{"id":"7e4bac72-03c6-4312-8427-872fbb20db04","arxiv_id":"2501.07901","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A two-flow residual network that fuses PolSAR polarization features with optical images outperforms six prior cloud-removal models on the authors' new airborne dataset, though data and code are not public.","lead":"This paper proposes PODF-CR, a two-flow network that fuses fully polarimetric SAR data with cloudy optical images to reconstruct cloud-free optical images. The method reports the best cloud-removal scores among seven compared deep learning models on a new airborne dataset, but the dataset and code are not released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3's headline PODF-CR score appears to be the PFSAR-only ablation, not the full BCFSAR+PFSAR two-flow model: Table 4's w PolSAR row is 34.874/0.890/0.921/0.922 while Table 3's 34.992/0.892/0.922/0.916 exactly matches w PFSAR.","rationale":"The reader's weakest assumption, that simulated clouds and single-date airborne data may not transfer to real cloud corruption, is a legitimate external-validity concern. However, the single most load-bearing issue is more basic: the paper's reported results are internally inconsistent with its own ablation. The strongest claim specifically cites Table 3's 34.992 PSNR. Table 4 shows that the full model (w PolSAR, both BCFSAR and PFSAR) scores 34.874, and that the PFSAR-only variant scores 34.992. The exact four-metric match between Table 3's PODF-CR row and Table 4's w PFSAR row means this is not a rounding difference; it is a labeling or copying error that must be resolved. The architecture contribution advertised in the abstract and introduction is precisely the two-flow fusion of BCFSAR and PFSAR, yet the ablation says adding BCFSAR reduces accuracy relative to PFSAR alone. This does not refute the broader claim that a model from this family beats the six listed baselines, because both 34.874 and 34.992 exceed the best baseline 33.036, and the PFSAR-only variant also outperforms all baselines. But it does mean the paper's advertised record numbers and its explanation of why the design works are both unsupported as written. A single reproducible rerun of the two configurations settles the issue. I therefore keep the reader's CONDITIONAL verdict: the internal inconsistency reinforces the need for code/data release, a clean validation protocol, and correction of the reported numbers, but it does not by itself prove the method is wrong.","tokens_in":24827,"tokens_out":6413,"duration_ms":60213,"concrete_test":"Run the full PODF-CR with both BCFSAR and PFSAR inputs on the same 1630-pair training / 400-pair test split, with the same loss (L_g + 10 L_l + 1 L_ssim) and optimizer settings, and compute PSNR/SSIM/CC/SAM on the test set. Compare to Table 3's PODF-CR row and Table 4's w PolSAR and w PFSAR rows. If the full model reproduces approximately 34.874/0.890/0.921/0.922 (w PolSAR) rather than 34.992/0.892/0.922/0.916, then Table 3's headline row is the PFSAR-only ablation and must be corrected. Also rerun the PFSAR-only variant to confirm it reproduces 34.992/0.892/0.922/0.916, which settles whether the advertised SOTA metric is attributable to the proposed fusion or to dropping BCFSAR.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing problem is internal: the headline result in Table 3 seems to be the PFSAR-only ablation, not the full two-flow model. Table 3 lists PODF-CR as PSNR 34.992, SSIM 0.892, CC 0.922, SAM 0.916. Table 4's 'w PFSAR' row is identical to four decimals, while the 'w PolSAR' row (both BCFSAR and PFSAR, i.e., the model described in Section 3.1) is 34.874/0.890/0.921/0.922. The text in Section 4.3.1.1 explicitly states that the PFSAR-only model achieves the highest accuracy and that introducing BCFSAR images 'may introduce noise to the residual learning of the entire network, leading to a decrease in reconstruction accuracy.' Thus the paper's own ablation contradicts the architecture claim: the two-flow BCFSAR+PFSAR fusion does not produce the reported SOTA metrics. Either Table 3 misattributes the PFSAR-only score to the full model, or the two tables were produced under different, unspecified conditions. Either way, the quantitative basis of the central claim is internally inconsistent and unreproducible as stated. The full model at 34.874 would still beat the best baseline USSRN-CR at 33.036, so the broad superiority claim may survive, but the claimed mechanism and the exact advertised numbers are not supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes PODF-CR, a two-flow encoder-decoder network for cloud removal that fuses cloudy optical