{"id":"450a7934-a935-42f7-abe7-ffc360e8ff0d","arxiv_id":"2501.05265","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Fine-tuning a pretrained masked autoencoder with a patch-based GAN improves cloud removal on RICE remote sensing images.","lead":"This paper proposes a cloud removal method that fine-tunes a masked autoencoder (MAE) pretrained on natural images inside a generative adversarial network with a patch-wise discriminator. It reports higher PSNR and SSIM than several GAN baselines on the RICE datasets, though comparisons are limited by missing details in prior work.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-time input resolution is unspecified: the ViT generator expects 224×224 (196 patches) while RICE images are 512×512, so Table I comparisons may mix resolutions; this is the load-bearing weak point.","rationale":"The reader's weakest assumption identifies protocol mismatch between the proposed method and baselines from [19]. My stress-test finds a more specific and more internal version of the same concern: even without consulting [19], the proposed architecture's fixed 224×224 input conflicts with the 512×512 dataset size, and the paper gives no test-time inference rule. This makes the headline numbers unverifiable as reported. The concern is load-bearing because the entire contribution is the quantitative superiority claim; if the evaluation protocol is inconsistent, the margins could be artifacts. However, the issue is fixable by disclosing and standardizing the protocol, so a conditional verdict is still appropriate rather than outright rejection. The reader's verdict is therefore not changed, but the condition should explicitly require a precise test-time preprocessing description and matched re-evaluation of baselines.","tokens_in":6222,"tokens_out":3453,"duration_ms":35006,"concrete_test":"Obtain or reimplement the trained model and define an explicit test-time protocol, then recompute Table I: run the generator on the RICE1/RICE2 test sets both at native 512×512 (e.g., overlapping 224×224 crops with stitching) and at directly resized 224×224, and re-evaluate the [19] baselines under the identical protocol. If the native-512 results drop below SpA GAN, or if the two protocols differ by more than the reported margins, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim (Table I: PSNR 33.659/34.056, SSIM 0.976/0.955) depends on the proposed model being evaluated under conditions comparable to baselines from [19]. Section III-A states the generator is a ViT operating on 196 patches of size 16×16, i.e., a fixed 224×224 input, and that during training images are randomly cropped to 224×224. Section IV-A states RICE1/RICE2 images are 512×512. The paper never specifies how 512×512 test images are fed to the model: resized to 224? center-cropped? tiled and stitched? If test images are resized, PSNR/SSIM are computed at reduced resolution and compared with baselines evaluated at native resolution; if passed at 512×512, the positional-embedding and patch-count assumptions break. This is an internal gap, not merely missing code. Without a stated test-time protocol, the reported margins over SpA GAN (30.232/0.954 and 28.368/0.906) could be artifacts of resolution mismatch or upsampling rather than evidence for the proposed transfer-learning architecture.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a cloud-removal method for remote sensing images that fine-tunes a pre-trained Masked Autoencoder (MAE) ViT-large model with a patch-wise GAN discriminator. The generator is initialized from an ImageNet-pretrained MAE with GAN loss, then fine-tuned on the RICE1/RICE2 datasets using a combination of MSE and patch-level GAN losses with layer-wise learning-rate decay. Experiments report PSNR/SSIM values in Table I showing large margins over Conditional GAN, Cycle GAN, and SpA GAN baselines taken from reference [19], and SSIM comparisons in Table II against several state-of-the-art methods, while acknowledging that the authors of those methods did not disclose their train/test splits. The paper claims significant improvements in cloud-removal performance and competitive results against state-of-the-art methods.","tokens_in":6465,"tokens_out":3377,"duration_ms":33046,"significance":"If the quantitative claims are reproducible and the comparisons are fair, the paper would make a useful contribution by demonstrating how a natural-image reconstruction model can be transferred to cloud removal with a patch-wise GAN objective. The idea is simple and potentially of interest to the remote sensing community, and the paper is honest about the limitations of the Table II comparison. However, the central evidence is currently not self-contained: the test-time input resolution is unspecified, the Table I baselines are copied from another paper without verification of protocol, and no variance or multiple-run information is provided. The multi-spectral language in the abstract and introduction is also not