{"id":"cc6f3fb3-f953-460f-a882-06608cd7cad5","arxiv_id":"2504.16870","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"CRSynthNet synthesizes a later cloud-free optical image from an earlier cloudy optical image plus radar images at both dates, reporting modest gains over prior methods on the new Zhengzhou flood dataset.","lead":"This paper introduces CRSynthNet, a neural network that creates cloud-free satellite photos by combining radar images and a cloudy photo from an earlier date. It also releases a new flood-region dataset, TCSEN12, to train and test such systems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reference images with up to 5% cloud and 6-day composite alignment bias every reported metric; without cloud-masked evaluation the claim of beating all baselines is not established.","rationale":"The paper's central assertion is the quantitative superiority in Table 1. The most fragile prerequisite is the ground-truth target. The dataset description in Section 2.1 explicitly allows residual cloud (<5%) and uses a 6-day composite to align dates; the Zhengzhou flood is chosen precisely because it produces large land-cover change. Thus the reference is not a true cloud-free instantaneous observation. All metrics in Section 3.2 compare model output against this reference, and the training loss in Equation 3 also uses it, so the network is optimized to match whatever is in the reference, including residual cloud or temporal artifacts. A separate, albeit visible, issue is the ablation contradiction: Table 2 shows 'No DownUp' improves PSNR, MAE, and RMSE over CRSynthNet while the text claims a drop. This undermines the component attribution, but it does not directly invalidate the headline comparison against baselines. The reference-label issue is therefore more load-bearing. The proposed check, cloud-masked re-evaluation plus a clean-reference subset, can settle whether the reported margins persist when the target is actually cloud-free. If they do, the central claim stands; if not, the result is conditional on label artifacts. This supports the reader's conditional verdict, so no change in verdict is recommended.","tokens_in":12673,"tokens_out":4494,"duration_ms":41773,"concrete_test":"Recompute all Table 1 metrics on the TCSEN12 test set with cloud and cloud-shadow pixels masked out of the reference before evaluation, using Sentinel-2 QA60 or s2cloudless masks, and re-rank all methods. Also repeat on a subset of test patches whose reference images have 0% cloud and no flood-related change within the 6-day composite. If CRSynthNet still achieves the highest PSNR and SSIM and the lowest MAE and RMSE on both the masked and clean-reference evaluations, the concern is resolved. If its rank drops or the margins shrink below meaningful thresholds, the headline result is an artifact of contaminated references.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that CRSynthNet beats all baselines depends on the validity of the reference images in Table 1. Section 2.1 defines the near cloud-free target as a Sentinel-2 image with less than 5% cloud coverage, aligned to the other dates through a 6-day composite. During the Zhengzhou flood period, land cover changed substantially between the two dates; the composite therefore cannot guarantee pixel-level correspondence between SAR and optical acquisitions. Because every metric in Table 1, including PSNR, SSIM, MAE, RMSE, and even FID, is computed against this contaminated reference, residual cloud pixels and misregistration are scored as ground truth. A model that reproduces residual cloud patterns can improve its metrics, while a model that correctly predicts the underlying surface is penalized. The paper reports no cloud masking, no elimination of cloud-containing test patches, and no sensitivity analysis to the 5% cloud threshold, so the numerical superiority over MTS2ONet (PSNR 26.978 vs 26.225) could be driven by artifacts of the label rather than by true cloud-free synthesis quality.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CRSynthNet, a GAN-based network for synthesizing a cloud-free optical image at a target date from two temporal Sentinel-1 SAR images and one cloud-contaminated Sentinel-2 image from an earlier date. The architecture combines convolutional blocks, a DownUp Block, FusionAttention, SwinBlocks, a decoder with channel and spatial attention, and a three-scale discriminator. The authors also introduce TCSEN12, a dataset of Sentinel-1/Sentinel-2 image triplets from the Zhengzhou July 2021 flood region. Quantitative evaluation reports that CRSynthNet achieves PSNR 26.978, SSIM 0.648, MAE 0.041, RMSE 0.050, and FID 72.789, outperforming eight comparison methods on the TCSEN12 test set. Ablation studies are presented for the DownUp Block, FusionAttention, the discriminator, and the attention modules.","tokens_in":12912,"tokens_out":3523,"duration_ms":32533,"significance":"If the claimed results are valid, the paper contributes a new dataset and a method that addresses a practical scenario: generating cloud-free optical images when the only available optical input is itself cloud-contaminated. The