{"id":"d0ffd24b-f67c-4d0f-9f6b-f76365d7098e","arxiv_id":"2506.04970","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Learned SAM prompts plus DSM elevation data improve tree crown segmentation on plantations, but the stated advantage over Mask R-CNN does not hold on all three test forests.","lead":"The paper tests a general-purpose image segmentation model (SAM) for picking out individual tree crowns in drone photos, and adds elevation maps to try to improve results. It finds that out-of-the-box SAM works poorly, but learning to prompt SAM with elevation data can beat a standard Mask R-CNN on plantations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BCI results contradict the central claim: Mask R-CNN+DSM encoder has the best wmAP on the tropical dataset, so 'outperform on all three datasets' is unsupported.","rationale":"The paper's headline contribution is a comparative claim: learned prompters integrated with SAM (BalSAM and RSPrompter) outperform custom-trained CNNs across three forest types. The most direct evidence for that claim should be the three main result tables. Tables 1 and 2 support it, but Table 3 does not: on the BCI tropical forest, Mask R-CNN+DSM encoder has the highest wmAP, and BalSAM is statistically worse. Because this is an internal contradiction, it does not rely on outside assumptions about label quality or baseline tuning, making it more load-bearing than the reader's stated weakest assumption. The reader did notice this in the rationale ('its own BCI results show a Mask R-CNN variant with the best weighted mAP'), but did not make it the central weakness; I partially agree. The correct fix is a statistical comparison with more seeds and possibly unmasked labels, plus per-dataset phrasing of the conclusion. The paper contains useful experiments and the abstract itself is more cautious, so the verdict remains CONDITIONAL rather than moving to REJECT; my recommendation is therefore UNCHANGED relative to the reader's verdict.","tokens_in":22485,"tokens_out":5699,"duration_ms":61757,"concrete_test":"Recompute BCI wmAP for RSPrompter, BalSAM, and Mask R-CNN+DSM encoder using the original unmasked BCI annotations (i.e., before the Section 3 manual masking), over 10 seeds with the same data splits and training protocol; then form paired bootstrap 95% CIs for (RSPrompter - Mask R-CNN+DSM encoder) and (BalSAM - Mask R-CNN+DSM encoder). If either CI excludes 0 in favor of the CNN or includes 0, Section 5.1 must be revised to per-dataset claims; only if the learned prompters are significantly and consistently ahead on wmAP would the 'all three datasets' claim be restored.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.1 states: 'RSPrompter and BalSAM models outperform Mask R-CNN-based models (integrating or not the DSM) in terms of multi-class mAP and wmAP on all three datasets.' Table 3 (BCI) directly contradicts this for wmAP: Mask R-CNN+DSM encoder = 11.86±0.27, RSPrompter = 11.53±0.34, BalSAM = 10.42±0.27. The custom-trained CNN variant is therefore the best weighted-mAP method on BCI, and BalSAM is about 1.4 wmAP points below it (roughly 3.8 standard errors of the difference). Multi-class mAP on BCI is a near tie (8.44, 8.48, 8.30) with overlapping errors. Thus the central comparative conclusion does not hold on the most challenging dataset; at best it holds on the plantation dataset. This is load-bearing because the paper's advertised contribution is that learned SAM prompting with optional DSM is the generally best recipe; if BCI is only comparable or worse, the result is dataset-specific. The reader's identified risks (BCI manual masking, per-dataset prompt parameters) are real but secondary: they concern baseline and prompt construction, while the BCI ranking contradiction is internal to the reported numbers and does not depend on those assumptions. Note also that the related claim that DSM integration 'generally improves' is weakened by SBL (RSPrompter wmAP 29.44 vs BalSAM 29.12) and BCI (11.53 vs 10.42), so the discussion should be per-dataset.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper benchmarks methods for individual tree crown instance segmentation from high-resolution drone RGB imagery on three datasets (Quebec Plantations, SBL, BCI), comparing SAM in automatic mode, SAM prompted by DSM local maxima, Mask R-CNN and Faster R-CNN variants with and without DSM input, RSPrompter, and a new BalSAM method that adds a trainable DSM encoder to RSPrompter. All trained comparisons use three seeds and report single-class and multi-class mAP/wmAP. The abstract and Section 5.1 claim that learned SAM prompting methods outperform Mask R-CNN-based models on all datasets and that DSM information generally improves performance, but the reported tables only support these claims for some datasets and metrics.","tokens_in":22803,"tokens_out":2815,"duration_ms":32337,"significance":"If the central claims held, the paper would provide a useful benchmark for tree crown instance segmentation and a strong argument for parameter-efficient SAM