supported by the RGB experiments. These gaps prevent the paper from being accepted in its current form.","major_comments":[{"comment":"The generator is described as a ViT that operates on 196 patches of 16×16 pixels, i.e., a fixed 224×224 input, and during training images are randomly cropped to 224×224. RICE1 and RICE2 images are 512×512. The paper never states how 512×512 test images are fed to the model: whether they are resized to 224×224, center-cropped, or tiled and stitched. This directly affects the PSNR/SSIM values in Table I and their comparability with baselines from [19]. If test images are resized, the metrics are computed at reduced resolution and compared with baselines evaluated at native resolution; if center-cropped, only a subregion is evaluated; if tiled, the stitching procedure is missing. Please specify the exact test-time protocol and, ideally, rerun the baselines under the same protocol.","section":"Section III-A and Section IV-A"},{"comment":"The baseline PSNR and SSIM values are taken directly from [19] with no verification that [19] used the same train/test split, preprocessing, or metric calculation on RICE1 and RICE2. The paper's own caveat about undisclosed data splits for the methods in Table II applies equally to the GAN baselines in Table I. No standard deviations or multiple-run results are reported for the proposed method, so the margins over SpA GAN (PSNR 33.659 vs 30.232 on RICE1, 34.056 vs 28.368 on RICE2) may be artifacts of protocol mismatch or evaluation variance. Please report the proposed method's results on the exact split used by [19], and provide error bars or at least confirm that the reported numbers are stable across runs.","section":"Table I and Section IV-C"},{"comment":"The paper repeatedly describes the input as 'multi-spectral remote sensing images,' but the discriminator is described as flattening 'RGB channels' (Section III-B, 16×16×3 = 768 values per patch) and the RICE dataset used in the experiments is a color-image dataset. No multi-spectral experiment is performed, so the claim that the method extracts information from multi-spectral imagery is unsupported. Please either restrict the claims to RGB inputs or add an experiment on a multi-spectral dataset (e.g., Sentinel-2 or Landsat imagery with spectral bands beyond RGB).","section":"Abstract and Section III-B"},{"comment":"The paper explicitly states that the authors of CMNet [20] and other methods 'did not disclose critical details about their dataset configurations, such as the specific splits used for training, testing, and validation.' Given this acknowledged protocol mismatch, the conclusion that the proposed method 'demonstrates competitive results' is stronger than the evidence supports. Please either obtain the exact evaluation protocols and splits from the authors of [20] and rerun the comparison, or soften the conclusion to state that the proposed method achieves favorable SSIM values on the paper's own splits under a best-effort comparison.","section":"Table II and Section V (Conclusion)"}],"minor_comments":[{"comment":"There is a typo in Equation (1): 'gks' should be 'g_k', and 'weight and height' should be 'width and height'. The summation indexing is also confusing; please clarify that the sum runs over all C×W×H pixels.","section":"Equation (1)"},{"comment":"The layer-wise learning rate decay schedule is described qualitatively but no concrete values (base learning rate, decay factor, number of layers) are given. For reproducibility, please include the exact schedule used in the experiments.","section":"Section III-C"},{"comment":"Reporting only single PSNR/SSIM values without standard deviations or the number of test images makes it impossible to assess the significance of the reported margins. Please add error bars or variance information.","section":"Table I"},{"comment":"The paper states the experiments ran on an NVIDIA 3060 GPU without giving training time or model size; providing these would help readers assess the practical cost of the method.","section":"Section IV-B"},{"comment":"The sentence 'Following [19], we used 1/5 training data for validation during the training phase' is ambiguous: it is unclear whether the validation split is taken from the training set (leaving fewer than 400/588 training pairs) or whether an additional validation set is used. Please clarify the exact split sizes.","section":"Section IV-A"}],"recommendation":"major_revision","confidential_remarks":"The test-time resolution issue is the single most important technical gap; if the authors can specify how 512×512 images are processed and confirm that baselines from [19] were evaluated under comparable conditions, the central claim may become solid. The paper's self-reported limitations in Table II are commendable but should be consistently reflected in the conclusions. The absence of code or detailed hyperparameters further hampers reproducibility; consider requesting them as part of the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a plausible idea — fine-tune an ImageNet-pretrained MAE inside a patch-based GAN for cloud removal — and reports large margins over GAN baselines on RICE1/RICE2. The combination of MAE, transfer learning, and a patch discriminator is natural but, as far as I know, new. The paper is readable and the idea is worth taking seriously. The problem is that the central numbers may not mean what they appear to mean.