Pith. sign in

REVIEW 4 major objections 6 minor 64 references

Large scale cross-regional remote sensing flood monitoring framework for operative mapping and impact analysis

T0 review · 4 major / 6 minor · reviewed 2026-07-31 · grok-4.5

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

Pith's one-line read Under limited labeled data across seven Russian regions, a supervised multimodal U-Net++ outperforms a fine-tuned self-supervised model for flood-water mapping, and the masks yield damage estimates that match official 2019 Tulun figures exc

desk verdict Solid applied Russia flood stack: supervised multimodal U-Net++ beats under-trained AnySat under region hold-out; the Tulun “close match” is mostly inundated area and may not be fully out-of-sample. read the letter →

arxiv 2607.28401 v1 pith:JB6AHMYR submitted 2026-07-30 cs.CV

Ilya Novikov , Svetlana Illarionova , Ruslan Dzharkinov , Maria Smirnova , Ayrat Abdullin , Anna Korotkova , Mariia Ulianova , Dmitrii Shadrin
show 1 more author
Evgeny Burnaev
This is my paper
classification cs.CV
keywords floodmonitoringremotesensingdeeplearningmultimodaldatadamageassessmentwatersegmentationself-supervisedSentinel-1/2
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 builds an end-to-end flood pipeline for Russia that fuses Sentinel-1 radar, Sentinel-2 optical bands and water indices, and DEM slope/aspect into a 21-channel input. With only about 1,259 labeled patches spanning seven climatically different regions, it pits a supervised U-Net++ against a self-supervised AnySat model pre-trained on thousands of unlabeled tiles then fine-tuned. Supervised learning wins on mean accuracy (F1 about 0.84), while the self-supervised path is more stable across held-out regions and remains useful when modalities go missing. The best water masks are morphologically cleaned, buffered, and intersected with population, cadastral buildings, land-cover and species layers exactly as the Russian emergency ministry requires. On the 2019 Tulun flood the resulting area and exposure numbers line up with official reports, showing that multimodal deep learning can turn scarce labels into operational maps and damage tallies at national scale.

What carries the argument

The 21-channel multimodal stack (S1 dual-pol and ratios, S2 ten bands, five water indices, DEM + slope + aspect) fed to a U-Net++ with ResNeSt-101e encoder, scored under region-held-out cross-validation, then post-processed by morphological closing, buffering and intersection with NSPD/OSM/GHS-POP layers following EMERCOM damage categories.

What would settle it

Re-run the Tulun 2019 pipeline with a complete, temporally aligned 2019 cadastral building layer and independent ground-surveyed inundation polygons; if material-damage totals then diverge sharply from the official figure while area and population stay matched, the damage-proxy claim fails.

Watch

Extended reading notes

Core claim

Under the study’s limited labeled multimodal data and strict region-based cross-validation, a supervised U-Net++ that ingests Sentinel-1, Sentinel-2 and DEM-aspect-slope channels reaches mean F1 0.84 ± 0.11 (IoU 0.75 ± 0.15) and outperforms both fine-tuned AnySat and the same architecture trained from scratch. The resulting masks, after closing and buffering, produce inundation-area and exposure estimates for the 2019 Tulun flood that closely match official assessments when processed by the Russian Ministry of Emergency Situations methodology, except material-damage figures limited by incomplete open cadastral sources.

Load-bearing premise

That intersecting a cleaned satellite water mask with today’s building footprints, population grids and static land-cover maps is a faithful stand-in for official damage categories, even though destroyed buildings, informal structures and cadastral-versus-replacement values are missing or mismatched.

Editorial extensions

If this is right

  • Multimodal S1+S2+DEM fusion can deliver cross-regional flood maps at F1 ~0.84 even when labeled data are scarce and regions span snowmelt to typhoon regimes.
  • Supervised U-Net++ remains the practical choice for single-task water segmentation under current labeled-data volumes; SSL gains appear mainly as fold stability and missing-modality robustness.
  • Official-style damage tallies (area, population exposure, ecological overlap) can be produced automatically from open satellite and cadastral layers once a reliable water mask exists.
  • Material-damage estimates stay unreliable until open building inventories capture informal structures and replacement costs rather than cadastral values.
  • The same pipeline can be re-parameterized (buffer size, land-cover classes) for evacuation planning versus agricultural-loss accounting.

