{"id":"81b6e303-b064-44d5-b6e5-67d60cebe01b","arxiv_id":"2501.13426","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"APSAM converts image-level landslide labels into SAM box and point prompts derived from class activation maps, yielding pseudo-masks that beat prior weakly supervised methods on two landslide datasets.","lead":"A landslide mapping method uses the Segment Anything Model, prompted automatically by class activation maps, to convert image-level labels into pixel-level pseudo-labels. It leads six weakly supervised baselines on a Turkey dataset, but gains only 0.52 F1 points over the best baseline on Hong Kong.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline margins are not trustworthy: T=120 is selected from a test-ground-truth IoU curve, results are single-run, and the abstract's at least 3.0% F1 / 3.69% IoU claim is contradicted by Hong Kong Table I (gains 0.52/0.70 over LGAGNet).","rationale":"The reader's CONDITIONAL verdict is well matched to the evidence. The APSAM pipeline is coherent and the ablation in Table II gives internal support for the central mechanism: combining point and box prompts outperforms either alone. My stress-test does not question the method's plausibility; it targets whether the reported margins are honestly established. The strongest reason for doubt is that T=120 is selected from an IoU curve computed on test labels (no validation split is described), and the final results are single-run. The abstract's 'at least 3.0% / 3.69%' claim is also contradicted by the Hong Kong row of Table I, where the margin over LGAGNet is 0.52 F1 and 0.70 IoU. A validation-based threshold search with multiple seeds would settle whether the true margin survives; if it does not, the paper remains publishable as a method contribution but with corrected claims. No code or data repository is currently available, so independent reproduction is not possible, which further supports keeping the verdict conditional rather than accepting the quantitative claims at face value.","tokens_in":7904,"tokens_out":7951,"duration_ms":72793,"concrete_test":"Hold out a validation split (e.g., 20% of the test images) from each dataset before any hyperparameter selection. For T in {80, 100, 120, 140, 160}, train the object localizer, generate pseudo-labels, retrain ResUNet with 5 random seeds, and evaluate only on the remaining held-out test images. Select T on validation IoU, not test IoU, then recompute the APSAM-versus-LGAGNet F1/IoU margins with mean ± std and a paired test over images. If the Hong Kong margin remains below 3.0% F1 or 3.69% IoU, or if the best validation T is not around 120, the abstract's 'at least' claim and the T=120 justification need revision. Running the same multi-seed protocol for LGAGNet provides a matched control for run-to-run variance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the central claim that APSAM beats state-of-the-art methods by at least 3.0% F1 and 3.69% IoU, the only free hyperparameter T must be set without access to the test labels. The paper describes only image-level training sets and pixel-level test sets (Sec. III-A), with no validation split. Fig. 5, used to justify T=120, is an IoU-versus-T curve; since pixel labels exist only for the test sets, that curve is computed on the same data used for Table I. Choosing T from test IoU is therefore a form of test-set leakage: the proposed column in Table I is conditional on a threshold selected to maximize IoU on the evaluation set. The stated plateau between T=50 and 170 may limit the damage, but the protocol is still not an independent evaluation. Table I also contradicts the abstract's 'at least' wording: on Hong Kong, the gain over the strongest baseline LGAGNet is 0.52 F1 and 0.70 IoU, not at least 3.0/3.69; only Turkey reaches those margins. All entries are single runs without error bars or significance tests, so even the Turkey 3.01 F1 margin could be within run-to-run noise. These issues do not invalidate the pipeline, but they mean the headline quantitative claim is not established as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes APSAM, a weakly supervised landslide extraction pipeline. An object localization network trained with image-level labels produces CAMs; an adaptive prompt generation (APG) algorithm binarizes the CAM at a fixed threshold T, finds contours, and derives bounding-box and centroid prompts, which are fed into a frozen SAM to synthesize pseudo-masks. A ResUNet is then trained on these pseudo-masks. Experiments on the Hong Kong and Turkey datasets compare APSAM against six weakly supervised baselines using OA, precision, recall, F1, and IoU. The authors claim improvements of at least 3.0% F1 and 3.69% IoU over state-of-the-art methods and report ablations showing that hybrid point-plus-box prompts outperform either prompt type