{"id":"ee6be76f-7449-4b60-8fcb-39ec2500d3cc","arxiv_id":"2504.19279","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper reports a wavelet-gradient band selection method for hyperspectral classification, but the experimental presentation is internally inconsistent and lacks baseline comparisons.","lead":"This paper proposes a wavelet-based band selection and undersampling strategy for hyperspectral image classification, claiming state-of-the-art accuracy on two benchmarks. The manuscript's internal inconsistencies, including swapped dataset class names and contradictory accuracy figures, make the central claim unverifiable.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified","rationale":"placeholder","tokens_in":10935,"tokens_out":3536,"duration_ms":37403,"concrete_test":"placeholder","verdict_should_be":"UNCHANGED","load_bearing_attack":"placeholder","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 'Iterative Wavelet-based Gradient Sampling' (IWGS, also spelled IGWS), a greedy band-selection algorithm that selects hyperspectral bands by descending a classifier loss in the wavelet-transformed domain, and claims that it consistently outperforms state-of-the-art band selection and classification methods in accuracy and computational efficiency. The experimental section presents a patch-size ablation, reporting overall accuracy of up to 97.8% on Indian Pines in the abstract, and includes an adversarial robustness experiment under PGD perturbations with atmospheric noise.","tokens_in":10858,"tokens_out":5133,"duration_ms":48984,"significance":"If the claimed results were reproducible, the method would be of genuine practical interest for onboard or edge hyperspectral classification, since it directly targets the memory and throughput bottlenecks of deep HSI classifiers. The paper makes falsifiable efficiency and accuracy claims, which is a strength in principle. However, the manuscript does not specify the algorithm's central gradient step in a computable way, compares against no external state-of-the-art baseline, and contains numerical and dataset contradictions that prevent verification. The paper ships no code and no machine-checked derivations, so these strengths are not realized in the current form.","major_comments":[{"comment":"The gradient ∂L/∂w_j used to select bands is not defined, because the selection vector w is binary and the reconstruction X̂_w is not differentiable with respect to an element of {0,1}^B; the paper states no relaxation, straight-through estimator, or surrogate gradient. Since every selected band and every reported accuracy depends on this step, the method cannot be reproduced or evaluated as written.","section":"Section 3.1, Algorithm 1, step 7"},{"comment":"The central comparison claim is unsupported: the experiments compare the proposed sampling only with 'random undersampling and no sampling' and never report a single run of any state-of-the-art band selection or classification method from the related-work list. The abstract's claim that IWGS 'consistently outperforms state-of-the-art band selection and classification techniques' therefore has no evidential basis in this manuscript.","section":"Sections 4.1–4.2 and Table 3"},{"comment":"The reported numbers are internally inconsistent: the text states P5 OA = 97.60% while Table 3 gives 97.10%; the text states P13 OA = 97.86% while Table 3 gives 97.36%; and the abstract's headline 97.8% appears in no row. Table 3's class names correspond to Houston 2013, while Table 1, Figure 1, and most of the experiments concern Indian Pines; the conclusion attributes 97.60% to Houston 2013. These contradictions make the experimental results unverifiable.","section":"Table 3 and Section 4.2"},{"comment":"The word 'optimal' in the title and the 'optimal configuration' language in Section 4.2 are not justified: the number of selected bands N_s, the wavelet transform W, the patch size, and the PGD parameters (ε and α) are selected from experimental results rather than derived or optimized with a guarantee. No optimality proof or parameter-free characterization is given.","section":"Title and Section 3.1"},{"comment":"The manuscript describes two incompatible methods under nearly the same name: the contribution bullet introduces 'IGWS' as an undersampling strategy that preserves minority-class samples, whereas the abstract and Section 3.1 define 'IWGS' as a band-selection algorithm; the experiments evaluate neither against an external baseline. It is therefore unclear what exactly was implemented and measured.","section":"Section 1 contribution list vs. Section 3.1"}],"minor_comments":[{"comment":"The dataset description is inconsistent: Section 4 says Indian Pines has 220 bands with 20 removed, while the Data Availability section says 224 bands with 200 retained; Table 1's total of 10,366 labeled samples also does not match the standard Indian Pines count, and no explanation is provided.","section":"Data Availability and Section 4"},{"comment":"The notation switches between IWGS and IGWS inconsistently (for example, the contribution list and Figure 1 caption use IGWS, while Algorithm 1 and the abstract use IWGS); please standardize the name and spelling.","section":"Throughout"},{"comment":"The Kappa coefficients in Table 2 are not tied to a specific dataset or to the PGD attack parameters used; the caption and text should state the exact attack settings, the dataset, and the classifier so that the robustness claim is checkable.","section":"Table 2"},{"comment":"The first sentence of Section 3.2 contains the typo 'HSP' where 'HSI' is intended; please correct.","section":"Section 3.2"}],"recommendation":"reject","confidential_remarks":"The numerical contradictions and undefined algorithm step are so extensive that a simple revision cannot fix them; the results would need to be regenerated from a clearly specified algorithm and compared with actual baselines. The authors may also want to clarify whether this is a single coherent paper, since the IWGS band-selection part and the IGWS class-undersampling part appear to be from different projects. There is no code repository or supplementary material that would allow reviewers to check the experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You asked for a read on arXiv:2504.19279. My take aligns with the desk report, and the stress-test placeholder doesn't change that. This paper is not close to publishable, and I don't think a serious referee should spend time on it in its current form.