Pith. sign in

REVIEW 3 major objections 7 minor 65 references

Explainable Flood Segmentation on Sentinel-1 SAR1 Imagery Using CNN and Transformer Architectures

T0 review · 3 major / 7 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read SegFormer-b2 beats U-Net on all seven ETCI flood scenes, but the advantage disappears after fine-tuning on Sen1Floods11.

desk verdict Useful benchmark paper with one credible significant result, one under-supported explainability story, and a missing sensitivity analysis on the ETCI tile filter. read the letter →

arxiv 2606.16302 v3 pith:H4R7MGYT submitted 2026-06-15 cs.CV

Arundhuti Banerjee , David Daou This is my paper · ORCID
classification cs.CV
keywords floodsegmentationSentinel-1SARSegFormerU-Netmulti-classexplainableAIGrad-CAMuncertaintyestimation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether a vision transformer (SegFormer) reliably outperforms the standard convolutional U-Net at the practically important task of separating newly flooded land from permanent water in Sentinel-1 radar images. On the ETCI benchmark, the paper reports that SegFormer-b2 beats U-Net on all seven test scenes, with a mean per-scene flood IoU advantage of 0.125 that a paired nonparametric test finds significant. After fine-tuning both models on Sen1Floods11, the advantage narrows to about 0.026 per test chip, within scene-to-scene variability, and appears mainly on spatially fragmented floods. The paper also finds complementary explainability: SegFormer-b2's Grad-CAM maps are more coherent and faithful, while U-Net offers more spatially precise explanations and marginally better calibration. If right, the conclusion is that architecture superiority in SAR flood mapping is conditional, depending on training domain and flood complexity rather than on a universal model-class advantage.

What carries the argument

The argument is carried by a head-to-head comparison of six segmentation models — three CNNs (U-Net, U-Net++, DeepLabV3) and three SegFormer variants (b0, b1, b2), where SegFormer is a hybrid vision transformer with a hierarchical encoder and a lightweight MLP decoder. Models are trained with a combined Dice and focal loss for a three-class problem (background, permanent water, flooded land), using scene-level data splits to prevent spatial leakage. Statistical claims rest on paired per-scene or per-chip flood IoU differences tested with a nonparametric signed-rank test and bootstrap confidence intervals. Explainability machinery includes entropy maps, Grad-CAM (gradient-weighted class activ

What would settle it

Recompute the per-scene flood IoU comparison on all 7,948 ETCI test tiles without the 600-flood-pixel filter, using a cluster-aware significance test that groups the four North Alabama acquisitions; the headline claim survives only if SegFormer-b2 still wins on all seven scenes with p<0.05.

Watch

Extended reading notes

Core claim

The central claim is that for multi-class flood segmentation on Sentinel-1 SAR imagery, the SegFormer-b2 transformer outperforms the U-Net baseline when both are trained from scratch on the ETCI dataset, achieving higher flood IoU on all 7 test scenes (mean +0.125, p=0.016). After in-domain fine-tuning on Sen1Floods11, however, the per-chip advantage drops to +0.026 without test-time augmentation and is no longer statistically significant (p=0.34), with the residual edge concentrated in spatially fragmented flood events such as India and Paraguay. The paper further asserts a division of explainability labor: SegFormer-b2's gradient-weighted activation maps are more spatially coherent and cau

Load-bearing premise

The load-bearing premise is that the seven ETCI test scenes can be treated as independent statistical replicates and that the 940 flood-heavy tiles used for evaluation represent each scene; if the four North Alabama acquisitions are autocorrelated, or if the 600-flood-pixel filter changes per-scene rankings on the excluded tiles, the 7-of-7 result could be overstated.