alone.","tokens_in":8126,"tokens_out":6438,"duration_ms":53740,"significance":"The idea is practically appealing: it avoids CAM-based pseudo-label refinement or SAM fine-tuning by using a frozen SAM with automatically generated prompts, making it a low-cost way to turn image-level labels into pixel-level landslide maps. If the claimed gains hold, the method would be a useful contribution to weakly supervised landslide extraction. The paper's strengths are its simple and reproducible pipeline, the planned release of code and data, and the ablation study in Table II, which supports the hybrid prompt design. The main concerns are evaluation-protocol issues rather than conceptual flaws: the headline quantitative claim is not fully supported by the reported tables, the threshold T is not validated independently of the test labels as described, and the results are single runs without uncertainty measures.","major_comments":[{"comment":"The abstract's claim of 'at least 3.0% in F1 score and 3.69% in IoU compared to other state-of-the-art methods' is not supported by Table I. On the Hong Kong dataset, the gain over the strongest baseline LGAGNet is 0.52 F1 points (79.14 vs. 78.62) and 0.70 IoU points (65.48 vs. 64.78); only the Turkey dataset reaches the stated margins (3.01 F1 and 3.69 IoU). The wording 'at least' is therefore inaccurate and should be revised, or the claim should be restricted to the dataset where it actually holds.","section":"Abstract; Table I"},{"comment":"The key hyperparameter T is not validated independently of the test labels. Section III-A describes image-level training sets and pixel-level test sets but no validation split. Section III-B fixes T=120, and Section III-D justifies this choice using the IoU-versus-threshold curve in Fig. 5. Since pixel-level labels are described as existing only for the test sets, the IoU values in Fig. 5 appear to be computed on the evaluation data, making the 'Ours' row in Table I conditional on a threshold selected to maximize test IoU. The reported plateau from T=50 to 170 may limit the practical impact, but the protocol as described is not an independent evaluation; please either fix T a priori, choose it on a held-out split, or explicitly report the sensitivity of Table I to T.","section":"Sec. III-A, III-B, III-D (Fig. 5)"},{"comment":"All quantitative results are reported as single runs without error bars, standard deviations, or significance tests. The Hong Kong F1 margin of 0.52 points over LGAGNet is small enough that it could arise from training stochasticity, and the conclusion that 'our network achieves the highest F1-Score and IoU values' on Hong Kong is not statistically established. Repeating the experiments over multiple seeds and reporting mean and standard deviation, or another measure of uncertainty, is needed to support the claimed superiority.","section":"Table I; Sec. III-C"}],"minor_comments":[{"comment":"The introduction contains a typo: 'zeo-shot' should be 'zero-shot.'","section":"Sec. I"},{"comment":"The dataset name is rendered inconsistently as both 'TurKey' and 'Turkey'; please unify the spelling.","section":"Sec. III-A"},{"comment":"There are minor textual inconsistencies: 'Tabel I' should be 'Table I,' and 'FilpCAM' in the experimental setup should be 'FlipCAM' to match the table and references.","section":"Sec. III-C and III-B"},{"comment":"The prose describing Table II mislabels the metrics: for the Turkey 'Point + Box' row, the table lists Precision 64.00, Recall 71.03, F1 67.33, and IoU 50.75, but the text says 'the precision 71.03%, recall 67.33%, and F1-score 50.75%.' Please correct the text or the table so the reported numbers align with the metric names.","section":"Sec. III-D; Table II"},{"comment":"In the threshold analysis paragraph, 'the grayscale threshold L is between 50 and 170' should refer to T, the variable defined in Eq. (2), not L.","section":"Sec. III-D"},{"comment":"Figure 5 lacks axis labels and a legend; please specify which dataset and metric the IoU-versus-threshold curve corresponds to.","section":"Fig. 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the scope of the journal and the proposed pipeline is a reasonable contribution if the evaluation is made rigorous. The main barriers are the abstract overclaim, the apparent test-set dependence of the threshold selection, and the lack of uncertainty estimates. These issues are fixable within the manuscript's scope: a held-out validation split or a sensitivity analysis, multiple runs, and a corrected abstract would substantially strengthen the paper. I do not see a fundamental flaw in the methodology itself."