An unsupervised semantic segmentation network for wood-leaf separation in 3D point clouds.

Area of Science:

• Forestry and Remote Sensing
• Computer Vision and Machine Learning

Background:

• Automated forest inventory and management rely on accurate separation of wood and leaf components in tree point clouds.
• Traditional supervised methods require extensive, costly, and time-consuming point-wise annotation, hindering widespread adoption.
• A need exists for unsupervised approaches to overcome the limitations of supervised learning in wood-leaf separation.

Purpose of the Study:

• To explore the feasibility of unsupervised wood-leaf separation in 3D tree point clouds.
• To propose and evaluate a novel unsupervised semantic segmentation network for direct wood and leaf component extraction.

Main Methods:

• Developed an unsupervised semantic segmentation network utilizing a sparse convolutional neural network backbone.
• Incorporated dual point attention (DPA) and point cloud feature convolutional integrator (PFCI) modules for enhanced feature extraction and fusion.
• Generated pseudolabels via super point clustering for semantic classification.

Main Results:

• Achieved an overall accuracy (oAcc) of 67.583% and mean intersection over union (mIoU) of 38.512% for forest-level wood-leaf separation.
• Attained oAcc of 80.856% and mIoU of 49.695% for tree-level wood and leaf separation.
• Outperformed state-of-the-art methods (GrowSP, PointDC) and demonstrated robustness under occlusion and strong generalization capability.

Conclusions:

• The proposed unsupervised network effectively separates wood and leaf components in 3D point clouds without annotated data.
• The DPA and PFCI modules significantly contribute to improved segmentation accuracy.
• This approach offers a viable and robust solution for automated forest inventory and management.