{"id":"268f46eb-5302-4a67-9d7e-163c12a71da3","arxiv_id":"2608.11064","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An entropy-based sampling explanation method for segmentation, paired with a high-salience removal test, is reported to outperform Grad-CAM, Score-CAM, and Seg-Sobol on WHU building footprint segmentation.","lead":"This paper introduces Entropy-Centric, an explanation method for image segmentation models that scores image regions by how much masking them raises the model's uncertainty, measured as entropy. It also proposes a new test for whether highlighted regions truly matter, and reports that the method beats three existing explanation techniques on a building segmentation benchmark.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"H-SIT is defined by the same equation as L-SIT (Eqs. 9 vs. 10), so the complementary test that underpins the claimed superiority is ambiguous as written.","rationale":"The central claim is comparative superiority, and the paper's quantitative support is Table I. The L-SIT columns alone are mixed: Score-CAM has a smaller confidence drop (0.6% vs. 2.0%), although Entropy-Centric has better IoU and entropy drops. Therefore the H-SIT columns carry much of the claimed advantage. As written, H-SIT is not a clearly distinct operation: Eq. (10) is the same expression as Eq. (9), with only a contradictory redefinition of φ. This is an internal inconsistency in the evaluation methodology, not merely a disagreement with consensus. The unspecified source of the mask M in Eq. (5) is a separate but related ambiguity that the reader already identified; it affects whether the method is a black-box tool and whether the perturbed-image experiments are meaningful. I am not claiming the method is wrong: the entropy-based attribution idea is plausible, the qualitative examples are suggestive, and the approach is computationally reasonable. But the paper as submitted does not make the evaluation reproducible, and no code or data are provided. This reinforces rather than changes the reader's CONDITIONAL verdict; I keep the verdict as UNCHANGED.","tokens_in":6626,"tokens_out":10438,"duration_ms":95512,"concrete_test":"Request the exact pseudocode or code for Eqs. (9)-(10) and Eq. (5). Rerun Table I on the WHU set with a corrected H-SIT that actually masks high-salience non-target pixels, e.g., I'_H-SIT = T ∪ (¬φ_high), and with M taken from the model's predicted mask rather than ground truth. If Entropy-Centric's H-SIT drops (37.4% confidence, 44% IoU, 48.5% entropy) change materially or fall below Score-CAM's (32.3%, 36.2%, 45.3%), the claimed superiority is not supported; if they persist, the notation issue is benign.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The quantitative case for superiority rests mainly on Table I, especially the H-SIT columns where Entropy-Centric shows the largest drops. But Section III-A defines L-SIT as I'_L-SIT = T ∪ φ with φ the set of highlighted pixels above a threshold, and H-SIT as I'_H-SIT = T ∪ φ, now saying φ is 'highlighted pixels below a specific threshold' while claiming that out-of-object high-saliency regions are removed. The formula does not encode that removal; if φ is high-salience the image keeps those regions, and if φ is low-salience the description is inverted. Thus the two tests differ only by an unstated redefinition of φ, making H-SIT not a well-defined complementary evaluation. The problem is compounded by Eq. (5): the object mask M used to isolate the target object is never specified as ground truth or model prediction. If M is ground truth, the method cannot explain unlabeled images despite the black-box framing; if M is the model's prediction, the explanation is contaminated by segmentation errors. Both ambiguities must be resolved before the reported H-SIT superiority can be trusted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Entropy-Centric, a sampling-based XAI method for semantic segmentation that computes patch-importance scores as the mean absolute difference in binary entropy of the target class under masked-input perturbations drawn from quasi-Monte Carlo sampling. It also introduces a new evaluation protocol, the High-Salience Influence Test (H-SIT), intended to complement the existing Low-Salience Irrelevance Test (L-SIT) by removing high-salience out-of-object regions. The authors report results on the WHU building segmentation dataset indicating that Entropy-Centric yields smaller L-SIT drops and larger H-SIT drops than Grad-CAM, Score-CAM, and Seg-Sobol, which they interpret as evidence that Entropy-Centric more accurately identifies decision-critical regions.","tokens_in":6851,"tokens_out":7992,"duration_ms":67922,"significance":"The proposed method is attractive in principle: it is a black-box, architecture-agnostic explainer requiring no internal activations, and it has no fitted parameters beyond a sampling grid and threshold. The H-SIT idea of testing whether out-of-object high-salience regions are truly influential is a useful complement to L-SIT. The paper also evaluates faithfulness through multiple metrics (confidence, IoU, and entropy), which is a strength. However, the current evidence is too thin to support the headline claim of superiority: the experiments use one dataset, one model, one threshold, and no error bars or significance tests, and key definitional ambiguities (especially the source of the object mask and the formal identity of the L-SIT and H-SIT equations) must be resolved before the method and its evaluation can be reproduced