design of the DownUp Block and FusionAttention, and the use of multi-temporal SAR with contaminated optical data, are plausible directions for further research. The paper also explicitly discusses limitations regarding feature-sparse scenes. The main value lies in the dataset and the empirical comparison, but the evaluation's validity depends on the quality of the reference images and the consistency of the ablation analysis.","major_comments":[{"comment":"The ablation narrative contradicts the reported numbers. The text states that removing the DownUpBlock caused \"a noticeable performance drop,\" yet Table 2 reports PSNR 27.053, which is higher than the full model's 26.978; it also reports a lower MAE (0.034 vs. 0.041). Similarly, the removal of Channel Attention improves FID to 71.875, better than the full model's 72.789, while the text calls it \"a significant reduction in performance.\" Because the paper's central contribution claims rest on the necessity of these components, this internal inconsistency must be resolved: either reinterpret the ablations with appropriate metrics or provide a corrected analysis.","section":"§3.4, Table 2"},{"comment":"The evaluation reference is defined as a Sentinel-2 image with less than 5% cloud coverage, aligned to the other dates through a 6-day composite. During the Zhengzhou flood period, land cover changed substantially between the two acquisition windows, so the composite cannot guarantee pixel-level correspondence. All metrics in Table 1 (PSNR, SSIM, MAE, RMSE, FID) are computed against this potentially contaminated reference, with no cloud masking, no exclusion of cloudy patches, and no sensitivity analysis to the 5% cloud threshold. Residual cloud or misregistration in the reference could bias every reported metric and even affect the ranking over baselines. The paper should include cloud-masked evaluation or otherwise demonstrate that the reference images are sufficiently clean to support the headline comparison.","section":"§2.1, §3.2"}],"minor_comments":[{"comment":"There are multiple grammatical and wording errors, e.g., \"spectral consist\" should be \"spectral consistency,\" and \"This study offer\" should be \"This study offers.\" Proofreading is needed throughout.","section":"Abstract"},{"comment":"The reference list contains an incomplete citation: \"[30, 31? ]\" appears on page 4. Also, \"deep learnig\" on page 3 is a typo.","section":"§1 (References)"},{"comment":"The values of the weighting factors alpha, beta, gamma, lambda, and lambda_gp are not reported. Since the paper states that the balance achieved by lambda is critical, please provide the actual values or a sensitivity study.","section":"§2.2.3 (Loss function)"},{"comment":"The dataset description would benefit from explicit details: the number of image pairs, the spatial resolution of the tiles, the number of training/validation/test samples, and the cloud coverage distribution. This information is important for reproducibility and for interpreting the comparison.","section":"§2.1 (TCSEN12 Dataset)"},{"comment":"The quantitative comparison reports a single run without standard deviations or statistical significance tests. The margin over MTS2ONet (PSNR 26.978 vs. 26.225, MAE 0.041 vs. 0.049) is modest, so providing variation across runs would strengthen the claim of superiority.","section":"§3.2 (Quantitative Evaluation)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the dataset is a useful contribution. The main concerns are the internal inconsistency in the ablation section and the validity of the reference used for evaluation. The reference issue is not circular, since the comparison is on a held-out test set, but it is a correctness-risk concern that should be addressed with cloud-masked metrics or a sensitivity analysis. The ablation inconsistency is directly checkable from the table and needs correction before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this as a dataset paper with a network attached. The genuinely new piece is TCSEN12: two temporal Sentinel-1 scenes plus a cloudy Sentinel-2 from the earlier date, with a near cloud-free Sentinel-2 from the later date as reference. That is more realistic than the common assumption of a clean auxiliary optical image. The comparison against eight baselines on a held-out test split is straightforward, and Table 1 supports the claim that CRSynthNet beats them on this dataset. The margin over MTS2ONet is modest (0.75 dB PSNR, 0.026 SSIM) but consistent across metrics.\n\nThe soft spots are real but manageable. First, the ablation section contradicts itself. Table 2 shows that removing the DownUpBlock yields PSNR 27.053, MAE 0.034, RMSE 0.049 — all better than the full model's 26.978/0.041/0.050 — yet the text calls that a \"noticeable performance drop.