{"id":"d294b9d6-f5fe-45de-a00f-cb497ac2fc9c","arxiv_id":"2506.08061","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"Per-tree canopy volume can be estimated from mobile LiDAR in both regular and overlapping orchards using DBSCAN plus spectral clustering for segmentation.","lead":"A robot-driven LiDAR system measures each tree's canopy volume while moving through orchards. The method combines density-based and graph-based clustering so it can handle both evenly spaced pistachio trees and crowded, overlapping almond trees.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central volume-accuracy claim is unsupported: Table I's error percentages do not match the listed volumes, so the paper's only per-tree validation is internally inconsistent.","rationale":"The reader's weakest assumption was the drone-based spherical ground truth, which is a valid concern. My stress-test identifies a more fundamental problem: Table I fails internal consistency checks before ground-truth fidelity even becomes relevant. The volume errors reported in Table I are internally contradictory, so the central claim of 'strong agreement' rests on uninterpretable data. The segmentation success percentages (93%, 80%, 46%) are also loosely defined without counts or a failure taxonomy, but the volume claim is the headline contribution. Since the evidence as published cannot support the stated quantitative conclusion, I agree with the reader's reject verdict and do not recommend changing it. However, my primary concern is the invalid table rather than the spherical ground-truth assumption; the two are related because a corrected table would still need to address the idealized ground-truth model before the volume claim could be accepted.","tokens_in":6311,"tokens_out":3410,"duration_ms":35508,"concrete_test":"Recompute every row of Table I by taking the absolute difference between each listed predicted volume and the listed ground truth, divided by the ground truth. If more than one row disagrees with the printed error percentage by more than a rounding tolerance (e.g., 1 percentage point), the table cannot support the claimed 'strong agreement,' and the authors should supply a corrected table or release the raw data for independent recalculation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim is 'strong agreement to drone derived canopy volume estimates,' and the only evidence for that claim is Table I. That table is internally inconsistent in ways that cannot be explained by rounding. For example, row 9 lists ground truth 31.78 m³, Alpha Shape 6.44 m³, and Alpha Shape Error 0.29%, but the true error is about 80%; row 13 lists ground truth 9.53 m³, Convex Hull 11.38 m³, and Convex Hull Error 10.89%, while the true error is 19.4%, and its Alpha Shape volume of 19.48 m³ is labeled 14.35% error when the true error is 104%. Many rows show similar mismatches (rows 8, 10, 11, and 15). Because no code, data, or derivation is provided, a reader cannot reconstruct which columns are intended. This is not a minor typo: the printed percentages are the only per-tree evidence for volume accuracy. In addition, the validation depends on drone orthomosaic diameters converted to spherical volumes, which Section III-A concedes is idealized and valid only for approximately symmetrical crowns, and no almond volume validation is presented. If Table I cannot be trusted, the volume-estimation claim is unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a mobile LiDAR-based pipeline for per-tree canopy volume estimation in orchards, combining LiDAR-inertial odometry, RANSAC ground removal, DBSCAN clustering (with an additional spectral clustering refinement for dense almond orchards), and geometric volume reconstruction via Convex Hull and Alpha Shape. The system is evaluated in two commercial orchards: a pistachio orchard where DBSCAN alone is reported to achieve 93% segmentation success, and an almond orchard where a hybrid DBSCAN-plus-spectral strategy is reported to raise segmentation success from 46% to 80%. The abstract claims \"strong agreement\" between LiDAR-derived volumes and drone-derived ground truth, with Table I presented as the supporting per-tree evidence. The paper also describes tree labeling, traversal speed, and discusses limitations and future work.","tokens_in":6572,"tokens_out":1695,"duration_ms":20721,"significance":"If the claims were substantiated, the work would provide a practically valuable contribution: real-time, per-tree canopy volume estimation from a mobile platform during routine navigation, without static scans, and with adaptation to structurally diverse orchards. The system targets an important precision-agriculture need (scalable canopy monitoring for yield and harvest optimization), and the combination of DBSCAN with spectral clustering for overlapping canopies is a sensible and potentially reusable idea. The paper also makes its intended scope clear by honestly acknowledging that almond volume validation was not performed. However, the actual quantitative evidence for the central volume-accuracy claim is internally inconsistent, and the segmentation parameters are selected after observing the same data on which success rates are measured; these issues currently leave the core claims unverified.","major_comments":[{"comment":"Table I contains multiple internally inconsistent error values that cannot be explained by rounding. For example, Tree 9 lists ground truth 31.78 m^3, convex hull 33.83 m^3, and convex hull error 31.87%, whereas the true error is (33.83-31.78)/31.78 = 6.4%; the same row lists alpha shape volume 6.44 m^3 with error 