faster-RCNN, CenterNet, and YOLO series. The findings can provide a strong reference for the grading detection and subsequent deployment of hole tray seedlings.

Xinjiang is the largest agricultural production base in China; however, due to the region's unique climate, early sowing is not possible. The use of seedling transplanting technology has been adapted to extreme climatic conditions, and this technology has improved the survival rate of tomato seedlings. As tomato production continues to expand and agricultural modernisation advances, the efficiency of transplanting machinery is becoming an essential development requirement. The present study proposes a non-singular fast terminal sliding mode control method based on the interference observer, with the aim of enhancing the system's seedling picking up and delivery mechanism. This method addresses the challenges posed by interference and modelling errors, which have been identified as contributing factors to the system's suboptimal performance, characterised by low position control accuracy, slow response speed and stability. 1)The paper analyses the constraints on the seedling pickup and delivery mechanism, and establishes a reduced order model of seedling pickup and delivery manipulator dynamics based on the Lagrange dynamics modelling method. 2)The non-singular fast terminal sliding mode controller is designed, and the improved fal function convergence law is adopted to ensure the fast and accurate position control of the seedling pickup and delivery mechanism. 3)Due to the jitter vibration caused by the nonlinear function and friction vibration in the controller, it is proposed to introduce a nonlinear disturbance observer to achieve effective suppression of the jitter vibration and stability of the system tracking. 4)In Matlab/Simulink, the trajectory tracking simulation experiment of transplanting machine seedling pickup and delivery mechanism under interference is carried out. The results demonstrate that the position tracking of the seedling pickup and delivery mechanism linkage 1 and linkage 2 converge at 0.7 and 0.8 s, respectively, when employing a non-singular fast terminal sliding mode controller. The maximum position tracking error is 0.32 and 0.45 rad, which verifies that the sliding mode controller effectively improves the convergence speed and control accuracy of the seedling pickup and delivery mechanism. After the introduction of the interference observer, the mean squared error of position tracking of connecting rod 1 and connecting rod 2 of the seedling pickup and delivery robot is 0.00772 and 0.00824 rad, which is a decrease of 45.7% and 52.8% in the tracking error mean squared error, and it verifies that the interference observer effectively improves the stability of the control. Concurrently, the convergence time is reduced by 42.9% and 25%, and the maximum position tracking error is decreased by 25% and 11%, signifying that the interference observer enhances convergence speed and control accuracy. Finally, a test platform has been constructed in order to verify the rate of seedling pickup and delivery and error accumulation through the use of high-speed photography. The results of this study demonstrate that the maximum attitude error in the rate test of seedling pickup and delivery is 2.63°, and that the error accumulation test is within 3°, which meets the actual working requirements of seedling pickup and delivery. This study provides a reference for the automation of pickup and delivery of tomato hole tray seedling transplanter.