HNGT-Net: Hard-Negative Guided Topology Transfer for Lightweight Hyperspectral Small-Target Detection

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Please refer to the PDF file.

Comments for author File: Comments.pdf

Author Response

Comment 1: The use of the term “absolute zero-shot” is difficult to justify. The GT mask of each evaluation scene is used to remove anomalous pixels and construct a purified normal image, after which the model is trained on that same scene. This setting is closer to scene-specific normal-only or transductive learning. The authors should either revise the terminology or add a strict cross-scene experiment in which the training and testing images come from different scenes or sensors. In this context, hyperspectral remote sensing object detection via cross domain learning from visible images may be discussed to distinguish genuine domain transfer from within-scene adaptation.

Response 1: We sincerely thank the Reviewer for this precise and important comment. We fully agree with the Reviewer’s analysis: because the purified training image and the test image originate from the same scene, our protocol should not have been described as “absolute zero-shot.” Following the Reviewer’s suggestion, we have chosen to revise the terminology throughout the manuscript: the setting is now consistently described as “scene-specific normal-only”  and all “absolute zero-shot” phrasing has been removed. Specifically:

 Terminology revised to “scene-specific normal-only” throughout the manuscript; explicit protocol characterization added in Section 3.1; scope and future-work statements added in the Conclusion; domain-transfer discussion expanded in the Introduction. All changes are highlighted in red.

Comment 2: More information is needed about the frozen teacher. Its pretraining dataset, objective function, optimizer, training epochs, initialization, and possible exposure to the evaluation scenes are not reported. Since the teacher determines both the graph topology and the objective used to generate hard negatives, these details are essential for judging reproducibility and experimental fairness. The training settings and tuning budgets of the competing methods should also be clarified.

Response 2:  We sincerely thank the Reviewer for this insightful comment, and we apologize for the misleading wording in the original manuscript, which caused an understandable ambiguity. We would like to clarify that the teacher network is not pre-trained at all: it is randomly initialized once and then kept frozen throughout the entire training process. There is therefore no pre-training dataset, no pre-training loss, and no pre-training epochs involved. The phrase “pre-trained and kept frozen” in the original text was inaccurate, and we are grateful to the Reviewer for catching it.

We have added an explicit clarification paragraph in Section 2 (Proposed HNGT-Net Framework), stating that “the teacher network T(·) is randomly initialized once and remains frozen throughout training,” and explaining the above rationale. Algorithm 1 (line 1) now also explicitly states “Randomly initialize teacher T(·) and freeze all of its parameters.” All changes are highlighted in red. We hope this clarification fully addresses the Reviewer’s concern and improves the reproducibility of our work.

Comment 3: A clear ablation study is missing. The authors should separately remove L_node, L_struct, L_lap, SG-HNS, the margin loss, and the multiplicative score fusion. Sensitivity experiments on K, p, ϵ, margin m, the number of PGD steps, and the loss weights would also help determine whether the reported improvement is stable or dependent on carefully selected parameters.

Response 3: We sincerely thank the Reviewer for this constructive comment, which has substantially strengthened the experimental section. We have added two new subsections that directly address the requested analyses:

1. Ablation study (new Section 3.5, Table 8).We conducted a cumulative ablation that effectively covers the separate removal of each requested component: starting from the node-consistency-only baseline (L_node alone), the structural term (L_struct, weight β), the Laplacian term (L_lap, weight γ), the SG-HNS hard-negative separation term (weight ρ), and the response-margin constraint (m) are successively enabled at their selected values. Each column of Table 8 therefore corresponds to the model with the remaining components removed, allowing the individual contribution of every term to be read off directly.
2. Sensitivity analysis (new Section 3.4, Figures 5–8, Table 7).For the loss weights β, γ, ρ and the margin m, we performed a systematic one-dimensional sensitivity analysis on each of the five benchmarks, varying one parameter over a prescribed grid while holding the others fixed. The resulting AUC curves are presented in Figures 5–8, the selected optimal values are summarized in Table 7, and the scene-dependent trends are discussed in detail.

