Abstract
Accurate acquisition of tobacco phenotypic traits is crucial for growth monitoring, cultivar selection, and other scientific management practices. Traditional manual measurements are time-consuming and labor-intensive, making them unsuitable for large-scale, high-throughput field phenotyping. The integration of 3D reconstruction and stem-leaf segmentation techniques offers an effective approach for crop phenotypic data acquisition. In this study, we propose a tobacco phenotyping method that combines unmanned aerial vehicle (UAV) remote sensing with an improved PointNet++ model. First, a 3D point-cloud dataset of field-grown tobacco plants was generated using multi-view UAV imagery. Next, the PointNet++ architecture was enhanced by incorporating a Local Spatial Encoding (LSE) module and a Density-Aware Pooling (DAP) module to improve the accuracy of stem and leaf segmentation. Finally, based on the segmentation results, an automated pipeline was developed to compute key phenotypic traits, including plant height, leaf length, leaf width, leaf number, and internode length. Experimental results demonstrated that the improved PointNet++ model achieved an overall accuracy (OA) of 95.25% and a mean intersection over union (mIoU) of 93.97% for tobacco plant segmentation-improvements of 5.12% and 5.55%, respectively, over the original PointNet++ model. Moreover, using the segmentation results from the improved PointNet++ model, the predicted phenotypic values exhibited strong agreement with ground-truth measurements, with coefficients of determination (R²) ranging from 0.86 to 0.95 and root mean square errors (RMSE) between 0.31 and 2.27 cm. This study provides a technical foundation for high-throughput phenotyping of tobacco and presents a transferable framework for phenotypic analysis in other crops.