{"id":"2bcb8806-06e3-4636-8554-5f841f152f5b","arxiv_id":"2501.02613","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"LWFNet, a Transformer plus KAN network, retrieves wind fields directly from coherent Doppler lidar spectra and reportedly beats the spectral centroid baseline, but the reported gains rely on a test-set-fitted bias correction.","lead":"A deep learning model called LWFNet directly retrieves wind speed and direction from coherent Doppler lidar spectra, using a line Transformer and a KAN decoder. It claims to beat the traditional spectral centroid algorithm and extend the usable altitude range, but the evaluation relies on a bias correction fitted to the very radiosonde measurements used for scoring.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"All reported accuracy numbers in Table I and Fig. 3 are computed after fitting a per-profile constant bias to the same radiosonde profiles used as ground truth (Section IV-D).","rationale":"The reader's REJECT verdict is correct. The central claim is the super-accuracy phenomenon, and it is evaluated in Section IV-D. The per-profile bias fitted to ground truth is the load-bearing step: without independent bias estimation, the absolute metrics and the relative comparison are in-sample. The model might still be useful, but the paper as written does not support the headline claim. This stress-test reinforces the reader's finding, so no verdict change is needed.","tokens_in":15973,"tokens_out":3952,"duration_ms":42050,"concrete_test":"Recompute all metrics in Table I and Fig. 3 from the raw network and spectral-centroid outputs with no per-profile bias fitted to the March radiosondes. Instead, either (a) apply no bias correction, or (b) estimate the bias from the December-February training/validation radiosondes (or a separate calibration campaign) and apply it fixed to the March test profiles. Then recompute the high-SNR speed RMSE difference between LWFNet and the spectral centroid with a paired bootstrap or t-test over the 32 profiles. If the gap does not remain significantly in LWFNet's favor, the super-accuracy claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-D states: \"we introduce a fixed bias to both the spectral centroid and LWFNet retrieval results. This bias is determined by minimizing the mean squared error between the wind retrieval sequence and the corresponding radiosonde measurement sequence in high-SNR regions. All subsequent results presented account for this bias adjustment.\" Thus every metric in Table I, the scatter plots in Fig. 3, and the abstract's \"super-accuracy\" claim are computed after a per-profile offset has been fitted to the very radiosonde data used as ground truth. That is an in-sample calibration: the reported RMSE/MAE are optimistically biased, and the comparison between LWFNet and the spectral centroid is not a blind test. The gap underlying the headline claim is small (speed RMSE 0.795 vs 0.885 m/s in high-SNR regions), and no per-instance confidence intervals are given. One could argue that a constant bias is needed to handle coordinate-system differences, but it must be estimated from independent data (e.g., training-period radiosondes or a held-out subset), not from the same 32 March profiles used for scoring. With 32 per-profile constants fitted to the test target, the paper does not establish that LWFNet is more accurate than its training labels.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes LWFNet, an end-to-end neural network for coherent Doppler wind lidar (CDWL) wind field retrieval. The network combines a line Transformer that treats each range-gate power spectrum as a token, a KAN decoder, and a median filter. It is trained on spectral-centroid retrieval labels filtered by hand-set criteria and evaluated against 32 radiosonde profiles from March 2024. The paper claims that LWFNet extends the valid detection range and achieves 'super-accuracy' relative to its training labels, and it reports comparisons with ResNet, ViT, and architectural ablations.","tokens_in":16177,"tokens_out":5628,"duration_ms":53137,"significance":"If the quantitative claims were supported, LWFNet would be a useful benchmark for deep-learning-based CDWL retrieval, and the line Transformer is a plausible inductive bias for this signal class. The paper also provides a fairly complete description of the data, architecture, and ablations. However, the evaluation protocol in Section IV-D fits a per-profile bias to the radiosonde ground truth before computing every reported metric, and the headline accuracy and super-accuracy claims rest on that in-sample adjustment. The current evidence is therefore not sufficient to support the central quantitative conclusions.","major_comments":[{"comment":"The reported metrics are not a blind evaluation. Section IV-D states that 'we introduce a fixed bias to both the spectral centroid and