Hierarchical segmentation framework with dynamic parameter optimization for accurate stem-leaf partitioning and phenotypic extraction in maize

• We propose a hierarchical framework integrating spatial normalization and dynamic parameter tuning. • Calculate the OBB spindle and rotate it to correct the tilt and eliminate the attitude change caused by the viewpoint. • Coarse segmentation is achieved by using dynamic root radius adjustment and clustering. • Introduce misclassification detection and re-clustering to improve segmentation by correcting the error of dense blade overlap at the top. Maize, as one of the most important crops, plays a key role in phenotypic research, which promotes the development of precision agriculture and the in-depth exploration of the gene-phenotypic association mechanism. However, traditional phenotyping methods relying on manual measurements or two-dimensional images have significant limitations in terms of efficiency and accuracy, particularly in effectively analyzing the complex three-dimensional structures of plants. To address these challenges, this paper proposes a high-precision automatic segmentation method for maize stem and leaf organs based on 3D point clouds. The method consists of the following three core modules: (1) point cloud rotation correction preprocessing using a directed bounding box; (2) coarse segmentation of the stem and leaf using a dynamic root-shoot radius adjustment strategy; (3) fine segmentation incorporating dynamic misclassification detection and re-clustering mechanisms. The proposed method was systematically evaluated on maize plant point cloud data at multiple growth stages, and compared with manually annotated results. Experimental results show that the method achieved an average precision of 0.944, average recall of 0.915, Micro-F1 score of 0.920, and average overall accuracy of 0.935, demonstrating excellent segmentation performance. Furthermore, seven key phenotypic parameters, including plant height, crown diameter, stem height, stem diameter, number of leaves, leaf length, and leaf width, were automatically extracted, with the results showing highly significant correlations with manual measurements. This study provides effective technical support for high-precision 3D segmentation of maize stem and leaf organs and automated phenotypic analysis, laying a solid foundation for high-throughput plant phenotyping research and 3D reconstruction applications.