images with PolSAR-derived backscatter coefficient images (BCFSAR) and polarization feature images (PFSAR). The PolSAR branch uses spatial and channel dynamic filters for denoising, and the network introduces multi-modality cross fusion (MMCF), multi-modality refinement fusion (MMRF), and atrous spatial pyramid pooling (ASPP). The method is evaluated on a new airborne OPT-BCFSAR-PFSAR dataset of 2030 image pairs with simulated cloud masks, and the paper claims state-of-the-art results in PSNR, SSIM, CC, and SAM relative to six baselines.","tokens_in":25090,"tokens_out":4338,"duration_ms":42410,"significance":"If the reported numbers are correct and reproducible, the paper would offer a useful architecture and a new PolSAR-feature dataset for SAR-optical cloud removal; the dynamic-filter denoising and cross-modal fusion components are sensible, and the per-category and per-cloud-cover analysis is more detailed than is common in this literature. However, the central quantitative claim is currently undermined by an internal inconsistency between Tables 3 and 4, and the evaluation protocol does not separate model selection from final testing. As a result, the significance of the contribution cannot be assessed from the manuscript as written. No code or dataset repository is indicated, which further limits immediate verification.","major_comments":[{"comment":"The headline result in Table 3 (PODF-CR: PSNR 34.992, SSIM 0.892, CC 0.922, SAM 0.916) is numerically identical to the 'w PFSAR' ablation row in Table 4, while the full model with both BCFSAR and PFSAR ('w PolSAR') is reported as 34.874/0.890/0.921/0.922. Section 4.3.1.1 further states that adding BCFSAR feature images may introduce noise to residual learning and decrease reconstruction accuracy. The main comparison table therefore appears to attribute the PFSAR-only variant to the full two-flow PODF-CR architecture described in Section 3.1. This must be reconciled: either Table 3 is mislabeled, or the two tables were produced under different experimental conditions. In either case, the current text does not support the claim that the proposed two-flow BCFSAR+PFSAR fusion produces the advertised state-of-the-art numbers.","section":"§4.2.1 / Table 3 and §4.3.1.1 / Table 4"},{"comment":"Hyperparameters are selected on the test split: the learning rate is found by 'systematic search', and lambda1 and lambda2 are chosen in Fig. 13 by maximizing PSNR, with no separate validation set described anywhere in the 80/20 train/test split. This makes the reported quantitative comparisons and final model selection optimistic and makes the protocol impossible to reproduce as a fixed evaluation. Please introduce a validation split or nested cross-validation, select hyperparameters and input configurations using validation data only, and report test-set metrics exactly once after fixing those choices.","section":"§4.1.3 and §4.3.2"},{"comment":"The dataset consists of a single airborne scene acquired on one date, with clouds simulated by randomly generated masks; no real cloudy optical images or independent acquisition dates are used. Since the reconstruction loss is evaluated only on these simulated masks, the claim that the method outperforms existing methods for actual cloud removal is not yet supported. I recommend evaluating on a public benchmark such as SEN12MS-CR or including real cloudy optical images with co-registered PolSAR data, and also reporting sensitivity to registration errors beyond the stated 2-pixel tolerance.","section":"§4.1.1"}],"minor_comments":[{"comment":"The CC column is marked with a down arrow, but Section 4.1.2 states that higher CC values indicate better performance; the arrow should be an up arrow.","section":"Table 3"},{"comment":"The attention unit is called SCRU in the text and SCAU in Figure 4, and the fusion block is called MMRF in the text but MRFB in Figure 4(a); please unify the notation.","section":"§3.2.4 and Figure 4"},{"comment":"The text says the model has a 'parameter number of 58.86' without specifying the unit; it should state 58.86M, and the FLOPs figures should state the exact input size used for measurement.","section":"§4.3.3"},{"comment":"The ASPP paragraph contains a bare '(?)' placeholder where a citation is expected; please insert the appropriate reference.","section":"§3.2.5"},{"comment":"The dataset name is given as OPT-FPBCSAR-PFSAR in the contribution list and OPT-BCFSAR-PFSAR elsewhere; please make the name consistent.","section":"§1.1 and throughout"},{"comment":"The reference list appears to contain the same Cheng et al. paper twice as 2013a and 2013b; please verify and merge duplicate entries.","section":"§2.1"},{"comment":"The axis labels in Figure 13 contain stray characters and are difficult to read; please regenerate the figure with clean mathematical notation.","section":"Figure 13"}],"recommendation":"major_revision","confidential_remarks":"The internal inconsistency between Tables 3 and 4 is the most serious issue and should be resolved before the paper is reconsidered. I also note that no code or dataset release is mentioned; for a paper whose main empirical contribution is a new dataset, this is a significant omission for reproducibility. The paper is otherwise within the scope of the journal, but the evaluation protocol and external validity need strengthening."