{"id":"a93ee353-7b72-44d5-b250-5130ceac05b2","arxiv_id":"2507.20099","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"HDST, built on SERT, adds FFT-based frequency-domain fusion and atrous multiscale modules to improve hyperspectral image denoising, achieving up to 0.94 dB PSNR over its baseline.","lead":"This paper presents HDST, a hyperspectral image denoising network that adds frequency-domain processing and multiscale convolution to the SERT transformer baseline. On real and synthetic HSI datasets it reports modest PSNR gains (up to 0.94 dB) over SERT, with mixed SSIM results.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-run evaluation: the 0.94 dB gain over SERT and the 0.05 dB gain from the frequency module alone are reported without seeds or significance tests, while SSIM regresses on 3 of 5 ICVL noise types.","rationale":"The paper is a plausible incremental extension of SERT, and the public code is a real asset; I am not asserting the gains are fabricated. However, the claim of 'significant' improvement is only as strong as the stability of the comparison, and the manuscript provides no seed variance, error bars, or paired tests. The ablation table makes this more acute because the frequency-domain component that defines the contribution moves PSNR by only 0.05 dB. The self-reported 'work-in-progress' statements and unfinished figure captions further support a conditional, not unconditional, reading. The reader's weakest assumption is exactly the right one; I agree with the CONDITIONAL verdict and would keep it until multi-seed results are supplied.","tokens_in":12741,"tokens_out":5355,"duration_ms":55116,"concrete_test":"Using the released code, retrain SERT, HDST, and Net1 with at least 5 independent random seeds under identical splits and hyperparameters, and report mean ± std PSNR/SSIM/SAM on Realistic and on each ICVL noise type. Compute a paired bootstrap or Wilcoxon signed-rank test across test images for (i) HDST vs SERT and (ii) Net1 vs SERT. The claim stands if the 95% confidence interval for the Realistic PSNR difference excludes 0 and the mean ICVL gains exceed roughly 2× the seed standard deviation; otherwise the reported gains should be treated as within-run variance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim ('HDST significantly improves HSI denoising over SERT') rests entirely on Tables I-III, which report one training run per configuration and no variance or paired statistics. On Realistic the gain is 0.94 dB; on ICVL the gains are 0.15-0.52 dB across noise types. These gaps are the size at which seed-to-seed variation in hyperspectral denoising typically matters, so the claimed superiority may not be robust. The ablation sharpens the risk: Net1, which isolates the paper's novel frequency-domain processing (FSGF+FSCA), is only 0.05 dB above baseline (29.73 vs 29.68), a difference that is plausibly pure noise. Additionally, HDST's SSIM is lower than SERT's on Gaussian+Stripe (0.9963 vs 0.9969), Non-i.i.d Gaussian (0.9970 vs 0.9971), and Gaussian+Deadline (0.9966 vs 0.9969), so the 'significant improvement' is not consistent across metrics. If those small differences are within run-to-run variance, the paper's central methodological claim is unsupported. This is exactly the condition that would have to be true for the claim to hold: the reported deltas must be reproducible and distinguishable from training noise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HDST, a hyperspectral image denoising network built on the SERT baseline. The method adds an FFT-based preprocessing module with multi-band ASPP convolution (ASPP-FFT), a gated frequency-spatial fusion module (FSGF), a frequency-spatial collaborative attention module (FSCA), and a hybrid-domain multi-scale module (HDMS). The authors evaluate on the Realistic dataset and the ICVL synthetic dataset, reporting PSNR/SSIM/SAM comparisons against traditional and deep-learning baselines, an ablation study, and a computational efficiency analysis. The central claim is that HDST significantly improves over SERT, with +0.94 dB PSNR on Realistic and up to +0.52 dB on ICVL, while increasing FLOPs by only 14.7%.","tokens_in":1417,"tokens_out":1863,"duration_ms":53424,"significance":"If the reported gains are reproducible, the architecture is a plausible incremental extension of SERT with a sensible frequency-domain branch, and the released code