Purpose <p>High-resolution characterisation of orchard trees is important for precision crop management. The partial canopy penetration of LiDAR enables access to internal branch architecture beyond the canopy surface. However, large-scale use of mobile LiDAR remains limited by scarce annotated datasets and poor model transferability. This study proposes a method for tree-level structural characterisation that reduces reliance on labelled field data and supports integration into monitoring workflows.</p> Methods <p>A multiscale framework combining synthetic data generation, semantic pretraining, descriptor-guided sampling, and contrastive representation learning is introduced to analyse unlabelled mobile LiDAR point clouds. Semantic segmentation is performed using a model trained exclusively on simulated data to classify points into ground, low vegetation, and tree classes. Individual trees are then isolated via density-based clustering applied to the tree-labelled subset. Two structural representations are extracted: skeleton-based geometric descriptors and deep metric embeddings learned via contrastive learning, supporting variability analysis, anomaly detection, and clustering.</p> Results <p>Experiments on simulated and real datasets demonstrate synthetic-to-field transfer without field-specific retraining and reliable tree isolation under operational conditions. Both descriptor types capture inter-tree variability and enable identification of atypical elements without manual annotation, supporting orchard-scale mapping of structural heterogeneity.</p> Conclusion <p>The approach establishes an annotation-efficient and transferable pathway for tree-level structural analysis by combining simulation-driven learning with label-free representation. While direct agronomic validation was beyond the scope of this study, the extracted descriptors act as phenotypic proxies that can be integrated with management variables in future research.</p>