Poplar-Studio: skeleton-guided reconstruction of editable 3D poplar seedling models from multi-view point clouds

Accurate, topologically consistent 3D plant models are important for quantitative phenotyping and for initializing functional–structural plant models (FSPMs). Structure-from-motion and multi-view stereo (SfM–MVS) reconstruction yields dense plant point clouds at low cost, but these clouds lack explicit organ connectivity, structural roles, and editable geometry. We developed Poplar-Studio, an integrated topology-preserving modelling framework that converts multi-view-derived poplar seedling point clouds into editable, organ-resolved structural models in standard Blender/OBJ formats. Three design properties define this framework. First, a two-class segmentation separates leaves from the aggregated stem system, which is contracted into a skeleton graph by Laplacian contraction and minimum-spanning-tree construction. Main-stem, petiole, junction, and terminal roles are recovered from this graph without point-level main-stem/petiole labels and converted into differentiated parametric geometry. Second, each lamina is reconstructed directly from its own point cloud as a boundary-constrained open surface in a PCA-aligned local frame, rather than instantiated from generic leaf templates. Third, the two branches are integrated into a common metric coordinate system in which the main stem, individual petioles, and individual leaves remain as separately editable organ-level objects, so that the framework terminates in a complete digital plant model rather than an intermediate structural representation. Across 240 fully expanded leaves, reconstructed leaf area agreed closely with independent measurements (R 2 = 0.943, relative RMSE = 10.3%, bias = 0.65 cm 2 ), with all leaves reconstructed as complete, hole-free laminae and the proposed method providing the best overall balance of accuracy, completeness, and efficiency among four reconstruction methods. Across four Populus genotypes, mean stem-system Chamfer distance to the source point clouds ranged from 0.143 to 0.227 cm. In a paired comparison on 40 plants at the 1-cm petiole-matching threshold, the graph-guided workflow showed higher mean petiole-branch precision than a geometry-first segment-then-fit baseline (78.4% vs 74.2%; Holm-adjusted p = 0.015). Recall and F1 did not differ significantly, whereas surface agreement with the source point clouds favoured the baseline (mean Chamfer distance 0.14 vs 0.20 cm). Starting from prepared point clouds, the proposed reconstruction workflow required less than 2.5 min per plant, corresponding to at least 24 plants h⁻ 1 on a single workstation. Poplar-Studio therefore provides a quantitatively characterized route from unordered plant point clouds to physically scaled, editable, topology-resolved models for phenotyping, structural analysis, and FSPM-oriented applications.