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing you should know: this is a clean, competent engineering paper that combines two existing ideas—CAM-based localization and SAM prompting—into a pipeline that produces pseudo-labels for landslide segmentation with only image-level supervision. The actual new content is the specific recipe and the APG algorithm that converts CAM contours into box and point prompts for frozen SAM. The ablation in Table II supports the design: point+box beats either alone by a clear margin on both datasets. The qualitative figures look reasonable. That is the good part.\n\nThe weak part is the quantitative claim. The abstract says “at least 3.0% F1 and 3.69% IoU” improvement, but on the Hong Kong dataset the gain over the strongest baseline (LGAGNet) is 0.52 F1 and 0.70 IoU. Only Turkey reaches the stated margins, and all results are single runs without error bars or significance tests. The bigger problem is the threshold T=120. It is selected from the IoU-versus-T curve in Fig. 5, and since pixel-level labels exist only for the test sets (no validation split is described), that curve is computed on the test data. So the measured IoU is partly a fitted quantity. The paper notes a plateau between T=50 and 170, which limits how much this matters, but the protocol is still not a clean independent evaluation.\n\nThe other practical gripe: the abstract promises source code and datasets at a GitHub link, but no link or code is actually provided in the manuscript, and the SAM checkpoint configuration is unspecified. That makes the result hard to reproduce or stress-test.\n\nI don't think the method is broken. The logic is coherent, the ablation is sensible, and the problem is genuinely important. But the headline numbers should be treated cautiously. I'd want the authors to release code, add error bars, use a validation split for T (or at least explain why the plateau makes test-set selection harmless), and rewrite the abstract to match the actual margins. These are referee-level requests, not fatal flaws.\n\nMy take: this paper deserves a serious referee, especially in an applied remote sensing venue. It is a useful, incremental contribution. I would cite it for the approach if I worked on landslide mapping, but I would not repeat the stated performance gains as established. For a reading group, it's a decent example of how to adapt SAM under weak supervision, but not a must-read.\n\nRecommendation: send to peer review with a request that the reviewers focus on the evaluation protocol and reproducibility.","headline":"Clean, plausible SAM-prompting recipe for landslide extraction, but the headline performance claim is overstated and T is selected on test data.","tokens_in":8749,"tokens_out":3554,"would_cite":true,"duration_ms":31827,"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":"Auto-prompting SAM lifts landslide F1 by at least 3 points","keywords":["weakly supervised landslide extraction","Segment Anything Model","prompt engineering","class activation maps","object localization","pseudo-labeling","remote sensing"],"falsifier":"A direct test would be to compute pseudo-masks with the same CAM and threshold but replace SAM's prompts with random boxes and centroids of the same size and count: if the performance gap over baselines persists, the prompts are not doing the work attributed to them.","tokens_in":7613,"feed_emoji":"🏔️","tokens_out":7705,"duration_ms":51196,"temperature":0.7,"pith_summary":"This paper aims to show that pixel-level landslide maps can be produced from image-level labels alone by feeding automatically generated prompts into a frozen Segment Anything Model (SAM). The method, APSAM, takes coarse class activation maps from an object localization network, binarizes them, and converts the resulting contours into box prompts and centroid point prompts. These prompts guide SAM to produce pseudo-masks, which then train a standard segmentation network. The authors report gains of at least 3.0% in F1 score and 3.69% in IoU over state-of-the-art weakly supervised methods on two high-resolution landslide datasets. If correct, this offers a low-annotation route to accurate landslide mapping.","feed_headline":"Auto-prompting SAM lifts landslide F1 by at least 3 points","feed_subtitle":"Weak labels plus a frozen SAM yield pixel-level landslide masks.","key_machinery":"The adaptive prompt generation (APG) algorithm is the central mechanism. It binarizes the CAM at a fixed threshold, extracts contours, computes each contour's bounding box as a box prompt and its moment-based centroid as a point prompt, and feeds the combined prompts into SAM's prompt encoder. This converts a coarse activation heatmap into the sparse, structured input SAM needs to produce clean segmentation masks without any parameter update.","core_discovery":"The central claim is that a frozen, prompt-only SAM can replace the usual pseudo-labeling pipeline in weakly supervised landslide extraction. Rather than refining CAMs or fine-tuning SAM, APSAM uses the CAM heatmap to guess where landslides