Improved 3D point cloud segmentation for accurate phenotypic analysis of cabbage plants using deep learning and clustering algorithms

Abstract

Plant phenotyping is essential for understanding and managing plant growth and development. 3D point clouds provide a better understanding of plant 3D structures. Point cloud segmentation is the basis for studying the 3D structure of plants through 3D point clouds, and accurate point cloud segmentation is crucial for extracting relevant phenotypic parameters. In this study, cabbage was used as an example, and a plant point cloud segmentation method combining deep learning algorithms and clustering algorithms was proposed. Specifically, a cabbage point cloud dataset was constructed using a 3D scanning platform. The ASAP attention module was incorporated into the PointNet++ model, resulting in the improved ASAP-PointNet model. Superior semantic segmentation performance on the cabbage point cloud dataset was demonstrated by this model. The workflow of the DBSCAN algorithm was also optimized, which exhibited enhanced performance in organ-level plant point cloud segmentation experiments. Subsequently, five phenotypic features were extracted. The experimental results revealed that an accuracy of 0.95 and an intersection over union (IoU) of 0.86 for semantic segmentation were achieved by the ASAP-PointNet model. The correlation coefficients between the four phenotype parameters (plant height, leaf length, leaf width, and leaf area) and their corresponding measured values were 0.96, 0.91, 0.95, and 0.94, respectively. An automated data analysis, from plant 3D point clouds to phenotypic parameters, is enabled by the proposed method, which serves as a valuable reference for plant phenotype research.