{"id":"fb569a7c-83a7-4a55-b0d0-10f923398075","arxiv_id":"2501.12487","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"fabSAM couples a Deeplabv3+ prompter with fine-tuned SAM decoders, improving mIOU on AI4Boundaries and AI4SmallFarms over zero-shot SAM and Deeplabv3+ by 4.9 to 23.5 percentage points.","lead":"The paper presents fabSAM, a system that combines a Deeplabv3+ model with the Segment Anything Model to outline farmland from satellite images. It reports higher boundary accuracy on two public field boundary datasets, suggesting a path toward automated global farmland maps from open data such as Sentinel-2.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline mIOU gains are not statistically anchored: a single 70/15/15 split with no error bars, plus crop-level preprocessing that may leak across train/test, leaves the 23.47%/15.10% and 4.87%/12.50% improvements unverified.","rationale":"The reader's weakest_assumption already identified the lack of error bars and single-split evaluation, and I agree that this is the main vulnerability. I would sharpen it: the more specific hazard is that the 256×256 cropping combined with an unspecified split level can create spatial overlap between training and test tiles, directly inflating the boundary IoU numbers that drive the mIOU improvement. This is not an allegation of dishonesty; it is a missing methodological detail that must be resolved before the numbers can be trusted. The paper does provide some independent support: the architecture is clearly described, the ablation in Table 3 confirms that decoder fine-tuning matters (removing it drops region IoU from 60.55 to 44.23 and boundary IoU from 37.4 to 6.85), and the paper explicitly acknowledges in Section 4.4.2 that the Prompter is the main accuracy bottleneck. Those are real strengths. But the central claim is a quantitative superiority claim, and the evaluation protocol currently lacks the statistical controls needed to support it. A scene-level cross-validation with multiple seeds is a concrete, feasible check that would settle whether the gains are genuine or an artifact of split leakage and variance. Since the reader already reached CONDITIONAL and my concern is a refinement of the same weakness rather than a new fatal flaw, the verdict should remain unchanged.","tokens_in":11124,"tokens_out":6219,"duration_ms":71438,"concrete_test":"Re-run the full pipeline with 5-fold cross-validation at the parent-image/scene level, ensuring all 256×256 crops from one original AI4B or AI4S image stay in the same fold. Repeat with at least 3 random seeds and report mean ± std for region IoU, boundary IoU, and mIOU. If the Deeplabv3+ mIOU advantage falls below ~1 percentage point or boundary IoU gains are inconsistent across folds and seeds, the headline quantitative claim is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: fabSAM improves mIOU by 23.47% and 15.10% over zero-shot SAM and by 4.87% and 12.50% over Deeplabv3+. For that claim to hold, the evaluation must be a stable, leakage-free estimate. Section 4.1 reports that images were cropped to 256×256 pixels and then 70/15/15 split, but it does not say whether the split was performed at the parent-image level or at the crop level. If crops from the same large scene appear in both training and test sets, the model can memorize local parcel geometry rather than learning transferable boundary delineation. This is especially dangerous for the boundary IoU metric, which is the largest contributor to the reported mIOU gain (e.g., 37.40 vs 28.99 on AI4B, 27.62 vs 3.25 on AI4S). The paper's own Section 4.4.2 admits that accuracy depends mainly on the Prompter and that fragmented farmland is missed, and Table 2 shows only tiny region-IoU gains over Deeplabv3+ (60.64 vs 59.30 on AI4B; 84.93 vs 84.30 on AI4S). Thus the 'significant improvement in farmland region identification and boundary delineation' is carried almost entirely by the boundary class, which is exactly the class most sensitive to spatial leakage and to run-to-run variation. With no error bars, no multiple seeds, and no clarified split protocol, the numerical superiority of fabSAM is not yet established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes fabSAM, a two-block framework for farmland boundary delineation from satellite imagery. A Deeplabv3+-based Prompter generates mask and point prompts, which are fed into a SAM-based block with a frozen image encoder, a fine-tuned prompt encoder, and two separately fine-tuned mask decoders for region identification and boundary delineation. The method is evaluated on the AI4Boundaries (AI4B) and AI4SmallFarms (AI4S) datasets. The abstract and Section 4.4.1 report mIOU improvements of 23.47% and 15.10% over zero-shot SAM, and 4.87% and 12.50% over Deeplabv3+, respectively. The paper also claims to be the first to combine a mask-prompt generator with a SAM-based block for this task.","tokens_in":11459,"tokens_out":4094,"duration_ms":40379,"significance":"The proposed architecture is simple and pragmatic: it reuses SAM's strong image features and adds lightweight, task-specific decoders, which is a promising recipe for remote sensing segmentation. The boundary IoU gains, if reproducible, would