Deep learning-based 3D morphological segmentation and quantitative growth analysis of field-grown cabbage across the full cycle.

Monitoring the growth dynamics in field-grown cabbage is critically important for ensuring stable vegetable production and advancing precision agricultural management. However, conventional two-dimensional (2D) image-based monitoring approaches are limited to planar projection information and lack representations of spatial structural characteristics, rendering them inadequate for supporting high-precision, full-cycle phenotypic monitoring of cabbage under open-field conditions. In this study, a high-precision three-dimensional (3D) point cloud dataset covering the period from the seedling stage to maturity was constructed using depth cameras in conjunction with multi-view spatial registration techniques. Building on this dataset, an adaptive point cloud segmentation network designed for the whole-cycle growth monitoring was proposed, incorporating a Head Refinement Module (HRM), a Leaf Instance Segmentation Module (LISM), and Cross Module Interaction (CMI) to address leaf adhesion and head boundary delineation. Experimental results demonstrated that the proposed method consistently outperformed state-of-the-art models in both semantic and instance segmentation tasks. For semantic segmentation, the mean Intersection over Union (mIoU) reached 0.767, with a point classification accuracy of 94.8%. The model comprises 54.25 million parameters and achieves an average response time of 0.76 s. For instance segmentation, the Average Precision (AP) improved by 2.3% for cabbage heads and 3.8% for leaves, while the Average Recall (AR) increased by 6.9%. Growth parameters, including plant height and canopy spread, extracted from the segmentation results showed strong agreement with ground-truth measurements, with correlation of coefficients (R