{"id":"71a5e76c-b814-4ea8-9dc0-d95c53188ce0","arxiv_id":"2501.11923","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ProCANet, a UNet-style model with progressive self- and cross-attention between RGB+NIR and NIR satellite inputs, reports IoU 0.815 on Sen1Floods11, ahead of the baselines in its comparison.","lead":"This paper introduces ProCANet, a deep learning model that combines visible-color and near-infrared satellite imagery with progressive attention mechanisms to map flooded areas. On a public flood benchmark it reports the best overlap score among the models tested, plus a test on new high-resolution imagery from Indonesia.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported Sen1Floods11 results may not come from a true held-out test set: Section III-A describes a 65/35 train/validation split, while Section III-D-1 reports 'test set' accuracy, with no seeds or error bars; this undermines the superiority claim.","rationale":"The evaluation validity is the linchpin of the paper's central claim. The text contains an unresolved contradiction between the dataset split described in Section III-A and the 'test set' reported in Section III-D-1; this is not a matter of style or consensus but of whether the numbers in Table I are meaningful. The paper also provides no reproducibility artifacts (code, seeds, or trained weights), and the Citarum evaluation uses a pseudo ground truth whose own limitations are acknowledged. These are correctable, so the appropriate verdict is CONDITIONAL rather than UNCHANGED or REJECT: if the authors provide official-split, multi-seed results and clarify the Citarum label construction, the claim can be assessed. The ablation study (Tables III and IV) provides some internal support, but it does not validate the comparison against external baselines. Overall, my read aligns with the reader's: the method may be sound, but the evidence for superiority is currently under-specified.","tokens_in":7520,"tokens_out":3953,"duration_ms":37397,"concrete_test":"Run ProCANet and the top three baselines (UNet, MANet, LinkNet) on the official Sen1Floods11 train/val/test splits using identical patch size, loss function, optimizer, and epoch count as described in Section III, repeating each training with at least 5 random seeds, and report mean plus or minus standard deviation IoU/F1 on the held-out test set. If ProCANet's mean IoU is not higher than UNet's by more than the pooled standard error, the claimed superiority is not established. As a secondary check, for the Citarum example, compute IoU against the raw NDWI threshold and against the manually corrected labels separately; if the gap between Ours and UNet shrinks, the generalization result is partly an artifact of label construction.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim of superior performance on Sen1Floods11 (Table I, IoU 0.815) rests on the assumption that the evaluation is on a true held-out test set and that baselines are trained under comparable, fair conditions. The paper does not establish this. Section III-A states the dataset was 'divided into two distinct subsets, with 65% allocated to the training set and the remaining 35% designated for validation,' yet Section III-D-1 reports results on 'Sen1Floods11's test set' without reconciling the two. Sen1Floods11 has an official benchmark split with geographically disjoint test regions; if the paper uses its own random 65/35 split, the test images may share locations with training, inflating IoU and making the comparison to published baselines not directly comparable. Additionally, no random seeds, repeated runs, or error bars are reported, and the IoU margin over the closest baseline (UNet, 0.791) is 0.024, plausibly within run-to-run variance for a 25-epoch training. The generalization claim on Citarum (Table II, IoU 0.659) is measured against a 'modified NDWI (pseudo ground truth)' that the authors explicitly state they are 'unable to accurately validate'; a model that relies heavily on NIR may align with this NDWI-derived label without truly segmenting flood extent. Both issues are load-bearing because the headline superiority and generalization claims collapse if the comparison is not on a valid held-out set or a valid reference.