be practically useful for parcel mapping from open Sentinel-2 data. The paper also honestly discloses in Section 4.4.2 that accuracy depends mainly on the Prompter and that fragmented farmland is missed, which is a genuine limitation. The central weakness is that the quantitative evidence is not statistically anchored: there is a single data split, no error bars, and the crop-level preprocessing raises a potential leakage concern. These issues must be resolved before the headline improvements can be accepted.","major_comments":[{"comment":"The manuscript states that images were cropped to 256×256 pixels and then split 70/15/15, but it does not state whether the split was performed at the level of the original scenes or at the crop level. If crops from the same large scene appear in both the training and test sets, the model can memorize local parcel geometry rather than learning transferable boundary delineation, which would particularly inflate the boundary IoU values in Table 2. Please clarify the split protocol and, if the split is at the crop level, repeat the evaluation with a scene-level split or provide evidence that the reported gains persist.","section":"Section 4.1"},{"comment":"The reported mIOU improvements are single-run point estimates without error bars, confidence intervals, or multiple seeds. This is a load-bearing issue because Section 3.1 introduces stochastic point-prompt generation and the fine-tuning involves random initialization, so the 4.87% and 12.50% gains over Deeplabv3+ may not be stable. Please provide at least three independent runs with mean and standard deviation, or a paired statistical test, for the key comparisons in Table 2.","section":"Table 2 / Section 4.4.1"},{"comment":"The point-prompt probability thresholds (0.7 for foreground, 0.3 for background) are described as 'selected based on the experiment results' without specifying which data were used for this selection. If the test split was used to choose these thresholds, the comparison in Table 2 is partially circular. Please state explicitly that the thresholds were chosen on the validation split only, and report the sensitivity of the results to these values.","section":"Section 3.1"},{"comment":"The abstract claims 'significant improvement in farmland region identification and boundary delineation,' but Table 2 shows that the region IoU gains over Deeplabv3+ are small (60.64 vs 59.30 on AI4B; 84.93 vs 84.30 on AI4S). The mIOU gains are driven almost entirely by the boundary class, and Section 4.4.2 itself concedes that 'the improvement in the performance of fabSAM on region identification is not significant.' Please revise the abstract, the introduction, and the conclusion to separate the region-identification claim from the boundary-delineation claim, and avoid the phrase 'significant improvement' for both tasks jointly.","section":"Abstract and Section 4.4.1"}],"minor_comments":[{"comment":"Table 1 is titled 'Parameters setting for model training and fine-tuning' but omits the weights wD and wF used in Eq. (6); please add these values for reproducibility.","section":"Table 1"},{"comment":"The loss formula for LP, referenced as Eq. (3), is missing from the rendered text; the equation should be inserted so that the main and auxiliary loss combination is fully specified.","section":"Section 3.3.1"},{"comment":"The definition of mIOU as the average of IoUr and IoUb (Eq. 8) is nonstandard; please clarify in the text that this is an average over the region and boundary classes only, not the usual mean over all semantic classes.","section":"Section 4.3"},{"comment":"The caption contains a typo: 'de-lineation' should be 'delineation'.","section":"Figure 1 caption"},{"comment":"The novelty claim of being 'the first to introduce a hybrid architecture including a mask-prompt generator and a SAM-based block' is difficult to verify given the related work on GeoSAM and Few-shot SLVM; please soften the claim or add a more explicit comparison of how the proposed mask-prompt formulation differs from these existing mask-prompt approaches.","section":"Section 1"}],"recommendation":"major_revision","confidential_remarks":"I concur with the reader's assessment that the method is plausible but the empirical evidence is not yet statistically grounded. The main fixes—clarifying the split level, adding repeated runs, and explicitly separating threshold selection from test evaluation—are within scope and do not require a fundamental redesign. The paper fits the journal's scope; however, the 'first hybrid architecture' claim should be toned down or carefully contextualized, as GeoSAM already uses a mask prompter with SAM, and the authors' distinction is not crisply stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid engineering contribution that combines a Deeplabv3+ prompter with two fine-tuned SAM decoders for farmland region and boundary delineation. The assembly is new in this specific form, though every component is borrowed—GeoSAM already did CNN mask prompting and decoder fine-tuning, and the paper cites it. What it does well: the architecture is clearly described, the ablation shows decoder fine-tuning is the main driver of the gain, and the authors are admirably explicit that region identification gains are small and that the prompter is the bottleneck. The visual evidence in Fig. 3 and 4 is consistent with the boundary improvements.