Editorial extensions

If this is right

  • On ETCI, SegFormer-b2's flood-IoU advantage over U-Net is statistically significant and holds across all three geographic regions in the test split.
  • After fine-tuning on Sen1Floods11, no architecture shows a statistically significant edge; the residual transformer advantage appears mainly on fragmented flood patterns.
  • Flooded-land segmentation is roughly twice as hard as permanent-water segmentation for every model, reflecting the SAR ambiguity of shallow water and wet soil.
  • For operational use, U-Net's better-calibrated, spatially precise uncertainty is better suited to communicating flood-risk boundaries, while SegFormer-b2's faithful attributions suit automated feature prioritization.
  • Test-time augmentation yields only small gains on Sen1Floods11, suggesting the models are already robust to geometric and intensity variations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The fragmented-flood concentration suggests a testable rule of thumb: transformer advantage should grow with the spatial complexity of inundation; a controlled comparison on deliberately fragmented versus contiguous floods would quantify this.
  • A practical deployment pattern suggested by the complementary explainability results is an ensemble that uses SegFormer-b2's faithful attributions for automated prioritization and U-Net's better-calibrated boundary uncertainty for human-facing risk communication.
  • Extending the same scene-based paired comparison to multi-temporal SAR (pre- and post-flood) would directly test whether the transformer's long-range context becomes decisive when change information is available.
  • The multi-class formulation that separates flooded land from permanent water could carry over to other SAR segmentation tasks, such as wetland or snowmelt mapping, where the same backscatter ambiguity arises.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper compares three CNN-based segmentation models (U-Net, U-Net++, DeepLabV3 with ResNet-34 backbone) and three SegFormer variants (b0, b1, b2) on the multiclass task of separating flooded land from permanent water in Sentinel-1 SAR imagery. The authors train on ETCI from scratch and fine-tune on Sen1Floods11, using scene-based splits on ETCI and chip-level evaluation on Sen1Floods11. The central empirical claims are: (1) on ETCI, SegFormer-b2 outperforms U-Net on all seven test scenes, with a mean per-scene flood-IoU advantage of +0.125 (bootstrap CI [+0.067, +0.181], Wilcoxon p = 0.016); (2) after fine-tuning on Sen1Floods11, this advantage shrinks to +0.026 without TTA (p = 0.34) and is not statistically significant; (3) qualitative and quantitative explainability analyses show complementary strengths, with U-Net yielding more precise and slightly better-calibrated explanations and SegFormer-b2 yielding more faithful and more globally coherent attributions.

Significance. If the empirical results hold, the paper offers a useful, reasonably rigorous comparison point for SAR flood segmentation: it is one of the few studies to separate flooded land from permanent water in a multiclass setting, it evaluates spatial generalization via scene-based splits, and it reports paired statistical tests with bootstrap confidence intervals rather than relying solely on pooled metrics. The Sen1Floods11 null result is especially valuable because it cautions against unconditional claims that transformer architectures dominate CNNs for SAR flood mapping. The explainability section also contributes quantified XAI metrics, although the statistical support for those claims is weaker. The paper does not involve fitted parameters masquerading as predictions; all headline numbers are direct measurements against external benchmark data.

major comments (3)
  1. [§3.1, §5.1, Fig. 3a] The ETCI significance claim is computed on only 940 of 7,948 test tiles, namely those containing at least 600 flood pixels. The paper does not report per-scene tile counts for this filtered subset, nor does it test whether the 7/7 per-scene ordering is stable when the threshold is lowered or removed. Four of the seven test scenes are North Alabama acquisitions (different months), so if the filter differentially drops tiles across scenes, or if some scenes contribute very few qualifying tiles, the per-scene IoU estimates become noise-sensitive. With n=7, p=0.016 is the minimum attainable Wilcoxon value; one scene flipping direction would make the result non-significant. Please provide per-scene tile counts and a sensitivity analysis with thresholds such as 0, 100, 300, 600, and 1000 flood pixels, reporting per-scene IoU and the paired test under each threshold.
  2. [§5.2, Fig. 3b, §6] The abstract and conclusion state that SegFormer-b2's residual advantage on Sen1Floods11 is 'concentrated in spatially fragmented flood events,' but the manuscript offers no quantitative definition of fragmentation, no fragmentation metric, and no statistical interaction or correlation test to support that claim. The observation appears to rest on the India and Paraguay chips having lower U-Net IoU in some cases. This is load-bearing for the paper's nuanced conclusion that the advantage depends on flood pattern complexity. Please operationalize fragmentation (e.g., edge density, number/area of connected components, patch-size distribution) and test whether per-chip advantage correlates with that metric; otherwise, soften the claim to an informal observation.
  3. [§4.3, Tables 5 and 6, Fig. 7] The explainability claims are presented as the paper's second main contribution, but the differences in Tables 5 and 6 are not accompanied by any uncertainty quantification or paired significance tests. Several differences are very small (ECE 0.0675 vs 0.0694; AUIC 0.1494 vs 0.1489), and the pointing-game difference (0.7407 vs 0.6667) is computed over 27 chips, where a few chips could change the conclusion. The statement that U-Net is 'better calibrated' or that SegFormer-b2 is 'more faithful' therefore lacks statistical support. Please report bootstrap confidence intervals or paired tests for all XAI metrics, and clarify how deletion/insertion tests were adapted to the multiclass setting, including what baseline image was used and how the Grad-CAM saliency was defined for class 2.
minor comments (7)
  1. [Abstract/§3.2] The dataset name is inconsistently rendered as 'SenFloods11' in the abstract/keywords and 'Sen1Floods11' elsewhere; Section 3.2's header appears truncated as '11 Dataset.' Please standardize.
  2. [Eq. (4)] Typo: 'P recison' should be 'Precision' in the F1 equation and surrounding text.
  3. [§5.2, Table 5 discussion] The text 'As seen in Table IV' should refer to Table V, which reports the quantitative XAI metrics.
  4. [§2, §4.1] The paper uses 'DeepLabV3' and 'DeepLabV3+' in different places; the experimental setup should state exactly which variant was used.
  5. [§3.1] The third input channel is described as '1−(V H/V V)' but the formula is not defined precisely in terms of the channel values used; please clarify the computation and normalization.
  6. [Figures 4 and 5] Both captions describe the scene as 'the severe monsoon flooding in Assam, India, on 12th August 2016'; please confirm that the two figures show distinct scenes and update the captions accordingly.
  7. [General] No random seed or number of repeated runs is reported, which matters for the small per-scene differences on Sen1Floods11; please state the seed and, if feasible, report results over multiple seeds.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical benchmark comparison with no fitted-parameter-as-prediction structure.

