Overall Workflow

≈160 km²

Forest Area

≈33B

LiDAR Points

An unlabeled pretraining corpus (≈15,900 ha) collected without any annotations.

EMA

Momentum Teacher

The student learns to match the teacher's representations — no labels required.

4

Downstream Tasks

The pretrained backbone is transferred to each downstream forestry task via frozen feature transfer or end-to-end fine-tuning.

Self-Distillation Pipeline

The student encodes masked, partial views of a forest point cloud; the teacher encodes the same scene in full. Two prediction heads align the student's representations with the momentum-updated teacher's — with no labels and no negative pairs.

Self-Distillation Loss

→

Local-to-global Distillation

The teacher encodes a global view and its local-global head predicts target P_t^l; the student encodes a local view and predicts P_s^l with the same head.

L_l2g = Σ H(P_t^l, P_s^l)

→

Masked-to-global Distillation

The teacher encodes the unmasked global view and its mask-global head predicts P_t^m; the student encodes the corresponding masked global view and predicts P_s^m.

L_m2g = Σ H(P_t^m, P_s^m)

Overall Workflow

1

Source Task at a Time

The backbone is pretrained with full supervision on one forestry task (species classification shown) — repeated independently for each of the four tasks.

3

Remaining Tasks for Transfer

Each resulting encoder is transferred to the remaining tasks, excluding its own source task from evaluation.

Learned Representations

Per-point features from our frozen self-supervised backbone, projected to RGB via the top-3 principal components (PCA).

Downstream Tasks

Semantic Segmentation

Instance Segmentation

Species Classification

Age regression

State-of-the-art semantic segmentation

We evaluate 3-class point-wise segmentation (ground / wood / leaf) on held-out forest plots, comparing a training-free kNN probe against linear probing, decoder probing, training from scratch, and full fine-tuning of the pretrained backbone.

Data efficiency

kNN alone scores decently with zero training; fine-tuning beats scratch throughout, and pretraining's edge is largest when labels are scarce.

Training dynamics

Pretraining accelerates convergence — fine-tuning for just 10 epochs already beats scratch trained for the full 100.

Qualitative comparison

Even the training-free kNN probe (leftmost column) produces decent segmentations that closely track ground truth across all three held-out plots — NIBIO, NIBIO_MLS, BlueCat.

Qualitative forest semantic segmentation results on FOR-instanceV2 test set samples when 0.81 ha labelled training data is used.

State-of-the-art instance segmentation

We evaluate individual-tree instance segmentation and its auxiliary semantic segmentation on the FOR-instanceV2 benchmark.

Benchmark comparison

On the full training set, LitePT-scratch slightly outperforms ForPT-finetune, while both variants surpass the previous best results in F1 score for individual tree segmentation and mIoU for semantic segmentation.

	Individual tree seg.					Semantic seg.
Method	mAP ↑	mAP25 ↑	mAP50 ↑	F1 ↑	Cov ↑	Ground ↑	Wood ↑	Leaf ↑	mIoU ↑
ForAINetV2_R8	34.5	65.1	55.2	72.4	73.3	98.3	68.1	94.3	86.9
ForAINetV2_R16	37.3	63.8	54.8	70.8	72.7	98.8	69.2	94.5	87.5
TreeLearn	19.6	41.4	30.9	50.3	52.5	–	–	–	–
OneFormer3D	48.7	71.8	65.7	73.1	80.0	97.8	63.1	93.4	84.8
ForestFormer3D	58.3	79.5	74.8	83.5	81.6	98.6	65.5	93.9	86.0
ForestMamba	54.9	77.7	72.0	82.0	76.6	98.5	65.4	93.9	85.9
SegmentAnyTreeV2	–	–	–	83.4	89.6	99.1	70.5	93.9	87.8
LitePT-scratch (ours)	61.3	82.0	78.4	85.5	81.6	98.4	71.9	94.6	88.3
ForPT-finetune (ours)	60.6	81.4	77.6	84.7	81.1	98.5	71.4	94.4	88.1

Qualitative comparison

In the low-data regime, fine-tuning still provides a clear advantage over training from scratch. When only 1.35 ha labelled training data is used, both training from scratch and fine-tuning produce reasonable segmentation results, while fine-tuning yields more accurate instance predictions, correcting under-segmented trees, reducing incorrect instance assignments, and better preserving fine-grained details.

Qualitative forest instance segmentation results on FOR-instanceV2 test set samples when only 1.35 ha labelled training data is used.

State-of-the-art species classification

We evaluate individual-tree species classification on the FOR-species20K benchmark (34 species).

Benchmark comparison

LitePT-scratch and ForPT-finetune outperform all previous methods on this benchmark. LitePT-scratch achieves an overall accuracy of 82.7%, while fine-tuning the pretrained ForPT backbone further increases the accuracy to 83.2%, achieving the best performance across all reported metrics.