are (boxes) and where their centers are (points), and relies on SAM's zero-shot segmentation to turn those guesses into fine-grained masks. The paper demonstrates this on the Hong Kong and Turkey datasets, where the approach surpasses six weakly supervised baselines in F1 and IoU, including the CAM-based LGAGNet and the SAM-based S2C.","pith_inferences":["Because the threshold T is tuned on the evaluation data, the method's edge over baselines may shrink if the CAM distribution shifts; a non-fixed threshold or a learned prompt generator would be a natural stress test.","The same auto-prompting recipe may transfer to other weakly supervised segmentation tasks with irregular, fragmented objects, such as building damage or flooded areas, provided SAM produces reasonable masks for those categories.","The fixed threshold and contour geometry imply that landslides that merge into one CAM blob will be treated as a single object; adding a splitting rule based on size or shape could refine the prompts."],"forward_implications":["APSAM yields pseudo-masks that train a ResUNet to outperform six weakly supervised baselines on both datasets.","The approach works on both aerial and satellite imagery (Hong Kong and Turkey) with 0.5–0.59 m resolution.","Combining box and point prompts outperforms either prompt type alone in the reported ablations.","The method requires no fine-tuning of SAM, so it inherits SAM's zero-shot segmentation behavior while adding only an off-the-shelf CAM module.","IoU gains range from 3.69 to 16.41 percentage points over the baselines on the Turkey dataset."],"supporting_citations":[{"why":"Supplies the class activation map (CAM) used for coarse object localization.","marker":"[4]"},{"why":"Provides the frozen Segment Anything Model whose prompt-conditioned inference produces the pseudo-masks.","marker":"[7]"},{"why":"Introduces the two high-resolution landslide datasets (Hong Kong and Turkey) used for training and evaluation.","marker":"[2]"},{"why":"The strongest CAM-based weakly supervised baseline that APSAM is compared against.","marker":"[5]"},{"why":"A SAM-based weakly supervised baseline (from-SAM-to-CAMs) that APSAM directly contrasts with.","marker":"[16]"}],"fun_headline_variants":["SAM auto-prompting boosts landslide masks without pixel labels","Weak supervision + SAM prompts = sharper landslide maps","Frozen SAM turns weak labels into landslide masks","Auto-prompted SAM beats CAM-only landslide methods","Prompt engineering unlocks SAM for landslide extraction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that binarizing the class activation map at threshold 120 produces contours whose bounding boxes and centroids point to actual landslide extents and centers; if the heatmap is incomplete or the threshold is miscalibrated, SAM is guided to the wrong regions.","fun_headline_variants_meta":{"raw":{"variants":["SAM auto-prompting boosts landslide masks without pixel labels","Weak supervision + SAM prompts = sharper landslide maps","Frozen SAM turns weak labels into landslide masks","Auto-prompted SAM beats CAM-only landslide methods","Prompt engineering unlocks SAM for landslide extraction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000369,"raw_usage":{"total_tokens":1957,"prompt_tokens":905,"completion_tokens":1052,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":980}},"tokens_in":521,"tokens_out":1052,"duration_ms":8785,"temperature":1.0,"reasoning_tokens":980,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:57:38.256430+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to compute pseudo-masks with the same CAM and threshold but replace SAM's prompts with random boxes and centroids of the same size and count: if the performance gap over baselines persists, the prompts are not doing the work attributed to them.","supporting_citations":[{"cited_title":"Learning deep features for discriminative localization,","cited_arxiv_id":null,"evidence_quote":"Supplies the class activation map (CAM) used for coarse object localization."},{"cited_title":"Cross-domain landslide mapping from large-scale remote sensing images using prototype-guided domain-aware progressive representation learning,","cited_arxiv_id":null,"evidence_quote":"Introduces the two high-resolution landslide datasets (Hong Kong and Turkey) used for training and evaluation."},{"cited_title":"Weakly supervised local– global anchor guidance network for landslide extraction with image-level annotations,","cited_arxiv_id":null,"evidence_quote":"The strongest CAM-based weakly supervised baseline that APSAM is compared against."},{"cited_title":"From sam to cams: Exploring segment anything model for weakly supervised semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"A SAM-based weakly supervised baseline (from-SAM-to-CAMs) that APSAM directly contrasts with."}],"review_version":1}