{"id":"79050cbf-9b3a-413d-8c1c-cc4f58df10c1","arxiv_id":"2607.06948","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":6,"one_line_summary":"Self-supervised pretraining of Point-M2AE on ShapeNet-55 plus 2,400 tree point clouds improves cross-site leaf-wood segmentation and halves QSM-derived volume estimation error versus algorithmic baselines.","lead":"This paper shows that self-supervised pretraining on a mix of synthetic 3D objects and real tree scans substantially improves leaf-wood segmentation in forest point clouds, cutting cross-site performance variation and halving wood-volume estimation error versus algorithmic baselines. A generalist reader might care because better wood-leaf separation directly improves aboveground biomass estimates from terrestrial laser scanning, which underpin carbon accounting.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"KR test-set trees appear in ShapeNet+KR pretraining but not ShapeNet-only pretraining, creating asymmetric representation leakage that inflates the primary quantitative evidence for adding tree data.","rationale":"The reader correctly identified the most load-bearing concern: the dual use of KR data for both pretraining and fine-tuning creates a leakage confound that specifically inflates the Table 2 comparison, which is the paper's primary quantitative evidence for the benefit of adding tree data to pretraining. I agree with the reader's CONDITIONAL verdict. The external benchmark (Table 3) and volume estimation (Table 5) provide independent support that is not affected by this leakage, which is why the paper is not fatally undermined — but the external benchmark gaps are smaller and lack significance testing, and the volume estimation shows negligible difference between the two pretrained models (2.45 vs 2.40 m³). The paper's strongest independent evidence is actually the comparison of any SSL-pretrained model vs the randomly-initialized model and vs algorithmic baselines, which holds up well. The weaker link is the marginal benefit of adding KR tree data specifically, which is the part most affected by the leakage concern. The concrete test proposed (excluding test trees from pretraining) is straightforward to implement given that the authors have released code and data, and would directly settle whether the Table 2 gains are real or inflated. If the gains persist, the paper's claims are fully supported; if they shrink to the external benchmark level, the paper should reframe the KR-data contribution as modest rather than substantial. Either way, the broader claim that SSL pretraining improves leaf-wood segmentation remains well-supported by the external benchmarks.","tokens_in":21330,"tokens_out":2664,"duration_ms":143385,"concrete_test":"Retrain the ShapeNet+KR encoder using only the KR training+validation split (excluding the KR test set trees) for pretraining augmentation. Re-evaluate on the KR test set and compare wood IoU against the ShapeNet-only model. If the wood IoU gap between ShapeNet+KR and ShapeNet-only shrinks to near the external benchmark gap (~1.8 points) or disappears, the Table 2 improvements were inflated by pretraining leakage. If the gap persists at 2.8–4.1 points, the concern does not land.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central quantitative claim — that adding KR tree data to pretraining improves wood IoU from 67.2%→70.0% (needleleaf) and 72.2%→76.3% (broadleaf) over ShapeNet-only (Table 2) — rests on a comparison where the ShapeNet+KR encoder has seen the KR test trees' geometry during self-supervised pretraining, while the ShapeNet-only encoder has not. The KR test set is held out from fine-tuning (7:2:1 split, §2.2), but all 2,400 KR trees were included in the ShapeNet+KR pretraining pool (§2.1). SSL pretraining learns representations from raw point coordinates without labels, so the encoder has already encoded the test trees' branching geometry, creating an asymmetric advantage for the ShapeNet+KR model on the KR test set. This does not affect the external benchmark (Table 3: 6 sites, 58 trees never seen during pretraining or fine-tuning), where the ShapeNet+KR model still outperforms ShapeNet-only but by a smaller margin (wood IoU 78.5% vs 76.7%, a 1.8-point gap vs the 2.8–4.1-point gaps on the KR test set). No significance testing is reported for any comparison, so it is unclear whether the external benchmark differences are statistically meaningful. The volume estimation results (Table 5) show essentially no difference between ShapeNet-only and ShapeNet+KR (MAE 2.45 vs 2.40 m³), further suggesting that the KR test set gains may be inflated. The paper acknowledges the KR data is always combined with ShapeNet (§4.3) but does not discuss the test-set leakage from pretraining overlap.