Method	Input	OA ↑	Precision ↑	Recall ↑	F1 ↑
SimpleView	Image	76.2	76.9	75.5	75.6
YOLOv5	Image	77.9	84.2	75.0	77.3
DetailView	Image	79.5	82.3	76.7	78.0
PointAugment + DGCNN	Point cloud	68.3	72.5	65.7	70.3
PointMixer	Point cloud	71.1	74.4	65.5	71.1
MinkNet	Point cloud	73.7	79.9	70.6	72.3
Ensemble PointNet++	Point cloud	75.6	78.2	73.5	74.9
LitePT-scratch (ours)	Point cloud	82.7	85.3	81.6	82.5
ForPT-linear probing (ours)	Point cloud	68.7	74.2	66.0	67.8
ForPT-finetune (ours)	Point cloud	83.2	86.0	81.8	83.0

Training dynamics

Pretraining accelerates convergence — fine-tuning matches scratch's full 100-epoch accuracy by epoch 54, nearly twice as fast.

Long-tail robustness

Fine-tuning helps most where data is scarce — classes with the fewest training samples see the largest accuracy gains, while common species are largely unaffected, mitigating the long-tail effect.

Per-class accuracy gain of fine-tuning over training from scratch at epoch 20.

State-of-the-art age regression

We evaluate individual-tree age regression on the FOR-age benchmark.

Benchmark comparison

LitePT-scratch achieves an RMSE of 20.41 years and an R² of 0.74, already outperforming the previously reported PTv3-scratch and ForestFormer3D-finetune results. Fine-tuning the pretrained ForPT backbone improves the performance further, reducing the RMSE to 20.03 years and increasing R² to 0.75, thereby achieving the best overall performance.

Method	RMSE (years) ↓	MD (years) → 0	R² ↑
Linear regression	34.15	-2.68	0.28
PTv3-scratch	22.80	-3.96	0.68
ForestFormer3D-finetune	20.71	2.37	0.74
LitePT-scratch (ours)	20.41	-0.55	0.74
ForPT-linear probing (ours)	30.65	-1.10	0.42
ForPT-finetune (ours)	20.03	1.39	0.75

Predicted vs. reference age

We show the scatterplots of the predicted vs reference age when 625 trees (50%) are used for training. Fine-tuning produces predictions most closely aligned with the 1:1 reference line, resulting in the lowest RMSE (21.61 years) and highest R² (0.75). Training from scratch exhibits a larger spread around the reference line, whereas linear probing shows greater underestimation of older trees.

Scatterplots of the predicted vs reference age on the FOR-age validation set when 625 trees are used for training.

We probe frozen representations from four task-specific supervised pretrained backbones (semantic segmentation, instance segmentation, species classification, age regression) on every other downstream task, and compare them against our self-supervised (SSL) pretrained backbone under the same protocol.

Transfer consistency

Across all four downstream tasks, self-supervised pretraining consistently produces the most transferable representations, outperforming task-specific supervised pretraining under the same probing protocol.

		Target task (frozen probing)
		Sem. Seg.(mIoU ↑)	Inst. Seg.*(mAP ↑)	Species Cls.(allAcc ↑)	Age Reg.(RMSE ↓)
Source (pretrained) model	SSL	84.2	38.7	72.8	29.1
	Sem. Seg.	—	—	68.8	30.9
	Inst. Seg.	—	—	70.1	30.6
	Species Cls.	82.2	22.4	—	30.8
	Age Reg.	79.5	24.6	70.9	—

Within-column rank:

1 (best) → 4 (worst)

Values are the raw metric at the full (100%) label budget. Blank cells are in-domain pairs (a source probed on its own pretraining task) or the semantic/instance segmentation pair, which share a backbone and are not an independent transfer test. * Instance segmentation is probed with a task head rather than a linear layer.

Data efficiency of frozen transfer

Self-supervised representations provide more consistent transfer than task-specific pretraining across diverse downstream objectives and annotation budgets, supporting their use as a general-purpose initializationfor forest understanding tasks.

Semantic segmentation

Instance segmentation

Species classification

Age regression

Part of the compute is supported by the Swiss AI Initiative under projects a0182 and a144 on the Alps supercomputer. The project is supported by the Circular Bio-based Europe Joint Undertaking (CBE JU) and its members under Grant Agreement No 101157488 (SingleTree). This project is also funded by the EU Horizon Europe program under Grant Agreement No 101213369 (DVPS) and Grant Agreement No 101131841 (Embed2Scale). Additional funding for Embed2Scale has been provided by the Swiss State Secretariat for Education, Research and Innovation (SERI) and UK Research and Innovation (UKRI). Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or CBE JU. Neither the European Union nor the CBE JU can be held responsible for them.