An automated, high-performance approach for detecting and characterizing broccoli based on UAV remote-sensing and transformers: A case study from Haining, China

Accurate canopy mapping and head-volume estimation of large areas of broccoli is an important prerequisite for precision farming since it provides important phenotypic traits associated with field management, environmental control, and yield prediction. Currently, the detection and characterization of broccoli mostly rely on ground surveys and human interpretation, which is often time- and labor-intensive. Recent developments based on unmanned aerial vehicle (UAV) remote sensing offer low cost, timely, and flexible data acquisition, thereby providing a potential alternative technique to enhance in situ field surveys. The combination of UAV data and deep learning has led to a series of breakthroughs in rapid and automated collection of simultaneous multisensor and multimodal plant phenotyping data. However, their application for monitoring broccoli remains problematic when faced with the significant spatial scale involved and the variety of vegetation species. To address this problem, we propose herein a fast and reliable semi-automatic workflow based on deep learning to process UAV RGB imagery and LiDAR point clouds and thereby remotely detect and characterize broccoli canopy and heads. First, we explore the use of TransUNet to differentiate canopy and non-canopy regions in RGB images at the individual-plant scale. The results demonstrate that TransUNet consistently achieves the highest accuracy (average returned Precision, Recall, F1 score, and IoU of 0.917, 0.864, 0.901, and 0.895, respectively) compared with three CNN-based and two shallow learning-based approaches. In addition, TransUNet performs best in terms of robustness against variations in training samples. Subsequently, to estimate the volume of broccoli heads, a point cloud transformer (PCT) network is developed for point cloud segmentation. Improving upon the results of three existing methods PointNet, PointNet++, and K-means that were applied to the same datasets, the best-performing PCT produced a precision of 0.914, an overall recall of 0.899, an overall F1 score of 0.901, and an overall IoU of 0.879. A regression analysis indicates that the PCT estimates had R2 = 0.875, RMSE = 18.62, and rRMSE = 3.64 %, which is also superior to the results from other comparison approaches. Collectively, the wide application of such technology would facilitate applied research in plant phenotyping and precision agro-ecological applications and field management.