full rationale

This paper reports direct measurements of model performance on two external benchmarks (ETCI and Sen1Floods11). There is no derivation chain in which an output is defined in terms of an input, no fitted parameter is later renamed as a prediction, and no load-bearing self-citation. The ETCI significance claim is based on per-scene flood IoU computed from model predictions against ground-truth labels; the 940-tile selection criterion is a preprocessing choice that could affect robustness, but it is not a fitted parameter and does not make the reported comparison true by construction. The Sen1Floods11 three-class label construction combines hand labels with the JRC permanent-water layer, an external data operation, not a model output. No author self-citations appear in the reference list, and no uniqueness theorem or ansatz is imported from prior work by the same authors. Statistical concerns about the small number of scenes and the tile filter are legitimate correctness/robustness questions, not circularity. Therefore the appropriate finding is no significant circularity, score 0.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central claims are empirical, so the ledger records hand-set hyperparameters that can shift rankings, dataset/label assumptions that ground the benchmark, and the XAI validity assumptions that support the explainability conclusions. No new entities, particles, forces, or theoretical constructs are introduced.

free parameters (8)
  • learning_rate = 1e-4 (UNet), 5e-5 (SegFormer/DeepLabV3)
    Hand-set hyperparameter that directly affects all reported segmentation scores and rankings.
  • weight_decay = 1e-4
    Hand-set AdamW regularization; affects generalization and calibration.
  • focal_loss_gamma = 2.0
    Hand-set loss parameter for class imbalance; influences flood-class learning dynamics.
  • positive_tile_flood_threshold = 600 flood pixels
    Chosen threshold that defines positive tiles for ETCI sampling and also defines the 940-tile evaluation subset.
  • batch_size = 16 (ETCI), 8 (fine-tune)
    Hand-set; affects optimization stability and comparison across architectures.
  • early_stopping_patience = 8 epochs
    Hand-set; determines when each model stops training.
  • tta_intensity_scales = 0.95, 1.0, 1.05
    Hand-chosen augmentation scales for test-time augmentation; matters for Sen1Floods11 results.
  • max_epochs = 100 (ETCI), 40 (fine-tune)
    Hand-set training caps; interact with early stopping.
assumptions (6)
  • domain assumption Scene-level splitting prevents spatial leakage because tiles within a scene are croppings of a single Sentinel-1 acquisition.
    Stated in Sections 3.1 and 4.2; the independence of the seven test scenes is the basis for the Wilcoxon p=0.016 claim.
  • domain assumption ETCI labels and Sen1Floods11 hand labels plus the JRC permanent-water layer correctly identify permanent water versus flooded land.
    All flood/water metrics and the multiclass formulation depend on label correctness; stated in Section 3.2.
  • domain assumption VV, VH, and 1-(VH/VV) SAR channels contain sufficient information to separate the three classes.
    The entire modeling approach presumes the input features encode flood-relevant structure; implicitly assumed throughout.
  • domain assumption Grad-CAM applied to both architectures yields meaningful localization, including for SegFormer's transformer/MLP design.
    Section 4.3 describes Grad-CAM from final convolutional layers, but SegFormer's decoder is a lightweight MLP; the paper does not specify which layer is used or justify the adaptation.
  • domain assumption Deletion/insertion AUDC/AUIC and the pointing game are accepted proxies for explanation faithfulness.
    Sections 4.3 and 5.2 interpret these metrics as faithfulness and spatial precision without external validation in this domain.
  • standard math Standard statistical procedures (IoU, ECE, Pearson correlation, Wilcoxon signed-rank, bootstrap) are correctly applied.
    The formulas in Section 4 are standard and the implementation is not independently verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Explainable Flood Segmentation on Sentinel-1 SAR1 Imagery Using CNN and Transformer Architectures." pith.science (2026). https://pith.science/paper/H4R7MGYT