0.29%, whereas the true error is about 79.7%. Tree 13 lists ground truth 9.53 m^3, convex hull 11.38 m^3, and convex hull error 10.89%, but the true error is 19.4%, and its alpha shape volume of 19.48 m^3 is labeled 14.35% when the true error is 104.4%. Similar mismatches appear in rows 8, 10, 11, and 15. Because Table I is the only per-tree evidence for the abstract's claim of \"strong agreement to drone derived canopy volume estimates,\" and because no code, data, or derivation is provided to reconstruct the intended columns, the central volume-accuracy claim is not supported by the presented evidence.","section":"Table I"},{"comment":"The segmentation parameters are selected after observing the data, and success rates are measured on the same data. Section II-C defines epsilon = 0.8 m, minPts = 1300, the 45,000-point cluster threshold, and k = 10 for spectral clustering as fixed values, while Section III-B reports that plain DBSCAN gives 46% success and the enhanced method gives 80% success in the almond orchard, with no held-out validation or cross-validation. Because the thresholds appear to be tuned to the specific orchards rather than derived from a general rule or an independent training set, the reported percentages may reflect overfitting and the claimed generalizability is not established. A concrete test would be to fix the pipeline parameters on one orchard and then evaluate on a second orchard without further tuning, or to report parameter sensitivity analysis.","section":"Section II-C and Section III-B"},{"comment":"The ground truth for canopy volume is based on drone photogrammetry processed into an orthomosaic, from which canopy diameters are converted to spherical volumes. Section III-A concedes this approximation is \"idealized\" and valid only when crowns are approximately symmetrical. Since many of the reported trees (e.g., Trees 8, 9, 13, 15) have volumes or errors that appear anomalous, and since no almond volume validation is presented at all, the numerical accuracy of the LiDAR volume estimates against a truly independent and physically meaningful reference remains unverified. The spherical assumption is a load-bearing limitation because the stated objective is per-tree canopy volume estimation, and the only numerical evidence for that objective depends on a model that is known to be inaccurate for non-symmetrical crowns.","section":"Section III-A"},{"comment":"The paper does not report any uncertainty quantification, repeatability analysis, or comparison with a direct manual or destructive measurement of canopy volume. Given that the stated contribution is quantitative volume estimation, the lack of any error bars, standard deviations, or repeated traversals leaves the reader unable to judge whether the reported \"errors\" are within acceptable agronomic tolerances. This is particularly concerning because the same mobile platform was used in both orchards, yet no information is given about how many runs were performed, how tree labels were matched across runs, or how partial scans and occlusion affected the completeness of individual tree point clouds.","section":"General (Section III)"}],"minor_comments":[{"comment":"The title contains a typo: \"Canopy V olume\" should be \"Canopy Volume.\"","section":"Title and Abstract"},{"comment":"The alpha shape radius is stated as alpha = 0.9 m without any sensitivity analysis or justification; a sentence explaining how this value was chosen (e.g., relative to point spacing or canopy size) would improve reproducibility.","section":"Section II-E"},{"comment":"The traversal speed paragraph reports times \"around 50 seconds\" and \"around 110 seconds\" but does not state the row lengths or the number of trees per row, making the speed claim difficult to interpret.","section":"Section II-F"},{"comment":"Table I reports \"seventeen representative trees\" but does not state how these were selected or whether they are consecutive trees in a row or a random sample; this selection bias could affect the reported error statistics.","section":"Section III-A"},{"comment":"Reference [17] appears to be an unrelated arXiv preprint on air-ground collaboration and language-specified missions; if it is cited as a related LiDAR or agricultural application, the citation is inaccurate or at least needs clarification.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript is a workshop-style paper with a potentially useful engineering idea, but the primary quantitative evidence is internally inconsistent (Table I), and the validation protocol is not sufficient to support the central claim. The issues are load-bearing and cannot be fixed with small edits: the authors would need to recompute the volume-error table, provide the data or code, and ideally add a held-out evaluation or a second dataset with proper ground truth. As submitted, the paper does not meet the bar for publication even after minor revision. I would suggest rejection, with encouragement for the authors to resubmit after substantial validation work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a reasonable idea — DBSCAN plus spectral clustering to split overlapping almond canopies from mobile LiDAR — and reports a nice jump in segmentation success. But the volume validation is the load-bearing part of the abstract, and Table I is arithmetically inconsistent in multiple rows. Until that table is fixed, the 'strong agreement' claim should not be taken at face value.