{"id":"b459ed47-34f2-455d-a831-88866941196a","arxiv_id":"2509.00835","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A Swin Transformer + U-Net hybrid with a watershed loss achieves slightly higher PSNR/SSIM than prior dehazing methods on the RICE and SateHaze1k benchmarks.","lead":"This paper proposes a hybrid neural network that combines Swin Transformer and U-Net blocks plus a watershed-based loss for removing haze from satellite images. On two standard datasets it reports modest gains in PSNR and SSIM over prior methods, though key implementation details are missing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 contradicts the abstract's unqualified 'outperforms state-of-the-art' claim: on SateHaze1K, SUFERNOBWA is not best on PSNR in Thin/Thick haze and not best on SSIM in Moderate haze.","rationale":"The reader's weakest assumption concerned non-differentiability of the watershed loss, which is a valid and serious issue: Eq. (12) is computed from label maps produced by local-minima detection, integer marker assignment, and four-directional region growing, all non-differentiable or zero-gradient almost everywhere, so the watershed term cannot provide a gradient signal unless a surrogate is described. That would invalidate the ablation attribution in Table 3. However, I find a more direct and decisive problem in the paper's own empirical results: the SateHaze1K table contradicts the central claim regardless of gradient issues. The abstract and conclusion say the method outperforms state-of-the-art on both datasets, but Table 2 shows SUFERNOBWA is not the best on PSNR in Thin or Thick haze, and not best on SSIM in Moderate haze. The conclusion even asserts highest PSNR and SSIM in Moderate, while Table 2 shows UDAVM-Net's SSIM is higher. This is an internal inconsistency that cannot be fixed by code release alone. The paper's Section 5 limitation also concedes that only synthetic datasets were used and real-world validation was not performed, so claims of practical applicability are further qualified. Because the central claim fails on the paper's own reported numbers, the existing REJECT verdict is appropriate; no verdict change is needed.","tokens_in":22779,"tokens_out":10359,"duration_ms":119529,"concrete_test":"Perform a per-condition, per-metric ranking of Table 2: for each of Thin/Moderate/Thick, count how many baselines have strictly higher PSNR and higher SSIM than Ours. If any non-Ours entry exceeds Ours on either metric in any condition — which the table shows for Thin PSNR, Thick PSNR, and Moderate SSIM — then the unqualified claim 'outperforms state-of-the-art' on SateHaze1K must be revised or the definition of 'outperforms' explicitly changed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, stated in both abstract and conclusion, is that SUFERNOBWA 'outperforms state-of-the-art models on both the RICE and SateHaze1K datasets.' Section 4.3's Table 2 directly undermines this for SateHaze1K. Under Thin Fog, Ours has PSNR 24.19 dB, below X. Chen et al. (25.84), M2SCN (25.21), DVKT (24.73), and UDAVM-Net (26.76). Under Thick Fog, Ours has PSNR 22.33 dB, below X. Chen et al. (25.20), B. Huang et al. (25.07), DVKT (23.31), and UDAVM-Net (23.48). Under Moderate Fog, Ours achieves the highest PSNR (28.15 dB) but its SSIM (0.950) is lower than UDAVM-Net's 0.952. Thus no haze level has SUFERNOBWA best on both metrics. The conclusion's statement that the method achieves 'the highest PSNR and SSIM in the Moderate Fog region' is contradicted by the table. This is an internal-evidence problem: no code, reimplementation, or differentiability analysis is needed to see that the headline claim is unsupported as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes SUFERNOBWA, an encoder–bottleneck–decoder network for single-image dehazing of satellite imagery. The encoder and decoder use SwinRRDB blocks that combine Swin Transformer windowed self-attention with RRDB-style residual dense connections, and the bottleneck uses lightweight Bottleneck blocks without self-attention. Training uses a composite loss (Eq. 13) comprising L2 loss, guided-filter loss, and a novel watershed loss based on watershed label maps of predicted and ground-truth images. The method is evaluated on RICE and SateHaze1K with PSNR/SSIM, including comparisons with several recent dehazing methods and ablation studies for the loss components and