@misc{pith2026260616302,
  author       = {Pith},
  title        = {Pith review of: Explainable Flood Segmentation on Sentinel-1 SAR1 Imagery Using CNN and Transformer Architectures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H4R7MGYT}},
  note         = {Machine review of arXiv:2606.16302}
}
read the original abstract

Rapid and accurate flood prediction is essential for disaster response and mitigation planning. Synthetic Aperture Radar (SAR) sensors in satellites are well-suited for this purpose because they operate independently of weather and daylight conditions. Although SAR-based data enable all-weather flood monitoring, distinguishing flooded land from permanent water remains a significant challenge, particularly when flooding is defined strictly as inundated land. This study provides a comprehensive comparison of convolutional neural network (CNN) and vision transformer architectures for multi-class flood segmentation using Sentinel-1 SAR imagery, specifically trained to separate flooded land from permanent water bodies and land. Three state-of-the-art (SOTA)CNN-based models, U-Net, U-Net++, and DeepLabV3 with ResNet-34 backbone, and three SegFormer variants (b0,b1,b2) were evaluated in two benchmark datasets, the ETCI NASA dataset and SenFloods11, using scene-based data splits to ensure a realistic assessment of spatial generalization. The results demonstrate that SegFormer-b2 significantly outperforms the U-Net baseline on the ETCI dataset (higher flood IoU across all 7 test scenes in the Wilcoxon signed-rank test), while after fine-tuning on Sen1Floods11, the advantage narrows to within the range of scene variability and is concentrated in spatially fragmented flood events. The study includes both qualitative and quantitative explainability techniques to visually comprehend model decisions and systematically assess prediction reliability. Qualitative analysis reveals that SegFormer-b2 produces more spatially coherent Grad-CAM activations focused on flood-relevant features, while U-Net generates more informative uncertainty estimates along flood boundaries.

Figures

Figures reproduced from arXiv: 2606.16302 by the authors.

Figure 1
Figure 1. Vanilla U-Net architecture with a symmetric encoder-decoder path, skip connections via concatenation, and a [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Vanilla ViT encoder paired with a decoder for semantic segmentation. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Paired per-scene (a) and per-chip (b) flood IoU comparison between U-Net and SegFormer-b2; panel (b) [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Visualization of segmentation performance of the models on the Sen1Floods11 test data scene 1, (a) VV [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Visualization of segmentation performance of the models on the Sen1Floods11 test data scene 2, (a) VV [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Reliability diagrams for U-Net (left) and SegFormer-b2 (right) on the Sen1Floods11 test set. Both models are [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Deletion and insertion faithfulness curves comparing U-Net and SegFormer-b2 Grad-CAM explanations [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

65 extracted references · 3 linked inside Pith

  1. [1]

    (2026) Floods

    World Meteorological Organization. (2026) Floods. [Online]. Available: https://wmo.int/topics/floods

    2026

  2. [2]

    Flood detection with SAR: A review of techniques and datasets,

    D. Amitranoet al., “Flood detection with SAR: A review of techniques and datasets,”Remote Sensing, 2024

    2024

  3. [3]

    Mapping global floods with 10 years of satellite radar data,

    A. Misra, K. White, S. Nsutezo, W. Straka III, and J. Lavista, “Mapping global floods with 10 years of satellite radar data,”Nature Communications, 2025

    2025

  4. [4]

    Sen1floods11: A georeferenced dataset to train and test deep learning flood algorithms for Sentinel-1,

    D. Bonafilia, B. Tellman, T. Anderson, and E. Issenberg, “Sen1floods11: A georeferenced dataset to train and test deep learning flood algorithms for Sentinel-1,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020

    2020

  5. [5]

    Sentinel-1-based flood mapping: a fully automated processing chain,

    A. Twele, W. Cao, S. Plank, and S. Martinis, “Sentinel-1-based flood mapping: a fully automated processing chain,”International Journal of Remote Sensing, 2016

    2016

  6. [6]

    A local thresholding approach to flood water delineation using Sentinel-1 SAR imagery,

    J. Liang and D. Liu, “A local thresholding approach to flood water delineation using Sentinel-1 SAR imagery,” ISPRS Journal of Photogrammetry and Remote Sensing, 2020

    2020

  7. [7]

