手持式结构光扫描系统设计与玉米叶面积提取

Handheld Structured Light Scanning System Design and Maize Leaf Area Extraction

摘要:

基于三维点云的叶面积提取方法具有非接触、高效率和高精度优势，能够更好地满足现代农业对叶面积快速获取和精准评估的需求。以大田全生育期夏玉米为研究对象,通过自主设计的手持式结构光作物三维扫描仪，采集夏玉米全生育期点云数据，并提出了点云配准、去噪和下采样等预处理流程。随后，应用点云分割网络对玉米作物器官点云进行了精确分割，成功提取了玉米叶片点云数据，并提取了叶面积。结果表明，分割网络在点云分割精度方面表现优异，叶片点云精确率、召回率、F1分数和交并比指标均超过95%，其他器官分割指标也均高于75%。不同生育期叶面积提取结果存在显著差异。在苗期、拔节期、全生育期模型表现较好，R2分别为0.906 2、0.983 8、0.994 9，均方根误差分别为221.34、172.77、206.64 cm2；但在成熟期，模型表现显著下降，R2降至0.517 8，RMSE上升至209.32 cm2。不同施肥量下，叶面积提取结果整体良好，R2均高于0.98。随着施肥量变化，均方根误差呈先下降后上升趋势，分别为176.38、106.36、110.18、270.34 cm2。基于本文设计的设备和方法，能够准确有效地提取大田单株玉米叶面积，为智慧农业和表型机器人提供技术支持。

Abstract:

The method of leaf area extraction based on 3D point clouds offers advantages such as non-contact, high efficiency, and high-precision, making it well-suited to meet the demands of modern agriculture for rapid acquisition and accurate assessment of leaf area. Focusing on summer maize during its full growth period in field conditions, with four different fertilization treatments, each containing two sample plots, a self-developed handheld structured light crop 3D scanner was used, point cloud data were collected throughout the entire growth period of summer maize, and a series of point cloud preprocessing processes, including point cloud registration, denoising, and downsampling, were proposed. Subsequently, the PCT deep learning point cloud segmentation network was applied to accurately segment the crop organ point clouds, extracting the maize leaf point cloud data and successfully calculating the leaf area. The segmentation results showed that the PCT network performed excellently in the point cloud segmentation accuracy for maize organs, with the precision, recall, F1-score, and IoU metrics for the leaf point cloud all exceeding 95%, and the segmentation metrics for other organs also being above 75%. Significant differences were observed in the leaf area extraction results across different growth stages. During the seedling, jointing, and full growth stages, the extraction results were excellent, with R2 values of 0.906 2, 0.983 8, and 0.994 9, and RMSE values of 221.34 cm2, 172.77 cm2, and 206.64 cm2, respectively. However, in the mature stage, the model’s performance significantly was decreased, with an R2 of 0.517 8 and an RMSE of 209.32 cm2. Under different fertilization levels, the leaf area extraction results were consistently good, with R2 values above 0.98. As the fertilization amount changed, the RMSE showed a trend of first decreasing and then increasing, with specific values of 176.38 cm2, 106.36 cm2, 110.18 cm2, and 270.34 cm2. In conclusion, the method proposed can accurately and effectively extract the leaf area of individual maize plants in field conditions, providing reliable data support for precision agriculture.