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"This manuscript investigates whether self-supervised pretraining (SSL) improves leaf-wood segmentation of terrestrial laser scanning (TLS) point clouds. The authors pretrain Point-M2AE on ShapeNet-55 augmented with 2,400 individual tree point clouds (KR dataset), then fine-tune for leaf-wood segmentation using voxel-based input partitioning. They evaluate on a multi-site benchmark (58 trees across 6 sites in 4 countries), at plot level, and on a downstream QSM-based volume estimation task (28 tropical trees). The pretrained model improves wood IoU, reduces cross-site variation, and achieves the lowest volume estimation error among compared methods. The central claim that SSL pretraining improves cross-site generalization is supported by the external benchmark, though the specific contribution of adding KR tree data to pretraining is confounded by data reuse between the pretraining pool and the fine-tuning/test split.","tokens_in":21813,"tokens_out":2359,"duration_ms":160514,"significance":"The paper addresses a practical and important problem: leaf-wood segmentation accuracy varies across forest types and sites, and existing methods rely on handcrafted geometric features with limited transferability. Applying SSL pretraining to this task is novel and well-motivated. The recursive voxel subdivision strategy enabling cross-scale inference without architecture change is a practical contribution. The multi-site benchmark spanning temperate, boreal, and tropical forests is broader than most prior evaluations. The public release of code, pretrained weights, and fine-tuned models is a notable strength that enhances reproducibility. The downstream volume estimation evaluation against destructive measurements provides end-to-end validation that is rare in this subfield.","major_comments":[{"comment":"§2.1, §2.2, Table 2: The KR dataset is used both for SSL pretraining augmentation (all 2,400 trees) and for fine-tuning/testing (7:2:1 split). The KR test set trees (10% of KR) therefore appear in the ShapeNet+KR pretraining pool but not in the ShapeNet-only pretraining pool. Since SSL pretraining learns representations from raw point coordinates without labels, the ShapeNet+KR encoder has encoded the test trees' geometry during pretraining, creating an asymmetric representation advantage on the KR test set. This inflates the primary quantitative evidence for adding tree data to pretraining (Table 2: wood IoU gains of 2.8-4.1 points on KR test set). The authors acknowledge in §4.3 that KR is always combined with ShapeNet, but do not discuss this specific test-set leakage from pretraining. The external benchmark (Table 3, 58 trees never seen during pretraining or fine-tuning) shows a much","section":null},{"comment":"§2.1, §2.2, Table 2: The KR dataset is used both for SSL pretraining augmentation (all 2,400 trees) and for fine-tuning/testing (7:2:1 split). The KR test set trees (10% of KR) therefore appear in the ShapeNet+KR pretraining pool but not in the ShapeNet-only pretraining pool. Since SSL pretraining learns representations from raw point coordinates without labels, the ShapeNet+KR encoder has encoded the test trees' geometry during pretraining, creating an asymmetric representation advantage on the KR test set. This inflates the primary quantitative evidence for adding tree data to pretraining (Table 2: wood IoU gains of 2.8-4.1 points on KR test set). The authors acknowledge in §4.3 that KR is always combined with ShapeNet, but do not discuss this specific test-set leakage from pretraining. The external benchmark (Table 3, 58 trees never seen during pretraining or fine-tuning) shows a much","section":null},{"comment":"smaller margin (wood IoU 78.5% vs 76.7%, a 1.8-point gap), and the volume estimation results (Table 5) show essentially no difference between ShapeNet-only and ShapeNet+KR (MAE 2.45 vs 2.40 m³). The authors should either (a) exclude KR test trees from the pretraining pool and re-evaluate, or (b) explicitly discuss this as a limitation and reframe the KR test set results accordingly, emphasizing the external benchmark as the primary evidence for cross-site generalization.","section":null},{"comment":"§2.3, Tables 2-5: No statistical significance testing is reported for any comparison. The external benchmark differences between ShapeNet+KR and ShapeNet-only (Table 3: wood IoU 78.5% vs 76.7%) and between ShapeNet+KR and PointTransformer (Table 3: wood IoU 78.5% vs 74.4%) are based on 58 trees across 6 sites, with some sites having only 4-6 trees (Table 1: SWDEU-NL has 4 trees, SWDEU-BL has 6 trees, FIN-NL has 12 trees). With such small per-site sample sizes, it is unclear whether the reported differences are statistically meaningful or within noise. The