prompting with auxiliary elevation data. The experiments are carefully run with three seeds, standard errors, and detailed hyperparameters, and the paper is among the first to benchmark SAM-based methods on these three drone datasets. However, the central generalization claim is directly contradicted by the authors' own numbers on the BCI dataset, and the DSM benefit is inconsistent across datasets. The empirical work is valuable, but the current framing overstates the conclusions.","major_comments":[{"comment":"The sentence 'RSPrompter and BalSAM models outperform Mask R-CNN-based models (integrating or not the DSM) in terms of multi-class mAP and wmAP on all three datasets' is contradicted by the BCI results in Table 3. For wmAP, Mask R-CNN+DSM encoder achieves 11.86±0.27, RSPrompter achieves 11.53±0.34, and BalSAM achieves 10.42±0.27; for multi-class mAP the three methods are essentially tied (8.30, 8.44, 8.48 with overlapping standard errors). The claim should be revised to a per-dataset statement, and the BCI result should be discussed as a counterexample rather than glossed over.","section":"Section 5.1, Table 3"},{"comment":"The claim that integrating DSM 'generally improves' model performance is not consistently supported. On SBL (Table 2), Mask R-CNN+DSM has wmAP 26.82±0.15 versus 27.27±0.18 for Mask R-CNN without DSM, and RSPrompter wmAP 29.44±0.83 versus BalSAM wmAP 29.12±0.81, so DSM integration does not help the strongest methods on this dataset. The discussion should report the effect per dataset and metric rather than as a general trend.","section":"Section 5.1, Tables 1-3"},{"comment":"The BCI test-set construction involves manual decisions: the text says 'We manually correct for missing annotations by masking out parts of the imagery that contain unannotated trees,' and the SAM+DSM prompt method uses a hand-set local-maxima minimum distance (50 for Quebec Plantations, 20 for SBL and BCI). Because BCI is the dataset where the central comparative claim fails, and because these choices affect both labels and prompts, the paper should provide a sensitivity analysis or a clear argument that these decisions do not bias the comparison in favor of the proposed methods.","section":"Section 3 and Appendix B.6"},{"comment":"The statement that BalSAM 'shows potential over other methods' is supported on Quebec Plantations, but the advantage over RSPrompter there is moderate (wmAP 64.84±0.86 versus 62.37±1.41) and on SBL BalSAM is slightly below RSPrompter. The conclusion should be phrased as a per-dataset observation, not as a general superiority claim.","section":"Section 5.1, Table 1"}],"minor_comments":[{"comment":"There is a typo in the Mask R-CNN row: '81.82(±0.21' is missing a closing parenthesis.","section":"Table 1"},{"comment":"The word 'derading' in 'leading to gridded segmentation patterns, derading the segmentations overall' should be 'degrading'.","section":"Appendix B.4"},{"comment":"The phrase 'UA V Canadian (Quebec) Plantations dataset' contains an unusual spacing in 'UA V'; this should be 'UAV'.","section":"Section 3"},{"comment":"The description of the first-layer initialization for Mask R-CNN+DSM and Faster R-CNN+DSM is clear, but the sentence 'Then, we copy back the ImageNet pre-trained backbone's weights of the first layer onto the RGB channels' is slightly confusing and could be reworded to say the weights are copied from the pre-trained first layer to the RGB channels of the new four-channel first layer.","section":"Section 4.3"},{"comment":"The paper states that only a 'representative sample' of code is provided in the supplementary material. For reproducibility of a benchmark paper, the full code and configuration files should be released.","section":"Section 4.3 and Appendix B.6"}],"recommendation":"major_revision","confidential_remarks":"The BCI contradiction is internal to the reported numbers and is the main reason I cannot recommend acceptance in the current form. The overclaim in Section 5.1 is straightforward to fix by rewriting the discussion, but the manual masking of BCI test labels and the per-dataset prompt hyperparameters should also be addressed, as they affect the credibility of the benchmark conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a genuinely useful empirical comparison and a simple, sensible extension of RSPrompter (BalSAM), but the central claim as written does not survive its own Table 3. On BCI the custom Mask R-CNN with a DSM encoder has the best wmAP, and BalSAM is about 1.4 points behind. The paper needs a corrected discussion and per-dataset conclusions.\n\nWhat's new: BalSAM is a modest but clean idea – a small trainable DSM encoder, element-wise addition to SAM image embeddings, fed as dense prompt. It's a natural extension of RSPrompter, not a big architectural leap, but it is a concrete contribution. The bigger value is the first instance-segmentation benchmark on three publicly available drone datasets with different canopy structures. The experimental discipline is solid: three seeds, standard errors, per-class results, ablations for prompt types, losses, and DSM integration. That level of detail will help anyone working on tree crown segmentation.