Pith. sign in

REVIEW 5 major objections 6 minor 33 references

Entropy-Centric Explainable AI for Remote Sensing Image Segmentation

T0 review · 5 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Entropy-Centric identifies the regions a segmentation model truly relies on, and outperforms Grad-CAM, Score-CAM, and Seg-Sobol on building footprint explanations.

desk verdict 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. read the letter →

arxiv 2608.11064 v1 pith:CRHKIT6L submitted 2026-08-11 cs.CV cs.AI

Ali Saleh , Abdul Karim Gizzini , Mohamad Ghassany , Ali J. Ghandour This is my paper · ORCID
classification cs.CVcs.AI
keywords ExplainableAIsemanticsegmentationremotesensingentropySobolsensitivityanalysisattributionmapsbuildingextractionblack-boxexplanation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

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.

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 (5)
  1. [Section III, Eq. (5)] 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.
  2. [Section III-A, Eqs. (9)-(10)] 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.
  3. [Section IV, Table I and Figure 3] 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.
  4. [Section III-A and Table I] 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).
  5. [Section III] 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.
minor comments (6)
  1. [Figure 1 caption] The word "Proposed" is misspelled as "Propsoed" in the caption.
  2. [Section III] The text "dimesniosn" should be "dimensions."
  3. [Section III, Eq. (4)] 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).
  4. [Section II] 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.
  5. [Table II] 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.
  6. [References] References [9] and [10] contain unusual volume/page data ("vol. 1, no. 0"); please verify these citations or use more standard bibliographic information.

Circularity Check

1 steps flagged · score 4.0 of 10

H-SIT's entropy metric is the same entropy-change objective that Entropy-Centric optimizes, so part of the reported superiority is self-confirming; confidence and IoU drops provide partial independent support.

  1. self definitional [Section III (method description before Eq. 4) and Section IV, metric (iii), Table I; Eq. (8)]
    "The greater the entropy change upon masking a region, the higher its attribution score. ... (iii) The increase in the entropy of the target object between the masked-image prediction, using L-SIT and H-SIT, and the original prediction."

    Entropy-Centric's importance score (Eq. 8) is the mean absolute difference in binary entropy between perturbed and base masked outputs: S_i = (1/N) Σ |H(P_{C_i,j}) - H(P_{A_j})|. H-SIT then removes the regions assigned the highest S_i and measures, as its primary quantitative evidence of 'decision-critical' relevance, the increase in that same binary entropy H. Thus the regions selected as important are exactly those that maximize the H-SIT entropy metric; a large entropy score is the optimization target, not an externally verified property. The confidence-drop and IoU-drop metrics are not identical to the objective, so the circularity is partial rather than total.

full rationale

The paper's central derivation chain is not circular in the sense of fitting parameters or importing a uniqueness theorem: Entropy-Centric is a parameter-free black-box attribution computed from perturbed forward passes, and its comparison against Grad-CAM, Score-CAM, and Seg-Sobol has independent content on the confidence and IoU metrics. However, one of the three headline metrics, the H-SIT entropy increase, is definitionally aligned with the method's objective: the method assigns high scores to regions whose masking changes entropy, and H-SIT scores the method by how much entropy increases when those regions are masked. The H-SIT/L-SIT equations are also printed identically (Eqs. 9 and 10) with only a verbal redefinition of φ, and the provenance of the target mask M in Eq. 5 is unspecified; these are validity concerns rather than circularity because they do not make the superiority claim reduce to a fit. Weighing the self-aligned entropy metric against the two independent metrics and the otherwise self-contained method, the circularity score is 4.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method is parameter-light, but it leans on an unstated object-mask source, an entropy-as-importance assumption, and a hand-picked threshold. The evaluation's entropy metric partly shares the method's objective, adding a mild circularity burden.

free parameters (3)
  • Salience threshold tau = 0.1
    Used for both L-SIT and H-SIT in Table I; no sensitivity analysis is reported, so all comparative numbers depend on this hand-picked value.
  • Perturbation sample count N = unspecified
    Equations (3) and (8) scale with N, but the results do not report N, so the computational cost and variance of the estimates cannot be audited.
  • Patch sampling grid and upsampling = unspecified
    The method is patch-based and depends on the chosen grid and interpolation in Eq. (6), but these choices are not stated.
assumptions (4)
  • standard math Sobol variance decomposition and Jansen estimator provide valid sensitivity estimates
    The method builds on Eqs. (1)-(3) from [11]; treated as background without proof.
  • domain assumption The object mask M in Eq. (5) is available and correctly identifies target object pixels
    Section III states the output is multiplied by M, but does not explain how M is obtained; if M is ground truth, the method cannot explain new images without labels.
  • domain assumption Binary entropy of the target class probability is a faithful proxy for the importance of an input region
    Section III assumes that larger entropy change upon masking means higher relevance; no theoretical or empirical justification beyond the method definition.
  • ad hoc to paper Absolute difference of entropies in Eq. (8) is an appropriate estimator
    The paper replaces the squared difference in the Jansen estimator with an absolute entropy difference without justifying the choice or comparing to other divergence measures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Entropy-Centric Explainable AI for Remote Sensing Image Segmentation." pith.science (2026). https://pith.science/paper/CRHKIT6L