authors should report confidence intervals or significance tests for the key cross-site comparisons, or at minimum acknowledge this limitation.","section":null},{"comment":"§2.1, Table 1: The benchmark dataset is small and imbalanced across sites. The 58 individual trees are distributed across 6 sites with 4-15 trees per site, and the plot-level evaluation uses only 1 plot per leaf type (CAM-BL-PL: 15 trees, FIN-NL-PL: 15 trees). The plot-level results (Table 4) are thus based on n=1 plot per leaf type, making it difficult to assess whether the reported plot-level performance generalizes. The authors should acknowledge this as a limitation in §4.3 and clarify that plot-level conclusions are preliminary.","section":null},{"comment":"§3.3, Table 4: For the needleleaf plot (FIN-NL-PL), the ShapeNet+KR-pretrained model performs worse than the ShapeNet-pretrained model across all metrics (e.g., mIoU 77.7% vs 78.5%, wood IoU 64.3% vs 65.2%). The authors attribute this to increased leaf-to-wood misclassification, but this result partially contradicts the central claim that adding KR tree data to pretraining improves segmentation. The abstract states 'mIoU of 84.7% for broadleaf and 77.7% for needleleaf plots' without noting that the ShapeNet-only model achieves higher mIoU (78.5%) on the needleleaf plot. The abstract should be amended to accurately reflect this mixed result.","section":null}],"minor_comments":[{"comment":"§2.2, Eq. (1): The Chamfer Distance loss is labeled as Equation (1), but the cross-entropy loss in the same section is also labeled as Equation (2). However, in §2.1, the truncated-cone formula is also labeled as Equation (1). The equation numbering is inconsistent across sections.","section":null},{"comment":"Abstract: The abstract states the pretrained model improved wood IoU 'from 60.5% to 70.0% for needleleaf and from 69.7% to 76.3% for broadleaf trees.' These are KR test set results (Table 2), which are subject to the pretraining data leakage concern raised in Major Comment 1. The abstract should clarify that these are in-domain results, or lead with the external benchmark results.","section":null},{"comment":"§2.1: The country codes in the benchmark description are introduced without a key. While Table 1 provides the mapping, the main text in §2.1 refers to 'SWDEU-BL,' 'DEU-BL,' 'FIN-NL,' etc. without immediately explaining the naming convention. A brief note on first use would improve readability.","section":null},{"comment":"§2.3: The TreeQSM cover set parameters are listed as 'PatchDiam1 = 1, PatchDiam2Min = 1, PatchDiam2Max = 1.' It would help to briefly justify why the default parameters were used and whether any sensitivity analysis was performed.","section":null},{"comment":"Figure 2: The figure caption refers to panels (a)-(l) but the description of zoomed panels is somewhat unclear. Consider explicitly stating which panels are zoomed views and which show the full tree.","section":null},{"comment":"§4.1: The statement 'Deep learning models are generally expected to improve logarithmically with pretraining data volume (Sun et al., 2017)' is used to argue that a 4% data increase should yield small gains. This scaling law applies to supervised learning with labeled data; its applicability to SSL pretraining with unlabeled data is less established. The argument could be softened.","section":null},{"comment":"Table 5: The relative MAE percentages are shown in parentheses but the column headers do not indicate these are relative MAE values. Adding '%' to the column headers or a footnote would improve clarity.","section":null},{"comment":"§2.1: The reference to 'Seely et al. (2026)' appears to be a future-dated publication. If this is a preprint or forthcoming paper, the citation format should be clarified.","section":null},{"comment":"Figure A2: The y-axis label 'patch coverage' is defined as 'the product of patch size and patch number' but the units or typical ranges are not immediately clear from the figure. A brief note in the caption would help.","section":null},{"comment":"§3.4: One tree (MDD01_006) from PER was excluded due to unstable QSM reconstruction. The exclusion criteria (SD = 2.70 m³, 3.6× the mean SD) are reasonable, but it would be useful to report how the results change if this tree is included, to confirm the exclusion does not materially affect the conclusions.","section":null}],"recommendation":"major_revision","confidential_remarks":"The data leakage concern is the most serious issue. The KR test set appearing in the ShapeNet+KR pretraining pool but not the ShapeNet-only pool creates an asymmetric advantage that directly inflates the paper's primary quantitative evidence (Table 2). The external benchmark (Table 3) and volume estimation (Table 5) partially mitigate this, but the margins there are smaller and lack significance testing. The authors can likely address this by either re-running with KR test trees excluded from pretraining, or by reframing the paper to emphasize the external benchmark as the primary evaluation. The core contribution — that SSL pretraining improves cross-site leaf-wood segmentation — is defensible based on the external benchmark, but the specific claim about the benefit of adding tree data to pretraining is not cleanly supported by the current experimental design. I would encourage the editor to require at minimum a clear acknowledgment and reframing if re-running experiments is not feasible before the next major revision cycle."