SwinRRDB.","tokens_in":23180,"tokens_out":8713,"duration_ms":101711,"significance":"If the results held, the architectural combination would be a modest incremental contribution to satellite-image dehazing, and the watershed-loss idea is interesting in principle. The paper provides quantitative comparisons on two datasets, a visual comparison, and an ablation study, and it explicitly acknowledges limitations such as the use of synthetic data and the lack of real-world validation. However, the central claimed contribution is not currently supported: the watershed loss is described through non-differentiable operations without a surrogate gradient, the ablation text contradicts the reported numbers, and the headline 'outperforms state-of-the-art on both datasets' is contradicted by the paper's own Table 2. The significance of the work is therefore contingent on substantial revision and re-evaluation.","major_comments":[{"comment":"The watershed loss is computed from label maps obtained by (i) Gaussian smoothing, (ii) local-minima detection and integer marker assignment, (iii) four-directional label propagation, and (iv) min-max normalization. Steps (ii)–(iii) are discrete assignment operations, so the loss is piecewise constant almost everywhere with respect to the predicted image; its gradient with respect to network parameters is zero (or undefined at boundaries). No surrogate gradient, straight-through estimator, or detached-target formulation is described in §3.2 or §4.2. Consequently, the training procedure as written cannot credit L_water for the improvements in Table 3. Because this loss is one of the paper's two main contributions, the claim that watershed loss improves boundary preservation is unsubstantiated as stated.","section":"§3.2, Eq. (12)"},{"comment":"The unqualified claim that SUFERNOBWA 'outperforms state-of-the-art models on both the RICE and SateHaze1K datasets' is contradicted by Table 2. Under Thin Fog, Ours has PSNR 24.19 dB, below X. Chen et al. (25.84), M2SCN (25.21), DVKT (24.73), and UDAVM-Net (26.76). Under Thick Fog, Ours has PSNR 22.33 dB, below X. Chen et al. (25.20), B. Huang et al. (25.07), DVKT (23.31), and UDAVM-Net (23.48). Under Moderate Fog, Ours has the highest PSNR (28.15 dB), but its SSIM of 0.950 is below UDAVM-Net's 0.952. The conclusion's statement that the method achieves 'the highest PSNR and SSIM in the Moderate Fog region' is false as written. The claims must be revised to reflect the per-condition results.","section":"Abstract and §5 vs Table 2"},{"comment":"The ablation narrative contradicts the table. The L2+Watershed row (O X O) reports PSNR 32.28 dB, which is lower than the L2-only baseline's 32.71 dB, yet the text states that 'Adding Watershed Loss significantly improved performance, increasing PSNR to 32.28 dB.' Similarly, L2+Guided (32.61 dB) also does not improve PSNR over L2-only; only the three-loss combination reaches 33.24 dB. Thus the ablations do not support the claim that adding guided or watershed loss individually improves pixel accuracy; at most they improve SSIM/UQI. The text needs to be corrected and the interaction effects analyzed, or the experiments need to be rerun with properly controlled comparisons.","section":"§4.4, Table 3"},{"comment":"All reported results appear to come from a single training run with no error bars, multiple seeds, or significance tests. The claimed RICE advantage over RSDformer is 0.23 dB PSNR and 0.014 SSIM; these margins are plausibly within run-to-run variation for deep image-restoration methods. The narrative also selects whichever metric favors the method in each haze condition (PSNR for Moderate, SSIM for Thin/Thick), which further weakens the superiority claim. The authors should report mean±std across multiple runs, or at minimum explicitly state that this is a single-run comparison and temper the comparative claims accordingly.","section":"§4.3, Tables 1–2"}],"minor_comments":[{"comment":"The citation numbering is inconsistent. For example, ICL-Net is cited as [14] in §2 but reference [14] is a different paper, and C. Li et al.'s efficient dehazing method is cited as [38] while reference [38] in the list is the Swin Transformer paper. Please renumber all references and verify each citation against the bibliography.","section":"References"},{"comment":"The guided-filter formula is ambiguous: products such as I_t·I_GT are not defined with local-window means, and the replacement t ∈ {GT, pred} is confusing. The