    A tool for pre-operational daily mapping of floods and permanent water using Sentinel-1 data,

    L. Pulvirentiet al., “A tool for pre-operational daily mapping of floods and permanent water using Sentinel-1 data,” Remote Sensing, 2021

    2021

  8. [8]

    Y . Bai, W. Wu, Z. Yang, J. Yu, B. Zhao, X. Liu, others, and S. Koshimura, “Enhancement of detecting permanent water and temporary water in flood disasters by fusing Sentinel-1 and Sentinel-2 imagery using deep learning algorithms: Demonstration of Sen1Floods11 benchmark datasets,”Remote Sensing, 2021. 15 APREPRINT- JULY7, 2026

    2021

Show all 65 references
  1. [9]

    Deep learning methods for flood mapping: A review of existing applications and future research directions,

    R. Bentivoglioet al., “Deep learning methods for flood mapping: A review of existing applications and future research directions,”Hydrology and Earth System Sciences, 2022

    2022

  2. [10]

    Automatic flood detection from Sentinel-1 data using a nested UNet model and a NASA benchmark dataset,

    B. Ghosh, S. Garg, M. Motagh, and S. Martinis, “Automatic flood detection from Sentinel-1 data using a nested UNet model and a NASA benchmark dataset,”PFG–Journal of Photogrammetry, Remote Sensing and Geoinformation Science, 2024

    2024

  3. [11]

    Understanding flood detection models across Sentinel-1 and Sentinel-2 modalities and benchmark datasets,

    E. Portales-Julia, G. Mateo-García, and L. Gómez-Chova, “Understanding flood detection models across Sentinel-1 and Sentinel-2 modalities and benchmark datasets,”Remote Sensing of Environment, 2025

    2025

  4. [12]

    S1S2-Water: A global dataset for semantic segmentation of water bodies from Sentinel-1 and Sentinel-2 satellite images,

    M. Wieland, F. Fichtner, S. Martinis, S. Groth, C. Krullikowski, S. Plank, and M. Motagh, “S1S2-Water: A global dataset for semantic segmentation of water bodies from Sentinel-1 and Sentinel-2 satellite images,”IEEE Journal of Selected Topics in Applied Earth Observations and ...

    2023

  5. [13]

    A CNN-based flood mapping approach using Sentinel-1 data,

    B. Tavus, R. Can, and S. Kocaman, “A CNN-based flood mapping approach using Sentinel-1 data,” inISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2022

    2022

  6. [14]

    U-Net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional networks for biomedical image segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention, 2015

    2015

  7. [15]

    UNet++: A nested U-Net architecture for medical image segmentation,

    Z. Zhou, M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang, “UNet++: A nested U-Net architecture for medical image segmentation,” inInternational Workshop on Deep Learning in Medical Image Analysis, 2018

    2018

  8. [16]

    Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs,

    L. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. Yuille, “Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017

    2017

  9. [17]

    Convolutional neural network-based deep learning approach for automatic flood mapping using NovaSAR-1 and Sentinel-1 data,

    O. Andrew, A. Apan, D. Paudyal, and K. Perera, “Convolutional neural network-based deep learning approach for automatic flood mapping using NovaSAR-1 and Sentinel-1 data,”ISPRS International Journal of Geo-Information, 2023

    2023

  10. [18]

    DAM-Net: Flood detection from SAR imagery using differential attention metric-based vision transformers,

    T. Saleh, X. Weng, S. Holail, C. Hao, and G. Xia, “DAM-Net: Flood detection from SAR imagery using differential attention metric-based vision transformers,”ISPRS Journal of Photogrammetry and Remote Sensing, 2024

    2024

  11. [19]

    Focal attention for long-range interactions in vision transformers,

    J. Yang, C. Li, P. Zhang, X. Dai, B. Xiao, L. Yuan, and J. Gao, “Focal attention for long-range interactions in vision transformers,”Advances in Neural Information Processing Systems, 2021

    2021

  12. [20]

    Self-attention and transformers: Driving the evolution of large language models,

    Q. Luo, W. Zeng, M. Chen, G. Peng, X. Yuan, and Q. Yin, “Self-attention and transformers: Driving the evolution of large language models,” in2023 IEEE 6th International Conference on Electronic Information and Communication Technology (ICEICT), 2023

    2023

  13. [21]

    Trans- formers for vision: A survey on innovative methods for computer vision,

    V . Hassija, B. Palanisamy, A. Chatterjee, A. Mandal, D. Chakraborty, A. Pandey, others, and D. Kumar, “Trans- formers for vision: A survey on innovative methods for computer vision,”IEEE Access, 2025

    2025

  14. [22]

