Self-Supervised Pretraining Improves Cross-Site and Cross-Scale Robustness of Point Cloud Leaf-Wood Segmentation

This study addresses the challenge of inconsistent leaf-wood segmentation in terrestrial laser scanning (TLS) point clouds across diverse forest types. The authors employ a self-supervised learning (SSL) strategy using the Point-M2AE architecture. By pretraining the model on a combination of synthetic ShapeNet-55 objects and 2,400 individual tree point clouds, the model learns to reconstruct masked point regions, allowing it to capture generalizable geometric features without manual labels. The researchers then fine-tune this pretrained encoder on annotated tree data using recursive voxel subdivision, which enables the model to handle varying point densities and operate at both individual-tree and plot scales without architectural changes.

SSL pretraining substantially outperformed models trained from scratch. Wood intersection-over-union (IoU) increased from 60.5% to 70.0% for needleleaf trees and from 69.7% to 76.3% for broadleaf trees. In a multi-site benchmark covering four countries, the pretrained model demonstrated the highest overall performance and the lowest cross-site variation among tested methods. Furthermore, the model maintained high accuracy at the plot level (mIoU of 84.7% for broadleaf and 77.7% for needleleaf) without requiring additional site-specific fine-tuning.

To validate the practical utility of the improved segmentation, the authors used the model to process TLS data from tropical forests in Guyana, Indonesia, and Peru. By feeding the segmented wood points into a quantitative structure model (QSM), they estimated wood volume for 28 trees. The SSL-pretrained model achieved a mean absolute error (MAE) of 2.40 m³, which is less than half the error of traditional algorithmic baselines (5.27–5.94 m³), confirming that better leaf-wood separation directly translates into more reliable biomass and structural assessments.