Reading between the lines

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

  • Scaling the unlabeled pre-training corpus by an order of magnitude, or adding thermal/LiDAR modalities, could reverse the supervised-versus-SSL ranking once the foundation model is no longer under-trained on the DEM and index branches.
  • The region-held-out protocol and 21-channel stack could serve as a public benchmark for any new multimodal flood foundation model that claims cross-climate generalization.
  • Temporal mismatch between flood-peak imagery and cadastral snapshots is a generic bottleneck; pairing the framework with pre/post building change detection would tighten material-damage estimates without new field surveys.
  • Because the damage module already intersects species-range maps, the same water masks could feed near-real-time biodiversity-impact alerts for protected areas during large floods.
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

4 major / 6 minor

Summary. The paper proposes an end-to-end multimodal flood monitoring and damage-assessment framework for Russian territories, using a 21-channel input (Sentinel-1, Sentinel-2, water indices, DEM + slope/aspect). On a self-collected labeled set of 1,259 patches across seven climatically diverse regions, it compares a supervised U-Net++ (ResNeSt-101e) baseline against AnySat (JEPA-style SSL pre-trained on a Russia-filtered SSL4EO-S12 subset, then fine-tuned) under strict region-based 4-fold CV and three modality configurations. Under these data conditions the supervised multimodal model is stronger (mean IoU 0.75±0.15, F1 0.84±0.11), while SSL improves fold stability. Best water masks are post-processed and intersected with OSM, NSPD, GHS-POP, land-cover, and iNaturalist layers following EMERCOM damage categories; applied to the 2019 Tulun flood, inundated-area estimates are reported to match official figures, with larger discrepancies for buildings and material damage attributed to open cadastral sources.

Significance. If the claims hold, the work is a useful operational contribution: a cross-regional Russian multimodal flood dataset and pipeline, an honest supervised-vs-SSL comparison under limited labels (showing foundation-model advantages are conditional), and an EMERCOM-aligned impact workflow that goes beyond water masks alone. Strengths include region-held-out CV, modality ablations with mean±std, qualitative appendices, and explicit acknowledgment that SSL did not dominate. The damage module’s attempt to link remote-sensing outputs to official methodology is practically relevant for emergency agencies, even if open-data limitations constrain property valuation.

major comments (4)
  1. [§6.2 Damage Assessment Application] §6.2 / Table 1: The Tulun 2019 case study is not shown to be out-of-sample. The labeled corpus includes 66 manually annotated Tulun patches from the same event. The text refers only to “best flood area predictions” / the baseline U-Net++ mask and never states that the mask used for damage assessment came from a model trained without those patches (e.g. Split-3 weights). If Tulun labels were seen, the close match of city inundated area (14.67 km² vs official ~13–16 km²) mainly validates annotation and morphological post-processing, not cross-regional operative generalization. Please state explicitly which fold/weights produced the mask, or re-run with Tulun fully held out, and revise the abstract/conclusion claim accordingly.
  2. [§6.2; Abstract] §6.2 and Abstract: The claim that results “closely matched official assessments, except for material damage” overstates agreement. Only inundated area tracks officials. NSPD yields 162 buildings in the flood extent vs official 2,451 damaged / 83 destroyed; cadastral value ~0.5B RUB vs ~17B official (OSM-based ~7B still far low). Population “~15% exposed” is not mapped to 26 fatalities or >2k evacuated, and the text itself notes evacuation protocols are not modeled. Reframe the end-to-end claim to what is actually validated (area and qualitative exposure maps) and treat property/casualty figures as illustrative given known open-data gaps, not as a near-match except one category.
  3. [§4 Flood Damage Assessment] §4 (Raw mask processing; Flood detection validation): Morphological closing, flood-boundary buffer size, and the permanent-water exceedance threshold that “confirms” a flood are load-bearing free parameters for the damage numbers, yet no numeric values, selection criteria, or sensitivity analysis are reported. Without these, the Tulun area figure and downstream intersections are not reproducible and the operational EMERCOM alignment cannot be audited. Specify parameters and show sensitivity of inundated area and building counts to buffer/threshold choices.
  4. [§5.2.3; §6.1.3 Comparative Analysis] §5.2.3 / §6.1.2–6.1.3: SSL pre-training was stopped at 12 epochs for compute, DAS/index encoders start from scratch, and fine-tuning uses reduced input geometry for memory. The paper correctly notes these limits, but the central comparative claim (“supervised proved more effective… AnySat retains advantages for larger unlabelled data”) then rests on an under-trained SSL pathway. Either complete a longer pre-train / ablate epoch count, or soften language so the result is clearly conditioned on this truncated regime rather than a general verdict on AnySat for flood segmentation.
minor comments (6)
  1. [§5.1.3; Tables 4–9] Metrics are restricted to images with >5% water (§5.1.3). State this limitation in the abstract/results and, if possible, report complementary scores on the full validation sets so low-water scenes are not invisible.
  2. [Table 2] Table 2 lists four CV splits but the text sometimes says “4-fold”; ensure fold definitions and image counts are consistent everywhere, including whether Orsk is nested under Orenburg.
  3. [§2.2] Eqs. (1)–(5) and channel lists: clarify exact S2 band mapping for each index (e.g. which SWIR) and whether VV/VH ratio is one of the three S1 channels in Table 3.
  4. [Appendix A] Figures A1–A4 show Unet++ and AnySat-from-scratch but not the SSL-pretrained AnySat predictions side-by-side; adding one row would make the stability-vs-accuracy discussion easier to inspect.
  5. Minor typos/wording: “mutimodal” in Fig. 1 caption; “W ater” spacing in index names; arXiv date “30 Jul 2026” looks like a placeholder; “casualties” in the abstract is stronger than what the GHS-POP intersection actually estimates (exposure).
  6. [Statements & Declarations] Data availability: “available from the corresponding author on reasonable request” is weak for a dataset-centric claim; consider releasing the 1,259-patch labels or a subset and the exact post-processing scripts for the Tulun mask.

