PlantSeg: Automated multi-temporal sweet pepper phenotyping through segmentation and quantitative parameter extraction

Semantic segmentation and quantitative phenotypic analysis of plant 3D point clouds are essential for high-throughput crop growth monitoring and digital plant phenomics. Nevertheless, complex plant architectures, severe organ occlusion, non-uniform point density, and incomplete local geometric details continue to limit the accuracy and robustness of fine-grained stem–leaf segmentation. To address these challenges, this study presents PlantSeg, a semantic segmentation network tailored for 3D point clouds of sweet pepper plants. PlantSeg incorporates a Dilated Multi-Scale (DMS) module that jointly encodes local geometric details and global spatial context with low redundancy via shared neighborhood indexing and dilated sampling strategies. In addition, a Graph Wavelet Transform (GWT) module is introduced to enhance high-frequency geometric representations through spectral-domain multi-scale wavelet decomposition, thereby strengthening the model's discriminative capacity for slender stems, overlapping leaves, and stem–leaf junctions. On a custom-built sweet pepper point cloud dataset, PlantSeg achieves 83.00% mIoU and 90.65% mAcc, surpassing several state-of-the-art methods. The applicability of PlantSeg to other crop species is further evaluated on public point cloud datasets covering sorghum, tomato, and strawberry. Moreover, an automated 3D phenotype extraction framework is developed upon the segmentation results, enabling efficient quantification of key traits including leaf length, leaf width, plant height, and stem diameter. The extracted measurements show high agreement with manual references, with R² exceeding 0.87 for all traits. These results demonstrate that PlantSeg, together with its phenotype extraction framework, provides accurate, robust, and generalizable technical support for organ-level plant point cloud segmentation and 3D phenotypic quantification.