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The headline: this paper applies SSL pretraining (Point-M2AE) to TLS leaf-wood segmentation for the first time, and the external evidence — six benchmark sites across four countries plus downstream QSM volume estimation on 28 tropical trees with destructive references — is solid enough to take seriously. Code and weights are public. The cross-site consistency story is the real contribution: the pretrained model has the lowest standard deviation across sites among all six methods compared, and the volume MAE (2.40 m³) is less than half the algorithmic baselines. That part holds up because those trees were never seen during pretraining or fine-tuning. Credit earned there. The stress-test concern about KR test-set leakage is real and lands hard on the paper's primary quantitative claim. All 2,400 KR trees went into the ShapeNet+KR pretraining pool, including the 10% held out from fine-tuning. SSL learns from raw coordinates without labels, so the encoder has already encoded the test trees' geometry. This means the KR test-set comparison (Table 2: wood IoU 67.2→70.0 needleleaf, 72.2→76.3 broadleaf) is asymmetric — the ShapeNet+KR model has seen those trees, the ShapeNet-only model has not. The paper never discusses this. The external benchmark (Table 3) is the cleaner test, and there the ShapeNet+KR advantage shrinks to 1.8 points (78.5 vs 76.7 wood IoU) with no significance testing reported. The volume results (Table 5) show essentially no difference between ShapeNet-only and ShapeNet+KR (2.45 vs 2.40 m³), which further suggests the KR test-set gains are inflated. The benchmark is also small — 58 trees, some sites with 4-6 trees — so those 1.8-point differences could be noise. The recursive voxel subdivision for cross-scale inference is a nice engineering detail that lets the same model handle individual-tree and plot-level inputs without architecture changes. The plot-level results (Table 4) are a useful proof of concept, though the needleleaf plot actually does worse with KR augmentation than with ShapeNet-only pretraining, which complicates the story. This paper is for researchers working on TLS-based forest structure and biomass estimation. It deserves a serious referee because the external benchmark and downstream volume evaluation provide genuine evidence that SSL pretraining helps, even if the KR test-set comparison needs to be either re-run with test trees excluded from pretraining or explicitly flagged as not a clean comparison. The authors should also add significance tests for the external benchmark differences. I'd accept this for peer review conditional on the data-leakage issue being addressed.","headline":"SSL pretraining for TLS leaf-wood segmentation works, but the KR test-set gains are inflated by pretraining overlap with test trees","tokens_in":22427,"tokens_out":629,"would_cite":true,"duration_ms":119286,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Pretraining on tree shapes halves wood volume error in forest scans","keywords":[],"falsifier":"If a model pretrained on ShapeNet-55 alone (without tree data) matched or exceeded the ShapeNet+KR-pretrained model on cross-site wood IoU and volumetric MAE, the central claim that domain-relevant pretraining data drives the improvement would be undermined.","tokens_in":21565,"feed_emoji":"🌳","tokens_out":838,"duration_ms":138096,"temperature":0.7,"pith_summary":"Separating leaf points from wood points in terrestrial laser scans of forests is a prerequisite for estimating tree volume and biomass, but existing methods lose accuracy when applied across different forest types, climates, and scanning densities. This paper shows that an encoder pretrained via masked reconstruction on a mix of synthetic 3D objects and 2,400 tree point clouds, then fine-tuned for leaf-wood segmentation, consistently outperforms both algorithmic and fully supervised baselines across six sites in four countries. The key finding is that even a small amount of domain-relevant data added to pretraining (roughly 4% of the pool) drives