    Transformer architectures for computer vision: A comprehensive review and future research directions,

    T. Ugile and N. Uke, “Transformer architectures for computer vision: A comprehensive review and future research directions,” inEPJ Web of Conferences, 2025

    2025

  15. [23]

    The robustness gap: Real-world challenges in AI-based flood severity assessment,

    S. Bharti, P. Das, and B. Das, “The robustness gap: Real-world challenges in AI-based flood severity assessment,” Asian Journal of Research in Computer Science, 2026

    2026

  16. [24]

    Adversarial robustness of deep learning models for inland water body segmentation from SAR images,

    S. Kothari, M. Srinivasan, S. Kothari, U. Verma, and J. Sreevalsan-Nair, “Adversarial robustness of deep learning models for inland water body segmentation from SAR images,”IEEE Access, 2026

    2026

  17. [25]

    Effectiveness of Sentinel-1 and Sentinel-2 for flood detection assessment in Europe,

    A. Tarpanelliet al., “Effectiveness of Sentinel-1 and Sentinel-2 for flood detection assessment in Europe,”Natural Hazards and Earth System Sciences, 2022

    2022

  18. [26]

    Introduction to RADAR remote sensing,

    D. Gupta, S. Prashar, S. Singh, P. Srivastava, and R. Prasad, “Introduction to RADAR remote sensing,” inRadar Remote Sensing, 2022

    2022

  19. [27]

    Water-body segmentation for SAR images: past, current, and future,

    Z. Guo, L. Wu, Y . Huang, Z. Guo, J. Zhao, and N. Li, “Water-body segmentation for SAR images: past, current, and future,”Remote Sensing, 2022

    2022

  20. [28]

    Towards operational near real-time flood detection using a split-based automatic thresholding procedure on high resolution TerraSAR-X data,

    S. Martinis, A. Twele, and S. V oigt, “Towards operational near real-time flood detection using a split-based automatic thresholding procedure on high resolution TerraSAR-X data,”Natural Hazards and Earth System Sciences, 2009

    2009

  21. [29]

    Flood segmentation on Sentinel-1 SAR imagery with semi-supervised learning,

    S. Paul and S. Ganju, “Flood segmentation on Sentinel-1 SAR imagery with semi-supervised learning,”arXiv preprint arXiv:2107.08369, 2021

    2021 arXiv

  22. [30]

    Flood change detection model based on an improved U-Net network and multi-head attention mechanism,

    F. Wang and X. Feng, “Flood change detection model based on an improved U-Net network and multi-head attention mechanism,”Scientific Reports, 2025

    2025

  23. [31]

    Flood detection in dual-polarization SAR images based on multi-scale Deeplab model,

    H. Wu, J. Song, J. Huang, H. Zhong, R. Zhan, X. Teng, Z. Qiu, M. He, and J. Cao, “Flood detection in dual-polarization SAR images based on multi-scale Deeplab model,”Remote Sensing, 2022. 16 APREPRINT- JULY7, 2026

    2022

  24. [32]

    U-Net-based semantic classification for flood extent extraction using SAR imagery and GEE platform: A case study for 2019 central US flooding,

    Z. Li and I. Demir, “U-Net-based semantic classification for flood extent extraction using SAR imagery and GEE platform: A case study for 2019 central US flooding,”Science of the Total Environment, 2023

    2019

  25. [33]

    A near-real-time flood detection method based on deep learning and SAR images,

    X. Wu, Z. Zhanget al., “A near-real-time flood detection method based on deep learning and SAR images,” Remote Sensing, 2023

    2023

  26. [34]

    Deep learning-based flood segmentation: evaluating EfficientNet and ResNet for UNet, SegNet and DeepLabV3+,

    ¸ S. Karcı, E. Özceylan Aslan, K. Yıldız, and Ö. Demir, “Deep learning-based flood segmentation: evaluating EfficientNet and ResNet for UNet, SegNet and DeepLabV3+,”Natural Hazards, 2026

    2026

  27. [35]

    Deep-learning integration of CNN–Transformer and U-Net for bi-temporal SAR flash-flood detection,

    A. Noori, A. Ziboon, and A. AL-Hameedawi, “Deep-learning integration of CNN–Transformer and U-Net for bi-temporal SAR flash-flood detection,”Applied Sciences, 2025

    2025

  28. [36]

    Residual wave vision U-Net for flood mapping using dual polarization Sentinel-1 SAR imagery,

    A. Jamali, S. Roy, L. Hashemi Beni, B. Pradhan, J. Li, and P. Ghamisi, “Residual wave vision U-Net for flood mapping using dual polarization Sentinel-1 SAR imagery,”International Journal of Applied Earth Observation and Geoinformation, 2024