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ProCANet, a two-encoder UNet-style segmentation network that progressively applies self-attention and cross-attention to intermediate features from an RGB+NIR encoder and an NIR-only encoder. The method is evaluated on the public Sen1Floods11 dataset, where it reports an IoU of 0.815, F1 of 0.8982, and accuracy of 0.9811 against six baselines, and on PlanetScope 5 m imagery of the Citarum River basin, where it reports IoU of 0.659 against a manually corrected modified-NDWI pseudo-ground-truth label set. The main claims are that the progressive attention fusion produces optimal multispectral feature combinations and that the method is state of the art on Sen1Floods11 while generalizing to higher-resolution imagery.","tokens_in":7898,"tokens_out":7098,"duration_ms":65090,"significance":"If the Sen1Floods11 numbers are obtained on a true geographically disjoint test set with fairly trained baselines, the contribution is a modest but plausible incremental advance: attention-based fusion of RGB+NIR and NIR features inside a U-shaped architecture, supported by an ablation showing attention improving IoU from 0.804 to 0.815. The paper clearly describes the attention block in equations and reports ablations over several modality configurations, which is a strength. However, the significance is currently limited by the absence of statistical rigor: all results are single-run table entries without error bars, the evaluation split is not reconciled with the official Sen1Floods11 benchmark, the generalization reference labels are explicitly unvalidated, and baseline training conditions are not described. The paper does not mention code or model releases, which limits reproducibility.","major_comments":[{"comment":"Section III-A states that Sen1Floods11 was divided into 65% training and 35% validation, but Section III-D-1 and Table I report results on \"Sen1Floods11's test set\" without reconciling the two. Sen1Floods11 includes an official benchmark split with geographically disjoint test regions; if a random 65/35 split was used instead, test images may share locations with training images, inflating IoU and making the comparison with published baselines not directly comparable. This is load-bearing because the headline IoU of 0.815 in Table I is the paper's central claim. Please explicitly state the split used; if it is not the official Sen1Floods11 split, rerun the comparison on the official test split and report those results.","section":"III-A / III-D-1, Table I"},{"comment":"Section III-C says all baselines were trained with MobileNetV2 backbones and RGB+NIR inputs, but it gives no training epochs, optimizer settings, loss functions, patch sizes, or data augmentation for the baselines. Table I is therefore a comparison of ProCANet with under-specified versions of UNet, PSPNet, LinkNet, MANet, PAN, and ConvNeXt V2. Moreover, only one run is reported for each model; the 0.024 IoU margin over UNet (0.815 vs 0.791) may be within run-to-run variance, and the attention ablation in Table III adds only 0.011 IoU (0.804 to 0.815). Please report the mean and standard deviation over at least three seeds for every method and provide a complete training protocol for all baselines.","section":"III-C, Table I, Table III"},{"comment":"Section III-D-2 evaluates generalization on the Citarum PlanetScope data against a \"modified NDWI (pseudo ground truth)\" that the authors explicitly state they are \"unable to accurately validate,\" yet Table II reports an IoU of 0.659 as a quantitative result and Fig. 3's caption calls the same data \"our ground truth data.\" Since ProCANet's second encoder consumes only the NIR band and NDWI is computed from NIR and green bands, agreement with this pseudo-label may partly reflect matching the spectral index rather than detecting true flood extent. Please present this experiment as an illustrative transfer check, or add independent validation labels that do not derive from the same spectral index used in the model's input.","section":"III-D-2, Table II, Fig. 3"},{"comment":"Table IV includes a configuration with \"SAR\" as the second-encoder input, but Section II-A and Section III-A state that only Sentinel-2 multispectral bands (RGB and NIR) were used. The reported numbers for that row (accuracy 0.972, F1 0.838, IoU 0.722) are identical to the Y. Bai et al. entry in Table I, suggesting it is a baseline replication rather than a ProCANet modality variant. Please clarify whether SAR data were used at all; if not, remove the row and correct the modality comparison in Table IV.","section":"Table IV"}],"minor_comments":[{"comment":"Section II-D first says original images are cut into non-overlapping 128x128 patches and then says the training data loader used \"64 patch steps\" to derive more data; please clarify whether the training patches were actually overlapping with stride 64.","section":"II-D"},{"comment":"The BCE loss expression in Section II-D is missing a closing bracket and uses y_pred inconsistently as logits and as