\n\nThe core idea is reasonable: select hyperspectral bands by looking at gradients in a wavelet-transformed domain, doing an iterative greedy search, and wrap it around an efficient CNN/state-space classifier. The components are established, and the self-citation to the MRI undersampling work [25] is legitimate as an inspiration. That's the new part, and it's plausible on its face.\n\nBut the paper undermines itself at every load-bearing point. The method is called IWGS in the abstract and Algorithm 1, IGWS in the contributions and most of the experiments, and Section 3.2 describes IGWS as a sample undersampling strategy while Section 3.1 defines IWGS as band selection. Those are not the same thing. Table 3 uses Houston 2013 class names (Healthy grass, Stressed grass, Residential, etc.) while the surrounding text says Indian Pines throughout; the abstract claims 97.8% on Indian Pines, but no OA row in Table 3 shows 97.8% — the closest are 97.99 at P11 and 97.90 Kappa at P5. The text reports P5 OA as 97.60% and P13 OA as 97.86%, but Table 3 shows 97.10% and 97.36% for those same patch sizes. The conclusion claims 97.60% on Houston, which no table supports.\n\nThe algorithm itself is underspecified: it computes a gradient of the loss with respect to a binary selection vector w, but binary masks are not differentiable and no relaxation, straight-through estimator, or surrogate is given. Every reported number depends on that unspecified implementation detail. There are no baseline comparisons, no error bars, no code, no ablation of the band-selection mechanism (the experiments are about patch size). These are not minor polish issues; they mean a reader cannot tell what was actually computed or what was actually achieved.\n\nThe paper does show some awareness of its limits — the abstract says \"for selected classes,\" and the patch-size sensitivity is at least explored. But the disclaimers don't fix the dataset contradictions or the missing gradient definition.\n\nBottom line: no one should cite this yet, and I'd desk-reject it. If the authors fix the nomenclature, align text and tables, specify the gradient approximation, add real baselines, and release code, there might be a useful small paper here about wavelet-domain gradient band selection. In its current form, it's a draft, not a submission.","headline":"Internally inconsistent to the point of being unassessable: the method name, dataset, and headline numbers all disagree, and no baseline or code is provided.","tokens_in":11457,"tokens_out":1617,"would_cite":false,"duration_ms":18985,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"IWGS, an iterative wavelet-domain gradient sampler, selects spectral bands that keep hyperspectral classification accuracy high (up to 97.8% on Indian Pines) while reducing computational cost.","keywords":["hyperspectral image classification","band selection","wavelet transform","iterative gradient sampling","undersampling","satellite imaging","edge computing","Indian Pines"],"falsifier":"Run the published Algorithm 1 on Indian Pines with a fixed classifier and a stated gradient estimator, and compare the resulting accuracy and selected bands against random band selection with the same classifier. If random bands match or beat IWGS, or if the reported 97.8% accuracy cannot be reproduced, the central claim is refuted.","tokens_in":10678,"feed_emoji":"🛰️","tokens_out":7686,"duration_ms":72256,"temperature":0.7,"pith_summary":"The paper introduces IWGS, an iterative band-selection algorithm that picks spectral bands one at a time by measuring the gradient of the classifier's loss in a wavelet-transformed domain. It claims this produces state-of-the-art hyperspectral image classification accuracy while cutting computational cost, reporting up to 97.8% overall accuracy on Indian Pines and 97.6% on Houston 2013. The motivation is practical: satellites and edge devices cannot process hundreds of bands, and a task-specific subset keeps accuracy high with less memory and latency. The paper frames IWGS as an undersampling strategy that preserves informative spectral structure and scales to resource-constrained deployment.","feed_headline":"Wavelet-gradient band picker hits 97.8% on hyperspectral images","feed_subtitle":"Selecting bands by loss gradients in a wavelet domain keeps accuracy high while cutting compute for satellite imaging.","key_machinery":"The mechanism is the gradient of the classification loss with respect to a binary band-selection vector $w \\in \\{0,1\\}^B$, evaluated through a wavelet transform and its inverse. At each iteration the method computes the wavelet representation $X_W = \\mathcal{W}(X)$, reconstructs the masked cube $\\hat{X}_w = \\mathcal{W}^{-1}(X_W \\cdot w)$, obtains predictions from the classifier $S$, and selects the band with the largest $\\left|\\partial L/\\partial w_j\\right|$. The wavelet transform, a multi-resolution decomposition that separates spectral detail at different scales, is doing the work of exposing