LWFNet retrieval results,' with the bias 'determined by minimizing the mean squared error between the wind retrieval sequence and the corresponding radiosonde measurement sequence in high-SNR regions,' and that 'all subsequent results presented account for this bias adjustment.' Because this bias is fit for each of the 32 test profiles to the same radiosonde data used as ground truth, the RMSE, MAE, Pearson correlations, and scatter quantities in Table I and Fig. 3 are optimistically biased, and the comparison against the spectral centroid is not a head-to-head test of the retrieval methods. Moreover, the same protocol appears to underlie Tables II and IV. Please re-run the evaluation with a bias estimated only from independent data (e.g., training-period radiosondes or a held-out subset), or report all metrics without any bias adjustment, and clearly state which numbers are affected.","section":"IV-D, Table I, Fig. 3"},{"comment":"The central 'super-accuracy' claim is not established by the reported numbers. In the high-SNR region the horizontal speed RMSE difference is only 0.795 m/s versus 0.885 m/s for the spectral centroid, the comparison is made after the test-set-fitted bias described above, and no per-instance confidence intervals or significance tests are provided. With only 32 radiosonde profiles, this margin is not sufficient to support the claim that LWFNet 'surpasses the labeled targets' without an unbiased evaluation and uncertainty quantification.","section":"Abstract, IV-D, Table I"},{"comment":"The training labels are produced by applying hand-set filtering thresholds (LOS speed < 42 m/s, SNR > -35 dB, spectral width between 0.5 and 7.5, and the -25 dB exception) to spectral centroid outputs. The paper does not test the sensitivity of the results to these thresholds, nor does it discuss the fact that the thresholds encode external assumptions about which spectral centroid outputs are credible. Because the super-accuracy claim is about surpassing these labels, a sensitivity analysis or a justification for the thresholds is needed.","section":"IV-A, V"}],"minor_comments":[{"comment":"The formula for the Pearson correlation coefficient appears to have a malformed denominator; it should be the square root of the product of the variances of v and v_hat, not the printed expression.","section":"IV-C, Eq. (16)"},{"comment":"There is a typo in the spline definition: 'B-sphinei(x)' should be 'B-spline_i(x)'.","section":"III-B"},{"comment":"Please clarify how the invalid range gates 20-40 are handled in the evaluation metrics (e.g., excluded, masked, or treated as NaN) and how the 'theoretical maximum detection range' is converted to range-gate indices.","section":"IV-B"},{"comment":"The section heading 'Aknowledgement' should be 'Acknowledgments', and there are small typographical issues elsewhere (e.g., 'e,g,' in Section IV-F and inconsistent capitalization).","section":"VII"}],"recommendation":"major_revision","confidential_remarks":"The evaluation flaw is central but fixable in principle: the authors could recalibrate using independent radiosonde data or report metrics without calibration. If no independent calibration source is available, the super-accuracy claim should be withdrawn. The paper also currently overstates confidence given 32 test profiles and no uncertainty intervals."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, read this one if you want a clean case study in how a test-set-fitted calibration can sink otherwise interesting work. The core idea is genuinely new: LWFNet is the first end-to-end network that takes raw coherent Doppler wind lidar spectra and directly outputs wind components, using a line Transformer that splits the spectrogram by range gate (vector embedding) rather than square patches, plus a KAN decoder. That is a sensible adaptation to the physics of the signal, and the ablation study supports it: LiT beats ViT, KAN helps a bit, median filter helps. The paper is also unusually transparent: it states the bias-adjustment procedure in Section IV-D and discusses data limitations in Section V.\n\nThe soft spot is load-bearing. All reported accuracy numbers, including the 'super-accuracy' claim, are computed after fitting a per-profile constant offset that minimizes MSE against the same radiosonde profiles used as ground truth. With 32 test instances, that is 32 nuisance parameters fit to the evaluation target. The headline gap is small (speed RMSE 0.795 vs 0.885 m/s) and no per-instance uncertainty is given. So the paper does not demonstrate that LWFNet is more accurate than its training labels. It may well be—the smoothing and multi-range-gate context could plausibly improve on a per-gate centroid estimate—but this protocol cannot show it. The fix is straightforward: estimate the bias on independent data (e.g., radiosondes from the training period) or show sensitivity to holding out a few test profiles. I'd also want error bars and ideally code/data, since 32 profiles is a thin test set.