Circularity Check

1 steps flagged · score 1.0 of 10

Empirical ML comparison and external Tulun check; no definitional or fitted-input circularity in the derivation chain.

  1. self citation load bearing [§4 Flood Damage Assessment, Land cover impact quantification; ref. [56]]
    "Utilizing a pre-developed land cover classification algorithm [56], inundated areas are categorized into: open water, settlements, bare soil, forest, grassland, agricultural land, and swamps."

    Land-cover stratification for ecological/agricultural impact cites the authors’ own prior algorithm. This is a minor self-citation in one damage sub-module; it does not define the water mask, the CV metrics, or the Tulun area match, so it is not load-bearing for the central claims.

full rationale

The paper’s load-bearing chain is (i) train/evaluate water segmentation under region-held-out CV on a self-collected labeled set, (ii) pick the best empirical mask (U-Net++ S1+S2+DAS), (iii) morphologically post-process and intersect with independent open layers (OSM hydrography, GHS-POP, NSPD, land-cover, iNaturalist), (iv) compare area/exposure numbers to published EMERCOM/official Tulun figures. None of these steps equates a fitted parameter to its own prediction by construction: IoU/F1 are computed on held-out geographic folds; water indices are fixed formulas; damage quantities are spatial intersections with external databases, not quantities the model was optimized to match. The only self-citation in the pipeline is the authors’ prior land-cover classifier [56] used for one ecological/agricultural sub-module; it does not force the segmentation leaderboard or the inundated-area agreement. Concerns that the Tulun case-study mask may have been trained with Tulun labels, or that only area (not buildings/rubles/casualties) tracks officials, are independence and completeness issues—not circular reductions of outputs to inputs. Score 1 for that minor non-load-bearing self-citation only.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

Load-bearing content is empirical and methodological rather than axiomatic physics. Claims rest on standard CV training assumptions, the adequacy of hand labels and open geospatial layers as ground truth for water and damage, and several hand-chosen preprocessing and postprocessing parameters (normalization, water-fraction filter, morphological close/buffer, flood-vs-permanent-water threshold).

free parameters (5)
  • Focal loss alpha/gamma and class weight 1/15 = γ=2.0, α=1/15
    Supervised training uses Focal Loss γ=2.0 and α=1/15 to handle class imbalance; these are chosen hyperparameters that affect reported F1/IoU.
  • Validation filter: images with >5% water only = >5% water coverage
    Metrics exclude low-water patches; changes the reported means and can hide false positives on dry scenes.
  • Morphological closing and flood-boundary buffer size
    Raw mask post-processing expands inundation to cover partial submersion and timing mismatch; buffer magnitude is task-tunable and directly changes intersected damage area.
  • Permanent-water exceedance threshold for flood confirmation
    Flood is declared when mask area minus OSM permanent water exceeds a predetermined threshold; threshold value is not derived from first principles.
  • SSL pre-training length (12 epochs) and reduced fine-tune geometry = 12 epochs pre-train
    Pre-training stopped early for compute; fine-tuning used memory-constrained patch/resolution settings that the authors say penalize edge accuracy—these choices shape the SSL vs supervised comparison.
assumptions (5)
  • domain assumption Region-held-out patches from seven selected Russian flood events are a valid test of cross-regional generalization for national-scale monitoring.
    Study design in §2.1 and experimental protocol; seven regions span regimes but are still a convenience sample of notable floods, not a random national sample.
  • domain assumption Manual pixel flood labels on Sentinel imagery are sufficiently accurate ground truth for training and metric computation.
    Supervised dataset construction §2.2; label noise and permanent vs flood water ambiguity are not quantified with multi-annotator agreement.
  • domain assumption Open layers (GHS-POP, NSPD, OSM, iNaturalist ranges, authors' land-cover model) plus EMERCOM category list adequately operationalize official damage components.
    §4 Flood Damage Assessment; paper itself shows material-damage failure mode when cadastre diverges from reality.
  • standard math Standard segmentation losses, AdamW, and ImageNet-initialized ResNeSt encoders are appropriate inductive biases for this remote-sensing task.
    Baseline training setup §5.1; conventional CV practice rather than paper-specific postulates.
  • ad hoc to paper Bicubic upsampling of 30 m DEM/derivatives to 10 m and per-channel min-max or z-score normalization preserve task-relevant topography and radiometry.
    Preprocessing §2.2 and §5; resampling can smooth floodplain edges that matter for IoU.
invented entities (2)
  • Self-collected 21-channel seven-region Russian flood segmentation dataset (1,259 labeled patches)
    purpose: Enable supervised training, fine-tuning, and region-based CV for the multimodal comparison.
    New curated resource central to all quantitative claims; not an external public benchmark with independent labels.
  • EMERCOM-aligned automated damage intersection workflow (mask → population/buildings/land cover/ecology) independent evidence
    purpose: Turn water masks into operational impact estimates matching Russian regulatory categories.
    Systems composition of existing open datasets and official methodology rather than a new physical entity; validated only on one case study with known database gaps.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large scale cross-regional remote sensing flood monitoring framework for operative mapping and impact analysis." pith.science (2026). https://pith.science/paper/JB6AHMYR