\" Removing channel attention also improves FID from 72.789 to 71.875. Either the narrative or the table is wrong, and as written a reader cannot tell which component actually helps. Second, all metrics come from a single run; no error bars or seeds, so the 0.75 dB lead over MTS2ONet is not established as significant. Third, the reference images are only \"near cloud-free\" (less than 5% cloud) and are aligned through a 6-day composite during a flood. Residual cloud or misregistration is then scored as ground truth. There is no cloud masking and no sensitivity analysis around the 5% threshold, so the absolute numbers could be biased. This affects all methods equally, so the relative ranking is less vulnerable, but the absolute quality claims need a robustness check. Code and data are promised, not provided, which adds uncertainty.\n\nThe stress-test worry about the reference is the one that would make me hesitate to publish as is. It is not fatal — the comparison is still meaningful within the dataset — but the authors should show that results hold when cloud-containing patches are excluded or masked. I would not demand that as a condition for review, but I would expect it to be addressed in revision.\n\nThe paper deserves a serious referee. The dataset alone gives the community a useful resource, and the architecture is workmanlike but not derivative in a problematic way. A careful revision can fix the ablation inconsistency, add repeated runs, and address the reference issue. I would send it to peer review and let the technical content be evaluated.","headline":"Useful new dataset and a workmanlike SAR-optical fusion architecture, but the ablation narrative contradicts its own table and the 'near cloud-free' reference deserves a robustness check before the metrics are taken at face value.","tokens_in":13437,"tokens_out":5831,"would_cite":false,"duration_ms":44424,"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":"CRSynthNet synthesizes a cloud-free optical image for a later date from two SAR images and one cloudy earlier optical image, beating eight methods on the new Zhengzhou flood dataset.","keywords":["SAR-optical data fusion","deep learning","missing information","information reconstruction","cloud cover","generative adversarial network","multi-temporal image synthesis"],"falsifier":"Take the TCSEN12 test pairs whose reference Sentinel-2 tiles are completely cloud-free by manual inspection, or mask out residual cloud pixels before computing PSNR, SSIM, MAE, and RMSE, and re-run the comparison against MTS2ONet; if CRSynthNet no longer leads on all metrics, the reported advantage is an artifact of scoring against a contaminated target.","tokens_in":12460,"feed_emoji":"🛰️","tokens_out":8093,"duration_ms":68290,"temperature":0.7,"pith_summary":"This paper tries to establish that cloud-free optical satellite imagery can be synthesized for a date when only cloudy optical data exist, by fusing two radar acquisitions with the contaminated optical image. The proposed network, CRSynthNet, is evaluated on a new two-temporal Sentinel-1/Sentinel-2 dataset, TCSEN12, built over the Zhengzhou July 2021 flood area, where land cover changed substantially between dates. On that dataset the paper reports that CRSynthNet beats eight comparison methods on PSNR, SSIM, MAE, RMSE, and FID, with PSNR 26.978 and SSIM 0.648. The claimed value is practical: flood response and agricultural monitoring often fail not from a lack of algorithms but from missing or cloudy optical observations, and a method that tolerates real cloud contamination would make those applications more reliable.","feed_headline":"Cloudy scenes become cloud-free optical images via SAR fusion","feed_subtitle":"Network fuses two radar dates and one cloudy optical date, beating eight baselines in tests.","key_machinery":"The load-bearing mechanism is the FusionAttention block, a joint attention module for two temporal feature maps. It derives queries, keys, and values from both inputs, concatenates Q and K along channels, L2-normalizes them, and computes Attention1 = Q · (K · V1) and Attention2 = Q · (K · V2); each attention output is scaled by a learnable γ and added back to its input as a residual. This lets the network decide, per location and channel, how much to trust the radar structure and how much to trust the earlier cloudy optical image when reconstructing the target scene. Around this block the architecture wraps a DownUpBlock for coarse-to-fine feature extraction, four scaled-up Swin Transformer blocks for multi-scale features, and a decoder with channel and spatial attention, while a three-scale discriminator with spectral normalization supplies adversarial feedback.","core_discovery":"The paper's central claim is that missing optical images can be synthesized at the target date even when the only optical image available is cloud-contaminated, provided SAR images from both the earlier and the target date are available. The generator consumes S1 T1, S1 T2, and S2 T1 and predicts S2 T2; the reference used for training and testing is the latest S2 image with less than 5% cloud cover. CRSynthNet fuses the two modalities with a DownUp Block, a FusionAttention block, Swin Transformer blocks, and an attention-equipped decoder, and is trained adversarially with a similarity loss plus a least-squares GAN loss and a WGAN gradient penalty on the discriminator. On the TCSEN12 test set the paper reports PSNR 26.978, SSIM 0.648, MAE 0.041, RMSE 0.050, and FID 72.789, which