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a modest but real contribution—first to push full-polarization PolSAR features (PFSAR) into cloud removal for optical imagery. The reported 34.99 PSNR is for the PFSAR-only variant, which is consistent with the architecture description in Section 3.1 (the PolSAR branch takes coherence-matrix polarization features). So the stress-test's claim of a Table 3/4 inconsistency does not land; there is no arithmetic error. The actual soft spot is framing: the title and abstract trumpet a two-flow 'BCFSAR+PFSAR' fusion, but the ablation shows adding BCFSAR hurts (34.874 vs 34.992). The authors are honest about this in Section 4.3.1.1, but it undercuts the headline claim that both feature types matter.\n\nWhat is genuinely new: using PFSAR (9-band coherence-matrix features) as auxiliary input, plus a two-stream architecture with gated convolution, dynamic filters, cross-skip fusion, and attention refinement. The component ablations are individually positive (e.g., SCDF +0.25, GC +0.76 PSNR), which gives some confidence that the design choices matter. The dataset, while new and carefully annotated, is single-scene, single-date, airborne, with simulated clouds, and is not released.\n\nMain weaknesses, in order: (1) Hyperparameters (lr, lambda1, lambda2) were tuned on the test split—no validation set is described; this biases the reported numbers. (2) No error bars; with 400 test patches, seed variance could easily be ~0.1 dB. (3) Simulated clouds over one scene; real cloud thickness, shadow, and the 2-pixel registration tolerance at 0.5 m resolution may erode the 1.96 dB advantage. (4) No code or data release, so the PFSAR-specific claim cannot be reproduced. (5) Former-CR is cited but not included as a baseline, which is a gap since it is the transformer-based SAR-optical method.\n\nThe paper deserves a serious referee because the PFSAR idea is novel and the architecture is plausible, but it needs major evaluation fixes: a proper validation split, multiple seeds, evaluation on real clouds or a public benchmark, and clearer positioning of PFSAR-only as the actual method. I would send it to review, with the expectation of heavy revision.\n\nFor a reading group, it is a decent case study in how ablations can conflict with framing.","headline":"The PFSAR-only variant is the actual method despite the BCFSAR+PFSAR framing; the main barrier is the evaluation protocol, not the architecture.","tokens_in":25735,"tokens_out":5672,"would_cite":false,"duration_ms":51947,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","68U10","94A08"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a two-flow residual network fusing full-polarimetric SAR features with cloudy optical images reconstructs cloud-free optical imagery more accurately than existing methods, reporting state-of-the-art PSNR, SSIM…","keywords":["cloud removal","PolSAR","optical-SAR fusion","two-flow residual network","dynamic filters","attention mechanism","image restoration","polarimetric features"],"falsifier":"Take the trained PODF-CR model and test it on real cloud-contaminated optical images from the same area where the cloud mask is not known and coregistration errors exceed the stated 2-pixel tolerance; if PSNR and SSIM against a later cloud-free acquisition drop to the level of the DSen2-CR or USSRN-CR baselines, then the reported state-of-the-art numbers depend on the simulated-mask training setup rather than on the fusion architecture itself.","tokens_in":24488,"feed_emoji":"🛰️","tokens_out":7543,"duration_ms":66279,"temperature":0.7,"pith_summary":"This paper attempts to establish that full-polarimetric SAR (PolSAR) data can serve as a stronger auxiliary source than single- or dual-polarization SAR for reconstructing optical satellite images hidden by clouds. To do this, it introduces PODF-CR, a two-flow residual network that encodes PolSAR and optical images in parallel, suppresses speckle with decoupled dynamic filters, and