or trusted. If these open points are addressed, the method could be a valuable contribution to XAI for semantic segmentation.","major_comments":[{"comment":"The object mask M in Eq. (5) is never specified. If M is the ground-truth segmentation mask, the method cannot explain an unlabeled image and is not a black-box explainer as claimed in the abstract and Table II. If M is the model's predicted mask, the entropy in Eqs. (7)-(8) is computed only over the model's own predicted object, so segmentation errors directly bias the attribution scores. This is load-bearing because the entire importance computation is masked by M. Please state explicitly where M comes from and, if it is the predicted mask, discuss the sensitivity of the results to segmentation errors.","section":"Section III, Eq. (5)"},{"comment":"As written, the equations for I'_L-SIT and I'_H-SIT are literally identical (T ∪ φ), with φ redefined in words from \"highlighted pixels above a specific threshold\" in Eq. (9) to \"highlighted pixels below a specific threshold\" in Eq. (10). This makes the complementary evaluation ambiguous and prone to misimplementation. Please use distinct symbols (e.g., φ_high and φ_low) and explicitly define the perturbation as keeping or removing the relevant set, for example I'_H-SIT = T ∪ φ_low means that high-salience out-of-object pixels are removed while low-salience pixels and the target are retained.","section":"Section III-A, Eqs. (9)-(10)"},{"comment":"The quantitative superiority claim rests on a single dataset, a single U-Net model, one threshold (0.1), and no error bars or significance tests. The test set appears to consist of 142 images (per the Figure 3 caption), and several comparisons are close (e.g., L-SIT confidence drop: 2% for Entropy-Centric vs. 0.6% for Score-CAM; H-SIT entropy increase: 48.5% vs. 45.3%). The qualitative samples in Figure 3 are explicitly selected, so they cannot support a general superiority claim. Please provide per-image distributions, confidence intervals or significance tests, and ideally results across multiple thresholds and architectures.","section":"Section IV, Table I and Figure 3"},{"comment":"The H-SIT entropy metric measures the same entropy change that Entropy-Centric is designed to maximize, so its large entropy increase is partly by construction. The confidence and IoU drops provide more independent grounding, but the paper should explicitly discuss this potential circularity and include at least one evaluation metric that does not share the method's objective (e.g., change in a separate segmentation quality metric or agreement with an independent explanation method).","section":"Section III-A and Table I"},{"comment":"The method description omits implementation details required for reproducibility: the number of perturbation samples N, the patch sampling grid and resolution, the upsampling procedure for the masks, and the exact meaning of the 0.1 threshold (absolute salience, quantile, or fraction of pixels). Without these details, the experiments cannot be repeated or fairly compared across methods.","section":"Section III"}],"minor_comments":[{"comment":"The word \"Proposed\" is misspelled as \"Propsoed\" in the caption.","section":"Figure 1 caption"},{"comment":"The text \"dimesniosn\" should be \"dimensions.\"","section":"Section III"},{"comment":"Equation (4) defines binary entropy H as a scalar function of a probability p_k, but Eq. (8) applies H to the entire probability map P. Please clarify that H is computed per-pixel and specify how the per-pixel values are aggregated (e.g., mean over the masked object region) to yield the scalar entropy score appearing in Eq. (8).","section":"Section III, Eq. (4)"},{"comment":"Section II reviews the conventional Sobol method, but the connection between that variance-based estimator and the proposed entropy-based method is not made explicit; please state that Entropy-Centric adopts the QMC sampling and Jansen-estimator framework but replaces variance with entropy, to avoid confusing the reader into thinking the conventional Sobol indices are actually computed.","section":"Section II"},{"comment":"The qualitative ratings (+++, ++++) in Table II are presented without a methodology for how they are assigned; please clarify the scale and how these ratings are supported by the experimental results, or rephrase them as qualitative attributes rather than scores.","section":"Table II"},{"comment":"References [9] and [10] contain unusual volume/page data (\"vol. 1, no. 0\"); please verify these citations or use more standard bibliographic information.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is in the spirit of a workshop or short paper. The entropy-based perturbation idea is a reasonable incremental extension of Sobol-based black-box explainers, and the H-SIT metric could be a useful evaluation addition. However, the central claim of superiority is not yet credible due to the undefined object mask, the ambiguous H-SIT definition, and the lack of statistical rigor. For a journal, the authors will need to substantially strengthen the experimental protocol and formalize the definitions. I would not reject outright because the core approach is sound in principle and the issues appear addressable within the scope of the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a modest but real extension: it swaps variance for binary entropy in the Sobol attribution scheme and tests a complementary high-salience removal test. The estimator in Eq. (8) is a concrete, implementable recipe, and the H-SIT idea (removing high-salience out-of-object regions) is a sensible evaluation angle that L-SIT misses. The authors also deserve credit for the black-box framing and for comparing against Seg-Sobol, Score-CAM, and Grad-CAM on a standard remote-sensing dataset. The qualitative samples show an honest effort to report failure cases (Score-CAM skipping buildings, Grad-CAM fragmenting them), which is more than many XAI papers bother to do.