Missed transplanting rate evaluation method for tobacco seedling transplanter based on UAV imagery and improved YOLOv5s

Missed transplanting remains a significant challenge in the application of tobacco seedling transplanters due to the specific agronomic requirements for successful transplanting. Currently, the detection of missed transplanting rate in large-scale field tests primarily relies on manual seedling counting, a process that is notably inefficient. Traditional online detection methods, including photoelectric sensors and machine vision, suffer from problems such as complex structures and high costs. They require sensor deployment on the machine itself, making it difficult to fully meet the actual detection needs of transplanters during the R&D and testing phase. To address these limitations, this paper proposes an automated evaluation method for detecting missed transplanting rates using UAV (unmanned aerial vehicle) imagery. The method integrates an improved YOLOv5s model, DeepSORT, and line-crossing counting approach. First, a second-order channel attention (SOCA) attention mechanism was incorporated into the YOLOv5s model to improve its ability to extract features for small targets. Additionally, the Spatial Pyramid Pooling Fast (SPPF) was replaced by the Simplified Spatial Pyramid Pooling-Fast (SimSPPF) to enhance the model's ability to extract multi-scale features for targets such as seedling-planted holes. The DeepSORT algorithm, combined with line-crossing counting principle, was then employed for visual tracking and dynamic counting of seedling-planted and missed-planting holes, enabling accurate evaluation of the missed transplanting rate. Test results showed that, in terms of target detection, the Precision and mAP of the improved YOLOv5s model increased by 3.9% and 5.3%, respectively, compared to the original YOLOv5s. In target tracking, the combination of the improved YOLOv5s and DeepSORT reduced the missed detection rate <i>M_m</i> and false detection rate <i>M_f</i> by 2.5% and 6.1%, respectively. Field experiments achieved an accuracy of 90.28% for the missed transplanting rate and a 10× higher detection efficiency compared to manual inspection. This method offers a novel automated solution for the rapid detection of missed transplanting rates in large-scale transplanting operations and provides valuable technical insights for evaluating the performance of other seedling transplanters.