\n\nSummary: the architecture contribution is worth keeping, the evaluation needs revision. For peer review, I would send it out rather than desk-reject; a serious referee could help them fix the protocol, and the 'first end-to-end CDWL retrieval' point is worth checking against the literature. As it stands, reject or major revision; don't let the headline claim stand as written.","headline":"A genuinely novel end-to-end CDWL retrieval architecture whose headline accuracy claims are undermined by per-profile bias fitting to the test ground truth.","tokens_in":16747,"tokens_out":1865,"would_cite":false,"duration_ms":17660,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural network trained on standard lidar wind estimates beats those estimates themselves.","keywords":["coherent Doppler wind lidar","wind field retrieval","deep learning","Transformer","Kolmogorov-Arnold network","super-accuracy","spectral centroid","radiosonde"],"falsifier":"Re-evaluate LWFNet with no per-profile bias adjustment, or with the offset estimated from an independent set of radiosondes collected at a different site or season, and compare against the spectral centroid estimator; if the super-accuracy gap disappears or reverses, the central claim is unsupported.","tokens_in":15772,"feed_emoji":"🌬","tokens_out":3652,"duration_ms":35396,"temperature":0.7,"pith_summary":"This paper claims that a neural network trained only on labels produced by a conventional spectral-centroid wind retrieval algorithm can retrieve tropospheric wind fields from coherent Doppler wind lidar spectra more accurately than those labels themselves, and can extend the altitude range over which valid winds are obtained. If true, it would mean lidar wind fields can be improved without any additional ground truth beyond what the standard algorithm already provides. The authors propose LWFNet, a hybrid of a line Transformer and a Kolmogorov-Arnold network decoder, and report that it beats the spectral centroid estimator on radiosonde comparisons within the high-SNR region while also producing credible winds where the conventional method fails. The quantitative evidence is limited to 32 radiosonde profiles from a single month, with a per-profile bias correction fitted using the same radiosonde data.","feed_headline":"Neural net beats its own wind-retrieval training labels","feed_subtitle":"LWFNet, trained on standard centroid estimates, also pushes valid lidar winds higher into the troposphere.","key_machinery":"The central object is the line Transformer (LiT), which embeds each range gate's full 128-bin power spectrum as a single token via a convolutional layer with kernel (1, W), so the complete Doppler spectrum of each range gate is preserved rather than chopped into square patches. A two-layer Kolmogorov-Arnold network decoder, whose learnable univariate spline activations approximate the mapping from token to wind component, and a median filter over seven range gates complete the pipeline. This vector embedding is what lets the network pool information across all range gates while respecting the physical structure of the lidar signal.","core_discovery":"LWFNet, built from a line Transformer encoder, a Kolmogorov-Arnold network decoder, and a median filter, retrieves horizontal wind speed and direction directly from the raw power spectra of a three-direction VAD scan. Trained on spectral centroid labels, it surpasses the spectral centroid estimator in high-SNR regions on RMSE, MAE, and Pearson correlation, and continues to produce meteorologically acceptable results in regions where the centroid method fails entirely. The authors call this 'super-accuracy' and attribute it to the model's global view across range gates, the inherent smoothing of deterministic models, and the masking of unreliable targets during training.","pith_inferences":["A natural test of super-accuracy would be to train on spectral-centroid labels from one lidar and evaluate against radiosondes from a different site or season; if the effect is real, it should survive without the per-profile offset fitted on the same test data.","The reported super-accuracy may be partly a statistical consequence of label noise: if the centroid labels are noisy around the truth and the network learns a conditional mean, its predictions can be closer to the true wind than any single noisy label, which is consistent with the paper's smoothing discussion and is testable by adding controlled noise to labels.","The architecture could be applied to other scanning geometries, such as four-beam or Doppler beam swinging, without structural change since the network consumes a fixed number of spectra as input channels."],"forward_implications":["If super-accuracy holds, deep models can be trained on existing operational