\n\nThe soft spots are real but not fatal. The biggest one is Eq. (5): the object mask M is never specified as ground truth or predicted mask. The reader flagged this and it is load-bearing. If M is ground truth, the method cannot explain an unlabeled image, contradicting the black-box promise. If M is the model's own prediction, the explanation is contaminated by segmentation errors. The paper must state the source and, ideally, run both variants. Second, the stress-test note is correct: Eqs. (9) and (10) are identical as written, and the difference between L-SIT and H-SIT is smuggled into the ambiguous phrase 'below a specific threshold.' That needs a clean formal definition, e.g., H-SIT masks out-of-object pixels whose attribution exceeds tau. Third, the evaluation uses one dataset, one model, one threshold, no error bars, and the entropy metric that H-SIT rewards is the same entropy change that Entropy-Centric maximizes. The confidence and IoU drops partly mitigate that circularity, but the paper does not report per-image variability or significance tests.\n\nThe quantitative claim of superiority is plausible but not fully supported. The method is a variation on [9] and [11], not a new framework, and the paper's significance is contained to XAI evaluation for segmentation. Still, the core idea is coherent, the math is simple enough to check, and there are no fitted parameters or hidden data. The citation pattern looks fine — the key related work is cited, and the self-citations are to the immediate precursors.\n\nWho gets value? Researchers working on black-box attribution for dense prediction, especially in remote sensing, and anyone building evaluation protocols for XAI faithfulness. It deserves a serious referee, though I would ask for the mask-source clarification, a corrected H-SIT definition, and error bars before acceptance.","headline":"A workable entropy-based Sobol attribution variant for segmentation, undermined by an unspecified object mask and an evaluation metric that partly encodes the method's own objective.","tokens_in":7362,"tokens_out":631,"would_cite":true,"duration_ms":7307,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Entropy-Centric identifies the regions a segmentation model truly relies on, and outperforms Grad-CAM, Score-CAM, and Seg-Sobol on building footprint explanations.","keywords":["Explainable AI","semantic segmentation","remote sensing","entropy","Sobol sensitivity analysis","attribution maps","building extraction","black-box explanation"],"falsifier":"Run Entropy-Centric on the same images twice, once with $M$ as the ground-truth building mask and once with $M$ as the model's predicted mask, and compare the H-SIT drops. If the attribution maps rank differently, or if the reported H-SIT advantage over Score-CAM disappears when predicted masks are used, then the claimed superiority depends on the unspecified mask source.","tokens_in":6433,"feed_emoji":"🛰️","tokens_out":8410,"duration_ms":72397,"temperature":0.7,"pith_summary":"This paper introduces Entropy-Centric, a black-box explainability method for semantic segmentation that ranks image regions by how much masking each region changes the binary entropy of the target class in the model's output. The intended contribution is an attribution map that pinpoints decision-critical regions without requiring gradients or internal activations, plus a test, H-SIT, that exposes false high-salience regions. On WHU building-footprint segmentation with a U-Net model, the paper reports that Entropy-Centric outperforms Grad-CAM, Score-CAM, and Seg-Sobol on both low-salience irrelevance and high-salience influence tests. If correct, this gives remote-sensing users a model-agnostic way to audit why a segmentation model labels a building. The core confidence figure is a 37.4% drop in prediction confidence when Entropy-Centric's high-salience regions are masked, against 32.3% for Score-CAM and 2.2% for Grad-CAM.","feed_headline":"Entropy method sharpens explanations of building-segmentation AI","feed_subtitle":"Perturbation-based attribution drops confidence 37.4% when true hot spots are masked, beating three rivals.","key_machinery":"The central object is the per-pixel binary entropy $H(p) = -p\\log p - (1-p)\\log(1-p)$ of the target class probability $p$, restricted to the target object by element-wise multiplication with the mask $M$. The importance score is the mean over paired samples of $|H(C_{i,j}) - H(A_j)|$, where $C_{i,j}$ and $A_j$ are softmax outputs from perturbed and base quasi-Monte Carlo masks. This converts Sobol's variance-based sensitivity into an uncertainty-based sensitivity: regions whose masking changes target-class certainty are judged decision-critical. The accompanying H-SIT test measures whether masking high-salience out-of-object regions actually degrades predictions, providing evidence that the attribution map is not simply highlighting the object itself.","core_discovery":"The paper claims that a segmentation model's decision-critical regions can be found by entropic sensitivity: generate paired base and perturbed