@misc{pith2026260811064,
  author       = {Pith},
  title        = {Pith review of: Entropy-Centric Explainable AI for Remote Sensing Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CRHKIT6L}},
  note         = {Machine review of arXiv:2608.11064}
}
read the original abstract

Artificial intelligence (AI) has become a powerful approach to solving complex problems in critical domains. Many concerns arise regarding the decision-making process of its models, mainly due to deep neural networks outperforming their peers at the cost of ambiguity in feature extraction and prediction. Consequently, in critical domains such as remote sensing, where high-resolution imagery must be analyzed using black-box models, the lack of transparency limits trust in these models and, thus, their adoption. In light of this reality, explaining and understanding the complex decision-making process of AI models has become essential. Explainable AI (XAI) aims to bridge this gap by providing insights into how and why certain decisions are made. While significant progress has been achieved in explaining image classification tasks, image segmentation still offers considerable room for improvement. In this context, this paper proposes an entropy-centric XAI method for semantic segmentation. Moreover, a new XAI evaluation methodology is proposed to efficiently measure the relevance of the regions highlighted by the proposed XAI method. Experimental results demonstrate the superiority of the proposed XAI method compared with recently adapted XAI methods for semantic segmentation.

Figures

Figures reproduced from arXiv: 2608.11064 by the authors.

Figure 1
Figure 1. Propsoed Entropy-centric XAI Method. • Performance evaluation using the WHU dataset for build￾ing footprint segmentation [10]. The remainder of this paper is organized as follows: Section II presents the proposed entropy-centric XAI methodology. The performance of the benchmarked XAI methods in terms of the proposed XAI evaluation methodology is analyzed and discussed in Sections III and IV, respectively. II. CONVEN… view at source ↗
Figure 2
Figure 2. The evaluation framework proposed through Irrelative Validation and False Sailency Detection methods applied on 3 heatmaps generated by the XAI [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Sample of XAI heatmaps generated by the 3 methods Grad-CAM, Score-CAM and Entropy-Centric that summarizes all the cases ion the testing [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 28 canonical work pages

  1. [9]

    The Moonlight Review , urldate =

    Gradient-based Feature Attribution in Explainable AI: A Technical Review , year =. The Moonlight Review , urldate =

    work page

  2. [11]

    Sensors , VOLUME =

    Fahim, Masud An Nur Islam and Saqib, Nazmus and Siam, Shafkat Khan and Jung, Ho Yub , TITLE =. Sensors , VOLUME =. 2022 , NUMBER =

    work page 2022

  3. [1]

    Advances in Neural Information Processing Systems , year =

    Thomas Fel and R. Advances in Neural Information Processing Systems , year =

    work page

  4. [2]

    2019 , volume=

    Ji, Shunping and Wei, Shiqing and Lu, Meng , journal=. 2019 , volume=

    work page 2019

  5. [3]

    Shreim, H and Gizzini, A and Ghandour, A , journal=

    work page

  6. [4]

    Gizzini, A and Shukor, M and Ghandour, A , journal=

    work page

  7. [5]

    2022 , publisher=

    Nasrallah, Hasan and Samhat, Abed Ellatif and Shi, Yilei and Zhu, Xiao Xiang and Faour, Ghaleb and Ghandour, Ali J , journal=. 2022 , publisher=

    work page 2022

  8. [6]

    Attribution-based XAI Methods in Computer Vision: A Review

    Kumar Abhishek and Deeksha Kamath , year=. 2211.14736 , archivePrefix=

    work page Pith review arXiv

Show all 33 references
  1. [7]

    Proceedings of the AAAI Conference on Artificial Intelligence , publisher=

    Vinogradova, Kira and Dibrov, Alexandr and Myers, Gene , year=. Proceedings of the AAAI Conference on Artificial Intelligence , publisher=. doi:10.1609/aaai.v34i10.7244 , number=

    doi

  2. [8]

    Interpretable Machine Learning , urldate =

    Molnar, Christoph , title =. Interpretable Machine Learning , urldate =. 2024 , url =

    2024

  3. [10]