@misc{pith2026260728401,
  author       = {Pith},
  title        = {Pith review of: Large scale cross-regional remote sensing flood monitoring framework for operative mapping and impact analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JB6AHMYR}},
  note         = {Machine review of arXiv:2607.28401}
}
read the original abstract

Effective flood monitoring is critical for minimizing the impacts of flood disasters on populations and infrastructure. Yet reliable remote sensing across extensive and environmentally diverse regions remains challenging, as most segmentation algorithms lack the generalisation capacity required for large-scale application, while annotated flood data are scarce and unevenly distributed. This study presents an end-to-end multimodal framework for Russian Federation territories sustainable flood monitoring and damage assessment based on synthetic aperture radar data, multispectral imagery, and digital elevation models with their derivatives, forming a 21-channel input. Using a self-collected multimodal dataset covering seven Russian regions, two strategies for water surface detection under limited data conditions were compared: a supervised U-Net++ model and the self-supervised AnySat architecture pre-trained and fine-tuned for the segmentation task. Under the data conditions of this study, supervised learning proved more effective, while the AnySat-based approach offered greater stability and retains advantages for settings where larger unlabelled data or missing modalities at inference are expected. The best flood area predictions were used to estimate flood impact in urban areas in terms of the area affected, material damage, casualties, and ecological and agricultural impact. The estimations were conducted following the official methodology of the Russian Ministry of Emergency Situations. Applied to the 2019 Tulun flood, the obtained results closely matched official assessments, except for material damage, due to the open-source databases usage. The results demonstrate the potential of deep learning and multimodal satellite data integration for scalable, reliable flood monitoring across diverse environmental and data-limited conditions.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 3 canonical work pages

  1. [1]

    IEEE Geoscience and Remote Sensing Magazine11(3), 98–106 (2023)

    Wang, Y., Braham, N.A.A., Xiong, Z., Liu, C., Albrecht, C.M., Zhu, X.X.: Ssl4eo- s12: A large-scale multimodal, multitemporal dataset for self-supervised learning in earth observation [software and data sets]. IEEE Geoscience and Remote Sensing Magazine11(3), 98–106 (2023)

    2023

  2. [2]

    Journal of Environmental Planning and Management57(8), 1252–1265 (2014) 28 Fig

    Brody, S., Blessing, R., Sebastian, A., Bedient, P.: Examining the impact of land use/land cover characteristics on flood losses. Journal of Environmental Planning and Management57(8), 1252–1265 (2014) 28 Fig. A1Qualitative examples of the flood-extent segmentation with the Unet++ and AnySat-based (trained from scratch) models under three input configurat...

    2014

  3. [3]

    Patil, P.S., Jain, S.S.: Survey on flood monitoring & alerting systems. Int. J. Sci. Res. Eng. Manag4(01), 1–3 (2020)

    2020

  4. [4]

    WMO Rep (2021)

    Douris, J., Kim, G.: The atlas of mortality and economic losses from weather, climate and water extremes (1970-2019). WMO Rep (2021)

    1970

  5. [5]

    Global Water Monitor Consortium (2024)

    Van Dijk, A., Beck, H., Boergens, E., Jeu, R., Dorigo, W., Frederikse, T., G¨ untner, A., Haas, J., Hou, J., Preimesberger, W., et al.: Global water monitor 2023, summary report. Global Water Monitor Consortium (2024)

    2023

  6. [6]

    Hydrology Research52(1), 125–141 (2021) 29 Fig

    Tananaev, N., Efremova, V., Gavrilyeva, T., Parfenova, O.: Assessment of the community vulnerability to extreme spring floods: the case of the amga river, central yakutia, siberia. Hydrology Research52(1), 125–141 (2021) 29 Fig. A2Qualitative examples of the flood-extent segmentation with the Unet++ and AnySat-based (trained from scratch) models under thr...

