Pith. sign in

REVIEW 3 major objections 5 minor 47 references

Leveraging Self-Supervised Features for Efficient Flooded Region Identification in UAV Aerial Images

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Frozen DINOv2 features, fused into the encoder bottleneck of U-Net and DeepLabV3, improve flood-region segmentation in UAV aerial images; DINOv2+DeepLabV3 gains 6.36 mIoU over DeepLabV3 on FloodNet.

desk verdict The reported DINOv2 gains are confounded by a loss change; the paper's own ablation shows the features don't help when the loss is held fixed. read the letter →

arxiv 2507.04915 v1 pith:S6SXX3YN submitted 2025-07-07 cs.CV

Dibyabha Deb , Ujjwal Verma This is my paper
classification cs.CV
keywords semanticsegmentationflooddetectionUAVaerialimageryself-supervisedlearningDINOv2featuretransferclassimbalanceencoder-decoderfusion
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 tries to show that self-supervised visual features, learned on ordinary natural images without labels, transfer to top-down UAV images of flooded areas. The authors fuse features from the frozen DINOv2 model into the encoder stages of U-Net and DeepLabV3 and evaluate on FloodNet, a post-Hurricane Harvey dataset. They report mean Intersection over Union (mIoU) scores of 47.39 for DINOv2+DeepLabV3, an absolute gain of 6.36 over DeepLabV3 and 7.55 over E-Net, and 54.26 for DINOv2+U-Net, the best overall score. If the claim holds, flood-region mapping can start from a general-purpose feature extractor trained on unlabeled data, reducing the need for expensive pixel-level annotation of aerial images.

What carries the argument

The load-bearing mechanism is DINOv2, a self-supervised vision transformer trained on unlabeled natural images to output patch-level visual features. Its frozen weights produce a feature map that is reshaped and convolved with a $1\times1$ kernel to match the encoder bottleneck, then concatenated with the U-Net or DeepLabV3 encoder features, and the combined tensor is passed to the decoder. The other machinery is the custom training loss: a focal loss reweighted by inverse class frequency per class, added with equal weight to dice loss, which shifts training emphasis toward rare flooded classes.

What would settle it

Retrain vanilla U-Net and DeepLabV3 on the same FloodNet split with the exact combined dice and class-frequency-weighted focal loss used for the proposed models; if they reach or exceed mIoU values of 54.26 and 47.39, the central claim that DINOv2 features drive the improvement is falsified. The paper's own ablation already shows DINOv2+U-Net with dice-only loss scoring 52.07, below vanilla U-Net's 53.47.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that a frozen self-supervised vision transformer pretrained on natural images produces features that improve flooded-region segmentation on aerial images when concatenated with the encoder bottleneck features of standard segmentation networks, without fine-tuning the feature extractor. On FloodNet, DINOv2+DeepLabV3 raises mIoU from 41.03 for DeepLabV3 to 47.39, while DINOv2+U-Net reaches 54.26 mIoU, the best score among the compared models. The authors attribute the transfer to DINOv2's pretraining on natural-image datasets such as ADE20k and Cityscapes, whose classes include buildings, roads, and vehicles that resemble flood-affected infrastructure, and they combine a class-frequency-weighted focal loss with dice loss to counter the dataset's severe class imbalance.

Load-bearing premise

The load-bearing premise is that the reported mIoU gains come from DINOv2 feature fusion, but the baseline models were not trained with the same combined dice and modified focal loss, so the improvement could be due to the loss function instead of the features.

Editorial extensions

If this is right

  • Fusing frozen DINOv2 features into DeepLabV3 raises FloodNet mIoU by 6.36 points over DeepLabV3 and by 7.55 over E-Net, with the largest per-class gains on flooded building and road classes.
  • DINOv2+U-Net reaches 54.26 mIoU, the best score among the compared models, improving over vanilla U-Net by 0.79 and over E-Net by 14.42.
  • DINOv2 features trained on non-aerial images can serve as a frozen feature extractor for aerial segmentation, making fine-tuning of the feature extractor unnecessary.
  • Combining dice loss with class-frequency-weighted focal loss yields higher mIoU than either loss alone for the proposed DINOv2+U-Net model.