would be a useful resource for the HSI denoising community. The paper also gives a rare explicit attempt to connect multiscale atrous convolution with frequency-band processing. However, the significance of the empirical claim is currently limited by the lack of multi-seed evaluation, by SSIM regressions on several noise types, and by the authors' own statements that figures and ablations are unfinished. The contribution is therefore better viewed as a promising system study than as a settled methodological result.","major_comments":[{"comment":"All comparisons rest on a single training run per configuration. The headline deltas are small: 0.94 dB on Realistic, 0.15-0.52 dB on ICVL, and only 0.05 dB for the frequency module alone in the ablation (Net1 vs Baseline, Table III). At these magnitudes, run-to-run variance is a plausible alternative explanation for the reported improvement. The authors should report mean and standard deviation over at least three to five seeds and, ideally, paired statistical tests over the same test images.","section":"Section IV-A, Table I; Section IV-B, Table II; Section IV-C, Table III"},{"comment":"HDST's SSIM is lower than SERT's on Gaussian+Deadline (0.9966 vs 0.9969), Gaussian+Stripe (0.9963 vs 0.9969), and Non-i.i.d Gaussian (0.9970 vs 0.9971). The claim that HDST 'significantly outperforms' the baseline is therefore not consistent across metrics. This should be discussed explicitly, and the authors should determine whether the PSNR gains come at the cost of structural fidelity in selected cases.","section":"Table II"},{"comment":"The captions state 'We are working on this figure, since we want to provide a more level of detail comparison for the visual comparison,' and the acknowledgments state 'This preprint reports work-in-progress. Full ablation studies and refined technical descriptions will be updated in subsequent versions.' These statements are incompatible with a journal submission. The visual evidence must be finalized and the work-in-progress language removed before the manuscript can be evaluated as a complete research article.","section":"Figures 3-4 and Acknowledgments"},{"comment":"The physical-interpretation claim that ASPP dilation rates [2,4,8] correspond to multi-band frequency processing is asserted but not demonstrated. The authors should provide evidence, for example, by visualizing the learned gating masks or frequency responses and showing that they align with the stated noise priors (e.g., stripe-noise frequency lines vs high-frequency sensor noise). Without such evidence, the 'frequency domain noise prior' rationale remains post hoc.","section":"Section III.B.2 and Equations (1)-(5)"}],"minor_comments":[{"comment":"The hyperparameter alpha controlling frequency-domain injection intensity is never given a value or sensitivity analysis; please specify its value and any effect on performance.","section":"Section III.B.2, Eq. (5)"},{"comment":"The FSCA window size M is not specified in the paper; please report the value used and justify it.","section":"Section III.B.3, Eq. (6)"},{"comment":"The column header 'ASPP/HDMS' is ambiguous for Net3 and Net4; clearly define which modules (ASPP-FFT, HDMS, dynamic fusion) are active in each ablation row.","section":"Table III"},{"comment":"Reference [16] is described as a cross-domain filtering denoising method (FCENet), but the cited paper title is about arbitrary-shaped text detection; either the intended reference is different or the citation is incorrect.","section":"Related Work, reference [16]"},{"comment":"The phrase 'slight increase in parameters such as PSNR' should be reworded to 'slight increase in metrics such as PSNR' for clarity.","section":"Section IV.C"},{"comment":"There are formatting glitches such as 'T ransformer' in the title line and a duplicated sentence in the abstract; please proofread the final version.","section":"Title and Abstract"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be an early preprint: the figure captions explicitly say the visual comparisons are unfinished, and the acknowledgments state that full ablations and technical descriptions will be updated in later versions. My recommendation assumes the authors are willing to add multi-seed evaluation, finalize all figures, and remove the work-in-progress statements. The reported improvement over the authors' own SERT baseline is modest and partially offset by SSIM regressions, so the novelty claim should be appropriately tempered in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this as a clearly labeled work-in-progress, not a finished result. The method is a sensible incremental extension of SERT: FFT preprocessing with multi-band convolution, an ASPP-FFT module, gated fusion, and a multiscale module. The authors are honest about the state of the paper—the acknowledgements say so directly, and the figure captions say the visual comparisons are unfinished. That matches the evidence.