\n\nThe main soft spot is one the authors never address: the ViT generator expects 224×224 inputs (196 16×16 patches), and training crops to 224×224. RICE test images are 512×512. The paper never says how test images are fed in — resized, cropped, tiled? If resized, PSNR/SSIM are computed on lower-resolution reconstructions and compared to baselines evaluated at native resolution; that is not apples-to-apples. If fed at 512, the patch and positional-embedding assumptions break. This is internal to the paper, not just missing code, and it undermines Table I. The stress-test note is right to flag it.\n\nOther soft spots are real but less severe: baselines are copied from [19] rather than re-run, there are no error bars or multiple seeds, and the abstract's \"multi-spectral\" framing doesn't match the RGB-only implementation. Those are reporting issues, not fatal flaws.\n\nWhat the paper does well: it identifies an apt analogy between MAE reconstruction and cloud removal, and the patch discriminator is a sensible match for remote sensing imagery. The reported gains are large enough that the idea is worth testing. There is no circularity — the results are measured against external data.\n\nBottom line: this is a credible engineering paper with a potentially useful contribution, but the key comparison needs clarification and ideally a re-run with matched protocols. A serious referee should see it. If I were editing, I'd send it to review with a request to specify the test-time pipeline and to include their own baseline runs or at least an explicit statement of resolution handling.\n\nRecommendation: engage with it, but require the test-time protocol question answered before trusting the numbers.","headline":"Plausible idea with large reported gains, but the test-time resolution gap between the 224×224 ViT generator and 512×512 RICE images makes the main comparison unverified.","tokens_in":6998,"tokens_out":1640,"would_cite":false,"duration_ms":16094,"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 masked autoencoder pre-trained on natural images, fine-tuned with a patch-wise GAN, removes clouds from remote sensing images and beats existing GAN-based methods on the RICE benchmarks.","keywords":["cloud removal","masked autoencoder","patch-GAN","transfer learning","remote sensing","image reconstruction"],"falsifier":"Rerun Conditional GAN, Cycle GAN, and SpA GAN on the same RICE1 and RICE2 splits, crops, and metric code used by the proposed method, then compare PSNR and SSIM. If the reported margins shrink or disappear, the central claim of improvement would be shown to rest on protocol mismatch instead of the model itself.","tokens_in":6014,"feed_emoji":"☁️","tokens_out":7520,"duration_ms":62367,"temperature":0.7,"pith_summary":"This paper claims that a masked autoencoder trained on natural images can be transferred to the task of removing clouds from satellite imagery. The approach fine-tunes an ImageNet-pretrained MAE as the generator of a generative adversarial network, with a patch-wise discriminator that judges each 16×16 patch separately. On the RICE1 and RICE2 benchmarks the authors report PSNR values of 33.66 and 34.06 and SSIM values of 0.976 and 0.955, above the GAN baselines they compare with. The paper also cautions that comparisons with some state-of-the-art methods rest on unclear train/test split details, so those numbers are less certain.","feed_headline":"Fine-tuned masked autoencoder outperforms cloud-removal GANs","feed_subtitle":"A natural-image pretrained model transfers to satellite imagery, topping GAN baselines on RICE1/RICE2.","key_machinery":"The load-bearing component is the masked autoencoder (MAE), a vision-transformer autoencoder that reconstructs a full image from randomly masked patches; here it serves as the generator of a GAN and is fine-tuned from an ImageNet-pretrained checkpoint. Input images are cropped to 224×224 and divided into 196 patches of 16×16 pixels, and the decoder outputs cloud-free patches. The second component is a patch-wise discriminator: a fully connected network that flattens each 16×16 RGB patch into a 768-dimensional vector and outputs a real/fake score per patch. The training objective is the sum of MSE between the prediction and the cloud-free ground truth and a patch-based GAN loss, with layer-wise learning rate decay to keep early general features intact while adapting later layers to the task.","core_discovery":"The central claim is that reconstructive pretraining transfers to cloud removal. An MAE, whose encoder is a vision transformer