\n\nWhat's actually new: the split-then-refine segmentation for dense, overlapping crowns is a sensible engineering adaptation. Standard components, but the combination for almond orchards is a legitimate extension. The fact that they run it on a moving platform during normal navigation is practically useful. The segmentation numbers (93% pistachio, 80% almond vs 46% with plain DBSCAN) are plausible, though I'd want a precise definition of 'success' and how the 17 trees were selected.\n\nSoft spots, in order of seriousness. First, Table I: row 9 lists ground truth 31.78, convex hull 33.83, and calls that 31.87% error; true error is about 6%. Alpha shape 6.44 with 0.29% error; true error is about 80%. Similar mismatches appear in rows 13, 10, 11, 14, and 15. This is not rounding. Without code or a derivation, a reader cannot tell what those percentages mean. That makes the central volume claim unverifiable as printed. Second, the ground truth is drone photogrammetry turned into spherical volumes from orthomosaic diameters; the paper itself concedes this is idealized and only valid for roughly symmetric crowns. That is a weak yardstick for canopy volume. Third, no almond volume validation is presented, so the harder environment is only shown for segmentation. Fourth, no code or data release, and parameters appear chosen after inspecting the same data. These last points are common in workshop papers, but they compound the table problem.\n\nWho gets value: people working on LiDAR-based orchard phenotyping and precision-ag robotics. The segmentation adaptation might be worth trying in their own systems. As a peer reviewer, I'd send it back for major revision: fix the table, clarify the success metric, report error bars, and ideally release a small dataset. The idea is not broken, but the evidence as printed does not support the abstract's volume claim.","headline":"The segmentation idea is reasonable, but Table I's volume numbers don't add up, so the paper's central claim is unsupported as printed.","tokens_in":7122,"tokens_out":2702,"would_cite":false,"duration_ms":29750,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Per-tree canopy volume can be estimated in real time from a robot's routine mobile LiDAR scans, provided segmentation adapts to orchard density — 93% success in pistachio, 80% in almond after a spectral-clustering split step.","keywords":["canopy volume estimation","mobile LiDAR","per-tree segmentation","DBSCAN","spectral clustering","precision agriculture","orchard mapping","Alpha Shape"],"falsifier":"Measure the same pistachio trees with a static, high-resolution terrestrial LiDAR scanner that sees the full crown, and check whether the mobile-LiDAR Alpha Shape volumes match the drone-sphere volumes within the errors reported in Table I; alternatively, run the almond two-stage pipeline on the pistachio rows and see if the added spectral splitting degrades the 93% segmentation, which would show the 46% to 80% gain is not a generalizable result.","tokens_in":6078,"feed_emoji":"🌳","tokens_out":6515,"duration_ms":72336,"temperature":0.7,"pith_summary":"This paper claims that per-tree canopy volume can be measured in real time from LiDAR data a robot already collects while driving through an orchard, eliminating the need for static scans. The key adaptivity is in segmentation: uniform pistachio rows are handled by density-based DBSCAN clustering alone (93% of trees correctly segmented), while dense almond orchards with overlapping crowns need an added spectral-clustering step that splits large merged clusters, raising segmentation success from 46% to 80%. Volumes built with the Alpha Shape method agreed with drone-photogrammetry estimates in the pistachio test, where the paper notes its aerial ground truth is only reliable for roughly symmetric crowns. If the claim holds, orchard operators could get per-tree growth, vigor, and yield-relevant volume data from ordinary navigation passes.","feed_headline":"Adaptive clustering lifts almond tree segmentation from 46% to 80%","feed_subtitle":"A DBSCAN-plus-spectral pipeline turns routine robot scans into per-tree canopy volume estimates in dense orchards.","key_machinery":"The load-bearing mechanism is a two-stage clustering cascade. First DBSCAN (ε=0.8 m, min points 1300) produces initial clusters, with voxel downsampling and RANSAC ground removal as preprocessing. Second, any cluster above 45,000 points is treated as a merged tree group; spectral clustering on a k-nearest-neighbor graph (k=10) embeds the points and k-means splits them into subclusters whose count is estimated from cluster size divided by the maximum allowed size. Volumes are then computed by Convex Hull (upper bound) and Alpha Shape with α=0.9 m, which captures concave canopy shape. The identity that carries the argument is that overly large density clusters correspond to overlapping crowns, and that graph connectivity reveals the individual trees within them.","core_discovery":"The central discovery is that the obstacle to per-tree canopy volume from mobile LiDAR is not geometry reconstruction but segmentation: in a dense almond orchard the same DBSCAN parameters that work in a uniform pistachio orchard group multiple interwoven trees into one cluster, failing on 54% of trees. Adding a graph-based spectral clustering refinement that splits any cluster larger than 45,000 points into subclusters on a k-nearest-neighbor connectivity graph recovers correct individual crowns in 80% of cases. With per-tree clusters in hand, the paper shows that Alpha Shape reconstruction (α=0.9 m) tracks drone-derived spherical volume estimates for pistachio trees with the errors shown in Table I, while Convex Hull systematically overestimates. The authors present this as evidence that