Reading between the lines

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

  • A decisive apples-to-apples test would retrain vanilla U-Net and DeepLabV3 under the same dice-plus-weighted-focal loss; the paper's loss ablation shows loss choice alone moves mIoU by several points, so the feature-specific contribution may be smaller than the headline comparison suggests.
  • The same frozen-feature fusion recipe could apply to other disaster-response segmentation tasks, such as building-damage or landslide mapping, where labeled aerial data are scarce.
  • Because DINOv2 was pretrained partly on urban scene datasets, its transfer may be strongest for classes that resemble infrastructure; evaluating on aerial datasets without buildings or roads would reveal how general the feature transfer really is.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes two segmentation models, DINOv2+U-Net and DINOv2+DeepLabV3, which concatenate frozen DINOv2 features with the conventional encoder bottleneck for flooded-region segmentation in UAV imagery. The authors evaluate on FloodNet, report mIoU values of 54.26 for DINOv2+U-Net and 47.39 for DINOv2+DeepLabV3, and claim that DINOv2 self-supervised features trained on natural images transfer to aerial images. They also introduce a class-frequency-weighted focal loss combined with Dice loss, and ablate loss choices in Section 5.3.

Significance. If the central claim is upheld, the paper would provide a practical demonstration that frozen self-supervised DINOv2 features serve as off-the-shelf extractors for aerial flood segmentation, reducing the need for domain-specific pretraining and large annotated UAV datasets. The paper has clear strengths: it uses a fixed public benchmark, describes the feature-fusion architectures explicitly, reports per-class IoU, and includes a loss ablation that partially exposes sensitivity to training choices. However, the current evidence is only partial: the headline comparison does not control for a simultaneous change in loss function, the loss ablation is internally inconsistent for DeepLabV3, and no error bars or multiple runs are reported. These issues are fixable within the manuscript's scope, but they must be addressed before the transferability claim can be accepted.

major comments (3)
  1. [Section 5.1, Table 2 vs. Section 4.3, Eq. (5)] The headline improvements are confounded with a change in the loss function. The proposed models are trained with Dice + modified focal loss (Eq. 5), while the baselines (E-Net, U-Net, DeepLabV3) are reported without stating their training loss or schedule. The 6.36 mIoU gain of DINOv2+DeepLabV3 over DeepLabV3 therefore cannot be attributed to DINOv2 features. The paper's own Table 3 shows the confound: with Dice alone, DINOv2+DeepLabV3 scores 37.46 mIoU, below vanilla DeepLabV3's 41.03, and with Focal alone it scores 53.66. To support the attribution, the authors need to retrain all baselines under identical loss, optimizer, resolution, and epochs, or at least report a same-loss control such as vanilla DeepLabV3 trained with the Dice+modified focal loss.
  2. [Section 5.3, Table 3] The choice of Dice+Modified Focal as the final loss for DINOv2-DeepLabV3 is not supported by the paper's own ablation. Table 3 shows that DINOv2+DeepLabV3 with Focal alone achieves 53.66 mIoU, while Dice+Modified Focal achieves only 47.39, a difference of 6.27 points in favor of Focal alone. The text states that 'a combination of dice and modified focal loss presents a higher mIoU compared to all other loss functions,' but that statement is made only for DINOv2-U-Net. For DeepLabV3, the reported final configuration is worse than the focal-only variant, so the method selection is internally inconsistent and requires either changing the reported final model or providing a justification for preferring the lower-scoring configuration.
  3. [Section 5.1 and Table 2] No error bars or multiple runs are reported for any configuration, which makes the quantitative claims unverifiable. The U-Net gain is only 0.79 mIoU (54.26 vs. 53.47 for DINOv2+U-Net vs. Vanilla U-Net), and this could easily fall within run-to-run variation. Table 3 also shows extreme instability across losses for the same architecture, for example Vehicle IoU ranges from 0.00 to 39.80 for DINOv2+U-Net and several classes drop to 0.00 for DINOv2+DeepLabV3 with Dice loss. Reporting single-run numbers without variance or a multi-seed average undermines the reliability of the improvements claimed in the paper.
minor comments (5)
  1. [Section 6] The conclusion contains an incomplete sentence: 'While the DINOv2-DeepLabV3 model shows a significant improvement of 6.36' ends without completing the comparison or specifying units; the sentence should be finished.
  2. [Abstract] The code availability statement says 'To access the repository, visit the link (GitHub Link),' which is a placeholder rather than a working URL; the repository should be cited with an actual link, DOI, or repository identifier.
  3. [Section 5.1] The claim that DINOv2 was trained on ADE20k and Cityscapes is inaccurate for the DINOv2 model of [10], which is pretrained on LVD-142M; this attribution should be corrected or qualified.
  4. [References] References [7] and [20] are the same paper (Majidizadeh et al., ISPRS Annals 2022) and should be merged or distinguished in the bibliography.
  5. [Abstract and Section 5.1] The abstract claims that the method enables 'high-accuracy segmentation with limited labeled aerial data,' but all experiments use the full 1445-image training set; no experiment varies the amount of labeled data, so this claim is not supported by the reported evaluation.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: feature-transfer claim rests on external FloodNet benchmarks; the loss-confound concern is a soundness gap, not a circular derivation.

