Toward A Foundation Model for Forest Point Clouds

Yuanwen Yue, Stefano Puliti^†, Damien Robert^†, Atakan Topaloğlu, Binbin Xiang, Maciej Wielgosz, Jan Dirk Wegner, Rasmus Astrup, Christian Rupprecht, Konrad Schindler

ETH Zurich, University of Oxford, Norwegian Institute of Bioeconomy Research (NIBIO), University of Zurich

^†Equal contribution

ForPT is a self-supervised pretrained point cloud model for forests. We pretrain a LitePT backbone on a large corpus of unlabeled airborne (ALS), UAV (ULS) and mobile (MLS) laser scans covering diverse forest ecosystems, and evaluate it on four forestry tasks: semantic segmentation, instance segmentation, tree species classification and tree age regression. Self-supervised pretraining accelerates convergence, improves performance when annotations are scarce, and transfers better than task-specific supervised pretraining.

Preparation

Environment

ForPT relies on LitePT. Follow the instructions in the LitePT repository to install the environment for LitePT. Then install the few additional packages used by ForPT:

pip install pandas matplotlib seaborn tqdm

Data

All data should be placed in ForPT/data, organized as expected by the configs:

ForPT/data/
├── forpt_ssl_corpus/              # unlabeled pretraining corpus (ALS / ULS / MLS)
│   └── knn_track/                 # labeled subset for kNN evaluation during pretraining
├── forinstancev2_sem/             # FOR-instanceV2 (semantic segmentation)
│   ├── train/ val/ test/
│   └── train_{5,10,20,50}.txt     # label-efficient training subsets
├── forspecies20k/                 # FOR-species20K (species classification)
│   ├── processed/                 # full-resolution trees: <species>/<tree>.txt (x,y,z), test/, train/val/test splits
│   └── processed_8192/            # same trees downsampled to 8,192 points, plus train_{1,5,10,20,50}p.txt subsets
└── forage/                        # tree age regression
    ├── FORage_tree_metadata_train_val.csv  # official FORage train/val metadata
    ├── FORage_tree_metadata_test.csv       # hidden test set: sampleID, file (no age)
    ├── pc/<name>/coord.npy        # one point cloud per tree (test: <name> = sampleID)
    ├── forage_{train,val,test}_8192points_uniform.pth  # cached 8,192-point samples
    └── train_{5,10,20,50}p.txt    # label-efficient training subsets

We provide the preprocessed datasets on Hugging Face:

Dataset	Task	Download	Local folder
ForPT corpus	Self-supervised pretraining	-	forpt_ssl_corpus
FOR-instanceV2	Semantic segmentation	prs-eth/forpt-forinstancev2-sem	forinstancev2_sem
	Instance segmentation	-	forinstancev2_ins
FOR-species20K	Species classification	prs-eth/forpt-forspecies20k	forspecies20k
FORage	Age regression	prs-eth/forpt-forage	forage

Point clouds are packed into .tar files, while split lists, metadata CSVs and the FORage .pth caches are stored as-is. To download everything and unpack it into the layout above, run the following from the ForPT root (requires pip install -U huggingface_hub):

mkdir -p data && cd data

# FOR-instanceV2 (~14 GB): train.tar, val.tar, test.tar
hf download prs-eth/forpt-forinstancev2-sem --repo-type dataset --local-dir forinstancev2_sem
# FOR-species20K (~168 GB): one tar per species folder in processed/ and processed_8192/
hf download prs-eth/forpt-forspecies20k --repo-type dataset --local-dir forspecies20k
# FORage (~13 GB): pc.tar
hf download prs-eth/forpt-forage --repo-type dataset --local-dir forage

# Unpack every tar next to itself, then delete the archive
find forinstancev2_sem forspecies20k forage -name "*.tar" | while read f; do
    tar -xf "$f" -C "$(dirname "$f")" && rm "$f"
done
rm -rf {forinstancev2_sem,forspecies20k,forage}/{.cache,.gitattributes}

FOR-species20K is the largest download. The full-resolution processed/ folder is only used by the *-allpts configs. If you only need the 8,192-point version used by all other configs, add --include "processed_8192/*" to its download command. Interrupted downloads resume when you rerun the same command.

