An agronomy-oriented framework for multiscale maize phenotyping from 3D point clouds

Accurate stem and leaf segmentation and phenotypic trait extraction from maize 3D point clouds are essential for understanding plant architectural responses to stress and their relationships with agronomic performance. This study established a full-season single‑plant maize 3D point cloud dataset under three planting densities (D1, 90000 plants ha⁻¹; D2, 110000 plants ha⁻¹; and D3, 130000 plants ha⁻¹) and three irrigation levels (I1, 100% ETc; I2, 65% ETc; and I3, 40% ETc). The point clouds were acquired by indoor multi-station laser scanning of individual plants excavated from the field plots at multiple growth stages. The dataset contained 408 labelled samples, including 285 for training and 123 for independent evaluation. Based on these data, CGHFormer, an optimized PTv3-based framework, was developed for maize stem and leaf semantic and instance segmentation by integrating cylindrical geometry enhancement and height aware dual task attention. Compared with the baseline PTv3, the full CGHFormer increased mIoU and AP@50 by 1.58 and 3.12 percentage points, respectively, and achieved the best performance, with mIoU, mPrec, mRec, mF1, and AP@50 reaching 91.34%, 94.09%, 95.21%, 94.61%, and 54.27%, respectively. A multiscale workflow was then developed to extract conventional whole- plant morphological traits and agronomy-oriented traits from the ear leaf and third internode. Whole plant traits showed high agreement with reference measurements (R² = 0.953–0.997), and organ traits showed reliable accuracy (R² = 0.913–0.989). Six key structural traits captured distinct agronomic responses, including plant height, plant volume, ear leaf area, ear leaf base angle, stem volume, and third internode diameter. Yield and RUE increased with planting density but decreased under reduced irrigation. D3I1 produced the highest grain yield of 19.28 t ha⁻¹, while D2I1 yielded 17.83 t ha⁻¹, with no significant difference between the two treatments. D2I1 also exhibited higher stem bending and puncture strengths than D3I1. Correlation and Mantel analyses further supported the agronomic relevance of the selected 3D traits in explaining treatment-related differences in maize performance. This framework supports organ segmentation and multiscale trait extraction from indoor scans of field-grown maize, providing structural indicators for the analysis of irrigation and planting density responses.