 

Comment  4: SG-HNS is central to the method, yet there is no evidence that its generated samples resemble meaningful hyperspectral anomalies rather than ordinary adversarial perturbations. Several generated spectra should be visualized and compared with normal and real anomalous spectra using spectral angle or spectral smoothness statistics. Comparisons with Gaussian noise, spectral mixing, and simple sparse perturbation baselines would make the contribution more convincing.

Response 4: We sincerely thank the Reviewer for this insightful comment, and we fully agree that direct spectral evidence would further strengthen the paper. We would like to respectfully clarify the design intent of SG-HNS, because it bears directly on how such evidence should be interpreted.

As now stated explicitly in Section 2.3, SG-HNS deliberately does not attempt to imitate realistic anomaly morphologies or spectra. Under the normal-only protocol, the nature of the anomalous threat is by definition unknown at training time; any attempt to make synthesized samples spectrally resemble “real anomalies” would implicitly reintroduce target priors—precisely what the framework is designed to avoid. The role of SG-HNS is instead to construct controlled out-of-distribution calibration samples near the normal manifold: the perturbation is spatially sparse (a small fraction p of pixels under the gating mask), bounded in magnitude (an ℓ∞ ball of radius ϵ), applied synchronously across all spectral bands so that spectral coherence is preserved, and optimized by a bounded number of PGD steps to maximize teacher–student divergence. The margin loss then requires only that the student respond to such deviations more strongly than to normal samples. The mechanism therefore relies on divergence-maximizing deviation from normality—not on spectral resemblance to real targets—and in this sense the synthesized samples are intentionally “structured yet subtle” perturbations rather than ordinary unconstrained adversarial noise.

We would also like to respectfully point out that the revised manuscript already contains indirect experimental evidence that the synthesized negatives carry useful discriminative signal: if they were uninformative noise, optimizing the hinge separation loss on them should not systematically improve the detection of real anomalies. Yet the cumulative ablation (Table 8) shows that enabling the SG-HNS separation term yields the largest single gains on PaviaC (+0.0032) and abu-beach-3 (+0.0009) and contributes to the improved low-false-alarm detection of 1–6-pixel targets reported in Tables 3 and 4.

We are nonetheless in full agreement with the Reviewer that the suggested spectral-angle/smoothness visualization and the comparison with Gaussian-noise, spectral-mixing, and sparse-perturbation baselines would constitute valuable additional validation. Given the scope and length of the present revision, we respectfully defer this systematic spectral characterization to future work—the Conclusion already identifies generative synthesis of hard negatives beyond gradient-based perturbation as an important future direction—and we have added an explicit clarification of the SG-HNS design objective in Section 2.3 so that readers can correctly interpret the role of the synthesized samples. We sincerely hope the Reviewer finds this clarification, together with the indirect evidence above, acceptable at this stage.

Comment  5: The current evaluation is too limited for a method aimed specifically at micro-scale targets. Overall AUC values are useful, but most results are already close to saturation and provide little information about performance at very low false-alarm rates. It would be better to report partial AUC or detection probability at fixed false-alarm rates, mean and standard deviation over repeated runs, and separate results for the 1–6 pixel targets in the Salinas-derived dataset. Cross-dataset testing is also needed to support the claimed generalization.

Response 5: We sincerely thank the Reviewer for this valuable suggestion, which has markedly improved the evaluation. Following the Reviewer’s advice, we have extended the “AUC-based quantitative evaluation” part of Section 3.2 with the following additions:

1. Detection probability at fixed false-alarm rates (new Table 3).We now report the detection probability P_d at P_fa ∈ {10⁻⁴, 10⁻³, 10⁻²} for all five benchmarks, together with a detailed analysis.
2. Per-size results for micro-scale targets (new Table 4).The Salinas-derived synthetic targets are stratified into six size classes (1×1 to 6×6 pixels, two targets per class), and we report per-class AUC and P_d at the three false-alarm levels.
3. Cross-dataset testing.We respectfully note that, as clarified in response to Comment  1, the scope of this work is the scene-specific normal-only (transductive) protocol; a strict cross-scene/cross-sensor evaluation—with fully separated training and testing scenes and no target-scene purification—is now explicitly stated as an important direction for future work in both Section 3.1 and the Conclusion. We completely agree with the Reviewer on its importance and have committed to it as the next step of this research line.