    2021

  7. [7]

    Water sector of Russia: problems, technologies, management (2019) https://doi.org/10.35567/1999-4508-2019-6-4

    Shalikovsky, A.V., Lepikhin, A., Tiunov, A., Kurganovich, K., Morozov, M.: The 2019 floods in irkutsk region. Water sector of Russia: problems, technologies, management (2019) https://doi.org/10.35567/1999-4508-2019-6-4

    work page doi:10.35567/1999-4508-2019-6-4 2019

  8. [8]

    Sustainability15(1), 765 (2022)

    Li, C., Sun, N., Lu, Y., Guo, B., Wang, Y., Sun, X., Yao, Y.: Review on urban flood risk assessment. Sustainability15(1), 765 (2022)

    2022

Show all 64 references
  1. [9]

    Patrick, M., Sandro, M.R., Wolfgang, W., Vahid, F., Peter, Z., Niall, M.C.: Feasibility assessment of an automated, global, satellite-based flood moni- toring product for the copernicus emergency management service. (2020). https://api.semanticscholar.org/CorpusID:229658605

    2020

  2. [10]

    A3Qualitative examples of the flood-extent segmentation with the Unet++ and AnySat-based (trained from scratch) models under three input configurations for validation split #3

    Kamau, G.: A real-time flood monitoring and alert system: Enhancing community 30 Fig. A3Qualitative examples of the flood-extent segmentation with the Unet++ and AnySat-based (trained from scratch) models under three input configurations for validation split #3. Columns from l...

    2024

  3. [11]

    Interna- tional Journal of Disaster Risk Reduction34, 18–27 (2019) https://doi.org/10

    Pangali Sharma, T.P., Zhang, J., Koju, U.A., Zhang, S., Bai, Y., Suwal, M.K.: Review of flood disaster studies in nepal: A remote sensing perspective. Interna- tional Journal of Disaster Risk Reduction34, 18–27 (2019) https://doi.org/10. 1016/j.ijdrr.2018.11.022

    2019

  4. [12]

    International handbook of disaster research, 1–19 (2023)

    Sadiq, R., Imran, M., Ofli, F.: Remote sensing for flood mapping and monitoring. International handbook of disaster research, 1–19 (2023)

    2023

  5. [13]

    A4Qualitative examples of the flood-extent segmentation with the Unet++ and AnySat-based (trained from scratch) models under three input configurations for validation split #4

    Li, M., Zhang, T., Tu, Y., Ren, Z., Xu, B.: Monitoring post-flood recovery of 31 Fig. A4Qualitative examples of the flood-extent segmentation with the Unet++ and AnySat-based (trained from scratch) models under three input configurations for validation split #4. Columns from l...

    2022

  6. [14]

    Natural Hazards 88(3), 1805–1823 (2017)

    Chowdhury, E.H., Hassan, Q.K.: Use of remote sensing data in comprehending an extremely unusual flooding event over southwest bangladesh. Natural Hazards 88(3), 1805–1823 (2017)

    2017

  7. [15]

    Environ- mental Monitoring and Assessment194(6), 402 (2022)

    Pandey, A.C., Bhattacharjee, S., Wasim, M., Salim, M., Ranjan Parida, B.: Extreme rainfall-induced urban flood monitoring and damage assessment in wuhan (china) and kumamoto (japan) cities using google earth engine. Environ- mental Monitoring and Assessment194(6), 402 (2022)

    2022

  8. [16]

    Engineering survey (2018)

    Zelentsov, V., Potryasaev, S.A., Pimanov, I.Y., Ponomarenko, M.R.: The use of 32 synthetic aperture radar (sar) data in the analysis of inundated areas during the spring flood. Engineering survey (2018)

    2018

  9. [17]

    Remote Sensing16(13), 2358 (2024)

    Lee, Y.J., Hwang, J.Y., Park, J., Jung, H.G., Suhr, J.K.: Deep neural network- based flood monitoring system fusing rgb and lwir cameras for embedded iot edge devices. Remote Sensing16(13), 2358 (2024)

    2024

  10. [18]

    International Journal of Remote Sensing27(14), 3025–3033 (2006) https://doi.org/10.1080/01431160600589179 https://doi.org/10.1080/01431160600589179

