A Two-Stage Leaf-Stem Separation Model for Maize With High Planting Density With Terrestrial, Backpack, and UAV-Based Laser Scanning

The accurate and high-throughput extraction of phenotypic traits is of great significance for crop breeding and growth monitoring. The segmentation of structural components (e.g., leaves and stems) is a prerequisite for extracting phenotypic traits. In the past decade, there has been an increase in methods attempting to separate leaves and stems in point clouds. However, previous researches mainly focus on plants at the individual level due to the interlocked and overlapped nature of leaves and the bottleneck existing for field plants to extract phenotypic traits. To address this issue, a novel two-stage leaf-stem separation model encompassing the initial separation of leaves and stems and optimization is presented in this article. The model is based on the different geometric features of leaves and stems of maize plants defined by neighborhood points, and a cylinder is used to find the neighborhood points by considering the elongated characteristic of maize stems. After that, another elongated cylinder (0.5 m high and 0.02 m diameter) is used to traverse the stem points to optimize the initially separated results. Maize plants with the planting density of 45 000 plants/ha in the filling stage (Exp. 2019) were used to train and test the model in the initial separation step (Experiment 1), showing that the separation accuracy (SA) could be up to 91.3%. It was concluded that a 0.11-m-high and 0.07-m diameter cylinder was the optimal searching parameter for the initial separation and 0.25 m was the optimal threshold for optimization. We also tested the transferability of the model (Experiment 2) for maize plants with different planting densities (45 000, 67 500, 90 000, and 105 000 plants/ha), different growth stages (jointing, silking and filling), and point clouds collected using multiple platforms [terrestrial laser scanning (TLS), light detection and ranging (LiDAR) Backpack (LiBackpack), and unmanned aerial vehicle-LiDAR (UAV-LiDAR)], suggesting that the model performed well for all the datasets. In addition, the simulated datasets of maize with different planting densities were used to assess the model performance at the point level, showing the SA values were 0.92, 0.91, 0.91, and 0.90 for maize with the planting densities of 45 000, 67 500, 90 000, and 105 000 plants/ha, respectively. The proposed model in this study is innovative, and it has promising prospects for the high-throughput extraction of the phenotypic traits in field maize plants and could facilitate genotype selection in crop breeding and 3-D plant modeling.