disproportionate gains in wood IoU and cross-site consistency, and that these segmentation gains propagate downstream: wood volume estimates from quantitative structure models fitted to the segmented wood points achieve the lowest mean absolute error (2.40 m³) among all tested methods, less than half that of algorithmic baselines.","feed_headline":"Pretraining on tree shapes halves wood volume error in forest scans","feed_subtitle":"Adding 2,400 tree point clouds to self-supervised pretraining cuts cross-site segmentation variance and halves volumetric estimation error,示","key_machinery":"Point-M2AE, a masked autoencoder for point clouds that reconstructs masked regions using Chamfer Distance loss, pretrained on ShapeNet-55 augmented with tree point clouds, then fine-tuned with a PointNet++-style segmentation head. At inference, recursive voxel subdivision partitions large or dense point clouds into fixed-size inputs without architecture change, enabling the same model to operate on both individual trees and full forest plots.","core_discovery":"Adding 2,400 tree point clouds (a 4% increase) to a synthetic-object pretraining pool yields wood IoU gains exceeding half of the total improvement from pretraining itself, and these gains transfer to unseen sites and to downstream volumetric estimation. The mechanism is that masked-reconstruction pretraining shifts the encoder toward preserving fine woody structures in dense canopy regions where leaf and wood points intermingle, reducing the systematic wood-to-leaf misclassification that propagates into volume overestimation or underestimation.","pith_inferences":[],"forward_implications":["Forest biomass inventories that rely on terrestrial laser scanning could adopt SSL-pretrained segmentation as a drop-in preprocessing step, reducing site-specific annotation requirements and improving volume estimates in tropical forests where destructive calibration data are scarce.","The disproportionate gain from a small domain-specific pretraining augmentation suggests that pretraining data curation for forestry 3D tasks should prioritize structural relevance over dataset volume, which has implications for other point cloud tasks in ecology and remote sensing.","Recursive voxel subdivision without retraining provides a practical pathway for deploying fixed-input-size architectures on operationally variable point clouds, a pattern that could extend to urban scanning, infrastructure inspection, or any domain where point density varies by orders of magnitude.","The finding that improved fine-branch recovery can increase volumetric error against destructive references that exclude small branches highlights a mismatch between segmentation quality and the ground truth used for evaluation, motivating reference standards that explicitly account for branch-size thresholds."],"fun_headline_variants":["Self-supervised pretraining boosts leaf-wood segmentation across forests","SSL pretraining cuts cross-site variance in tree point cloud segmentation","Pretrained model halves wood volume error in tropical forest scans","Tree point clouds in SSL pretraining improve wood segmentation IoU","Voxel subdivision lets one model segment trees at plot and individual scales"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The claim that adding tree point clouds to pretraining drives the improvement rests on a confound: the tree data was always combined with synthetic objects during pretraining, never tested alone, so the independent contribution of tree geometry cannot be separated from the combined effect. Additionally, the same tree point clouds used for pretraining augmentation were also used for fine-tuning, which means the pretraining stage may have already seen the fine-tuning domain.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised pretraining boosts leaf-wood segmentation across forests","SSL pretraining cuts cross-site variance in tree point cloud segmentation","Pretrained model halves wood volume error in tropical forest scans","Tree point clouds in SSL pretraining improve wood segmentation IoU","Voxel subdivision lets one model segment trees at plot and individual scales"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":785,"prompt_tokens":716,"completion_tokens":69,"prompt_tokens_details":null},"tokens_in":716,"tokens_out":69,"duration_ms":67640,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T22:21:58.959406+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If a model pretrained on ShapeNet-55 alone (without tree data) matched or exceeded the ShapeNet+KR-pretrained model on cross-site wood IoU and volumetric MAE, the central claim that domain-relevant pretraining data drives the improvement would be undermined.","supporting_citations":[],"review_version":1}