    Xu, H.: Modification of normalised difference water index (ndwi) to enhance open water features in remotely sensed imagery. International Journal of Remote Sensing27(14), 3025–3033 (2006) https://doi.org/10.1080/01431160600589179 https://doi.org/10.1080/01431160600589179

    2006 doi

  11. [19]

    Hydrology and Earth System Sciences Discussions2022, 1–50 (2022)

    Bentivoglio, R., Isufi, E., Jonkman, S.N., Taormina, R.: Deep learning methods for flood mapping: a review of existing applications and future research directions. Hydrology and Earth System Sciences Discussions2022, 1–50 (2022)

    2022

  12. [20]

    Remote Sensing of Environment (2023)

    Shastry, A., Carter, E., Coltin, B., Sleeter, R., McMichael, S., Eggleston, J.R.: Mapping floods from remote sensing data and quantifying the effects of surface obstruction by clouds and vegetation. Remote Sensing of Environment (2023)

    2023

  13. [21]

    Nature627, 559–563 (2024)

    Nearing, G.S., Cohen, D., Dube, V., Gauch, M., Gilon, O., Harrigan, S., Hassidim, A., Klotz, D., Kratzert, F., Metzger, A., Nevo, S., Pappenberger, F., Prudhomme, C., Shalev, G., Shenzis, S., Tekalign, T.Y., Weitzner, D., Matias, Y.: Global prediction of extreme floods in unga...

    2024

  14. [22]

    https://sites.research.google/gr/floodforecasting/

    Google Research: Flood Forecasting / Flood Hub (AI-driven global flood fore- casting). https://sites.research.google/gr/floodforecasting/. accessed: 2025-10-16 (2024)

    2025

  15. [23]

    Remote Sensing15(18), 4463 (2023)

    Popandopulo, G., Illarionova, S., Shadrin, D., Evteeva, K., Sotiriadi, N., Bur- naev, E.: Flood extent and volume estimation using remote sensing data. Remote Sensing15(18), 4463 (2023)

    2023

  16. [24]

    Amitrano, D., Martino, G.D., Simone, A.D., Imperatore, P.: Flood detection with sar: A review of techniques and datasets. Remote. Sens.16, 656 (2024)

    2024

  17. [25]

    Earth Systems and Environment (2026) https: //doi.org/10.1007/s41748-026-01328-8

    Illarionova, S., Tasuev, U., Tregubova, P., Rubin, I., Shadrin, D., Zaytsev, A., Katalevsky, D., Marusov, A., Bernstein, A.V., Burnaev, E.: From spectral indices to foundation models: A review of the ai revolution in satellite-based envi- ronmental hazard monitoring. Earth Sys...

    2026 doi

  18. [26]

    In: European Conference on Computer Vision (2024).https://api.semanticscholar.org/CorpusID:274023985

    Bourcier, J., Dashyan, G., Karteek, A., Chanussot, J.: Learning representa- tions of satellite images from metadata supervision. In: European Conference on Computer Vision (2024).https://api.semanticscholar.org/CorpusID:274023985

    2024

  19. [27]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    He, K., Chen, X., Xie, S., Li, Y., Doll´ ar, P., Girshick, R.: Masked autoencoders 33 are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 16000–16009 (2022)

    2022

  20. [28]

    Nakayama, Y., Su, J., Pazos-Out’on, L.M.: Satswinmae: Efficient autoencoding for multiscale time-series satellite imagery. (2024). https://api.semanticscholar.org/CorpusID:269605361

    2024

  21. [29]

    2023 IEEE/CVF International Conference on Computer Vision (ICCV), 4065–4076 (2022)

    Reed, C., Gupta, R., Li, S., Brockman, S., Funk, C., Clipp, B., Candido, S., Uyttendaele, M., Darrell, T.: Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 4065–4076 (2022)

    2023

  22. [30]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., LeCun, Y., Ballas, N.: Self-supervised learning from images with a joint-embedding pre- dictive architecture. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1...

    2023

  23. [31]

    ISPRS Journal of Photogrammetry and Remote Sensing (2023)

    Li, W.-J., Yang, W., Liu, T., Hou, Y., Li, Y., Liu, Z., Liu, Y., Liu, L.: Pre- dicting gradient is better: Exploring self-supervised learning for sar atr with a joint-embedding predictive architecture. ISPRS Journal of Photogrammetry and Remote Sensing (2023)

    2023

  24. [32]

    In: European Conference on Computer Vision (2024).https://api.semanticscholar.org/CorpusID:269137486

    Astruc, G., Gonthier, N., Mallet, C., Landrieu, L.: Omnisat: Self-supervised modality fusion for earth observation. In: European Conference on Computer Vision (2024).https://api.semanticscholar.org/CorpusID:269137486