localized, sparse structure; the gradient is doing the work of making the choice task-specific rather than purely statistical.","core_discovery":"The paper's central claim is that a greedy, loss-driven band-selection rule operating in the wavelet domain outperforms existing band selection and classification methods on standard hyperspectral benchmarks. IWGS starts with the central band set to one, then, for each of $N_s$ steps, transforms the hyperspectral cube into the wavelet domain, masks it with the current binary selection vector, reconstructs the input, classifies, computes the loss, and takes the gradient of the loss with respect to the selection vector. The next band chosen is the unselected one with the largest absolute gradient component, so each addition is the band that most reduces the current classification loss. The paper reports that this strategy maintains or improves accuracy while greatly reducing the number of bands, and that the resulting pipeline survives PGD adversarial perturbations combined with atmospheric noise, with patch size 5 identified as the best accuracy/efficiency trade-off.","pith_inferences":["The paper does not specify the classifier architecture, the number of selected bands, or the gradient estimator for the binary mask; if the reported numbers only hold for one hidden configuration, the method's generality is untested.","If the gradient rule is the true driver, then swapping the wavelet transform for another sparsifying transform, such as a discrete cosine transform or learned bases, should preserve much of the benefit; a comparison would isolate what wavelets contribute.","The iterative loss-driven selection is a general recipe that could transfer to other high-dimensional sensing tasks, including multispectral imaging, MRI undersampling, and audio spectrograms, wherever a differentiable loss over a binary channel mask can be defined.","The paper's undersampling framing suggests a testable extension: measure how the selected band subset changes as the training set shrinks and whether the accuracy advantage over random band selection grows exactly in the low-label regime it targets."],"forward_implications":["If IWGS works as reported, hyperspectral classifiers can run on a small, task-specific subset of bands, cutting memory and latency enough for onboard satellite processing.","Because selection is driven by the classifier's own loss, the chosen bands are coupled to the downstream model; a different classifier would be expected to select a different subset.","The reported patch-size analysis gives a concrete deployment rule: patch size 5 preserves most of the accuracy (97.10% OA) while reducing compute relative to larger patches.","The perturbation results imply that band selection by wavelet-domain gradients retains enough spectral structure to survive PGD attacks under additive atmospheric noise."],"supporting_citations":[{"why":"Supplies the iterative undersampling pattern, originally for MRI, that IWGS adapts to spectral band selection.","marker":"[25]"},{"why":"Introduces the efficient state-space model whose design principles motivate the hybrid classifier used with IWGS.","marker":"[23]"},{"why":"Defines the PGD attack used to test IWGS under adversarial perturbations with atmospheric noise.","marker":"[45]"},{"why":"Describes the AVIRIS sensor and provides the Indian Pines dataset on which the 97.8% accuracy claim is reported.","marker":"[46]"},{"why":"Reviews hyperspectral band selection and frames dimensionality reduction as the problem IWGS targets.","marker":"[18]"}],"fun_headline_variants":["Wavelet gradient sampling picks optimal bands for satellite HSIs","Iterative wavelet-gradient band selection tops 97.8% on HSI","Optimal undersampling: wavelet-gradient band picker for HSI","Wavelet gradients pick smartest bands for satellite imaging","Wavelet-gradient band picker cuts compute, keeps HSI accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the classifier's loss can be differentiated with respect to a yes/no band-selection mask, even though such a mask has no ordinary derivative and the paper never says how that derivative is approximated.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet gradient sampling picks optimal bands for satellite HSIs","Iterative wavelet-gradient band selection tops 97.8% on HSI","Optimal undersampling: wavelet-gradient band picker for HSI","Wavelet gradients pick smartest bands for satellite imaging","Wavelet-gradient band picker cuts compute, keeps HSI accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000985,"raw_usage":{"total_tokens":4177,"prompt_tokens":944,"completion_tokens":3233,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":3145}},"tokens_in":560,"tokens_out":3233,"duration_ms":20693,"temperature":1.0,"reasoning_tokens":3145,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:56:10.576879+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the published Algorithm 1 on Indian Pines with a fixed classifier and a stated gradient estimator, and compare the resulting accuracy and selected bands against random band selection with the same classifier. If random bands match or beat IWGS, or if the reported 97.8% accuracy cannot be reproduced, the central claim is refuted.","supporting_citations":[{"cited_title":"Optimal mri undersampling patterns for ultimate benefit of medical vision tasks,","cited_arxiv_id":null,"evidence_quote":"Supplies the iterative undersampling pattern, originally for MRI, that IWGS adapts to spectral band selection."},{"cited_title":"Hyperspectral band selection: A review,","cited_arxiv_id":null,"evidence_quote":"Reviews hyperspectral band selection and frames dimensionality reduction as the problem IWGS targets."}],"review_version":1}