full rationale

The paper's central claim is that frozen DINOv2 features improve flood segmentation on FloodNet. The evidence is an empirical comparison against external baselines (U-Net, DeepLabV3, E-Net) on a fixed public benchmark (Table 2), so the claim is not derived from the fitted loss parameters or from a self-citation. The inverse-frequency alpha values in Eqs. (2)-(3) and Table 1 are hyperparameters of the proposed loss, and the mIoU numbers are measurements, not algebraic consequences of those alphas. The only self-citations (refs [39]-[40]) are contextual related-work references about UAV disaster assessment; they are not load-bearing for the feature-transfer conclusion. The soundness issue identified by the skeptical reading - that Table 2 baselines may have been trained with a different loss than the proposed dice+modified focal loss, while Table 3 shows DINOv2+DeepLabV3 with Dice alone reaching 37.46 mIoU versus DeepLabV3's 41.03 - is a missing-control/attribution concern, not circularity. It does not reduce any equation to its inputs or rename a fitted parameter as a prediction. Accordingly no circular step is exhibited, and the score reflects only the minor, non-load-bearing self-citations.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on tuned loss hyperparameters and an untested transferability assumption; no new physical or conceptual entities are introduced.

free parameters (5)
  • Per-class alpha weights for weighted focal loss = Table 1: Background 0.08, Building Flooded 0.14, Building Non-Flooded 0.10, Road Flooded 0.10, Road Non-Flooded 0.05…
    Alpha values weight the focal loss by inverse class frequency, but the reported values cannot be derived from Equation 3 without the per-class beta offsets, so they are effectively fitted to validation performance.
  • Per-class beta offsets = Unreported; selected between -0.20 and 0.20 for each class
    Equation 3 introduces a per-class beta tuned by hand; the paper does not report the chosen values or the validation procedure, making the loss a tuned component.
  • Loss combination weight w = 0.5
    Fixed weight balancing Dice and focal loss in Equation 5; chosen without sensitivity analysis.
  • Focal loss gamma = 4.0
    Set without ablation or justification; standard focal loss uses 2.0.
  • Input resolution = 448x448
    All images resized to 448x448 via bicubic interpolation, which discards the native 3000x4000 resolution and affects segmentation of small objects.
assumptions (4)
  • domain assumption DINOv2 features trained on natural images transfer to aerial imagery without fine-tuning.
    The paper's central premise (Section 3) is that frozen DINOv2 features are useful for UAV images; this is an empirical assumption tested only indirectly through the final mIoU.
  • domain assumption The fixed 60/20/20 split of FloodNet is a representative evaluation of flood segmentation performance.
    Section 4.1 describes a single fixed split with no cross-validation or repeated sampling, so the reported mIoU is a point estimate on one test set.
  • domain assumption The combined Dice and weighted focal loss is appropriate for the class imbalance in FloodNet.
    Section 4.3 motivates the loss but does not compare against weighted cross-entropy or other standard losses on the baseline models.
  • domain assumption Ground truth annotations in FloodNet are correct.
    The evaluation relies entirely on the FloodNet labels (Section 4.1) with no quality assessment.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Leveraging Self-Supervised Features for Efficient Flooded Region Identification in UAV Aerial Images." pith.science (2026). https://pith.science/paper/S6SXX3YN