probabilities; please align the formula with the implementation.","section":"II-D"},{"comment":"Reference [4] in the Introduction is described as using \"multispectral Sentinel-1 and SAR Sentinel-2,\" but Sentinel-1 is the SAR sensor and Sentinel-2 is the multispectral sensor; the sensor labels are reversed.","section":"I"},{"comment":"Section III-B describes a cosine-annealing scheduler with ten restarts and doubling period lengths while training for only 25 epochs; these numbers are inconsistent and the actual restart schedule should be described precisely.","section":"III-B"},{"comment":"Table II evaluates only UNet and ProCANet on the Citarum data; since generalization is a central claim, please also report the other baselines on the same data or state explicitly why they are omitted.","section":"Table II"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the unresolved ambiguity between the 65/35 validation split described in Section III-A and the \"test set\" results reported in Section III-D-1. If the authors cannot provide results on the official Sen1Floods11 test split, the headline superiority claim should be substantially weakened. The additional issues of missing statistical replication, under-specified baseline training, and the unvalidated pseudo-ground-truth generalization experiment all need to be addressed before I would consider the paper suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's central idea—progressively applying self- and cross-attention at each encoder scale to fuse an RGB+NIR stream with a NIR-only stream—is a reasonable extension of existing attention-based segmentation work. The modality ablation in Table IV is genuinely useful: it shows NIR in the second encoder helps and that RGB alone is far weaker. That kind of systematic input-band comparison is more than many flood-segmentation papers bother to do.\n\nThe problem is that the headline results are not demonstrated robustly. Section III-A describes a 65/35 train/validation split, yet Section III-D-1 reports performance on 'Sen1Floods11's test set' without reconciling the two. If the authors used their own random split rather than the official benchmark split, the comparison to published baselines is not apples-to-apples. Even taking the numbers at face value, there are no seeds, no repeated runs, and no error bars. The margin over UNet is 0.024 IoU, which is well within run-to-run variance for 25 epochs. The attention ablation makes this worse: removing attention still gives IoU 0.804, so the proposed mechanism contributes only 0.011—again, noise-level without error bars.\n\nThe generalization experiment is honestly labeled as a pseudo ground truth, and the authors explicitly say they cannot validate it. That makes the Citarum IoU a case study, not a validation. It's fine to include, but the wording in the conclusions overstates it.\n\nThe baselines are under-specified: all use MobileNetV2, but there is no account of training patches, learning rates, or early stopping for the competitors. So the reader cannot tell if the comparison was fair.\n\nThat said, the paper is clearly written and the core idea is plausible. The fixes are straightforward: define the test split against the official benchmark, run multiple seeds and report mean and std, release code, and describe baseline training in enough detail to reproduce. With those changes, the claimed gain might survive, but as submitted the evidence only supports a modest, possibly within-noise improvement.\n\nThis deserves peer review because the architecture is sensible and the question matters for operational flood mapping, but it needs major revision before acceptance. I'd bring it to a reading group as an example of common validation pitfalls in applied remote-sensing deep learning.","headline":"ProCANet is a sensible attention-based fusion idea with a plausible but unproven empirical claim; the split ambiguity and single-run numbers prevent me from trusting the headline IoU.","tokens_in":8399,"tokens_out":2865,"would_cite":false,"duration_ms":27867,"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 flood-segmentation network that progressively fuses RGB and near-infrared features with self- and cross-attention reports IoU of 0.815 on the Sen1Floods11 test set.","keywords":["flood segmentation","multispectral satellite imagery","cross attention","self-attention","semantic segmentation","near-infrared","Sen1Floods11","remote sensing"],"falsifier":"Re-run the comparison on a properly held-out Sen1Floods11 test split with multiple random seeds and report the mean and standard deviation of IoU; if the 0.815 versus 0.791 gap over UNet falls within one