\n\nSoft spots: the main overclaim is in Section 5.1. It states RSPrompter and BalSAM outperform Mask R-CNN-based models on all three datasets, but Table 3 shows Mask R-CNN+DSM encoder at 11.86 wmAP vs RSPrompter 11.53 and BalSAM 10.42 on BCI. That's not a small gap – about 3.8 standard errors of the difference for BalSAM. The multi-class mAP on BCI is a near tie (8.30 vs 8.48). So the general claim collapses; at best it holds for plantations, and mixed for the temperate forest. Also, the claim that DSM 'generally improves' is too broad: on SBL, BalSAM (29.12) is slightly below RSPrompter without DSM (29.44), and on BCI it's clearly worse. The discussion should be per-dataset.\n\nSecondary concerns are real but lower priority: BCI manual masking of unannotated trees is a label-quality risk, and the SAM+DSM prompt baseline uses per-dataset tuned local-max distances. Those affect the evaluation, but they don't drive the central contradiction, which is internal to the reported numbers. Full code isn't released yet, only a representative sample; that's standard for under-review papers.\n\nWho it's for: people working on tree crown segmentation from drone RGB or RGB+DSM, and anyone using SAM for instance segmentation in remote sensing. It's not a method paper with a clear win everywhere, but the benchmark and the analysis of when DSM helps (or hurts) are genuinely informative.\n\nRecommendation: it deserves a serious referee – the empirical work is careful and the overclaim is fixable. I'd ask for a substantial revision: correct the summary claims to match the tables, report per-dataset conclusions, and discuss the BCI result honestly. Don't desk reject.","headline":"Useful benchmark and a clean but modest method, undermined by an overclaim that its own Table 3 contradicts; worth reviewing after corrections.","tokens_in":23370,"tokens_out":4046,"would_cite":true,"duration_ms":42803,"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":"The paper argues that learning task-specific prompts for SAM, with elevation data added, gives the best tree crown instance segmentation from drone imagery.","keywords":["tree crown segmentation","instance segmentation","Segment Anything Model","Digital Surface Model","drone imagery","prompt learning","forest monitoring","remote sensing"],"falsifier":"An independent evaluation with fresh, unmodified annotations and prompt parameters chosen only by validation (not by dataset-specific hand tuning) in which a custom Mask R-CNN with DSM matches or beats the learned-prompt SAM methods would show the central claim does not hold.","tokens_in":22279,"feed_emoji":"🌲","tokens_out":3558,"duration_ms":38387,"temperature":0.7,"pith_summary":"This paper asks whether the Segment Anything Model (SAM), a foundation model for image segmentation, can be turned into a reliable tool for mapping individual tree crowns from high-resolution drone imagery. The authors compare SAM used out of the box, SAM given hand-crafted or predicted prompts, and custom convolutional networks such as Mask R-CNN, across plantations, a temperate forest, and a tropical forest. They find that SAM alone underperforms a custom Mask R-CNN, but that a module which learns task-specific prompts for SAM outperforms the CNN baselines. They also show that adding a Digital Surface Model (DSM) computed from the same drone images improves results, most strongly where trees are well separated. Their proposed model, BalSAM, fuses elevation information into SAM's prompt learning and achieves the best results in the plantation setting.","feed_headline":"Learned SAM prompts beat custom CNNs for tree crowns","feed_subtitle":"Adding drone-derived elevation data makes the learned prompting approach win across three forest types.","key_machinery":"The load-bearing mechanism is the learnable prompt generator coupled to SAM's frozen image encoder and mask decoder. BalSAM extends the RSPrompter architecture by adding a small three-layer CNN that encodes the Digital Surface Model and fuses its output with SAM's image embedding through an element-wise sum; this fused embedding is fed to the mask decoder alongside the learned prompts. The DSM encoder lets the model exploit canopy height and crown shape as dense information without retraining SAM's core.","core_discovery":"The central claim is that the best recipe for tree crown instance segmentation from drone imagery is not SAM used directly, nor a CNN alone, but a learned prompting module integrated with SAM, optionally fed with elevation information. Out-of-the-box SAM, even when prompted with treetop locations derived from the DSM, produces many false positives and merges overlapping crowns, and scores far below a custom Mask R-CNN. However, methods that train a small module to generate prompts for SAM (RSPrompter, and BalSAM with DSM) outperform Mask R-CNN on all three datasets in multi-class mean average precision. Adding