{"id":"02c6d108-980d-414c-9308-b5fd27ccc6a2","arxiv_id":"2502.00784","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The authors propose MSwin-Pix2Pix, a Swin-Transformer-based Pix2Pix variant with mask and median filter, to estimate forest carbon stocks from multi-temporal satellite imagery, reporting R2 of 0.71 and an increasing carbon trend.","lead":"This paper trains a deep learning model to estimate forest carbon stocks from satellite images of a county in China, using style transfer to harmonize images from different years and sensors. It reports improved accuracy over older methods and an overall carbon increase in the study area from 2005 to 2020.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 2005–2020 trend is unsupported: the style-transfer step that maps historical imagery into the 2020 domain is never validated against any historical ground truth, so the reported 44.04% increase is not established.","rationale":"The reader's weakest_assumption identified exactly the same load-bearing issue: the 2005–2020 trend depends on style transfer preserving carbon-stock-relevant features across domains, and this is never validated against historical ground truth. My reading agrees that this is the central weakness because the trend claim is the most policy-relevant and least supported part of the paper. The paper's other issues, such as the possible feature-selection leakage and the unit-mixing in literature comparisons, further weaken the headline accuracy, but they are secondary to the trend claim. I considered whether the feature-selection leakage is more load-bearing: it directly undermines the reported 2020 accuracy, which is also a central claim. However, even a perfectly accurate 2020 model would not justify the historical trend without transferability evidence, so the unvalidated transfer assumption is the single most decisive gap. The proposed concrete test is feasible without historical field data: by simulating the very domain shift the style-transfer step is meant to correct and checking whether carbon-stock predictions survive, the authors can empirically establish whether their transfer step preserves the signal needed for the trend analysis. Until such a test is run, or historical inventory data are compared, the paper's central trend conclusion should not be accepted. The reader's REJECT verdict is therefore unchanged, and the underlying idea of Swin-Transformer-based style transfer for multi-sensor carbon mapping remains worth revisiting with corrected validation.","tokens_in":17183,"tokens_out":3282,"duration_ms":35350,"concrete_test":"Run a controlled transferability experiment on the 2020 data: apply a simulated Landsat TM degradation (30 m resampling and spectral band matching) to the 2020 GF-1 WFV image, pass the degraded image through Swin-Pix2Pix back into the 2020 domain, then evaluate MSwin-Pix2Pix against the original 2020 field plots. If MAE or RMSE degrades materially relative to the reported 16.29/29.38, the style-transfer step does not preserve carbon-stock-relevant signal, and the historical trend cannot be trusted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is the multi-year trend claim. MSwin-Pix2Pix is trained only on 2020 field data; for 2005, 2013, and 2015, the historical imagery is first translated into the 2020 GF-1 WFV style using Swin-Pix2Pix (§3.1), and the 2020-trained estimator is then applied to these translated images. This requires the style transfer to preserve exactly the spectral and textural features that predict carbon stock, yet §3.1 provides only visual examples and no quantitative fidelity metric, and no comparison against any historical field measurements is made. Because the style-transfer ground truth in Figure 5 is the 2020 image itself, the network is trained to make historical images resemble 2020, which can erase genuine temporal change in forest structure or reflectance. The trend percentages in §3.3 therefore rest on an unvalidated transferability assumption. A second, independent flaw is that feature screening in §2.3.1 selects the three GLCM-Mean bands using correlations with measured carbon stock across the full dataset, seemingly before the five-fold cross-validation described in §3.2; if feature selection is not nested inside the folds, the reported MAE, RMSE, R2, and SSIM are optimistically biased. Both issues are real, but the trend claim fails even if the 2020 accuracy were accepted as reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a deep-learning pipeline for estimating forest carbon stocks in Huize County, China, from multispectral satellite imagery (GF-1 WFV and Landsat TM). The authors first use a Swin-Transformer-based Pix2Pix variant (Swin-Pix2Pix) for style transfer to harmonize multi-temporal images into a 2020 GF-1 WFV reference domain, then apply a masked SUNet-based regression model (MSwin-Pix2Pix) to map the harmonized images to carbon stock values. The reported results include a best 2020 accuracy of MAE = 16.2891, RMSE = 29.3763, R2 = 0.7105, SSIM = 0.7510, and a claimed 2005–2020 trend