@misc{pith2026250704915,
  author       = {Pith},
  title        = {Pith review of: Leveraging Self-Supervised Features for Efficient Flooded Region Identification in UAV Aerial Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S6SXX3YN}},
  note         = {Machine review of arXiv:2507.04915}
}
read the original abstract

Identifying regions affected by disasters is a vital step in effectively managing and planning relief and rescue efforts. Unlike the traditional approaches of manually assessing post-disaster damage, analyzing images of Unmanned Aerial Vehicles (UAVs) offers an objective and reliable way to assess the damage. In the past, segmentation techniques have been adopted to identify post-flood damage in UAV aerial images. However, most of these supervised learning approaches rely on manually annotated datasets. Indeed, annotating images is a time-consuming and error-prone task that requires domain expertise. This work focuses on leveraging self-supervised features to accurately identify flooded regions in UAV aerial images. This work proposes two encoder-decoder-based segmentation approaches, which integrate the visual features learned from DINOv2 with the traditional encoder backbone. This study investigates the generalization of self-supervised features for UAV aerial images. Specifically, we evaluate the effectiveness of features from the DINOv2 model, trained on non-aerial images, for segmenting aerial images, noting the distinct perspectives between the two image types. Our results demonstrate that DINOv2's self-supervised pretraining on natural images generates transferable, general-purpose visual features that streamline the development of aerial segmentation workflows. By leveraging these features as a foundation, we significantly reduce reliance on labor-intensive manual annotation processes, enabling high-accuracy segmentation with limited labeled aerial data.

Figures

Figures reproduced from arXiv: 2507.04915 by the authors.

Figure 1
Figure 1. Integrating Self-Supervised Features in DeepLab: The features from DINOv2 module is integrated into the DeepLab encoder and then processed through a decoder module to generate the final segmentation map. hence, cannot yield a valid estimate of the risks of error. Determining bias using this procedure is also opposed by the fact that there is a general lack of explicit description of the sampling methodology in tradi… view at source ↗
Figure 2
Figure 2. Integrating Self-Supervised Features in U-Net: The features from DINOv2 module and U-Net encoder are fused together and then fed to the decoder module to create the final segmentation. The details of the two encoder-decoder architectures, viz., DINOv2-U-Net and DINOv2-DeepLabV3, are discussed below. 3.1 DINOv2-U-Net Feature Extraction The DINOv2 pre-trained model was employed for feature extraction due to its abilit… view at source ↗
Figure 3
Figure 3. Qualitative comparison of the predicted masks across various models. (a) Input image to different models and (b) corresponding ground truth images. (c) Results from DINOv2 with U-Net model, (d) Results from Vanilla U-Net (e), Results from DINOv2 with DeepLabV3 model and (f) Results from DeepLabV3. (a) Relative changes observed between proposed model and Vanilla U-Net/E-Net (b) Relative changes observed between propo… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Relative changes observed in Proposed model using U-Net and DeepLabV3 with E-Net and U-Net/DeepLabV3 The proposed method integrating DINOv2 features in DeepLabV3 achieved an improvement of 7.55 and 6.36 in terms of absolute mIoU performance when compared with E-Net and…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 28 canonical work pages

  1. [1]

    Delforge, D. et al. Em-dat: The emergency events database. Tech. Rep., UCLouvain, USAID and Centre for Research on the Epidemiology of Disasters (CRED) (2023). DOI: 10.21203/rs.3.rs-3807553/v1

    work page doi:10.21203/rs.3.rs-3807553/v1 2023

  2. [2]

    & Srinivasa, G.Flood Disaster Management (National Remote Sensing Centre, 2010)

    Bhanumurthy, V ., Manjusree, P. & Srinivasa, G.Flood Disaster Management (National Remote Sensing Centre, 2010)

    work page 2010

  3. [3]

    The limitations of traditional surveying techniques in a forested environment

    Alexander, D. The limitations of traditional surveying techniques in a forested environment. J. Field Archaeol. 10, 177–186 (1983)

    work page 1983

  4. [4]