    Sensors , VOLUME =

    Cheng, Zhihan and Wu, Yue and Li, Yule and Cai, Lingfeng and Ihnaini, Baha , TITLE =. Sensors , VOLUME =. 2025 , NUMBER =

    2025

  4. [12]

    Guttag AND Arvind Satyanarayan , booktitle =

    Angie Boggust AND Harini Suresh AND Hendrik Strobelt AND John V. Guttag AND Arvind Satyanarayan , booktitle =. 2023 , doi =

    2023

  5. [13]

    Sangeetha and L

    V. Sangeetha and L. Agilandeeswari , keywords =. The Egyptian Journal of Remote Sensing and Space Sciences , volume =. 2025 , issn =. doi:https://doi.org/10.1016/j.ejrs.2025.06.004 , url =

    2025 doi

  6. [14]

    Journal of Environmental Management , volume =

    Márcia. Journal of Environmental Management , volume =. 2025 , issn =. doi:https://doi.org/10.1016/j.jenvman.2024.123864 , url =

    2025

  7. [15]

    Electronics , VOLUME =

    Miller, Tymoteusz and Durlik, Irmina and Kostecka, Ewelina and Kozlovska, Polina and Łobodzińska, Adrianna and Sokołowska, Sylwia and Nowy, Agnieszka , TITLE =. Electronics , VOLUME =. 2025 , NUMBER =

    2025

  8. [16]

    Kamu Yönetimi ve Teknoloji Dergisi , volume=

    Eren, Veysel and Duman, Hasret , keywords=. Kamu Yönetimi ve Teknoloji Dergisi , volume=. 2025 , DOI=

    2025

  9. [17]

    PAMM , volume =

    Raj, Rohit and Tismer, Alexander and Gaisser, Lukas and Riedelbauch, Stefan , title =. PAMM , volume =. doi:https://doi.org/10.1002/pamm.202400104 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1002/pamm.202400104 , abstract =

    doi

  10. [18]

    Morris , title =

    Max D. Morris , title =. Technometrics , volume =. 1991 , publisher =. doi:10.1080/00401706.1991.10484804 , URL=

    1991

  11. [19]

    Speith, Timo , booktitle=

  12. [20]

    P. J. Brockwell and R. A. Davis , publisher =. Time Series: Theory and Methods , year =

  13. [21]

    and Ferro-Famil, L

    Pottier, E. and Ferro-Famil, L. and Allain, S. and Cloude, S. and Hajnsek, I. and Papathanassiou, K. and Moreira, A. and Williams, M. and Minchella, A. and Lavalle, M. and Desnos, Y.-L. , booktitle =. Overview of the. 2009 , pages =. doi:10.1109/IGARSS.2009.5417532 , keywords =

    2009

  14. [22]

    R. C. Beal and A. D. Goldfinger and A. G. Tilly and W. J. Geckle , title =. 1981 , number =

    1981

  15. [23]

    A. C. Frery , journal =. How To Successfully Make a Scientific Contribution Through. 2015 , month =. doi:10.1109/LGRS.2015.2404211 , groups =

    2015

  16. [24]

    1998 , Edition =

    How to Write & Publish a Scientific Paper , Author =. 1998 , Edition =

    1998

  17. [25]

    The densities and distributions of the largest eigenvalue and the trace of a

    Drensky, Vesselin and Edelman, Alan and Genoar, Tierney and Kan, Raymond and Koev, Plamen , journal =. The densities and distributions of the largest eigenvalue and the trace of a. 2019 , issn =. doi:10.1142/s2010326321500106 , publisher =

    2019 doi

  18. [26]

    Hicham Bou Habib , url=

  19. [27]

    2024 , url =

    R: A Language and Environment for Statistical Computing , author =. 2024 , url =

    2024

  20. [28]

    2016 , isbn =

    Hadley Wickham , title =. 2016 , isbn =

    2016

  21. [29]

    2024 , note =

    ggthemes: Extra Themes, Scales and Geoms for ``ggplot2'' , author =. 2024 , note =

    2024

  22. [30]

    and Gambini, Juliana and Lucini, Magdalena , journal =

    Rey, Andrea and Frery, Alejandro C. and Gambini, Juliana and Lucini, Magdalena , journal =. Asymptotic distribution of entropies and. 2024 , issn =. doi:10.1016/j.chaos.2024.115481 , groups =

    2024

  23. [31]

    A. C. Frery and L. Gomez and A. C. Medeiros , journal =. A Badging System for Reproducibility and Replicability in Remote Sensing Research , year =. doi:10.1109/JSTARS.2020.3019418 , groups =

    2020

  24. [32]

    2001 , Edition =

    The Visual Display of Quantitative Information , Author =. 2001 , Edition =

    2001

  25. [33]

    E. R. Tufte , title =

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.