of 44.04% of the area increasing, 10.22% decreasing, and 45.74% unchanged in carbon stock. The central claims are (i) that MSwin-Pix2Pix outperforms several statistical, machine-learning, and deep-learning baselines on the 2020 data, and (ii) that the style transfer enables reliable multi-temporal carbon-stock trend analysis without historical ground truth.","tokens_in":17467,"tokens_out":4618,"duration_ms":46412,"significance":"If validated, the 2020 carbon-stock estimation result would be a useful addition to the growing literature on optical-image-based forest carbon mapping, and the cross-domain style-transfer idea is conceptually appealing for dealing with multi-sensor, multi-temporal imagery. The paper includes a fairly broad comparison with OLS, GWR, RF, SVR, CNN, and Pix2Pix, and the ablation study for the proposed modules is detailed. However, the paper's most advertised contribution is the 2005–2020 trend analysis, and that claim is not supported by the presented evidence. The feature-selection procedure also raises a leakage concern that affects even the 2020 accuracy numbers. These issues make the central claims unreliable as they stand.","major_comments":[{"comment":"The 2005–2020 trend percentages rest on an unvalidated transferability assumption. The style-transfer model Swin-Pix2Pix is trained with the 2020 GF-1 WFV image as ground truth, and the 2020-trained MSwin-Pix2Pix estimator is then applied to the translated 2005, 2013, and 2015 images. No quantitative fidelity metric is reported for the style transfer, and no comparison against historical field measurements is made. Because the style-transfer objective is to make historical images resemble the 2020 domain, genuine temporal changes in forest structure, reflectance, or carbon stock can be erased by the translation itself. The trend percentages in Section 3.3 therefore cannot be interpreted as evidence of an actual carbon-stock increase. A concrete remedy would be to validate the pipeline against any historical inventory data (e.g., the 2005 standing-volume statistics cited in Section 3.3) or to demonstrate on held-out temporal data that the style transfer preserves carbon-stock-relevant features.","section":"Section 3.1 and Section 3.3"},{"comment":"The feature-screening step selects the three GLCM-Mean bands (19, 35, 27) using correlations with measured carbon stock computed across the full dataset, but the five-fold cross-validation in Section 3.2 is described only after this selection. If the band selection is not nested inside the cross-validation folds, the test-fold carbon-stock labels have influenced the choice of predictor bands, which biases the reported MAE, RMSE, R2, and SSIM optimistically. The feature selection must be performed independently on each training fold, or the band set must be fixed a priori without access to the test labels. Without this correction, Table 4 does not provide an unbiased estimate of generalization performance.","section":"Section 2.3.1 and Section 3.2"},{"comment":"The spatial-resolution mismatch between the 2005 Landsat TM (30 m) and the 2020 GF-1 WFV (16 m) images is not addressed in the style-transfer description. The paper says only that the four common bands are selected and that Swin-Pix2Pix is used for style transfer; it is not stated whether the 30 m images are resampled to 16 m before translation or whether the 2020-trained model is applied to 30 m imagery directly. If the resolution remains 30 m, the feature scale presented to the 2020-trained estimator is inconsistent, and the trend analysis in Section 3.3 is further compromised.","section":"Section 2.2 and Section 3.1"}],"minor_comments":[{"comment":"Equation numbering restarts at (1) in both Section 2.3.5 and Section 2.3.6, which makes cross-references confusing; the equations should be numbered continuously throughout the paper.","section":"Section 2.3.5 and Section 2.3.6"},{"comment":"The term 'high-resolution' is used for 16 m and 30 m imagery; these are more commonly called medium resolution in the remote-sensing literature, and the wording should be adjusted to avoid overstatement.","section":"Abstract and Introduction"},{"comment":"The R2 values for the L2-only and SmoothL1-only rows (0.4197 and 0.3609) are notably low relative to their MAE and RMSE values; the authors should verify these numbers or provide an explanation, as the inconsistency is distracting.","section":"Table 3"},{"comment":"The NDVI mask threshold is computed from the 2020 image (M = μ − 2σ) and then applied to the other years without re-estimation; sensor and illumination differences between Landsat TM and GF-1 WFV may make a fixed threshold inappropriate across years.","section":"Section 2.3.1"},{"comment":"The style-transfer results are evaluated only visually (Figure 5); reporting a quantitative metric such as FID, SSIM, or a vegetation-index-preservation statistic would strengthen the claim that Swin-Pix2Pix outperforms Pix2Pix.","section":"Section 3.1"}],"recommendation":"reject","confidential_remarks":"The two main concerns — unvalidated temporal