Comment  6: It may be useful to broaden the discussion of spectral guidance mechanisms. For example, the Spectral Objectness Prior Module provides an explicit way to exploit spectral targetness for feature enhancement, whereas HNGT-Net learns anomaly responses indirectly through normal-topology transfer and teacher–student disagreement. Discussing their different assumptions and applicable scenarios would better position the proposed method within recent hyperspectral target and anomaly detection research.

Response 6: We sincerely thank the Reviewer for this excellent suggestion, which has helped us position HNGT-Net much more clearly within the recent literature. Following the Reviewer’s advice, we have broadened the discussion in two places:

1. A new discussion now contrasts the two paradigms explicitly. The Spectral Objectness Prior Module (SOPM) in SPG-OD [43] exploits known spectral signatures to directly modulate feature responses toward target-like regions—a principled choice when the target spectrum is available a priori. HNGT-Net instead addresses the fundamentally different setting of complete target ignorance, where no spectral prior exists; it therefore learns anomaly discriminability indirectly, by constructing a compact manifold of normal spectral–spatial patterns through graph-structured topology transfer and flagging deviations from it. The discussion also clarifies how spectral information is nonetheless exploited within HNGT-Net—through feature-space nearest-neighbor graph construction, which encodes spectral affinity, and through graph-Laplacian response consistency, which regularizes high-frequency spectral variations—and frames the design as a deliberate trade-off between target-directed sensitivity and generalization to genuinely unknown anomalies.
2. Conclusion and Future Directions.We further emphasize that the two paradigms occupy complementary niches, and we identify hybrid architectures that combine the generalization benefits of topology transfer with adaptive spectral priors learned from limited target exposure as a promising future direction.

Reviewer 2 Report

Comments and Suggestions for Authors

This paper proposes HNGT-Net, a lightweight teacher-student framework for hyperspectral small-target anomaly detection under the normal-only training paradigm. The method addresses representation learning and discriminative learning jointly. On the representation side, a dual-neighborhood similarity graph fusing spatial 8-connectivity and feature-space K-nearest neighbors is constructed, and a three-level consistency objective over node embedding, edge structure, and graph-Laplacian response transfers the spectral-spatial topology from a frozen teacher to a compact student. On the discrimination side, a sparsity-gated hard-negative synthesis module combined with a response-margin separation loss forces the student to distinguish normal samples from synthesized hard negatives. Experiments on four public hyperspectral scenes and one Salinas-derived synthetic benchmark demonstrate a mean AUC of 0.9961 and a 10.5× student-side parameter compression.

The paper is well organized and presents clear engineering value. I recommend revising the following issues.