fuses the two streams with cross-skip fusion blocks and an attention-based refinement stage. On the OPT-BCFSAR-PFSAR airborne dataset, the method reports the best quantitative results among seven methods, with PSNR 34.992, SSIM 0.892, correlation coefficient 0.922, and SAM 0.916, and it retains an advantage at heavy cloud coverage up to 80-100%. A sympathetic reader would care because the paper shows that polarization and scattering features extracted from the coherence matrix can carry enough ground information to restore fine texture details that optical-only or single-polarization SAR methods miss.","feed_headline":"Two-flow fusion network sets cloud-removal record","feed_subtitle":"Polarimetric SAR features help reconstruct cloud-obscured pixels, reaching PSNR 34.99 and SSIM 0.89.","key_machinery":"The load-bearing mechanism is the two-flow fusion of PolSAR-derived features with optical features. The PolSAR branch is built around SCDF (coupled spatial and channel dynamic filters), which generates per-pixel spatial filters and per-channel filters from the input to suppress speckle before features propagate; MMCF (multi-modality cross fusion) blocks use cross-modality dense skip connections to let each branch see the other's feature maps at multiple encoder scales; and MMRF (multi-modality refinement fusion) blocks apply spatial-channel attention plus a cross-modal weighted refinement unit that computes pixel-wise correlations between optical, PolSAR, and fusion features to refine the fused representation globally. This combination is what lets the network use polarization and scattering information to fill cloud gaps while preserving cloud-free regions.","core_discovery":"The central claim is that the proposed PODF-CR network reconstructs cloud-free optical images from four aligned inputs: the cloudy optical image, a 3-band backscatter coefficient image (BCFSAR), a 9-band polarization feature image (PFSAR), and a cloud mask used in the loss. The architecture is an encoder-decoder with two parallel branches; the optical branch uses gated convolutions to avoid propagating cloud pixels, the PolSAR branch uses coupled spatial-channel dynamic filters to suppress speckle, and multi-modality cross fusion (MMCF) blocks pass each modality's features into the other branch at several scales via cross-skip connections. The resulting fusion features are refined by a multi-modality refinement fusion (MMRF) block with spatial-channel attention and a cross-modal weighted refinement unit, and the decoder applies atrous spatial pyramid pooling to combine multi-scale context before output. The loss combines global L1, a cloud-masked local L1, and SSIM. The paper reports that this configuration outperforms six comparison algorithms on the dataset, and ablation experiments attribute the gains to the polarization features, the dynamic filters, the gated convolutions, the fusion modules, and the multi-scale pooling.","pith_inferences":["If the PFSAR-only result generalizes, a testable extension is that BCFSAR backscatter images may be redundant or even detrimental for this setting; future datasets could treat them as an auxiliary denoising channel rather than primary input.","The method's reliance on a coherence-matrix representation suggests that other polarimetric decompositions (for example, Cloude-Pottier, Freeman-Durden, or Yamaguchi components) could be substituted for the current nine-band PFSAR input and compared directly.","Because the training data is a single airborne scene with simulated random cloud masks, real-world deployment would require validation on multi-temporal, satellite-acquired cloudy images; the reported 2-pixel registration tolerance between optical and PolSAR is a likely practical constraint.","The 58.86 million parameters and roughly 419 GFLOPs per 256x256 tile imply that satellite-scale processing would need tiling or a lighter variant, a direction the paper does not address."],"forward_implications":["Full-polarimetric SAR input, represented through backscatter coefficients and polarization features from the coherence matrix, becomes an effective auxiliary modality for optical cloud removal.","The reported architecture is a concrete recipe: parallel optical and PolSAR encoders, gated convolutions on the optical side, dynamic filters on the PolSAR side, cross-skip fusion blocks, attention-based refinement, and multi-scale decoding.","Ablation results indicate that polarization feature images (PFSAR) carry more restoration value than backscatter coefficient images (BCFSAR), with PFSAR-only reaching the same top PSNR of 34.992 as the full model.","Performance degrades gracefully with