retrieval outputs without needing simulated spectra or manual labels, and still beat the outputs they train on.","LWFNet extends the credible wind detection range beyond the altitude at which spectral centroid estimates become unreliable, up to the 220 range gates of the lidar.","The line Transformer's vector embedding outperforms patch-based ViT and ResNet baselines on this task, suggesting that preserving spectrum completeness matters for lidar retrieval.","Median filtering and target masking both contribute to accuracy, indicating that smoothness priors are useful for meter-scale wind fields.","LWFNet fills in missing wind data across time, producing continuous wind fields without abrupt spatial or temporal changes."],"supporting_citations":[{"why":"The spectral centroid frequency estimator supplies the training labels and the baseline that LWFNet is compared against.","marker":"[26]"},{"why":"The vision Transformer architecture is the starting point for the line Transformer's encoder design.","marker":"[36]"},{"why":"The Kolmogorov-Arnold network provides the decoder architecture with learnable spline activations.","marker":"[37]"},{"why":"A convolutional-neural-network denoiser for lidar spectra is an earlier deep-learning approach that LWFNet outperforms in the comparison.","marker":"[28]"},{"why":"A U-Net based lidar spectral denoiser is another prior deep-learning baseline that requires simulated data, in contrast to LWFNet.","marker":"[29]"},{"why":"ResNet serves as the convolutional deep-learning baseline in the model comparison.","marker":"[56]"},{"why":"The attention mechanism and Transformer encoder blocks underpin the line Transformer and the patch-based ViT baseline.","marker":"[50]"},{"why":"Masked autoencoders inspire the target masking strategy used in LWFNet training.","marker":"[57]"},{"why":"Velocity azimuth display scanning is the measurement technique that defines the input format of the lidar power spectra.","marker":"[54]"}],"fun_headline_variants":["Wind lidar net beats its training labels","Neural net extends wind retrieval range and accuracy","LWFNet outperforms centroid-based wind estimation","Transformer-based wind retrieval surpasses traditional method","Lidar wind net achieves super-accuracy over centroid"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The quantitative claims assume that a fixed per-profile offset, fitted by minimizing the error between the retrieval sequence and the radiosonde sequence in high-SNR regions, does not itself manufacture the reported accuracy gains.","fun_headline_variants_meta":{"raw":{"variants":["Wind lidar net beats its training labels","Neural net extends wind retrieval range and accuracy","LWFNet outperforms centroid-based wind estimation","Transformer-based wind retrieval surpasses traditional method","Lidar wind net achieves super-accuracy over centroid"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000318,"raw_usage":{"total_tokens":1805,"prompt_tokens":961,"completion_tokens":844,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":773}},"tokens_in":577,"tokens_out":844,"duration_ms":8656,"temperature":1.0,"reasoning_tokens":773,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:08:21.495328+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-evaluate LWFNet with no per-profile bias adjustment, or with the offset estimated from an independent set of radiosondes collected at a different site or season, and compare against the spectral centroid estimator; if the super-accuracy gap disappears or reverses, the central claim is unsupported.","supporting_citations":[{"cited_title":"Perceptual effects of spectral modifications on musical timbres,","cited_arxiv_id":null,"evidence_quote":"The spectral centroid frequency estimator supplies the training labels and the baseline that LWFNet is compared against."},{"cited_title":"Coherent doppler wind lidar with real-time wind pro- cessing and low signal-to-noise ratio reconstruction based on a convolutional neural network,","cited_arxiv_id":null,"evidence_quote":"A convolutional-neural-network denoiser for lidar spectra is an earlier deep-learning approach that LWFNet outperforms in the comparison."},{"cited_title":"Denoising coherent doppler lidar data based on a u-net convolutional neural network,","cited_arxiv_id":null,"evidence_quote":"A U-Net based lidar spectral denoiser is another prior deep-learning baseline that requires simulated data, in contrast to LWFNet."},{"cited_title":"Masked autoencoders are scalable vision learners,","cited_arxiv_id":null,"evidence_quote":"Masked autoencoders inspire the target masking strategy used in LWFNet training."},{"cited_title":"Precipitation motion by pulse doppler","cited_arxiv_id":null,"evidence_quote":"Velocity azimuth display scanning is the measurement technique that defines the input format of the lidar power spectra."}],"review_version":1}