Wood-leaf semantic segmentation of TLS forest point clouds

Reflectance-Modulated Anisotropic Convolutions for Leaf-Wood Segmentation Across Diverse Forest TLS Data

Owen, H. J. F. · Allen, M. J. A. · Grieve, S. W. D. · Wilkes, P. · Flynn, W. R. M. · Lines, E. R.

Under review

Drag to reveal wood classification

Reflectance-modulated convolution

LiDAR return intensity acts as a spatial modulator — weighting geometric neighbourhood contributions by material properties rather than as a flat input feature. Falls back gracefully to geometry when reflectance is absent.

Biome-specific distillation

Lightweight student models (~1M params) distilled from the EU teacher (~20M params) for Finnish, Polish, and Spanish forests. Auto-selected with --region. Same architecture, 20× smaller.

H4-MCC evaluation

Models are ranked by the harmonic mean of MCC across four conditions — pure/edge voxels × with/without reflectance. A model must perform well in every setting to score well.

Quick start

git clone https://github.com/harryjfowen/PointsToWood
cd PointsToWood
bash install.sh && conda activate ptw
python predict.py --point-cloud your_plot.ply

Output: your_plot_p2w.ply with per-point prediction (0 = leaf, 1 = wood) and pwood probability columns.

Pre-trained models

Model	Region	Params	Flag
`h4mcc-eu.pth`	Pan-European (default)	~20M	—
`h4mcc-finland.pth`	Finnish forest	~1M	`--region finland`
`h4mcc-poland.pth`	Polish forest	~1M	`--region poland`
`h4mcc-spain.pth`	Spanish forest	~1M	`--region spain`