\n\nWhat is genuinely good: the ablation is informative. Net3 (multiscale HDMS alone) gives a 0.59 dB gain, Net4 (frequency + multiscale) gives 0.78 dB, and full HDST gives 0.94 dB on the Realistic set. That ordering is coherent and suggests the multiscale branch is doing most of the work, with the frequency processing contributing a smaller positive increment. Code is public, datasets are standard, and the related-work coverage of frequency-domain denoising is adequate. No citation red flags.\n\nThe soft spots are proportional: the headline claim is not statistically supported. All tables show one run per configuration. On ICVL the gains are 0.15–0.52 dB; on the Realistic set, 0.94 dB. Those are exactly the magnitudes where seed-to-seed variation matters in HSI denoising. The ablation sharpens the concern: Net1, the pure frequency module, is only 0.05 dB above baseline—easily noise. Also, SSIM regresses against SERT on three of the five ICVL noise types (Gaussian+Stripe, Non-i.i.d Gaussian, Gaussian+Deadline). The paper's discussion of noise frequency characteristics is fine as a post hoc hypothesis, but without repeated runs or paired tests it is just a story.\n\nBottom line: the architecture is plausible, the code is available, and the ablation ordering is coherent. But the paper cannot support \"significantly improves\" until there are multiple seeds with means and variances, or paired statistical tests, and the figures are finished. A serious referee could engage with this—it is a legitimate incremental technique that deserves the time. The current version is not accept-ready, but it is a solid starting point for a revision.","headline":"Incremental but honest work-in-progress: the architecture and ablation order are plausible, but the headline gains rest on single runs and the paper itself says it is unfinished.","tokens_in":13558,"tokens_out":2570,"would_cite":false,"duration_ms":25931,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hyperspectral denoising network that routes features through an FFT-based gating unit and multiscale pyramids reports 0.94 dB PSNR over its spatial-spectral transformer baseline on real noisy images.","keywords":["hyperspectral image denoising","frequency domain","transformer","atrous spatial pyramid pooling","cross-domain attention","gating mechanism","multiscale convolution","spatial-spectral modeling"],"falsifier":"Train SERT and HDST on the same Realistic split with at least five random seeds and compute the 95% confidence interval of the per-image PSNR difference; if the interval contains zero, the claimed 0.94 dB improvement is not statistically established. As a second check, run HDST on pure i.i.d. Gaussian noise at several levels: the paper's own mechanism predicts little or no gain there, so a large gain would contradict the proposed frequency-gating explanation.","tokens_in":12576,"feed_emoji":"🛰️","tokens_out":8015,"duration_ms":73775,"temperature":0.7,"pith_summary":"This paper claims that hyperspectral image denoising improves when a spatial-spectral transformer is augmented with a frequency-domain branch that decouples noise from signal and a multiscale branch that captures local detail and global statistics. The proposed network, HDST, builds on the SERT transformer and adds an FFT-based preprocessing stage, a gated cross-domain attention that lets spatial features query frequency-domain noise patterns, and a multiscale atrous pyramid. On the Realistic benchmark it reports 30.62 dB PSNR against 29.68 dB for its baseline, a 0.94 dB gain, and on the ICVL synthetic benchmark a maximum 0.52 dB gain, largest under mixture noise. The paper argues that the improvement comes from a closed loop of spatial feature extraction, frequency-domain noise decoupling, and cross-domain calibration, not from any single module. A sympathetic reader would care because