Leaf–Wood Separation in Tree Point Clouds via Unlearning-Driven Supervoxel Random Forest

Accurate separation of wood and leaf components from tree point clouds is essential for precise tree parameter estimation and high-fidelity 3D model reconstruction. However, deep learning approaches often suffer from poor generalization across species, while traditional machine learning methods based on original point-level features are highly sensitive to noise, resolution inconsistencies, and data incompleteness. To address these challenges, we propose a novelUnlearning-Driven Super-voxel Random Forest (USRF)framework for robust and efficient leaf–wood classification. Our method reformulates the classification task at the supervoxel level, enabling more stable feature representation and topological context modeling. A hierarchical feature extraction module is developed to integrate low-level geometric descriptors, PCA-based shape features, and supervoxel neighborhood characteristics, facilitating multi-scale structural perception. To further enhance classification performance and scalability, a machine unlearning-based feature selection strategy is introduced. This strategy iteratively removes redundant or detrimental features by evaluating their marginal contribution to classification outcomes. Experimental results show that, among 27 extracted features, retaining only the top six can still preserve 94% of the full model’s accuracy while significantly reducing training and inference time—providing critical insights for large-scale tree segmentation applications. A random forest classifier is trained using the selected features, and a category-aware distance refinement module is applied post-classification to correct residual misclassifications near trunk–canopy boundaries. Extensive evaluations on multiple public and in-house datasets demonstrate that the proposed USRF framework achieves an average classification accuracy of 93.7%, outperforming existing baselines in both precision and robustness.