Automated extraction of architectural traits from apple tree branches via 3D instance segmentation and skeletonisation

Accurate and efficient extraction of 3D phenotypic traits of apple tree branches is crucial for apple tree breeding and growth monitoring. Current methods, however, rely heavily on manual measurement, which are time-consuming, costly, and difficult to scale. Moreover, existing approaches based on deep learning and skeleton extraction still face critical challenges when processing high-density, large-scale point cloud data, including insufficient branch segmentation accuracy, suboptimal skeleton extraction quality, and a lack of category-specific fine-grained phenotypic quantification capabilities. To overcome these limitations, a dedicated framework for automated phenotypic trait extraction was presented. This approach consists of three key steps: (1) instance segmentation of branches from 3D data using the Relation3D model to distinguish individual branches; (2) skeletonisation of each segmented branch based on geometric characteristics and the L1-median algorithm to obtain fine-grained skeletal structures; and (3) quantitative trait extraction, where trunk height, branch count by category, and branch length are automatically computed. Experimental results demonstrate that Relation3D achieves high-precision instance segmentation, and our skeletonisation method reliably recovers branch geometry. Consequently, the framework accurately extracts key phenotypic traits, including branch number and length. This study provides an efficient and accurate solution for branch-level phenotyping, with significant potential to enhance apple tree breeding and precision horticulture.