    2024

  29. [37]

    Transformer-based semantic segmentation for flood region recognition in SAR images,

    L. Zhou, X. Zhou, H. Feng, W. Liu, and H. Liu, “Transformer-based semantic segmentation for flood region recognition in SAR images,”IEEE Journal on Miniaturization for Air and Space Systems, 2025

    2025

  30. [38]

    DeepSARFlood: Rapid and automated SAR-based flood inundation mapping using vision transformer-based deep ensembles with uncertainty estimates,

    N. Sharma and M. Saharia, “DeepSARFlood: Rapid and automated SAR-based flood inundation mapping using vision transformer-based deep ensembles with uncertainty estimates,”Science of Remote Sensing, 2025

    2025

  31. [39]

    SegFormer: Simple and efficient design for semantic segmentation with transformers,

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. Alvarez, and P. Luo, “SegFormer: Simple and efficient design for semantic segmentation with transformers,”Advances in Neural Information Processing Systems, 2021

    2021

  32. [40]

    Exploring SAR-centered multi-source data fusion for flood mapping with transformer and convolution-based ensemble models,

    S. Zou and Q. Yang, “Exploring SAR-centered multi-source data fusion for flood mapping with transformer and convolution-based ensemble models,” inIGARSS 2024–2024 IEEE International Geoscience and Remote Sensing Symposium, 2024

    2024

  33. [41]

    Sen2GF3Floods: A benchmark multi-source flood dataset with dual-temporal and active learning annotation,

    W. Chen, Y . Zhu, W. Hanet al., “Sen2GF3Floods: A benchmark multi-source flood dataset with dual-temporal and active learning annotation,”Scientific Data, 2026

    2026

  34. [42]

    Flood inundation monitoring with multi-source satellite imagery based on deep learning and explainable frameworks,

    H. Yu, Y . Xu, Y . Huang, and Y . Chiang, “Flood inundation monitoring with multi-source satellite imagery based on deep learning and explainable frameworks,”Journal of Hydrology, 2026

    2026

  35. [43]

    An XGBoost-SHAP framework for interpretable and probabilistic flood susceptibility mapping,

    H. Rezvani, A. Arfa, H. Shafizadeh-Moghadam, and M. Minaei, “An XGBoost-SHAP framework for interpretable and probabilistic flood susceptibility mapping,”Natural Hazards, 2026

    2026

  36. [44]

    Challenges and opportunities in flood mapping and modeling of next-generation geospatial intelligence: a review,

    M. Shirmohammadi, S. Pirasteh, H. Li, M. Akhavan, V . Isazadeh, J. Ji, others, and Y . Muhammad, “Challenges and opportunities in flood mapping and modeling of next-generation geospatial intelligence: a review,”Geomatics, Natural Hazards and Risk, 2026

    2026

  37. [45]

    The transparency revolution in geohazard science: A systematic review and research roadmap for explainable artificial intelligence,

    M. Tosan, V . Nourani, O. Kisi, Y . Zhang, S. Kantoush, M. Gebremichael, others, and J. Huang, “The transparency revolution in geohazard science: A systematic review and research roadmap for explainable artificial intelligence,” Computer Modeling in Engineering & Sciences, 2026

    2026

  38. [46]

    XFIMNet: an explainable deep learning architecture for versatile flood inundation mapping with synthetic aperture radar and multi-spectral optical images,

    J. Sanderson, N. Tengtrairat, W. Woo, H. Mao, and R. Al-Nima, “XFIMNet: an explainable deep learning architecture for versatile flood inundation mapping with synthetic aperture radar and multi-spectral optical images,” International Journal of Remote Sensing, 2023

    2023

  39. [47]

    ConvExNet: Deep learning-based flood detection utilizing Shapley additive explanations,

    K. Bathe and N. Patil, “ConvExNet: Deep learning-based flood detection utilizing Shapley additive explanations,” Journal of Earth System Science, 2025

    2025

  40. [48]

    Individual comparisons by ranking methods,

    F. Wilcoxon, “Individual comparisons by ranking methods,”Biometrics Bulletin, vol. 1, no. 6, pp. 80–83, 1945

    1945

  41. [49]

    Statistical comparisons of classifiers over multiple data sets,

    J. Demšar, “Statistical comparisons of classifiers over multiple data sets,”Journal of Machine Learning Research, vol. 7, pp. 1–30, 2006

    2006

  42. [50]

    A mathematical theory of communication,

    C. Shannon, “A mathematical theory of communication,”The Bell System Technical Journal, 1948

    1948

  43. [51]