\n\nThe soft spots are in the evaluation, not the method. The reported 23.47%/15.10% mIOU gains over zero-shot SAM and 4.87%/12.50% over Deeplabv3+ come from a single run with a single fixed split and no error bars. More concerning, Section 4.1 says images were cropped to 256×256 and then split 70/15/15, but does not say whether the split was at the parent-image level or the crop level. If crops from the same scene appear in both training and test, the boundary IoU—the metric carrying most of the gain—can be inflated by memorization. The 0.7/0.3 point-prompt thresholds are also selected on the same datasets, adding a post-hoc tuning element. The abstract's claim of \"significant improvement in farmland region identification\" overstates what Table 2 shows: region IoU gains over Deeplabv3+ are 1.3 and 0.6 points.\n\nNone of this is fatal. The architecture is plausible, the ablation is informative, and the authors genuinely flag the prompter-dependence limitation in Section 4.4.2. The paper deserves to go to peer review, but it needs a revision that clarifies the split protocol, reports variance across seeds or folds, and tempers the abstract. If the split is at the parent-image level and the numbers hold with error bars, this is a useful baseline for SAM-based field boundary mapping. Without that, the quantitative claim is unverified.\n\nFor whom: researchers working on SAM adaptation for remote sensing, especially field boundary delineation. It is a reasonable read for a reading group, but not a must-cite until the evaluation is tightened.\n\nRecommendation: engage with it; send to review with a request for major revision.","headline":"A competent SAM-based farmland delineation pipeline with honest self-assessment, but the headline mIOU gains rest on a single split and an ambiguous crop-level protocol, so the numbers are not yet trustworthy.","tokens_in":12018,"tokens_out":2763,"would_cite":false,"duration_ms":26265,"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":"A Deeplabv3+-based prompter feeding mask and point prompts into fine-tuned SAM decoders improves farmland boundary delineation by up to 23 percentage points on two benchmarks.","keywords":["fabSAM","farmland boundary delineation","Segment Anything Model","semantic segmentation","prompt engineering","remote sensing","AI4Boundaries","AI4SmallFarms"],"falsifier":"Take a third Sentinel-2 farmland region not represented in AI4B or AI4S, run fabSAM and Deeplabv3+ over several random seeds, and compare boundary IoU; if the gap is within seed noise or reverses, then the reported gains are tied to the two benchmark distributions rather than to the architecture.","tokens_in":10919,"feed_emoji":"🌾","tokens_out":7832,"duration_ms":71082,"temperature":0.7,"pith_summary":"This paper proposes fabSAM, a farmland-boundary delineation pipeline that combines a Deeplabv3+-based prompter with a Segment Anything Model (SAM) block to map agricultural parcels from satellite images. The prompter produces both a coarse mask and selected foreground and background points, and two fine-tuned SAM decoders use those prompts to predict farmland regions and boundaries separately. On the AI4Boundaries and AI4SmallFarms datasets, fabSAM raises mean IoU by 23.47 and 15.10 percentage points over zero-shot SAM and by 4.87 and 12.50 points over Deeplabv3+ alone. The larger gains come on boundary delineation, especially for smallholder farms, where boundary IoU rises from 3.25 to 27.62. The paper argues this makes automated, continent-scale farmland maps from open 10-meter Sentinel-2 imagery more practical.","feed_headline":"Hybrid SAM pipeline lifts farmland-boundary accuracy by 23 points","feed_subtitle":"A Deeplabv3+ prompter feeds mask and point prompts into SAM, beating zero-shot SAM on two farmland datasets.","key_machinery":"The load-bearing mechanism is the prompt pathway from Deeplabv3+ to SAM. Deeplabv3+ predicts a farmland logit map; that map is fed to SAM's prompt encoder as a mask prompt, and a sampler converts the sigmoid probability map into point prompts, choosing foreground pixels with probability above 0.7, background pixels below 0.3, and favoring higher-confidence pixels. SAM's frozen image encoder supplies high-level visual features, and two separately fine-tuned decoders combine those features with the prompt embeddings to output region and boundary masks. A post-processor takes the symmetric difference between the two masks to produce closed parcel boundaries. The design lets any segmentation model be reused as an automatic prompt generator and treats SAM's block as a trainable post-processor.","core_discovery":"fabSAM's central claim is that a classical segmentation model can act as an automatic prompt generator for SAM, and that fine-tuning SAM's prompt encoder and two lightweight decoders on those prompts yields better farmland region and boundary masks than either the prompter or zero-shot SAM alone. The prompter is Deeplabv3+; its logits are used directly as a mask prompt, and a point generator samples reliable