adaptive, geometry-aware segmentation is the missing piece for scalable per-tree canopy monitoring.","pith_inferences":["The paper limits quantitative volume validation to the pistachio orchard and admits the drone orthomosaic-to-sphere ground truth is idealized for symmetric crowns; for asymmetric almond canopies the reported accuracy should be treated as untested until a true 3D reference is used.","The 20% of almond trees still mis-segmented could skew per-tree averages in yield or biomass models; a natural extension is measuring how much the volume distribution changes when remaining merges are corrected manually.","The heuristic of splitting clusters by a point-count threshold on a k-NN graph is a transferable recipe: any LiDAR instance-segmentation problem with touching objects (street trees, shrubs, stacked material) could adopt the same over-cluster refinement."],"forward_implications":["During a normal robot pass at up to about 1.5 m/s, every tree in a row can receive a label and a volume estimate, so canopy monitoring requires no stationary scanning or offline processing.","The same DBSCAN-plus-spectral pipeline should apply to other dense plantings with overlapping canopies, such as olives or vineyards, without retuning.","With per-tree volumes and sequential row labeling, repeated passes enable growth, pruning, and health tracking across the season.","Volume estimates from Alpha Shape, not Convex Hull, should be used when the goal is fidelity to concave canopy shape."],"supporting_citations":[{"why":"Supplies the canopy volume measurement model from LiDAR point clouds that motivates the volume estimation stage.","marker":"[1]"},{"why":"Compares canopy volume measurements from imagery and LiDAR, supporting the use of aerial-derived volumes as a validation reference.","marker":"[8]"},{"why":"Shows UAV photogrammetry can replicate LiDAR-derived forest structure, backing the drone orthomosaic ground truth used for validation.","marker":"[14]"},{"why":"Provides the prior laser-scanner measurement system for tree geometry that this work extends from static to mobile, real-time operation.","marker":"[18]"},{"why":"Establishes terrestrial LiDAR scanning as the reference method for orchard 3D structure, the static-scan approach the mobile pipeline replaces.","marker":"[19]"}],"fun_headline_variants":["Adaptive clustering lifts almond tree segmentation to 80%","Mobile LiDAR per-tree volume via adaptive segmentation","DBSCAN plus spectral step recovers 80% of dense almond trees","From 46% to 80%: adaptive clustering for orchard monitoring"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The validation assumes that drone photogrammetry turned into spherical volumes from canopy diameters is accurate ground truth for canopy volume; the paper itself notes this only holds for roughly symmetric crowns (Section III-A), so the volume-error claims rest on that simplification.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive clustering lifts almond tree segmentation to 80%","Mobile LiDAR per-tree volume via adaptive segmentation","DBSCAN plus spectral step recovers 80% of dense almond trees","From 46% to 80%: adaptive clustering for orchard monitoring"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000545,"raw_usage":{"total_tokens":2561,"prompt_tokens":850,"completion_tokens":1711,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":1639}},"tokens_in":466,"tokens_out":1711,"duration_ms":14201,"temperature":1.0,"reasoning_tokens":1639,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:30:43.317312+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the same pistachio trees with a static, high-resolution terrestrial LiDAR scanner that sees the full crown, and check whether the mobile-LiDAR Alpha Shape volumes match the drone-sphere volumes within the errors reported in Table I; alternatively, run the almond two-stage pipeline on the pistachio rows and see if the added spectral splitting degrades the 93% segmentation, which would show the 46% to 80% gain is not a generalizable result.","supporting_citations":[{"cited_title":"A study on canopy volume measurement model for fruit tree application based on lidar point cloud,","cited_arxiv_id":null,"evidence_quote":"Supplies the canopy volume measurement model from LiDAR point clouds that motivates the volume estimation stage."},{"cited_title":"Comparison of canopy volume measurements of scattered eucalypt farm trees derived from high spatial resolution imagery and lidar,","cited_arxiv_id":null,"evidence_quote":"Compares canopy volume measurements from imagery and LiDAR, supporting the use of aerial-derived volumes as a validation reference."},{"cited_title":"To what extent can uav photogrammetry replicate uav lidar to determine forest structure? a test in two contrasting tropical forests,","cited_arxiv_id":null,"evidence_quote":"Shows UAV photogrammetry can replicate LiDAR-derived forest structure, backing the drone orthomosaic ground truth used for validation."},{"cited_title":"A laser scanner based measurement system for quantification of citrus tree geometric characteristics,","cited_arxiv_id":null,"evidence_quote":"Provides the prior laser-scanner measurement system for tree geometry that this work extends from static to mobile, real-time operation."},{"cited_title":"Obtaining the three- dimensional structure of tree orchards from remote 2d terrestrial lidar scanning,","cited_arxiv_id":null,"evidence_quote":"Establishes terrestrial LiDAR scanning as the reference method for orchard 3D structure, the static-scan approach the mobile pipeline replaces."}],"review_version":1}