1. In Figure 1, the lower-left module is labeled "SLDM Pseudo-Anomaly Generator," whereas the text consistently uses "SG-HNS" (Sparsity-Gated Hard-Negative Synthesis). Please align the figure label with the text. The terms "micro-anomaly," "micro-scale," and "small-target" are used interchangeably throughout the paper, and I suggest adopting "small-target" consistently to match the paper title. In Algorithm 1, line 4, the symbol `\bar{F}_{T}(x)` should be corrected to `F_T(x)`, and line 6, where `A_T, A_S` are listed, lacks the explicit assignment reference from Eq. (5)–(7). A careful proofreading pass is needed to catch these inconsistencies.
2. Section 2.1 states that the teacher is "pre-trained and kept frozen," but the paper does not explain how this pre-training is performed. It remains unclear whether the teacher was pre-trained on ImageNet, through self-supervised learning on hyperspectral data, or simply initialized randomly and trained on normal samples. Since the teacher's quality directly determines the upper bound of the student's performance, this detail is essential for reproducibility. Please add a short paragraph specifying the pre-training data, the loss function used, and the number of training epochs. If the teacher was trained on the purified normal set with the node consistency loss, stating this explicitly would be perfectly acceptable and would strengthen the paper.
3. The Highlights and Abstract contain truncated sentences that need to be completed. In the first bullet of the Highlights, the phrase "yields an average AUC of" is left unfinished, and I suggest appending "0.9961 across five benchmarks." In the second bullet, the phrase "supporting resource-constrained remote sensing" is a sentence fragment, which should be completed, for example, as "supporting resource-constrained remote sensing platforms and edge deployment." The Abstract also ends abruptly, and the final sentence should be finished to provide a complete summary of the paper's findings and implications.
4. The runtime discrepancies in Table 4 should be explained. The table reports that CRD takes 1,989.67 seconds on abu-beach-3 but only 1,041.13 seconds on abu-urban-2, despite both scenes having the same spatial size of 100×100 pixels. This inconsistency suggests that window sizes or other implementation parameters may have varied across scenes. Please verify that all competing methods were executed under uniform settings on the same hardware, and briefly state the sliding-window size used for LRX and CRD so that readers can reproduce the timing results and assess the fairness of the comparison.

Author Response

Comment 1: In Figure 1, the lower-left module is labeled “SLDM Pseudo-Anomaly Generator,” whereas the text consistently uses “SG-HNS” (Sparsity-Gated Hard-Negative Synthesis). Please align the figure label with the text. The terms “micro-anomaly,” “micro-scale,” and “small-target” are used interchangeably throughout the paper, and I suggest adopting “small-target” consistently to match the paper title. In Algorithm 1, line 4, the symbol \bar{F}_{T}(x) should be corrected to F_T(x), and line 6, where A_T, A_S are listed, lacks the explicit assignment reference from Eq. (5)–(7). A careful proofreading pass is needed to catch these inconsistencies.

Response 1: We sincerely thank the Reviewer for this careful reading and for pointing out these inconsistencies. We fully agree that consistent notation and terminology are essential for readability. All of these issues have been corrected as follows:

1. Figure 1 label.We have completely redrawn Figure 1 from scratch and thoroughly checked and revised all figure captions to ensure full consistency with the terminology used throughout the text.
2. Following the Reviewer’s suggestion, we have unified the terminology throughout the manuscript and now consistently use “small-target”, in agreement with the paper title. All occurrences of “micro-anomaly” and “micro-scale” have been replaced, and the changes are highlighted in red.
3. Algorithm 1. We have corrected the symbol F̄_T(x) to F_T(x) in line 4. In line 6, we have added the explicit reference that the adjacency matrices A_T and A_S are constructed according to Eq. (5)--(7).

 

Comment 2: Section 2.1 states that the teacher is “pre-trained and kept frozen,” but the paper does not explain how this pre-training is performed. It remains unclear whether the teacher was pre-trained on ImageNet, through self-supervised learning on hyperspectral data, or simply initialized randomly and trained on normal samples. Since the teacher’s quality directly determines the upper bound of the student’s performance, this detail is essential for reproducibility. Please add a short paragraph specifying the pre-training data, the loss function used, and the number of training epochs. If the teacher was trained on the purified normal set with the node consistency loss, stating this explicitly would be perfectly acceptable and would strengthen the paper.

Response 2 : We sincerely thank the Reviewer for this insightful comment, and we apologize for the misleading wording in the original manuscript, which caused an understandable ambiguity. We would like to clarify that the teacher network is not pre-trained at all: it is randomly initialized once and then kept frozen throughout the entire training process. There is therefore no pre-training dataset, no pre-training loss, and no pre-training epochs involved. The phrase “pre-trained and kept frozen” in the original text was inaccurate, and we are grateful to the Reviewer for catching it.