    S., Abdu Rahiman, K

    Malakeel, G. S., Abdu Rahiman, K. U. & Vishnudas, S. Flood risk assessment methods—a review. In Thomas, J., Jayalekshmi, B. & Nagarajan, P. (eds.) Current Trends in Civil Engineering, 197–208, DOI: 10.1007/978-981-15-8151-9 19 (Springer Singapore, Singapore, 2021)

    work page doi:10.1007/978-981-15-8151-9

  5. [5]

    & Houet, T

    Emilien, A.-V ., Thomas, C. & Houet, T. Uav and satellite synergies for optical remote sensing applications: a literature review. Sci. Remote. Sens. 3, 100019, DOI: 10.1016/j.srs.2021.100019 (2021)

    arXiv 2021

  6. [6]

    & Dinka, M

    Matsimbe, J., Mdolo, W., Kapachika, C., Musonda, I. & Dinka, M. Comparative utilization of drone technology vs. traditional methods in open pit stockpile volumetric computation: A case of njuli quarry, malawi. Front. Built Environ. 8, DOI: 10.3389/fbuil.2022.1037487 (2022)

    work page arXiv 2022

  7. [8]

    Ozbulak, U. et al. Know your self-supervised learning: A survey on image-based generative and discriminative training. Transactions on Mach. Learn. Res. (2023). Survey Certification

    work page 2023

  8. [9]

    & Tian, Y

    Jing, L. & Tian, Y . Self-supervised visual feature learning with deep neural networks: A survey.IEEE Transactions on Pattern Analysis Mach. Intell. 43, 4037–4058 (2021)

    work page 2021

Show all 47 references
  1. [10]

    Oquab, M. et al. DINOv2: Learning robust visual features without supervision. Transactions on Mach. Learn. Res. (2024)

    2024

  2. [11]

    & Brox, T

    Ronneberger, O., Fischer, P. & Brox, T. U-net: Convolutional networks for biomedical image segmentation. vol. 9351, 234–241, DOI: 10.1007/978-3-319-24574-4 28 (2015)

    2015 doi

  3. [12]

    Yang, R. & Yu, Y . Artificial convolutional neural network in object detection and semantic segmentation for medical imaging analysis. Front. Oncol. 11, 638182, DOI: 10.3389/fonc.2021.638182 (2021)

    2021

  4. [13]

    Li, W. et al. Semantic segmentation-based building footprint extraction using very high-resolution satellite images and multi-source gis data. Remote. Sens. 11, 403, DOI: 10.3390/rs11040403 (2019)

    2019 doi

  5. [14]

    Feng, D. et al. Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intell. Transp. Syst. PP, 1–20, DOI: 10.1109/TITS.2020.2972974 (2020). 11/13

    2020

  6. [15]

    Bentivoglio, R., Isufi, E., Jonkman, S. N. & Taormina, R. Deep learning methods for flood mapping: a review of existing applications and future research directions. Hydrol. Earth Syst. Sci. 26, 4345–4378 (2022)

    2022

  7. [16]

    V ., Dobbs, D

    Rahnemoonfar, M., Murphy, R., Miquel, M. V ., Dobbs, D. & Adams, A. Flooded area detection from uav images based on densely connected recurrent neural networks. In IGARSS 2018 - 2018 IEEE International Geoscience and Remote Sensing Symposium , 1788–1791, DOI: 10.1109/IGARSS.20...

    2018

  8. [17]

    L., Ayele, E

    Edamo, M. L., Ayele, E. G., Yisihak Ukumo, T., Alemayehu Kassaye, A. & Paulos Haile, A. Capability of logistic regression in identifying flood-susceptible areas in a small watershed. H2Open J. 7, 351–374, DOI: 10.2166/h2oj.2024.024 (2024). https: //iwaponline.com/h2open/articl...

