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

PointsToWood — wood-leaf semantic segmentation of TLS forest point clouds.

_{Probability of wood (blue = low, red = high). Data: Wang et al., 2021.}

---

Project Page  ·  GitHub

---

Paper

PointsToWood: 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)

For the implementation described in the preprint, see the version1.0-paper branch.

---

Quick Start

cd pointstowood
python predict.py --point-cloud your_plot.ply

Output: your_plot_p2w.ply with per-point columns prediction (0 = leaf, 1 = wood) and pwood (0.0–1.0).

---

Installation

Requirements: Linux, CUDA 12.9, NVIDIA driver 570+, conda/mamba

bash pointstowood/install.sh
conda activate ptw

The script creates a ptw conda environment with Python 3.11, PyTorch 2.8.0, and all PyG extensions.

---

Model Weights (Git LFS)

# Install Git LFS if needed
sudo apt-get install git-lfs
git lfs install

# Pull weights after cloning
git lfs pull

File	Description
model/h4mcc-eu.pth	EU teacher model (default, ~20M params)
model/h4mcc-finland.pth	Distilled — Finnish forest (~1M params)
model/h4mcc-poland.pth	Distilled — Polish forest (~1M params)
model/h4mcc-spain.pth	Distilled — Spanish forest (~1M params)

---

Inference

# Default (EU teacher, standard quality)
python predict.py --point-cloud your_plot.ply

# Biome-specific model (auto-selected by region)
python predict.py --point-cloud your_plot.ply --region finland
python predict.py --point-cloud your_plot.ply --region poland
python predict.py --point-cloud your_plot.ply --region spain

# Maximum accuracy
python predict.py --point-cloud your_plot.ply --thorough

# Geometry only (no reflectance)
python predict.py --point-cloud your_plot.ply --no-refl

Key options:

--region finland|poland|spain   auto-select distilled biome model
--thorough                      maximum accuracy: 8 overlap offsets, 4× TTA, dual-perspective
--any-wood 0.5                  wood if any point in voxel exceeds threshold (higher recall)
--no-refl                       geometry-only mode (zeros out reflectance channel)

By default, inference runs at standard quality: two scales (multi-perspective), 2× z-rotation TTA, no overlap offsets. Add --thorough for the highest accuracy — dense overlap, more augmentation, and dual-perspective inference — at significantly higher compute cost.

---

Training

1. Prepare data

Split raw .ply files into train/test/eval along the x-axis:

python split_ply.py /path/to/plots/ --prefix fin --output data/

Files are named by region prefix (fin, spa, pol, gbr, etc.) and written to data/train/, data/test/, data/eval/.

2. Train

# EU model (all European data)
python train.py --region eu --preprocess --device cuda

# Single biome
python train.py --region fin --preprocess --device cuda

---

Knowledge Distillation

Distil a lightweight biome-specific student from the EU teacher:

python distill.py --region fin --preprocess --teacher-model h4mcc-eu.pth

---

Architecture

A 3-stage encoder-decoder with a custom anisotropic convolution operator. LiDAR reflectance acts as a spatial modulator — weighting geometric neighbour contributions based on material properties rather than being treated as a flat input feature. The model falls back to geometry alone when reflectance is absent.

Supervision uses a multi-scale contrastive boundary loss at every encoder stage and a cyclical focal loss schedule (0 → peak → 0 over training) to stabilise early gradients while focusing on hard boundary points mid-training. Decoder unpooling uses encoder cluster indices rather than k-NN interpolation.

Student models use the same architecture compressed by channel width and block depth, trained via knowledge distillation from the teacher.

---

Evaluation Metric — H4-MCC

Models are selected and compared using H4-MCC: the harmonic mean of Matthews Correlation Coefficient (MCC) across four conditions:

Condition	Description
Pure + reflectance	Unambiguous voxels, full sensor data
Edge + reflectance	Wood/leaf boundary voxels, full sensor data
Pure + geometry	Unambiguous voxels, XYZ only
Edge + geometry	Wood/leaf boundary voxels, XYZ only

MCC is preferred over F1 because it accounts for both wood and leaf errors — F1 ignores true negatives and can appear high even when leaf classification is poor. The harmonic mean across all four conditions means a model must perform well in every setting: strong reflectance performance cannot mask failure on geometry-only inputs, and easy voxels cannot hide poor boundary classification.

---

Region Prefixes

--region	Data used
eu	All European prefixes
global	All files in pool
fin / spa / pol / ...	That prefix only

---

License

Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)

Free for academic research, personal use, and non-commercial applications. Commercial use is prohibited without explicit permission from the authors.

---

References

_{Mspace Lab (2024) ForestSemantic: A Dataset for Semantic Learning of Forest from Close-Range Sensing. Zenodo. https://doi.org/10.5281/zenodo.13285640.}

_{Wang, Di; Takoudjou, Stéphane Momo; Casella, Eric (2021). LeWoS: A universal leaf-wood classification method to facilitate the 3D modelling of large tropical trees using terrestrial LiDAR. Dryad. https://doi.org/10.5061/dryad.np5hqbzp6.}

_{Wan, Peng; Zhang, Wuming; Jin, Shuangna (2021). Plot-level wood-leaf separation for terrestrial laser scanning point clouds. Dryad. https://doi.org/10.5061/dryad.rfj6q5799.}

_{Weiser, Hannah et al. (2024). Manually labeled terrestrial laser scanning point clouds of individual trees for leaf-wood separation. https://doi.org/10.11588/data/UUMEDI.}

_{Owen, H. J. F., Lines, E., & Grieve, S. (2024). Plot-level semantically labelled terrestrial laser scanning point clouds (1.0). Zenodo. https://doi.org/10.5281/zenodo.13268500.}

_{Van den Broeck, W.A.J., Terryn, L., Chen, S., Cherlet, W., Cooper, Z.T. and Calders, K. (2025). Pointwise deep learning for leaf-wood segmentation of tropical tree point clouds from terrestrial laser scanning. ISPRS Journal of Photogrammetry and Remote Sensing, 227, pp.366–382. https://doi.org/10.1016/j.isprsjprs.2025.06.023}

_{Van den Broeck, W.A.J., Terryn, L. and Calders, K. Leaf-wood annotated tropical tree point clouds from terrestrial laser scanning. [Dataset]. Associated publication: Van den Broeck et al. (2025). https://doi.org/10.1016/j.isprsjprs.2025.06.023}

_{Ali, M., Biswas, A., Iglseder, A., Kumar, V., Kumar, S., Gupta, S., Hollaus, M., Pfeifer, N. and Lohani, B. (2026). Terrestrial and Airborne Laser Scanning Dataset of Trees in the Shivalik Range, India with Field Measurements and Leaf–Wood Classifications. Scientific Data, 13, Article 420.}