standard guided-filter expression should be written with explicit windowed mean/covariance notation.","section":"Eq. (9)"},{"comment":"Figure 2 and its caption describe an L1 difference between 'input and output' images and a sum of remaining label values, but Eq. (12) defines L2 loss between normalized watershed label maps of predicted and ground-truth images. The figure, caption, and equation need to be aligned.","section":"Fig. 2 and Eq. (12)"},{"comment":"Subscript conventions are inconsistent: Eq. (2) indexes F_l for l ∈ {1,2,3,4}, while Eq. (3) uses F_{l-1} and F_l in ways that do not clearly match the encoder/decoder hierarchy. The text also says 'three consecutive SwinRRDB' but the equation suggests four layers. Please clarify.","section":"Eqs. (2)–(3)"},{"comment":"There are numerous typos and inconsistent spellings, e.g., 'StateHaze1K' vs 'SateHaze1k', 'applyting', 'iscrucial', 'his configuration', and inconsistent use of 'watershed loss' vs 'Watershed loss'. A careful proofreading pass is needed.","section":"Global"}],"recommendation":"major_revision","confidential_remarks":"I considered reject because the watershed-gradient problem and the Table 2 contradiction are serious. I ultimately recommend major_revision because both issues are addressable in principle: the loss can be redefined with a differentiable surrogate and the claims can be rewritten to match the actual per-condition results. However, the revision would require re-running all experiments with the corrected loss; this is not a simple text edit. If the authors cannot provide a working differentiable watershed loss, the paper's main novelty disappears and rejection would be appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis paper is mostly a recombination of existing building blocks. The SwinRRDB-in-U-Net architecture is a straightforward extension of DehazeFormer and UAVD-Net, and the reported gains over RSDformer on RICE are about 0.23 dB PSNR and 0.014 SSIM. The one genuinely new piece, the watershed-based loss, is also the weakest part.\n\nCredit where due: the paper compares against a broad set of baselines, evaluates on two datasets, runs ablations of both the loss terms and the SwinRRDB module, and explicitly notes that only synthetic data was used and that real-world validation is missing. That is honest and useful.\n\nThe soft spots are load-bearing. Section 3.2 describes the watershed label map via local minima detection, integer marker assignment, and four-directional region growing. None of those operations are differentiable, and no surrogate gradient or approximation is mentioned. Equation (12) is then minimized by backpropagation. Either there is an unstated relaxation (which must be named), or the loss contributes no gradient and cannot explain the reported improvement. This is not a minor detail. Figure 2 also describes the loss as a sum of label differences, which does not match the L2 form in Eq. (12).\n\nSecond, the ablation text contradicts its own table: the text says \"Adding Watershed Loss ... increasing PSNR to 32.28 dB,\" but the table shows L2+Watershed at 32.28 dB, which is below L2-only at 32.71 dB. That is a direct internal contradiction.\n\nThird, Table 2 undercuts the abstract's unqualified \"outperforms state-of-the-art\" claim. Under Thin and Thick Fog the method's PSNR is behind several baselines, and under Moderate Fog it wins PSNR but not SSIM (UDAVM-Net has 0.952 vs 0.950). The conclusion's statement that the method achieves the highest PSNR and SSIM in Moderate Fog is false as written. There are also no error bars; all comparisons appear to be single runs, so I would not trust small metric gaps.\n\nIn short: the empirical shell is plausible, but the novel component is ill-defined and the headline claim is contradicted by the paper's own data. If the authors provide code, specify a differentiable implementation of the watershed loss, and fix the internal contradictions, there is a revisable core. As it stands, I would not cite it.