Separating structural components is important but also challenging for plant phenotyping and precision agriculture. Light detection and ranging (LiDAR) technology can potentially overcome these difficulties by providing high quality data. However, there are difficulties in automatically classifying and segmenting components of interest. Deep learning can extract complex features, but it is mostly used with images. Here, we propose a voxel-based convolutional neural network (VCNN) for maize stem and leaf classification and segmentation. Maize plants at three different growth stages were scanned with a terrestrial LiDAR and the voxelized LiDAR data were used as inputs. A total of 3000 individual plants (22 004 leaves and 3000 stems) were prepared for training through data augmentation, and 103 maize plants were used to evaluate the accuracy of classification and segmentation at both instance and point levels. The VCNN was compared with traditional clustering methods ( $K$ -means and density-based spatial clustering of applications with noise), a geometry-based segmentation method, and state-of-the-art deep learning methods (PointNet and PointNet++). The results showed that: 1) at the instance level, the mean accuracy of classification and segmentation (F-score) were 1.00 and 0.96, respectively; 2) at the point level, the mean accuracy of classification and segmentation (F-score) were 0.91 and 0.89, respectively; 3) the VCNN method outperformed traditional clustering methods; and 4) the VCNN was on par with PointNet and PointNet++ in classification, and performed the best in segmentation. The proposed method demonstrated LiDAR’s ability to separate structural components for crop phenotyping using deep learning, which can be useful for other fields.

中文翻译：

---

使用地面激光雷达数据和深度卷积神经网络分离玉米的结构成分以进行田间表型分析

分离结构成分很重要，但对植物表型分析和精准农业也具有挑战性。光探测和测距 (LiDAR) 技术可以通过提供高质量数据来克服这些困难。然而，在自动分类和分割感兴趣的组件方面存在困难。深度学习可以提取复杂的特征，但主要用于图像。在这里，我们提出了一种基于体素的卷积神经网络 (VCNN)，用于玉米茎叶分类和分割。使用地面 LiDAR 扫描处于三个不同生长阶段的玉米植株，并将体素化的 LiDAR 数据用作输入。通过数据增强，总共准备了 3000 株单株植物（22 004 片叶子和 3000 根茎）用于训练，和 103 株玉米植物被用来评估实例和点级别的分类和分割的准确性。将 VCNN 与传统的聚类方法（$K$ -means 和基于密度的有噪声应用空间聚类）、基于几何的分割方法和最先进的深度学习方法（PointNet 和 PointNet++）进行了比较。结果表明：1）在实例层面，分类和分割的平均准确率（F-score）分别为1.00和0.96；2）在点层面，分类和分割的平均准确率（F-score）分别为0.91和0.89；3）VCNN方法优于传统聚类方法；4）VCNN在分类上与PointNet和PointNet++相当，在分割方面表现最好。