operating on visible patches and whose decoder fills masked ones, is fine-tuned with a patch-wise GAN loss so that the cloud-obscured areas are treated as masks to reconstruct. The paper reports performance on RICE1 and RICE2: PSNR 33.659 and 34.056, SSIM 0.976 and 0.955, exceeding the Conditional GAN, Cycle GAN, and SpA GAN baselines, and SSIM scores competitive with or above published state-of-the-art methods such as CMNet.","pith_inferences":["If the transfer works as claimed, MAE-style masked reconstruction could become a generic pretraining backbone for other structured-occlusion tasks in remote sensing, such as shadow or haze removal, because the encoder already learns to reason about missing patches.","The paper's own caveat about unclear train/test splits in the state-of-the-art comparison suggests the SSIM rankings in Table II could shift when all methods are evaluated under one shared protocol.","A natural testable extension is to swap the fully connected patch discriminator for a convolutional or transformer-based discriminator and measure whether local realism and PSNR improve further.","Because training crops are 224×224 while RICE images are 512×512, full-resolution behavior is untested; evaluating on full images could show whether the patch-level realism generalizes beyond the crop size."],"forward_implications":["The method's reported PSNR and SSIM on RICE1 and RICE2 beat the three GAN baselines it is directly compared with, indicating a transferable advantage from reconstructive pretraining.","The patch-wise discriminator, instead of an image-level one, allows the adversarial loss to act on local 16×16 regions, which is how the paper handles the spatial complexity of remote sensing scenes.","Because the pre-trained MAE checkpoint is public and the fine-tuning is light enough for a single NVIDIA 3060 GPU, the approach can be reproduced without large compute budgets.","The SSIM values on RICE1 (0.976) and RICE2 (0.955) place the method at or above the published state-of-the-art values compared in Table II, though those comparisons are qualified by unclear data splits in prior work."],"supporting_citations":[{"why":"Supplies the masked autoencoder architecture and the ImageNet-pretrained reconstruction model that the paper fine-tunes as its generator.","marker":"[5]"},{"why":"Provides the vision transformer design used for both the MAE encoder and decoder.","marker":"[14]"},{"why":"Provides the RICE1 and RICE2 cloud-removal datasets used for all training and evaluation.","marker":"[18]"},{"why":"Supplies the baseline GAN results that the paper compares against and the train/validation split protocol it follows.","marker":"[19]"},{"why":"Provides the CMNet SSIM values used for the state-of-the-art comparison in Table II.","marker":"[20]"}],"fun_headline_variants":["MAE transfer learning beats GANs at cloud removal","Patch-GAN + MAE: cloud removal without the clouds","Fine-tuned MAE clears clouds better than GAN baselines","Reconstructive transfer learning for cloud removal beats GANs","Masked autoencoder transfer outperforms GANs on RICE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed superiority over other GAN-based methods assumes that the baseline PSNR and SSIM values taken from [19] were measured under the same train/test splits, preprocessing, and metric implementation as the proposed method's numbers.","fun_headline_variants_meta":{"raw":{"variants":["MAE transfer learning beats GANs at cloud removal","Patch-GAN + MAE: cloud removal without the clouds","Fine-tuned MAE clears clouds better than GAN baselines","Reconstructive transfer learning for cloud removal beats GANs","Masked autoencoder transfer outperforms GANs on RICE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000767,"raw_usage":{"total_tokens":3359,"prompt_tokens":862,"completion_tokens":2497,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":478,"completion_tokens_details":{"reasoning_tokens":2410}},"tokens_in":478,"tokens_out":2497,"duration_ms":16881,"temperature":1.0,"reasoning_tokens":2410,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:14:01.976911+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun Conditional GAN, Cycle GAN, and SpA GAN on the same RICE1 and RICE2 splits, crops, and metric code used by the proposed method, then compare PSNR and SSIM. If the reported margins shrink or disappear, the central claim of improvement would be shown to rest on protocol mismatch instead of the model itself.","supporting_citations":[{"cited_title":"Masked autoencoders are scalable vision learners,","cited_arxiv_id":null,"evidence_quote":"Supplies the masked autoencoder architecture and the ImageNet-pretrained reconstruction model that the paper fine-tunes as its generator."},{"cited_title":"Cascaded memory network for opti- cal remote sensing imagery cloud removal,","cited_arxiv_id":null,"evidence_quote":"Provides the CMNet SSIM values used for the state-of-the-art comparison in Table II."}],"review_version":1}