Automated Organ Segmentation of Sunflower Plants from 3D Point Clouds Using an Improved Region-Growing Algorithm

Accurate segmentation of plant organs is a key foundation for the automatic extraction of three-dimensional phenotypic traits of crops and the intelligent monitoring of growth status, and it plays an important supporting role in the screening of superior traits and precision cultivation management. However, for crops with complex plant architecture, the complex spatial relationships among organs and significant morphological differences still make 3D point cloud segmentation face considerable challenges. Taking sunflower as an example, its stems, leaves, and flower discs are interwoven with one another in 3D space, and the leaf morphology is irregular. Existing plant organ segmentation methods are prone to problems such as confusion between flower discs and stems, fragmentation of stem point clouds, and leaf oversegmentation. To address these issues, this study proposes an improved region-growing algorithm for organ segmentation of sunflower plants. The proposed method first performs denoising and downsampling preprocessing on the original point cloud data to remove noise and redundant points. Using the plant base as the initial seed point, the main-stem skeleton is iteratively extracted by incorporating local neighborhood compactness, representative candidate points selection, growth direction continuity, and local linearity constraints. Subsequently, a buffer-based reconstruction strategy is employed to recover the complete stem point cloud, while the remaining point clouds are processed through density-based clustering and organ classification. Experiments were conducted on point cloud datasets of 30 tall single-head sunflower plants (5 plants for parameter tuning, 25 plants for quantitative evaluation). The proposed method achieved segmentation precision values of 93.22%, 99.00%, and 91.86% for sunflower stems, leaves, and flower, respectively. The corresponding recall values were 98.11%, 97.38%, and 98.77%, while the F1-scores reached 95.55%, 97.81%, and 94.20%, respectively. Comparative experiments with the traditional region-growing method and the skeleton extraction method on four representative plants showed that the proposed method achieved the highest F1-scores for all three organs, improving the average F1-score by 48.28 and 12.35 percentage points, respectively. The results demonstrate that the proposed method can accurately segment individual sunflower organs from single-plant point clouds, providing reliable technical support for efficient extraction of key phenotypic traits.