    2024 doi

  9. [18]

    & Darrell, T

    Shelhamer, E., Long, J. & Darrell, T. Fully convolutional networks for semantic segmentation. IEEE Transactions on Pattern Analysis Mach. Intell. 39, 640–651, DOI: 10.1109/TPAMI.2016.2572683 (2017)

    2017

  10. [19]

    & Cipolla, R

    Badrinarayanan, V ., Kendall, A. & Cipolla, R. Segnet: A deep convolutional encoder-decoder architecture for image segmentation.IEEE transactions on pattern analysis machine intelligence 39, 2481–2495 (2017)

    2017

  11. [20]

    & Jafari, M

    Majidizadeh, A., Hasani, H. & Jafari, M. Semantic segmentation of uav images based on u-net in urban area. ISPRS Annals Photogramm. Remote. Sens. Spatial Inf. Sci. X-4/W1-2022, 451–457, DOI: 10.5194/isprs-annals-X-4-W1-2022-451-2023 (2023)

    2023 doi

  12. [21]

    & Adam, H

    Chen, L.-C., Zhu, Y ., Papandreou, G., Schroff, F. & Adam, H. Encoder-decoder with atrous separable convolution for semantic image segmentation. arxiv 1–18, DOI: 10.48550/arXiv.1802.02611 (2018)

    2018 doi arXiv

  13. [22]

    & Adam, H

    Chen, L.-C., Papandreou, G., Schroff, F. & Adam, H. Rethinking atrous convolution for semantic image segmentation (2017). 1706.05587

    2017 arXiv

  14. [23]

    A progressive image semantic segmentation method using recurrent neural network

    Yi, L. A progressive image semantic segmentation method using recurrent neural network. In 2021 6th International Conference on Intelligent Computing and Signal Processing (ICSP) , 765–768, DOI: 10.1109/ICSP51882.2021.9408920 (2021)

    2021

  15. [24]

    Dosovitskiy, A. et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (2021)

    2021

  16. [25]

    Touvron, H. et al. Training data-efficient image transformers and; distillation through attention. In Meila, M. & Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning , vol. 139 of Proceedings of Machine Learning Research , 10347–10357 (PMLR, 2021)

    2021

  17. [26]

    Zheng, S. et al. Rethinking Semantic Segmentation from a Sequence-to-Sequence Perspective with Transformers . In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 6877–6886, DOI: 10.1109/CVPR46437.2021.00681 (IEEE Computer Society, Los Alamitos, CA, ...

    2021

  18. [27]

    & Schmid, C

    Strudel, R., Garcia, R., Laptev, I. & Schmid, C. Segmenter: Transformer for Semantic Segmentation . In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 7242–7252, DOI: 10.1109/ICCV48922.2021.00717 (IEEE Computer Society, Los Alamitos, CA, USA, 2021)

    2021

  19. [28]

    Fully self-supervised learning for semantic segmentation

    Wang, Y .et al. Fully self-supervised learning for semantic segmentation. arXiv preprint arXiv:2202.11981 (2022)

    2022 arXiv

  20. [29]

    & Wang, Q

    Cheng, J., Deng, C., Su, Y ., An, Z. & Wang, Q. Methods and datasets on semantic segmentation for unmanned aerial vehicle remote sensing images: A review. ISPRS J. Photogramm. Remote. Sens. 211, 1–34, DOI: https://doi.org/10.1016/j.isprsjprs.2024.03.012 (2024)

    2024 doi

  21. [30]

    Wang, L. et al. Unetformer: A unet-like transformer for efficient semantic segmentation of remote sensing urban scene imagery. ISPRS J. Photogramm. Remote. Sens. 190, 196–214, DOI: 10.1016/j.isprsjprs.2022.06.008 (2022)

    2022 doi

  22. [31]

    Wei, Y ., Wang, Z. & Xu, M. Road structure refined cnn for road extraction in aerial image. IEEE Geosci. Remote. Sens. Lett. 14, 709–713, DOI: 10.1109/LGRS.2017.2672734 (2017)

    2017

  23. [32]

    Marmanis, D. et al. Classification with an edge: Improving semantic image segmentation with boundary detection.ISPRS J. Photogramm. Remote. Sens. 135, DOI: 10.1016/j.isprsjprs.2017.11.009 (2016)

    2016 doi

  24. [33]