\n\nMy recommendation: send it to peer review rather than desk-reject, because the topic is legitimate and the flaws are specific enough to be checked and potentially fixed. Expect major revision or likely rejection unless the differentiability issue is resolved.","headline":"A modest architectural recombination whose one novel piece, the watershed loss, is not shown to be differentiable, and whose headline claim is contradicted by its own Table 2.","tokens_in":23580,"tokens_out":4136,"would_cite":false,"duration_ms":45870,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims SUFERNOBWA, a Swin-Transformer-hybrid U-Net trained with a watershed loss, outperforms existing dehazing methods on RICE and SateHaze1K, achieving 33.24 dB PSNR and 0.967 SSIM on RICE.","keywords":["satellite image dehazing","Swin Transformer","U-Net","watershed loss","guided filter","remote sensing image restoration","haze removal","cloud removal"],"falsifier":"Compute the gradient of the watershed loss with respect to the predicted image at a random early checkpoint; if it is zero at every pixel, the loss cannot drive learning. A cleaner experiment: train the same network with the watershed term replaced by a differentiable boundary loss, or omitted entirely, and compare PSNR and SSIM—identical curves would show the watershed loss is inert.","tokens_in":22728,"feed_emoji":"🛰️","tokens_out":10855,"duration_ms":126233,"temperature":0.7,"pith_summary":"This paper proposes SUFERNOBWA, a dehazing network for satellite images built by inserting Swin Transformer blocks into a U-Net encoder-decoder, and claims it outperforms existing methods on the RICE and SateHaze1K benchmarks. The central assertion is that a composite loss—L2, guided-filter loss, and a novel watershed-based loss—lets the network remove haze and clouds while keeping object boundaries such as roads and buildings intact. On RICE the method reports PSNR 33.24 dB and SSIM 0.967; on SateHaze1K it reports the highest SSIM under thin and thick haze and the highest PSNR under moderate haze. The authors argue this matters because clear satellite images are required for environmental monitoring, agriculture, disaster response, and urban analysis.","feed_headline":"Satellite dehazing hits 33.24 dB with watershed-guided U-Net","feed_subtitle":"Hybrid Swin Transformer and U-Net preserves road and building edges while removing haze, topping RICE and SateHaze1K.","key_machinery":"Two mechanisms carry the argument. SwinRRDB is a block that wraps Swin Transformer self-attention layers inside the Residual-in-Residual Dense Block structure, so a single module can model long-range context and local spatial detail; it is used in every encoder and decoder stage. The watershed loss converts ground-truth and predicted images into normalized segmentation label maps via Gaussian smoothing, local-minima markers, and four-directional region growing, then penalizes their L2 difference, which is meant to force the network to keep object boundaries and internal region consistency.","core_discovery":"The paper's discovery claim is that combining the shifted-window global attention of the Swin Transformer with a U-Net's multi-scale local reconstruction, plus a boundary-aware watershed loss, yields state-of-the-art dehazing for satellite images. The network, called SUFERNOBWA, uses a block the authors name SwinRRDB—Swin Transformer layers embedded in residual-in-residual dense blocks—in the encoder and decoder, with a lightweight bottleneck. The loss compares three quantities between prediction and ground truth: L2 pixel error, a guided-filtered version of each image, and normalized label maps produced by the watershed segmentation algorithm. The authors report that the full loss combinati","pith_inferences":["The paper's own limitations note only synthetic datasets were used; the strongest next test is validation on real hazy/clear satellite pairs, since synthetic haze may not reproduce real reflectance and atmosphere.","The reported RICE gain over the closest Transformer baseline is about 0.23 dB in PSNR, so the practical case for the method likely rests on structural metrics and downstream tasks rather than on pixel-error margin.","If the watershed loss genuinely contributes, a differentiable surrogate—for example, comparing edge maps or soft segmentation maps—could give similar boundary preservation while making end-to-end training cleaner.","Because the watershed label maps are extracted independently from each image, a sensitivity test varying the Gaussian smoothing scale and marker density would show how brittle the method is."],"forward_implications":["If the reported numbers hold, a single RGB-image network can beat methods that use extra sensors or physical priors for satellite dehazing on standard benchmarks.","The watershed-loss recipe should carry over to other restoration tasks where