transfer and feature-selection leakage — are load-bearing for the paper's central claims. The trend analysis cannot be repaired without historical ground truth, which the authors appear not to have, and the accuracy comparison is not trustworthy without nested feature selection. I therefore recommend rejection rather than major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — my quick read:\n\nThe paper is a workmanlike application of Swin-Transformer style transfer to forest carbon stock mapping, and the 2020 estimation results are plausible. The real problem is the headline 2005–2020 trend. The style transfer is trained with 2020 as ground truth, then the 2020-trained estimator is applied to translated historical images. No historical field data are used to check that the transfer preserves the carbon-relevant features. So the 44.04% increase / 10.22% decrease numbers are simply outputs of an unvalidated pipeline. That conclusion (ecological improvement) is the paper's boldest claim, and it does not hold up as stated.\n\nWhat's good: the ablation table is internally consistent and informative. Comparing OLS, GWR, RF, SVR, CNN, Pix2Pix, and Swin-Pix2Pix with/without mask and median filter shows where the gains come from. The architecture descriptions are explicit enough to reproduce. The idea of using GAN-based style transfer to harmonize multi-sensor, multi-date imagery before regression is practical and worth building on.\n\nSoft spots, in order:\n1. The temporal trend is unsupported (above).\n2. Feature screening (§2.3.1) selects bands using full-data carbon correlations, and the paper never says this is nested inside the five-fold CV. If it's not, the reported MAE/RMSE/R2 are optimistically biased. This is a legitimate concern, not a nitpick.\n3. The literature comparison mixes biomass and carbon units; 'strong correlation' does not make them interchangeable.\nMinor: no code/data release, so independent verification is hard.\n\nFor a remote sensing audience, this is a useful case study in domain adaptation for carbon mapping, but it needs major revision. I'd send it to review because the method and ablation are worth engaging, but I'd expect the trend claim to be removed or reworked with actual historical validation.","headline":"A useful ablation study with an unvalidated multi-year trend and a possible feature-selection leak; the 2020 results are plausible but the headline conclusion is not.","tokens_in":18011,"tokens_out":3975,"would_cite":false,"duration_ms":46681,"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 Swin-Transformer-based style transfer model estimates forest carbon stock more accurately than existing regression methods.","keywords":["carbon stock","forest","style transfer","remote sensing","deep learning","Swin Transformer","Pix2Pix","domain shift"],"falsifier":"Measure carbon stock from historical field plots in the study area for 2005, 2013, or 2015 and compare them to the style-transfer model's maps for those years; if the errors are much larger than the 2020 MAE and RMSE, the transfer does not preserve carbon-relevant features.","tokens_in":16959,"feed_emoji":"🌲","tokens_out":4448,"duration_ms":38551,"temperature":0.7,"pith_summary":"The paper tries to show that forest carbon stock can be estimated more accurately from high-resolution satellite images if the images are first put into a common visual domain by style transfer. It proposes a model family based on Swin Transformer and Pix2Pix, converting carbon stock estimation into an image-to-image translation task. On a mountainous county in Yunnan, China, the proposed MSwin-Pix2Pix model achieves the best accuracy among tested statistical, machine-learning, and deep-learning baselines, and it is used to map carbon stock changes from 2005 to 2020. The central practical payoff would be a way to monitor forest carbon over long time periods using only one field campaign, by translating older or cloudy images into the style of a reference year.","feed_headline":"Style transfer sharpens forest carbon stock estimates","feed_subtitle":"A Swin-Transformer model maps carbon gains from 2005 to 2020 by stylizing old satellite images to match new ones.","key_machinery":"The load-bearing mechanism is the style-transfer step: Swin-Pix2Pix, a conditional GAN whose generator is a Swin Transformer U-Net, translates source-domain images (Landsat TM 2005, GF-1 WFV 2013 and 2015) into the target domain of the 2020 GF-1 WFV image by matching global features via shifted-window self-attention. For estimation, the same generator backbone is extended with an NDVI-derived mask to exclude non-forest areas and a median filter to remove high-intensity anomalies, producing the MSwin-Pix2Pix model.","core_discovery":"The central claim is that replacing the generator of Pix2Pix with a Swin-Transformer-based U-Net, adding a vegetation mask and a median filter, yields a model whose carbon stock estimates are accurate enough for regional monitoring (MAE = 16.2891, RMSE = 29.3763, R2 = 0.7105, SSIM = 0.7510) and that the same style-transfer step