standard deviation, the claimed advantage is not established. Independently, replace the Citarum pseudo ground truth with flood extent from an independent source such as Sentinel-1 SAR or multi-annotator manual labelling; if the 0.659 IoU is mostly agreement with NDWI-like water signals rather than true flood boundaries, the generalization claim is weakened.","tokens_in":7387,"feed_emoji":"🌊","tokens_out":9366,"duration_ms":81086,"temperature":0.7,"pith_summary":"The paper sets out to establish that ProCANet, a segmentation model which progressively applies self-attention and cross-attention to multispectral satellite features, outperforms seven baselines on Sen1Floods11 with an IoU of 0.815, an F1 score of 0.8982, and an accuracy of 0.9811. The motivation is that naively concatenating spectral bands ignores complementary and contrastive information between modalities, while attention lets each band group filter itself and then modulate the other before fusion. The paper also claims the model generalizes without retraining to 5 m PlanetScope imagery of the Citarum River basin, reaching an IoU of 0.659 against a manually corrected NDWI pseudo ground truth. If these results hold, attention-based fusion of RGB and near-infrared data is a practical route to more accurate flood mapping from publicly available satellite imagery.","feed_headline":"Progressive attention fusion maps floods at IoU 0.815","feed_subtitle":"The model edges out UNet and other baselines by fusing RGB and near-infrared bands progressively.","key_machinery":"The load-bearing component is the progressive cross-attention block, inserted after each pooling operation in a UNet-like dual-encoder architecture. For features $\\mathbf{X}_R^i$ from the RGB+NIR encoder and $\\mathbf{X}_N^i$ from the NIR encoder, the block computes self-attention masks $\\mathbf{a}_R^i = \\sigma(\\mathbf{W}_R^i(\\mathbf{X}_R^i))$ and $\\mathbf{a}_N^i$, multiplies them back element-wise to obtain $\\hat{\\mathbf{X}}_R^i$ and $\\hat{\\mathbf{X}}_N^i$, then computes cross-attention masks $\\mathbf{a}_{R\\to N}^i = \\sigma(\\mathbf{W}_{R\\to N}^i(\\hat{\\mathbf{X}}_R^i))$ and $\\mathbf{a}_{N\\to R}^i$ to modulate the opposite stream, and finally fuses the results by element-wise addition, $\\tilde{\\mathbf{X}}_{R+N}^i = \\tilde{\\mathbf{X}}_R^i \\oplus \\tilde{\\mathbf{X}}_N^i$. Applying this block at every scale is what makes the attention 'progressive'; the mechanism lets the water-sensitive NIR channel re-weight the RGB+NIR stream and vice versa before the decoder reconstructs the flood map.","core_discovery":"The paper's central claim is that progressively applying self-attention followed by cross-attention to intermediate features from two encoders—one fed with RGB+NIR and one with NIR alone—produces feature combinations that segment floods more accurately than state-of-the-art models. In the self-attention stage, each modality's features pass through a 3x3 convolution and sigmoid to generate an attention mask, which is multiplied back to amplify relevant and suppress irrelevant content. In the cross-attention stage, each self-attended stream generates a mask that modulates the other stream, and the two modulated streams are merged by element-wise addition at every pooling scale. On Sen1Floods11 this yields IoU 0.815, F1 0.8982, and accuracy 0.9811, above the best baseline UNet at IoU 0.791; the ablation attributes a 0.011 IoU gain to the attention mechanism (0.804 to 0.815) and identifies RGB+NIR with a NIR-only second encoder as the best modality combination.","pith_inferences":["A direct comparison against a single channel-wise attention module at the bottleneck is not reported, so whether the progressive multi-scale application is the specific source of the gain, rather than attention in general, remains an open question this paper does not settle.","Because the Citarum pseudo ground truth was built partly from NDWI, the IoU of 0.659 may partly reflect agreement with NDWI water signals; independent flood-extent labels would be needed to confirm that the model generalizes to a new region and resolution.","The progressive cross-attention block is a generic fusion operator, so the same design could be tested on other modality pairs such as Sentinel-1 SAR with Sentinel-2 optical, or optical with elevation data, which the paper leaves as future work."],"forward_implications":["If the reported results are reproducible, ProCANet provides a consistent accuracy gain for flood segmentation on Sen1Floods11 without changing the backbone or the training data, moving IoU from 0.791 (UNet) to 0.815.","The ablation indicates that the NIR-only second encoder is