Long Zhao, Hui Wang, Dongling Wu, Jianjun Liu, Zhe Li, Yi Shi
Missed transplants, lodged seedlings, and weak seedlings may occur during mechanized transplanting or shortly afterward, making timely post-transplant inspection essential. Plastic mulching is widely used after tobacco transplanting to improve early growth conditions. However, occlusion by the mulch film and reflections from the film surface make tobacco seedling detection more challenging in plastic-mulched tobacco fields. In this study, multispectral images of plastic-mulched tobacco fields were collected using an unmanned aerial vehicle (UAV). The normalized difference vegetation index (NDVI), green normalized difference vegetation index (GNDVI), and optimized soil-adjusted vegetation index (OSAVI) were calculated and assigned to the RGB channels in different combinations, yielding nine candidate input configurations, including six multi-index channel assignments and three single-index configurations. Screening results showed that the GNDVI-OSAVI-NDVI channel configuration enabled YOLOv11 to achieve the best overall balance of detection performance, with the highest precision, F1 score, and mAP@0.5 among the nine input configurations. Based on this optimal input strategy, depthwise separable convolution (DWSConv) and parallelized patch-aware attention (PPA) were incorporated, and the resulting model was referred to as YOLOv11-PD. Ablation experiments showed that PPA improved the main detection metrics, whereas DWSConv reduced the number of parameters and inference latency. YOLOv11-PD achieved a favorable balance between detection accuracy and computational efficiency, obtaining the highest recall, F1 score, and mAP@0.5 among all evaluated models, with values of 83.9%, 83.7%, and 0.876, respectively, while maintaining a precision of 83.5%. These results provide technical support for early-stage tobacco seedling detection under plastic mulching, evaluation of transplanting quality, and decisions on replanting missing seedlings.