the DSM as an extra channel or as a learned embedding reliably improves the CNN and SAM-based models, with the largest gains in the plantation dataset where trees are distinct and the ground is visible.","pith_inferences":["The DSM encoder design suggests that other cheap structural modalities (e.g., canopy height models from LiDAR, or multi-season imagery) could be fused into SAM the same way, potentially improving species classification in closed forests.","Because the learned prompter is task- but not site-specific, the same approach may transfer to other fine-grained instance segmentation problems with scarce labels, such as mapping shrubs, crops, or individual buildings.","An implicit implication of the work is that benchmark conclusions in this area are sensitive to annotation completeness; future studies should report how missing crowns are handled."],"forward_implications":["Forest monitoring pipelines can skip manual tree crown labeling at scale, using a trained prompt module around SAM instead of a fully custom CNN.","Publishing drone orthomosaics together with photogrammetric DSMs is a low-cost addition that improves segmentation accuracy, especially in plantations and other open canopies.","SAM out-of-the-box is not a drop-in solution for tree crown segmentation; it needs task-specific adaptation to compete.","Learned prompting helps most on rare classes, which are the ones ecologists often care about."],"supporting_citations":[{"why":"Introduces the Segment Anything Model whose frozen encoder and mask decoder all methods build on.","marker":"[14]"},{"why":"Presents RSPrompter, the learned prompt generation approach that BalSAM extends.","marker":"[20]"},{"why":"Defines Mask R-CNN, the custom CNN baseline that the paper compares against.","marker":"[8]"},{"why":"Supplies the Quebec Plantations dataset with RGB orthomosaics, DSMs, and tree crown annotations.","marker":"[21]"},{"why":"Supplies the SBL temperate forest dataset with imagery and 22,933 labeled tree crowns.","marker":"[22]"},{"why":"Supplies the BCI tropical forest dataset with crown maps and an accompanying DSM.","marker":"[23]"},{"why":"Showed DSM alongside RGB drone imagery is promising for semantic tree segmentation, a cue for the instance segmentation here.","marker":"[58]"}],"fun_headline_variants":["Elevation data boosts SAM for tree crown mapping","Learned SAM prompts with elevation beat custom CNNs","BalSAM: SAM with elevation outshines CNN for tree crowns","Tuned SAM plus elevation wins tree crown segmentation","Drone tree crowns: learned SAM prompts top CNN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes that manually correcting missing tree annotations and hand-picking per-dataset prompt parameters (like the minimum distance between treetop prompts) does not unfairly advantage the proposed methods.","fun_headline_variants_meta":{"raw":{"variants":["Elevation data boosts SAM for tree crown mapping","Learned SAM prompts with elevation beat custom CNNs","BalSAM: SAM with elevation outshines CNN for tree crowns","Tuned SAM plus elevation wins tree crown segmentation","Drone tree crowns: learned SAM prompts top CNN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00046,"raw_usage":{"total_tokens":2288,"prompt_tokens":912,"completion_tokens":1376,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":1299}},"tokens_in":528,"tokens_out":1376,"duration_ms":11783,"temperature":1.0,"reasoning_tokens":1299,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:28:55.600972+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An independent evaluation with fresh, unmodified annotations and prompt parameters chosen only by validation (not by dataset-specific hand tuning) in which a custom Mask R-CNN with DSM matches or beats the learned-prompt SAM methods would show the central claim does not hold.","supporting_citations":[{"cited_title":"Mask R-CNN","cited_arxiv_id":null,"evidence_quote":"Defines Mask R-CNN, the custom CNN baseline that the paper compares against."},{"cited_title":"Segment Anything","cited_arxiv_id":null,"evidence_quote":"Introduces the Segment Anything Model whose frozen encoder and mask decoder all methods build on."},{"cited_title":"RSPrompter: Learning to prompt for remote sensing instance segmentation based on visual foundation model.IEEE Transactions on Geoscience and Remote Sensing, 2024","cited_arxiv_id":null,"evidence_quote":"Presents RSPrompter, the learned prompt generation approach that BalSAM extends."},{"cited_title":"Influence of temperate forest autumn leaf phenology on segmentation of tree species from UA V imagery using deep learning.Remote Sensing of Environment, 311:114283, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies the SBL temperate forest dataset with imagery and 22,933 labeled tree crowns."},{"cited_title":"Barro Colorado Island 50-ha plot crown maps: manually segmented and instance segmented","cited_arxiv_id":null,"evidence_quote":"Supplies the BCI tropical forest dataset with crown maps and an accompanying DSM."}],"review_version":1}