Model Zoo

Pretrained backbone

Model	Params	Pretraining data	Config	Checkpoint
ForPT (LitePT-S)	12.4M	ForPT corpus	link	Download

Downstream tasks (100% labels)

Task	Benchmark	Model	Config	Checkpoint
Semantic segmentation	FOR-instanceV2	LitePT-scratch	link	Download
		ForPT-linear probing	link	Download
		ForPT-decoder probing	link	Download
		ForPT-finetune	link	Download
Instance segmentation	FOR-instanceV2	Coming soon	-	-
Species classification	FOR-species20K	LitePT-scratch	link	Download
		ForPT-linear probing	link	Download
		ForPT-finetune	link	Download
Age regression	FORage	LitePT-scratch	link	Download
		ForPT-linear probing	link	Download
		ForPT-finetune	link	Download

All checkpoints are hosted on prs-eth/ForPT under the <task>/<config name>/model/ layout. To download them into ForPT/weights, where the configs expect them (e.g. weights/00_ssl_pretrain/pretrain-forpt/model/epoch_30.pth), run from the ForPT root:

hf download prs-eth/ForPT --local-dir weights --exclude ".gitattributes" --exclude "README.md"

Training

Configs are grouped by task under configs/. The -d flag of scripts/train.sh selects the config folder, and experiments are saved to exp/${FOLDER}/${EXP_NAME}.

Config names follow the pattern {task}-{init}-{protocol}-{labels}:

• Initialization and protocol
  • litepts-scratch: LitePT-S trained from scratch (supervised baseline)
  • forpt-lin: linear probing on frozen ForPT features
  • forpt-dec: initialize from ForPT, freeze the encoder and train the decoder
  • forpt-ft: initialize from ForPT and fine-tune the full model
  • {sem,ins,species,age}-pretrain-lin: linear probing on a backbone pretrained with supervision on another task
• Labels: p001, p005, p010, p020, p050 and p100 train on 1%–100% of the training labels. allpts (species only) trains on train+val at full point density.

Self-supervised pretraining

We pretrain on 16 nodes with 4 GPUs each. On a SLURM cluster, launch the script once per node; the node rank and master address are read from SLURM:

srun --nodes=16 --ntasks-per-node=1 sh scripts/train.sh -g 4 -m 16 -d 00_ssl_pretrain -c pretrain-forpt -n pretrain-forpt

batch_size in the configs is the total across all GPUs, so running on fewer GPUs increases the batch per GPU.

All downstream experiments use a single node with 4 GPUs.

Semantic segmentation (FOR-instanceV2)

# LitePT-scratch
sh scripts/train.sh -g 4 -d 01_forinstancev2_sem -c semseg-litepts-scratch-p100 -n semseg-litepts-scratch-p100
# ForPT-linear probing
sh scripts/train.sh -g 4 -d 01_forinstancev2_sem -c semseg-forpt-lin-p100 -n semseg-forpt-lin-p100
# ForPT-decoder probing
sh scripts/train.sh -g 4 -d 01_forinstancev2_sem -c semseg-forpt-dec-p100 -n semseg-forpt-dec-p100
# ForPT-finetune
sh scripts/train.sh -g 4 -d 01_forinstancev2_sem -c semseg-forpt-ft-p100 -n semseg-forpt-ft-p100

Species classification (FOR-species20K)

# LitePT-scratch
sh scripts/train.sh -g 4 -d 03_forspecies20k -c cls-litepts-scratch-allpts -n cls-litepts-scratch-allpts
# ForPT-linear probing
sh scripts/train.sh -g 4 -d 03_forspecies20k -c cls-forpt-lin-allpts -n cls-forpt-lin-allpts
# ForPT-finetune
sh scripts/train.sh -g 4 -d 03_forspecies20k -c cls-forpt-ft-allpts -n cls-forpt-ft-allpts

Age regression (FORage)

# LitePT-scratch
sh scripts/train.sh -g 4 -d 04_forage -c age-litepts-scratch-p100 -n age-litepts-scratch-p100
# ForPT-linear probing
sh scripts/train.sh -g 4 -d 04_forage -c age-forpt-lin-p100 -n age-forpt-lin-p100
# ForPT-finetune
sh scripts/train.sh -g 4 -d 04_forage -c age-forpt-ft-p100 -n age-forpt-ft-p100

To run another protocol or label fraction, change the config name, e.g. -c semseg-forpt-lin-p010 for linear probing with 10% of the labels.

Checklist

• Release pretrained models.
• Release pretraining code and configs.
• Release semantic segmentation, species classification and age regression code.
• Release instance segmentation code.

Acknowledgment

This work uses parts of the code from LitePT, Pointcept, Sonata, ForestFormer3D, and FORage. We thank the authors for making their code publicly available.

Citation

If you find this project useful, please cite:

@article{yue2026forpt,
    title={Toward a foundation model for forest point clouds},
    author={Yue, Yuanwen and Puliti, Stefano and Robert, Damien and Topalo{\u{g}}lu, Atakan and Xiang, Binbin and Wielgosz, Maciej and Wegner, Jan Dirk and Astrup, Rasmus and Rupprecht, Christian and Schindler, Konrad},
    journal={arXiv preprint arXiv:2609.24787},
    year={2026}
}