real HSI noise is spatially non-uniform and coupled across bands, a setting where RGB-oriented denoisers and pure spatial models tend to blur weak spectral features.","feed_headline":"HSI denoiser gains 0.94 dB fusing frequency and spatial domains","feed_subtitle":"The gains are largest on real and mixed noise, at only 14.7% more compute than the baseline.","key_machinery":"The central mechanism is the frequency post-processing (FPP) unit inserted into the deepest residual transformer layers. Its forward pass is $F = \\mathrm{FFT}(S)$, $F_c = \\mathrm{Concat}(\\mathrm{Re}\\,F, \\mathrm{Im}\\,F)$, $F_{\\mathrm{proc}} = \\mathrm{ASPP\\text{-}FFT}(F_c)$ with dilation rates $[2,4,8]$, $S' = \\mathrm{Conv}_{3\\times3}(\\mathrm{IFFT}(F_{\\mathrm{proc}}))$, $\\mathrm{Gate} = \\sigma(\\mathrm{Conv}_{1\\times1}(F_{\\mathrm{proc}}))$, and fusion $F' = S \\odot \\mathrm{Gate} + \\alpha\\, S' \\odot (1 - \\mathrm{Gate})$. The gate is intended to keep clean areas ($\\mathrm{Gate} \\to 1$) and replace noisy regions ($\\mathrm{Gate} \\to 0$), so the frequency branch injects noise-prior information only where needed. A frequency-spatial collaborative attention module then uses spatial windows as queries against frequency-derived keys and values, and a learnable coefficient $\\beta$ scales the residual. The HDMS multiscale module, applied after each transformer block with dilation rates $[2,4,8]$, is the second pillar: it supplies edge-to-global spatial features that give the frequency pyramid content to act on.","core_discovery":"HDST's central claim is that the coupled spatial-spectral noise of hyperspectral images can be separated by routing features through a frequency-domain post-processing unit and recombining them with spatial features under learned control. The denoiser keeps SERT's rectangular-window attention and spectral enhancement, then inserts, in the deepest transformer layers, an FFT-scale gated fusion module that transforms features with FFT, processes the concatenated real and imaginary parts with atrous convolutions at dilation rates 2, 4, and 8, and fuses the reconstructed spatial signal with the original using a learned gate and a scalar intensity parameter. A frequency-spatial collaborative attention module then treats spatial features as queries and frequency-domain features as keys and values, so spatial positions can retrieve the noise pattern that corrupts them, and a learnable coefficient controls the residual. The paper reports that the complete model reaches 30.62 dB PSNR on the Realistic dataset versus 29.68 dB for the SERT baseline, outperforms the baseline on most synthetic noise patterns, and that the ablation attributes the largest single gain to the multiscale pyramid while the full gain comes from the combination.","pith_inferences":["The gating mechanism predicts a clean ordering of gains: largest when noise is spatially non-uniform and frequency-localized (real sensor noise, stripe, mixture), smaller for i.i.d. Gaussian; a direct test would be to measure the gate's activation statistics on each noise type and confirm that the gate approaches 0 in noisy regions and 1 in clean ones.","Because the new modules add 61% parameters but only about 15% FLOPs, the design suggests the same frequency-gating branch could be transplanted onto other spatial-spectral backbones; if the gain transfers, the framework would be a general plug-in rather than a SERT-specific fix.","The ablation's single-run gaps of 0.05 to 0.15 dB between the frequency-only and baseline configurations mean the independent contribution of the frequency module is not yet separated from training noise; multi-seed runs would clarify which module actually carries the gain."],"forward_implications":["On the Realistic benchmark, HDST improves PSNR from 29.68 to 30.62 dB over SERT, with SSIM rising from 0.9533 to 0.9555 and SAM falling from 2.536 to 2.417.","On the ICVL synthetic benchmark, HDST beats SERT by up to 0.52 dB under Gaussian+Mixture noise (40.52 vs 40.00 dB) and improves most other noise patterns, with the smallest gains under Non-i.i.d Gaussian.","The ablation shows the full gain is synergistic: multiscale-only gives +0.59 dB, frequency-plus-multiscale without dynamic fusion gives +0.78 dB, and the complete model gives +0.94 dB, while frequency processing alone gives only +0.05 dB.","The architecture adds 61.2% parameters but only 14.7% FLOPs, so the