    Grad-CAM: Visual explanations from deep networks via gradient-based localization,

    R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-CAM: Visual explanations from deep networks via gradient-based localization,” inProceedings of the IEEE International Conference on Computer Vision, 2017

    2017

  44. [52]

    On calibration of modern neural networks,

    C. Guo, G. Pleiss, Y . Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” inProceedings of the International Conference on Machine Learning (ICML), 2017, pp. 1321–1330

    2017

  45. [53]

    Obtaining well calibrated probabilities using bayesian binning,

    M. P. Naeini, G. Cooper, and M. Hauskrecht, “Obtaining well calibrated probabilities using bayesian binning,” in Proceedings of the AAAI conference on artificial intelligence, vol. 29, no. 1, 2015

    2015

  46. [54]

    RISE: Randomized input sampling for explanation of black-box models,

    V . Petsiuk, A. Das, and K. Saenko, “RISE: Randomized input sampling for explanation of black-box models,” in Proceedings of the British Machine Vision Conference (BMVC), 2018. 17 APREPRINT- JULY7, 2026

    2018

  47. [55]

    Top-down neural attention by excitation backprop,

    J. Zhang, S. A. Bargal, Z. Lin, J. Brandt, X. Shen, and S. Sclaroff, “Top-down neural attention by excitation backprop,”International Journal of Computer Vision, vol. 126, no. 10, pp. 1084–1102, 2018

    2018

  48. [56]

    Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks,

    G. Wang, W. Li, M. Aertsen, J. Deprest, S. Ourselin, and T. Vercauteren, “Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks,”Neurocomputing, 2019

    2019

  49. [57]

    A hierarchical split-based approach for parametric thresholding of SAR images: Flood inundation as a test case,

    M. Chini, R. Hostache, L. Giustarini, and P. Matgen, “A hierarchical split-based approach for parametric thresholding of SAR images: Flood inundation as a test case,”IEEE Transactions on Geoscience and Remote Sensing, 2017

    2017

  50. [58]

    Reliability of uncertainty quantification methods for deep learning auto-segmentation in head and neck organs at risk,

    J. Breitlinget al., “Reliability of uncertainty quantification methods for deep learning auto-segmentation in head and neck organs at risk,”Physics in Medicine & Biology, vol. 70, no. 2, p. 025003, 2025

    2025

  51. [59]

    We care each pixel: Calibrating on medical segmentation model,

    S. Wanget al., “We care each pixel: Calibrating on medical segmentation model,”arXiv preprint arXiv:2503.05107, 2025

    2025 arXiv

  52. [60]

    Improving interpretability of deep active learning for flood inundation mapping through class ambiguity indices using multi-spectral satellite imagery,

    B. I. Nasutionet al., “Improving interpretability of deep active learning for flood inundation mapping through class ambiguity indices using multi-spectral satellite imagery,”Remote Sensing of Environment, vol. 308, p. 114191, 2024

    2024

  53. [61]

    Towards reliable WMH segmentation under domain shift: An application study using maximum entropy regularization to improve uncertainty estimation,

    A. Gonzalez-Jimenezet al., “Towards reliable WMH segmentation under domain shift: An application study using maximum entropy regularization to improve uncertainty estimation,”Computers in Biology and Medicine, 2025

    2025

  54. [62]

    A comprehensive review of explainable artificial intelligence (XAI) in computer vision,

    I. Abdelhalimet al., “A comprehensive review of explainable artificial intelligence (XAI) in computer vision,” Sensors, vol. 25, no. 14, p. 4322, 2025

    2025

  55. [63]

    Contextual attention network: Transformer meets U-Net,

    R. Azad, M. Heidari, Y . Wu, and D. Merhof, “Contextual attention network: Transformer meets U-Net,” in International Workshop on Machine Learning in Medical Imaging (MLMI). Springer, 2022, pp. 377–386

    2022

  56. [64]

    ConvFormer: Combining CNN and transformer for medical image segmentation,

    P. Gu, Y . Zhang, C. Wang, and D. Z. Chen, “ConvFormer: Combining CNN and transformer for medical image segmentation,”arXiv preprint arXiv:2211.08564, 2022

    2022 arXiv

  57. [65]

    Drowning overconfidence with uncertainty: Mitigating deep learning overconfidence in flood depth super-resolution through maximum entropy regularization,

    A. Zeineldinet al., “Drowning overconfidence with uncertainty: Mitigating deep learning overconfidence in flood depth super-resolution through maximum entropy regularization,”Stochastic Environmental Research and Risk Assessment, vol. 39, pp. 895–912, 2025. 18

    2025

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.