images from quasi-Monte Carlo masks, run the model, isolate the target class probability map by multiplying the output by the object mask $M$, compute per-pixel binary entropy of the target class probability, and assign each patch an importance score equal to the mean absolute entropy difference between perturbed and base samples. The claim is that this score identifies irrelevant areas (removing them barely changes confidence, IoU, or entropy) and critical areas (removing them sharply degrades all three) more faithfully than Grad-CAM, Score-CAM, and Seg-Sobol. The paper also proposes H-SIT, a complementary evaluation that masks high-salience regions outside the target object to catch false positives in attribution maps.","pith_inferences":["Because the paper leaves the source of $M$ unspecified, a deployment-ready version of Entropy-Centric would need to derive the mask from the model's own prediction; comparing that variant against a ground-truth-mask version is an immediate next step.","Because binary entropy generalizes to multi-class probability vectors, the same per-patch entropy-difference score should extend to multi-class segmentation and object detection, not just binary building masks.","A direct stress test would replace the perturbation masks with random masks of equal area: if random masking produces comparable H-SIT drops, then the attribution map is not discriminating critical regions.","The patch-based attribution could be combined with a finer-grained gradient method, using the entropy signal as a global constraint to recover pixel-level maps while keeping the black-box property."],"forward_implications":["Entropy-Centric explanations are computable for any segmentation model that accepts image patches, including pipelines with non-differentiable components, because they rely only on input perturbations and output probabilities.","The H-SIT test gives a quantitative handle on false-positive salience, complementing L-SIT's check for false-negative salience; together they can rank XAI methods by fidelity.","The reported drop pattern implies that buildings plus surrounding context, not just building pixels, drive the U-Net's segmentation decisions.","Grad-CAM's nearly zero H-SIT drop (2.2% confidence) indicates that its heatmaps miss cross-region context, while Entropy-Centric and Score-CAM capture it."],"supporting_citations":[{"why":"Supplies the rooftop U-Net segmentation model used as the case study for the XAI comparison.","marker":"[8]"},{"why":"Supplies the CAM-based XAI baselines adapted to segmentation and the L-SIT evaluation protocol that H-SIT extends.","marker":"[9]"},{"why":"Supplies the WHU building footprint dataset on which the segmentation model is trained and evaluated.","marker":"[10]"},{"why":"Supplies Sobol sensitivity analysis, quasi-Monte Carlo sampling, and the Jansen estimator that Entropy-Centric adapts and Seg-Sobol uses as baseline.","marker":"[11]"}],"fun_headline_variants":["Entropy spotlight beats rival XAI in building-segmentation maps","Entropy exposes true decision zones in remote-sensing AI","Segmentation AI self-explains via entropy, besting Grad-CAM","Entropy pinpoints AI's true focus in satellite-image segmentation","Entropy-based XAI trumps rivals for building segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the object mask $M$ used to isolate the target object's output is available and correct; the paper never states whether $M$ is the ground-truth mask or the model's own predicted mask, and the entropy scores are computed only inside this mask.","fun_headline_variants_meta":{"raw":{"variants":["Entropy spotlight beats rival XAI in building-segmentation maps","Entropy exposes true decision zones in remote-sensing AI","Segmentation AI self-explains via entropy, besting Grad-CAM","Entropy pinpoints AI's true focus in satellite-image segmentation","Entropy-based XAI trumps rivals for building segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000969,"raw_usage":{"total_tokens":4101,"prompt_tokens":901,"completion_tokens":3200,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":3113}},"tokens_in":517,"tokens_out":3200,"duration_ms":19665,"temperature":1.0,"reasoning_tokens":3113,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:57:18.279277+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Entropy-Centric on the same images twice, once with $M$ as the ground-truth building mask and once with $M$ as the model's predicted mask, and compare the H-SIT drops. If the attribution maps rank differently, or if the reported H-SIT advantage over Score-CAM disappears when predicted masks are used, then the claimed superiority depends on the unspecified mask source.","supporting_citations":[{"cited_title":"Interpretable Machine Learning , urldate =","cited_arxiv_id":null,"evidence_quote":"Supplies the rooftop U-Net segmentation model used as the case study for the XAI comparison."},{"cited_title":"The Moonlight Review , urldate =","cited_arxiv_id":null,"evidence_quote":"Supplies the CAM-based XAI baselines adapted to segmentation and the L-SIT evaluation protocol that H-SIT extends."},{"cited_title":"Sensors , VOLUME =","cited_arxiv_id":null,"evidence_quote":"Supplies the WHU building footprint dataset on which the segmentation model is trained and evaluated."},{"cited_title":"Sensors , VOLUME =","cited_arxiv_id":null,"evidence_quote":"Supplies Sobol sensitivity analysis, quasi-Monte Carlo sampling, and the Jansen estimator that Entropy-Centric adapts and Seg-Sobol uses as baseline."}],"review_version":1}