frequency and multiscale branches are claimed to be computationally inexpensive relative to their parameter cost.","Because the framework is designed as post-processing on a spatial-domain denoiser with matched input and output dimensions, the authors claim it can be transplanted to other spatial restoration models and high-dimensional visual data."],"supporting_citations":[{"why":"Supplies the SERT spatial-spectral transformer baseline whose architecture HDST extends and against which all reported gains are measured.","marker":"[20]"},{"why":"Provides the atrous spatial pyramid pooling idea, adapted with frequency-domain dilation rates [2,4,8] to build the ASPP-FFT module.","marker":"[17]"},{"why":"BM4D, the classic volumetric filtering baseline used for comparison on real and synthetic noise.","marker":"[35]"},{"why":"HSID-CNN, a deep convolutional baseline showing the gap that transformer-based methods aim to close.","marker":"[38]"},{"why":"SSRT-UNet, the strongest recent deep baseline that HDST is compared against on synthetic noise.","marker":"[40]"},{"why":"Supplies the Realistic dataset of 59 noisy/clean hyperspectral images and the train/test split used for the headline 0.94 dB gain.","marker":"[43]"},{"why":"Defines the synthetic noise patterns (Non-i.i.d Gaussian, stripe, deadline, impulse, mixture) on which the ICVL experiments are built.","marker":"[44]"},{"why":"Provides the ICVL hyperspectral image collection used as the synthetic benchmark source.","marker":"[45]"}],"fun_headline_variants":["Hybrid-domain transformer lifts HSI denoising by 0.94 dB","Gated FFT fusion strips spatial-spectral noise in HSI","Frequency-spatial attention boosts HSI denoising","Multiscale FFT transformer improves HSI denoising"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central evidence rests on single-run comparisons: each configuration is trained once, and the reported gains, especially the 0.05 to 0.15 dB differences in ablation steps and some synthetic noise types, are treated as meaningful without error bars or multi-seed statistics.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid-domain transformer lifts HSI denoising by 0.94 dB","Gated FFT fusion strips spatial-spectral noise in HSI","Frequency-spatial attention boosts HSI denoising","Multiscale FFT transformer improves HSI denoising"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000654,"raw_usage":{"total_tokens":3043,"prompt_tokens":1040,"completion_tokens":2003,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":1929}},"tokens_in":656,"tokens_out":2003,"duration_ms":15523,"temperature":1.0,"reasoning_tokens":1929,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:48:37.793480+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SERT and HDST on the same Realistic split with at least five random seeds and compute the 95% confidence interval of the per-image PSNR difference; if the interval contains zero, the claimed 0.94 dB improvement is not statistically established. As a second check, run HDST on pure i.i.d. Gaussian noise at several levels: the paper's own mechanism predicts little or no gain there, so a large gain would contradict the proposed frequency-gating explanation.","supporting_citations":[{"cited_title":"Gao , author X","cited_arxiv_id":null,"evidence_quote":"Supplies the SERT spatial-spectral transformer baseline whose architecture HDST extends and against which all reported gains are measured."},{"cited_title":"Ozdemir \\ and\\ author K","cited_arxiv_id":null,"evidence_quote":"Provides the atrous spatial pyramid pooling idea, adapted with frequency-domain dilation rates [2,4,8] to build the ASPP-FFT module."},{"cited_title":"Wang , author H","cited_arxiv_id":null,"evidence_quote":"BM4D, the classic volumetric filtering baseline used for comparison on real and synthetic noise."},{"cited_title":"Zhang , author Y","cited_arxiv_id":null,"evidence_quote":"HSID-CNN, a deep convolutional baseline showing the gap that transformer-based methods aim to close."},{"cited_title":"Maggioni , author V","cited_arxiv_id":null,"evidence_quote":"SSRT-UNet, the strongest recent deep baseline that HDST is compared against on synthetic noise."},{"cited_title":"Xiong , author J","cited_arxiv_id":null,"evidence_quote":"Defines the synthetic noise patterns (Non-i.i.d Gaussian, stripe, deadline, impulse, mixture) on which the ICVL experiments are built."}],"review_version":2}