We have added an explicit clarification paragraph in Section 2 (Proposed HNGT-Net Framework), stating that “the teacher network T(·) is randomly initialized once and remains frozen throughout training,” and explaining the above rationale. Algorithm 1 (line 1) now also explicitly states “Randomly initialize teacher T(·) and freeze all of its parameters.” All changes are highlighted in red. We hope this clarification fully addresses the Reviewer’s concern and improves the reproducibility of our work.

 

Comment 3: The Highlights and Abstract contain truncated sentences that need to be completed. In the first bullet of the Highlights, the phrase “yields an average AUC of” is left unfinished, and I suggest appending “0.9961 across five benchmarks.” In the second bullet, the phrase “supporting resource-constrained remote sensing” is a sentence fragment, which should be completed, for example, as “supporting resource-constrained remote sensing platforms and edge deployment.” The Abstract also ends abruptly, and the final sentence should be finished to provide a complete summary of the paper’s findings and implications.

Response 3: Thank you for catching these oversights. We have completed all incomplete sentences in the Highlights and Abstract:

Highlights bullet 1. Completed to "yields an average AUC of 0.9953 across five benchmarks."

Highlights bullet 2. Completed to "supporting resource-constrained remote sensing platforms and edge deployment."

Abstract. The final sentence has been rewritten to provide a complete summary: "Comprehensive evaluation on four public benchmark scenes and one Salinas-derived synthetic dataset shows that HNGT-Net attains a mean AUC of 0.9953 together with the smallest cross-dataset variance among all compared methods, indicating both high accuracy and stable generalization across diverse sensing scenes."

 

Comment 4: The runtime discrepancies in Table 4 should be explained. The table reports that CRD takes 1,989.67 seconds on abu-beach-3 but only 1,041.13 seconds on abu-urban-2, despite both scenes having the same spatial size of 100×100 pixels. This inconsistency suggests that window sizes or other implementation parameters may have varied across scenes. Please verify that all competing methods were executed under uniform settings on the same hardware, and briefly state the sliding-window size used for LRX and CRD so that readers can reproduce the timing results and assess the fairness of the comparison.

Response 4: Thank you for raising this important point. We confirm that all methods were executed on the same hardware configuration (the Wuhan node of the National Supercomputing Internet, with identical container allocation: seven cores of a Hygon C86-7380 processor, 60 GB DDR4 memory, and one Hygon DCU accelerator). The runtime variation for CRD is attributable to the dual-window size configuration, which was tuned per scene to optimize detection accuracy:

LRX: 7×11 (abu-urban-2), 7×13 (abu-urban-4), 23×27 (abu-beach-3 and PaviaC), 5×27 (Salinas simulate).

CRD: 7×11 (abu-urban-2), 7×13 (abu-urban-4), 9×13 (abu-beach-3), 3×5 (PaviaC), 3×15 (Salinas simulate).

CRD's runtime on abu-beach-3 is higher because its optimal dual-window configuration (9×13) is larger than that on abu-urban-2 (7×11), leading to a proportionally higher computational cost per pixel. We have added these dual-window sizes to the "Measured runtime" paragraph in the revised manuscript (highlighted in red) to ensure full reproducibility.

 

Reviewer 3 Report

Comments and Suggestions for Authors

Paper proposes the Hard-Negative-Guided Topology Transfer Network (HNGT-Net), a novel teacher–student topology transfer framework devised for HS micro-anomaly detection when training is restricted to normal samples.

A similarity graph is built by fusing spatial 8-neighborhood adjacency with feature-space KNN relations, providing a compact description of the topological organization and intrinsic regularities of normal spectral–spatial backgrounds; Node-feature alignment, pairwise structure alignment, and graph-Laplacian response alignment are jointly imposed so that the compact student can absorb normal-feature knowledge from the frozen teacher and become more responsive to localized structural disruptions; The SG-HNS module applies sparsity-gated projected gradient perturbations to purified normal pixels, producing micro-scale hard negatives without relying on anomaly spectra, target templates, or morphological priors; Hierarchical consistency transfer;  A normal–negative separation constraint expands the response gap between purified background samples and synthesized hard negatives, alleviating normal-distribution overfitting and improving micro-anomaly discrimination under normal-only training.