it presents as the best among nine methods, with MTS2ONet the closest competitor. The supporting ablation study attributes the gains to the designed modules and discriminator, with spatial attention having the largest individual effect.","pith_inferences":["If the reference images retain residual cloud or geometric misalignment, the absolute PSNR and SSIM numbers in the table are optimistic; the relative ordering of methods is still informative, but the magnitudes should be treated as upper bounds rather than true reconstruction fidelity.","Because the network predicts a later date from earlier data, the same pipeline could support near-real-time flood monitoring: as soon as a new SAR pass arrives, a cloud-free optical estimate can be produced without waiting for a later clear optical overpass.","A direct stress test would be to apply CRSynthNet to a non-flood region with stable land cover, or to train with reference labels built from multi-date compositing; this would separate the benefit of multi-temporal fusion from the benefit of learning flood-induced change."],"forward_implications":["Users can obtain a usable optical image for a flooded area on a target date even when that date's optical acquisition is fully obscured, because the method only needs two SAR acquisitions and an earlier cloudy optical scene.","On the TCSEN12 test set, CRSynthNet reports the best scores on all five metrics: PSNR 26.978, SSIM 0.648, MAE 0.041, RMSE 0.050, and FID 72.789, ahead of the previous best MTS2ONet with PSNR 26.225, SSIM 0.622, and FID 81.150.","Ablations show the designed components carry the result: replacing the discriminator with the one from [13] lowers SSIM from 0.648 to 0.632, and removing spatial attention lowers PSNR to 25.808 and raises FID to 84.289.","Because the reference target is allowed up to 5 percent cloud cover, the dataset reflects realistic operational conditions rather than requiring perfectly cloud-free training targets."],"supporting_citations":[{"why":"Supplies the closest baseline and the similarity-loss structure this work builds on; its discriminator is also swapped into the ablation.","marker":"[13]"},{"why":"Provides the scaled-up transformer blocks that generate multi-scale features in the generator.","marker":"[33]"},{"why":"Defines the least-squares adversarial loss combined into the generator objective.","marker":"[34]"},{"why":"Defines the perceptual feature loss term used to compare generated and reference images.","marker":"[35]"},{"why":"Defines the multi-scale SSIM used as a structural similarity loss term.","marker":"[36]"},{"why":"Defines the SSIM metric used to rank all methods in the experiments.","marker":"[37]"},{"why":"Supplies the Wasserstein GAN loss with gradient penalty that stabilizes the discriminator.","marker":"[38]"},{"why":"Defines the FID metric used to compare perceptual quality across methods.","marker":"[45]"}],"fun_headline_variants":["Cloudy optical plus SAR yields cloud-free scenes","AI uses radar to strip clouds from satellite images","Multi-temporal SAR and dirty optical produce clean images","New method synthesizes cloud-free views from cloudy data","Fusing SAR and cloudy optical beats eight baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation treats a satellite image that may still contain up to 5 percent cloud cover as the perfect truth, so any leftover cloud or slight misalignment in those reference images would directly inflate every reported quality score.","fun_headline_variants_meta":{"raw":{"variants":["Cloudy optical plus SAR yields cloud-free scenes","AI uses radar to strip clouds from satellite images","Multi-temporal SAR and dirty optical produce clean images","New method synthesizes cloud-free views from cloudy data","Fusing SAR and cloudy optical beats eight baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1486,"prompt_tokens":990,"completion_tokens":496,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":423}},"tokens_in":606,"tokens_out":496,"duration_ms":4392,"temperature":1.0,"reasoning_tokens":423,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:53:32.188483+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the TCSEN12 test pairs whose reference Sentinel-2 tiles are completely cloud-free by manual inspection, or mask out residual cloud pixels before computing PSNR, SSIM, MAE, and RMSE, and re-run the comparison against MTS2ONet; if CRSynthNet no longer leads on all metrics, the reported advantage is an artifact of scoring against a contaminated target.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the closest baseline and the similarity-loss structure this work builds on; its discriminator is also swapped into the ablation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the least-squares adversarial loss combined into the generator objective."},{"cited_title":"Arjovsky, S","cited_arxiv_id":null,"evidence_quote":"Supplies the Wasserstein GAN loss with gradient penalty that stabilizes the discriminator."},{"cited_title":"Heusel, H","cited_arxiv_id":null,"evidence_quote":"Defines the FID metric used to compare perceptual quality across methods."}],"review_version":1}