essential: removing it drops IoU from 0.815 to 0.804, and using RGB alone drops IoU to 0.483, so multispectral input is doing real work.","The claimed cross-resolution result implies a model trained on 10 m Sentinel-2 patches can delineate floods on 5 m PlanetScope imagery without retraining, which would be useful for rapid response when local labels do not exist.","Because the attention block uses only 3x3 convolutions and sigmoids at each scale, the accuracy gain could be added to existing UNet-style flood segmentation pipelines at modest computational cost.","Band-combination experiments suggest that the same architecture can be adjusted by swapping encoder inputs (e.g., RGB+NDWI, or RGB with SAR) with only small changes in performance."],"supporting_citations":[{"why":"Supplies the Sen1Floods11 imagery and hand-labelled flood masks used for training and for the main test comparison.","marker":"[13]"},{"why":"Defines the Sen1Floods11 benchmark and provides a Sentinel-1 plus Sentinel-2 fusion baseline included in the comparison table.","marker":"[4]"},{"why":"Provides the UNet encoding-decoding backbone that ProCANet builds on and is the strongest baseline at IoU 0.791.","marker":"[12]"},{"why":"Supplies MANet, an attention-based remote-sensing segmentation baseline that the method must beat.","marker":"[16]"},{"why":"Supplies PAN, a pyramid attention baseline used in the performance comparison.","marker":"[17]"},{"why":"Supplies PSPNet, a pyramid scene parsing baseline used in the performance comparison.","marker":"[14]"},{"why":"Supplies LinkNet, an efficient encoder-decoder baseline used in the performance comparison.","marker":"[15]"},{"why":"Supplies ConvNeXt V2, a recent convolutional baseline used in the performance comparison.","marker":"[18]"}],"fun_headline_variants":["Progressive attention beats UNet on flood segmentation","Cross-attention network hits IoU 0.815 for flood maps","Self-then-cross attention improves flood segmentation","ProCANet: progressive cross-attention for flood mapping"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes the comparison models were trained and evaluated fairly on a genuinely held-out test set and that the manually corrected NDWI labels for the Citarum imagery are a valid proxy for true flood extent.","fun_headline_variants_meta":{"raw":{"variants":["Progressive attention beats UNet on flood segmentation","Cross-attention network hits IoU 0.815 for flood maps","Self-then-cross attention improves flood segmentation","ProCANet: progressive cross-attention for flood mapping"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001071,"raw_usage":{"total_tokens":4475,"prompt_tokens":925,"completion_tokens":3550,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":3485}},"tokens_in":541,"tokens_out":3550,"duration_ms":24953,"temperature":1.0,"reasoning_tokens":3485,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:42:57.285821+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the comparison on a properly held-out Sen1Floods11 test split with multiple random seeds and report the mean and standard deviation of IoU; if the 0.815 versus 0.791 gap over UNet falls within one standard deviation, the claimed advantage is not established. Independently, replace the Citarum pseudo ground truth with flood extent from an independent source such as Sentinel-1 SAR or multi-annotator manual labelling; if the 0.659 IoU is mostly agreement with NDWI-like water signals rather than true flood boundaries, the generalization claim is weakened.","supporting_citations":[{"cited_title":"Sen1Floods11: A georeferenced dataset to train and test deep learning flood algorithms for sentinel -1,","cited_arxiv_id":null,"evidence_quote":"Supplies the Sen1Floods11 imagery and hand-labelled flood masks used for training and for the main test comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Sen1Floods11 benchmark and provides a Sentinel-1 plus Sentinel-2 fusion baseline included in the comparison table."},{"cited_title":"U -net: Convolutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the UNet encoding-decoding backbone that ProCANet builds on and is the strongest baseline at IoU 0.791."},{"cited_title":"Multiattention network for semantic segmentation of fine -resolution remote sensing images,","cited_arxiv_id":null,"evidence_quote":"Supplies MANet, an attention-based remote-sensing segmentation baseline that the method must beat."},{"cited_title":"Linknet: Exploiting encoder representations for efficient semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies LinkNet, an efficient encoder-decoder baseline used in the performance comparison."}],"review_version":1}