Automatic organ-level point cloud segmentation of maize shoots by integrating high-throughput data acquisition and deep learning

Point cloud segmentation is essential for studying the 3D spatial characteristics of plants. Notably, the segmentation accuracy greatly impacts subsequent 3D plant phenotypes extraction and 3D plant reconstruction. Automated segmentation approaches for plant point clouds are a bottleneck in achieving big data processing of 3D plant phenotypes. Using maize as a representative crop, this study developed DeepSeg3DMaize, a technique for plant point cloud segmentation that integrates high-throughput data acquisition and deep learning. A high-throughput data acquisition platform for individual plants and an association mapping panel containing 515 inbred lines were used to construct the training dataset. Specifically, the MVS-Pheno platform was used to acquire high-throughput data, and Label3DMaize was used for point cloud data labeling. Based on the dataset, PointNet was introduced to implement stem-leaf and organ instance segmentation, and six phenotypes were extracted. According to the results, the mean precision and F1-Score of stem-leaf segmentation were 0.91 and 0.85, respectively. Meanwhile, the mean precision and F1-Score for organ instance segmentation were 0.94 and 0.93, respectively. The correlations of the six parameters (leaf length, leaf width, leaf inclination, leaf growth height, plant height, and stem height) extracted from the segmentation results with the measured values were 0.90, 0.82, 0.94, 0.95, 0.99, and 0.94, respectively. High-throughput data acquisition, automatic organ segmentation, and phenotypic data extraction form an automatic phenotypic data processing pipeline, which is practical for dealing with large amounts of initial data. Besides, it provides a systematic reference for the automated analysis of 3D phenotypic features at the individual plant level.