ForestSeg3D: A semantically guided framework for individual tree segmentation from LiDAR point clouds

Area of Science:

• Forestry
• Remote Sensing
• Computer Vision

Background:

• Accurate individual tree segmentation from LiDAR point clouds is vital for forest inventory and ecological monitoring.
• Challenges include dense crown overlap, occlusion, and complex vertical structures in forests.
• Existing methods struggle with precise delineation in these complex environments.

Purpose of the Study:

• To propose ForestSeg3D, a novel framework for improved individual tree segmentation from LiDAR point clouds.
• To enhance segmentation accuracy in challenging forest conditions through semantic guidance.
• To enable more reliable forest structural attribute estimation.

Main Methods:

• Developed ForestSeg3D, a semantically guided framework utilizing hierarchical semantic supervision (HSS) and bidirectional cross-task distillation (BCTD).
• HSS employs a coarse-to-fine semantic learning scheme (Tree vs. Non-Tree, then Ground, Wood, Leaf) for structured priors.
• BCTD couples semantic prediction and instance partition to reduce label-boundary conflicts, enhanced by semantic-aware region merging (SRM) for large-scale inference.

Main Results:

• ForestSeg3D achieved state-of-the-art performance on the FOR-instanceV2 dataset, excelling in detection completeness, false-positive suppression, and F1-score.
• Demonstrated strong performance on the ForestSemantic dataset via cross-validation.
• Enabled more reliable estimation of tree-level attributes like height, crown diameter, and volume.

Conclusions:

• ForestSeg3D effectively addresses challenges in individual tree segmentation from LiDAR data.
• The framework offers practical value for forest inventory and structural assessment.
• The proposed methods significantly improve delineation accuracy in complex forest scenes.