    2024

  25. [33]

    Egyptian Journal of Remote Sensing and Space Science25, 659–671 (2022) https://doi

    Fouad, M.A., Elbohy, A., Mashaly, A., Abosekeen, A., Abdalla, A., Azouz, A.: Spaceborne sar image formation enhancement using moco techniques. Egyptian Journal of Remote Sensing and Space Science25, 659–671 (2022) https://doi. org/10.1016/j.ejrs.2022.06.001

    2022 doi

  26. [34]

    IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium, 6953–6958 (2024)

    Chan-To-Hing, H., Veeravalli, B.: Fus-mae: A cross-attention-based data fusion approach for masked autoencoders in remote sensing. IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium, 6953–6958 (2024)

    2024

  27. [35]

    CoRR (2024)

    Li, X., Hong, D., Li, C., Chanussot, J.: Seamo: A multi-seasonal and multimodal remote sensing foundation model. CoRR (2024)

    2024

  28. [36]

    ArXivabs/2309.05300(2023)

    Wang, Y., Albrecht, C.M., Braham, N.A.A., Liu, C., Xiong, Z., Zhu, X.X.: Decur: decoupling common & unique representations for multimodal self-supervision. ArXivabs/2309.05300(2023)

    2023 arXiv

  29. [37]

    analysis of causes and consequences

    Markin, D., Klokova, C.: Flooding in the irkutsk region. analysis of causes and consequences. Molodezhnyi innovatsionnyi vestnik9(2S), 218–219 (2020)

    2020

  30. [38]

    https://en.wikipedia.org/wiki/Orsk Dam collapse

    Wikipedia contributors: Orsk Dam collapse. https://en.wikipedia.org/wiki/Orsk Dam collapse. Accessed: 2025-12-16 (2024) 34

    2025

  31. [39]

    Palaeogeography, Palaeoclimatology, Palaeoecology 556, 109901 (2020)

    Razjigaeva, N., Grebennikova, T., Ganzey, L., Ponomarev, V., Gorbunov, A., Klimin, M., Arslanov, K., Maksimov, F., Petrov, A.: Recurrence of extreme floods in southern sakhalin island as evidence of paleo-typhoon variability in the north- western pacific since 6.6 ka. Palaeoge...

    2020

  32. [40]

    Scientific reports12(1), 8463 (2022)

    Medvedev, I.P., Rabinovich, A.B., ˇSepi´ c, J.: Destructive coastal sea level oscil- lations generated by typhoon maysak in the sea of japan in september 2020. Scientific reports12(1), 8463 (2022)

    2020

  33. [41]

    International journal of remote sensing17(7), 1425–1432 (1996)

    McFeeters, S.K.: The use of the normalized difference water index (ndwi) in the delineation of open water features. International journal of remote sensing17(7), 1425–1432 (1996)

    1996

  34. [42]

    International journal of remote sensing 27(14), 3025–3033 (2006)

    Xu, H.: Modification of normalised difference water index (ndwi) to enhance open water features in remotely sensed imagery. International journal of remote sensing 27(14), 3025–3033 (2006)

    2006

  35. [43]

    (2004).https://api.semanticscholar.org/CorpusID:18238283

    Bhuiyan, C.: Various drought indices for monitoring drought condition in aravalli terrain of india. (2004).https://api.semanticscholar.org/CorpusID:18238283

    2004

  36. [44]

    Remote Sensing of Environment140, 23–35 (2014)

    Feyisa, G.L., Meilby, H., Fensholt, R., Proud, S.R.: Automated water extraction index: A new technique for surface water mapping using landsat imagery. Remote Sensing of Environment140, 23–35 (2014)

    2014

  37. [45]

    In: International Workshop on Deep Learning in Medical Image Analysis, pp

    Zhou, Z., Rahman Siddiquee, M.M., Tajbakhsh, N., Liang, J.: Unet++: A nested u-net architecture for medical image segmentation. In: International Workshop on Deep Learning in Medical Image Analysis, pp. 3–11 (2018). Springer

    2018

  38. [46]

    https://arxiv

    Astruc, G., Gonthier, N., Mallet, C., Landrieu, L.: AnySat: One Earth Observa- tion Model for Many Resolutions, Scales, and Modalities (2024). https://arxiv. org/abs/2412.14123

    2024 arXiv

  39. [47]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsu- pervised visual representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 9729–9738 (2020)

    2020

  40. [48]

    Registered with the Ministry of Justice of Russia under No

    Ministry of Emergency Situations of Russia: Methodology for Assessing Damage from Emergency Situations. Registered with the Ministry of Justice of Russia under No. 61087 on November 25, 2020 (2020)