aligns images from different years and sensors so that a model trained on 2020 field data can be applied to 2005, 2013, and 2015 imagery. Applying this pipeline, the paper reports that from 2005 to 2020 carbon stock increased in 44.04% of the study area, decreased in 10.22%, and stayed unchanged in 45.74%, implying an overall rise.","pith_inferences":["A direct test of the transfer assumption would be to acquire a small set of historical field plots and compare the model's 2005 and 2013 estimates against them; the paper does not report such a validation.","Because the mask is based on NDVI of the reference year, areas that changed from non-vegetation to vegetation between 2005 and 2020 may be systematically included or excluded depending on which year's mask is used, and this could bias the reported increase.","The method's reliance on a single reference domain suggests it could be extended to other regions by training on one contemporary image and field campaign, but sensor differences beyond Landsat and GF-1 would need separate testing.","The median filter removes pixels above intensity 240; if large carbon stock values genuinely occur in those bright pixels, the filter could truncate real extremes rather than only noise."],"forward_implications":["If the style-transfer step works as claimed, multi-year carbon stock mapping no longer requires ground truth for every year; one field campaign can anchor a time series.","The reported accuracy gaps (MSwin-Pix2Pix MAE 16.29 vs Pix2Pix 21.12, RF 52.98) suggest that deep global-feature extraction and targeted post-filtering matter more than traditional spectral regression.","Combining mask and median filter improved accuracy more than either alone, pointing to boundary handling as a key remaining source of error.","The reported trend, with 44.04% of the area gaining carbon and 45.74% unchanged, is the kind of result that could inform regional carbon accounting and ecological policy.","De-clouding through style transfer extends usable optical imagery in persistently cloudy highland regions."],"supporting_citations":[{"why":"Supplies the generative adversarial network framework that the style-transfer and estimation models build on.","marker":"Goodfellow et al. (2014)"},{"why":"Supplies the Pix2Pix conditional GAN with L1 loss and PatchGAN discriminator that the proposed model adapts.","marker":"Isola et al. (2017)"},{"why":"Supplies the Swin Transformer with shifted-window self-attention, the backbone used for global feature extraction.","marker":"Liu et al. (2021)"},{"why":"Supplies the Transformer attention mechanism that Swin Transformer and the paper's architecture rely on.","marker":"Vaswani et al. (2017)"},{"why":"Supplies the SUNet U-Net structure with Swin blocks and dual up-sampling that the paper extends into Mask SUNet.","marker":"Fan et al. (2022)"},{"why":"Supplies the Swin-Unet pure-Transformer U-Net and patch-merging design used for down-sampling.","marker":"Cao et al. (2021)"},{"why":"Provides the coarse-resolution biomass carbon map baseline with RMSE around 30 that the paper compares against.","marker":"Chen et al. (2023)"}],"fun_headline_variants":["Style transfer boosts forest carbon mapping accuracy","Swin Transformer and style transfer refine carbon stock maps","Reducing domain shift sharpens forest carbon estimates","Style transfer aligns images to track forest carbon","Carbon stock maps improved by style-transfer domain adaptation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole multi-year trend rests on the assumption that translating historical images into the 2020 reference style preserves exactly the spectral and textural features that predict carbon stock, so a model trained only on 2020 field data remains valid for 2005, 2013, and 2015 without any historical ground truth to check it.","fun_headline_variants_meta":{"raw":{"variants":["Style transfer boosts forest carbon mapping accuracy","Swin Transformer and style transfer refine carbon stock maps","Reducing domain shift sharpens forest carbon estimates","Style transfer aligns images to track forest carbon","Carbon stock maps improved by style-transfer domain adaptation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000404,"raw_usage":{"total_tokens":2041,"prompt_tokens":821,"completion_tokens":1220,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":1151}},"tokens_in":437,"tokens_out":1220,"duration_ms":9601,"temperature":1.0,"reasoning_tokens":1151,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T17:45:01.701561+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure carbon stock from historical field plots in the study area for 2005, 2013, or 2015 and compare them to the style-transfer model's maps for those years; if the errors are much larger than the 2020 MAE and RMSE, the transfer does not preserve carbon-relevant features.","supporting_citations":[{"cited_title":"M., Crowther, T","cited_arxiv_id":null,"evidence_quote":"Provides the coarse-resolution biomass carbon map baseline with RMSE around 30 that the paper compares against."}],"review_version":1}