Comments:

• In opinion of this expert, the presented explications of all blocks of fig.1 and algorithm 1 that have been exposed in sect. 2. HNGT-Net Framework did not permit a potential reader to understand how this technique works. In fig. 1, there were seven steps (the authors forgot to put step 4), but the authors never explained all these steps into the text. This reviewer thinks that the authors should redact a text and connect it with explications of all the steps presented in fig.1.
• Authors wrote (lines: 358-360): “SG-HNS adopts a weighted aggregation of the three distillation objectives as the divergence functional to be maximized”, presenting eqs. 15, 16, where several parameters lambda are introduced. This reviewer thinks that the authors should provide theoretical or experimental justification on how they select values of these parameters that can significantly change output results. It is important for a better understanding by a potential reader
• This reviewer presents similar comments relating to eq. 18 where several parameters (betta, gamma, ro) are used. Please explain how you chose the values of these parameters that can change quality results of designed framework.
• In subsect. 1. Benchmark Datasets and Training Data Preparation, the authors exposed some short information (lines: 488-503). This reviewer thinks that information presented in subsect. Normal-only training data preparation cannot permit a potential reader to understand mentioned process. Please present in your manuscript more details about training process.
• This reviewer thinks that authors should discuss the visual perception details presented in fig. 3, mentioning visual anomalies in perception comparison between a method image against GT. The discussions presented in text (lines: 559-593) maybe could use some zoomed parts of the images, explaining better visual perception on an image detail. Without such detailed explications, a potential reader cannot understand better visual performance of your framework. Please change name of your system for HNGT-Net (not proposed) in sub images.

Author Response

We are sincerely grateful to Reviewer   for the rigorous and constructive evaluation. The comments on the framework illustration, the hyperparameter justification, the training-data preparation, and the visual analysis have prompted us to substantially strengthen the manuscript, and we believe the revised version is now considerably clearer and more complete.

Comment 1: In opinion of this expert, the presented explications of all blocks of fig.1 and algorithm 1 that have been exposed in sect. 2. HNGT-Net Framework did not permit a potential reader to understand how this technique works. In fig. 1, there were seven steps (the authors forgot to put step 4), but the authors never explained all these steps into the text. This reviewer thinks that the authors should redact a text and connect it with explications of all the steps presented in fig.1.

Response1： We deeply appreciate this fundamental critique and have thoroughly restructured the explanatory text in Section2 (Proposed HNGT-Net Framework) to achieve a one-to-one correspondence with the seven steps depicted in Figure 1 and Algorithm 1.

First, we have completely redrawn Figure 1 to ensure that all seven steps are explicitly labeled and visually distinct, including the previously overlooked Step 4 (Sparsity-Gated Hard-Negative Synthesis). Each step is now clearly demarcated with numbered labels that match the subsection headings in the text. 

Second, we have reorganized Section 2 into seven explicitly titled subsections—Step 1 through Step 7—each of which opens with a concise descriptive paragraph (highlighted in red) that states the purpose of the step, its inputs from the preceding step, and its outputs for the subsequent step.

Third, we have revised Algorithm 1 to ensure that its line ordering exactly mirrors the execution flow depicted in Figure 1 and described in the seven subsections.

Comment 2: Authors wrote (lines: 358-360): “SG-HNS adopts a weighted aggregation of the three distillation objectives as the divergence functional to be maximized”, presenting eqs. 15, 16, where several parameters lambda are introduced. This reviewer thinks that the authors should provide theoretical or experimental justification on how they select values of these parameters that can significantly change output results. It is important for a better understanding by a potential reader.

Response 2: We sincerely thank the Reviewer for this important comment. We first owe the Reviewer an apology for a notation inconsistency that created the impression of additional independent parameters: the λ weights appearing in the original Eq. (15) were in fact intended to be exactly the same coefficients β and γ used in the joint training objective of Eq. (18), with the node-level coefficient fixed to one. In the revised manuscript, Eq. (15) has been rewritten as D(x) = L_node(x) + β·L_struct(x) + γ·L_lap(x), and the text now explicitly states that β and γ are shared with the corresponding terms in Eq. (18). Consequently, the divergence functional introduces no additional independent weighting parameters.