increasing cloud coverage, with the method still reconstructing building and traffic-route structure at 80-100% cloud cover where other models blur.","Removing the PolSAR auxiliary data entirely drops PSNR by about 2.2 dB, supporting the claim that the PolSAR branch is the source of the missing-area information."],"supporting_citations":[{"why":"Optical-only SpAGAN baseline that PODF-CR must beat; using only cloudy optical data gives the weakest results among compared methods.","marker":"Pan, 2020"},{"why":"Pix2pix baseline that translates PolSAR images directly to optical, showing that PolSAR-only conversion distorts cloud-free areas.","marker":"Isola et al., 2017"},{"why":"SAR-Opt-cGAN baseline, the first joint SAR-optical fusion approach that combines damaged optical and SAR inputs.","marker":"Grohnfeldt et al., 2018"},{"why":"DSen2-CR deep residual SAR-optical cloud removal baseline, also the source of the residual scaling constant 0.1 used in RB-GC.","marker":"Meraner et al., 2020"},{"why":"GLF-CR baseline using global-local fusion and attention, the direct predecessor for the cross-modality fusion strategy.","marker":"Xu et al., 2022"},{"why":"USSRN-CR baseline with gated convolution and self-attention, the strongest numerical competitor reported.","marker":"Wang et al., 2023"},{"why":"Decoupled dynamic filters (DDF) that the SCDF module adapts for spatial and channel dynamic filtering.","marker":"Zhou et al., 2021"},{"why":"Encoder-decoder with atrous separable convolution (DeepLabv3+), source of the ASPP multi-scale pooling module used in the decoder.","marker":"Chen et al., 2018"}],"fun_headline_variants":["PolSAR-optical fusion lifts cloud removal to new high","Two-flow net uses radar data to erase clouds from optical images","Cross-skip fusion boosts cloud-free image reconstruction","Dynamic filters polish PolSAR for sharper cloud removal","Fusion of radar and optical beats cloud-removal benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that randomly generated cloud masks over a single-date airborne scene faithfully mimic real cloud corruption, with optical and PolSAR images registered within about 2 pixels, so the measured reconstruction gains transfer to genuine cloud-covered imagery.","fun_headline_variants_meta":{"raw":{"variants":["PolSAR-optical fusion lifts cloud removal to new high","Two-flow net uses radar data to erase clouds from optical images","Cross-skip fusion boosts cloud-free image reconstruction","Dynamic filters polish PolSAR for sharper cloud removal","Fusion of radar and optical beats cloud-removal benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000685,"raw_usage":{"total_tokens":3163,"prompt_tokens":1057,"completion_tokens":2106,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":673,"completion_tokens_details":{"reasoning_tokens":2027}},"tokens_in":673,"tokens_out":2106,"duration_ms":16778,"temperature":1.0,"reasoning_tokens":2027,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:31:13.119622+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained PODF-CR model and test it on real cloud-contaminated optical images from the same area where the cloud mask is not known and coregistration errors exceed the stated 2-pixel tolerance; if PSNR and SSIM against a later cloud-free acquisition drop to the level of the DSen2-CR or USSRN-CR baselines, then the reported state-of-the-art numbers depend on the simulated-mask training setup rather than on the fusion architecture itself.","supporting_citations":[{"cited_title":"and Efros, A","cited_arxiv_id":null,"evidence_quote":"Pix2pix baseline that translates PolSAR images directly to optical, showing that PolSAR-only conversion distorts cloud-free areas."},{"cited_title":"and Zhu, X., 2018","cited_arxiv_id":null,"evidence_quote":"SAR-Opt-cGAN baseline, the first joint SAR-optical fusion approach that combines damaged optical and SAR inputs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DSen2-CR deep residual SAR-optical cloud removal baseline, also the source of the residual scaling constant 0.1 used in RB-GC."},{"cited_title":"and Li, Z., 2023","cited_arxiv_id":null,"evidence_quote":"USSRN-CR baseline with gated convolution and self-attention, the strongest numerical competitor reported."},{"cited_title":"and Yang, M.-H., 2021","cited_arxiv_id":null,"evidence_quote":"Decoupled dynamic filters (DDF) that the SCDF module adapts for spatial and channel dynamic filtering."},{"cited_title":"and Adam, H., 2018","cited_arxiv_id":null,"evidence_quote":"Encoder-decoder with atrous separable convolution (DeepLabv3+), source of the ASPP multi-scale pooling module used in the decoder."}],"review_version":1}