boundaries matter, such as cloud removal, super-resolution, or map generation from aerial images.","On SateHaze1K, the method's SSIM advantage in thin and thick haze suggests structure preservation can matter more than pixel-error improvements when haze is extreme.","The ablations indicate the SwinRRDB module adds roughly 2.7 dB on RICE, so both the architecture and the loss contribute to the reported gain."],"supporting_citations":[{"why":"Provides the RICE paired cloudy/cloud-free benchmark used for the headline result and ablations.","marker":"[44]"},{"why":"Provides the SateHaze1k paired hazy/clear benchmark with thin, moderate, and thick haze splits.","marker":"[22]"},{"why":"Defines the guided filtering operation used by the guided-loss term in the composite loss.","marker":"[43]"},{"why":"RSDformer is the closest Transformer baseline on RICE, at 33.01 dB PSNR, against which the improvement is measured.","marker":"[24]"},{"why":"DehazeFormer is a Swin Transformer dehazing baseline compared on both benchmarks.","marker":"[35]"},{"why":"UDAVM-Net is a strong U-Net-style baseline on SateHaze1k whose moderate-haze PSNR and SSIM provide the main contrast.","marker":"[40]"},{"why":"The dark-channel-prior physics baseline shows the failure of classical dehazing on these satellite datasets.","marker":"[45]"}],"fun_headline_variants":["Watershed loss sharpens satellite dehazing to 33.24 dB","Swin-U-Net fusion clears haze, bests RICE and SateHaze1K","Hybrid Swin Transformer dehazes satellite imagery with edge focus","SUFERNOBWA: Swin-U-Net dehazing with boundary-aware loss"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The watershed label maps are produced by a non-differentiable pipeline of smoothing, local-minima detection, integer labeling, and region growing, yet the watershed loss is minimized with backpropagation; if gradients through that loss are zero or undefined, the watershed term cannot be what improves training.","fun_headline_variants_meta":{"raw":{"variants":["Watershed loss sharpens satellite dehazing to 33.24 dB","Swin-U-Net fusion clears haze, bests RICE and SateHaze1K","Hybrid Swin Transformer dehazes satellite imagery with edge focus","SUFERNOBWA: Swin-U-Net dehazing with boundary-aware loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000166,"raw_usage":{"total_tokens":1112,"prompt_tokens":787,"completion_tokens":325,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":238}},"tokens_in":531,"tokens_out":325,"duration_ms":4248,"temperature":1.0,"reasoning_tokens":238,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:09:21.240635+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the gradient of the watershed loss with respect to the predicted image at a random early checkpoint; if it is zero at every pixel, the loss cannot drive learning. A cleaner experiment: train the same network with the watershed term replaced by a differentiable boundary loss, or omitted entirely, and compare PSNR and SSIM—identical curves would show the watershed loss is inert.","supporting_citations":[{"cited_title":"Single satellite optical imagery dehazing using SAR image prior based on conditional generative adversarial networks","cited_arxiv_id":null,"evidence_quote":"Provides the SateHaze1k paired hazy/clear benchmark with thin, moderate, and thick haze splits."},{"cited_title":"Guided image filtering","cited_arxiv_id":null,"evidence_quote":"Defines the guided filtering operation used by the guided-loss term in the composite loss."},{"cited_title":"Learning an effective transformer for remote sensing satellite image dehazing","cited_arxiv_id":null,"evidence_quote":"RSDformer is the closest Transformer baseline on RICE, at 33.01 dB PSNR, against which the improvement is measured."},{"cited_title":"Vision transformers for single image dehazing","cited_arxiv_id":null,"evidence_quote":"DehazeFormer is a Swin Transformer dehazing baseline compared on both benchmarks."},{"cited_title":"A Dehazing Method for UAV Remote Sensing Based on Global and Local Feature Collaboration","cited_arxiv_id":null,"evidence_quote":"UDAVM-Net is a strong U-Net-style baseline on SateHaze1k whose moderate-haze PSNR and SSIM provide the main contrast."},{"cited_title":"Single image haze removal using dark channel prior","cited_arxiv_id":null,"evidence_quote":"The dark-channel-prior physics baseline shows the failure of classical dehazing on these satellite datasets."}],"review_version":1}