This design has a deliberate theoretical motivation: since SG-HNS synthesizes hard negatives by maximizing the same divergence that the student is trained to minimize on normal samples and to separate through the margin loss, sharing the coefficients guarantees that the synthesized negatives lie precisely along the directions to which the student is being trained to respond, rather than along an arbitrary direction defined by a separate set of weights.

Regarding the selection of the parameter values,  we have added a new Section 3.4 “Hyperparameter Sensitivity and Selection”, which provides a systematic experimental justification: for each of the key hyperparameters (the structural consistency weight β, the Laplacian response weight γ, the hard-negative separation weight ρ, and the response margin m), we conducted a one-dimensional sensitivity analysis on each of the five benchmark datasets, varying one parameter over a prescribed grid while holding all others fixed at their reference values. The resulting AUC curves are presented in Figures 5–8, the selected optimal values are summarized in Table 7, and the scene-dependent trends are discussed in detail.

Comment 3: This reviewer presents similar comments relating to eq. 18 where several parameters (betta, gamma, ro) are used. Please explain how you chose the values of these parameters that can change quality results of designed framework.

Response 3: We sincerely thank the Reviewer for raising this point, which is indeed closely related to Comment 2.2. The values of β, γ, and ρ in Eq. (18) were not chosen arbitrarily: they were selected through the systematic sensitivity analysis described in the newly added Section 3.4. Specifically, for each benchmark dataset we varied each coefficient over a prescribed grid while keeping the remaining hyperparameters fixed at their reference values, and we selected the value that maximized the detection AUC. The complete AUC-versus-parameter curves are shown in Figures 5–8, and the selected per-dataset optima are listed in Table 7.

Comment 4: In subsect. 1. Benchmark Datasets and Training Data Preparation, the authors exposed some short information (lines: 488-503). This reviewer thinks that information presented in subsect. Normal-only training data preparation cannot permit a potential reader to understand mentioned process. Please present in your manuscript more details about training process.

Response 4: We sincerely thank the Reviewer for this constructive suggestion, and we agree that the original description was too brief to support reproducibility. The Section 3.1 (“Normal-only training data preparation”) expanded with the complete mathematical formulation (Eqs. (22)–(25)) and the full implementation-level training details described above. All changes are highlighted in red. We hope the expanded description now allows a reader to fully understand and reproduce the training process.

Comment 5: This reviewer thinks that authors should discuss the visual perception details presented in fig. 3, mentioning visual anomalies in perception comparison between a method image against GT. The discussions presented in text (lines: 559-593) maybe could use some zoomed parts of the images, explaining better visual perception on an image detail. Without such detailed explications, a potential reader cannot understand better visual performance of your framework. Please change name of your system for HNGT-Net (not proposed) in sub images.

Response 5: We sincerely thank the Reviewer for this helpful suggestion, which has enabled us to present the visual evidence much more convincingly. The following revisions have been made:

1. New zoomed comparison figure.We have added a new Figure 4, which provides enlarged region-of-interest (ROI) comparisons between the ground-truth (GT) maps and the HNGT-Net detection results for all five datasets. Each row corresponds to one benchmark, and the three columns show, respectively, the complete HNGT-Net detection map with the selected ROI marked by a red rectangle, the enlarged GT region, and the corresponding enlarged HNGT-Net response.
2. Detailed pixel-level visual analysis.A new paragraph, “Detailed visual analysis of the selected regions of interest,” has been added to accompany Figure 4. It examines target localization, boundary preservation, response continuity, and background interference at the pixel level for every scene.
3. Sub-image labels corrected.Following the Reviewer’s request, the system name in all sub-image annotations has been unified to “HNGT-Net” instead of “Proposed”.

 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I have no further questions and recommend accepting this manuscript.

Reviewer 3 Report

Comments and Suggestions for Authors

No comments.