    Li, R. et al. Abcnet: Attentive bilateral contextual network for efficient semantic segmentation of fine-resolution remotely sensed imagery. ISPRS J. Photogramm. Remote. Sens. 181, 84–98, DOI: https://doi.org/10.1016/j.isprsjprs.2021.09.005 (2021)

    2021 doi

  25. [34]

    Yang, J. et al. Dinov2-based uav visual self-localization in low-altitude urban environments. IEEE Robotics Autom. Lett. 10, 2080–2087, DOI: 10.1109/LRA.2025.3527762 (2025)

    2025

  26. [35]

    & Chen, G

    Shi, L., Yang, K., Chen, Y . & Chen, G. An interactive prompt based network for urban floods area segmentation using uav images.IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens. 18, 935–948, DOI: 10.1109/JSTARS.2024.3498865 (2025). 12/13

    2025

  27. [36]

    A review of flood damage analysis for a building structure and contents

    Tavanaie Marvi, M. A review of flood damage analysis for a building structure and contents. Nat. Hazards 102, DOI: 10.1007/ s11069-020-03941-w (2020)

    2020

  28. [37]

    & Fukushi, K

    Kefi, M., Mishra, B., Masago, Y . & Fukushi, K. Analysis of flood damage and influencing factors in urban catchments: case studies in manila, philippines, and jakarta, indonesia. Nat. Hazards 104, DOI: 10.1007/s11069-020-04281-5 (2020)

    2020 doi

  29. [38]

    & Pang, G

    Doshi, J., Basu, S. & Pang, G. From satellite imagery to disaster insights. arXiv preprint arXiv:1812.07033 (2018)

    2018 arXiv

  30. [39]

    Recent trends and challenges in analysis of uav aerial images for post-disaster scene understanding

    Verma, U. Recent trends and challenges in analysis of uav aerial images for post-disaster scene understanding. In IGARSS 2022 - 2022 IEEE International Geoscience and Remote Sensing Symposium , 4647–4649, DOI: 10.1109/IGARSS46834.2022.9883265 (2022)

    2022

  31. [40]

    & Verma, U

    Sharma, A. & Verma, U. Flood magnitude assessment from uav aerial videos based on image segmentation and similarity. In TENCON 2021 - 2021 IEEE Region 10 Conference (TENCON) , 476–481, DOI: 10.1109/TENCON54134.2021.9707250 (2021)

    2021

  32. [41]

    & Hauptmann, A

    Zhu, X., Liang, J. & Hauptmann, A. Msnet: A multilevel instance segmentation network for natural disaster damage assessment in aerial videos. 2022–2031, DOI: 10.1109/W ACV48630.2021.00207 (2021)

    2021

  33. [42]

    Wu, C. et al. Building damage detection using u-net with attention mechanism from pre- and post-disaster remote sensing datasets. Remote. Sens. 13, 905, DOI: 10.3390/rs13050905 (2021)

    2021 doi

  34. [43]

    Rahnemoonfar, M. et al. Floodnet: A high resolution aerial imagery dataset for post flood scene understanding. IEEE Access 9, 89644–89654, DOI: 10.1109/ACCESS.2021.3090981 (2021)

    2021

  35. [44]

    & Doll´ar, P

    Lin, T.-Y ., Goyal, P., Girshick, R., He, K. & Doll´ar, P. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, 2980–2988 (2017)

    2017

  36. [45]

    H., Li, W., Vercauteren, T., Ourselin, S

    Sudre, C. H., Li, W., Vercauteren, T., Ourselin, S. & Jorge Cardoso, M. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third Intern...

    2017

  37. [46]

    & Culurciello, E

    Paszke, A., Chaurasia, A., Kim, S. & Culurciello, E. Enet: A deep neural network architecture for real-time semantic segmentation. arXiv preprint arXiv:1606.02147 (2016)

    2016 arXiv

  38. [47]

    Zhou, B. et al. Scene parsing through ade20k dataset. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 5122–5130, DOI: 10.1109/CVPR.2017.544 (2017)

    2017 doi

  39. [48]

    Cordts, M. et al. The cityscapes dataset for semantic urban scene understanding. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 3213–3223, DOI: 10.1109/CVPR.2016.350 (2016). 13/13

    2016 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.