TreeSeg-Net: An End-to-End Instance Segmentation Network for Leaf-Off Forest Point Clouds Using Global Context and...

Area of Science:

• Forestry science
• Computer vision
• Remote sensing

Background:

• Forest ecosystems are vital for carbon cycling and biodiversity.
• Unmanned aerial vehicle (UAV) technologies provide cost-effective forest data.
• Leaf-off conditions in complex forests present segmentation challenges due to interlaced crowns and fuzzy boundaries.

Purpose of the Study:

• To develop an automated method for precise individual tree segmentation from UAV-derived point clouds.
• To address limitations in existing segmentation techniques for complex, leaf-off forest environments.
• To improve forest resource management through accurate tree parameter extraction.

Main Methods:

• Proposed TreeSeg-Net, an end-to-end instance segmentation network for raw point clouds.
• Incorporated a global context attention module (GCAM) to capture long-range dependencies.
• Introduced a spatial proximity weighting module (SPWM) with geometric constraints to reduce under-segmentation.

Main Results:

• TreeSeg-Net achieved 97.2% average precision (AP) for instance segmentation.
• The network attained a 99.7% mean intersection over union (mIoU) for semantic segmentation.
• Demonstrated superior accuracy compared to mainstream segmentation networks.

Conclusions:

• TreeSeg-Net offers an efficient and automated solution for precise forest resource inventory.
• The method effectively handles complex forest structures and segmentation challenges.
• Enables accurate individual tree separation directly from point cloud data.