    2020

  41. [49]

    Remote Sensing of Environment264(2021)

    Kucharczyk, M., Hugenholtz, C.H.: Remote sensing of natural hazard-related disasters with small drones: Global trends, biases, and research opportunities. Remote Sensing of Environment264(2021)

    2021

  42. [50]

    Agriculture10(4) (2020) https://doi.org/10

    Rahman, M.S., Di, L.: A systematic review on case studies of remote-sensing- based flood crop loss assessment. Agriculture10(4) (2020) https://doi.org/10. 35 3390/agriculture10040131

    2020

  43. [51]

    Remote Sensing Applications: Society and Environment30(2023)

    Billah, M.e.a.: Random forest classifications for landuse mapping to assess rapid flood damage using sentinel-1 and sentinel-2 data. Remote Sensing Applications: Society and Environment30(2023)

    2023

  44. [52]

    Remote Sensing12(4), 714 (2020)

    Sajjad, A.e.a.: Operational monitoring and damage assessment of riverine flood- 2014 in the lower chenab plain, punjab, pakistan, using remote sensing and gis techniques. Remote Sensing12(4), 714 (2020)

    2014

  45. [53]

    https://www

    OpenStreetMap contributors: OpenStreetMap data. https://www. openstreetmap.org. [Accessed: 2025-09-15] (2025)

    2025

  46. [54]

    European Commission, Joint Research Centre (JRC)

    Carioli, A., Schiavina, M., Freire, S., MacManus, K.: GHS-POP R2023A - GHS population grid multitemporal (1975-2030). European Commission, Joint Research Centre (JRC). Accessed September 22, 2025 (2023). https://doi.org/ 10.2905/2FF68A52-5B5B-4A22-8F40-C41DA8332CFE . http://da...

    1975 doi

  47. [55]

    https://nspd.gov.ru

    Spatial Data Portal: National Spatial Data System. https://nspd.gov.ru. [Accessed: 2025-09-15] (2025)

    2025

  48. [56]

    IEEE Access13, 97454–97471 (2025) https://doi.org/10.1109/ACCESS.2025.3575699

    Makarov, A., Mirpulatov, I., Firsov, N., Lobanov, V., Illarionova, S., Podlip- nov, V., Vybornova, Y., Shadrin, D., Rastorguev, A., Skidanov, R., Burnaev, E., Nikonorov, A.: Deep spectral-spatial transformer for robust hyperspectral image segmentation in varying field conditio...

    2025

  49. [57]

    Accessed September 22, 2025 (2025)

    iNaturalist: iNaturalist Open Range Map Dataset. Accessed September 22, 2025 (2025). https://www.inaturalist.org Accessed 2025-09-22

    2025

  50. [58]

    arXiv preprint arXiv:2004.08955 (2020)

    Zhang, H., Wu, C., Zhang, Z., Zhu, Y., Zhang, Z., Lin, H., Sun, Y., He, T., Muller, J., Manmatha, R., Li, M., Smola, A.: Resnest: Split-attention networks. arXiv preprint arXiv:2004.08955 (2020)

    2004 arXiv

  51. [59]

    In: 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., Fei-Fei, L.: Imagenet: A large- scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp. 248–255 (2009). Ieee

    2009

  52. [60]

    status as of july 3, 2019

    TerraTech: Space flood monitoring in tulun. status as of july 3, 2019. Techni- cal report, TerraTech (2019). [Accessed: 2025-09-29]. https://www.roscosmos.ru/ media/img/geoanaliticeskii.otcet.pdf Accessed 2025-09-29

    2019

  53. [61]

    [Accessed: 2025-09-29] (2019)

    IrCity News Agency: The area of the flood zone in Tulun was almost 16 square kilometers. [Accessed: 2025-09-29] (2019). https://ircity.ru/text/society/2019/ 08/13/70654961/ Accessed 2025-09-29 36

    2025

  54. [62]

    [Accessed: 2025-09-29] (2020)

    TASS News Agency: Damage from flood in Irkutsk region estimated at 35 bil- lion rubles. [Accessed: 2025-09-29] (2020). https://tass.ru/proisshestviya/7060618 Accessed 2025-09-29

    2025

  55. [63]

    [Accessed: 2025-09-29] (2019)

    TASS News Agency: Death toll from flood in Irkutsk region rises to 25. [Accessed: 2025-09-29] (2019). https://tass.ru/proisshestviya/6618825 Accessed 2025-09-29

    2025

  56. [64]

    https: //www.irk.ru/news/20240628/water/

    IRK.ru: June 28 marks five years since the devastating flood in Tulun. https: //www.irk.ru/news/20240628/water/. Accessed: 2025-12-16 (2024) 37

    2025

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.