foreground pixels (probability above 0.7) and background pixels (below 0.3) from the sigmoid of those logits. The SAM image encoder is kept frozen, while the mask decoder and boundary decoder are trained separately with a Dice-plus-Focal loss. On the AI4Boundaries and AI4SmallFarms benchmarks this improves mIOU by 23.47 and 15.10 points over zero-shot SAM and by 4.87 and 12.50 points over Deeplabv3+, with the largest absolute improvement in boundary IoU. The paper further claims this is the first hybrid architecture to combine a mask-prompt generator with a SAM-based block specifically for farmland boundary delineation.","pith_inferences":["Because the paper reports a single train/validation/test split, a fair reading is that the headline margins are untested for variance; a natural extension is to rerun with several seeds and report error bars.","The ablation suggests the point prompt contributes about twice as much as the mask prompt, so a prompt generator optimized for reliable foreground and background points may matter more than denser masks.","The stated Prompter bottleneck implies fabSAM's gains should scale with Prompter quality: pairing the same SAM block with a stronger or domain-adapted segmenter is a cheap testable path to further improvement.","Since SAM's image encoder is frozen at its native patch size, applying fabSAM to very high-resolution imagery may require a boundary-refinement decoder or tiling, a limitation the paper does not address."],"forward_implications":["Farmland region and boundary maps can be generated automatically from 10-meter Sentinel-2 imagery without human annotation, which supports crop monitoring and agricultural census at large scale.","The approach generalizes beyond farmland: any task where a classical segmenter produces rough masks can use the same mask-and-point prompt recipe to sharpen outputs with a fine-tuned SAM decoder.","Boundary delineation, not just region overlap, is where the hybrid method pays off, so precision-agriculture applications that need parcel perimeters gain the most.","Separate decoders for region and boundary with task-specific fine-tuning allow each objective to be optimized without compromising the other.","On the AI4SmallFarms benchmark, boundary IoU improves from 3.25 with Deeplabv3+ to 27.62 with fabSAM, suggesting the method is especially useful in fragmented smallholder landscapes."],"supporting_citations":[{"why":"Supplies the frozen SAM image encoder, prompt encoder, and decoding structure that fabSAM adapts.","marker":"[17]"},{"why":"Defines the Deeplabv3+ encoder-decoder used as the Prompter that generates mask and point prompts.","marker":"[32]"},{"why":"Provides the AI4Boundaries benchmark and Sentinel-2 subset on which fabSAM is trained and evaluated.","marker":"[19]"},{"why":"Provides the AI4SmallFarms benchmark for fragmented smallholder farmland where the largest boundary gains appear.","marker":"[20]"},{"why":"Contributes the sparse-plus-dense prompting and decoder fine-tuning recipe, including the Dice-plus-Focal loss, that fabSAM adapts.","marker":"[25]"},{"why":"Shows a prior point-prompt SAM application to crop-field boundary delineation, the direct predecessor line fabSAM extends.","marker":"[18]"}],"fun_headline_variants":["SAM + Deeplabv3+ prompter boosts farmland boundary mIOU by 23%","Prompter-fed SAM lifts farmland boundary accuracy 23% over zero-shot","Deeplabv3+ prompts SAM: farmland boundary mIOU jumps 23%","Hybrid Deeplabv3+ SAM model sharpens farmland boundaries by 23%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method inherits the errors of its initial Deeplabv3+ stage: if that stage misses an area of farmland, the SAM block only refines the prompts it receives and cannot recover the missed area.","fun_headline_variants_meta":{"raw":{"variants":["SAM + Deeplabv3+ prompter boosts farmland boundary mIOU by 23%","Prompter-fed SAM lifts farmland boundary accuracy 23% over zero-shot","Deeplabv3+ prompts SAM: farmland boundary mIOU jumps 23%","Hybrid Deeplabv3+ SAM model sharpens farmland boundaries by 23%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001019,"raw_usage":{"total_tokens":4332,"prompt_tokens":1008,"completion_tokens":3324,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":624,"completion_tokens_details":{"reasoning_tokens":3228}},"tokens_in":624,"tokens_out":3324,"duration_ms":23290,"temperature":1.0,"reasoning_tokens":3228,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:07:55.483111+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a third Sentinel-2 farmland region not represented in AI4B or AI4S, run fabSAM and Deeplabv3+ over several random seeds, and compare boundary IoU; if the gap is within seed noise or reverses, then the reported gains are tied to the two benchmark distributions rather than to the architecture.","supporting_citations":[{"cited_title":"Persello, J","cited_arxiv_id":null,"evidence_quote":"Provides the AI4SmallFarms benchmark for fragmented smallholder farmland where the largest boundary gains appear."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows a prior point-prompt SAM application to crop-field boundary delineation, the direct predecessor line fabSAM extends."}],"review_version":1}